diff --git a/.gitattributes b/.gitattributes index bed0738c7eeb449bca98b5d2f33c89a1ee56349a..5cc799d5b0780c0f83c1a1e5fe09b985964c3341 100644 --- a/.gitattributes +++ b/.gitattributes @@ -1,60 +1,5022 @@ *.7z filter=lfs diff=lfs merge=lfs -text *.arrow filter=lfs diff=lfs merge=lfs -text -*.avro filter=lfs diff=lfs merge=lfs -text *.bin filter=lfs diff=lfs merge=lfs -text +*.bin.* filter=lfs diff=lfs merge=lfs -text *.bz2 filter=lfs diff=lfs merge=lfs -text -*.ckpt filter=lfs diff=lfs merge=lfs -text *.ftz filter=lfs diff=lfs merge=lfs -text *.gz filter=lfs diff=lfs merge=lfs -text *.h5 filter=lfs diff=lfs merge=lfs -text *.joblib filter=lfs diff=lfs merge=lfs -text *.lfs.* filter=lfs diff=lfs merge=lfs -text -*.lz4 filter=lfs diff=lfs merge=lfs -text -*.mds filter=lfs diff=lfs merge=lfs -text -*.mlmodel filter=lfs diff=lfs merge=lfs -text *.model filter=lfs diff=lfs merge=lfs -text *.msgpack filter=lfs diff=lfs merge=lfs -text -*.npy filter=lfs diff=lfs merge=lfs -text -*.npz filter=lfs diff=lfs merge=lfs -text *.onnx filter=lfs diff=lfs merge=lfs -text -*.ot filter=lfs diff=lfs merge=lfs -text + *.parquet filter=lfs diff=lfs merge=lfs -text *.pb filter=lfs diff=lfs merge=lfs -text -*.pickle filter=lfs diff=lfs merge=lfs -text -*.pkl filter=lfs diff=lfs merge=lfs -text *.pt filter=lfs diff=lfs merge=lfs -text *.pth filter=lfs diff=lfs merge=lfs -text *.rar filter=lfs diff=lfs merge=lfs -text -*.safetensors filter=lfs diff=lfs merge=lfs -text saved_model/**/* filter=lfs diff=lfs merge=lfs -text -*.tar.* filter=lfs diff=lfs merge=lfs -text *.tar filter=lfs diff=lfs merge=lfs -text +*.tar.* filter=lfs diff=lfs merge=lfs -text +*.mat filter=lfs diff=lfs merge=lfs -text +*.npz filter=lfs diff=lfs merge=lfs -text +*.npy filter=lfs diff=lfs merge=lfs -text +*.h5 filter=lfs diff=lfs merge=lfs -text +*.hdf5 filter=lfs diff=lfs merge=lfs -text +*.pickle filter=lfs diff=lfs merge=lfs -text +*.pkl filter=lfs diff=lfs merge=lfs -text *.tflite filter=lfs diff=lfs merge=lfs -text *.tgz filter=lfs diff=lfs merge=lfs -text -*.wasm filter=lfs diff=lfs merge=lfs -text *.xz filter=lfs diff=lfs merge=lfs -text *.zip filter=lfs diff=lfs merge=lfs -text -*.zst filter=lfs diff=lfs merge=lfs -text -*tfevents* filter=lfs diff=lfs merge=lfs -text -# Audio files - uncompressed -*.pcm filter=lfs diff=lfs merge=lfs -text -*.sam filter=lfs diff=lfs merge=lfs -text -*.raw filter=lfs diff=lfs merge=lfs -text -# Audio files - compressed -*.aac filter=lfs diff=lfs merge=lfs -text -*.flac filter=lfs diff=lfs merge=lfs -text -*.mp3 filter=lfs diff=lfs merge=lfs -text -*.ogg filter=lfs diff=lfs merge=lfs -text -*.wav filter=lfs diff=lfs merge=lfs -text -# Image files - uncompressed -*.bmp filter=lfs diff=lfs merge=lfs -text -*.gif filter=lfs diff=lfs merge=lfs -text -*.png filter=lfs diff=lfs merge=lfs -text -*.tiff filter=lfs diff=lfs merge=lfs -text -# Image files - compressed +*.zstandard filter=lfs diff=lfs merge=lfs -text +*.tfevents* filter=lfs diff=lfs merge=lfs -text +*.db* filter=lfs diff=lfs merge=lfs -text +*.ark* filter=lfs diff=lfs merge=lfs -text +**/*ckpt*data* filter=lfs diff=lfs merge=lfs -text +**/*ckpt*.meta filter=lfs diff=lfs merge=lfs -text +**/*ckpt*.index filter=lfs diff=lfs merge=lfs -text +*.safetensors filter=lfs diff=lfs merge=lfs -text +*.ckpt filter=lfs diff=lfs merge=lfs -text *.jpg filter=lfs diff=lfs merge=lfs -text +*.png filter=lfs diff=lfs merge=lfs -text *.jpeg filter=lfs diff=lfs merge=lfs -text +*.bmp filter=lfs diff=lfs merge=lfs -text +*.gif filter=lfs diff=lfs merge=lfs -text *.webp filter=lfs diff=lfs merge=lfs -text -# Video files - compressed +*.mp3 filter=lfs diff=lfs merge=lfs -text +*.wav filter=lfs diff=lfs merge=lfs -text +*.wma filter=lfs diff=lfs merge=lfs -text +*.aac filter=lfs diff=lfs merge=lfs -text +*.ogg filter=lfs diff=lfs merge=lfs -text +*.m4a filter=lfs diff=lfs merge=lfs -text +*.m3u8 filter=lfs diff=lfs merge=lfs -text +*.amr filter=lfs diff=lfs merge=lfs -text +*.audio filter=lfs diff=lfs merge=lfs -text +*.avi filter=lfs diff=lfs merge=lfs -text +*.flv filter=lfs diff=lfs merge=lfs -text *.mp4 filter=lfs diff=lfs merge=lfs -text -*.webm filter=lfs diff=lfs merge=lfs -text +*.mpg filter=lfs diff=lfs merge=lfs -text +*.asf filter=lfs diff=lfs merge=lfs -text +*.mov filter=lfs diff=lfs merge=lfs -text +*.mpeg filter=lfs diff=lfs merge=lfs -text +*.3gp filter=lfs diff=lfs merge=lfs -text +*.wmv filter=lfs diff=lfs merge=lfs -text +*.rmvb filter=lfs diff=lfs merge=lfs -text +*.rm filter=lfs diff=lfs merge=lfs -text +*.ts filter=lfs diff=lfs merge=lfs -text +*.mkv filter=lfs diff=lfs merge=lfs -text +*.flash filter=lfs diff=lfs merge=lfs -text +*.vob filter=lfs diff=lfs merge=lfs -text +*.pdf filter=lfs diff=lfs merge=lfs -text +*.ost filter=lfs diff=lfs merge=lfs -text +*.pst filter=lfs diff=lfs merge=lfs -text +*.doc filter=lfs diff=lfs merge=lfs -text +*.docx filter=lfs diff=lfs merge=lfs -text +*.txt filter=lfs diff=lfs merge=lfs -text +*.ppt filter=lfs diff=lfs merge=lfs -text +*.pptx filter=lfs diff=lfs merge=lfs -text +*.xls filter=lfs diff=lfs merge=lfs -text +*.xlsx filter=lfs diff=lfs merge=lfs -text +*.vsd filter=lfs diff=lfs merge=lfs -text +*.vsdx filter=lfs diff=lfs merge=lfs -text +*.jsonl filter=lfs diff=lfs merge=lfs -text +*.json filter=lfs diff=lfs merge=lfs -text +dataset_infos.json ignore +*.csv filter=lfs diff=lfs merge=lfs -text +*.tsv filter=lfs diff=lfs merge=lfs -text + +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst filter=lfs diff=lfs merge=lfs -text + +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst filter=lfs diff=lfs merge=lfs -text + +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst filter=lfs diff=lfs merge=lfs -text +lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst filter=lfs diff=lfs merge=lfs -text + +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst filter=lfs diff=lfs merge=lfs -text +lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst filter=lfs diff=lfs merge=lfs -text + +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst filter=lfs diff=lfs merge=lfs -text + +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst filter=lfs diff=lfs merge=lfs -text + +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst filter=lfs diff=lfs merge=lfs -text + +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst filter=lfs diff=lfs merge=lfs -text + +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst filter=lfs diff=lfs merge=lfs -text + +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst filter=lfs diff=lfs merge=lfs -text +lambda0.02/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst filter=lfs diff=lfs merge=lfs -text \ No newline at end of file diff --git a/README.md b/README.md new file mode 100644 index 0000000000000000000000000000000000000000..1c931342748f396aff46f7910bd373612f360a3f --- /dev/null +++ b/README.md @@ -0,0 +1,9 @@ +--- +license: mit +tags: + - large-model-feature-coding +language: + - en +models: + - moonshotai/Kimi-Audio-7B-Instruct +--- diff --git a/lambda0.001/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.001/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..e9bf6ae871e99d1dd596b756720a0888286c7607 --- /dev/null +++ b/lambda0.001/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 286 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- -------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench +Output output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench +---------------- -------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,316B, BPFP=0.2539 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,756B, BPFP=0.5316 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,176B, BPFP=0.6127 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,168B, BPFP=0.4182 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,712B, BPFP=0.5231 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,184B, BPFP=0.4213 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,200B, BPFP=0.4244 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,572B, BPFP=0.4961 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,044B, BPFP=0.5872 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,856B, BPFP=0.5509 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,368B, BPFP=0.0377 +⌛️ [2/4] FRONTEND: Frontend time: 2.185s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.227s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12089630 342.71248071 + layer.0.v_cache 0.00001394 0.03912840 + layer.1.k_cache 0.01174029 26.04195903 + layer.1.v_cache 0.00000579 0.01541970 + layer.2.k_cache 0.00835781 4.11488210 + layer.2.v_cache 0.00001852 0.04635139 + layer.3.k_cache 0.02896961 16.83106976 + layer.3.v_cache 0.00001872 0.05226893 + layer.4.k_cache 0.00063445 1.22381752 + layer.4.v_cache 0.00005063 0.10015474 + layer.4.output 0.17246709 670.31398810 + ------------------------------------------------------------------------------------- + TOTAL 0.08105739 299.02208523 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 26352 +BPFP 0.2990 bits/point +EBPFP 0.5980 equivalent bits/point +MSE 299.022085 +---------------------- -------------------------------------------------------- +Time: 3.416s Load: 0.004s, Pack+Encode: 2.185s, Decode+Unpack: 1.227s +---------------------- -------------------------------------------------------- +💾 Converting with 299.0221 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,248B, BPFP=0.2437 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,884B, BPFP=0.5633 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,952B, BPFP=0.5766 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,412B, BPFP=0.4711 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,508B, BPFP=0.4898 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,168B, BPFP=0.4234 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,012B, BPFP=0.3930 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,660B, BPFP=0.5195 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,956B, BPFP=0.5773 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,968B, BPFP=0.5797 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0391 +⌛️ [2/4] FRONTEND: Frontend time: 1.686s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.170s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08029544 321.96032715 + layer.0.v_cache 0.00001376 0.04022255 + layer.1.k_cache 0.03632395 25.30027466 + layer.1.v_cache 0.00000563 0.01646130 + layer.2.k_cache 0.00522719 4.17606659 + layer.2.v_cache 0.00001997 0.05015263 + layer.3.k_cache 0.02707962 16.59712067 + layer.3.v_cache 0.00001843 0.05449881 + layer.4.k_cache 0.00063057 1.19642916 + layer.4.v_cache 0.00005085 0.10492008 + layer.4.output 0.18451144 678.66478795 + ------------------------------------------------------------------------------------- + TOTAL 0.08477915 301.18529348 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 26168 +BPFP 0.3006 bits/point +EBPFP 0.6013 equivalent bits/point +MSE 301.185293 +---------------------- -------------------------------------------------------- +Time: 2.860s Load: 0.004s, Pack+Encode: 1.686s, Decode+Unpack: 1.170s +---------------------- -------------------------------------------------------- +💾 Converting with 301.1853 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,296B, BPFP=0.2355 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,904B, BPFP=0.5276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,780B, BPFP=0.5051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,208B, BPFP=0.4012 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,332B, BPFP=0.4237 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,536B, BPFP=0.4608 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,168B, BPFP=0.3939 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,400B, BPFP=0.4360 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,884B, BPFP=0.5240 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,808B, BPFP=0.5102 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,516B, BPFP=0.0393 +⌛️ [2/4] FRONTEND: Frontend time: 1.651s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.164s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10874708 336.31386265 + layer.0.v_cache 0.00001608 0.04131184 + layer.1.k_cache 0.03270756 24.72106366 + layer.1.v_cache 0.00000572 0.01610014 + layer.2.k_cache 0.00777350 3.89486623 + layer.2.v_cache 0.00002080 0.04627290 + layer.3.k_cache 0.05920834 15.70795387 + layer.3.v_cache 0.00001846 0.05246993 + layer.4.k_cache 0.00062493 1.21008017 + layer.4.v_cache 0.00005036 0.09511404 + layer.4.output 0.17137358 631.29879568 + ------------------------------------------------------------------------------------- + TOTAL 0.08286988 282.42298031 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 25832 +BPFP 0.2761 bits/point +EBPFP 0.5522 equivalent bits/point +MSE 282.422980 +---------------------- -------------------------------------------------------- +Time: 2.819s Load: 0.004s, Pack+Encode: 1.651s, Decode+Unpack: 1.164s +---------------------- -------------------------------------------------------- +💾 Converting with 282.4230 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,240B, BPFP=0.2484 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,756B, BPFP=0.5521 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,984B, BPFP=0.5978 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,984B, BPFP=0.3974 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,432B, BPFP=0.4872 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,932B, BPFP=0.3870 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,024B, BPFP=0.4054 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,508B, BPFP=0.5024 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,828B, BPFP=0.5665 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,920B, BPFP=0.5849 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,432B, BPFP=0.0410 +⌛️ [2/4] FRONTEND: Frontend time: 1.687s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.241s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12073612 338.89863782 + layer.0.v_cache 0.00001348 0.03896025 + layer.1.k_cache 0.03279715 25.32394800 + layer.1.v_cache 0.00000568 0.01666466 + layer.2.k_cache 0.00696802 4.00323212 + layer.2.v_cache 0.00001833 0.04818709 + layer.3.k_cache 0.03227850 16.63837139 + layer.3.v_cache 0.00001834 0.05162611 + layer.4.k_cache 0.00060115 1.19579374 + layer.4.v_cache 0.00005191 0.09836342 + layer.4.output 0.18320683 696.14514652 + ------------------------------------------------------------------------------------- + TOTAL 0.08681979 309.37234178 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 25040 +BPFP 0.2951 bits/point +EBPFP 0.5901 equivalent bits/point +MSE 309.372342 +---------------------- -------------------------------------------------------- +Time: 2.932s Load: 0.004s, Pack+Encode: 1.687s, Decode+Unpack: 1.241s +---------------------- -------------------------------------------------------- +💾 Converting with 309.3723 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,228B, BPFP=0.2460 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,852B, BPFP=0.5713 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,036B, BPFP=0.6082 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,008B, BPFP=0.4022 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,424B, BPFP=0.4856 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,064B, BPFP=0.4135 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,260B, BPFP=0.4527 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,356B, BPFP=0.4720 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,996B, BPFP=0.6002 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,724B, BPFP=0.5457 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,404B, BPFP=0.0402 +⌛️ [2/4] FRONTEND: Frontend time: 1.775s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10595027 335.51562500 + layer.0.v_cache 0.00001376 0.03916374 + layer.1.k_cache 0.03307601 25.18965345 + layer.1.v_cache 0.00000550 0.01613666 + layer.2.k_cache 0.00695119 4.03080710 + layer.2.v_cache 0.00001929 0.04684169 + layer.3.k_cache 0.03450640 16.60051082 + layer.3.v_cache 0.00001844 0.05259078 + layer.4.k_cache 0.00061739 1.18839460 + layer.4.v_cache 0.00005297 0.09798013 + layer.4.output 0.18589628 695.93332189 + ------------------------------------------------------------------------------------- + TOTAL 0.08720501 309.07711513 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 25352 +BPFP 0.2987 bits/point +EBPFP 0.5975 equivalent bits/point +MSE 309.077115 +---------------------- -------------------------------------------------------- +Time: 2.925s Load: 0.005s, Pack+Encode: 1.775s, Decode+Unpack: 1.145s +---------------------- -------------------------------------------------------- +💾 Converting with 309.0771 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,256B, BPFP=0.2484 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,912B, BPFP=0.5759 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,952B, BPFP=0.5839 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,420B, BPFP=0.4786 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,836B, BPFP=0.5609 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,256B, BPFP=0.4462 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,132B, BPFP=0.4217 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,672B, BPFP=0.5285 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,840B, BPFP=0.5617 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,892B, BPFP=0.5720 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,528B, BPFP=0.0432 +⌛️ [2/4] FRONTEND: Frontend time: 1.702s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07772056 338.33865704 + layer.0.v_cache 0.00001351 0.04075260 + layer.1.k_cache 0.03518496 25.43567358 + layer.1.v_cache 0.00000556 0.01580806 + layer.2.k_cache 0.00230807 4.10718121 + layer.2.v_cache 0.00001833 0.04739976 + layer.3.k_cache 0.04514034 16.98403815 + layer.3.v_cache 0.00001901 0.05395030 + layer.4.k_cache 0.00062178 1.22929885 + layer.4.v_cache 0.00004990 0.10267591 + layer.4.output 0.18426369 687.25090416 + ------------------------------------------------------------------------------------- + TOTAL 0.08534870 305.71245674 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 26696 +BPFP 0.3106 bits/point +EBPFP 0.6212 equivalent bits/point +MSE 305.712457 +---------------------- -------------------------------------------------------- +Time: 2.919s Load: 0.004s, Pack+Encode: 1.702s, Decode+Unpack: 1.213s +---------------------- -------------------------------------------------------- +💾 Converting with 305.7125 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,240B, BPFP=0.2516 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,884B, BPFP=0.5852 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,244B, BPFP=0.6583 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,088B, BPFP=0.4237 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,700B, BPFP=0.5479 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,080B, BPFP=0.4221 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,228B, BPFP=0.4521 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,628B, BPFP=0.5333 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,992B, BPFP=0.6071 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,068B, BPFP=0.6226 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,340B, BPFP=0.0388 +⌛️ [2/4] FRONTEND: Frontend time: 1.733s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14568528 336.63002232 + layer.0.v_cache 0.00001322 0.03861331 + layer.1.k_cache 0.01379078 26.03942078 + layer.1.v_cache 0.00000548 0.01639459 + layer.2.k_cache 0.00727733 3.98731301 + layer.2.v_cache 0.00001786 0.04850694 + layer.3.k_cache 0.03082054 16.18624918 + layer.3.v_cache 0.00001879 0.05511598 + layer.4.k_cache 0.00061820 1.18609916 + layer.4.v_cache 0.00005377 0.10518418 + layer.4.output 0.19662610 705.10789657 + ------------------------------------------------------------------------------------- + TOTAL 0.09262847 312.94401150 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 26492 +BPFP 0.3162 bits/point +EBPFP 0.6324 equivalent bits/point +MSE 312.944011 +---------------------- -------------------------------------------------------- +Time: 2.935s Load: 0.004s, Pack+Encode: 1.733s, Decode+Unpack: 1.199s +---------------------- -------------------------------------------------------- +💾 Converting with 312.9440 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,228B, BPFP=0.2398 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,624B, BPFP=0.5125 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,056B, BPFP=0.5969 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,816B, BPFP=0.5500 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,676B, BPFP=0.5227 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,096B, BPFP=0.4094 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,456B, BPFP=0.4797 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,436B, BPFP=0.4758 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,884B, BPFP=0.5633 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,960B, BPFP=0.5781 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,420B, BPFP=0.0396 +⌛️ [2/4] FRONTEND: Frontend time: 1.659s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.179s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10149734 321.66181641 + layer.0.v_cache 0.00001421 0.04215872 + layer.1.k_cache 0.01324784 26.47544861 + layer.1.v_cache 0.00000586 0.01640189 + layer.2.k_cache 0.00706157 4.15174789 + layer.2.v_cache 0.00001822 0.04937894 + layer.3.k_cache 0.02712160 17.39275513 + layer.3.v_cache 0.00002016 0.05744398 + layer.4.k_cache 0.00063687 1.19263773 + layer.4.v_cache 0.00005451 0.10473888 + layer.4.output 0.17858309 678.72271205 + ------------------------------------------------------------------------------------- + TOTAL 0.08233881 301.30608897 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 26652 +BPFP 0.3062 bits/point +EBPFP 0.6124 equivalent bits/point +MSE 301.306089 +---------------------- -------------------------------------------------------- +Time: 2.841s Load: 0.003s, Pack+Encode: 1.659s, Decode+Unpack: 1.179s +---------------------- -------------------------------------------------------- +💾 Converting with 301.3061 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,600B, BPFP=0.2874 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,452B, BPFP=0.4404 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,728B, BPFP=0.4899 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,252B, BPFP=0.4045 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,496B, BPFP=0.4483 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,244B, BPFP=0.4030 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,416B, BPFP=0.4339 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,256B, BPFP=0.4052 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,888B, BPFP=0.5187 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,672B, BPFP=0.4799 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,424B, BPFP=0.0365 +⌛️ [2/4] FRONTEND: Frontend time: 1.689s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.192s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10772817 361.74836117 + layer.0.v_cache 0.00001348 0.04103419 + layer.1.k_cache 0.03276526 24.46841045 + layer.1.v_cache 0.00000562 0.01689594 + layer.2.k_cache 0.01020141 3.96928809 + layer.2.v_cache 0.00001934 0.04891458 + layer.3.k_cache 0.04694761 15.64968907 + layer.3.v_cache 0.00001910 0.05839158 + layer.4.k_cache 0.00062564 1.24147156 + layer.4.v_cache 0.00005323 0.10179293 + layer.4.output 0.17166949 623.95740969 + ------------------------------------------------------------------------------------- + TOTAL 0.08235678 280.88506573 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 25428 +BPFP 0.2686 bits/point +EBPFP 0.5373 equivalent bits/point +MSE 280.885066 +---------------------- -------------------------------------------------------- +Time: 2.885s Load: 0.004s, Pack+Encode: 1.689s, Decode+Unpack: 1.192s +---------------------- -------------------------------------------------------- +💾 Converting with 280.8851 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,212B, BPFP=0.2282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,608B, BPFP=0.4910 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,124B, BPFP=0.5881 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,624B, BPFP=0.4940 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,480B, BPFP=0.4669 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,032B, BPFP=0.3825 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,340B, BPFP=0.4405 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,748B, BPFP=0.5173 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,852B, BPFP=0.5369 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,888B, BPFP=0.5437 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,380B, BPFP=0.0371 +⌛️ [2/4] FRONTEND: Frontend time: 1.803s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12335577 330.04513366 + layer.0.v_cache 0.00001400 0.04072651 + layer.1.k_cache 0.01519604 26.07469527 + layer.1.v_cache 0.00000547 0.01520960 + layer.2.k_cache 0.01098718 4.07066106 + layer.2.v_cache 0.00001779 0.04874439 + layer.3.k_cache 0.07416311 16.47909215 + layer.3.v_cache 0.00001944 0.05401263 + layer.4.k_cache 0.00061719 1.20720994 + layer.4.v_cache 0.00006444 0.10069155 + layer.4.output 0.17504946 654.15205465 + ------------------------------------------------------------------------------------- + TOTAL 0.08528157 291.60003290 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 26288 +BPFP 0.2911 bits/point +EBPFP 0.5822 equivalent bits/point +MSE 291.600033 +---------------------- -------------------------------------------------------- +Time: 3.018s Load: 0.004s, Pack+Encode: 1.803s, Decode+Unpack: 1.212s +---------------------- -------------------------------------------------------- +💾 Converting with 291.6000 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,240B, BPFP=0.2516 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,960B, BPFP=0.6006 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,200B, BPFP=0.6494 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,304B, BPFP=0.4675 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,452B, BPFP=0.4976 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,876B, BPFP=0.3807 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,896B, BPFP=0.3847 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,124B, BPFP=0.4310 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,996B, BPFP=0.6080 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,824B, BPFP=0.5731 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,396B, BPFP=0.0405 +⌛️ [2/4] FRONTEND: Frontend time: 2.074s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13955514 338.68108259 + layer.0.v_cache 0.00001392 0.03983589 + layer.1.k_cache 0.03454666 25.91739169 + layer.1.v_cache 0.00000526 0.01581584 + layer.2.k_cache 0.00236268 4.06254597 + layer.2.v_cache 0.00001660 0.04602565 + layer.3.k_cache 0.02874016 15.60524997 + layer.3.v_cache 0.00001899 0.05380108 + layer.4.k_cache 0.00062191 1.18954270 + layer.4.v_cache 0.00005265 0.10173468 + layer.4.output 0.18728454 705.08998145 + ------------------------------------------------------------------------------------- + TOTAL 0.08923092 313.02017036 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 25268 +BPFP 0.3016 bits/point +EBPFP 0.6032 equivalent bits/point +MSE 313.020170 +---------------------- -------------------------------------------------------- +Time: 3.281s Load: 0.004s, Pack+Encode: 2.074s, Decode+Unpack: 1.203s +---------------------- -------------------------------------------------------- +💾 Converting with 313.0202 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,240B, BPFP=0.2516 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,900B, BPFP=0.5885 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,760B, BPFP=0.5601 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,320B, BPFP=0.4708 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,636B, BPFP=0.5349 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,952B, BPFP=0.3961 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,208B, BPFP=0.4481 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,748B, BPFP=0.5576 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,916B, BPFP=0.5917 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,000B, BPFP=0.6088 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,344B, BPFP=0.0390 +⌛️ [2/4] FRONTEND: Frontend time: 1.690s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11454884 339.52843446 + layer.0.v_cache 0.00001351 0.04203328 + layer.1.k_cache 0.01665594 25.61509804 + layer.1.v_cache 0.00000570 0.01733140 + layer.2.k_cache 0.00836552 4.17778471 + layer.2.v_cache 0.00001974 0.05155736 + layer.3.k_cache 0.03124422 16.56257293 + layer.3.v_cache 0.00001876 0.05391338 + layer.4.k_cache 0.00062231 1.16761582 + layer.4.v_cache 0.00005487 0.10277539 + layer.4.output 0.19663499 705.12633349 + ------------------------------------------------------------------------------------- + TOTAL 0.09105849 313.12961477 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 26024 +BPFP 0.3106 bits/point +EBPFP 0.6213 equivalent bits/point +MSE 313.129615 +---------------------- -------------------------------------------------------- +Time: 2.916s Load: 0.004s, Pack+Encode: 1.690s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 313.1296 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,216B, BPFP=0.2262 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,720B, BPFP=0.5060 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,108B, BPFP=0.5781 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,028B, BPFP=0.3772 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,476B, BPFP=0.4606 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,320B, BPFP=0.4315 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,112B, BPFP=0.3929 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,248B, BPFP=0.4182 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,772B, BPFP=0.5156 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,804B, BPFP=0.5216 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0372 +⌛️ [2/4] FRONTEND: Frontend time: 1.672s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12290487 320.11232794 + layer.0.v_cache 0.00001422 0.04081209 + layer.1.k_cache 0.01283375 24.55873617 + layer.1.v_cache 0.00000579 0.01684582 + layer.2.k_cache 0.01100613 3.68482971 + layer.2.v_cache 0.00001929 0.04628950 + layer.3.k_cache 0.03011658 16.40779768 + layer.3.v_cache 0.00001858 0.05584252 + layer.4.k_cache 0.00063020 1.21728970 + layer.4.v_cache 0.00005223 0.10041584 + layer.4.output 0.17536174 646.36787840 + ------------------------------------------------------------------------------------- + TOTAL 0.08265493 287.69507857 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 25204 +BPFP 0.2758 bits/point +EBPFP 0.5516 equivalent bits/point +MSE 287.695079 +---------------------- -------------------------------------------------------- +Time: 2.872s Load: 0.004s, Pack+Encode: 1.672s, Decode+Unpack: 1.197s +---------------------- -------------------------------------------------------- +💾 Converting with 287.6951 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,240B, BPFP=0.2279 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,548B, BPFP=0.4684 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,904B, BPFP=0.5338 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,280B, BPFP=0.4191 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,544B, BPFP=0.4676 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,252B, BPFP=0.4140 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,292B, BPFP=0.4213 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,344B, BPFP=0.4309 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,652B, BPFP=0.4875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,620B, BPFP=0.4816 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,404B, BPFP=0.0369 +⌛️ [2/4] FRONTEND: Frontend time: 1.727s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.180s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11218034 327.80615809 + layer.0.v_cache 0.00001438 0.04070062 + layer.1.k_cache 0.03157643 24.86049517 + layer.1.v_cache 0.00000545 0.01609267 + layer.2.k_cache 0.00802605 3.99199614 + layer.2.v_cache 0.00001826 0.04706156 + layer.3.k_cache 0.04519741 16.46256893 + layer.3.v_cache 0.00001794 0.05097445 + layer.4.k_cache 0.00061137 1.14707336 + layer.4.v_cache 0.00005913 0.10013993 + layer.4.output 0.17010492 638.54254202 + ------------------------------------------------------------------------------------- + TOTAL 0.08167301 284.96006206 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 25080 +BPFP 0.2712 bits/point +EBPFP 0.5424 equivalent bits/point +MSE 284.960062 +---------------------- -------------------------------------------------------- +Time: 2.913s Load: 0.005s, Pack+Encode: 1.727s, Decode+Unpack: 1.180s +---------------------- -------------------------------------------------------- +💾 Converting with 284.9601 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,240B, BPFP=0.2253 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,660B, BPFP=0.4833 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,752B, BPFP=0.5000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,176B, BPFP=0.3953 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,484B, BPFP=0.4513 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,448B, BPFP=0.4448 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,372B, BPFP=0.4310 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,500B, BPFP=0.4542 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,860B, BPFP=0.5196 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,780B, BPFP=0.5051 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,520B, BPFP=0.0395 +⌛️ [2/4] FRONTEND: Frontend time: 1.724s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10146660 325.30552780 + layer.0.v_cache 0.00001344 0.04018178 + layer.1.k_cache 0.03090366 25.32400300 + layer.1.v_cache 0.00000566 0.01586273 + layer.2.k_cache 0.01188999 3.93568137 + layer.2.v_cache 0.00001845 0.04536481 + layer.3.k_cache 0.02542871 15.55100268 + layer.3.v_cache 0.00001884 0.05348467 + layer.4.k_cache 0.00060243 1.21632545 + layer.4.v_cache 0.00005224 0.09903305 + layer.4.output 0.16839122 631.28535091 + ------------------------------------------------------------------------------------- + TOTAL 0.07936109 281.79905434 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 25792 +BPFP 0.2756 bits/point +EBPFP 0.5513 equivalent bits/point +MSE 281.799054 +---------------------- -------------------------------------------------------- +Time: 3.103s Load: 0.004s, Pack+Encode: 1.724s, Decode+Unpack: 1.375s +---------------------- -------------------------------------------------------- +💾 Converting with 281.7991 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,392B, BPFP=0.2364 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,368B, BPFP=0.4022 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,724B, BPFP=0.4626 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,452B, BPFP=0.4164 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,364B, BPFP=0.4015 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,512B, BPFP=0.4266 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,824B, BPFP=0.3098 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,276B, BPFP=0.3865 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,504B, BPFP=0.4253 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,400B, BPFP=0.4076 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,312B, BPFP=0.0318 +⌛️ [2/4] FRONTEND: Frontend time: 1.769s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.169s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12981709 359.81479280 + layer.0.v_cache 0.00001628 0.04054676 + layer.1.k_cache 0.06761242 25.35759967 + layer.1.v_cache 0.00000547 0.01557810 + layer.2.k_cache 0.00384689 3.62700421 + layer.2.v_cache 0.00001812 0.04727204 + layer.3.k_cache 0.05327560 15.91282322 + layer.3.v_cache 0.00001896 0.05564627 + layer.4.k_cache 0.00060800 1.16273300 + layer.4.v_cache 0.00005116 0.10183586 + layer.4.output 0.15518735 590.16750776 + ------------------------------------------------------------------------------------- + TOTAL 0.07891655 266.90049331 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 24128 +BPFP 0.2410 bits/point +EBPFP 0.4821 equivalent bits/point +MSE 266.900493 +---------------------- -------------------------------------------------------- +Time: 2.942s Load: 0.005s, Pack+Encode: 1.769s, Decode+Unpack: 1.169s +---------------------- -------------------------------------------------------- +💾 Converting with 266.9005 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,192B, BPFP=0.2299 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,756B, BPFP=0.5316 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,148B, BPFP=0.6073 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,348B, BPFP=0.4529 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,712B, BPFP=0.5231 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,200B, BPFP=0.4244 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,140B, BPFP=0.4128 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,468B, BPFP=0.4761 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,020B, BPFP=0.5826 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,984B, BPFP=0.5756 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,364B, BPFP=0.0376 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.179s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887507 328.07262731 + layer.0.v_cache 0.00001345 0.04126784 + layer.1.k_cache 0.03259810 26.02413978 + layer.1.v_cache 0.00000523 0.01552801 + layer.2.k_cache 0.00244350 4.18318459 + layer.2.v_cache 0.00001834 0.04723636 + layer.3.k_cache 0.10864470 16.31486003 + layer.3.v_cache 0.00001904 0.05367420 + layer.4.k_cache 0.00060864 1.19143724 + layer.4.v_cache 0.00004940 0.10199775 + layer.4.output 0.18628448 670.17901235 + ------------------------------------------------------------------------------------- + TOTAL 0.09395687 298.07641409 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 26332 +BPFP 0.2988 bits/point +EBPFP 0.5976 equivalent bits/point +MSE 298.076414 +---------------------- -------------------------------------------------------- +Time: 2.912s Load: 0.004s, Pack+Encode: 1.730s, Decode+Unpack: 1.179s +---------------------- -------------------------------------------------------- +💾 Converting with 298.0764 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,312B, BPFP=0.2303 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,724B, BPFP=0.4782 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,640B, BPFP=0.4635 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,560B, BPFP=0.4494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,360B, BPFP=0.4143 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,408B, BPFP=0.4228 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,048B, BPFP=0.3596 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,344B, BPFP=0.4115 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,612B, BPFP=0.4586 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,872B, BPFP=0.5042 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0351 +⌛️ [2/4] FRONTEND: Frontend time: 1.750s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.264s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422350 326.80517029 + layer.0.v_cache 0.00001493 0.04115438 + layer.1.k_cache 0.01422887 25.63874320 + layer.1.v_cache 0.00000595 0.01601590 + layer.2.k_cache 0.00668612 3.91917642 + layer.2.v_cache 0.00001902 0.04828947 + layer.3.k_cache 0.02497406 17.14236004 + layer.3.v_cache 0.00001869 0.05621188 + layer.4.k_cache 0.00061935 1.17209008 + layer.4.v_cache 0.00005212 0.10101671 + layer.4.output 0.16899524 609.98033708 + ------------------------------------------------------------------------------------- + TOTAL 0.07904761 273.22368164 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 25280 +BPFP 0.2611 bits/point +EBPFP 0.5221 equivalent bits/point +MSE 273.223682 +---------------------- -------------------------------------------------------- +Time: 3.019s Load: 0.004s, Pack+Encode: 1.750s, Decode+Unpack: 1.264s +---------------------- -------------------------------------------------------- +💾 Converting with 273.2237 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 113, 128) +Output shape: (1, 113, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.output: torch.Size([1, 113, 3584]) -> torch.Size([1, 1, 113, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,728B, BPFP=0.2389 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,812B, BPFP=0.3888 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,136B, BPFP=0.4336 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,784B, BPFP=0.3850 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,612B, BPFP=0.3612 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,684B, BPFP=0.3711 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,464B, BPFP=0.3407 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,604B, BPFP=0.3601 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,352B, BPFP=0.4635 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,724B, BPFP=0.3767 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,488B, BPFP=0.0294 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.329s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10214825 362.72151549 + layer.0.v_cache 0.00001637 0.04062372 + layer.1.k_cache 0.04284562 24.21918859 + layer.1.v_cache 0.00000578 0.01621645 + layer.2.k_cache 0.00913706 3.75300612 + layer.2.v_cache 0.00001928 0.04774137 + layer.3.k_cache 0.05585717 14.72654846 + layer.3.v_cache 0.00001870 0.05206507 + layer.4.k_cache 0.00065968 1.25408503 + layer.4.v_cache 0.00005753 0.09817642 + layer.4.output 10.09706179 475.73245891 + ------------------------------------------------------------------------------------- + TOTAL 4.17001165 219.82684583 + (elements=983,552) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 983552 +Total Bytes 28388 +BPFP 0.2309 bits/point +EBPFP 0.4618 equivalent bits/point +MSE 219.826846 +---------------------- -------------------------------------------------------- +Time: 3.167s Load: 0.005s, Pack+Encode: 1.833s, Decode+Unpack: 1.329s +---------------------- -------------------------------------------------------- +💾 Converting with 219.8268 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,448B, BPFP=0.2486 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,532B, BPFP=0.4348 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,640B, BPFP=0.4533 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,384B, BPFP=0.4093 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,436B, BPFP=0.4183 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,144B, BPFP=0.3681 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,000B, BPFP=0.3434 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,256B, BPFP=0.3874 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,456B, BPFP=0.4217 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,636B, BPFP=0.4526 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,288B, BPFP=0.0316 +⌛️ [2/4] FRONTEND: Frontend time: 1.960s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.396s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11481400 344.48860319 + layer.0.v_cache 0.00001396 0.04135568 + layer.1.k_cache 0.03135788 25.44525079 + layer.1.v_cache 0.00000545 0.01602885 + layer.2.k_cache 0.00370715 3.91869446 + layer.2.v_cache 0.00001965 0.04824948 + layer.3.k_cache 0.03930106 16.51566793 + layer.3.v_cache 0.00001886 0.05748711 + layer.4.k_cache 0.00060819 1.21462376 + layer.4.v_cache 0.00005437 0.10287096 + layer.4.output 0.16591480 596.60665228 + ------------------------------------------------------------------------------------- + TOTAL 0.07948848 268.71149401 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 24220 +BPFP 0.2446 bits/point +EBPFP 0.4893 equivalent bits/point +MSE 268.711494 +---------------------- -------------------------------------------------------- +Time: 3.362s Load: 0.006s, Pack+Encode: 1.960s, Decode+Unpack: 1.396s +---------------------- -------------------------------------------------------- +💾 Converting with 268.7115 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,324B, BPFP=0.2324 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,688B, BPFP=0.4719 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,660B, BPFP=0.4670 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,440B, BPFP=0.4284 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,312B, BPFP=0.4059 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,580B, BPFP=0.4529 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,832B, BPFP=0.3216 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,564B, BPFP=0.4501 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,772B, BPFP=0.4867 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,672B, BPFP=0.4691 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,388B, BPFP=0.0348 +⌛️ [2/4] FRONTEND: Frontend time: 1.862s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.249s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09788724 341.86955758 + layer.0.v_cache 0.00001308 0.03950488 + layer.1.k_cache 0.03299916 24.94927636 + layer.1.v_cache 0.00000544 0.01575070 + layer.2.k_cache 0.00367894 3.81157076 + layer.2.v_cache 0.00002012 0.04637001 + layer.3.k_cache 0.03939044 17.14542277 + layer.3.v_cache 0.00001875 0.05262605 + layer.4.k_cache 0.00061788 1.15851482 + layer.4.v_cache 0.00005010 0.10071645 + layer.4.output 0.17021053 609.96890048 + ------------------------------------------------------------------------------------- + TOTAL 0.08036205 274.05715375 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 25232 +BPFP 0.2606 bits/point +EBPFP 0.5212 equivalent bits/point +MSE 274.057154 +---------------------- -------------------------------------------------------- +Time: 3.116s Load: 0.005s, Pack+Encode: 1.862s, Decode+Unpack: 1.249s +---------------------- -------------------------------------------------------- +💾 Converting with 274.0572 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,684B, BPFP=0.2436 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,764B, BPFP=0.3999 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,232B, BPFP=0.4676 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,792B, BPFP=0.4039 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,676B, BPFP=0.3872 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,600B, BPFP=0.3762 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,512B, BPFP=0.3634 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,600B, BPFP=0.3762 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,228B, BPFP=0.4670 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,872B, BPFP=0.4155 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,692B, BPFP=0.0350 +⌛️ [2/4] FRONTEND: Frontend time: 1.690s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12672405 365.21383102 + layer.0.v_cache 0.00001401 0.03890890 + layer.1.k_cache 0.12044146 24.67720540 + layer.1.v_cache 0.00000568 0.01590250 + layer.2.k_cache 0.00570256 3.31591062 + layer.2.v_cache 0.00001866 0.04801099 + layer.3.k_cache 0.03611955 14.68941696 + layer.3.v_cache 0.00001983 0.05741297 + layer.4.k_cache 0.00066233 1.20245828 + layer.4.v_cache 0.00005331 0.10089592 + layer.4.output 10.56037249 498.05468750 + ------------------------------------------------------------------------------------- + TOTAL 4.36543346 229.16133918 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 28652 +BPFP 0.2438 bits/point +EBPFP 0.4877 equivalent bits/point +MSE 229.161339 +---------------------- -------------------------------------------------------- +Time: 2.900s Load: 0.005s, Pack+Encode: 1.690s, Decode+Unpack: 1.205s +---------------------- -------------------------------------------------------- +💾 Converting with 229.1613 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,292B, BPFP=0.2294 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,752B, BPFP=0.4886 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,624B, BPFP=0.4659 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,280B, BPFP=0.4048 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,336B, BPFP=0.4148 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,236B, BPFP=0.3970 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,960B, BPFP=0.3480 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,680B, BPFP=0.4759 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,568B, BPFP=0.4560 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,656B, BPFP=0.4716 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,560B, BPFP=0.0396 +⌛️ [2/4] FRONTEND: Frontend time: 1.707s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.157s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13733016 322.77747692 + layer.0.v_cache 0.00001373 0.04147889 + layer.1.k_cache 0.03155638 25.34754250 + layer.1.v_cache 0.00000617 0.01661289 + layer.2.k_cache 0.00751131 3.76872427 + layer.2.v_cache 0.00001754 0.04597597 + layer.3.k_cache 0.02548880 16.09541460 + layer.3.v_cache 0.00001765 0.05185706 + layer.4.k_cache 0.00061011 1.14277961 + layer.4.v_cache 0.00005045 0.09939974 + layer.4.output 0.16239834 617.00771104 + ------------------------------------------------------------------------------------- + TOTAL 0.07878769 275.79066116 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 24944 +BPFP 0.2605 bits/point +EBPFP 0.5211 equivalent bits/point +MSE 275.790661 +---------------------- -------------------------------------------------------- +Time: 2.868s Load: 0.004s, Pack+Encode: 1.707s, Decode+Unpack: 1.157s +---------------------- -------------------------------------------------------- +💾 Converting with 275.7907 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,284B, BPFP=0.2540 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,732B, BPFP=0.5403 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,964B, BPFP=0.5862 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,668B, BPFP=0.5277 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,520B, BPFP=0.4984 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,328B, BPFP=0.4604 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,312B, BPFP=0.4573 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,368B, BPFP=0.4684 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,952B, BPFP=0.5839 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,968B, BPFP=0.5870 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,488B, BPFP=0.0420 +⌛️ [2/4] FRONTEND: Frontend time: 1.660s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.165s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10979201 339.85324367 + layer.0.v_cache 0.00001339 0.03976907 + layer.1.k_cache 0.01160004 26.15602749 + layer.1.v_cache 0.00000567 0.01549947 + layer.2.k_cache 0.00714037 4.10604858 + layer.2.v_cache 0.00001829 0.04629585 + layer.3.k_cache 0.04622446 16.56926640 + layer.3.v_cache 0.00001896 0.05419836 + layer.4.k_cache 0.00064378 1.19290123 + layer.4.v_cache 0.00005251 0.10105065 + layer.4.output 0.18415536 687.26005877 + ------------------------------------------------------------------------------------- + TOTAL 0.08615276 305.82086542 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 26584 +BPFP 0.3093 bits/point +EBPFP 0.6186 equivalent bits/point +MSE 305.820865 +---------------------- -------------------------------------------------------- +Time: 2.830s Load: 0.004s, Pack+Encode: 1.660s, Decode+Unpack: 1.165s +---------------------- -------------------------------------------------------- +💾 Converting with 305.8209 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,220B, BPFP=0.2476 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,700B, BPFP=0.5479 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,996B, BPFP=0.6080 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,124B, BPFP=0.4310 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,336B, BPFP=0.4740 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,816B, BPFP=0.3685 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,048B, BPFP=0.4156 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,820B, BPFP=0.3693 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,944B, BPFP=0.5974 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,576B, BPFP=0.5227 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,348B, BPFP=0.0391 +⌛️ [2/4] FRONTEND: Frontend time: 1.724s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11290005 335.11038961 + layer.0.v_cache 0.00001570 0.04145262 + layer.1.k_cache 0.03434501 25.15299056 + layer.1.v_cache 0.00000534 0.01576473 + layer.2.k_cache 0.00238675 3.82309118 + layer.2.v_cache 0.00001838 0.04832066 + layer.3.k_cache 0.02838124 16.45770343 + layer.3.v_cache 0.00002047 0.05750548 + layer.4.k_cache 0.00062757 1.22375369 + layer.4.v_cache 0.00005455 0.10855434 + layer.4.output 0.18420308 705.20182050 + ------------------------------------------------------------------------------------- + TOTAL 0.08636333 312.85013352 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 23928 +BPFP 0.2856 bits/point +EBPFP 0.5712 equivalent bits/point +MSE 312.850134 +---------------------- -------------------------------------------------------- +Time: 2.931s Load: 0.003s, Pack+Encode: 1.724s, Decode+Unpack: 1.204s +---------------------- -------------------------------------------------------- +💾 Converting with 312.8501 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,208B, BPFP=0.2420 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,784B, BPFP=0.5577 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,916B, BPFP=0.5841 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,428B, BPFP=0.4864 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,508B, BPFP=0.5024 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,292B, BPFP=0.4591 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,052B, BPFP=0.4111 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,324B, BPFP=0.4655 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,856B, BPFP=0.5721 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,908B, BPFP=0.5825 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,412B, BPFP=0.0404 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09752406 334.71058694 + layer.0.v_cache 0.00001393 0.03960847 + layer.1.k_cache 0.01493649 25.77142803 + layer.1.v_cache 0.00000588 0.01556849 + layer.2.k_cache 0.01003601 4.10092593 + layer.2.v_cache 0.00001805 0.04672460 + layer.3.k_cache 0.04451986 17.24561799 + layer.3.v_cache 0.00001914 0.05549949 + layer.4.k_cache 0.00061869 1.18596121 + layer.4.v_cache 0.00005251 0.10368770 + layer.4.output 0.18409089 696.13118132 + ------------------------------------------------------------------------------------- + TOTAL 0.08566946 309.18787518 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 25688 +BPFP 0.3027 bits/point +EBPFP 0.6054 equivalent bits/point +MSE 309.187875 +---------------------- -------------------------------------------------------- +Time: 3.077s Load: 0.004s, Pack+Encode: 1.730s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 309.1879 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,244B, BPFP=0.2342 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,992B, BPFP=0.5633 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,296B, BPFP=0.6205 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,496B, BPFP=0.4699 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,452B, BPFP=0.4616 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,152B, BPFP=0.4051 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,196B, BPFP=0.4134 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,120B, BPFP=0.3991 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,968B, BPFP=0.5587 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,964B, BPFP=0.5580 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,376B, BPFP=0.0370 +⌛️ [2/4] FRONTEND: Frontend time: 2.056s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.421s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13658243 330.19935994 + layer.0.v_cache 0.00001379 0.04069756 + layer.1.k_cache 0.01234693 25.76836349 + layer.1.v_cache 0.00000566 0.01564593 + layer.2.k_cache 0.00375735 4.02836848 + layer.2.v_cache 0.00001852 0.04581674 + layer.3.k_cache 0.02700986 16.61372394 + layer.3.v_cache 0.00001930 0.05306019 + layer.4.k_cache 0.00062007 1.21819572 + layer.4.v_cache 0.00005146 0.09985908 + layer.4.output 0.17614663 654.14925775 + ------------------------------------------------------------------------------------- + TOTAL 0.08314423 291.59575855 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 26256 +BPFP 0.2908 bits/point +EBPFP 0.5815 equivalent bits/point +MSE 291.595759 +---------------------- -------------------------------------------------------- +Time: 3.483s Load: 0.005s, Pack+Encode: 2.056s, Decode+Unpack: 1.421s +---------------------- -------------------------------------------------------- +💾 Converting with 291.5958 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,140B, BPFP=0.2440 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,448B, BPFP=0.5240 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,280B, BPFP=0.4880 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,788B, BPFP=0.3827 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,108B, BPFP=0.4512 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,800B, BPFP=0.3853 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,924B, BPFP=0.4118 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,104B, BPFP=0.4503 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,280B, BPFP=0.4880 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,464B, BPFP=0.5274 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,436B, BPFP=0.0439 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08738674 328.89228382 + layer.0.v_cache 0.00001368 0.04043939 + layer.1.k_cache 0.01286346 25.45067356 + layer.1.v_cache 0.00000597 0.01635258 + layer.2.k_cache 0.00251921 3.93543484 + layer.2.v_cache 0.00001769 0.04870948 + layer.3.k_cache 0.08324167 16.26524708 + layer.3.v_cache 0.00002021 0.05541611 + layer.4.k_cache 0.00062021 1.23856511 + layer.4.v_cache 0.00005375 0.11295355 + layer.4.output 0.19567458 743.71673190 + ------------------------------------------------------------------------------------- + TOTAL 0.09155674 328.35724699 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 21772 +BPFP 0.2741 bits/point +EBPFP 0.5482 equivalent bits/point +MSE 328.357247 +---------------------- -------------------------------------------------------- +Time: 2.950s Load: 0.004s, Pack+Encode: 1.730s, Decode+Unpack: 1.217s +---------------------- -------------------------------------------------------- +💾 Converting with 328.3572 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,224B, BPFP=0.2304 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,760B, BPFP=0.5196 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,008B, BPFP=0.5663 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,296B, BPFP=0.4322 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,440B, BPFP=0.4593 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,228B, BPFP=0.4194 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,500B, BPFP=0.4706 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,488B, BPFP=0.4684 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,000B, BPFP=0.5648 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,664B, BPFP=0.5015 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,388B, BPFP=0.0373 +⌛️ [2/4] FRONTEND: Frontend time: 1.697s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.242s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12176956 331.64904932 + layer.0.v_cache 0.00001396 0.03930027 + layer.1.k_cache 0.01120206 25.92815206 + layer.1.v_cache 0.00000562 0.01615202 + layer.2.k_cache 0.00678912 3.94678470 + layer.2.v_cache 0.00001849 0.04715057 + layer.3.k_cache 0.02622183 17.50668592 + layer.3.v_cache 0.00001965 0.05433699 + layer.4.k_cache 0.00062413 1.24193784 + layer.4.v_cache 0.00005248 0.10228874 + layer.4.output 0.17940697 654.13269148 + ------------------------------------------------------------------------------------- + TOTAL 0.08368034 291.73298111 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 25996 +BPFP 0.2879 bits/point +EBPFP 0.5757 equivalent bits/point +MSE 291.732981 +---------------------- -------------------------------------------------------- +Time: 2.943s Load: 0.004s, Pack+Encode: 1.697s, Decode+Unpack: 1.242s +---------------------- -------------------------------------------------------- +💾 Converting with 291.7330 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,676B, BPFP=0.2494 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,764B, BPFP=0.4113 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,272B, BPFP=0.4869 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,820B, BPFP=0.4196 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,512B, BPFP=0.3738 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,628B, BPFP=0.3911 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,584B, BPFP=0.3845 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,652B, BPFP=0.3946 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,076B, BPFP=0.4577 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,868B, BPFP=0.4268 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,396B, BPFP=0.0297 +⌛️ [2/4] FRONTEND: Frontend time: 1.932s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120816 382.69654018 + layer.0.v_cache 0.00001607 0.04011159 + layer.1.k_cache 0.10678518 25.15455729 + layer.1.v_cache 0.00000585 0.01618315 + layer.2.k_cache 0.00813519 3.45630580 + layer.2.v_cache 0.00001919 0.04799064 + layer.3.k_cache 0.06560528 14.69322684 + layer.3.v_cache 0.00001973 0.05662306 + layer.4.k_cache 0.00062300 1.18735003 + layer.4.v_cache 0.00005287 0.10166902 + layer.4.output 10.86395687 512.24340986 + ------------------------------------------------------------------------------------- + TOTAL 4.49059815 236.06790745 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 28248 +BPFP 0.2473 bits/point +EBPFP 0.4945 equivalent bits/point +MSE 236.067907 +---------------------- -------------------------------------------------------- +Time: 3.143s Load: 0.005s, Pack+Encode: 1.932s, Decode+Unpack: 1.206s +---------------------- -------------------------------------------------------- +💾 Converting with 236.0679 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,612B, BPFP=0.2895 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,600B, BPFP=0.4670 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,744B, BPFP=0.4928 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,256B, BPFP=0.4052 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,336B, BPFP=0.4195 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,508B, BPFP=0.4504 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,348B, BPFP=0.4217 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,544B, BPFP=0.4569 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,684B, BPFP=0.4820 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,636B, BPFP=0.4734 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,420B, BPFP=0.0364 +⌛️ [2/4] FRONTEND: Frontend time: 1.687s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.178s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12848519 355.17688129 + layer.0.v_cache 0.00001371 0.03909719 + layer.1.k_cache 0.03007136 25.07680327 + layer.1.v_cache 0.00000570 0.01587283 + layer.2.k_cache 0.00936044 3.58710839 + layer.2.v_cache 0.00001842 0.04583079 + layer.3.k_cache 0.04065019 16.05590680 + layer.3.v_cache 0.00001877 0.05494013 + layer.4.k_cache 0.00062204 1.21742363 + layer.4.v_cache 0.00010837 0.09785859 + layer.4.output 0.16426703 624.13172209 + ------------------------------------------------------------------------------------- + TOTAL 0.07995432 280.60528103 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 25688 +BPFP 0.2714 bits/point +EBPFP 0.5428 equivalent bits/point +MSE 280.605281 +---------------------- -------------------------------------------------------- +Time: 2.869s Load: 0.004s, Pack+Encode: 1.687s, Decode+Unpack: 1.178s +---------------------- -------------------------------------------------------- +💾 Converting with 280.6053 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,256B, BPFP=0.2453 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,748B, BPFP=0.5367 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,220B, BPFP=0.6289 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,076B, BPFP=0.4055 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,500B, BPFP=0.4883 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,156B, BPFP=0.4211 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,484B, BPFP=0.4852 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,352B, BPFP=0.4594 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,884B, BPFP=0.5633 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,944B, BPFP=0.5750 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,396B, BPFP=0.0390 +⌛️ [2/4] FRONTEND: Frontend time: 1.699s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.306s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09257430 327.61303711 + layer.0.v_cache 0.00001329 0.04015619 + layer.1.k_cache 0.01202901 25.77683105 + layer.1.v_cache 0.00000544 0.01606285 + layer.2.k_cache 0.00549439 4.07233849 + layer.2.v_cache 0.00001849 0.04745887 + layer.3.k_cache 0.11014032 17.36524353 + layer.3.v_cache 0.00001916 0.05299085 + layer.4.k_cache 0.00063444 1.17884502 + layer.4.v_cache 0.00005494 0.09898395 + layer.4.output 0.19143520 678.51774554 + ------------------------------------------------------------------------------------- + TOTAL 0.09182531 301.52271569 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 26016 +BPFP 0.2989 bits/point +EBPFP 0.5978 equivalent bits/point +MSE 301.522716 +---------------------- -------------------------------------------------------- +Time: 3.008s Load: 0.004s, Pack+Encode: 1.699s, Decode+Unpack: 1.306s +---------------------- -------------------------------------------------------- +💾 Converting with 301.5227 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,228B, BPFP=0.2284 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,696B, BPFP=0.5015 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,100B, BPFP=0.5766 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,388B, BPFP=0.4442 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,508B, BPFP=0.4665 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,160B, BPFP=0.4018 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,128B, BPFP=0.3958 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,504B, BPFP=0.4658 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,720B, BPFP=0.5060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,820B, BPFP=0.5246 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0372 +⌛️ [2/4] FRONTEND: Frontend time: 1.780s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.173s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12365316 314.63550967 + layer.0.v_cache 0.00001332 0.04038837 + layer.1.k_cache 0.03511774 25.16326032 + layer.1.v_cache 0.00000569 0.01707263 + layer.2.k_cache 0.01236608 3.91434806 + layer.2.v_cache 0.00001835 0.04603103 + layer.3.k_cache 0.02851222 16.01804897 + layer.3.v_cache 0.00001844 0.05147086 + layer.4.k_cache 0.00063207 1.19027365 + layer.4.v_cache 0.00005066 0.09745381 + layer.4.output 0.17964420 646.23469388 + ------------------------------------------------------------------------------------- + TOTAL 0.08575865 287.34215968 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 25652 +BPFP 0.2807 bits/point +EBPFP 0.5614 equivalent bits/point +MSE 287.342160 +---------------------- -------------------------------------------------------- +Time: 2.958s Load: 0.005s, Pack+Encode: 1.780s, Decode+Unpack: 1.173s +---------------------- -------------------------------------------------------- +💾 Converting with 287.3422 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,228B, BPFP=0.2460 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,708B, BPFP=0.5425 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,888B, BPFP=0.5785 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,260B, BPFP=0.4527 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,204B, BPFP=0.4415 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,376B, BPFP=0.4760 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,144B, BPFP=0.4295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,200B, BPFP=0.4407 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,900B, BPFP=0.5809 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,748B, BPFP=0.5505 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,464B, BPFP=0.0419 +⌛️ [2/4] FRONTEND: Frontend time: 1.661s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12096892 326.35156250 + layer.0.v_cache 0.00001393 0.04018680 + layer.1.k_cache 0.03247837 26.36333133 + layer.1.v_cache 0.00000613 0.01594055 + layer.2.k_cache 0.00394396 3.92551207 + layer.2.v_cache 0.00001920 0.04726283 + layer.3.k_cache 0.02751121 16.77561599 + layer.3.v_cache 0.00001919 0.05750094 + layer.4.k_cache 0.00062481 1.20201483 + layer.4.v_cache 0.00005293 0.10125855 + layer.4.output 0.18564256 696.09895833 + ------------------------------------------------------------------------------------- + TOTAL 0.08736098 308.68075851 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 25120 +BPFP 0.2960 bits/point +EBPFP 0.5920 equivalent bits/point +MSE 308.680759 +---------------------- -------------------------------------------------------- +Time: 2.864s Load: 0.003s, Pack+Encode: 1.661s, Decode+Unpack: 1.199s +---------------------- -------------------------------------------------------- +💾 Converting with 308.6808 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,208B, BPFP=0.2451 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,884B, BPFP=0.5852 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,928B, BPFP=0.5942 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,248B, BPFP=0.4562 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,532B, BPFP=0.5138 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,188B, BPFP=0.4440 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,120B, BPFP=0.4302 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,032B, BPFP=0.4123 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,896B, BPFP=0.5877 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,848B, BPFP=0.5779 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,320B, BPFP=0.0383 +⌛️ [2/4] FRONTEND: Frontend time: 1.755s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.333s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08564486 343.44338474 + layer.0.v_cache 0.00001408 0.04115993 + layer.1.k_cache 0.03507188 25.81873034 + layer.1.v_cache 0.00000566 0.01607569 + layer.2.k_cache 0.00239554 3.96839726 + layer.2.v_cache 0.00001938 0.04707432 + layer.3.k_cache 0.03313774 16.43325766 + layer.3.v_cache 0.00001993 0.05785865 + layer.4.k_cache 0.00061714 1.14991423 + layer.4.v_cache 0.00005134 0.10106618 + layer.4.output 0.18226704 705.20587894 + ------------------------------------------------------------------------------------- + TOTAL 0.08428511 313.38341598 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 25204 +BPFP 0.3008 bits/point +EBPFP 0.6017 equivalent bits/point +MSE 313.383416 +---------------------- -------------------------------------------------------- +Time: 3.093s Load: 0.005s, Pack+Encode: 1.755s, Decode+Unpack: 1.333s +---------------------- -------------------------------------------------------- +💾 Converting with 313.3834 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 117, 128) +Output shape: (1, 117, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.output: torch.Size([1, 117, 3584]) -> torch.Size([1, 1, 117, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,832B, BPFP=0.2447 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,780B, BPFP=0.3713 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,164B, BPFP=0.4225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,708B, BPFP=0.3616 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,512B, BPFP=0.3355 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,684B, BPFP=0.3584 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,388B, BPFP=0.3189 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,404B, BPFP=0.4546 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,052B, BPFP=0.4076 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,252B, BPFP=0.4343 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,200B, BPFP=0.0420 +⌛️ [2/4] FRONTEND: Frontend time: 1.789s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.173s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11900647 392.67821848 + layer.0.v_cache 0.00001428 0.04335551 + layer.1.k_cache 0.12535267 25.72907068 + layer.1.v_cache 0.00000579 0.01769688 + layer.2.k_cache 0.00517510 3.63681004 + layer.2.v_cache 0.00001992 0.05122067 + layer.3.k_cache 0.09315677 15.17075028 + layer.3.v_cache 0.00001978 0.06255134 + layer.4.k_cache 0.00061896 1.25668009 + layer.4.v_cache 0.00005133 0.10614576 + layer.4.output 9.75369281 459.55200702 + ------------------------------------------------------------------------------------- + TOTAL 4.03642769 215.03626758 + (elements=1,018,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1018368 +Total Bytes 29976 +BPFP 0.2355 bits/point +EBPFP 0.4710 equivalent bits/point +MSE 215.036268 +---------------------- -------------------------------------------------------- +Time: 2.967s Load: 0.005s, Pack+Encode: 1.789s, Decode+Unpack: 1.173s +---------------------- -------------------------------------------------------- +💾 Converting with 215.0363 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,700B, BPFP=0.2530 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,544B, BPFP=0.3786 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,168B, BPFP=0.4714 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,584B, BPFP=0.3845 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,656B, BPFP=0.3952 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,424B, BPFP=0.3607 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,592B, BPFP=0.3857 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,680B, BPFP=0.3988 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,044B, BPFP=0.4530 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,812B, BPFP=0.4185 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,444B, BPFP=0.0307 +⌛️ [2/4] FRONTEND: Frontend time: 1.673s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.263s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14131550 368.23515625 + layer.0.v_cache 0.00001383 0.03983206 + layer.1.k_cache 0.02978072 24.81055618 + layer.1.v_cache 0.00000537 0.01572983 + layer.2.k_cache 0.00578868 3.32083798 + layer.2.v_cache 0.00001872 0.04814524 + layer.3.k_cache 0.04020810 14.13465867 + layer.3.v_cache 0.00001850 0.05456211 + layer.4.k_cache 0.00063358 1.19532979 + layer.4.v_cache 0.00005294 0.09826342 + layer.4.output 10.86925024 512.18150510 + ------------------------------------------------------------------------------------- + TOTAL 4.48838751 235.13080043 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 27648 +BPFP 0.2420 bits/point +EBPFP 0.4840 equivalent bits/point +MSE 235.130800 +---------------------- -------------------------------------------------------- +Time: 2.941s Load: 0.005s, Pack+Encode: 1.673s, Decode+Unpack: 1.263s +---------------------- -------------------------------------------------------- +💾 Converting with 235.1308 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,232B, BPFP=0.2292 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,944B, BPFP=0.5476 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,948B, BPFP=0.5484 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,716B, BPFP=0.5052 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,412B, BPFP=0.4487 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,528B, BPFP=0.4702 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,396B, BPFP=0.4457 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,640B, BPFP=0.4911 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,828B, BPFP=0.5260 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,724B, BPFP=0.5067 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,416B, BPFP=0.0376 +⌛️ [2/4] FRONTEND: Frontend time: 1.585s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09514355 315.31656901 + layer.0.v_cache 0.00001423 0.04087478 + layer.1.k_cache 0.03283414 24.20808193 + layer.1.v_cache 0.00000553 0.01560296 + layer.2.k_cache 0.00489674 3.98186675 + layer.2.v_cache 0.00001883 0.04644003 + layer.3.k_cache 0.07199075 16.54286702 + layer.3.v_cache 0.00001932 0.05520795 + layer.4.k_cache 0.00061592 1.17416736 + layer.4.v_cache 0.00006816 0.09952634 + layer.4.output 0.17253765 646.27832696 + ------------------------------------------------------------------------------------- + TOTAL 0.08313946 287.37820546 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 26784 +BPFP 0.2931 bits/point +EBPFP 0.5861 equivalent bits/point +MSE 287.378205 +---------------------- -------------------------------------------------------- +Time: 2.733s Load: 0.004s, Pack+Encode: 1.585s, Decode+Unpack: 1.144s +---------------------- -------------------------------------------------------- +💾 Converting with 287.3782 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,484B, BPFP=0.2665 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,624B, BPFP=0.4713 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,764B, BPFP=0.4964 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,208B, BPFP=0.3966 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,396B, BPFP=0.4303 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,684B, BPFP=0.4820 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,476B, BPFP=0.4447 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,572B, BPFP=0.4619 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,748B, BPFP=0.4935 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,640B, BPFP=0.4741 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,412B, BPFP=0.0362 +⌛️ [2/4] FRONTEND: Frontend time: 1.681s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11660385 335.31939206 + layer.0.v_cache 0.00001381 0.04044855 + layer.1.k_cache 0.05175667 25.13211656 + layer.1.v_cache 0.00000569 0.01658010 + layer.2.k_cache 0.00227456 3.78358512 + layer.2.v_cache 0.00001743 0.04444057 + layer.3.k_cache 0.02781237 16.56440823 + layer.3.v_cache 0.00001883 0.05258235 + layer.4.k_cache 0.00061846 1.19433611 + layer.4.v_cache 0.00005344 0.10263792 + layer.4.output 0.17410791 623.92698071 + ------------------------------------------------------------------------------------- + TOTAL 0.08340767 279.39643485 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 26008 +BPFP 0.2748 bits/point +EBPFP 0.5495 equivalent bits/point +MSE 279.396435 +---------------------- -------------------------------------------------------- +Time: 2.885s Load: 0.004s, Pack+Encode: 1.681s, Decode+Unpack: 1.199s +---------------------- -------------------------------------------------------- +💾 Converting with 279.3964 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,224B, BPFP=0.2304 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,652B, BPFP=0.4992 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,032B, BPFP=0.5708 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,232B, BPFP=0.4202 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,512B, BPFP=0.4729 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,876B, BPFP=0.3532 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,132B, BPFP=0.4014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,300B, BPFP=0.4330 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,840B, BPFP=0.5346 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,640B, BPFP=0.4970 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,372B, BPFP=0.0369 +⌛️ [2/4] FRONTEND: Frontend time: 1.736s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.191s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212080 331.64434300 + layer.0.v_cache 0.00001360 0.04000950 + layer.1.k_cache 0.03288655 25.74645849 + layer.1.v_cache 0.00000536 0.01537474 + layer.2.k_cache 0.00973156 4.07730286 + layer.2.v_cache 0.00001762 0.04663392 + layer.3.k_cache 0.02906220 16.03406497 + layer.3.v_cache 0.00001891 0.05394134 + layer.4.k_cache 0.00063934 1.19796597 + layer.4.v_cache 0.00004866 0.09932139 + layer.4.output 0.18425958 654.05061317 + ------------------------------------------------------------------------------------- + TOTAL 0.08496245 291.60645343 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 24812 +BPFP 0.2748 bits/point +EBPFP 0.5495 equivalent bits/point +MSE 291.606453 +---------------------- -------------------------------------------------------- +Time: 2.931s Load: 0.004s, Pack+Encode: 1.736s, Decode+Unpack: 1.191s +---------------------- -------------------------------------------------------- +💾 Converting with 291.6065 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,244B, BPFP=0.2287 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,996B, BPFP=0.5507 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,988B, BPFP=0.5493 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,196B, BPFP=0.4037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,392B, BPFP=0.4397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,464B, BPFP=0.4529 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,136B, BPFP=0.3926 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,404B, BPFP=0.4419 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,868B, BPFP=0.5272 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,716B, BPFP=0.4993 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0368 +⌛️ [2/4] FRONTEND: Frontend time: 1.731s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.184s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10758395 319.76819853 + layer.0.v_cache 0.00001397 0.04031603 + layer.1.k_cache 0.03293324 24.95282054 + layer.1.v_cache 0.00000575 0.01655972 + layer.2.k_cache 0.00823421 3.82595179 + layer.2.v_cache 0.00001745 0.04513477 + layer.3.k_cache 0.04116619 16.30146484 + layer.3.v_cache 0.00001985 0.05486224 + layer.4.k_cache 0.00062779 1.22726476 + layer.4.v_cache 0.00005042 0.10095803 + layer.4.output 0.16954060 638.70210084 + ------------------------------------------------------------------------------------- + TOTAL 0.08102571 284.54401395 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 25804 +BPFP 0.2790 bits/point +EBPFP 0.5580 equivalent bits/point +MSE 284.544014 +---------------------- -------------------------------------------------------- +Time: 2.920s Load: 0.005s, Pack+Encode: 1.731s, Decode+Unpack: 1.184s +---------------------- -------------------------------------------------------- +💾 Converting with 284.5440 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,296B, BPFP=0.2250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,616B, BPFP=0.4542 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,696B, BPFP=0.4681 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,688B, BPFP=0.4667 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,592B, BPFP=0.4500 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,544B, BPFP=0.4417 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,980B, BPFP=0.3438 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,408B, BPFP=0.4181 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,568B, BPFP=0.4458 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,700B, BPFP=0.4688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,384B, BPFP=0.0343 +⌛️ [2/4] FRONTEND: Frontend time: 1.706s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.192s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12307767 322.18602431 + layer.0.v_cache 0.00001410 0.04115238 + layer.1.k_cache 0.03088827 24.53192817 + layer.1.v_cache 0.00000584 0.01625923 + layer.2.k_cache 0.00371391 3.92641127 + layer.2.v_cache 0.00001950 0.04879658 + layer.3.k_cache 0.03079298 16.60471056 + layer.3.v_cache 0.00001950 0.05423112 + layer.4.k_cache 0.00063014 1.16005342 + layer.4.v_cache 0.00005282 0.10057197 + layer.4.output 0.15767360 603.00947421 + ------------------------------------------------------------------------------------- + TOTAL 0.07605470 269.98449756 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 25472 +BPFP 0.2601 bits/point +EBPFP 0.5203 equivalent bits/point +MSE 269.984498 +---------------------- -------------------------------------------------------- +Time: 2.901s Load: 0.004s, Pack+Encode: 1.706s, Decode+Unpack: 1.192s +---------------------- -------------------------------------------------------- +💾 Converting with 269.9845 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,168B, BPFP=0.2226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,944B, BPFP=0.5610 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,100B, BPFP=0.5907 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,188B, BPFP=0.4169 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,452B, BPFP=0.4672 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,568B, BPFP=0.4893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,116B, BPFP=0.4032 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,808B, BPFP=0.5351 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,000B, BPFP=0.5716 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,860B, BPFP=0.5450 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,412B, BPFP=0.0384 +⌛️ [2/4] FRONTEND: Frontend time: 1.700s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.160s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10295349 310.26929306 + layer.0.v_cache 0.00001348 0.03902900 + layer.1.k_cache 0.03334363 25.93804187 + layer.1.v_cache 0.00000550 0.01582389 + layer.2.k_cache 0.01126997 3.97774226 + layer.2.v_cache 0.00001933 0.04637903 + layer.3.k_cache 0.09131574 17.33967107 + layer.3.v_cache 0.00001769 0.05099473 + layer.4.k_cache 0.00062275 1.20982566 + layer.4.v_cache 0.00006541 0.09949117 + layer.4.output 0.17512874 661.95121951 + ------------------------------------------------------------------------------------- + TOTAL 0.08620754 293.68498990 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 26616 +BPFP 0.2983 bits/point +EBPFP 0.5967 equivalent bits/point +MSE 293.684990 +---------------------- -------------------------------------------------------- +Time: 2.864s Load: 0.004s, Pack+Encode: 1.700s, Decode+Unpack: 1.160s +---------------------- -------------------------------------------------------- +💾 Converting with 293.6850 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,576B, BPFP=0.2830 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,488B, BPFP=0.4468 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,612B, BPFP=0.4691 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,320B, BPFP=0.4167 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,448B, BPFP=0.4397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,264B, BPFP=0.4066 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,412B, BPFP=0.4332 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,692B, BPFP=0.4835 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,488B, BPFP=0.4468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,880B, BPFP=0.5172 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,408B, BPFP=0.0361 +⌛️ [2/4] FRONTEND: Frontend time: 1.688s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.249s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12314748 354.85223599 + layer.0.v_cache 0.00001636 0.04008162 + layer.1.k_cache 0.01190477 25.43221871 + layer.1.v_cache 0.00000588 0.01625774 + layer.2.k_cache 0.01614088 4.02182042 + layer.2.v_cache 0.00001834 0.04740704 + layer.3.k_cache 0.05604688 16.45261708 + layer.3.v_cache 0.00001882 0.05564897 + layer.4.k_cache 0.00063774 1.19201932 + layer.4.v_cache 0.00005162 0.09877874 + layer.4.output 0.17348555 624.00533662 + ------------------------------------------------------------------------------------- + TOTAL 0.08366986 280.60273188 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 25588 +BPFP 0.2703 bits/point +EBPFP 0.5407 equivalent bits/point +MSE 280.602732 +---------------------- -------------------------------------------------------- +Time: 2.941s Load: 0.004s, Pack+Encode: 1.688s, Decode+Unpack: 1.249s +---------------------- -------------------------------------------------------- +💾 Converting with 280.6027 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,204B, BPFP=0.2323 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,848B, BPFP=0.5494 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,188B, BPFP=0.6150 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,192B, BPFP=0.4228 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,452B, BPFP=0.4730 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,032B, BPFP=0.3920 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,200B, BPFP=0.4244 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,052B, BPFP=0.3958 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,976B, BPFP=0.5741 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,616B, BPFP=0.5046 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,356B, BPFP=0.0374 +⌛️ [2/4] FRONTEND: Frontend time: 1.719s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.237s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11560643 312.13920235 + layer.0.v_cache 0.00001402 0.04160842 + layer.1.k_cache 0.03272506 26.35968244 + layer.1.v_cache 0.00000563 0.01608155 + layer.2.k_cache 0.00645547 4.07951751 + layer.2.v_cache 0.00001830 0.04918036 + layer.3.k_cache 0.02731901 16.88281250 + layer.3.v_cache 0.00001960 0.05703612 + layer.4.k_cache 0.00061302 1.19516782 + layer.4.v_cache 0.00004978 0.10160065 + layer.4.output 0.17251868 670.37235450 + ------------------------------------------------------------------------------------- + TOTAL 0.08179159 297.26637478 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 25116 +BPFP 0.2850 bits/point +EBPFP 0.5700 equivalent bits/point +MSE 297.266375 +---------------------- -------------------------------------------------------- +Time: 2.960s Load: 0.005s, Pack+Encode: 1.719s, Decode+Unpack: 1.237s +---------------------- -------------------------------------------------------- +💾 Converting with 297.2664 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,288B, BPFP=0.2287 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,776B, BPFP=0.4929 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,700B, BPFP=0.4794 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,584B, BPFP=0.4588 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,424B, BPFP=0.4304 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,680B, BPFP=0.4759 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,084B, BPFP=0.3700 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,596B, BPFP=0.4609 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,672B, BPFP=0.4744 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,708B, BPFP=0.4808 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,468B, BPFP=0.0372 +⌛️ [2/4] FRONTEND: Frontend time: 1.665s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.391s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11690623 335.10395952 + layer.0.v_cache 0.00001418 0.04148493 + layer.1.k_cache 0.03305150 25.61567272 + layer.1.v_cache 0.00000545 0.01640076 + layer.2.k_cache 0.01095151 3.90091254 + layer.2.v_cache 0.00001854 0.04584217 + layer.3.k_cache 0.04293454 16.98166171 + layer.3.v_cache 0.00002007 0.05620664 + layer.4.k_cache 0.00062489 1.15850795 + layer.4.v_cache 0.00005451 0.09834873 + layer.4.output 0.16672201 616.91193182 + ------------------------------------------------------------------------------------- + TOTAL 0.08068444 276.55308943 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 25980 +BPFP 0.2713 bits/point +EBPFP 0.5427 equivalent bits/point +MSE 276.553089 +---------------------- -------------------------------------------------------- +Time: 3.060s Load: 0.004s, Pack+Encode: 1.665s, Decode+Unpack: 1.391s +---------------------- -------------------------------------------------------- +💾 Converting with 276.5531 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,612B, BPFP=0.2768 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,536B, BPFP=0.4354 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,676B, BPFP=0.4595 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,312B, BPFP=0.3970 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,380B, BPFP=0.4087 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,356B, BPFP=0.4045 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,048B, BPFP=0.3516 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,612B, BPFP=0.4485 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,712B, BPFP=0.4657 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,644B, BPFP=0.4540 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,288B, BPFP=0.0316 +⌛️ [2/4] FRONTEND: Frontend time: 1.722s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.190s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11067300 349.92247596 + layer.0.v_cache 0.00001409 0.04052865 + layer.1.k_cache 0.05152405 25.61350661 + layer.1.v_cache 0.00000569 0.01609416 + layer.2.k_cache 0.01362713 3.82993392 + layer.2.v_cache 0.00001889 0.04817707 + layer.3.k_cache 0.08227781 16.32869913 + layer.3.v_cache 0.00001889 0.05418361 + layer.4.k_cache 0.00062414 1.19803921 + layer.4.v_cache 0.00005257 0.09803145 + layer.4.output 0.15746453 596.64197410 + ------------------------------------------------------------------------------------- + TOTAL 0.08006400 269.03785226 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 25176 +BPFP 0.2543 bits/point +EBPFP 0.5086 equivalent bits/point +MSE 269.037852 +---------------------- -------------------------------------------------------- +Time: 2.917s Load: 0.004s, Pack+Encode: 1.722s, Decode+Unpack: 1.190s +---------------------- -------------------------------------------------------- +💾 Converting with 269.0379 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 116, 128) +Output shape: (1, 116, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.output: torch.Size([1, 116, 3584]) -> torch.Size([1, 1, 116, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,752B, BPFP=0.2360 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,960B, BPFP=0.3987 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,088B, BPFP=0.4159 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,748B, BPFP=0.3702 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,528B, BPFP=0.3405 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,600B, BPFP=0.3502 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,468B, BPFP=0.3324 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,796B, BPFP=0.3766 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,984B, BPFP=0.4019 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,768B, BPFP=0.3728 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,776B, BPFP=0.0342 +⌛️ [2/4] FRONTEND: Frontend time: 1.707s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.114s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16314862 381.51882408 + layer.0.v_cache 0.00001373 0.04041816 + layer.1.k_cache 0.05481785 24.69897671 + layer.1.v_cache 0.00000557 0.01620894 + layer.2.k_cache 0.00968372 3.64556069 + layer.2.v_cache 0.00001954 0.04930674 + layer.3.k_cache 0.02418142 14.81266416 + layer.3.v_cache 0.00001860 0.05275354 + layer.4.k_cache 0.00062333 1.22848484 + layer.4.v_cache 0.00005467 0.10110192 + layer.4.output 9.83417319 463.61868842 + ------------------------------------------------------------------------------------- + TOTAL 4.06422232 215.97030110 + (elements=1,009,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1009664 +Total Bytes 28468 +BPFP 0.2256 bits/point +EBPFP 0.4511 equivalent bits/point +MSE 215.970301 +---------------------- -------------------------------------------------------- +Time: 2.825s Load: 0.004s, Pack+Encode: 1.707s, Decode+Unpack: 1.114s +---------------------- -------------------------------------------------------- +💾 Converting with 215.9703 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,224B, BPFP=0.2452 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,820B, BPFP=0.5649 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,904B, BPFP=0.5817 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,996B, BPFP=0.3998 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,388B, BPFP=0.4784 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,840B, BPFP=0.3686 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,116B, BPFP=0.4239 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,828B, BPFP=0.5665 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,956B, BPFP=0.5921 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,812B, BPFP=0.5633 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,448B, BPFP=0.0414 +⌛️ [2/4] FRONTEND: Frontend time: 1.699s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11749315 334.45472756 + layer.0.v_cache 0.00001389 0.04040134 + layer.1.k_cache 0.03399578 25.22715094 + layer.1.v_cache 0.00000555 0.01656451 + layer.2.k_cache 0.00854440 4.12251829 + layer.2.v_cache 0.00001777 0.04608536 + layer.3.k_cache 0.02740619 16.45648350 + layer.3.v_cache 0.00001935 0.05363112 + layer.4.k_cache 0.00063531 1.24734497 + layer.4.v_cache 0.00005166 0.10024486 + layer.4.output 0.18490290 696.09930174 + ------------------------------------------------------------------------------------- + TOTAL 0.08720608 309.08589792 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 25332 +BPFP 0.2985 bits/point +EBPFP 0.5970 equivalent bits/point +MSE 309.085898 +---------------------- -------------------------------------------------------- +Time: 3.059s Load: 0.003s, Pack+Encode: 1.699s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 309.0859 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,336B, BPFP=0.2221 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,348B, BPFP=0.3903 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,524B, BPFP=0.4195 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,344B, BPFP=0.3896 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,296B, BPFP=0.3816 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,208B, BPFP=0.3670 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,252B, BPFP=0.3743 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,336B, BPFP=0.3883 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,656B, BPFP=0.4415 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,584B, BPFP=0.4295 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,296B, BPFP=0.0308 +⌛️ [2/4] FRONTEND: Frontend time: 1.854s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11492289 334.91994265 + layer.0.v_cache 0.00001440 0.04005466 + layer.1.k_cache 0.01112092 24.93576504 + layer.1.v_cache 0.00000555 0.01559772 + layer.2.k_cache 0.01096075 3.68822463 + layer.2.v_cache 0.00001882 0.04797003 + layer.3.k_cache 0.03781782 15.49919486 + layer.3.v_cache 0.00001905 0.05459706 + layer.4.k_cache 0.00062716 1.14768300 + layer.4.v_cache 0.00006343 0.09490003 + layer.4.output 0.15224552 577.65021847 + ------------------------------------------------------------------------------------- + TOTAL 0.07301703 260.23502700 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 24180 +BPFP 0.2364 bits/point +EBPFP 0.4729 equivalent bits/point +MSE 260.235027 +---------------------- -------------------------------------------------------- +Time: 3.062s Load: 0.005s, Pack+Encode: 1.854s, Decode+Unpack: 1.203s +---------------------- -------------------------------------------------------- +💾 Converting with 260.2350 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,204B, BPFP=0.2443 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,688B, BPFP=0.5455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,892B, BPFP=0.5869 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,000B, BPFP=0.4058 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,428B, BPFP=0.4927 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,852B, BPFP=0.3758 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,100B, BPFP=0.4261 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,092B, BPFP=0.4245 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,984B, BPFP=0.6055 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,888B, BPFP=0.5860 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,336B, BPFP=0.0387 +⌛️ [2/4] FRONTEND: Frontend time: 1.745s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.189s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10405910 341.58205662 + layer.0.v_cache 0.00001453 0.04003772 + layer.1.k_cache 0.03397077 25.44632711 + layer.1.v_cache 0.00000547 0.01662521 + layer.2.k_cache 0.00568739 4.06131814 + layer.2.v_cache 0.00001916 0.05057688 + layer.3.k_cache 0.09581309 16.29704146 + layer.3.v_cache 0.00001862 0.05442781 + layer.4.k_cache 0.00061896 1.16570193 + layer.4.v_cache 0.00005653 0.09860169 + layer.4.output 0.19123089 705.12244898 + ------------------------------------------------------------------------------------- + TOTAL 0.09287528 313.21587397 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 24464 +BPFP 0.2920 bits/point +EBPFP 0.5840 equivalent bits/point +MSE 313.215874 +---------------------- -------------------------------------------------------- +Time: 2.937s Load: 0.004s, Pack+Encode: 1.745s, Decode+Unpack: 1.189s +---------------------- -------------------------------------------------------- +💾 Converting with 313.2159 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,260B, BPFP=0.2316 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,336B, BPFP=0.4294 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,888B, BPFP=0.5309 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,120B, BPFP=0.3897 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,444B, BPFP=0.4493 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,584B, BPFP=0.4750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,120B, BPFP=0.3897 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,192B, BPFP=0.4029 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,948B, BPFP=0.5419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,764B, BPFP=0.5081 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,396B, BPFP=0.0367 +⌛️ [2/4] FRONTEND: Frontend time: 1.744s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.183s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09384689 328.58690257 + layer.0.v_cache 0.00001364 0.03968805 + layer.1.k_cache 0.01313462 24.85412454 + layer.1.v_cache 0.00000592 0.01632516 + layer.2.k_cache 0.00813576 3.87699980 + layer.2.v_cache 0.00001762 0.04563350 + layer.3.k_cache 0.02628646 16.19026884 + layer.3.v_cache 0.00002044 0.05782053 + layer.4.k_cache 0.00061596 1.24731374 + layer.4.v_cache 0.00005334 0.10501871 + layer.4.output 0.17343949 638.66302521 + ------------------------------------------------------------------------------------- + TOTAL 0.07977689 285.03889835 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 25052 +BPFP 0.2709 bits/point +EBPFP 0.5418 equivalent bits/point +MSE 285.038898 +---------------------- -------------------------------------------------------- +Time: 2.930s Load: 0.004s, Pack+Encode: 1.744s, Decode+Unpack: 1.183s +---------------------- -------------------------------------------------------- +💾 Converting with 285.0389 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 126, 128) +Output shape: (1, 126, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.output: torch.Size([1, 126, 3584]) -> torch.Size([1, 1, 126, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,148B, BPFP=0.2664 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,472B, BPFP=0.3065 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,568B, BPFP=0.3185 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,416B, BPFP=0.2996 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,056B, BPFP=0.2550 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,360B, BPFP=0.2927 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,788B, BPFP=0.2217 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,684B, BPFP=0.3328 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,436B, BPFP=0.3021 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,316B, BPFP=0.2872 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,028B, BPFP=0.0359 +⌛️ [2/4] FRONTEND: Frontend time: 1.695s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.150s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13141363 420.22832961 + layer.0.v_cache 0.00001429 0.04051268 + layer.1.k_cache 0.06496696 24.38500589 + layer.1.v_cache 0.00000583 0.01678822 + layer.2.k_cache 0.01057190 3.14652652 + layer.2.v_cache 0.00002020 0.04793636 + layer.3.k_cache 0.05250616 13.93926905 + layer.3.v_cache 0.00001853 0.05363189 + layer.4.k_cache 0.00062861 1.19876583 + layer.4.v_cache 0.00004930 0.09414782 + layer.4.output 9.05470577 426.67176871 + ------------------------------------------------------------------------------------- + TOTAL 3.74371387 202.93254675 + (elements=1,096,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1096704 +Total Bytes 25272 +BPFP 0.1843 bits/point +EBPFP 0.3687 equivalent bits/point +MSE 202.932547 +---------------------- -------------------------------------------------------- +Time: 2.850s Load: 0.004s, Pack+Encode: 1.695s, Decode+Unpack: 1.150s +---------------------- -------------------------------------------------------- +💾 Converting with 202.9325 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,476B, BPFP=0.2651 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,572B, BPFP=0.4619 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,724B, BPFP=0.4892 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,180B, BPFP=0.3915 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,380B, BPFP=0.4274 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,624B, BPFP=0.4713 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,144B, BPFP=0.3851 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,532B, BPFP=0.4547 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,644B, BPFP=0.4749 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,896B, BPFP=0.5201 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,404B, BPFP=0.0360 +⌛️ [2/4] FRONTEND: Frontend time: 1.691s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.152s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13059476 341.88811063 + layer.0.v_cache 0.00001380 0.03977048 + layer.1.k_cache 0.03034586 24.47292565 + layer.1.v_cache 0.00000531 0.01588600 + layer.2.k_cache 0.00374181 3.88503046 + layer.2.v_cache 0.00001886 0.04517914 + layer.3.k_cache 0.04454902 16.49671813 + layer.3.v_cache 0.00001968 0.05449041 + layer.4.k_cache 0.00060938 1.17353523 + layer.4.v_cache 0.00007316 0.09998155 + layer.4.output 0.16485390 624.04495074 + ------------------------------------------------------------------------------------- + TOTAL 0.08023229 279.79331076 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 25576 +BPFP 0.2702 bits/point +EBPFP 0.5404 equivalent bits/point +MSE 279.793311 +---------------------- -------------------------------------------------------- +Time: 2.848s Load: 0.005s, Pack+Encode: 1.691s, Decode+Unpack: 1.152s +---------------------- -------------------------------------------------------- +💾 Converting with 279.7933 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.2255 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,424B, BPFP=0.4117 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,744B, BPFP=0.4660 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,396B, BPFP=0.4069 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,152B, BPFP=0.3655 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,372B, BPFP=0.4029 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,012B, BPFP=0.3417 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,392B, BPFP=0.4062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,576B, BPFP=0.4375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,628B, BPFP=0.4463 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,280B, BPFP=0.0311 +⌛️ [2/4] FRONTEND: Frontend time: 1.636s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15436328 349.52407439 + layer.0.v_cache 0.00001586 0.04015397 + layer.1.k_cache 0.01080767 25.14285676 + layer.1.v_cache 0.00000575 0.01627514 + layer.2.k_cache 0.00786338 3.72335948 + layer.2.v_cache 0.00001961 0.04554984 + layer.3.k_cache 0.05652578 16.51600183 + layer.3.v_cache 0.00002042 0.05550869 + layer.4.k_cache 0.00062828 1.15972204 + layer.4.v_cache 0.00005674 0.10569328 + layer.4.output 0.15192759 590.22840644 + ------------------------------------------------------------------------------------- + TOTAL 0.07610588 266.34870827 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 24304 +BPFP 0.2428 bits/point +EBPFP 0.4856 equivalent bits/point +MSE 266.348708 +---------------------- -------------------------------------------------------- +Time: 2.841s Load: 0.004s, Pack+Encode: 1.636s, Decode+Unpack: 1.201s +---------------------- -------------------------------------------------------- +💾 Converting with 266.3487 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,208B, BPFP=0.2420 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,384B, BPFP=0.4776 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,988B, BPFP=0.5986 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,980B, BPFP=0.3966 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,628B, BPFP=0.5264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,752B, BPFP=0.3510 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,040B, BPFP=0.4087 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,340B, BPFP=0.4688 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,792B, BPFP=0.5593 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,768B, BPFP=0.5545 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,468B, BPFP=0.0420 +⌛️ [2/4] FRONTEND: Frontend time: 1.684s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.105s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10254851 342.31009615 + layer.0.v_cache 0.00001399 0.04113419 + layer.1.k_cache 0.05476840 25.34934958 + layer.1.v_cache 0.00000542 0.01615747 + layer.2.k_cache 0.00244555 4.09485568 + layer.2.v_cache 0.00001844 0.05116430 + layer.3.k_cache 0.02811479 16.59567808 + layer.3.v_cache 0.00001905 0.05389799 + layer.4.k_cache 0.00060111 1.15170542 + layer.4.v_cache 0.00005313 0.10357766 + layer.4.output 0.19635826 696.00972985 + ------------------------------------------------------------------------------------- + TOTAL 0.09194683 309.51974856 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 24348 +BPFP 0.2869 bits/point +EBPFP 0.5738 equivalent bits/point +MSE 309.519749 +---------------------- -------------------------------------------------------- +Time: 2.793s Load: 0.004s, Pack+Encode: 1.684s, Decode+Unpack: 1.105s +---------------------- -------------------------------------------------------- +💾 Converting with 309.5197 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,168B, BPFP=0.2226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,692B, BPFP=0.5130 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,148B, BPFP=0.5998 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,308B, BPFP=0.4398 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,536B, BPFP=0.4832 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,412B, BPFP=0.4596 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,328B, BPFP=0.4436 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,748B, BPFP=0.5236 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,048B, BPFP=0.5808 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,956B, BPFP=0.5633 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0381 +⌛️ [2/4] FRONTEND: Frontend time: 1.716s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.178s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11987135 313.22256098 + layer.0.v_cache 0.00001390 0.04072364 + layer.1.k_cache 0.03139273 26.03267912 + layer.1.v_cache 0.00000562 0.01577093 + layer.2.k_cache 0.00816743 4.18004850 + layer.2.v_cache 0.00001857 0.04498309 + layer.3.k_cache 0.02640573 15.64377799 + layer.3.v_cache 0.00001968 0.05387901 + layer.4.k_cache 0.00062091 1.22456062 + layer.4.v_cache 0.00004972 0.09813648 + layer.4.output 0.16983549 661.93417901 + ------------------------------------------------------------------------------------- + TOTAL 0.08090671 293.77037491 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 26744 +BPFP 0.2998 bits/point +EBPFP 0.5995 equivalent bits/point +MSE 293.770375 +---------------------- -------------------------------------------------------- +Time: 2.898s Load: 0.004s, Pack+Encode: 1.716s, Decode+Unpack: 1.178s +---------------------- -------------------------------------------------------- +💾 Converting with 293.7704 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,276B, BPFP=0.2292 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,452B, BPFP=0.4404 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,512B, BPFP=0.4511 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,684B, BPFP=0.4820 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,436B, BPFP=0.4375 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,564B, BPFP=0.4605 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,996B, BPFP=0.3585 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,200B, BPFP=0.3951 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,680B, BPFP=0.4813 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,904B, BPFP=0.5216 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,404B, BPFP=0.0360 +⌛️ [2/4] FRONTEND: Frontend time: 1.737s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14089697 326.23657507 + layer.0.v_cache 0.00001393 0.04059970 + layer.1.k_cache 0.01618495 25.59393802 + layer.1.v_cache 0.00000582 0.01608453 + layer.2.k_cache 0.00661465 3.99366918 + layer.2.v_cache 0.00002083 0.04746174 + layer.3.k_cache 0.07087009 15.88691799 + layer.3.v_cache 0.00001822 0.05745469 + layer.4.k_cache 0.00061891 1.26258552 + layer.4.v_cache 0.00005311 0.10244694 + layer.4.output 0.16425455 624.11879105 + ------------------------------------------------------------------------------------- + TOTAL 0.08147525 278.94525122 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 25108 +BPFP 0.2653 bits/point +EBPFP 0.5305 equivalent bits/point +MSE 278.945251 +---------------------- -------------------------------------------------------- +Time: 2.943s Load: 0.006s, Pack+Encode: 1.737s, Decode+Unpack: 1.200s +---------------------- -------------------------------------------------------- +💾 Converting with 278.9453 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,288B, BPFP=0.2261 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,608B, BPFP=0.4579 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,628B, BPFP=0.4614 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,592B, BPFP=0.4551 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,444B, BPFP=0.4291 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,420B, BPFP=0.4249 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,076B, BPFP=0.3645 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,520B, BPFP=0.4424 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,740B, BPFP=0.4810 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,772B, BPFP=0.4867 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0351 +⌛️ [2/4] FRONTEND: Frontend time: 1.959s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12786581 337.44149403 + layer.0.v_cache 0.00001426 0.04110161 + layer.1.k_cache 0.03183105 25.25395837 + layer.1.v_cache 0.00000593 0.01686753 + layer.2.k_cache 0.00241907 4.02154815 + layer.2.v_cache 0.00001785 0.04517736 + layer.3.k_cache 0.02946699 16.49363863 + layer.3.v_cache 0.00001957 0.05779734 + layer.4.k_cache 0.00063747 1.19897024 + layer.4.v_cache 0.00005003 0.10125477 + layer.4.output 0.16568538 609.87545144 + ------------------------------------------------------------------------------------- + TOTAL 0.07953681 273.75293930 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 25488 +BPFP 0.2632 bits/point +EBPFP 0.5264 equivalent bits/point +MSE 273.752939 +---------------------- -------------------------------------------------------- +Time: 3.350s Load: 0.004s, Pack+Encode: 1.959s, Decode+Unpack: 1.387s +---------------------- -------------------------------------------------------- +💾 Converting with 273.7529 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,272B, BPFP=0.2548 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,708B, BPFP=0.5425 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,928B, BPFP=0.5865 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,396B, BPFP=0.4800 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,476B, BPFP=0.4960 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,400B, BPFP=0.4808 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,076B, BPFP=0.4159 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,200B, BPFP=0.4407 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,848B, BPFP=0.5705 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,916B, BPFP=0.5841 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,452B, BPFP=0.0416 +⌛️ [2/4] FRONTEND: Frontend time: 2.027s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.288s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12225728 331.69621394 + layer.0.v_cache 0.00001337 0.04122282 + layer.1.k_cache 0.03480693 26.14238406 + layer.1.v_cache 0.00000545 0.01639983 + layer.2.k_cache 0.00853881 4.03104890 + layer.2.v_cache 0.00001924 0.04825947 + layer.3.k_cache 0.06130744 16.42966872 + layer.3.v_cache 0.00001928 0.06104704 + layer.4.k_cache 0.00058936 1.15725884 + layer.4.v_cache 0.00007197 0.10639663 + layer.4.output 0.18657821 696.06250000 + ------------------------------------------------------------------------------------- + TOTAL 0.09021627 308.95102354 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 25672 +BPFP 0.3025 bits/point +EBPFP 0.6050 equivalent bits/point +MSE 308.951024 +---------------------- -------------------------------------------------------- +Time: 3.318s Load: 0.003s, Pack+Encode: 2.027s, Decode+Unpack: 1.288s +---------------------- -------------------------------------------------------- +💾 Converting with 308.9510 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,248B, BPFP=0.2532 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,732B, BPFP=0.5544 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,088B, BPFP=0.6266 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,340B, BPFP=0.4748 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,288B, BPFP=0.4643 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,856B, BPFP=0.3766 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,944B, BPFP=0.3945 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,892B, BPFP=0.3839 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,036B, BPFP=0.6161 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,592B, BPFP=0.5260 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,356B, BPFP=0.0393 +⌛️ [2/4] FRONTEND: Frontend time: 1.705s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.185s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582910 336.30073052 + layer.0.v_cache 0.00001386 0.03988482 + layer.1.k_cache 0.03408930 25.96744845 + layer.1.v_cache 0.00000538 0.01552666 + layer.2.k_cache 0.00228146 4.12826340 + layer.2.v_cache 0.00001859 0.04940092 + layer.3.k_cache 0.06276476 16.54813375 + layer.3.v_cache 0.00001913 0.05580436 + layer.4.k_cache 0.00060520 1.18609322 + layer.4.v_cache 0.00005125 0.10987797 + layer.4.output 0.19356344 705.11439007 + ------------------------------------------------------------------------------------- + TOTAL 0.09238954 312.95305262 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 24372 +BPFP 0.2909 bits/point +EBPFP 0.5818 equivalent bits/point +MSE 312.953053 +---------------------- -------------------------------------------------------- +Time: 2.894s Load: 0.004s, Pack+Encode: 1.705s, Decode+Unpack: 1.185s +---------------------- -------------------------------------------------------- +💾 Converting with 312.9531 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,236B, BPFP=0.2327 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,596B, BPFP=0.4887 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,076B, BPFP=0.5791 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,128B, BPFP=0.4006 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,372B, BPFP=0.4465 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,800B, BPFP=0.5271 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,164B, BPFP=0.4074 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,424B, BPFP=0.4563 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,004B, BPFP=0.5655 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,876B, BPFP=0.5414 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,380B, BPFP=0.0371 +⌛️ [2/4] FRONTEND: Frontend time: 1.689s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.151s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09700135 337.24687029 + layer.0.v_cache 0.00001355 0.04086774 + layer.1.k_cache 0.01103399 25.59804452 + layer.1.v_cache 0.00000598 0.01668890 + layer.2.k_cache 0.00794396 3.82746538 + layer.2.v_cache 0.00001935 0.04751353 + layer.3.k_cache 0.04207947 16.71171993 + layer.3.v_cache 0.00002010 0.05744639 + layer.4.k_cache 0.00062246 1.21051329 + layer.4.v_cache 0.00005226 0.10281641 + layer.4.output 0.17535226 654.14850473 + ------------------------------------------------------------------------------------- + TOTAL 0.08154460 291.99408703 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 26056 +BPFP 0.2885 bits/point +EBPFP 0.5771 equivalent bits/point +MSE 291.994087 +---------------------- -------------------------------------------------------- +Time: 2.843s Load: 0.004s, Pack+Encode: 1.689s, Decode+Unpack: 1.151s +---------------------- -------------------------------------------------------- +💾 Converting with 291.9941 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,264B, BPFP=0.2324 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,556B, BPFP=0.4699 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,976B, BPFP=0.5471 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,244B, BPFP=0.4125 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,656B, BPFP=0.4882 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,544B, BPFP=0.4676 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,532B, BPFP=0.4654 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,500B, BPFP=0.4596 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,728B, BPFP=0.5015 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,784B, BPFP=0.5118 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,392B, BPFP=0.0366 +⌛️ [2/4] FRONTEND: Frontend time: 1.801s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.171s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11944316 323.02513787 + layer.0.v_cache 0.00001525 0.03936966 + layer.1.k_cache 0.01376067 25.08731330 + layer.1.v_cache 0.00000561 0.01570132 + layer.2.k_cache 0.00522648 4.35206478 + layer.2.v_cache 0.00001966 0.04765133 + layer.3.k_cache 0.04600189 16.35436724 + layer.3.v_cache 0.00001912 0.05583483 + layer.4.k_cache 0.00062079 1.20115330 + layer.4.v_cache 0.00005114 0.10031181 + layer.4.output 0.17175110 638.71964286 + ------------------------------------------------------------------------------------- + TOTAL 0.08161303 284.78331797 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 26176 +BPFP 0.2830 bits/point +EBPFP 0.5661 equivalent bits/point +MSE 284.783318 +---------------------- -------------------------------------------------------- +Time: 2.976s Load: 0.004s, Pack+Encode: 1.801s, Decode+Unpack: 1.171s +---------------------- -------------------------------------------------------- +💾 Converting with 284.7833 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,276B, BPFP=0.2556 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,668B, BPFP=0.5345 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,860B, BPFP=0.5729 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,872B, BPFP=0.3750 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,404B, BPFP=0.4816 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,012B, BPFP=0.4030 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,952B, BPFP=0.3910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,364B, BPFP=0.4736 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,784B, BPFP=0.5577 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,724B, BPFP=0.5457 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,420B, BPFP=0.0406 +⌛️ [2/4] FRONTEND: Frontend time: 1.716s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120220 345.58168069 + layer.0.v_cache 0.00001368 0.03920817 + layer.1.k_cache 0.03273370 25.82927684 + layer.1.v_cache 0.00000573 0.01600879 + layer.2.k_cache 0.00892877 4.01685901 + layer.2.v_cache 0.00001953 0.04905958 + layer.3.k_cache 0.02974504 17.07262871 + layer.3.v_cache 0.00001879 0.05400583 + layer.4.k_cache 0.00063874 1.23684986 + layer.4.v_cache 0.00005200 0.10430631 + layer.4.output 0.17853031 696.22035256 + ------------------------------------------------------------------------------------- + TOTAL 0.08429825 309.85543245 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 24336 +BPFP 0.2868 bits/point +EBPFP 0.5735 equivalent bits/point +MSE 309.855432 +---------------------- -------------------------------------------------------- +Time: 2.991s Load: 0.003s, Pack+Encode: 1.716s, Decode+Unpack: 1.272s +---------------------- -------------------------------------------------------- +💾 Converting with 309.8554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,220B, BPFP=0.2297 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,724B, BPFP=0.5128 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,172B, BPFP=0.5971 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,424B, BPFP=0.4563 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,552B, BPFP=0.4804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,188B, BPFP=0.4119 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,548B, BPFP=0.4797 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,192B, BPFP=0.4127 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,884B, BPFP=0.5429 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,920B, BPFP=0.5497 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,376B, BPFP=0.0370 +⌛️ [2/4] FRONTEND: Frontend time: 1.765s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09802409 320.63445971 + layer.0.v_cache 0.00001402 0.04181485 + layer.1.k_cache 0.03453328 25.24683794 + layer.1.v_cache 0.00000594 0.01665175 + layer.2.k_cache 0.01027090 3.95728385 + layer.2.v_cache 0.00001933 0.04816021 + layer.3.k_cache 0.02613793 17.00317089 + layer.3.v_cache 0.00001931 0.05698122 + layer.4.k_cache 0.00063818 1.21231336 + layer.4.v_cache 0.00005393 0.10413053 + layer.4.output 0.18084440 654.10402324 + ------------------------------------------------------------------------------------- + TOTAL 0.08444869 291.00293923 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 26200 +BPFP 0.2901 bits/point +EBPFP 0.5803 equivalent bits/point +MSE 291.002939 +---------------------- -------------------------------------------------------- +Time: 2.970s Load: 0.004s, Pack+Encode: 1.765s, Decode+Unpack: 1.201s +---------------------- -------------------------------------------------------- +💾 Converting with 291.0029 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,472B, BPFP=0.2233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,608B, BPFP=0.3956 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,164B, BPFP=0.4800 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,724B, BPFP=0.4132 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,616B, BPFP=0.3968 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,428B, BPFP=0.3683 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,492B, BPFP=0.3780 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,600B, BPFP=0.3944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,156B, BPFP=0.4788 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,744B, BPFP=0.4163 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,488B, BPFP=0.0322 +⌛️ [2/4] FRONTEND: Frontend time: 1.745s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13030120 351.95885164 + layer.0.v_cache 0.00001385 0.03929067 + layer.1.k_cache 0.02890817 24.71634178 + layer.1.v_cache 0.00000611 0.01695973 + layer.2.k_cache 0.00721959 3.26072204 + layer.2.v_cache 0.00001830 0.04686958 + layer.3.k_cache 0.02195997 14.92415049 + layer.3.v_cache 0.00001903 0.05237209 + layer.4.k_cache 0.00063766 1.17715350 + layer.4.v_cache 0.00005061 0.09732417 + layer.4.output 11.07892277 522.16535194 + ------------------------------------------------------------------------------------- + TOTAL 4.57303494 238.32044113 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 27492 +BPFP 0.2453 bits/point +EBPFP 0.4906 equivalent bits/point +MSE 238.320441 +---------------------- -------------------------------------------------------- +Time: 2.967s Load: 0.005s, Pack+Encode: 1.745s, Decode+Unpack: 1.217s +---------------------- -------------------------------------------------------- +💾 Converting with 238.3204 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,248B, BPFP=0.2349 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,296B, BPFP=0.4322 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,052B, BPFP=0.5745 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,236B, BPFP=0.4209 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,616B, BPFP=0.4925 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,496B, BPFP=0.4699 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,112B, BPFP=0.3976 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,264B, BPFP=0.4262 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,940B, BPFP=0.5535 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,916B, BPFP=0.5489 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,380B, BPFP=0.0371 +⌛️ [2/4] FRONTEND: Frontend time: 1.678s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.239s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12633393 331.51524849 + layer.0.v_cache 0.00001372 0.04059049 + layer.1.k_cache 0.01411889 25.76079513 + layer.1.v_cache 0.00000549 0.01628628 + layer.2.k_cache 0.00833277 3.75983696 + layer.2.v_cache 0.00001851 0.04555388 + layer.3.k_cache 0.05797521 16.09193365 + layer.3.v_cache 0.00001879 0.05357626 + layer.4.k_cache 0.00064672 1.24487066 + layer.4.v_cache 0.00005349 0.10375361 + layer.4.output 0.17447430 654.11268287 + ------------------------------------------------------------------------------------- + TOTAL 0.08404927 291.61301327 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 25556 +BPFP 0.2830 bits/point +EBPFP 0.5660 equivalent bits/point +MSE 291.613013 +---------------------- -------------------------------------------------------- +Time: 2.921s Load: 0.004s, Pack+Encode: 1.678s, Decode+Unpack: 1.239s +---------------------- -------------------------------------------------------- +💾 Converting with 291.6130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,048B, BPFP=0.2339 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,944B, BPFP=0.4339 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,416B, BPFP=0.5393 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,748B, BPFP=0.3902 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,360B, BPFP=0.5268 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,536B, BPFP=0.3429 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,616B, BPFP=0.3607 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,604B, BPFP=0.3580 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,148B, BPFP=0.4795 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,516B, BPFP=0.5616 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,352B, BPFP=0.0431 +⌛️ [2/4] FRONTEND: Frontend time: 1.702s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.188s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08086312 324.59185268 + layer.0.v_cache 0.00001331 0.04195726 + layer.1.k_cache 0.01551799 26.38672224 + layer.1.v_cache 0.00000513 0.01644723 + layer.2.k_cache 0.00232868 3.96669268 + layer.2.v_cache 0.00002262 0.05215634 + layer.3.k_cache 0.10623371 16.71844482 + layer.3.v_cache 0.00001836 0.05720034 + layer.4.k_cache 0.00061897 1.28389653 + layer.4.v_cache 0.00005922 0.10732754 + layer.4.output 0.20964142 775.47710459 + ------------------------------------------------------------------------------------- + TOTAL 0.09842183 341.26837822 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 20288 +BPFP 0.2664 bits/point +EBPFP 0.5328 equivalent bits/point +MSE 341.268378 +---------------------- -------------------------------------------------------- +Time: 2.893s Load: 0.003s, Pack+Encode: 1.702s, Decode+Unpack: 1.188s +---------------------- -------------------------------------------------------- +💾 Converting with 341.2684 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,336B, BPFP=0.2269 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,488B, BPFP=0.4226 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,600B, BPFP=0.4416 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,344B, BPFP=0.3981 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,312B, BPFP=0.3927 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,432B, BPFP=0.4130 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,044B, BPFP=0.3471 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,376B, BPFP=0.4035 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,424B, BPFP=0.4117 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,608B, BPFP=0.4429 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,304B, BPFP=0.0316 +⌛️ [2/4] FRONTEND: Frontend time: 1.701s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.193s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13670269 335.75144361 + layer.0.v_cache 0.00001568 0.04031686 + layer.1.k_cache 0.03139175 24.56341287 + layer.1.v_cache 0.00000602 0.01643305 + layer.2.k_cache 0.00629326 3.66560762 + layer.2.v_cache 0.00001871 0.04424202 + layer.3.k_cache 0.03837063 16.32703167 + layer.3.v_cache 0.00001860 0.05121888 + layer.4.k_cache 0.00062597 1.17600599 + layer.4.v_cache 0.00005354 0.09576437 + layer.4.output 0.15974679 590.04692352 + ------------------------------------------------------------------------------------- + TOTAL 0.07833673 265.41529068 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 24268 +BPFP 0.2424 bits/point +EBPFP 0.4849 equivalent bits/point +MSE 265.415291 +---------------------- -------------------------------------------------------- +Time: 2.898s Load: 0.003s, Pack+Encode: 1.701s, Decode+Unpack: 1.193s +---------------------- -------------------------------------------------------- +💾 Converting with 265.4153 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,732B, BPFP=0.2529 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,816B, BPFP=0.4112 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,152B, BPFP=0.4603 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,584B, BPFP=0.3773 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,656B, BPFP=0.3879 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,584B, BPFP=0.3773 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,548B, BPFP=0.3721 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,772B, BPFP=0.4048 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,000B, BPFP=0.4381 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,876B, BPFP=0.4200 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,564B, BPFP=0.0326 +⌛️ [2/4] FRONTEND: Frontend time: 1.903s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.178s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12281527 372.19538551 + layer.0.v_cache 0.00001417 0.04037404 + layer.1.k_cache 0.04442539 24.51168224 + layer.1.v_cache 0.00000572 0.01607205 + layer.2.k_cache 0.01033308 3.31874042 + layer.2.v_cache 0.00001923 0.04682413 + layer.3.k_cache 0.04555132 14.40695732 + layer.3.v_cache 0.00001841 0.05319937 + layer.4.k_cache 0.00060580 1.14011283 + layer.4.v_cache 0.00007068 0.09655197 + layer.4.output 10.66341841 502.63009012 + ------------------------------------------------------------------------------------- + TOTAL 4.40398753 231.42567828 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 28284 +BPFP 0.2430 bits/point +EBPFP 0.4859 equivalent bits/point +MSE 231.425678 +---------------------- -------------------------------------------------------- +Time: 3.086s Load: 0.006s, Pack+Encode: 1.903s, Decode+Unpack: 1.178s +---------------------- -------------------------------------------------------- +💾 Converting with 231.4257 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,240B, BPFP=0.2253 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,572B, BPFP=0.4673 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,892B, BPFP=0.5254 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,408B, BPFP=0.4375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,420B, BPFP=0.4397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,300B, BPFP=0.4179 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,344B, BPFP=0.4259 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,456B, BPFP=0.4462 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,804B, BPFP=0.5094 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,544B, BPFP=0.4622 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,536B, BPFP=0.0399 +⌛️ [2/4] FRONTEND: Frontend time: 1.678s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11362342 326.60778525 + layer.0.v_cache 0.00001418 0.04066138 + layer.1.k_cache 0.05429294 24.89556459 + layer.1.v_cache 0.00000580 0.01640282 + layer.2.k_cache 0.00513919 3.74509057 + layer.2.v_cache 0.00001816 0.04707657 + layer.3.k_cache 0.05885953 16.02366461 + layer.3.v_cache 0.00001841 0.05291419 + layer.4.k_cache 0.00061763 1.19890435 + layer.4.v_cache 0.00005080 0.10069475 + layer.4.output 0.16727475 631.34567068 + ------------------------------------------------------------------------------------- + TOTAL 0.08256255 281.89108552 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 25516 +BPFP 0.2727 bits/point +EBPFP 0.5454 equivalent bits/point +MSE 281.891086 +---------------------- -------------------------------------------------------- +Time: 2.907s Load: 0.004s, Pack+Encode: 1.678s, Decode+Unpack: 1.225s +---------------------- -------------------------------------------------------- +💾 Converting with 281.8911 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.2231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,472B, BPFP=0.4153 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,560B, BPFP=0.4301 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,304B, BPFP=0.3871 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,420B, BPFP=0.4066 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,312B, BPFP=0.3884 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,176B, BPFP=0.3656 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,344B, BPFP=0.3938 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,568B, BPFP=0.4315 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,604B, BPFP=0.4375 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,312B, BPFP=0.0315 +⌛️ [2/4] FRONTEND: Frontend time: 1.745s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.300s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702597 328.84122984 + layer.0.v_cache 0.00001402 0.04186350 + layer.1.k_cache 0.01332537 25.29153541 + layer.1.v_cache 0.00000581 0.01622186 + layer.2.k_cache 0.00507878 3.78849185 + layer.2.v_cache 0.00001944 0.04605893 + layer.3.k_cache 0.04051446 16.48182859 + layer.3.v_cache 0.00001931 0.05249695 + layer.4.k_cache 0.00062870 1.13238558 + layer.4.v_cache 0.00006514 0.09914386 + layer.4.output 0.16549673 583.76752112 + ------------------------------------------------------------------------------------- + TOTAL 0.07795142 262.48022966 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 24400 +BPFP 0.2411 bits/point +EBPFP 0.4823 equivalent bits/point +MSE 262.480230 +---------------------- -------------------------------------------------------- +Time: 3.049s Load: 0.004s, Pack+Encode: 1.745s, Decode+Unpack: 1.300s +---------------------- -------------------------------------------------------- +💾 Converting with 262.4802 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,252B, BPFP=0.2301 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,448B, BPFP=0.4500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,928B, BPFP=0.5382 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,092B, BPFP=0.3846 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,560B, BPFP=0.4706 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,352B, BPFP=0.4324 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,132B, BPFP=0.3919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,308B, BPFP=0.4243 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,756B, BPFP=0.5066 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,596B, BPFP=0.4772 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,384B, BPFP=0.0363 +⌛️ [2/4] FRONTEND: Frontend time: 1.752s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13925676 322.84839154 + layer.0.v_cache 0.00001387 0.04039542 + layer.1.k_cache 0.03290265 24.57932273 + layer.1.v_cache 0.00000589 0.01583172 + layer.2.k_cache 0.00800494 3.94728358 + layer.2.v_cache 0.00001921 0.04672774 + layer.3.k_cache 0.08301921 16.24743796 + layer.3.v_cache 0.00001931 0.05523094 + layer.4.k_cache 0.00065893 1.22520644 + layer.4.v_cache 0.00004891 0.09970739 + layer.4.output 0.17367665 638.46706933 + ------------------------------------------------------------------------------------- + TOTAL 0.08704037 284.61029534 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 24808 +BPFP 0.2683 bits/point +EBPFP 0.5365 equivalent bits/point +MSE 284.610295 +---------------------- -------------------------------------------------------- +Time: 2.955s Load: 0.004s, Pack+Encode: 1.752s, Decode+Unpack: 1.199s +---------------------- -------------------------------------------------------- +💾 Converting with 284.6103 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,504B, BPFP=0.2282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,648B, BPFP=0.4017 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,128B, BPFP=0.4745 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,764B, BPFP=0.4193 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,660B, BPFP=0.4035 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,820B, BPFP=0.4278 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,448B, BPFP=0.3714 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,968B, BPFP=0.4502 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,048B, BPFP=0.4624 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,724B, BPFP=0.4132 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,536B, BPFP=0.0333 +⌛️ [2/4] FRONTEND: Frontend time: 1.772s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.444s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10583916 358.88000607 + layer.0.v_cache 0.00001406 0.04029521 + layer.1.k_cache 0.01351924 24.18864011 + layer.1.v_cache 0.00000536 0.01638083 + layer.2.k_cache 0.00567923 3.36701773 + layer.2.v_cache 0.00001856 0.04839610 + layer.3.k_cache 0.05293757 14.92915063 + layer.3.v_cache 0.00001841 0.05429319 + layer.4.k_cache 0.00064880 1.14128424 + layer.4.v_cache 0.00004836 0.09448979 + layer.4.output 11.07886250 522.07871012 + ------------------------------------------------------------------------------------- + TOTAL 4.57239802 238.66534852 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 28248 +BPFP 0.2521 bits/point +EBPFP 0.5041 equivalent bits/point +MSE 238.665349 +---------------------- -------------------------------------------------------- +Time: 3.221s Load: 0.005s, Pack+Encode: 1.772s, Decode+Unpack: 1.444s +---------------------- -------------------------------------------------------- +💾 Converting with 238.6653 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,212B, BPFP=0.2282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,868B, BPFP=0.5399 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,084B, BPFP=0.5806 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,996B, BPFP=0.3758 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,440B, BPFP=0.4593 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,316B, BPFP=0.4360 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,224B, BPFP=0.4187 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,184B, BPFP=0.4111 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,960B, BPFP=0.5572 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,920B, BPFP=0.5497 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,380B, BPFP=0.0371 +⌛️ [2/4] FRONTEND: Frontend time: 1.843s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.228s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860243 323.77727316 + layer.0.v_cache 0.00001403 0.04044838 + layer.1.k_cache 0.01400426 26.56990658 + layer.1.v_cache 0.00000584 0.01697703 + layer.2.k_cache 0.01277778 3.98119345 + layer.2.v_cache 0.00001980 0.04822583 + layer.3.k_cache 0.07376247 16.82295098 + layer.3.v_cache 0.00001877 0.05331146 + layer.4.k_cache 0.00062152 1.21612034 + layer.4.v_cache 0.00005140 0.10221080 + layer.4.output 0.18179212 654.11596386 + ------------------------------------------------------------------------------------- + TOTAL 0.08837783 291.26119794 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 25584 +BPFP 0.2833 bits/point +EBPFP 0.5666 equivalent bits/point +MSE 291.261198 +---------------------- -------------------------------------------------------- +Time: 3.075s Load: 0.004s, Pack+Encode: 1.843s, Decode+Unpack: 1.228s +---------------------- -------------------------------------------------------- +💾 Converting with 291.2612 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,288B, BPFP=0.2236 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,620B, BPFP=0.4549 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,528B, BPFP=0.4389 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,440B, BPFP=0.4236 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,420B, BPFP=0.4201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,448B, BPFP=0.4250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,984B, BPFP=0.3444 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,492B, BPFP=0.4326 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,796B, BPFP=0.4854 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,700B, BPFP=0.4688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,384B, BPFP=0.0343 +⌛️ [2/4] FRONTEND: Frontend time: 1.837s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.154s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13781073 336.59787326 + layer.0.v_cache 0.00001414 0.04265152 + layer.1.k_cache 0.01252146 24.80973307 + layer.1.v_cache 0.00000561 0.01624289 + layer.2.k_cache 0.00238173 3.86458672 + layer.2.v_cache 0.00001849 0.04936317 + layer.3.k_cache 0.04360520 16.77264811 + layer.3.v_cache 0.00001973 0.05599135 + layer.4.k_cache 0.00061668 1.19859907 + layer.4.v_cache 0.00005448 0.10752767 + layer.4.output 0.16772948 603.21378968 + ------------------------------------------------------------------------------------- + TOTAL 0.08065615 270.94186733 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 25100 +BPFP 0.2563 bits/point +EBPFP 0.5127 equivalent bits/point +MSE 270.941867 +---------------------- -------------------------------------------------------- +Time: 2.995s Load: 0.004s, Pack+Encode: 1.837s, Decode+Unpack: 1.154s +---------------------- -------------------------------------------------------- +💾 Converting with 270.9419 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,264B, BPFP=0.2500 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,948B, BPFP=0.5831 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,944B, BPFP=0.5823 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,152B, BPFP=0.4256 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,592B, BPFP=0.5127 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,016B, BPFP=0.3987 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,252B, BPFP=0.4454 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,320B, BPFP=0.4589 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,936B, BPFP=0.5807 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,952B, BPFP=0.5839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,536B, BPFP=0.0434 +⌛️ [2/4] FRONTEND: Frontend time: 1.877s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.396s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07550411 334.30458861 + layer.0.v_cache 0.00001381 0.04077800 + layer.1.k_cache 0.01271453 25.82413840 + layer.1.v_cache 0.00000531 0.01552602 + layer.2.k_cache 0.00681068 3.91902721 + layer.2.v_cache 0.00001904 0.05150517 + layer.3.k_cache 0.02844424 16.46207476 + layer.3.v_cache 0.00001798 0.05330026 + layer.4.k_cache 0.00061127 1.18691099 + layer.4.v_cache 0.00009323 0.10504858 + layer.4.output 0.18255964 687.16811709 + ------------------------------------------------------------------------------------- + TOTAL 0.08247951 305.41998339 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 25912 +BPFP 0.3015 bits/point +EBPFP 0.6029 equivalent bits/point +MSE 305.419983 +---------------------- -------------------------------------------------------- +Time: 3.276s Load: 0.003s, Pack+Encode: 1.877s, Decode+Unpack: 1.396s +---------------------- -------------------------------------------------------- +💾 Converting with 305.4200 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,240B, BPFP=0.2422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,904B, BPFP=0.5672 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,012B, BPFP=0.5883 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,400B, BPFP=0.4688 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,480B, BPFP=0.4844 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,500B, BPFP=0.4883 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,148B, BPFP=0.4195 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,080B, BPFP=0.4062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,960B, BPFP=0.5781 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,016B, BPFP=0.5891 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,424B, BPFP=0.0397 +⌛️ [2/4] FRONTEND: Frontend time: 1.990s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.195s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08590809 329.76870117 + layer.0.v_cache 0.00001388 0.04017139 + layer.1.k_cache 0.01594085 25.34622498 + layer.1.v_cache 0.00000552 0.01587543 + layer.2.k_cache 0.00791870 4.17404022 + layer.2.v_cache 0.00001801 0.04569499 + layer.3.k_cache 0.02757127 16.01318665 + layer.3.v_cache 0.00001862 0.05622661 + layer.4.k_cache 0.00063773 1.22168007 + layer.4.v_cache 0.00005492 0.10409987 + layer.4.output 0.18408036 678.65881696 + ------------------------------------------------------------------------------------- + TOTAL 0.08392059 301.61162471 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 26164 +BPFP 0.3006 bits/point +EBPFP 0.6012 equivalent bits/point +MSE 301.611625 +---------------------- -------------------------------------------------------- +Time: 3.191s Load: 0.005s, Pack+Encode: 1.990s, Decode+Unpack: 1.195s +---------------------- -------------------------------------------------------- +💾 Converting with 301.6116 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,288B, BPFP=0.2648 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,964B, BPFP=0.6094 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,412B, BPFP=0.4959 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,924B, BPFP=0.3956 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,464B, BPFP=0.5066 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,864B, BPFP=0.3832 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,084B, BPFP=0.4285 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,876B, BPFP=0.3857 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,796B, BPFP=0.5748 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,420B, BPFP=0.4975 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,328B, BPFP=0.0390 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.268s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08991440 337.34503495 + layer.0.v_cache 0.00001347 0.03964173 + layer.1.k_cache 0.03345199 24.87604402 + layer.1.v_cache 0.00000518 0.01645542 + layer.2.k_cache 0.00398020 4.08379966 + layer.2.v_cache 0.00001815 0.04793345 + layer.3.k_cache 0.04607732 16.14757980 + layer.3.v_cache 0.00001896 0.05559890 + layer.4.k_cache 0.00062119 1.21096701 + layer.4.v_cache 0.00005040 0.10372878 + layer.4.output 0.18612516 714.49089521 + ------------------------------------------------------------------------------------- + TOTAL 0.08688396 316.78606177 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 23420 +BPFP 0.2832 bits/point +EBPFP 0.5665 equivalent bits/point +MSE 316.786062 +---------------------- -------------------------------------------------------- +Time: 3.002s Load: 0.003s, Pack+Encode: 1.730s, Decode+Unpack: 1.268s +---------------------- -------------------------------------------------------- +💾 Converting with 316.7861 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,560B, BPFP=0.2679 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,428B, BPFP=0.4169 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,572B, BPFP=0.4416 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,352B, BPFP=0.4038 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,400B, BPFP=0.4121 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,352B, BPFP=0.4038 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,968B, BPFP=0.3379 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,148B, BPFP=0.3688 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,628B, BPFP=0.4512 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,576B, BPFP=0.4423 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,292B, BPFP=0.0317 +⌛️ [2/4] FRONTEND: Frontend time: 1.758s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.172s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11186004 338.58791209 + layer.0.v_cache 0.00001364 0.04059017 + layer.1.k_cache 0.01386243 25.28424676 + layer.1.v_cache 0.00000608 0.01593015 + layer.2.k_cache 0.00506424 3.95043610 + layer.2.v_cache 0.00001872 0.04848653 + layer.3.k_cache 0.02496567 16.07612627 + layer.3.v_cache 0.00001876 0.05753840 + layer.4.k_cache 0.00064518 1.23505947 + layer.4.v_cache 0.00005436 0.09912246 + layer.4.output 0.16682192 596.55165816 + ------------------------------------------------------------------------------------- + TOTAL 0.07789780 268.30923856 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 24276 +BPFP 0.2452 bits/point +EBPFP 0.4904 equivalent bits/point +MSE 268.309239 +---------------------- -------------------------------------------------------- +Time: 2.935s Load: 0.004s, Pack+Encode: 1.758s, Decode+Unpack: 1.172s +---------------------- -------------------------------------------------------- +💾 Converting with 268.3092 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,224B, BPFP=0.2304 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,584B, BPFP=0.4864 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,216B, BPFP=0.6054 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,412B, BPFP=0.4541 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,512B, BPFP=0.4729 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,064B, BPFP=0.3886 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,284B, BPFP=0.4300 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,124B, BPFP=0.3998 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,044B, BPFP=0.5730 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,848B, BPFP=0.5361 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,384B, BPFP=0.0372 +⌛️ [2/4] FRONTEND: Frontend time: 1.710s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.151s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08542032 336.80073419 + layer.0.v_cache 0.00001402 0.04040312 + layer.1.k_cache 0.01083013 26.16158580 + layer.1.v_cache 0.00000556 0.01563054 + layer.2.k_cache 0.00382075 4.12762856 + layer.2.v_cache 0.00001860 0.05038920 + layer.3.k_cache 0.02610962 16.77666280 + layer.3.v_cache 0.00002051 0.05681885 + layer.4.k_cache 0.00063312 1.21683052 + layer.4.v_cache 0.00005614 0.10592049 + layer.4.output 0.16811641 654.24591222 + ------------------------------------------------------------------------------------- + TOTAL 0.07669080 292.06317586 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 25696 +BPFP 0.2845 bits/point +EBPFP 0.5691 equivalent bits/point +MSE 292.063176 +---------------------- -------------------------------------------------------- +Time: 2.865s Load: 0.005s, Pack+Encode: 1.710s, Decode+Unpack: 1.151s +---------------------- -------------------------------------------------------- +💾 Converting with 292.0632 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,300B, BPFP=0.2508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,648B, BPFP=0.5108 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,096B, BPFP=0.5972 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,436B, BPFP=0.4699 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,780B, BPFP=0.5363 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,036B, BPFP=0.3927 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,312B, BPFP=0.4460 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,412B, BPFP=0.4653 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,976B, BPFP=0.5741 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,964B, BPFP=0.5718 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,352B, BPFP=0.0373 +⌛️ [2/4] FRONTEND: Frontend time: 1.608s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.119s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10540552 334.71402392 + layer.0.v_cache 0.00001541 0.04163060 + layer.1.k_cache 0.03402772 26.05720727 + layer.1.v_cache 0.00000555 0.01621667 + layer.2.k_cache 0.00394259 4.27915446 + layer.2.v_cache 0.00001868 0.05160806 + layer.3.k_cache 0.02873192 16.54863824 + layer.3.v_cache 0.00001983 0.05723303 + layer.4.k_cache 0.00063741 1.26639679 + layer.4.v_cache 0.00005806 0.10667890 + layer.4.output 0.17961645 670.19824735 + ------------------------------------------------------------------------------------- + TOTAL 0.08412811 298.50155997 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 26312 +BPFP 0.2986 bits/point +EBPFP 0.5971 equivalent bits/point +MSE 298.501560 +---------------------- -------------------------------------------------------- +Time: 2.731s Load: 0.004s, Pack+Encode: 1.608s, Decode+Unpack: 1.119s +---------------------- -------------------------------------------------------- +💾 Converting with 298.5016 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,240B, BPFP=0.2422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,732B, BPFP=0.5336 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,912B, BPFP=0.5687 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,040B, BPFP=0.3984 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,596B, BPFP=0.5070 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,948B, BPFP=0.3805 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,132B, BPFP=0.4164 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,512B, BPFP=0.4906 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,024B, BPFP=0.5906 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,016B, BPFP=0.5891 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,416B, BPFP=0.0395 +⌛️ [2/4] FRONTEND: Frontend time: 1.727s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.273s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10022384 311.24853516 + layer.0.v_cache 0.00001350 0.04112488 + layer.1.k_cache 0.03445883 25.14728546 + layer.1.v_cache 0.00000556 0.01727578 + layer.2.k_cache 0.00550650 4.13407402 + layer.2.v_cache 0.00001848 0.04899079 + layer.3.k_cache 0.03016714 16.29971008 + layer.3.v_cache 0.00001937 0.05767179 + layer.4.k_cache 0.00063435 1.24029951 + layer.4.v_cache 0.00006510 0.10601292 + layer.4.output 0.18593751 678.68973214 + ------------------------------------------------------------------------------------- + TOTAL 0.08662796 300.53935914 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 25568 +BPFP 0.2938 bits/point +EBPFP 0.5875 equivalent bits/point +MSE 300.539359 +---------------------- -------------------------------------------------------- +Time: 3.003s Load: 0.003s, Pack+Encode: 1.727s, Decode+Unpack: 1.273s +---------------------- -------------------------------------------------------- +💾 Converting with 300.5394 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,472B, BPFP=0.2371 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,344B, BPFP=0.3776 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,900B, BPFP=0.4671 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,236B, BPFP=0.3602 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,428B, BPFP=0.3911 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,180B, BPFP=0.3512 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,316B, BPFP=0.3731 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,416B, BPFP=0.3892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,844B, BPFP=0.4581 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,688B, BPFP=0.4330 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,428B, BPFP=0.0329 +⌛️ [2/4] FRONTEND: Frontend time: 1.712s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.168s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11898830 352.53825709 + layer.0.v_cache 0.00001403 0.04058395 + layer.1.k_cache 0.01243949 25.70895921 + layer.1.v_cache 0.00000551 0.01609118 + layer.2.k_cache 0.00346032 3.34759396 + layer.2.v_cache 0.00001901 0.04945774 + layer.3.k_cache 0.02403709 15.14003856 + layer.3.v_cache 0.00001994 0.05489445 + layer.4.k_cache 0.00063949 1.13048750 + layer.4.v_cache 0.00006591 0.10100120 + layer.4.output 0.03773703 574.69316090 + ------------------------------------------------------------------------------------- + TOTAL 0.02493225 260.05761712 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 25252 +BPFP 0.2393 bits/point +EBPFP 0.4785 equivalent bits/point +MSE 260.057617 +---------------------- -------------------------------------------------------- +Time: 2.886s Load: 0.005s, Pack+Encode: 1.712s, Decode+Unpack: 1.168s +---------------------- -------------------------------------------------------- +💾 Converting with 260.0576 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,164B, BPFP=0.2218 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,848B, BPFP=0.5427 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,048B, BPFP=0.5808 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,236B, BPFP=0.4261 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,528B, BPFP=0.4817 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,492B, BPFP=0.4748 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,128B, BPFP=0.4055 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,732B, BPFP=0.5206 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,952B, BPFP=0.5625 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,852B, BPFP=0.5434 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,416B, BPFP=0.0385 +⌛️ [2/4] FRONTEND: Frontend time: 1.665s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.188s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10693085 325.92725800 + layer.0.v_cache 0.00001341 0.03956838 + layer.1.k_cache 0.01172146 25.77887409 + layer.1.v_cache 0.00000553 0.01595746 + layer.2.k_cache 0.00236719 4.03626456 + layer.2.v_cache 0.00001835 0.04468734 + layer.3.k_cache 0.02622769 16.15522133 + layer.3.v_cache 0.00001864 0.05250020 + layer.4.k_cache 0.00061713 1.17437000 + layer.4.v_cache 0.00004943 0.10072835 + layer.4.output 0.18273007 662.06386106 + ------------------------------------------------------------------------------------- + TOTAL 0.08394589 294.57485042 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 26396 +BPFP 0.2959 bits/point +EBPFP 0.5917 equivalent bits/point +MSE 294.574850 +---------------------- -------------------------------------------------------- +Time: 2.857s Load: 0.004s, Pack+Encode: 1.665s, Decode+Unpack: 1.188s +---------------------- -------------------------------------------------------- +💾 Converting with 294.5749 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,444B, BPFP=0.2234 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,684B, BPFP=0.4152 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,160B, BPFP=0.4889 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,856B, BPFP=0.4418 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,728B, BPFP=0.4220 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,888B, BPFP=0.4468 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,440B, BPFP=0.3775 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,612B, BPFP=0.4041 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,148B, BPFP=0.4870 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,860B, BPFP=0.4425 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,508B, BPFP=0.0333 +⌛️ [2/4] FRONTEND: Frontend time: 1.677s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13156438 342.47555693 + layer.0.v_cache 0.00001398 0.03956262 + layer.1.k_cache 0.04507210 25.33850460 + layer.1.v_cache 0.00000603 0.01723127 + layer.2.k_cache 0.00848899 3.42084768 + layer.2.v_cache 0.00001983 0.05009758 + layer.3.k_cache 0.06093804 14.74074925 + layer.3.v_cache 0.00001891 0.05180162 + layer.4.k_cache 0.00062755 1.17670932 + layer.4.v_cache 0.00004781 0.09361320 + layer.4.output 11.29576791 532.43153289 + ------------------------------------------------------------------------------------- + TOTAL 4.66571606 242.02502378 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 28328 +BPFP 0.2578 bits/point +EBPFP 0.5156 equivalent bits/point +MSE 242.025024 +---------------------- -------------------------------------------------------- +Time: 2.882s Load: 0.004s, Pack+Encode: 1.677s, Decode+Unpack: 1.201s +---------------------- -------------------------------------------------------- +💾 Converting with 242.0250 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,224B, BPFP=0.2304 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,420B, BPFP=0.4556 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,124B, BPFP=0.5881 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,060B, BPFP=0.3878 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,768B, BPFP=0.5211 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,200B, BPFP=0.4142 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,348B, BPFP=0.4420 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,992B, BPFP=0.3750 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,864B, BPFP=0.5392 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,948B, BPFP=0.5550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,372B, BPFP=0.0369 +⌛️ [2/4] FRONTEND: Frontend time: 1.670s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.167s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12340720 335.45627824 + layer.0.v_cache 0.00001379 0.04113383 + layer.1.k_cache 0.01365945 26.42765790 + layer.1.v_cache 0.00000586 0.01692139 + layer.2.k_cache 0.00515996 4.30527947 + layer.2.v_cache 0.00002118 0.05098422 + layer.3.k_cache 0.04682968 16.70342356 + layer.3.v_cache 0.00001899 0.05767763 + layer.4.k_cache 0.00062596 1.23037867 + layer.4.v_cache 0.00004975 0.10222035 + layer.4.output 0.18155291 654.10235585 + ------------------------------------------------------------------------------------- + TOTAL 0.08592130 291.94755566 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 25320 +BPFP 0.2804 bits/point +EBPFP 0.5608 equivalent bits/point +MSE 291.947556 +---------------------- -------------------------------------------------------- +Time: 2.841s Load: 0.004s, Pack+Encode: 1.670s, Decode+Unpack: 1.167s +---------------------- -------------------------------------------------------- +💾 Converting with 291.9476 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,232B, BPFP=0.2292 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,452B, BPFP=0.4561 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,096B, BPFP=0.5759 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,708B, BPFP=0.5037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,368B, BPFP=0.4405 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,624B, BPFP=0.4881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,452B, BPFP=0.4561 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,740B, BPFP=0.5097 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,928B, BPFP=0.5446 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,992B, BPFP=0.5565 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0372 +⌛️ [2/4] FRONTEND: Frontend time: 1.649s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.109s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10400932 323.26997303 + layer.0.v_cache 0.00001381 0.04091156 + layer.1.k_cache 0.01300498 25.86860003 + layer.1.v_cache 0.00000559 0.01587462 + layer.2.k_cache 0.00535372 3.83740016 + layer.2.v_cache 0.00001862 0.04563412 + layer.3.k_cache 0.04415199 16.37880452 + layer.3.v_cache 0.00001841 0.05547093 + layer.4.k_cache 0.00060249 1.18748047 + layer.4.v_cache 0.00005274 0.10694326 + layer.4.output 0.16694923 646.44424957 + ------------------------------------------------------------------------------------- + TOTAL 0.07858096 287.99510822 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 26992 +BPFP 0.2953 bits/point +EBPFP 0.5907 equivalent bits/point +MSE 287.995108 +---------------------- -------------------------------------------------------- +Time: 2.762s Load: 0.004s, Pack+Encode: 1.649s, Decode+Unpack: 1.109s +---------------------- -------------------------------------------------------- +💾 Converting with 287.9951 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,224B, BPFP=0.2391 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,968B, BPFP=0.5797 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,024B, BPFP=0.5906 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,528B, BPFP=0.4938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,816B, BPFP=0.5500 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,256B, BPFP=0.4406 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,304B, BPFP=0.4500 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,020B, BPFP=0.3945 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,704B, BPFP=0.5281 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,768B, BPFP=0.5406 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,428B, BPFP=0.0398 +⌛️ [2/4] FRONTEND: Frontend time: 1.632s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.165s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09321334 324.15083008 + layer.0.v_cache 0.00001862 0.04194016 + layer.1.k_cache 0.03257937 25.70877380 + layer.1.v_cache 0.00000532 0.01539957 + layer.2.k_cache 0.01442847 4.14296112 + layer.2.v_cache 0.00001865 0.04835315 + layer.3.k_cache 0.04385926 16.40492249 + layer.3.v_cache 0.00001828 0.05501630 + layer.4.k_cache 0.00061064 1.19310513 + layer.4.v_cache 0.00005353 0.10352201 + layer.4.output 0.18500509 678.63247768 + ------------------------------------------------------------------------------------- + TOTAL 0.08704948 301.31130397 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 26040 +BPFP 0.2992 bits/point +EBPFP 0.5983 equivalent bits/point +MSE 301.311304 +---------------------- -------------------------------------------------------- +Time: 2.802s Load: 0.004s, Pack+Encode: 1.632s, Decode+Unpack: 1.165s +---------------------- -------------------------------------------------------- +💾 Converting with 301.3113 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,532B, BPFP=0.2324 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,576B, BPFP=0.3908 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,000B, BPFP=0.4551 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,704B, BPFP=0.4102 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,652B, BPFP=0.4023 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,480B, BPFP=0.3762 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,436B, BPFP=0.3695 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,840B, BPFP=0.4308 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,216B, BPFP=0.4879 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,656B, BPFP=0.4029 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,500B, BPFP=0.0325 +⌛️ [2/4] FRONTEND: Frontend time: 1.748s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.289s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11527186 373.18810680 + layer.0.v_cache 0.00001510 0.04011076 + layer.1.k_cache 0.02920971 23.89354047 + layer.1.v_cache 0.00000551 0.01650613 + layer.2.k_cache 0.00857845 3.28157592 + layer.2.v_cache 0.00001837 0.04615874 + layer.3.k_cache 0.03550095 14.54085207 + layer.3.v_cache 0.00001950 0.05570231 + layer.4.k_cache 0.00063242 1.17513845 + layer.4.v_cache 0.00005150 0.10118805 + layer.4.output 11.08123415 522.15234917 + ------------------------------------------------------------------------------------- + TOTAL 4.57399661 239.49443082 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 27592 +BPFP 0.2462 bits/point +EBPFP 0.4924 equivalent bits/point +MSE 239.494431 +---------------------- -------------------------------------------------------- +Time: 3.042s Load: 0.005s, Pack+Encode: 1.748s, Decode+Unpack: 1.289s +---------------------- -------------------------------------------------------- +💾 Converting with 239.4944 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,336B, BPFP=0.2269 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,300B, BPFP=0.3906 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,636B, BPFP=0.4477 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,528B, BPFP=0.4293 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,280B, BPFP=0.3872 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,320B, BPFP=0.3940 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,936B, BPFP=0.3288 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,364B, BPFP=0.4015 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,728B, BPFP=0.4633 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,668B, BPFP=0.4531 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,296B, BPFP=0.0314 +⌛️ [2/4] FRONTEND: Frontend time: 1.877s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.408s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09673090 362.92756454 + layer.0.v_cache 0.00001344 0.03988630 + layer.1.k_cache 0.04918879 25.30067245 + layer.1.v_cache 0.00000572 0.01633927 + layer.2.k_cache 0.00880511 3.76933819 + layer.2.v_cache 0.00001944 0.04719219 + layer.3.k_cache 0.06586690 16.13360330 + layer.3.v_cache 0.00001823 0.05344796 + layer.4.k_cache 0.00062484 1.18425510 + layer.4.v_cache 0.00006325 0.10058974 + layer.4.output 0.15294319 590.16731366 + ------------------------------------------------------------------------------------- + TOTAL 0.07599641 267.10259322 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 24392 +BPFP 0.2437 bits/point +EBPFP 0.4874 equivalent bits/point +MSE 267.102593 +---------------------- -------------------------------------------------------- +Time: 3.290s Load: 0.004s, Pack+Encode: 1.877s, Decode+Unpack: 1.408s +---------------------- -------------------------------------------------------- +💾 Converting with 267.1026 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,316B, BPFP=0.2235 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,492B, BPFP=0.4232 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,724B, BPFP=0.4626 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,200B, BPFP=0.3736 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,440B, BPFP=0.4144 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,324B, BPFP=0.3947 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,984B, BPFP=0.3370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,356B, BPFP=0.4001 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,592B, BPFP=0.4402 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,544B, BPFP=0.4321 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,312B, BPFP=0.0318 +⌛️ [2/4] FRONTEND: Frontend time: 1.774s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.168s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12676482 346.94974949 + layer.0.v_cache 0.00001339 0.03985360 + layer.1.k_cache 0.03133728 25.35483717 + layer.1.v_cache 0.00000590 0.01601258 + layer.2.k_cache 0.00372887 3.57487587 + layer.2.v_cache 0.00001820 0.04740415 + layer.3.k_cache 0.02418065 16.15211022 + layer.3.v_cache 0.00002034 0.05590452 + layer.4.k_cache 0.00062996 1.19168945 + layer.4.v_cache 0.00006053 0.09821891 + layer.4.output 0.15389095 590.23641304 + ------------------------------------------------------------------------------------- + TOTAL 0.07435274 266.18444396 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 24284 +BPFP 0.2426 bits/point +EBPFP 0.4852 equivalent bits/point +MSE 266.184444 +---------------------- -------------------------------------------------------- +Time: 2.947s Load: 0.005s, Pack+Encode: 1.774s, Decode+Unpack: 1.168s +---------------------- -------------------------------------------------------- +💾 Converting with 266.1844 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,292B, BPFP=0.2243 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,592B, BPFP=0.4500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,560B, BPFP=0.4444 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,588B, BPFP=0.4493 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,376B, BPFP=0.4125 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,256B, BPFP=0.3917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,940B, BPFP=0.3368 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,604B, BPFP=0.4521 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,516B, BPFP=0.4368 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,800B, BPFP=0.4861 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,396B, BPFP=0.0346 +⌛️ [2/4] FRONTEND: Frontend time: 1.708s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12863920 333.87660590 + layer.0.v_cache 0.00001425 0.04124285 + layer.1.k_cache 0.01389826 24.87823079 + layer.1.v_cache 0.00000571 0.01654086 + layer.2.k_cache 0.00659207 3.67374166 + layer.2.v_cache 0.00001883 0.04721144 + layer.3.k_cache 0.05540012 17.01956923 + layer.3.v_cache 0.00001922 0.05239524 + layer.4.k_cache 0.00063279 1.18170844 + layer.4.v_cache 0.00005416 0.09975380 + layer.4.output 0.16196846 603.00099206 + ------------------------------------------------------------------------------------- + TOTAL 0.07876787 270.69964380 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 24920 +BPFP 0.2545 bits/point +EBPFP 0.5090 equivalent bits/point +MSE 270.699644 +---------------------- -------------------------------------------------------- +Time: 2.911s Load: 0.006s, Pack+Encode: 1.708s, Decode+Unpack: 1.197s +---------------------- -------------------------------------------------------- +💾 Converting with 270.6996 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,668B, BPFP=0.2413 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,644B, BPFP=0.3825 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,176B, BPFP=0.4595 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,540B, BPFP=0.3675 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,680B, BPFP=0.3877 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,600B, BPFP=0.3762 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,548B, BPFP=0.3686 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,864B, BPFP=0.4144 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,196B, BPFP=0.4624 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,732B, BPFP=0.3953 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,696B, BPFP=0.0351 +⌛️ [2/4] FRONTEND: Frontend time: 1.666s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13160179 371.13505498 + layer.0.v_cache 0.00001577 0.03930400 + layer.1.k_cache 0.05775581 24.71941461 + layer.1.v_cache 0.00000549 0.01563908 + layer.2.k_cache 0.01027601 3.58775556 + layer.2.v_cache 0.00002197 0.04638192 + layer.3.k_cache 0.07329086 14.31014563 + layer.3.v_cache 0.00001983 0.05602733 + layer.4.k_cache 0.00064477 1.18189791 + layer.4.v_cache 0.00005260 0.09703124 + layer.4.output 10.56448284 497.98127480 + ------------------------------------------------------------------------------------- + TOTAL 4.36618028 229.47397505 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 28344 +BPFP 0.2412 bits/point +EBPFP 0.4824 equivalent bits/point +MSE 229.473975 +---------------------- -------------------------------------------------------- +Time: 2.876s Load: 0.004s, Pack+Encode: 1.666s, Decode+Unpack: 1.206s +---------------------- -------------------------------------------------------- +💾 Converting with 229.4740 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,236B, BPFP=0.2327 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,552B, BPFP=0.4804 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,044B, BPFP=0.5730 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,980B, BPFP=0.3727 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,700B, BPFP=0.5083 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,128B, BPFP=0.4006 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,396B, BPFP=0.4511 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,236B, BPFP=0.4209 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,932B, BPFP=0.5520 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,004B, BPFP=0.5655 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,376B, BPFP=0.0370 +⌛️ [2/4] FRONTEND: Frontend time: 1.900s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10139442 331.12594127 + layer.0.v_cache 0.00001346 0.04016699 + layer.1.k_cache 0.03266963 25.58890248 + layer.1.v_cache 0.00000526 0.01581637 + layer.2.k_cache 0.01283502 3.98124676 + layer.2.v_cache 0.00001884 0.04703213 + layer.3.k_cache 0.02666297 16.32801175 + layer.3.v_cache 0.00001780 0.05272885 + layer.4.k_cache 0.00061512 1.17978153 + layer.4.v_cache 0.00004947 0.09792759 + layer.4.output 0.18393325 654.07976549 + ------------------------------------------------------------------------------------- + TOTAL 0.08598910 291.58917142 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 25584 +BPFP 0.2833 bits/point +EBPFP 0.5666 equivalent bits/point +MSE 291.589171 +---------------------- -------------------------------------------------------- +Time: 3.115s Load: 0.005s, Pack+Encode: 1.900s, Decode+Unpack: 1.210s +---------------------- -------------------------------------------------------- +💾 Converting with 291.5892 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,400B, BPFP=0.2166 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,920B, BPFP=0.4517 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,192B, BPFP=0.4938 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,708B, BPFP=0.4189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,676B, BPFP=0.4140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,444B, BPFP=0.3781 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,576B, BPFP=0.3985 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,800B, BPFP=0.4332 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,168B, BPFP=0.4901 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,696B, BPFP=0.4171 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,436B, BPFP=0.0317 +⌛️ [2/4] FRONTEND: Frontend time: 1.744s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.336s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11149573 346.50719369 + layer.0.v_cache 0.00001357 0.03980013 + layer.1.k_cache 0.08092193 25.56008035 + layer.1.v_cache 0.00000571 0.01640515 + layer.2.k_cache 0.01109533 3.48528932 + layer.2.v_cache 0.00001920 0.04747857 + layer.3.k_cache 0.03617269 15.06978071 + layer.3.v_cache 0.00001858 0.05276336 + layer.4.k_cache 0.00062559 1.13708647 + layer.4.v_cache 0.00005684 0.09649898 + layer.4.output 11.30219499 532.41027228 + ------------------------------------------------------------------------------------- + TOTAL 4.66798765 242.28731075 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 28016 +BPFP 0.2550 bits/point +EBPFP 0.5099 equivalent bits/point +MSE 242.287311 +---------------------- -------------------------------------------------------- +Time: 3.085s Load: 0.004s, Pack+Encode: 1.744s, Decode+Unpack: 1.336s +---------------------- -------------------------------------------------------- +💾 Converting with 242.2873 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,220B, BPFP=0.2444 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,496B, BPFP=0.5000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,792B, BPFP=0.5593 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,960B, BPFP=0.3926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,664B, BPFP=0.5337 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,708B, BPFP=0.3421 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,100B, BPFP=0.4207 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,372B, BPFP=0.4752 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,740B, BPFP=0.5489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,960B, BPFP=0.5929 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,468B, BPFP=0.0420 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.046s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10167113 336.84440104 + layer.0.v_cache 0.00001386 0.04277596 + layer.1.k_cache 0.03545215 25.40411220 + layer.1.v_cache 0.00000569 0.01704319 + layer.2.k_cache 0.01047806 4.05569849 + layer.2.v_cache 0.00001755 0.05110362 + layer.3.k_cache 0.06687863 17.16592798 + layer.3.v_cache 0.00001960 0.05492571 + layer.4.k_cache 0.00059181 1.18564156 + layer.4.v_cache 0.00004882 0.10030775 + layer.4.output 0.19633365 696.04035027 + ------------------------------------------------------------------------------------- + TOTAL 0.09350076 309.24731703 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 24480 +BPFP 0.2885 bits/point +EBPFP 0.5769 equivalent bits/point +MSE 309.247317 +---------------------- -------------------------------------------------------- +Time: 2.776s Load: 0.004s, Pack+Encode: 1.726s, Decode+Unpack: 1.046s +---------------------- -------------------------------------------------------- +💾 Converting with 309.2473 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,592B, BPFP=0.2859 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,840B, BPFP=0.5101 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,584B, BPFP=0.4641 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,664B, BPFP=0.4784 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,432B, BPFP=0.4368 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,564B, BPFP=0.4605 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,184B, BPFP=0.3922 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,584B, BPFP=0.4641 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,524B, BPFP=0.4533 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,876B, BPFP=0.5165 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,416B, BPFP=0.0363 +⌛️ [2/4] FRONTEND: Frontend time: 1.704s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11445767 358.22712374 + layer.0.v_cache 0.00001583 0.04028823 + layer.1.k_cache 0.03092440 24.68022910 + layer.1.v_cache 0.00000551 0.01573114 + layer.2.k_cache 0.00639992 3.85552979 + layer.2.v_cache 0.00001897 0.04719915 + layer.3.k_cache 0.05613507 15.78208906 + layer.3.v_cache 0.00002026 0.05572168 + layer.4.k_cache 0.00063356 1.16344733 + layer.4.v_cache 0.00005123 0.10114653 + layer.4.output 0.16055210 624.04110222 + ------------------------------------------------------------------------------------- + TOTAL 0.07838395 280.72095419 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 26260 +BPFP 0.2774 bits/point +EBPFP 0.5549 equivalent bits/point +MSE 280.720954 +---------------------- -------------------------------------------------------- +Time: 2.928s Load: 0.004s, Pack+Encode: 1.704s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 280.7210 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,208B, BPFP=0.2247 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,428B, BPFP=0.4516 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,104B, BPFP=0.5774 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,320B, BPFP=0.4315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,516B, BPFP=0.4680 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,140B, BPFP=0.3981 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,200B, BPFP=0.4092 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,612B, BPFP=0.4859 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,732B, BPFP=0.5082 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,756B, BPFP=0.5126 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,396B, BPFP=0.0371 +⌛️ [2/4] FRONTEND: Frontend time: 1.754s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.148s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13956541 326.34351749 + layer.0.v_cache 0.00001349 0.04058654 + layer.1.k_cache 0.03336124 24.66747466 + layer.1.v_cache 0.00000558 0.01599800 + layer.2.k_cache 0.00698305 3.86920021 + layer.2.v_cache 0.00001831 0.04729616 + layer.3.k_cache 0.07289400 16.55157761 + layer.3.v_cache 0.00001855 0.05169862 + layer.4.k_cache 0.00061006 1.18638357 + layer.4.v_cache 0.00005429 0.10234664 + layer.4.output 0.17891866 646.33280187 + ------------------------------------------------------------------------------------- + TOTAL 0.08858556 288.07092309 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 25412 +BPFP 0.2781 bits/point +EBPFP 0.5561 equivalent bits/point +MSE 288.070923 +---------------------- -------------------------------------------------------- +Time: 2.905s Load: 0.004s, Pack+Encode: 1.754s, Decode+Unpack: 1.148s +---------------------- -------------------------------------------------------- +💾 Converting with 288.0709 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,320B, BPFP=0.2546 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,976B, BPFP=0.5741 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,160B, BPFP=0.6096 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,044B, BPFP=0.3943 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,660B, BPFP=0.5131 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,128B, BPFP=0.4105 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,356B, BPFP=0.4545 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,184B, BPFP=0.4213 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,932B, BPFP=0.5656 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,744B, BPFP=0.5293 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,348B, BPFP=0.0371 +⌛️ [2/4] FRONTEND: Frontend time: 1.762s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.295s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10837562 340.85619213 + layer.0.v_cache 0.00001644 0.04010423 + layer.1.k_cache 0.01241452 25.67288472 + layer.1.v_cache 0.00000532 0.01568597 + layer.2.k_cache 0.01724591 4.20847349 + layer.2.v_cache 0.00001958 0.04637698 + layer.3.k_cache 0.04641535 16.52711769 + layer.3.v_cache 0.00001847 0.05232093 + layer.4.k_cache 0.00063678 1.19146389 + layer.4.v_cache 0.00004974 0.10062510 + layer.4.output 0.18014439 670.06354718 + ------------------------------------------------------------------------------------- + TOTAL 0.08507109 298.77388679 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 25852 +BPFP 0.2933 bits/point +EBPFP 0.5867 equivalent bits/point +MSE 298.773887 +---------------------- -------------------------------------------------------- +Time: 3.061s Load: 0.003s, Pack+Encode: 1.762s, Decode+Unpack: 1.295s +---------------------- -------------------------------------------------------- +💾 Converting with 298.7739 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.2743 bits/point +Avg EBPFP 0.5487 equivalent bits/point +Avg MSE 282.580770 +Avg Time 2.974s +------------------------ ---------------------------- diff --git a/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..af837f6f09ecc830aa2159d26d2c6930637a06b1 --- /dev/null +++ b/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 286 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean +Output output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean +---------------- ------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,920B, BPFP=0.2797 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,792B, BPFP=0.3419 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,108B, BPFP=0.3644 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,536B, BPFP=0.3236 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,904B, BPFP=0.2785 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,504B, BPFP=0.3213 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,712B, BPFP=0.2648 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,492B, BPFP=0.3205 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,228B, BPFP=0.3730 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,408B, BPFP=0.3858 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,116B, BPFP=0.0216 +⌛️ [2/4] FRONTEND: Frontend time: 2.063s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.267s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12261864 400.54601884 + layer.0.v_cache 0.00001806 0.04726847 + layer.1.k_cache 0.30198812 24.93378104 + layer.1.v_cache 0.00000607 0.01708835 + layer.2.k_cache 0.01966049 2.96936063 + layer.2.v_cache 0.00002021 0.05033399 + layer.3.k_cache 0.01413035 13.01194617 + layer.3.v_cache 0.00002054 0.05949129 + layer.4.k_cache 0.00067868 1.25035590 + layer.4.v_cache 0.00004987 0.10220191 + layer.4.output 1.39792533 247.40837003 + ------------------------------------------------------------------------------------- + TOTAL 0.60262755 127.93214334 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 47720 +BPFP 0.2003 bits/point +EBPFP 0.4006 equivalent bits/point +MSE 127.932143 +---------------------- -------------------------------------------------------- +Time: 3.337s Load: 0.007s, Pack+Encode: 2.063s, Decode+Unpack: 1.267s +---------------------- -------------------------------------------------------- +💾 Converting with 127.9321 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,556B, BPFP=0.2572 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,920B, BPFP=0.3559 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,276B, BPFP=0.3817 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,384B, BPFP=0.3171 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,996B, BPFP=0.2891 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,416B, BPFP=0.3194 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,728B, BPFP=0.2697 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,452B, BPFP=0.3220 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,220B, BPFP=0.3776 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,988B, BPFP=0.3608 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,208B, BPFP=0.0228 +⌛️ [2/4] FRONTEND: Frontend time: 1.705s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11787021 392.04824942 + layer.0.v_cache 0.00001738 0.04365968 + layer.1.k_cache 0.25265321 24.95723018 + layer.1.v_cache 0.00000586 0.01688085 + layer.2.k_cache 0.01010228 2.88725535 + layer.2.v_cache 0.00002017 0.05189366 + layer.3.k_cache 0.01159736 14.11739886 + layer.3.v_cache 0.00002197 0.06257765 + layer.4.k_cache 0.00067244 1.32036463 + layer.4.v_cache 0.00004845 0.09989458 + layer.4.output 1.41728832 250.82798032 + ------------------------------------------------------------------------------------- + TOTAL 0.60670750 128.90595689 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 47144 +BPFP 0.2006 bits/point +EBPFP 0.4012 equivalent bits/point +MSE 128.905957 +---------------------- -------------------------------------------------------- +Time: 2.929s Load: 0.007s, Pack+Encode: 1.705s, Decode+Unpack: 1.217s +---------------------- -------------------------------------------------------- +💾 Converting with 128.9060 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,680B, BPFP=0.2578 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,548B, BPFP=0.3187 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,020B, BPFP=0.3517 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,288B, BPFP=0.3004 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,728B, BPFP=0.2612 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,684B, BPFP=0.2581 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,420B, BPFP=0.2396 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,380B, BPFP=0.3069 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,688B, BPFP=0.3285 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,320B, BPFP=0.3027 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,016B, BPFP=0.0202 +⌛️ [2/4] FRONTEND: Frontend time: 1.711s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13311632 392.61455998 + layer.0.v_cache 0.00001526 0.04349769 + layer.1.k_cache 0.36474110 25.76120200 + layer.1.v_cache 0.00000573 0.01652285 + layer.2.k_cache 0.01296863 2.80255921 + layer.2.v_cache 0.00001984 0.04898807 + layer.3.k_cache 0.01490937 13.96715378 + layer.3.v_cache 0.00002017 0.05980047 + layer.4.k_cache 0.00067370 1.23905021 + layer.4.v_cache 0.00005228 0.09882973 + layer.4.output 1.37278975 242.95361547 + ------------------------------------------------------------------------------------- + TOTAL 0.59623828 125.72514543 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 43772 +BPFP 0.1804 bits/point +EBPFP 0.3608 equivalent bits/point +MSE 125.725145 +---------------------- -------------------------------------------------------- +Time: 2.930s Load: 0.007s, Pack+Encode: 1.711s, Decode+Unpack: 1.212s +---------------------- -------------------------------------------------------- +💾 Converting with 125.7251 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,020B, BPFP=0.2574 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,984B, BPFP=0.3192 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,292B, BPFP=0.3389 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,396B, BPFP=0.2815 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,960B, BPFP=0.2536 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,140B, BPFP=0.2651 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,772B, BPFP=0.2415 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,364B, BPFP=0.2795 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,320B, BPFP=0.3407 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,604B, BPFP=0.2948 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,068B, BPFP=0.0189 +⌛️ [2/4] FRONTEND: Frontend time: 1.707s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11263794 402.94675333 + layer.0.v_cache 0.00001639 0.04545133 + layer.1.k_cache 0.42101050 24.26187484 + layer.1.v_cache 0.00000577 0.01732024 + layer.2.k_cache 0.00499437 3.05286120 + layer.2.v_cache 0.00001973 0.05068959 + layer.3.k_cache 0.03029772 13.12204430 + layer.3.v_cache 0.00002073 0.06185139 + layer.4.k_cache 0.00068836 1.29991488 + layer.4.v_cache 0.00005112 0.10123195 + layer.4.output 1.25471843 222.06462237 + ------------------------------------------------------------------------------------- + TOTAL 0.55016304 117.61249115 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 46920 +BPFP 0.1767 bits/point +EBPFP 0.3535 equivalent bits/point +MSE 117.612491 +---------------------- -------------------------------------------------------- +Time: 2.928s Load: 0.008s, Pack+Encode: 1.707s, Decode+Unpack: 1.213s +---------------------- -------------------------------------------------------- +💾 Converting with 117.6125 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,692B, BPFP=0.2646 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,180B, BPFP=0.3713 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,216B, BPFP=0.3739 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,508B, BPFP=0.3231 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,792B, BPFP=0.2718 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,200B, BPFP=0.3010 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,716B, BPFP=0.2663 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,188B, BPFP=0.3002 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,476B, BPFP=0.3925 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,984B, BPFP=0.3572 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,040B, BPFP=0.0209 +⌛️ [2/4] FRONTEND: Frontend time: 1.706s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10233516 398.32249857 + layer.0.v_cache 0.00001822 0.04697778 + layer.1.k_cache 0.31275415 25.13709280 + layer.1.v_cache 0.00000587 0.01716796 + layer.2.k_cache 0.01225604 2.89853073 + layer.2.v_cache 0.00002018 0.05302528 + layer.3.k_cache 0.02847941 14.00294649 + layer.3.v_cache 0.00002101 0.06336364 + layer.4.k_cache 0.00067306 1.34889879 + layer.4.v_cache 0.00005947 0.10503617 + layer.4.output 1.40430263 248.53505898 + ------------------------------------------------------------------------------------- + TOTAL 0.60510241 128.33770300 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 46992 +BPFP 0.1981 bits/point +EBPFP 0.3962 equivalent bits/point +MSE 128.337703 +---------------------- -------------------------------------------------------- +Time: 2.927s Load: 0.007s, Pack+Encode: 1.706s, Decode+Unpack: 1.213s +---------------------- -------------------------------------------------------- +💾 Converting with 128.3377 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,652B, BPFP=0.2472 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,768B, BPFP=0.3065 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,520B, BPFP=0.3465 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,456B, BPFP=0.2900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,912B, BPFP=0.2611 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,048B, BPFP=0.2683 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,676B, BPFP=0.2485 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,376B, BPFP=0.2857 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,976B, BPFP=0.3707 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,644B, BPFP=0.3000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,456B, BPFP=0.0186 +⌛️ [2/4] FRONTEND: Frontend time: 2.063s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16298727 398.17293793 + layer.0.v_cache 0.00001770 0.04667150 + layer.1.k_cache 0.61736183 24.39728987 + layer.1.v_cache 0.00000612 0.01754698 + layer.2.k_cache 0.02593859 2.98622993 + layer.2.v_cache 0.00002024 0.05046750 + layer.3.k_cache 0.01326758 13.24382590 + layer.3.v_cache 0.00002270 0.06393641 + layer.4.k_cache 0.00075427 1.34311141 + layer.4.v_cache 0.00005043 0.09830985 + layer.4.output 0.04538776 184.58863277 + ------------------------------------------------------------------------------------- + TOTAL 0.06694947 101.91416216 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 57484 +BPFP 0.1797 bits/point +EBPFP 0.3594 equivalent bits/point +MSE 101.914162 +---------------------- -------------------------------------------------------- +Time: 3.416s Load: 0.010s, Pack+Encode: 2.063s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 101.9142 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 239, 128) +Output shape: (1, 239, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.output: torch.Size([1, 239, 3584]) -> torch.Size([1, 1, 239, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,900B, BPFP=0.2550 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,824B, BPFP=0.3154 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,284B, BPFP=0.3454 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,468B, BPFP=0.2921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,164B, BPFP=0.2722 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,440B, BPFP=0.2903 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,828B, BPFP=0.2503 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,572B, BPFP=0.2989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,444B, BPFP=0.3559 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,748B, BPFP=0.3104 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,272B, BPFP=0.0212 +⌛️ [2/4] FRONTEND: Frontend time: 1.709s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.264s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11863919 400.21541579 + layer.0.v_cache 0.00001702 0.04657469 + layer.1.k_cache 0.43747612 25.43036782 + layer.1.v_cache 0.00000797 0.01845752 + layer.2.k_cache 0.02286501 3.11269567 + layer.2.v_cache 0.00002174 0.05082043 + layer.3.k_cache 0.01903599 13.19606944 + layer.3.v_cache 0.00002123 0.06152881 + layer.4.k_cache 0.00068597 1.33316334 + layer.4.v_cache 0.00006921 0.10068623 + layer.4.output 1.28101462 226.73929692 + ------------------------------------------------------------------------------------- + TOTAL 0.56270246 119.45534460 + (elements=2,080,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2080256 +Total Bytes 47944 +BPFP 0.1844 bits/point +EBPFP 0.3688 equivalent bits/point +MSE 119.455345 +---------------------- -------------------------------------------------------- +Time: 2.982s Load: 0.009s, Pack+Encode: 1.709s, Decode+Unpack: 1.264s +---------------------- -------------------------------------------------------- +💾 Converting with 119.4553 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,016B, BPFP=0.2386 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,540B, BPFP=0.3291 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,280B, BPFP=0.3731 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,300B, BPFP=0.3149 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,536B, BPFP=0.2695 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,152B, BPFP=0.3061 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,340B, BPFP=0.2578 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,684B, BPFP=0.3377 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,604B, BPFP=0.3923 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,048B, BPFP=0.3593 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,376B, BPFP=0.0202 +⌛️ [2/4] FRONTEND: Frontend time: 1.822s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15151563 393.75068322 + layer.0.v_cache 0.00001631 0.04579299 + layer.1.k_cache 0.46755480 24.78897895 + layer.1.v_cache 0.00000667 0.01777505 + layer.2.k_cache 0.02171374 2.98817148 + layer.2.v_cache 0.00002115 0.05245424 + layer.3.k_cache 0.01385514 13.39639340 + layer.3.v_cache 0.00002107 0.06229900 + layer.4.k_cache 0.00069081 1.28958466 + layer.4.v_cache 0.00005638 0.10343248 + layer.4.output 0.00504555 211.16811515 + ------------------------------------------------------------------------------------- + TOTAL 0.04063356 112.62778656 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 55876 +BPFP 0.1953 bits/point +EBPFP 0.3905 equivalent bits/point +MSE 112.627787 +---------------------- -------------------------------------------------------- +Time: 3.172s Load: 0.009s, Pack+Encode: 1.822s, Decode+Unpack: 1.341s +---------------------- -------------------------------------------------------- +💾 Converting with 112.6278 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,828B, BPFP=0.2658 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,440B, BPFP=0.3083 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,176B, BPFP=0.3594 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,504B, BPFP=0.3128 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,832B, BPFP=0.2661 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,068B, BPFP=0.2825 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,548B, BPFP=0.2464 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,492B, BPFP=0.3119 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,560B, BPFP=0.3861 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,684B, BPFP=0.3253 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,128B, BPFP=0.0211 +⌛️ [2/4] FRONTEND: Frontend time: 1.710s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14529029 402.84520833 + layer.0.v_cache 0.00001764 0.04517299 + layer.1.k_cache 0.31089128 26.03524957 + layer.1.v_cache 0.00000630 0.01758219 + layer.2.k_cache 0.01355181 3.02979709 + layer.2.v_cache 0.00002032 0.05075910 + layer.3.k_cache 0.02426839 13.83109375 + layer.3.v_cache 0.00002043 0.06035880 + layer.4.k_cache 0.00068925 1.28496989 + layer.4.v_cache 0.00005165 0.10046835 + layer.4.output 1.36066019 240.81216270 + ------------------------------------------------------------------------------------- + TOTAL 0.58937816 125.46975288 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 46260 +BPFP 0.1890 bits/point +EBPFP 0.3779 equivalent bits/point +MSE 125.469753 +---------------------- -------------------------------------------------------- +Time: 2.924s Load: 0.008s, Pack+Encode: 1.710s, Decode+Unpack: 1.206s +---------------------- -------------------------------------------------------- +💾 Converting with 125.4698 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 262, 128) +Output shape: (1, 262, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.output: torch.Size([1, 262, 3584]) -> torch.Size([1, 1, 262, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,692B, BPFP=0.2202 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,776B, BPFP=0.3445 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,264B, BPFP=0.3736 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,480B, BPFP=0.3268 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,660B, BPFP=0.2779 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,376B, BPFP=0.3206 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,004B, BPFP=0.2388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,232B, BPFP=0.3120 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,512B, BPFP=0.3884 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,968B, BPFP=0.3559 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,448B, BPFP=0.0209 +⌛️ [2/4] FRONTEND: Frontend time: 1.822s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.331s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12917162 380.45252863 + layer.0.v_cache 0.00001830 0.04586899 + layer.1.k_cache 0.40634074 25.77096814 + layer.1.v_cache 0.00000633 0.01724441 + layer.2.k_cache 0.02363680 2.83911785 + layer.2.v_cache 0.00002248 0.05120441 + layer.3.k_cache 0.02343226 13.28835524 + layer.3.v_cache 0.00002042 0.06127165 + layer.4.k_cache 0.00071170 1.28572746 + layer.4.v_cache 0.00005093 0.10103128 + layer.4.output 0.00505940 211.96711423 + ------------------------------------------------------------------------------------- + TOTAL 0.03640161 112.21665398 + (elements=2,280,448) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2280448 +Total Bytes 55412 +BPFP 0.1944 bits/point +EBPFP 0.3888 equivalent bits/point +MSE 112.216654 +---------------------- -------------------------------------------------------- +Time: 3.164s Load: 0.011s, Pack+Encode: 1.822s, Decode+Unpack: 1.331s +---------------------- -------------------------------------------------------- +💾 Converting with 112.2167 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,824B, BPFP=0.2656 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,732B, BPFP=0.3286 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,000B, BPFP=0.3472 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,568B, BPFP=0.3172 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,888B, BPFP=0.2700 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,532B, BPFP=0.3147 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,704B, BPFP=0.2572 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,304B, BPFP=0.2989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,656B, BPFP=0.3928 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,184B, BPFP=0.3600 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,076B, BPFP=0.0206 +⌛️ [2/4] FRONTEND: Frontend time: 1.703s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18074451 402.68034722 + layer.0.v_cache 0.00001764 0.04647379 + layer.1.k_cache 0.33421082 23.88479601 + layer.1.v_cache 0.00000655 0.01751627 + layer.2.k_cache 0.01487562 2.88359131 + layer.2.v_cache 0.00002082 0.05483724 + layer.3.k_cache 0.01319740 13.22576172 + layer.3.v_cache 0.00002096 0.06299534 + layer.4.k_cache 0.00067848 1.31150024 + layer.4.v_cache 0.00005239 0.10262529 + layer.4.output 1.36066546 240.80988095 + ------------------------------------------------------------------------------------- + TOTAL 0.59226373 125.29056536 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 47468 +BPFP 0.1939 bits/point +EBPFP 0.3878 equivalent bits/point +MSE 125.290565 +---------------------- -------------------------------------------------------- +Time: 2.915s Load: 0.010s, Pack+Encode: 1.703s, Decode+Unpack: 1.202s +---------------------- -------------------------------------------------------- +💾 Converting with 125.2906 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 232, 128) +Output shape: (1, 232, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.output: torch.Size([1, 232, 3584]) -> torch.Size([1, 1, 232, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,424B, BPFP=0.2306 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,616B, BPFP=0.3109 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,428B, BPFP=0.3656 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,404B, BPFP=0.2966 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,236B, BPFP=0.2853 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,416B, BPFP=0.2974 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,972B, BPFP=0.2675 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,532B, BPFP=0.3052 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,532B, BPFP=0.3726 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,004B, BPFP=0.3370 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,476B, BPFP=0.0238 +⌛️ [2/4] FRONTEND: Frontend time: 1.708s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11978169 386.84007947 + layer.0.v_cache 0.00001746 0.04596760 + layer.1.k_cache 0.32299256 25.87822644 + layer.1.v_cache 0.00000596 0.01691151 + layer.2.k_cache 0.02091339 3.03827010 + layer.2.v_cache 0.00002092 0.05152641 + layer.3.k_cache 0.01769928 12.96960239 + layer.3.v_cache 0.00002180 0.06213462 + layer.4.k_cache 0.00071208 1.33046643 + layer.4.v_cache 0.00005763 0.10207596 + layer.4.output 1.31963615 233.57108220 + ------------------------------------------------------------------------------------- + TOTAL 0.57174564 121.49016685 + (elements=2,019,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2019328 +Total Bytes 48040 +BPFP 0.1903 bits/point +EBPFP 0.3806 equivalent bits/point +MSE 121.490167 +---------------------- -------------------------------------------------------- +Time: 2.921s Load: 0.010s, Pack+Encode: 1.708s, Decode+Unpack: 1.203s +---------------------- -------------------------------------------------------- +💾 Converting with 121.4902 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,768B, BPFP=0.2676 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,664B, BPFP=0.3312 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,960B, BPFP=0.3523 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,380B, BPFP=0.3111 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,832B, BPFP=0.2722 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,300B, BPFP=0.3054 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,624B, BPFP=0.2574 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,516B, BPFP=0.3207 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,556B, BPFP=0.3946 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,484B, BPFP=0.3185 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,040B, BPFP=0.0207 +⌛️ [2/4] FRONTEND: Frontend time: 1.708s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15639258 405.52176847 + layer.0.v_cache 0.00001700 0.04697749 + layer.1.k_cache 0.33150226 24.41044256 + layer.1.v_cache 0.00000613 0.01751150 + layer.2.k_cache 0.01694729 2.88830372 + layer.2.v_cache 0.00002208 0.05110118 + layer.3.k_cache 0.03858858 14.05314719 + layer.3.v_cache 0.00002090 0.06297487 + layer.4.k_cache 0.00067984 1.31582489 + layer.4.v_cache 0.00005217 0.10231344 + layer.4.output 1.39161012 246.30971997 + ------------------------------------------------------------------------------------- + TOTAL 0.60502939 127.80225912 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 46124 +BPFP 0.1927 bits/point +EBPFP 0.3854 equivalent bits/point +MSE 127.802259 +---------------------- -------------------------------------------------------- +Time: 2.919s Load: 0.008s, Pack+Encode: 1.708s, Decode+Unpack: 1.204s +---------------------- -------------------------------------------------------- +💾 Converting with 127.8023 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,352B, BPFP=0.2491 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,052B, BPFP=0.3464 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,524B, BPFP=0.3734 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,700B, BPFP=0.3262 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,764B, BPFP=0.2727 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,584B, BPFP=0.3196 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,552B, BPFP=0.2605 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,036B, BPFP=0.3455 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,556B, BPFP=0.3752 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,260B, BPFP=0.3583 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,408B, BPFP=0.0197 +⌛️ [2/4] FRONTEND: Frontend time: 1.818s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.331s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13605902 392.25280449 + layer.0.v_cache 0.00001757 0.04606648 + layer.1.k_cache 0.58683352 25.50445355 + layer.1.v_cache 0.00000623 0.01808385 + layer.2.k_cache 0.01500385 2.97039057 + layer.2.v_cache 0.00002229 0.05127514 + layer.3.k_cache 0.00978726 12.95000215 + layer.3.v_cache 0.00001998 0.05908071 + layer.4.k_cache 0.00067570 1.26514298 + layer.4.v_cache 0.00005472 0.09974213 + layer.4.output 0.00490606 203.44206240 + ------------------------------------------------------------------------------------- + TOTAL 0.04604839 109.37126346 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 58788 +BPFP 0.1979 bits/point +EBPFP 0.3958 equivalent bits/point +MSE 109.371263 +---------------------- -------------------------------------------------------- +Time: 3.161s Load: 0.012s, Pack+Encode: 1.818s, Decode+Unpack: 1.331s +---------------------- -------------------------------------------------------- +💾 Converting with 109.3713 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,140B, BPFP=0.2537 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,524B, BPFP=0.2772 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,292B, BPFP=0.2630 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,556B, BPFP=0.2792 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,180B, BPFP=0.1949 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,316B, BPFP=0.2645 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,132B, BPFP=0.1919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,504B, BPFP=0.2760 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,912B, BPFP=0.3010 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,228B, BPFP=0.2591 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,008B, BPFP=0.0176 +⌛️ [2/4] FRONTEND: Frontend time: 1.707s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12967140 406.25214461 + layer.0.v_cache 0.00001601 0.04737696 + layer.1.k_cache 0.49353925 22.91349380 + layer.1.v_cache 0.00000620 0.01755977 + layer.2.k_cache 0.03189284 2.85194020 + layer.2.v_cache 0.00002040 0.05174430 + layer.3.k_cache 0.04285583 12.82024548 + layer.3.v_cache 0.00002221 0.06310954 + layer.4.k_cache 0.00071132 1.25574030 + layer.4.v_cache 0.00005107 0.09473902 + layer.4.output 1.20063613 212.46558123 + ------------------------------------------------------------------------------------- + TOTAL 0.53548467 113.74277427 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 43792 +BPFP 0.1578 bits/point +EBPFP 0.3157 equivalent bits/point +MSE 113.742774 +---------------------- -------------------------------------------------------- +Time: 2.921s Load: 0.009s, Pack+Encode: 1.707s, Decode+Unpack: 1.206s +---------------------- -------------------------------------------------------- +💾 Converting with 113.7428 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,108B, BPFP=0.2578 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,900B, BPFP=0.3075 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,824B, BPFP=0.3027 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,592B, BPFP=0.2882 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,720B, BPFP=0.2334 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,428B, BPFP=0.2779 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,544B, BPFP=0.2224 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,524B, BPFP=0.2839 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,432B, BPFP=0.3409 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,348B, BPFP=0.2728 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,552B, BPFP=0.0229 +⌛️ [2/4] FRONTEND: Frontend time: 1.702s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16529650 409.40304970 + layer.0.v_cache 0.00001613 0.04687052 + layer.1.k_cache 0.54577502 24.66148186 + layer.1.v_cache 0.00000662 0.01844513 + layer.2.k_cache 0.01453520 3.13700334 + layer.2.v_cache 0.00002096 0.05151226 + layer.3.k_cache 0.02409829 13.93008264 + layer.3.v_cache 0.00002071 0.06165527 + layer.4.k_cache 0.00068699 1.37535763 + layer.4.v_cache 0.00004971 0.10093407 + layer.4.output 1.22953407 217.40253873 + ------------------------------------------------------------------------------------- + TOTAL 0.55042615 116.15318609 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 46972 +BPFP 0.1734 bits/point +EBPFP 0.3468 equivalent bits/point +MSE 116.153186 +---------------------- -------------------------------------------------------- +Time: 2.914s Load: 0.011s, Pack+Encode: 1.702s, Decode+Unpack: 1.202s +---------------------- -------------------------------------------------------- +💾 Converting with 116.1532 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,692B, BPFP=0.2242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,932B, BPFP=0.3312 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,292B, BPFP=0.3484 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,440B, BPFP=0.3077 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,288B, BPFP=0.2527 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,524B, BPFP=0.3117 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,192B, BPFP=0.2481 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,720B, BPFP=0.3211 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,608B, BPFP=0.3635 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,264B, BPFP=0.3471 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,564B, BPFP=0.0175 +⌛️ [2/4] FRONTEND: Frontend time: 2.170s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.521s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19043710 382.81522362 + layer.0.v_cache 0.00001603 0.04591967 + layer.1.k_cache 0.64205214 25.55977040 + layer.1.v_cache 0.00000630 0.01822333 + layer.2.k_cache 0.01513053 2.80290796 + layer.2.v_cache 0.00002191 0.05413620 + layer.3.k_cache 0.02775778 13.87052334 + layer.3.v_cache 0.00002125 0.06353764 + layer.4.k_cache 0.00071623 1.28633458 + layer.4.v_cache 0.00005146 0.09915768 + layer.4.output 0.04089398 165.97814275 + ------------------------------------------------------------------------------------- + TOTAL 0.06838051 93.43898433 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 66516 +BPFP 0.1870 bits/point +EBPFP 0.3739 equivalent bits/point +MSE 93.438984 +---------------------- -------------------------------------------------------- +Time: 3.703s Load: 0.012s, Pack+Encode: 2.170s, Decode+Unpack: 1.521s +---------------------- -------------------------------------------------------- +💾 Converting with 93.4390 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,956B, BPFP=0.2736 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,136B, BPFP=0.3388 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,984B, BPFP=0.3304 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,500B, BPFP=0.3037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,760B, BPFP=0.2628 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,748B, BPFP=0.3174 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,508B, BPFP=0.2489 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,820B, BPFP=0.3213 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,288B, BPFP=0.3472 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,228B, BPFP=0.3439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,344B, BPFP=0.0185 +⌛️ [2/4] FRONTEND: Frontend time: 1.862s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.452s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14392723 404.99489289 + layer.0.v_cache 0.00001682 0.04675006 + layer.1.k_cache 0.53349789 24.26624614 + layer.1.v_cache 0.00000638 0.01820074 + layer.2.k_cache 0.01523476 2.85904977 + layer.2.v_cache 0.00002109 0.05222321 + layer.3.k_cache 0.03867486 12.57330344 + layer.3.v_cache 0.00001994 0.06140273 + layer.4.k_cache 0.00070525 1.26899574 + layer.4.v_cache 0.00005078 0.09859115 + layer.4.output 0.00472849 196.24593009 + ------------------------------------------------------------------------------------- + TOTAL 0.04501497 107.05653920 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 58272 +BPFP 0.1893 bits/point +EBPFP 0.3785 equivalent bits/point +MSE 107.056539 +---------------------- -------------------------------------------------------- +Time: 3.322s Load: 0.009s, Pack+Encode: 1.862s, Decode+Unpack: 1.452s +---------------------- -------------------------------------------------------- +💾 Converting with 107.0565 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,536B, BPFP=0.2456 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,648B, BPFP=0.3228 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,184B, BPFP=0.3600 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,328B, BPFP=0.3006 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,892B, BPFP=0.2703 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,288B, BPFP=0.2978 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,460B, BPFP=0.2403 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,540B, BPFP=0.3153 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,304B, BPFP=0.3683 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,612B, BPFP=0.3203 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,088B, BPFP=0.0207 +⌛️ [2/4] FRONTEND: Frontend time: 1.739s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14673459 395.06944444 + layer.0.v_cache 0.00001655 0.04539775 + layer.1.k_cache 0.37581726 26.14684896 + layer.1.v_cache 0.00000603 0.01732081 + layer.2.k_cache 0.01062915 2.96238688 + layer.2.v_cache 0.00002006 0.05132259 + layer.3.k_cache 0.04027582 13.21561740 + layer.3.v_cache 0.00002066 0.06018438 + layer.4.k_cache 0.00066373 1.26802517 + layer.4.v_cache 0.00004861 0.09926506 + layer.4.output 1.36064577 240.80156746 + ------------------------------------------------------------------------------------- + TOTAL 0.59404429 124.97334033 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 45880 +BPFP 0.1874 bits/point +EBPFP 0.3748 equivalent bits/point +MSE 124.973340 +---------------------- -------------------------------------------------------- +Time: 2.967s Load: 0.008s, Pack+Encode: 1.739s, Decode+Unpack: 1.219s +---------------------- -------------------------------------------------------- +💾 Converting with 124.9733 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,712B, BPFP=0.2479 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,724B, BPFP=0.3154 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,608B, BPFP=0.3745 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,328B, BPFP=0.2890 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,196B, BPFP=0.2802 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,220B, BPFP=0.2818 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,980B, BPFP=0.2658 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,488B, BPFP=0.2997 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,372B, BPFP=0.3587 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,624B, BPFP=0.3088 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,332B, BPFP=0.0222 +⌛️ [2/4] FRONTEND: Frontend time: 1.988s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.410s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12385076 392.26138488 + layer.0.v_cache 0.00002004 0.04950307 + layer.1.k_cache 0.38583452 25.52458517 + layer.1.v_cache 0.00000646 0.01866288 + layer.2.k_cache 0.02500911 3.04778270 + layer.2.v_cache 0.00002014 0.05246257 + layer.3.k_cache 0.03425702 13.79059725 + layer.3.v_cache 0.00002057 0.06516043 + layer.4.k_cache 0.00068928 1.31750280 + layer.4.v_cache 0.00004882 0.10116149 + layer.4.output 1.30836928 231.57541590 + ------------------------------------------------------------------------------------- + TOTAL 0.57225539 121.01510086 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 47584 +BPFP 0.1869 bits/point +EBPFP 0.3738 equivalent bits/point +MSE 121.015101 +---------------------- -------------------------------------------------------- +Time: 3.408s Load: 0.009s, Pack+Encode: 1.988s, Decode+Unpack: 1.410s +---------------------- -------------------------------------------------------- +💾 Converting with 121.0151 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 247, 128) +Output shape: (1, 247, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.output: torch.Size([1, 247, 3584]) -> torch.Size([1, 1, 247, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,764B, BPFP=0.2381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,116B, BPFP=0.3236 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,208B, BPFP=0.3295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,412B, BPFP=0.2791 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,796B, BPFP=0.2401 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,284B, BPFP=0.2710 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,640B, BPFP=0.2303 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,696B, BPFP=0.2971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,476B, BPFP=0.3464 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,588B, BPFP=0.2902 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,944B, BPFP=0.0266 +⌛️ [2/4] FRONTEND: Frontend time: 2.146s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.426s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13907220 399.49003669 + layer.0.v_cache 0.00001735 0.04732050 + layer.1.k_cache 0.47208408 24.85363819 + layer.1.v_cache 0.00000611 0.01763109 + layer.2.k_cache 0.03408500 3.04671092 + layer.2.v_cache 0.00001989 0.04975534 + layer.3.k_cache 0.02387486 13.16203821 + layer.3.v_cache 0.00002031 0.06082233 + layer.4.k_cache 0.00069623 1.34443263 + layer.4.v_cache 0.00004922 0.09699859 + layer.4.output 1.23951819 219.38195850 + ------------------------------------------------------------------------------------- + TOTAL 0.54979721 116.34371141 + (elements=2,149,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2149888 +Total Bytes 47924 +BPFP 0.1783 bits/point +EBPFP 0.3567 equivalent bits/point +MSE 116.343711 +---------------------- -------------------------------------------------------- +Time: 3.581s Load: 0.009s, Pack+Encode: 2.146s, Decode+Unpack: 1.426s +---------------------- -------------------------------------------------------- +💾 Converting with 116.3437 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 427, 128) +Output shape: (1, 427, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.output: torch.Size([1, 427, 3584]) -> torch.Size([1, 1, 427, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,968B, BPFP=0.2550 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,008B, BPFP=0.2930 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,764B, BPFP=0.3207 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,604B, BPFP=0.2782 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,528B, BPFP=0.2389 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,832B, BPFP=0.2500 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,132B, BPFP=0.2244 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,756B, BPFP=0.2838 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,504B, BPFP=0.3478 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,916B, BPFP=0.2897 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,160B, BPFP=0.0165 +⌛️ [2/4] FRONTEND: Frontend time: 2.579s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.796s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15738957 402.47438525 + layer.0.v_cache 0.00001693 0.04443773 + layer.1.k_cache 1.00627419 25.10104792 + layer.1.v_cache 0.00000663 0.01758400 + layer.2.k_cache 0.03590691 2.81826046 + layer.2.v_cache 0.00002108 0.04872560 + layer.3.k_cache 0.02176077 13.18444910 + layer.3.v_cache 0.00002004 0.05701970 + layer.4.k_cache 0.00078997 1.28782579 + layer.4.v_cache 0.00005212 0.09432508 + layer.4.output 0.00621442 129.91226163 + ------------------------------------------------------------------------------------- + TOTAL 0.07445524 79.67728776 + (elements=3,716,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3716608 +Total Bytes 79172 +BPFP 0.1704 bits/point +EBPFP 0.3408 equivalent bits/point +MSE 79.677288 +---------------------- -------------------------------------------------------- +Time: 4.390s Load: 0.015s, Pack+Encode: 2.579s, Decode+Unpack: 1.796s +---------------------- -------------------------------------------------------- +💾 Converting with 79.6773 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,480B, BPFP=0.2583 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,224B, BPFP=0.3589 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,344B, BPFP=0.3658 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,992B, BPFP=0.3455 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,756B, BPFP=0.2742 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,288B, BPFP=0.3049 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,336B, BPFP=0.2500 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,572B, BPFP=0.3213 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,664B, BPFP=0.3842 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,748B, BPFP=0.3314 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,532B, BPFP=0.0209 +⌛️ [2/4] FRONTEND: Frontend time: 2.133s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.556s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15504150 389.87205950 + layer.0.v_cache 0.00001776 0.04496486 + layer.1.k_cache 0.59434689 25.34503647 + layer.1.v_cache 0.00000612 0.01687142 + layer.2.k_cache 0.01687181 2.99167422 + layer.2.v_cache 0.00002031 0.04887810 + layer.3.k_cache 0.01346795 13.07044043 + layer.3.v_cache 0.00002023 0.05918736 + layer.4.k_cache 0.00071492 1.26052733 + layer.4.v_cache 0.00006004 0.09283639 + layer.4.output 0.00488892 204.93715406 + ------------------------------------------------------------------------------------- + TOTAL 0.04792882 109.84485615 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 57936 +BPFP 0.1965 bits/point +EBPFP 0.3930 equivalent bits/point +MSE 109.844856 +---------------------- -------------------------------------------------------- +Time: 3.698s Load: 0.009s, Pack+Encode: 2.133s, Decode+Unpack: 1.556s +---------------------- -------------------------------------------------------- +💾 Converting with 109.8449 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 361, 128) +Output shape: (1, 361, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.output: torch.Size([1, 361, 3584]) -> torch.Size([1, 1, 361, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,692B, BPFP=0.2464 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,584B, BPFP=0.3283 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,596B, BPFP=0.3288 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,768B, BPFP=0.2929 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,616B, BPFP=0.2431 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,300B, BPFP=0.2727 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,372B, BPFP=0.2325 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,824B, BPFP=0.2954 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,236B, BPFP=0.3565 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,624B, BPFP=0.2867 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,980B, BPFP=0.0184 +⌛️ [2/4] FRONTEND: Frontend time: 2.120s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.625s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21057716 399.74896122 + layer.0.v_cache 0.00001554 0.04188485 + layer.1.k_cache 0.82271198 24.35568516 + layer.1.v_cache 0.00000592 0.01615557 + layer.2.k_cache 0.01184860 2.97559439 + layer.2.v_cache 0.00001987 0.04851757 + layer.3.k_cache 0.01345186 13.21370217 + layer.3.v_cache 0.00002111 0.05618560 + layer.4.k_cache 0.00076976 1.27116390 + layer.4.v_cache 0.00005077 0.09112569 + layer.4.output 0.03703259 150.34387366 + ------------------------------------------------------------------------------------- + TOTAL 0.07757063 87.89565246 + (elements=3,142,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3142144 +Total Bytes 69592 +BPFP 0.1772 bits/point +EBPFP 0.3544 equivalent bits/point +MSE 87.895652 +---------------------- -------------------------------------------------------- +Time: 3.757s Load: 0.012s, Pack+Encode: 2.120s, Decode+Unpack: 1.625s +---------------------- -------------------------------------------------------- +💾 Converting with 87.8957 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,624B, BPFP=0.2562 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,384B, BPFP=0.3537 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,120B, BPFP=0.3391 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,736B, BPFP=0.3178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,732B, BPFP=0.2622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,868B, BPFP=0.3251 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,420B, BPFP=0.2449 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,652B, BPFP=0.3132 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,568B, BPFP=0.3639 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,904B, BPFP=0.3271 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,320B, BPFP=0.0184 +⌛️ [2/4] FRONTEND: Frontend time: 2.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.875s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21757852 398.02316046 + layer.0.v_cache 0.00001544 0.04016749 + layer.1.k_cache 0.59705023 24.42136248 + layer.1.v_cache 0.00000603 0.01598520 + layer.2.k_cache 0.00787334 2.90000061 + layer.2.v_cache 0.00001970 0.04610023 + layer.3.k_cache 0.01914395 13.58231695 + layer.3.v_cache 0.00001916 0.05304856 + layer.4.k_cache 0.00071861 1.23096926 + layer.4.v_cache 0.00005003 0.09233286 + layer.4.output 0.00471288 196.93461879 + ------------------------------------------------------------------------------------- + TOTAL 0.05149795 106.99692798 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 58328 +BPFP 0.1901 bits/point +EBPFP 0.3802 equivalent bits/point +MSE 106.996928 +---------------------- -------------------------------------------------------- +Time: 4.041s Load: 0.010s, Pack+Encode: 2.157s, Decode+Unpack: 1.875s +---------------------- -------------------------------------------------------- +💾 Converting with 106.9969 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 372, 128) +Output shape: (1, 372, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.output: torch.Size([1, 372, 3584]) -> torch.Size([1, 1, 372, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,924B, BPFP=0.2488 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,084B, BPFP=0.2975 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,472B, BPFP=0.3138 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,628B, BPFP=0.2784 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,664B, BPFP=0.2379 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,980B, BPFP=0.2512 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,328B, BPFP=0.2238 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,600B, BPFP=0.2772 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,144B, BPFP=0.3421 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,944B, BPFP=0.2917 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,652B, BPFP=0.0159 +⌛️ [2/4] FRONTEND: Frontend time: 2.827s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.940s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18192377 399.34236391 + layer.0.v_cache 0.00001786 0.04462377 + layer.1.k_cache 0.82695811 24.89450500 + layer.1.v_cache 0.00000687 0.01794541 + layer.2.k_cache 0.01216373 3.05055614 + layer.2.v_cache 0.00002073 0.05023322 + layer.3.k_cache 0.01743993 12.58155872 + layer.3.v_cache 0.00002103 0.05856653 + layer.4.k_cache 0.00069676 1.30183706 + layer.4.v_cache 0.00005544 0.09421671 + layer.4.output 0.03606073 145.93373176 + ------------------------------------------------------------------------------------- + TOTAL 0.07598408 86.05720758 + (elements=3,237,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3237888 +Total Bytes 68420 +BPFP 0.1690 bits/point +EBPFP 0.3381 equivalent bits/point +MSE 86.057208 +---------------------- -------------------------------------------------------- +Time: 4.783s Load: 0.016s, Pack+Encode: 2.827s, Decode+Unpack: 1.940s +---------------------- -------------------------------------------------------- +💾 Converting with 86.0572 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,880B, BPFP=0.2361 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,452B, BPFP=0.3605 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,184B, BPFP=0.3475 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,388B, BPFP=0.3090 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,608B, BPFP=0.2713 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,836B, BPFP=0.3307 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,908B, BPFP=0.2374 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,444B, BPFP=0.3117 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,736B, BPFP=0.3742 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,248B, BPFP=0.3506 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,756B, BPFP=0.0190 +⌛️ [2/4] FRONTEND: Frontend time: 2.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.831s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16470559 387.21768576 + layer.0.v_cache 0.00001558 0.04418040 + layer.1.k_cache 0.70577214 24.80680994 + layer.1.v_cache 0.00000663 0.01709140 + layer.2.k_cache 0.01487649 2.94332404 + layer.2.v_cache 0.00002061 0.05095726 + layer.3.k_cache 0.03012228 13.33752782 + layer.3.v_cache 0.00002030 0.05833951 + layer.4.k_cache 0.00074575 1.25631582 + layer.4.v_cache 0.00005065 0.09323551 + layer.4.output 0.04139163 168.03290856 + ------------------------------------------------------------------------------------- + TOTAL 0.07094573 94.47387220 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 67440 +BPFP 0.1919 bits/point +EBPFP 0.3838 equivalent bits/point +MSE 94.473872 +---------------------- -------------------------------------------------------- +Time: 4.334s Load: 0.012s, Pack+Encode: 2.491s, Decode+Unpack: 1.831s +---------------------- -------------------------------------------------------- +💾 Converting with 94.4739 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,288B, BPFP=0.2528 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,104B, BPFP=0.3599 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,456B, BPFP=0.3807 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,776B, BPFP=0.3406 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,388B, BPFP=0.2587 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,464B, BPFP=0.3222 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,148B, BPFP=0.2446 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,228B, BPFP=0.3083 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,696B, BPFP=0.3948 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,820B, BPFP=0.3432 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,384B, BPFP=0.0201 +⌛️ [2/4] FRONTEND: Frontend time: 2.151s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.549s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16585089 393.81193986 + layer.0.v_cache 0.00001549 0.04388631 + layer.1.k_cache 0.52552974 25.59198113 + layer.1.v_cache 0.00000676 0.01697261 + layer.2.k_cache 0.01473994 2.94974273 + layer.2.v_cache 0.00002188 0.05141228 + layer.3.k_cache 0.05129043 13.60296470 + layer.3.v_cache 0.00002012 0.05914326 + layer.4.k_cache 0.00075129 1.29320702 + layer.4.v_cache 0.00005717 0.09664430 + layer.4.output 0.00501330 209.52129380 + ------------------------------------------------------------------------------------- + TOTAL 0.04666922 112.00982064 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 56752 +BPFP 0.1968 bits/point +EBPFP 0.3937 equivalent bits/point +MSE 112.009821 +---------------------- -------------------------------------------------------- +Time: 3.710s Load: 0.010s, Pack+Encode: 2.151s, Decode+Unpack: 1.549s +---------------------- -------------------------------------------------------- +💾 Converting with 112.0098 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 368, 128) +Output shape: (1, 368, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.output: torch.Size([1, 368, 3584]) -> torch.Size([1, 1, 368, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,056B, BPFP=0.2571 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,344B, BPFP=0.3118 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,488B, BPFP=0.3179 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,716B, BPFP=0.2852 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,568B, BPFP=0.2364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,048B, BPFP=0.2568 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,360B, BPFP=0.2276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,488B, BPFP=0.2755 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,176B, BPFP=0.3471 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,208B, BPFP=0.3060 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,672B, BPFP=0.0162 +⌛️ [2/4] FRONTEND: Frontend time: 2.205s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.628s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14147143 402.52819293 + layer.0.v_cache 0.00001611 0.04448400 + layer.1.k_cache 0.89062865 24.77717391 + layer.1.v_cache 0.00000637 0.01692831 + layer.2.k_cache 0.02201829 2.98631021 + layer.2.v_cache 0.00002088 0.05011040 + layer.3.k_cache 0.00965744 12.64885944 + layer.3.v_cache 0.00002101 0.05999777 + layer.4.k_cache 0.00077426 1.29011361 + layer.4.v_cache 0.00005366 0.09792840 + layer.4.output 0.03638816 147.50029115 + ------------------------------------------------------------------------------------- + TOTAL 0.07761090 86.88247865 + (elements=3,203,072) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3203072 +Total Bytes 69124 +BPFP 0.1726 bits/point +EBPFP 0.3453 equivalent bits/point +MSE 86.882479 +---------------------- -------------------------------------------------------- +Time: 3.846s Load: 0.013s, Pack+Encode: 2.205s, Decode+Unpack: 1.628s +---------------------- -------------------------------------------------------- +💾 Converting with 86.8825 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 453, 128) +Output shape: (1, 453, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.output: torch.Size([1, 453, 3584]) -> torch.Size([1, 1, 453, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,248B, BPFP=0.2500 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,808B, BPFP=0.3383 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,416B, BPFP=0.3248 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,812B, BPFP=0.3039 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,484B, BPFP=0.2236 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,652B, BPFP=0.2639 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,396B, BPFP=0.2206 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,676B, BPFP=0.2993 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,820B, BPFP=0.3732 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,764B, BPFP=0.3368 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,176B, BPFP=0.0156 +⌛️ [2/4] FRONTEND: Frontend time: 2.935s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.924s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13665763 401.55539459 + layer.0.v_cache 0.00001897 0.04601798 + layer.1.k_cache 1.07282356 25.11622603 + layer.1.v_cache 0.00000630 0.01695111 + layer.2.k_cache 0.02241544 2.72007849 + layer.2.v_cache 0.00001979 0.05009233 + layer.3.k_cache 0.01806638 12.88441100 + layer.3.v_cache 0.00002079 0.06142458 + layer.4.k_cache 0.00076567 1.30071631 + layer.4.v_cache 0.00005104 0.09819400 + layer.4.output 0.00584071 122.43945128 + ------------------------------------------------------------------------------------- + TOTAL 0.07598415 76.52503914 + (elements=3,942,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3942912 +Total Bytes 88252 +BPFP 0.1791 bits/point +EBPFP 0.3581 equivalent bits/point +MSE 76.525039 +---------------------- -------------------------------------------------------- +Time: 4.874s Load: 0.015s, Pack+Encode: 2.935s, Decode+Unpack: 1.924s +---------------------- -------------------------------------------------------- +💾 Converting with 76.5250 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 402, 128) +Output shape: (1, 402, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.output: torch.Size([1, 402, 3584]) -> torch.Size([1, 1, 402, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,436B, BPFP=0.2502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,508B, BPFP=0.3307 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,460B, BPFP=0.3288 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,792B, BPFP=0.3029 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,272B, BPFP=0.2438 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,492B, BPFP=0.2912 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,920B, BPFP=0.2301 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,400B, BPFP=0.2876 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,088B, BPFP=0.3532 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,012B, BPFP=0.3114 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,028B, BPFP=0.0168 +⌛️ [2/4] FRONTEND: Frontend time: 2.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.853s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18062694 394.09060168 + layer.0.v_cache 0.00001752 0.04477367 + layer.1.k_cache 0.89169403 24.12958401 + layer.1.v_cache 0.00000593 0.01664606 + layer.2.k_cache 0.03854401 2.81770894 + layer.2.v_cache 0.00001967 0.04931101 + layer.3.k_cache 0.02336847 13.65792619 + layer.3.v_cache 0.00001924 0.05749033 + layer.4.k_cache 0.00085529 1.25499547 + layer.4.v_cache 0.00004959 0.09178620 + layer.4.output 0.00648942 137.96135394 + ------------------------------------------------------------------------------------- + TOTAL 0.06944863 82.46707654 + (elements=3,499,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3499008 +Total Bytes 78408 +BPFP 0.1793 bits/point +EBPFP 0.3585 equivalent bits/point +MSE 82.467077 +---------------------- -------------------------------------------------------- +Time: 4.366s Load: 0.013s, Pack+Encode: 2.500s, Decode+Unpack: 1.853s +---------------------- -------------------------------------------------------- +💾 Converting with 82.4671 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,284B, BPFP=0.2399 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,932B, BPFP=0.3322 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,352B, BPFP=0.3557 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,852B, BPFP=0.3277 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,852B, BPFP=0.2717 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,708B, BPFP=0.3197 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,532B, BPFP=0.2538 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,792B, BPFP=0.3244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,392B, BPFP=0.3580 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,832B, BPFP=0.3266 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,304B, BPFP=0.0184 +⌛️ [2/4] FRONTEND: Frontend time: 2.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.581s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13958113 388.71973006 + layer.0.v_cache 0.00001688 0.04611034 + layer.1.k_cache 0.56864935 25.26161024 + layer.1.v_cache 0.00000621 0.01796416 + layer.2.k_cache 0.01018704 2.91683665 + layer.2.v_cache 0.00002046 0.05233484 + layer.3.k_cache 0.04374025 13.11464984 + layer.3.v_cache 0.00002178 0.06199598 + layer.4.k_cache 0.00068741 1.30519656 + layer.4.v_cache 0.00005272 0.10277187 + layer.4.output 0.00479460 199.06573221 + ------------------------------------------------------------------------------------- + TOTAL 0.04685444 107.35643094 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 57832 +BPFP 0.1905 bits/point +EBPFP 0.3810 equivalent bits/point +MSE 107.356431 +---------------------- -------------------------------------------------------- +Time: 3.728s Load: 0.016s, Pack+Encode: 2.132s, Decode+Unpack: 1.581s +---------------------- -------------------------------------------------------- +💾 Converting with 107.3564 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 381, 128) +Output shape: (1, 381, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.output: torch.Size([1, 381, 3584]) -> torch.Size([1, 1, 381, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,092B, BPFP=0.2498 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,896B, BPFP=0.2828 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,940B, BPFP=0.2846 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,620B, BPFP=0.2715 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,904B, BPFP=0.2011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,808B, BPFP=0.2792 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,640B, BPFP=0.1903 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,816B, BPFP=0.2795 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,516B, BPFP=0.3082 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,408B, BPFP=0.2628 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,940B, BPFP=0.0172 +⌛️ [2/4] FRONTEND: Frontend time: 2.596s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15713709 413.05027887 + layer.0.v_cache 0.00001719 0.04583837 + layer.1.k_cache 0.86767426 24.88322260 + layer.1.v_cache 0.00000622 0.01772180 + layer.2.k_cache 0.01635739 2.83610475 + layer.2.v_cache 0.00002098 0.05285308 + layer.3.k_cache 0.01709203 12.82101224 + layer.3.v_cache 0.00002073 0.06193924 + layer.4.k_cache 0.00073405 1.31938688 + layer.4.v_cache 0.00005083 0.09722634 + layer.4.output 0.03519400 142.43092660 + ------------------------------------------------------------------------------------- + TOTAL 0.07679228 85.42365120 + (elements=3,316,224) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3316224 +Total Bytes 66580 +BPFP 0.1606 bits/point +EBPFP 0.3212 equivalent bits/point +MSE 85.423651 +---------------------- -------------------------------------------------------- +Time: 4.628s Load: 0.013s, Pack+Encode: 2.596s, Decode+Unpack: 2.018s +---------------------- -------------------------------------------------------- +💾 Converting with 85.4237 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,904B, BPFP=0.2607 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,612B, BPFP=0.3080 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,500B, BPFP=0.3673 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,404B, BPFP=0.2941 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,188B, BPFP=0.2796 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,224B, BPFP=0.2821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,872B, BPFP=0.2585 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,264B, BPFP=0.2847 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,568B, BPFP=0.3718 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,684B, BPFP=0.3128 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,316B, BPFP=0.0221 +⌛️ [2/4] FRONTEND: Frontend time: 2.012s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.460s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422791 402.08877537 + layer.0.v_cache 0.00001930 0.04647983 + layer.1.k_cache 0.38928963 25.41390391 + layer.1.v_cache 0.00000631 0.01786765 + layer.2.k_cache 0.01623082 3.01131915 + layer.2.v_cache 0.00002095 0.05144937 + layer.3.k_cache 0.01313608 12.82385254 + layer.3.v_cache 0.00002100 0.06172898 + layer.4.k_cache 0.00066994 1.28911610 + layer.4.v_cache 0.00005661 0.09973921 + layer.4.output 1.30836332 231.55420101 + ------------------------------------------------------------------------------------- + TOTAL 0.57013069 121.51668466 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 47536 +BPFP 0.1867 bits/point +EBPFP 0.3734 equivalent bits/point +MSE 121.516685 +---------------------- -------------------------------------------------------- +Time: 3.479s Load: 0.007s, Pack+Encode: 2.012s, Decode+Unpack: 1.460s +---------------------- -------------------------------------------------------- +💾 Converting with 121.5167 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,256B, BPFP=0.2818 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,612B, BPFP=0.3053 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,456B, BPFP=0.3612 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,412B, BPFP=0.2921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,144B, BPFP=0.2744 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,236B, BPFP=0.2805 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,964B, BPFP=0.2624 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,532B, BPFP=0.3001 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,388B, BPFP=0.3567 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,624B, BPFP=0.3061 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,120B, BPFP=0.0201 +⌛️ [2/4] FRONTEND: Frontend time: 2.550s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.743s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13591524 400.60679290 + layer.0.v_cache 0.00001656 0.04534358 + layer.1.k_cache 0.34520291 25.72793424 + layer.1.v_cache 0.00000598 0.01714954 + layer.2.k_cache 0.01250870 3.06167447 + layer.2.v_cache 0.00001982 0.04833531 + layer.3.k_cache 0.02633334 13.48203808 + layer.3.v_cache 0.00002177 0.06161663 + layer.4.k_cache 0.00067446 1.28407805 + layer.4.v_cache 0.00005295 0.09840534 + layer.4.output 1.29725540 229.60839134 + ------------------------------------------------------------------------------------- + TOTAL 0.56479644 120.68777103 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 47744 +BPFP 0.1859 bits/point +EBPFP 0.3719 equivalent bits/point +MSE 120.687771 +---------------------- -------------------------------------------------------- +Time: 4.304s Load: 0.011s, Pack+Encode: 2.550s, Decode+Unpack: 1.743s +---------------------- -------------------------------------------------------- +💾 Converting with 120.6878 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,812B, BPFP=0.2745 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,836B, BPFP=0.3482 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,388B, BPFP=0.3880 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,620B, BPFP=0.3327 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,112B, BPFP=0.2961 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,424B, BPFP=0.3185 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,708B, BPFP=0.2670 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,748B, BPFP=0.3419 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,568B, BPFP=0.4009 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,556B, BPFP=0.3281 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,064B, BPFP=0.0212 +⌛️ [2/4] FRONTEND: Frontend time: 2.462s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.703s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150146 404.20074885 + layer.0.v_cache 0.00001564 0.04486188 + layer.1.k_cache 0.36060091 25.32709893 + layer.1.v_cache 0.00000598 0.01701377 + layer.2.k_cache 0.01419523 2.79029220 + layer.2.v_cache 0.00002081 0.04951445 + layer.3.k_cache 0.03680107 13.30395395 + layer.3.v_cache 0.00002092 0.05812176 + layer.4.k_cache 0.00067164 1.25238895 + layer.4.v_cache 0.00004838 0.09660243 + layer.4.output 1.41076765 249.66647465 + ------------------------------------------------------------------------------------- + TOTAL 0.61289739 129.10623057 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 47836 +BPFP 0.2026 bits/point +EBPFP 0.4052 equivalent bits/point +MSE 129.106231 +---------------------- -------------------------------------------------------- +Time: 4.175s Load: 0.010s, Pack+Encode: 2.462s, Decode+Unpack: 1.703s +---------------------- -------------------------------------------------------- +💾 Converting with 129.1062 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 343, 128) +Output shape: (1, 343, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.output: torch.Size([1, 343, 3584]) -> torch.Size([1, 1, 343, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,508B, BPFP=0.2509 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,364B, BPFP=0.3355 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,168B, BPFP=0.3265 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,440B, BPFP=0.2934 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,736B, BPFP=0.2613 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,672B, BPFP=0.3039 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,184B, BPFP=0.2362 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,964B, BPFP=0.3172 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,148B, BPFP=0.3712 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,940B, BPFP=0.3161 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,488B, BPFP=0.0162 +⌛️ [2/4] FRONTEND: Frontend time: 2.330s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.681s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16677478 398.97603863 + layer.0.v_cache 0.00001698 0.04582331 + layer.1.k_cache 0.72144124 24.88214627 + layer.1.v_cache 0.00000659 0.01817571 + layer.2.k_cache 0.01351720 2.95062701 + layer.2.v_cache 0.00002117 0.05190730 + layer.3.k_cache 0.02347133 12.88135762 + layer.3.v_cache 0.00002006 0.06049622 + layer.4.k_cache 0.00069292 1.28889052 + layer.4.v_cache 0.00005144 0.09767555 + layer.4.output 0.03905215 158.25756195 + ------------------------------------------------------------------------------------- + TOTAL 0.07055169 91.12094540 + (elements=2,985,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2985472 +Total Bytes 68612 +BPFP 0.1839 bits/point +EBPFP 0.3677 equivalent bits/point +MSE 91.120945 +---------------------- -------------------------------------------------------- +Time: 4.024s Load: 0.014s, Pack+Encode: 2.330s, Decode+Unpack: 1.681s +---------------------- -------------------------------------------------------- +💾 Converting with 91.1209 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,032B, BPFP=0.2291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,352B, BPFP=0.3609 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,116B, BPFP=0.3475 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,784B, BPFP=0.3286 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,836B, BPFP=0.2748 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,668B, BPFP=0.3220 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,452B, BPFP=0.2530 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,836B, BPFP=0.3316 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,840B, BPFP=0.3886 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,236B, BPFP=0.3543 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,432B, BPFP=0.0197 +⌛️ [2/4] FRONTEND: Frontend time: 2.278s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.619s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13317820 378.58818182 + layer.0.v_cache 0.00001784 0.04466621 + layer.1.k_cache 0.51782770 24.68484197 + layer.1.v_cache 0.00000612 0.01699607 + layer.2.k_cache 0.01694398 2.95086936 + layer.2.v_cache 0.00002016 0.05027533 + layer.3.k_cache 0.01920085 12.94425071 + layer.3.v_cache 0.00002016 0.05874049 + layer.4.k_cache 0.00071488 1.28244052 + layer.4.v_cache 0.00005349 0.09978642 + layer.4.output 0.00484358 201.96358766 + ------------------------------------------------------------------------------------- + TOTAL 0.04246403 107.90977427 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 58584 +BPFP 0.1958 bits/point +EBPFP 0.3916 equivalent bits/point +MSE 107.909774 +---------------------- -------------------------------------------------------- +Time: 3.906s Load: 0.010s, Pack+Encode: 2.278s, Decode+Unpack: 1.619s +---------------------- -------------------------------------------------------- +💾 Converting with 107.9098 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 374, 128) +Output shape: (1, 374, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.output: torch.Size([1, 374, 3584]) -> torch.Size([1, 1, 374, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,988B, BPFP=0.2502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,176B, BPFP=0.2998 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,340B, BPFP=0.3067 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,060B, BPFP=0.2950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,284B, BPFP=0.2208 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,008B, BPFP=0.2510 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,172B, BPFP=0.2161 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,688B, BPFP=0.2794 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,960B, BPFP=0.3326 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,584B, BPFP=0.2751 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,260B, BPFP=0.0195 +⌛️ [2/4] FRONTEND: Frontend time: 2.154s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.640s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18781370 404.74176972 + layer.0.v_cache 0.00001670 0.04498223 + layer.1.k_cache 0.85217587 25.24568641 + layer.1.v_cache 0.00000605 0.01741112 + layer.2.k_cache 0.02919085 2.94057983 + layer.2.v_cache 0.00002159 0.04783233 + layer.3.k_cache 0.02976600 13.27426523 + layer.3.v_cache 0.00002022 0.05776372 + layer.4.k_cache 0.00072409 1.28770496 + layer.4.v_cache 0.00004949 0.09361200 + layer.4.output 0.03578218 145.10945856 + ------------------------------------------------------------------------------------- + TOTAL 0.07942705 86.08928338 + (elements=3,255,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3255296 +Total Bytes 68520 +BPFP 0.1684 bits/point +EBPFP 0.3368 equivalent bits/point +MSE 86.089283 +---------------------- -------------------------------------------------------- +Time: 3.806s Load: 0.011s, Pack+Encode: 2.154s, Decode+Unpack: 1.640s +---------------------- -------------------------------------------------------- +💾 Converting with 86.0893 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,960B, BPFP=0.2567 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,992B, BPFP=0.3237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,440B, BPFP=0.3527 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,808B, BPFP=0.3117 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,176B, BPFP=0.2707 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,156B, BPFP=0.2695 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,900B, BPFP=0.2529 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,676B, BPFP=0.3032 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,972B, BPFP=0.3872 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,816B, BPFP=0.3122 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,064B, BPFP=0.0191 +⌛️ [2/4] FRONTEND: Frontend time: 2.019s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.434s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13979621 398.16688278 + layer.0.v_cache 0.00001743 0.04451973 + layer.1.k_cache 0.46959303 24.70947671 + layer.1.v_cache 0.00000602 0.01744880 + layer.2.k_cache 0.01642122 3.22046760 + layer.2.v_cache 0.00002053 0.05090944 + layer.3.k_cache 0.03876302 13.75541364 + layer.3.v_cache 0.00002036 0.06036839 + layer.4.k_cache 0.00069112 1.32531890 + layer.4.v_cache 0.00005399 0.10163161 + layer.4.output 1.27033806 224.84525044 + ------------------------------------------------------------------------------------- + TOTAL 0.56222055 118.55112886 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 48960 +BPFP 0.1867 bits/point +EBPFP 0.3734 equivalent bits/point +MSE 118.551129 +---------------------- -------------------------------------------------------- +Time: 3.462s Load: 0.009s, Pack+Encode: 2.019s, Decode+Unpack: 1.434s +---------------------- -------------------------------------------------------- +💾 Converting with 118.5511 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,724B, BPFP=0.2563 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,916B, BPFP=0.3384 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,256B, BPFP=0.3618 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,640B, BPFP=0.3194 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,128B, BPFP=0.2841 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,472B, BPFP=0.3078 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,872B, BPFP=0.2665 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,568B, BPFP=0.3144 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,904B, BPFP=0.3376 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,820B, BPFP=0.3318 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,224B, BPFP=0.0219 +⌛️ [2/4] FRONTEND: Frontend time: 1.986s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13893990 394.19878855 + layer.0.v_cache 0.00001711 0.04563156 + layer.1.k_cache 0.36820191 25.39057338 + layer.1.v_cache 0.00000637 0.01792019 + layer.2.k_cache 0.01566778 2.95896018 + layer.2.v_cache 0.00002161 0.05207543 + layer.3.k_cache 0.01884081 13.19372182 + layer.3.v_cache 0.00002076 0.06379500 + layer.4.k_cache 0.00070054 1.29204220 + layer.4.v_cache 0.00005210 0.10131440 + layer.4.output 1.34866800 238.69125629 + ------------------------------------------------------------------------------------- + TOTAL 0.58724382 124.00903628 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 47524 +BPFP 0.1924 bits/point +EBPFP 0.3848 equivalent bits/point +MSE 124.009036 +---------------------- -------------------------------------------------------- +Time: 3.376s Load: 0.010s, Pack+Encode: 1.986s, Decode+Unpack: 1.381s +---------------------- -------------------------------------------------------- +💾 Converting with 124.0090 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,484B, BPFP=0.2624 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,456B, BPFP=0.3778 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,476B, BPFP=0.3790 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,836B, BPFP=0.3415 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,584B, BPFP=0.2683 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,684B, BPFP=0.3326 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,520B, BPFP=0.2645 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,740B, BPFP=0.3359 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,648B, BPFP=0.3890 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,356B, BPFP=0.3720 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,448B, BPFP=0.0205 +⌛️ [2/4] FRONTEND: Frontend time: 2.134s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.588s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11943251 392.99215824 + layer.0.v_cache 0.00001561 0.04447255 + layer.1.k_cache 0.49619702 25.36473695 + layer.1.v_cache 0.00000608 0.01707473 + layer.2.k_cache 0.01349212 2.89782578 + layer.2.v_cache 0.00002006 0.04929125 + layer.3.k_cache 0.01080775 13.30230461 + layer.3.v_cache 0.00002073 0.05842878 + layer.4.k_cache 0.00068886 1.27719996 + layer.4.v_cache 0.00007406 0.10204830 + layer.4.output 0.00494179 207.99312801 + ------------------------------------------------------------------------------------- + TOTAL 0.03972631 111.29749631 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 59232 +BPFP 0.2039 bits/point +EBPFP 0.4078 equivalent bits/point +MSE 111.297496 +---------------------- -------------------------------------------------------- +Time: 3.733s Load: 0.011s, Pack+Encode: 2.134s, Decode+Unpack: 1.588s +---------------------- -------------------------------------------------------- +💾 Converting with 111.2975 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,712B, BPFP=0.2624 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,804B, BPFP=0.3396 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,840B, BPFP=0.3422 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,556B, BPFP=0.3221 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,872B, BPFP=0.2738 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,612B, BPFP=0.3261 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,560B, BPFP=0.2517 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,492B, BPFP=0.3176 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,172B, BPFP=0.3657 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,748B, BPFP=0.3357 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,076B, BPFP=0.0210 +⌛️ [2/4] FRONTEND: Frontend time: 1.970s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.418s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11815372 390.74872738 + layer.0.v_cache 0.00001627 0.04461164 + layer.1.k_cache 0.37623358 24.76730858 + layer.1.v_cache 0.00000616 0.01749245 + layer.2.k_cache 0.01429399 2.84758610 + layer.2.v_cache 0.00002051 0.05108082 + layer.3.k_cache 0.01320818 13.93508511 + layer.3.v_cache 0.00002004 0.06082444 + layer.4.k_cache 0.00067885 1.30938527 + layer.4.v_cache 0.00005318 0.10401557 + layer.4.output 1.38524649 245.16475436 + ------------------------------------------------------------------------------------- + TOTAL 0.60114176 126.47290576 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 46444 +BPFP 0.1932 bits/point +EBPFP 0.3863 equivalent bits/point +MSE 126.472906 +---------------------- -------------------------------------------------------- +Time: 3.397s Load: 0.008s, Pack+Encode: 1.970s, Decode+Unpack: 1.418s +---------------------- -------------------------------------------------------- +💾 Converting with 126.4729 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,236B, BPFP=0.2381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,040B, BPFP=0.3395 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,252B, BPFP=0.3514 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,704B, BPFP=0.3206 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,732B, BPFP=0.2660 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,284B, BPFP=0.2970 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,384B, BPFP=0.2464 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,948B, BPFP=0.3343 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,596B, BPFP=0.3707 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,912B, BPFP=0.3323 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,312B, BPFP=0.0186 +⌛️ [2/4] FRONTEND: Frontend time: 2.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.580s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09820673 384.87997415 + layer.0.v_cache 0.00001666 0.04325154 + layer.1.k_cache 0.50544519 25.17626251 + layer.1.v_cache 0.00000613 0.01699522 + layer.2.k_cache 0.01930715 2.86747160 + layer.2.v_cache 0.00002033 0.04912170 + layer.3.k_cache 0.01337509 13.54021031 + layer.3.v_cache 0.00001987 0.05858542 + layer.4.k_cache 0.00074230 1.29184136 + layer.4.v_cache 0.00004869 0.09733757 + layer.4.output 0.00473525 199.75252120 + ------------------------------------------------------------------------------------- + TOTAL 0.03943146 107.42874704 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 57400 +BPFP 0.1898 bits/point +EBPFP 0.3795 equivalent bits/point +MSE 107.428747 +---------------------- -------------------------------------------------------- +Time: 3.840s Load: 0.009s, Pack+Encode: 2.251s, Decode+Unpack: 1.580s +---------------------- -------------------------------------------------------- +💾 Converting with 107.4287 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,636B, BPFP=0.2606 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,656B, BPFP=0.3337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,140B, BPFP=0.3684 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,504B, BPFP=0.3228 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,052B, BPFP=0.2904 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,276B, BPFP=0.3065 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,668B, BPFP=0.2629 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,404B, BPFP=0.3157 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,136B, BPFP=0.3681 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,640B, BPFP=0.3326 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,136B, BPFP=0.0219 +⌛️ [2/4] FRONTEND: Frontend time: 2.526s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.494s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13853801 389.49014478 + layer.0.v_cache 0.00001803 0.04485001 + layer.1.k_cache 0.23740530 25.56708716 + layer.1.v_cache 0.00000581 0.01662454 + layer.2.k_cache 0.01896443 2.92424669 + layer.2.v_cache 0.00001977 0.04938768 + layer.3.k_cache 0.00876656 14.17659319 + layer.3.v_cache 0.00002034 0.06115921 + layer.4.k_cache 0.00073522 1.22475118 + layer.4.v_cache 0.00004914 0.09901468 + layer.4.output 1.40427464 248.52508601 + ------------------------------------------------------------------------------------- + TOTAL 0.60202618 127.84290948 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 46248 +BPFP 0.1950 bits/point +EBPFP 0.3900 equivalent bits/point +MSE 127.842909 +---------------------- -------------------------------------------------------- +Time: 4.028s Load: 0.008s, Pack+Encode: 2.526s, Decode+Unpack: 1.494s +---------------------- -------------------------------------------------------- +💾 Converting with 127.8429 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,180B, BPFP=0.2366 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,336B, BPFP=0.3587 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,372B, BPFP=0.3607 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,832B, BPFP=0.3302 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,664B, BPFP=0.2640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,068B, BPFP=0.2869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,444B, BPFP=0.2516 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,296B, BPFP=0.2998 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,716B, BPFP=0.3802 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,260B, BPFP=0.3544 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,364B, BPFP=0.0191 +⌛️ [2/4] FRONTEND: Frontend time: 2.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.542s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10827675 379.91148664 + layer.0.v_cache 0.00001706 0.04400792 + layer.1.k_cache 0.56919524 25.03224956 + layer.1.v_cache 0.00000592 0.01697150 + layer.2.k_cache 0.01911232 2.85637300 + layer.2.v_cache 0.00002018 0.05005576 + layer.3.k_cache 0.01342056 13.58206486 + layer.3.v_cache 0.00002333 0.06254718 + layer.4.k_cache 0.00074072 1.28688580 + layer.4.v_cache 0.00005062 0.09903929 + layer.4.output 0.00480139 201.22397451 + ------------------------------------------------------------------------------------- + TOTAL 0.04379250 107.73585312 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 57532 +BPFP 0.1916 bits/point +EBPFP 0.3832 equivalent bits/point +MSE 107.735853 +---------------------- -------------------------------------------------------- +Time: 3.680s Load: 0.010s, Pack+Encode: 2.128s, Decode+Unpack: 1.542s +---------------------- -------------------------------------------------------- +💾 Converting with 107.7359 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,516B, BPFP=0.2643 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,940B, BPFP=0.3476 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,512B, BPFP=0.3811 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,308B, BPFP=0.3106 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,648B, BPFP=0.2720 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,100B, BPFP=0.2985 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,504B, BPFP=0.2636 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,284B, BPFP=0.3092 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,992B, BPFP=0.4092 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,028B, BPFP=0.3528 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,448B, BPFP=0.0205 +⌛️ [2/4] FRONTEND: Frontend time: 2.340s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.880s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15963619 398.39984199 + layer.0.v_cache 0.00001721 0.04906269 + layer.1.k_cache 0.49929055 25.93183813 + layer.1.v_cache 0.00000624 0.01824201 + layer.2.k_cache 0.03270629 3.09933277 + layer.2.v_cache 0.00002062 0.05218807 + layer.3.k_cache 0.00946706 14.16086215 + layer.3.v_cache 0.00002056 0.06239167 + layer.4.k_cache 0.00068536 1.28980544 + layer.4.v_cache 0.00004941 0.09765647 + layer.4.output 0.00495045 208.00344436 + ------------------------------------------------------------------------------------- + TOTAL 0.04332662 111.71678423 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 57280 +BPFP 0.1972 bits/point +EBPFP 0.3944 equivalent bits/point +MSE 111.716784 +---------------------- -------------------------------------------------------- +Time: 4.229s Load: 0.009s, Pack+Encode: 2.340s, Decode+Unpack: 1.880s +---------------------- -------------------------------------------------------- +💾 Converting with 111.7168 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,904B, BPFP=0.2585 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,836B, BPFP=0.3202 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,432B, BPFP=0.3596 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,480B, BPFP=0.2966 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,192B, BPFP=0.2775 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,164B, BPFP=0.2757 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,900B, BPFP=0.2582 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,572B, BPFP=0.3027 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,416B, BPFP=0.3586 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,776B, BPFP=0.3162 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,224B, BPFP=0.0210 +⌛️ [2/4] FRONTEND: Frontend time: 2.266s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.658s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14874478 395.98497087 + layer.0.v_cache 0.00001714 0.04511415 + layer.1.k_cache 0.45078779 25.61484127 + layer.1.v_cache 0.00000615 0.01739295 + layer.2.k_cache 0.01724745 3.07114941 + layer.2.v_cache 0.00001987 0.04813879 + layer.3.k_cache 0.02849353 13.24881240 + layer.3.v_cache 0.00002056 0.05958793 + layer.4.k_cache 0.00066912 1.27148696 + layer.4.v_cache 0.00005068 0.09764500 + layer.4.output 1.29722614 229.60177058 + ------------------------------------------------------------------------------------- + TOTAL 0.57215530 120.39244316 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 47896 +BPFP 0.1865 bits/point +EBPFP 0.3731 equivalent bits/point +MSE 120.392443 +---------------------- -------------------------------------------------------- +Time: 3.935s Load: 0.011s, Pack+Encode: 2.266s, Decode+Unpack: 1.658s +---------------------- -------------------------------------------------------- +💾 Converting with 120.3924 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 341, 128) +Output shape: (1, 341, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.output: torch.Size([1, 341, 3584]) -> torch.Size([1, 1, 341, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,652B, BPFP=0.2590 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,600B, BPFP=0.3482 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,260B, BPFP=0.3327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,720B, BPFP=0.3079 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,368B, BPFP=0.2460 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,608B, BPFP=0.3028 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,212B, BPFP=0.2388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,816B, BPFP=0.3123 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,896B, BPFP=0.3618 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,112B, BPFP=0.3259 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,696B, BPFP=0.0176 +⌛️ [2/4] FRONTEND: Frontend time: 2.683s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.932s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17009405 402.38567632 + layer.0.v_cache 0.00001631 0.04589105 + layer.1.k_cache 0.79816426 24.22628471 + layer.1.v_cache 0.00000611 0.01739444 + layer.2.k_cache 0.02878924 2.84051138 + layer.2.v_cache 0.00002016 0.04978930 + layer.3.k_cache 0.01035325 13.27337306 + layer.3.v_cache 0.00002082 0.05909438 + layer.4.k_cache 0.00072160 1.28379576 + layer.4.v_cache 0.00005837 0.09663250 + layer.4.output 0.03919861 159.15971931 + ------------------------------------------------------------------------------------- + TOTAL 0.07544909 91.67038106 + (elements=2,968,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2968064 +Total Bytes 68940 +BPFP 0.1858 bits/point +EBPFP 0.3716 equivalent bits/point +MSE 91.670381 +---------------------- -------------------------------------------------------- +Time: 4.630s Load: 0.015s, Pack+Encode: 2.683s, Decode+Unpack: 1.932s +---------------------- -------------------------------------------------------- +💾 Converting with 91.6704 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 335, 128) +Output shape: (1, 335, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.output: torch.Size([1, 335, 3584]) -> torch.Size([1, 1, 335, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,396B, BPFP=0.2517 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,616B, BPFP=0.3552 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,916B, BPFP=0.3226 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,564B, BPFP=0.3062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,544B, BPFP=0.2586 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,724B, BPFP=0.3136 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,068B, BPFP=0.2364 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,828B, BPFP=0.3185 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,992B, BPFP=0.3728 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,664B, BPFP=0.3108 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,756B, BPFP=0.0184 +⌛️ [2/4] FRONTEND: Frontend time: 2.684s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.883s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15195324 392.08745336 + layer.0.v_cache 0.00001666 0.04613634 + layer.1.k_cache 0.76788093 23.79170651 + layer.1.v_cache 0.00000624 0.01816455 + layer.2.k_cache 0.02686723 2.89304163 + layer.2.v_cache 0.00001960 0.04893229 + layer.3.k_cache 0.01080669 13.32150187 + layer.3.v_cache 0.00002024 0.05825117 + layer.4.k_cache 0.00071790 1.28520699 + layer.4.v_cache 0.00005106 0.09506092 + layer.4.output 3.38143948 160.47963753 + ------------------------------------------------------------------------------------- + TOTAL 1.44873036 91.58840696 + (elements=2,915,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2915840 +Total Bytes 68068 +BPFP 0.1868 bits/point +EBPFP 0.3735 equivalent bits/point +MSE 91.588407 +---------------------- -------------------------------------------------------- +Time: 4.580s Load: 0.012s, Pack+Encode: 2.684s, Decode+Unpack: 1.883s +---------------------- -------------------------------------------------------- +💾 Converting with 91.5884 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,380B, BPFP=0.2583 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,240B, BPFP=0.3679 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,276B, BPFP=0.3700 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,476B, BPFP=0.3229 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,432B, BPFP=0.2613 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,560B, BPFP=0.3278 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,292B, BPFP=0.2531 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,472B, BPFP=0.3226 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,856B, BPFP=0.4042 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,536B, BPFP=0.3854 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,304B, BPFP=0.0194 +⌛️ [2/4] FRONTEND: Frontend time: 2.521s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.743s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17469128 401.27329009 + layer.0.v_cache 0.00001775 0.04757349 + layer.1.k_cache 0.51265420 25.25564011 + layer.1.v_cache 0.00000620 0.01799880 + layer.2.k_cache 0.03428403 3.02983537 + layer.2.v_cache 0.00002024 0.05097295 + layer.3.k_cache 0.01443097 13.43907448 + layer.3.v_cache 0.00002182 0.06534900 + layer.4.k_cache 0.00070119 1.28507817 + layer.4.v_cache 0.00005097 0.10157106 + layer.4.output 0.00503371 209.54349730 + ------------------------------------------------------------------------------------- + TOTAL 0.04541851 112.43358027 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 57824 +BPFP 0.2006 bits/point +EBPFP 0.4011 equivalent bits/point +MSE 112.433580 +---------------------- -------------------------------------------------------- +Time: 4.276s Load: 0.012s, Pack+Encode: 2.521s, Decode+Unpack: 1.743s +---------------------- -------------------------------------------------------- +💾 Converting with 112.4336 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,056B, BPFP=0.2392 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,332B, BPFP=0.3733 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,348B, BPFP=0.3743 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,836B, BPFP=0.3441 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,512B, BPFP=0.2660 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,780B, BPFP=0.3408 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,124B, BPFP=0.2432 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,040B, BPFP=0.3561 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,760B, BPFP=0.3986 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,320B, BPFP=0.3726 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,380B, BPFP=0.0200 +⌛️ [2/4] FRONTEND: Frontend time: 2.549s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.755s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18043976 396.15383255 + layer.0.v_cache 0.00001814 0.04413481 + layer.1.k_cache 0.58196952 24.82675597 + layer.1.v_cache 0.00000608 0.01719356 + layer.2.k_cache 0.02409397 2.89318456 + layer.2.v_cache 0.00002020 0.04945222 + layer.3.k_cache 0.01875559 13.41098356 + layer.3.v_cache 0.00002085 0.05901417 + layer.4.k_cache 0.00069073 1.26154186 + layer.4.v_cache 0.00005389 0.10033146 + layer.4.output 0.00498411 209.54465970 + ------------------------------------------------------------------------------------- + TOTAL 0.04946809 112.09582604 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 58488 +BPFP 0.2029 bits/point +EBPFP 0.4057 equivalent bits/point +MSE 112.095826 +---------------------- -------------------------------------------------------- +Time: 4.313s Load: 0.010s, Pack+Encode: 2.549s, Decode+Unpack: 1.755s +---------------------- -------------------------------------------------------- +💾 Converting with 112.0958 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 386, 128) +Output shape: (1, 386, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.output: torch.Size([1, 386, 3584]) -> torch.Size([1, 1, 386, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,356B, BPFP=0.2573 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,180B, BPFP=0.3311 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,016B, BPFP=0.3245 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,660B, BPFP=0.3101 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,208B, BPFP=0.2513 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,376B, BPFP=0.2986 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,616B, BPFP=0.2273 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,960B, BPFP=0.3222 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,968B, BPFP=0.3630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,388B, BPFP=0.3395 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,984B, BPFP=0.0173 +⌛️ [2/4] FRONTEND: Frontend time: 2.879s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.096s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16808204 397.49923089 + layer.0.v_cache 0.00001827 0.04649570 + layer.1.k_cache 0.89383947 24.08814362 + layer.1.v_cache 0.00000605 0.01801504 + layer.2.k_cache 0.02643975 2.71207471 + layer.2.v_cache 0.00002094 0.05450705 + layer.3.k_cache 0.04368651 12.97837268 + layer.3.v_cache 0.00002149 0.06572174 + layer.4.k_cache 0.00075087 1.33247621 + layer.4.v_cache 0.00005208 0.10217913 + layer.4.output 0.03469837 140.57433151 + ------------------------------------------------------------------------------------- + TOTAL 0.08092977 83.70103161 + (elements=3,359,744) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3359744 +Total Bytes 77712 +BPFP 0.1850 bits/point +EBPFP 0.3701 equivalent bits/point +MSE 83.701032 +---------------------- -------------------------------------------------------- +Time: 4.990s Load: 0.016s, Pack+Encode: 2.879s, Decode+Unpack: 2.096s +---------------------- -------------------------------------------------------- +💾 Converting with 83.7010 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,116B, BPFP=0.2233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,552B, BPFP=0.3012 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,932B, BPFP=0.3218 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,492B, BPFP=0.2980 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,528B, BPFP=0.2457 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,288B, BPFP=0.2869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,372B, BPFP=0.2372 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,464B, BPFP=0.2964 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,656B, BPFP=0.3611 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,780B, BPFP=0.3136 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,200B, BPFP=0.0171 +⌛️ [2/4] FRONTEND: Frontend time: 2.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.771s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14364293 383.54779731 + layer.0.v_cache 0.00001600 0.04593484 + layer.1.k_cache 0.62366660 24.90395101 + layer.1.v_cache 0.00000628 0.01849774 + layer.2.k_cache 0.01981349 2.75633876 + layer.2.v_cache 0.00002090 0.05367292 + layer.3.k_cache 0.02158341 13.84509871 + layer.3.v_cache 0.00002097 0.06141418 + layer.4.k_cache 0.00071696 1.33444712 + layer.4.v_cache 0.00005883 0.10314646 + layer.4.output 0.00463133 192.84221540 + ------------------------------------------------------------------------------------- + TOTAL 0.04952739 104.50387099 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 55380 +BPFP 0.1767 bits/point +EBPFP 0.3535 equivalent bits/point +MSE 104.503871 +---------------------- -------------------------------------------------------- +Time: 4.281s Load: 0.011s, Pack+Encode: 2.500s, Decode+Unpack: 1.771s +---------------------- -------------------------------------------------------- +💾 Converting with 104.5039 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,372B, BPFP=0.2466 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,016B, BPFP=0.3394 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,432B, BPFP=0.3628 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,496B, BPFP=0.3100 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,776B, BPFP=0.2694 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,256B, BPFP=0.2965 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,588B, BPFP=0.2588 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,420B, BPFP=0.3057 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,868B, BPFP=0.3874 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,104B, BPFP=0.3443 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,352B, BPFP=0.0190 +⌛️ [2/4] FRONTEND: Frontend time: 2.512s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.835s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12300906 391.51379174 + layer.0.v_cache 0.00001720 0.04452750 + layer.1.k_cache 0.58060381 24.49449494 + layer.1.v_cache 0.00000644 0.01708405 + layer.2.k_cache 0.02305070 2.85203558 + layer.2.v_cache 0.00002053 0.05321082 + layer.3.k_cache 0.05279791 13.09065461 + layer.3.v_cache 0.00002021 0.06346033 + layer.4.k_cache 0.00074809 1.32956296 + layer.4.v_cache 0.00005348 0.10259943 + layer.4.output 0.00483522 200.49953262 + ------------------------------------------------------------------------------------- + TOTAL 0.04789259 108.06224414 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 57680 +BPFP 0.1914 bits/point +EBPFP 0.3828 equivalent bits/point +MSE 108.062244 +---------------------- -------------------------------------------------------- +Time: 4.357s Load: 0.010s, Pack+Encode: 2.512s, Decode+Unpack: 1.835s +---------------------- -------------------------------------------------------- +💾 Converting with 108.0622 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,708B, BPFP=0.2598 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,704B, BPFP=0.3296 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,016B, BPFP=0.3515 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,276B, BPFP=0.2996 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,672B, BPFP=0.2573 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,316B, BPFP=0.3024 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,456B, BPFP=0.2422 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,220B, BPFP=0.2957 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,852B, BPFP=0.3400 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,464B, BPFP=0.3128 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,948B, BPFP=0.0195 +⌛️ [2/4] FRONTEND: Frontend time: 2.362s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.623s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16177035 392.08113789 + layer.0.v_cache 0.00001782 0.04792840 + layer.1.k_cache 0.31952496 25.87902011 + layer.1.v_cache 0.00000626 0.01780474 + layer.2.k_cache 0.01271096 2.95658019 + layer.2.v_cache 0.00002040 0.05386207 + layer.3.k_cache 0.01363577 13.74250171 + layer.3.v_cache 0.00002033 0.06236521 + layer.4.k_cache 0.00066992 1.30839395 + layer.4.v_cache 0.00005030 0.10154095 + layer.4.output 1.37284199 242.97505605 + ------------------------------------------------------------------------------------- + TOTAL 0.59519535 125.71038398 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 44632 +BPFP 0.1840 bits/point +EBPFP 0.3679 equivalent bits/point +MSE 125.710384 +---------------------- -------------------------------------------------------- +Time: 3.994s Load: 0.009s, Pack+Encode: 2.362s, Decode+Unpack: 1.623s +---------------------- -------------------------------------------------------- +💾 Converting with 125.7104 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.2630 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,720B, BPFP=0.3176 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,120B, BPFP=0.3518 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,276B, BPFP=0.2797 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,108B, BPFP=0.2654 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,520B, BPFP=0.3005 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,972B, BPFP=0.2538 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,440B, BPFP=0.2937 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,252B, BPFP=0.3630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,588B, BPFP=0.3064 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,636B, BPFP=0.0322 +⌛️ [2/4] FRONTEND: Frontend time: 2.517s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.469s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09962382 405.20927254 + layer.0.v_cache 0.00001686 0.04738453 + layer.1.k_cache 0.10988800 24.42934063 + layer.1.v_cache 0.00000590 0.01769673 + layer.2.k_cache 0.01214521 3.42132135 + layer.2.v_cache 0.00002216 0.05585336 + layer.3.k_cache 0.02260447 13.93469705 + layer.3.v_cache 0.00002006 0.06835238 + layer.4.k_cache 0.00069723 1.42154481 + layer.4.v_cache 0.00005242 0.10980480 + layer.4.output 0.00882482 300.22945941 + ------------------------------------------------------------------------------------- + TOTAL 0.01804999 150.01891083 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 37712 +BPFP 0.1894 bits/point +EBPFP 0.3788 equivalent bits/point +MSE 150.018911 +---------------------- -------------------------------------------------------- +Time: 3.994s Load: 0.008s, Pack+Encode: 2.517s, Decode+Unpack: 1.469s +---------------------- -------------------------------------------------------- +💾 Converting with 150.0189 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,116B, BPFP=0.2289 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,480B, BPFP=0.3603 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,216B, BPFP=0.3456 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,584B, BPFP=0.3105 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,780B, BPFP=0.2658 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,672B, BPFP=0.3154 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,624B, BPFP=0.2571 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,592B, BPFP=0.3109 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,732B, BPFP=0.3743 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,260B, BPFP=0.3481 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,192B, BPFP=0.0174 +⌛️ [2/4] FRONTEND: Frontend time: 2.508s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.812s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11148910 385.55933052 + layer.0.v_cache 0.00001684 0.04657701 + layer.1.k_cache 0.60671574 24.69127419 + layer.1.v_cache 0.00000630 0.01720801 + layer.2.k_cache 0.04149003 3.07873188 + layer.2.v_cache 0.00001976 0.05009644 + layer.3.k_cache 0.02200458 13.84219132 + layer.3.v_cache 0.00001954 0.06015907 + layer.4.k_cache 0.00078895 1.31768973 + layer.4.v_cache 0.00004864 0.09630978 + layer.4.output 0.00472744 197.63233986 + ------------------------------------------------------------------------------------- + TOTAL 0.04798185 106.59917335 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 58248 +BPFP 0.1905 bits/point +EBPFP 0.3810 equivalent bits/point +MSE 106.599173 +---------------------- -------------------------------------------------------- +Time: 4.333s Load: 0.013s, Pack+Encode: 2.508s, Decode+Unpack: 1.812s +---------------------- -------------------------------------------------------- +💾 Converting with 106.5992 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,452B, BPFP=0.2697 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,140B, BPFP=0.4016 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,136B, BPFP=0.4012 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,536B, BPFP=0.3544 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,804B, BPFP=0.2972 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,320B, BPFP=0.3375 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,348B, BPFP=0.2616 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,072B, BPFP=0.3181 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,636B, BPFP=0.4403 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,108B, BPFP=0.3991 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,272B, BPFP=0.0254 +⌛️ [2/4] FRONTEND: Frontend time: 2.336s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.574s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14565937 399.90941406 + layer.0.v_cache 0.00001550 0.04309331 + layer.1.k_cache 0.22559690 25.36848877 + layer.1.v_cache 0.00000590 0.01672301 + layer.2.k_cache 0.00679684 2.90720978 + layer.2.v_cache 0.00002314 0.05113273 + layer.3.k_cache 0.06866968 12.93830078 + layer.3.v_cache 0.00002119 0.06110564 + layer.4.k_cache 0.00065507 1.28930984 + layer.4.v_cache 0.00005191 0.10340112 + layer.4.output 1.53062134 270.89236607 + ------------------------------------------------------------------------------------- + TOTAL 0.65657911 137.58439656 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 46824 +BPFP 0.2152 bits/point +EBPFP 0.4304 equivalent bits/point +MSE 137.584397 +---------------------- -------------------------------------------------------- +Time: 3.919s Load: 0.009s, Pack+Encode: 2.336s, Decode+Unpack: 1.574s +---------------------- -------------------------------------------------------- +💾 Converting with 137.5844 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,144B, BPFP=0.2444 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,608B, BPFP=0.3582 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,236B, BPFP=0.4070 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,264B, BPFP=0.3315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,892B, BPFP=0.3025 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,084B, BPFP=0.3952 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,544B, BPFP=0.2755 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,636B, BPFP=0.3604 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,084B, BPFP=0.3952 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,068B, BPFP=0.3940 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,296B, BPFP=0.0255 +⌛️ [2/4] FRONTEND: Frontend time: 2.446s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.649s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10663509 386.75610230 + layer.0.v_cache 0.00001613 0.04414802 + layer.1.k_cache 0.26620468 26.15443048 + layer.1.v_cache 0.00000602 0.01723443 + layer.2.k_cache 0.00649704 3.06549406 + layer.2.v_cache 0.00002002 0.04853818 + layer.3.k_cache 0.02565914 13.79730255 + layer.3.v_cache 0.00002186 0.06124184 + layer.4.k_cache 0.00070101 1.27093749 + layer.4.v_cache 0.00005421 0.10362478 + layer.4.output 1.52303164 269.54821873 + ------------------------------------------------------------------------------------- + TOTAL 0.65100216 136.36215207 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 46856 +BPFP 0.2143 bits/point +EBPFP 0.4285 equivalent bits/point +MSE 136.362152 +---------------------- -------------------------------------------------------- +Time: 4.103s Load: 0.008s, Pack+Encode: 2.446s, Decode+Unpack: 1.649s +---------------------- -------------------------------------------------------- +💾 Converting with 136.3622 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,448B, BPFP=0.2603 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,508B, BPFP=0.3809 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,536B, BPFP=0.3825 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,880B, BPFP=0.3441 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,552B, BPFP=0.2664 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,032B, BPFP=0.3530 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,144B, BPFP=0.2425 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,916B, BPFP=0.3462 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,732B, BPFP=0.3940 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,208B, BPFP=0.3633 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,468B, BPFP=0.0206 +⌛️ [2/4] FRONTEND: Frontend time: 2.473s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.782s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258070 394.85802903 + layer.0.v_cache 0.00001713 0.04621290 + layer.1.k_cache 0.49703808 24.58557723 + layer.1.v_cache 0.00000670 0.01849146 + layer.2.k_cache 0.01458295 2.87933853 + layer.2.v_cache 0.00002181 0.05245946 + layer.3.k_cache 0.02940617 13.32109887 + layer.3.v_cache 0.00002120 0.06410709 + layer.4.k_cache 0.00070157 1.34604305 + layer.4.v_cache 0.00005095 0.10304890 + layer.4.output 0.00497170 208.01033307 + ------------------------------------------------------------------------------------- + TOTAL 0.04171936 111.37333753 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 59424 +BPFP 0.2046 bits/point +EBPFP 0.4091 equivalent bits/point +MSE 111.373338 +---------------------- -------------------------------------------------------- +Time: 4.266s Load: 0.012s, Pack+Encode: 2.473s, Decode+Unpack: 1.782s +---------------------- -------------------------------------------------------- +💾 Converting with 111.3733 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,824B, BPFP=0.2644 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,860B, BPFP=0.3360 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,224B, BPFP=0.3612 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,460B, BPFP=0.3084 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,968B, BPFP=0.2743 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,832B, BPFP=0.3341 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,672B, BPFP=0.2539 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,656B, BPFP=0.3219 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,152B, BPFP=0.3562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,880B, BPFP=0.3374 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,064B, BPFP=0.0204 +⌛️ [2/4] FRONTEND: Frontend time: 2.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.654s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15433978 401.09516731 + layer.0.v_cache 0.00001575 0.04372423 + layer.1.k_cache 0.29362690 26.02119486 + layer.1.v_cache 0.00000607 0.01723068 + layer.2.k_cache 0.00940348 2.89433842 + layer.2.v_cache 0.00002159 0.05247389 + layer.3.k_cache 0.02320663 14.09796629 + layer.3.v_cache 0.00002067 0.06412498 + layer.4.k_cache 0.00067785 1.31950527 + layer.4.v_cache 0.00005016 0.10126266 + layer.4.output 1.35461534 239.73354535 + ------------------------------------------------------------------------------------- + TOTAL 0.58609860 124.93187094 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 47592 +BPFP 0.1936 bits/point +EBPFP 0.3871 equivalent bits/point +MSE 124.931871 +---------------------- -------------------------------------------------------- +Time: 3.997s Load: 0.010s, Pack+Encode: 2.334s, Decode+Unpack: 1.654s +---------------------- -------------------------------------------------------- +💾 Converting with 124.9319 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,708B, BPFP=0.2581 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,652B, BPFP=0.3099 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,720B, BPFP=0.3136 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,612B, BPFP=0.3077 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,636B, BPFP=0.2542 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,528B, BPFP=0.3031 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,316B, BPFP=0.2366 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,668B, BPFP=0.3107 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,456B, BPFP=0.3539 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,820B, BPFP=0.3191 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,336B, BPFP=0.0183 +⌛️ [2/4] FRONTEND: Frontend time: 2.518s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.782s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14751465 392.88141447 + layer.0.v_cache 0.00001719 0.04700336 + layer.1.k_cache 0.52245269 25.31239892 + layer.1.v_cache 0.00000618 0.01784416 + layer.2.k_cache 0.02707210 2.92133553 + layer.2.v_cache 0.00002227 0.05258837 + layer.3.k_cache 0.03364062 13.27368250 + layer.3.v_cache 0.00002083 0.06377984 + layer.4.k_cache 0.00069748 1.30535621 + layer.4.v_cache 0.00005305 0.09725423 + layer.4.output 0.00471531 194.88068609 + ------------------------------------------------------------------------------------- + TOTAL 0.04497084 105.89043884 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 56452 +BPFP 0.1821 bits/point +EBPFP 0.3641 equivalent bits/point +MSE 105.890439 +---------------------- -------------------------------------------------------- +Time: 4.310s Load: 0.011s, Pack+Encode: 2.518s, Decode+Unpack: 1.782s +---------------------- -------------------------------------------------------- +💾 Converting with 105.8904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,148B, BPFP=0.2542 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,460B, BPFP=0.2733 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,452B, BPFP=0.2728 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,156B, BPFP=0.2547 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,240B, BPFP=0.1985 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,176B, BPFP=0.2559 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,064B, BPFP=0.1877 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,136B, BPFP=0.2534 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,808B, BPFP=0.2946 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,192B, BPFP=0.2569 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,032B, BPFP=0.0178 +⌛️ [2/4] FRONTEND: Frontend time: 2.354s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.680s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17432048 407.40287990 + layer.0.v_cache 0.00001704 0.04254221 + layer.1.k_cache 0.39057378 24.08935355 + layer.1.v_cache 0.00000668 0.01735045 + layer.2.k_cache 0.01299327 2.85138107 + layer.2.v_cache 0.00002114 0.04844433 + layer.3.k_cache 0.02499930 12.92641697 + layer.3.v_cache 0.00002096 0.05881488 + layer.4.k_cache 0.00069121 1.26124639 + layer.4.v_cache 0.00005416 0.09355628 + layer.4.output 1.20070616 212.51025910 + ------------------------------------------------------------------------------------- + TOTAL 0.52992007 113.90375293 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 42864 +BPFP 0.1545 bits/point +EBPFP 0.3090 equivalent bits/point +MSE 113.903753 +---------------------- -------------------------------------------------------- +Time: 4.044s Load: 0.010s, Pack+Encode: 2.354s, Decode+Unpack: 1.680s +---------------------- -------------------------------------------------------- +💾 Converting with 113.9038 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,780B, BPFP=0.2579 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,052B, BPFP=0.3447 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,340B, BPFP=0.3644 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,376B, BPFP=0.2986 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,104B, BPFP=0.2800 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,244B, BPFP=0.2896 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,084B, BPFP=0.2787 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,664B, BPFP=0.3182 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,588B, BPFP=0.3813 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,016B, BPFP=0.3422 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,240B, BPFP=0.0218 +⌛️ [2/4] FRONTEND: Frontend time: 2.385s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.576s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16239141 403.49996588 + layer.0.v_cache 0.00001645 0.04434068 + layer.1.k_cache 0.33751179 25.90658902 + layer.1.v_cache 0.00000668 0.01742822 + layer.2.k_cache 0.00660169 2.88991177 + layer.2.v_cache 0.00002045 0.04987933 + layer.3.k_cache 0.02213682 13.14252269 + layer.3.v_cache 0.00002049 0.06151007 + layer.4.k_cache 0.00067824 1.31422071 + layer.4.v_cache 0.00005183 0.09848456 + layer.4.output 1.33690934 236.61759201 + ------------------------------------------------------------------------------------- + TOTAL 0.58163537 123.72635277 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 48488 +BPFP 0.1946 bits/point +EBPFP 0.3892 equivalent bits/point +MSE 123.726353 +---------------------- -------------------------------------------------------- +Time: 3.971s Load: 0.011s, Pack+Encode: 2.385s, Decode+Unpack: 1.576s +---------------------- -------------------------------------------------------- +💾 Converting with 123.7264 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,552B, BPFP=0.2654 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,952B, BPFP=0.4053 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,760B, BPFP=0.3941 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,756B, BPFP=0.3356 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,564B, BPFP=0.2661 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,552B, BPFP=0.3237 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,120B, BPFP=0.2402 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,428B, BPFP=0.3165 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,740B, BPFP=0.3930 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,392B, BPFP=0.3727 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,524B, BPFP=0.0210 +⌛️ [2/4] FRONTEND: Frontend time: 2.558s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.755s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15796844 394.68085354 + layer.0.v_cache 0.00001527 0.04272235 + layer.1.k_cache 0.50235566 25.08291307 + layer.1.v_cache 0.00000626 0.01728762 + layer.2.k_cache 0.02709456 2.91930310 + layer.2.v_cache 0.00002063 0.04615179 + layer.3.k_cache 0.00856834 12.93266091 + layer.3.v_cache 0.00001945 0.05553754 + layer.4.k_cache 0.00071483 1.29168587 + layer.4.v_cache 0.00005464 0.09519947 + layer.4.output 0.00491706 207.22104877 + ------------------------------------------------------------------------------------- + TOTAL 0.04301397 111.04186216 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 59340 +BPFP 0.2035 bits/point +EBPFP 0.4070 equivalent bits/point +MSE 111.041862 +---------------------- -------------------------------------------------------- +Time: 4.323s Load: 0.011s, Pack+Encode: 2.558s, Decode+Unpack: 1.755s +---------------------- -------------------------------------------------------- +💾 Converting with 111.0419 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,000B, BPFP=0.2520 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,704B, BPFP=0.2964 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,180B, BPFP=0.3264 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,436B, BPFP=0.2795 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,784B, BPFP=0.2384 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,628B, BPFP=0.2916 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,580B, BPFP=0.2256 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,024B, BPFP=0.3165 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,540B, BPFP=0.3490 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,720B, BPFP=0.2974 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,944B, BPFP=0.0265 +⌛️ [2/4] FRONTEND: Frontend time: 2.365s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.651s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13184587 402.35754788 + layer.0.v_cache 0.00001706 0.04810749 + layer.1.k_cache 0.45546043 24.64467892 + layer.1.v_cache 0.00000678 0.01905710 + layer.2.k_cache 0.02272066 3.20399180 + layer.2.v_cache 0.00001996 0.05656957 + layer.3.k_cache 0.01643563 13.46088040 + layer.3.v_cache 0.00002069 0.06836579 + layer.4.k_cache 0.00071781 1.40566992 + layer.4.v_cache 0.00005262 0.10285743 + layer.4.output 1.23455937 218.49953197 + ------------------------------------------------------------------------------------- + TOTAL 0.54524783 116.16849706 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 48540 +BPFP 0.1799 bits/point +EBPFP 0.3598 equivalent bits/point +MSE 116.168497 +---------------------- -------------------------------------------------------- +Time: 4.026s Load: 0.011s, Pack+Encode: 2.365s, Decode+Unpack: 1.651s +---------------------- -------------------------------------------------------- +💾 Converting with 116.1685 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,500B, BPFP=0.2663 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,276B, BPFP=0.3714 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,468B, BPFP=0.3828 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,420B, BPFP=0.3208 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,632B, BPFP=0.2741 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,192B, BPFP=0.3073 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,276B, BPFP=0.2531 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,676B, BPFP=0.3359 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,740B, BPFP=0.3989 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,196B, BPFP=0.3667 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,380B, BPFP=0.0201 +⌛️ [2/4] FRONTEND: Frontend time: 2.505s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.789s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17705467 400.15074574 + layer.0.v_cache 0.00001633 0.04198399 + layer.1.k_cache 0.50406676 25.02319151 + layer.1.v_cache 0.00000599 0.01677567 + layer.2.k_cache 0.00944760 3.04270288 + layer.2.v_cache 0.00001883 0.04742959 + layer.3.k_cache 0.04120483 13.42568507 + layer.3.v_cache 0.00002392 0.06186318 + layer.4.k_cache 0.00070100 1.29745356 + layer.4.v_cache 0.00005002 0.09665244 + layer.4.output 0.00499618 210.36569940 + ------------------------------------------------------------------------------------- + TOTAL 0.04515078 112.69202232 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 57756 +BPFP 0.2011 bits/point +EBPFP 0.4022 equivalent bits/point +MSE 112.692022 +---------------------- -------------------------------------------------------- +Time: 4.306s Load: 0.012s, Pack+Encode: 2.505s, Decode+Unpack: 1.789s +---------------------- -------------------------------------------------------- +💾 Converting with 112.6920 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,488B, BPFP=0.2597 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,508B, BPFP=0.3766 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,448B, BPFP=0.3731 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,564B, BPFP=0.3220 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,496B, BPFP=0.2602 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,836B, BPFP=0.3377 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,192B, BPFP=0.2426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,724B, BPFP=0.3312 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,336B, BPFP=0.3667 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,276B, BPFP=0.3632 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,608B, BPFP=0.0216 +⌛️ [2/4] FRONTEND: Frontend time: 2.510s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.789s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20501094 394.80706019 + layer.0.v_cache 0.00001516 0.04175842 + layer.1.k_cache 0.57085758 25.12727141 + layer.1.v_cache 0.00000574 0.01627442 + layer.2.k_cache 0.01259345 2.88602069 + layer.2.v_cache 0.00002099 0.04799036 + layer.3.k_cache 0.01648921 13.35876013 + layer.3.v_cache 0.00001970 0.05716310 + layer.4.k_cache 0.00070444 1.19964747 + layer.4.v_cache 0.00005121 0.09665837 + layer.4.output 0.00487421 205.67981151 + ------------------------------------------------------------------------------------- + TOTAL 0.04940517 110.43513442 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 58476 +BPFP 0.1991 bits/point +EBPFP 0.3981 equivalent bits/point +MSE 110.435134 +---------------------- -------------------------------------------------------- +Time: 4.311s Load: 0.011s, Pack+Encode: 2.510s, Decode+Unpack: 1.789s +---------------------- -------------------------------------------------------- +💾 Converting with 110.4351 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,332B, BPFP=0.2498 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,116B, BPFP=0.3526 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,404B, BPFP=0.3692 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,684B, BPFP=0.3277 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,692B, BPFP=0.2705 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,396B, BPFP=0.3111 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,416B, BPFP=0.2546 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,708B, BPFP=0.3291 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,852B, BPFP=0.3951 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,916B, BPFP=0.3411 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,448B, BPFP=0.0202 +⌛️ [2/4] FRONTEND: Frontend time: 2.518s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.740s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14466163 391.89466098 + layer.0.v_cache 0.00001631 0.04628595 + layer.1.k_cache 0.53789613 26.54460476 + layer.1.v_cache 0.00000653 0.01751049 + layer.2.k_cache 0.02095415 2.92689069 + layer.2.v_cache 0.00002169 0.05070157 + layer.3.k_cache 0.03083320 13.42120744 + layer.3.v_cache 0.00002083 0.06127533 + layer.4.k_cache 0.00069737 1.31827762 + layer.4.v_cache 0.00004943 0.09794489 + layer.4.output 0.00490546 204.94066289 + ------------------------------------------------------------------------------------- + TOTAL 0.04526444 110.05670588 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 57964 +BPFP 0.1966 bits/point +EBPFP 0.3932 equivalent bits/point +MSE 110.056706 +---------------------- -------------------------------------------------------- +Time: 4.269s Load: 0.011s, Pack+Encode: 2.518s, Decode+Unpack: 1.740s +---------------------- -------------------------------------------------------- +💾 Converting with 110.0567 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,872B, BPFP=0.2546 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,028B, BPFP=0.3150 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,440B, BPFP=0.3365 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,620B, BPFP=0.2937 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,024B, BPFP=0.2625 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,864B, BPFP=0.2542 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,612B, BPFP=0.2410 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,648B, BPFP=0.2952 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,960B, BPFP=0.3637 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,364B, BPFP=0.3326 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,560B, BPFP=0.0191 +⌛️ [2/4] FRONTEND: Frontend time: 1.928s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14240686 399.74947742 + layer.0.v_cache 0.00001673 0.04397843 + layer.1.k_cache 0.61438330 25.54666270 + layer.1.v_cache 0.00000638 0.01689686 + layer.2.k_cache 0.02027981 2.90383962 + layer.2.v_cache 0.00002012 0.04898483 + layer.3.k_cache 0.02199794 13.36168250 + layer.3.v_cache 0.00002003 0.05738921 + layer.4.k_cache 0.00078937 1.27404112 + layer.4.v_cache 0.00005038 0.09319168 + layer.4.output 0.04464151 181.49985069 + ------------------------------------------------------------------------------------- + TOTAL 0.06543891 100.79971172 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 58992 +BPFP 0.1813 bits/point +EBPFP 0.3627 equivalent bits/point +MSE 100.799712 +---------------------- -------------------------------------------------------- +Time: 3.286s Load: 0.010s, Pack+Encode: 1.928s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 100.7997 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 399, 128) +Output shape: (1, 399, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.output: torch.Size([1, 399, 3584]) -> torch.Size([1, 1, 399, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,476B, BPFP=0.2536 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,264B, BPFP=0.3236 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,332B, BPFP=0.3263 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,116B, BPFP=0.3178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,308B, BPFP=0.2470 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,128B, BPFP=0.2791 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,708B, BPFP=0.2235 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,928B, BPFP=0.3105 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,744B, BPFP=0.3424 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,344B, BPFP=0.3268 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,012B, BPFP=0.0169 +⌛️ [2/4] FRONTEND: Frontend time: 2.117s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.573s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17889012 399.73390508 + layer.0.v_cache 0.00001635 0.04353927 + layer.1.k_cache 0.94984295 25.32402295 + layer.1.v_cache 0.00000590 0.01667239 + layer.2.k_cache 0.02305104 2.78356566 + layer.2.v_cache 0.00002114 0.05065306 + layer.3.k_cache 0.01620252 13.04071947 + layer.3.v_cache 0.00002049 0.06094958 + layer.4.k_cache 0.00073175 1.25084432 + layer.4.v_cache 0.00005255 0.09653465 + layer.4.output 0.00656733 139.00943206 + ------------------------------------------------------------------------------------- + TOTAL 0.07145918 83.26279005 + (elements=3,472,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3472896 +Total Bytes 78360 +BPFP 0.1805 bits/point +EBPFP 0.3610 equivalent bits/point +MSE 83.262790 +---------------------- -------------------------------------------------------- +Time: 3.703s Load: 0.013s, Pack+Encode: 2.117s, Decode+Unpack: 1.573s +---------------------- -------------------------------------------------------- +💾 Converting with 83.2628 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,756B, BPFP=0.2597 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,776B, BPFP=0.3302 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,016B, BPFP=0.3468 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,220B, BPFP=0.2918 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,020B, BPFP=0.2779 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,168B, BPFP=0.2882 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,784B, BPFP=0.2616 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,396B, BPFP=0.3039 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,348B, BPFP=0.3697 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,652B, BPFP=0.3216 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,040B, BPFP=0.0201 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.434s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10670873 393.41827987 + layer.0.v_cache 0.00001672 0.04585507 + layer.1.k_cache 0.29393762 26.17927052 + layer.1.v_cache 0.00000627 0.01770547 + layer.2.k_cache 0.01243114 2.98362799 + layer.2.v_cache 0.00002105 0.05471980 + layer.3.k_cache 0.00919804 13.42548742 + layer.3.v_cache 0.00002002 0.06460722 + layer.4.k_cache 0.00070102 1.28655182 + layer.4.v_cache 0.00005024 0.10545617 + layer.4.output 1.35463108 239.74257269 + ------------------------------------------------------------------------------------- + TOTAL 0.58267697 124.45762178 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 46176 +BPFP 0.1878 bits/point +EBPFP 0.3756 equivalent bits/point +MSE 124.457622 +---------------------- -------------------------------------------------------- +Time: 3.173s Load: 0.009s, Pack+Encode: 1.730s, Decode+Unpack: 1.434s +---------------------- -------------------------------------------------------- +💾 Converting with 124.4576 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,148B, BPFP=0.2552 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,820B, BPFP=0.2965 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,548B, BPFP=0.2798 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,540B, BPFP=0.2793 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,260B, BPFP=0.2005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,048B, BPFP=0.2490 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,128B, BPFP=0.1924 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,476B, BPFP=0.2753 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,216B, BPFP=0.3209 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,288B, BPFP=0.2638 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,472B, BPFP=0.0217 +⌛️ [2/4] FRONTEND: Frontend time: 1.921s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.444s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633087 415.67002953 + layer.0.v_cache 0.00001891 0.04647008 + layer.1.k_cache 0.36012592 23.70673521 + layer.1.v_cache 0.00000581 0.01701105 + layer.2.k_cache 0.01407751 2.78485684 + layer.2.v_cache 0.00002023 0.05125283 + layer.3.k_cache 0.01230825 12.33056545 + layer.3.v_cache 0.00002000 0.06233489 + layer.4.k_cache 0.00071033 1.30316571 + layer.4.v_cache 0.00004856 0.09766293 + layer.4.output 1.20534739 213.29638287 + ------------------------------------------------------------------------------------- + TOTAL 0.52535871 114.65557439 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 44944 +BPFP 0.1626 bits/point +EBPFP 0.3253 equivalent bits/point +MSE 114.655574 +---------------------- -------------------------------------------------------- +Time: 3.373s Load: 0.008s, Pack+Encode: 1.921s, Decode+Unpack: 1.444s +---------------------- -------------------------------------------------------- +💾 Converting with 114.6556 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,040B, BPFP=0.2619 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,088B, BPFP=0.3299 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,436B, BPFP=0.3524 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,660B, BPFP=0.3021 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,044B, BPFP=0.2622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,512B, BPFP=0.2925 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,888B, BPFP=0.2521 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,548B, BPFP=0.2949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,556B, BPFP=0.3602 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,704B, BPFP=0.3050 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,024B, BPFP=0.0187 +⌛️ [2/4] FRONTEND: Frontend time: 2.026s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.552s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11142051 404.65469398 + layer.0.v_cache 0.00001650 0.04347613 + layer.1.k_cache 0.42032433 24.68948554 + layer.1.v_cache 0.00000619 0.01708560 + layer.2.k_cache 0.01443043 2.98048977 + layer.2.v_cache 0.00002030 0.05067147 + layer.3.k_cache 0.02895402 13.73384316 + layer.3.v_cache 0.00001914 0.05788054 + layer.4.k_cache 0.00067703 1.31680121 + layer.4.v_cache 0.00005693 0.10295907 + layer.4.output 1.27033357 224.83969324 + ------------------------------------------------------------------------------------- + TOTAL 0.55695649 118.91324936 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 48500 +BPFP 0.1850 bits/point +EBPFP 0.3699 equivalent bits/point +MSE 118.913249 +---------------------- -------------------------------------------------------- +Time: 3.586s Load: 0.008s, Pack+Encode: 2.026s, Decode+Unpack: 1.552s +---------------------- -------------------------------------------------------- +💾 Converting with 118.9132 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 302, 128) +Output shape: (1, 302, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.output: torch.Size([1, 302, 3584]) -> torch.Size([1, 1, 302, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,856B, BPFP=0.2512 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,972B, BPFP=0.3090 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,276B, BPFP=0.3247 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,876B, BPFP=0.3040 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,760B, BPFP=0.2463 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,252B, BPFP=0.2717 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,452B, BPFP=0.2303 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,468B, BPFP=0.2829 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,544B, BPFP=0.3386 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,632B, BPFP=0.2914 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,460B, BPFP=0.0182 +⌛️ [2/4] FRONTEND: Frontend time: 2.519s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.566s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15012774 399.91628725 + layer.0.v_cache 0.00001504 0.04121771 + layer.1.k_cache 0.60100308 25.09843718 + layer.1.v_cache 0.00000575 0.01598645 + layer.2.k_cache 0.00814273 2.89520910 + layer.2.v_cache 0.00002032 0.04785670 + layer.3.k_cache 0.01797006 13.30591629 + layer.3.v_cache 0.00001899 0.05578164 + layer.4.k_cache 0.00070977 1.26665633 + layer.4.v_cache 0.00004868 0.09195373 + layer.4.output 0.04414053 179.67939924 + ------------------------------------------------------------------------------------- + TOTAL 0.06394387 100.02888806 + (elements=2,628,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2628608 +Total Bytes 57548 +BPFP 0.1751 bits/point +EBPFP 0.3503 equivalent bits/point +MSE 100.028888 +---------------------- -------------------------------------------------------- +Time: 4.095s Load: 0.010s, Pack+Encode: 2.519s, Decode+Unpack: 1.566s +---------------------- -------------------------------------------------------- +💾 Converting with 100.0289 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,516B, BPFP=0.2633 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,212B, BPFP=0.3622 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,588B, BPFP=0.3841 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,476B, BPFP=0.3193 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,748B, BPFP=0.2768 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,808B, BPFP=0.2803 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,496B, BPFP=0.2621 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,780B, BPFP=0.3370 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,864B, BPFP=0.4002 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,912B, BPFP=0.3447 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,552B, BPFP=0.0213 +⌛️ [2/4] FRONTEND: Frontend time: 2.087s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.574s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14350806 394.80996385 + layer.0.v_cache 0.00001655 0.04214367 + layer.1.k_cache 0.46812200 25.39508694 + layer.1.v_cache 0.00000598 0.01685288 + layer.2.k_cache 0.01096032 2.87835716 + layer.2.v_cache 0.00002019 0.05055531 + layer.3.k_cache 0.00658789 12.81704211 + layer.3.v_cache 0.00002014 0.06011627 + layer.4.k_cache 0.00067528 1.27900935 + layer.4.v_cache 0.00005785 0.10111407 + layer.4.output 0.00496594 207.22258129 + ------------------------------------------------------------------------------------- + TOTAL 0.03910211 111.05931239 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 57952 +BPFP 0.1987 bits/point +EBPFP 0.3975 equivalent bits/point +MSE 111.059312 +---------------------- -------------------------------------------------------- +Time: 3.670s Load: 0.009s, Pack+Encode: 2.087s, Decode+Unpack: 1.574s +---------------------- -------------------------------------------------------- +💾 Converting with 111.0593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 260, 128) +Output shape: (1, 260, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.output: torch.Size([1, 260, 3584]) -> torch.Size([1, 1, 260, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,124B, BPFP=0.2478 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,880B, BPFP=0.3534 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,244B, BPFP=0.3752 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,492B, BPFP=0.3300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,404B, BPFP=0.2647 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,324B, BPFP=0.3200 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,076B, BPFP=0.2450 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,512B, BPFP=0.3312 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,420B, BPFP=0.3858 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,976B, BPFP=0.3591 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,460B, BPFP=0.0211 +⌛️ [2/4] FRONTEND: Frontend time: 2.093s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.493s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14548651 395.53843149 + layer.0.v_cache 0.00001584 0.04269323 + layer.1.k_cache 0.47062912 25.75539363 + layer.1.v_cache 0.00000604 0.01640600 + layer.2.k_cache 0.00619981 2.78269747 + layer.2.v_cache 0.00002003 0.05125454 + layer.3.k_cache 0.01299995 13.84994554 + layer.3.v_cache 0.00001981 0.05795134 + layer.4.k_cache 0.00067777 1.28293692 + layer.4.v_cache 0.00006192 0.10049517 + layer.4.output 0.00504520 213.57906937 + ------------------------------------------------------------------------------------- + TOTAL 0.03949607 113.79598182 + (elements=2,263,040) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2263040 +Total Bytes 55912 +BPFP 0.1977 bits/point +EBPFP 0.3953 equivalent bits/point +MSE 113.795982 +---------------------- -------------------------------------------------------- +Time: 3.595s Load: 0.009s, Pack+Encode: 2.093s, Decode+Unpack: 1.493s +---------------------- -------------------------------------------------------- +💾 Converting with 113.7960 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,176B, BPFP=0.2399 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,300B, BPFP=0.3619 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,556B, BPFP=0.3766 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,832B, BPFP=0.3350 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,800B, BPFP=0.2757 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,480B, BPFP=0.3148 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,332B, BPFP=0.2489 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,532B, BPFP=0.3178 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,808B, BPFP=0.3911 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,904B, BPFP=0.3392 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,484B, BPFP=0.0204 +⌛️ [2/4] FRONTEND: Frontend time: 2.172s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.542s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14555165 387.37149586 + layer.0.v_cache 0.00001661 0.04525558 + layer.1.k_cache 0.55468144 25.58089851 + layer.1.v_cache 0.00000621 0.01758810 + layer.2.k_cache 0.02243149 3.01512303 + layer.2.v_cache 0.00002103 0.04995484 + layer.3.k_cache 0.02028967 13.17772989 + layer.3.v_cache 0.00002036 0.06035049 + layer.4.k_cache 0.00072975 1.29952060 + layer.4.v_cache 0.00005124 0.10045255 + layer.4.output 0.00486695 204.18026195 + ------------------------------------------------------------------------------------- + TOTAL 0.04575695 109.41060018 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 58204 +BPFP 0.1967 bits/point +EBPFP 0.3934 equivalent bits/point +MSE 109.410600 +---------------------- -------------------------------------------------------- +Time: 3.722s Load: 0.009s, Pack+Encode: 2.172s, Decode+Unpack: 1.542s +---------------------- -------------------------------------------------------- +💾 Converting with 109.4106 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 356, 128) +Output shape: (1, 356, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.output: torch.Size([1, 356, 3584]) -> torch.Size([1, 1, 356, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,840B, BPFP=0.2563 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,656B, BPFP=0.3360 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,492B, BPFP=0.3288 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,940B, BPFP=0.3046 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,724B, BPFP=0.2512 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,220B, BPFP=0.2730 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,316B, BPFP=0.2333 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,992B, BPFP=0.3069 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,356B, BPFP=0.3667 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,296B, BPFP=0.3202 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,908B, BPFP=0.0182 +⌛️ [2/4] FRONTEND: Frontend time: 2.731s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.903s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15223843 396.91217521 + layer.0.v_cache 0.00001740 0.04485701 + layer.1.k_cache 0.82972649 24.67124956 + layer.1.v_cache 0.00000631 0.01807119 + layer.2.k_cache 0.03086275 2.79978823 + layer.2.v_cache 0.00001960 0.05073592 + layer.3.k_cache 0.01100341 12.57642150 + layer.3.v_cache 0.00002062 0.05920648 + layer.4.k_cache 0.00073857 1.28592382 + layer.4.v_cache 0.00005279 0.09469210 + layer.4.output 0.03758925 152.46192817 + ------------------------------------------------------------------------------------- + TOTAL 0.07575360 88.57333048 + (elements=3,098,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3098624 +Total Bytes 70740 +BPFP 0.1826 bits/point +EBPFP 0.3653 equivalent bits/point +MSE 88.573330 +---------------------- -------------------------------------------------------- +Time: 4.646s Load: 0.011s, Pack+Encode: 2.731s, Decode+Unpack: 1.903s +---------------------- -------------------------------------------------------- +💾 Converting with 88.5733 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,456B, BPFP=0.2617 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,948B, BPFP=0.3494 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,548B, BPFP=0.3846 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,456B, BPFP=0.3205 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,724B, BPFP=0.2775 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,564B, BPFP=0.3268 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,272B, BPFP=0.2509 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,176B, BPFP=0.3040 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,744B, BPFP=0.3961 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,392B, BPFP=0.3755 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,356B, BPFP=0.0198 +⌛️ [2/4] FRONTEND: Frontend time: 2.085s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.573s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12347319 398.32630404 + layer.0.v_cache 0.00001696 0.04564525 + layer.1.k_cache 0.44284580 24.51059350 + layer.1.v_cache 0.00000660 0.01792701 + layer.2.k_cache 0.00882101 2.87278621 + layer.2.v_cache 0.00002210 0.05481831 + layer.3.k_cache 0.01145097 13.90290913 + layer.3.v_cache 0.00001984 0.06098392 + layer.4.k_cache 0.00068754 1.31315154 + layer.4.v_cache 0.00006483 0.10201728 + layer.4.output 0.00500964 208.71403397 + ------------------------------------------------------------------------------------- + TOTAL 0.03661626 111.89443376 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 57636 +BPFP 0.1992 bits/point +EBPFP 0.3983 equivalent bits/point +MSE 111.894434 +---------------------- -------------------------------------------------------- +Time: 3.668s Load: 0.009s, Pack+Encode: 2.085s, Decode+Unpack: 1.573s +---------------------- -------------------------------------------------------- +💾 Converting with 111.8944 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,836B, BPFP=0.2651 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,948B, BPFP=0.3261 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,640B, BPFP=0.3092 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,544B, BPFP=0.3039 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,492B, BPFP=0.2463 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,016B, BPFP=0.2750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,180B, BPFP=0.2292 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,428B, BPFP=0.2976 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,628B, BPFP=0.3634 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,948B, BPFP=0.3261 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,320B, BPFP=0.0182 +⌛️ [2/4] FRONTEND: Frontend time: 2.110s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.583s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13002123 397.69873904 + layer.0.v_cache 0.00001833 0.04589698 + layer.1.k_cache 0.50397591 24.24724164 + layer.1.v_cache 0.00000608 0.01697842 + layer.2.k_cache 0.01926375 2.85213238 + layer.2.v_cache 0.00002060 0.05009440 + layer.3.k_cache 0.04992905 12.60834104 + layer.3.v_cache 0.00002096 0.06218411 + layer.4.k_cache 0.00069540 1.30712120 + layer.4.v_cache 0.00005467 0.10209978 + layer.4.output 0.00466491 194.84968672 + ------------------------------------------------------------------------------------- + TOTAL 0.04333296 106.05521388 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 55980 +BPFP 0.1805 bits/point +EBPFP 0.3611 equivalent bits/point +MSE 106.055214 +---------------------- -------------------------------------------------------- +Time: 3.702s Load: 0.009s, Pack+Encode: 2.110s, Decode+Unpack: 1.583s +---------------------- -------------------------------------------------------- +💾 Converting with 106.0552 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,440B, BPFP=0.2478 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,680B, BPFP=0.3170 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,136B, BPFP=0.3424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,676B, BPFP=0.3167 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,732B, BPFP=0.2641 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,804B, BPFP=0.3239 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,552B, BPFP=0.2540 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,488B, BPFP=0.3063 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,604B, BPFP=0.3685 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,124B, BPFP=0.3417 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,268B, BPFP=0.0181 +⌛️ [2/4] FRONTEND: Frontend time: 2.151s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.824s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15974760 396.35479911 + layer.0.v_cache 0.00001739 0.04506941 + layer.1.k_cache 0.45644542 24.32957241 + layer.1.v_cache 0.00000643 0.01746778 + layer.2.k_cache 0.01631587 2.94921177 + layer.2.v_cache 0.00002005 0.05017069 + layer.3.k_cache 0.01511352 13.62213309 + layer.3.v_cache 0.00002063 0.06112598 + layer.4.k_cache 0.00073517 1.29333725 + layer.4.v_cache 0.00005091 0.09987362 + layer.4.output 0.00481773 198.36139987 + ------------------------------------------------------------------------------------- + TOTAL 0.04012924 107.49132707 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 57504 +BPFP 0.1888 bits/point +EBPFP 0.3775 equivalent bits/point +MSE 107.491327 +---------------------- -------------------------------------------------------- +Time: 3.985s Load: 0.010s, Pack+Encode: 2.151s, Decode+Unpack: 1.824s +---------------------- -------------------------------------------------------- +💾 Converting with 107.4913 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,492B, BPFP=0.2629 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,340B, BPFP=0.3710 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,536B, BPFP=0.3825 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,364B, BPFP=0.3139 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,656B, BPFP=0.2725 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,788B, BPFP=0.3387 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,296B, BPFP=0.2514 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,320B, BPFP=0.3113 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,536B, BPFP=0.3825 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,660B, BPFP=0.3312 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,452B, BPFP=0.0205 +⌛️ [2/4] FRONTEND: Frontend time: 2.120s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.533s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14148096 397.04421231 + layer.0.v_cache 0.00001588 0.04345173 + layer.1.k_cache 0.55266568 25.76652841 + layer.1.v_cache 0.00000632 0.01700930 + layer.2.k_cache 0.00687984 3.15580835 + layer.2.v_cache 0.00001885 0.05039615 + layer.3.k_cache 0.01675770 13.36541451 + layer.3.v_cache 0.00001964 0.05870735 + layer.4.k_cache 0.00068991 1.27827311 + layer.4.v_cache 0.00004859 0.09596607 + layer.4.output 0.00492290 207.99481675 + ------------------------------------------------------------------------------------- + TOTAL 0.04429669 111.57879321 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 57440 +BPFP 0.1977 bits/point +EBPFP 0.3955 equivalent bits/point +MSE 111.578793 +---------------------- -------------------------------------------------------- +Time: 3.664s Load: 0.011s, Pack+Encode: 2.120s, Decode+Unpack: 1.533s +---------------------- -------------------------------------------------------- +💾 Converting with 111.5788 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,844B, BPFP=0.2343 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,356B, BPFP=0.3075 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,264B, BPFP=0.3514 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,484B, BPFP=0.3137 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,360B, BPFP=0.2593 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,176B, BPFP=0.2988 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,940B, BPFP=0.2390 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,168B, BPFP=0.2984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,552B, BPFP=0.3653 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,436B, BPFP=0.3113 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,764B, BPFP=0.0191 +⌛️ [2/4] FRONTEND: Frontend time: 2.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.643s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13144862 393.96693595 + layer.0.v_cache 0.00001769 0.04598504 + layer.1.k_cache 0.64513603 25.62151400 + layer.1.v_cache 0.00000607 0.01704692 + layer.2.k_cache 0.02726801 2.75829455 + layer.2.v_cache 0.00001956 0.05094209 + layer.3.k_cache 0.02271994 14.49490555 + layer.3.v_cache 0.00002018 0.05695741 + layer.4.k_cache 0.00073057 1.28272448 + layer.4.v_cache 0.00004940 0.08974331 + layer.4.output 0.04135688 168.01459531 + ------------------------------------------------------------------------------------- + TOTAL 0.06570084 94.96983626 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 64344 +BPFP 0.1831 bits/point +EBPFP 0.3662 equivalent bits/point +MSE 94.969836 +---------------------- -------------------------------------------------------- +Time: 3.783s Load: 0.010s, Pack+Encode: 2.130s, Decode+Unpack: 1.643s +---------------------- -------------------------------------------------------- +💾 Converting with 94.9698 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,672B, BPFP=0.2669 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,924B, BPFP=0.3578 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,392B, BPFP=0.3919 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,312B, BPFP=0.3134 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,888B, BPFP=0.2826 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,376B, BPFP=0.3180 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,740B, BPFP=0.2718 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,528B, BPFP=0.3291 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,732B, BPFP=0.4166 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,956B, BPFP=0.3602 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,180B, BPFP=0.0226 +⌛️ [2/4] FRONTEND: Frontend time: 2.416s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.706s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12396691 389.13615552 + layer.0.v_cache 0.00001598 0.04649904 + layer.1.k_cache 0.24854917 25.99825127 + layer.1.v_cache 0.00000584 0.01714439 + layer.2.k_cache 0.02290044 3.00221089 + layer.2.v_cache 0.00002299 0.04935124 + layer.3.k_cache 0.01294550 13.03294763 + layer.3.v_cache 0.00002077 0.06000902 + layer.4.k_cache 0.00065144 1.29585131 + layer.4.v_cache 0.00005054 0.09929689 + layer.4.output 1.42385976 251.96519934 + ------------------------------------------------------------------------------------- + TOTAL 0.61036164 129.20553603 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 47700 +BPFP 0.2039 bits/point +EBPFP 0.4078 equivalent bits/point +MSE 129.205536 +---------------------- -------------------------------------------------------- +Time: 4.130s Load: 0.008s, Pack+Encode: 2.416s, Decode+Unpack: 1.706s +---------------------- -------------------------------------------------------- +💾 Converting with 129.2055 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,404B, BPFP=0.2381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,140B, BPFP=0.3320 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,036B, BPFP=0.3263 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,580B, BPFP=0.3017 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,652B, BPFP=0.2515 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,440B, BPFP=0.2941 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,364B, BPFP=0.2359 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,204B, BPFP=0.2814 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,688B, BPFP=0.3616 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,904B, BPFP=0.3192 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,284B, BPFP=0.0176 +⌛️ [2/4] FRONTEND: Frontend time: 2.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.570s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14783513 391.07612457 + layer.0.v_cache 0.00001600 0.04583024 + layer.1.k_cache 0.54442752 25.29683445 + layer.1.v_cache 0.00000628 0.01732919 + layer.2.k_cache 0.01848349 2.89345881 + layer.2.v_cache 0.00002061 0.05288940 + layer.3.k_cache 0.01777397 12.87051085 + layer.3.v_cache 0.00002092 0.06361062 + layer.4.k_cache 0.00070447 1.32020226 + layer.4.v_cache 0.00005049 0.09866058 + layer.4.output 0.04615090 187.76560183 + ------------------------------------------------------------------------------------- + TOTAL 0.06190560 102.82909787 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 56696 +BPFP 0.1803 bits/point +EBPFP 0.3606 equivalent bits/point +MSE 102.829098 +---------------------- -------------------------------------------------------- +Time: 3.904s Load: 0.014s, Pack+Encode: 2.319s, Decode+Unpack: 1.570s +---------------------- -------------------------------------------------------- +💾 Converting with 102.8291 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,852B, BPFP=0.2544 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,008B, BPFP=0.3150 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,508B, BPFP=0.3412 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,776B, BPFP=0.3029 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,024B, BPFP=0.2634 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,188B, BPFP=0.2720 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,628B, BPFP=0.2427 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,360B, BPFP=0.2810 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,772B, BPFP=0.3551 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,660B, BPFP=0.2968 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,524B, BPFP=0.0189 +⌛️ [2/4] FRONTEND: Frontend time: 2.084s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16274841 402.99252831 + layer.0.v_cache 0.00001794 0.04535095 + layer.1.k_cache 0.56737488 25.51793205 + layer.1.v_cache 0.00000586 0.01670934 + layer.2.k_cache 0.02027899 2.88743929 + layer.2.v_cache 0.00002117 0.04892455 + layer.3.k_cache 0.01891868 13.11964693 + layer.3.v_cache 0.00002140 0.05888612 + layer.4.k_cache 0.00069658 1.27137808 + layer.4.v_cache 0.00005005 0.09333570 + layer.4.output 0.04476916 182.09580237 + ------------------------------------------------------------------------------------- + TOTAL 0.06373636 101.21898517 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 58300 +BPFP 0.1798 bits/point +EBPFP 0.3596 equivalent bits/point +MSE 101.218985 +---------------------- -------------------------------------------------------- +Time: 3.442s Load: 0.009s, Pack+Encode: 2.084s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 101.2190 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,040B, BPFP=0.2312 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,064B, BPFP=0.3471 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,384B, BPFP=0.3654 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,832B, BPFP=0.3338 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,720B, BPFP=0.2701 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,368B, BPFP=0.3072 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,380B, BPFP=0.2507 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,964B, BPFP=0.3413 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,660B, BPFP=0.3812 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,428B, BPFP=0.3679 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,408B, BPFP=0.0197 +⌛️ [2/4] FRONTEND: Frontend time: 1.845s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13366390 383.58173077 + layer.0.v_cache 0.00001719 0.04539381 + layer.1.k_cache 0.57618574 25.42466697 + layer.1.v_cache 0.00000658 0.01691531 + layer.2.k_cache 0.02858028 2.78292031 + layer.2.v_cache 0.00002021 0.04776567 + layer.3.k_cache 0.00873641 12.98234228 + layer.3.v_cache 0.00002023 0.05739753 + layer.4.k_cache 0.00067337 1.27219232 + layer.4.v_cache 0.00005011 0.09575256 + layer.4.output 0.00484872 203.42932365 + ------------------------------------------------------------------------------------- + TOTAL 0.04599383 108.84190254 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 58248 +BPFP 0.1961 bits/point +EBPFP 0.3922 equivalent bits/point +MSE 108.841903 +---------------------- -------------------------------------------------------- +Time: 3.184s Load: 0.012s, Pack+Encode: 1.845s, Decode+Unpack: 1.328s +---------------------- -------------------------------------------------------- +💾 Converting with 108.8419 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,808B, BPFP=0.2692 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,620B, BPFP=0.3266 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,820B, BPFP=0.3408 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,128B, BPFP=0.2919 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,872B, BPFP=0.2738 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,292B, BPFP=0.3035 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,688B, BPFP=0.2607 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,644B, BPFP=0.3283 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,252B, BPFP=0.3713 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,588B, BPFP=0.3244 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,060B, BPFP=0.0208 +⌛️ [2/4] FRONTEND: Frontend time: 1.748s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.226s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13057878 396.97617364 + layer.0.v_cache 0.00001797 0.04749787 + layer.1.k_cache 0.31658528 25.82415468 + layer.1.v_cache 0.00000647 0.01791117 + layer.2.k_cache 0.02849985 2.87338519 + layer.2.v_cache 0.00001991 0.05247614 + layer.3.k_cache 0.03599559 14.12088054 + layer.3.v_cache 0.00002239 0.06579486 + layer.4.k_cache 0.00078811 1.31444097 + layer.4.v_cache 0.00005540 0.10533368 + layer.4.output 1.38530015 245.18214690 + ------------------------------------------------------------------------------------- + TOTAL 0.60056887 126.92194571 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 45772 +BPFP 0.1904 bits/point +EBPFP 0.3807 equivalent bits/point +MSE 126.921946 +---------------------- -------------------------------------------------------- +Time: 2.982s Load: 0.008s, Pack+Encode: 1.748s, Decode+Unpack: 1.226s +---------------------- -------------------------------------------------------- +💾 Converting with 126.9219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,616B, BPFP=0.2604 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,016B, BPFP=0.3612 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,180B, BPFP=0.3730 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,724B, BPFP=0.3401 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,108B, BPFP=0.2958 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,220B, BPFP=0.3039 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,844B, BPFP=0.2768 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,684B, BPFP=0.3373 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,624B, BPFP=0.4050 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,848B, BPFP=0.3491 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,096B, BPFP=0.0216 +⌛️ [2/4] FRONTEND: Frontend time: 1.711s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12725546 390.57682892 + layer.0.v_cache 0.00001736 0.04766864 + layer.1.k_cache 0.32262681 25.12535552 + layer.1.v_cache 0.00000608 0.01779392 + layer.2.k_cache 0.02157613 2.85353644 + layer.2.v_cache 0.00002208 0.05201151 + layer.3.k_cache 0.01760322 13.71590469 + layer.3.v_cache 0.00002192 0.06643609 + layer.4.k_cache 0.00069785 1.31225783 + layer.4.v_cache 0.00005172 0.10451544 + layer.4.output 1.41081570 249.69033904 + ------------------------------------------------------------------------------------- + TOTAL 0.60974050 128.33556955 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 47960 +BPFP 0.2031 bits/point +EBPFP 0.4063 equivalent bits/point +MSE 128.335570 +---------------------- -------------------------------------------------------- +Time: 3.045s Load: 0.008s, Pack+Encode: 1.711s, Decode+Unpack: 1.327s +---------------------- -------------------------------------------------------- +💾 Converting with 128.3356 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,556B, BPFP=0.2609 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,720B, BPFP=0.3462 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,396B, BPFP=0.3958 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,612B, BPFP=0.3383 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,280B, BPFP=0.3140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,456B, BPFP=0.3269 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,832B, BPFP=0.2811 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,876B, BPFP=0.3577 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,268B, BPFP=0.3864 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,180B, BPFP=0.3800 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,368B, BPFP=0.0248 +⌛️ [2/4] FRONTEND: Frontend time: 1.931s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.540s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13788442 391.48752934 + layer.0.v_cache 0.00001717 0.04913031 + layer.1.k_cache 0.30723858 25.15837735 + layer.1.v_cache 0.00000664 0.01870417 + layer.2.k_cache 0.01337219 3.00223566 + layer.2.v_cache 0.00002159 0.05227836 + layer.3.k_cache 0.03522867 13.13147947 + layer.3.v_cache 0.00002126 0.06537453 + layer.4.k_cache 0.00073009 1.30938105 + layer.4.v_cache 0.00005119 0.10704408 + layer.4.output 1.43730973 254.39987844 + ------------------------------------------------------------------------------------- + TOTAL 0.62092588 130.30474608 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 48544 +BPFP 0.2095 bits/point +EBPFP 0.4189 equivalent bits/point +MSE 130.304746 +---------------------- -------------------------------------------------------- +Time: 3.478s Load: 0.007s, Pack+Encode: 1.931s, Decode+Unpack: 1.540s +---------------------- -------------------------------------------------------- +💾 Converting with 130.3047 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,508B, BPFP=0.2687 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,536B, BPFP=0.3474 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,384B, BPFP=0.4124 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,272B, BPFP=0.3272 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,120B, BPFP=0.3156 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,228B, BPFP=0.3238 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,840B, BPFP=0.2941 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,564B, BPFP=0.3496 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,652B, BPFP=0.4329 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,460B, BPFP=0.4182 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,356B, BPFP=0.0258 +⌛️ [2/4] FRONTEND: Frontend time: 1.986s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.475s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11546621 400.35309436 + layer.0.v_cache 0.00001823 0.04813745 + layer.1.k_cache 0.26924565 25.19944374 + layer.1.v_cache 0.00000737 0.01861861 + layer.2.k_cache 0.01761903 3.16484728 + layer.2.v_cache 0.00002145 0.05353373 + layer.3.k_cache 0.02374693 14.00368844 + layer.3.v_cache 0.00002104 0.06435727 + layer.4.k_cache 0.00067272 1.31188815 + layer.4.v_cache 0.00005153 0.10406825 + layer.4.output 1.50071327 265.61458333 + ------------------------------------------------------------------------------------- + TOTAL 0.64305077 135.50728004 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 47920 +BPFP 0.2159 bits/point +EBPFP 0.4318 equivalent bits/point +MSE 135.507280 +---------------------- -------------------------------------------------------- +Time: 3.468s Load: 0.007s, Pack+Encode: 1.986s, Decode+Unpack: 1.475s +---------------------- -------------------------------------------------------- +💾 Converting with 135.5073 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,888B, BPFP=0.2480 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,840B, BPFP=0.2963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,396B, BPFP=0.3245 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,272B, BPFP=0.2675 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,820B, BPFP=0.2445 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,172B, BPFP=0.2624 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,640B, BPFP=0.2354 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,460B, BPFP=0.2770 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,404B, BPFP=0.3249 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,140B, BPFP=0.3115 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,340B, BPFP=0.0170 +⌛️ [2/4] FRONTEND: Frontend time: 2.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.584s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18206066 399.37281859 + layer.0.v_cache 0.00001652 0.04663504 + layer.1.k_cache 0.62405802 25.68377131 + layer.1.v_cache 0.00000616 0.01769782 + layer.2.k_cache 0.02036299 2.85908409 + layer.2.v_cache 0.00002087 0.05215883 + layer.3.k_cache 0.02546444 13.04844606 + layer.3.v_cache 0.00002064 0.06089793 + layer.4.k_cache 0.00069880 1.30402117 + layer.4.v_cache 0.00007146 0.10285780 + layer.4.output 0.04336799 176.20832850 + ------------------------------------------------------------------------------------- + TOTAL 0.06802097 98.58862871 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 57372 +BPFP 0.1712 bits/point +EBPFP 0.3424 equivalent bits/point +MSE 98.588629 +---------------------- -------------------------------------------------------- +Time: 3.847s Load: 0.011s, Pack+Encode: 2.253s, Decode+Unpack: 1.584s +---------------------- -------------------------------------------------------- +💾 Converting with 98.5886 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,220B, BPFP=0.2197 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,756B, BPFP=0.3245 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,416B, BPFP=0.3695 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,288B, BPFP=0.2926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,048B, BPFP=0.2762 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,572B, BPFP=0.3120 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,860B, BPFP=0.2634 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,728B, BPFP=0.3226 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,876B, BPFP=0.4009 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,784B, BPFP=0.3264 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,228B, BPFP=0.0217 +⌛️ [2/4] FRONTEND: Frontend time: 2.005s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.510s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14018573 378.40778521 + layer.0.v_cache 0.00001653 0.04763121 + layer.1.k_cache 0.40671090 25.64263783 + layer.1.v_cache 0.00000662 0.01857442 + layer.2.k_cache 0.03036383 2.98051486 + layer.2.v_cache 0.00002088 0.05365044 + layer.3.k_cache 0.05940984 12.38973799 + layer.3.v_cache 0.00002025 0.06139792 + layer.4.k_cache 0.00071161 1.32102040 + layer.4.v_cache 0.00005094 0.09707741 + layer.4.output 1.33692960 236.62835309 + ------------------------------------------------------------------------------------- + TOTAL 0.58800025 122.20108819 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 47776 +BPFP 0.1918 bits/point +EBPFP 0.3835 equivalent bits/point +MSE 122.201088 +---------------------- -------------------------------------------------------- +Time: 3.523s Load: 0.008s, Pack+Encode: 2.005s, Decode+Unpack: 1.510s +---------------------- -------------------------------------------------------- +💾 Converting with 122.2011 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,788B, BPFP=0.2703 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,876B, BPFP=0.3479 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,120B, BPFP=0.3653 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,456B, BPFP=0.3179 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,776B, BPFP=0.2694 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,136B, BPFP=0.2951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,672B, BPFP=0.2620 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,284B, BPFP=0.3057 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,292B, BPFP=0.3776 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,900B, BPFP=0.3496 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,052B, BPFP=0.0209 +⌛️ [2/4] FRONTEND: Frontend time: 2.044s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.453s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351684 397.71646689 + layer.0.v_cache 0.00001872 0.04892137 + layer.1.k_cache 0.28219827 26.40165035 + layer.1.v_cache 0.00000627 0.01830177 + layer.2.k_cache 0.01639774 2.92200710 + layer.2.v_cache 0.00002172 0.05467003 + layer.3.k_cache 0.02752649 13.60310315 + layer.3.v_cache 0.00002090 0.06410810 + layer.4.k_cache 0.00067264 1.29875162 + layer.4.v_cache 0.00005058 0.10488140 + layer.4.output 1.39794763 247.41344586 + ------------------------------------------------------------------------------------- + TOTAL 0.60212139 127.88982252 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 46352 +BPFP 0.1945 bits/point +EBPFP 0.3891 equivalent bits/point +MSE 127.889823 +---------------------- -------------------------------------------------------- +Time: 3.505s Load: 0.008s, Pack+Encode: 2.044s, Decode+Unpack: 1.453s +---------------------- -------------------------------------------------------- +💾 Converting with 127.8898 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,572B, BPFP=0.2666 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,236B, BPFP=0.3636 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,416B, BPFP=0.3741 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,880B, BPFP=0.3428 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,744B, BPFP=0.2766 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,532B, BPFP=0.3225 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,216B, BPFP=0.2458 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,900B, BPFP=0.3440 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,708B, BPFP=0.3911 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,496B, BPFP=0.3787 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,572B, BPFP=0.0214 +⌛️ [2/4] FRONTEND: Frontend time: 2.619s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.558s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15646637 395.93094100 + layer.0.v_cache 0.00001727 0.04574507 + layer.1.k_cache 0.53079696 25.48989003 + layer.1.v_cache 0.00000635 0.01795688 + layer.2.k_cache 0.02340859 2.97617443 + layer.2.v_cache 0.00002061 0.05016084 + layer.3.k_cache 0.01050839 13.79215616 + layer.3.v_cache 0.00002156 0.05883560 + layer.4.k_cache 0.00068596 1.27624432 + layer.4.v_cache 0.00006072 0.09940922 + layer.4.output 0.00497423 207.23349214 + ------------------------------------------------------------------------------------- + TOTAL 0.04451838 111.19835050 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 59272 +BPFP 0.2033 bits/point +EBPFP 0.4066 equivalent bits/point +MSE 111.198351 +---------------------- -------------------------------------------------------- +Time: 4.188s Load: 0.011s, Pack+Encode: 2.619s, Decode+Unpack: 1.558s +---------------------- -------------------------------------------------------- +💾 Converting with 111.1984 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,936B, BPFP=0.2504 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,956B, BPFP=0.3022 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,440B, BPFP=0.3267 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,744B, BPFP=0.2914 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,848B, BPFP=0.2459 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,628B, BPFP=0.2855 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,608B, BPFP=0.2338 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,036B, BPFP=0.3062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,072B, BPFP=0.3588 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,968B, BPFP=0.3028 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,368B, BPFP=0.0172 +⌛️ [2/4] FRONTEND: Frontend time: 2.138s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.568s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15292851 403.97661323 + layer.0.v_cache 0.00001614 0.04690350 + layer.1.k_cache 0.61736927 25.23286894 + layer.1.v_cache 0.00000669 0.01865489 + layer.2.k_cache 0.02003183 3.00742806 + layer.2.v_cache 0.00002344 0.05099644 + layer.3.k_cache 0.01028479 12.38263970 + layer.3.v_cache 0.00002212 0.05894036 + layer.4.k_cache 0.00070434 1.35283601 + layer.4.v_cache 0.00006490 0.09856345 + layer.4.output 0.04342050 176.22572182 + ------------------------------------------------------------------------------------- + TOTAL 0.06502327 98.81214690 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 59604 +BPFP 0.1779 bits/point +EBPFP 0.3557 equivalent bits/point +MSE 98.812147 +---------------------- -------------------------------------------------------- +Time: 3.718s Load: 0.011s, Pack+Encode: 2.138s, Decode+Unpack: 1.568s +---------------------- -------------------------------------------------------- +💾 Converting with 98.8121 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,460B, BPFP=0.2591 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,964B, BPFP=0.3464 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,208B, BPFP=0.3606 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,804B, BPFP=0.3371 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,832B, BPFP=0.2807 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,664B, BPFP=0.3290 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,140B, BPFP=0.2405 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,280B, BPFP=0.3067 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,872B, BPFP=0.3992 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,520B, BPFP=0.3787 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,504B, BPFP=0.0208 +⌛️ [2/4] FRONTEND: Frontend time: 2.630s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.871s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12950333 392.37836896 + layer.0.v_cache 0.00001690 0.04559215 + layer.1.k_cache 0.52417032 25.12611815 + layer.1.v_cache 0.00000602 0.01727048 + layer.2.k_cache 0.01578268 2.82086431 + layer.2.v_cache 0.00002268 0.05348956 + layer.3.k_cache 0.03345619 13.64677480 + layer.3.v_cache 0.00002007 0.06280688 + layer.4.k_cache 0.00067301 1.27780458 + layer.4.v_cache 0.00005077 0.10036097 + layer.4.output 0.00489311 206.44825412 + ------------------------------------------------------------------------------------- + TOTAL 0.04340905 110.62748410 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 58248 +BPFP 0.1990 bits/point +EBPFP 0.3980 equivalent bits/point +MSE 110.627484 +---------------------- -------------------------------------------------------- +Time: 4.510s Load: 0.010s, Pack+Encode: 2.630s, Decode+Unpack: 1.871s +---------------------- -------------------------------------------------------- +💾 Converting with 110.6275 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,712B, BPFP=0.2624 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,460B, BPFP=0.3153 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,552B, BPFP=0.3218 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,284B, BPFP=0.3029 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,852B, BPFP=0.2723 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,264B, BPFP=0.3015 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,668B, BPFP=0.2593 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,400B, BPFP=0.3111 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,964B, BPFP=0.3510 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,552B, BPFP=0.3218 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,120B, BPFP=0.0214 +⌛️ [2/4] FRONTEND: Frontend time: 2.477s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.693s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13888381 402.09990102 + layer.0.v_cache 0.00001731 0.04556724 + layer.1.k_cache 0.27925331 23.34016191 + layer.1.v_cache 0.00000589 0.01694409 + layer.2.k_cache 0.03172064 2.90602713 + layer.2.v_cache 0.00001980 0.04918258 + layer.3.k_cache 0.01738505 14.32370175 + layer.3.v_cache 0.00001961 0.05854267 + layer.4.k_cache 0.00068340 1.22614050 + layer.4.v_cache 0.00005063 0.09945983 + layer.4.output 1.38525280 245.16824903 + ------------------------------------------------------------------------------------- + TOTAL 0.59792994 127.07902188 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 44828 +BPFP 0.1864 bits/point +EBPFP 0.3729 equivalent bits/point +MSE 127.079022 +---------------------- -------------------------------------------------------- +Time: 4.178s Load: 0.008s, Pack+Encode: 2.477s, Decode+Unpack: 1.693s +---------------------- -------------------------------------------------------- +💾 Converting with 127.0790 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.1890 bits/point +Avg EBPFP 0.3780 equivalent bits/point +Avg MSE 111.614219 +Avg Time 3.773s +------------------------ ---------------------------- diff --git a/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..384a452bef398de7a9576bf935f9d0603162f511 --- /dev/null +++ b/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 286 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other +Output output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other +---------------- ------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,372B, BPFP=0.2621 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,664B, BPFP=0.3626 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,972B, BPFP=0.3865 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,380B, BPFP=0.3405 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,108B, BPFP=0.3193 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,296B, BPFP=0.3340 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,600B, BPFP=0.2799 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,124B, BPFP=0.3206 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,500B, BPFP=0.4275 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,976B, BPFP=0.3868 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,192B, BPFP=0.0243 +⌛️ [2/4] FRONTEND: Frontend time: 2.707s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.423s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09475397 392.20794465 + layer.0.v_cache 0.00001927 0.04776344 + layer.1.k_cache 0.17401673 25.73004606 + layer.1.v_cache 0.00000670 0.01834416 + layer.2.k_cache 0.00814445 3.05812490 + layer.2.v_cache 0.00002157 0.05559423 + layer.3.k_cache 0.01691350 13.75242319 + layer.3.v_cache 0.00002161 0.06862465 + layer.4.k_cache 0.00069969 1.41061872 + layer.4.v_cache 0.00005477 0.11154268 + layer.4.output 1.52309445 269.48391969 + ------------------------------------------------------------------------------------- + TOTAL 0.64448903 136.63814497 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 46184 +BPFP 0.2112 bits/point +EBPFP 0.4224 equivalent bits/point +MSE 136.638145 +---------------------- -------------------------------------------------------- +Time: 4.137s Load: 0.008s, Pack+Encode: 2.707s, Decode+Unpack: 1.423s +---------------------- -------------------------------------------------------- +💾 Converting with 136.6381 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 212, 128) +Output shape: (1, 212, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.output: torch.Size([1, 212, 3584]) -> torch.Size([1, 1, 212, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,564B, BPFP=0.2627 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,752B, BPFP=0.3502 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,408B, BPFP=0.3986 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,508B, BPFP=0.3323 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,112B, BPFP=0.3031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,548B, BPFP=0.3352 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,880B, BPFP=0.2860 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,636B, BPFP=0.3417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,216B, BPFP=0.3844 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,108B, BPFP=0.3765 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,352B, BPFP=0.0248 +⌛️ [2/4] FRONTEND: Frontend time: 1.876s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.412s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09403566 389.31876474 + layer.0.v_cache 0.00001799 0.04808104 + layer.1.k_cache 0.24957257 24.95809764 + layer.1.v_cache 0.00000621 0.01765275 + layer.2.k_cache 0.01169950 2.97403530 + layer.2.v_cache 0.00002031 0.05296281 + layer.3.k_cache 0.03645871 13.73117353 + layer.3.v_cache 0.00002058 0.06577713 + layer.4.k_cache 0.00069161 1.35832675 + layer.4.v_cache 0.00005130 0.10475135 + layer.4.output 1.44404146 255.58275775 + ------------------------------------------------------------------------------------- + TOTAL 0.61769792 130.68876043 + (elements=1,845,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1845248 +Total Bytes 48084 +BPFP 0.2085 bits/point +EBPFP 0.4169 equivalent bits/point +MSE 130.688760 +---------------------- -------------------------------------------------------- +Time: 3.294s Load: 0.007s, Pack+Encode: 1.876s, Decode+Unpack: 1.412s +---------------------- -------------------------------------------------------- +💾 Converting with 130.6888 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 250, 128) +Output shape: (1, 250, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.output: torch.Size([1, 250, 3584]) -> torch.Size([1, 1, 250, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,564B, BPFP=0.2228 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,436B, BPFP=0.2772 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,032B, BPFP=0.3145 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,308B, BPFP=0.2692 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,664B, BPFP=0.2290 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,188B, BPFP=0.2617 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,492B, BPFP=0.2182 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,676B, BPFP=0.2923 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,856B, BPFP=0.3035 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,364B, BPFP=0.2727 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,540B, BPFP=0.0227 +⌛️ [2/4] FRONTEND: Frontend time: 2.420s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.668s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09121272 388.57831250 + layer.0.v_cache 0.00001793 0.04640640 + layer.1.k_cache 0.34710925 25.32284180 + layer.1.v_cache 0.00000615 0.01765014 + layer.2.k_cache 0.02211576 3.20649316 + layer.2.v_cache 0.00002179 0.05346134 + layer.3.k_cache 0.02174057 14.17977148 + layer.3.v_cache 0.00001938 0.05981859 + layer.4.k_cache 0.00073173 1.34753320 + layer.4.v_cache 0.00005145 0.10256290 + layer.4.output 1.22463198 216.73875000 + ------------------------------------------------------------------------------------- + TOTAL 0.53267356 114.71094715 + (elements=2,176,000) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2176000 +Total Bytes 45120 +BPFP 0.1659 bits/point +EBPFP 0.3318 equivalent bits/point +MSE 114.710947 +---------------------- -------------------------------------------------------- +Time: 4.099s Load: 0.011s, Pack+Encode: 2.420s, Decode+Unpack: 1.668s +---------------------- -------------------------------------------------------- +💾 Converting with 114.7109 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,680B, BPFP=0.2533 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,544B, BPFP=0.3128 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,128B, BPFP=0.3530 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,420B, BPFP=0.3042 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,948B, BPFP=0.2718 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,908B, BPFP=0.2690 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,792B, BPFP=0.2610 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,284B, BPFP=0.2949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,104B, BPFP=0.3513 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,572B, BPFP=0.3147 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,240B, BPFP=0.0220 +⌛️ [2/4] FRONTEND: Frontend time: 1.943s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.434s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12780563 397.72301762 + layer.0.v_cache 0.00001798 0.04721587 + layer.1.k_cache 0.36413978 24.54823875 + layer.1.v_cache 0.00000619 0.01753752 + layer.2.k_cache 0.01529831 3.01565861 + layer.2.v_cache 0.00002052 0.05164087 + layer.3.k_cache 0.01269315 13.37291674 + layer.3.v_cache 0.00002023 0.06160152 + layer.4.k_cache 0.00069261 1.26866291 + layer.4.v_cache 0.00005545 0.09896834 + layer.4.output 1.34865724 238.68736233 + ------------------------------------------------------------------------------------- + TOTAL 0.58596180 124.17747030 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 45620 +BPFP 0.1847 bits/point +EBPFP 0.3694 equivalent bits/point +MSE 124.177470 +---------------------- -------------------------------------------------------- +Time: 3.385s Load: 0.008s, Pack+Encode: 1.943s, Decode+Unpack: 1.434s +---------------------- -------------------------------------------------------- +💾 Converting with 124.1775 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 214, 128) +Output shape: (1, 214, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.output: torch.Size([1, 214, 3584]) -> torch.Size([1, 1, 214, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,580B, BPFP=0.2614 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,732B, BPFP=0.3455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,376B, BPFP=0.3925 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,480B, BPFP=0.3271 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,968B, BPFP=0.2897 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,184B, BPFP=0.3055 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,828B, BPFP=0.2795 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,240B, BPFP=0.3096 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,156B, BPFP=0.3765 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,932B, BPFP=0.3601 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,308B, BPFP=0.0241 +⌛️ [2/4] FRONTEND: Frontend time: 2.006s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.455s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12820032 391.22495619 + layer.0.v_cache 0.00001693 0.04567141 + layer.1.k_cache 0.24875343 25.20829759 + layer.1.v_cache 0.00000587 0.01672155 + layer.2.k_cache 0.01453762 2.93550295 + layer.2.v_cache 0.00001988 0.04959375 + layer.3.k_cache 0.02276520 13.37121810 + layer.3.v_cache 0.00001932 0.05886954 + layer.4.k_cache 0.00071399 1.26211305 + layer.4.v_cache 0.00004895 0.09894790 + layer.4.output 1.43051836 253.17331442 + ------------------------------------------------------------------------------------- + TOTAL 0.61345353 129.79324076 + (elements=1,862,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1862656 +Total Bytes 46784 +BPFP 0.2009 bits/point +EBPFP 0.4019 equivalent bits/point +MSE 129.793241 +---------------------- -------------------------------------------------------- +Time: 3.469s Load: 0.008s, Pack+Encode: 2.006s, Decode+Unpack: 1.455s +---------------------- -------------------------------------------------------- +💾 Converting with 129.7932 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,288B, BPFP=0.2419 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,564B, BPFP=0.3139 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,348B, BPFP=0.3581 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,576B, BPFP=0.3145 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,680B, BPFP=0.2640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,196B, BPFP=0.2931 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,532B, BPFP=0.2556 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,732B, BPFP=0.3233 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,444B, BPFP=0.3635 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,696B, BPFP=0.3213 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,376B, BPFP=0.0191 +⌛️ [2/4] FRONTEND: Frontend time: 2.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.625s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11875103 389.61287229 + layer.0.v_cache 0.00001689 0.04584120 + layer.1.k_cache 0.49171178 24.96344413 + layer.1.v_cache 0.00000616 0.01768815 + layer.2.k_cache 0.02532451 2.92237298 + layer.2.v_cache 0.00002086 0.05216498 + layer.3.k_cache 0.02370731 13.78042151 + layer.3.v_cache 0.00002129 0.06360267 + layer.4.k_cache 0.00073268 1.27348157 + layer.4.v_cache 0.00004879 0.10123552 + layer.4.output 0.00480151 200.48868618 + ------------------------------------------------------------------------------------- + TOTAL 0.04082070 108.01493696 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 56432 +BPFP 0.1872 bits/point +EBPFP 0.3745 equivalent bits/point +MSE 108.014937 +---------------------- -------------------------------------------------------- +Time: 4.124s Load: 0.010s, Pack+Encode: 2.489s, Decode+Unpack: 1.625s +---------------------- -------------------------------------------------------- +💾 Converting with 108.0149 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,040B, BPFP=0.2577 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,020B, BPFP=0.3202 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,368B, BPFP=0.3423 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,844B, BPFP=0.3089 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,948B, BPFP=0.2518 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,332B, BPFP=0.2763 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,792B, BPFP=0.2418 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,480B, BPFP=0.2857 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,080B, BPFP=0.3240 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,652B, BPFP=0.2967 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,500B, BPFP=0.0228 +⌛️ [2/4] FRONTEND: Frontend time: 1.970s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.438s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10657067 406.76919643 + layer.0.v_cache 0.00001604 0.04574337 + layer.1.k_cache 0.39270948 26.15376475 + layer.1.v_cache 0.00000611 0.01787597 + layer.2.k_cache 0.01541229 3.13918557 + layer.2.v_cache 0.00002285 0.05345167 + layer.3.k_cache 0.01321528 13.15769990 + layer.3.v_cache 0.00001985 0.06194056 + layer.4.k_cache 0.00076706 1.35463618 + layer.4.v_cache 0.00005205 0.10175811 + layer.4.output 1.24960368 221.13458455 + ------------------------------------------------------------------------------------- + TOTAL 0.54564808 117.57631437 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 48056 +BPFP 0.1803 bits/point +EBPFP 0.3606 equivalent bits/point +MSE 117.576314 +---------------------- -------------------------------------------------------- +Time: 3.417s Load: 0.009s, Pack+Encode: 1.970s, Decode+Unpack: 1.438s +---------------------- -------------------------------------------------------- +💾 Converting with 117.5763 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 197, 128) +Output shape: (1, 197, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.output: torch.Size([1, 197, 3584]) -> torch.Size([1, 1, 197, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,808B, BPFP=0.2227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,376B, BPFP=0.3471 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,020B, BPFP=0.3982 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,248B, BPFP=0.3369 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,948B, BPFP=0.3131 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,780B, BPFP=0.2998 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,312B, BPFP=0.2627 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,308B, BPFP=0.3417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,452B, BPFP=0.4324 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,204B, BPFP=0.4128 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,352B, BPFP=0.0266 +⌛️ [2/4] FRONTEND: Frontend time: 1.931s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.430s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09219642 384.35532995 + layer.0.v_cache 0.00001621 0.04535280 + layer.1.k_cache 0.17840836 24.99894660 + layer.1.v_cache 0.00000639 0.01710339 + layer.2.k_cache 0.00922235 3.03847662 + layer.2.v_cache 0.00002142 0.05214770 + layer.3.k_cache 0.01957878 13.74212554 + layer.3.v_cache 0.00002124 0.06263611 + layer.4.k_cache 0.00068177 1.32273470 + layer.4.v_cache 0.00005086 0.10389507 + layer.4.output 1.55401793 275.04280729 + ------------------------------------------------------------------------------------- + TOTAL 0.65754878 138.41402350 + (elements=1,714,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1714688 +Total Bytes 44808 +BPFP 0.2091 bits/point +EBPFP 0.4181 equivalent bits/point +MSE 138.414023 +---------------------- -------------------------------------------------------- +Time: 3.368s Load: 0.007s, Pack+Encode: 1.931s, Decode+Unpack: 1.430s +---------------------- -------------------------------------------------------- +💾 Converting with 138.4140 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,068B, BPFP=0.2554 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,828B, BPFP=0.3442 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,212B, BPFP=0.3131 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,512B, BPFP=0.2778 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,604B, BPFP=0.2321 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,864B, BPFP=0.2956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,320B, BPFP=0.2177 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,304B, BPFP=0.2673 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,124B, BPFP=0.3591 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,720B, BPFP=0.2883 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,792B, BPFP=0.0201 +⌛️ [2/4] FRONTEND: Frontend time: 2.111s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.562s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11169702 404.86398690 + layer.0.v_cache 0.00001595 0.04571422 + layer.1.k_cache 0.54955656 24.89651115 + layer.1.v_cache 0.00000603 0.01694460 + layer.2.k_cache 0.01751511 3.05742778 + layer.2.v_cache 0.00002011 0.05398238 + layer.3.k_cache 0.01650532 13.61234564 + layer.3.v_cache 0.00001967 0.05818542 + layer.4.k_cache 0.00077751 1.35905791 + layer.4.v_cache 0.00005008 0.10009806 + layer.4.output 0.04306356 175.05771889 + ------------------------------------------------------------------------------------- + TOTAL 0.05868284 98.43931096 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 59348 +BPFP 0.1760 bits/point +EBPFP 0.3519 equivalent bits/point +MSE 98.439311 +---------------------- -------------------------------------------------------- +Time: 3.686s Load: 0.013s, Pack+Encode: 2.111s, Decode+Unpack: 1.562s +---------------------- -------------------------------------------------------- +💾 Converting with 98.4393 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,676B, BPFP=0.2635 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,880B, BPFP=0.3498 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,304B, BPFP=0.3802 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,300B, BPFP=0.3082 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,088B, BPFP=0.2930 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,292B, BPFP=0.3076 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,860B, BPFP=0.2767 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,376B, BPFP=0.3136 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,668B, BPFP=0.4062 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,040B, BPFP=0.3612 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,144B, BPFP=0.0220 +⌛️ [2/4] FRONTEND: Frontend time: 2.477s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.541s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09832749 392.12342317 + layer.0.v_cache 0.00001658 0.04801895 + layer.1.k_cache 0.33493462 25.62718607 + layer.1.v_cache 0.00000656 0.01830952 + layer.2.k_cache 0.02762392 3.07048945 + layer.2.v_cache 0.00002012 0.05123115 + layer.3.k_cache 0.01647770 13.84504910 + layer.3.v_cache 0.00001951 0.06263645 + layer.4.k_cache 0.00067889 1.32353476 + layer.4.v_cache 0.00004874 0.10020379 + layer.4.output 1.40435175 248.51714040 + ------------------------------------------------------------------------------------- + TOTAL 0.60638920 127.99353325 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 47628 +BPFP 0.2008 bits/point +EBPFP 0.4016 equivalent bits/point +MSE 127.993533 +---------------------- -------------------------------------------------------- +Time: 4.026s Load: 0.008s, Pack+Encode: 2.477s, Decode+Unpack: 1.541s +---------------------- -------------------------------------------------------- +💾 Converting with 127.9935 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,764B, BPFP=0.2602 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,884B, BPFP=0.3377 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,120B, BPFP=0.3540 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,568B, BPFP=0.3158 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,028B, BPFP=0.2785 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,328B, BPFP=0.2992 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,768B, BPFP=0.2605 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,420B, BPFP=0.3056 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,636B, BPFP=0.3897 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,908B, BPFP=0.3393 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,016B, BPFP=0.0199 +⌛️ [2/4] FRONTEND: Frontend time: 1.921s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.400s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102250 401.25974834 + layer.0.v_cache 0.00001721 0.04400707 + layer.1.k_cache 0.37371607 25.43338418 + layer.1.v_cache 0.00000658 0.01710878 + layer.2.k_cache 0.01127678 3.02522669 + layer.2.v_cache 0.00001998 0.05203626 + layer.3.k_cache 0.02136800 12.74111587 + layer.3.v_cache 0.00001963 0.06096256 + layer.4.k_cache 0.00069450 1.26619295 + layer.4.v_cache 0.00005147 0.10326483 + layer.4.output 1.35464589 239.74435051 + ------------------------------------------------------------------------------------- + TOTAL 0.58945376 124.83608830 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 47440 +BPFP 0.1929 bits/point +EBPFP 0.3859 equivalent bits/point +MSE 124.836088 +---------------------- -------------------------------------------------------- +Time: 3.330s Load: 0.008s, Pack+Encode: 1.921s, Decode+Unpack: 1.400s +---------------------- -------------------------------------------------------- +💾 Converting with 124.8361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,720B, BPFP=0.2316 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,864B, BPFP=0.3028 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,736B, BPFP=0.2948 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,680B, BPFP=0.2913 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,532B, BPFP=0.2199 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,124B, BPFP=0.2567 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,400B, BPFP=0.2117 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,556B, BPFP=0.2836 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,408B, BPFP=0.3367 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,424B, BPFP=0.2754 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,460B, BPFP=0.0219 +⌛️ [2/4] FRONTEND: Frontend time: 2.182s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.425s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11871604 398.25728337 + layer.0.v_cache 0.00001755 0.04666931 + layer.1.k_cache 0.46315337 24.82942838 + layer.1.v_cache 0.00000588 0.01716651 + layer.2.k_cache 0.02202094 3.06340458 + layer.2.v_cache 0.00001989 0.05233065 + layer.3.k_cache 0.01278766 13.59594824 + layer.3.v_cache 0.00002092 0.06379621 + layer.4.k_cache 0.00074254 1.34996197 + layer.4.v_cache 0.00004925 0.09791123 + layer.4.output 1.21975157 215.88716562 + ------------------------------------------------------------------------------------- + TOTAL 0.53857618 114.85788587 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 45904 +BPFP 0.1681 bits/point +EBPFP 0.3362 equivalent bits/point +MSE 114.857886 +---------------------- -------------------------------------------------------- +Time: 3.615s Load: 0.008s, Pack+Encode: 2.182s, Decode+Unpack: 1.425s +---------------------- -------------------------------------------------------- +💾 Converting with 114.8579 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 195, 128) +Output shape: (1, 195, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.output: torch.Size([1, 195, 3584]) -> torch.Size([1, 1, 195, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,768B, BPFP=0.2218 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,728B, BPFP=0.3788 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,848B, BPFP=0.3885 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,756B, BPFP=0.3010 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,036B, BPFP=0.3234 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,232B, BPFP=0.3391 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,360B, BPFP=0.2692 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,260B, BPFP=0.3413 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,968B, BPFP=0.3981 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,088B, BPFP=0.4077 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,484B, BPFP=0.0284 +⌛️ [2/4] FRONTEND: Frontend time: 1.913s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605874 379.15564904 + layer.0.v_cache 0.00001676 0.04602286 + layer.1.k_cache 0.13872500 24.39611128 + layer.1.v_cache 0.00000593 0.01731659 + layer.2.k_cache 0.01222654 2.87779979 + layer.2.v_cache 0.00001968 0.05245399 + layer.3.k_cache 0.01415822 13.66233348 + layer.3.v_cache 0.00002127 0.06441220 + layer.4.k_cache 0.00071127 1.28494513 + layer.4.v_cache 0.00005419 0.10578883 + layer.4.output 1.56987251 277.84487179 + ------------------------------------------------------------------------------------- + TOTAL 0.66300619 139.21040799 + (elements=1,697,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1697280 +Total Bytes 44528 +BPFP 0.2099 bits/point +EBPFP 0.4198 equivalent bits/point +MSE 139.210408 +---------------------- -------------------------------------------------------- +Time: 3.282s Load: 0.007s, Pack+Encode: 1.913s, Decode+Unpack: 1.362s +---------------------- -------------------------------------------------------- +💾 Converting with 139.2104 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,204B, BPFP=0.2751 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,536B, BPFP=0.3036 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,176B, BPFP=0.3585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,300B, BPFP=0.2833 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,156B, BPFP=0.2709 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,276B, BPFP=0.2812 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,080B, BPFP=0.2644 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,116B, BPFP=0.2675 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,984B, BPFP=0.3420 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,600B, BPFP=0.3091 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,592B, BPFP=0.0318 +⌛️ [2/4] FRONTEND: Frontend time: 2.100s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11176265 409.90921188 + layer.0.v_cache 0.00001802 0.04897469 + layer.1.k_cache 0.18491925 24.53636086 + layer.1.v_cache 0.00000608 0.01833402 + layer.2.k_cache 0.01999094 3.53993678 + layer.2.v_cache 0.00002176 0.05390646 + layer.3.k_cache 0.03910406 14.37885662 + layer.3.v_cache 0.00002004 0.06453953 + layer.4.k_cache 0.00067358 1.38274576 + layer.4.v_cache 0.00005212 0.10647423 + layer.4.output 0.00886795 301.87151688 + ------------------------------------------------------------------------------------- + TOTAL 0.02462613 151.00823288 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 37020 +BPFP 0.1870 bits/point +EBPFP 0.3739 equivalent bits/point +MSE 151.008233 +---------------------- -------------------------------------------------------- +Time: 3.413s Load: 0.006s, Pack+Encode: 2.100s, Decode+Unpack: 1.307s +---------------------- -------------------------------------------------------- +💾 Converting with 151.0082 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,864B, BPFP=0.2835 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,212B, BPFP=0.3823 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,424B, BPFP=0.3979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,724B, BPFP=0.3465 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,080B, BPFP=0.2993 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,448B, BPFP=0.3263 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,848B, BPFP=0.2823 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,848B, BPFP=0.3556 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,480B, BPFP=0.4020 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,900B, BPFP=0.3594 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,368B, BPFP=0.0248 +⌛️ [2/4] FRONTEND: Frontend time: 1.953s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.406s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15089922 398.86465669 + layer.0.v_cache 0.00001686 0.04696478 + layer.1.k_cache 0.31217133 24.78090843 + layer.1.v_cache 0.00000593 0.01746050 + layer.2.k_cache 0.01305961 2.98870076 + layer.2.v_cache 0.00001999 0.05150593 + layer.3.k_cache 0.01178671 13.18757565 + layer.3.v_cache 0.00002039 0.06220214 + layer.4.k_cache 0.00067606 1.29061574 + layer.4.v_cache 0.00005285 0.10288655 + layer.4.output 1.43724915 254.36820926 + ------------------------------------------------------------------------------------- + TOTAL 0.62055606 130.70417306 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 49196 +BPFP 0.2123 bits/point +EBPFP 0.4246 equivalent bits/point +MSE 130.704173 +---------------------- -------------------------------------------------------- +Time: 3.367s Load: 0.008s, Pack+Encode: 1.953s, Decode+Unpack: 1.406s +---------------------- -------------------------------------------------------- +💾 Converting with 130.7042 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 187, 128) +Output shape: (1, 187, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.output: torch.Size([1, 187, 3584]) -> torch.Size([1, 1, 187, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,800B, BPFP=0.2340 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,680B, BPFP=0.3075 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,808B, BPFP=0.3182 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,628B, BPFP=0.3031 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,840B, BPFP=0.2373 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,308B, BPFP=0.2764 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,724B, BPFP=0.2276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,432B, BPFP=0.2868 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,136B, BPFP=0.3456 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,444B, BPFP=0.2878 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,132B, BPFP=0.0254 +⌛️ [2/4] FRONTEND: Frontend time: 1.832s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.292s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08258281 392.13410762 + layer.0.v_cache 0.00001733 0.05097924 + layer.1.k_cache 0.10049684 25.66832909 + layer.1.v_cache 0.00000647 0.01881993 + layer.2.k_cache 0.01472783 3.27942678 + layer.2.v_cache 0.00002140 0.06079746 + layer.3.k_cache 0.02125903 14.26286634 + layer.3.v_cache 0.00002180 0.07358542 + layer.4.k_cache 0.00070831 1.47400784 + layer.4.v_cache 0.00005280 0.11345827 + layer.4.output 0.00866398 293.75747231 + ------------------------------------------------------------------------------------- + TOTAL 0.01650250 146.67286377 + (elements=1,627,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1627648 +Total Bytes 35932 +BPFP 0.1766 bits/point +EBPFP 0.3532 equivalent bits/point +MSE 146.672864 +---------------------- -------------------------------------------------------- +Time: 3.130s Load: 0.006s, Pack+Encode: 1.832s, Decode+Unpack: 1.292s +---------------------- -------------------------------------------------------- +💾 Converting with 146.6729 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,488B, BPFP=0.2583 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,764B, BPFP=0.3528 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,648B, BPFP=0.4182 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,608B, BPFP=0.3412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,960B, BPFP=0.2932 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,216B, BPFP=0.3122 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,004B, BPFP=0.2965 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,480B, BPFP=0.3318 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,460B, BPFP=0.4043 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,336B, BPFP=0.3951 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,372B, BPFP=0.0251 +⌛️ [2/4] FRONTEND: Frontend time: 1.938s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.437s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11071491 394.37229710 + layer.0.v_cache 0.00001804 0.04775774 + layer.1.k_cache 0.26784508 25.08521086 + layer.1.v_cache 0.00000610 0.01786392 + layer.2.k_cache 0.01369454 3.10339703 + layer.2.v_cache 0.00002102 0.05244311 + layer.3.k_cache 0.01444715 13.18645170 + layer.3.v_cache 0.00002041 0.06381629 + layer.4.k_cache 0.00067564 1.31630096 + layer.4.v_cache 0.00005248 0.10534026 + layer.4.output 1.45091435 256.80090555 + ------------------------------------------------------------------------------------- + TOTAL 0.62140564 131.46807164 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 48336 +BPFP 0.2106 bits/point +EBPFP 0.4211 equivalent bits/point +MSE 131.468072 +---------------------- -------------------------------------------------------- +Time: 3.383s Load: 0.007s, Pack+Encode: 1.938s, Decode+Unpack: 1.437s +---------------------- -------------------------------------------------------- +💾 Converting with 131.4681 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,716B, BPFP=0.2604 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,792B, BPFP=0.3358 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,976B, BPFP=0.3487 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,356B, BPFP=0.3052 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,736B, BPFP=0.2618 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,156B, BPFP=0.2912 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,580B, BPFP=0.2508 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,436B, BPFP=0.3108 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,348B, BPFP=0.3747 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,468B, BPFP=0.3131 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,948B, BPFP=0.0195 +⌛️ [2/4] FRONTEND: Frontend time: 1.962s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.435s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13389712 397.56600336 + layer.0.v_cache 0.00001748 0.04599992 + layer.1.k_cache 0.36644789 25.60185153 + layer.1.v_cache 0.00000587 0.01712054 + layer.2.k_cache 0.02260907 2.98383718 + layer.2.v_cache 0.00001979 0.04986402 + layer.3.k_cache 0.02550721 13.06012647 + layer.3.v_cache 0.00001968 0.06176843 + layer.4.k_cache 0.00067529 1.27675721 + layer.4.v_cache 0.00006705 0.10319018 + layer.4.output 1.37284074 242.95389574 + ------------------------------------------------------------------------------------- + TOTAL 0.59759716 125.96728171 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 45512 +BPFP 0.1876 bits/point +EBPFP 0.3752 equivalent bits/point +MSE 125.967282 +---------------------- -------------------------------------------------------- +Time: 3.405s Load: 0.008s, Pack+Encode: 1.962s, Decode+Unpack: 1.435s +---------------------- -------------------------------------------------------- +💾 Converting with 125.9673 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 406, 128) +Output shape: (1, 406, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.output: torch.Size([1, 406, 3584]) -> torch.Size([1, 1, 406, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,596B, BPFP=0.2538 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,096B, BPFP=0.3116 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,672B, BPFP=0.3337 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,844B, BPFP=0.3019 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,416B, BPFP=0.2469 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,416B, BPFP=0.2854 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,000B, BPFP=0.2309 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,424B, BPFP=0.2857 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,536B, BPFP=0.3670 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,796B, BPFP=0.3000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,748B, BPFP=0.0151 +⌛️ [2/4] FRONTEND: Frontend time: 2.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.929s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10474718 399.97806342 + layer.0.v_cache 0.00001775 0.04670557 + layer.1.k_cache 0.89873824 24.64295336 + layer.1.v_cache 0.00000687 0.01813936 + layer.2.k_cache 0.02602503 2.76954869 + layer.2.v_cache 0.00002164 0.05507266 + layer.3.k_cache 0.01558245 12.32570524 + layer.3.v_cache 0.00002201 0.06309280 + layer.4.k_cache 0.00072958 1.33271707 + layer.4.v_cache 0.00005327 0.10044863 + layer.4.output 0.00651461 136.63225721 + ------------------------------------------------------------------------------------- + TOTAL 0.06420861 82.22107337 + (elements=3,533,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3533824 +Total Bytes 78544 +BPFP 0.1778 bits/point +EBPFP 0.3556 equivalent bits/point +MSE 82.221073 +---------------------- -------------------------------------------------------- +Time: 4.790s Load: 0.014s, Pack+Encode: 2.847s, Decode+Unpack: 1.929s +---------------------- -------------------------------------------------------- +💾 Converting with 82.2211 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,884B, BPFP=0.2477 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,576B, BPFP=0.2918 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,392B, BPFP=0.3439 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,560B, BPFP=0.2908 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,944B, BPFP=0.2515 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,320B, BPFP=0.2755 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,808B, BPFP=0.2429 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,412B, BPFP=0.2814 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,408B, BPFP=0.3449 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,664B, BPFP=0.2974 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,504B, BPFP=0.0228 +⌛️ [2/4] FRONTEND: Frontend time: 2.337s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.635s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14414167 400.59445153 + layer.0.v_cache 0.00001717 0.04782522 + layer.1.k_cache 0.44179790 24.96643814 + layer.1.v_cache 0.00000667 0.01883616 + layer.2.k_cache 0.03016641 3.30398572 + layer.2.v_cache 0.00002043 0.05392686 + layer.3.k_cache 0.03726396 13.32026467 + layer.3.v_cache 0.00002099 0.06582491 + layer.4.k_cache 0.00070653 1.32370107 + layer.4.v_cache 0.00004910 0.10333060 + layer.4.output 1.24963187 221.15049198 + ------------------------------------------------------------------------------------- + TOTAL 0.55303611 117.16776640 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 47472 +BPFP 0.1781 bits/point +EBPFP 0.3562 equivalent bits/point +MSE 117.167766 +---------------------- -------------------------------------------------------- +Time: 3.981s Load: 0.009s, Pack+Encode: 2.337s, Decode+Unpack: 1.635s +---------------------- -------------------------------------------------------- +💾 Converting with 117.1678 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,124B, BPFP=0.2453 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,800B, BPFP=0.3769 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,008B, BPFP=0.3932 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,396B, BPFP=0.3452 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,788B, BPFP=0.2974 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,884B, BPFP=0.3050 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,424B, BPFP=0.2688 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,344B, BPFP=0.3411 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,448B, BPFP=0.4278 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,860B, BPFP=0.3816 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,216B, BPFP=0.0249 +⌛️ [2/4] FRONTEND: Frontend time: 2.070s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.506s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15985105 384.18742148 + layer.0.v_cache 0.00001905 0.04790436 + layer.1.k_cache 0.22745673 25.72894747 + layer.1.v_cache 0.00000590 0.01757830 + layer.2.k_cache 0.01549998 2.81046130 + layer.2.v_cache 0.00002003 0.05273667 + layer.3.k_cache 0.01333464 13.91861407 + layer.3.v_cache 0.00002034 0.06392055 + layer.4.k_cache 0.00067076 1.29548223 + layer.4.v_cache 0.00004949 0.10182247 + layer.4.output 1.53832061 272.26637653 + ------------------------------------------------------------------------------------- + TOTAL 0.65795131 137.29938380 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 45292 +BPFP 0.2092 bits/point +EBPFP 0.4184 equivalent bits/point +MSE 137.299384 +---------------------- -------------------------------------------------------- +Time: 3.584s Load: 0.008s, Pack+Encode: 2.070s, Decode+Unpack: 1.506s +---------------------- -------------------------------------------------------- +💾 Converting with 137.2994 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 237, 128) +Output shape: (1, 237, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.output: torch.Size([1, 237, 3584]) -> torch.Size([1, 1, 237, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,868B, BPFP=0.2550 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,572B, BPFP=0.3014 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,492B, BPFP=0.3621 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,196B, BPFP=0.2766 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,196B, BPFP=0.2766 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,136B, BPFP=0.2727 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,920B, BPFP=0.2584 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,248B, BPFP=0.2801 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,596B, BPFP=0.3689 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,636B, BPFP=0.3056 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,268B, BPFP=0.0214 +⌛️ [2/4] FRONTEND: Frontend time: 1.990s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.480s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351868 394.18532437 + layer.0.v_cache 0.00001655 0.04554621 + layer.1.k_cache 0.34375901 26.45242863 + layer.1.v_cache 0.00000619 0.01786495 + layer.2.k_cache 0.01282014 3.11012802 + layer.2.v_cache 0.00002104 0.05237990 + layer.3.k_cache 0.03241006 13.37098456 + layer.3.v_cache 0.00001985 0.06025155 + layer.4.k_cache 0.00067667 1.35308748 + layer.4.v_cache 0.00005279 0.09957388 + layer.4.output 1.29178675 228.63187538 + ------------------------------------------------------------------------------------- + TOTAL 0.56210637 119.95121748 + (elements=2,062,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2062848 +Total Bytes 47128 +BPFP 0.1828 bits/point +EBPFP 0.3655 equivalent bits/point +MSE 119.951217 +---------------------- -------------------------------------------------------- +Time: 3.479s Load: 0.009s, Pack+Encode: 1.990s, Decode+Unpack: 1.480s +---------------------- -------------------------------------------------------- +💾 Converting with 119.9512 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,576B, BPFP=0.2599 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,680B, BPFP=0.3401 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,268B, BPFP=0.3828 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,480B, BPFP=0.3256 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,988B, BPFP=0.2898 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,444B, BPFP=0.3230 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,776B, BPFP=0.2744 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,344B, BPFP=0.3157 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,324B, BPFP=0.3869 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,952B, BPFP=0.3599 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,092B, BPFP=0.0217 +⌛️ [2/4] FRONTEND: Frontend time: 2.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.542s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11023432 395.24280523 + layer.0.v_cache 0.00001677 0.04531197 + layer.1.k_cache 0.30291741 24.85242097 + layer.1.v_cache 0.00000597 0.01711188 + layer.2.k_cache 0.03464115 2.84850166 + layer.2.v_cache 0.00001998 0.05214512 + layer.3.k_cache 0.01330193 13.75605923 + layer.3.v_cache 0.00002112 0.05965297 + layer.4.k_cache 0.00069379 1.28627163 + layer.4.v_cache 0.00005021 0.09796188 + layer.4.output 1.42388847 252.00315615 + ------------------------------------------------------------------------------------- + TOTAL 0.61347776 129.54590209 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 46924 +BPFP 0.2006 bits/point +EBPFP 0.4012 equivalent bits/point +MSE 129.545902 +---------------------- -------------------------------------------------------- +Time: 3.779s Load: 0.008s, Pack+Encode: 2.230s, Decode+Unpack: 1.542s +---------------------- -------------------------------------------------------- +💾 Converting with 129.5459 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 231, 128) +Output shape: (1, 231, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.output: torch.Size([1, 231, 3584]) -> torch.Size([1, 1, 231, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,588B, BPFP=0.2427 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,672B, BPFP=0.3160 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,420B, BPFP=0.3666 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,356B, BPFP=0.2946 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,156B, BPFP=0.2811 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,176B, BPFP=0.2825 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,004B, BPFP=0.2708 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,228B, BPFP=0.2860 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,712B, BPFP=0.3864 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,840B, BPFP=0.3274 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,228B, BPFP=0.0215 +⌛️ [2/4] FRONTEND: Frontend time: 2.135s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.617s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13257201 388.04924242 + layer.0.v_cache 0.00001701 0.04801471 + layer.1.k_cache 0.40462213 25.75882711 + layer.1.v_cache 0.00000658 0.01886607 + layer.2.k_cache 0.02649644 2.93811537 + layer.2.v_cache 0.00002056 0.05401046 + layer.3.k_cache 0.06147716 13.02067903 + layer.3.v_cache 0.00002105 0.06691625 + layer.4.k_cache 0.00069437 1.32549837 + layer.4.v_cache 0.00005136 0.10585865 + layer.4.output 1.32536713 234.58916976 + ------------------------------------------------------------------------------------- + TOTAL 0.58256168 121.97118922 + (elements=2,010,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2010624 +Total Bytes 47380 +BPFP 0.1885 bits/point +EBPFP 0.3770 equivalent bits/point +MSE 121.971189 +---------------------- -------------------------------------------------------- +Time: 3.761s Load: 0.008s, Pack+Encode: 2.135s, Decode+Unpack: 1.617s +---------------------- -------------------------------------------------------- +💾 Converting with 121.9712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,180B, BPFP=0.2561 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,912B, BPFP=0.3010 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,368B, BPFP=0.2676 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,304B, BPFP=0.2637 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,148B, BPFP=0.1929 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,964B, BPFP=0.2429 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,020B, BPFP=0.1850 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,056B, BPFP=0.2485 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,616B, BPFP=0.2828 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,188B, BPFP=0.2566 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,976B, BPFP=0.0173 +⌛️ [2/4] FRONTEND: Frontend time: 2.348s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.501s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12841698 409.70435049 + layer.0.v_cache 0.00001718 0.04781984 + layer.1.k_cache 0.45450200 23.43863358 + layer.1.v_cache 0.00000627 0.01833296 + layer.2.k_cache 0.01434461 2.70752528 + layer.2.v_cache 0.00002215 0.05612906 + layer.3.k_cache 0.03530993 12.59065755 + layer.3.v_cache 0.00002139 0.06544735 + layer.4.k_cache 0.00068344 1.28671061 + layer.4.v_cache 0.00005073 0.10314968 + layer.4.output 1.20064817 212.48067227 + ------------------------------------------------------------------------------------- + TOTAL 0.53164187 113.96373307 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 42732 +BPFP 0.1540 bits/point +EBPFP 0.3080 equivalent bits/point +MSE 113.963733 +---------------------- -------------------------------------------------------- +Time: 3.860s Load: 0.011s, Pack+Encode: 2.348s, Decode+Unpack: 1.501s +---------------------- -------------------------------------------------------- +💾 Converting with 113.9637 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,784B, BPFP=0.2651 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,120B, BPFP=0.3391 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,156B, BPFP=0.3411 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,664B, BPFP=0.3138 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,764B, BPFP=0.2640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,400B, BPFP=0.2992 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,612B, BPFP=0.2555 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,408B, BPFP=0.2996 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,876B, BPFP=0.3810 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,916B, BPFP=0.3278 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,312B, BPFP=0.0183 +⌛️ [2/4] FRONTEND: Frontend time: 2.127s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.613s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13731835 407.42012965 + layer.0.v_cache 0.00001709 0.04587076 + layer.1.k_cache 0.55879569 24.32709822 + layer.1.v_cache 0.00000637 0.01757775 + layer.2.k_cache 0.01348169 2.93362816 + layer.2.v_cache 0.00002375 0.05264430 + layer.3.k_cache 0.01306405 14.36776062 + layer.3.v_cache 0.00002005 0.06084100 + layer.4.k_cache 0.00072533 1.32160225 + layer.4.v_cache 0.00005384 0.10401891 + layer.4.output 0.00472113 196.93406472 + ------------------------------------------------------------------------------------- + TOTAL 0.04450318 107.59938969 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 58012 +BPFP 0.1891 bits/point +EBPFP 0.3782 equivalent bits/point +MSE 107.599390 +---------------------- -------------------------------------------------------- +Time: 3.750s Load: 0.010s, Pack+Encode: 2.127s, Decode+Unpack: 1.613s +---------------------- -------------------------------------------------------- +💾 Converting with 107.5994 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,848B, BPFP=0.2517 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,992B, BPFP=0.3110 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,396B, BPFP=0.3320 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,856B, BPFP=0.3040 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,840B, BPFP=0.2512 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,448B, BPFP=0.3347 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,676B, BPFP=0.2427 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,716B, BPFP=0.2967 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,508B, BPFP=0.3378 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,960B, BPFP=0.3094 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,568B, BPFP=0.0190 +⌛️ [2/4] FRONTEND: Frontend time: 2.056s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.789s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13972237 398.53457226 + layer.0.v_cache 0.00001689 0.04681191 + layer.1.k_cache 0.51719427 24.33065725 + layer.1.v_cache 0.00000613 0.01803157 + layer.2.k_cache 0.02019965 2.95057207 + layer.2.v_cache 0.00002053 0.05563510 + layer.3.k_cache 0.01759115 13.02269291 + layer.3.v_cache 0.00002125 0.06410101 + layer.4.k_cache 0.00069604 1.28561229 + layer.4.v_cache 0.00005872 0.10133947 + layer.4.output 0.04434862 180.27301851 + ------------------------------------------------------------------------------------- + TOTAL 0.05917455 100.13653855 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 59808 +BPFP 0.1826 bits/point +EBPFP 0.3653 equivalent bits/point +MSE 100.136539 +---------------------- -------------------------------------------------------- +Time: 3.855s Load: 0.010s, Pack+Encode: 2.056s, Decode+Unpack: 1.789s +---------------------- -------------------------------------------------------- +💾 Converting with 100.1365 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,144B, BPFP=0.2469 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,592B, BPFP=0.3606 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,068B, BPFP=0.3979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,264B, BPFP=0.3348 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,912B, BPFP=0.3072 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,608B, BPFP=0.3618 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,492B, BPFP=0.2742 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,648B, BPFP=0.3649 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,060B, BPFP=0.3973 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,060B, BPFP=0.3973 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,228B, BPFP=0.0250 +⌛️ [2/4] FRONTEND: Frontend time: 1.959s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.613s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13044639 385.86718750 + layer.0.v_cache 0.00001737 0.04649711 + layer.1.k_cache 0.22565776 25.47592454 + layer.1.v_cache 0.00000620 0.01782293 + layer.2.k_cache 0.01998205 2.88474477 + layer.2.v_cache 0.00001989 0.05079226 + layer.3.k_cache 0.00629339 13.37436573 + layer.3.v_cache 0.00002050 0.06250272 + layer.4.k_cache 0.00068118 1.26825028 + layer.4.v_cache 0.00005692 0.10112545 + layer.4.output 1.53833376 272.25634871 + ------------------------------------------------------------------------------------- + TOTAL 0.65597165 137.34962672 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 46076 +BPFP 0.2128 bits/point +EBPFP 0.4256 equivalent bits/point +MSE 137.349627 +---------------------- -------------------------------------------------------- +Time: 3.579s Load: 0.007s, Pack+Encode: 1.959s, Decode+Unpack: 1.613s +---------------------- -------------------------------------------------------- +💾 Converting with 137.3496 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 246, 128) +Output shape: (1, 246, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.output: torch.Size([1, 246, 3584]) -> torch.Size([1, 1, 246, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,000B, BPFP=0.2541 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,116B, BPFP=0.3249 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,360B, BPFP=0.3404 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,484B, BPFP=0.2848 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,816B, BPFP=0.2424 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,508B, BPFP=0.2863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,804B, BPFP=0.2416 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,524B, BPFP=0.2873 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,604B, BPFP=0.3559 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,512B, BPFP=0.2866 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,604B, BPFP=0.0236 +⌛️ [2/4] FRONTEND: Frontend time: 1.916s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.410s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131485 407.71220783 + layer.0.v_cache 0.00001613 0.04587264 + layer.1.k_cache 0.49701033 25.60942462 + layer.1.v_cache 0.00000620 0.01744699 + layer.2.k_cache 0.01298508 2.95459332 + layer.2.v_cache 0.00002120 0.05114411 + layer.3.k_cache 0.02762915 13.17642534 + layer.3.v_cache 0.00002136 0.06242688 + layer.4.k_cache 0.00068392 1.32321762 + layer.4.v_cache 0.00006372 0.09910592 + layer.4.output 1.24453647 220.26061629 + ------------------------------------------------------------------------------------- + TOTAL 0.55302984 117.22801055 + (elements=2,141,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2141184 +Total Bytes 48332 +BPFP 0.1806 bits/point +EBPFP 0.3612 equivalent bits/point +MSE 117.228011 +---------------------- -------------------------------------------------------- +Time: 3.334s Load: 0.008s, Pack+Encode: 1.916s, Decode+Unpack: 1.410s +---------------------- -------------------------------------------------------- +💾 Converting with 117.2280 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,372B, BPFP=0.2634 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,004B, BPFP=0.3909 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,100B, BPFP=0.3984 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,544B, BPFP=0.3550 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,840B, BPFP=0.3000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,336B, BPFP=0.3387 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,364B, BPFP=0.2628 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,488B, BPFP=0.3506 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,192B, BPFP=0.4056 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,024B, BPFP=0.3925 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,208B, BPFP=0.0246 +⌛️ [2/4] FRONTEND: Frontend time: 2.117s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.429s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11599884 398.16242188 + layer.0.v_cache 0.00001671 0.04325551 + layer.1.k_cache 0.22559120 25.27189697 + layer.1.v_cache 0.00000595 0.01683764 + layer.2.k_cache 0.01146694 3.00999268 + layer.2.v_cache 0.00002201 0.05145952 + layer.3.k_cache 0.03668072 13.61911865 + layer.3.v_cache 0.00002131 0.05826447 + layer.4.k_cache 0.00066762 1.25771667 + layer.4.v_cache 0.00005146 0.10123376 + layer.4.output 1.53061454 270.87997768 + ------------------------------------------------------------------------------------- + TOTAL 0.65322497 137.51482597 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 46472 +BPFP 0.2136 bits/point +EBPFP 0.4271 equivalent bits/point +MSE 137.514826 +---------------------- -------------------------------------------------------- +Time: 3.556s Load: 0.009s, Pack+Encode: 2.117s, Decode+Unpack: 1.429s +---------------------- -------------------------------------------------------- +💾 Converting with 137.5148 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,472B, BPFP=0.2369 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,728B, BPFP=0.3226 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,272B, BPFP=0.3597 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,560B, BPFP=0.3111 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,208B, BPFP=0.2871 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,308B, BPFP=0.2939 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,916B, BPFP=0.2672 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,744B, BPFP=0.3237 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,908B, BPFP=0.4031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,740B, BPFP=0.3234 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,196B, BPFP=0.0214 +⌛️ [2/4] FRONTEND: Frontend time: 2.064s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605523 391.75337746 + layer.0.v_cache 0.00001866 0.04717150 + layer.1.k_cache 0.31881074 26.27265071 + layer.1.v_cache 0.00000653 0.01834630 + layer.2.k_cache 0.02091764 2.98706161 + layer.2.v_cache 0.00002113 0.05452338 + layer.3.k_cache 0.03221333 12.74679739 + layer.3.v_cache 0.00002185 0.06403970 + layer.4.k_cache 0.00068277 1.34404212 + layer.4.v_cache 0.00005596 0.10457628 + layer.4.output 1.33695214 236.63363615 + ------------------------------------------------------------------------------------- + TOTAL 0.57926287 123.04870821 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 48052 +BPFP 0.1929 bits/point +EBPFP 0.3857 equivalent bits/point +MSE 123.048708 +---------------------- -------------------------------------------------------- +Time: 3.543s Load: 0.009s, Pack+Encode: 2.064s, Decode+Unpack: 1.470s +---------------------- -------------------------------------------------------- +💾 Converting with 123.0487 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 192, 128) +Output shape: (1, 192, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.output: torch.Size([1, 192, 3584]) -> torch.Size([1, 1, 192, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,696B, BPFP=0.2194 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,528B, BPFP=0.2871 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,196B, BPFP=0.2601 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,016B, BPFP=0.2454 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,180B, BPFP=0.1774 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,852B, BPFP=0.2321 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,044B, BPFP=0.1663 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,832B, BPFP=0.2305 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,116B, BPFP=0.2536 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,992B, BPFP=0.2435 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,004B, BPFP=0.0117 +⌛️ [2/4] FRONTEND: Frontend time: 1.861s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.288s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12444629 403.15295410 + layer.0.v_cache 0.00001540 0.04176321 + layer.1.k_cache 0.18411521 23.48416138 + layer.1.v_cache 0.00000585 0.01703002 + layer.2.k_cache 0.01124177 2.63625463 + layer.2.v_cache 0.00002179 0.05093457 + layer.3.k_cache 0.01616090 12.38894145 + layer.3.v_cache 0.00002117 0.06110372 + layer.4.k_cache 0.00068507 1.26691898 + layer.4.v_cache 0.00005275 0.09445309 + layer.4.output 0.00839530 286.12041946 + ------------------------------------------------------------------------------------- + TOTAL 0.02326666 143.88455596 + (elements=1,671,168) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1671168 +Total Bytes 29456 +BPFP 0.1410 bits/point +EBPFP 0.2820 equivalent bits/point +MSE 143.884556 +---------------------- -------------------------------------------------------- +Time: 3.155s Load: 0.006s, Pack+Encode: 1.861s, Decode+Unpack: 1.288s +---------------------- -------------------------------------------------------- +💾 Converting with 143.8846 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,476B, BPFP=0.2331 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,560B, BPFP=0.3351 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,196B, BPFP=0.3950 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,272B, BPFP=0.3080 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,356B, BPFP=0.3159 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,384B, BPFP=0.3185 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,020B, BPFP=0.2843 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,348B, BPFP=0.3151 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,324B, BPFP=0.4070 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,620B, BPFP=0.3407 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,188B, BPFP=0.0294 +⌛️ [2/4] FRONTEND: Frontend time: 1.776s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.332s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09999813 385.88248306 + layer.0.v_cache 0.00001899 0.04392323 + layer.1.k_cache 0.12989934 25.03233834 + layer.1.v_cache 0.00000610 0.01687397 + layer.2.k_cache 0.00966061 2.96084668 + layer.2.v_cache 0.00001994 0.05023840 + layer.3.k_cache 0.03384685 13.86240617 + layer.3.v_cache 0.00002040 0.05889249 + layer.4.k_cache 0.00066502 1.24272818 + layer.4.v_cache 0.00005200 0.10162745 + layer.4.output 0.00959846 330.95188791 + ------------------------------------------------------------------------------------- + TOTAL 0.02008098 161.52444549 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 36744 +BPFP 0.2034 bits/point +EBPFP 0.4069 equivalent bits/point +MSE 161.524445 +---------------------- -------------------------------------------------------- +Time: 3.114s Load: 0.006s, Pack+Encode: 1.776s, Decode+Unpack: 1.332s +---------------------- -------------------------------------------------------- +💾 Converting with 161.5244 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 428, 128) +Output shape: (1, 428, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.output: torch.Size([1, 428, 3584]) -> torch.Size([1, 1, 428, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,916B, BPFP=0.2525 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,032B, BPFP=0.2932 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,852B, BPFP=0.3232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,520B, BPFP=0.2745 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,684B, BPFP=0.2440 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,832B, BPFP=0.2494 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,116B, BPFP=0.2233 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,436B, BPFP=0.2715 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,248B, BPFP=0.3376 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,752B, BPFP=0.2830 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,980B, BPFP=0.0155 +⌛️ [2/4] FRONTEND: Frontend time: 2.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.930s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13760449 403.81118575 + layer.0.v_cache 0.00001624 0.04564187 + layer.1.k_cache 0.89392254 25.39384674 + layer.1.v_cache 0.00000615 0.01706546 + layer.2.k_cache 0.03917003 2.85671598 + layer.2.v_cache 0.00002113 0.05216139 + layer.3.k_cache 0.03749864 13.11966313 + layer.3.v_cache 0.00002059 0.06008589 + layer.4.k_cache 0.00086919 1.32357560 + layer.4.v_cache 0.00005177 0.09746346 + layer.4.output 0.00617707 129.60696554 + ------------------------------------------------------------------------------------- + TOTAL 0.06778943 79.64859788 + (elements=3,725,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3725312 +Total Bytes 78368 +BPFP 0.1683 bits/point +EBPFP 0.3366 equivalent bits/point +MSE 79.648598 +---------------------- -------------------------------------------------------- +Time: 4.784s Load: 0.013s, Pack+Encode: 2.840s, Decode+Unpack: 1.930s +---------------------- -------------------------------------------------------- +💾 Converting with 79.6486 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,640B, BPFP=0.2571 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,352B, BPFP=0.3520 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,040B, BPFP=0.3347 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,824B, BPFP=0.3227 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,604B, BPFP=0.2551 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,796B, BPFP=0.3211 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,488B, BPFP=0.2487 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,400B, BPFP=0.2992 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,868B, BPFP=0.3805 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,932B, BPFP=0.3287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,328B, BPFP=0.0184 +⌛️ [2/4] FRONTEND: Frontend time: 2.106s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.566s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385377 401.83203125 + layer.0.v_cache 0.00001705 0.04635638 + layer.1.k_cache 0.49372572 24.47741439 + layer.1.v_cache 0.00000626 0.01779869 + layer.2.k_cache 0.02422354 2.72272010 + layer.2.v_cache 0.00002046 0.05370351 + layer.3.k_cache 0.04153839 13.55374210 + layer.3.v_cache 0.00002075 0.06240984 + layer.4.k_cache 0.00074884 1.33308032 + layer.4.v_cache 0.00004954 0.09778855 + layer.4.output 0.00474379 196.95010132 + ------------------------------------------------------------------------------------- + TOTAL 0.04161240 107.22633849 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 58272 +BPFP 0.1899 bits/point +EBPFP 0.3798 equivalent bits/point +MSE 107.226338 +---------------------- -------------------------------------------------------- +Time: 3.680s Load: 0.009s, Pack+Encode: 2.106s, Decode+Unpack: 1.566s +---------------------- -------------------------------------------------------- +💾 Converting with 107.2263 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,544B, BPFP=0.2198 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,868B, BPFP=0.3322 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,316B, BPFP=0.3539 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,132B, BPFP=0.2966 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,340B, BPFP=0.2583 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,172B, BPFP=0.2986 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,088B, BPFP=0.2461 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,104B, BPFP=0.2953 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,964B, BPFP=0.3853 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,768B, BPFP=0.3274 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,772B, BPFP=0.0192 +⌛️ [2/4] FRONTEND: Frontend time: 2.499s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.655s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12985682 382.09573336 + layer.0.v_cache 0.00001645 0.04562948 + layer.1.k_cache 0.61935935 24.78287055 + layer.1.v_cache 0.00000656 0.01734560 + layer.2.k_cache 0.01492825 2.84757802 + layer.2.v_cache 0.00002178 0.05410946 + layer.3.k_cache 0.02579401 13.35017626 + layer.3.v_cache 0.00002068 0.06019486 + layer.4.k_cache 0.00071613 1.32912972 + layer.4.v_cache 0.00005057 0.09843871 + layer.4.output 0.04141478 168.03308824 + ------------------------------------------------------------------------------------- + TOTAL 0.06356906 94.17134257 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 65068 +BPFP 0.1852 bits/point +EBPFP 0.3703 equivalent bits/point +MSE 94.171343 +---------------------- -------------------------------------------------------- +Time: 4.165s Load: 0.011s, Pack+Encode: 2.499s, Decode+Unpack: 1.655s +---------------------- -------------------------------------------------------- +💾 Converting with 94.1713 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,772B, BPFP=0.2607 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,340B, BPFP=0.2917 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,792B, BPFP=0.3164 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,196B, BPFP=0.2839 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,572B, BPFP=0.2498 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,064B, BPFP=0.2767 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,132B, BPFP=0.2257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,712B, BPFP=0.3121 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,384B, BPFP=0.3488 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,256B, BPFP=0.2872 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,140B, BPFP=0.0167 +⌛️ [2/4] FRONTEND: Frontend time: 2.106s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.626s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12359551 400.04381556 + layer.0.v_cache 0.00001766 0.04735913 + layer.1.k_cache 0.52692803 24.73744810 + layer.1.v_cache 0.00000637 0.01825967 + layer.2.k_cache 0.02102186 2.88429570 + layer.2.v_cache 0.00002187 0.05355461 + layer.3.k_cache 0.03868174 13.03608586 + layer.3.v_cache 0.00002097 0.06581981 + layer.4.k_cache 0.00070799 1.32496504 + layer.4.v_cache 0.00005204 0.10152594 + layer.4.output 0.00469152 194.19246379 + ------------------------------------------------------------------------------------- + TOTAL 0.04375851 105.98002211 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 54360 +BPFP 0.1747 bits/point +EBPFP 0.3494 equivalent bits/point +MSE 105.980022 +---------------------- -------------------------------------------------------- +Time: 3.741s Load: 0.010s, Pack+Encode: 2.106s, Decode+Unpack: 1.626s +---------------------- -------------------------------------------------------- +💾 Converting with 105.9800 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,904B, BPFP=0.2280 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,740B, BPFP=0.3722 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,264B, BPFP=0.4133 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,820B, BPFP=0.3785 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,976B, BPFP=0.3122 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,264B, BPFP=0.3348 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,500B, BPFP=0.2748 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,388B, BPFP=0.3445 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,304B, BPFP=0.4165 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,196B, BPFP=0.4080 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,200B, BPFP=0.0247 +⌛️ [2/4] FRONTEND: Frontend time: 1.935s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10671347 381.23280465 + layer.0.v_cache 0.00001788 0.04812820 + layer.1.k_cache 0.15257932 25.72669009 + layer.1.v_cache 0.00000642 0.01829479 + layer.2.k_cache 0.02197224 3.14736463 + layer.2.v_cache 0.00002094 0.05311781 + layer.3.k_cache 0.03325213 13.94893290 + layer.3.v_cache 0.00002126 0.06613267 + layer.4.k_cache 0.00067229 1.32655986 + layer.4.v_cache 0.00005023 0.10256853 + layer.4.output 1.53834953 272.26568108 + ------------------------------------------------------------------------------------- + TOTAL 0.65198546 137.14884481 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 46556 +BPFP 0.2150 bits/point +EBPFP 0.4301 equivalent bits/point +MSE 137.148845 +---------------------- -------------------------------------------------------- +Time: 3.256s Load: 0.007s, Pack+Encode: 1.935s, Decode+Unpack: 1.314s +---------------------- -------------------------------------------------------- +💾 Converting with 137.1488 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,672B, BPFP=0.2286 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,636B, BPFP=0.2886 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,864B, BPFP=0.3028 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,624B, BPFP=0.2878 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,628B, BPFP=0.2258 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,120B, BPFP=0.2565 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,420B, BPFP=0.2129 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,828B, BPFP=0.3005 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,616B, BPFP=0.3496 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,596B, BPFP=0.2861 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,456B, BPFP=0.0218 +⌛️ [2/4] FRONTEND: Frontend time: 2.301s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.404s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15214830 396.88439990 + layer.0.v_cache 0.00001722 0.04835789 + layer.1.k_cache 0.49754258 24.33364394 + layer.1.v_cache 0.00000639 0.01961583 + layer.2.k_cache 0.01862443 3.12446017 + layer.2.v_cache 0.00002219 0.05620964 + layer.3.k_cache 0.01200796 13.64827798 + layer.3.v_cache 0.00002210 0.06815881 + layer.4.k_cache 0.00075226 1.41000257 + layer.4.v_cache 0.00005409 0.10721201 + layer.4.output 1.21979868 215.89668113 + ------------------------------------------------------------------------------------- + TOTAL 0.54234049 114.76335922 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 46460 +BPFP 0.1701 bits/point +EBPFP 0.3403 equivalent bits/point +MSE 114.763359 +---------------------- -------------------------------------------------------- +Time: 3.715s Load: 0.010s, Pack+Encode: 2.301s, Decode+Unpack: 1.404s +---------------------- -------------------------------------------------------- +💾 Converting with 114.7634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,928B, BPFP=0.2525 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,228B, BPFP=0.3191 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,384B, BPFP=0.3270 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,728B, BPFP=0.2934 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,952B, BPFP=0.2537 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,088B, BPFP=0.2607 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,612B, BPFP=0.2363 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,436B, BPFP=0.2785 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,732B, BPFP=0.3449 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,856B, BPFP=0.3000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,240B, BPFP=0.0164 +⌛️ [2/4] FRONTEND: Frontend time: 2.057s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.617s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18432382 400.13217213 + layer.0.v_cache 0.00001629 0.04589252 + layer.1.k_cache 0.53789903 24.25685515 + layer.1.v_cache 0.00000607 0.01794694 + layer.2.k_cache 0.01514517 3.05865559 + layer.2.v_cache 0.00002151 0.05245999 + layer.3.k_cache 0.04182221 12.79687100 + layer.3.v_cache 0.00002017 0.06322288 + layer.4.k_cache 0.00068627 1.33274376 + layer.4.v_cache 0.00005063 0.10154622 + layer.4.output 0.04375917 177.93212822 + ------------------------------------------------------------------------------------- + TOTAL 0.06390032 99.25783904 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 58184 +BPFP 0.1753 bits/point +EBPFP 0.3507 equivalent bits/point +MSE 99.257839 +---------------------- -------------------------------------------------------- +Time: 3.685s Load: 0.011s, Pack+Encode: 2.057s, Decode+Unpack: 1.617s +---------------------- -------------------------------------------------------- +💾 Converting with 99.2578 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 371, 128) +Output shape: (1, 371, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.output: torch.Size([1, 371, 3584]) -> torch.Size([1, 1, 371, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,984B, BPFP=0.2520 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,528B, BPFP=0.3170 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,512B, BPFP=0.3164 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,744B, BPFP=0.2840 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,680B, BPFP=0.2392 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,180B, BPFP=0.2603 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,360B, BPFP=0.2257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,300B, BPFP=0.3074 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,292B, BPFP=0.3492 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,772B, BPFP=0.2852 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,588B, BPFP=0.0156 +⌛️ [2/4] FRONTEND: Frontend time: 2.185s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.674s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16952868 404.88872978 + layer.0.v_cache 0.00001698 0.04775239 + layer.1.k_cache 0.82091874 24.56119177 + layer.1.v_cache 0.00000666 0.01831844 + layer.2.k_cache 0.02418445 2.97261775 + layer.2.v_cache 0.00002238 0.05472397 + layer.3.k_cache 0.03055777 12.86599509 + layer.3.v_cache 0.00002117 0.06276455 + layer.4.k_cache 0.00071103 1.36488346 + layer.4.v_cache 0.00005915 0.10181976 + layer.4.output 0.03610923 146.30789132 + ------------------------------------------------------------------------------------- + TOTAL 0.07639951 86.53494331 + (elements=3,229,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3229184 +Total Bytes 69940 +BPFP 0.1733 bits/point +EBPFP 0.3465 equivalent bits/point +MSE 86.534943 +---------------------- -------------------------------------------------------- +Time: 3.872s Load: 0.013s, Pack+Encode: 2.185s, Decode+Unpack: 1.674s +---------------------- -------------------------------------------------------- +💾 Converting with 86.5349 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,628B, BPFP=0.2600 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,880B, BPFP=0.3498 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,180B, BPFP=0.3713 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,624B, BPFP=0.3314 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,016B, BPFP=0.2878 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,480B, BPFP=0.3211 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,900B, BPFP=0.2795 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,564B, BPFP=0.3271 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,304B, BPFP=0.3802 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,624B, BPFP=0.3314 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,168B, BPFP=0.0222 +⌛️ [2/4] FRONTEND: Frontend time: 2.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.505s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13674593 390.08246130 + layer.0.v_cache 0.00001768 0.04916037 + layer.1.k_cache 0.26206151 26.09684095 + layer.1.v_cache 0.00000624 0.01837844 + layer.2.k_cache 0.02938290 2.92311866 + layer.2.v_cache 0.00002158 0.05424887 + layer.3.k_cache 0.05653211 13.37990633 + layer.3.v_cache 0.00002129 0.06662831 + layer.4.k_cache 0.00070776 1.34185987 + layer.4.v_cache 0.00005193 0.10485443 + layer.4.output 1.40434551 248.54916858 + ------------------------------------------------------------------------------------- + TOTAL 0.60682162 127.88009633 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 47368 +BPFP 0.1997 bits/point +EBPFP 0.3994 equivalent bits/point +MSE 127.880096 +---------------------- -------------------------------------------------------- +Time: 3.742s Load: 0.008s, Pack+Encode: 2.228s, Decode+Unpack: 1.505s +---------------------- -------------------------------------------------------- +💾 Converting with 127.8801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 193, 128) +Output shape: (1, 193, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.output: torch.Size([1, 193, 3584]) -> torch.Size([1, 1, 193, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,892B, BPFP=0.2341 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,344B, BPFP=0.3517 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,816B, BPFP=0.3899 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,868B, BPFP=0.3131 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,536B, BPFP=0.2863 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,460B, BPFP=0.2801 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,164B, BPFP=0.2562 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,948B, BPFP=0.3196 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,056B, BPFP=0.4093 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,044B, BPFP=0.4084 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,416B, BPFP=0.0279 +⌛️ [2/4] FRONTEND: Frontend time: 2.215s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.708s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14305212 386.52396373 + layer.0.v_cache 0.00001770 0.04793048 + layer.1.k_cache 0.14179654 25.85032029 + layer.1.v_cache 0.00000608 0.01837809 + layer.2.k_cache 0.01555968 3.08333607 + layer.2.v_cache 0.00002047 0.05599653 + layer.3.k_cache 0.02655400 13.84231172 + layer.3.v_cache 0.00002031 0.06642462 + layer.4.k_cache 0.00069483 1.36781287 + layer.4.v_cache 0.00005004 0.10778319 + layer.4.output 1.58615237 280.63536269 + ------------------------------------------------------------------------------------- + TOTAL 0.67240225 140.90657626 + (elements=1,679,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1679872 +Total Bytes 42544 +BPFP 0.2026 bits/point +EBPFP 0.4052 equivalent bits/point +MSE 140.906576 +---------------------- -------------------------------------------------------- +Time: 3.932s Load: 0.009s, Pack+Encode: 2.215s, Decode+Unpack: 1.708s +---------------------- -------------------------------------------------------- +💾 Converting with 140.9066 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,584B, BPFP=0.2249 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,144B, BPFP=0.3228 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,036B, BPFP=0.3160 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,380B, BPFP=0.2748 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,816B, BPFP=0.2395 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,076B, BPFP=0.2558 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,580B, BPFP=0.2246 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,444B, BPFP=0.2789 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,552B, BPFP=0.3484 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,772B, BPFP=0.2994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,536B, BPFP=0.0227 +⌛️ [2/4] FRONTEND: Frontend time: 2.423s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.461s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10821181 390.47960592 + layer.0.v_cache 0.00001942 0.04773361 + layer.1.k_cache 0.30751975 24.39938269 + layer.1.v_cache 0.00000599 0.01766846 + layer.2.k_cache 0.01421910 3.21371815 + layer.2.v_cache 0.00002020 0.05433351 + layer.3.k_cache 0.01730646 13.19117780 + layer.3.v_cache 0.00002068 0.06597346 + layer.4.k_cache 0.00075739 1.41415663 + layer.4.v_cache 0.00005209 0.10392633 + layer.4.output 1.22958295 217.57279116 + ------------------------------------------------------------------------------------- + TOTAL 0.53265962 115.05865969 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 46920 +BPFP 0.1732 bits/point +EBPFP 0.3464 equivalent bits/point +MSE 115.058660 +---------------------- -------------------------------------------------------- +Time: 3.896s Load: 0.011s, Pack+Encode: 2.423s, Decode+Unpack: 1.461s +---------------------- -------------------------------------------------------- +💾 Converting with 115.0587 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 202, 128) +Output shape: (1, 202, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.output: torch.Size([1, 202, 3584]) -> torch.Size([1, 1, 202, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,184B, BPFP=0.2463 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,040B, BPFP=0.3899 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,304B, BPFP=0.4103 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,364B, BPFP=0.3376 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,016B, BPFP=0.3106 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,580B, BPFP=0.3543 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,684B, BPFP=0.2850 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,228B, BPFP=0.3270 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,176B, BPFP=0.4004 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,924B, BPFP=0.3809 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,324B, BPFP=0.0257 +⌛️ [2/4] FRONTEND: Frontend time: 2.031s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.431s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15282421 379.20645111 + layer.0.v_cache 0.00001801 0.04996990 + layer.1.k_cache 0.08597872 25.26703908 + layer.1.v_cache 0.00000649 0.01876136 + layer.2.k_cache 0.01425038 3.11694245 + layer.2.v_cache 0.00002047 0.05486376 + layer.3.k_cache 0.05704349 14.31839926 + layer.3.v_cache 0.00002256 0.06804161 + layer.4.k_cache 0.00069767 1.34620591 + layer.4.v_cache 0.00005056 0.10205902 + layer.4.output 1.51553867 268.22221977 + ------------------------------------------------------------------------------------- + TOTAL 0.64233431 135.35907481 + (elements=1,758,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1758208 +Total Bytes 46824 +BPFP 0.2131 bits/point +EBPFP 0.4261 equivalent bits/point +MSE 135.359075 +---------------------- -------------------------------------------------------- +Time: 3.470s Load: 0.008s, Pack+Encode: 2.031s, Decode+Unpack: 1.431s +---------------------- -------------------------------------------------------- +💾 Converting with 135.3591 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,860B, BPFP=0.2683 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,500B, BPFP=0.3589 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,188B, BPFP=0.3417 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,488B, BPFP=0.3030 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,760B, BPFP=0.2628 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,364B, BPFP=0.2962 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,424B, BPFP=0.2443 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,824B, BPFP=0.3216 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,828B, BPFP=0.3770 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,164B, BPFP=0.3403 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,280B, BPFP=0.0180 +⌛️ [2/4] FRONTEND: Frontend time: 2.386s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.786s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14837397 402.75151833 + layer.0.v_cache 0.00001603 0.04150311 + layer.1.k_cache 0.60911565 25.36027737 + layer.1.v_cache 0.00000597 0.01588996 + layer.2.k_cache 0.02477714 3.05400511 + layer.2.v_cache 0.00001938 0.04976570 + layer.3.k_cache 0.01052417 14.05610058 + layer.3.v_cache 0.00002039 0.05965232 + layer.4.k_cache 0.00073933 1.25599741 + layer.4.v_cache 0.00005963 0.09936843 + layer.4.output 0.00470031 196.23968324 + ------------------------------------------------------------------------------------- + TOTAL 0.04862081 107.08363888 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 58680 +BPFP 0.1906 bits/point +EBPFP 0.3812 equivalent bits/point +MSE 107.083639 +---------------------- -------------------------------------------------------- +Time: 4.184s Load: 0.012s, Pack+Encode: 2.386s, Decode+Unpack: 1.786s +---------------------- -------------------------------------------------------- +💾 Converting with 107.0836 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 210, 128) +Output shape: (1, 210, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.output: torch.Size([1, 210, 3584]) -> torch.Size([1, 1, 210, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,612B, BPFP=0.2687 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,104B, BPFP=0.3798 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,556B, BPFP=0.4134 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,460B, BPFP=0.3318 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,148B, BPFP=0.3086 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,292B, BPFP=0.3193 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,916B, BPFP=0.2914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,636B, BPFP=0.3449 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,676B, BPFP=0.4223 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,956B, BPFP=0.3688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,368B, BPFP=0.0252 +⌛️ [2/4] FRONTEND: Frontend time: 2.280s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.674s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11415643 398.19776786 + layer.0.v_cache 0.00001882 0.04720868 + layer.1.k_cache 0.27145578 25.60229492 + layer.1.v_cache 0.00000623 0.01789650 + layer.2.k_cache 0.01201470 3.09704328 + layer.2.v_cache 0.00002208 0.05294946 + layer.3.k_cache 0.00887682 13.26833612 + layer.3.v_cache 0.00002165 0.06516049 + layer.4.k_cache 0.00072611 1.33589071 + layer.4.v_cache 0.00005233 0.10578115 + layer.4.output 1.45781997 258.01894133 + ------------------------------------------------------------------------------------- + TOTAL 0.62424063 132.23075991 + (elements=1,827,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1827840 +Total Bytes 48724 +BPFP 0.2133 bits/point +EBPFP 0.4265 equivalent bits/point +MSE 132.230760 +---------------------- -------------------------------------------------------- +Time: 3.963s Load: 0.008s, Pack+Encode: 2.280s, Decode+Unpack: 1.674s +---------------------- -------------------------------------------------------- +💾 Converting with 132.2308 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 165, 128) +Output shape: (1, 165, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.output: torch.Size([1, 165, 3584]) -> torch.Size([1, 1, 165, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,036B, BPFP=0.2875 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,332B, BPFP=0.3155 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,192B, BPFP=0.3970 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,184B, BPFP=0.3015 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,380B, BPFP=0.3201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,108B, BPFP=0.2943 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,244B, BPFP=0.3072 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,200B, BPFP=0.3030 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,996B, BPFP=0.3784 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,724B, BPFP=0.3527 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,168B, BPFP=0.0293 +⌛️ [2/4] FRONTEND: Frontend time: 2.217s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10245823 400.11482008 + layer.0.v_cache 0.00002070 0.04995487 + layer.1.k_cache 0.13127365 24.41094934 + layer.1.v_cache 0.00000615 0.01815524 + layer.2.k_cache 0.01458293 2.96735082 + layer.2.v_cache 0.00002074 0.05313745 + layer.3.k_cache 0.02375244 13.65689364 + layer.3.v_cache 0.00002214 0.06706165 + layer.4.k_cache 0.00067489 1.35773362 + layer.4.v_cache 0.00005196 0.10669694 + layer.4.output 0.00973386 332.97410714 + ------------------------------------------------------------------------------------- + TOTAL 0.02005887 163.15420610 + (elements=1,436,160) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1436160 +Total Bytes 36564 +BPFP 0.2037 bits/point +EBPFP 0.4074 equivalent bits/point +MSE 163.154206 +---------------------- -------------------------------------------------------- +Time: 3.572s Load: 0.008s, Pack+Encode: 2.217s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 163.1542 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,124B, BPFP=0.2667 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,516B, BPFP=0.3002 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,076B, BPFP=0.3480 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,168B, BPFP=0.2705 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,112B, BPFP=0.2657 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,248B, BPFP=0.2773 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,868B, BPFP=0.2449 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,480B, BPFP=0.2971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,344B, BPFP=0.3709 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,472B, BPFP=0.2964 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,696B, BPFP=0.0329 +⌛️ [2/4] FRONTEND: Frontend time: 2.343s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.584s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12273688 406.25409836 + layer.0.v_cache 0.00001959 0.05049046 + layer.1.k_cache 0.11126613 24.95043545 + layer.1.v_cache 0.00000674 0.01924967 + layer.2.k_cache 0.01416154 3.28061030 + layer.2.v_cache 0.00002164 0.05538260 + layer.3.k_cache 0.04491856 14.08143757 + layer.3.v_cache 0.00002114 0.06774181 + layer.4.k_cache 0.00068813 1.45928005 + layer.4.v_cache 0.00006151 0.10968987 + layer.4.output 0.00884247 300.23358216 + ------------------------------------------------------------------------------------- + TOTAL 0.02092936 150.11549949 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 37104 +BPFP 0.1864 bits/point +EBPFP 0.3727 equivalent bits/point +MSE 150.115499 +---------------------- -------------------------------------------------------- +Time: 3.936s Load: 0.008s, Pack+Encode: 2.343s, Decode+Unpack: 1.584s +---------------------- -------------------------------------------------------- +💾 Converting with 150.1155 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 257, 128) +Output shape: (1, 257, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.output: torch.Size([1, 257, 3584]) -> torch.Size([1, 1, 257, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,332B, BPFP=0.2634 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,132B, BPFP=0.3728 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,824B, BPFP=0.3541 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,364B, BPFP=0.3261 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,296B, BPFP=0.2612 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,580B, BPFP=0.2785 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,096B, BPFP=0.2490 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,092B, BPFP=0.3096 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,380B, BPFP=0.3879 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,172B, BPFP=0.3752 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,556B, BPFP=0.0222 +⌛️ [2/4] FRONTEND: Frontend time: 2.093s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.557s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12694686 394.96133268 + layer.0.v_cache 0.00002055 0.04869266 + layer.1.k_cache 0.54844232 25.16728858 + layer.1.v_cache 0.00000676 0.01762663 + layer.2.k_cache 0.01835482 2.82286315 + layer.2.v_cache 0.00002099 0.05181897 + layer.3.k_cache 0.02141366 13.02885324 + layer.3.v_cache 0.00002047 0.06482905 + layer.4.k_cache 0.00070767 1.31511002 + layer.4.v_cache 0.00005046 0.09814243 + layer.4.output 0.00513101 215.99796762 + ------------------------------------------------------------------------------------- + TOTAL 0.04422951 114.68013710 + (elements=2,236,928) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2236928 +Total Bytes 54824 +BPFP 0.1961 bits/point +EBPFP 0.3921 equivalent bits/point +MSE 114.680137 +---------------------- -------------------------------------------------------- +Time: 3.659s Load: 0.009s, Pack+Encode: 2.093s, Decode+Unpack: 1.557s +---------------------- -------------------------------------------------------- +💾 Converting with 114.6801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,340B, BPFP=0.2867 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,464B, BPFP=0.3832 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,176B, BPFP=0.3585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,396B, BPFP=0.2916 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,192B, BPFP=0.2740 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,344B, BPFP=0.2871 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,976B, BPFP=0.2555 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,428B, BPFP=0.2943 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,396B, BPFP=0.3774 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,536B, BPFP=0.3036 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,612B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 1.895s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12731336 415.89684924 + layer.0.v_cache 0.00001750 0.05075943 + layer.1.k_cache 0.10052182 25.01032903 + layer.1.v_cache 0.00000572 0.01743365 + layer.2.k_cache 0.01141533 3.41353934 + layer.2.v_cache 0.00001926 0.05043473 + layer.3.k_cache 0.03745206 14.19529909 + layer.3.v_cache 0.00002083 0.06564549 + layer.4.k_cache 0.00066429 1.39960186 + layer.4.v_cache 0.00005101 0.10447996 + layer.4.output 0.00884527 301.77818387 + ------------------------------------------------------------------------------------- + TOTAL 0.01996459 151.33245052 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 38860 +BPFP 0.1962 bits/point +EBPFP 0.3925 equivalent bits/point +MSE 151.332451 +---------------------- -------------------------------------------------------- +Time: 3.269s Load: 0.007s, Pack+Encode: 1.895s, Decode+Unpack: 1.367s +---------------------- -------------------------------------------------------- +💾 Converting with 151.3325 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,616B, BPFP=0.2545 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,032B, BPFP=0.3542 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,872B, BPFP=0.3429 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,200B, BPFP=0.2956 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,788B, BPFP=0.2666 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,240B, BPFP=0.2984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,512B, BPFP=0.2472 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,296B, BPFP=0.3024 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,220B, BPFP=0.3674 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,352B, BPFP=0.3063 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,968B, BPFP=0.0198 +⌛️ [2/4] FRONTEND: Frontend time: 2.019s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.423s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12752426 394.69052646 + layer.0.v_cache 0.00002004 0.04975103 + layer.1.k_cache 0.30660921 25.40838348 + layer.1.v_cache 0.00000613 0.01728218 + layer.2.k_cache 0.01808885 2.87704743 + layer.2.v_cache 0.00002107 0.05142825 + layer.3.k_cache 0.01678786 13.73075248 + layer.3.v_cache 0.00002050 0.06456618 + layer.4.k_cache 0.00068290 1.32703022 + layer.4.v_cache 0.00004851 0.09966131 + layer.4.output 1.37905047 244.07074485 + ------------------------------------------------------------------------------------- + TOTAL 0.59548016 126.28303782 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 45096 +BPFP 0.1867 bits/point +EBPFP 0.3734 equivalent bits/point +MSE 126.283038 +---------------------- -------------------------------------------------------- +Time: 3.450s Load: 0.008s, Pack+Encode: 2.019s, Decode+Unpack: 1.423s +---------------------- -------------------------------------------------------- +💾 Converting with 126.2830 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,616B, BPFP=0.2604 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,912B, BPFP=0.3537 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,232B, BPFP=0.3767 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,520B, BPFP=0.3255 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,016B, BPFP=0.2892 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,584B, BPFP=0.3301 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,844B, BPFP=0.2768 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,684B, BPFP=0.3373 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,612B, BPFP=0.4041 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,020B, BPFP=0.3615 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,048B, BPFP=0.0211 +⌛️ [2/4] FRONTEND: Frontend time: 2.007s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.465s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12766143 388.70870536 + layer.0.v_cache 0.00001653 0.04740766 + layer.1.k_cache 0.34631467 25.15282528 + layer.1.v_cache 0.00000659 0.01808337 + layer.2.k_cache 0.01786991 2.99968751 + layer.2.v_cache 0.00002041 0.05093128 + layer.3.k_cache 0.06874437 13.01858394 + layer.3.v_cache 0.00002276 0.06949147 + layer.4.k_cache 0.00067882 1.32292843 + layer.4.v_cache 0.00005621 0.10641303 + layer.4.output 1.41085444 249.69326037 + ------------------------------------------------------------------------------------- + TOTAL 0.61396310 128.19693411 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 48088 +BPFP 0.2037 bits/point +EBPFP 0.4074 equivalent bits/point +MSE 128.196934 +---------------------- -------------------------------------------------------- +Time: 3.480s Load: 0.008s, Pack+Encode: 2.007s, Decode+Unpack: 1.465s +---------------------- -------------------------------------------------------- +💾 Converting with 128.1969 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,412B, BPFP=0.2359 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,948B, BPFP=0.3421 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,136B, BPFP=0.3551 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,748B, BPFP=0.3283 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,080B, BPFP=0.2821 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,200B, BPFP=0.2904 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,876B, BPFP=0.2680 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,328B, BPFP=0.2992 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,456B, BPFP=0.3772 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,612B, BPFP=0.3189 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,976B, BPFP=0.0195 +⌛️ [2/4] FRONTEND: Frontend time: 2.037s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12070838 390.93985066 + layer.0.v_cache 0.00001763 0.04933136 + layer.1.k_cache 0.31015487 24.81283704 + layer.1.v_cache 0.00000677 0.01959380 + layer.2.k_cache 0.03544329 3.04725633 + layer.2.v_cache 0.00002212 0.05636695 + layer.3.k_cache 0.05682086 13.67004827 + layer.3.v_cache 0.00002136 0.06966083 + layer.4.k_cache 0.00073636 1.38456294 + layer.4.v_cache 0.00005031 0.10400465 + layer.4.output 1.35470685 239.76999052 + ------------------------------------------------------------------------------------- + TOTAL 0.58864294 124.26726156 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 46772 +BPFP 0.1902 bits/point +EBPFP 0.3804 equivalent bits/point +MSE 124.267262 +---------------------- -------------------------------------------------------- +Time: 3.483s Load: 0.007s, Pack+Encode: 2.037s, Decode+Unpack: 1.439s +---------------------- -------------------------------------------------------- +💾 Converting with 124.2673 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,808B, BPFP=0.2793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,992B, BPFP=0.3662 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,320B, BPFP=0.3903 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,384B, BPFP=0.3216 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,128B, BPFP=0.3028 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,788B, BPFP=0.3512 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,024B, BPFP=0.2952 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,048B, BPFP=0.3703 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,664B, BPFP=0.4155 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,168B, BPFP=0.3791 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,368B, BPFP=0.0248 +⌛️ [2/4] FRONTEND: Frontend time: 2.049s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.447s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13227103 395.01540493 + layer.0.v_cache 0.00002050 0.04908602 + layer.1.k_cache 0.22864332 24.85857495 + layer.1.v_cache 0.00000662 0.01885521 + layer.2.k_cache 0.02950178 3.08600226 + layer.2.v_cache 0.00002095 0.05391142 + layer.3.k_cache 0.01835468 14.00139836 + layer.3.v_cache 0.00002123 0.06856746 + layer.4.k_cache 0.00070999 1.37866698 + layer.4.v_cache 0.00005457 0.10555299 + layer.4.output 1.43734575 254.40809440 + ------------------------------------------------------------------------------------- + TOTAL 0.61594264 130.55839302 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 49692 +BPFP 0.2144 bits/point +EBPFP 0.4289 equivalent bits/point +MSE 130.558393 +---------------------- -------------------------------------------------------- +Time: 3.504s Load: 0.008s, Pack+Encode: 2.049s, Decode+Unpack: 1.447s +---------------------- -------------------------------------------------------- +💾 Converting with 130.5584 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,736B, BPFP=0.2630 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,692B, BPFP=0.3302 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,924B, BPFP=0.3466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,236B, BPFP=0.2981 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,724B, BPFP=0.2621 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,816B, BPFP=0.2686 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,404B, BPFP=0.2396 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,112B, BPFP=0.2894 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,332B, BPFP=0.3753 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,152B, BPFP=0.2922 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,972B, BPFP=0.0198 +⌛️ [2/4] FRONTEND: Frontend time: 1.972s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.425s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12518616 397.20706644 + layer.0.v_cache 0.00001676 0.04637634 + layer.1.k_cache 0.33007166 24.89311699 + layer.1.v_cache 0.00000630 0.01813308 + layer.2.k_cache 0.01131002 3.06287721 + layer.2.v_cache 0.00002125 0.05422531 + layer.3.k_cache 0.02232737 14.19768792 + layer.3.v_cache 0.00001983 0.06289302 + layer.4.k_cache 0.00071526 1.38334724 + layer.4.v_cache 0.00005006 0.09906040 + layer.4.output 1.37904434 244.07412323 + ------------------------------------------------------------------------------------- + TOTAL 0.59664912 126.44374392 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 44100 +BPFP 0.1826 bits/point +EBPFP 0.3652 equivalent bits/point +MSE 126.443744 +---------------------- -------------------------------------------------------- +Time: 3.405s Load: 0.008s, Pack+Encode: 1.972s, Decode+Unpack: 1.425s +---------------------- -------------------------------------------------------- +💾 Converting with 126.4437 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,828B, BPFP=0.2451 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,828B, BPFP=0.3092 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,248B, BPFP=0.3361 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,844B, BPFP=0.3102 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,108B, BPFP=0.2631 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,176B, BPFP=0.2674 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,892B, BPFP=0.2492 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,400B, BPFP=0.2818 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,492B, BPFP=0.3517 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,920B, BPFP=0.3151 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,100B, BPFP=0.0192 +⌛️ [2/4] FRONTEND: Frontend time: 1.943s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.440s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14655781 398.06771901 + layer.0.v_cache 0.00001709 0.04563245 + layer.1.k_cache 0.36683986 24.55253426 + layer.1.v_cache 0.00000651 0.01799340 + layer.2.k_cache 0.01755252 3.20354324 + layer.2.v_cache 0.00002060 0.05272042 + layer.3.k_cache 0.02713263 13.70695521 + layer.3.v_cache 0.00001997 0.06062931 + layer.4.k_cache 0.00071102 1.33685415 + layer.4.v_cache 0.00005162 0.09799533 + layer.4.output 1.25479176 222.09574429 + ------------------------------------------------------------------------------------- + TOTAL 0.54955600 117.40075217 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 47836 +BPFP 0.1802 bits/point +EBPFP 0.3604 equivalent bits/point +MSE 117.400752 +---------------------- -------------------------------------------------------- +Time: 3.392s Load: 0.009s, Pack+Encode: 1.943s, Decode+Unpack: 1.440s +---------------------- -------------------------------------------------------- +💾 Converting with 117.4008 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 155, 128) +Output shape: (1, 155, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.output: torch.Size([1, 155, 3584]) -> torch.Size([1, 1, 155, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,476B, BPFP=0.2496 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,696B, BPFP=0.3726 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,232B, BPFP=0.4266 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,272B, BPFP=0.3298 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,252B, BPFP=0.3278 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,472B, BPFP=0.3500 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,864B, BPFP=0.2887 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,548B, BPFP=0.3577 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,688B, BPFP=0.3718 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,724B, BPFP=0.3754 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,916B, BPFP=0.0276 +⌛️ [2/4] FRONTEND: Frontend time: 2.032s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.463s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102877 383.70690524 + layer.0.v_cache 0.00001822 0.04634147 + layer.1.k_cache 0.12773433 25.44214340 + layer.1.v_cache 0.00000585 0.01728070 + layer.2.k_cache 0.00931232 3.07310909 + layer.2.v_cache 0.00002110 0.05230974 + layer.3.k_cache 0.01537714 13.54187878 + layer.3.v_cache 0.00002091 0.06536611 + layer.4.k_cache 0.00066471 1.30129198 + layer.4.v_cache 0.00005180 0.10834472 + layer.4.output 0.01742802 357.66538018 + ------------------------------------------------------------------------------------- + TOTAL 0.02389596 172.41250779 + (elements=1,349,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1349120 +Total Bytes 36140 +BPFP 0.2143 bits/point +EBPFP 0.4286 equivalent bits/point +MSE 172.412508 +---------------------- -------------------------------------------------------- +Time: 3.501s Load: 0.006s, Pack+Encode: 2.032s, Decode+Unpack: 1.463s +---------------------- -------------------------------------------------------- +💾 Converting with 172.4125 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,596B, BPFP=0.2613 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,804B, BPFP=0.3491 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,080B, BPFP=0.3692 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,540B, BPFP=0.3299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,972B, BPFP=0.2887 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,480B, BPFP=0.3256 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,888B, BPFP=0.2826 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,548B, BPFP=0.3305 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,636B, BPFP=0.4096 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,760B, BPFP=0.3459 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,128B, BPFP=0.0221 +⌛️ [2/4] FRONTEND: Frontend time: 2.183s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.704s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11659910 391.42594477 + layer.0.v_cache 0.00001673 0.04873702 + layer.1.k_cache 0.29571861 25.45900481 + layer.1.v_cache 0.00000615 0.01809456 + layer.2.k_cache 0.04707494 2.93579613 + layer.2.v_cache 0.00002002 0.05061272 + layer.3.k_cache 0.03274950 12.96538313 + layer.3.v_cache 0.00002078 0.06340739 + layer.4.k_cache 0.00068608 1.32968750 + layer.4.v_cache 0.00005136 0.10244637 + layer.4.output 1.42391751 252.01598837 + ------------------------------------------------------------------------------------- + TOTAL 0.61531564 129.32417841 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 47432 +BPFP 0.2028 bits/point +EBPFP 0.4055 equivalent bits/point +MSE 129.324178 +---------------------- -------------------------------------------------------- +Time: 3.895s Load: 0.008s, Pack+Encode: 2.183s, Decode+Unpack: 1.704s +---------------------- -------------------------------------------------------- +💾 Converting with 129.3242 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,924B, BPFP=0.2472 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,004B, BPFP=0.3153 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,168B, BPFP=0.3256 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,316B, BPFP=0.2719 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,776B, BPFP=0.2379 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,192B, BPFP=0.2641 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,612B, BPFP=0.2276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,860B, BPFP=0.3062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,728B, BPFP=0.3609 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,540B, BPFP=0.2860 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,936B, BPFP=0.0264 +⌛️ [2/4] FRONTEND: Frontend time: 2.079s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.454s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10631144 400.70315650 + layer.0.v_cache 0.00001780 0.04913712 + layer.1.k_cache 0.46192560 25.29357123 + layer.1.v_cache 0.00000619 0.01881637 + layer.2.k_cache 0.02890189 3.25943879 + layer.2.v_cache 0.00002115 0.05400213 + layer.3.k_cache 0.04994748 13.31476421 + layer.3.v_cache 0.00002140 0.06742319 + layer.4.k_cache 0.00068365 1.41312445 + layer.4.v_cache 0.00005040 0.10769360 + layer.4.output 1.23456514 218.50855055 + ------------------------------------------------------------------------------------- + TOTAL 0.54646135 116.10829302 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 48056 +BPFP 0.1781 bits/point +EBPFP 0.3562 equivalent bits/point +MSE 116.108293 +---------------------- -------------------------------------------------------- +Time: 3.540s Load: 0.008s, Pack+Encode: 2.079s, Decode+Unpack: 1.454s +---------------------- -------------------------------------------------------- +💾 Converting with 116.1083 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 194, 128) +Output shape: (1, 194, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.output: torch.Size([1, 194, 3584]) -> torch.Size([1, 1, 194, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,720B, BPFP=0.2191 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,548B, BPFP=0.3663 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,072B, BPFP=0.4085 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,016B, BPFP=0.3235 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,556B, BPFP=0.2864 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,104B, BPFP=0.3305 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,080B, BPFP=0.2481 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,168B, BPFP=0.3357 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,936B, BPFP=0.3976 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,028B, BPFP=0.4050 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,472B, BPFP=0.0284 +⌛️ [2/4] FRONTEND: Frontend time: 1.964s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.457s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15027005 381.31797680 + layer.0.v_cache 0.00001782 0.04906519 + layer.1.k_cache 0.21402168 26.50258487 + layer.1.v_cache 0.00000588 0.01796866 + layer.2.k_cache 0.02177384 2.98177157 + layer.2.v_cache 0.00002019 0.05325680 + layer.3.k_cache 0.04024272 14.45893026 + layer.3.v_cache 0.00002076 0.06478656 + layer.4.k_cache 0.00069591 1.31512294 + layer.4.v_cache 0.00005073 0.10631977 + layer.4.output 1.57794378 279.17880155 + ------------------------------------------------------------------------------------- + TOTAL 0.67486624 140.06584672 + (elements=1,688,576) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1688576 +Total Bytes 43700 +BPFP 0.2070 bits/point +EBPFP 0.4141 equivalent bits/point +MSE 140.065847 +---------------------- -------------------------------------------------------- +Time: 3.428s Load: 0.007s, Pack+Encode: 1.964s, Decode+Unpack: 1.457s +---------------------- -------------------------------------------------------- +💾 Converting with 140.0658 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 258, 128) +Output shape: (1, 258, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.output: torch.Size([1, 258, 3584]) -> torch.Size([1, 1, 258, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,080B, BPFP=0.2471 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,988B, BPFP=0.3626 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,924B, BPFP=0.3588 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,112B, BPFP=0.3096 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,316B, BPFP=0.2614 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,876B, BPFP=0.2953 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,128B, BPFP=0.2500 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,084B, BPFP=0.3079 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,308B, BPFP=0.3820 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,404B, BPFP=0.3878 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,592B, BPFP=0.0224 +⌛️ [2/4] FRONTEND: Frontend time: 2.317s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.653s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09035230 391.57152374 + layer.0.v_cache 0.00001726 0.04629776 + layer.1.k_cache 0.42455262 25.15235700 + layer.1.v_cache 0.00000600 0.01658595 + layer.2.k_cache 0.01891134 2.88959024 + layer.2.v_cache 0.00001996 0.05137679 + layer.3.k_cache 0.02093011 13.37458647 + layer.3.v_cache 0.00001928 0.06045882 + layer.4.k_cache 0.00072082 1.25655607 + layer.4.v_cache 0.00004951 0.10367541 + layer.4.output 0.00505962 215.16715116 + ------------------------------------------------------------------------------------- + TOTAL 0.03476450 114.15841567 + (elements=2,245,632) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2245632 +Total Bytes 54812 +BPFP 0.1953 bits/point +EBPFP 0.3905 equivalent bits/point +MSE 114.158416 +---------------------- -------------------------------------------------------- +Time: 3.979s Load: 0.009s, Pack+Encode: 2.317s, Decode+Unpack: 1.653s +---------------------- -------------------------------------------------------- +💾 Converting with 114.1584 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,472B, BPFP=0.2359 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,016B, BPFP=0.3408 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,332B, BPFP=0.3622 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,560B, BPFP=0.3098 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,252B, BPFP=0.2889 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,224B, BPFP=0.2870 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,924B, BPFP=0.2666 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,656B, BPFP=0.3163 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,788B, BPFP=0.3932 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,700B, BPFP=0.3193 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,180B, BPFP=0.0212 +⌛️ [2/4] FRONTEND: Frontend time: 2.062s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.429s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15008528 386.08675272 + layer.0.v_cache 0.00001847 0.04735919 + layer.1.k_cache 0.25612781 25.40091287 + layer.1.v_cache 0.00000678 0.01841681 + layer.2.k_cache 0.03931687 2.99472630 + layer.2.v_cache 0.00002107 0.05386521 + layer.3.k_cache 0.03786138 12.69926970 + layer.3.v_cache 0.00002177 0.06595544 + layer.4.k_cache 0.00069674 1.35143884 + layer.4.v_cache 0.00005115 0.09998136 + layer.4.output 1.33117570 235.61523680 + ------------------------------------------------------------------------------------- + TOTAL 0.57661395 122.24266683 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 48104 +BPFP 0.1922 bits/point +EBPFP 0.3845 equivalent bits/point +MSE 122.242667 +---------------------- -------------------------------------------------------- +Time: 3.499s Load: 0.008s, Pack+Encode: 2.062s, Decode+Unpack: 1.429s +---------------------- -------------------------------------------------------- +💾 Converting with 122.2427 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,912B, BPFP=0.2528 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,700B, BPFP=0.3212 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,364B, BPFP=0.3788 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,324B, BPFP=0.2885 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,368B, BPFP=0.2924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,144B, BPFP=0.2729 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,192B, BPFP=0.2771 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,464B, BPFP=0.3007 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,108B, BPFP=0.3566 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,628B, BPFP=0.3149 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,868B, BPFP=0.0232 +⌛️ [2/4] FRONTEND: Frontend time: 2.008s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.332s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15161699 399.75989583 + layer.0.v_cache 0.00001786 0.04940335 + layer.1.k_cache 0.08662784 24.85829536 + layer.1.v_cache 0.00000652 0.01889602 + layer.2.k_cache 0.02063684 3.44724291 + layer.2.v_cache 0.00002327 0.05507731 + layer.3.k_cache 0.08087930 13.63656820 + layer.3.v_cache 0.00002116 0.06726729 + layer.4.k_cache 0.00074161 1.43403219 + layer.4.v_cache 0.00006096 0.11107127 + layer.4.output 0.00898438 305.25262897 + ------------------------------------------------------------------------------------- + TOTAL 0.02373665 151.77683250 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 37072 +BPFP 0.1893 bits/point +EBPFP 0.3786 equivalent bits/point +MSE 151.776832 +---------------------- -------------------------------------------------------- +Time: 3.347s Load: 0.007s, Pack+Encode: 2.008s, Decode+Unpack: 1.332s +---------------------- -------------------------------------------------------- +💾 Converting with 151.7768 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,760B, BPFP=0.2670 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,752B, BPFP=0.3375 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,100B, BPFP=0.3622 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,540B, BPFP=0.3224 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,844B, BPFP=0.2730 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,188B, BPFP=0.2974 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,672B, BPFP=0.2608 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,196B, BPFP=0.2980 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,280B, BPFP=0.3750 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,496B, BPFP=0.3193 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,008B, BPFP=0.0204 +⌛️ [2/4] FRONTEND: Frontend time: 2.410s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.725s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14327769 402.24964489 + layer.0.v_cache 0.00001763 0.04835776 + layer.1.k_cache 0.34925704 25.59072710 + layer.1.v_cache 0.00000642 0.01856729 + layer.2.k_cache 0.02352169 2.93510160 + layer.2.v_cache 0.00002271 0.05367489 + layer.3.k_cache 0.00845094 14.06505682 + layer.3.v_cache 0.00002026 0.06588389 + layer.4.k_cache 0.00069169 1.34571977 + layer.4.v_cache 0.00005086 0.10603622 + layer.4.output 1.39158238 246.29653003 + ------------------------------------------------------------------------------------- + TOTAL 0.60390550 127.67967532 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 45836 +BPFP 0.1915 bits/point +EBPFP 0.3830 equivalent bits/point +MSE 127.679675 +---------------------- -------------------------------------------------------- +Time: 4.143s Load: 0.008s, Pack+Encode: 2.410s, Decode+Unpack: 1.725s +---------------------- -------------------------------------------------------- +💾 Converting with 127.6797 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,540B, BPFP=0.2580 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,092B, BPFP=0.3461 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,316B, BPFP=0.3589 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,432B, BPFP=0.3086 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,768B, BPFP=0.2709 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,108B, BPFP=0.2902 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,244B, BPFP=0.2411 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,640B, BPFP=0.3205 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,852B, BPFP=0.3893 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,984B, BPFP=0.3400 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,484B, BPFP=0.0202 +⌛️ [2/4] FRONTEND: Frontend time: 2.295s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.550s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12716211 397.46181818 + layer.0.v_cache 0.00001677 0.04686552 + layer.1.k_cache 0.51313377 25.63226740 + layer.1.v_cache 0.00000586 0.01750252 + layer.2.k_cache 0.02995576 2.87881348 + layer.2.v_cache 0.00002133 0.05430395 + layer.3.k_cache 0.03339291 12.62645774 + layer.3.v_cache 0.00002012 0.06490974 + layer.4.k_cache 0.00068476 1.29537065 + layer.4.v_cache 0.00005875 0.10180308 + layer.4.output 0.00481428 201.94936688 + ------------------------------------------------------------------------------------- + TOTAL 0.04342071 109.04856944 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 57460 +BPFP 0.1920 bits/point +EBPFP 0.3841 equivalent bits/point +MSE 109.048569 +---------------------- -------------------------------------------------------- +Time: 3.854s Load: 0.009s, Pack+Encode: 2.295s, Decode+Unpack: 1.550s +---------------------- -------------------------------------------------------- +💾 Converting with 109.0486 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,048B, BPFP=0.2490 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,544B, BPFP=0.2795 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,632B, BPFP=0.2849 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,212B, BPFP=0.2591 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,308B, BPFP=0.2035 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,736B, BPFP=0.2298 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,204B, BPFP=0.1971 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,368B, BPFP=0.2687 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,008B, BPFP=0.3081 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,340B, BPFP=0.2670 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,524B, BPFP=0.0222 +⌛️ [2/4] FRONTEND: Frontend time: 1.903s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.425s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13396681 415.10980561 + layer.0.v_cache 0.00001641 0.04678882 + layer.1.k_cache 0.47111181 23.79970088 + layer.1.v_cache 0.00000608 0.01808994 + layer.2.k_cache 0.02933135 2.93317371 + layer.2.v_cache 0.00002255 0.05375410 + layer.3.k_cache 0.03744583 12.53860305 + layer.3.v_cache 0.00002260 0.06396765 + layer.4.k_cache 0.00071777 1.31004346 + layer.4.v_cache 0.00004917 0.09717945 + layer.4.output 1.20538323 213.29803501 + ------------------------------------------------------------------------------------- + TOTAL 0.53590429 114.65043246 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 43924 +BPFP 0.1589 bits/point +EBPFP 0.3179 equivalent bits/point +MSE 114.650432 +---------------------- -------------------------------------------------------- +Time: 3.336s Load: 0.009s, Pack+Encode: 1.903s, Decode+Unpack: 1.425s +---------------------- -------------------------------------------------------- +💾 Converting with 114.6504 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,064B, BPFP=0.2205 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,904B, BPFP=0.3203 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,908B, BPFP=0.3205 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,396B, BPFP=0.2928 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,524B, BPFP=0.2454 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,244B, BPFP=0.2845 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,268B, BPFP=0.2316 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,660B, BPFP=0.3071 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,968B, BPFP=0.3780 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,420B, BPFP=0.3483 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,160B, BPFP=0.0167 +⌛️ [2/4] FRONTEND: Frontend time: 2.069s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.816s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14060297 382.23361545 + layer.0.v_cache 0.00001598 0.04695972 + layer.1.k_cache 0.58610762 25.29157003 + layer.1.v_cache 0.00000669 0.01764145 + layer.2.k_cache 0.02834670 2.69204330 + layer.2.v_cache 0.00002256 0.05391077 + layer.3.k_cache 0.01962393 13.45048269 + layer.3.v_cache 0.00002003 0.06318224 + layer.4.k_cache 0.00067797 1.32156022 + layer.4.v_cache 0.00005035 0.10110825 + layer.4.output 0.00463834 192.83668155 + ------------------------------------------------------------------------------------- + TOTAL 0.04752607 104.41934382 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 56516 +BPFP 0.1804 bits/point +EBPFP 0.3607 equivalent bits/point +MSE 104.419344 +---------------------- -------------------------------------------------------- +Time: 3.895s Load: 0.010s, Pack+Encode: 2.069s, Decode+Unpack: 1.816s +---------------------- -------------------------------------------------------- +💾 Converting with 104.4193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,812B, BPFP=0.2473 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,908B, BPFP=0.3037 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,248B, BPFP=0.3211 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,292B, BPFP=0.2720 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,788B, BPFP=0.2461 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,048B, BPFP=0.2595 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,588B, BPFP=0.2358 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,980B, BPFP=0.2560 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,656B, BPFP=0.3421 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,736B, BPFP=0.2948 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,248B, BPFP=0.0165 +⌛️ [2/4] FRONTEND: Frontend time: 2.068s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.549s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12404394 396.47126850 + layer.0.v_cache 0.00001645 0.04603944 + layer.1.k_cache 0.57408182 23.94005223 + layer.1.v_cache 0.00000616 0.01734875 + layer.2.k_cache 0.03522489 2.90665115 + layer.2.v_cache 0.00002326 0.05273919 + layer.3.k_cache 0.02908128 13.51368633 + layer.3.v_cache 0.00002078 0.06384732 + layer.4.k_cache 0.00072493 1.33488796 + layer.4.v_cache 0.00004967 0.09842560 + layer.4.output 0.04390477 178.51713757 + ------------------------------------------------------------------------------------- + TOTAL 0.06297686 99.29793585 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 56304 +BPFP 0.1702 bits/point +EBPFP 0.3405 equivalent bits/point +MSE 99.297936 +---------------------- -------------------------------------------------------- +Time: 3.627s Load: 0.010s, Pack+Encode: 2.068s, Decode+Unpack: 1.549s +---------------------- -------------------------------------------------------- +💾 Converting with 99.2979 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 242, 128) +Output shape: (1, 242, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.output: torch.Size([1, 242, 3584]) -> torch.Size([1, 1, 242, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,160B, BPFP=0.2686 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,092B, BPFP=0.3288 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,392B, BPFP=0.3481 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,588B, BPFP=0.2962 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,140B, BPFP=0.2673 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,252B, BPFP=0.2745 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,980B, BPFP=0.2570 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,380B, BPFP=0.2828 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,508B, BPFP=0.3556 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,844B, BPFP=0.3128 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,084B, BPFP=0.0192 +⌛️ [2/4] FRONTEND: Frontend time: 2.414s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.664s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11466674 401.86005294 + layer.0.v_cache 0.00001794 0.04644676 + layer.1.k_cache 0.39062487 25.06590990 + layer.1.v_cache 0.00000616 0.01743925 + layer.2.k_cache 0.02414889 3.00671841 + layer.2.v_cache 0.00002046 0.05105398 + layer.3.k_cache 0.03495786 13.21812250 + layer.3.v_cache 0.00002053 0.06389538 + layer.4.k_cache 0.00069125 1.35774080 + layer.4.v_cache 0.00005034 0.10328558 + layer.4.output 1.26514320 223.92918019 + ------------------------------------------------------------------------------------- + TOTAL 0.55418868 118.37028981 + (elements=2,106,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2106368 +Total Bytes 48420 +BPFP 0.1839 bits/point +EBPFP 0.3678 equivalent bits/point +MSE 118.370290 +---------------------- -------------------------------------------------------- +Time: 4.087s Load: 0.009s, Pack+Encode: 2.414s, Decode+Unpack: 1.664s +---------------------- -------------------------------------------------------- +💾 Converting with 118.3703 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 189, 128) +Output shape: (1, 189, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.output: torch.Size([1, 189, 3584]) -> torch.Size([1, 1, 189, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,764B, BPFP=0.2285 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,848B, BPFP=0.3181 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,740B, BPFP=0.3092 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,576B, BPFP=0.2956 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,688B, BPFP=0.2222 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,424B, BPFP=0.2831 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,580B, BPFP=0.2133 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,576B, BPFP=0.2956 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,040B, BPFP=0.3340 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,956B, BPFP=0.3271 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,044B, BPFP=0.0241 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.331s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14361043 409.59321263 + layer.0.v_cache 0.00001648 0.04988067 + layer.1.k_cache 0.18932698 24.73425874 + layer.1.v_cache 0.00000663 0.01950109 + layer.2.k_cache 0.00768016 2.98697045 + layer.2.v_cache 0.00002453 0.05724089 + layer.3.k_cache 0.01421528 14.04283828 + layer.3.v_cache 0.00002024 0.06780794 + layer.4.k_cache 0.00072804 1.43957939 + layer.4.v_cache 0.00005086 0.10788262 + layer.4.output 0.00858909 290.62658258 + ------------------------------------------------------------------------------------- + TOTAL 0.02445902 146.32266181 + (elements=1,645,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1645056 +Total Bytes 36236 +BPFP 0.1762 bits/point +EBPFP 0.3524 equivalent bits/point +MSE 146.322662 +---------------------- -------------------------------------------------------- +Time: 3.196s Load: 0.008s, Pack+Encode: 1.857s, Decode+Unpack: 1.331s +---------------------- -------------------------------------------------------- +💾 Converting with 146.3227 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,668B, BPFP=0.2691 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,104B, BPFP=0.3744 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,460B, BPFP=0.4005 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,488B, BPFP=0.3292 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,024B, BPFP=0.2952 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,520B, BPFP=0.3316 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,840B, BPFP=0.2817 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,608B, BPFP=0.3380 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,396B, BPFP=0.3958 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,120B, BPFP=0.3756 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,344B, BPFP=0.0246 +⌛️ [2/4] FRONTEND: Frontend time: 1.912s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16894778 395.73173415 + layer.0.v_cache 0.00001779 0.04733377 + layer.1.k_cache 0.37628758 25.37878246 + layer.1.v_cache 0.00000652 0.01769109 + layer.2.k_cache 0.02181128 2.99786721 + layer.2.v_cache 0.00002030 0.05112174 + layer.3.k_cache 0.01815432 13.43762608 + layer.3.v_cache 0.00002063 0.06345984 + layer.4.k_cache 0.00071780 1.31721461 + layer.4.v_cache 0.00005059 0.10302901 + layer.4.output 1.43727796 254.38786888 + ------------------------------------------------------------------------------------- + TOTAL 0.62629296 130.58005542 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 48572 +BPFP 0.2096 bits/point +EBPFP 0.4192 equivalent bits/point +MSE 130.580055 +---------------------- -------------------------------------------------------- +Time: 3.275s Load: 0.007s, Pack+Encode: 1.912s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 130.5801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,668B, BPFP=0.2653 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,256B, BPFP=0.3802 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,200B, BPFP=0.3762 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,436B, BPFP=0.3209 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,816B, BPFP=0.2760 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,576B, BPFP=0.3310 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,876B, BPFP=0.2804 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,016B, BPFP=0.3628 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,532B, BPFP=0.4002 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,124B, BPFP=0.3707 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,124B, BPFP=0.0219 +⌛️ [2/4] FRONTEND: Frontend time: 1.954s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.464s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09877178 394.60633681 + layer.0.v_cache 0.00001652 0.04693836 + layer.1.k_cache 0.28824160 25.03052436 + layer.1.v_cache 0.00000619 0.01763847 + layer.2.k_cache 0.02137197 2.85417684 + layer.2.v_cache 0.00001995 0.05039239 + layer.3.k_cache 0.03811082 13.80799470 + layer.3.v_cache 0.00002058 0.06099901 + layer.4.k_cache 0.00070031 1.31018123 + layer.4.v_cache 0.00004695 0.09635592 + layer.4.output 1.41729720 250.83072917 + ------------------------------------------------------------------------------------- + TOTAL 0.60990512 129.04097896 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 48624 +BPFP 0.2069 bits/point +EBPFP 0.4138 equivalent bits/point +MSE 129.040979 +---------------------- -------------------------------------------------------- +Time: 3.426s Load: 0.008s, Pack+Encode: 1.954s, Decode+Unpack: 1.464s +---------------------- -------------------------------------------------------- +💾 Converting with 129.0410 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,868B, BPFP=0.2674 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,836B, BPFP=0.3343 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,020B, BPFP=0.3471 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,480B, BPFP=0.3097 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,032B, BPFP=0.2788 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,572B, BPFP=0.3161 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,656B, BPFP=0.2528 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,784B, BPFP=0.3308 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,736B, BPFP=0.3966 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,064B, BPFP=0.3501 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,028B, BPFP=0.0200 +⌛️ [2/4] FRONTEND: Frontend time: 1.983s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.443s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12161526 406.30693446 + layer.0.v_cache 0.00001916 0.04633568 + layer.1.k_cache 0.26028682 24.65665618 + layer.1.v_cache 0.00000628 0.01777227 + layer.2.k_cache 0.01296707 2.99300553 + layer.2.v_cache 0.00002040 0.05126142 + layer.3.k_cache 0.04595671 13.28121111 + layer.3.v_cache 0.00002073 0.06060732 + layer.4.k_cache 0.00067352 1.30830707 + layer.4.v_cache 0.00005047 0.10192488 + layer.4.output 1.35465500 239.73635035 + ------------------------------------------------------------------------------------- + TOTAL 0.58377655 125.11638049 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 48076 +BPFP 0.1955 bits/point +EBPFP 0.3910 equivalent bits/point +MSE 125.116380 +---------------------- -------------------------------------------------------- +Time: 3.435s Load: 0.009s, Pack+Encode: 1.983s, Decode+Unpack: 1.443s +---------------------- -------------------------------------------------------- +💾 Converting with 125.1164 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,868B, BPFP=0.2674 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,548B, BPFP=0.3144 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,112B, BPFP=0.3534 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,404B, BPFP=0.3045 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,956B, BPFP=0.2735 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,172B, BPFP=0.2884 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,740B, BPFP=0.2586 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,524B, BPFP=0.3128 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,688B, BPFP=0.3933 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,580B, BPFP=0.3166 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,988B, BPFP=0.0196 +⌛️ [2/4] FRONTEND: Frontend time: 1.948s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.457s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10229862 387.46608822 + layer.0.v_cache 0.00001727 0.04613040 + layer.1.k_cache 0.34669285 25.36874093 + layer.1.v_cache 0.00000643 0.01760533 + layer.2.k_cache 0.01639456 2.89691648 + layer.2.v_cache 0.00002164 0.05363661 + layer.3.k_cache 0.03343437 13.35679424 + layer.3.v_cache 0.00002043 0.06329504 + layer.4.k_cache 0.00071218 1.29657590 + layer.4.v_cache 0.00005088 0.09933713 + layer.4.output 1.35462105 239.74290850 + ------------------------------------------------------------------------------------- + TOTAL 0.58717627 124.05091058 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 46580 +BPFP 0.1894 bits/point +EBPFP 0.3789 equivalent bits/point +MSE 124.050911 +---------------------- -------------------------------------------------------- +Time: 3.413s Load: 0.007s, Pack+Encode: 1.948s, Decode+Unpack: 1.457s +---------------------- -------------------------------------------------------- +💾 Converting with 124.0509 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,788B, BPFP=0.2607 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,848B, BPFP=0.3337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,208B, BPFP=0.3585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,500B, BPFP=0.3097 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,008B, BPFP=0.2759 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,160B, BPFP=0.2863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,956B, BPFP=0.2723 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,572B, BPFP=0.3147 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,732B, BPFP=0.3945 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,480B, BPFP=0.3084 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,256B, BPFP=0.0222 +⌛️ [2/4] FRONTEND: Frontend time: 1.934s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.416s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11800324 393.86350496 + layer.0.v_cache 0.00001819 0.04654100 + layer.1.k_cache 0.35011090 25.54516494 + layer.1.v_cache 0.00000612 0.01718953 + layer.2.k_cache 0.02017111 2.91321957 + layer.2.v_cache 0.00002006 0.05217007 + layer.3.k_cache 0.03026835 12.94235485 + layer.3.v_cache 0.00001975 0.06288944 + layer.4.k_cache 0.00074701 1.31762400 + layer.4.v_cache 0.00005118 0.09701626 + layer.4.output 1.34863711 238.68152140 + ------------------------------------------------------------------------------------- + TOTAL 0.58587504 123.97813673 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 47508 +BPFP 0.1924 bits/point +EBPFP 0.3847 equivalent bits/point +MSE 123.978137 +---------------------- -------------------------------------------------------- +Time: 3.357s Load: 0.007s, Pack+Encode: 1.934s, Decode+Unpack: 1.416s +---------------------- -------------------------------------------------------- +💾 Converting with 123.9781 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,472B, BPFP=0.2369 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,972B, BPFP=0.3392 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,248B, BPFP=0.3581 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,472B, BPFP=0.3051 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,132B, BPFP=0.2819 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,688B, BPFP=0.3199 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,956B, BPFP=0.2699 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,528B, BPFP=0.3090 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,508B, BPFP=0.3758 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,008B, BPFP=0.3417 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,184B, BPFP=0.0213 +⌛️ [2/4] FRONTEND: Frontend time: 1.962s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.460s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10957497 394.14458242 + layer.0.v_cache 0.00001697 0.04639714 + layer.1.k_cache 0.31262030 25.13482106 + layer.1.v_cache 0.00000629 0.01823863 + layer.2.k_cache 0.01311818 2.98163082 + layer.2.v_cache 0.00002149 0.05610083 + layer.3.k_cache 0.06227440 12.92064257 + layer.3.v_cache 0.00002139 0.06858465 + layer.4.k_cache 0.00068925 1.35465264 + layer.4.v_cache 0.00005302 0.11003555 + layer.4.output 1.33688878 236.62733936 + ------------------------------------------------------------------------------------- + TOTAL 0.57980104 123.13100364 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 48168 +BPFP 0.1933 bits/point +EBPFP 0.3867 equivalent bits/point +MSE 123.131004 +---------------------- -------------------------------------------------------- +Time: 3.430s Load: 0.008s, Pack+Encode: 1.962s, Decode+Unpack: 1.460s +---------------------- -------------------------------------------------------- +💾 Converting with 123.1310 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,604B, BPFP=0.2644 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,948B, BPFP=0.3630 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,552B, BPFP=0.4073 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,468B, BPFP=0.3278 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,036B, BPFP=0.2961 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,420B, BPFP=0.3242 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,840B, BPFP=0.2817 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,228B, BPFP=0.3102 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,524B, BPFP=0.4052 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,048B, BPFP=0.3703 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,320B, BPFP=0.0243 +⌛️ [2/4] FRONTEND: Frontend time: 1.945s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.392s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13987277 396.00029343 + layer.0.v_cache 0.00001887 0.04778325 + layer.1.k_cache 0.23498858 26.38544188 + layer.1.v_cache 0.00000603 0.01743901 + layer.2.k_cache 0.02028183 3.10874316 + layer.2.v_cache 0.00001992 0.05261771 + layer.3.k_cache 0.05552547 13.61223935 + layer.3.v_cache 0.00002044 0.06349376 + layer.4.k_cache 0.00067773 1.30814602 + layer.4.v_cache 0.00004777 0.09853646 + layer.4.output 1.43727485 254.38849765 + ------------------------------------------------------------------------------------- + TOTAL 0.61837549 130.67142456 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 47988 +BPFP 0.2071 bits/point +EBPFP 0.4141 equivalent bits/point +MSE 130.671425 +---------------------- -------------------------------------------------------- +Time: 3.345s Load: 0.008s, Pack+Encode: 1.945s, Decode+Unpack: 1.392s +---------------------- -------------------------------------------------------- +💾 Converting with 130.6714 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,516B, BPFP=0.2604 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,260B, BPFP=0.3895 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,324B, BPFP=0.3943 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,332B, BPFP=0.3208 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,052B, BPFP=0.3001 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,736B, BPFP=0.3507 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,692B, BPFP=0.2734 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,184B, BPFP=0.3098 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,576B, BPFP=0.4129 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,932B, BPFP=0.3652 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,340B, BPFP=0.0248 +⌛️ [2/4] FRONTEND: Frontend time: 1.962s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.429s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17004660 392.51488448 + layer.0.v_cache 0.00001740 0.04476392 + layer.1.k_cache 0.16740718 25.59419200 + layer.1.v_cache 0.00000654 0.01759182 + layer.2.k_cache 0.01191964 2.86785947 + layer.2.v_cache 0.00001991 0.05129215 + layer.3.k_cache 0.00660476 13.43252809 + layer.3.v_cache 0.00002076 0.06149071 + layer.4.k_cache 0.00070551 1.31383323 + layer.4.v_cache 0.00005179 0.10141639 + layer.4.output 1.45088344 256.79214624 + ------------------------------------------------------------------------------------- + TOTAL 0.61841083 131.38499270 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 47944 +BPFP 0.2088 bits/point +EBPFP 0.4177 equivalent bits/point +MSE 131.384993 +---------------------- -------------------------------------------------------- +Time: 3.399s Load: 0.008s, Pack+Encode: 1.962s, Decode+Unpack: 1.429s +---------------------- -------------------------------------------------------- +💾 Converting with 131.3850 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,644B, BPFP=0.2476 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,864B, BPFP=0.3304 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,324B, BPFP=0.3617 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,516B, BPFP=0.3068 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,180B, BPFP=0.2840 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,372B, BPFP=0.2970 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,828B, BPFP=0.2601 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,352B, BPFP=0.2957 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,856B, BPFP=0.3978 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,572B, BPFP=0.3106 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,228B, BPFP=0.0216 +⌛️ [2/4] FRONTEND: Frontend time: 1.960s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.463s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14712874 393.01919158 + layer.0.v_cache 0.00001826 0.04365977 + layer.1.k_cache 0.36788522 25.56423021 + layer.1.v_cache 0.00000598 0.01725094 + layer.2.k_cache 0.01156470 2.96009097 + layer.2.v_cache 0.00002030 0.05135875 + layer.3.k_cache 0.01315050 12.42462636 + layer.3.v_cache 0.00002017 0.05961463 + layer.4.k_cache 0.00068370 1.28838236 + layer.4.v_cache 0.00004817 0.09415347 + layer.4.output 1.33105469 235.58594720 + ------------------------------------------------------------------------------------- + TOTAL 0.57987697 122.62495232 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 47736 +BPFP 0.1908 bits/point +EBPFP 0.3815 equivalent bits/point +MSE 122.624952 +---------------------- -------------------------------------------------------- +Time: 3.431s Load: 0.008s, Pack+Encode: 1.960s, Decode+Unpack: 1.463s +---------------------- -------------------------------------------------------- +💾 Converting with 122.6250 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 188, 128) +Output shape: (1, 188, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.output: torch.Size([1, 188, 3584]) -> torch.Size([1, 1, 188, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,700B, BPFP=0.2244 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,500B, BPFP=0.2909 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,844B, BPFP=0.3195 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,592B, BPFP=0.2985 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,832B, BPFP=0.2354 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,380B, BPFP=0.2809 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,668B, BPFP=0.2217 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,200B, BPFP=0.2660 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,032B, BPFP=0.3351 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,520B, BPFP=0.2926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,340B, BPFP=0.0278 +⌛️ [2/4] FRONTEND: Frontend time: 1.848s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005002 402.94946809 + layer.0.v_cache 0.00001663 0.04618972 + layer.1.k_cache 0.14149688 24.51154733 + layer.1.v_cache 0.00000584 0.01766922 + layer.2.k_cache 0.00881058 3.33200723 + layer.2.v_cache 0.00002285 0.05457823 + layer.3.k_cache 0.03192479 13.92887456 + layer.3.v_cache 0.00001958 0.06327576 + layer.4.k_cache 0.00066612 1.37515194 + layer.4.v_cache 0.00005162 0.10639854 + layer.4.output 0.00857985 292.17389343 + ------------------------------------------------------------------------------------- + TOTAL 0.02253670 146.56484792 + (elements=1,636,352) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1636352 +Total Bytes 35608 +BPFP 0.1741 bits/point +EBPFP 0.3482 equivalent bits/point +MSE 146.564848 +---------------------- -------------------------------------------------------- +Time: 3.222s Load: 0.007s, Pack+Encode: 1.848s, Decode+Unpack: 1.367s +---------------------- -------------------------------------------------------- +💾 Converting with 146.5648 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,208B, BPFP=0.2519 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,668B, BPFP=0.3665 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,024B, BPFP=0.3945 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,184B, BPFP=0.3285 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,920B, BPFP=0.3078 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,184B, BPFP=0.3285 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,548B, BPFP=0.2786 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,012B, BPFP=0.3150 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,080B, BPFP=0.3989 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,340B, BPFP=0.4193 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,236B, BPFP=0.0251 +⌛️ [2/4] FRONTEND: Frontend time: 2.050s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.456s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11435634 388.83605528 + layer.0.v_cache 0.00001764 0.04844500 + layer.1.k_cache 0.25025516 24.82549221 + layer.1.v_cache 0.00000690 0.01765520 + layer.2.k_cache 0.01528106 2.94331337 + layer.2.v_cache 0.00002105 0.05322955 + layer.3.k_cache 0.02533426 13.83741952 + layer.3.v_cache 0.00002127 0.06493197 + layer.4.k_cache 0.00070809 1.31238667 + layer.4.v_cache 0.00005072 0.10319751 + layer.4.output 1.53833902 272.25598977 + ------------------------------------------------------------------------------------- + TOTAL 0.65731915 137.51965027 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 45404 +BPFP 0.2097 bits/point +EBPFP 0.4194 equivalent bits/point +MSE 137.519650 +---------------------- -------------------------------------------------------- +Time: 3.514s Load: 0.008s, Pack+Encode: 2.050s, Decode+Unpack: 1.456s +---------------------- -------------------------------------------------------- +💾 Converting with 137.5197 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,984B, BPFP=0.2590 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,576B, BPFP=0.3104 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,324B, BPFP=0.3753 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,244B, BPFP=0.2816 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,380B, BPFP=0.2934 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,032B, BPFP=0.2632 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,160B, BPFP=0.2743 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,100B, BPFP=0.2691 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,448B, BPFP=0.3861 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,548B, BPFP=0.3080 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,884B, BPFP=0.0234 +⌛️ [2/4] FRONTEND: Frontend time: 1.902s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258550 402.30707465 + layer.0.v_cache 0.00001767 0.04924814 + layer.1.k_cache 0.14790291 25.32267253 + layer.1.v_cache 0.00000629 0.01786522 + layer.2.k_cache 0.00999225 3.50368856 + layer.2.v_cache 0.00002110 0.05558583 + layer.3.k_cache 0.02117847 14.27552355 + layer.3.v_cache 0.00002045 0.06851995 + layer.4.k_cache 0.00067024 1.41447576 + layer.4.v_cache 0.00006105 0.11002485 + layer.4.output 0.00897616 305.24139385 + ------------------------------------------------------------------------------------- + TOTAL 0.02207583 151.98908447 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 36680 +BPFP 0.1873 bits/point +EBPFP 0.3746 equivalent bits/point +MSE 151.989084 +---------------------- -------------------------------------------------------- +Time: 3.227s Load: 0.007s, Pack+Encode: 1.902s, Decode+Unpack: 1.319s +---------------------- -------------------------------------------------------- +💾 Converting with 151.9891 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,068B, BPFP=0.2350 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,772B, BPFP=0.3655 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,440B, BPFP=0.4167 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,156B, BPFP=0.3183 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,068B, BPFP=0.3116 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,196B, BPFP=0.3214 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,528B, BPFP=0.2702 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,124B, BPFP=0.3159 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,572B, BPFP=0.4268 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,016B, BPFP=0.3842 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,456B, BPFP=0.0269 +⌛️ [2/4] FRONTEND: Frontend time: 2.010s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.442s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09364704 385.43975950 + layer.0.v_cache 0.00001853 0.04751582 + layer.1.k_cache 0.25163280 24.78889017 + layer.1.v_cache 0.00000592 0.01732287 + layer.2.k_cache 0.01595125 3.07461129 + layer.2.v_cache 0.00002050 0.05236744 + layer.3.k_cache 0.04325477 14.17479152 + layer.3.v_cache 0.00002016 0.06161585 + layer.4.k_cache 0.00069210 1.30974609 + layer.4.v_cache 0.00005026 0.10298993 + layer.4.output 1.50065788 265.59692314 + ------------------------------------------------------------------------------------- + TOTAL 0.64175873 134.60282779 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 46396 +BPFP 0.2090 bits/point +EBPFP 0.4181 equivalent bits/point +MSE 134.602828 +---------------------- -------------------------------------------------------- +Time: 3.459s Load: 0.008s, Pack+Encode: 2.010s, Decode+Unpack: 1.442s +---------------------- -------------------------------------------------------- +💾 Converting with 134.6028 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,596B, BPFP=0.2486 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,956B, BPFP=0.3426 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,956B, BPFP=0.3426 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,608B, BPFP=0.3186 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,036B, BPFP=0.2790 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,704B, BPFP=0.3252 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,800B, BPFP=0.2627 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,416B, BPFP=0.3053 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,728B, BPFP=0.3960 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,908B, BPFP=0.3393 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,020B, BPFP=0.0200 +⌛️ [2/4] FRONTEND: Frontend time: 1.995s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.428s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510114 401.86324668 + layer.0.v_cache 0.00001717 0.04573520 + layer.1.k_cache 0.31797088 24.85781509 + layer.1.v_cache 0.00000673 0.01770589 + layer.2.k_cache 0.01711024 2.99998542 + layer.2.v_cache 0.00002022 0.05371832 + layer.3.k_cache 0.01307247 13.28297411 + layer.3.v_cache 0.00002089 0.06237083 + layer.4.k_cache 0.00069413 1.30951238 + layer.4.v_cache 0.00005461 0.10031639 + layer.4.output 1.35466395 239.73729851 + ------------------------------------------------------------------------------------- + TOTAL 0.58392448 124.86790999 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 47728 +BPFP 0.1941 bits/point +EBPFP 0.3882 equivalent bits/point +MSE 124.867910 +---------------------- -------------------------------------------------------- +Time: 3.431s Load: 0.008s, Pack+Encode: 1.995s, Decode+Unpack: 1.428s +---------------------- -------------------------------------------------------- +💾 Converting with 124.8679 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,888B, BPFP=0.2450 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,436B, BPFP=0.3425 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,060B, BPFP=0.3188 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,988B, BPFP=0.3143 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,772B, BPFP=0.2377 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,556B, BPFP=0.2870 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,604B, BPFP=0.2271 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,092B, BPFP=0.2578 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,524B, BPFP=0.3480 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,548B, BPFP=0.2865 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,868B, BPFP=0.0258 +⌛️ [2/4] FRONTEND: Frontend time: 1.974s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.462s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14860459 396.72794859 + layer.0.v_cache 0.00001629 0.04817508 + layer.1.k_cache 0.38266554 25.50789716 + layer.1.v_cache 0.00000669 0.02011403 + layer.2.k_cache 0.01628031 3.22149584 + layer.2.v_cache 0.00002106 0.05902541 + layer.3.k_cache 0.01842327 13.63542126 + layer.3.v_cache 0.00002289 0.06747394 + layer.4.k_cache 0.00072690 1.41541893 + layer.4.v_cache 0.00004923 0.10306025 + layer.4.output 1.23454798 218.49891993 + ------------------------------------------------------------------------------------- + TOTAL 0.54168545 115.89991000 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 48336 +BPFP 0.1791 bits/point +EBPFP 0.3583 equivalent bits/point +MSE 115.899910 +---------------------- -------------------------------------------------------- +Time: 3.444s Load: 0.008s, Pack+Encode: 1.974s, Decode+Unpack: 1.462s +---------------------- -------------------------------------------------------- +💾 Converting with 115.8999 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 435, 128) +Output shape: (1, 435, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.output: torch.Size([1, 435, 3584]) -> torch.Size([1, 1, 435, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,276B, BPFP=0.2614 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,132B, BPFP=0.2921 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,816B, BPFP=0.3167 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,016B, BPFP=0.2879 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,564B, BPFP=0.2358 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,708B, BPFP=0.2769 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,124B, BPFP=0.2200 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,180B, BPFP=0.2938 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,408B, BPFP=0.3379 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,040B, BPFP=0.2888 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,864B, BPFP=0.0147 +⌛️ [2/4] FRONTEND: Frontend time: 2.286s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.723s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15864419 404.10531609 + layer.0.v_cache 0.00001607 0.04605272 + layer.1.k_cache 0.94329666 25.47922953 + layer.1.v_cache 0.00000639 0.01799445 + layer.2.k_cache 0.02505282 2.87325987 + layer.2.v_cache 0.00002131 0.05310506 + layer.3.k_cache 0.01724711 13.13493400 + layer.3.v_cache 0.00002032 0.06056703 + layer.4.k_cache 0.00074513 1.32641209 + layer.4.v_cache 0.00005173 0.09687304 + layer.4.output 0.00608059 127.52325534 + ------------------------------------------------------------------------------------- + TOTAL 0.06986270 78.81509007 + (elements=3,786,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3786240 +Total Bytes 81128 +BPFP 0.1714 bits/point +EBPFP 0.3428 equivalent bits/point +MSE 78.815090 +---------------------- -------------------------------------------------------- +Time: 4.024s Load: 0.015s, Pack+Encode: 2.286s, Decode+Unpack: 1.723s +---------------------- -------------------------------------------------------- +💾 Converting with 78.8151 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 437, 128) +Output shape: (1, 437, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.output: torch.Size([1, 437, 3584]) -> torch.Size([1, 1, 437, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,068B, BPFP=0.2527 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,740B, BPFP=0.3125 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,824B, BPFP=0.3155 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,548B, BPFP=0.3056 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,388B, BPFP=0.2284 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,252B, BPFP=0.2951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,056B, BPFP=0.2165 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,236B, BPFP=0.2945 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,764B, BPFP=0.3491 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,740B, BPFP=0.3125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,464B, BPFP=0.0177 +⌛️ [2/4] FRONTEND: Frontend time: 2.423s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.783s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14454122 404.35068650 + layer.0.v_cache 0.00001702 0.04601515 + layer.1.k_cache 0.94267751 24.61811490 + layer.1.v_cache 0.00000668 0.01889554 + layer.2.k_cache 0.01808156 2.88209209 + layer.2.v_cache 0.00002154 0.05455564 + layer.3.k_cache 0.03034366 12.84007392 + layer.3.v_cache 0.00002210 0.06462072 + layer.4.k_cache 0.00074160 1.38175675 + layer.4.v_cache 0.00005552 0.10349006 + layer.4.output 0.00611241 126.89989580 + ------------------------------------------------------------------------------------- + TOTAL 0.06937031 78.50938658 + (elements=3,803,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3803648 +Total Bytes 84080 +BPFP 0.1768 bits/point +EBPFP 0.3537 equivalent bits/point +MSE 78.509387 +---------------------- -------------------------------------------------------- +Time: 4.220s Load: 0.014s, Pack+Encode: 2.423s, Decode+Unpack: 1.783s +---------------------- -------------------------------------------------------- +💾 Converting with 78.5094 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,988B, BPFP=0.2530 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,280B, BPFP=0.3186 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,436B, BPFP=0.3265 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,972B, BPFP=0.3030 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,000B, BPFP=0.2537 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,828B, BPFP=0.2957 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,668B, BPFP=0.2368 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,088B, BPFP=0.3088 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,008B, BPFP=0.3555 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,020B, BPFP=0.3054 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,380B, BPFP=0.0172 +⌛️ [2/4] FRONTEND: Frontend time: 2.055s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.648s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12827574 396.80628551 + layer.0.v_cache 0.00001866 0.04729933 + layer.1.k_cache 0.53653341 24.95181710 + layer.1.v_cache 0.00000626 0.01732172 + layer.2.k_cache 0.01950284 3.03298158 + layer.2.v_cache 0.00002083 0.05228087 + layer.3.k_cache 0.04340698 13.75893333 + layer.3.v_cache 0.00002181 0.06533027 + layer.4.k_cache 0.00072091 1.33433652 + layer.4.v_cache 0.00005393 0.10264049 + layer.4.output 0.04342448 176.21906888 + ------------------------------------------------------------------------------------- + TOTAL 0.06073722 98.45310052 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 60668 +BPFP 0.1810 bits/point +EBPFP 0.3621 equivalent bits/point +MSE 98.453101 +---------------------- -------------------------------------------------------- +Time: 3.713s Load: 0.011s, Pack+Encode: 2.055s, Decode+Unpack: 1.648s +---------------------- -------------------------------------------------------- +💾 Converting with 98.4531 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,496B, BPFP=0.2641 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,000B, BPFP=0.3524 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,364B, BPFP=0.3738 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,640B, BPFP=0.3313 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,724B, BPFP=0.2775 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,208B, BPFP=0.3059 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,392B, BPFP=0.2580 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,528B, BPFP=0.3247 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,584B, BPFP=0.3867 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,316B, BPFP=0.3710 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,356B, BPFP=0.0198 +⌛️ [2/4] FRONTEND: Frontend time: 2.641s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.665s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702504 396.03824013 + layer.0.v_cache 0.00001641 0.04728049 + layer.1.k_cache 0.51130361 24.57684923 + layer.1.v_cache 0.00000656 0.01872484 + layer.2.k_cache 0.01810413 2.89168692 + layer.2.v_cache 0.00002122 0.05477965 + layer.3.k_cache 0.02476560 12.89158871 + layer.3.v_cache 0.00002197 0.06496072 + layer.4.k_cache 0.00071356 1.33594668 + layer.4.v_cache 0.00005290 0.10382021 + layer.4.output 0.00501064 208.78210593 + ------------------------------------------------------------------------------------- + TOTAL 0.04100620 111.73521289 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 57608 +BPFP 0.1991 bits/point +EBPFP 0.3981 equivalent bits/point +MSE 111.735213 +---------------------- -------------------------------------------------------- +Time: 4.315s Load: 0.009s, Pack+Encode: 2.641s, Decode+Unpack: 1.665s +---------------------- -------------------------------------------------------- +💾 Converting with 111.7352 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,612B, BPFP=0.2412 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,704B, BPFP=0.3141 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,480B, BPFP=0.3659 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,480B, BPFP=0.2991 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,280B, BPFP=0.2858 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,120B, BPFP=0.2751 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,876B, BPFP=0.2588 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,244B, BPFP=0.2834 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,436B, BPFP=0.3630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,796B, BPFP=0.3202 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,344B, BPFP=0.0224 +⌛️ [2/4] FRONTEND: Frontend time: 1.950s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.433s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11892304 397.63635150 + layer.0.v_cache 0.00001803 0.04804266 + layer.1.k_cache 0.31174104 25.61143663 + layer.1.v_cache 0.00000657 0.01824741 + layer.2.k_cache 0.01452858 2.97908085 + layer.2.v_cache 0.00002161 0.05458554 + layer.3.k_cache 0.02263702 12.76895429 + layer.3.v_cache 0.00002141 0.06748750 + layer.4.k_cache 0.00068312 1.33065144 + layer.4.v_cache 0.00005251 0.10645139 + layer.4.output 1.30838121 231.56011523 + ------------------------------------------------------------------------------------- + TOTAL 0.56631185 121.26718211 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 47372 +BPFP 0.1861 bits/point +EBPFP 0.3721 equivalent bits/point +MSE 121.267182 +---------------------- -------------------------------------------------------- +Time: 3.392s Load: 0.009s, Pack+Encode: 1.950s, Decode+Unpack: 1.433s +---------------------- -------------------------------------------------------- +💾 Converting with 121.2672 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 233, 128) +Output shape: (1, 233, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.output: torch.Size([1, 233, 3584]) -> torch.Size([1, 1, 233, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,820B, BPFP=0.2562 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,864B, BPFP=0.3262 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,520B, BPFP=0.3702 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,296B, BPFP=0.2881 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,192B, BPFP=0.2811 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,288B, BPFP=0.2876 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,996B, BPFP=0.2680 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,532B, BPFP=0.3039 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,788B, BPFP=0.3881 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,760B, BPFP=0.3192 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,464B, BPFP=0.0236 +⌛️ [2/4] FRONTEND: Frontend time: 1.933s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13365312 396.29325376 + layer.0.v_cache 0.00001735 0.04710349 + layer.1.k_cache 0.40872042 24.99335896 + layer.1.v_cache 0.00000792 0.01876555 + layer.2.k_cache 0.01308113 3.10136132 + layer.2.v_cache 0.00002137 0.05282781 + layer.3.k_cache 0.00601081 13.19426154 + layer.3.v_cache 0.00002094 0.06185261 + layer.4.k_cache 0.00068546 1.31741320 + layer.4.v_cache 0.00005338 0.10100614 + layer.4.output 1.31399239 232.55715435 + ------------------------------------------------------------------------------------- + TOTAL 0.57413051 121.59301676 + (elements=2,028,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2028032 +Total Bytes 48520 +BPFP 0.1914 bits/point +EBPFP 0.3828 equivalent bits/point +MSE 121.593017 +---------------------- -------------------------------------------------------- +Time: 3.315s Load: 0.007s, Pack+Encode: 1.933s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 121.5930 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 148, 128) +Output shape: (1, 148, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.output: torch.Size([1, 148, 3584]) -> torch.Size([1, 1, 148, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,528B, BPFP=0.2669 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,780B, BPFP=0.3991 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,408B, BPFP=0.4654 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,372B, BPFP=0.3560 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,120B, BPFP=0.3294 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,596B, BPFP=0.3796 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,964B, BPFP=0.3129 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,328B, BPFP=0.3514 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,968B, BPFP=0.4189 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,824B, BPFP=0.4037 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,156B, BPFP=0.0325 +⌛️ [2/4] FRONTEND: Frontend time: 1.837s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.280s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08608804 382.66659628 + layer.0.v_cache 0.00001767 0.04661176 + layer.1.k_cache 0.12276106 24.50979862 + layer.1.v_cache 0.00000594 0.01728603 + layer.2.k_cache 0.00886028 3.21293867 + layer.2.v_cache 0.00002088 0.05548205 + layer.3.k_cache 0.01513371 14.15557696 + layer.3.v_cache 0.00002085 0.06507613 + layer.4.k_cache 0.00068430 1.30598811 + layer.4.v_cache 0.00005331 0.10937845 + layer.4.output 0.08957551 366.54174710 + ------------------------------------------------------------------------------------- + TOTAL 0.05062792 175.99629193 + (elements=1,288,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1288192 +Total Bytes 37044 +BPFP 0.2301 bits/point +EBPFP 0.4601 equivalent bits/point +MSE 175.996292 +---------------------- -------------------------------------------------------- +Time: 3.122s Load: 0.005s, Pack+Encode: 1.837s, Decode+Unpack: 1.280s +---------------------- -------------------------------------------------------- +💾 Converting with 175.9963 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 170, 128) +Output shape: (1, 170, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.output: torch.Size([1, 170, 3584]) -> torch.Size([1, 1, 170, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,008B, BPFP=0.2765 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,660B, BPFP=0.3364 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,364B, BPFP=0.4011 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,500B, BPFP=0.3217 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,508B, BPFP=0.3224 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,604B, BPFP=0.3312 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,296B, BPFP=0.3029 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,804B, BPFP=0.3496 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,172B, BPFP=0.3835 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,740B, BPFP=0.3438 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,264B, BPFP=0.0297 +⌛️ [2/4] FRONTEND: Frontend time: 1.793s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.545s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13519069 396.61015625 + layer.0.v_cache 0.00001779 0.04766726 + layer.1.k_cache 0.14803200 25.13891314 + layer.1.v_cache 0.00000678 0.01784455 + layer.2.k_cache 0.02267574 3.10239904 + layer.2.v_cache 0.00002005 0.04931071 + layer.3.k_cache 0.04618364 14.22061409 + layer.3.v_cache 0.00002282 0.06480945 + layer.4.k_cache 0.00077641 1.28987732 + layer.4.v_cache 0.00005264 0.10400170 + layer.4.output 0.00945594 323.16867122 + ------------------------------------------------------------------------------------- + TOTAL 0.02465707 158.98978188 + (elements=1,479,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1479680 +Total Bytes 38920 +BPFP 0.2104 bits/point +EBPFP 0.4208 equivalent bits/point +MSE 158.989782 +---------------------- -------------------------------------------------------- +Time: 3.344s Load: 0.006s, Pack+Encode: 1.793s, Decode+Unpack: 1.545s +---------------------- -------------------------------------------------------- +💾 Converting with 158.9898 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,440B, BPFP=0.2297 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,440B, BPFP=0.3238 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,348B, BPFP=0.4093 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,436B, BPFP=0.3234 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,416B, BPFP=0.3215 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,360B, BPFP=0.3163 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,140B, BPFP=0.2956 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,496B, BPFP=0.3291 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,156B, BPFP=0.3912 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,572B, BPFP=0.3362 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,180B, BPFP=0.0293 +⌛️ [2/4] FRONTEND: Frontend time: 1.823s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.268s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12520085 384.34476186 + layer.0.v_cache 0.00001716 0.04927601 + layer.1.k_cache 0.15152623 25.18919722 + layer.1.v_cache 0.00000633 0.01816627 + layer.2.k_cache 0.01605446 3.18158584 + layer.2.v_cache 0.00002200 0.05412999 + layer.3.k_cache 0.04332644 13.43215979 + layer.3.v_cache 0.00002349 0.06720029 + layer.4.k_cache 0.00069336 1.34694047 + layer.4.v_cache 0.00005614 0.10483284 + layer.4.output 0.00966471 330.96191910 + ------------------------------------------------------------------------------------- + TOTAL 0.02379879 161.44245202 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 36984 +BPFP 0.2048 bits/point +EBPFP 0.4095 equivalent bits/point +MSE 161.442452 +---------------------- -------------------------------------------------------- +Time: 3.099s Load: 0.008s, Pack+Encode: 1.823s, Decode+Unpack: 1.268s +---------------------- -------------------------------------------------------- +💾 Converting with 161.4425 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 181, 128) +Output shape: (1, 181, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.output: torch.Size([1, 181, 3584]) -> torch.Size([1, 1, 181, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,360B, BPFP=0.2901 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,684B, BPFP=0.3180 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,232B, BPFP=0.3653 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,484B, BPFP=0.3008 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,264B, BPFP=0.2818 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,592B, BPFP=0.3101 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,020B, BPFP=0.2607 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,464B, BPFP=0.2990 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,060B, BPFP=0.3505 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,532B, BPFP=0.3049 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,420B, BPFP=0.0298 +⌛️ [2/4] FRONTEND: Frontend time: 1.771s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.304s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13566811 414.40806285 + layer.0.v_cache 0.00002016 0.04795840 + layer.1.k_cache 0.14782082 24.98436960 + layer.1.v_cache 0.00000634 0.01828430 + layer.2.k_cache 0.01546242 3.41134391 + layer.2.v_cache 0.00002121 0.05242684 + layer.3.k_cache 0.02895848 13.67749293 + layer.3.v_cache 0.00002197 0.06551436 + layer.4.k_cache 0.00067433 1.36585973 + layer.4.v_cache 0.00005229 0.10462135 + layer.4.output 0.00892717 303.54972376 + ------------------------------------------------------------------------------------- + TOTAL 0.02301155 151.94023533 + (elements=1,575,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1575424 +Total Bytes 38112 +BPFP 0.1935 bits/point +EBPFP 0.3871 equivalent bits/point +MSE 151.940235 +---------------------- -------------------------------------------------------- +Time: 3.081s Load: 0.006s, Pack+Encode: 1.771s, Decode+Unpack: 1.304s +---------------------- -------------------------------------------------------- +💾 Converting with 151.9402 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 186, 128) +Output shape: (1, 186, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.output: torch.Size([1, 186, 3584]) -> torch.Size([1, 1, 186, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,624B, BPFP=0.2204 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,580B, BPFP=0.3007 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,948B, BPFP=0.3317 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,468B, BPFP=0.2913 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,956B, BPFP=0.2483 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,796B, BPFP=0.3189 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,748B, BPFP=0.2308 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,404B, BPFP=0.2860 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,112B, BPFP=0.3454 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,380B, BPFP=0.2839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,360B, BPFP=0.0283 +⌛️ [2/4] FRONTEND: Frontend time: 1.863s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.516s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09795722 385.89814348 + layer.0.v_cache 0.00001815 0.04967403 + layer.1.k_cache 0.13430934 25.46776819 + layer.1.v_cache 0.00000632 0.01854807 + layer.2.k_cache 0.01830538 3.51377623 + layer.2.v_cache 0.00002120 0.05637406 + layer.3.k_cache 0.04892764 14.30168300 + layer.3.v_cache 0.00002113 0.06723856 + layer.4.k_cache 0.00067235 1.43599201 + layer.4.v_cache 0.00008605 0.10783521 + layer.4.output 0.00869856 295.30369144 + ------------------------------------------------------------------------------------- + TOTAL 0.02124792 146.94369840 + (elements=1,618,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1618944 +Total Bytes 36376 +BPFP 0.1798 bits/point +EBPFP 0.3595 equivalent bits/point +MSE 146.943698 +---------------------- -------------------------------------------------------- +Time: 3.385s Load: 0.006s, Pack+Encode: 1.863s, Decode+Unpack: 1.516s +---------------------- -------------------------------------------------------- +💾 Converting with 146.9437 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 163, 128) +Output shape: (1, 163, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.output: torch.Size([1, 163, 3584]) -> torch.Size([1, 1, 163, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,852B, BPFP=0.2734 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,416B, BPFP=0.3275 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,192B, BPFP=0.4018 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,432B, BPFP=0.3290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,408B, BPFP=0.3267 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,452B, BPFP=0.3309 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,232B, BPFP=0.3098 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,412B, BPFP=0.3271 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,404B, BPFP=0.4222 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,644B, BPFP=0.3493 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,260B, BPFP=0.0309 +⌛️ [2/4] FRONTEND: Frontend time: 1.989s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11530171 396.76691910 + layer.0.v_cache 0.00001813 0.04788417 + layer.1.k_cache 0.14319220 25.31908431 + layer.1.v_cache 0.00000612 0.01775182 + layer.2.k_cache 0.02368733 2.96103336 + layer.2.v_cache 0.00002106 0.05133762 + layer.3.k_cache 0.02525310 13.46947793 + layer.3.v_cache 0.00002161 0.06668541 + layer.4.k_cache 0.00069984 1.32277244 + layer.4.v_cache 0.00005709 0.10493142 + layer.4.output 0.00982084 337.03943909 + ------------------------------------------------------------------------------------- + TOTAL 0.02217671 164.67082066 + (elements=1,418,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1418752 +Total Bytes 37704 +BPFP 0.2126 bits/point +EBPFP 0.4252 equivalent bits/point +MSE 164.670821 +---------------------- -------------------------------------------------------- +Time: 3.302s Load: 0.006s, Pack+Encode: 1.989s, Decode+Unpack: 1.307s +---------------------- -------------------------------------------------------- +💾 Converting with 164.6708 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 191, 128) +Output shape: (1, 191, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.output: torch.Size([1, 191, 3584]) -> torch.Size([1, 1, 191, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,164B, BPFP=0.2588 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,236B, BPFP=0.2647 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,404B, BPFP=0.2785 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,004B, BPFP=0.2457 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,448B, BPFP=0.2003 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,424B, BPFP=0.2801 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,280B, BPFP=0.1865 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,184B, BPFP=0.2605 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,684B, BPFP=0.3014 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,124B, BPFP=0.2556 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,732B, BPFP=0.0202 +⌛️ [2/4] FRONTEND: Frontend time: 1.843s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15112510 399.79667048 + layer.0.v_cache 0.00002069 0.04836094 + layer.1.k_cache 0.18491925 24.21316928 + layer.1.v_cache 0.00000629 0.01792789 + layer.2.k_cache 0.02122488 2.65711280 + layer.2.v_cache 0.00002327 0.05754396 + layer.3.k_cache 0.01259250 12.40883840 + layer.3.v_cache 0.00002182 0.06560641 + layer.4.k_cache 0.00068006 1.38022106 + layer.4.v_cache 0.00005885 0.10626019 + layer.4.output 0.00845948 287.63346111 + ------------------------------------------------------------------------------------- + TOTAL 0.02528759 144.36387877 + (elements=1,662,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1662464 +Total Bytes 32684 +BPFP 0.1573 bits/point +EBPFP 0.3146 equivalent bits/point +MSE 144.363879 +---------------------- -------------------------------------------------------- +Time: 3.160s Load: 0.007s, Pack+Encode: 1.843s, Decode+Unpack: 1.309s +---------------------- -------------------------------------------------------- +💾 Converting with 144.3639 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 174, 128) +Output shape: (1, 174, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.output: torch.Size([1, 174, 3584]) -> torch.Size([1, 1, 174, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,000B, BPFP=0.2694 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,428B, BPFP=0.3078 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,184B, BPFP=0.3757 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,156B, BPFP=0.2834 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,288B, BPFP=0.2953 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,088B, BPFP=0.2773 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,004B, BPFP=0.2698 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,108B, BPFP=0.2791 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,900B, BPFP=0.3502 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,504B, BPFP=0.3147 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,120B, BPFP=0.0272 +⌛️ [2/4] FRONTEND: Frontend time: 2.085s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.323s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14146894 394.01620869 + layer.0.v_cache 0.00001664 0.04777014 + layer.1.k_cache 0.16292955 24.45760372 + layer.1.v_cache 0.00000602 0.01791259 + layer.2.k_cache 0.01834387 3.30983094 + layer.2.v_cache 0.00002001 0.05273716 + layer.3.k_cache 0.03015787 13.88024762 + layer.3.v_cache 0.00002180 0.06648591 + layer.4.k_cache 0.00067479 1.35487059 + layer.4.v_cache 0.00005130 0.10727819 + layer.4.output 0.00923280 315.75413075 + ------------------------------------------------------------------------------------- + TOTAL 0.02460708 155.74058005 + (elements=1,514,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1514496 +Total Bytes 35780 +BPFP 0.1890 bits/point +EBPFP 0.3780 equivalent bits/point +MSE 155.740580 +---------------------- -------------------------------------------------------- +Time: 3.415s Load: 0.007s, Pack+Encode: 2.085s, Decode+Unpack: 1.323s +---------------------- -------------------------------------------------------- +💾 Converting with 155.7406 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.1912 bits/point +Avg EBPFP 0.3825 equivalent bits/point +Avg MSE 126.456220 +Avg Time 3.593s +------------------------ ---------------------------- diff --git a/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..6f47034da6b618d42fb7a725b6eda0cb1a3e7b5e --- /dev/null +++ b/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 286 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- -------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa +Output output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa +---------------- -------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,084B, BPFP=0.2216 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,032B, BPFP=0.3273 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,004B, BPFP=0.3257 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,956B, BPFP=0.3231 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,604B, BPFP=0.2498 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,784B, BPFP=0.3138 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,072B, BPFP=0.2209 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,092B, BPFP=0.3305 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,808B, BPFP=0.3694 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,740B, BPFP=0.3114 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,176B, BPFP=0.0169 +⌛️ [2/4] FRONTEND: Frontend time: 2.615s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.609s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14134069 384.94314236 + layer.0.v_cache 0.00001405 0.03736394 + layer.1.k_cache 0.62039534 25.08938938 + layer.1.v_cache 0.00000582 0.01447121 + layer.2.k_cache 0.00676800 2.79624452 + layer.2.v_cache 0.00001893 0.04548793 + layer.3.k_cache 0.03896382 13.63423665 + layer.3.v_cache 0.00001942 0.05338694 + layer.4.k_cache 0.00069280 1.12565168 + layer.4.v_cache 0.00005238 0.09266394 + layer.4.output 0.00804578 193.00620040 + ------------------------------------------------------------------------------------- + TOTAL 0.05085834 104.63973184 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 57352 +BPFP 0.1830 bits/point +EBPFP 0.3661 equivalent bits/point +MSE 104.639732 +---------------------- -------------------------------------------------------- +Time: 4.235s Load: 0.010s, Pack+Encode: 2.615s, Decode+Unpack: 1.609s +---------------------- -------------------------------------------------------- +💾 Converting with 104.6397 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,568B, BPFP=0.2436 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,076B, BPFP=0.3240 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,396B, BPFP=0.3411 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,664B, BPFP=0.3020 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,988B, BPFP=0.2660 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,768B, BPFP=0.3076 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,472B, BPFP=0.2385 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,836B, BPFP=0.3112 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,968B, BPFP=0.3716 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,768B, BPFP=0.3076 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,516B, BPFP=0.0192 +⌛️ [2/4] FRONTEND: Frontend time: 2.514s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.606s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15682776 393.05023464 + layer.0.v_cache 0.00001501 0.03818129 + layer.1.k_cache 0.54447479 25.46734182 + layer.1.v_cache 0.00000580 0.01510614 + layer.2.k_cache 0.00853899 2.91812353 + layer.2.v_cache 0.00001972 0.04644572 + layer.3.k_cache 0.02401569 12.59464324 + layer.3.v_cache 0.00001939 0.05510314 + layer.4.k_cache 0.00071059 1.15701471 + layer.4.v_cache 0.00005329 0.09545406 + layer.4.output 0.05121508 185.37725500 + ------------------------------------------------------------------------------------- + TOTAL 0.06430510 101.94579019 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 59020 +BPFP 0.1851 bits/point +EBPFP 0.3703 equivalent bits/point +MSE 101.945790 +---------------------- -------------------------------------------------------- +Time: 4.132s Load: 0.011s, Pack+Encode: 2.514s, Decode+Unpack: 1.606s +---------------------- -------------------------------------------------------- +💾 Converting with 101.9458 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,624B, BPFP=0.2457 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,868B, BPFP=0.3119 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,356B, BPFP=0.3378 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,820B, BPFP=0.3093 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,864B, BPFP=0.2585 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,564B, BPFP=0.2957 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,540B, BPFP=0.2413 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,736B, BPFP=0.3048 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,872B, BPFP=0.3652 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,740B, BPFP=0.3051 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,428B, BPFP=0.0184 +⌛️ [2/4] FRONTEND: Frontend time: 2.088s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.556s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15363285 393.74946854 + layer.0.v_cache 0.00001412 0.03660232 + layer.1.k_cache 0.60852461 25.09994154 + layer.1.v_cache 0.00000580 0.01476932 + layer.2.k_cache 0.01182756 2.85186352 + layer.2.v_cache 0.00001863 0.04372529 + layer.3.k_cache 0.03467803 12.61781861 + layer.3.v_cache 0.00001919 0.05380338 + layer.4.k_cache 0.00072319 1.12019172 + layer.4.v_cache 0.00005301 0.09066595 + layer.4.output 0.05035121 184.75580053 + ------------------------------------------------------------------------------------- + TOTAL 0.06835032 101.70408552 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 58412 +BPFP 0.1826 bits/point +EBPFP 0.3652 equivalent bits/point +MSE 101.704086 +---------------------- -------------------------------------------------------- +Time: 3.654s Load: 0.010s, Pack+Encode: 2.088s, Decode+Unpack: 1.556s +---------------------- -------------------------------------------------------- +💾 Converting with 101.7041 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,932B, BPFP=0.2723 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,172B, BPFP=0.3408 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,936B, BPFP=0.3277 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,748B, BPFP=0.3174 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,740B, BPFP=0.2617 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,500B, BPFP=0.3037 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,356B, BPFP=0.2405 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,928B, BPFP=0.3273 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,624B, BPFP=0.3657 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,672B, BPFP=0.3132 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,348B, BPFP=0.0185 +⌛️ [2/4] FRONTEND: Frontend time: 2.126s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.529s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12197856 401.83245914 + layer.0.v_cache 0.00001365 0.03730462 + layer.1.k_cache 0.62524759 24.18126104 + layer.1.v_cache 0.00000551 0.01431737 + layer.2.k_cache 0.01486346 2.81893414 + layer.2.v_cache 0.00001981 0.04621465 + layer.3.k_cache 0.07387851 13.69457749 + layer.3.v_cache 0.00001943 0.05408428 + layer.4.k_cache 0.00069560 1.13915584 + layer.4.v_cache 0.00005405 0.09744356 + layer.4.output 0.01098318 196.39823637 + ------------------------------------------------------------------------------------- + TOTAL 0.05374461 106.98255334 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 57956 +BPFP 0.1882 bits/point +EBPFP 0.3765 equivalent bits/point +MSE 106.982553 +---------------------- -------------------------------------------------------- +Time: 3.664s Load: 0.009s, Pack+Encode: 2.126s, Decode+Unpack: 1.529s +---------------------- -------------------------------------------------------- +💾 Converting with 106.9826 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,716B, BPFP=0.2586 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,988B, BPFP=0.3283 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,004B, BPFP=0.3292 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,608B, BPFP=0.3075 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,588B, BPFP=0.2515 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,808B, BPFP=0.3184 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,244B, BPFP=0.2327 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,632B, BPFP=0.3088 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,612B, BPFP=0.3625 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,312B, BPFP=0.2912 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,312B, BPFP=0.0181 +⌛️ [2/4] FRONTEND: Frontend time: 2.061s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.532s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15362121 402.63670504 + layer.0.v_cache 0.00001420 0.03694092 + layer.1.k_cache 0.52952420 25.07721354 + layer.1.v_cache 0.00000567 0.01455940 + layer.2.k_cache 0.00939641 2.97811044 + layer.2.v_cache 0.00001891 0.04530159 + layer.3.k_cache 0.04339437 13.38341471 + layer.3.v_cache 0.00002732 0.05397340 + layer.4.k_cache 0.00068326 1.15468300 + layer.4.v_cache 0.00005156 0.09102448 + layer.4.output 0.00776138 195.05167607 + ------------------------------------------------------------------------------------- + TOTAL 0.04653334 106.51962700 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 56824 +BPFP 0.1833 bits/point +EBPFP 0.3665 equivalent bits/point +MSE 106.519627 +---------------------- -------------------------------------------------------- +Time: 3.602s Load: 0.009s, Pack+Encode: 2.061s, Decode+Unpack: 1.532s +---------------------- -------------------------------------------------------- +💾 Converting with 106.5196 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,708B, BPFP=0.2502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,944B, BPFP=0.3159 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,424B, BPFP=0.3414 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,508B, BPFP=0.2927 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,872B, BPFP=0.2589 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,476B, BPFP=0.2910 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,640B, BPFP=0.2466 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,780B, BPFP=0.3072 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,756B, BPFP=0.3591 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,676B, BPFP=0.3017 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,552B, BPFP=0.0194 +⌛️ [2/4] FRONTEND: Frontend time: 2.090s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.775s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14921317 397.16395621 + layer.0.v_cache 0.00001395 0.03711864 + layer.1.k_cache 0.64083395 25.29363972 + layer.1.v_cache 0.00000574 0.01496174 + layer.2.k_cache 0.00578804 2.95878466 + layer.2.v_cache 0.00001912 0.04584691 + layer.3.k_cache 0.02848778 12.59981864 + layer.3.v_cache 0.00001956 0.05403368 + layer.4.k_cache 0.00070939 1.15564214 + layer.4.v_cache 0.00005206 0.09317059 + layer.4.output 0.04916091 184.73451166 + ------------------------------------------------------------------------------------- + TOTAL 0.06878054 101.91520909 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 58336 +BPFP 0.1824 bits/point +EBPFP 0.3647 equivalent bits/point +MSE 101.915209 +---------------------- -------------------------------------------------------- +Time: 3.875s Load: 0.010s, Pack+Encode: 2.090s, Decode+Unpack: 1.775s +---------------------- -------------------------------------------------------- +💾 Converting with 101.9152 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,632B, BPFP=0.2470 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,148B, BPFP=0.3279 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,396B, BPFP=0.3411 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,616B, BPFP=0.2995 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,964B, BPFP=0.2647 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,244B, BPFP=0.2797 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,512B, BPFP=0.2406 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,924B, BPFP=0.3159 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,780B, BPFP=0.3616 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,952B, BPFP=0.3174 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,484B, BPFP=0.0189 +⌛️ [2/4] FRONTEND: Frontend time: 2.431s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.533s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12806672 392.18094070 + layer.0.v_cache 0.00001388 0.03772723 + layer.1.k_cache 0.56701790 25.02277757 + layer.1.v_cache 0.00000557 0.01458874 + layer.2.k_cache 0.01020011 2.93575102 + layer.2.v_cache 0.00001881 0.04505527 + layer.3.k_cache 0.04301871 12.51643325 + layer.3.v_cache 0.00002051 0.05479496 + layer.4.k_cache 0.00071755 1.13038974 + layer.4.v_cache 0.00005284 0.09196410 + layer.4.output 0.04940383 185.39712335 + ------------------------------------------------------------------------------------- + TOTAL 0.06440938 101.87119330 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 58652 +BPFP 0.1840 bits/point +EBPFP 0.3680 equivalent bits/point +MSE 101.871193 +---------------------- -------------------------------------------------------- +Time: 3.974s Load: 0.010s, Pack+Encode: 2.431s, Decode+Unpack: 1.533s +---------------------- -------------------------------------------------------- +💾 Converting with 101.8712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,832B, BPFP=0.2649 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,116B, BPFP=0.3353 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,760B, BPFP=0.3158 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,764B, BPFP=0.3160 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,516B, BPFP=0.2476 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,112B, BPFP=0.3351 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,096B, BPFP=0.2246 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,956B, BPFP=0.3265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,352B, BPFP=0.3482 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,728B, BPFP=0.3140 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,300B, BPFP=0.0180 +⌛️ [2/4] FRONTEND: Frontend time: 2.134s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.496s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12903612 402.69610746 + layer.0.v_cache 0.00001458 0.03749297 + layer.1.k_cache 0.59860808 24.88958162 + layer.1.v_cache 0.00000581 0.01465074 + layer.2.k_cache 0.00940504 2.82623634 + layer.2.v_cache 0.00001954 0.04677212 + layer.3.k_cache 0.02644065 12.94442760 + layer.3.v_cache 0.00002029 0.05446346 + layer.4.k_cache 0.00068693 1.13708560 + layer.4.v_cache 0.00005376 0.09365776 + layer.4.output 0.00862506 195.04083647 + ------------------------------------------------------------------------------------- + TOTAL 0.04850978 106.47213711 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 57532 +BPFP 0.1855 bits/point +EBPFP 0.3711 equivalent bits/point +MSE 106.472137 +---------------------- -------------------------------------------------------- +Time: 3.640s Load: 0.010s, Pack+Encode: 2.134s, Decode+Unpack: 1.496s +---------------------- -------------------------------------------------------- +💾 Converting with 106.4721 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,884B, BPFP=0.2478 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,228B, BPFP=0.3159 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,396B, BPFP=0.3245 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,684B, BPFP=0.2884 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,880B, BPFP=0.2476 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,280B, BPFP=0.2679 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,712B, BPFP=0.2390 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,904B, BPFP=0.2995 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,684B, BPFP=0.3391 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,928B, BPFP=0.3007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,340B, BPFP=0.0170 +⌛️ [2/4] FRONTEND: Frontend time: 2.393s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.824s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16469599 401.89843750 + layer.0.v_cache 0.00001357 0.03717777 + layer.1.k_cache 0.70445866 25.10992035 + layer.1.v_cache 0.00000558 0.01477654 + layer.2.k_cache 0.01447187 3.10628183 + layer.2.v_cache 0.00001924 0.04505543 + layer.3.k_cache 0.02617162 12.13412119 + layer.3.v_cache 0.00001974 0.05427583 + layer.4.k_cache 0.00070375 1.16408331 + layer.4.v_cache 0.00005076 0.09279460 + layer.4.output 0.04814868 176.33137175 + ------------------------------------------------------------------------------------- + TOTAL 0.07339127 98.70450157 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 58920 +BPFP 0.1758 bits/point +EBPFP 0.3517 equivalent bits/point +MSE 98.704502 +---------------------- -------------------------------------------------------- +Time: 4.227s Load: 0.011s, Pack+Encode: 2.393s, Decode+Unpack: 1.824s +---------------------- -------------------------------------------------------- +💾 Converting with 98.7045 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,368B, BPFP=0.2537 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,264B, BPFP=0.3638 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,400B, BPFP=0.3717 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,752B, BPFP=0.3341 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,540B, BPFP=0.2637 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,684B, BPFP=0.3302 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,284B, BPFP=0.2488 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,832B, BPFP=0.3388 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,668B, BPFP=0.3873 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,048B, BPFP=0.3513 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,472B, BPFP=0.0205 +⌛️ [2/4] FRONTEND: Frontend time: 2.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.573s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12132761 392.11140799 + layer.0.v_cache 0.00001366 0.03636865 + layer.1.k_cache 0.57241889 25.76555058 + layer.1.v_cache 0.00000556 0.01400646 + layer.2.k_cache 0.00926234 2.78762556 + layer.2.v_cache 0.00002145 0.04463046 + layer.3.k_cache 0.03109839 13.14582002 + layer.3.v_cache 0.00002257 0.05201837 + layer.4.k_cache 0.00070701 1.13067400 + layer.4.v_cache 0.00004920 0.09032541 + layer.4.output 0.01031505 206.62765534 + ------------------------------------------------------------------------------------- + TOTAL 0.04747836 110.68070676 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 58312 +BPFP 0.1992 bits/point +EBPFP 0.3985 equivalent bits/point +MSE 110.680707 +---------------------- -------------------------------------------------------- +Time: 3.744s Load: 0.008s, Pack+Encode: 2.162s, Decode+Unpack: 1.573s +---------------------- -------------------------------------------------------- +💾 Converting with 110.6807 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,796B, BPFP=0.2515 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,908B, BPFP=0.3098 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,460B, BPFP=0.3387 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,724B, BPFP=0.3001 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,892B, BPFP=0.2565 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,236B, BPFP=0.2745 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,724B, BPFP=0.2477 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,624B, BPFP=0.2949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,908B, BPFP=0.3622 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,764B, BPFP=0.3022 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,508B, BPFP=0.0188 +⌛️ [2/4] FRONTEND: Frontend time: 2.117s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.570s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14357947 398.39363989 + layer.0.v_cache 0.00001420 0.03762528 + layer.1.k_cache 0.63723908 25.08383691 + layer.1.v_cache 0.00000573 0.01458622 + layer.2.k_cache 0.00765126 2.72902889 + layer.2.v_cache 0.00002073 0.04480164 + layer.3.k_cache 0.02843528 12.84699428 + layer.3.v_cache 0.00002112 0.05402009 + layer.4.k_cache 0.00069171 1.16011600 + layer.4.v_cache 0.00005234 0.09174078 + layer.4.output 0.04871205 182.27672579 + ------------------------------------------------------------------------------------- + TOTAL 0.06815855 100.96432180 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 58544 +BPFP 0.1806 bits/point +EBPFP 0.3611 equivalent bits/point +MSE 100.964322 +---------------------- -------------------------------------------------------- +Time: 3.698s Load: 0.011s, Pack+Encode: 2.117s, Decode+Unpack: 1.570s +---------------------- -------------------------------------------------------- +💾 Converting with 100.9643 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,968B, BPFP=0.2545 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,136B, BPFP=0.3143 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,416B, BPFP=0.3287 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,772B, BPFP=0.2957 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,864B, BPFP=0.2492 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,468B, BPFP=0.2801 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,692B, BPFP=0.2404 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,032B, BPFP=0.3090 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,996B, BPFP=0.3584 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,872B, BPFP=0.3008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,268B, BPFP=0.0166 +⌛️ [2/4] FRONTEND: Frontend time: 2.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.523s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16857396 401.70540471 + layer.0.v_cache 0.00001478 0.03791152 + layer.1.k_cache 0.58716421 23.89080110 + layer.1.v_cache 0.00000600 0.01494070 + layer.2.k_cache 0.00827454 3.01751289 + layer.2.v_cache 0.00002177 0.04625642 + layer.3.k_cache 0.02531170 12.85050029 + layer.3.v_cache 0.00001983 0.05316338 + layer.4.k_cache 0.00069463 1.16577789 + layer.4.v_cache 0.00005316 0.09336528 + layer.4.output 0.04836898 178.03760246 + ------------------------------------------------------------------------------------- + TOTAL 0.06639514 99.36110891 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 59484 +BPFP 0.1793 bits/point +EBPFP 0.3585 equivalent bits/point +MSE 99.361109 +---------------------- -------------------------------------------------------- +Time: 3.756s Load: 0.011s, Pack+Encode: 2.222s, Decode+Unpack: 1.523s +---------------------- -------------------------------------------------------- +💾 Converting with 99.3611 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,792B, BPFP=0.2513 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,852B, BPFP=0.3068 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,428B, BPFP=0.3370 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,704B, BPFP=0.2991 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,916B, BPFP=0.2578 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,260B, BPFP=0.2758 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,748B, BPFP=0.2490 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,776B, BPFP=0.3029 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,160B, BPFP=0.3754 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,884B, BPFP=0.3085 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,624B, BPFP=0.0197 +⌛️ [2/4] FRONTEND: Frontend time: 2.117s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.544s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14028978 401.25537437 + layer.0.v_cache 0.00001397 0.03726867 + layer.1.k_cache 0.59966084 25.18579757 + layer.1.v_cache 0.00000578 0.01469142 + layer.2.k_cache 0.01765286 2.84698179 + layer.2.v_cache 0.00001884 0.04418863 + layer.3.k_cache 0.02653811 13.39381521 + layer.3.v_cache 0.00001935 0.05323875 + layer.4.k_cache 0.00071527 1.17074677 + layer.4.v_cache 0.00005402 0.09160733 + layer.4.output 0.04907703 182.27828380 + ------------------------------------------------------------------------------------- + TOTAL 0.06638282 101.17892336 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 59144 +BPFP 0.1824 bits/point +EBPFP 0.3648 equivalent bits/point +MSE 101.178923 +---------------------- -------------------------------------------------------- +Time: 3.671s Load: 0.010s, Pack+Encode: 2.117s, Decode+Unpack: 1.544s +---------------------- -------------------------------------------------------- +💾 Converting with 101.1789 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,316B, BPFP=0.2479 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,304B, BPFP=0.3621 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,340B, BPFP=0.3642 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,488B, BPFP=0.3153 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,688B, BPFP=0.2693 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,100B, BPFP=0.2930 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,320B, BPFP=0.2482 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,912B, BPFP=0.3396 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,692B, BPFP=0.3844 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,908B, BPFP=0.3394 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,456B, BPFP=0.0202 +⌛️ [2/4] FRONTEND: Frontend time: 2.081s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.541s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13255837 392.96596392 + layer.0.v_cache 0.00001394 0.03731827 + layer.1.k_cache 0.54921044 25.46896901 + layer.1.v_cache 0.00000578 0.01441303 + layer.2.k_cache 0.01292971 2.84434150 + layer.2.v_cache 0.00001901 0.04649694 + layer.3.k_cache 0.07190662 13.58608829 + layer.3.v_cache 0.00001944 0.05401921 + layer.4.k_cache 0.00068159 1.08770370 + layer.4.v_cache 0.00005125 0.09198705 + layer.4.output 0.00790709 204.35262933 + ------------------------------------------------------------------------------------- + TOTAL 0.04839681 109.80386507 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 57524 +BPFP 0.1944 bits/point +EBPFP 0.3888 equivalent bits/point +MSE 109.803865 +---------------------- -------------------------------------------------------- +Time: 3.632s Load: 0.010s, Pack+Encode: 2.081s, Decode+Unpack: 1.541s +---------------------- -------------------------------------------------------- +💾 Converting with 109.8039 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,744B, BPFP=0.2629 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,100B, BPFP=0.3380 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,060B, BPFP=0.3358 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,696B, BPFP=0.3156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,860B, BPFP=0.2693 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,884B, BPFP=0.3260 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,356B, BPFP=0.2414 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,396B, BPFP=0.3544 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,676B, BPFP=0.3699 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,792B, BPFP=0.3209 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,352B, BPFP=0.0186 +⌛️ [2/4] FRONTEND: Frontend time: 2.124s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.567s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13695534 401.81155807 + layer.0.v_cache 0.00001452 0.03813546 + layer.1.k_cache 0.52272531 24.92413633 + layer.1.v_cache 0.00000574 0.01512674 + layer.2.k_cache 0.00793556 2.87331071 + layer.2.v_cache 0.00001907 0.04620139 + layer.3.k_cache 0.04682703 13.37227463 + layer.3.v_cache 0.00002008 0.05420896 + layer.4.k_cache 0.00068945 1.15265304 + layer.4.v_cache 0.00005221 0.09633639 + layer.4.output 0.00941004 197.10932751 + ------------------------------------------------------------------------------------- + TOTAL 0.04594792 107.30289613 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 58916 +BPFP 0.1920 bits/point +EBPFP 0.3840 equivalent bits/point +MSE 107.302896 +---------------------- -------------------------------------------------------- +Time: 3.701s Load: 0.010s, Pack+Encode: 2.124s, Decode+Unpack: 1.567s +---------------------- -------------------------------------------------------- +💾 Converting with 107.3029 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,524B, BPFP=0.2525 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,324B, BPFP=0.3529 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,140B, BPFP=0.3426 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,740B, BPFP=0.3203 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,668B, BPFP=0.2605 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,504B, BPFP=0.3071 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,320B, BPFP=0.2411 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,172B, BPFP=0.3444 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,644B, BPFP=0.3708 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,864B, BPFP=0.3272 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,308B, BPFP=0.0184 +⌛️ [2/4] FRONTEND: Frontend time: 2.093s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.508s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14340984 388.99369420 + layer.0.v_cache 0.00001397 0.03828787 + layer.1.k_cache 0.51622396 24.81761126 + layer.1.v_cache 0.00000557 0.01473880 + layer.2.k_cache 0.00797994 2.82080994 + layer.2.v_cache 0.00001813 0.04526270 + layer.3.k_cache 0.02756977 13.22647269 + layer.3.v_cache 0.00001890 0.05520572 + layer.4.k_cache 0.00068956 1.13209730 + layer.4.v_cache 0.00005177 0.09305230 + layer.4.output 0.00947078 198.52160395 + ------------------------------------------------------------------------------------- + TOTAL 0.04483982 107.11108591 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 58208 +BPFP 0.1911 bits/point +EBPFP 0.3821 equivalent bits/point +MSE 107.111086 +---------------------- -------------------------------------------------------- +Time: 3.610s Load: 0.010s, Pack+Encode: 2.093s, Decode+Unpack: 1.508s +---------------------- -------------------------------------------------------- +💾 Converting with 107.1111 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,700B, BPFP=0.2498 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,084B, BPFP=0.3233 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,304B, BPFP=0.3350 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,820B, BPFP=0.3093 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,916B, BPFP=0.2613 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,100B, BPFP=0.2710 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,520B, BPFP=0.2402 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,820B, BPFP=0.3093 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,848B, BPFP=0.3639 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,908B, BPFP=0.3140 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,464B, BPFP=0.0187 +⌛️ [2/4] FRONTEND: Frontend time: 2.051s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.535s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13622742 399.12731186 + layer.0.v_cache 0.00001380 0.03672135 + layer.1.k_cache 0.61520739 24.55928797 + layer.1.v_cache 0.00000584 0.01466034 + layer.2.k_cache 0.00900942 2.78018438 + layer.2.v_cache 0.00001916 0.04514611 + layer.3.k_cache 0.02667915 13.08790059 + layer.3.v_cache 0.00001881 0.05251688 + layer.4.k_cache 0.00071466 1.13926977 + layer.4.v_cache 0.00005760 0.09187941 + layer.4.output 0.04972813 184.73076105 + ------------------------------------------------------------------------------------- + TOTAL 0.06682648 102.00295330 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 58484 +BPFP 0.1828 bits/point +EBPFP 0.3657 equivalent bits/point +MSE 102.002953 +---------------------- -------------------------------------------------------- +Time: 3.596s Load: 0.010s, Pack+Encode: 2.051s, Decode+Unpack: 1.535s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0030 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,796B, BPFP=0.2506 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,904B, BPFP=0.3085 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,468B, BPFP=0.3380 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,488B, BPFP=0.2868 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,116B, BPFP=0.2673 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,080B, BPFP=0.2655 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,808B, BPFP=0.2513 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,764B, BPFP=0.3012 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,012B, BPFP=0.3664 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,888B, BPFP=0.3077 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,552B, BPFP=0.0191 +⌛️ [2/4] FRONTEND: Frontend time: 2.116s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.501s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11478914 397.90264423 + layer.0.v_cache 0.00001377 0.03667097 + layer.1.k_cache 0.64218915 25.24939414 + layer.1.v_cache 0.00000556 0.01421677 + layer.2.k_cache 0.00898698 2.90073683 + layer.2.v_cache 0.00001980 0.04540224 + layer.3.k_cache 0.05442990 13.04039588 + layer.3.v_cache 0.00001964 0.05516778 + layer.4.k_cache 0.00070461 1.16413966 + layer.4.v_cache 0.00005015 0.09161320 + layer.4.output 0.05017195 181.65840002 + ------------------------------------------------------------------------------------- + TOTAL 0.06896543 100.71230482 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 58876 +BPFP 0.1810 bits/point +EBPFP 0.3620 equivalent bits/point +MSE 100.712305 +---------------------- -------------------------------------------------------- +Time: 3.628s Load: 0.011s, Pack+Encode: 2.116s, Decode+Unpack: 1.501s +---------------------- -------------------------------------------------------- +💾 Converting with 100.7123 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,500B, BPFP=0.2653 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,696B, BPFP=0.3358 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,360B, BPFP=0.3750 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,828B, BPFP=0.3436 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,552B, BPFP=0.2684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,628B, BPFP=0.3318 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,096B, BPFP=0.2415 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,400B, BPFP=0.3184 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,428B, BPFP=0.3790 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,180B, BPFP=0.3644 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,392B, BPFP=0.0201 +⌛️ [2/4] FRONTEND: Frontend time: 2.102s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.549s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14755792 398.98755896 + layer.0.v_cache 0.00001375 0.03651319 + layer.1.k_cache 0.51940866 25.89671285 + layer.1.v_cache 0.00000586 0.01483885 + layer.2.k_cache 0.00816260 2.94380159 + layer.2.v_cache 0.00001900 0.04517427 + layer.3.k_cache 0.02921419 13.14777141 + layer.3.v_cache 0.00001979 0.05256952 + layer.4.k_cache 0.00071017 1.14318249 + layer.4.v_cache 0.00005507 0.09104195 + layer.4.output 0.00937099 209.73202493 + ------------------------------------------------------------------------------------- + TOTAL 0.04533906 112.38137292 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 57060 +BPFP 0.1979 bits/point +EBPFP 0.3958 equivalent bits/point +MSE 112.381373 +---------------------- -------------------------------------------------------- +Time: 3.660s Load: 0.009s, Pack+Encode: 2.102s, Decode+Unpack: 1.549s +---------------------- -------------------------------------------------------- +💾 Converting with 112.3814 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,440B, BPFP=0.2598 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,228B, BPFP=0.3645 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,440B, BPFP=0.3769 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,744B, BPFP=0.3361 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,596B, BPFP=0.2690 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,804B, BPFP=0.3397 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,188B, BPFP=0.2451 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,904B, BPFP=0.3455 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,692B, BPFP=0.3916 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,260B, BPFP=0.3663 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,476B, BPFP=0.0207 +⌛️ [2/4] FRONTEND: Frontend time: 2.269s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.580s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12840328 396.68647589 + layer.0.v_cache 0.00001372 0.03579977 + layer.1.k_cache 0.60081110 25.02695605 + layer.1.v_cache 0.00000562 0.01413446 + layer.2.k_cache 0.00586274 3.01292711 + layer.2.v_cache 0.00001839 0.04383589 + layer.3.k_cache 0.02491910 12.84513529 + layer.3.v_cache 0.00001957 0.05169314 + layer.4.k_cache 0.00072047 1.12249413 + layer.4.v_cache 0.00005229 0.09195063 + layer.4.output 0.01050496 208.17536116 + ------------------------------------------------------------------------------------- + TOTAL 0.04908006 111.53876061 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 58772 +BPFP 0.2023 bits/point +EBPFP 0.4046 equivalent bits/point +MSE 111.538761 +---------------------- -------------------------------------------------------- +Time: 3.861s Load: 0.011s, Pack+Encode: 2.269s, Decode+Unpack: 1.580s +---------------------- -------------------------------------------------------- +💾 Converting with 111.5388 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,876B, BPFP=0.2692 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,080B, BPFP=0.3357 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,952B, BPFP=0.3286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,692B, BPFP=0.3143 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,696B, BPFP=0.2593 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,528B, BPFP=0.3052 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,360B, BPFP=0.2407 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,752B, BPFP=0.3176 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,512B, BPFP=0.3595 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,576B, BPFP=0.3079 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,344B, BPFP=0.0185 +⌛️ [2/4] FRONTEND: Frontend time: 2.064s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.548s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14595435 399.85697328 + layer.0.v_cache 0.00001364 0.03668674 + layer.1.k_cache 0.59759904 24.41545660 + layer.1.v_cache 0.00000566 0.01417284 + layer.2.k_cache 0.01409306 2.93779849 + layer.2.v_cache 0.00002047 0.04504389 + layer.3.k_cache 0.06079736 13.33624893 + layer.3.v_cache 0.00002215 0.05371084 + layer.4.k_cache 0.00070776 1.13956034 + layer.4.v_cache 0.00005221 0.09274950 + layer.4.output 0.00769303 196.39257320 + ------------------------------------------------------------------------------------- + TOTAL 0.05135982 106.86331846 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 57368 +BPFP 0.1863 bits/point +EBPFP 0.3726 equivalent bits/point +MSE 106.863318 +---------------------- -------------------------------------------------------- +Time: 3.621s Load: 0.010s, Pack+Encode: 2.064s, Decode+Unpack: 1.548s +---------------------- -------------------------------------------------------- +💾 Converting with 106.8633 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 315, 128) +Output shape: (1, 315, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.output: torch.Size([1, 315, 3584]) -> torch.Size([1, 1, 315, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,796B, BPFP=0.2379 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,884B, BPFP=0.2919 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,020B, BPFP=0.2986 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,868B, BPFP=0.2911 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,376B, BPFP=0.2171 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,224B, BPFP=0.2591 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,084B, BPFP=0.2026 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,148B, BPFP=0.3050 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,356B, BPFP=0.3153 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,640B, BPFP=0.2798 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,664B, BPFP=0.0189 +⌛️ [2/4] FRONTEND: Frontend time: 2.337s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.522s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15500355 403.92279266 + layer.0.v_cache 0.00001366 0.03814536 + layer.1.k_cache 0.68167124 24.12823041 + layer.1.v_cache 0.00000571 0.01508916 + layer.2.k_cache 0.01007247 3.11080109 + layer.2.v_cache 0.00001936 0.04746156 + layer.3.k_cache 0.03272899 13.32391028 + layer.3.v_cache 0.00002006 0.05465245 + layer.4.k_cache 0.00071175 1.19953080 + layer.4.v_cache 0.00005427 0.09220587 + layer.4.output 0.04584261 172.40962302 + ------------------------------------------------------------------------------------- + TOTAL 0.07065879 97.22354004 + (elements=2,741,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2741760 +Total Bytes 57060 +BPFP 0.1665 bits/point +EBPFP 0.3330 equivalent bits/point +MSE 97.223540 +---------------------- -------------------------------------------------------- +Time: 3.869s Load: 0.010s, Pack+Encode: 2.337s, Decode+Unpack: 1.522s +---------------------- -------------------------------------------------------- +💾 Converting with 97.2235 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,556B, BPFP=0.2627 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,652B, BPFP=0.3835 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,388B, BPFP=0.3683 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,640B, BPFP=0.3252 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,436B, BPFP=0.2558 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,440B, BPFP=0.3137 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,352B, BPFP=0.2509 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,904B, BPFP=0.3404 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,664B, BPFP=0.3842 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,992B, BPFP=0.3455 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,592B, BPFP=0.0213 +⌛️ [2/4] FRONTEND: Frontend time: 2.036s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.458s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12889319 397.65728782 + layer.0.v_cache 0.00001350 0.03685810 + layer.1.k_cache 0.58775566 24.81356485 + layer.1.v_cache 0.00000553 0.01388402 + layer.2.k_cache 0.00723753 2.87296354 + layer.2.v_cache 0.00001924 0.04351073 + layer.3.k_cache 0.01761036 12.95236015 + layer.3.v_cache 0.00001847 0.05215821 + layer.4.k_cache 0.00071814 1.10735239 + layer.4.v_cache 0.00005270 0.08846047 + layer.4.output 0.01035140 205.08587572 + ------------------------------------------------------------------------------------- + TOTAL 0.04792848 110.30820767 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 58616 +BPFP 0.1988 bits/point +EBPFP 0.3976 equivalent bits/point +MSE 110.308208 +---------------------- -------------------------------------------------------- +Time: 3.503s Load: 0.008s, Pack+Encode: 2.036s, Decode+Unpack: 1.458s +---------------------- -------------------------------------------------------- +💾 Converting with 110.3082 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,860B, BPFP=0.2674 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,232B, BPFP=0.3429 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,980B, BPFP=0.3290 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,576B, BPFP=0.3068 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,696B, BPFP=0.2584 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,508B, BPFP=0.3030 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,296B, BPFP=0.2364 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,712B, BPFP=0.3143 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,436B, BPFP=0.3541 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,480B, BPFP=0.3015 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,284B, BPFP=0.0180 +⌛️ [2/4] FRONTEND: Frontend time: 2.366s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.603s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13033964 402.53243288 + layer.0.v_cache 0.00001368 0.03648673 + layer.1.k_cache 0.63029007 24.31679309 + layer.1.v_cache 0.00000551 0.01411010 + layer.2.k_cache 0.00637632 2.82071675 + layer.2.v_cache 0.00001960 0.04374489 + layer.3.k_cache 0.02368569 13.18914021 + layer.3.v_cache 0.00001974 0.05083499 + layer.4.k_cache 0.00068050 1.11776099 + layer.4.v_cache 0.00005209 0.08997003 + layer.4.output 0.00765416 195.69058099 + ------------------------------------------------------------------------------------- + TOTAL 0.04970952 106.70859162 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 57060 +BPFP 0.1847 bits/point +EBPFP 0.3693 equivalent bits/point +MSE 106.708592 +---------------------- -------------------------------------------------------- +Time: 3.978s Load: 0.009s, Pack+Encode: 2.366s, Decode+Unpack: 1.603s +---------------------- -------------------------------------------------------- +💾 Converting with 106.7086 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,356B, BPFP=0.2502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,312B, BPFP=0.3626 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,256B, BPFP=0.3594 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,996B, BPFP=0.3444 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,824B, BPFP=0.2771 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,592B, BPFP=0.3212 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,408B, BPFP=0.2532 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,184B, BPFP=0.3552 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,452B, BPFP=0.3706 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,372B, BPFP=0.3660 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,472B, BPFP=0.0203 +⌛️ [2/4] FRONTEND: Frontend time: 2.022s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.508s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15644295 395.31991039 + layer.0.v_cache 0.00001405 0.03677817 + layer.1.k_cache 0.56163917 25.51522648 + layer.1.v_cache 0.00000554 0.01451537 + layer.2.k_cache 0.01059573 2.86414562 + layer.2.v_cache 0.00001807 0.04437273 + layer.3.k_cache 0.04944291 12.99449786 + layer.3.v_cache 0.00001950 0.05482598 + layer.4.k_cache 0.00071267 1.10887584 + layer.4.v_cache 0.00005817 0.09401116 + layer.4.output 0.00956290 204.33109572 + ------------------------------------------------------------------------------------- + TOTAL 0.04975818 109.90381351 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 59224 +BPFP 0.2001 bits/point +EBPFP 0.4002 equivalent bits/point +MSE 109.903814 +---------------------- -------------------------------------------------------- +Time: 3.540s Load: 0.009s, Pack+Encode: 2.022s, Decode+Unpack: 1.508s +---------------------- -------------------------------------------------------- +💾 Converting with 109.9038 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 312, 128) +Output shape: (1, 312, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.output: torch.Size([1, 312, 3584]) -> torch.Size([1, 1, 312, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,968B, BPFP=0.2488 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,652B, BPFP=0.3331 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,248B, BPFP=0.3129 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,872B, BPFP=0.2941 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,548B, BPFP=0.2278 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,460B, BPFP=0.2734 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,344B, BPFP=0.2175 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,400B, BPFP=0.3205 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,620B, BPFP=0.3315 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,704B, BPFP=0.2857 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,212B, BPFP=0.0230 +⌛️ [2/4] FRONTEND: Frontend time: 2.099s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.618s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14984696 406.36318109 + layer.0.v_cache 0.00001474 0.03913899 + layer.1.k_cache 0.67859346 24.45313126 + layer.1.v_cache 0.00000581 0.01536145 + layer.2.k_cache 0.01626646 3.11766248 + layer.2.v_cache 0.00001951 0.04763992 + layer.3.k_cache 0.02914289 13.44189453 + layer.3.v_cache 0.00001919 0.05602047 + layer.4.k_cache 0.00071665 1.22055631 + layer.4.v_cache 0.00005096 0.09397023 + layer.4.output 0.04642455 174.10326522 + ------------------------------------------------------------------------------------- + TOTAL 0.07056756 98.09243608 + (elements=2,715,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2715648 +Total Bytes 60028 +BPFP 0.1768 bits/point +EBPFP 0.3537 equivalent bits/point +MSE 98.092436 +---------------------- -------------------------------------------------------- +Time: 3.727s Load: 0.010s, Pack+Encode: 2.099s, Decode+Unpack: 1.618s +---------------------- -------------------------------------------------------- +💾 Converting with 98.0924 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,632B, BPFP=0.2691 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,408B, BPFP=0.3141 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,480B, BPFP=0.3764 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,680B, BPFP=0.3299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,932B, BPFP=0.2865 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,712B, BPFP=0.3318 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,060B, BPFP=0.2358 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,912B, BPFP=0.3434 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,760B, BPFP=0.3927 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,100B, BPFP=0.3543 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,472B, BPFP=0.0205 +⌛️ [2/4] FRONTEND: Frontend time: 2.085s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.557s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15823836 399.80146375 + layer.0.v_cache 0.00001502 0.03695822 + layer.1.k_cache 0.53668020 25.02229576 + layer.1.v_cache 0.00000576 0.01462625 + layer.2.k_cache 0.01018995 3.02346836 + layer.2.v_cache 0.00002218 0.04618491 + layer.3.k_cache 0.05410197 13.13486455 + layer.3.v_cache 0.00002001 0.05385964 + layer.4.k_cache 0.00069143 1.14624897 + layer.4.v_cache 0.00005304 0.09343829 + layer.4.output 0.01041225 206.63150558 + ------------------------------------------------------------------------------------- + TOTAL 0.04899433 111.10552634 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 58148 +BPFP 0.1987 bits/point +EBPFP 0.3974 equivalent bits/point +MSE 111.105526 +---------------------- -------------------------------------------------------- +Time: 3.652s Load: 0.009s, Pack+Encode: 2.085s, Decode+Unpack: 1.557s +---------------------- -------------------------------------------------------- +💾 Converting with 111.1055 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 329, 128) +Output shape: (1, 329, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.output: torch.Size([1, 329, 3584]) -> torch.Size([1, 1, 329, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,768B, BPFP=0.2264 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,432B, BPFP=0.3530 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,860B, BPFP=0.3258 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,980B, BPFP=0.3315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,388B, BPFP=0.2559 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,752B, BPFP=0.3207 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,924B, BPFP=0.2339 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,984B, BPFP=0.3317 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,984B, BPFP=0.3792 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,076B, BPFP=0.3361 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,636B, BPFP=0.0179 +⌛️ [2/4] FRONTEND: Frontend time: 2.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.742s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11391156 384.86426672 + layer.0.v_cache 0.00001522 0.03812088 + layer.1.k_cache 0.63933222 24.31298531 + layer.1.v_cache 0.00000652 0.01604127 + layer.2.k_cache 0.01643764 2.89679215 + layer.2.v_cache 0.00001873 0.04573073 + layer.3.k_cache 0.02826017 12.60270083 + layer.3.v_cache 0.00001876 0.05343508 + layer.4.k_cache 0.00074607 1.17569536 + layer.4.v_cache 0.00005483 0.09462272 + layer.4.output 3.44632065 163.54315024 + ------------------------------------------------------------------------------------- + TOTAL 1.46606155 92.40602604 + (elements=2,863,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2863616 +Total Bytes 67784 +BPFP 0.1894 bits/point +EBPFP 0.3787 equivalent bits/point +MSE 92.406026 +---------------------- -------------------------------------------------------- +Time: 4.247s Load: 0.011s, Pack+Encode: 2.494s, Decode+Unpack: 1.742s +---------------------- -------------------------------------------------------- +💾 Converting with 92.4060 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,536B, BPFP=0.2540 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,428B, BPFP=0.3600 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,104B, BPFP=0.3418 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,620B, BPFP=0.3147 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,640B, BPFP=0.2599 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,640B, BPFP=0.3159 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,264B, BPFP=0.2388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,080B, BPFP=0.3405 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,396B, BPFP=0.3582 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,700B, BPFP=0.3192 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,248B, BPFP=0.0180 +⌛️ [2/4] FRONTEND: Frontend time: 2.273s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.503s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13536828 404.97373432 + layer.0.v_cache 0.00001433 0.03677995 + layer.1.k_cache 0.57185238 23.99423688 + layer.1.v_cache 0.00000558 0.01418137 + layer.2.k_cache 0.00766427 2.98018523 + layer.2.v_cache 0.00001891 0.04487839 + layer.3.k_cache 0.04143807 13.71841048 + layer.3.v_cache 0.00001947 0.05240133 + layer.4.k_cache 0.00070404 1.12175278 + layer.4.v_cache 0.00005498 0.09150622 + layer.4.output 0.00948056 199.08739759 + ------------------------------------------------------------------------------------- + TOTAL 0.04844142 108.27293236 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 57656 +BPFP 0.1899 bits/point +EBPFP 0.3799 equivalent bits/point +MSE 108.272932 +---------------------- -------------------------------------------------------- +Time: 3.788s Load: 0.012s, Pack+Encode: 2.273s, Decode+Unpack: 1.503s +---------------------- -------------------------------------------------------- +💾 Converting with 108.2729 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,564B, BPFP=0.2584 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,132B, BPFP=0.3471 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,252B, BPFP=0.3539 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,688B, BPFP=0.3220 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,732B, BPFP=0.2679 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,476B, BPFP=0.3100 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,332B, BPFP=0.2452 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,112B, BPFP=0.3460 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,724B, BPFP=0.3807 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,816B, BPFP=0.3293 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,432B, BPFP=0.0197 +⌛️ [2/4] FRONTEND: Frontend time: 2.049s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.556s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15132187 396.05471581 + layer.0.v_cache 0.00001456 0.03651946 + layer.1.k_cache 0.61798344 24.77416815 + layer.1.v_cache 0.00000569 0.01407002 + layer.2.k_cache 0.01686249 2.98487677 + layer.2.v_cache 0.00001898 0.04560308 + layer.3.k_cache 0.04401700 13.63893194 + layer.3.v_cache 0.00001935 0.05229874 + layer.4.k_cache 0.00071189 1.12319360 + layer.4.v_cache 0.00005111 0.09148001 + layer.4.output 0.01058244 201.36340256 + ------------------------------------------------------------------------------------- + TOTAL 0.05324020 108.72703974 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 58260 +BPFP 0.1940 bits/point +EBPFP 0.3880 equivalent bits/point +MSE 108.727040 +---------------------- -------------------------------------------------------- +Time: 3.615s Load: 0.010s, Pack+Encode: 2.049s, Decode+Unpack: 1.556s +---------------------- -------------------------------------------------------- +💾 Converting with 108.7270 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,292B, BPFP=0.2430 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,092B, BPFP=0.3449 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,136B, BPFP=0.3474 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,924B, BPFP=0.3354 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,792B, BPFP=0.2713 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,432B, BPFP=0.3075 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,356B, BPFP=0.2466 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,708B, BPFP=0.3231 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,644B, BPFP=0.3761 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,408B, BPFP=0.3628 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,388B, BPFP=0.0193 +⌛️ [2/4] FRONTEND: Frontend time: 2.066s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.573s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12581717 385.50506680 + layer.0.v_cache 0.00001375 0.03764492 + layer.1.k_cache 0.57814745 24.38292572 + layer.1.v_cache 0.00000592 0.01462907 + layer.2.k_cache 0.00733882 2.78249812 + layer.2.v_cache 0.00002058 0.04625321 + layer.3.k_cache 0.02353939 13.53292183 + layer.3.v_cache 0.00002077 0.05349845 + layer.4.k_cache 0.00069182 1.16505841 + layer.4.v_cache 0.00005027 0.09117688 + layer.4.output 0.01015823 201.33318776 + ------------------------------------------------------------------------------------- + TOTAL 0.04745609 108.05552869 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 58172 +BPFP 0.1937 bits/point +EBPFP 0.3874 equivalent bits/point +MSE 108.055529 +---------------------- -------------------------------------------------------- +Time: 3.649s Load: 0.010s, Pack+Encode: 2.066s, Decode+Unpack: 1.573s +---------------------- -------------------------------------------------------- +💾 Converting with 108.0555 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,232B, BPFP=0.2413 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,244B, BPFP=0.3561 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,264B, BPFP=0.3572 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,844B, BPFP=0.3333 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,724B, BPFP=0.2694 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,804B, BPFP=0.3310 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,368B, BPFP=0.2491 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,168B, BPFP=0.3517 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,956B, BPFP=0.3967 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,908B, BPFP=0.3369 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,580B, BPFP=0.0210 +⌛️ [2/4] FRONTEND: Frontend time: 2.114s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.624s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13625057 393.37283303 + layer.0.v_cache 0.00001362 0.03656532 + layer.1.k_cache 0.57512581 25.59761526 + layer.1.v_cache 0.00000584 0.01468482 + layer.2.k_cache 0.00872732 3.01637680 + layer.2.v_cache 0.00001851 0.04507443 + layer.3.k_cache 0.04362779 12.65478427 + layer.3.v_cache 0.00001925 0.05564785 + layer.4.k_cache 0.00073619 1.17178077 + layer.4.v_cache 0.00005148 0.09478457 + layer.4.output 0.00783937 202.83058525 + ------------------------------------------------------------------------------------- + TOTAL 0.04820306 109.16907317 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 59092 +BPFP 0.1982 bits/point +EBPFP 0.3964 equivalent bits/point +MSE 109.169073 +---------------------- -------------------------------------------------------- +Time: 3.749s Load: 0.010s, Pack+Encode: 2.114s, Decode+Unpack: 1.624s +---------------------- -------------------------------------------------------- +💾 Converting with 109.1691 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,016B, BPFP=0.2578 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,900B, BPFP=0.3032 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,368B, BPFP=0.3273 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,652B, BPFP=0.2905 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,928B, BPFP=0.2533 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,352B, BPFP=0.2751 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,628B, BPFP=0.2379 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,000B, BPFP=0.3084 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,008B, BPFP=0.3602 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,808B, BPFP=0.2985 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,232B, BPFP=0.0164 +⌛️ [2/4] FRONTEND: Frontend time: 2.072s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.517s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16730670 402.22882401 + layer.0.v_cache 0.00001388 0.03668256 + layer.1.k_cache 0.64158324 24.52775815 + layer.1.v_cache 0.00000571 0.01459379 + layer.2.k_cache 0.02207795 2.85687176 + layer.2.v_cache 0.00001922 0.04416667 + layer.3.k_cache 0.04503629 13.22301443 + layer.3.v_cache 0.00001955 0.05494002 + layer.4.k_cache 0.00070478 1.15815293 + layer.4.v_cache 0.00005260 0.09211344 + layer.4.output 0.04763387 178.64780604 + ------------------------------------------------------------------------------------- + TOTAL 0.07119159 99.69245647 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 58892 +BPFP 0.1781 bits/point +EBPFP 0.3561 equivalent bits/point +MSE 99.692456 +---------------------- -------------------------------------------------------- +Time: 3.600s Load: 0.011s, Pack+Encode: 2.072s, Decode+Unpack: 1.517s +---------------------- -------------------------------------------------------- +💾 Converting with 99.6925 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,656B, BPFP=0.2626 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,976B, BPFP=0.3371 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,368B, BPFP=0.3592 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,668B, BPFP=0.3197 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,708B, BPFP=0.2656 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,036B, BPFP=0.3405 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,340B, BPFP=0.2448 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,124B, BPFP=0.3454 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,632B, BPFP=0.3741 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,972B, BPFP=0.3369 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,352B, BPFP=0.0190 +⌛️ [2/4] FRONTEND: Frontend time: 2.078s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.558s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15225300 405.96099391 + layer.0.v_cache 0.00001427 0.03594528 + layer.1.k_cache 0.52731114 24.95428489 + layer.1.v_cache 0.00000556 0.01428106 + layer.2.k_cache 0.01483519 2.97217216 + layer.2.v_cache 0.00001864 0.04635697 + layer.3.k_cache 0.04699480 14.20908397 + layer.3.v_cache 0.00001968 0.05401568 + layer.4.k_cache 0.00070142 1.12588798 + layer.4.v_cache 0.00005177 0.09299677 + layer.4.output 0.01077793 200.64541968 + ------------------------------------------------------------------------------------- + TOTAL 0.04809712 109.05787979 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 58832 +BPFP 0.1952 bits/point +EBPFP 0.3904 equivalent bits/point +MSE 109.057880 +---------------------- -------------------------------------------------------- +Time: 3.645s Load: 0.008s, Pack+Encode: 2.078s, Decode+Unpack: 1.558s +---------------------- -------------------------------------------------------- +💾 Converting with 109.0579 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,620B, BPFP=0.2606 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,396B, BPFP=0.3608 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,284B, BPFP=0.3545 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,876B, BPFP=0.3315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,536B, BPFP=0.2559 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,856B, BPFP=0.3303 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,504B, BPFP=0.2541 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,772B, BPFP=0.3256 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,436B, BPFP=0.3630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,720B, BPFP=0.3227 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,404B, BPFP=0.0194 +⌛️ [2/4] FRONTEND: Frontend time: 2.125s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.527s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15700974 403.05590027 + layer.0.v_cache 0.00001374 0.03654423 + layer.1.k_cache 0.52255304 24.04341825 + layer.1.v_cache 0.00000566 0.01427296 + layer.2.k_cache 0.01360668 2.99969042 + layer.2.v_cache 0.00001926 0.04570852 + layer.3.k_cache 0.08085050 12.80850032 + layer.3.v_cache 0.00001943 0.05489045 + layer.4.k_cache 0.00071029 1.10214377 + layer.4.v_cache 0.00005450 0.09284837 + layer.4.output 0.01014708 200.66949781 + ------------------------------------------------------------------------------------- + TOTAL 0.04975720 108.76120013 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 58404 +BPFP 0.1938 bits/point +EBPFP 0.3876 equivalent bits/point +MSE 108.761200 +---------------------- -------------------------------------------------------- +Time: 3.663s Load: 0.010s, Pack+Encode: 2.125s, Decode+Unpack: 1.527s +---------------------- -------------------------------------------------------- +💾 Converting with 108.7612 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,876B, BPFP=0.2673 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,868B, BPFP=0.3217 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,736B, BPFP=0.3145 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,708B, BPFP=0.3129 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,516B, BPFP=0.2476 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,496B, BPFP=0.3013 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,168B, BPFP=0.2285 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,620B, BPFP=0.3081 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,516B, BPFP=0.3572 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,640B, BPFP=0.3092 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,288B, BPFP=0.0179 +⌛️ [2/4] FRONTEND: Frontend time: 2.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.799s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16436556 403.39369518 + layer.0.v_cache 0.00001441 0.03647006 + layer.1.k_cache 0.62741763 24.76960321 + layer.1.v_cache 0.00000560 0.01441179 + layer.2.k_cache 0.01487184 2.88347254 + layer.2.v_cache 0.00002167 0.04573902 + layer.3.k_cache 0.02142572 12.98111294 + layer.3.v_cache 0.00001943 0.05359542 + layer.4.k_cache 0.00070776 1.12558326 + layer.4.v_cache 0.00005251 0.09044738 + layer.4.output 0.00894435 195.01495927 + ------------------------------------------------------------------------------------- + TOTAL 0.05244192 106.49993210 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 56432 +BPFP 0.1820 bits/point +EBPFP 0.3640 equivalent bits/point +MSE 106.499932 +---------------------- -------------------------------------------------------- +Time: 4.037s Load: 0.010s, Pack+Encode: 2.228s, Decode+Unpack: 1.799s +---------------------- -------------------------------------------------------- +💾 Converting with 106.4999 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,492B, BPFP=0.2498 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,036B, BPFP=0.3356 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,164B, BPFP=0.3427 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,784B, BPFP=0.3216 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,852B, BPFP=0.2698 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,000B, BPFP=0.2780 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,552B, BPFP=0.2531 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,976B, BPFP=0.3323 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,744B, BPFP=0.3750 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,604B, BPFP=0.3116 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,236B, BPFP=0.0178 +⌛️ [2/4] FRONTEND: Frontend time: 2.457s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.819s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11705062 399.02694061 + layer.0.v_cache 0.00001445 0.03679411 + layer.1.k_cache 0.56620099 25.04473768 + layer.1.v_cache 0.00000566 0.01437784 + layer.2.k_cache 0.01299965 2.91734493 + layer.2.v_cache 0.00001972 0.04407309 + layer.3.k_cache 0.05069733 13.44245058 + layer.3.v_cache 0.00001944 0.05267389 + layer.4.k_cache 0.00068883 1.11701124 + layer.4.v_cache 0.00005144 0.09079742 + layer.4.output 0.00838925 197.79054398 + ------------------------------------------------------------------------------------- + TOTAL 0.04743958 107.43064760 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 57440 +BPFP 0.1879 bits/point +EBPFP 0.3758 equivalent bits/point +MSE 107.430648 +---------------------- -------------------------------------------------------- +Time: 4.287s Load: 0.010s, Pack+Encode: 2.457s, Decode+Unpack: 1.819s +---------------------- -------------------------------------------------------- +💾 Converting with 107.4306 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,884B, BPFP=0.2697 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,844B, BPFP=0.3227 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,968B, BPFP=0.3295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,956B, BPFP=0.3288 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,780B, BPFP=0.2639 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,520B, BPFP=0.3048 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,304B, BPFP=0.2376 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,800B, BPFP=0.3202 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,728B, BPFP=0.3715 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,724B, BPFP=0.3160 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,352B, BPFP=0.0186 +⌛️ [2/4] FRONTEND: Frontend time: 2.236s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.495s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14370025 403.50676347 + layer.0.v_cache 0.00001382 0.03673072 + layer.1.k_cache 0.61448195 24.55634731 + layer.1.v_cache 0.00000570 0.01431568 + layer.2.k_cache 0.01208573 2.82041861 + layer.2.v_cache 0.00001823 0.04428721 + layer.3.k_cache 0.01319265 13.85799470 + layer.3.v_cache 0.00002242 0.05244209 + layer.4.k_cache 0.00071920 1.14373488 + layer.4.v_cache 0.00005145 0.09183349 + layer.4.output 0.00992508 196.41437405 + ------------------------------------------------------------------------------------- + TOTAL 0.05022159 107.11914627 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 57860 +BPFP 0.1879 bits/point +EBPFP 0.3758 equivalent bits/point +MSE 107.119146 +---------------------- -------------------------------------------------------- +Time: 3.743s Load: 0.011s, Pack+Encode: 2.236s, Decode+Unpack: 1.495s +---------------------- -------------------------------------------------------- +💾 Converting with 107.1191 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 259, 128) +Output shape: (1, 259, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.output: torch.Size([1, 259, 3584]) -> torch.Size([1, 1, 259, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,708B, BPFP=0.2237 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,096B, BPFP=0.3678 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,156B, BPFP=0.3714 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,560B, BPFP=0.3354 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,624B, BPFP=0.2790 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,404B, BPFP=0.3260 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,844B, BPFP=0.2319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,548B, BPFP=0.3347 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,532B, BPFP=0.3941 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,280B, BPFP=0.3789 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,568B, BPFP=0.0221 +⌛️ [2/4] FRONTEND: Frontend time: 1.993s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.543s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13349269 384.06654199 + layer.0.v_cache 0.00001436 0.03780731 + layer.1.k_cache 0.50562990 25.46917607 + layer.1.v_cache 0.00000590 0.01514126 + layer.2.k_cache 0.00820730 2.77812425 + layer.2.v_cache 0.00001794 0.04465791 + layer.3.k_cache 0.03685973 13.34371041 + layer.3.v_cache 0.00002061 0.05134222 + layer.4.k_cache 0.00069539 1.15764026 + layer.4.v_cache 0.00005422 0.09356005 + layer.4.output 0.01018517 214.61372725 + ------------------------------------------------------------------------------------- + TOTAL 0.04448790 113.49139956 + (elements=2,254,336) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2254336 +Total Bytes 56320 +BPFP 0.1999 bits/point +EBPFP 0.3997 equivalent bits/point +MSE 113.491400 +---------------------- -------------------------------------------------------- +Time: 3.544s Load: 0.008s, Pack+Encode: 1.993s, Decode+Unpack: 1.543s +---------------------- -------------------------------------------------------- +💾 Converting with 113.4914 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,472B, BPFP=0.2588 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,404B, BPFP=0.3706 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,436B, BPFP=0.3725 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,644B, BPFP=0.3266 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,852B, BPFP=0.2808 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,588B, BPFP=0.3234 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,276B, BPFP=0.2475 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,768B, BPFP=0.3338 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,728B, BPFP=0.3894 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,068B, BPFP=0.3512 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,644B, BPFP=0.0219 +⌛️ [2/4] FRONTEND: Frontend time: 2.175s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.598s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12649573 397.34649884 + layer.0.v_cache 0.00001373 0.03717131 + layer.1.k_cache 0.51574391 25.13158818 + layer.1.v_cache 0.00000589 0.01458043 + layer.2.k_cache 0.01194659 2.98110781 + layer.2.v_cache 0.00001984 0.04784910 + layer.3.k_cache 0.04962628 13.43674045 + layer.3.v_cache 0.00002059 0.05415155 + layer.4.k_cache 0.00068541 1.15274376 + layer.4.v_cache 0.00005704 0.09598311 + layer.4.output 0.00772497 205.85443122 + ------------------------------------------------------------------------------------- + TOTAL 0.04462881 110.66349606 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 58880 +BPFP 0.2004 bits/point +EBPFP 0.4009 equivalent bits/point +MSE 110.663496 +---------------------- -------------------------------------------------------- +Time: 3.783s Load: 0.010s, Pack+Encode: 2.175s, Decode+Unpack: 1.598s +---------------------- -------------------------------------------------------- +💾 Converting with 110.6635 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,456B, BPFP=0.2588 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,188B, BPFP=0.3594 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,428B, BPFP=0.3734 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,792B, BPFP=0.3364 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,764B, BPFP=0.2767 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,192B, BPFP=0.3016 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,048B, BPFP=0.2351 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,564B, BPFP=0.3232 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,488B, BPFP=0.3769 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,204B, BPFP=0.3604 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,508B, BPFP=0.0208 +⌛️ [2/4] FRONTEND: Frontend time: 2.102s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.554s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12210708 391.76748374 + layer.0.v_cache 0.00001342 0.03591920 + layer.1.k_cache 0.54942129 25.37122444 + layer.1.v_cache 0.00000547 0.01433213 + layer.2.k_cache 0.01053270 2.98376419 + layer.2.v_cache 0.00001820 0.04571263 + layer.3.k_cache 0.02850239 12.88651182 + layer.3.v_cache 0.00002491 0.05527327 + layer.4.k_cache 0.00067987 1.13476063 + layer.4.v_cache 0.00005179 0.09149827 + layer.4.output 0.01126663 206.58394185 + ------------------------------------------------------------------------------------- + TOTAL 0.04648373 110.61612196 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 57632 +BPFP 0.1969 bits/point +EBPFP 0.3938 equivalent bits/point +MSE 110.616122 +---------------------- -------------------------------------------------------- +Time: 3.665s Load: 0.010s, Pack+Encode: 2.102s, Decode+Unpack: 1.554s +---------------------- -------------------------------------------------------- +💾 Converting with 110.6161 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,876B, BPFP=0.2514 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,152B, BPFP=0.3172 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,304B, BPFP=0.3251 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,004B, BPFP=0.3096 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,864B, BPFP=0.2508 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,876B, BPFP=0.3030 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,580B, BPFP=0.2362 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,928B, BPFP=0.3057 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,940B, BPFP=0.3579 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,956B, BPFP=0.3071 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,460B, BPFP=0.0181 +⌛️ [2/4] FRONTEND: Frontend time: 2.086s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.517s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14443027 407.27931621 + layer.0.v_cache 0.00001375 0.03705597 + layer.1.k_cache 0.62837083 24.50232054 + layer.1.v_cache 0.00000570 0.01429850 + layer.2.k_cache 0.01557094 2.84573676 + layer.2.v_cache 0.00001871 0.04456348 + layer.3.k_cache 0.04676386 12.81967032 + layer.3.v_cache 0.00001981 0.05171402 + layer.4.k_cache 0.00070582 1.15009602 + layer.4.v_cache 0.00005437 0.08907103 + layer.4.output 0.04838977 179.25960632 + ------------------------------------------------------------------------------------- + TOTAL 0.06909897 100.21476983 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 59940 +BPFP 0.1818 bits/point +EBPFP 0.3636 equivalent bits/point +MSE 100.214770 +---------------------- -------------------------------------------------------- +Time: 3.613s Load: 0.011s, Pack+Encode: 2.086s, Decode+Unpack: 1.517s +---------------------- -------------------------------------------------------- +💾 Converting with 100.2148 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,880B, BPFP=0.2694 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,116B, BPFP=0.3377 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,076B, BPFP=0.3355 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,524B, BPFP=0.3050 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,672B, BPFP=0.2580 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,584B, BPFP=0.3083 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,272B, BPFP=0.2359 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,892B, BPFP=0.3253 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,652B, BPFP=0.3673 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,836B, BPFP=0.3222 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,368B, BPFP=0.0187 +⌛️ [2/4] FRONTEND: Frontend time: 2.513s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.742s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13293209 399.41179881 + layer.0.v_cache 0.00001388 0.03737379 + layer.1.k_cache 0.56267701 25.08262478 + layer.1.v_cache 0.00000566 0.01429167 + layer.2.k_cache 0.01145046 2.84633806 + layer.2.v_cache 0.00001921 0.04524499 + layer.3.k_cache 0.02674564 13.87197197 + layer.3.v_cache 0.00001926 0.05326810 + layer.4.k_cache 0.00069135 1.14079064 + layer.4.v_cache 0.00005493 0.09295697 + layer.4.output 0.00973303 196.40897905 + ------------------------------------------------------------------------------------- + TOTAL 0.04722004 106.90938312 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 57872 +BPFP 0.1880 bits/point +EBPFP 0.3759 equivalent bits/point +MSE 106.909383 +---------------------- -------------------------------------------------------- +Time: 4.266s Load: 0.010s, Pack+Encode: 2.513s, Decode+Unpack: 1.742s +---------------------- -------------------------------------------------------- +💾 Converting with 106.9094 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,696B, BPFP=0.2649 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,040B, BPFP=0.3407 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,216B, BPFP=0.3506 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,996B, BPFP=0.3382 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,692B, BPFP=0.2647 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,540B, BPFP=0.3125 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,440B, BPFP=0.2505 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,152B, BPFP=0.3470 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,592B, BPFP=0.3718 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,836B, BPFP=0.3292 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,492B, BPFP=0.0201 +⌛️ [2/4] FRONTEND: Frontend time: 2.313s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.593s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15182159 406.07409183 + layer.0.v_cache 0.00001420 0.03727857 + layer.1.k_cache 0.56224567 25.35319128 + layer.1.v_cache 0.00000588 0.01469254 + layer.2.k_cache 0.00652889 2.83604745 + layer.2.v_cache 0.00001844 0.04361632 + layer.3.k_cache 0.06011557 13.72256106 + layer.3.v_cache 0.00001921 0.05410559 + layer.4.k_cache 0.00068317 1.13893849 + layer.4.v_cache 0.00005096 0.09080606 + layer.4.output 0.01130176 200.65186630 + ------------------------------------------------------------------------------------- + TOTAL 0.05062446 109.05461137 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 58692 +BPFP 0.1947 bits/point +EBPFP 0.3895 equivalent bits/point +MSE 109.054611 +---------------------- -------------------------------------------------------- +Time: 3.916s Load: 0.010s, Pack+Encode: 2.313s, Decode+Unpack: 1.593s +---------------------- -------------------------------------------------------- +💾 Converting with 109.0546 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,816B, BPFP=0.2640 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,000B, BPFP=0.3289 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,708B, BPFP=0.3129 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,696B, BPFP=0.3123 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,556B, BPFP=0.2498 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,848B, BPFP=0.3206 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,292B, BPFP=0.2353 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,708B, BPFP=0.3129 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,472B, BPFP=0.3548 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,000B, BPFP=0.3289 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,308B, BPFP=0.0181 +⌛️ [2/4] FRONTEND: Frontend time: 2.274s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.565s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13053865 401.18338816 + layer.0.v_cache 0.00001414 0.03871215 + layer.1.k_cache 0.56363723 25.13103927 + layer.1.v_cache 0.00000590 0.01515301 + layer.2.k_cache 0.01033439 2.91585394 + layer.2.v_cache 0.00002015 0.04585098 + layer.3.k_cache 0.03655195 13.91607045 + layer.3.v_cache 0.00002026 0.05348019 + layer.4.k_cache 0.00068516 1.14338529 + layer.4.v_cache 0.00005378 0.09428870 + layer.4.output 0.00863618 195.04736842 + ------------------------------------------------------------------------------------- + TOTAL 0.04719499 106.46287065 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 57404 +BPFP 0.1851 bits/point +EBPFP 0.3703 equivalent bits/point +MSE 106.462871 +---------------------- -------------------------------------------------------- +Time: 3.849s Load: 0.010s, Pack+Encode: 2.274s, Decode+Unpack: 1.565s +---------------------- -------------------------------------------------------- +💾 Converting with 106.4629 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,680B, BPFP=0.2640 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,124B, BPFP=0.3454 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,160B, BPFP=0.3475 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,888B, BPFP=0.3321 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,756B, BPFP=0.2683 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,912B, BPFP=0.3335 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,572B, BPFP=0.2579 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,904B, BPFP=0.3330 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,572B, BPFP=0.3707 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,828B, BPFP=0.3287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,372B, BPFP=0.0191 +⌛️ [2/4] FRONTEND: Frontend time: 2.065s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.812s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15908699 403.80206453 + layer.0.v_cache 0.00001433 0.03660285 + layer.1.k_cache 0.53822481 25.22429208 + layer.1.v_cache 0.00000570 0.01460015 + layer.2.k_cache 0.01139215 2.92442988 + layer.2.v_cache 0.00001845 0.04531836 + layer.3.k_cache 0.02907990 13.25795264 + layer.3.v_cache 0.00001963 0.05489352 + layer.4.k_cache 0.00069750 1.16782630 + layer.4.v_cache 0.00005397 0.09215539 + layer.4.output 0.01069233 200.65802282 + ------------------------------------------------------------------------------------- + TOTAL 0.04784940 108.89566444 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 58768 +BPFP 0.1950 bits/point +EBPFP 0.3900 equivalent bits/point +MSE 108.895664 +---------------------- -------------------------------------------------------- +Time: 3.886s Load: 0.009s, Pack+Encode: 2.065s, Decode+Unpack: 1.812s +---------------------- -------------------------------------------------------- +💾 Converting with 108.8957 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,784B, BPFP=0.2614 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,868B, BPFP=0.3206 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,828B, BPFP=0.3184 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,348B, BPFP=0.2922 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,512B, BPFP=0.2465 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,432B, BPFP=0.2968 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,140B, BPFP=0.2262 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,924B, BPFP=0.3236 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,528B, BPFP=0.3566 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,756B, BPFP=0.3145 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,176B, BPFP=0.0170 +⌛️ [2/4] FRONTEND: Frontend time: 2.515s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.592s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16084659 405.79288680 + layer.0.v_cache 0.00001440 0.03655101 + layer.1.k_cache 0.59028764 24.77183266 + layer.1.v_cache 0.00000559 0.01459062 + layer.2.k_cache 0.00929020 2.75922783 + layer.2.v_cache 0.00002205 0.04671442 + layer.3.k_cache 0.04382113 13.00992099 + layer.3.v_cache 0.00002103 0.05387845 + layer.4.k_cache 0.00069271 1.13360617 + layer.4.v_cache 0.00005052 0.09158128 + layer.4.output 0.00984751 194.35617507 + ------------------------------------------------------------------------------------- + TOTAL 0.05141085 106.36494210 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 56296 +BPFP 0.1809 bits/point +EBPFP 0.3618 equivalent bits/point +MSE 106.364942 +---------------------- -------------------------------------------------------- +Time: 4.116s Load: 0.010s, Pack+Encode: 2.515s, Decode+Unpack: 1.592s +---------------------- -------------------------------------------------------- +💾 Converting with 106.3649 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,124B, BPFP=0.2237 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,108B, BPFP=0.3314 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,980B, BPFP=0.3244 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,788B, BPFP=0.3140 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,564B, BPFP=0.2476 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,196B, BPFP=0.3362 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,196B, BPFP=0.2276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,260B, BPFP=0.3396 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,640B, BPFP=0.3602 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,548B, BPFP=0.3010 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,192B, BPFP=0.0170 +⌛️ [2/4] FRONTEND: Frontend time: 2.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.515s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14211286 377.18286133 + layer.0.v_cache 0.00001516 0.03687588 + layer.1.k_cache 0.60291523 24.73338657 + layer.1.v_cache 0.00000570 0.01510910 + layer.2.k_cache 0.01041983 2.71257125 + layer.2.v_cache 0.00001835 0.04533685 + layer.3.k_cache 0.03896446 13.31467353 + layer.3.v_cache 0.00001926 0.05339774 + layer.4.k_cache 0.00071470 1.12926038 + layer.4.v_cache 0.00005236 0.09319180 + layer.4.output 0.00754456 193.02430556 + ------------------------------------------------------------------------------------- + TOTAL 0.04988528 104.14628255 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 57596 +BPFP 0.1838 bits/point +EBPFP 0.3676 equivalent bits/point +MSE 104.146283 +---------------------- -------------------------------------------------------- +Time: 3.776s Load: 0.009s, Pack+Encode: 2.251s, Decode+Unpack: 1.515s +---------------------- -------------------------------------------------------- +💾 Converting with 104.1463 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 296, 128) +Output shape: (1, 296, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.output: torch.Size([1, 296, 3584]) -> torch.Size([1, 1, 296, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,740B, BPFP=0.2502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,740B, BPFP=0.3030 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,392B, BPFP=0.3374 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,608B, BPFP=0.2960 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,992B, BPFP=0.2635 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,116B, BPFP=0.2701 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,652B, BPFP=0.2456 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,536B, BPFP=0.2922 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,972B, BPFP=0.3680 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,744B, BPFP=0.3032 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,712B, BPFP=0.0205 +⌛️ [2/4] FRONTEND: Frontend time: 2.053s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.580s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14616069 392.93237965 + layer.0.v_cache 0.00001366 0.03662069 + layer.1.k_cache 0.61022733 24.80583397 + layer.1.v_cache 0.00000576 0.01424024 + layer.2.k_cache 0.01378242 2.87955083 + layer.2.v_cache 0.00001907 0.04541979 + layer.3.k_cache 0.03184220 12.61850718 + layer.3.v_cache 0.00001934 0.05470246 + layer.4.k_cache 0.00072724 1.16086980 + layer.4.v_cache 0.00005080 0.09070346 + layer.4.output 0.04784788 183.47844776 + ------------------------------------------------------------------------------------- + TOTAL 0.06692845 101.11693896 + (elements=2,576,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2576384 +Total Bytes 58204 +BPFP 0.1807 bits/point +EBPFP 0.3615 equivalent bits/point +MSE 101.116939 +---------------------- -------------------------------------------------------- +Time: 3.641s Load: 0.009s, Pack+Encode: 2.053s, Decode+Unpack: 1.580s +---------------------- -------------------------------------------------------- +💾 Converting with 101.1169 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,872B, BPFP=0.2555 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,148B, BPFP=0.3224 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,372B, BPFP=0.3341 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,756B, BPFP=0.3018 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,000B, BPFP=0.2622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,208B, BPFP=0.2731 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,736B, BPFP=0.2483 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,864B, BPFP=0.3075 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,880B, BPFP=0.3607 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,764B, BPFP=0.3022 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,496B, BPFP=0.0187 +⌛️ [2/4] FRONTEND: Frontend time: 2.037s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.506s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15387250 399.49058830 + layer.0.v_cache 0.00001382 0.03725312 + layer.1.k_cache 0.63844893 24.46819782 + layer.1.v_cache 0.00000583 0.01495815 + layer.2.k_cache 0.01105453 2.99136967 + layer.2.v_cache 0.00001885 0.04562836 + layer.3.k_cache 0.03893430 13.01300499 + layer.3.v_cache 0.00001923 0.05518029 + layer.4.k_cache 0.00070257 1.16653084 + layer.4.v_cache 0.00005146 0.09165615 + layer.4.output 0.04873301 182.27657598 + ------------------------------------------------------------------------------------- + TOTAL 0.06966195 101.01825880 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 59096 +BPFP 0.1823 bits/point +EBPFP 0.3645 equivalent bits/point +MSE 101.018259 +---------------------- -------------------------------------------------------- +Time: 3.552s Load: 0.010s, Pack+Encode: 2.037s, Decode+Unpack: 1.506s +---------------------- -------------------------------------------------------- +💾 Converting with 101.0183 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,504B, BPFP=0.2597 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,956B, BPFP=0.3434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,236B, BPFP=0.3595 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,960B, BPFP=0.3436 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,812B, BPFP=0.2774 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,676B, BPFP=0.3273 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,360B, BPFP=0.2514 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,284B, BPFP=0.3623 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,496B, BPFP=0.3745 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,244B, BPFP=0.3600 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,484B, BPFP=0.0205 +⌛️ [2/4] FRONTEND: Frontend time: 2.108s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.693s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12629252 400.16146794 + layer.0.v_cache 0.00001451 0.03809394 + layer.1.k_cache 0.49097243 25.92846409 + layer.1.v_cache 0.00000577 0.01505583 + layer.2.k_cache 0.01060468 2.96874865 + layer.2.v_cache 0.00001909 0.04563563 + layer.3.k_cache 0.02748298 13.03045812 + layer.3.v_cache 0.00001876 0.05430354 + layer.4.k_cache 0.00069967 1.13573948 + layer.4.v_cache 0.00005414 0.09231072 + layer.4.output 0.01047103 205.09332169 + ------------------------------------------------------------------------------------- + TOTAL 0.04290952 110.53667822 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 59012 +BPFP 0.2001 bits/point +EBPFP 0.4003 equivalent bits/point +MSE 110.536678 +---------------------- -------------------------------------------------------- +Time: 3.811s Load: 0.010s, Pack+Encode: 2.108s, Decode+Unpack: 1.693s +---------------------- -------------------------------------------------------- +💾 Converting with 110.5367 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,324B, BPFP=0.2422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,300B, BPFP=0.3528 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,916B, BPFP=0.3313 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,824B, BPFP=0.3262 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,784B, BPFP=0.2679 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,640B, BPFP=0.3159 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,388B, BPFP=0.2457 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,700B, BPFP=0.3192 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,692B, BPFP=0.3748 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,512B, BPFP=0.3087 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,236B, BPFP=0.0179 +⌛️ [2/4] FRONTEND: Frontend time: 2.098s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.573s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15022702 389.41577061 + layer.0.v_cache 0.00001395 0.03640431 + layer.1.k_cache 0.54271635 24.83401763 + layer.1.v_cache 0.00000573 0.01446478 + layer.2.k_cache 0.01332666 2.79085943 + layer.2.v_cache 0.00001832 0.04471788 + layer.3.k_cache 0.02579262 13.02541250 + layer.3.v_cache 0.00001838 0.05191492 + layer.4.k_cache 0.00069325 1.13222873 + layer.4.v_cache 0.00005155 0.09014259 + layer.4.output 0.01006869 199.22070212 + ------------------------------------------------------------------------------------- + TOTAL 0.04725557 107.41063813 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 57316 +BPFP 0.1888 bits/point +EBPFP 0.3776 equivalent bits/point +MSE 107.410638 +---------------------- -------------------------------------------------------- +Time: 3.680s Load: 0.009s, Pack+Encode: 2.098s, Decode+Unpack: 1.573s +---------------------- -------------------------------------------------------- +💾 Converting with 107.4106 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,044B, BPFP=0.2672 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,740B, BPFP=0.3040 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,320B, BPFP=0.3347 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,780B, BPFP=0.3061 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,044B, BPFP=0.2672 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,712B, BPFP=0.3025 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,404B, BPFP=0.2333 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,796B, BPFP=0.3070 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,920B, BPFP=0.3665 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,728B, BPFP=0.3034 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,500B, BPFP=0.0189 +⌛️ [2/4] FRONTEND: Frontend time: 2.116s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.614s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12416320 399.49295551 + layer.0.v_cache 0.00001370 0.03631976 + layer.1.k_cache 0.62725913 24.53569750 + layer.1.v_cache 0.00000565 0.01460359 + layer.2.k_cache 0.01102555 2.97166458 + layer.2.v_cache 0.00002033 0.04542161 + layer.3.k_cache 0.04118381 12.96005859 + layer.3.v_cache 0.00001949 0.05424649 + layer.4.k_cache 0.00070768 1.14187250 + layer.4.v_cache 0.00005397 0.09160866 + layer.4.output 0.05105524 184.11109262 + ------------------------------------------------------------------------------------- + TOTAL 0.06834348 101.77188806 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 58988 +BPFP 0.1838 bits/point +EBPFP 0.3676 equivalent bits/point +MSE 101.771888 +---------------------- -------------------------------------------------------- +Time: 3.740s Load: 0.010s, Pack+Encode: 2.116s, Decode+Unpack: 1.614s +---------------------- -------------------------------------------------------- +💾 Converting with 101.7719 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,540B, BPFP=0.2455 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,128B, BPFP=0.3313 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,156B, BPFP=0.3328 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,660B, BPFP=0.3060 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,644B, BPFP=0.2511 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,544B, BPFP=0.2997 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,240B, BPFP=0.2292 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,736B, BPFP=0.3101 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,660B, BPFP=0.3601 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,788B, BPFP=0.3129 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,356B, BPFP=0.0182 +⌛️ [2/4] FRONTEND: Frontend time: 2.199s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.598s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15516999 390.08188257 + layer.0.v_cache 0.00001361 0.03795715 + layer.1.k_cache 0.56212566 24.71348366 + layer.1.v_cache 0.00000565 0.01508069 + layer.2.k_cache 0.00839357 2.94754081 + layer.2.v_cache 0.00001980 0.04609122 + layer.3.k_cache 0.02402391 13.28447874 + layer.3.v_cache 0.00001987 0.05417458 + layer.4.k_cache 0.00071212 1.14845176 + layer.4.v_cache 0.00005188 0.09340918 + layer.4.output 0.05073151 187.90896873 + ------------------------------------------------------------------------------------- + TOTAL 0.06503863 102.81090185 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 57452 +BPFP 0.1827 bits/point +EBPFP 0.3654 equivalent bits/point +MSE 102.810902 +---------------------- -------------------------------------------------------- +Time: 3.807s Load: 0.010s, Pack+Encode: 2.199s, Decode+Unpack: 1.598s +---------------------- -------------------------------------------------------- +💾 Converting with 102.8109 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,900B, BPFP=0.2578 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,000B, BPFP=0.3157 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,472B, BPFP=0.3405 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,600B, BPFP=0.2946 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,900B, BPFP=0.2578 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,176B, BPFP=0.2723 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,700B, BPFP=0.2473 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,880B, BPFP=0.3093 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,904B, BPFP=0.3632 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,852B, BPFP=0.3079 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,572B, BPFP=0.0193 +⌛️ [2/4] FRONTEND: Frontend time: 2.015s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.481s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005571 400.06081650 + layer.0.v_cache 0.00001459 0.03701966 + layer.1.k_cache 0.64038677 24.07197463 + layer.1.v_cache 0.00000545 0.01442168 + layer.2.k_cache 0.01623841 2.91149430 + layer.2.v_cache 0.00002030 0.04568963 + layer.3.k_cache 0.01774169 12.71171678 + layer.3.v_cache 0.00001948 0.05560343 + layer.4.k_cache 0.00069419 1.16024862 + layer.4.v_cache 0.00005057 0.09256873 + layer.4.output 0.05052170 182.87254990 + ------------------------------------------------------------------------------------- + TOTAL 0.06875759 101.25114137 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 58956 +BPFP 0.1824 bits/point +EBPFP 0.3649 equivalent bits/point +MSE 101.251141 +---------------------- -------------------------------------------------------- +Time: 3.505s Load: 0.009s, Pack+Encode: 2.015s, Decode+Unpack: 1.481s +---------------------- -------------------------------------------------------- +💾 Converting with 101.2511 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,340B, BPFP=0.2431 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,284B, BPFP=0.3519 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,108B, BPFP=0.3421 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,736B, BPFP=0.3212 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,780B, BPFP=0.2677 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,148B, BPFP=0.2883 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,336B, BPFP=0.2428 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,088B, BPFP=0.3409 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,664B, BPFP=0.3732 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,980B, BPFP=0.3349 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,280B, BPFP=0.0182 +⌛️ [2/4] FRONTEND: Frontend time: 2.090s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.596s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14010120 389.80642361 + layer.0.v_cache 0.00001437 0.03738270 + layer.1.k_cache 0.58692467 25.09774201 + layer.1.v_cache 0.00000569 0.01462629 + layer.2.k_cache 0.00806839 2.80420097 + layer.2.v_cache 0.00001934 0.04529018 + layer.3.k_cache 0.03832107 13.97870114 + layer.3.v_cache 0.00001992 0.05371657 + layer.4.k_cache 0.00069804 1.14285836 + layer.4.v_cache 0.00005131 0.09300799 + layer.4.output 0.00759059 199.17788338 + ------------------------------------------------------------------------------------- + TOTAL 0.04866812 107.48936079 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 57744 +BPFP 0.1902 bits/point +EBPFP 0.3805 equivalent bits/point +MSE 107.489361 +---------------------- -------------------------------------------------------- +Time: 3.697s Load: 0.011s, Pack+Encode: 2.090s, Decode+Unpack: 1.596s +---------------------- -------------------------------------------------------- +💾 Converting with 107.4894 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,460B, BPFP=0.2600 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,988B, BPFP=0.3491 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,512B, BPFP=0.3797 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,512B, BPFP=0.3214 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,704B, BPFP=0.2743 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,656B, BPFP=0.3298 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,268B, BPFP=0.2488 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,940B, BPFP=0.3463 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,744B, BPFP=0.3932 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,088B, BPFP=0.3549 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,484B, BPFP=0.0207 +⌛️ [2/4] FRONTEND: Frontend time: 2.055s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.482s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13199043 399.74830924 + layer.0.v_cache 0.00001386 0.03684195 + layer.1.k_cache 0.56989129 25.79042714 + layer.1.v_cache 0.00000577 0.01487693 + layer.2.k_cache 0.00892691 2.88660829 + layer.2.v_cache 0.00001816 0.04452566 + layer.3.k_cache 0.04626510 13.18824426 + layer.3.v_cache 0.00001927 0.05300770 + layer.4.k_cache 0.00069392 1.14572474 + layer.4.v_cache 0.00005195 0.09182410 + layer.4.output 0.00911696 207.41108076 + ------------------------------------------------------------------------------------- + TOTAL 0.04833502 111.46340914 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 58356 +BPFP 0.2001 bits/point +EBPFP 0.4003 equivalent bits/point +MSE 111.463409 +---------------------- -------------------------------------------------------- +Time: 3.546s Load: 0.009s, Pack+Encode: 2.055s, Decode+Unpack: 1.482s +---------------------- -------------------------------------------------------- +💾 Converting with 111.4634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,016B, BPFP=0.2290 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,960B, BPFP=0.3399 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,372B, BPFP=0.3634 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,880B, BPFP=0.3353 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,788B, BPFP=0.2730 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,008B, BPFP=0.3426 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,400B, BPFP=0.2509 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,876B, BPFP=0.3351 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,740B, BPFP=0.3844 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,000B, BPFP=0.3422 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,600B, BPFP=0.0212 +⌛️ [2/4] FRONTEND: Frontend time: 2.198s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.673s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16500157 382.98645643 + layer.0.v_cache 0.00001391 0.03778322 + layer.1.k_cache 0.57858722 24.39797417 + layer.1.v_cache 0.00000575 0.01443589 + layer.2.k_cache 0.01018238 3.04518139 + layer.2.v_cache 0.00001877 0.04580098 + layer.3.k_cache 0.03425542 14.00115120 + layer.3.v_cache 0.00002021 0.05434232 + layer.4.k_cache 0.00069807 1.14586734 + layer.4.v_cache 0.00005194 0.09042651 + layer.4.output 0.01023613 202.85168470 + ------------------------------------------------------------------------------------- + TOTAL 0.05061695 108.57536543 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 58640 +BPFP 0.1967 bits/point +EBPFP 0.3934 equivalent bits/point +MSE 108.575365 +---------------------- -------------------------------------------------------- +Time: 3.880s Load: 0.009s, Pack+Encode: 2.198s, Decode+Unpack: 1.673s +---------------------- -------------------------------------------------------- +💾 Converting with 108.5754 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,472B, BPFP=0.2588 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,956B, BPFP=0.3447 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,304B, BPFP=0.3648 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,644B, BPFP=0.3266 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,532B, BPFP=0.2623 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,336B, BPFP=0.3088 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,360B, BPFP=0.2523 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,868B, BPFP=0.3396 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,616B, BPFP=0.3829 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,240B, BPFP=0.3611 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,536B, BPFP=0.0210 +⌛️ [2/4] FRONTEND: Frontend time: 1.866s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.334s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385483 392.72867477 + layer.0.v_cache 0.00001387 0.03659261 + layer.1.k_cache 0.55871571 25.72655165 + layer.1.v_cache 0.00000565 0.01405077 + layer.2.k_cache 0.00818202 3.04344641 + layer.2.v_cache 0.00001875 0.04438187 + layer.3.k_cache 0.04697755 13.84912109 + layer.3.v_cache 0.00001978 0.05380807 + layer.4.k_cache 0.00071636 1.12605014 + layer.4.v_cache 0.00005146 0.09192311 + layer.4.output 0.01106052 205.85958995 + ------------------------------------------------------------------------------------- + TOTAL 0.04741057 110.45480765 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 57864 +BPFP 0.1970 bits/point +EBPFP 0.3940 equivalent bits/point +MSE 110.454808 +---------------------- -------------------------------------------------------- +Time: 3.209s Load: 0.010s, Pack+Encode: 1.866s, Decode+Unpack: 1.334s +---------------------- -------------------------------------------------------- +💾 Converting with 110.4548 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,332B, BPFP=0.2461 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,068B, BPFP=0.3448 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,288B, BPFP=0.3573 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,800B, BPFP=0.3295 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,752B, BPFP=0.2700 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,524B, BPFP=0.3139 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,496B, BPFP=0.2555 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,184B, BPFP=0.3514 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,804B, BPFP=0.3866 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,944B, BPFP=0.3377 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,480B, BPFP=0.0201 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.332s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203608 393.67960227 + layer.0.v_cache 0.00001420 0.03645111 + layer.1.k_cache 0.56673284 24.69349254 + layer.1.v_cache 0.00000552 0.01416534 + layer.2.k_cache 0.00530624 2.91456277 + layer.2.v_cache 0.00001825 0.04475597 + layer.3.k_cache 0.02188354 12.71514737 + layer.3.v_cache 0.00001854 0.05246053 + layer.4.k_cache 0.00069888 1.13523204 + layer.4.v_cache 0.00004918 0.09105118 + layer.4.output 0.00799176 202.12407468 + ------------------------------------------------------------------------------------- + TOTAL 0.04662974 108.83796729 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 58672 +BPFP 0.1961 bits/point +EBPFP 0.3922 equivalent bits/point +MSE 108.837967 +---------------------- -------------------------------------------------------- +Time: 3.176s Load: 0.011s, Pack+Encode: 1.833s, Decode+Unpack: 1.332s +---------------------- -------------------------------------------------------- +💾 Converting with 108.8380 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,900B, BPFP=0.2705 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,308B, BPFP=0.3483 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,788B, BPFP=0.3196 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,624B, BPFP=0.3105 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,624B, BPFP=0.2553 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,452B, BPFP=0.3010 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,188B, BPFP=0.2312 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,040B, BPFP=0.3335 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,492B, BPFP=0.3584 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,764B, BPFP=0.3182 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,272B, BPFP=0.0179 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.336s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12732459 404.45668617 + layer.0.v_cache 0.00001480 0.03684704 + layer.1.k_cache 0.57322790 24.11230814 + layer.1.v_cache 0.00000557 0.01427884 + layer.2.k_cache 0.00950604 2.97089809 + layer.2.v_cache 0.00002039 0.04549522 + layer.3.k_cache 0.01510038 12.92754803 + layer.3.v_cache 0.00001916 0.05287829 + layer.4.k_cache 0.00068141 1.11350037 + layer.4.v_cache 0.00005165 0.09152938 + layer.4.output 0.00936729 196.42663112 + ------------------------------------------------------------------------------------- + TOTAL 0.04656017 107.10637573 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 57452 +BPFP 0.1866 bits/point +EBPFP 0.3732 equivalent bits/point +MSE 107.106376 +---------------------- -------------------------------------------------------- +Time: 3.177s Load: 0.010s, Pack+Encode: 1.830s, Decode+Unpack: 1.336s +---------------------- -------------------------------------------------------- +💾 Converting with 107.1064 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,504B, BPFP=0.2550 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,900B, BPFP=0.3340 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,276B, BPFP=0.3553 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,776B, BPFP=0.3270 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,852B, BPFP=0.2747 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,648B, BPFP=0.3197 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,284B, BPFP=0.2425 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,980B, BPFP=0.3385 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,852B, BPFP=0.3879 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,628B, BPFP=0.3186 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,376B, BPFP=0.0192 +⌛️ [2/4] FRONTEND: Frontend time: 1.822s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.336s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13630490 393.01907835 + layer.0.v_cache 0.00001544 0.03706146 + layer.1.k_cache 0.58776402 25.08200648 + layer.1.v_cache 0.00000559 0.01418522 + layer.2.k_cache 0.01000112 2.89838841 + layer.2.v_cache 0.00001940 0.04609103 + layer.3.k_cache 0.04424953 12.96458015 + layer.3.v_cache 0.00002139 0.05376673 + layer.4.k_cache 0.00071093 1.13019971 + layer.4.v_cache 0.00005158 0.09503434 + layer.4.output 0.01101471 201.37545290 + ------------------------------------------------------------------------------------- + TOTAL 0.05036746 108.52756248 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 58076 +BPFP 0.1934 bits/point +EBPFP 0.3868 equivalent bits/point +MSE 108.527562 +---------------------- -------------------------------------------------------- +Time: 3.168s Load: 0.010s, Pack+Encode: 1.822s, Decode+Unpack: 1.336s +---------------------- -------------------------------------------------------- +💾 Converting with 108.5276 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,656B, BPFP=0.2617 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,988B, BPFP=0.3366 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,180B, BPFP=0.3473 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,836B, BPFP=0.3280 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,836B, BPFP=0.2718 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,636B, BPFP=0.3168 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,448B, BPFP=0.2500 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,944B, BPFP=0.3341 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,504B, BPFP=0.3656 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,716B, BPFP=0.3213 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,260B, BPFP=0.0181 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.340s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12958164 400.02807442 + layer.0.v_cache 0.00001430 0.03633120 + layer.1.k_cache 0.54121377 25.56221195 + layer.1.v_cache 0.00000577 0.01433234 + layer.2.k_cache 0.01080061 2.79196650 + layer.2.v_cache 0.00001903 0.04574288 + layer.3.k_cache 0.01783211 13.72038785 + layer.3.v_cache 0.00002028 0.05180627 + layer.4.k_cache 0.00068935 1.13923074 + layer.4.v_cache 0.00005053 0.09163231 + layer.4.output 0.00833506 199.95755717 + ------------------------------------------------------------------------------------- + TOTAL 0.04462193 108.42262451 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 58004 +BPFP 0.1918 bits/point +EBPFP 0.3835 equivalent bits/point +MSE 108.422625 +---------------------- -------------------------------------------------------- +Time: 3.181s Load: 0.011s, Pack+Encode: 1.830s, Decode+Unpack: 1.340s +---------------------- -------------------------------------------------------- +💾 Converting with 108.4226 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,264B, BPFP=0.2533 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,388B, BPFP=0.3795 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,284B, BPFP=0.3733 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,856B, BPFP=0.3479 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,520B, BPFP=0.2685 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,380B, BPFP=0.3196 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,064B, BPFP=0.2414 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,668B, BPFP=0.3367 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,716B, BPFP=0.3990 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,012B, BPFP=0.3572 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,428B, BPFP=0.0206 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12547956 406.21263070 + layer.0.v_cache 0.00001492 0.03835415 + layer.1.k_cache 0.50446328 24.58785906 + layer.1.v_cache 0.00000633 0.01489336 + layer.2.k_cache 0.00671430 2.76608265 + layer.2.v_cache 0.00001992 0.04741260 + layer.3.k_cache 0.02142675 13.45486926 + layer.3.v_cache 0.00001986 0.05561562 + layer.4.k_cache 0.00069859 1.14937623 + layer.4.v_cache 0.00005414 0.09545971 + layer.4.output 0.01038299 211.35834804 + ------------------------------------------------------------------------------------- + TOTAL 0.04303403 113.40770527 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 57580 +BPFP 0.2012 bits/point +EBPFP 0.4025 equivalent bits/point +MSE 113.407705 +---------------------- -------------------------------------------------------- +Time: 3.179s Load: 0.008s, Pack+Encode: 1.830s, Decode+Unpack: 1.341s +---------------------- -------------------------------------------------------- +💾 Converting with 113.4077 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,856B, BPFP=0.2521 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,288B, BPFP=0.3264 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,308B, BPFP=0.3275 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,756B, BPFP=0.2988 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,876B, BPFP=0.2531 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,260B, BPFP=0.2730 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,724B, BPFP=0.2452 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,912B, BPFP=0.3069 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,916B, BPFP=0.3590 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,788B, BPFP=0.3005 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,544B, BPFP=0.0189 +⌛️ [2/4] FRONTEND: Frontend time: 1.843s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13970643 399.77917359 + layer.0.v_cache 0.00001441 0.03722838 + layer.1.k_cache 0.60015088 24.57441828 + layer.1.v_cache 0.00000584 0.01483243 + layer.2.k_cache 0.00648033 2.82729484 + layer.2.v_cache 0.00001840 0.04545940 + layer.3.k_cache 0.03288257 13.35345641 + layer.3.v_cache 0.00002024 0.05388758 + layer.4.k_cache 0.00070287 1.17596963 + layer.4.v_cache 0.00005060 0.09204813 + layer.4.output 0.04874230 180.35834421 + ------------------------------------------------------------------------------------- + TOTAL 0.06595463 100.26248107 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 59228 +BPFP 0.1809 bits/point +EBPFP 0.3617 equivalent bits/point +MSE 100.262481 +---------------------- -------------------------------------------------------- +Time: 3.197s Load: 0.011s, Pack+Encode: 1.843s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 100.2625 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,928B, BPFP=0.2355 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,704B, BPFP=0.3681 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,196B, BPFP=0.3438 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,668B, BPFP=0.3186 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,568B, BPFP=0.2661 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,128B, BPFP=0.3406 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,116B, BPFP=0.2445 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,368B, BPFP=0.3043 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,784B, BPFP=0.3719 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,640B, BPFP=0.3651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,540B, BPFP=0.0173 +⌛️ [2/4] FRONTEND: Frontend time: 1.951s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.455s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13960611 385.46820050 + layer.0.v_cache 0.00001491 0.03824288 + layer.1.k_cache 0.68890666 25.30678995 + layer.1.v_cache 0.00000583 0.01478631 + layer.2.k_cache 0.00940153 2.74909772 + layer.2.v_cache 0.00001882 0.04672471 + layer.3.k_cache 0.04667800 13.14910049 + layer.3.v_cache 0.00001939 0.05453992 + layer.4.k_cache 0.00070827 1.13884117 + layer.4.v_cache 0.00005291 0.09423411 + layer.4.output 0.04427218 166.09805046 + ------------------------------------------------------------------------------------- + TOTAL 0.07031281 93.57334770 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 68640 +BPFP 0.1929 bits/point +EBPFP 0.3859 equivalent bits/point +MSE 93.573348 +---------------------- -------------------------------------------------------- +Time: 3.418s Load: 0.012s, Pack+Encode: 1.951s, Decode+Unpack: 1.455s +---------------------- -------------------------------------------------------- +💾 Converting with 93.5733 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,820B, BPFP=0.2519 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,144B, BPFP=0.3211 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,484B, BPFP=0.3388 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,832B, BPFP=0.3048 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,920B, BPFP=0.2571 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,664B, BPFP=0.2960 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,728B, BPFP=0.2471 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,600B, BPFP=0.2926 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,968B, BPFP=0.3641 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,756B, BPFP=0.3008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,552B, BPFP=0.0191 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15817952 398.84769544 + layer.0.v_cache 0.00001474 0.03680925 + layer.1.k_cache 0.61609703 24.26468110 + layer.1.v_cache 0.00000566 0.01455391 + layer.2.k_cache 0.00844678 2.82145951 + layer.2.v_cache 0.00001876 0.04488337 + layer.3.k_cache 0.04682518 12.97923580 + layer.3.v_cache 0.00001896 0.05290585 + layer.4.k_cache 0.00069859 1.13782432 + layer.4.v_cache 0.00005216 0.09156118 + layer.4.output 0.04649196 181.70008660 + ------------------------------------------------------------------------------------- + TOTAL 0.06798830 100.71718917 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 59468 +BPFP 0.1828 bits/point +EBPFP 0.3656 equivalent bits/point +MSE 100.717189 +---------------------- -------------------------------------------------------- +Time: 3.183s Load: 0.009s, Pack+Encode: 1.830s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 100.7172 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,664B, BPFP=0.2539 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,976B, BPFP=0.3253 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,980B, BPFP=0.3256 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,596B, BPFP=0.3047 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,540B, BPFP=0.2472 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,884B, BPFP=0.2659 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,148B, BPFP=0.2258 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,748B, BPFP=0.3129 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,300B, BPFP=0.3430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,572B, BPFP=0.3034 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,212B, BPFP=0.0172 +⌛️ [2/4] FRONTEND: Frontend time: 1.880s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14002854 399.60080030 + layer.0.v_cache 0.00001601 0.03787770 + layer.1.k_cache 0.54217949 25.14309022 + layer.1.v_cache 0.00000579 0.01494175 + layer.2.k_cache 0.01082110 2.77038106 + layer.2.v_cache 0.00001971 0.04619069 + layer.3.k_cache 0.03102738 13.20149343 + layer.3.v_cache 0.00001907 0.05533948 + layer.4.k_cache 0.00066782 1.12784186 + layer.4.v_cache 0.00005141 0.09201034 + layer.4.output 0.00856273 193.68067135 + ------------------------------------------------------------------------------------- + TOTAL 0.04616326 105.75615684 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 55620 +BPFP 0.1781 bits/point +EBPFP 0.3562 equivalent bits/point +MSE 105.756157 +---------------------- -------------------------------------------------------- +Time: 3.235s Load: 0.011s, Pack+Encode: 1.880s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 105.7562 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,848B, BPFP=0.2551 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,984B, BPFP=0.3148 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,580B, BPFP=0.3462 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,676B, BPFP=0.2986 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,960B, BPFP=0.2609 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,516B, BPFP=0.2902 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,668B, BPFP=0.2456 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,876B, BPFP=0.3091 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,928B, BPFP=0.3645 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,924B, BPFP=0.3117 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,680B, BPFP=0.0201 +⌛️ [2/4] FRONTEND: Frontend time: 1.842s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.342s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14705232 399.88733691 + layer.0.v_cache 0.00001374 0.03613149 + layer.1.k_cache 0.65666707 24.41808383 + layer.1.v_cache 0.00000579 0.01437818 + layer.2.k_cache 0.01062412 2.89767795 + layer.2.v_cache 0.00001918 0.04473866 + layer.3.k_cache 0.03205616 13.36154596 + layer.3.v_cache 0.00001999 0.05402810 + layer.4.k_cache 0.00071739 1.12642559 + layer.4.v_cache 0.00005111 0.08907135 + layer.4.output 0.04976816 182.88434945 + ------------------------------------------------------------------------------------- + TOTAL 0.07032965 101.30116848 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 59640 +BPFP 0.1846 bits/point +EBPFP 0.3691 equivalent bits/point +MSE 101.301168 +---------------------- -------------------------------------------------------- +Time: 3.196s Load: 0.012s, Pack+Encode: 1.842s, Decode+Unpack: 1.342s +---------------------- -------------------------------------------------------- +💾 Converting with 101.3012 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,272B, BPFP=0.2375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,176B, BPFP=0.3434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,188B, BPFP=0.3441 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,612B, BPFP=0.3121 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,740B, BPFP=0.2636 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,936B, BPFP=0.3301 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,472B, BPFP=0.2487 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,844B, BPFP=0.3250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,628B, BPFP=0.3685 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,012B, BPFP=0.3343 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,300B, BPFP=0.0183 +⌛️ [2/4] FRONTEND: Frontend time: 1.851s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131647 398.34794818 + layer.0.v_cache 0.00001392 0.03680138 + layer.1.k_cache 0.51967005 24.86035330 + layer.1.v_cache 0.00000557 0.01443071 + layer.2.k_cache 0.01013049 2.91901221 + layer.2.v_cache 0.00001908 0.04608842 + layer.3.k_cache 0.05186962 13.68032956 + layer.3.v_cache 0.00001891 0.05375642 + layer.4.k_cache 0.00069958 1.12873400 + layer.4.v_cache 0.00005017 0.09100094 + layer.4.output 0.01073789 197.79357842 + ------------------------------------------------------------------------------------- + TOTAL 0.04758583 107.39608847 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 58180 +BPFP 0.1903 bits/point +EBPFP 0.3806 equivalent bits/point +MSE 107.396088 +---------------------- -------------------------------------------------------- +Time: 3.207s Load: 0.011s, Pack+Encode: 1.851s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 107.3961 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,224B, BPFP=0.2292 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,836B, BPFP=0.3166 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,012B, BPFP=0.3262 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,724B, BPFP=0.3105 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,616B, BPFP=0.2504 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,452B, BPFP=0.2958 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,300B, BPFP=0.2333 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,460B, BPFP=0.2962 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,548B, BPFP=0.3553 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,408B, BPFP=0.2934 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,208B, BPFP=0.0171 +⌛️ [2/4] FRONTEND: Frontend time: 1.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15566496 391.02197266 + layer.0.v_cache 0.00001437 0.03583181 + layer.1.k_cache 0.62643353 24.24716865 + layer.1.v_cache 0.00000573 0.01471088 + layer.2.k_cache 0.01010078 2.85252126 + layer.2.v_cache 0.00001868 0.04478281 + layer.3.k_cache 0.06271097 12.75845167 + layer.3.v_cache 0.00001958 0.05438002 + layer.4.k_cache 0.00073784 1.11350462 + layer.4.v_cache 0.00005294 0.09128907 + layer.4.output 0.00677609 193.03514075 + ------------------------------------------------------------------------------------- + TOTAL 0.05312894 104.91062345 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 55788 +BPFP 0.1780 bits/point +EBPFP 0.3561 equivalent bits/point +MSE 104.910623 +---------------------- -------------------------------------------------------- +Time: 3.201s Load: 0.009s, Pack+Encode: 1.847s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 104.9106 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,744B, BPFP=0.2592 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,768B, BPFP=0.3151 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,692B, BPFP=0.3110 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,556B, BPFP=0.3035 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,488B, BPFP=0.2452 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,744B, BPFP=0.3138 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,064B, BPFP=0.2220 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,780B, BPFP=0.3158 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,460B, BPFP=0.3529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,480B, BPFP=0.2994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,232B, BPFP=0.0174 +⌛️ [2/4] FRONTEND: Frontend time: 1.849s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13918158 399.51753715 + layer.0.v_cache 0.00001566 0.03723435 + layer.1.k_cache 0.57997750 24.66966237 + layer.1.v_cache 0.00000571 0.01447298 + layer.2.k_cache 0.01637170 3.09698678 + layer.2.v_cache 0.00001887 0.04644627 + layer.3.k_cache 0.04393929 12.79808375 + layer.3.v_cache 0.00001903 0.05395768 + layer.4.k_cache 0.00071104 1.12224189 + layer.4.v_cache 0.00005175 0.09227501 + layer.4.output 0.01065216 194.34384366 + ------------------------------------------------------------------------------------- + TOTAL 0.05028572 105.99151787 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 56008 +BPFP 0.1800 bits/point +EBPFP 0.3600 equivalent bits/point +MSE 105.991518 +---------------------- -------------------------------------------------------- +Time: 3.207s Load: 0.012s, Pack+Encode: 1.849s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 105.9915 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,040B, BPFP=0.2540 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,996B, BPFP=0.3022 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,292B, BPFP=0.3171 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,920B, BPFP=0.2984 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,632B, BPFP=0.2335 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,088B, BPFP=0.2565 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,472B, BPFP=0.2254 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,784B, BPFP=0.2915 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,496B, BPFP=0.3274 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,708B, BPFP=0.2877 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,876B, BPFP=0.0207 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16496690 407.09034778 + layer.0.v_cache 0.00001433 0.03780651 + layer.1.k_cache 0.65265572 25.09091009 + layer.1.v_cache 0.00000578 0.01480206 + layer.2.k_cache 0.01241827 3.02191379 + layer.2.v_cache 0.00001920 0.04540229 + layer.3.k_cache 0.08444468 13.26553994 + layer.3.v_cache 0.00001975 0.05601471 + layer.4.k_cache 0.00068928 1.19827290 + layer.4.v_cache 0.00005259 0.09361219 + layer.4.output 0.04785494 175.21453053 + ------------------------------------------------------------------------------------- + TOTAL 0.07354536 98.61272565 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 58304 +BPFP 0.1729 bits/point +EBPFP 0.3457 equivalent bits/point +MSE 98.612726 +---------------------- -------------------------------------------------------- +Time: 3.195s Load: 0.011s, Pack+Encode: 1.840s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 98.6127 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,684B, BPFP=0.2489 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,096B, BPFP=0.3240 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,344B, BPFP=0.3372 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,664B, BPFP=0.3010 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,952B, BPFP=0.2632 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,440B, BPFP=0.2891 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,564B, BPFP=0.2426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,876B, BPFP=0.3123 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,924B, BPFP=0.3680 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,808B, BPFP=0.3087 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,460B, BPFP=0.0187 +⌛️ [2/4] FRONTEND: Frontend time: 1.843s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.342s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15001445 402.73022959 + layer.0.v_cache 0.00001441 0.03691378 + layer.1.k_cache 0.57475359 24.27060082 + layer.1.v_cache 0.00000584 0.01456199 + layer.2.k_cache 0.01434258 2.99439162 + layer.2.v_cache 0.00001954 0.04571247 + layer.3.k_cache 0.02649183 12.43065327 + layer.3.v_cache 0.00001945 0.05299806 + layer.4.k_cache 0.00069414 1.13038293 + layer.4.v_cache 0.00005306 0.09095232 + layer.4.output 0.05035525 184.74552053 + ------------------------------------------------------------------------------------- + TOTAL 0.06581739 102.17741415 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 58812 +BPFP 0.1839 bits/point +EBPFP 0.3677 equivalent bits/point +MSE 102.177414 +---------------------- -------------------------------------------------------- +Time: 3.197s Load: 0.011s, Pack+Encode: 1.843s, Decode+Unpack: 1.342s +---------------------- -------------------------------------------------------- +💾 Converting with 102.1774 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,720B, BPFP=0.2615 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,128B, BPFP=0.3395 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,032B, BPFP=0.3342 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,392B, BPFP=0.2988 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,716B, BPFP=0.2613 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,476B, BPFP=0.3034 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,408B, BPFP=0.2442 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,944B, BPFP=0.3293 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,692B, BPFP=0.3708 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,700B, BPFP=0.3158 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,344B, BPFP=0.0186 +⌛️ [2/4] FRONTEND: Frontend time: 1.832s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.339s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11854591 402.10688165 + layer.0.v_cache 0.00001566 0.03706448 + layer.1.k_cache 0.56985614 24.80487277 + layer.1.v_cache 0.00000555 0.01434026 + layer.2.k_cache 0.01267573 2.84430278 + layer.2.v_cache 0.00001902 0.04525929 + layer.3.k_cache 0.04477839 14.05011116 + layer.3.v_cache 0.00001896 0.05335140 + layer.4.k_cache 0.00071754 1.15121579 + layer.4.v_cache 0.00005321 0.09308310 + layer.4.output 0.00761333 197.08757599 + ------------------------------------------------------------------------------------- + TOTAL 0.04705761 107.34197145 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 57552 +BPFP 0.1876 bits/point +EBPFP 0.3752 equivalent bits/point +MSE 107.341971 +---------------------- -------------------------------------------------------- +Time: 3.183s Load: 0.011s, Pack+Encode: 1.832s, Decode+Unpack: 1.339s +---------------------- -------------------------------------------------------- +💾 Converting with 107.3420 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,576B, BPFP=0.2432 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,028B, BPFP=0.3204 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,388B, BPFP=0.3395 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,664B, BPFP=0.3010 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,924B, BPFP=0.2617 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,908B, BPFP=0.3140 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,640B, BPFP=0.2466 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,884B, BPFP=0.3127 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,804B, BPFP=0.3616 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,952B, BPFP=0.3163 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,424B, BPFP=0.0184 +⌛️ [2/4] FRONTEND: Frontend time: 1.824s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.339s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582189 396.47996386 + layer.0.v_cache 0.00001509 0.03787272 + layer.1.k_cache 0.58311327 24.64312453 + layer.1.v_cache 0.00000606 0.01483178 + layer.2.k_cache 0.00657600 2.66565149 + layer.2.v_cache 0.00001909 0.04694349 + layer.3.k_cache 0.02234380 12.35652752 + layer.3.v_cache 0.00001905 0.05375731 + layer.4.k_cache 0.00071828 1.13904411 + layer.4.v_cache 0.00005370 0.09505928 + layer.4.output 0.05044473 184.75672680 + ------------------------------------------------------------------------------------- + TOTAL 0.06363526 101.81352139 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 59192 +BPFP 0.1850 bits/point +EBPFP 0.3701 equivalent bits/point +MSE 101.813521 +---------------------- -------------------------------------------------------- +Time: 3.173s Load: 0.010s, Pack+Encode: 1.824s, Decode+Unpack: 1.339s +---------------------- -------------------------------------------------------- +💾 Converting with 101.8135 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 292, 128) +Output shape: (1, 292, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.output: torch.Size([1, 292, 3584]) -> torch.Size([1, 1, 292, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,544B, BPFP=0.2432 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,308B, BPFP=0.3375 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,464B, BPFP=0.3459 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,860B, BPFP=0.3136 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,064B, BPFP=0.2710 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,820B, BPFP=0.3114 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,700B, BPFP=0.2515 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,584B, BPFP=0.2988 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,876B, BPFP=0.3679 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,676B, BPFP=0.3037 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,632B, BPFP=0.0201 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.340s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887579 391.37853168 + layer.0.v_cache 0.00001384 0.03749002 + layer.1.k_cache 0.57277983 25.06460362 + layer.1.v_cache 0.00000568 0.01458555 + layer.2.k_cache 0.01314942 2.84296010 + layer.2.v_cache 0.00002105 0.04567606 + layer.3.k_cache 0.03125376 12.50488950 + layer.3.v_cache 0.00001978 0.05421568 + layer.4.k_cache 0.00068035 1.14819096 + layer.4.v_cache 0.00005123 0.09222214 + layer.4.output 0.04859251 185.99338001 + ------------------------------------------------------------------------------------- + TOTAL 0.06511755 102.06688385 + (elements=2,541,568) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2541568 +Total Bytes 59528 +BPFP 0.1874 bits/point +EBPFP 0.3747 equivalent bits/point +MSE 102.066884 +---------------------- -------------------------------------------------------- +Time: 3.184s Load: 0.010s, Pack+Encode: 1.834s, Decode+Unpack: 1.340s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0669 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 338, 128) +Output shape: (1, 338, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.output: torch.Size([1, 338, 3584]) -> torch.Size([1, 1, 338, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,732B, BPFP=0.2650 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,624B, BPFP=0.3524 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,400B, BPFP=0.3421 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,348B, BPFP=0.3397 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,676B, BPFP=0.2624 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,184B, BPFP=0.3321 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,036B, BPFP=0.2328 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,572B, BPFP=0.3500 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,896B, BPFP=0.3650 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,976B, BPFP=0.3225 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,772B, BPFP=0.0183 +⌛️ [2/4] FRONTEND: Frontend time: 1.966s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.486s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13976314 399.15509430 + layer.0.v_cache 0.00001472 0.03781490 + layer.1.k_cache 0.73750278 25.41140151 + layer.1.v_cache 0.00000589 0.01520949 + layer.2.k_cache 0.02011468 2.83361925 + layer.2.v_cache 0.00002041 0.04427145 + layer.3.k_cache 0.02722352 13.33869094 + layer.3.v_cache 0.00002029 0.05343492 + layer.4.k_cache 0.00072609 1.16632513 + layer.4.v_cache 0.00005477 0.09275677 + layer.4.output 0.04270584 160.72368977 + ------------------------------------------------------------------------------------- + TOTAL 0.07202277 92.18908512 + (elements=2,941,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2941952 +Total Bytes 71216 +BPFP 0.1937 bits/point +EBPFP 0.3873 equivalent bits/point +MSE 92.189085 +---------------------- -------------------------------------------------------- +Time: 3.464s Load: 0.012s, Pack+Encode: 1.966s, Decode+Unpack: 1.486s +---------------------- -------------------------------------------------------- +💾 Converting with 92.1891 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,608B, BPFP=0.2474 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,396B, BPFP=0.3434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,172B, BPFP=0.3314 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,460B, BPFP=0.2932 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,980B, BPFP=0.2674 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,680B, BPFP=0.3050 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,536B, BPFP=0.2436 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,040B, BPFP=0.3243 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,924B, BPFP=0.3718 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,728B, BPFP=0.3076 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,476B, BPFP=0.0190 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150811 394.79231100 + layer.0.v_cache 0.00001374 0.03715229 + layer.1.k_cache 0.66021251 25.15878705 + layer.1.v_cache 0.00000574 0.01456710 + layer.2.k_cache 0.00963001 2.68454409 + layer.2.v_cache 0.00001988 0.04492706 + layer.3.k_cache 0.08055985 12.00332736 + layer.3.v_cache 0.00001942 0.05169295 + layer.4.k_cache 0.00069026 1.13856690 + layer.4.v_cache 0.00005126 0.09265502 + layer.4.output 0.05093874 186.62187040 + ------------------------------------------------------------------------------------- + TOTAL 0.07289894 102.49244845 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 59000 +BPFP 0.1864 bits/point +EBPFP 0.3727 equivalent bits/point +MSE 102.492448 +---------------------- -------------------------------------------------------- +Time: 3.205s Load: 0.011s, Pack+Encode: 1.838s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 102.4924 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,888B, BPFP=0.2521 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,120B, BPFP=0.3156 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,280B, BPFP=0.3238 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,424B, BPFP=0.2797 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,860B, BPFP=0.2506 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,660B, BPFP=0.2919 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,656B, BPFP=0.2401 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,896B, BPFP=0.3040 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,812B, BPFP=0.3513 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,716B, BPFP=0.2948 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,556B, BPFP=0.0188 +⌛️ [2/4] FRONTEND: Frontend time: 1.848s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14893842 401.23094575 + layer.0.v_cache 0.00001415 0.03655607 + layer.1.k_cache 0.65110169 24.00172429 + layer.1.v_cache 0.00000575 0.01450149 + layer.2.k_cache 0.01411552 2.87736325 + layer.2.v_cache 0.00001904 0.04492636 + layer.3.k_cache 0.03939451 12.95544232 + layer.3.v_cache 0.00001876 0.05201688 + layer.4.k_cache 0.00070558 1.16402856 + layer.4.v_cache 0.00005345 0.09030539 + layer.4.output 0.04819407 179.27146688 + ------------------------------------------------------------------------------------- + TOTAL 0.07010149 99.84518109 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 58868 +BPFP 0.1786 bits/point +EBPFP 0.3571 equivalent bits/point +MSE 99.845181 +---------------------- -------------------------------------------------------- +Time: 3.212s Load: 0.010s, Pack+Encode: 1.848s, Decode+Unpack: 1.354s +---------------------- -------------------------------------------------------- +💾 Converting with 99.8452 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,340B, BPFP=0.2578 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,112B, BPFP=0.3631 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,988B, BPFP=0.3558 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,716B, BPFP=0.3396 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,464B, BPFP=0.2652 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,388B, BPFP=0.3201 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,944B, BPFP=0.2343 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,912B, BPFP=0.3512 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,496B, BPFP=0.3859 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,476B, BPFP=0.3847 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,324B, BPFP=0.0197 +⌛️ [2/4] FRONTEND: Frontend time: 1.831s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14492303 406.13872386 + layer.0.v_cache 0.00001375 0.03686424 + layer.1.k_cache 0.50395557 24.57437656 + layer.1.v_cache 0.00000564 0.01430568 + layer.2.k_cache 0.00856250 2.84116262 + layer.2.v_cache 0.00001902 0.04682521 + layer.3.k_cache 0.02692018 13.20244363 + layer.3.v_cache 0.00002118 0.05359673 + layer.4.k_cache 0.00070790 1.12760038 + layer.4.v_cache 0.00005130 0.09289309 + layer.4.output 0.01131245 211.34658474 + ------------------------------------------------------------------------------------- + TOTAL 0.04496278 113.38558148 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 57160 +BPFP 0.1998 bits/point +EBPFP 0.3995 equivalent bits/point +MSE 113.385581 +---------------------- -------------------------------------------------------- +Time: 3.199s Load: 0.010s, Pack+Encode: 1.831s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 113.3856 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,632B, BPFP=0.2622 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,064B, BPFP=0.3433 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,192B, BPFP=0.3505 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,696B, BPFP=0.3225 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,564B, BPFP=0.2584 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,304B, BPFP=0.3003 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,428B, BPFP=0.2507 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,816B, BPFP=0.3293 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,552B, BPFP=0.3709 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,836B, BPFP=0.3304 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,360B, BPFP=0.0191 +⌛️ [2/4] FRONTEND: Frontend time: 1.842s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14113761 394.10580842 + layer.0.v_cache 0.00001369 0.03676580 + layer.1.k_cache 0.59212339 25.56448674 + layer.1.v_cache 0.00000552 0.01418534 + layer.2.k_cache 0.01129283 2.96917128 + layer.2.v_cache 0.00001853 0.04441394 + layer.3.k_cache 0.01358872 12.91200146 + layer.3.v_cache 0.00001883 0.05317712 + layer.4.k_cache 0.00070763 1.11168040 + layer.4.v_cache 0.00005153 0.09112362 + layer.4.output 0.01019159 201.38719591 + ------------------------------------------------------------------------------------- + TOTAL 0.04884114 108.62430503 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 57444 +BPFP 0.1913 bits/point +EBPFP 0.3826 equivalent bits/point +MSE 108.624305 +---------------------- -------------------------------------------------------- +Time: 3.194s Load: 0.010s, Pack+Encode: 1.842s, Decode+Unpack: 1.341s +---------------------- -------------------------------------------------------- +💾 Converting with 108.6243 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,652B, BPFP=0.2533 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,980B, BPFP=0.3256 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,024B, BPFP=0.3280 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,360B, BPFP=0.2918 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,556B, BPFP=0.2480 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,312B, BPFP=0.2892 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,012B, BPFP=0.2184 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,548B, BPFP=0.3020 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,404B, BPFP=0.3486 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,452B, BPFP=0.2968 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,172B, BPFP=0.0169 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.340s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11028925 402.24814895 + layer.0.v_cache 0.00001466 0.03708033 + layer.1.k_cache 0.56032453 25.60746781 + layer.1.v_cache 0.00000595 0.01458532 + layer.2.k_cache 0.00809042 2.96229601 + layer.2.v_cache 0.00001929 0.04646288 + layer.3.k_cache 0.03501172 13.18638988 + layer.3.v_cache 0.00001861 0.05271543 + layer.4.k_cache 0.00070074 1.15171436 + layer.4.v_cache 0.00005246 0.09157526 + layer.4.output 0.00889761 193.68519786 + ------------------------------------------------------------------------------------- + TOTAL 0.04569476 105.95263655 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 55472 +BPFP 0.1776 bits/point +EBPFP 0.3553 equivalent bits/point +MSE 105.952637 +---------------------- -------------------------------------------------------- +Time: 3.184s Load: 0.011s, Pack+Encode: 1.833s, Decode+Unpack: 1.340s +---------------------- -------------------------------------------------------- +💾 Converting with 105.9526 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 311, 128) +Output shape: (1, 311, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.output: torch.Size([1, 311, 3584]) -> torch.Size([1, 1, 311, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,976B, BPFP=0.2500 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,424B, BPFP=0.3227 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,212B, BPFP=0.3121 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,680B, BPFP=0.2854 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,660B, BPFP=0.2341 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,520B, BPFP=0.2773 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,336B, BPFP=0.2178 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,032B, BPFP=0.3031 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,364B, BPFP=0.3197 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,788B, BPFP=0.2908 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,240B, BPFP=0.0233 +⌛️ [2/4] FRONTEND: Frontend time: 1.836s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.340s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14227538 404.34573453 + layer.0.v_cache 0.00001500 0.03805163 + layer.1.k_cache 0.64922478 24.55886694 + layer.1.v_cache 0.00000572 0.01499208 + layer.2.k_cache 0.00828172 3.08099777 + layer.2.v_cache 0.00001922 0.04582140 + layer.3.k_cache 0.02636586 13.07465679 + layer.3.v_cache 0.00001996 0.05723623 + layer.4.k_cache 0.00069494 1.20032209 + layer.4.v_cache 0.00005248 0.09473740 + layer.4.output 0.04686997 174.66361392 + ------------------------------------------------------------------------------------- + TOTAL 0.06794382 98.18568908 + (elements=2,706,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2706944 +Total Bytes 59232 +BPFP 0.1751 bits/point +EBPFP 0.3501 equivalent bits/point +MSE 98.185689 +---------------------- -------------------------------------------------------- +Time: 3.186s Load: 0.010s, Pack+Encode: 1.836s, Decode+Unpack: 1.340s +---------------------- -------------------------------------------------------- +💾 Converting with 98.1857 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,168B, BPFP=0.2554 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,352B, BPFP=0.2667 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,416B, BPFP=0.2706 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,144B, BPFP=0.2539 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,332B, BPFP=0.2042 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,916B, BPFP=0.2400 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,036B, BPFP=0.1860 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,536B, BPFP=0.2779 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,696B, BPFP=0.2877 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,220B, BPFP=0.2586 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,004B, BPFP=0.0175 +⌛️ [2/4] FRONTEND: Frontend time: 1.945s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.216s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13149661 404.07408088 + layer.0.v_cache 0.00001370 0.03640407 + layer.1.k_cache 0.50214604 23.68327206 + layer.1.v_cache 0.00000625 0.01462244 + layer.2.k_cache 0.00807812 2.65197730 + layer.2.v_cache 0.00001969 0.04570082 + layer.3.k_cache 0.06061829 12.14904354 + layer.3.v_cache 0.00001868 0.05433345 + layer.4.k_cache 0.00068906 1.12352690 + layer.4.v_cache 0.00005044 0.09142401 + layer.4.output 1.20679682 212.62605042 + ------------------------------------------------------------------------------------- + TOTAL 0.53827733 113.66510226 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 42820 +BPFP 0.1543 bits/point +EBPFP 0.3087 equivalent bits/point +MSE 113.665102 +---------------------- -------------------------------------------------------- +Time: 3.169s Load: 0.009s, Pack+Encode: 1.945s, Decode+Unpack: 1.216s +---------------------- -------------------------------------------------------- +💾 Converting with 113.6651 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 300, 128) +Output shape: (1, 300, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.output: torch.Size([1, 300, 3584]) -> torch.Size([1, 1, 300, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,856B, BPFP=0.2529 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,772B, BPFP=0.3006 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,392B, BPFP=0.3329 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,580B, BPFP=0.2906 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,064B, BPFP=0.2637 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,208B, BPFP=0.2712 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,612B, BPFP=0.2402 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,492B, BPFP=0.2860 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,056B, BPFP=0.3675 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,820B, BPFP=0.3031 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,484B, BPFP=0.0185 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13491969 397.85091146 + layer.0.v_cache 0.00001558 0.03653471 + layer.1.k_cache 0.64590566 25.04375651 + layer.1.v_cache 0.00000556 0.01420106 + layer.2.k_cache 0.01395306 2.87955851 + layer.2.v_cache 0.00001972 0.04540236 + layer.3.k_cache 0.02379198 13.19572184 + layer.3.v_cache 0.00001942 0.05489421 + layer.4.k_cache 0.00073123 1.16362406 + layer.4.v_cache 0.00005274 0.09121257 + layer.4.output 0.05011183 181.05089286 + ------------------------------------------------------------------------------------- + TOTAL 0.06883514 100.45482749 + (elements=2,611,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2611200 +Total Bytes 58336 +BPFP 0.1787 bits/point +EBPFP 0.3575 equivalent bits/point +MSE 100.454827 +---------------------- -------------------------------------------------------- +Time: 3.189s Load: 0.010s, Pack+Encode: 1.834s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 100.4548 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,512B, BPFP=0.2423 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,956B, BPFP=0.3198 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,272B, BPFP=0.3368 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,796B, BPFP=0.3112 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,852B, BPFP=0.2605 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,656B, BPFP=0.3037 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,468B, BPFP=0.2399 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,828B, BPFP=0.3129 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,808B, BPFP=0.3655 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,724B, BPFP=0.3073 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,504B, BPFP=0.0192 +⌛️ [2/4] FRONTEND: Frontend time: 1.832s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14826349 393.04786834 + layer.0.v_cache 0.00001415 0.03670433 + layer.1.k_cache 0.61440683 25.47325360 + layer.1.v_cache 0.00000562 0.01480254 + layer.2.k_cache 0.00551110 2.82588180 + layer.2.v_cache 0.00001912 0.04474221 + layer.3.k_cache 0.01294636 13.31015088 + layer.3.v_cache 0.00002032 0.05136562 + layer.4.k_cache 0.00070633 1.11831141 + layer.4.v_cache 0.00005074 0.09043344 + layer.4.output 0.04972342 186.64250430 + ------------------------------------------------------------------------------------- + TOTAL 0.06647106 102.50064966 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 58376 +BPFP 0.1844 bits/point +EBPFP 0.3688 equivalent bits/point +MSE 102.500650 +---------------------- -------------------------------------------------------- +Time: 3.191s Load: 0.011s, Pack+Encode: 1.832s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 102.5006 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,580B, BPFP=0.2583 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,396B, BPFP=0.3608 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,176B, BPFP=0.3484 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,952B, BPFP=0.3357 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,628B, BPFP=0.2611 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,376B, BPFP=0.3032 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,336B, BPFP=0.2446 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,936B, BPFP=0.3348 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,608B, BPFP=0.3727 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,984B, BPFP=0.3375 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,416B, BPFP=0.0195 +⌛️ [2/4] FRONTEND: Frontend time: 1.835s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13861307 407.67243908 + layer.0.v_cache 0.00001419 0.03678924 + layer.1.k_cache 0.53867822 24.44155961 + layer.1.v_cache 0.00000573 0.01449087 + layer.2.k_cache 0.00871976 2.85621616 + layer.2.v_cache 0.00001877 0.04384540 + layer.3.k_cache 0.02195022 12.90868259 + layer.3.v_cache 0.00001967 0.05526675 + layer.4.k_cache 0.00069227 1.14380533 + layer.4.v_cache 0.00004961 0.08914191 + layer.4.output 0.00954030 200.64037519 + ------------------------------------------------------------------------------------- + TOTAL 0.04562021 109.04381549 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 58388 +BPFP 0.1937 bits/point +EBPFP 0.3875 equivalent bits/point +MSE 109.043815 +---------------------- -------------------------------------------------------- +Time: 3.213s Load: 0.010s, Pack+Encode: 1.835s, Decode+Unpack: 1.367s +---------------------- -------------------------------------------------------- +💾 Converting with 109.0438 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,764B, BPFP=0.2603 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,776B, BPFP=0.3156 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,724B, BPFP=0.3127 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,756B, BPFP=0.3145 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,520B, BPFP=0.2469 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,128B, BPFP=0.2802 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,116B, BPFP=0.2249 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,820B, BPFP=0.3180 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,488B, BPFP=0.3545 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,476B, BPFP=0.2992 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,184B, BPFP=0.0170 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13928204 395.78652207 + layer.0.v_cache 0.00001373 0.03571829 + layer.1.k_cache 0.60860016 25.17150452 + layer.1.v_cache 0.00000555 0.01440279 + layer.2.k_cache 0.01026906 2.73160580 + layer.2.v_cache 0.00001872 0.04479054 + layer.3.k_cache 0.05166252 12.82059420 + layer.3.v_cache 0.00001937 0.05509056 + layer.4.k_cache 0.00073448 1.12207682 + layer.4.v_cache 0.00005026 0.09191518 + layer.4.output 0.00697322 194.35935939 + ------------------------------------------------------------------------------------- + TOTAL 0.05055697 105.78763156 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 55752 +BPFP 0.1792 bits/point +EBPFP 0.3583 equivalent bits/point +MSE 105.787632 +---------------------- -------------------------------------------------------- +Time: 3.195s Load: 0.010s, Pack+Encode: 1.840s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 105.7876 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,996B, BPFP=0.2646 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,852B, BPFP=0.3100 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,372B, BPFP=0.3375 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,652B, BPFP=0.2994 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,948B, BPFP=0.2621 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,044B, BPFP=0.2672 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,588B, BPFP=0.2430 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,700B, BPFP=0.3019 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,684B, BPFP=0.3540 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,612B, BPFP=0.2972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,448B, BPFP=0.0185 +⌛️ [2/4] FRONTEND: Frontend time: 1.844s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13820392 395.42256356 + layer.0.v_cache 0.00001454 0.03684704 + layer.1.k_cache 0.61684219 23.59878509 + layer.1.v_cache 0.00000570 0.01410722 + layer.2.k_cache 0.01285210 2.84431649 + layer.2.v_cache 0.00002011 0.04405326 + layer.3.k_cache 0.05980752 12.75844230 + layer.3.v_cache 0.00002263 0.05329419 + layer.4.k_cache 0.00071023 1.14683755 + layer.4.v_cache 0.00005833 0.09250899 + layer.4.output 0.05086143 184.11923426 + ------------------------------------------------------------------------------------- + TOTAL 0.06968043 101.46155268 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 57896 +BPFP 0.1804 bits/point +EBPFP 0.3608 equivalent bits/point +MSE 101.461553 +---------------------- -------------------------------------------------------- +Time: 3.196s Load: 0.011s, Pack+Encode: 1.844s, Decode+Unpack: 1.341s +---------------------- -------------------------------------------------------- +💾 Converting with 101.4616 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,908B, BPFP=0.2548 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,052B, BPFP=0.3142 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,308B, BPFP=0.3275 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,792B, BPFP=0.3007 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,952B, BPFP=0.2571 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,688B, BPFP=0.2953 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,692B, BPFP=0.2436 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,196B, BPFP=0.3216 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,108B, BPFP=0.3690 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,000B, BPFP=0.3115 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,540B, BPFP=0.0188 +⌛️ [2/4] FRONTEND: Frontend time: 1.842s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14825445 398.01723422 + layer.0.v_cache 0.00001396 0.03744319 + layer.1.k_cache 0.63210902 24.27426158 + layer.1.v_cache 0.00000581 0.01487254 + layer.2.k_cache 0.01184132 2.90511568 + layer.2.v_cache 0.00001890 0.04577893 + layer.3.k_cache 0.01767834 13.37905874 + layer.3.v_cache 0.00002009 0.05411840 + layer.4.k_cache 0.00068272 1.16184029 + layer.4.v_cache 0.00005247 0.09351435 + layer.4.output 0.04867053 180.45566860 + ------------------------------------------------------------------------------------- + TOTAL 0.06772769 100.18664224 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 60236 +BPFP 0.1839 bits/point +EBPFP 0.3679 equivalent bits/point +MSE 100.186642 +---------------------- -------------------------------------------------------- +Time: 3.198s Load: 0.010s, Pack+Encode: 1.842s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 100.1866 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,416B, BPFP=0.2614 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,892B, BPFP=0.3487 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,176B, BPFP=0.3655 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,824B, BPFP=0.3447 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,584B, BPFP=0.2713 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,892B, BPFP=0.3487 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,084B, BPFP=0.2417 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,628B, BPFP=0.3331 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,508B, BPFP=0.3852 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,212B, BPFP=0.3677 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,388B, BPFP=0.0202 +⌛️ [2/4] FRONTEND: Frontend time: 1.837s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14315909 400.01532907 + layer.0.v_cache 0.00001391 0.03734213 + layer.1.k_cache 0.49711875 23.98093669 + layer.1.v_cache 0.00000560 0.01435266 + layer.2.k_cache 0.00646762 2.70795764 + layer.2.v_cache 0.00001905 0.04542320 + layer.3.k_cache 0.04707547 13.51882472 + layer.3.v_cache 0.00001924 0.05425143 + layer.4.k_cache 0.00071833 1.13339430 + layer.4.v_cache 0.00005197 0.09067897 + layer.4.output 0.00997522 210.52994792 + ------------------------------------------------------------------------------------- + TOTAL 0.04496915 112.66518390 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 57604 +BPFP 0.2005 bits/point +EBPFP 0.4011 equivalent bits/point +MSE 112.665184 +---------------------- -------------------------------------------------------- +Time: 3.186s Load: 0.009s, Pack+Encode: 1.837s, Decode+Unpack: 1.341s +---------------------- -------------------------------------------------------- +💾 Converting with 112.6652 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,444B, BPFP=0.2489 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,168B, BPFP=0.3454 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,116B, BPFP=0.3425 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,924B, BPFP=0.3318 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,748B, BPFP=0.2659 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,176B, BPFP=0.3459 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,324B, BPFP=0.2422 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,792B, BPFP=0.3244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,784B, BPFP=0.3799 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,052B, BPFP=0.3389 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,268B, BPFP=0.0181 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.356s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15581964 396.97538642 + layer.0.v_cache 0.00001703 0.03710820 + layer.1.k_cache 0.56298232 24.93487833 + layer.1.v_cache 0.00000550 0.01421319 + layer.2.k_cache 0.00979712 2.96700820 + layer.2.v_cache 0.00002023 0.04736460 + layer.3.k_cache 0.04960944 12.99426313 + layer.3.v_cache 0.00001971 0.05466029 + layer.4.k_cache 0.00068561 1.13237016 + layer.4.v_cache 0.00005245 0.09271192 + layer.4.output 0.01122868 199.18613991 + ------------------------------------------------------------------------------------- + TOTAL 0.05044764 107.85605552 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 58796 +BPFP 0.1937 bits/point +EBPFP 0.3874 equivalent bits/point +MSE 107.856056 +---------------------- -------------------------------------------------------- +Time: 3.206s Load: 0.010s, Pack+Encode: 1.840s, Decode+Unpack: 1.356s +---------------------- -------------------------------------------------------- +💾 Converting with 107.8561 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,904B, BPFP=0.2708 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,356B, BPFP=0.3509 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,824B, BPFP=0.3216 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,736B, BPFP=0.3167 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,768B, BPFP=0.2633 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,400B, BPFP=0.2981 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,456B, BPFP=0.2460 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,652B, BPFP=0.3121 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,656B, BPFP=0.3675 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,452B, BPFP=0.3010 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,384B, BPFP=0.0188 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13599485 402.16839664 + layer.0.v_cache 0.00001418 0.03673887 + layer.1.k_cache 0.56373192 23.00117326 + layer.1.v_cache 0.00000588 0.01476278 + layer.2.k_cache 0.01549433 2.92025358 + layer.2.v_cache 0.00001867 0.04472423 + layer.3.k_cache 0.03475572 13.86288355 + layer.3.v_cache 0.00001907 0.05147685 + layer.4.k_cache 0.00072111 1.12927624 + layer.4.v_cache 0.00005295 0.08987788 + layer.4.output 0.00820773 196.42871340 + ------------------------------------------------------------------------------------- + TOTAL 0.04754487 106.96003281 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 57588 +BPFP 0.1870 bits/point +EBPFP 0.3741 equivalent bits/point +MSE 106.960033 +---------------------- -------------------------------------------------------- +Time: 3.201s Load: 0.009s, Pack+Encode: 1.841s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 106.9600 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,916B, BPFP=0.2494 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,980B, BPFP=0.3034 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,388B, BPFP=0.3241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,792B, BPFP=0.2938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,880B, BPFP=0.2476 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,776B, BPFP=0.2930 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,484B, BPFP=0.2275 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,800B, BPFP=0.2942 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,900B, BPFP=0.3500 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,812B, BPFP=0.2948 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,376B, BPFP=0.0172 +⌛️ [2/4] FRONTEND: Frontend time: 1.845s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13331154 404.94308036 + layer.0.v_cache 0.00001460 0.03694697 + layer.1.k_cache 0.64535681 24.25127302 + layer.1.v_cache 0.00000563 0.01471309 + layer.2.k_cache 0.00762887 3.00062620 + layer.2.v_cache 0.00001910 0.04658693 + layer.3.k_cache 0.03538683 13.39574403 + layer.3.v_cache 0.00002103 0.05449504 + layer.4.k_cache 0.00072257 1.19757744 + layer.4.v_cache 0.00005300 0.09489581 + layer.4.output 0.04522907 176.27587256 + ------------------------------------------------------------------------------------- + TOTAL 0.06700726 98.88041452 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 59104 +BPFP 0.1764 bits/point +EBPFP 0.3528 equivalent bits/point +MSE 98.880415 +---------------------- -------------------------------------------------------- +Time: 3.205s Load: 0.012s, Pack+Encode: 1.845s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 98.8804 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,944B, BPFP=0.2720 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,044B, BPFP=0.3325 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,912B, BPFP=0.3253 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,768B, BPFP=0.3173 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,620B, BPFP=0.2542 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,344B, BPFP=0.2940 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,264B, BPFP=0.2346 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,616B, BPFP=0.3090 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,616B, BPFP=0.3640 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,500B, BPFP=0.3026 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,412B, BPFP=0.0190 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14806804 408.75861026 + layer.0.v_cache 0.00001376 0.03672199 + layer.1.k_cache 0.60662186 25.48195423 + layer.1.v_cache 0.00000571 0.01496173 + layer.2.k_cache 0.01239752 2.89397570 + layer.2.v_cache 0.00002027 0.04512660 + layer.3.k_cache 0.02456893 13.42667529 + layer.3.v_cache 0.00001975 0.05448198 + layer.4.k_cache 0.00069718 1.14285590 + layer.4.v_cache 0.00005379 0.09204495 + layer.4.output 0.00818182 195.73778609 + ------------------------------------------------------------------------------------- + TOTAL 0.04998468 107.18305360 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 57040 +BPFP 0.1846 bits/point +EBPFP 0.3692 equivalent bits/point +MSE 107.183054 +---------------------- -------------------------------------------------------- +Time: 3.189s Load: 0.011s, Pack+Encode: 1.833s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 107.1831 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,952B, BPFP=0.2734 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,408B, BPFP=0.3538 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,952B, BPFP=0.3286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,852B, BPFP=0.3231 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,804B, BPFP=0.2652 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,348B, BPFP=0.2953 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,392B, BPFP=0.2425 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,768B, BPFP=0.3185 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,704B, BPFP=0.3701 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,776B, BPFP=0.3189 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,284B, BPFP=0.0180 +⌛️ [2/4] FRONTEND: Frontend time: 1.842s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14592592 402.33552341 + layer.0.v_cache 0.00001391 0.03807608 + layer.1.k_cache 0.53254910 24.39871701 + layer.1.v_cache 0.00000583 0.01526220 + layer.2.k_cache 0.01029040 3.01282213 + layer.2.v_cache 0.00002009 0.04735873 + layer.3.k_cache 0.03848921 14.13703794 + layer.3.v_cache 0.00001972 0.05633647 + layer.4.k_cache 0.00069417 1.17956716 + layer.4.v_cache 0.00005427 0.09664397 + layer.4.output 0.01004909 196.41918539 + ------------------------------------------------------------------------------------- + TOTAL 0.04696507 107.07362605 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 58240 +BPFP 0.1891 bits/point +EBPFP 0.3783 equivalent bits/point +MSE 107.073626 +---------------------- -------------------------------------------------------- +Time: 3.195s Load: 0.010s, Pack+Encode: 1.842s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 107.0736 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,432B, BPFP=0.2546 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,360B, BPFP=0.3653 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,440B, BPFP=0.3699 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,152B, BPFP=0.3534 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,776B, BPFP=0.2744 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,476B, BPFP=0.3146 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,368B, BPFP=0.2509 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,176B, BPFP=0.3548 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,364B, BPFP=0.3656 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,100B, BPFP=0.3504 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,480B, BPFP=0.0204 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15384994 400.71906595 + layer.0.v_cache 0.00001404 0.03825306 + layer.1.k_cache 0.48768111 24.99818331 + layer.1.v_cache 0.00000584 0.01498154 + layer.2.k_cache 0.00538594 2.94535020 + layer.2.v_cache 0.00001948 0.04607011 + layer.3.k_cache 0.02347840 12.97233043 + layer.3.v_cache 0.00001899 0.05318390 + layer.4.k_cache 0.00068094 1.11936019 + layer.4.v_cache 0.00005116 0.09459173 + layer.4.output 0.00909345 204.36605502 + ------------------------------------------------------------------------------------- + TOTAL 0.04322588 110.20963268 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 59124 +BPFP 0.1998 bits/point +EBPFP 0.3996 equivalent bits/point +MSE 110.209633 +---------------------- -------------------------------------------------------- +Time: 3.214s Load: 0.009s, Pack+Encode: 1.840s, Decode+Unpack: 1.366s +---------------------- -------------------------------------------------------- +💾 Converting with 110.2096 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,660B, BPFP=0.2477 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,192B, BPFP=0.3291 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,368B, BPFP=0.3384 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,824B, BPFP=0.3095 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,036B, BPFP=0.2676 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,488B, BPFP=0.2917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,652B, BPFP=0.2472 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,580B, BPFP=0.2966 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,676B, BPFP=0.3548 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,916B, BPFP=0.3144 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,528B, BPFP=0.0192 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13453574 394.64963329 + layer.0.v_cache 0.00001386 0.03602743 + layer.1.k_cache 0.70462929 24.77795493 + layer.1.v_cache 0.00000580 0.01436334 + layer.2.k_cache 0.00896267 2.94570321 + layer.2.v_cache 0.00001881 0.04401866 + layer.3.k_cache 0.02468209 11.94247582 + layer.3.v_cache 0.00001940 0.05316148 + layer.4.k_cache 0.00068809 1.13414074 + layer.4.v_cache 0.00005209 0.08852855 + layer.4.output 0.04953035 184.74067663 + ------------------------------------------------------------------------------------- + TOTAL 0.07178355 101.69827905 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 58920 +BPFP 0.1842 bits/point +EBPFP 0.3684 equivalent bits/point +MSE 101.698279 +---------------------- -------------------------------------------------------- +Time: 3.198s Load: 0.011s, Pack+Encode: 1.834s, Decode+Unpack: 1.353s +---------------------- -------------------------------------------------------- +💾 Converting with 101.6983 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,452B, BPFP=0.2596 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,004B, BPFP=0.3500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,452B, BPFP=0.3762 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,816B, BPFP=0.3391 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,528B, BPFP=0.2640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,664B, BPFP=0.3302 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,992B, BPFP=0.2327 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,108B, BPFP=0.3561 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,632B, BPFP=0.3867 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,204B, BPFP=0.3617 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,524B, BPFP=0.0210 +⌛️ [2/4] FRONTEND: Frontend time: 1.848s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150386 393.34823927 + layer.0.v_cache 0.00001401 0.03683094 + layer.1.k_cache 0.57442873 25.52511551 + layer.1.v_cache 0.00000554 0.01463375 + layer.2.k_cache 0.00779028 2.91762008 + layer.2.v_cache 0.00001875 0.04583685 + layer.3.k_cache 0.05176373 12.80465015 + layer.3.v_cache 0.00002027 0.05458636 + layer.4.k_cache 0.00069872 1.10910854 + layer.4.v_cache 0.00005023 0.09394198 + layer.4.output 0.01044900 207.39983675 + ------------------------------------------------------------------------------------- + TOTAL 0.04937866 111.04408357 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 58376 +BPFP 0.2002 bits/point +EBPFP 0.4004 equivalent bits/point +MSE 111.044084 +---------------------- -------------------------------------------------------- +Time: 3.214s Load: 0.008s, Pack+Encode: 1.848s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 111.0441 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.1875 bits/point +Avg EBPFP 0.3751 equivalent bits/point +Avg MSE 105.354782 +Avg Time 3.534s +------------------------ ---------------------------- diff --git a/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..b974ff5c3191e4aeb91c0645ed04c430a7faeaea --- /dev/null +++ b/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 286 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- -------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.001_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa +Output output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa +---------------- -------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,208B, BPFP=0.2247 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,456B, BPFP=0.4568 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,760B, BPFP=0.5134 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,280B, BPFP=0.4241 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,428B, BPFP=0.4516 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,440B, BPFP=0.4539 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,324B, BPFP=0.4323 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,000B, BPFP=0.3720 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,096B, BPFP=0.5759 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,704B, BPFP=0.5030 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,404B, BPFP=0.0373 +⌛️ [2/4] FRONTEND: Frontend time: 1.954s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.035s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14073507 332.56826637 + layer.0.v_cache 0.00001426 0.03987856 + layer.1.k_cache 0.05042761 24.61150251 + layer.1.v_cache 0.00000517 0.01430769 + layer.2.k_cache 0.00244098 3.81171381 + layer.2.v_cache 0.00001715 0.04200668 + layer.3.k_cache 0.02726505 16.41191755 + layer.3.v_cache 0.00001796 0.05137641 + layer.4.k_cache 0.00069123 1.26465534 + layer.4.v_cache 0.00005084 0.09840502 + layer.4.output 0.16186065 645.73634141 + ------------------------------------------------------------------------------------- + TOTAL 0.07968764 288.18049528 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 25100 +BPFP 0.2746 bits/point +EBPFP 0.5493 equivalent bits/point +MSE 288.180495 +---------------------- -------------------------------------------------------- +Time: 2.993s Load: 0.004s, Pack+Encode: 1.954s, Decode+Unpack: 1.035s +---------------------- -------------------------------------------------------- +💾 Converting with 288.1805 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,192B, BPFP=0.2244 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,084B, BPFP=0.5806 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,984B, BPFP=0.5617 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,168B, BPFP=0.4081 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,536B, BPFP=0.4774 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,508B, BPFP=0.4721 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,276B, BPFP=0.4285 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,376B, BPFP=0.4473 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,016B, BPFP=0.5678 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,812B, BPFP=0.5294 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,396B, BPFP=0.0375 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.002s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11392649 310.20705478 + layer.0.v_cache 0.00001670 0.03754311 + layer.1.k_cache 0.05249626 25.73112764 + layer.1.v_cache 0.00000506 0.01422943 + layer.2.k_cache 0.00416583 3.90614999 + layer.2.v_cache 0.00001711 0.04063819 + layer.3.k_cache 0.05931453 16.63464797 + layer.3.v_cache 0.00001789 0.04866602 + layer.4.k_cache 0.00069248 1.21148277 + layer.4.v_cache 0.00004729 0.09468103 + layer.4.output 0.16383441 653.51780336 + ------------------------------------------------------------------------------------- + TOTAL 0.08103179 290.15004967 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 26348 +BPFP 0.2918 bits/point +EBPFP 0.5835 equivalent bits/point +MSE 290.150050 +---------------------- -------------------------------------------------------- +Time: 2.491s Load: 0.005s, Pack+Encode: 1.484s, Decode+Unpack: 1.002s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1500 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,360B, BPFP=0.2261 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,612B, BPFP=0.4342 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,624B, BPFP=0.4362 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,084B, BPFP=0.3464 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,380B, BPFP=0.3956 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,192B, BPFP=0.3644 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,232B, BPFP=0.3710 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,576B, BPFP=0.4282 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,800B, BPFP=0.4654 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,568B, BPFP=0.4269 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,280B, BPFP=0.0304 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.000s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17268721 337.82047872 + layer.0.v_cache 0.00001453 0.03785097 + layer.1.k_cache 0.08276152 24.87691936 + layer.1.v_cache 0.00000545 0.01368669 + layer.2.k_cache 0.00739249 3.83041122 + layer.2.v_cache 0.00001625 0.03853497 + layer.3.k_cache 0.05929063 15.94789675 + layer.3.v_cache 0.00001766 0.04726248 + layer.4.k_cache 0.00068371 1.14325876 + layer.4.v_cache 0.00005055 0.08875183 + layer.4.output 0.14466690 577.01923442 + ------------------------------------------------------------------------------------- + TOTAL 0.07856402 260.17527604 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 24708 +BPFP 0.2416 bits/point +EBPFP 0.4832 equivalent bits/point +MSE 260.175276 +---------------------- -------------------------------------------------------- +Time: 2.482s Load: 0.004s, Pack+Encode: 1.478s, Decode+Unpack: 1.000s +---------------------- -------------------------------------------------------- +💾 Converting with 260.1753 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 112, 128) +Output shape: (1, 112, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.output: torch.Size([1, 112, 3584]) -> torch.Size([1, 1, 112, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,672B, BPFP=0.2333 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,984B, BPFP=0.4163 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,108B, BPFP=0.4336 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,712B, BPFP=0.3783 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,616B, BPFP=0.3650 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,552B, BPFP=0.3560 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,540B, BPFP=0.3544 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,812B, BPFP=0.3923 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,488B, BPFP=0.4866 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,748B, BPFP=0.3834 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,660B, BPFP=0.0331 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17321212 366.53428432 + layer.0.v_cache 0.00001450 0.04028177 + layer.1.k_cache 0.08643453 24.58380127 + layer.1.v_cache 0.00000659 0.01539149 + layer.2.k_cache 0.00690480 3.43548856 + layer.2.v_cache 0.00001899 0.04328391 + layer.3.k_cache 0.01205242 15.11976188 + layer.3.v_cache 0.00001923 0.05316761 + layer.4.k_cache 0.00071999 1.26088769 + layer.4.v_cache 0.00004947 0.08993302 + layer.4.output 10.17892331 479.73317921 + ------------------------------------------------------------------------------------- + TOTAL 4.20775858 221.72403153 + (elements=974,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 974848 +Total Bytes 28892 +BPFP 0.2371 bits/point +EBPFP 0.4742 equivalent bits/point +MSE 221.724032 +---------------------- -------------------------------------------------------- +Time: 2.489s Load: 0.006s, Pack+Encode: 1.484s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 221.7240 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,308B, BPFP=0.2322 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,512B, BPFP=0.4460 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,608B, BPFP=0.4631 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,276B, BPFP=0.4041 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,376B, BPFP=0.4219 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,632B, BPFP=0.4673 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,384B, BPFP=0.4233 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,428B, BPFP=0.4311 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,992B, BPFP=0.5312 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,824B, BPFP=0.5014 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,528B, BPFP=0.0388 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.000s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11687154 327.99298651 + layer.0.v_cache 0.00001423 0.03962402 + layer.1.k_cache 0.05324238 25.62157648 + layer.1.v_cache 0.00000539 0.01449703 + layer.2.k_cache 0.00397935 3.77885680 + layer.2.v_cache 0.00001766 0.04026114 + layer.3.k_cache 0.01210797 16.93039773 + layer.3.v_cache 0.00001803 0.05151519 + layer.4.k_cache 0.00075037 1.22700856 + layer.4.v_cache 0.00004481 0.08845878 + layer.4.output 0.15448949 616.36196226 + ------------------------------------------------------------------------------------- + TOTAL 0.07461636 275.90111283 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 25868 +BPFP 0.2702 bits/point +EBPFP 0.5404 equivalent bits/point +MSE 275.901113 +---------------------- -------------------------------------------------------- +Time: 2.487s Load: 0.005s, Pack+Encode: 1.482s, Decode+Unpack: 1.000s +---------------------- -------------------------------------------------------- +💾 Converting with 275.9011 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,256B, BPFP=0.2423 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,564B, BPFP=0.4946 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,112B, BPFP=0.6003 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,248B, BPFP=0.4336 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,440B, BPFP=0.4707 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,372B, BPFP=0.4576 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,364B, BPFP=0.4560 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,396B, BPFP=0.4622 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,084B, BPFP=0.5949 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,012B, BPFP=0.5810 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,368B, BPFP=0.0377 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203883 332.19854360 + layer.0.v_cache 0.00001438 0.04108609 + layer.1.k_cache 0.05633916 26.08766758 + layer.1.v_cache 0.00000602 0.01643631 + layer.2.k_cache 0.00244657 3.92033933 + layer.2.v_cache 0.00001847 0.04592289 + layer.3.k_cache 0.05492775 17.20913809 + layer.3.v_cache 0.00001883 0.05313809 + layer.4.k_cache 0.00062984 1.21212769 + layer.4.v_cache 0.00005178 0.10459723 + layer.4.output 0.18538215 670.16468254 + ------------------------------------------------------------------------------------- + TOTAL 0.09142157 298.35539851 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 26216 +BPFP 0.2975 bits/point +EBPFP 0.5950 equivalent bits/point +MSE 298.355399 +---------------------- -------------------------------------------------------- +Time: 2.487s Load: 0.005s, Pack+Encode: 1.485s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 298.3554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,352B, BPFP=0.2347 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,592B, BPFP=0.4500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,616B, BPFP=0.4542 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,328B, BPFP=0.4042 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,624B, BPFP=0.4556 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,260B, BPFP=0.3924 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,012B, BPFP=0.3493 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,596B, BPFP=0.4507 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,788B, BPFP=0.4840 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,768B, BPFP=0.4806 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0347 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13028670 339.96631944 + layer.0.v_cache 0.00001462 0.03823502 + layer.1.k_cache 0.04946813 25.07871094 + layer.1.v_cache 0.00000624 0.01401285 + layer.2.k_cache 0.00411291 3.85309720 + layer.2.v_cache 0.00001729 0.04108397 + layer.3.k_cache 0.01793825 16.39147542 + layer.3.v_cache 0.00001895 0.05315004 + layer.4.k_cache 0.00077120 1.23687456 + layer.4.v_cache 0.00004532 0.08990432 + layer.4.output 0.15109633 602.57802579 + ------------------------------------------------------------------------------------- + TOTAL 0.07413847 270.87112025 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 25336 +BPFP 0.2587 bits/point +EBPFP 0.5175 equivalent bits/point +MSE 270.871120 +---------------------- -------------------------------------------------------- +Time: 2.502s Load: 0.005s, Pack+Encode: 1.496s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 270.8711 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,408B, BPFP=0.3860 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,672B, BPFP=0.4583 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,856B, BPFP=0.5088 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,472B, BPFP=0.4035 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,428B, BPFP=0.3914 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,424B, BPFP=0.3904 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,256B, BPFP=0.3443 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,712B, BPFP=0.4693 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,676B, BPFP=0.4594 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,644B, BPFP=0.4507 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,408B, BPFP=0.0943 +⌛️ [2/4] FRONTEND: Frontend time: 1.734s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.980s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11683119 462.07846765 + layer.0.v_cache 0.00001514 0.05030798 + layer.1.k_cache 0.01218004 29.83317057 + layer.1.v_cache 0.00000521 0.01809211 + layer.2.k_cache 0.00491684 4.99360844 + layer.2.v_cache 0.00001706 0.05094347 + layer.3.k_cache 0.10823302 18.12976717 + layer.3.v_cache 0.00002043 0.07012984 + layer.4.k_cache 0.00062874 1.60063011 + layer.4.v_cache 0.00004794 0.12380128 + layer.4.output 0.23833110 951.22924499 + ------------------------------------------------------------------------------------- + TOTAL 0.11242431 422.09139021 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 17956 +BPFP 0.2895 bits/point +EBPFP 0.5791 equivalent bits/point +MSE 422.091390 +---------------------- -------------------------------------------------------- +Time: 2.717s Load: 0.002s, Pack+Encode: 1.734s, Decode+Unpack: 0.980s +---------------------- -------------------------------------------------------- +💾 Converting with 422.0914 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,692B, BPFP=0.5288 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,872B, BPFP=0.5850 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,992B, BPFP=0.6225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,404B, BPFP=0.4387 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,620B, BPFP=0.5062 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,796B, BPFP=0.5613 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,320B, BPFP=0.4125 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,848B, BPFP=0.5775 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,048B, BPFP=0.6400 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,892B, BPFP=0.5913 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,832B, BPFP=0.0818 +⌛️ [2/4] FRONTEND: Frontend time: 1.532s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.971s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14598725 432.92437500 + layer.0.v_cache 0.00001449 0.04764938 + layer.1.k_cache 0.01347294 27.08503906 + layer.1.v_cache 0.00000521 0.01620118 + layer.2.k_cache 0.00488209 4.50988495 + layer.2.v_cache 0.00001649 0.04665264 + layer.3.k_cache 0.04421538 16.23677490 + layer.3.v_cache 0.00001905 0.06282004 + layer.4.k_cache 0.00073779 1.47419891 + layer.4.v_cache 0.00005089 0.11538689 + layer.4.output 0.27158585 1084.46669643 + ------------------------------------------------------------------------------------- + TOTAL 0.12414721 474.92857988 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 19316 +BPFP 0.3551 bits/point +EBPFP 0.7101 equivalent bits/point +MSE 474.928580 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.002s, Pack+Encode: 1.532s, Decode+Unpack: 0.971s +---------------------- -------------------------------------------------------- +💾 Converting with 474.9286 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 46, 128) +Output shape: (1, 46, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.output: torch.Size([1, 46, 3584]) -> torch.Size([1, 1, 46, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,320B, BPFP=0.4484 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,796B, BPFP=0.6101 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,024B, BPFP=0.6875 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,708B, BPFP=0.5802 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,684B, BPFP=0.5720 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,428B, BPFP=0.4851 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,708B, BPFP=0.5802 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,776B, BPFP=0.6033 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,068B, BPFP=0.7024 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,868B, BPFP=0.6345 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,736B, BPFP=0.0842 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.967s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14129612 423.71679688 + layer.0.v_cache 0.00001532 0.04679297 + layer.1.k_cache 0.01315293 26.49132770 + layer.1.v_cache 0.00000553 0.01597712 + layer.2.k_cache 0.00779452 4.26212941 + layer.2.v_cache 0.00001723 0.04425094 + layer.3.k_cache 0.01828384 17.95194278 + layer.3.v_cache 0.00001944 0.06300858 + layer.4.k_cache 0.00060559 1.37459830 + layer.4.v_cache 0.00005113 0.11407965 + layer.4.output 0.29517000 1179.04260481 + ------------------------------------------------------------------------------------- + TOTAL 0.13220186 513.37524341 + (elements=400,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 400384 +Total Bytes 19116 +BPFP 0.3820 bits/point +EBPFP 0.7639 equivalent bits/point +MSE 513.375243 +---------------------- -------------------------------------------------------- +Time: 2.452s Load: 0.003s, Pack+Encode: 1.482s, Decode+Unpack: 0.967s +---------------------- -------------------------------------------------------- +💾 Converting with 513.3752 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 110, 128) +Output shape: (1, 110, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.output: torch.Size([1, 110, 3584]) -> torch.Size([1, 1, 110, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,604B, BPFP=0.2278 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,732B, BPFP=0.3881 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,004B, BPFP=0.4267 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,716B, BPFP=0.3858 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,600B, BPFP=0.3693 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,512B, BPFP=0.3568 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,364B, BPFP=0.3358 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,640B, BPFP=0.3750 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,188B, BPFP=0.4528 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,852B, BPFP=0.4051 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,608B, BPFP=0.0326 +⌛️ [2/4] FRONTEND: Frontend time: 1.479s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.000s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14476244 352.72038352 + layer.0.v_cache 0.00001476 0.04103338 + layer.1.k_cache 0.05665370 24.89688388 + layer.1.v_cache 0.00000525 0.01426403 + layer.2.k_cache 0.00786287 3.59603937 + layer.2.v_cache 0.00001801 0.04078280 + layer.3.k_cache 0.01410385 14.64342152 + layer.3.v_cache 0.00002027 0.05378119 + layer.4.k_cache 0.00066873 1.21066534 + layer.4.v_cache 0.00004626 0.08923427 + layer.4.output 10.36401748 488.45584416 + ------------------------------------------------------------------------------------- + TOTAL 4.28072226 224.49984696 + (elements=957,440) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 957440 +Total Bytes 27820 +BPFP 0.2325 bits/point +EBPFP 0.4649 equivalent bits/point +MSE 224.499847 +---------------------- -------------------------------------------------------- +Time: 2.484s Load: 0.005s, Pack+Encode: 1.479s, Decode+Unpack: 1.000s +---------------------- -------------------------------------------------------- +💾 Converting with 224.4998 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,188B, BPFP=0.2292 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,776B, BPFP=0.5355 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,056B, BPFP=0.5895 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,160B, BPFP=0.4167 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,780B, BPFP=0.5363 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,096B, BPFP=0.4043 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,648B, BPFP=0.5108 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,752B, BPFP=0.5309 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,172B, BPFP=0.6119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,056B, BPFP=0.5895 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,372B, BPFP=0.0378 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10533771 315.19077932 + layer.0.v_cache 0.00001610 0.04854797 + layer.1.k_cache 0.05553475 26.39364812 + layer.1.v_cache 0.00000748 0.01845092 + layer.2.k_cache 0.02166171 4.17278525 + layer.2.v_cache 0.00002000 0.05232245 + layer.3.k_cache 0.06386927 17.76906859 + layer.3.v_cache 0.00001944 0.06146557 + layer.4.k_cache 0.00064315 1.35061872 + layer.4.v_cache 0.00005385 0.10873399 + layer.4.output 0.16801396 669.62604718 + ------------------------------------------------------------------------------------- + TOTAL 0.08372125 297.20875007 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 27056 +BPFP 0.3070 bits/point +EBPFP 0.6140 equivalent bits/point +MSE 297.208750 +---------------------- -------------------------------------------------------- +Time: 2.486s Load: 0.005s, Pack+Encode: 1.484s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 297.2088 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 49, 128) +Output shape: (1, 49, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.output: torch.Size([1, 49, 3584]) -> torch.Size([1, 1, 49, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,308B, BPFP=0.4171 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,924B, BPFP=0.6135 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,016B, BPFP=0.6429 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,552B, BPFP=0.4949 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,868B, BPFP=0.5957 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,752B, BPFP=0.5587 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,156B, BPFP=0.3686 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,872B, BPFP=0.5969 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,980B, BPFP=0.6314 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,876B, BPFP=0.5982 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,864B, BPFP=0.0849 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.966s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12410851 391.17594069 + layer.0.v_cache 0.00001740 0.04692048 + layer.1.k_cache 0.01401279 27.81705646 + layer.1.v_cache 0.00000579 0.01752338 + layer.2.k_cache 0.00545398 4.85749070 + layer.2.v_cache 0.00001955 0.05114914 + layer.3.k_cache 0.07546502 16.54327766 + layer.3.v_cache 0.00001978 0.06548524 + layer.4.k_cache 0.00063206 1.48121316 + layer.4.v_cache 0.00005205 0.11766294 + layer.4.output 0.27729562 1106.73742711 + ------------------------------------------------------------------------------------- + TOTAL 0.12710919 481.72562998 + (elements=426,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 426496 +Total Bytes 19168 +BPFP 0.3595 bits/point +EBPFP 0.7191 equivalent bits/point +MSE 481.725630 +---------------------- -------------------------------------------------------- +Time: 2.454s Load: 0.002s, Pack+Encode: 1.485s, Decode+Unpack: 0.966s +---------------------- -------------------------------------------------------- +💾 Converting with 481.7256 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,356B, BPFP=0.3717 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,748B, BPFP=0.4792 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,880B, BPFP=0.5154 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,628B, BPFP=0.4463 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,420B, BPFP=0.3893 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,236B, BPFP=0.3388 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,288B, BPFP=0.3531 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,468B, BPFP=0.4024 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,712B, BPFP=0.4693 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,744B, BPFP=0.4781 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,376B, BPFP=0.0930 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.965s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19197876 450.01754386 + layer.0.v_cache 0.00001449 0.05052251 + layer.1.k_cache 0.01466701 29.47093656 + layer.1.v_cache 0.00000519 0.01851260 + layer.2.k_cache 0.00720649 5.10088522 + layer.2.v_cache 0.00001785 0.04893718 + layer.3.k_cache 0.10851213 17.85943068 + layer.3.v_cache 0.00001996 0.06474988 + layer.4.k_cache 0.00062085 1.64748222 + layer.4.v_cache 0.00004770 0.12278229 + layer.4.output 0.23832821 951.25454261 + ------------------------------------------------------------------------------------- + TOTAL 0.11714047 421.36374007 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 17856 +BPFP 0.2879 bits/point +EBPFP 0.5759 equivalent bits/point +MSE 421.363740 +---------------------- -------------------------------------------------------- +Time: 2.447s Load: 0.002s, Pack+Encode: 1.480s, Decode+Unpack: 0.965s +---------------------- -------------------------------------------------------- +💾 Converting with 421.3637 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 856B, BPFP=0.2623 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,584B, BPFP=0.4853 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,032B, BPFP=0.6225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,344B, BPFP=0.4118 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,516B, BPFP=0.4645 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 968B, BPFP=0.2966 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,300B, BPFP=0.3983 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,200B, BPFP=0.3676 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,004B, BPFP=0.6140 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,756B, BPFP=0.5380 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,956B, BPFP=0.0856 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.968s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12665685 344.43830423 + layer.0.v_cache 0.00001436 0.04482917 + layer.1.k_cache 0.01497617 27.82992494 + layer.1.v_cache 0.00000537 0.01576014 + layer.2.k_cache 0.00986420 4.72193310 + layer.2.v_cache 0.00001802 0.04582413 + layer.3.k_cache 0.09846177 16.92407227 + layer.3.v_cache 0.00001923 0.06107238 + layer.4.k_cache 0.00062397 1.42264452 + layer.4.v_cache 0.00005014 0.11408897 + layer.4.output 0.26631049 1063.43995098 + ------------------------------------------------------------------------------------- + TOTAL 0.12440374 461.15871239 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 16516 +BPFP 0.2976 bits/point +EBPFP 0.5953 equivalent bits/point +MSE 461.158712 +---------------------- -------------------------------------------------------- +Time: 2.451s Load: 0.002s, Pack+Encode: 1.481s, Decode+Unpack: 0.968s +---------------------- -------------------------------------------------------- +💾 Converting with 461.1587 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,448B, BPFP=0.4714 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,844B, BPFP=0.6003 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,036B, BPFP=0.6628 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,872B, BPFP=0.6094 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,836B, BPFP=0.5977 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 984B, BPFP=0.3203 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,404B, BPFP=0.4570 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,736B, BPFP=0.5651 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.6536 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,912B, BPFP=0.6224 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,640B, BPFP=0.0763 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.966s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18531116 419.50524902 + layer.0.v_cache 0.00001874 0.04989905 + layer.1.k_cache 0.01459585 26.72410583 + layer.1.v_cache 0.00000674 0.01873790 + layer.2.k_cache 0.01042949 4.50177542 + layer.2.v_cache 0.00001841 0.05085448 + layer.3.k_cache 0.15260293 17.08067576 + layer.3.v_cache 0.00002390 0.07138410 + layer.4.k_cache 0.00061422 1.42036645 + layer.4.v_cache 0.00004881 0.12087185 + layer.4.output 0.28297247 1129.93480283 + ------------------------------------------------------------------------------------- + TOTAL 0.13791044 492.88750233 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 18720 +BPFP 0.3585 bits/point +EBPFP 0.7169 equivalent bits/point +MSE 492.887502 +---------------------- -------------------------------------------------------- +Time: 2.449s Load: 0.002s, Pack+Encode: 1.481s, Decode+Unpack: 0.966s +---------------------- -------------------------------------------------------- +💾 Converting with 492.8875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,344B, BPFP=0.3559 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,528B, BPFP=0.4047 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,624B, BPFP=0.4301 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,440B, BPFP=0.3814 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,344B, BPFP=0.3559 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,024B, BPFP=0.2712 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,288B, BPFP=0.3411 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,572B, BPFP=0.4163 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,612B, BPFP=0.4269 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,656B, BPFP=0.4386 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,508B, BPFP=0.0571 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.968s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12052289 446.62327860 + layer.0.v_cache 0.00001546 0.04862587 + layer.1.k_cache 0.01434235 29.12183031 + layer.1.v_cache 0.00000574 0.01884473 + layer.2.k_cache 0.00254285 5.31918956 + layer.2.v_cache 0.00001710 0.05064731 + layer.3.k_cache 0.03737745 18.93817863 + layer.3.v_cache 0.00001946 0.07293818 + layer.4.k_cache 0.00063766 1.61580283 + layer.4.v_cache 0.00005039 0.12618514 + layer.4.output 0.23032015 919.00431295 + ------------------------------------------------------------------------------------- + TOTAL 0.10516308 407.93915952 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 15940 +BPFP 0.2483 bits/point +EBPFP 0.4966 equivalent bits/point +MSE 407.939160 +---------------------- -------------------------------------------------------- +Time: 2.450s Load: 0.002s, Pack+Encode: 1.480s, Decode+Unpack: 0.968s +---------------------- -------------------------------------------------------- +💾 Converting with 407.9392 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,188B, BPFP=0.2508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,884B, BPFP=0.6090 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,284B, BPFP=0.4823 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,108B, BPFP=0.4451 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,276B, BPFP=0.4806 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,804B, BPFP=0.3809 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,128B, BPFP=0.4493 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,396B, BPFP=0.5059 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,836B, BPFP=0.5988 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,036B, BPFP=0.6410 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,304B, BPFP=0.0393 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11335979 336.91992188 + layer.0.v_cache 0.00001513 0.04457422 + layer.1.k_cache 0.01655370 26.36279462 + layer.1.v_cache 0.00000667 0.01605189 + layer.2.k_cache 0.00428703 4.09756635 + layer.2.v_cache 0.00001789 0.04857033 + layer.3.k_cache 0.04938591 16.65365683 + layer.3.v_cache 0.00002111 0.05992128 + layer.4.k_cache 0.00066117 1.31693866 + layer.4.v_cache 0.00004716 0.09656985 + layer.4.output 0.18372514 732.94280888 + ------------------------------------------------------------------------------------- + TOTAL 0.08649597 324.48330753 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 24244 +BPFP 0.3011 bits/point +EBPFP 0.6022 equivalent bits/point +MSE 324.483308 +---------------------- -------------------------------------------------------- +Time: 2.477s Load: 0.003s, Pack+Encode: 1.476s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 324.4833 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,436B, BPFP=0.5099 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,848B, BPFP=0.6562 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,208B, BPFP=0.7841 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,524B, BPFP=0.5412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,992B, BPFP=0.7074 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,240B, BPFP=0.4403 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,528B, BPFP=0.5426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,792B, BPFP=0.6364 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,120B, BPFP=0.7528 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,952B, BPFP=0.6932 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,708B, BPFP=0.0866 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.965s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14777968 415.58527166 + layer.0.v_cache 0.00001531 0.04984305 + layer.1.k_cache 0.01728138 26.08903920 + layer.1.v_cache 0.00000547 0.01769109 + layer.2.k_cache 0.00823456 4.23413363 + layer.2.v_cache 0.00001932 0.05419491 + layer.3.k_cache 0.16770924 16.92115506 + layer.3.v_cache 0.00002046 0.06866205 + layer.4.k_cache 0.00061569 1.38003809 + layer.4.v_cache 0.00005592 0.13817607 + layer.4.output 0.30867824 1232.65969968 + ------------------------------------------------------------------------------------- + TOTAL 0.14720498 534.89153544 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 19348 +BPFP 0.4042 bits/point +EBPFP 0.8083 equivalent bits/point +MSE 534.891535 +---------------------- -------------------------------------------------------- +Time: 2.451s Load: 0.002s, Pack+Encode: 1.485s, Decode+Unpack: 0.965s +---------------------- -------------------------------------------------------- +💾 Converting with 534.8915 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,192B, BPFP=0.2388 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,788B, BPFP=0.5585 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,944B, BPFP=0.5897 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,928B, BPFP=0.5865 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,452B, BPFP=0.4912 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,296B, BPFP=0.4599 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,284B, BPFP=0.4575 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,560B, BPFP=0.5128 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,004B, BPFP=0.6018 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,876B, BPFP=0.5761 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0401 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.996s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14408422 339.32925180 + layer.0.v_cache 0.00002097 0.04595315 + layer.1.k_cache 0.05735342 26.11393542 + layer.1.v_cache 0.00000575 0.01647180 + layer.2.k_cache 0.01678689 4.08216975 + layer.2.v_cache 0.00001894 0.04748894 + layer.3.k_cache 0.03029843 16.40008075 + layer.3.v_cache 0.00002125 0.06206844 + layer.4.k_cache 0.00062582 1.28755168 + layer.4.v_cache 0.00005050 0.10922635 + layer.4.output 0.17433185 695.42439332 + ------------------------------------------------------------------------------------- + TOTAL 0.08644642 309.14499714 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 26724 +BPFP 0.3149 bits/point +EBPFP 0.6298 equivalent bits/point +MSE 309.144997 +---------------------- -------------------------------------------------------- +Time: 2.477s Load: 0.003s, Pack+Encode: 1.478s, Decode+Unpack: 0.996s +---------------------- -------------------------------------------------------- +💾 Converting with 309.1450 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,420B, BPFP=0.5043 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,880B, BPFP=0.6676 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,220B, BPFP=0.7884 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,772B, BPFP=0.6293 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,988B, BPFP=0.7060 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,476B, BPFP=0.5241 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,444B, BPFP=0.5128 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,072B, BPFP=0.3807 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,124B, BPFP=0.7543 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,896B, BPFP=0.6733 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,708B, BPFP=0.0866 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.965s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13243668 424.68554688 + layer.0.v_cache 0.00001576 0.04903831 + layer.1.k_cache 0.01981235 26.38778409 + layer.1.v_cache 0.00000577 0.01734806 + layer.2.k_cache 0.01756107 4.33143477 + layer.2.v_cache 0.00001728 0.04979955 + layer.3.k_cache 0.04920139 17.12866211 + layer.3.v_cache 0.00001837 0.06118183 + layer.4.k_cache 0.00061593 1.42532851 + layer.4.v_cache 0.00004907 0.11747267 + layer.4.output 0.30854983 1232.61201299 + ------------------------------------------------------------------------------------- + TOTAL 0.13997544 535.44339339 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 19000 +BPFP 0.3969 bits/point +EBPFP 0.7938 equivalent bits/point +MSE 535.443393 +---------------------- -------------------------------------------------------- +Time: 2.453s Load: 0.003s, Pack+Encode: 1.485s, Decode+Unpack: 0.965s +---------------------- -------------------------------------------------------- +💾 Converting with 535.4434 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,828B, BPFP=0.5713 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,600B, BPFP=0.5000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,116B, BPFP=0.6613 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,848B, BPFP=0.5775 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,788B, BPFP=0.5587 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,192B, BPFP=0.3725 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,364B, BPFP=0.4263 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,768B, BPFP=0.5525 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,944B, BPFP=0.6075 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,912B, BPFP=0.5975 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,840B, BPFP=0.0821 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.972s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20666794 437.92363281 + layer.0.v_cache 0.00002565 0.04928163 + layer.1.k_cache 0.01518770 28.15884033 + layer.1.v_cache 0.00000581 0.01787339 + layer.2.k_cache 0.01031687 4.81987030 + layer.2.v_cache 0.00001873 0.05249388 + layer.3.k_cache 0.09638391 16.20867676 + layer.3.v_cache 0.00001967 0.06826891 + layer.4.k_cache 0.00062172 1.41588043 + layer.4.v_cache 0.00004827 0.12520308 + layer.4.output 0.27162739 1084.60705357 + ------------------------------------------------------------------------------------- + TOTAL 0.13121694 475.35819980 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 19200 +BPFP 0.3529 bits/point +EBPFP 0.7059 equivalent bits/point +MSE 475.358200 +---------------------- -------------------------------------------------------- +Time: 2.456s Load: 0.002s, Pack+Encode: 1.483s, Decode+Unpack: 0.972s +---------------------- -------------------------------------------------------- +💾 Converting with 475.3582 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 996B, BPFP=0.2289 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,136B, BPFP=0.4908 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,236B, BPFP=0.5138 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,808B, BPFP=0.4154 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,056B, BPFP=0.4724 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,672B, BPFP=0.3842 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,632B, BPFP=0.3750 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,536B, BPFP=0.3529 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,620B, BPFP=0.6020 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,468B, BPFP=0.5671 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,408B, BPFP=0.0462 +⌛️ [2/4] FRONTEND: Frontend time: 1.580s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13431002 328.08030790 + layer.0.v_cache 0.00001720 0.04686797 + layer.1.k_cache 0.01179634 26.69375072 + layer.1.v_cache 0.00000537 0.01587391 + layer.2.k_cache 0.00798416 3.99694375 + layer.2.v_cache 0.00001783 0.04956611 + layer.3.k_cache 0.05795463 16.64015647 + layer.3.v_cache 0.00001974 0.06480896 + layer.4.k_cache 0.00062059 1.29564409 + layer.4.v_cache 0.00005204 0.11237377 + layer.4.output 0.19989206 797.63366597 + ------------------------------------------------------------------------------------- + TOTAL 0.09482484 350.61364444 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 20568 +BPFP 0.2780 bits/point +EBPFP 0.5560 equivalent bits/point +MSE 350.613644 +---------------------- -------------------------------------------------------- +Time: 2.580s Load: 0.003s, Pack+Encode: 1.580s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 350.6136 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.5775 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,984B, BPFP=0.6200 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,884B, BPFP=0.5887 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,832B, BPFP=0.5725 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,796B, BPFP=0.5613 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,508B, BPFP=0.4713 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,264B, BPFP=0.3950 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,552B, BPFP=0.4850 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,000B, BPFP=0.6250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,856B, BPFP=0.5800 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,892B, BPFP=0.0845 +⌛️ [2/4] FRONTEND: Frontend time: 1.486s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.966s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15467889 446.84371094 + layer.0.v_cache 0.00001741 0.04889569 + layer.1.k_cache 0.01620592 27.18839844 + layer.1.v_cache 0.00000548 0.01709120 + layer.2.k_cache 0.00723796 4.72693542 + layer.2.v_cache 0.00001737 0.04949209 + layer.3.k_cache 0.06661982 16.50785767 + layer.3.v_cache 0.00002040 0.06453677 + layer.4.k_cache 0.00058776 1.48055008 + layer.4.v_cache 0.00004656 0.12226120 + layer.4.output 0.27162074 1084.61821429 + ------------------------------------------------------------------------------------- + TOTAL 0.12628134 475.84571938 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 19416 +BPFP 0.3569 bits/point +EBPFP 0.7138 equivalent bits/point +MSE 475.845719 +---------------------- -------------------------------------------------------- +Time: 2.455s Load: 0.003s, Pack+Encode: 1.486s, Decode+Unpack: 0.966s +---------------------- -------------------------------------------------------- +💾 Converting with 475.8457 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 800B, BPFP=0.2451 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,700B, BPFP=0.5208 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,896B, BPFP=0.5809 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,364B, BPFP=0.4179 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,748B, BPFP=0.5355 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,124B, BPFP=0.3444 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,244B, BPFP=0.3811 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,732B, BPFP=0.5306 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,584B, BPFP=0.4853 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,936B, BPFP=0.5931 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,992B, BPFP=0.0872 +⌛️ [2/4] FRONTEND: Frontend time: 1.479s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.968s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11261671 337.67696844 + layer.0.v_cache 0.00001829 0.04472096 + layer.1.k_cache 0.01190703 26.70086311 + layer.1.v_cache 0.00000528 0.01531809 + layer.2.k_cache 0.00973116 4.59817475 + layer.2.v_cache 0.00001668 0.04407364 + layer.3.k_cache 0.04200072 16.26589906 + layer.3.v_cache 0.00001874 0.06184646 + layer.4.k_cache 0.00062040 1.27985696 + layer.4.v_cache 0.00004776 0.11236338 + layer.4.output 0.26626884 1063.44248950 + ------------------------------------------------------------------------------------- + TOTAL 0.12005086 460.64103008 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 17120 +BPFP 0.3085 bits/point +EBPFP 0.6171 equivalent bits/point +MSE 460.641030 +---------------------- -------------------------------------------------------- +Time: 2.449s Load: 0.002s, Pack+Encode: 1.479s, Decode+Unpack: 0.968s +---------------------- -------------------------------------------------------- +💾 Converting with 460.6410 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,252B, BPFP=0.3762 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,436B, BPFP=0.4315 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,048B, BPFP=0.6154 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,376B, BPFP=0.4135 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,740B, BPFP=0.5228 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,420B, BPFP=0.4267 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,276B, BPFP=0.3834 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,772B, BPFP=0.5325 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,912B, BPFP=0.5745 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,888B, BPFP=0.5673 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,120B, BPFP=0.0910 +⌛️ [2/4] FRONTEND: Frontend time: 1.487s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.966s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12932784 424.70150992 + layer.0.v_cache 0.00001491 0.04973414 + layer.1.k_cache 0.01488123 28.84390728 + layer.1.v_cache 0.00000563 0.01737005 + layer.2.k_cache 0.00939683 4.84774105 + layer.2.v_cache 0.00001981 0.05326659 + layer.3.k_cache 0.07051506 16.70832942 + layer.3.v_cache 0.00001927 0.06682576 + layer.4.k_cache 0.00064413 1.55200210 + layer.4.v_cache 0.00005939 0.12649305 + layer.4.output 0.26127321 1042.97149725 + ------------------------------------------------------------------------------------- + TOTAL 0.12081156 457.51574471 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 18240 +BPFP 0.3224 bits/point +EBPFP 0.6448 equivalent bits/point +MSE 457.515745 +---------------------- -------------------------------------------------------- +Time: 2.456s Load: 0.002s, Pack+Encode: 1.487s, Decode+Unpack: 0.966s +---------------------- -------------------------------------------------------- +💾 Converting with 457.5157 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 936B, BPFP=0.2812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,836B, BPFP=0.5517 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,016B, BPFP=0.6058 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,644B, BPFP=0.4940 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,524B, BPFP=0.4579 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,448B, BPFP=0.4351 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,288B, BPFP=0.3870 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,832B, BPFP=0.5505 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,884B, BPFP=0.5661 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,840B, BPFP=0.5529 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,100B, BPFP=0.0901 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.967s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13190371 364.36560998 + layer.0.v_cache 0.00001552 0.05288925 + layer.1.k_cache 0.01668521 28.21166523 + layer.1.v_cache 0.00000581 0.01923680 + layer.2.k_cache 0.00490881 4.48292718 + layer.2.v_cache 0.00001873 0.05439675 + layer.3.k_cache 0.11824917 16.38628211 + layer.3.v_cache 0.00002115 0.07206264 + layer.4.k_cache 0.00061206 1.45594406 + layer.4.v_cache 0.00004844 0.12454037 + layer.4.output 0.26122982 1043.04103709 + ------------------------------------------------------------------------------------- + TOTAL 0.12359279 453.91251847 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 18348 +BPFP 0.3243 bits/point +EBPFP 0.6486 equivalent bits/point +MSE 453.912518 +---------------------- -------------------------------------------------------- +Time: 2.451s Load: 0.002s, Pack+Encode: 1.482s, Decode+Unpack: 0.967s +---------------------- -------------------------------------------------------- +💾 Converting with 453.9125 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,508B, BPFP=0.2561 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,540B, BPFP=0.4314 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,600B, BPFP=0.4416 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,424B, BPFP=0.4117 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,412B, BPFP=0.4096 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,388B, BPFP=0.4056 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,884B, BPFP=0.3200 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,296B, BPFP=0.3899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,796B, BPFP=0.4749 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,632B, BPFP=0.4470 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,276B, BPFP=0.0310 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10326390 373.19985564 + layer.0.v_cache 0.00001541 0.04063771 + layer.1.k_cache 0.04976616 24.68295686 + layer.1.v_cache 0.00000534 0.01418091 + layer.2.k_cache 0.01168035 3.48993517 + layer.2.v_cache 0.00001798 0.04352346 + layer.3.k_cache 0.02907468 17.39264845 + layer.3.v_cache 0.00001946 0.05684953 + layer.4.k_cache 0.00073861 1.23160487 + layer.4.v_cache 0.00005077 0.09711757 + layer.4.output 0.14788401 589.56298525 + ------------------------------------------------------------------------------------- + TOTAL 0.07234240 267.48177688 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 24756 +BPFP 0.2473 bits/point +EBPFP 0.4946 equivalent bits/point +MSE 267.481777 +---------------------- -------------------------------------------------------- +Time: 2.482s Load: 0.003s, Pack+Encode: 1.481s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 267.4818 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 64, 128) +Output shape: (1, 64, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.output: torch.Size([1, 64, 3584]) -> torch.Size([1, 1, 64, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 992B, BPFP=0.2422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,200B, BPFP=0.2930 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,004B, BPFP=0.2451 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,004B, BPFP=0.2451 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 844B, BPFP=0.2061 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 988B, BPFP=0.2412 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 784B, BPFP=0.1914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,152B, BPFP=0.2812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,092B, BPFP=0.2666 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,196B, BPFP=0.2920 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 452B, BPFP=0.0158 +⌛️ [2/4] FRONTEND: Frontend time: 1.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.967s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09722494 402.09945679 + layer.0.v_cache 0.00001391 0.03799081 + layer.1.k_cache 0.01806639 22.79320145 + layer.1.v_cache 0.00000557 0.01442282 + layer.2.k_cache 0.00423524 2.78364325 + layer.2.v_cache 0.00001868 0.04225035 + layer.3.k_cache 0.03759564 13.38555145 + layer.3.v_cache 0.00001825 0.05046747 + layer.4.k_cache 0.00061753 1.00965631 + layer.4.v_cache 0.00004894 0.08633809 + layer.4.output 0.21408452 847.19217355 + ------------------------------------------------------------------------------------- + TOTAL 0.09743746 374.86165845 + (elements=557,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 557056 +Total Bytes 10708 +BPFP 0.1538 bits/point +EBPFP 0.3076 equivalent bits/point +MSE 374.861658 +---------------------- -------------------------------------------------------- +Time: 2.459s Load: 0.004s, Pack+Encode: 1.488s, Decode+Unpack: 0.967s +---------------------- -------------------------------------------------------- +💾 Converting with 374.8617 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,408B, BPFP=0.2157 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,816B, BPFP=0.4314 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,212B, BPFP=0.4920 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,500B, BPFP=0.3830 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,588B, BPFP=0.3964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,428B, BPFP=0.3719 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,468B, BPFP=0.3781 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,612B, BPFP=0.4001 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,296B, BPFP=0.5049 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,788B, BPFP=0.4271 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,432B, BPFP=0.0313 +⌛️ [2/4] FRONTEND: Frontend time: 1.477s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.995s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11648470 334.51960784 + layer.0.v_cache 0.00001454 0.03803177 + layer.1.k_cache 0.02721847 25.60798675 + layer.1.v_cache 0.00000531 0.01425497 + layer.2.k_cache 0.01045542 3.43526444 + layer.2.v_cache 0.00001699 0.04002234 + layer.3.k_cache 0.01744144 14.96703384 + layer.3.v_cache 0.00001839 0.05087022 + layer.4.k_cache 0.00076342 1.17103921 + layer.4.v_cache 0.00004795 0.09131299 + layer.4.output 11.17676328 526.73188025 + ------------------------------------------------------------------------------------- + TOTAL 4.61234174 239.23874036 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 27548 +BPFP 0.2482 bits/point +EBPFP 0.4965 equivalent bits/point +MSE 239.238740 +---------------------- -------------------------------------------------------- +Time: 2.476s Load: 0.004s, Pack+Encode: 1.477s, Decode+Unpack: 0.995s +---------------------- -------------------------------------------------------- +💾 Converting with 239.2387 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,220B, BPFP=0.2383 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,988B, BPFP=0.5836 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,692B, BPFP=0.5258 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,024B, BPFP=0.3953 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,628B, BPFP=0.5133 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,000B, BPFP=0.3906 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,308B, BPFP=0.4508 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,356B, BPFP=0.4602 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,008B, BPFP=0.5875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,876B, BPFP=0.5617 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,404B, BPFP=0.0392 +⌛️ [2/4] FRONTEND: Frontend time: 1.474s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.996s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11433966 336.31279297 + layer.0.v_cache 0.00001437 0.03939706 + layer.1.k_cache 0.05315235 25.24947205 + layer.1.v_cache 0.00000524 0.01377849 + layer.2.k_cache 0.00678293 3.97927170 + layer.2.v_cache 0.00001704 0.04039334 + layer.3.k_cache 0.01951189 17.07492371 + layer.3.v_cache 0.00001758 0.05103771 + layer.4.k_cache 0.00076624 1.27304821 + layer.4.v_cache 0.00004856 0.09347643 + layer.4.output 0.16992235 678.02092634 + ------------------------------------------------------------------------------------- + TOTAL 0.08141837 301.78082800 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 25504 +BPFP 0.2930 bits/point +EBPFP 0.5860 equivalent bits/point +MSE 301.780828 +---------------------- -------------------------------------------------------- +Time: 2.474s Load: 0.005s, Pack+Encode: 1.474s, Decode+Unpack: 0.996s +---------------------- -------------------------------------------------------- +💾 Converting with 301.7808 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,520B, BPFP=0.2639 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,764B, BPFP=0.4799 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,576B, BPFP=0.4472 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,660B, BPFP=0.4618 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,352B, BPFP=0.4083 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,608B, BPFP=0.4528 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,008B, BPFP=0.3486 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,520B, BPFP=0.4375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,840B, BPFP=0.4931 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,744B, BPFP=0.4764 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,396B, BPFP=0.0346 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11570581 368.57604167 + layer.0.v_cache 0.00001517 0.04042382 + layer.1.k_cache 0.03236232 25.58970812 + layer.1.v_cache 0.00000662 0.01435973 + layer.2.k_cache 0.00641356 3.72062242 + layer.2.v_cache 0.00001676 0.04103394 + layer.3.k_cache 0.02648412 17.24973823 + layer.3.v_cache 0.00001896 0.05422000 + layer.4.k_cache 0.00074524 1.24769914 + layer.4.v_cache 0.00004843 0.09617855 + layer.4.output 0.15113260 602.68308532 + ------------------------------------------------------------------------------------- + TOTAL 0.07292619 272.67127193 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 25988 +BPFP 0.2654 bits/point +EBPFP 0.5308 equivalent bits/point +MSE 272.671272 +---------------------- -------------------------------------------------------- +Time: 2.480s Load: 0.003s, Pack+Encode: 1.480s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 272.6713 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 114, 128) +Output shape: (1, 114, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.output: torch.Size([1, 114, 3584]) -> torch.Size([1, 1, 114, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.2533 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,232B, BPFP=0.4430 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,292B, BPFP=0.4512 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,876B, BPFP=0.3942 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,592B, BPFP=0.3553 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,728B, BPFP=0.3739 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,484B, BPFP=0.3405 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,704B, BPFP=0.3706 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,296B, BPFP=0.4518 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,952B, BPFP=0.4046 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,580B, BPFP=0.0309 +⌛️ [2/4] FRONTEND: Frontend time: 1.477s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10081583 363.02388295 + layer.0.v_cache 0.00001633 0.04079760 + layer.1.k_cache 0.08473034 25.08366100 + layer.1.v_cache 0.00000563 0.01452817 + layer.2.k_cache 0.00792928 3.59312894 + layer.2.v_cache 0.00001730 0.04282744 + layer.3.k_cache 0.01019665 15.54049736 + layer.3.v_cache 0.00001810 0.04879561 + layer.4.k_cache 0.00077074 1.26219606 + layer.4.v_cache 0.00005022 0.09341879 + layer.4.output 10.00037814 471.30635182 + ------------------------------------------------------------------------------------- + TOTAL 4.12983514 218.11107039 + (elements=992,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 992256 +Total Bytes 29584 +BPFP 0.2385 bits/point +EBPFP 0.4770 equivalent bits/point +MSE 218.111070 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.006s, Pack+Encode: 1.477s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 218.1111 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 98, 128) +Output shape: (1, 98, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.output: torch.Size([1, 98, 3584]) -> torch.Size([1, 1, 98, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,540B, BPFP=0.2455 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,768B, BPFP=0.4413 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,036B, BPFP=0.4841 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,720B, BPFP=0.4337 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,404B, BPFP=0.3833 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,616B, BPFP=0.4171 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,416B, BPFP=0.3852 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,460B, BPFP=0.3922 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,120B, BPFP=0.4974 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,716B, BPFP=0.4330 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,476B, BPFP=0.0336 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11008905 357.42498406 + layer.0.v_cache 0.00001718 0.03768773 + layer.1.k_cache 0.06240322 24.68442084 + layer.1.v_cache 0.00000548 0.01381694 + layer.2.k_cache 0.00473963 3.72485756 + layer.2.v_cache 0.00001718 0.04100618 + layer.3.k_cache 0.01141277 15.96356824 + layer.3.v_cache 0.00001769 0.04654796 + layer.4.k_cache 0.00076610 1.20031746 + layer.4.v_cache 0.00004814 0.08879487 + layer.4.output 0.02426045 568.27172923 + ------------------------------------------------------------------------------------- + TOTAL 0.02113762 257.71341803 + (elements=852,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 852992 +Total Bytes 27272 +BPFP 0.2558 bits/point +EBPFP 0.5116 equivalent bits/point +MSE 257.713418 +---------------------- -------------------------------------------------------- +Time: 2.483s Load: 0.005s, Pack+Encode: 1.481s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 257.7134 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,628B, BPFP=0.2765 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,564B, BPFP=0.4355 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,516B, BPFP=0.4273 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,548B, BPFP=0.4327 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,220B, BPFP=0.3770 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,512B, BPFP=0.4266 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,056B, BPFP=0.3492 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,352B, BPFP=0.3995 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,432B, BPFP=0.4130 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,604B, BPFP=0.4423 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,332B, BPFP=0.0323 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08919133 381.87572181 + layer.0.v_cache 0.00001414 0.03854609 + layer.1.k_cache 0.08591949 25.60543691 + layer.1.v_cache 0.00000528 0.01441025 + layer.2.k_cache 0.00389855 3.54868483 + layer.2.v_cache 0.00001738 0.04344869 + layer.3.k_cache 0.07679920 17.06357608 + layer.3.v_cache 0.00002028 0.05825600 + layer.4.k_cache 0.00072407 1.22512917 + layer.4.v_cache 0.00004779 0.09845095 + layer.4.output 0.14786680 589.55803571 + ------------------------------------------------------------------------------------- + TOTAL 0.07598265 268.02811240 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 24764 +BPFP 0.2474 bits/point +EBPFP 0.4948 equivalent bits/point +MSE 268.028112 +---------------------- -------------------------------------------------------- +Time: 2.485s Load: 0.004s, Pack+Encode: 1.476s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 268.0281 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,640B, BPFP=0.2395 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,864B, BPFP=0.4182 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,196B, BPFP=0.4667 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,572B, BPFP=0.3756 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,616B, BPFP=0.3820 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,484B, BPFP=0.3627 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,604B, BPFP=0.3803 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,576B, BPFP=0.3762 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,384B, BPFP=0.4942 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,784B, BPFP=0.4065 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,592B, BPFP=0.0332 +⌛️ [2/4] FRONTEND: Frontend time: 1.475s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11353527 357.82279498 + layer.0.v_cache 0.00001674 0.03996818 + layer.1.k_cache 0.04276301 24.45912812 + layer.1.v_cache 0.00000542 0.01424407 + layer.2.k_cache 0.00247095 3.34020739 + layer.2.v_cache 0.00001885 0.04288525 + layer.3.k_cache 0.02722364 14.46071161 + layer.3.v_cache 0.00001942 0.05186532 + layer.4.k_cache 0.00068256 1.22535891 + layer.4.v_cache 0.00004843 0.09394031 + layer.4.output 10.65454330 502.16142356 + ------------------------------------------------------------------------------------- + TOTAL 4.39815220 230.39300406 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 28312 +BPFP 0.2432 bits/point +EBPFP 0.4864 equivalent bits/point +MSE 230.393004 +---------------------- -------------------------------------------------------- +Time: 2.477s Load: 0.004s, Pack+Encode: 1.475s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 230.3930 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,228B, BPFP=0.2284 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,020B, BPFP=0.5618 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,004B, BPFP=0.5588 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,068B, BPFP=0.3847 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,464B, BPFP=0.4583 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,676B, BPFP=0.4978 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,296B, BPFP=0.4271 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,604B, BPFP=0.4844 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,036B, BPFP=0.5647 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,444B, BPFP=0.4546 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,404B, BPFP=0.0373 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12550225 330.22979446 + layer.0.v_cache 0.00001468 0.03920431 + layer.1.k_cache 0.05234316 25.31819080 + layer.1.v_cache 0.00000508 0.01365191 + layer.2.k_cache 0.00733058 3.84243702 + layer.2.v_cache 0.00001689 0.04045103 + layer.3.k_cache 0.01592888 16.82417661 + layer.3.v_cache 0.00001847 0.05122473 + layer.4.k_cache 0.00069460 1.25328255 + layer.4.v_cache 0.00005205 0.09656724 + layer.4.output 0.16184913 645.72868835 + ------------------------------------------------------------------------------------- + TOTAL 0.07852062 288.10645877 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 26244 +BPFP 0.2872 bits/point +EBPFP 0.5743 equivalent bits/point +MSE 288.106459 +---------------------- -------------------------------------------------------- +Time: 2.485s Load: 0.005s, Pack+Encode: 1.484s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 288.1065 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,208B, BPFP=0.2420 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,840B, BPFP=0.5689 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,052B, BPFP=0.6114 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,876B, BPFP=0.3758 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,608B, BPFP=0.5224 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,092B, BPFP=0.4191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,268B, BPFP=0.4543 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,140B, BPFP=0.4287 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,972B, BPFP=0.5954 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,592B, BPFP=0.5192 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,396B, BPFP=0.0399 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11879376 333.02128405 + layer.0.v_cache 0.00001438 0.04051835 + layer.1.k_cache 0.05518859 26.14899777 + layer.1.v_cache 0.00000514 0.01536113 + layer.2.k_cache 0.00243519 4.08351331 + layer.2.v_cache 0.00001742 0.04132494 + layer.3.k_cache 0.07959401 16.95958847 + layer.3.v_cache 0.00001812 0.05249921 + layer.4.k_cache 0.00069358 1.25818086 + layer.4.v_cache 0.00004620 0.09607907 + layer.4.output 0.17426339 695.39119734 + ------------------------------------------------------------------------------------- + TOTAL 0.08686177 308.79151345 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 25044 +BPFP 0.2951 bits/point +EBPFP 0.5902 equivalent bits/point +MSE 308.791513 +---------------------- -------------------------------------------------------- +Time: 2.484s Load: 0.003s, Pack+Encode: 1.482s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 308.7915 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,152B, BPFP=0.2195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,924B, BPFP=0.5572 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,960B, BPFP=0.5640 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,460B, BPFP=0.4688 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,580B, BPFP=0.4916 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,240B, BPFP=0.4268 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,396B, BPFP=0.4566 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,700B, BPFP=0.5145 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,160B, BPFP=0.6021 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,832B, BPFP=0.5396 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,408B, BPFP=0.0383 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11325797 308.65439215 + layer.0.v_cache 0.00001565 0.03799074 + layer.1.k_cache 0.07651102 25.80252001 + layer.1.v_cache 0.00000530 0.01455998 + layer.2.k_cache 0.00872486 4.09771915 + layer.2.v_cache 0.00001696 0.04142294 + layer.3.k_cache 0.12876377 16.67692157 + layer.3.v_cache 0.00001905 0.04928807 + layer.4.k_cache 0.00069190 1.27373467 + layer.4.v_cache 0.00004968 0.09470024 + layer.4.output 0.16578155 661.43554007 + ------------------------------------------------------------------------------------- + TOTAL 0.08756041 293.34070765 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 26812 +BPFP 0.3005 bits/point +EBPFP 0.6011 equivalent bits/point +MSE 293.340708 +---------------------- -------------------------------------------------------- +Time: 2.478s Load: 0.003s, Pack+Encode: 1.476s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 293.3407 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,184B, BPFP=0.2403 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,508B, BPFP=0.5089 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,216B, BPFP=0.6526 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,308B, BPFP=0.4683 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,684B, BPFP=0.5446 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,364B, BPFP=0.4797 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,300B, BPFP=0.4667 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,048B, BPFP=0.4156 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,860B, BPFP=0.5804 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,924B, BPFP=0.5933 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,336B, BPFP=0.0387 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12835005 344.40160816 + layer.0.v_cache 0.00001616 0.03851225 + layer.1.k_cache 0.05672838 26.15539075 + layer.1.v_cache 0.00000535 0.01431989 + layer.2.k_cache 0.00240846 4.08709003 + layer.2.v_cache 0.00001644 0.03946781 + layer.3.k_cache 0.04570918 16.03410201 + layer.3.v_cache 0.00001733 0.04902570 + layer.4.k_cache 0.00072090 1.17540176 + layer.4.v_cache 0.00004908 0.09355046 + layer.4.output 0.18444509 705.16883117 + ------------------------------------------------------------------------------------- + TOTAL 0.08971394 313.42766394 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 25732 +BPFP 0.3072 bits/point +EBPFP 0.6143 equivalent bits/point +MSE 313.427664 +---------------------- -------------------------------------------------------- +Time: 2.478s Load: 0.004s, Pack+Encode: 1.476s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 313.4277 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,068B, BPFP=0.2350 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,480B, BPFP=0.5458 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,184B, BPFP=0.4806 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,724B, BPFP=0.3794 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,088B, BPFP=0.4595 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,584B, BPFP=0.3486 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,540B, BPFP=0.3389 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,848B, BPFP=0.4067 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,504B, BPFP=0.5511 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,444B, BPFP=0.5379 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,344B, BPFP=0.0423 +⌛️ [2/4] FRONTEND: Frontend time: 1.474s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10085608 320.01523988 + layer.0.v_cache 0.00001554 0.03703062 + layer.1.k_cache 0.06589249 26.25698380 + layer.1.v_cache 0.00000499 0.01392566 + layer.2.k_cache 0.00422055 4.07069676 + layer.2.v_cache 0.00001579 0.03873271 + layer.3.k_cache 0.05337993 16.37854348 + layer.3.v_cache 0.00001747 0.04843003 + layer.4.k_cache 0.00075412 1.24620207 + layer.4.v_cache 0.00004712 0.09443025 + layer.4.output 0.19135183 763.93976358 + ------------------------------------------------------------------------------------- + TOTAL 0.09203923 336.22226825 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 20808 +BPFP 0.2694 bits/point +EBPFP 0.5387 equivalent bits/point +MSE 336.222268 +---------------------- -------------------------------------------------------- +Time: 2.475s Load: 0.004s, Pack+Encode: 1.474s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 336.2223 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,220B, BPFP=0.2269 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,748B, BPFP=0.5112 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,992B, BPFP=0.5565 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,096B, BPFP=0.3899 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,488B, BPFP=0.4628 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,276B, BPFP=0.4234 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,460B, BPFP=0.4576 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,500B, BPFP=0.4650 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,148B, BPFP=0.5856 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,780B, BPFP=0.5171 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,400B, BPFP=0.0372 +⌛️ [2/4] FRONTEND: Frontend time: 1.475s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09057366 333.01316034 + layer.0.v_cache 0.00001789 0.03616113 + layer.1.k_cache 0.03614600 25.66434733 + layer.1.v_cache 0.00000481 0.01307919 + layer.2.k_cache 0.00554357 3.91071174 + layer.2.v_cache 0.00001714 0.03883500 + layer.3.k_cache 0.02742965 16.80830601 + layer.3.v_cache 0.00001729 0.04759206 + layer.4.k_cache 0.00093027 1.22317741 + layer.4.v_cache 0.00004906 0.09317011 + layer.4.output 0.16441821 645.76078869 + ------------------------------------------------------------------------------------- + TOTAL 0.07715628 288.30435654 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 26108 +BPFP 0.2857 bits/point +EBPFP 0.5713 equivalent bits/point +MSE 288.304357 +---------------------- -------------------------------------------------------- +Time: 2.478s Load: 0.004s, Pack+Encode: 1.475s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 288.3044 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,180B, BPFP=0.2394 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,008B, BPFP=0.6104 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,208B, BPFP=0.6510 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,324B, BPFP=0.4716 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,472B, BPFP=0.5016 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,284B, BPFP=0.4635 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,144B, BPFP=0.4351 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,272B, BPFP=0.4610 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,056B, BPFP=0.6201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,552B, BPFP=0.5179 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,384B, BPFP=0.0401 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09325875 341.13103693 + layer.0.v_cache 0.00001406 0.03808750 + layer.1.k_cache 0.05775181 26.18995092 + layer.1.v_cache 0.00000524 0.01346257 + layer.2.k_cache 0.00413042 3.83223338 + layer.2.v_cache 0.00001743 0.04055634 + layer.3.k_cache 0.04387648 17.06023615 + layer.3.v_cache 0.00002497 0.05444002 + layer.4.k_cache 0.00071512 1.24486611 + layer.4.v_cache 0.00005147 0.10133941 + layer.4.output 0.17652170 704.42086039 + ------------------------------------------------------------------------------------- + TOTAL 0.08444104 312.97954306 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 25884 +BPFP 0.3090 bits/point +EBPFP 0.6179 equivalent bits/point +MSE 312.979543 +---------------------- -------------------------------------------------------- +Time: 2.476s Load: 0.003s, Pack+Encode: 1.476s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 312.9795 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,564B, BPFP=0.2777 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,648B, BPFP=0.4702 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,584B, BPFP=0.4588 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,288B, BPFP=0.4062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,208B, BPFP=0.3920 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,416B, BPFP=0.4290 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,848B, BPFP=0.3281 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,468B, BPFP=0.4382 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,868B, BPFP=0.5092 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,852B, BPFP=0.5064 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,500B, BPFP=0.0380 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212358 367.30988104 + layer.0.v_cache 0.00001682 0.03126154 + layer.1.k_cache 0.05918076 25.06673639 + layer.1.v_cache 0.00000473 0.01105995 + layer.2.k_cache 0.00260097 3.90007192 + layer.2.v_cache 0.00001544 0.03360407 + layer.3.k_cache 0.02838851 16.23235807 + layer.3.v_cache 0.00001657 0.04151217 + layer.4.k_cache 0.00106615 1.20002296 + layer.4.v_cache 0.00004456 0.08154994 + layer.4.output 0.17173813 616.58715503 + ------------------------------------------------------------------------------------- + TOTAL 0.08091912 278.23636137 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 25244 +BPFP 0.2637 bits/point +EBPFP 0.5273 equivalent bits/point +MSE 278.236361 +---------------------- -------------------------------------------------------- +Time: 2.476s Load: 0.003s, Pack+Encode: 1.476s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 278.2364 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,280B, BPFP=0.2273 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,600B, BPFP=0.4616 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,792B, BPFP=0.4957 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,304B, BPFP=0.4091 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,192B, BPFP=0.3892 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,276B, BPFP=0.4041 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,228B, BPFP=0.3956 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,260B, BPFP=0.4013 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,012B, BPFP=0.5348 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,716B, BPFP=0.4822 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,484B, BPFP=0.0376 +⌛️ [2/4] FRONTEND: Frontend time: 1.475s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10731998 327.00905540 + layer.0.v_cache 0.00001618 0.03864839 + layer.1.k_cache 0.07375650 25.62700306 + layer.1.v_cache 0.00000540 0.01449460 + layer.2.k_cache 0.00246326 3.71760316 + layer.2.v_cache 0.00001746 0.04147593 + layer.3.k_cache 0.01642703 16.51483709 + layer.3.v_cache 0.00001778 0.04965968 + layer.4.k_cache 0.00066607 1.17691005 + layer.4.v_cache 0.00004728 0.08933885 + layer.4.output 0.15452130 616.34603287 + ------------------------------------------------------------------------------------- + TOTAL 0.07543447 275.80595625 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 25144 +BPFP 0.2626 bits/point +EBPFP 0.5252 equivalent bits/point +MSE 275.805956 +---------------------- -------------------------------------------------------- +Time: 2.478s Load: 0.005s, Pack+Encode: 1.475s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 275.8060 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 99, 128) +Output shape: (1, 99, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.output: torch.Size([1, 99, 3584]) -> torch.Size([1, 1, 99, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,476B, BPFP=0.2330 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,760B, BPFP=0.4356 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,128B, BPFP=0.4937 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,620B, BPFP=0.4135 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,576B, BPFP=0.4066 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,800B, BPFP=0.4419 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,580B, BPFP=0.4072 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,508B, BPFP=0.3958 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,396B, BPFP=0.5360 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,868B, BPFP=0.4527 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,504B, BPFP=0.0339 +⌛️ [2/4] FRONTEND: Frontend time: 1.477s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09061131 361.67140152 + layer.0.v_cache 0.00001461 0.03729295 + layer.1.k_cache 0.05062923 26.15571733 + layer.1.v_cache 0.00000709 0.01356233 + layer.2.k_cache 0.00392854 3.77363263 + layer.2.v_cache 0.00001793 0.04102347 + layer.3.k_cache 0.05082532 14.82671441 + layer.3.v_cache 0.00001820 0.04997300 + layer.4.k_cache 0.00080398 1.20961122 + layer.4.v_cache 0.00005053 0.08865487 + layer.4.output 0.02403493 562.55699856 + ------------------------------------------------------------------------------------- + TOTAL 0.02147948 255.63332786 + (elements=861,696) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 861696 +Total Bytes 28216 +BPFP 0.2620 bits/point +EBPFP 0.5239 equivalent bits/point +MSE 255.633328 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.005s, Pack+Encode: 1.477s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 255.6333 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,080B, BPFP=0.2377 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,268B, BPFP=0.4991 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,452B, BPFP=0.5396 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,796B, BPFP=0.3952 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,828B, BPFP=0.4023 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,596B, BPFP=0.3512 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,512B, BPFP=0.3327 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,832B, BPFP=0.4032 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,488B, BPFP=0.5475 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,436B, BPFP=0.5361 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,360B, BPFP=0.0428 +⌛️ [2/4] FRONTEND: Frontend time: 1.475s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11136492 322.73674076 + layer.0.v_cache 0.00001955 0.04255207 + layer.1.k_cache 0.06280328 25.62546765 + layer.1.v_cache 0.00000544 0.01588519 + layer.2.k_cache 0.00235174 3.95120411 + layer.2.v_cache 0.00001859 0.04601812 + layer.3.k_cache 0.03326307 16.19980503 + layer.3.v_cache 0.00001926 0.05603641 + layer.4.k_cache 0.00064473 1.26418251 + layer.4.v_cache 0.00005254 0.10343100 + layer.4.output 0.19494371 764.40926811 + ------------------------------------------------------------------------------------- + TOTAL 0.09265583 336.52389409 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 20648 +BPFP 0.2673 bits/point +EBPFP 0.5346 equivalent bits/point +MSE 336.523894 +---------------------- -------------------------------------------------------- +Time: 2.475s Load: 0.003s, Pack+Encode: 1.475s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 336.5239 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,172B, BPFP=0.2475 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,724B, BPFP=0.5752 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,996B, BPFP=0.4215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,044B, BPFP=0.4316 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,160B, BPFP=0.4561 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,996B, BPFP=0.4215 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,596B, BPFP=0.3370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,020B, BPFP=0.4265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,600B, BPFP=0.5490 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,504B, BPFP=0.5287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,328B, BPFP=0.0401 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.003s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08829598 325.48131334 + layer.0.v_cache 0.00001455 0.03938243 + layer.1.k_cache 0.05950350 25.89996173 + layer.1.v_cache 0.00000529 0.01466925 + layer.2.k_cache 0.00411868 4.13024902 + layer.2.v_cache 0.00001680 0.04228713 + layer.3.k_cache 0.04877868 17.05363506 + layer.3.v_cache 0.00001896 0.05441471 + layer.4.k_cache 0.00063741 1.24653089 + layer.4.v_cache 0.00004764 0.09841619 + layer.4.output 0.18365649 732.91807432 + ------------------------------------------------------------------------------------- + TOTAL 0.08747253 323.79337530 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 22140 +BPFP 0.2750 bits/point +EBPFP 0.5500 equivalent bits/point +MSE 323.793375 +---------------------- -------------------------------------------------------- +Time: 2.489s Load: 0.003s, Pack+Encode: 1.483s, Decode+Unpack: 1.003s +---------------------- -------------------------------------------------------- +💾 Converting with 323.7934 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,224B, BPFP=0.2484 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,576B, BPFP=0.5227 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,064B, BPFP=0.6218 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,148B, BPFP=0.4359 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,320B, BPFP=0.4708 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,648B, BPFP=0.5373 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,000B, BPFP=0.4058 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,184B, BPFP=0.4432 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,952B, BPFP=0.5990 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,732B, BPFP=0.5544 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,344B, BPFP=0.0390 +⌛️ [2/4] FRONTEND: Frontend time: 1.477s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11195674 346.26105925 + layer.0.v_cache 0.00001346 0.03733083 + layer.1.k_cache 0.03755230 25.98739664 + layer.1.v_cache 0.00000535 0.01516235 + layer.2.k_cache 0.00239600 4.02769252 + layer.2.v_cache 0.00001753 0.04055006 + layer.3.k_cache 0.02867049 16.59378805 + layer.3.v_cache 0.00001751 0.04952595 + layer.4.k_cache 0.00071811 1.21005150 + layer.4.v_cache 0.00005486 0.09786064 + layer.4.output 0.19495771 705.10146104 + ------------------------------------------------------------------------------------- + TOTAL 0.09094743 313.53121442 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 25192 +BPFP 0.3007 bits/point +EBPFP 0.6014 equivalent bits/point +MSE 313.531214 +---------------------- -------------------------------------------------------- +Time: 2.477s Load: 0.003s, Pack+Encode: 1.477s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 313.5312 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,348B, BPFP=0.2340 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,604B, BPFP=0.4521 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,464B, BPFP=0.4278 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,608B, BPFP=0.4528 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,392B, BPFP=0.4153 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,360B, BPFP=0.4097 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,940B, BPFP=0.3368 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,692B, BPFP=0.4674 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,816B, BPFP=0.4889 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,516B, BPFP=0.4368 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,364B, BPFP=0.0338 +⌛️ [2/4] FRONTEND: Frontend time: 1.486s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.996s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08657128 343.31627604 + layer.0.v_cache 0.00001574 0.03527718 + layer.1.k_cache 0.05618976 24.93204753 + layer.1.v_cache 0.00000506 0.01272126 + layer.2.k_cache 0.00392605 3.89687161 + layer.2.v_cache 0.00001570 0.03726875 + layer.3.k_cache 0.02849737 16.36458333 + layer.3.v_cache 0.00001769 0.04802674 + layer.4.k_cache 0.00099738 1.24317983 + layer.4.v_cache 0.00004751 0.08912574 + layer.4.output 0.15107292 602.61716270 + ------------------------------------------------------------------------------------- + TOTAL 0.07257612 271.07620688 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 25104 +BPFP 0.2564 bits/point +EBPFP 0.5127 equivalent bits/point +MSE 271.076207 +---------------------- -------------------------------------------------------- +Time: 2.486s Load: 0.004s, Pack+Encode: 1.486s, Decode+Unpack: 0.996s +---------------------- -------------------------------------------------------- +💾 Converting with 271.0762 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,256B, BPFP=0.2453 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,016B, BPFP=0.5891 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,940B, BPFP=0.5742 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,136B, BPFP=0.4172 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,680B, BPFP=0.5234 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,084B, BPFP=0.4070 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,244B, BPFP=0.4383 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,164B, BPFP=0.4227 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,240B, BPFP=0.6328 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,572B, BPFP=0.5023 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,424B, BPFP=0.0397 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.996s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09254595 333.02934570 + layer.0.v_cache 0.00001776 0.03480692 + layer.1.k_cache 0.05775445 26.42014465 + layer.1.v_cache 0.00000504 0.01263103 + layer.2.k_cache 0.00891839 4.08069763 + layer.2.v_cache 0.00001732 0.03860091 + layer.3.k_cache 0.02836140 16.36163025 + layer.3.v_cache 0.00001672 0.04701345 + layer.4.k_cache 0.00083566 1.21163692 + layer.4.v_cache 0.00004876 0.09251938 + layer.4.output 0.16993865 678.47991071 + ------------------------------------------------------------------------------------- + TOTAL 0.08106423 301.80520011 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 25756 +BPFP 0.2959 bits/point +EBPFP 0.5918 equivalent bits/point +MSE 301.805200 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.004s, Pack+Encode: 1.480s, Decode+Unpack: 0.996s +---------------------- -------------------------------------------------------- +💾 Converting with 301.8052 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,400B, BPFP=0.2255 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,612B, BPFP=0.4207 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,856B, BPFP=0.4601 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,360B, BPFP=0.3802 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,308B, BPFP=0.3718 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,144B, BPFP=0.3454 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,368B, BPFP=0.3814 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,196B, BPFP=0.3537 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,072B, BPFP=0.4948 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,680B, BPFP=0.4317 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,436B, BPFP=0.0330 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13093194 342.76115496 + layer.0.v_cache 0.00001635 0.03711004 + layer.1.k_cache 0.05235285 25.29595884 + layer.1.v_cache 0.00000496 0.01304816 + layer.2.k_cache 0.00512381 3.75119365 + layer.2.v_cache 0.00001647 0.03857902 + layer.3.k_cache 0.01138228 15.80305402 + layer.3.v_cache 0.00001850 0.04770814 + layer.4.k_cache 0.00081880 1.16441062 + layer.4.v_cache 0.00004796 0.08863707 + layer.4.output 0.02447439 574.15818299 + ------------------------------------------------------------------------------------- + TOTAL 0.02188439 259.30047856 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 25432 +BPFP 0.2410 bits/point +EBPFP 0.4820 equivalent bits/point +MSE 259.300479 +---------------------- -------------------------------------------------------- +Time: 2.478s Load: 0.004s, Pack+Encode: 1.476s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 259.3005 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,236B, BPFP=0.2445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,828B, BPFP=0.5593 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,832B, BPFP=0.5601 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,012B, BPFP=0.3979 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,588B, BPFP=0.5119 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,248B, BPFP=0.4446 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,064B, BPFP=0.4082 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,216B, BPFP=0.4383 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,864B, BPFP=0.5665 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,864B, BPFP=0.5665 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,464B, BPFP=0.0414 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510638 328.70824268 + layer.0.v_cache 0.00001399 0.03820584 + layer.1.k_cache 0.05467150 26.14139760 + layer.1.v_cache 0.00000511 0.01395359 + layer.2.k_cache 0.00545186 4.00537341 + layer.2.v_cache 0.00001652 0.04103884 + layer.3.k_cache 0.02986256 16.47390631 + layer.3.v_cache 0.00001747 0.04600915 + layer.4.k_cache 0.00068266 1.22545112 + layer.4.v_cache 0.00005277 0.09483206 + layer.4.output 0.17642128 687.36612794 + ------------------------------------------------------------------------------------- + TOTAL 0.08357822 305.19713566 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 25216 +BPFP 0.2934 bits/point +EBPFP 0.5867 equivalent bits/point +MSE 305.197136 +---------------------- -------------------------------------------------------- +Time: 2.481s Load: 0.004s, Pack+Encode: 1.478s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 305.1971 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,212B, BPFP=0.2459 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,636B, BPFP=0.5349 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,096B, BPFP=0.6282 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,976B, BPFP=0.4010 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,664B, BPFP=0.5406 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,992B, BPFP=0.4042 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,928B, BPFP=0.3912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,292B, BPFP=0.4651 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,092B, BPFP=0.6274 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,880B, BPFP=0.5844 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,344B, BPFP=0.0390 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08730992 341.57279830 + layer.0.v_cache 0.00001543 0.03966480 + layer.1.k_cache 0.05635426 26.19021091 + layer.1.v_cache 0.00000520 0.01509149 + layer.2.k_cache 0.00390536 3.99067985 + layer.2.v_cache 0.00001661 0.03964233 + layer.3.k_cache 0.06483487 17.15242777 + layer.3.v_cache 0.00002010 0.04882911 + layer.4.k_cache 0.00072652 1.21675367 + layer.4.v_cache 0.00004842 0.09166720 + layer.4.output 0.18845482 705.15364100 + ------------------------------------------------------------------------------------- + TOTAL 0.09014238 313.31960309 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 25112 +BPFP 0.2998 bits/point +EBPFP 0.5995 equivalent bits/point +MSE 313.319603 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.004s, Pack+Encode: 1.476s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 313.3196 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 996B, BPFP=0.2289 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,600B, BPFP=0.5974 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,372B, BPFP=0.5450 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,600B, BPFP=0.3676 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,612B, BPFP=0.3704 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,476B, BPFP=0.3392 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,432B, BPFP=0.3290 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,332B, BPFP=0.3061 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,484B, BPFP=0.5708 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,380B, BPFP=0.5469 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,448B, BPFP=0.0475 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10784986 320.03349035 + layer.0.v_cache 0.00001387 0.03879076 + layer.1.k_cache 0.01409798 27.08183378 + layer.1.v_cache 0.00000489 0.01427707 + layer.2.k_cache 0.00254294 4.00925042 + layer.2.v_cache 0.00001653 0.04228400 + layer.3.k_cache 0.05709259 16.71292114 + layer.3.v_cache 0.00001736 0.05243935 + layer.4.k_cache 0.00062715 1.23612393 + layer.4.v_cache 0.00004758 0.09466358 + layer.4.output 0.22376562 798.37309611 + ------------------------------------------------------------------------------------- + TOTAL 0.10286295 350.46633807 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 19732 +BPFP 0.2667 bits/point +EBPFP 0.5334 equivalent bits/point +MSE 350.466338 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.003s, Pack+Encode: 1.478s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 350.4663 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,268B, BPFP=0.2508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,796B, BPFP=0.5530 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,084B, BPFP=0.6100 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,164B, BPFP=0.4280 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,448B, BPFP=0.4842 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,324B, BPFP=0.4597 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,116B, BPFP=0.4185 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,888B, BPFP=0.3734 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,036B, BPFP=0.6005 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,624B, BPFP=0.5190 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,480B, BPFP=0.0418 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14563359 339.94763647 + layer.0.v_cache 0.00001368 0.03395541 + layer.1.k_cache 0.05947259 25.70331506 + layer.1.v_cache 0.00000496 0.01258447 + layer.2.k_cache 0.00553216 3.90940992 + layer.2.v_cache 0.00001557 0.03513724 + layer.3.k_cache 0.04377548 15.95444151 + layer.3.v_cache 0.00001730 0.04714579 + layer.4.k_cache 0.00079035 1.19313262 + layer.4.v_cache 0.00005153 0.09315861 + layer.4.output 0.17347779 687.41551763 + ------------------------------------------------------------------------------------- + TOTAL 0.08645010 305.81403179 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 25228 +BPFP 0.2935 bits/point +EBPFP 0.5870 equivalent bits/point +MSE 305.814032 +---------------------- -------------------------------------------------------- +Time: 2.477s Load: 0.004s, Pack+Encode: 1.476s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 305.8140 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 936B, BPFP=0.2216 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,124B, BPFP=0.5028 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,696B, BPFP=0.4015 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,692B, BPFP=0.4006 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,612B, BPFP=0.3816 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,400B, BPFP=0.3314 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,204B, BPFP=0.2850 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,356B, BPFP=0.3210 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,212B, BPFP=0.5237 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,044B, BPFP=0.4839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,412B, BPFP=0.0478 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12373332 324.98606179 + layer.0.v_cache 0.00001436 0.03970382 + layer.1.k_cache 0.01270404 25.36909254 + layer.1.v_cache 0.00000490 0.01389639 + layer.2.k_cache 0.00252672 4.01687345 + layer.2.v_cache 0.00001629 0.04324368 + layer.3.k_cache 0.05185594 16.82909416 + layer.3.v_cache 0.00001849 0.05459722 + layer.4.k_cache 0.00068458 1.26471826 + layer.4.v_cache 0.00004975 0.10618880 + layer.4.output 0.22283003 822.47862554 + ------------------------------------------------------------------------------------- + TOTAL 0.10302462 360.59257935 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 17688 +BPFP 0.2463 bits/point +EBPFP 0.4926 equivalent bits/point +MSE 360.592579 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.003s, Pack+Encode: 1.476s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 360.5926 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,220B, BPFP=0.2269 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,768B, BPFP=0.5149 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,128B, BPFP=0.5818 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,016B, BPFP=0.3750 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,720B, BPFP=0.5060 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,236B, BPFP=0.4159 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,440B, BPFP=0.4539 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,240B, BPFP=0.4167 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,064B, BPFP=0.5699 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,968B, BPFP=0.5521 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,404B, BPFP=0.0373 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12002775 325.68173363 + layer.0.v_cache 0.00001466 0.04148486 + layer.1.k_cache 0.05271155 26.10100156 + layer.1.v_cache 0.00000524 0.01524109 + layer.2.k_cache 0.00808975 3.99791972 + layer.2.v_cache 0.00001826 0.04213212 + layer.3.k_cache 0.07501186 16.74262637 + layer.3.v_cache 0.00001895 0.05572420 + layer.4.k_cache 0.00071216 1.23688525 + layer.4.v_cache 0.00005170 0.09892296 + layer.4.output 0.17053346 646.39019983 + ------------------------------------------------------------------------------------- + TOTAL 0.08531742 288.16147474 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 26204 +BPFP 0.2867 bits/point +EBPFP 0.5734 equivalent bits/point +MSE 288.161475 +---------------------- -------------------------------------------------------- +Time: 2.478s Load: 0.003s, Pack+Encode: 1.476s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 288.1615 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 75, 128) +Output shape: (1, 75, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.output: torch.Size([1, 75, 3584]) -> torch.Size([1, 1, 75, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,180B, BPFP=0.2458 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,452B, BPFP=0.5108 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,324B, BPFP=0.4842 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,828B, BPFP=0.3808 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,632B, BPFP=0.5483 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,780B, BPFP=0.3708 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,884B, BPFP=0.3925 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,432B, BPFP=0.5067 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,076B, BPFP=0.6408 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,624B, BPFP=0.5467 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,328B, BPFP=0.0395 +⌛️ [2/4] FRONTEND: Frontend time: 1.474s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08149677 322.59442708 + layer.0.v_cache 0.00001391 0.04002148 + layer.1.k_cache 0.05841089 25.90164388 + layer.1.v_cache 0.00000541 0.01448914 + layer.2.k_cache 0.00252637 3.95624105 + layer.2.v_cache 0.00001693 0.04173621 + layer.3.k_cache 0.02876982 16.43719238 + layer.3.v_cache 0.00001932 0.05022994 + layer.4.k_cache 0.00075246 1.25883158 + layer.4.v_cache 0.00005041 0.09952442 + layer.4.output 0.19266937 723.95011905 + ------------------------------------------------------------------------------------- + TOTAL 0.08945576 319.88501003 + (elements=652,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 652800 +Total Bytes 23540 +BPFP 0.2885 bits/point +EBPFP 0.5770 equivalent bits/point +MSE 319.885010 +---------------------- -------------------------------------------------------- +Time: 2.478s Load: 0.004s, Pack+Encode: 1.474s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 319.8850 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 984B, BPFP=0.2261 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,060B, BPFP=0.4733 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,092B, BPFP=0.4807 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,580B, BPFP=0.3631 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,000B, BPFP=0.4596 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,388B, BPFP=0.3189 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,472B, BPFP=0.3382 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,464B, BPFP=0.3364 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,752B, BPFP=0.6324 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,248B, BPFP=0.5165 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,440B, BPFP=0.0473 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.000s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10404026 314.94094669 + layer.0.v_cache 0.00001395 0.03988309 + layer.1.k_cache 0.01458383 26.62893677 + layer.1.v_cache 0.00000499 0.01421397 + layer.2.k_cache 0.00236740 3.79777437 + layer.2.v_cache 0.00001641 0.04309944 + layer.3.k_cache 0.03645969 16.06423232 + layer.3.v_cache 0.00002420 0.05228043 + layer.4.k_cache 0.00069996 1.26179415 + layer.4.v_cache 0.00005119 0.10568250 + layer.4.output 0.20980707 798.48667279 + ------------------------------------------------------------------------------------- + TOTAL 0.09570067 350.13856196 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 19480 +BPFP 0.2633 bits/point +EBPFP 0.5266 equivalent bits/point +MSE 350.138562 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.003s, Pack+Encode: 1.476s, Decode+Unpack: 1.000s +---------------------- -------------------------------------------------------- +💾 Converting with 350.1386 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,068B, BPFP=0.2384 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,252B, BPFP=0.5027 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,068B, BPFP=0.4616 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,548B, BPFP=0.3455 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,736B, BPFP=0.3875 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,416B, BPFP=0.3161 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,512B, BPFP=0.3375 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,504B, BPFP=0.3357 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,612B, BPFP=0.5830 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,512B, BPFP=0.5607 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,380B, BPFP=0.0440 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11725240 333.23660714 + layer.0.v_cache 0.00001567 0.04013352 + layer.1.k_cache 0.03693308 26.64233573 + layer.1.v_cache 0.00000533 0.01521543 + layer.2.k_cache 0.00240566 3.92098563 + layer.2.v_cache 0.00001850 0.04690847 + layer.3.k_cache 0.06751672 17.16937430 + layer.3.v_cache 0.00001909 0.05206807 + layer.4.k_cache 0.00067984 1.24958311 + layer.4.v_cache 0.00004901 0.10086719 + layer.4.output 0.20094273 775.73195153 + ------------------------------------------------------------------------------------- + TOTAL 0.09597026 341.91751408 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 19608 +BPFP 0.2575 bits/point +EBPFP 0.5149 equivalent bits/point +MSE 341.917514 +---------------------- -------------------------------------------------------- +Time: 2.481s Load: 0.003s, Pack+Encode: 1.480s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 341.9175 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,088B, BPFP=0.2429 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,064B, BPFP=0.4607 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,376B, BPFP=0.5304 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,608B, BPFP=0.3589 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,732B, BPFP=0.3866 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,428B, BPFP=0.3187 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,528B, BPFP=0.3411 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,708B, BPFP=0.3812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,364B, BPFP=0.5277 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,404B, BPFP=0.5366 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,368B, BPFP=0.0436 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07685234 337.51875000 + layer.0.v_cache 0.00001390 0.04078992 + layer.1.k_cache 0.01187498 26.28254743 + layer.1.v_cache 0.00000934 0.01454872 + layer.2.k_cache 0.00239603 4.17644828 + layer.2.v_cache 0.00001621 0.04000897 + layer.3.k_cache 0.07241479 16.85792062 + layer.3.v_cache 0.00001716 0.05030307 + layer.4.k_cache 0.00072067 1.25356118 + layer.4.v_cache 0.00004958 0.09395982 + layer.4.output 0.21417375 775.52066327 + ------------------------------------------------------------------------------------- + TOTAL 0.09785772 342.05726358 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 19668 +BPFP 0.2582 bits/point +EBPFP 0.5165 equivalent bits/point +MSE 342.057264 +---------------------- -------------------------------------------------------- +Time: 2.486s Load: 0.004s, Pack+Encode: 1.483s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 342.0573 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,200B, BPFP=0.2467 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,732B, BPFP=0.5617 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,436B, BPFP=0.5008 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,200B, BPFP=0.4523 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,408B, BPFP=0.4951 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,728B, BPFP=0.3553 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,084B, BPFP=0.4285 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,616B, BPFP=0.5378 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,000B, BPFP=0.6168 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,840B, BPFP=0.5839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,352B, BPFP=0.0397 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.997s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09264422 343.89879729 + layer.0.v_cache 0.00001680 0.04075490 + layer.1.k_cache 0.05606840 25.59249396 + layer.1.v_cache 0.00000542 0.01523508 + layer.2.k_cache 0.00241068 4.08331178 + layer.2.v_cache 0.00001702 0.04375393 + layer.3.k_cache 0.15270803 17.15351466 + layer.3.v_cache 0.00001849 0.05189733 + layer.4.k_cache 0.00070970 1.23860540 + layer.4.v_cache 0.00005539 0.09728479 + layer.4.output 0.18838861 714.45758929 + ------------------------------------------------------------------------------------- + TOTAL 0.09549261 317.25992789 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 24596 +BPFP 0.2975 bits/point +EBPFP 0.5949 equivalent bits/point +MSE 317.259928 +---------------------- -------------------------------------------------------- +Time: 2.483s Load: 0.003s, Pack+Encode: 1.483s, Decode+Unpack: 0.997s +---------------------- -------------------------------------------------------- +💾 Converting with 317.2599 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,204B, BPFP=0.2443 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,592B, BPFP=0.5260 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,060B, BPFP=0.6209 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,124B, BPFP=0.4310 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,652B, BPFP=0.5381 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,916B, BPFP=0.3888 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,000B, BPFP=0.4058 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,204B, BPFP=0.4472 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,152B, BPFP=0.6396 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,692B, BPFP=0.5463 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,352B, BPFP=0.0392 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10144231 337.04274046 + layer.0.v_cache 0.00001330 0.03702134 + layer.1.k_cache 0.09893422 26.35687335 + layer.1.v_cache 0.00000513 0.01378846 + layer.2.k_cache 0.00368564 4.08926511 + layer.2.v_cache 0.00001777 0.04054033 + layer.3.k_cache 0.06341467 16.56780610 + layer.3.v_cache 0.00001739 0.04968796 + layer.4.k_cache 0.00073693 1.24177591 + layer.4.v_cache 0.00004994 0.09583370 + layer.4.output 0.18261438 705.20002319 + ------------------------------------------------------------------------------------- + TOTAL 0.09097753 313.05502912 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 24948 +BPFP 0.2978 bits/point +EBPFP 0.5956 equivalent bits/point +MSE 313.055029 +---------------------- -------------------------------------------------------- +Time: 2.485s Load: 0.004s, Pack+Encode: 1.483s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 313.0550 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,580B, BPFP=0.2838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,548B, BPFP=0.4576 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,608B, BPFP=0.4684 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,240B, BPFP=0.4023 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,372B, BPFP=0.4260 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,324B, BPFP=0.4174 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,144B, BPFP=0.3851 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,436B, BPFP=0.4375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,932B, BPFP=0.5266 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,700B, BPFP=0.4849 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,452B, BPFP=0.0373 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08947531 359.70689655 + layer.0.v_cache 0.00001432 0.03432594 + layer.1.k_cache 0.05320675 25.64185412 + layer.1.v_cache 0.00000497 0.01238136 + layer.2.k_cache 0.00562141 4.05927356 + layer.2.v_cache 0.00001574 0.03677744 + layer.3.k_cache 0.02742601 16.77326071 + layer.3.v_cache 0.00001759 0.04557594 + layer.4.k_cache 0.00114861 1.29490521 + layer.4.v_cache 0.00004734 0.08552408 + layer.4.output 0.15628038 623.42708333 + ------------------------------------------------------------------------------------- + TOTAL 0.07476122 280.68707990 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 25336 +BPFP 0.2677 bits/point +EBPFP 0.5353 equivalent bits/point +MSE 280.687080 +---------------------- -------------------------------------------------------- +Time: 2.482s Load: 0.003s, Pack+Encode: 1.481s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 280.6871 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,720B, BPFP=0.2921 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,540B, BPFP=0.4314 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,752B, BPFP=0.4674 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,376B, BPFP=0.4035 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,416B, BPFP=0.4103 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,376B, BPFP=0.4035 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,852B, BPFP=0.3145 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,284B, BPFP=0.3879 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,728B, BPFP=0.4633 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,484B, BPFP=0.4219 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,296B, BPFP=0.0314 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12311893 399.88892663 + layer.0.v_cache 0.00001357 0.03407569 + layer.1.k_cache 0.03743907 24.93955927 + layer.1.v_cache 0.00000515 0.01187384 + layer.2.k_cache 0.00511221 4.05981279 + layer.2.v_cache 0.00001545 0.03484005 + layer.3.k_cache 0.04225797 16.55741949 + layer.3.v_cache 0.00001799 0.04330254 + layer.4.k_cache 0.00119426 1.13348588 + layer.4.v_cache 0.00004621 0.08304279 + layer.4.output 0.14783020 589.57958075 + ------------------------------------------------------------------------------------- + TOTAL 0.07317837 269.04961201 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 24824 +BPFP 0.2480 bits/point +EBPFP 0.4960 equivalent bits/point +MSE 269.049612 +---------------------- -------------------------------------------------------- +Time: 2.481s Load: 0.004s, Pack+Encode: 1.478s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 269.0496 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,424B, BPFP=0.2181 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,044B, BPFP=0.4663 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,048B, BPFP=0.4669 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,328B, BPFP=0.3566 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,788B, BPFP=0.4271 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,340B, BPFP=0.3585 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,540B, BPFP=0.3891 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,372B, BPFP=0.3634 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,284B, BPFP=0.5031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,676B, BPFP=0.4099 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,420B, BPFP=0.0311 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09293266 351.56824449 + layer.0.v_cache 0.00001503 0.03845785 + layer.1.k_cache 0.07854857 24.85870242 + layer.1.v_cache 0.00000559 0.01410196 + layer.2.k_cache 0.00650115 3.40402341 + layer.2.v_cache 0.00001763 0.03964354 + layer.3.k_cache 0.03715762 14.98887245 + layer.3.v_cache 0.00001788 0.04906997 + layer.4.k_cache 0.00073365 1.11233947 + layer.4.v_cache 0.00004630 0.08640324 + layer.4.output 11.17680568 526.73603817 + ------------------------------------------------------------------------------------- + TOTAL 4.61491858 240.19483094 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 27264 +BPFP 0.2457 bits/point +EBPFP 0.4913 equivalent bits/point +MSE 240.194831 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.006s, Pack+Encode: 1.476s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 240.1948 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 96, 128) +Output shape: (1, 96, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.output: torch.Size([1, 96, 3584]) -> torch.Size([1, 1, 96, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,440B, BPFP=0.2344 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,600B, BPFP=0.4232 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,872B, BPFP=0.4674 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,196B, BPFP=0.3574 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,396B, BPFP=0.3900 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,260B, BPFP=0.3678 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,348B, BPFP=0.3822 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,668B, BPFP=0.4342 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,872B, BPFP=0.4674 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,668B, BPFP=0.4342 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,428B, BPFP=0.0332 +⌛️ [2/4] FRONTEND: Frontend time: 1.477s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10158994 339.79703776 + layer.0.v_cache 0.00001467 0.04078374 + layer.1.k_cache 0.06416908 24.27071126 + layer.1.v_cache 0.00000542 0.01474584 + layer.2.k_cache 0.00400108 3.54166222 + layer.2.v_cache 0.00001872 0.04286884 + layer.3.k_cache 0.01491791 15.05536270 + layer.3.v_cache 0.00001877 0.05469393 + layer.4.k_cache 0.00070988 1.19906123 + layer.4.v_cache 0.00004832 0.09511282 + layer.4.output 0.14169503 565.00511533 + ------------------------------------------------------------------------------------- + TOTAL 0.06925641 255.24399104 + (elements=835,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 835584 +Total Bytes 25748 +BPFP 0.2465 bits/point +EBPFP 0.4930 equivalent bits/point +MSE 255.243991 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.004s, Pack+Encode: 1.477s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 255.2440 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,104B, BPFP=0.2396 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,040B, BPFP=0.4427 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,228B, BPFP=0.4835 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,796B, BPFP=0.3898 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,016B, BPFP=0.4375 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,496B, BPFP=0.3247 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,888B, BPFP=0.4097 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,708B, BPFP=0.3707 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,704B, BPFP=0.5868 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,552B, BPFP=0.5538 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,308B, BPFP=0.0406 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.003s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10321087 326.41851128 + layer.0.v_cache 0.00001557 0.03981194 + layer.1.k_cache 0.06175087 26.00962321 + layer.1.v_cache 0.00000571 0.01427729 + layer.2.k_cache 0.01329466 3.97450214 + layer.2.v_cache 0.00001784 0.04408331 + layer.3.k_cache 0.05039416 16.15573968 + layer.3.v_cache 0.00001745 0.05037991 + layer.4.k_cache 0.00079786 1.27628952 + layer.4.v_cache 0.00005071 0.09818758 + layer.4.output 0.18870974 753.27963790 + ------------------------------------------------------------------------------------- + TOTAL 0.09120729 332.17875713 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 20840 +BPFP 0.2660 bits/point +EBPFP 0.5321 equivalent bits/point +MSE 332.178757 +---------------------- -------------------------------------------------------- +Time: 2.487s Load: 0.004s, Pack+Encode: 1.480s, Decode+Unpack: 1.003s +---------------------- -------------------------------------------------------- +💾 Converting with 332.1788 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,116B, BPFP=0.2422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,156B, BPFP=0.4679 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,408B, BPFP=0.5226 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,652B, BPFP=0.3585 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,068B, BPFP=0.4488 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,476B, BPFP=0.3203 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,772B, BPFP=0.3845 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,300B, BPFP=0.2821 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,760B, BPFP=0.5990 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,340B, BPFP=0.5078 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,328B, BPFP=0.0412 +⌛️ [2/4] FRONTEND: Frontend time: 1.487s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09666174 336.27658420 + layer.0.v_cache 0.00001715 0.03797723 + layer.1.k_cache 0.08476152 26.13342624 + layer.1.v_cache 0.00000506 0.01386276 + layer.2.k_cache 0.00238475 3.96062554 + layer.2.v_cache 0.00001567 0.03890662 + layer.3.k_cache 0.07332428 16.66539171 + layer.3.v_cache 0.00001796 0.05094508 + layer.4.k_cache 0.00072290 1.26689000 + layer.4.v_cache 0.00005092 0.09452808 + layer.4.output 0.18875189 753.33990575 + ------------------------------------------------------------------------------------- + TOTAL 0.09289560 332.81873399 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 20376 +BPFP 0.2601 bits/point +EBPFP 0.5202 equivalent bits/point +MSE 332.818734 +---------------------- -------------------------------------------------------- +Time: 2.502s Load: 0.003s, Pack+Encode: 1.487s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 332.8187 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 992B, BPFP=0.2279 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,108B, BPFP=0.4844 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,292B, BPFP=0.5267 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,564B, BPFP=0.3594 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,704B, BPFP=0.3915 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,396B, BPFP=0.3208 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,380B, BPFP=0.3171 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,480B, BPFP=0.3401 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,124B, BPFP=0.4881 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,412B, BPFP=0.5542 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,436B, BPFP=0.0471 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10601801 324.08895335 + layer.0.v_cache 0.00001817 0.04117969 + layer.1.k_cache 0.01696859 26.16561172 + layer.1.v_cache 0.00000501 0.01535804 + layer.2.k_cache 0.00274148 4.12069433 + layer.2.v_cache 0.00001757 0.04513342 + layer.3.k_cache 0.01706722 17.08964090 + layer.3.v_cache 0.00001843 0.05430673 + layer.4.k_cache 0.00080906 1.33241250 + layer.4.v_cache 0.00005076 0.09878314 + layer.4.output 0.19983917 797.62710084 + ------------------------------------------------------------------------------------- + TOTAL 0.09074050 350.37892822 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 18888 +BPFP 0.2553 bits/point +EBPFP 0.5106 equivalent bits/point +MSE 350.378928 +---------------------- -------------------------------------------------------- +Time: 2.496s Load: 0.003s, Pack+Encode: 1.484s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 350.3789 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 65, 128) +Output shape: (1, 65, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.output: torch.Size([1, 65, 3584]) -> torch.Size([1, 1, 65, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 932B, BPFP=0.2240 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,096B, BPFP=0.5038 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,012B, BPFP=0.4837 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,496B, BPFP=0.3596 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,556B, BPFP=0.3740 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,356B, BPFP=0.3260 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,200B, BPFP=0.2885 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,500B, BPFP=0.3606 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,028B, BPFP=0.4875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,308B, BPFP=0.5548 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,324B, BPFP=0.0455 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09579474 324.22121394 + layer.0.v_cache 0.00001312 0.03846317 + layer.1.k_cache 0.01648773 27.35360577 + layer.1.v_cache 0.00000712 0.01404418 + layer.2.k_cache 0.00431219 4.17178392 + layer.2.v_cache 0.00001803 0.04343366 + layer.3.k_cache 0.03460605 16.17528358 + layer.3.v_cache 0.00001721 0.05352320 + layer.4.k_cache 0.00068162 1.34558575 + layer.4.v_cache 0.00004727 0.10452683 + layer.4.output 0.20898957 834.37671703 + ------------------------------------------------------------------------------------- + TOTAL 0.09499483 365.53873431 + (elements=565,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 565760 +Total Bytes 17808 +BPFP 0.2518 bits/point +EBPFP 0.5036 equivalent bits/point +MSE 365.538734 +---------------------- -------------------------------------------------------- +Time: 2.497s Load: 0.003s, Pack+Encode: 1.483s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 365.5387 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 41, 128) +Output shape: (1, 41, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.output: torch.Size([1, 41, 3584]) -> torch.Size([1, 1, 41, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,308B, BPFP=0.4985 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,820B, BPFP=0.6936 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,104B, BPFP=0.8018 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,504B, BPFP=0.5732 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,932B, BPFP=0.7363 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,348B, BPFP=0.5137 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,172B, BPFP=0.4466 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,608B, BPFP=0.6128 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,116B, BPFP=0.8064 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,908B, BPFP=0.7271 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,588B, BPFP=0.0865 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.970s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10098721 422.87852515 + layer.0.v_cache 0.00001362 0.04613280 + layer.1.k_cache 0.01708067 27.45321432 + layer.1.v_cache 0.00000533 0.01616826 + layer.2.k_cache 0.00250472 4.51584234 + layer.2.v_cache 0.00001755 0.04986355 + layer.3.k_cache 0.08199174 17.77173745 + layer.3.v_cache 0.00001815 0.05649756 + layer.4.k_cache 0.00059437 1.36439570 + layer.4.v_cache 0.00005232 0.12377630 + layer.4.output 0.35770382 1323.90755662 + ------------------------------------------------------------------------------------- + TOTAL 0.15924661 573.03700293 + (elements=356,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 356864 +Total Bytes 18408 +BPFP 0.4127 bits/point +EBPFP 0.8253 equivalent bits/point +MSE 573.037003 +---------------------- -------------------------------------------------------- +Time: 2.454s Load: 0.002s, Pack+Encode: 1.482s, Decode+Unpack: 0.970s +---------------------- -------------------------------------------------------- +💾 Converting with 573.0370 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 936B, BPFP=0.2812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,764B, BPFP=0.5300 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,064B, BPFP=0.6202 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,736B, BPFP=0.5216 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,588B, BPFP=0.4772 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,748B, BPFP=0.5252 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,276B, BPFP=0.3834 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,820B, BPFP=0.5469 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,816B, BPFP=0.5457 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,792B, BPFP=0.5385 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,052B, BPFP=0.0881 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.967s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09210638 355.02246094 + layer.0.v_cache 0.00001369 0.04368912 + layer.1.k_cache 0.01682195 28.48412382 + layer.1.v_cache 0.00000520 0.01643402 + layer.2.k_cache 0.00487203 4.39924680 + layer.2.v_cache 0.00001755 0.05003485 + layer.3.k_cache 0.04965206 16.02892127 + layer.3.v_cache 0.00001944 0.05993151 + layer.4.k_cache 0.00060442 1.44575750 + layer.4.v_cache 0.00005207 0.11186441 + layer.4.output 0.26182200 1044.26304945 + ------------------------------------------------------------------------------------- + TOTAL 0.11746581 453.85316532 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 18592 +BPFP 0.3286 bits/point +EBPFP 0.6572 equivalent bits/point +MSE 453.853165 +---------------------- -------------------------------------------------------- +Time: 2.461s Load: 0.002s, Pack+Encode: 1.491s, Decode+Unpack: 0.967s +---------------------- -------------------------------------------------------- +💾 Converting with 453.8532 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 47, 128) +Output shape: (1, 47, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.output: torch.Size([1, 47, 3584]) -> torch.Size([1, 1, 47, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,532B, BPFP=0.5093 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,764B, BPFP=0.5864 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,080B, BPFP=0.6915 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,784B, BPFP=0.5931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,728B, BPFP=0.5745 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,764B, BPFP=0.5864 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,432B, BPFP=0.4761 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,580B, BPFP=0.5253 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,900B, BPFP=0.6316 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,964B, BPFP=0.6529 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,892B, BPFP=0.0899 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.973s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08461995 429.41755319 + layer.0.v_cache 0.00001366 0.04273476 + layer.1.k_cache 0.01838745 26.63430851 + layer.1.v_cache 0.00000609 0.01711190 + layer.2.k_cache 0.00780689 4.30127051 + layer.2.v_cache 0.00001932 0.04811774 + layer.3.k_cache 0.04292001 16.68184191 + layer.3.v_cache 0.00001942 0.05798606 + layer.4.k_cache 0.00059704 1.29400237 + layer.4.v_cache 0.00004850 0.11180318 + layer.4.output 0.29441516 1155.27925532 + ------------------------------------------------------------------------------------- + TOTAL 0.13031438 503.85655985 + (elements=409,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 409088 +Total Bytes 19420 +BPFP 0.3798 bits/point +EBPFP 0.7595 equivalent bits/point +MSE 503.856560 +---------------------- -------------------------------------------------------- +Time: 2.478s Load: 0.002s, Pack+Encode: 1.503s, Decode+Unpack: 0.973s +---------------------- -------------------------------------------------------- +💾 Converting with 503.8566 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,356B, BPFP=0.3717 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,536B, BPFP=0.4211 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,872B, BPFP=0.5132 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,560B, BPFP=0.4276 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,452B, BPFP=0.3980 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,580B, BPFP=0.4331 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,288B, BPFP=0.3531 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,624B, BPFP=0.4452 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,680B, BPFP=0.4605 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,600B, BPFP=0.4386 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,292B, BPFP=0.0898 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.970s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10653571 447.68523849 + layer.0.v_cache 0.00001396 0.04566411 + layer.1.k_cache 0.01513677 28.85791658 + layer.1.v_cache 0.00000574 0.01715373 + layer.2.k_cache 0.00241650 5.15730901 + layer.2.v_cache 0.00001997 0.05859008 + layer.3.k_cache 0.04250543 17.92357328 + layer.3.v_cache 0.00001889 0.06177187 + layer.4.k_cache 0.00060099 1.56856001 + layer.4.v_cache 0.00005317 0.12697233 + layer.4.output 0.25650722 951.30083020 + ------------------------------------------------------------------------------------- + TOTAL 0.11546222 421.21226829 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 17840 +BPFP 0.2877 bits/point +EBPFP 0.5753 equivalent bits/point +MSE 421.212268 +---------------------- -------------------------------------------------------- +Time: 2.469s Load: 0.002s, Pack+Encode: 1.497s, Decode+Unpack: 0.970s +---------------------- -------------------------------------------------------- +💾 Converting with 421.2123 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,508B, BPFP=0.4909 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,820B, BPFP=0.5924 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,872B, BPFP=0.6094 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,788B, BPFP=0.5820 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,832B, BPFP=0.5964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,764B, BPFP=0.5742 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,448B, BPFP=0.4714 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,840B, BPFP=0.5990 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,004B, BPFP=0.6523 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,796B, BPFP=0.5846 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,612B, BPFP=0.0750 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.969s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10088900 426.03458659 + layer.0.v_cache 0.00001321 0.04162998 + layer.1.k_cache 0.01601078 25.62658183 + layer.1.v_cache 0.00000529 0.01602267 + layer.2.k_cache 0.00240732 4.37183889 + layer.2.v_cache 0.00001858 0.04948333 + layer.3.k_cache 0.04236705 17.81002299 + layer.3.v_cache 0.00001968 0.05823201 + layer.4.k_cache 0.00061965 1.38210948 + layer.4.v_cache 0.00005235 0.12206501 + layer.4.output 0.30498679 1130.78906250 + ------------------------------------------------------------------------------------- + TOTAL 0.13513591 493.59035355 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 19284 +BPFP 0.3693 bits/point +EBPFP 0.7385 equivalent bits/point +MSE 493.590354 +---------------------- -------------------------------------------------------- +Time: 2.456s Load: 0.002s, Pack+Encode: 1.485s, Decode+Unpack: 0.969s +---------------------- -------------------------------------------------------- +💾 Converting with 493.5904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 53, 128) +Output shape: (1, 53, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.output: torch.Size([1, 53, 3584]) -> torch.Size([1, 1, 53, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 928B, BPFP=0.2736 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,712B, BPFP=0.5047 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,008B, BPFP=0.5920 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,576B, BPFP=0.4646 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,468B, BPFP=0.4328 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,736B, BPFP=0.5118 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,188B, BPFP=0.3502 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,536B, BPFP=0.4528 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,880B, BPFP=0.5542 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,860B, BPFP=0.5483 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,212B, BPFP=0.0932 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.970s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10150213 348.56227889 + layer.0.v_cache 0.00001370 0.04712537 + layer.1.k_cache 0.01753690 28.94739921 + layer.1.v_cache 0.00000536 0.01807332 + layer.2.k_cache 0.00234856 4.40360620 + layer.2.v_cache 0.00001946 0.05615285 + layer.3.k_cache 0.03772523 16.48583063 + layer.3.v_cache 0.00001792 0.05961553 + layer.4.k_cache 0.00061365 1.52677802 + layer.4.v_cache 0.00006478 0.12365120 + layer.4.output 0.25806696 1024.27181604 + ------------------------------------------------------------------------------------- + TOTAL 0.11566567 445.30195432 + (elements=461,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 461312 +Total Bytes 18104 +BPFP 0.3140 bits/point +EBPFP 0.6279 equivalent bits/point +MSE 445.301954 +---------------------- -------------------------------------------------------- +Time: 2.457s Load: 0.002s, Pack+Encode: 1.485s, Decode+Unpack: 0.970s +---------------------- -------------------------------------------------------- +💾 Converting with 445.3020 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,276B, BPFP=0.3323 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,484B, BPFP=0.3865 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,628B, BPFP=0.4240 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,328B, BPFP=0.3458 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,324B, BPFP=0.3448 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,448B, BPFP=0.3771 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,332B, BPFP=0.3469 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,576B, BPFP=0.4104 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,564B, BPFP=0.4073 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,524B, BPFP=0.3969 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,548B, BPFP=0.0576 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.970s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11850098 435.45009766 + layer.0.v_cache 0.00001385 0.04281216 + layer.1.k_cache 0.01468890 28.01074219 + layer.1.v_cache 0.00000523 0.01667481 + layer.2.k_cache 0.01477869 5.04079030 + layer.2.v_cache 0.00001915 0.05493811 + layer.3.k_cache 0.14661819 18.32333781 + layer.3.v_cache 0.00001925 0.06766455 + layer.4.k_cache 0.00060183 1.46952235 + layer.4.v_cache 0.00005752 0.12488545 + layer.4.output 0.22966646 905.00327381 + ------------------------------------------------------------------------------------- + TOTAL 0.11193934 401.38966953 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 16032 +BPFP 0.2456 bits/point +EBPFP 0.4912 equivalent bits/point +MSE 401.389670 +---------------------- -------------------------------------------------------- +Time: 2.457s Load: 0.002s, Pack+Encode: 1.484s, Decode+Unpack: 0.970s +---------------------- -------------------------------------------------------- +💾 Converting with 401.3897 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,516B, BPFP=0.2575 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,628B, BPFP=0.4463 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,736B, BPFP=0.4647 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,272B, BPFP=0.3859 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,548B, BPFP=0.4327 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,388B, BPFP=0.4056 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,260B, BPFP=0.3838 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,548B, BPFP=0.4327 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,648B, BPFP=0.4497 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,620B, BPFP=0.4450 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,252B, BPFP=0.0304 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10836371 366.00611413 + layer.0.v_cache 0.00001445 0.03806100 + layer.1.k_cache 0.06846340 24.89815090 + layer.1.v_cache 0.00000523 0.01341220 + layer.2.k_cache 0.00506382 3.41833828 + layer.2.v_cache 0.00001740 0.04095441 + layer.3.k_cache 0.03973729 16.06075121 + layer.3.v_cache 0.00001777 0.04933540 + layer.4.k_cache 0.00066797 1.14865419 + layer.4.v_cache 0.00005099 0.09168608 + layer.4.output 0.14786712 589.56177213 + ------------------------------------------------------------------------------------- + TOTAL 0.07396893 266.98222722 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 25416 +BPFP 0.2539 bits/point +EBPFP 0.5078 equivalent bits/point +MSE 266.982227 +---------------------- -------------------------------------------------------- +Time: 2.487s Load: 0.004s, Pack+Encode: 1.482s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 266.9822 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 932B, BPFP=0.2206 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,144B, BPFP=0.5076 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,208B, BPFP=0.5227 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,588B, BPFP=0.3759 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,540B, BPFP=0.3646 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,452B, BPFP=0.3438 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,296B, BPFP=0.3068 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,472B, BPFP=0.3485 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,936B, BPFP=0.4583 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,296B, BPFP=0.5436 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,416B, BPFP=0.0479 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08728886 324.57306463 + layer.0.v_cache 0.00001734 0.04403128 + layer.1.k_cache 0.01499850 27.34375740 + layer.1.v_cache 0.00000528 0.01575874 + layer.2.k_cache 0.00240521 4.10595934 + layer.2.v_cache 0.00001739 0.04672467 + layer.3.k_cache 0.03126823 16.01873964 + layer.3.v_cache 0.00001843 0.05941733 + layer.4.k_cache 0.00061395 1.36123230 + layer.4.v_cache 0.00004830 0.10539082 + layer.4.output 0.21953388 822.59645563 + ------------------------------------------------------------------------------------- + TOTAL 0.09843639 360.69701562 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 18280 +BPFP 0.2546 bits/point +EBPFP 0.5091 equivalent bits/point +MSE 360.697016 +---------------------- -------------------------------------------------------- +Time: 2.485s Load: 0.004s, Pack+Encode: 1.480s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 360.6970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 936B, BPFP=0.2216 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,056B, BPFP=0.4867 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,536B, BPFP=0.6004 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,724B, BPFP=0.4081 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,652B, BPFP=0.3911 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,516B, BPFP=0.3589 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,216B, BPFP=0.2879 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,372B, BPFP=0.3248 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,216B, BPFP=0.5246 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,232B, BPFP=0.5284 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,428B, BPFP=0.0483 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.000s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08223097 329.72085109 + layer.0.v_cache 0.00001442 0.04386817 + layer.1.k_cache 0.01480860 27.10246878 + layer.1.v_cache 0.00000502 0.01546672 + layer.2.k_cache 0.00576455 4.07815136 + layer.2.v_cache 0.00002304 0.04601576 + layer.3.k_cache 0.03215047 16.66188928 + layer.3.v_cache 0.00001933 0.05991285 + layer.4.k_cache 0.00061510 1.31878708 + layer.4.v_cache 0.00004916 0.10588221 + layer.4.output 0.22077460 822.47362013 + ------------------------------------------------------------------------------------- + TOTAL 0.09888840 360.96874319 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 18884 +BPFP 0.2630 bits/point +EBPFP 0.5260 equivalent bits/point +MSE 360.968743 +---------------------- -------------------------------------------------------- +Time: 2.481s Load: 0.003s, Pack+Encode: 1.478s, Decode+Unpack: 1.000s +---------------------- -------------------------------------------------------- +💾 Converting with 360.9687 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,132B, BPFP=0.2423 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,004B, BPFP=0.6430 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,312B, BPFP=0.4949 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,232B, BPFP=0.4777 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,192B, BPFP=0.4692 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,800B, BPFP=0.3853 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,816B, BPFP=0.3887 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,104B, BPFP=0.4503 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,828B, BPFP=0.6053 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,548B, BPFP=0.5454 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,428B, BPFP=0.0437 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10114760 332.57587757 + layer.0.v_cache 0.00001420 0.03864249 + layer.1.k_cache 0.03683303 26.46520662 + layer.1.v_cache 0.00000544 0.01444627 + layer.2.k_cache 0.00248627 4.05105424 + layer.2.v_cache 0.00001713 0.04210615 + layer.3.k_cache 0.02909129 16.47352914 + layer.3.v_cache 0.00001766 0.05116499 + layer.4.k_cache 0.00072613 1.25509706 + layer.4.v_cache 0.00005154 0.09814984 + layer.4.output 0.19731793 743.69208659 + ------------------------------------------------------------------------------------- + TOTAL 0.09127152 328.64175768 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 23396 +BPFP 0.2946 bits/point +EBPFP 0.5891 equivalent bits/point +MSE 328.641758 +---------------------- -------------------------------------------------------- +Time: 2.491s Load: 0.004s, Pack+Encode: 1.483s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 328.6418 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,212B, BPFP=0.2492 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,992B, BPFP=0.6151 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,428B, BPFP=0.4992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,204B, BPFP=0.4531 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,408B, BPFP=0.4951 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,856B, BPFP=0.3816 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,968B, BPFP=0.4046 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,904B, BPFP=0.3914 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,084B, BPFP=0.6340 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,716B, BPFP=0.5584 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,368B, BPFP=0.0402 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10230959 326.65851151 + layer.0.v_cache 0.00001409 0.04010123 + layer.1.k_cache 0.07973289 25.70710192 + layer.1.v_cache 0.00000506 0.01375743 + layer.2.k_cache 0.00553133 4.07440507 + layer.2.v_cache 0.00001928 0.04415303 + layer.3.k_cache 0.03946657 17.11547209 + layer.3.v_cache 0.00001980 0.05623595 + layer.4.k_cache 0.00066525 1.24444078 + layer.4.v_cache 0.00005119 0.10025745 + layer.4.output 0.17985767 714.16823308 + ------------------------------------------------------------------------------------- + TOTAL 0.08745993 316.13129812 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 24140 +BPFP 0.2919 bits/point +EBPFP 0.5839 equivalent bits/point +MSE 316.131298 +---------------------- -------------------------------------------------------- +Time: 2.488s Load: 0.004s, Pack+Encode: 1.483s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 316.1313 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,196B, BPFP=0.2427 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,920B, BPFP=0.5925 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,896B, BPFP=0.5877 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,408B, BPFP=0.4886 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,104B, BPFP=0.4269 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,860B, BPFP=0.3774 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,056B, BPFP=0.4172 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,444B, BPFP=0.4959 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,032B, BPFP=0.6153 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,436B, BPFP=0.4943 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,332B, BPFP=0.0386 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633976 334.41259131 + layer.0.v_cache 0.00001461 0.03879194 + layer.1.k_cache 0.07748085 25.88162826 + layer.1.v_cache 0.00000515 0.01295549 + layer.2.k_cache 0.00505460 3.97134994 + layer.2.v_cache 0.00001651 0.03846043 + layer.3.k_cache 0.02793909 16.77468040 + layer.3.v_cache 0.00001809 0.04539782 + layer.4.k_cache 0.00070819 1.23135634 + layer.4.v_cache 0.00004716 0.09299654 + layer.4.output 0.18765952 705.18297774 + ------------------------------------------------------------------------------------- + TOTAL 0.09007298 312.86947368 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 24684 +BPFP 0.2946 bits/point +EBPFP 0.5893 equivalent bits/point +MSE 312.869474 +---------------------- -------------------------------------------------------- +Time: 2.483s Load: 0.004s, Pack+Encode: 1.480s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 312.8695 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,068B, BPFP=0.2384 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,420B, BPFP=0.5402 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,480B, BPFP=0.5536 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,728B, BPFP=0.3857 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,720B, BPFP=0.3839 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,560B, BPFP=0.3482 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,664B, BPFP=0.3714 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,748B, BPFP=0.3902 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,580B, BPFP=0.5759 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,352B, BPFP=0.5250 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,344B, BPFP=0.0429 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08399124 332.52087054 + layer.0.v_cache 0.00001671 0.04085742 + layer.1.k_cache 0.03849935 26.05573382 + layer.1.v_cache 0.00000503 0.01432048 + layer.2.k_cache 0.00738276 4.03642055 + layer.2.v_cache 0.00001692 0.04264802 + layer.3.k_cache 0.04914829 16.21256278 + layer.3.v_cache 0.00001713 0.05127693 + layer.4.k_cache 0.00063731 1.23807046 + layer.4.v_cache 0.00004711 0.09647427 + layer.4.output 0.21328995 775.58781888 + ------------------------------------------------------------------------------------- + TOTAL 0.09839950 341.73082161 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 20664 +BPFP 0.2713 bits/point +EBPFP 0.5426 equivalent bits/point +MSE 341.730822 +---------------------- -------------------------------------------------------- +Time: 2.483s Load: 0.003s, Pack+Encode: 1.481s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 341.7308 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 56, 128) +Output shape: (1, 56, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.output: torch.Size([1, 56, 3584]) -> torch.Size([1, 1, 56, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 948B, BPFP=0.2645 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,620B, BPFP=0.4520 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,940B, BPFP=0.5413 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,576B, BPFP=0.4397 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,432B, BPFP=0.3996 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 996B, BPFP=0.2779 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,296B, BPFP=0.3616 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,644B, BPFP=0.4587 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,688B, BPFP=0.4710 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,632B, BPFP=0.4554 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,280B, BPFP=0.0909 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.967s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13768739 361.59116908 + layer.0.v_cache 0.00001432 0.04641489 + layer.1.k_cache 0.01638686 29.47940281 + layer.1.v_cache 0.00000533 0.01776659 + layer.2.k_cache 0.00251956 4.91015407 + layer.2.v_cache 0.00001745 0.04680534 + layer.3.k_cache 0.08599430 17.17974636 + layer.3.v_cache 0.00002012 0.06754610 + layer.4.k_cache 0.00061995 1.66650309 + layer.4.v_cache 0.00004798 0.11956932 + layer.4.output 0.24256272 968.48174426 + ------------------------------------------------------------------------------------- + TOTAL 0.11419131 423.20572279 + (elements=487,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 487424 +Total Bytes 17052 +BPFP 0.2799 bits/point +EBPFP 0.5597 equivalent bits/point +MSE 423.205723 +---------------------- -------------------------------------------------------- +Time: 2.455s Load: 0.003s, Pack+Encode: 1.485s, Decode+Unpack: 0.967s +---------------------- -------------------------------------------------------- +💾 Converting with 423.2057 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,232B, BPFP=0.2406 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,032B, BPFP=0.5922 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,944B, BPFP=0.5750 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,864B, BPFP=0.3641 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,556B, BPFP=0.4992 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,084B, BPFP=0.4070 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,344B, BPFP=0.4578 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,076B, BPFP=0.4055 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,096B, BPFP=0.6047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,964B, BPFP=0.5789 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,396B, BPFP=0.0390 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09278157 321.59721680 + layer.0.v_cache 0.00001504 0.03897539 + layer.1.k_cache 0.07523044 25.99208984 + layer.1.v_cache 0.00000517 0.01447491 + layer.2.k_cache 0.00381728 4.00992584 + layer.2.v_cache 0.00001789 0.04135511 + layer.3.k_cache 0.03327554 17.38752441 + layer.3.v_cache 0.00001939 0.04958190 + layer.4.k_cache 0.00067710 1.28365927 + layer.4.v_cache 0.00005192 0.09723191 + layer.4.output 0.16994183 678.00574777 + ------------------------------------------------------------------------------------- + TOTAL 0.08208730 300.97366293 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 25588 +BPFP 0.2940 bits/point +EBPFP 0.5880 equivalent bits/point +MSE 300.973663 +---------------------- -------------------------------------------------------- +Time: 2.503s Load: 0.003s, Pack+Encode: 1.502s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 300.9737 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,296B, BPFP=0.3432 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,540B, BPFP=0.4078 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,704B, BPFP=0.4513 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,432B, BPFP=0.3792 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,356B, BPFP=0.3591 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,400B, BPFP=0.3708 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,412B, BPFP=0.3739 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,516B, BPFP=0.4015 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,640B, BPFP=0.4343 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,580B, BPFP=0.4184 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,516B, BPFP=0.0574 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.967s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12468526 452.87943591 + layer.0.v_cache 0.00001451 0.05079203 + layer.1.k_cache 0.01313243 28.03491625 + layer.1.v_cache 0.00000580 0.01908658 + layer.2.k_cache 0.00906488 5.28839991 + layer.2.v_cache 0.00002099 0.05666248 + layer.3.k_cache 0.01569995 18.82567946 + layer.3.v_cache 0.00002002 0.06983362 + layer.4.k_cache 0.00064964 1.54308022 + layer.4.v_cache 0.00005478 0.12760522 + layer.4.output 0.23031524 919.12378935 + ------------------------------------------------------------------------------------- + TOTAL 0.10444441 408.28011865 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 16392 +BPFP 0.2554 bits/point +EBPFP 0.5107 equivalent bits/point +MSE 408.280119 +---------------------- -------------------------------------------------------- +Time: 2.453s Load: 0.003s, Pack+Encode: 1.483s, Decode+Unpack: 0.967s +---------------------- -------------------------------------------------------- +💾 Converting with 408.2801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,296B, BPFP=0.3375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,492B, BPFP=0.3885 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,648B, BPFP=0.4292 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,512B, BPFP=0.3937 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,340B, BPFP=0.3490 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,360B, BPFP=0.3542 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,308B, BPFP=0.3406 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,464B, BPFP=0.3812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,548B, BPFP=0.4031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,600B, BPFP=0.4167 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,564B, BPFP=0.0582 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.981s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15164433 432.27503255 + layer.0.v_cache 0.00001435 0.04867161 + layer.1.k_cache 0.01704682 27.61416829 + layer.1.v_cache 0.00000573 0.01885505 + layer.2.k_cache 0.00240009 5.11493835 + layer.2.v_cache 0.00001845 0.05355507 + layer.3.k_cache 0.02161660 19.60649007 + layer.3.v_cache 0.00001985 0.06656426 + layer.4.k_cache 0.00061459 1.51572393 + layer.4.v_cache 0.00005275 0.12641054 + layer.4.output 0.22653077 903.84724702 + ------------------------------------------------------------------------------------- + TOTAL 0.10465582 400.78653758 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 16132 +BPFP 0.2471 bits/point +EBPFP 0.4942 equivalent bits/point +MSE 400.786538 +---------------------- -------------------------------------------------------- +Time: 2.481s Load: 0.003s, Pack+Encode: 1.498s, Decode+Unpack: 0.981s +---------------------- -------------------------------------------------------- +💾 Converting with 400.7865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,112B, BPFP=0.2380 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,304B, BPFP=0.4932 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,136B, BPFP=0.4572 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,764B, BPFP=0.3776 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,120B, BPFP=0.4538 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,704B, BPFP=0.3647 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,164B, BPFP=0.4632 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,644B, BPFP=0.3519 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,984B, BPFP=0.6387 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,424B, BPFP=0.5188 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,416B, BPFP=0.0433 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08869729 325.79826627 + layer.0.v_cache 0.00001594 0.03808249 + layer.1.k_cache 0.08105310 25.39895588 + layer.1.v_cache 0.00000494 0.01406390 + layer.2.k_cache 0.00256802 4.05612183 + layer.2.v_cache 0.00001661 0.03860593 + layer.3.k_cache 0.01835501 16.44572219 + layer.3.v_cache 0.00001861 0.05248822 + layer.4.k_cache 0.00078324 1.29584869 + layer.4.v_cache 0.00004769 0.09187401 + layer.4.output 0.18611241 742.90790117 + ------------------------------------------------------------------------------------- + TOTAL 0.08790278 327.85796104 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 21772 +BPFP 0.2741 bits/point +EBPFP 0.5482 equivalent bits/point +MSE 327.857961 +---------------------- -------------------------------------------------------- +Time: 2.495s Load: 0.003s, Pack+Encode: 1.491s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 327.8580 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,400B, BPFP=0.2255 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,580B, BPFP=0.4156 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,880B, BPFP=0.4639 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,204B, BPFP=0.3550 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,320B, BPFP=0.3737 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,268B, BPFP=0.3653 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,372B, BPFP=0.3821 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,216B, BPFP=0.3570 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,968B, BPFP=0.4781 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,644B, BPFP=0.4259 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,408B, BPFP=0.0324 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17228427 356.28213595 + layer.0.v_cache 0.00001543 0.03956976 + layer.1.k_cache 0.04621154 24.34079766 + layer.1.v_cache 0.00000545 0.01440873 + layer.2.k_cache 0.00739691 3.68257818 + layer.2.v_cache 0.00001797 0.04212705 + layer.3.k_cache 0.03096496 14.75853234 + layer.3.v_cache 0.00001890 0.05366249 + layer.4.k_cache 0.00071091 1.25669908 + layer.4.v_cache 0.00004900 0.08934182 + layer.4.output 0.02450962 574.16596097 + ------------------------------------------------------------------------------------- + TOTAL 0.02524957 259.98362234 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 25260 +BPFP 0.2393 bits/point +EBPFP 0.4787 equivalent bits/point +MSE 259.983622 +---------------------- -------------------------------------------------------- +Time: 2.495s Load: 0.004s, Pack+Encode: 1.490s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 259.9836 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.2231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,524B, BPFP=0.4241 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,584B, BPFP=0.4341 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,460B, BPFP=0.4133 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,304B, BPFP=0.3871 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,388B, BPFP=0.4012 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,344B, BPFP=0.3938 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,392B, BPFP=0.4019 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,672B, BPFP=0.4489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,640B, BPFP=0.4435 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,344B, BPFP=0.0323 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16174046 322.83896169 + layer.0.v_cache 0.00001612 0.04085876 + layer.1.k_cache 0.03352992 25.05597908 + layer.1.v_cache 0.00000563 0.01502586 + layer.2.k_cache 0.00496730 3.75750339 + layer.2.v_cache 0.00001740 0.04407053 + layer.3.k_cache 0.01521110 16.08016738 + layer.3.v_cache 0.00001939 0.05348360 + layer.4.k_cache 0.00066431 1.11354016 + layer.4.v_cache 0.00005067 0.09416375 + layer.4.output 0.14624557 583.24270353 + ------------------------------------------------------------------------------------- + TOTAL 0.07293772 261.87015758 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 24980 +BPFP 0.2469 bits/point +EBPFP 0.4938 equivalent bits/point +MSE 261.870158 +---------------------- -------------------------------------------------------- +Time: 2.492s Load: 0.004s, Pack+Encode: 1.483s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 261.8702 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,244B, BPFP=0.2260 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,440B, BPFP=0.4433 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,800B, BPFP=0.5087 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,044B, BPFP=0.3714 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,572B, BPFP=0.4673 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,396B, BPFP=0.4353 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,324B, BPFP=0.4222 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,444B, BPFP=0.4440 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,992B, BPFP=0.5436 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,712B, BPFP=0.4927 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,540B, BPFP=0.0400 +⌛️ [2/4] FRONTEND: Frontend time: 1.486s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.002s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07121617 321.32449128 + layer.0.v_cache 0.00001681 0.04005738 + layer.1.k_cache 0.06877367 24.81586119 + layer.1.v_cache 0.00000530 0.01413104 + layer.2.k_cache 0.00486504 3.78537803 + layer.2.v_cache 0.00001749 0.03920898 + layer.3.k_cache 0.08833314 16.28684536 + layer.3.v_cache 0.00001723 0.04977792 + layer.4.k_cache 0.00077278 1.25853623 + layer.4.v_cache 0.00004595 0.09241569 + layer.4.output 0.15812202 630.70675872 + ------------------------------------------------------------------------------------- + TOTAL 0.07887751 281.33258907 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 25508 +BPFP 0.2726 bits/point +EBPFP 0.5452 equivalent bits/point +MSE 281.332589 +---------------------- -------------------------------------------------------- +Time: 2.492s Load: 0.004s, Pack+Encode: 1.486s, Decode+Unpack: 1.002s +---------------------- -------------------------------------------------------- +💾 Converting with 281.3326 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,400B, BPFP=0.2166 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,872B, BPFP=0.4443 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,228B, BPFP=0.4994 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,404B, BPFP=0.3719 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,592B, BPFP=0.4010 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,696B, BPFP=0.4171 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,544B, BPFP=0.3936 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,904B, BPFP=0.4493 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,972B, BPFP=0.4598 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,824B, BPFP=0.4369 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,408B, BPFP=0.0311 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.002s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13964315 346.01794554 + layer.0.v_cache 0.00001432 0.04014608 + layer.1.k_cache 0.04810369 26.65324296 + layer.1.v_cache 0.00000572 0.01477160 + layer.2.k_cache 0.00492308 3.56631681 + layer.2.v_cache 0.00001777 0.04190021 + layer.3.k_cache 0.04255116 15.22540464 + layer.3.v_cache 0.00001880 0.04831967 + layer.4.k_cache 0.00073971 1.17343729 + layer.4.v_cache 0.00004573 0.09069739 + layer.4.output 11.28739116 531.92278112 + ------------------------------------------------------------------------------------- + TOTAL 4.66163537 242.13715588 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 27844 +BPFP 0.2534 bits/point +EBPFP 0.5068 equivalent bits/point +MSE 242.137156 +---------------------- -------------------------------------------------------- +Time: 2.488s Load: 0.005s, Pack+Encode: 1.480s, Decode+Unpack: 1.002s +---------------------- -------------------------------------------------------- +💾 Converting with 242.1372 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,556B, BPFP=0.2763 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,756B, BPFP=0.4893 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,736B, BPFP=0.4858 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,688B, BPFP=0.4773 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,296B, BPFP=0.4077 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,460B, BPFP=0.4368 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,032B, BPFP=0.3608 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,552B, BPFP=0.4531 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,016B, BPFP=0.5355 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,584B, BPFP=0.4588 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,516B, BPFP=0.0385 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.003s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14145413 348.12770774 + layer.0.v_cache 0.00001516 0.04008749 + layer.1.k_cache 0.10647225 25.88795610 + layer.1.v_cache 0.00000560 0.01449528 + layer.2.k_cache 0.00244946 3.82730865 + layer.2.v_cache 0.00001685 0.04020501 + layer.3.k_cache 0.01581429 15.76096691 + layer.3.v_cache 0.00001841 0.04867811 + layer.4.k_cache 0.00072023 1.26412279 + layer.4.v_cache 0.00004652 0.08903457 + layer.4.output 0.15451367 616.38189935 + ------------------------------------------------------------------------------------- + TOTAL 0.07932992 277.04552106 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 26192 +BPFP 0.2736 bits/point +EBPFP 0.5471 equivalent bits/point +MSE 277.045521 +---------------------- -------------------------------------------------------- +Time: 2.492s Load: 0.003s, Pack+Encode: 1.485s, Decode+Unpack: 1.003s +---------------------- -------------------------------------------------------- +💾 Converting with 277.0455 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,160B, BPFP=0.2449 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,968B, BPFP=0.6267 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,168B, BPFP=0.4578 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,364B, BPFP=0.4992 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,996B, BPFP=0.4215 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,672B, BPFP=0.3530 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,020B, BPFP=0.4265 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,168B, BPFP=0.4578 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,600B, BPFP=0.5490 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,372B, BPFP=0.5008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,328B, BPFP=0.0401 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.002s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10237078 336.25541068 + layer.0.v_cache 0.00001391 0.03743682 + layer.1.k_cache 0.05953315 25.56115393 + layer.1.v_cache 0.00000517 0.01367513 + layer.2.k_cache 0.00728044 3.99290508 + layer.2.v_cache 0.00001648 0.04002871 + layer.3.k_cache 0.03101032 16.38839144 + layer.3.v_cache 0.00001666 0.04576682 + layer.4.k_cache 0.00072555 1.18401192 + layer.4.v_cache 0.00004870 0.09268429 + layer.4.output 0.18364459 732.89581322 + ------------------------------------------------------------------------------------- + TOTAL 0.08744314 324.34600926 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 22816 +BPFP 0.2834 bits/point +EBPFP 0.5668 equivalent bits/point +MSE 324.346009 +---------------------- -------------------------------------------------------- +Time: 2.487s Load: 0.004s, Pack+Encode: 1.481s, Decode+Unpack: 1.002s +---------------------- -------------------------------------------------------- +💾 Converting with 324.3460 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 111, 128) +Output shape: (1, 111, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.output: torch.Size([1, 111, 3584]) -> torch.Size([1, 1, 111, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,664B, BPFP=0.2342 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,728B, BPFP=0.3840 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,044B, BPFP=0.4285 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,516B, BPFP=0.3542 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,548B, BPFP=0.3587 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,348B, BPFP=0.3305 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,428B, BPFP=0.3418 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,644B, BPFP=0.3722 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,020B, BPFP=0.4251 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,640B, BPFP=0.3716 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,620B, BPFP=0.0326 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11107565 361.72314893 + layer.0.v_cache 0.00001597 0.03866243 + layer.1.k_cache 0.04333284 24.12896344 + layer.1.v_cache 0.00000558 0.01407045 + layer.2.k_cache 0.00477623 3.36754959 + layer.2.v_cache 0.00001758 0.04111622 + layer.3.k_cache 0.01020195 14.89237248 + layer.3.v_cache 0.00001870 0.04700912 + layer.4.k_cache 0.00079488 1.23498700 + layer.4.v_cache 0.00005041 0.09035515 + layer.4.output 10.27057757 484.04339607 + ------------------------------------------------------------------------------------- + TOTAL 4.23907840 223.16952984 + (elements=966,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 966144 +Total Bytes 27200 +BPFP 0.2252 bits/point +EBPFP 0.4505 equivalent bits/point +MSE 223.169530 +---------------------- -------------------------------------------------------- +Time: 2.493s Load: 0.005s, Pack+Encode: 1.484s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 223.1695 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,040B, BPFP=0.2289 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,028B, BPFP=0.6664 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,312B, BPFP=0.5088 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,772B, BPFP=0.3900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,436B, BPFP=0.5361 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,564B, BPFP=0.3442 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,904B, BPFP=0.4190 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,544B, BPFP=0.3398 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,636B, BPFP=0.5801 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,560B, BPFP=0.5634 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,356B, BPFP=0.0426 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.998s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09375787 323.13003411 + layer.0.v_cache 0.00001622 0.03925933 + layer.1.k_cache 0.05984993 25.83497985 + layer.1.v_cache 0.00000512 0.01465601 + layer.2.k_cache 0.00243208 3.96136647 + layer.2.v_cache 0.00001865 0.04426689 + layer.3.k_cache 0.04944374 16.57226391 + layer.3.v_cache 0.00001908 0.05479130 + layer.4.k_cache 0.00084877 1.26244376 + layer.4.v_cache 0.00005139 0.09992210 + layer.4.output 0.19371969 764.44749748 + ------------------------------------------------------------------------------------- + TOTAL 0.09191063 336.59685095 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 22152 +BPFP 0.2868 bits/point +EBPFP 0.5735 equivalent bits/point +MSE 336.596851 +---------------------- -------------------------------------------------------- +Time: 2.486s Load: 0.004s, Pack+Encode: 1.484s, Decode+Unpack: 0.998s +---------------------- -------------------------------------------------------- +💾 Converting with 336.5969 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 109, 128) +Output shape: (1, 109, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.output: torch.Size([1, 109, 3584]) -> torch.Size([1, 1, 109, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,656B, BPFP=0.2374 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,840B, BPFP=0.4071 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,144B, BPFP=0.4507 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,560B, BPFP=0.3670 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,556B, BPFP=0.3664 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,496B, BPFP=0.3578 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,416B, BPFP=0.3463 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,664B, BPFP=0.3819 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,268B, BPFP=0.4685 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,728B, BPFP=0.3911 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,680B, BPFP=0.0344 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.999s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860953 363.09758458 + layer.0.v_cache 0.00001500 0.03783925 + layer.1.k_cache 0.05917679 24.50711368 + layer.1.v_cache 0.00000528 0.01404390 + layer.2.k_cache 0.01253135 3.62293040 + layer.2.v_cache 0.00001869 0.04147998 + layer.3.k_cache 0.07406004 14.40345918 + layer.3.v_cache 0.00001854 0.04997910 + layer.4.k_cache 0.00071234 1.17454137 + layer.4.v_cache 0.00004916 0.09066905 + layer.4.output 10.45909253 492.95978047 + ------------------------------------------------------------------------------------- + TOTAL 4.32287320 226.92694728 + (elements=948,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 948736 +Total Bytes 28008 +BPFP 0.2362 bits/point +EBPFP 0.4723 equivalent bits/point +MSE 226.926947 +---------------------- -------------------------------------------------------- +Time: 2.503s Load: 0.004s, Pack+Encode: 1.500s, Decode+Unpack: 0.999s +---------------------- -------------------------------------------------------- +💾 Converting with 226.9269 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst + to output-fixed/kimiaudio/lambda0.001/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.2834 bits/point +Avg EBPFP 0.5668 equivalent bits/point +Avg MSE 338.762744 +Avg Time 2.486s +------------------------ ---------------------------- diff --git a/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..d4cbeb91ecd838d1e66c2ae6520c70cfdd7691b1 --- /dev/null +++ b/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 405 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench +Output output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 696B, BPFP=0.1343 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 792B, BPFP=0.1528 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 656B, BPFP=0.1265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 720B, BPFP=0.1389 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 668B, BPFP=0.1289 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 668B, BPFP=0.1289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 728B, BPFP=0.1404 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,372B, BPFP=0.0654 +⌛️ [2/4] FRONTEND: Frontend time: 0.383s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.240s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12089630 528.99064429 + layer.0.v_cache 0.00001394 0.04654895 + layer.1.k_cache 0.01174029 37.40573158 + layer.1.v_cache 0.00000579 0.01731843 + layer.2.k_cache 0.00835781 6.67573905 + layer.2.v_cache 0.00001852 0.05206705 + layer.3.k_cache 0.02896961 23.40296917 + layer.3.v_cache 0.00001872 0.06196806 + layer.4.k_cache 0.00063445 1.68015054 + layer.4.v_cache 0.00005063 0.13275630 + layer.4.output 0.17246709 670.35807981 + ------------------------------------------------------------------------------------- + TOTAL 0.08105739 311.23367365 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 9388 +BPFP 0.1065 bits/point +EBPFP 0.2131 equivalent bits/point +MSE 311.233674 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.005s, Pack+Encode: 0.383s, Decode+Unpack: 0.240s +---------------------- -------------------------------------------------------- +💾 Converting with 311.2337 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 628B, BPFP=0.1227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 680B, BPFP=0.1328 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1570 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 644B, BPFP=0.1258 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 712B, BPFP=0.1391 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 644B, BPFP=0.1258 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 672B, BPFP=0.1313 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 660B, BPFP=0.1289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 764B, BPFP=0.1492 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 728B, BPFP=0.1422 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,284B, BPFP=0.0637 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08029544 539.49960937 + layer.0.v_cache 0.00001376 0.04829406 + layer.1.k_cache 0.03632395 37.71667786 + layer.1.v_cache 0.00000563 0.01822627 + layer.2.k_cache 0.00522719 6.84393463 + layer.2.v_cache 0.00001997 0.05626415 + layer.3.k_cache 0.02707962 23.61040344 + layer.3.v_cache 0.00001843 0.06378412 + layer.4.k_cache 0.00063057 1.68656654 + layer.4.v_cache 0.00005085 0.13907855 + layer.4.output 0.18451144 678.53046875 + ------------------------------------------------------------------------------------- + TOTAL 0.08477915 315.25859531 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 9220 +BPFP 0.1059 bits/point +EBPFP 0.2119 equivalent bits/point +MSE 315.258595 +---------------------- -------------------------------------------------------- +Time: 0.386s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.218s +---------------------- -------------------------------------------------------- +💾 Converting with 315.2586 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 652B, BPFP=0.1185 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 772B, BPFP=0.1403 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1483 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 712B, BPFP=0.1294 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 764B, BPFP=0.1388 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 724B, BPFP=0.1315 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 716B, BPFP=0.1301 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 724B, BPFP=0.1315 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 808B, BPFP=0.1468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 816B, BPFP=0.1483 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,480B, BPFP=0.0644 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10874708 528.30627725 + layer.0.v_cache 0.00001608 0.04955172 + layer.1.k_cache 0.03270756 37.38690895 + layer.1.v_cache 0.00000572 0.01814928 + layer.2.k_cache 0.00777350 6.75532212 + layer.2.v_cache 0.00002080 0.05409673 + layer.3.k_cache 0.05920834 23.43540493 + layer.3.v_cache 0.00001846 0.06119248 + layer.4.k_cache 0.00062493 1.65973628 + layer.4.v_cache 0.00005036 0.12823908 + layer.4.output 0.17137358 637.75643688 + ------------------------------------------------------------------------------------- + TOTAL 0.08286988 297.77352570 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 9984 +BPFP 0.1067 bits/point +EBPFP 0.2134 equivalent bits/point +MSE 297.773526 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.005s, Pack+Encode: 0.168s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 297.7735 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 696B, BPFP=0.1394 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1635 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 644B, BPFP=0.1290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1434 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 656B, BPFP=0.1314 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1378 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 696B, BPFP=0.1394 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 776B, BPFP=0.1554 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 732B, BPFP=0.1466 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,648B, BPFP=0.0758 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12073612 536.71033654 + layer.0.v_cache 0.00001348 0.04743523 + layer.1.k_cache 0.03279715 38.10130271 + layer.1.v_cache 0.00000568 0.01821075 + layer.2.k_cache 0.00696802 6.89926539 + layer.2.v_cache 0.00001833 0.05368573 + layer.3.k_cache 0.03227850 23.64010229 + layer.3.v_cache 0.00001834 0.06220592 + layer.4.k_cache 0.00060115 1.72342936 + layer.4.v_cache 0.00005191 0.13560376 + layer.4.output 0.18320683 696.69614240 + ------------------------------------------------------------------------------------- + TOTAL 0.08681979 322.60379850 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9708 +BPFP 0.1144 bits/point +EBPFP 0.2288 equivalent bits/point +MSE 322.603798 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.003s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 322.6038 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1274 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 708B, BPFP=0.1418 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1619 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 672B, BPFP=0.1346 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 704B, BPFP=0.1410 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 660B, BPFP=0.1322 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 660B, BPFP=0.1322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1579 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 736B, BPFP=0.1474 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,628B, BPFP=0.0752 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.234s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10595027 536.68594752 + layer.0.v_cache 0.00001376 0.04782567 + layer.1.k_cache 0.03307601 38.21824920 + layer.1.v_cache 0.00000550 0.01767405 + layer.2.k_cache 0.00695119 6.91251471 + layer.2.v_cache 0.00001929 0.05344933 + layer.3.k_cache 0.03450640 23.61895908 + layer.3.v_cache 0.00001844 0.06340456 + layer.4.k_cache 0.00061739 1.67565155 + layer.4.v_cache 0.00005297 0.13322366 + layer.4.output 0.18589628 696.41277473 + ------------------------------------------------------------------------------------- + TOTAL 0.08720501 322.48919544 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9684 +BPFP 0.1141 bits/point +EBPFP 0.2282 equivalent bits/point +MSE 322.489195 +---------------------- -------------------------------------------------------- +Time: 0.392s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.234s +---------------------- -------------------------------------------------------- +💾 Converting with 322.4892 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 620B, BPFP=0.1226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 676B, BPFP=0.1337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 792B, BPFP=0.1566 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 644B, BPFP=0.1274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1416 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 636B, BPFP=0.1258 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 672B, BPFP=0.1329 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 656B, BPFP=0.1297 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 764B, BPFP=0.1511 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 724B, BPFP=0.1432 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,532B, BPFP=0.0715 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07772056 541.69284019 + layer.0.v_cache 0.00001351 0.04810716 + layer.1.k_cache 0.03518496 38.05576914 + layer.1.v_cache 0.00000556 0.01810548 + layer.2.k_cache 0.00230807 6.93455177 + layer.2.v_cache 0.00001833 0.05319134 + layer.3.k_cache 0.04514034 23.67247608 + layer.3.v_cache 0.00001901 0.06264332 + layer.4.k_cache 0.00062178 1.74756458 + layer.4.v_cache 0.00004990 0.13818094 + layer.4.output 0.18426369 687.63884494 + ------------------------------------------------------------------------------------- + TOTAL 0.08534870 319.17031438 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 9432 +BPFP 0.1097 bits/point +EBPFP 0.2195 equivalent bits/point +MSE 319.170314 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 319.1703 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1299 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 712B, BPFP=0.1445 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1648 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 640B, BPFP=0.1299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 708B, BPFP=0.1437 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 664B, BPFP=0.1347 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1380 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 684B, BPFP=0.1388 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 760B, BPFP=0.1542 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 728B, BPFP=0.1477 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,700B, BPFP=0.0783 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14568528 537.62784091 + layer.0.v_cache 0.00001322 0.04743470 + layer.1.k_cache 0.01379078 38.47718078 + layer.1.v_cache 0.00000548 0.01786173 + layer.2.k_cache 0.00727733 6.93292316 + layer.2.v_cache 0.00001786 0.05561982 + layer.3.k_cache 0.03082054 23.64744191 + layer.3.v_cache 0.00001879 0.06598769 + layer.4.k_cache 0.00061820 1.70169503 + layer.4.v_cache 0.00005377 0.14047568 + layer.4.output 0.19662610 705.93344156 + ------------------------------------------------------------------------------------- + TOTAL 0.09262847 326.48520896 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9728 +BPFP 0.1161 bits/point +EBPFP 0.2322 equivalent bits/point +MSE 326.485209 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 326.4852 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 628B, BPFP=0.1227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1352 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1562 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 648B, BPFP=0.1266 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 656B, BPFP=0.1281 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1328 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 656B, BPFP=0.1281 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1508 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 740B, BPFP=0.1445 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,312B, BPFP=0.0645 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10149734 538.66005859 + layer.0.v_cache 0.00001421 0.04908478 + layer.1.k_cache 0.01324784 37.84090881 + layer.1.v_cache 0.00000586 0.01861755 + layer.2.k_cache 0.00706157 6.77251740 + layer.2.v_cache 0.00001822 0.05316283 + layer.3.k_cache 0.02712160 23.50325012 + layer.3.v_cache 0.00002016 0.06397625 + layer.4.k_cache 0.00063687 1.70543575 + layer.4.v_cache 0.00005451 0.13878714 + layer.4.output 0.17858309 678.42460938 + ------------------------------------------------------------------------------------- + TOTAL 0.08233881 315.16341558 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 9300 +BPFP 0.1068 bits/point +EBPFP 0.2137 equivalent bits/point +MSE 315.163416 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 315.1634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 664B, BPFP=0.1193 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1365 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 704B, BPFP=0.1264 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 772B, BPFP=0.1386 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 744B, BPFP=0.1336 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 716B, BPFP=0.1286 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 724B, BPFP=0.1300 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 832B, BPFP=0.1494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 804B, BPFP=0.1444 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,412B, BPFP=0.0619 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10772817 527.05801006 + layer.0.v_cache 0.00001348 0.04764130 + layer.1.k_cache 0.03276526 37.29000539 + layer.1.v_cache 0.00000562 0.01804965 + layer.2.k_cache 0.01020141 6.78698450 + layer.2.v_cache 0.00001934 0.05374062 + layer.3.k_cache 0.04694761 23.39052678 + layer.3.v_cache 0.00001910 0.06592355 + layer.4.k_cache 0.00062564 1.67452372 + layer.4.v_cache 0.00005323 0.13803285 + layer.4.output 0.17166949 628.65137521 + ------------------------------------------------------------------------------------- + TOTAL 0.08235678 293.94606264 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 9948 +BPFP 0.1051 bits/point +EBPFP 0.2102 equivalent bits/point +MSE 293.946063 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 293.9461 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 688B, BPFP=0.1295 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1506 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 656B, BPFP=0.1235 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 720B, BPFP=0.1355 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 664B, BPFP=0.1250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 644B, BPFP=0.1212 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 768B, BPFP=0.1446 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,160B, BPFP=0.0581 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12335577 531.81108810 + layer.0.v_cache 0.00001400 0.04832661 + layer.1.k_cache 0.01519604 37.29605139 + layer.1.v_cache 0.00000547 0.01743494 + layer.2.k_cache 0.01098718 6.72561094 + layer.2.v_cache 0.00001779 0.05248801 + layer.3.k_cache 0.07416311 23.39515337 + layer.3.v_cache 0.00001944 0.06293142 + layer.4.k_cache 0.00061719 1.67841826 + layer.4.v_cache 0.00006444 0.13069596 + layer.4.output 0.17504946 654.90506670 + ------------------------------------------------------------------------------------- + TOTAL 0.08528157 305.03256858 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9204 +BPFP 0.1019 bits/point +EBPFP 0.2038 equivalent bits/point +MSE 305.032569 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 305.0326 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1404 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1640 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 640B, BPFP=0.1299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 704B, BPFP=0.1429 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 668B, BPFP=0.1356 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1372 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 664B, BPFP=0.1347 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 776B, BPFP=0.1575 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 744B, BPFP=0.1510 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,640B, BPFP=0.0765 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13955514 539.16061282 + layer.0.v_cache 0.00001392 0.04727672 + layer.1.k_cache 0.03454666 38.28050490 + layer.1.v_cache 0.00000526 0.01781995 + layer.2.k_cache 0.00236268 6.94303755 + layer.2.v_cache 0.00001660 0.05117874 + layer.3.k_cache 0.02874016 23.67033248 + layer.3.v_cache 0.00001899 0.06215525 + layer.4.k_cache 0.00062191 1.72526015 + layer.4.v_cache 0.00005265 0.13665404 + layer.4.output 0.18728454 705.88189935 + ------------------------------------------------------------------------------------- + TOTAL 0.08923092 326.54518400 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9648 +BPFP 0.1152 bits/point +EBPFP 0.2303 equivalent bits/point +MSE 326.545184 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 326.5452 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 708B, BPFP=0.1437 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1640 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 652B, BPFP=0.1323 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 704B, BPFP=0.1429 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 668B, BPFP=0.1356 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 668B, BPFP=0.1356 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1567 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 760B, BPFP=0.1542 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,652B, BPFP=0.0769 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11454884 536.20941558 + layer.0.v_cache 0.00001351 0.05033792 + layer.1.k_cache 0.01665594 38.40528612 + layer.1.v_cache 0.00000570 0.01943167 + layer.2.k_cache 0.00836552 6.91334514 + layer.2.v_cache 0.00001974 0.05632294 + layer.3.k_cache 0.03124422 23.64508929 + layer.3.v_cache 0.00001876 0.06399466 + layer.4.k_cache 0.00062231 1.73455097 + layer.4.v_cache 0.00005487 0.13888800 + layer.4.output 0.19663499 704.98156308 + ------------------------------------------------------------------------------------- + TOTAL 0.09105849 326.00632964 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9712 +BPFP 0.1159 bits/point +EBPFP 0.2319 equivalent bits/point +MSE 326.006330 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 326.0063 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 652B, BPFP=0.1213 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 700B, BPFP=0.1302 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1525 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1280 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 744B, BPFP=0.1384 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 676B, BPFP=0.1257 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1287 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 688B, BPFP=0.1280 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 804B, BPFP=0.1496 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 780B, BPFP=0.1451 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,176B, BPFP=0.0578 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12290487 524.80431548 + layer.0.v_cache 0.00001422 0.04815104 + layer.1.k_cache 0.01283375 37.31614467 + layer.1.v_cache 0.00000579 0.01855606 + layer.2.k_cache 0.01100613 6.69285656 + layer.2.v_cache 0.00001929 0.05249591 + layer.3.k_cache 0.03011658 23.36422439 + layer.3.v_cache 0.00001858 0.06139903 + layer.4.k_cache 0.00063020 1.65184820 + layer.4.v_cache 0.00005223 0.13245461 + layer.4.output 0.17536174 647.90284864 + ------------------------------------------------------------------------------------- + TOTAL 0.08265493 301.73308155 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 9420 +BPFP 0.1031 bits/point +EBPFP 0.2061 equivalent bits/point +MSE 301.733082 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 301.7331 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1206 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1397 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1485 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 696B, BPFP=0.1279 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1353 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 700B, BPFP=0.1287 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 700B, BPFP=0.1287 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 688B, BPFP=0.1265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 800B, BPFP=0.1471 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 784B, BPFP=0.1441 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,252B, BPFP=0.0591 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11218034 523.07123162 + layer.0.v_cache 0.00001438 0.04826465 + layer.1.k_cache 0.03157643 37.21565947 + layer.1.v_cache 0.00000545 0.01785449 + layer.2.k_cache 0.00802605 6.65629955 + layer.2.v_cache 0.00001826 0.05303902 + layer.3.k_cache 0.04519741 23.57493394 + layer.3.v_cache 0.00001794 0.05973400 + layer.4.k_cache 0.00061137 1.63905783 + layer.4.v_cache 0.00005913 0.13391661 + layer.4.output 0.17010492 641.51864496 + ------------------------------------------------------------------------------------- + TOTAL 0.08167301 299.00591211 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 9580 +BPFP 0.1036 bits/point +EBPFP 0.2072 equivalent bits/point +MSE 299.005912 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 299.0059 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 660B, BPFP=0.1199 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1381 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1475 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 764B, BPFP=0.1388 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 712B, BPFP=0.1294 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 720B, BPFP=0.1308 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 720B, BPFP=0.1308 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 820B, BPFP=0.1490 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 796B, BPFP=0.1446 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,396B, BPFP=0.0622 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.216s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10146660 526.01898619 + layer.0.v_cache 0.00001344 0.04782826 + layer.1.k_cache 0.03090366 37.50807935 + layer.1.v_cache 0.00000566 0.01780072 + layer.2.k_cache 0.01188999 6.77156954 + layer.2.v_cache 0.00001845 0.05291960 + layer.3.k_cache 0.02542871 23.54198651 + layer.3.v_cache 0.00001884 0.06332269 + layer.4.k_cache 0.00060243 1.66540510 + layer.4.v_cache 0.00005224 0.13263055 + layer.4.output 0.16839122 637.39856728 + ------------------------------------------------------------------------------------- + TOTAL 0.07936109 297.50649997 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 9848 +BPFP 0.1052 bits/point +EBPFP 0.2105 equivalent bits/point +MSE 297.506500 +---------------------- -------------------------------------------------------- +Time: 0.385s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.216s +---------------------- -------------------------------------------------------- +💾 Converting with 297.5065 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 676B, BPFP=0.1148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 772B, BPFP=0.1311 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 844B, BPFP=0.1433 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 724B, BPFP=0.1230 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 792B, BPFP=0.1345 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 716B, BPFP=0.1216 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 836B, BPFP=0.1420 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 808B, BPFP=0.1372 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,032B, BPFP=0.0493 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12981709 520.23743207 + layer.0.v_cache 0.00001628 0.04857932 + layer.1.k_cache 0.06761242 36.80238674 + layer.1.v_cache 0.00000547 0.01765096 + layer.2.k_cache 0.00384689 6.53157309 + layer.2.v_cache 0.00001812 0.05336074 + layer.3.k_cache 0.05327560 23.20420771 + layer.3.v_cache 0.00001896 0.06348383 + layer.4.k_cache 0.00060800 1.61170495 + layer.4.v_cache 0.00005116 0.13368248 + layer.4.output 0.15518735 590.50048525 + ------------------------------------------------------------------------------------- + TOTAL 0.07891655 277.77690933 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 9704 +BPFP 0.0969 bits/point +EBPFP 0.1939 equivalent bits/point +MSE 277.776909 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 277.7769 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 628B, BPFP=0.1211 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 688B, BPFP=0.1327 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 792B, BPFP=0.1528 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 644B, BPFP=0.1242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1381 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 636B, BPFP=0.1227 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 672B, BPFP=0.1296 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 656B, BPFP=0.1265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 736B, BPFP=0.1420 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,328B, BPFP=0.0642 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887507 535.80555556 + layer.0.v_cache 0.00001345 0.04787195 + layer.1.k_cache 0.03259810 37.38654393 + layer.1.v_cache 0.00000523 0.01740432 + layer.2.k_cache 0.00244350 6.77712221 + layer.2.v_cache 0.00001834 0.05262101 + layer.3.k_cache 0.10864470 23.31863968 + layer.3.v_cache 0.00001904 0.06116599 + layer.4.k_cache 0.00060864 1.64824234 + layer.4.v_cache 0.00004940 0.13348704 + layer.4.output 0.18628448 670.26251102 + ------------------------------------------------------------------------------------- + TOTAL 0.09395687 311.59330772 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 9268 +BPFP 0.1052 bits/point +EBPFP 0.2103 equivalent bits/point +MSE 311.593308 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 311.5933 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 672B, BPFP=0.1180 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1334 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 828B, BPFP=0.1454 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 700B, BPFP=0.1229 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 776B, BPFP=0.1362 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 752B, BPFP=0.1320 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 740B, BPFP=0.1299 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 752B, BPFP=0.1320 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 828B, BPFP=0.1454 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 808B, BPFP=0.1419 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,272B, BPFP=0.0570 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422350 524.02554424 + layer.0.v_cache 0.00001493 0.04883509 + layer.1.k_cache 0.01422887 36.86887728 + layer.1.v_cache 0.00000595 0.01820862 + layer.2.k_cache 0.00668612 6.65239587 + layer.2.v_cache 0.00001902 0.05501289 + layer.3.k_cache 0.02497406 23.44784717 + layer.3.v_cache 0.00001869 0.06476153 + layer.4.k_cache 0.00061935 1.62895906 + layer.4.v_cache 0.00005212 0.13494447 + layer.4.output 0.16899524 613.16056380 + ------------------------------------------------------------------------------------- + TOTAL 0.07904761 287.35701958 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 9888 +BPFP 0.1021 bits/point +EBPFP 0.2042 equivalent bits/point +MSE 287.357020 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 287.3570 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 113, 128) +Output shape: (1, 113, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.output: torch.Size([1, 113, 3584]) -> torch.Size([1, 1, 113, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 748B, BPFP=0.1034 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 796B, BPFP=0.1101 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 860B, BPFP=0.1189 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 788B, BPFP=0.1090 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 832B, BPFP=0.1150 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 780B, BPFP=0.1079 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 816B, BPFP=0.1128 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 768B, BPFP=0.1062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 876B, BPFP=0.1211 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 820B, BPFP=0.1134 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,124B, BPFP=0.0420 +⌛️ [2/4] FRONTEND: Frontend time: 0.271s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.266s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10214825 524.23879978 + layer.0.v_cache 0.00001637 0.04761118 + layer.1.k_cache 0.04284562 36.76696453 + layer.1.v_cache 0.00000578 0.01802237 + layer.2.k_cache 0.00913706 6.61528029 + layer.2.v_cache 0.00001928 0.05311713 + layer.3.k_cache 0.05585717 23.13725456 + layer.3.v_cache 0.00001870 0.05996660 + layer.4.k_cache 0.00065968 1.67923892 + layer.4.v_cache 0.00005753 0.13313179 + layer.4.output 10.09706179 477.03717604 + ------------------------------------------------------------------------------------- + TOTAL 4.17001165 231.29468350 + (elements=983,552) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 983552 +Total Bytes 10208 +BPFP 0.0830 bits/point +EBPFP 0.1661 equivalent bits/point +MSE 231.294683 +---------------------- -------------------------------------------------------- +Time: 0.543s Load: 0.006s, Pack+Encode: 0.271s, Decode+Unpack: 0.266s +---------------------- -------------------------------------------------------- +💾 Converting with 231.2947 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 680B, BPFP=0.1168 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1305 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 836B, BPFP=0.1435 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 716B, BPFP=0.1229 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 792B, BPFP=0.1360 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 736B, BPFP=0.1264 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 740B, BPFP=0.1271 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 724B, BPFP=0.1243 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 840B, BPFP=0.1442 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 792B, BPFP=0.1360 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,100B, BPFP=0.0515 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11481400 515.10448146 + layer.0.v_cache 0.00001396 0.04867229 + layer.1.k_cache 0.03135788 36.65877189 + layer.1.v_cache 0.00000545 0.01788767 + layer.2.k_cache 0.00370715 6.63793811 + layer.2.v_cache 0.00001965 0.05470240 + layer.3.k_cache 0.03930106 23.25015024 + layer.3.v_cache 0.00001886 0.06466549 + layer.4.k_cache 0.00060819 1.64546338 + layer.4.v_cache 0.00005437 0.13830712 + layer.4.output 0.16591480 598.55347331 + ------------------------------------------------------------------------------------- + TOTAL 0.07948848 280.79384431 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 9716 +BPFP 0.0981 bits/point +EBPFP 0.1963 equivalent bits/point +MSE 280.793844 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.005s, Pack+Encode: 0.166s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 280.7938 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1173 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1334 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 828B, BPFP=0.1454 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 704B, BPFP=0.1236 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 780B, BPFP=0.1369 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 732B, BPFP=0.1285 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 732B, BPFP=0.1285 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 720B, BPFP=0.1264 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 840B, BPFP=0.1475 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 796B, BPFP=0.1397 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,216B, BPFP=0.0556 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09788724 525.49482093 + layer.0.v_cache 0.00001308 0.04684345 + layer.1.k_cache 0.03299916 36.74086530 + layer.1.v_cache 0.00000544 0.01758307 + layer.2.k_cache 0.00367894 6.63511538 + layer.2.v_cache 0.00002012 0.05330989 + layer.3.k_cache 0.03939044 23.23346427 + layer.3.v_cache 0.00001875 0.06091722 + layer.4.k_cache 0.00061788 1.63780778 + layer.4.v_cache 0.00005010 0.13537205 + layer.4.output 0.17021053 613.30818620 + ------------------------------------------------------------------------------------- + TOTAL 0.08036205 287.48314134 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 9776 +BPFP 0.1010 bits/point +EBPFP 0.2019 equivalent bits/point +MSE 287.483141 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 287.4831 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 764B, BPFP=0.1105 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 812B, BPFP=0.1175 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 868B, BPFP=0.1256 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 796B, BPFP=0.1152 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 832B, BPFP=0.1204 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 780B, BPFP=0.1128 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 828B, BPFP=0.1198 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 780B, BPFP=0.1128 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 892B, BPFP=0.1291 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 832B, BPFP=0.1204 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,408B, BPFP=0.0498 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12672405 524.90350116 + layer.0.v_cache 0.00001401 0.04847082 + layer.1.k_cache 0.12044146 37.27694137 + layer.1.v_cache 0.00000568 0.01812521 + layer.2.k_cache 0.00570256 6.70171328 + layer.2.v_cache 0.00001866 0.05467576 + layer.3.k_cache 0.03611955 23.03782371 + layer.3.v_cache 0.00001983 0.06807083 + layer.4.k_cache 0.00066233 1.67362496 + layer.4.v_cache 0.00005331 0.14045831 + layer.4.output 10.56037249 498.41546792 + ------------------------------------------------------------------------------------- + TOTAL 4.36543346 240.16656946 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 10592 +BPFP 0.0901 bits/point +EBPFP 0.1803 equivalent bits/point +MSE 240.166569 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.006s, Pack+Encode: 0.156s, Decode+Unpack: 0.222s +---------------------- -------------------------------------------------------- +💾 Converting with 240.1666 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1186 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 772B, BPFP=0.1371 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1449 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 704B, BPFP=0.1250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 756B, BPFP=0.1342 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 744B, BPFP=0.1321 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 732B, BPFP=0.1300 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 820B, BPFP=0.1456 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 812B, BPFP=0.1442 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,428B, BPFP=0.0616 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13733016 527.65465199 + layer.0.v_cache 0.00001373 0.04905276 + layer.1.k_cache 0.03155638 37.10081343 + layer.1.v_cache 0.00000617 0.01873959 + layer.2.k_cache 0.00751131 6.70668446 + layer.2.v_cache 0.00001754 0.05265812 + layer.3.k_cache 0.02548880 23.39083030 + layer.3.v_cache 0.00001765 0.06005669 + layer.4.k_cache 0.00061011 1.60344973 + layer.4.v_cache 0.00005045 0.13184311 + layer.4.output 0.16239834 619.51324067 + ------------------------------------------------------------------------------------- + TOTAL 0.07878769 290.19773323 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 10020 +BPFP 0.1047 bits/point +EBPFP 0.2093 equivalent bits/point +MSE 290.197733 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1977 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 676B, BPFP=0.1337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 792B, BPFP=0.1566 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 648B, BPFP=0.1282 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 704B, BPFP=0.1392 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 648B, BPFP=0.1282 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 664B, BPFP=0.1313 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 664B, BPFP=0.1313 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 780B, BPFP=0.1543 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 740B, BPFP=0.1464 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,548B, BPFP=0.0720 +⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10979201 536.71281646 + layer.0.v_cache 0.00001339 0.04772852 + layer.1.k_cache 0.01160004 38.09728540 + layer.1.v_cache 0.00000567 0.01802576 + layer.2.k_cache 0.00714037 6.82009926 + layer.2.v_cache 0.00001829 0.05376530 + layer.3.k_cache 0.04622446 23.73294570 + layer.3.v_cache 0.00001896 0.06359128 + layer.4.k_cache 0.00064378 1.69506063 + layer.4.v_cache 0.00005251 0.13829612 + layer.4.output 0.18415536 687.29447333 + ------------------------------------------------------------------------------------- + TOTAL 0.08615276 318.73181928 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 9496 +BPFP 0.1105 bits/point +EBPFP 0.2210 equivalent bits/point +MSE 318.731819 +---------------------- -------------------------------------------------------- +Time: 0.406s Load: 0.004s, Pack+Encode: 0.199s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 318.7318 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 676B, BPFP=0.1372 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1631 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 652B, BPFP=0.1323 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 712B, BPFP=0.1445 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 644B, BPFP=0.1307 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 644B, BPFP=0.1307 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1591 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 748B, BPFP=0.1518 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,648B, BPFP=0.0768 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11290005 539.03809862 + layer.0.v_cache 0.00001570 0.04867531 + layer.1.k_cache 0.03434501 38.46739930 + layer.1.v_cache 0.00000534 0.01755915 + layer.2.k_cache 0.00238675 6.94100041 + layer.2.v_cache 0.00001838 0.05333289 + layer.3.k_cache 0.02838124 23.58846768 + layer.3.v_cache 0.00002047 0.06421329 + layer.4.k_cache 0.00062757 1.73716557 + layer.4.v_cache 0.00005455 0.14000546 + layer.4.output 0.18420308 704.93941327 + ------------------------------------------------------------------------------------- + TOTAL 0.08636333 326.15716533 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9632 +BPFP 0.1150 bits/point +EBPFP 0.2299 equivalent bits/point +MSE 326.157165 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 326.1572 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1266 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 680B, BPFP=0.1362 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1611 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 644B, BPFP=0.1290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 708B, BPFP=0.1418 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 648B, BPFP=0.1298 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 664B, BPFP=0.1330 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 780B, BPFP=0.1562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 740B, BPFP=0.1482 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,624B, BPFP=0.0751 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09752406 538.51061699 + layer.0.v_cache 0.00001393 0.04778118 + layer.1.k_cache 0.01493649 38.14490059 + layer.1.v_cache 0.00000588 0.01805089 + layer.2.k_cache 0.01003601 6.92576247 + layer.2.v_cache 0.00001805 0.05347248 + layer.3.k_cache 0.04451986 23.73654723 + layer.3.v_cache 0.00001914 0.06443777 + layer.4.k_cache 0.00061869 1.71662296 + layer.4.v_cache 0.00005251 0.13956390 + layer.4.output 0.18409089 695.76734203 + ------------------------------------------------------------------------------------- + TOTAL 0.08566946 322.33700886 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9608 +BPFP 0.1132 bits/point +EBPFP 0.2264 equivalent bits/point +MSE 322.337009 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 322.3370 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1197 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 700B, BPFP=0.1318 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 796B, BPFP=0.1498 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 652B, BPFP=0.1227 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1378 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 660B, BPFP=0.1242 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1310 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 680B, BPFP=0.1280 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1423 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,172B, BPFP=0.0584 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13658243 529.48366905 + layer.0.v_cache 0.00001379 0.04841363 + layer.1.k_cache 0.01234693 37.28862128 + layer.1.v_cache 0.00000566 0.01775942 + layer.2.k_cache 0.00375735 6.65187606 + layer.2.v_cache 0.00001852 0.05141921 + layer.3.k_cache 0.02700986 23.28616517 + layer.3.v_cache 0.00001930 0.06195203 + layer.4.k_cache 0.00062007 1.69494758 + layer.4.v_cache 0.00005146 0.13115189 + layer.4.output 0.17614663 654.66490964 + ------------------------------------------------------------------------------------- + TOTAL 0.08314423 304.78649075 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9264 +BPFP 0.1026 bits/point +EBPFP 0.2052 equivalent bits/point +MSE 304.786491 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 304.7865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 648B, BPFP=0.1387 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 700B, BPFP=0.1498 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 780B, BPFP=0.1670 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 656B, BPFP=0.1404 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1567 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 656B, BPFP=0.1404 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1447 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 648B, BPFP=0.1387 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 768B, BPFP=0.1644 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 752B, BPFP=0.1610 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,968B, BPFP=0.0908 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08738674 523.14372860 + layer.0.v_cache 0.00001368 0.04759829 + layer.1.k_cache 0.01286346 37.81722228 + layer.1.v_cache 0.00000597 0.01777030 + layer.2.k_cache 0.00251921 6.89722934 + layer.2.v_cache 0.00001769 0.05240641 + layer.3.k_cache 0.08324167 23.52443747 + layer.3.v_cache 0.00002021 0.06283307 + layer.4.k_cache 0.00062021 1.67588367 + layer.4.v_cache 0.00005375 0.14025055 + layer.4.output 0.19567458 743.90918542 + ------------------------------------------------------------------------------------- + TOTAL 0.09155674 341.22021517 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 9984 +BPFP 0.1257 bits/point +EBPFP 0.2514 equivalent bits/point +MSE 341.220215 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 341.2202 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1205 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 708B, BPFP=0.1333 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1529 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 672B, BPFP=0.1265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1378 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 676B, BPFP=0.1273 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1303 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 688B, BPFP=0.1295 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 792B, BPFP=0.1491 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 764B, BPFP=0.1438 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,204B, BPFP=0.0593 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12176956 528.75258848 + layer.0.v_cache 0.00001396 0.04753517 + layer.1.k_cache 0.01120206 37.25873612 + layer.1.v_cache 0.00000562 0.01782708 + layer.2.k_cache 0.00678912 6.60573201 + layer.2.v_cache 0.00001849 0.05261025 + layer.3.k_cache 0.02622183 23.24083002 + layer.3.v_cache 0.00001965 0.06311687 + layer.4.k_cache 0.00062413 1.68806770 + layer.4.v_cache 0.00005248 0.13320677 + layer.4.output 0.17940697 654.85445353 + ------------------------------------------------------------------------------------- + TOTAL 0.08368034 304.81420148 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9380 +BPFP 0.1039 bits/point +EBPFP 0.2077 equivalent bits/point +MSE 304.814201 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 304.8142 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 724B, BPFP=0.1077 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 820B, BPFP=0.1220 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 864B, BPFP=0.1286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 776B, BPFP=0.1155 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 844B, BPFP=0.1256 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 812B, BPFP=0.1208 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 816B, BPFP=0.1214 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 796B, BPFP=0.1185 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 884B, BPFP=0.1315 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 828B, BPFP=0.1232 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,428B, BPFP=0.0516 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120816 522.41529018 + layer.0.v_cache 0.00001607 0.04871363 + layer.1.k_cache 0.10678518 36.82490699 + layer.1.v_cache 0.00000585 0.01777321 + layer.2.k_cache 0.00813519 6.66069975 + layer.2.v_cache 0.00001919 0.05404857 + layer.3.k_cache 0.06560528 22.94239676 + layer.3.v_cache 0.00001973 0.06747199 + layer.4.k_cache 0.00062300 1.64889294 + layer.4.v_cache 0.00005287 0.14037314 + layer.4.output 10.86395687 512.38082483 + ------------------------------------------------------------------------------------- + TOTAL 4.49059815 245.73449065 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 10592 +BPFP 0.0927 bits/point +EBPFP 0.1854 equivalent bits/point +MSE 245.734491 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 245.7345 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 672B, BPFP=0.1207 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1379 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 704B, BPFP=0.1264 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 772B, BPFP=0.1386 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 732B, BPFP=0.1315 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 720B, BPFP=0.1293 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 720B, BPFP=0.1293 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 832B, BPFP=0.1494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 828B, BPFP=0.1487 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,496B, BPFP=0.0640 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12848519 528.96794181 + layer.0.v_cache 0.00001371 0.04660147 + layer.1.k_cache 0.03007136 37.34749910 + layer.1.v_cache 0.00000570 0.01788218 + layer.2.k_cache 0.00936044 6.71475816 + layer.2.v_cache 0.00001842 0.05295244 + layer.3.k_cache 0.04065019 23.42808515 + layer.3.v_cache 0.00001877 0.06439643 + layer.4.k_cache 0.00062204 1.65508069 + layer.4.v_cache 0.00010837 0.13500517 + layer.4.output 0.16426703 628.66989943 + ------------------------------------------------------------------------------------- + TOTAL 0.07995432 294.06585286 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 10060 +BPFP 0.1063 bits/point +EBPFP 0.2126 equivalent bits/point +MSE 294.065853 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 294.0659 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 624B, BPFP=0.1219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1352 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1570 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 652B, BPFP=0.1273 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 640B, BPFP=0.1250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1328 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 648B, BPFP=0.1266 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1531 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 740B, BPFP=0.1445 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,300B, BPFP=0.0642 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09257430 534.50922852 + layer.0.v_cache 0.00001329 0.04716557 + layer.1.k_cache 0.01202901 37.72550659 + layer.1.v_cache 0.00000544 0.01740756 + layer.2.k_cache 0.00549439 6.81236115 + layer.2.v_cache 0.00001849 0.05395011 + layer.3.k_cache 0.11014032 23.50902863 + layer.3.v_cache 0.00001916 0.06164876 + layer.4.k_cache 0.00063444 1.68829136 + layer.4.v_cache 0.00005494 0.13530850 + layer.4.output 0.19143520 678.26668527 + ------------------------------------------------------------------------------------- + TOTAL 0.09182531 314.84862904 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 9280 +BPFP 0.1066 bits/point +EBPFP 0.2132 equivalent bits/point +MSE 314.848629 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 314.8486 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 644B, BPFP=0.1198 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 708B, BPFP=0.1317 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 824B, BPFP=0.1533 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 672B, BPFP=0.1250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 744B, BPFP=0.1384 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 668B, BPFP=0.1243 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1287 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 668B, BPFP=0.1243 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 804B, BPFP=0.1496 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 776B, BPFP=0.1443 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,192B, BPFP=0.0582 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12365316 526.87425595 + layer.0.v_cache 0.00001332 0.04771436 + layer.1.k_cache 0.03511774 37.52179827 + layer.1.v_cache 0.00000569 0.01825115 + layer.2.k_cache 0.01236608 6.68774414 + layer.2.v_cache 0.00001835 0.05163446 + layer.3.k_cache 0.02851222 23.32480876 + layer.3.v_cache 0.00001844 0.06018300 + layer.4.k_cache 0.00063207 1.63174547 + layer.4.v_cache 0.00005066 0.13120488 + layer.4.output 0.17964420 647.91198980 + ------------------------------------------------------------------------------------- + TOTAL 0.08575865 301.86666288 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 9392 +BPFP 0.1028 bits/point +EBPFP 0.2055 equivalent bits/point +MSE 301.866663 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 301.8667 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1266 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 696B, BPFP=0.1394 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1627 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 644B, BPFP=0.1290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1434 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 656B, BPFP=0.1314 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1378 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 660B, BPFP=0.1322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 776B, BPFP=0.1554 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 736B, BPFP=0.1474 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,596B, BPFP=0.0743 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12096892 537.60792268 + layer.0.v_cache 0.00001393 0.04835918 + layer.1.k_cache 0.03247837 38.18309921 + layer.1.v_cache 0.00000613 0.01793652 + layer.2.k_cache 0.00394396 6.89773716 + layer.2.v_cache 0.00001920 0.05521800 + layer.3.k_cache 0.02751121 23.78190417 + layer.3.v_cache 0.00001919 0.06583417 + layer.4.k_cache 0.00062481 1.69761560 + layer.4.v_cache 0.00005293 0.13779154 + layer.4.output 0.18564256 696.51488095 + ------------------------------------------------------------------------------------- + TOTAL 0.08736098 322.59397558 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9612 +BPFP 0.1133 bits/point +EBPFP 0.2265 equivalent bits/point +MSE 322.593976 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 322.5940 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1404 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 788B, BPFP=0.1599 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 632B, BPFP=0.1282 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1453 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 656B, BPFP=0.1331 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1372 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 652B, BPFP=0.1323 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 768B, BPFP=0.1558 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 740B, BPFP=0.1502 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,696B, BPFP=0.0782 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08564486 542.48554180 + layer.0.v_cache 0.00001408 0.04982537 + layer.1.k_cache 0.03507188 38.32434241 + layer.1.v_cache 0.00000566 0.01832416 + layer.2.k_cache 0.00239554 6.88767153 + layer.2.v_cache 0.00001938 0.05355727 + layer.3.k_cache 0.03313774 23.31515860 + layer.3.v_cache 0.00001993 0.06507926 + layer.4.k_cache 0.00061714 1.67309511 + layer.4.v_cache 0.00005134 0.13739073 + layer.4.output 0.18226704 705.23167904 + ------------------------------------------------------------------------------------- + TOTAL 0.08428511 326.44892585 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9648 +BPFP 0.1152 bits/point +EBPFP 0.2303 equivalent bits/point +MSE 326.448926 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 326.4489 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 117, 128) +Output shape: (1, 117, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.output: torch.Size([1, 117, 3584]) -> torch.Size([1, 1, 117, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 756B, BPFP=0.1010 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 780B, BPFP=0.1042 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 884B, BPFP=0.1181 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 764B, BPFP=0.1020 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 840B, BPFP=0.1122 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 796B, BPFP=0.1063 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 828B, BPFP=0.1106 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 772B, BPFP=0.1031 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 884B, BPFP=0.1181 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 856B, BPFP=0.1143 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,112B, BPFP=0.0403 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11900647 523.30722489 + layer.0.v_cache 0.00001428 0.05056057 + layer.1.k_cache 0.12535267 36.77262787 + layer.1.v_cache 0.00000579 0.01907984 + layer.2.k_cache 0.00517510 6.76577368 + layer.2.v_cache 0.00001992 0.05643263 + layer.3.k_cache 0.09315677 23.38517879 + layer.3.v_cache 0.00001978 0.06663628 + layer.4.k_cache 0.00061896 1.65973213 + layer.4.v_cache 0.00005133 0.13461358 + layer.4.output 9.75369281 462.00808913 + ------------------------------------------------------------------------------------- + TOTAL 4.03642769 225.07496966 + (elements=1,018,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1018368 +Total Bytes 10272 +BPFP 0.0807 bits/point +EBPFP 0.1614 equivalent bits/point +MSE 225.074970 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.006s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 225.0750 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 736B, BPFP=0.1095 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 788B, BPFP=0.1173 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 864B, BPFP=0.1286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 752B, BPFP=0.1119 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 836B, BPFP=0.1244 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 780B, BPFP=0.1161 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 816B, BPFP=0.1214 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 760B, BPFP=0.1131 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 872B, BPFP=0.1298 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 848B, BPFP=0.1262 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,368B, BPFP=0.0503 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14131550 524.83727679 + layer.0.v_cache 0.00001383 0.04763553 + layer.1.k_cache 0.02978072 36.92155180 + layer.1.v_cache 0.00000537 0.01759889 + layer.2.k_cache 0.00578868 6.59804339 + layer.2.v_cache 0.00001872 0.05262354 + layer.3.k_cache 0.04020810 23.12067290 + layer.3.v_cache 0.00001850 0.06231612 + layer.4.k_cache 0.00063358 1.67555920 + layer.4.v_cache 0.00005294 0.13855876 + layer.4.output 10.86925024 512.30310374 + ------------------------------------------------------------------------------------- + TOTAL 4.48838751 245.85844489 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 10420 +BPFP 0.0912 bits/point +EBPFP 0.1824 equivalent bits/point +MSE 245.858445 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 245.8584 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 708B, BPFP=0.1317 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1525 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 672B, BPFP=0.1250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 748B, BPFP=0.1391 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 684B, BPFP=0.1272 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 696B, BPFP=0.1295 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1466 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 780B, BPFP=0.1451 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,164B, BPFP=0.0575 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09514355 523.74497768 + layer.0.v_cache 0.00001423 0.04872142 + layer.1.k_cache 0.03283414 37.33908226 + layer.1.v_cache 0.00000553 0.01768040 + layer.2.k_cache 0.00489674 6.67360360 + layer.2.v_cache 0.00001883 0.05316739 + layer.3.k_cache 0.07199075 23.38563465 + layer.3.v_cache 0.00001932 0.06478229 + layer.4.k_cache 0.00061592 1.67422231 + layer.4.v_cache 0.00006816 0.13139200 + layer.4.output 0.17253765 647.96423257 + ------------------------------------------------------------------------------------- + TOTAL 0.08313946 301.69899365 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 9396 +BPFP 0.1028 bits/point +EBPFP 0.2056 equivalent bits/point +MSE 301.698994 +---------------------- -------------------------------------------------------- +Time: 0.375s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 301.6990 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1372 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1473 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 700B, BPFP=0.1257 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 760B, BPFP=0.1365 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 708B, BPFP=0.1272 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 716B, BPFP=0.1286 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 736B, BPFP=0.1322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 808B, BPFP=0.1451 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 812B, BPFP=0.1458 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,404B, BPFP=0.0617 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11660385 526.27047414 + layer.0.v_cache 0.00001381 0.04823413 + layer.1.k_cache 0.05175667 37.27935861 + layer.1.v_cache 0.00000569 0.01821840 + layer.2.k_cache 0.00227456 6.75026799 + layer.2.v_cache 0.00001743 0.05146653 + layer.3.k_cache 0.02781237 23.63608455 + layer.3.v_cache 0.00001883 0.06017816 + layer.4.k_cache 0.00061846 1.66099601 + layer.4.v_cache 0.00005344 0.13424058 + layer.4.output 0.17410791 628.53525246 + ------------------------------------------------------------------------------------- + TOTAL 0.08340767 293.86213449 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 9896 +BPFP 0.1045 bits/point +EBPFP 0.2091 equivalent bits/point +MSE 293.862134 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.005s, Pack+Encode: 0.164s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 293.8621 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 688B, BPFP=0.1295 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1506 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 660B, BPFP=0.1242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 724B, BPFP=0.1363 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 672B, BPFP=0.1265 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 700B, BPFP=0.1318 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 684B, BPFP=0.1288 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1483 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 772B, BPFP=0.1453 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,184B, BPFP=0.0587 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212080 524.83974962 + layer.0.v_cache 0.00001360 0.04706885 + layer.1.k_cache 0.03288655 37.20275732 + layer.1.v_cache 0.00000536 0.01700958 + layer.2.k_cache 0.00973156 6.66393161 + layer.2.v_cache 0.00001762 0.05190555 + layer.3.k_cache 0.02906220 23.38829831 + layer.3.v_cache 0.00001891 0.06219813 + layer.4.k_cache 0.00063934 1.65475464 + layer.4.v_cache 0.00004866 0.13415888 + layer.4.output 0.18425958 654.76839501 + ------------------------------------------------------------------------------------- + TOTAL 0.08496245 304.55532927 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9304 +BPFP 0.1030 bits/point +EBPFP 0.2061 equivalent bits/point +MSE 304.555329 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.003s, Pack+Encode: 0.163s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 304.5553 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1206 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 740B, BPFP=0.1360 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1493 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 676B, BPFP=0.1243 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1346 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 688B, BPFP=0.1265 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 704B, BPFP=0.1294 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 692B, BPFP=0.1272 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 816B, BPFP=0.1500 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 764B, BPFP=0.1404 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,212B, BPFP=0.0581 +⌛️ [2/4] FRONTEND: Frontend time: 0.171s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10758395 526.01755515 + layer.0.v_cache 0.00001397 0.04823701 + layer.1.k_cache 0.03293324 37.26179343 + layer.1.v_cache 0.00000575 0.01803247 + layer.2.k_cache 0.00823421 6.75379567 + layer.2.v_cache 0.00001745 0.05310605 + layer.3.k_cache 0.04116619 23.40201631 + layer.3.v_cache 0.00001985 0.06452835 + layer.4.k_cache 0.00062779 1.65063279 + layer.4.v_cache 0.00005042 0.13388326 + layer.4.output 0.16954060 641.64537815 + ------------------------------------------------------------------------------------- + TOTAL 0.08102571 299.23066044 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 9492 +BPFP 0.1026 bits/point +EBPFP 0.2053 equivalent bits/point +MSE 299.230660 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.004s, Pack+Encode: 0.171s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 299.2307 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 676B, BPFP=0.1174 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1326 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 828B, BPFP=0.1437 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 704B, BPFP=0.1222 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 772B, BPFP=0.1340 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 724B, BPFP=0.1257 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 724B, BPFP=0.1257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 760B, BPFP=0.1319 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 848B, BPFP=0.1472 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 800B, BPFP=0.1389 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,188B, BPFP=0.0543 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.284s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12307767 524.90208333 + layer.0.v_cache 0.00001410 0.04911942 + layer.1.k_cache 0.03088827 36.85610352 + layer.1.v_cache 0.00000584 0.01829213 + layer.2.k_cache 0.00371391 6.69135200 + layer.2.v_cache 0.00001950 0.05499980 + layer.3.k_cache 0.03079298 23.17874891 + layer.3.v_cache 0.00001950 0.06426676 + layer.4.k_cache 0.00063014 1.64291297 + layer.4.v_cache 0.00005282 0.13778504 + layer.4.output 0.15767360 606.19737103 + ------------------------------------------------------------------------------------- + TOTAL 0.07605470 284.52807418 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 9788 +BPFP 0.1000 bits/point +EBPFP 0.1999 equivalent bits/point +MSE 284.528074 +---------------------- -------------------------------------------------------- +Time: 0.446s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.284s +---------------------- -------------------------------------------------------- +💾 Converting with 284.5281 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 688B, BPFP=0.1311 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1524 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 664B, BPFP=0.1265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1387 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 660B, BPFP=0.1258 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1311 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 680B, BPFP=0.1296 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1471 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 752B, BPFP=0.1433 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,236B, BPFP=0.0609 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10295349 529.83212652 + layer.0.v_cache 0.00001348 0.04639900 + layer.1.k_cache 0.03334363 37.37404726 + layer.1.v_cache 0.00000550 0.01736013 + layer.2.k_cache 0.01126997 6.66152731 + layer.2.v_cache 0.00001933 0.05275433 + layer.3.k_cache 0.09131574 23.22854689 + layer.3.v_cache 0.00001769 0.05961708 + layer.4.k_cache 0.00062275 1.64932567 + layer.4.v_cache 0.00006541 0.13168258 + layer.4.output 0.17512874 662.86285932 + ------------------------------------------------------------------------------------- + TOTAL 0.08620754 308.18196482 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 9308 +BPFP 0.1043 bits/point +EBPFP 0.2087 equivalent bits/point +MSE 308.181965 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.162s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 308.1820 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1372 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1473 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 712B, BPFP=0.1279 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1379 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 728B, BPFP=0.1307 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 720B, BPFP=0.1293 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 736B, BPFP=0.1322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 832B, BPFP=0.1494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 800B, BPFP=0.1437 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,396B, BPFP=0.0615 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12314748 526.61669361 + layer.0.v_cache 0.00001636 0.04785976 + layer.1.k_cache 0.01190477 37.34577609 + layer.1.v_cache 0.00000588 0.01786578 + layer.2.k_cache 0.01614088 6.70679412 + layer.2.v_cache 0.00001834 0.05179878 + layer.3.k_cache 0.05604688 23.45595507 + layer.3.v_cache 0.00001882 0.06493929 + layer.4.k_cache 0.00063774 1.70714139 + layer.4.v_cache 0.00005162 0.13447752 + layer.4.output 0.17348555 628.52001232 + ------------------------------------------------------------------------------------- + TOTAL 0.08366986 293.86996398 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 9944 +BPFP 0.1051 bits/point +EBPFP 0.2101 equivalent bits/point +MSE 293.869964 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 293.8700 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 684B, BPFP=0.1319 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 788B, BPFP=0.1520 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 628B, BPFP=0.1211 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1381 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 644B, BPFP=0.1242 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 672B, BPFP=0.1296 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 648B, BPFP=0.1250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 736B, BPFP=0.1420 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,360B, BPFP=0.0650 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11560643 534.95408951 + layer.0.v_cache 0.00001402 0.04893083 + layer.1.k_cache 0.03272506 37.43153815 + layer.1.v_cache 0.00000563 0.01816657 + layer.2.k_cache 0.00645547 6.69356961 + layer.2.v_cache 0.00001830 0.05252163 + layer.3.k_cache 0.02731901 23.31338313 + layer.3.v_cache 0.00001960 0.06180323 + layer.4.k_cache 0.00061302 1.63475903 + layer.4.v_cache 0.00004978 0.13355976 + layer.4.output 0.17251868 670.35973325 + ------------------------------------------------------------------------------------- + TOTAL 0.08179159 311.58002672 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 9280 +BPFP 0.1053 bits/point +EBPFP 0.2106 equivalent bits/point +MSE 311.580027 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 311.5800 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 664B, BPFP=0.1179 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1357 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 824B, BPFP=0.1463 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 724B, BPFP=0.1286 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 780B, BPFP=0.1385 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 752B, BPFP=0.1335 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 732B, BPFP=0.1300 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 752B, BPFP=0.1335 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 852B, BPFP=0.1513 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 828B, BPFP=0.1470 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,500B, BPFP=0.0634 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11690623 528.11874112 + layer.0.v_cache 0.00001418 0.04855231 + layer.1.k_cache 0.03305150 37.01536144 + layer.1.v_cache 0.00000545 0.01794179 + layer.2.k_cache 0.01095151 6.72316811 + layer.2.v_cache 0.00001854 0.05270274 + layer.3.k_cache 0.04293454 23.20353283 + layer.3.v_cache 0.00002007 0.06538363 + layer.4.k_cache 0.00062489 1.62690145 + layer.4.v_cache 0.00005451 0.13466635 + layer.4.output 0.16672201 619.33411120 + ------------------------------------------------------------------------------------- + TOTAL 0.08068444 290.13798413 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 10172 +BPFP 0.1062 bits/point +EBPFP 0.2125 equivalent bits/point +MSE 290.137984 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1380 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 680B, BPFP=0.1168 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1319 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 828B, BPFP=0.1422 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 720B, BPFP=0.1236 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 764B, BPFP=0.1312 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 700B, BPFP=0.1202 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 732B, BPFP=0.1257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 712B, BPFP=0.1223 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 840B, BPFP=0.1442 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 784B, BPFP=0.1346 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,064B, BPFP=0.0506 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11067300 519.92041552 + layer.0.v_cache 0.00001409 0.04791206 + layer.1.k_cache 0.05152405 36.71401206 + layer.1.v_cache 0.00000569 0.01804768 + layer.2.k_cache 0.01362713 6.74818773 + layer.2.v_cache 0.00001889 0.05364833 + layer.3.k_cache 0.08227781 23.06860620 + layer.3.v_cache 0.00001889 0.06310881 + layer.4.k_cache 0.00062414 1.66077700 + layer.4.v_cache 0.00005257 0.13294215 + layer.4.output 0.15746453 598.16718995 + ------------------------------------------------------------------------------------- + TOTAL 0.08006400 280.91752866 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 9592 +BPFP 0.0969 bits/point +EBPFP 0.1938 equivalent bits/point +MSE 280.917529 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 280.9175 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 116, 128) +Output shape: (1, 116, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.output: torch.Size([1, 116, 3584]) -> torch.Size([1, 1, 116, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 776B, BPFP=0.1045 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 832B, BPFP=0.1121 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 900B, BPFP=0.1212 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 788B, BPFP=0.1061 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 852B, BPFP=0.1148 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 804B, BPFP=0.1083 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 832B, BPFP=0.1121 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 772B, BPFP=0.1040 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 908B, BPFP=0.1223 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 840B, BPFP=0.1131 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,032B, BPFP=0.0391 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16314862 523.46534887 + layer.0.v_cache 0.00001373 0.04753088 + layer.1.k_cache 0.05481785 36.81400694 + layer.1.v_cache 0.00000557 0.01808324 + layer.2.k_cache 0.00968372 6.69796648 + layer.2.v_cache 0.00001954 0.05610187 + layer.3.k_cache 0.02418142 23.17446584 + layer.3.v_cache 0.00001860 0.06214084 + layer.4.k_cache 0.00062333 1.65478621 + layer.4.v_cache 0.00005467 0.13623190 + layer.4.output 9.83417319 467.02378387 + ------------------------------------------------------------------------------------- + TOTAL 4.06422232 227.13489118 + (elements=1,009,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1009664 +Total Bytes 10336 +BPFP 0.0819 bits/point +EBPFP 0.1638 equivalent bits/point +MSE 227.134891 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 227.1349 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 688B, BPFP=0.1378 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1611 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 656B, BPFP=0.1314 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 720B, BPFP=0.1442 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 668B, BPFP=0.1338 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 668B, BPFP=0.1338 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1571 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1514 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,672B, BPFP=0.0765 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11749315 536.12294671 + layer.0.v_cache 0.00001389 0.04896806 + layer.1.k_cache 0.03399578 38.25269494 + layer.1.v_cache 0.00000555 0.01799861 + layer.2.k_cache 0.00854440 6.89139733 + layer.2.v_cache 0.00001777 0.05265640 + layer.3.k_cache 0.02740619 23.73572716 + layer.3.v_cache 0.00001935 0.06453601 + layer.4.k_cache 0.00063531 1.73991551 + layer.4.v_cache 0.00005166 0.13668021 + layer.4.output 0.18490290 696.30963828 + ------------------------------------------------------------------------------------- + TOTAL 0.08720608 322.42535229 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9740 +BPFP 0.1148 bits/point +EBPFP 0.2295 equivalent bits/point +MSE 322.425352 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 322.4254 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 684B, BPFP=0.1137 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 796B, BPFP=0.1323 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 840B, BPFP=0.1396 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 692B, BPFP=0.1150 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 792B, BPFP=0.1316 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 756B, BPFP=0.1257 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 756B, BPFP=0.1257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 772B, BPFP=0.1283 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 808B, BPFP=0.1343 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 824B, BPFP=0.1370 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,004B, BPFP=0.0476 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11492289 522.11411237 + layer.0.v_cache 0.00001440 0.04783504 + layer.1.k_cache 0.01112092 36.79663086 + layer.1.v_cache 0.00000555 0.01742718 + layer.2.k_cache 0.01096075 6.59537392 + layer.2.v_cache 0.00001882 0.05317413 + layer.3.k_cache 0.03781782 22.93623774 + layer.3.v_cache 0.00001905 0.06450111 + layer.4.k_cache 0.00062716 1.63056296 + layer.4.v_cache 0.00006343 0.12959533 + layer.4.output 0.15224552 577.70578457 + ------------------------------------------------------------------------------------- + TOTAL 0.07301703 272.60740839 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 9724 +BPFP 0.0951 bits/point +EBPFP 0.1902 equivalent bits/point +MSE 272.607408 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 272.6074 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 688B, BPFP=0.1396 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1631 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 628B, BPFP=0.1274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 712B, BPFP=0.1445 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 660B, BPFP=0.1339 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1380 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 644B, BPFP=0.1307 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 792B, BPFP=0.1607 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 724B, BPFP=0.1469 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,624B, BPFP=0.0761 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10405910 538.79281656 + layer.0.v_cache 0.00001453 0.04758551 + layer.1.k_cache 0.03397077 38.32696771 + layer.1.v_cache 0.00000547 0.01812170 + layer.2.k_cache 0.00568739 6.88701521 + layer.2.v_cache 0.00001916 0.05444104 + layer.3.k_cache 0.09581309 23.47117080 + layer.3.v_cache 0.00001862 0.06171487 + layer.4.k_cache 0.00061896 1.67978341 + layer.4.v_cache 0.00005653 0.13531212 + layer.4.output 0.19123089 705.01507421 + ------------------------------------------------------------------------------------- + TOTAL 0.09287528 326.15179108 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9592 +BPFP 0.1145 bits/point +EBPFP 0.2290 equivalent bits/point +MSE 326.151791 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 326.1518 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1206 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 728B, BPFP=0.1338 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1485 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 672B, BPFP=0.1235 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1346 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 680B, BPFP=0.1250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 704B, BPFP=0.1294 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 704B, BPFP=0.1294 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 800B, BPFP=0.1471 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 780B, BPFP=0.1434 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,228B, BPFP=0.0585 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09384689 528.13621324 + layer.0.v_cache 0.00001364 0.04746269 + layer.1.k_cache 0.01313462 37.25091625 + layer.1.v_cache 0.00000592 0.01826821 + layer.2.k_cache 0.00813576 6.69415499 + layer.2.v_cache 0.00001762 0.05254270 + layer.3.k_cache 0.02628646 23.21123047 + layer.3.v_cache 0.00002044 0.06471053 + layer.4.k_cache 0.00061596 1.65389907 + layer.4.v_cache 0.00005334 0.13834098 + layer.4.output 0.17343949 641.50477941 + ------------------------------------------------------------------------------------- + TOTAL 0.07977689 299.28242324 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 9492 +BPFP 0.1026 bits/point +EBPFP 0.2053 equivalent bits/point +MSE 299.282423 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.003s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 299.2824 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 126, 128) +Output shape: (1, 126, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.output: torch.Size([1, 126, 3584]) -> torch.Size([1, 1, 126, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 600B, BPFP=0.0744 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 668B, BPFP=0.0828 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 732B, BPFP=0.0908 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 592B, BPFP=0.0734 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 672B, BPFP=0.0833 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 616B, BPFP=0.0764 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 636B, BPFP=0.0789 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 620B, BPFP=0.0769 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 744B, BPFP=0.0923 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 716B, BPFP=0.0888 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,964B, BPFP=0.0348 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13141363 504.55871776 + layer.0.v_cache 0.00001429 0.04760779 + layer.1.k_cache 0.06496696 34.66675967 + layer.1.v_cache 0.00000583 0.01822494 + layer.2.k_cache 0.01057190 6.16787865 + layer.2.v_cache 0.00002020 0.05247047 + layer.3.k_cache 0.05250616 22.08099656 + layer.3.v_cache 0.00001853 0.06138927 + layer.4.k_cache 0.00062861 1.51818690 + layer.4.v_cache 0.00004930 0.12556858 + layer.4.output 9.05470577 426.76654620 + ------------------------------------------------------------------------------------- + TOTAL 3.74371387 209.21550729 + (elements=1,096,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1096704 +Total Bytes 8560 +BPFP 0.0624 bits/point +EBPFP 0.1249 equivalent bits/point +MSE 209.215507 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.006s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 209.2155 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1372 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 700B, BPFP=0.1257 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1379 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 724B, BPFP=0.1300 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 716B, BPFP=0.1286 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 728B, BPFP=0.1307 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 844B, BPFP=0.1516 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 800B, BPFP=0.1437 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,408B, BPFP=0.0618 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13059476 531.28906250 + layer.0.v_cache 0.00001380 0.04772796 + layer.1.k_cache 0.03034586 37.28196278 + layer.1.v_cache 0.00000531 0.01758092 + layer.2.k_cache 0.00374181 6.67390951 + layer.2.v_cache 0.00001886 0.05282864 + layer.3.k_cache 0.04454902 23.30609903 + layer.3.v_cache 0.00001968 0.06368258 + layer.4.k_cache 0.00060938 1.64663135 + layer.4.v_cache 0.00007316 0.13572276 + layer.4.output 0.16485390 628.58364122 + ------------------------------------------------------------------------------------- + TOTAL 0.08023229 294.15298215 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 9936 +BPFP 0.1050 bits/point +EBPFP 0.2099 equivalent bits/point +MSE 294.152982 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 294.1530 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 676B, BPFP=0.1148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 772B, BPFP=0.1311 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 844B, BPFP=0.1433 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 736B, BPFP=0.1250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 792B, BPFP=0.1345 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 748B, BPFP=0.1270 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 764B, BPFP=0.1298 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 848B, BPFP=0.1440 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 788B, BPFP=0.1338 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,064B, BPFP=0.0501 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15436328 518.59973675 + layer.0.v_cache 0.00001586 0.04799926 + layer.1.k_cache 0.01080767 36.74886687 + layer.1.v_cache 0.00000575 0.01833867 + layer.2.k_cache 0.00786338 6.57673844 + layer.2.v_cache 0.00001961 0.05321140 + layer.3.k_cache 0.05652578 23.12958560 + layer.3.v_cache 0.00002042 0.06534303 + layer.4.k_cache 0.00062828 1.64830349 + layer.4.v_cache 0.00005674 0.14187018 + layer.4.output 0.15192759 590.51489713 + ------------------------------------------------------------------------------------- + TOTAL 0.07610588 277.68436903 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 9784 +BPFP 0.0977 bits/point +EBPFP 0.1955 equivalent bits/point +MSE 277.684369 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.005s, Pack+Encode: 0.223s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 277.6844 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1266 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 708B, BPFP=0.1418 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1619 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 636B, BPFP=0.1274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 724B, BPFP=0.1450 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 660B, BPFP=0.1322 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 644B, BPFP=0.1290 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1571 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 764B, BPFP=0.1530 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,636B, BPFP=0.0754 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10254851 537.84510216 + layer.0.v_cache 0.00001399 0.04887408 + layer.1.k_cache 0.05476840 38.03910632 + layer.1.v_cache 0.00000542 0.01785932 + layer.2.k_cache 0.00244555 6.93315633 + layer.2.v_cache 0.00001844 0.05413544 + layer.3.k_cache 0.02811479 23.74996244 + layer.3.v_cache 0.00001905 0.06232146 + layer.4.k_cache 0.00060111 1.70800116 + layer.4.v_cache 0.00005313 0.13781038 + layer.4.output 0.19635826 696.18858745 + ------------------------------------------------------------------------------------- + TOTAL 0.09194683 322.46567302 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9680 +BPFP 0.1141 bits/point +EBPFP 0.2281 equivalent bits/point +MSE 322.465673 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 322.4657 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1212 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 700B, BPFP=0.1334 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 796B, BPFP=0.1517 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 656B, BPFP=0.1250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1402 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 644B, BPFP=0.1227 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 676B, BPFP=0.1288 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 780B, BPFP=0.1486 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 744B, BPFP=0.1418 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,196B, BPFP=0.0598 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11987135 532.12633384 + layer.0.v_cache 0.00001390 0.04691870 + layer.1.k_cache 0.03139273 37.22687214 + layer.1.v_cache 0.00000562 0.01746555 + layer.2.k_cache 0.00816743 6.58747082 + layer.2.v_cache 0.00001857 0.05123832 + layer.3.k_cache 0.02640573 23.42503245 + layer.3.v_cache 0.00001968 0.06211718 + layer.4.k_cache 0.00062091 1.65374384 + layer.4.v_cache 0.00004972 0.13185651 + layer.4.output 0.16983549 662.92503267 + ------------------------------------------------------------------------------------- + TOTAL 0.08090671 308.34142812 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 9256 +BPFP 0.1037 bits/point +EBPFP 0.2075 equivalent bits/point +MSE 308.341428 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.162s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 308.3414 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1372 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1473 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 736B, BPFP=0.1322 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 760B, BPFP=0.1365 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 708B, BPFP=0.1272 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 724B, BPFP=0.1300 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 732B, BPFP=0.1315 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 832B, BPFP=0.1494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 788B, BPFP=0.1415 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,412B, BPFP=0.0619 +⌛️ [2/4] FRONTEND: Frontend time: 0.154s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14089697 531.43372845 + layer.0.v_cache 0.00001393 0.04739929 + layer.1.k_cache 0.01618495 37.38053385 + layer.1.v_cache 0.00000582 0.01815118 + layer.2.k_cache 0.00661465 6.78477040 + layer.2.v_cache 0.00002083 0.05289528 + layer.3.k_cache 0.07087009 23.49412519 + layer.3.v_cache 0.00001822 0.06338687 + layer.4.k_cache 0.00061891 1.69703236 + layer.4.v_cache 0.00005311 0.13892461 + layer.4.output 0.16425455 628.59339080 + ------------------------------------------------------------------------------------- + TOTAL 0.08147525 294.19204018 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 9944 +BPFP 0.1051 bits/point +EBPFP 0.2101 equivalent bits/point +MSE 294.192040 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.154s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 294.1920 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1173 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1334 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 828B, BPFP=0.1454 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 720B, BPFP=0.1264 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 788B, BPFP=0.1383 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 748B, BPFP=0.1313 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 736B, BPFP=0.1292 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 748B, BPFP=0.1313 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 848B, BPFP=0.1489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 820B, BPFP=0.1440 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,300B, BPFP=0.0577 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12786581 518.98683287 + layer.0.v_cache 0.00001426 0.04958009 + layer.1.k_cache 0.03183105 36.84038141 + layer.1.v_cache 0.00000593 0.01935907 + layer.2.k_cache 0.00241907 6.66888359 + layer.2.v_cache 0.00001785 0.05275155 + layer.3.k_cache 0.02946699 23.40851036 + layer.3.v_cache 0.00001957 0.06584203 + layer.4.k_cache 0.00063747 1.66767420 + layer.4.v_cache 0.00005003 0.13303152 + layer.4.output 0.16568538 613.26374398 + ------------------------------------------------------------------------------------- + TOTAL 0.07953681 287.10229733 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 9964 +BPFP 0.1029 bits/point +EBPFP 0.2058 equivalent bits/point +MSE 287.102297 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 287.1023 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 680B, BPFP=0.1362 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1619 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 644B, BPFP=0.1290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 712B, BPFP=0.1426 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 652B, BPFP=0.1306 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1362 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 656B, BPFP=0.1314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 764B, BPFP=0.1530 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 732B, BPFP=0.1466 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,616B, BPFP=0.0749 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12225728 534.93970353 + layer.0.v_cache 0.00001337 0.04787531 + layer.1.k_cache 0.03480693 38.25730857 + layer.1.v_cache 0.00000545 0.01823434 + layer.2.k_cache 0.00853881 6.90431800 + layer.2.v_cache 0.00001924 0.05597683 + layer.3.k_cache 0.06130744 23.75619742 + layer.3.v_cache 0.00001928 0.06846727 + layer.4.k_cache 0.00058936 1.67426828 + layer.4.v_cache 0.00007197 0.14057155 + layer.4.output 0.18657821 695.10382326 + ------------------------------------------------------------------------------------- + TOTAL 0.09021627 321.85821670 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9584 +BPFP 0.1129 bits/point +EBPFP 0.2259 equivalent bits/point +MSE 321.858217 +---------------------- -------------------------------------------------------- +Time: 0.372s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 321.8582 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1404 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1623 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 648B, BPFP=0.1315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1453 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 672B, BPFP=0.1364 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1372 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 632B, BPFP=0.1282 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 768B, BPFP=0.1558 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 724B, BPFP=0.1469 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,664B, BPFP=0.0772 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582910 538.65006088 + layer.0.v_cache 0.00001386 0.04837446 + layer.1.k_cache 0.03408930 38.47282112 + layer.1.v_cache 0.00000538 0.01793792 + layer.2.k_cache 0.00228146 6.90793996 + layer.2.v_cache 0.00001859 0.05332275 + layer.3.k_cache 0.06276476 23.52293971 + layer.3.v_cache 0.00001913 0.06219875 + layer.4.k_cache 0.00060520 1.70366540 + layer.4.v_cache 0.00005125 0.13777665 + layer.4.output 0.19356344 704.85923006 + ------------------------------------------------------------------------------------- + TOTAL 0.09238954 326.09362635 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9628 +BPFP 0.1149 bits/point +EBPFP 0.2299 equivalent bits/point +MSE 326.093626 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.003s, Pack+Encode: 0.162s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 326.0936 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1197 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 704B, BPFP=0.1325 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1521 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 660B, BPFP=0.1242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1370 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 672B, BPFP=0.1265 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1310 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 680B, BPFP=0.1280 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 780B, BPFP=0.1468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 740B, BPFP=0.1393 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,216B, BPFP=0.0596 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09700135 529.38031815 + layer.0.v_cache 0.00001355 0.04779561 + layer.1.k_cache 0.01103399 37.30606410 + layer.1.v_cache 0.00000598 0.01774982 + layer.2.k_cache 0.00794396 6.72022661 + layer.2.v_cache 0.00001935 0.05512959 + layer.3.k_cache 0.04207947 23.25159427 + layer.3.v_cache 0.00002010 0.06663185 + layer.4.k_cache 0.00062246 1.63968024 + layer.4.v_cache 0.00005226 0.13685213 + layer.4.output 0.17535226 654.90157057 + ------------------------------------------------------------------------------------- + TOTAL 0.08154460 304.87841390 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9320 +BPFP 0.1032 bits/point +EBPFP 0.2064 equivalent bits/point +MSE 304.878414 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.004s, Pack+Encode: 0.168s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 304.8784 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1206 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 744B, BPFP=0.1368 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1485 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 684B, BPFP=0.1257 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 744B, BPFP=0.1368 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 696B, BPFP=0.1279 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 704B, BPFP=0.1294 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 712B, BPFP=0.1309 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 796B, BPFP=0.1463 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 784B, BPFP=0.1441 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,220B, BPFP=0.0583 +⌛️ [2/4] FRONTEND: Frontend time: 0.175s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11944316 522.46686581 + layer.0.v_cache 0.00001525 0.04755794 + layer.1.k_cache 0.01376067 37.20102826 + layer.1.v_cache 0.00000561 0.01774334 + layer.2.k_cache 0.00522648 6.74928625 + layer.2.v_cache 0.00001966 0.05465518 + layer.3.k_cache 0.04600189 23.27751608 + layer.3.v_cache 0.00001912 0.06499060 + layer.4.k_cache 0.00062079 1.63635577 + layer.4.v_cache 0.00005114 0.13506353 + layer.4.output 0.17175110 641.62247899 + ------------------------------------------------------------------------------------- + TOTAL 0.08161303 299.00049504 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 9548 +BPFP 0.1032 bits/point +EBPFP 0.2065 equivalent bits/point +MSE 299.000495 +---------------------- -------------------------------------------------------- +Time: 0.388s Load: 0.005s, Pack+Encode: 0.175s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 299.0005 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1386 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1619 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 636B, BPFP=0.1274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1458 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 664B, BPFP=0.1330 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 656B, BPFP=0.1314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1579 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 740B, BPFP=0.1482 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,628B, BPFP=0.0752 +⌛️ [2/4] FRONTEND: Frontend time: 0.174s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120220 534.84685497 + layer.0.v_cache 0.00001368 0.04691788 + layer.1.k_cache 0.03273370 38.17488607 + layer.1.v_cache 0.00000573 0.01754827 + layer.2.k_cache 0.00892877 6.85215564 + layer.2.v_cache 0.00001953 0.05360953 + layer.3.k_cache 0.02974504 23.85665659 + layer.3.v_cache 0.00001879 0.06440235 + layer.4.k_cache 0.00063874 1.75092629 + layer.4.v_cache 0.00005200 0.14010819 + layer.4.output 0.17853031 696.49748168 + ------------------------------------------------------------------------------------- + TOTAL 0.08429825 322.42861397 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9664 +BPFP 0.1139 bits/point +EBPFP 0.2278 equivalent bits/point +MSE 322.428614 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.004s, Pack+Encode: 0.174s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 322.4286 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1197 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 680B, BPFP=0.1280 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1529 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 648B, BPFP=0.1220 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1370 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 648B, BPFP=0.1220 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1303 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 644B, BPFP=0.1212 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1423 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,200B, BPFP=0.0592 +⌛️ [2/4] FRONTEND: Frontend time: 0.170s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.227s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09802409 530.07577184 + layer.0.v_cache 0.00001402 0.04899996 + layer.1.k_cache 0.03453328 37.25733010 + layer.1.v_cache 0.00000594 0.01850985 + layer.2.k_cache 0.01027090 6.70968885 + layer.2.v_cache 0.00001933 0.05332164 + layer.3.k_cache 0.02613793 23.36867440 + layer.3.v_cache 0.00001931 0.06440102 + layer.4.k_cache 0.00063818 1.69604382 + layer.4.v_cache 0.00005393 0.13413567 + layer.4.output 0.18084440 654.76376936 + ------------------------------------------------------------------------------------- + TOTAL 0.08444869 304.86901545 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9228 +BPFP 0.1022 bits/point +EBPFP 0.2044 equivalent bits/point +MSE 304.869015 +---------------------- -------------------------------------------------------- +Time: 0.401s Load: 0.004s, Pack+Encode: 0.170s, Decode+Unpack: 0.227s +---------------------- -------------------------------------------------------- +💾 Converting with 304.8690 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 708B, BPFP=0.1074 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 796B, BPFP=0.1208 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 848B, BPFP=0.1286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 752B, BPFP=0.1141 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 800B, BPFP=0.1214 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 756B, BPFP=0.1147 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 800B, BPFP=0.1214 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 772B, BPFP=0.1171 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 868B, BPFP=0.1317 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 816B, BPFP=0.1238 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,256B, BPFP=0.0489 +⌛️ [2/4] FRONTEND: Frontend time: 0.173s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13030120 516.63797027 + layer.0.v_cache 0.00001385 0.04728524 + layer.1.k_cache 0.02890817 37.18142019 + layer.1.v_cache 0.00000611 0.01888562 + layer.2.k_cache 0.00721959 6.65935687 + layer.2.v_cache 0.00001830 0.05372862 + layer.3.k_cache 0.02195997 23.10237314 + layer.3.v_cache 0.00001903 0.06308115 + layer.4.k_cache 0.00063766 1.68214224 + layer.4.v_cache 0.00005061 0.13441123 + layer.4.output 11.07892277 523.07606623 + ------------------------------------------------------------------------------------- + TOTAL 4.57303494 249.83018342 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 10172 +BPFP 0.0908 bits/point +EBPFP 0.1815 equivalent bits/point +MSE 249.830183 +---------------------- -------------------------------------------------------- +Time: 0.402s Load: 0.005s, Pack+Encode: 0.173s, Decode+Unpack: 0.225s +---------------------- -------------------------------------------------------- +💾 Converting with 249.8302 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1197 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1303 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1529 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 664B, BPFP=0.1250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1386 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 656B, BPFP=0.1235 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1288 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 676B, BPFP=0.1273 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 776B, BPFP=0.1461 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 744B, BPFP=0.1401 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,144B, BPFP=0.0577 +⌛️ [2/4] FRONTEND: Frontend time: 0.171s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.262s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12633393 530.29141566 + layer.0.v_cache 0.00001372 0.04725963 + layer.1.k_cache 0.01411889 37.32132436 + layer.1.v_cache 0.00000549 0.01792480 + layer.2.k_cache 0.00833277 6.71177728 + layer.2.v_cache 0.00001851 0.05220203 + layer.3.k_cache 0.05797521 23.27409344 + layer.3.v_cache 0.00001879 0.06199025 + layer.4.k_cache 0.00064672 1.72832884 + layer.4.v_cache 0.00005349 0.13740587 + layer.4.output 0.17447430 654.88495052 + ------------------------------------------------------------------------------------- + TOTAL 0.08404927 304.93166916 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9220 +BPFP 0.1021 bits/point +EBPFP 0.2042 equivalent bits/point +MSE 304.931669 +---------------------- -------------------------------------------------------- +Time: 0.437s Load: 0.005s, Pack+Encode: 0.171s, Decode+Unpack: 0.262s +---------------------- -------------------------------------------------------- +💾 Converting with 304.9317 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 648B, BPFP=0.1446 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 720B, BPFP=0.1607 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 772B, BPFP=0.1723 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 680B, BPFP=0.1518 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1625 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 680B, BPFP=0.1518 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1545 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 668B, BPFP=0.1491 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 760B, BPFP=0.1696 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,484B, BPFP=0.0792 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08086312 524.96752232 + layer.0.v_cache 0.00001331 0.04753908 + layer.1.k_cache 0.01551799 38.30032785 + layer.1.v_cache 0.00000513 0.01746664 + layer.2.k_cache 0.00232868 7.09125977 + layer.2.v_cache 0.00002262 0.05603851 + layer.3.k_cache 0.10623371 23.52268764 + layer.3.v_cache 0.00001836 0.06422056 + layer.4.k_cache 0.00061897 1.74526106 + layer.4.v_cache 0.00005922 0.13914615 + layer.4.output 0.20964142 776.63705357 + ------------------------------------------------------------------------------------- + TOTAL 0.09842183 354.84769674 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 9588 +BPFP 0.1259 bits/point +EBPFP 0.2518 equivalent bits/point +MSE 354.847697 +---------------------- -------------------------------------------------------- +Time: 0.386s Load: 0.004s, Pack+Encode: 0.168s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 354.8477 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 680B, BPFP=0.1155 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1304 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 848B, BPFP=0.1440 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 684B, BPFP=0.1162 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 780B, BPFP=0.1325 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 716B, BPFP=0.1216 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 828B, BPFP=0.1406 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 808B, BPFP=0.1372 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,092B, BPFP=0.0508 +⌛️ [2/4] FRONTEND: Frontend time: 0.153s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13670269 515.31810462 + layer.0.v_cache 0.00001568 0.04777662 + layer.1.k_cache 0.03139175 36.73743472 + layer.1.v_cache 0.00000602 0.01842971 + layer.2.k_cache 0.00629326 6.55112092 + layer.2.v_cache 0.00001871 0.05148959 + layer.3.k_cache 0.03837063 23.25575323 + layer.3.v_cache 0.00001860 0.06031938 + layer.4.k_cache 0.00062597 1.62812026 + layer.4.v_cache 0.00005354 0.13196325 + layer.4.output 0.15974679 590.11738160 + ------------------------------------------------------------------------------------- + TOTAL 0.07833673 277.33071668 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 9708 +BPFP 0.0970 bits/point +EBPFP 0.1940 equivalent bits/point +MSE 277.330717 +---------------------- -------------------------------------------------------- +Time: 0.356s Load: 0.004s, Pack+Encode: 0.153s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 277.3307 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 756B, BPFP=0.1104 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 804B, BPFP=0.1174 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 868B, BPFP=0.1268 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 784B, BPFP=0.1145 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 824B, BPFP=0.1203 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 788B, BPFP=0.1151 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 808B, BPFP=0.1180 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 768B, BPFP=0.1121 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 872B, BPFP=0.1273 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 836B, BPFP=0.1221 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,308B, BPFP=0.0481 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12281527 527.49886828 + layer.0.v_cache 0.00001417 0.04907800 + layer.1.k_cache 0.04442539 37.32305509 + layer.1.v_cache 0.00000572 0.01849131 + layer.2.k_cache 0.01033308 6.70334461 + layer.2.v_cache 0.00001923 0.05529838 + layer.3.k_cache 0.04555132 23.15389530 + layer.3.v_cache 0.00001841 0.06448133 + layer.4.k_cache 0.00060580 1.64696681 + layer.4.v_cache 0.00007068 0.13489066 + layer.4.output 10.66341841 503.99102971 + ------------------------------------------------------------------------------------- + TOTAL 4.40398753 242.62268104 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 10416 +BPFP 0.0895 bits/point +EBPFP 0.1789 equivalent bits/point +MSE 242.622681 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.005s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 242.6227 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1192 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1388 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1483 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 692B, BPFP=0.1257 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 764B, BPFP=0.1388 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 700B, BPFP=0.1272 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 724B, BPFP=0.1315 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 716B, BPFP=0.1301 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 812B, BPFP=0.1475 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 788B, BPFP=0.1432 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,364B, BPFP=0.0614 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11362342 529.79033430 + layer.0.v_cache 0.00001418 0.04876235 + layer.1.k_cache 0.05429294 37.51542060 + layer.1.v_cache 0.00000580 0.01858960 + layer.2.k_cache 0.00513919 6.71505737 + layer.2.v_cache 0.00001816 0.05397574 + layer.3.k_cache 0.05885953 23.45970828 + layer.3.v_cache 0.00001841 0.06239332 + layer.4.k_cache 0.00061763 1.68068145 + layer.4.v_cache 0.00005080 0.13042177 + layer.4.output 0.16727475 637.62053571 + ------------------------------------------------------------------------------------- + TOTAL 0.08256255 297.81288793 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 9796 +BPFP 0.1047 bits/point +EBPFP 0.2094 equivalent bits/point +MSE 297.812888 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 297.8129 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 684B, BPFP=0.1149 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 796B, BPFP=0.1337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 852B, BPFP=0.1431 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 792B, BPFP=0.1331 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 716B, BPFP=0.1203 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 740B, BPFP=0.1243 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 736B, BPFP=0.1237 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 820B, BPFP=0.1378 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 804B, BPFP=0.1351 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,988B, BPFP=0.0477 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702597 516.28167003 + layer.0.v_cache 0.00001402 0.04835392 + layer.1.k_cache 0.01332537 36.88309602 + layer.1.v_cache 0.00000581 0.01826325 + layer.2.k_cache 0.00507878 6.60461295 + layer.2.v_cache 0.00001944 0.05314810 + layer.3.k_cache 0.04051446 23.14269888 + layer.3.v_cache 0.00001931 0.06188341 + layer.4.k_cache 0.00062870 1.63957395 + layer.4.v_cache 0.00006514 0.13543523 + layer.4.output 0.16549673 584.23507104 + ------------------------------------------------------------------------------------- + TOTAL 0.07795142 274.97142547 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 9616 +BPFP 0.0950 bits/point +EBPFP 0.1901 equivalent bits/point +MSE 274.971425 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 274.9714 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1206 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 724B, BPFP=0.1331 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1493 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 664B, BPFP=0.1221 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1353 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 684B, BPFP=0.1257 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1272 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 688B, BPFP=0.1265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 796B, BPFP=0.1463 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1390 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,268B, BPFP=0.0596 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13925676 523.31516544 + layer.0.v_cache 0.00001387 0.04762986 + layer.1.k_cache 0.03290265 37.28080193 + layer.1.v_cache 0.00000589 0.01774994 + layer.2.k_cache 0.00800494 6.71708123 + layer.2.v_cache 0.00001921 0.05277070 + layer.3.k_cache 0.08301921 23.19653464 + layer.3.v_cache 0.00001931 0.06260127 + layer.4.k_cache 0.00065893 1.65811875 + layer.4.v_cache 0.00004891 0.13320154 + layer.4.output 0.17367665 641.61060924 + ------------------------------------------------------------------------------------- + TOTAL 0.08704037 299.04446588 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 9476 +BPFP 0.1025 bits/point +EBPFP 0.2049 equivalent bits/point +MSE 299.044466 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 299.0445 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 700B, BPFP=0.1062 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 788B, BPFP=0.1195 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 840B, BPFP=0.1274 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 756B, BPFP=0.1147 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 800B, BPFP=0.1214 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 740B, BPFP=0.1123 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 788B, BPFP=0.1195 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 748B, BPFP=0.1135 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 856B, BPFP=0.1299 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 832B, BPFP=0.1262 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,212B, BPFP=0.0479 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10583916 521.74408374 + layer.0.v_cache 0.00001406 0.04805083 + layer.1.k_cache 0.01351924 37.03773750 + layer.1.v_cache 0.00000536 0.01787121 + layer.2.k_cache 0.00567923 6.65172392 + layer.2.v_cache 0.00001856 0.05293642 + layer.3.k_cache 0.05293757 23.24878641 + layer.3.v_cache 0.00001841 0.06104648 + layer.4.k_cache 0.00064880 1.69982169 + layer.4.v_cache 0.00004836 0.13090910 + layer.4.output 11.07886250 523.01248266 + ------------------------------------------------------------------------------------- + TOTAL 4.57239802 250.10472623 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 10060 +BPFP 0.0898 bits/point +EBPFP 0.1795 equivalent bits/point +MSE 250.104726 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 250.1047 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 628B, BPFP=0.1182 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1303 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1506 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 676B, BPFP=0.1273 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1370 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 668B, BPFP=0.1258 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1303 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 668B, BPFP=0.1258 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1483 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 768B, BPFP=0.1446 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,176B, BPFP=0.0585 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860243 529.44606551 + layer.0.v_cache 0.00001403 0.04814137 + layer.1.k_cache 0.01400426 37.24476421 + layer.1.v_cache 0.00000584 0.01830134 + layer.2.k_cache 0.01277778 6.66641934 + layer.2.v_cache 0.00001980 0.05472361 + layer.3.k_cache 0.07376247 23.38456561 + layer.3.v_cache 0.00001877 0.06258160 + layer.4.k_cache 0.00062152 1.66700175 + layer.4.v_cache 0.00005140 0.13674759 + layer.4.output 0.18179212 654.88226119 + ------------------------------------------------------------------------------------- + TOTAL 0.08837783 304.87677296 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9284 +BPFP 0.1028 bits/point +EBPFP 0.2056 equivalent bits/point +MSE 304.876773 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 304.8768 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1160 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1326 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 824B, BPFP=0.1431 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1194 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 772B, BPFP=0.1340 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 700B, BPFP=0.1215 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 732B, BPFP=0.1271 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 736B, BPFP=0.1278 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 832B, BPFP=0.1444 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 816B, BPFP=0.1417 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,144B, BPFP=0.0532 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13781073 521.82977431 + layer.0.v_cache 0.00001414 0.04899782 + layer.1.k_cache 0.01252146 36.82716743 + layer.1.v_cache 0.00000561 0.01812224 + layer.2.k_cache 0.00238173 6.69856771 + layer.2.v_cache 0.00001849 0.05413117 + layer.3.k_cache 0.04360520 23.39977485 + layer.3.v_cache 0.00001973 0.06606562 + layer.4.k_cache 0.00061668 1.64114956 + layer.4.v_cache 0.00005448 0.14024779 + layer.4.output 0.16772948 606.65158730 + ------------------------------------------------------------------------------------- + TOTAL 0.08065615 284.54618292 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 9676 +BPFP 0.0988 bits/point +EBPFP 0.1976 equivalent bits/point +MSE 284.546183 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 284.5462 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 624B, BPFP=0.1234 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 676B, BPFP=0.1337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1582 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 652B, BPFP=0.1290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 708B, BPFP=0.1400 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 652B, BPFP=0.1290 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 672B, BPFP=0.1329 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 644B, BPFP=0.1274 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 776B, BPFP=0.1535 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 736B, BPFP=0.1456 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,588B, BPFP=0.0731 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07550411 538.97577136 + layer.0.v_cache 0.00001381 0.04959022 + layer.1.k_cache 0.01271453 38.10403481 + layer.1.v_cache 0.00000531 0.01743364 + layer.2.k_cache 0.00681068 6.87864125 + layer.2.v_cache 0.00001904 0.05411280 + layer.3.k_cache 0.02844424 23.84491508 + layer.3.v_cache 0.00001798 0.06262384 + layer.4.k_cache 0.00061127 1.73546436 + layer.4.v_cache 0.00009323 0.14187173 + layer.4.output 0.18255964 687.23632459 + ------------------------------------------------------------------------------------- + TOTAL 0.08247951 318.85404301 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 9528 +BPFP 0.1109 bits/point +EBPFP 0.2217 equivalent bits/point +MSE 318.854043 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 318.8540 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 624B, BPFP=0.1219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 684B, BPFP=0.1336 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1586 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 652B, BPFP=0.1273 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1422 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 648B, BPFP=0.1266 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1352 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 664B, BPFP=0.1297 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1531 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 740B, BPFP=0.1445 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,320B, BPFP=0.0647 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08590809 535.72314453 + layer.0.v_cache 0.00001388 0.04807695 + layer.1.k_cache 0.01594085 37.71688232 + layer.1.v_cache 0.00000552 0.01787165 + layer.2.k_cache 0.00791870 6.74448395 + layer.2.v_cache 0.00001801 0.05289150 + layer.3.k_cache 0.02757127 23.69024353 + layer.3.v_cache 0.00001862 0.06337931 + layer.4.k_cache 0.00063773 1.67445374 + layer.4.v_cache 0.00005492 0.13902227 + layer.4.output 0.18408036 678.24648438 + ------------------------------------------------------------------------------------- + TOTAL 0.08392059 314.91740238 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 9348 +BPFP 0.1074 bits/point +EBPFP 0.2148 equivalent bits/point +MSE 314.917402 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 314.9174 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 644B, BPFP=0.1324 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 712B, BPFP=0.1464 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 796B, BPFP=0.1637 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 648B, BPFP=0.1332 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 724B, BPFP=0.1488 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 668B, BPFP=0.1373 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1398 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 680B, BPFP=0.1398 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1612 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 744B, BPFP=0.1530 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,664B, BPFP=0.0782 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08991440 531.60762747 + layer.0.v_cache 0.00001347 0.04780081 + layer.1.k_cache 0.03345199 38.35521176 + layer.1.v_cache 0.00000518 0.01742771 + layer.2.k_cache 0.00398020 6.91675929 + layer.2.v_cache 0.00001815 0.05306507 + layer.3.k_cache 0.04607732 23.59482293 + layer.3.v_cache 0.00001896 0.06331028 + layer.4.k_cache 0.00062119 1.70034750 + layer.4.v_cache 0.00005040 0.13528750 + layer.4.output 0.18612516 715.33276551 + ------------------------------------------------------------------------------------- + TOTAL 0.08688396 329.98947170 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 9744 +BPFP 0.1178 bits/point +EBPFP 0.2357 equivalent bits/point +MSE 329.989472 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 329.9895 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 676B, BPFP=0.1161 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1319 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 840B, BPFP=0.1442 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 680B, BPFP=0.1168 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 776B, BPFP=0.1332 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 712B, BPFP=0.1223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 732B, BPFP=0.1257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 716B, BPFP=0.1229 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 840B, BPFP=0.1442 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 800B, BPFP=0.1374 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,088B, BPFP=0.0512 +⌛️ [2/4] FRONTEND: Frontend time: 0.154s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11186004 520.44037603 + layer.0.v_cache 0.00001364 0.04815795 + layer.1.k_cache 0.01386243 36.63096669 + layer.1.v_cache 0.00000608 0.01771230 + layer.2.k_cache 0.00506424 6.64804178 + layer.2.v_cache 0.00001872 0.05357482 + layer.3.k_cache 0.02496567 23.14975210 + layer.3.v_cache 0.00001876 0.06443645 + layer.4.k_cache 0.00064518 1.64932285 + layer.4.v_cache 0.00005436 0.13749756 + layer.4.output 0.16682192 598.27374411 + ------------------------------------------------------------------------------------- + TOTAL 0.07789780 280.98564984 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 9628 +BPFP 0.0972 bits/point +EBPFP 0.1945 equivalent bits/point +MSE 280.985650 +---------------------- -------------------------------------------------------- +Time: 0.355s Load: 0.003s, Pack+Encode: 0.154s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 280.9856 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1197 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 720B, BPFP=0.1355 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1506 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 656B, BPFP=0.1235 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1378 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 664B, BPFP=0.1250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 692B, BPFP=0.1303 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 780B, BPFP=0.1468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 748B, BPFP=0.1408 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,172B, BPFP=0.0584 +⌛️ [2/4] FRONTEND: Frontend time: 0.153s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08542032 527.31306476 + layer.0.v_cache 0.00001402 0.04824764 + layer.1.k_cache 0.01083013 37.29023908 + layer.1.v_cache 0.00000556 0.01773891 + layer.2.k_cache 0.00382075 6.67058664 + layer.2.v_cache 0.00001860 0.05313532 + layer.3.k_cache 0.02610962 23.26641625 + layer.3.v_cache 0.00002051 0.06306522 + layer.4.k_cache 0.00063312 1.64613728 + layer.4.v_cache 0.00005614 0.14197641 + layer.4.output 0.16811641 654.75672332 + ------------------------------------------------------------------------------------- + TOTAL 0.07669080 304.69456887 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9288 +BPFP 0.1029 bits/point +EBPFP 0.2057 equivalent bits/point +MSE 304.694569 +---------------------- -------------------------------------------------------- +Time: 0.355s Load: 0.005s, Pack+Encode: 0.153s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 304.6946 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 624B, BPFP=0.1204 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 684B, BPFP=0.1319 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 796B, BPFP=0.1535 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 640B, BPFP=0.1235 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 712B, BPFP=0.1373 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 660B, BPFP=0.1273 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1304 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 648B, BPFP=0.1250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 780B, BPFP=0.1505 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 720B, BPFP=0.1389 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,348B, BPFP=0.0647 +⌛️ [2/4] FRONTEND: Frontend time: 0.153s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10540552 536.35869985 + layer.0.v_cache 0.00001541 0.04823255 + layer.1.k_cache 0.03402772 37.37279369 + layer.1.v_cache 0.00000555 0.01789963 + layer.2.k_cache 0.00394259 6.70302252 + layer.2.v_cache 0.00001868 0.05477339 + layer.3.k_cache 0.02873192 23.35452384 + layer.3.v_cache 0.00001983 0.06567087 + layer.4.k_cache 0.00063741 1.74940114 + layer.4.v_cache 0.00005806 0.14257975 + layer.4.output 0.17961645 670.06751543 + ------------------------------------------------------------------------------------- + TOTAL 0.08412811 311.54942384 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 9288 +BPFP 0.1054 bits/point +EBPFP 0.2108 equivalent bits/point +MSE 311.549424 +---------------------- -------------------------------------------------------- +Time: 0.356s Load: 0.005s, Pack+Encode: 0.153s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 311.5494 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 628B, BPFP=0.1227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1352 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1562 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 664B, BPFP=0.1297 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 652B, BPFP=0.1273 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1328 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 660B, BPFP=0.1289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 780B, BPFP=0.1523 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 752B, BPFP=0.1469 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,324B, BPFP=0.0648 +⌛️ [2/4] FRONTEND: Frontend time: 0.154s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10022384 533.85849609 + layer.0.v_cache 0.00001350 0.04855855 + layer.1.k_cache 0.03445883 37.69783936 + layer.1.v_cache 0.00000556 0.01851501 + layer.2.k_cache 0.00550650 6.78411179 + layer.2.v_cache 0.00001848 0.05407590 + layer.3.k_cache 0.03016714 23.43065643 + layer.3.v_cache 0.00001937 0.06609976 + layer.4.k_cache 0.00063435 1.72673016 + layer.4.v_cache 0.00006510 0.14301088 + layer.4.output 0.18593751 677.98152902 + ------------------------------------------------------------------------------------- + TOTAL 0.08662796 314.68816453 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 9348 +BPFP 0.1074 bits/point +EBPFP 0.2148 equivalent bits/point +MSE 314.688165 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.003s, Pack+Encode: 0.154s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 314.6882 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 684B, BPFP=0.1102 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 812B, BPFP=0.1308 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 824B, BPFP=0.1327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1108 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 776B, BPFP=0.1250 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 720B, BPFP=0.1160 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 756B, BPFP=0.1218 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 752B, BPFP=0.1211 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 824B, BPFP=0.1327 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 820B, BPFP=0.1321 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,008B, BPFP=0.0462 +⌛️ [2/4] FRONTEND: Frontend time: 0.154s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11898830 519.98018686 + layer.0.v_cache 0.00001403 0.04817438 + layer.1.k_cache 0.01243949 37.10552664 + layer.1.v_cache 0.00000551 0.01790108 + layer.2.k_cache 0.00346032 6.78031245 + layer.2.v_cache 0.00001901 0.05335630 + layer.3.k_cache 0.02403709 23.00318138 + layer.3.v_cache 0.00001994 0.06321141 + layer.4.k_cache 0.00063949 1.67392471 + layer.4.v_cache 0.00006591 0.13558120 + layer.4.output 0.03773703 579.19969624 + ------------------------------------------------------------------------------------- + TOTAL 0.02493225 273.13289589 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 9664 +BPFP 0.0916 bits/point +EBPFP 0.1831 equivalent bits/point +MSE 273.132896 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.005s, Pack+Encode: 0.154s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 273.1329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1204 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 704B, BPFP=0.1341 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 792B, BPFP=0.1509 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 652B, BPFP=0.1242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1387 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 656B, BPFP=0.1250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1326 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 660B, BPFP=0.1258 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1471 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 744B, BPFP=0.1418 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,236B, BPFP=0.0609 +⌛️ [2/4] FRONTEND: Frontend time: 0.154s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10693085 530.19107279 + layer.0.v_cache 0.00001341 0.04708822 + layer.1.k_cache 0.01172146 37.30401462 + layer.1.v_cache 0.00000553 0.01780589 + layer.2.k_cache 0.00236719 6.63158361 + layer.2.v_cache 0.00001835 0.05105127 + layer.3.k_cache 0.02622769 23.26505931 + layer.3.v_cache 0.00001864 0.06172165 + layer.4.k_cache 0.00061713 1.64429623 + layer.4.v_cache 0.00004943 0.13244604 + layer.4.output 0.18273007 662.88888284 + ------------------------------------------------------------------------------------- + TOTAL 0.08394589 308.20990115 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 9272 +BPFP 0.1039 bits/point +EBPFP 0.2079 equivalent bits/point +MSE 308.209901 +---------------------- -------------------------------------------------------- +Time: 0.356s Load: 0.005s, Pack+Encode: 0.154s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 308.2099 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 700B, BPFP=0.1083 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 788B, BPFP=0.1219 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 836B, BPFP=0.1293 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 712B, BPFP=0.1101 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1188 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 752B, BPFP=0.1163 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 756B, BPFP=0.1170 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 744B, BPFP=0.1151 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 856B, BPFP=0.1324 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 820B, BPFP=0.1269 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,172B, BPFP=0.0480 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13156438 515.08098700 + layer.0.v_cache 0.00001398 0.04843139 + layer.1.k_cache 0.04507210 37.02709236 + layer.1.v_cache 0.00000603 0.01850368 + layer.2.k_cache 0.00848899 6.62400893 + layer.2.v_cache 0.00001983 0.05367983 + layer.3.k_cache 0.06093804 23.18770788 + layer.3.v_cache 0.00001891 0.06100395 + layer.4.k_cache 0.00062755 1.66752217 + layer.4.v_cache 0.00004781 0.12791844 + layer.4.output 11.29576791 532.89763083 + ------------------------------------------------------------------------------------- + TOTAL 4.66571606 253.77531009 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 9904 +BPFP 0.0901 bits/point +EBPFP 0.1803 equivalent bits/point +MSE 253.775310 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 253.7753 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1303 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1506 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 660B, BPFP=0.1242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1370 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 656B, BPFP=0.1235 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 684B, BPFP=0.1288 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 744B, BPFP=0.1401 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,216B, BPFP=0.0596 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12340720 530.22016190 + layer.0.v_cache 0.00001379 0.04829477 + layer.1.k_cache 0.01365945 37.31043510 + layer.1.v_cache 0.00000586 0.01833798 + layer.2.k_cache 0.00515996 6.70962341 + layer.2.v_cache 0.00002118 0.05545695 + layer.3.k_cache 0.04682968 23.48677376 + layer.3.v_cache 0.00001899 0.06560676 + layer.4.k_cache 0.00062596 1.67808992 + layer.4.v_cache 0.00004975 0.13323853 + layer.4.output 0.18155291 654.69416954 + ------------------------------------------------------------------------------------- + TOTAL 0.08592130 304.85795328 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9284 +BPFP 0.1028 bits/point +EBPFP 0.2056 equivalent bits/point +MSE 304.857953 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 304.8580 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 644B, BPFP=0.1198 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 716B, BPFP=0.1332 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 824B, BPFP=0.1533 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 668B, BPFP=0.1243 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1369 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 684B, BPFP=0.1272 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1287 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 680B, BPFP=0.1265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 792B, BPFP=0.1473 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 772B, BPFP=0.1436 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,200B, BPFP=0.0585 +⌛️ [2/4] FRONTEND: Frontend time: 0.154s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10400932 529.27659970 + layer.0.v_cache 0.00001381 0.04742079 + layer.1.k_cache 0.01300498 37.38158308 + layer.1.v_cache 0.00000559 0.01779873 + layer.2.k_cache 0.00535372 6.68532744 + layer.2.v_cache 0.00001862 0.05252197 + layer.3.k_cache 0.04415199 23.26846749 + layer.3.v_cache 0.00001841 0.06423513 + layer.4.k_cache 0.00060249 1.65497989 + layer.4.v_cache 0.00005274 0.13894070 + layer.4.output 0.16694923 648.03454507 + ------------------------------------------------------------------------------------- + TOTAL 0.07858096 302.04880532 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 9408 +BPFP 0.1029 bits/point +EBPFP 0.2059 equivalent bits/point +MSE 302.048805 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.005s, Pack+Encode: 0.154s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 302.0488 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 620B, BPFP=0.1211 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 708B, BPFP=0.1383 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1562 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 652B, BPFP=0.1273 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 720B, BPFP=0.1406 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 672B, BPFP=0.1313 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1320 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 660B, BPFP=0.1289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1508 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1477 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,308B, BPFP=0.0644 +⌛️ [2/4] FRONTEND: Frontend time: 0.153s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09321334 540.12656250 + layer.0.v_cache 0.00001862 0.05013551 + layer.1.k_cache 0.03257937 37.71296692 + layer.1.v_cache 0.00000532 0.01764544 + layer.2.k_cache 0.01442847 6.80963669 + layer.2.v_cache 0.00001865 0.05525402 + layer.3.k_cache 0.04385926 23.48516541 + layer.3.v_cache 0.00001828 0.06309397 + layer.4.k_cache 0.00061064 1.70591965 + layer.4.v_cache 0.00005353 0.13542700 + layer.4.output 0.18500509 677.88175223 + ------------------------------------------------------------------------------------- + TOTAL 0.08704948 315.01965134 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 9344 +BPFP 0.1074 bits/point +EBPFP 0.2147 equivalent bits/point +MSE 315.019651 +---------------------- -------------------------------------------------------- +Time: 0.356s Load: 0.005s, Pack+Encode: 0.153s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 315.0197 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 708B, BPFP=0.1074 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 788B, BPFP=0.1195 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 848B, BPFP=0.1286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 756B, BPFP=0.1147 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 804B, BPFP=0.1220 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 776B, BPFP=0.1177 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 780B, BPFP=0.1183 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 756B, BPFP=0.1147 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 872B, BPFP=0.1323 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 848B, BPFP=0.1286 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,252B, BPFP=0.0488 +⌛️ [2/4] FRONTEND: Frontend time: 0.154s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11527186 522.10550667 + layer.0.v_cache 0.00001510 0.04742463 + layer.1.k_cache 0.02920971 37.07701096 + layer.1.v_cache 0.00000551 0.01807557 + layer.2.k_cache 0.00857845 6.60621658 + layer.2.v_cache 0.00001837 0.05320982 + layer.3.k_cache 0.03550095 23.29180256 + layer.3.v_cache 0.00001950 0.06331050 + layer.4.k_cache 0.00063242 1.69670475 + layer.4.v_cache 0.00005150 0.13715925 + layer.4.output 11.08123415 523.01738037 + ------------------------------------------------------------------------------------- + TOTAL 4.57399661 250.13047553 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 10188 +BPFP 0.0909 bits/point +EBPFP 0.1818 equivalent bits/point +MSE 250.130476 +---------------------- -------------------------------------------------------- +Time: 0.356s Load: 0.004s, Pack+Encode: 0.154s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 250.1305 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 676B, BPFP=0.1148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 784B, BPFP=0.1332 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 848B, BPFP=0.1440 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 708B, BPFP=0.1202 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 792B, BPFP=0.1345 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 736B, BPFP=0.1250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 764B, BPFP=0.1298 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 740B, BPFP=0.1257 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 828B, BPFP=0.1406 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 796B, BPFP=0.1352 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,996B, BPFP=0.0484 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09673090 514.83738111 + layer.0.v_cache 0.00001344 0.04748420 + layer.1.k_cache 0.04918879 36.69584855 + layer.1.v_cache 0.00000572 0.01807559 + layer.2.k_cache 0.00880511 6.56535472 + layer.2.v_cache 0.00001944 0.05350475 + layer.3.k_cache 0.06586690 23.23092519 + layer.3.v_cache 0.00001823 0.06185066 + layer.4.k_cache 0.00062484 1.67778513 + layer.4.v_cache 0.00006325 0.13393999 + layer.4.output 0.15294319 590.49791343 + ------------------------------------------------------------------------------------- + TOTAL 0.07599641 277.45926729 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 9668 +BPFP 0.0966 bits/point +EBPFP 0.1932 equivalent bits/point +MSE 277.459267 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 277.4593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 684B, BPFP=0.1162 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 772B, BPFP=0.1311 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 852B, BPFP=0.1447 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 720B, BPFP=0.1223 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 792B, BPFP=0.1345 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 760B, BPFP=0.1291 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 840B, BPFP=0.1427 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 812B, BPFP=0.1379 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,000B, BPFP=0.0485 +⌛️ [2/4] FRONTEND: Frontend time: 0.154s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12676482 515.38136889 + layer.0.v_cache 0.00001339 0.04701965 + layer.1.k_cache 0.03133728 36.79694400 + layer.1.v_cache 0.00000590 0.01778081 + layer.2.k_cache 0.00372887 6.67214966 + layer.2.v_cache 0.00001820 0.05367080 + layer.3.k_cache 0.02418065 23.05169412 + layer.3.v_cache 0.00002034 0.06390192 + layer.4.k_cache 0.00062996 1.63766280 + layer.4.v_cache 0.00006053 0.13260319 + layer.4.output 0.15389095 590.91401398 + ------------------------------------------------------------------------------------- + TOTAL 0.07435274 277.66193492 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 9736 +BPFP 0.0973 bits/point +EBPFP 0.1945 equivalent bits/point +MSE 277.661935 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.005s, Pack+Encode: 0.154s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 277.6619 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 672B, BPFP=0.1167 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 784B, BPFP=0.1361 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 832B, BPFP=0.1444 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 712B, BPFP=0.1236 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 776B, BPFP=0.1347 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 732B, BPFP=0.1271 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 736B, BPFP=0.1278 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 736B, BPFP=0.1278 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 828B, BPFP=0.1437 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 796B, BPFP=0.1382 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,172B, BPFP=0.0539 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12863920 523.80503472 + layer.0.v_cache 0.00001425 0.04883023 + layer.1.k_cache 0.01389826 36.79219293 + layer.1.v_cache 0.00000571 0.01866979 + layer.2.k_cache 0.00659207 6.71659342 + layer.2.v_cache 0.00001883 0.05358273 + layer.3.k_cache 0.05540012 23.42501085 + layer.3.v_cache 0.00001922 0.06177930 + layer.4.k_cache 0.00063279 1.66215515 + layer.4.v_cache 0.00005416 0.13399117 + layer.4.output 0.16196846 606.05431548 + ------------------------------------------------------------------------------------- + TOTAL 0.07876787 284.41753227 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 9776 +BPFP 0.0998 bits/point +EBPFP 0.1997 equivalent bits/point +MSE 284.417532 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 284.4175 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 768B, BPFP=0.1111 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 812B, BPFP=0.1175 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 864B, BPFP=0.1250 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 768B, BPFP=0.1111 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 828B, BPFP=0.1198 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 788B, BPFP=0.1140 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 824B, BPFP=0.1192 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 792B, BPFP=0.1146 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 884B, BPFP=0.1279 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 860B, BPFP=0.1244 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,228B, BPFP=0.0460 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13160179 528.79260706 + layer.0.v_cache 0.00001577 0.04796790 + layer.1.k_cache 0.05775581 37.21766719 + layer.1.v_cache 0.00000549 0.01784132 + layer.2.k_cache 0.01027601 6.68472516 + layer.2.v_cache 0.00002197 0.05528181 + layer.3.k_cache 0.07329086 23.21975821 + layer.3.v_cache 0.00001983 0.06516788 + layer.4.k_cache 0.00064477 1.71126443 + layer.4.v_cache 0.00005260 0.13715507 + layer.4.output 10.56448284 498.36375661 + ------------------------------------------------------------------------------------- + TOTAL 4.36618028 240.38210190 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 10416 +BPFP 0.0886 bits/point +EBPFP 0.1773 equivalent bits/point +MSE 240.382102 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 240.3821 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 628B, BPFP=0.1182 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 696B, BPFP=0.1310 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 796B, BPFP=0.1498 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 660B, BPFP=0.1242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1386 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 656B, BPFP=0.1235 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1303 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 680B, BPFP=0.1280 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 736B, BPFP=0.1386 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,216B, BPFP=0.0596 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10139442 528.96056099 + layer.0.v_cache 0.00001346 0.04703884 + layer.1.k_cache 0.03266963 37.39775508 + layer.1.v_cache 0.00000526 0.01756671 + layer.2.k_cache 0.01283502 6.70674152 + layer.2.v_cache 0.00001884 0.05282808 + layer.3.k_cache 0.02666297 23.26098192 + layer.3.v_cache 0.00001780 0.05935428 + layer.4.k_cache 0.00061512 1.64173944 + layer.4.v_cache 0.00004947 0.13319020 + layer.4.output 0.18393325 654.83379948 + ------------------------------------------------------------------------------------- + TOTAL 0.08598910 304.83025609 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9268 +BPFP 0.1026 bits/point +EBPFP 0.2053 equivalent bits/point +MSE 304.830256 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 304.8303 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 700B, BPFP=0.1083 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 788B, BPFP=0.1219 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 828B, BPFP=0.1281 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 700B, BPFP=0.1083 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1188 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 728B, BPFP=0.1126 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 752B, BPFP=0.1163 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 748B, BPFP=0.1157 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 856B, BPFP=0.1324 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 824B, BPFP=0.1275 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,160B, BPFP=0.0477 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11149573 520.47926980 + layer.0.v_cache 0.00001357 0.04828700 + layer.1.k_cache 0.08092193 37.09024018 + layer.1.v_cache 0.00000571 0.01809180 + layer.2.k_cache 0.01109533 6.66861294 + layer.2.v_cache 0.00001920 0.05500251 + layer.3.k_cache 0.03617269 23.18561456 + layer.3.v_cache 0.00001858 0.06296653 + layer.4.k_cache 0.00062559 1.65889974 + layer.4.v_cache 0.00005684 0.13412094 + layer.4.output 11.30219499 532.94099187 + ------------------------------------------------------------------------------------- + TOTAL 4.66798765 254.11694406 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 9852 +BPFP 0.0897 bits/point +EBPFP 0.1793 equivalent bits/point +MSE 254.116944 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.005s, Pack+Encode: 0.155s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 254.1169 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1266 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 668B, BPFP=0.1338 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 796B, BPFP=0.1595 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 632B, BPFP=0.1266 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 712B, BPFP=0.1426 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 664B, BPFP=0.1330 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 648B, BPFP=0.1298 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 776B, BPFP=0.1554 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 724B, BPFP=0.1450 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,616B, BPFP=0.0749 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10167113 539.25150240 + layer.0.v_cache 0.00001386 0.04953837 + layer.1.k_cache 0.03545215 38.21045548 + layer.1.v_cache 0.00000569 0.01860239 + layer.2.k_cache 0.01047806 6.90041645 + layer.2.v_cache 0.00001755 0.05416305 + layer.3.k_cache 0.06687863 23.53961651 + layer.3.v_cache 0.00001960 0.06376310 + layer.4.k_cache 0.00059181 1.70045862 + layer.4.v_cache 0.00004882 0.13719219 + layer.4.output 0.19633365 696.41145833 + ------------------------------------------------------------------------------------- + TOTAL 0.09350076 322.63564217 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9552 +BPFP 0.1126 bits/point +EBPFP 0.2251 equivalent bits/point +MSE 322.635642 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 322.6356 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1379 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 720B, BPFP=0.1293 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 772B, BPFP=0.1386 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 736B, BPFP=0.1322 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 712B, BPFP=0.1279 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 760B, BPFP=0.1365 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 832B, BPFP=0.1494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 808B, BPFP=0.1451 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,472B, BPFP=0.0634 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11445767 528.28825431 + layer.0.v_cache 0.00001583 0.04853084 + layer.1.k_cache 0.03092440 37.40491985 + layer.1.v_cache 0.00000551 0.01774107 + layer.2.k_cache 0.00639992 6.73266321 + layer.2.v_cache 0.00001897 0.05465039 + layer.3.k_cache 0.05613507 23.42611800 + layer.3.v_cache 0.00002026 0.06424191 + layer.4.k_cache 0.00063356 1.63907632 + layer.4.v_cache 0.00005123 0.13603950 + layer.4.output 0.16055210 628.66646141 + ------------------------------------------------------------------------------------- + TOTAL 0.07838395 294.02808619 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 10064 +BPFP 0.1063 bits/point +EBPFP 0.2126 equivalent bits/point +MSE 294.028086 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.003s, Pack+Encode: 0.155s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 294.0281 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 700B, BPFP=0.1302 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 824B, BPFP=0.1533 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 672B, BPFP=0.1250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 752B, BPFP=0.1399 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 688B, BPFP=0.1280 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1280 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 664B, BPFP=0.1235 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1466 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 780B, BPFP=0.1451 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,156B, BPFP=0.0573 +⌛️ [2/4] FRONTEND: Frontend time: 0.153s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13956541 527.34421503 + layer.0.v_cache 0.00001349 0.04768429 + layer.1.k_cache 0.03336124 37.35347203 + layer.1.v_cache 0.00000558 0.01800419 + layer.2.k_cache 0.00698305 6.60512070 + layer.2.v_cache 0.00001831 0.05197777 + layer.3.k_cache 0.07289400 23.41263108 + layer.3.v_cache 0.00001855 0.05917513 + layer.4.k_cache 0.00061006 1.68956230 + layer.4.v_cache 0.00005429 0.13661469 + layer.4.output 0.17891866 647.92346939 + ------------------------------------------------------------------------------------- + TOTAL 0.08858556 301.89310253 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 9352 +BPFP 0.1023 bits/point +EBPFP 0.2047 equivalent bits/point +MSE 301.893103 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.005s, Pack+Encode: 0.153s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 301.8931 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 692B, BPFP=0.1335 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1543 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 644B, BPFP=0.1242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1381 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 668B, BPFP=0.1289 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1312 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 692B, BPFP=0.1335 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 736B, BPFP=0.1420 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,364B, BPFP=0.0651 +⌛️ [2/4] FRONTEND: Frontend time: 0.153s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10837562 534.94627701 + layer.0.v_cache 0.00001644 0.04786722 + layer.1.k_cache 0.01241452 37.43124879 + layer.1.v_cache 0.00000532 0.01756351 + layer.2.k_cache 0.01724591 6.73455961 + layer.2.v_cache 0.00001958 0.05313876 + layer.3.k_cache 0.04641535 23.47397642 + layer.3.v_cache 0.00001847 0.05974046 + layer.4.k_cache 0.00063678 1.66406344 + layer.4.v_cache 0.00004974 0.13436737 + layer.4.output 0.18014439 670.25760582 + ------------------------------------------------------------------------------------- + TOTAL 0.08507109 311.55094372 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 9396 +BPFP 0.1066 bits/point +EBPFP 0.2132 equivalent bits/point +MSE 311.550944 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.153s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 311.5509 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.1031 bits/point +Avg EBPFP 0.2062 equivalent bits/point +Avg MSE 295.623516 +Avg Time 0.374s +------------------------ ---------------------------- diff --git a/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..dd33f182dd0edb2597bbf5919b36f62e36fdfe45 --- /dev/null +++ b/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 405 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean +Output output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,200B, BPFP=0.0856 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,456B, BPFP=0.1039 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,488B, BPFP=0.1062 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,228B, BPFP=0.0876 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,408B, BPFP=0.1005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,264B, BPFP=0.0902 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,364B, BPFP=0.0973 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,284B, BPFP=0.0916 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,552B, BPFP=0.1107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,472B, BPFP=0.1050 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,420B, BPFP=0.0349 +⌛️ [2/4] FRONTEND: Frontend time: 0.673s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.356s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12261864 508.71989155 + layer.0.v_cache 0.00001806 0.05460194 + layer.1.k_cache 0.30198812 35.15982627 + layer.1.v_cache 0.00000607 0.01918016 + layer.2.k_cache 0.01966049 6.35249001 + layer.2.v_cache 0.00002021 0.05831191 + layer.3.k_cache 0.01413035 22.57413403 + layer.3.v_cache 0.00002054 0.06704202 + layer.4.k_cache 0.00067868 1.79109087 + layer.4.v_cache 0.00004987 0.13688860 + layer.4.output 1.39792533 248.05638454 + ------------------------------------------------------------------------------------- + TOTAL 0.60262755 135.96047936 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 17136 +BPFP 0.0719 bits/point +EBPFP 0.1438 equivalent bits/point +MSE 135.960479 +---------------------- -------------------------------------------------------- +Time: 1.038s Load: 0.009s, Pack+Encode: 0.673s, Decode+Unpack: 0.356s +---------------------- -------------------------------------------------------- +💾 Converting with 135.9605 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,304B, BPFP=0.0943 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,420B, BPFP=0.1027 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,540B, BPFP=0.1114 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,288B, BPFP=0.0932 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,452B, BPFP=0.1050 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,332B, BPFP=0.0964 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,396B, BPFP=0.1010 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,320B, BPFP=0.0955 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,612B, BPFP=0.1166 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,436B, BPFP=0.1039 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,564B, BPFP=0.0368 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11787021 510.70059317 + layer.0.v_cache 0.00001738 0.05132079 + layer.1.k_cache 0.25265321 35.30089428 + layer.1.v_cache 0.00000586 0.01863935 + layer.2.k_cache 0.01010228 6.25428433 + layer.2.v_cache 0.00002017 0.05811278 + layer.3.k_cache 0.01159736 22.66441515 + layer.3.v_cache 0.00002197 0.06939987 + layer.4.k_cache 0.00067244 1.81322338 + layer.4.v_cache 0.00004845 0.13361864 + layer.4.output 1.41728832 251.82864170 + ------------------------------------------------------------------------------------- + TOTAL 0.60670750 137.63911727 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 17664 +BPFP 0.0752 bits/point +EBPFP 0.1503 equivalent bits/point +MSE 137.639117 +---------------------- -------------------------------------------------------- +Time: 0.556s Load: 0.009s, Pack+Encode: 0.230s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 137.6391 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,204B, BPFP=0.0844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,388B, BPFP=0.0973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,500B, BPFP=0.1051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,220B, BPFP=0.0855 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,372B, BPFP=0.0961 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,252B, BPFP=0.0877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,344B, BPFP=0.0942 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,244B, BPFP=0.0872 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,588B, BPFP=0.1113 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,416B, BPFP=0.0992 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,164B, BPFP=0.0317 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.322s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13311632 507.01975897 + layer.0.v_cache 0.00001526 0.05092835 + layer.1.k_cache 0.36474110 35.37765161 + layer.1.v_cache 0.00000573 0.01843770 + layer.2.k_cache 0.01296863 6.34250138 + layer.2.v_cache 0.00001984 0.05634215 + layer.3.k_cache 0.01490937 22.49179337 + layer.3.v_cache 0.00002017 0.06765056 + layer.4.k_cache 0.00067370 1.78314127 + layer.4.v_cache 0.00005228 0.13586059 + layer.4.output 1.37278975 242.98322389 + ------------------------------------------------------------------------------------- + TOTAL 0.59623828 133.77803725 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 16692 +BPFP 0.0688 bits/point +EBPFP 0.1376 equivalent bits/point +MSE 133.778037 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.008s, Pack+Encode: 0.224s, Decode+Unpack: 0.322s +---------------------- -------------------------------------------------------- +💾 Converting with 133.7780 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,356B, BPFP=0.0868 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,488B, BPFP=0.0953 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,596B, BPFP=0.1022 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,336B, BPFP=0.0856 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,508B, BPFP=0.0966 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,348B, BPFP=0.0863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,472B, BPFP=0.0943 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,380B, BPFP=0.0884 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,644B, BPFP=0.1053 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,532B, BPFP=0.0981 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,328B, BPFP=0.0304 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11263794 508.10021773 + layer.0.v_cache 0.00001639 0.05319890 + layer.1.k_cache 0.42101050 35.23094102 + layer.1.v_cache 0.00000577 0.01928326 + layer.2.k_cache 0.00499437 6.38324425 + layer.2.v_cache 0.00001973 0.05822881 + layer.3.k_cache 0.03029772 22.78996502 + layer.3.v_cache 0.00002073 0.07031094 + layer.4.k_cache 0.00068836 1.84393123 + layer.4.v_cache 0.00005112 0.13962672 + layer.4.output 1.25471843 223.57098946 + ------------------------------------------------------------------------------------- + TOTAL 0.55016304 125.86387495 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 17988 +BPFP 0.0678 bits/point +EBPFP 0.1355 equivalent bits/point +MSE 125.863875 +---------------------- -------------------------------------------------------- +Time: 0.556s Load: 0.010s, Pack+Encode: 0.225s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 125.8639 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,208B, BPFP=0.0866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,368B, BPFP=0.0981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,500B, BPFP=0.1075 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,248B, BPFP=0.0894 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,424B, BPFP=0.1021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,288B, BPFP=0.0923 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,384B, BPFP=0.0992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,304B, BPFP=0.0935 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,540B, BPFP=0.1104 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,440B, BPFP=0.1032 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,480B, BPFP=0.0356 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10233516 507.64012328 + layer.0.v_cache 0.00001822 0.05440820 + layer.1.k_cache 0.31275415 35.11112430 + layer.1.v_cache 0.00000587 0.01921996 + layer.2.k_cache 0.01225604 6.37209495 + layer.2.v_cache 0.00002018 0.05926424 + layer.3.k_cache 0.02847941 22.75994929 + layer.3.v_cache 0.00002101 0.07007836 + layer.4.k_cache 0.00067306 1.81856600 + layer.4.v_cache 0.00005947 0.14099849 + layer.4.output 1.40430263 249.51456012 + ------------------------------------------------------------------------------------- + TOTAL 0.60510241 136.50869106 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 17184 +BPFP 0.0725 bits/point +EBPFP 0.1449 equivalent bits/point +MSE 136.508691 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.008s, Pack+Encode: 0.227s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 136.5087 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,560B, BPFP=0.0829 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,728B, BPFP=0.0918 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,904B, BPFP=0.1012 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,564B, BPFP=0.0831 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,792B, BPFP=0.0952 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,636B, BPFP=0.0869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,704B, BPFP=0.0906 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,684B, BPFP=0.0895 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,988B, BPFP=0.1057 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,780B, BPFP=0.0946 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,040B, BPFP=0.0307 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.435s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16298727 503.78518282 + layer.0.v_cache 0.00001770 0.05488147 + layer.1.k_cache 0.61736183 35.16384991 + layer.1.v_cache 0.00000612 0.02003617 + layer.2.k_cache 0.02593859 6.33759085 + layer.2.v_cache 0.00002024 0.05873827 + layer.3.k_cache 0.01326758 22.60336947 + layer.3.v_cache 0.00002270 0.07301675 + layer.4.k_cache 0.00075427 1.86903101 + layer.4.v_cache 0.00005043 0.13718704 + layer.4.output 0.04538776 184.91080539 + ------------------------------------------------------------------------------------- + TOTAL 0.06694947 109.67520715 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 21380 +BPFP 0.0668 bits/point +EBPFP 0.1337 equivalent bits/point +MSE 109.675207 +---------------------- -------------------------------------------------------- +Time: 0.764s Load: 0.010s, Pack+Encode: 0.319s, Decode+Unpack: 0.435s +---------------------- -------------------------------------------------------- +💾 Converting with 109.6752 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 239, 128) +Output shape: (1, 239, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.output: torch.Size([1, 239, 3584]) -> torch.Size([1, 1, 239, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,324B, BPFP=0.0866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,464B, BPFP=0.0957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,596B, BPFP=0.1043 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,376B, BPFP=0.0900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,536B, BPFP=0.1004 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,400B, BPFP=0.0915 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,456B, BPFP=0.0952 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,416B, BPFP=0.0926 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,640B, BPFP=0.1072 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,496B, BPFP=0.0978 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,672B, BPFP=0.0343 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11863919 507.28415926 + layer.0.v_cache 0.00001702 0.05463723 + layer.1.k_cache 0.43747612 35.36661546 + layer.1.v_cache 0.00000797 0.02037130 + layer.2.k_cache 0.02286501 6.38418643 + layer.2.v_cache 0.00002174 0.05925229 + layer.3.k_cache 0.01903599 22.96396117 + layer.3.v_cache 0.00002123 0.07138591 + layer.4.k_cache 0.00068597 1.84786387 + layer.4.v_cache 0.00006921 0.13842448 + layer.4.output 1.28101462 227.36816348 + ------------------------------------------------------------------------------------- + TOTAL 0.56270246 127.39811775 + (elements=2,080,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2080256 +Total Bytes 18376 +BPFP 0.0707 bits/point +EBPFP 0.1413 equivalent bits/point +MSE 127.398118 +---------------------- -------------------------------------------------------- +Time: 0.537s Load: 0.009s, Pack+Encode: 0.221s, Decode+Unpack: 0.307s +---------------------- -------------------------------------------------------- +💾 Converting with 127.3981 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,556B, BPFP=0.0924 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,704B, BPFP=0.1012 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,864B, BPFP=0.1107 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,568B, BPFP=0.0932 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.1053 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,652B, BPFP=0.0981 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,696B, BPFP=0.1008 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,636B, BPFP=0.0972 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,996B, BPFP=0.1186 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,832B, BPFP=0.1088 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,560B, BPFP=0.0387 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15151563 504.32254040 + layer.0.v_cache 0.00001631 0.05422511 + layer.1.k_cache 0.46755480 35.20709809 + layer.1.v_cache 0.00000667 0.02002267 + layer.2.k_cache 0.02171374 6.24180644 + layer.2.v_cache 0.00002115 0.05934519 + layer.3.k_cache 0.01385514 22.68281399 + layer.3.v_cache 0.00002107 0.07161657 + layer.4.k_cache 0.00069081 1.82417907 + layer.4.v_cache 0.00005638 0.14387477 + layer.4.output 0.00504555 211.43132129 + ------------------------------------------------------------------------------------- + TOTAL 0.04063356 120.62628067 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 21836 +BPFP 0.0763 bits/point +EBPFP 0.1526 equivalent bits/point +MSE 120.626281 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.369s +---------------------- -------------------------------------------------------- +💾 Converting with 120.6263 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,200B, BPFP=0.0833 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,384B, BPFP=0.0961 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,504B, BPFP=0.1044 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,252B, BPFP=0.0869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,404B, BPFP=0.0975 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,300B, BPFP=0.0903 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,356B, BPFP=0.0942 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,312B, BPFP=0.0911 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,588B, BPFP=0.1103 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,456B, BPFP=0.1011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,328B, BPFP=0.0330 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14529029 503.87118056 + layer.0.v_cache 0.00001764 0.05326634 + layer.1.k_cache 0.31089128 35.40442708 + layer.1.v_cache 0.00000630 0.01966471 + layer.2.k_cache 0.01355181 6.30118381 + layer.2.v_cache 0.00002032 0.05840846 + layer.3.k_cache 0.02426839 22.48955729 + layer.3.v_cache 0.00002043 0.06982197 + layer.4.k_cache 0.00068925 1.77590875 + layer.4.v_cache 0.00005165 0.13540983 + layer.4.output 1.36066019 242.50859127 + ------------------------------------------------------------------------------------- + TOTAL 0.58937816 133.39640986 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 17084 +BPFP 0.0698 bits/point +EBPFP 0.1396 equivalent bits/point +MSE 133.396410 +---------------------- -------------------------------------------------------- +Time: 0.540s Load: 0.008s, Pack+Encode: 0.221s, Decode+Unpack: 0.311s +---------------------- -------------------------------------------------------- +💾 Converting with 133.3964 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 262, 128) +Output shape: (1, 262, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.output: torch.Size([1, 262, 3584]) -> torch.Size([1, 1, 262, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,544B, BPFP=0.0921 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,712B, BPFP=0.1021 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,884B, BPFP=0.1124 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,540B, BPFP=0.0918 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,760B, BPFP=0.1050 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,568B, BPFP=0.0935 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,696B, BPFP=0.1011 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,592B, BPFP=0.0949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.1183 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,816B, BPFP=0.1083 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,400B, BPFP=0.0375 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12917162 505.03691555 + layer.0.v_cache 0.00001830 0.05338347 + layer.1.k_cache 0.40634074 35.09667969 + layer.1.v_cache 0.00000633 0.01949600 + layer.2.k_cache 0.02363680 6.28940914 + layer.2.v_cache 0.00002248 0.05875804 + layer.3.k_cache 0.02343226 22.66823339 + layer.3.v_cache 0.00002042 0.06989576 + layer.4.k_cache 0.00071170 1.84129776 + layer.4.v_cache 0.00005093 0.13942401 + layer.4.output 0.00505940 212.29926050 + ------------------------------------------------------------------------------------- + TOTAL 0.03640161 121.02166566 + (elements=2,280,448) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2280448 +Total Bytes 21496 +BPFP 0.0754 bits/point +EBPFP 0.1508 equivalent bits/point +MSE 121.021666 +---------------------- -------------------------------------------------------- +Time: 0.626s Load: 0.009s, Pack+Encode: 0.248s, Decode+Unpack: 0.369s +---------------------- -------------------------------------------------------- +💾 Converting with 121.0217 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,192B, BPFP=0.0828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,336B, BPFP=0.0928 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,500B, BPFP=0.1042 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,240B, BPFP=0.0861 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,388B, BPFP=0.0964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,292B, BPFP=0.0897 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,368B, BPFP=0.0950 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,292B, BPFP=0.0897 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,568B, BPFP=0.1089 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,448B, BPFP=0.1006 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,324B, BPFP=0.0330 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18074451 508.04250000 + layer.0.v_cache 0.00001764 0.05404404 + layer.1.k_cache 0.33421082 35.32955946 + layer.1.v_cache 0.00000655 0.01978881 + layer.2.k_cache 0.01487562 6.31355957 + layer.2.v_cache 0.00002082 0.06001269 + layer.3.k_cache 0.01319740 22.47611111 + layer.3.v_cache 0.00002096 0.07148293 + layer.4.k_cache 0.00067848 1.83132338 + layer.4.v_cache 0.00005239 0.13945712 + layer.4.output 1.36066546 242.58101190 + ------------------------------------------------------------------------------------- + TOTAL 0.59226373 133.67087779 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 16948 +BPFP 0.0692 bits/point +EBPFP 0.1385 equivalent bits/point +MSE 133.670878 +---------------------- -------------------------------------------------------- +Time: 0.540s Load: 0.009s, Pack+Encode: 0.220s, Decode+Unpack: 0.311s +---------------------- -------------------------------------------------------- +💾 Converting with 133.6709 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 232, 128) +Output shape: (1, 232, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.output: torch.Size([1, 232, 3584]) -> torch.Size([1, 1, 232, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.0894 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,444B, BPFP=0.0973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,596B, BPFP=0.1075 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,348B, BPFP=0.0908 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,520B, BPFP=0.1024 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,396B, BPFP=0.0940 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,440B, BPFP=0.0970 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,420B, BPFP=0.0956 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,644B, BPFP=0.1107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,504B, BPFP=0.1013 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,700B, BPFP=0.0356 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.310s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11978169 506.07145744 + layer.0.v_cache 0.00001746 0.05371362 + layer.1.k_cache 0.32299256 35.32443763 + layer.1.v_cache 0.00000596 0.01891100 + layer.2.k_cache 0.02091339 6.27235676 + layer.2.v_cache 0.00002092 0.05982096 + layer.3.k_cache 0.01769928 22.65817366 + layer.3.v_cache 0.00002180 0.07127414 + layer.4.k_cache 0.00071208 1.85819823 + layer.4.v_cache 0.00005763 0.14167343 + layer.4.output 1.31963615 233.61578279 + ------------------------------------------------------------------------------------- + TOTAL 0.57174564 129.87297038 + (elements=2,019,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2019328 +Total Bytes 18340 +BPFP 0.0727 bits/point +EBPFP 0.1453 equivalent bits/point +MSE 129.872970 +---------------------- -------------------------------------------------------- +Time: 0.539s Load: 0.009s, Pack+Encode: 0.220s, Decode+Unpack: 0.310s +---------------------- -------------------------------------------------------- +💾 Converting with 129.8730 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,188B, BPFP=0.0844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,352B, BPFP=0.0960 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,476B, BPFP=0.1048 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,208B, BPFP=0.0858 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,392B, BPFP=0.0989 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,240B, BPFP=0.0881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,332B, BPFP=0.0946 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,272B, BPFP=0.0903 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,552B, BPFP=0.1102 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,432B, BPFP=0.1017 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,388B, BPFP=0.0344 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15639258 507.88448153 + layer.0.v_cache 0.00001700 0.05510403 + layer.1.k_cache 0.33150226 35.18063743 + layer.1.v_cache 0.00000613 0.01958464 + layer.2.k_cache 0.01694729 6.19276789 + layer.2.v_cache 0.00002208 0.05840217 + layer.3.k_cache 0.03858858 22.68453924 + layer.3.v_cache 0.00002090 0.07226743 + layer.4.k_cache 0.00067984 1.77079496 + layer.4.v_cache 0.00005217 0.14247082 + layer.4.output 1.39161012 246.37611607 + ------------------------------------------------------------------------------------- + TOTAL 0.60502939 135.21728604 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 16832 +BPFP 0.0703 bits/point +EBPFP 0.1406 equivalent bits/point +MSE 135.217286 +---------------------- -------------------------------------------------------- +Time: 0.539s Load: 0.007s, Pack+Encode: 0.221s, Decode+Unpack: 0.311s +---------------------- -------------------------------------------------------- +💾 Converting with 135.2173 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,516B, BPFP=0.0868 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,736B, BPFP=0.0994 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1092 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,584B, BPFP=0.0907 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,756B, BPFP=0.1005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,684B, BPFP=0.0964 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,720B, BPFP=0.0984 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,672B, BPFP=0.0957 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,012B, BPFP=0.1152 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,852B, BPFP=0.1060 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,312B, BPFP=0.0353 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13605902 504.00314789 + layer.0.v_cache 0.00001757 0.05369622 + layer.1.k_cache 0.58683352 35.06313673 + layer.1.v_cache 0.00000623 0.01975151 + layer.2.k_cache 0.01500385 6.18772134 + layer.2.v_cache 0.00002229 0.05962594 + layer.3.k_cache 0.00978726 22.51640840 + layer.3.v_cache 0.00001998 0.06738563 + layer.4.k_cache 0.00067570 1.76931651 + layer.4.v_cache 0.00005472 0.13942020 + layer.4.output 0.00490606 203.36332418 + ------------------------------------------------------------------------------------- + TOTAL 0.04604839 117.26016939 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 21752 +BPFP 0.0732 bits/point +EBPFP 0.1465 equivalent bits/point +MSE 117.260169 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.011s, Pack+Encode: 0.250s, Decode+Unpack: 0.368s +---------------------- -------------------------------------------------------- +💾 Converting with 117.2602 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,132B, BPFP=0.0694 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,268B, BPFP=0.0777 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,416B, BPFP=0.0868 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,100B, BPFP=0.0674 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,312B, BPFP=0.0804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,140B, BPFP=0.0699 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,236B, BPFP=0.0757 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,152B, BPFP=0.0706 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,488B, BPFP=0.0912 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,332B, BPFP=0.0816 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,576B, BPFP=0.0225 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12967140 503.32549020 + layer.0.v_cache 0.00001601 0.05454301 + layer.1.k_cache 0.49353925 34.26547181 + layer.1.v_cache 0.00000620 0.01945184 + layer.2.k_cache 0.03189284 6.08853114 + layer.2.v_cache 0.00002040 0.05785626 + layer.3.k_cache 0.04285583 22.25701402 + layer.3.v_cache 0.00002221 0.07038561 + layer.4.k_cache 0.00071132 1.75053687 + layer.4.v_cache 0.00005107 0.13722658 + layer.4.output 1.20063613 212.50343137 + ------------------------------------------------------------------------------------- + TOTAL 0.53548467 120.91473688 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 15152 +BPFP 0.0546 bits/point +EBPFP 0.1092 equivalent bits/point +MSE 120.914737 +---------------------- -------------------------------------------------------- +Time: 0.545s Load: 0.010s, Pack+Encode: 0.223s, Decode+Unpack: 0.312s +---------------------- -------------------------------------------------------- +💾 Converting with 120.9147 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,300B, BPFP=0.0816 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,432B, BPFP=0.0899 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,552B, BPFP=0.0974 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,304B, BPFP=0.0818 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,472B, BPFP=0.0924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,364B, BPFP=0.0856 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,416B, BPFP=0.0889 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,376B, BPFP=0.0863 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,600B, BPFP=0.1004 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,488B, BPFP=0.0934 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,188B, BPFP=0.0286 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16529650 510.11263805 + layer.0.v_cache 0.00001613 0.05473332 + layer.1.k_cache 0.54577502 35.33621596 + layer.1.v_cache 0.00000662 0.02017713 + layer.2.k_cache 0.01453520 6.38194967 + layer.2.v_cache 0.00002096 0.05962648 + layer.3.k_cache 0.02409829 22.81147245 + layer.3.v_cache 0.00002071 0.07068730 + layer.4.k_cache 0.00068699 1.84037444 + layer.4.v_cache 0.00004971 0.13555305 + layer.4.output 1.22953407 223.93798408 + ------------------------------------------------------------------------------------- + TOTAL 0.55042615 126.14054802 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 17492 +BPFP 0.0646 bits/point +EBPFP 0.1291 equivalent bits/point +MSE 126.140548 +---------------------- -------------------------------------------------------- +Time: 0.537s Load: 0.009s, Pack+Encode: 0.221s, Decode+Unpack: 0.307s +---------------------- -------------------------------------------------------- +💾 Converting with 126.1405 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.0883 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,048B, BPFP=0.0979 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,216B, BPFP=0.1059 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,848B, BPFP=0.0883 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,068B, BPFP=0.0988 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,940B, BPFP=0.0927 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,996B, BPFP=0.0954 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,908B, BPFP=0.0912 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,360B, BPFP=0.1128 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,108B, BPFP=0.1007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,160B, BPFP=0.0352 +⌛️ [2/4] FRONTEND: Frontend time: 0.531s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.498s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19043710 505.84656919 + layer.0.v_cache 0.00001603 0.05426883 + layer.1.k_cache 0.64205214 34.99754814 + layer.1.v_cache 0.00000630 0.02023385 + layer.2.k_cache 0.01513053 6.27353456 + layer.2.v_cache 0.00002191 0.06089664 + layer.3.k_cache 0.02775778 22.43863335 + layer.3.v_cache 0.00002125 0.07148679 + layer.4.k_cache 0.00071623 1.82310798 + layer.4.v_cache 0.00005146 0.13773442 + layer.4.output 0.04089398 166.25993884 + ------------------------------------------------------------------------------------- + TOTAL 0.06838051 102.09079915 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 25500 +BPFP 0.0717 bits/point +EBPFP 0.1433 equivalent bits/point +MSE 102.090799 +---------------------- -------------------------------------------------------- +Time: 1.040s Load: 0.012s, Pack+Encode: 0.531s, Decode+Unpack: 0.498s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0908 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,484B, BPFP=0.0819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,680B, BPFP=0.0928 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,832B, BPFP=0.1011 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,496B, BPFP=0.0826 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,724B, BPFP=0.0952 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,552B, BPFP=0.0857 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,668B, BPFP=0.0921 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,564B, BPFP=0.0864 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,932B, BPFP=0.1067 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,708B, BPFP=0.0943 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,924B, BPFP=0.0310 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14392723 504.90729903 + layer.0.v_cache 0.00001682 0.05481610 + layer.1.k_cache 0.53349789 34.85363295 + layer.1.v_cache 0.00000638 0.02023966 + layer.2.k_cache 0.01523476 6.28232189 + layer.2.v_cache 0.00002109 0.06027131 + layer.3.k_cache 0.03867486 22.32505038 + layer.3.v_cache 0.00001994 0.06987634 + layer.4.k_cache 0.00070525 1.80507916 + layer.4.v_cache 0.00005078 0.13799971 + layer.4.output 0.00472849 196.74028269 + ------------------------------------------------------------------------------------- + TOTAL 0.04501497 114.57050384 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 20564 +BPFP 0.0668 bits/point +EBPFP 0.1336 equivalent bits/point +MSE 114.570504 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.009s, Pack+Encode: 0.248s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 114.5705 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,196B, BPFP=0.0831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,316B, BPFP=0.0914 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,488B, BPFP=0.1033 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,220B, BPFP=0.0847 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,396B, BPFP=0.0969 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,272B, BPFP=0.0883 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,352B, BPFP=0.0939 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,268B, BPFP=0.0881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,552B, BPFP=0.1078 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,388B, BPFP=0.0964 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,284B, BPFP=0.0326 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14673459 509.98118056 + layer.0.v_cache 0.00001655 0.05314040 + layer.1.k_cache 0.37581726 35.37748698 + layer.1.v_cache 0.00000603 0.01928383 + layer.2.k_cache 0.01062915 6.37152669 + layer.2.v_cache 0.00002006 0.05807925 + layer.3.k_cache 0.04027582 22.55832899 + layer.3.v_cache 0.00002066 0.06984252 + layer.4.k_cache 0.00066373 1.75245334 + layer.4.v_cache 0.00004861 0.13483181 + layer.4.output 1.36064577 242.77513889 + ------------------------------------------------------------------------------------- + TOTAL 0.59404429 133.87071333 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 16732 +BPFP 0.0683 bits/point +EBPFP 0.1367 equivalent bits/point +MSE 133.870713 +---------------------- -------------------------------------------------------- +Time: 0.566s Load: 0.009s, Pack+Encode: 0.230s, Decode+Unpack: 0.327s +---------------------- -------------------------------------------------------- +💾 Converting with 133.8707 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,332B, BPFP=0.0889 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,468B, BPFP=0.0980 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,612B, BPFP=0.1076 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,380B, BPFP=0.0921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,536B, BPFP=0.1026 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,392B, BPFP=0.0929 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,472B, BPFP=0.0983 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,392B, BPFP=0.0929 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,680B, BPFP=0.1122 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,500B, BPFP=0.1002 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,628B, BPFP=0.0346 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12385076 505.39139290 + layer.0.v_cache 0.00002004 0.05755862 + layer.1.k_cache 0.38583452 35.49054320 + layer.1.v_cache 0.00000646 0.02072085 + layer.2.k_cache 0.02500911 6.43678949 + layer.2.v_cache 0.00002014 0.06028625 + layer.3.k_cache 0.03425702 22.76996528 + layer.3.v_cache 0.00002057 0.07348889 + layer.4.k_cache 0.00068928 1.90512672 + layer.4.v_cache 0.00004882 0.13898892 + layer.4.output 1.30836928 232.13076160 + ------------------------------------------------------------------------------------- + TOTAL 0.57225539 129.25059955 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 18392 +BPFP 0.0722 bits/point +EBPFP 0.1445 equivalent bits/point +MSE 129.250600 +---------------------- -------------------------------------------------------- +Time: 0.557s Load: 0.008s, Pack+Encode: 0.227s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 129.2506 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 247, 128) +Output shape: (1, 247, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.output: torch.Size([1, 247, 3584]) -> torch.Size([1, 1, 247, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,336B, BPFP=0.0845 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,416B, BPFP=0.0896 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,568B, BPFP=0.0992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,348B, BPFP=0.0853 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,504B, BPFP=0.0951 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,360B, BPFP=0.0860 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,444B, BPFP=0.0913 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,392B, BPFP=0.0881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,632B, BPFP=0.1032 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,496B, BPFP=0.0946 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,316B, BPFP=0.0300 +⌛️ [2/4] FRONTEND: Frontend time: 0.234s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13907220 510.52963689 + layer.0.v_cache 0.00001735 0.05537356 + layer.1.k_cache 0.47208408 35.58408717 + layer.1.v_cache 0.00000611 0.01998109 + layer.2.k_cache 0.03408500 6.43291075 + layer.2.v_cache 0.00001989 0.05858397 + layer.3.k_cache 0.02387486 22.98957806 + layer.3.v_cache 0.00002031 0.06875668 + layer.4.k_cache 0.00069623 1.88445748 + layer.4.v_cache 0.00004922 0.13313752 + layer.4.output 1.23951819 226.41512435 + ------------------------------------------------------------------------------------- + TOTAL 0.54979721 127.21543374 + (elements=2,149,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2149888 +Total Bytes 17812 +BPFP 0.0663 bits/point +EBPFP 0.1326 equivalent bits/point +MSE 127.215434 +---------------------- -------------------------------------------------------- +Time: 0.591s Load: 0.008s, Pack+Encode: 0.234s, Decode+Unpack: 0.349s +---------------------- -------------------------------------------------------- +💾 Converting with 127.2154 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 427, 128) +Output shape: (1, 427, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.output: torch.Size([1, 427, 3584]) -> torch.Size([1, 1, 427, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,168B, BPFP=0.0793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,400B, BPFP=0.0878 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,648B, BPFP=0.0969 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,172B, BPFP=0.0795 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,492B, BPFP=0.0912 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,236B, BPFP=0.0818 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,360B, BPFP=0.0864 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,288B, BPFP=0.0837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,760B, BPFP=0.1010 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,492B, BPFP=0.0912 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,396B, BPFP=0.0282 +⌛️ [2/4] FRONTEND: Frontend time: 0.384s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.499s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15738957 504.63487998 + layer.0.v_cache 0.00001693 0.05347323 + layer.1.k_cache 1.00627419 34.90934207 + layer.1.v_cache 0.00000663 0.01986382 + layer.2.k_cache 0.03590691 6.21639550 + layer.2.v_cache 0.00002108 0.05659695 + layer.3.k_cache 0.02176077 22.64030024 + layer.3.v_cache 0.00002004 0.06693876 + layer.4.k_cache 0.00078997 1.83678828 + layer.4.v_cache 0.00005212 0.13565064 + layer.4.output 0.00621442 130.21187061 + ------------------------------------------------------------------------------------- + TOTAL 0.07445524 87.17960728 + (elements=3,716,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3716608 +Total Bytes 29412 +BPFP 0.0633 bits/point +EBPFP 0.1266 equivalent bits/point +MSE 87.179607 +---------------------- -------------------------------------------------------- +Time: 0.898s Load: 0.016s, Pack+Encode: 0.384s, Decode+Unpack: 0.499s +---------------------- -------------------------------------------------------- +💾 Converting with 87.1796 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.017s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,508B, BPFP=0.0869 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,756B, BPFP=0.1012 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,916B, BPFP=0.1105 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,616B, BPFP=0.0932 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,752B, BPFP=0.1010 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,672B, BPFP=0.0964 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,728B, BPFP=0.0996 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,712B, BPFP=0.0987 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,012B, BPFP=0.1160 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,872B, BPFP=0.1079 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,316B, BPFP=0.0355 +⌛️ [2/4] FRONTEND: Frontend time: 0.274s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15504150 510.65688423 + layer.0.v_cache 0.00001776 0.05316159 + layer.1.k_cache 0.59434689 35.24013348 + layer.1.v_cache 0.00000612 0.01929424 + layer.2.k_cache 0.01687181 6.28050091 + layer.2.v_cache 0.00002031 0.05651153 + layer.3.k_cache 0.01346795 22.67284976 + layer.3.v_cache 0.00002023 0.06808205 + layer.4.k_cache 0.00071492 1.82064324 + layer.4.v_cache 0.00006004 0.13174049 + layer.4.output 0.00488892 204.96113930 + ------------------------------------------------------------------------------------- + TOTAL 0.04792882 118.33692804 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 21860 +BPFP 0.0741 bits/point +EBPFP 0.1483 equivalent bits/point +MSE 118.336928 +---------------------- -------------------------------------------------------- +Time: 0.674s Load: 0.017s, Pack+Encode: 0.274s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 118.3369 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 361, 128) +Output shape: (1, 361, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.output: torch.Size([1, 361, 3584]) -> torch.Size([1, 1, 361, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,884B, BPFP=0.0815 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,124B, BPFP=0.0919 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,296B, BPFP=0.0994 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,920B, BPFP=0.0831 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,160B, BPFP=0.0935 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,008B, BPFP=0.0869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,052B, BPFP=0.0888 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,024B, BPFP=0.0876 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,420B, BPFP=0.1047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,244B, BPFP=0.0971 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,804B, BPFP=0.0297 +⌛️ [2/4] FRONTEND: Frontend time: 0.301s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.441s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21057716 506.72740651 + layer.0.v_cache 0.00001554 0.05073732 + layer.1.k_cache 0.82271198 34.93577411 + layer.1.v_cache 0.00000592 0.01849169 + layer.2.k_cache 0.01184860 6.10810628 + layer.2.v_cache 0.00001987 0.05630031 + layer.3.k_cache 0.01345186 22.31998788 + layer.3.v_cache 0.00002111 0.06630959 + layer.4.k_cache 0.00076976 1.81205923 + layer.4.v_cache 0.00005077 0.12986369 + layer.4.output 0.03703259 150.39367333 + ------------------------------------------------------------------------------------- + TOTAL 0.07757063 95.58710294 + (elements=3,142,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3142144 +Total Bytes 25936 +BPFP 0.0660 bits/point +EBPFP 0.1321 equivalent bits/point +MSE 95.587103 +---------------------- -------------------------------------------------------- +Time: 0.755s Load: 0.014s, Pack+Encode: 0.301s, Decode+Unpack: 0.441s +---------------------- -------------------------------------------------------- +💾 Converting with 95.5871 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,496B, BPFP=0.0829 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,768B, BPFP=0.0980 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,840B, BPFP=0.1020 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,548B, BPFP=0.0858 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,716B, BPFP=0.0951 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,588B, BPFP=0.0880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,672B, BPFP=0.0926 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,604B, BPFP=0.0889 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,928B, BPFP=0.1068 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,792B, BPFP=0.0993 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,060B, BPFP=0.0321 +⌛️ [2/4] FRONTEND: Frontend time: 0.284s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21757852 507.76235594 + layer.0.v_cache 0.00001544 0.04884241 + layer.1.k_cache 0.59705023 35.01474886 + layer.1.v_cache 0.00000603 0.01829365 + layer.2.k_cache 0.00787334 6.15068411 + layer.2.v_cache 0.00001970 0.05333207 + layer.3.k_cache 0.01914395 22.46427236 + layer.3.v_cache 0.00001916 0.06177886 + layer.4.k_cache 0.00071861 1.72694754 + layer.4.v_cache 0.00005003 0.12668033 + layer.4.output 0.00471288 197.76318706 + ------------------------------------------------------------------------------------- + TOTAL 0.05149795 115.16295562 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 21012 +BPFP 0.0685 bits/point +EBPFP 0.1370 equivalent bits/point +MSE 115.162956 +---------------------- -------------------------------------------------------- +Time: 0.661s Load: 0.009s, Pack+Encode: 0.284s, Decode+Unpack: 0.368s +---------------------- -------------------------------------------------------- +💾 Converting with 115.1630 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 372, 128) +Output shape: (1, 372, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.output: torch.Size([1, 372, 3584]) -> torch.Size([1, 1, 372, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,900B, BPFP=0.0798 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,124B, BPFP=0.0892 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,284B, BPFP=0.0959 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,960B, BPFP=0.0823 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,156B, BPFP=0.0906 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,000B, BPFP=0.0840 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,064B, BPFP=0.0867 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,992B, BPFP=0.0837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,376B, BPFP=0.0998 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,200B, BPFP=0.0924 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,972B, BPFP=0.0298 +⌛️ [2/4] FRONTEND: Frontend time: 0.285s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.438s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18192377 506.60055444 + layer.0.v_cache 0.00001786 0.05287864 + layer.1.k_cache 0.82695811 34.91290480 + layer.1.v_cache 0.00000687 0.01968060 + layer.2.k_cache 0.01216373 6.17436103 + layer.2.v_cache 0.00002073 0.05794784 + layer.3.k_cache 0.01743993 22.57324219 + layer.3.v_cache 0.00002103 0.06770840 + layer.4.k_cache 0.00069676 1.78186987 + layer.4.v_cache 0.00005544 0.13202075 + layer.4.output 0.03606073 146.84326997 + ------------------------------------------------------------------------------------- + TOTAL 0.07598408 94.13388578 + (elements=3,237,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3237888 +Total Bytes 26028 +BPFP 0.0643 bits/point +EBPFP 0.1286 equivalent bits/point +MSE 94.133886 +---------------------- -------------------------------------------------------- +Time: 0.737s Load: 0.014s, Pack+Encode: 0.285s, Decode+Unpack: 0.438s +---------------------- -------------------------------------------------------- +💾 Converting with 94.1339 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,824B, BPFP=0.0882 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,036B, BPFP=0.0985 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,192B, BPFP=0.1060 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,796B, BPFP=0.0869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,060B, BPFP=0.0997 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,892B, BPFP=0.0915 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,960B, BPFP=0.0948 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,956B, BPFP=0.0946 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,288B, BPFP=0.1107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,172B, BPFP=0.1051 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,916B, BPFP=0.0340 +⌛️ [2/4] FRONTEND: Frontend time: 0.289s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.434s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16470559 503.50503096 + layer.0.v_cache 0.00001558 0.05241565 + layer.1.k_cache 0.70577214 35.02002709 + layer.1.v_cache 0.00000663 0.01950912 + layer.2.k_cache 0.01487649 6.22565283 + layer.2.v_cache 0.00002061 0.05776759 + layer.3.k_cache 0.03012228 22.39452369 + layer.3.v_cache 0.00002030 0.06811660 + layer.4.k_cache 0.00074575 1.80750268 + layer.4.v_cache 0.00005065 0.13039878 + layer.4.output 0.04139163 169.75352444 + ------------------------------------------------------------------------------------- + TOTAL 0.07094573 103.38562447 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 25092 +BPFP 0.0714 bits/point +EBPFP 0.1428 equivalent bits/point +MSE 103.385624 +---------------------- -------------------------------------------------------- +Time: 0.735s Load: 0.012s, Pack+Encode: 0.289s, Decode+Unpack: 0.434s +---------------------- -------------------------------------------------------- +💾 Converting with 103.3856 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,560B, BPFP=0.0920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,736B, BPFP=0.1024 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,892B, BPFP=0.1116 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,632B, BPFP=0.0962 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,756B, BPFP=0.1035 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,660B, BPFP=0.0979 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,728B, BPFP=0.1019 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,652B, BPFP=0.0974 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.1170 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,836B, BPFP=0.1083 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,508B, BPFP=0.0380 +⌛️ [2/4] FRONTEND: Frontend time: 0.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16585089 504.74027123 + layer.0.v_cache 0.00001549 0.05183730 + layer.1.k_cache 0.52552974 35.11788031 + layer.1.v_cache 0.00000676 0.01914371 + layer.2.k_cache 0.01473994 6.14944723 + layer.2.v_cache 0.00002188 0.05819236 + layer.3.k_cache 0.05129043 22.61173533 + layer.3.v_cache 0.00002012 0.06794729 + layer.4.k_cache 0.00075129 1.81086702 + layer.4.v_cache 0.00005717 0.13190492 + layer.4.output 0.00501330 209.58175539 + ------------------------------------------------------------------------------------- + TOTAL 0.04666922 119.87244203 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 21944 +BPFP 0.0761 bits/point +EBPFP 0.1522 equivalent bits/point +MSE 119.872442 +---------------------- -------------------------------------------------------- +Time: 0.656s Load: 0.013s, Pack+Encode: 0.267s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 119.8724 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 368, 128) +Output shape: (1, 368, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.output: torch.Size([1, 368, 3584]) -> torch.Size([1, 1, 368, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,868B, BPFP=0.0793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,172B, BPFP=0.0922 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,280B, BPFP=0.0968 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,936B, BPFP=0.0822 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,184B, BPFP=0.0927 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,964B, BPFP=0.0834 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,076B, BPFP=0.0881 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,008B, BPFP=0.0853 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,416B, BPFP=0.1026 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,156B, BPFP=0.0915 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,936B, BPFP=0.0299 +⌛️ [2/4] FRONTEND: Frontend time: 0.284s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.426s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14147143 504.81207541 + layer.0.v_cache 0.00001611 0.05406490 + layer.1.k_cache 0.89062865 34.88547416 + layer.1.v_cache 0.00000637 0.01937269 + layer.2.k_cache 0.02201829 6.22281680 + layer.2.v_cache 0.00002088 0.05907976 + layer.3.k_cache 0.00965744 22.63262143 + layer.3.v_cache 0.00002101 0.06937049 + layer.4.k_cache 0.00077426 1.83387259 + layer.4.v_cache 0.00005366 0.14033371 + layer.4.output 0.03638816 147.85192644 + ------------------------------------------------------------------------------------- + TOTAL 0.07761090 94.45250394 + (elements=3,203,072) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3203072 +Total Bytes 25996 +BPFP 0.0649 bits/point +EBPFP 0.1299 equivalent bits/point +MSE 94.452504 +---------------------- -------------------------------------------------------- +Time: 0.723s Load: 0.013s, Pack+Encode: 0.284s, Decode+Unpack: 0.426s +---------------------- -------------------------------------------------------- +💾 Converting with 94.4525 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 453, 128) +Output shape: (1, 453, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.output: torch.Size([1, 453, 3584]) -> torch.Size([1, 1, 453, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,404B, BPFP=0.0829 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,724B, BPFP=0.0940 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,904B, BPFP=0.1002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,412B, BPFP=0.0832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,700B, BPFP=0.0931 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,464B, BPFP=0.0850 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,616B, BPFP=0.0902 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,548B, BPFP=0.0879 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,092B, BPFP=0.1067 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,776B, BPFP=0.0958 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,356B, BPFP=0.0313 +⌛️ [2/4] FRONTEND: Frontend time: 0.449s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.561s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13665763 498.37368929 + layer.0.v_cache 0.00001897 0.05448473 + layer.1.k_cache 1.07282356 34.66389004 + layer.1.v_cache 0.00000630 0.01923993 + layer.2.k_cache 0.02241544 6.21789200 + layer.2.v_cache 0.00001979 0.05709132 + layer.3.k_cache 0.01806638 22.45916762 + layer.3.v_cache 0.00002079 0.06966561 + layer.4.k_cache 0.00076567 1.83318445 + layer.4.v_cache 0.00005104 0.13665734 + layer.4.output 0.00584071 122.50368575 + ------------------------------------------------------------------------------------- + TOTAL 0.07598415 83.61239780 + (elements=3,942,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3942912 +Total Bytes 32996 +BPFP 0.0669 bits/point +EBPFP 0.1339 equivalent bits/point +MSE 83.612398 +---------------------- -------------------------------------------------------- +Time: 1.025s Load: 0.015s, Pack+Encode: 0.449s, Decode+Unpack: 0.561s +---------------------- -------------------------------------------------------- +💾 Converting with 83.6124 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.017s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 402, 128) +Output shape: (1, 402, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.output: torch.Size([1, 402, 3584]) -> torch.Size([1, 1, 402, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,080B, BPFP=0.0808 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,488B, BPFP=0.0967 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,604B, BPFP=0.1012 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,236B, BPFP=0.0869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,444B, BPFP=0.0950 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,268B, BPFP=0.0882 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,340B, BPFP=0.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,272B, BPFP=0.0883 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,784B, BPFP=0.1082 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,600B, BPFP=0.1011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,388B, BPFP=0.0299 +⌛️ [2/4] FRONTEND: Frontend time: 0.335s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.490s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18062694 506.79699160 + layer.0.v_cache 0.00001752 0.05393200 + layer.1.k_cache 0.89169403 34.72012496 + layer.1.v_cache 0.00000593 0.01842651 + layer.2.k_cache 0.03854401 6.12136340 + layer.2.v_cache 0.00001967 0.05450844 + layer.3.k_cache 0.02336847 22.43733967 + layer.3.v_cache 0.00001924 0.06575246 + layer.4.k_cache 0.00085529 1.82008870 + layer.4.v_cache 0.00004959 0.12887624 + layer.4.output 0.00648942 138.04776342 + ------------------------------------------------------------------------------------- + TOTAL 0.06944863 90.50304399 + (elements=3,499,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3499008 +Total Bytes 29504 +BPFP 0.0675 bits/point +EBPFP 0.1349 equivalent bits/point +MSE 90.503044 +---------------------- -------------------------------------------------------- +Time: 0.843s Load: 0.017s, Pack+Encode: 0.335s, Decode+Unpack: 0.490s +---------------------- -------------------------------------------------------- +💾 Converting with 90.5030 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,568B, BPFP=0.0878 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,732B, BPFP=0.0970 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,884B, BPFP=0.1055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,544B, BPFP=0.0865 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,776B, BPFP=0.0995 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,600B, BPFP=0.0896 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,696B, BPFP=0.0950 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,612B, BPFP=0.0903 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,948B, BPFP=0.1091 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,816B, BPFP=0.1017 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,192B, BPFP=0.0335 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13958113 507.89745744 + layer.0.v_cache 0.00001688 0.05413921 + layer.1.k_cache 0.56864935 35.04827509 + layer.1.v_cache 0.00000621 0.01991497 + layer.2.k_cache 0.01018704 6.30409859 + layer.2.v_cache 0.00002046 0.05816228 + layer.3.k_cache 0.04374025 22.53744540 + layer.3.v_cache 0.00002178 0.07080155 + layer.4.k_cache 0.00068741 1.81565895 + layer.4.v_cache 0.00005272 0.13848595 + layer.4.output 0.00479460 200.53227407 + ------------------------------------------------------------------------------------- + TOTAL 0.04685444 116.33355046 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 21368 +BPFP 0.0704 bits/point +EBPFP 0.1408 equivalent bits/point +MSE 116.333550 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 116.3336 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.017s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 381, 128) +Output shape: (1, 381, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.output: torch.Size([1, 381, 3584]) -> torch.Size([1, 1, 381, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,732B, BPFP=0.0710 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,948B, BPFP=0.0799 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,132B, BPFP=0.0874 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,728B, BPFP=0.0709 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,996B, BPFP=0.0819 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,804B, BPFP=0.0740 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,872B, BPFP=0.0768 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,824B, BPFP=0.0748 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,232B, BPFP=0.0915 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,068B, BPFP=0.0848 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,388B, BPFP=0.0257 +⌛️ [2/4] FRONTEND: Frontend time: 0.293s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.425s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15713709 501.47276903 + layer.0.v_cache 0.00001719 0.05425166 + layer.1.k_cache 0.86767426 34.35395136 + layer.1.v_cache 0.00000622 0.01965826 + layer.2.k_cache 0.01635739 6.21267404 + layer.2.v_cache 0.00002098 0.05907425 + layer.3.k_cache 0.01709203 22.25986559 + layer.3.v_cache 0.00002073 0.06981687 + layer.4.k_cache 0.00073405 1.81543545 + layer.4.v_cache 0.00005083 0.13594071 + layer.4.output 0.03519400 142.47184336 + ------------------------------------------------------------------------------------- + TOTAL 0.07679228 91.98566710 + (elements=3,316,224) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3316224 +Total Bytes 23724 +BPFP 0.0572 bits/point +EBPFP 0.1145 equivalent bits/point +MSE 91.985667 +---------------------- -------------------------------------------------------- +Time: 0.735s Load: 0.017s, Pack+Encode: 0.293s, Decode+Unpack: 0.425s +---------------------- -------------------------------------------------------- +💾 Converting with 91.9857 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.0887 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,468B, BPFP=0.0980 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,612B, BPFP=0.1076 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,328B, BPFP=0.0887 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,524B, BPFP=0.1018 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,408B, BPFP=0.0940 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,468B, BPFP=0.0980 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,412B, BPFP=0.0943 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,660B, BPFP=0.1108 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,512B, BPFP=0.1010 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,628B, BPFP=0.0346 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422791 515.06163194 + layer.0.v_cache 0.00001930 0.05497968 + layer.1.k_cache 0.38928963 35.47030248 + layer.1.v_cache 0.00000631 0.01981968 + layer.2.k_cache 0.01623082 6.38434255 + layer.2.v_cache 0.00002095 0.05981377 + layer.3.k_cache 0.01313608 22.84062834 + layer.3.v_cache 0.00002100 0.07011123 + layer.4.k_cache 0.00066994 1.80936269 + layer.4.v_cache 0.00005661 0.13997495 + layer.4.output 1.30836332 232.13438645 + ------------------------------------------------------------------------------------- + TOTAL 0.57013069 129.81480426 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 18348 +BPFP 0.0721 bits/point +EBPFP 0.1441 equivalent bits/point +MSE 129.814804 +---------------------- -------------------------------------------------------- +Time: 0.540s Load: 0.008s, Pack+Encode: 0.223s, Decode+Unpack: 0.309s +---------------------- -------------------------------------------------------- +💾 Converting with 129.8148 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,332B, BPFP=0.0882 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,472B, BPFP=0.0975 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,608B, BPFP=0.1065 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,328B, BPFP=0.0879 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,524B, BPFP=0.1009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,376B, BPFP=0.0911 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,468B, BPFP=0.0972 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,392B, BPFP=0.0922 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,668B, BPFP=0.1104 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,540B, BPFP=0.1020 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,676B, BPFP=0.0348 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13591524 514.21659825 + layer.0.v_cache 0.00001656 0.05359369 + layer.1.k_cache 0.34520291 35.48274878 + layer.1.v_cache 0.00000598 0.01938183 + layer.2.k_cache 0.01250870 6.35729360 + layer.2.v_cache 0.00001982 0.05707039 + layer.3.k_cache 0.02633334 22.65624379 + layer.3.v_cache 0.00002177 0.07122163 + layer.4.k_cache 0.00067446 1.82005944 + layer.4.v_cache 0.00005295 0.13750351 + layer.4.output 1.29725540 229.77608581 + ------------------------------------------------------------------------------------- + TOTAL 0.56479644 128.78260680 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 18384 +BPFP 0.0716 bits/point +EBPFP 0.1432 equivalent bits/point +MSE 128.782607 +---------------------- -------------------------------------------------------- +Time: 0.544s Load: 0.010s, Pack+Encode: 0.226s, Decode+Unpack: 0.309s +---------------------- -------------------------------------------------------- +💾 Converting with 128.7826 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,276B, BPFP=0.0919 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,420B, BPFP=0.1022 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,512B, BPFP=0.1089 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,252B, BPFP=0.0901 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,400B, BPFP=0.1008 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,284B, BPFP=0.0925 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,376B, BPFP=0.0991 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,284B, BPFP=0.0925 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,576B, BPFP=0.1135 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,464B, BPFP=0.1054 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,496B, BPFP=0.0360 +⌛️ [2/4] FRONTEND: Frontend time: 0.217s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150146 506.40272177 + layer.0.v_cache 0.00001564 0.05291583 + layer.1.k_cache 0.36060091 35.23110329 + layer.1.v_cache 0.00000598 0.01906326 + layer.2.k_cache 0.01419523 6.34325553 + layer.2.v_cache 0.00002081 0.05686292 + layer.3.k_cache 0.03680107 22.64680390 + layer.3.v_cache 0.00002092 0.06657131 + layer.4.k_cache 0.00067164 1.76280487 + layer.4.v_cache 0.00004838 0.13131051 + layer.4.output 1.41076765 250.84675362 + ------------------------------------------------------------------------------------- + TOTAL 0.61289739 136.97886403 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 17340 +BPFP 0.0734 bits/point +EBPFP 0.1469 equivalent bits/point +MSE 136.978864 +---------------------- -------------------------------------------------------- +Time: 0.536s Load: 0.007s, Pack+Encode: 0.217s, Decode+Unpack: 0.312s +---------------------- -------------------------------------------------------- +💾 Converting with 136.9789 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 343, 128) +Output shape: (1, 343, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.output: torch.Size([1, 343, 3584]) -> torch.Size([1, 1, 343, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,820B, BPFP=0.0829 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,120B, BPFP=0.0966 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,232B, BPFP=0.1017 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,928B, BPFP=0.0878 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,092B, BPFP=0.0953 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,960B, BPFP=0.0893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,988B, BPFP=0.0906 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,996B, BPFP=0.0909 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,356B, BPFP=0.1073 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,148B, BPFP=0.0978 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,020B, BPFP=0.0327 +⌛️ [2/4] FRONTEND: Frontend time: 0.281s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.430s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16677478 508.74275692 + layer.0.v_cache 0.00001698 0.05421033 + layer.1.k_cache 0.72144124 34.94740229 + layer.1.v_cache 0.00000659 0.02034887 + layer.2.k_cache 0.01351720 6.24814154 + layer.2.v_cache 0.00002117 0.05977500 + layer.3.k_cache 0.02347133 22.51201485 + layer.3.v_cache 0.00002006 0.06959464 + layer.4.k_cache 0.00069292 1.76282664 + layer.4.v_cache 0.00005144 0.13598983 + layer.4.output 0.03905215 159.45013796 + ------------------------------------------------------------------------------------- + TOTAL 0.07055169 99.45317804 + (elements=2,985,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2985472 +Total Bytes 25660 +BPFP 0.0688 bits/point +EBPFP 0.1375 equivalent bits/point +MSE 99.453178 +---------------------- -------------------------------------------------------- +Time: 0.722s Load: 0.012s, Pack+Encode: 0.281s, Decode+Unpack: 0.430s +---------------------- -------------------------------------------------------- +💾 Converting with 99.4532 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,524B, BPFP=0.0866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,768B, BPFP=0.1005 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,920B, BPFP=0.1091 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,608B, BPFP=0.0914 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,760B, BPFP=0.1000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,668B, BPFP=0.0948 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,716B, BPFP=0.0975 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,680B, BPFP=0.0955 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.1141 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,848B, BPFP=0.1050 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,200B, BPFP=0.0341 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13317820 506.60068182 + layer.0.v_cache 0.00001784 0.05381309 + layer.1.k_cache 0.51782770 34.86946733 + layer.1.v_cache 0.00000612 0.01925541 + layer.2.k_cache 0.01694398 6.26116655 + layer.2.v_cache 0.00002016 0.05716848 + layer.3.k_cache 0.01920085 22.50090199 + layer.3.v_cache 0.00002016 0.06680595 + layer.4.k_cache 0.00071488 1.80807462 + layer.4.v_cache 0.00005349 0.13736237 + layer.4.output 0.00484358 202.14446429 + ------------------------------------------------------------------------------------- + TOTAL 0.04246403 116.90505574 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 21700 +BPFP 0.0725 bits/point +EBPFP 0.1451 equivalent bits/point +MSE 116.905056 +---------------------- -------------------------------------------------------- +Time: 0.625s Load: 0.009s, Pack+Encode: 0.245s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 116.9051 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 374, 128) +Output shape: (1, 374, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.output: torch.Size([1, 374, 3584]) -> torch.Size([1, 1, 374, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,884B, BPFP=0.0787 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,120B, BPFP=0.0886 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,280B, BPFP=0.0953 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,936B, BPFP=0.0809 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,148B, BPFP=0.0897 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,960B, BPFP=0.0819 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,044B, BPFP=0.0854 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,012B, BPFP=0.0841 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,396B, BPFP=0.1001 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,208B, BPFP=0.0922 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,448B, BPFP=0.0265 +⌛️ [2/4] FRONTEND: Frontend time: 0.278s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.442s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18781370 506.25108623 + layer.0.v_cache 0.00001670 0.05374622 + layer.1.k_cache 0.85217587 34.98931787 + layer.1.v_cache 0.00000605 0.01945151 + layer.2.k_cache 0.02919085 6.30202937 + layer.2.v_cache 0.00002159 0.05547614 + layer.3.k_cache 0.02976600 22.68182602 + layer.3.v_cache 0.00002022 0.06686603 + layer.4.k_cache 0.00072409 1.78714365 + layer.4.v_cache 0.00004949 0.13403408 + layer.4.output 0.03578218 147.15244223 + ------------------------------------------------------------------------------------- + TOTAL 0.07942705 94.25929839 + (elements=3,255,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3255296 +Total Bytes 25436 +BPFP 0.0625 bits/point +EBPFP 0.1250 equivalent bits/point +MSE 94.259298 +---------------------- -------------------------------------------------------- +Time: 0.734s Load: 0.014s, Pack+Encode: 0.278s, Decode+Unpack: 0.442s +---------------------- -------------------------------------------------------- +💾 Converting with 94.2593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.0861 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,492B, BPFP=0.0967 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,596B, BPFP=0.1035 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,380B, BPFP=0.0895 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,516B, BPFP=0.0983 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,392B, BPFP=0.0902 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,476B, BPFP=0.0957 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,416B, BPFP=0.0918 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,656B, BPFP=0.1074 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,540B, BPFP=0.0998 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,456B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 0.235s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.308s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13979621 511.85107624 + layer.0.v_cache 0.00001743 0.05242040 + layer.1.k_cache 0.46959303 35.22840622 + layer.1.v_cache 0.00000602 0.01925113 + layer.2.k_cache 0.01642122 6.26535705 + layer.2.v_cache 0.00002053 0.05700454 + layer.3.k_cache 0.03876302 22.52048755 + layer.3.v_cache 0.00002036 0.06967158 + layer.4.k_cache 0.00069112 1.79956865 + layer.4.v_cache 0.00005399 0.13934480 + layer.4.output 1.27033806 225.34284232 + ------------------------------------------------------------------------------------- + TOTAL 0.56222055 126.78838144 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 18248 +BPFP 0.0696 bits/point +EBPFP 0.1392 equivalent bits/point +MSE 126.788381 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.009s, Pack+Encode: 0.235s, Decode+Unpack: 0.308s +---------------------- -------------------------------------------------------- +💾 Converting with 126.7884 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,212B, BPFP=0.0834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,404B, BPFP=0.0966 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,516B, BPFP=0.1044 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,268B, BPFP=0.0873 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,420B, BPFP=0.0977 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,316B, BPFP=0.0906 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,352B, BPFP=0.0931 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,332B, BPFP=0.0917 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,596B, BPFP=0.1099 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,428B, BPFP=0.0983 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,240B, BPFP=0.0319 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.310s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13893990 506.00523128 + layer.0.v_cache 0.00001711 0.05398613 + layer.1.k_cache 0.36820191 35.45542659 + layer.1.v_cache 0.00000637 0.02001668 + layer.2.k_cache 0.01566778 6.37836903 + layer.2.v_cache 0.00002161 0.06153326 + layer.3.k_cache 0.01884081 22.76356648 + layer.3.v_cache 0.00002076 0.07346530 + layer.4.k_cache 0.00070054 1.85271585 + layer.4.v_cache 0.00005210 0.14103591 + layer.4.output 1.34866800 239.21729468 + ------------------------------------------------------------------------------------- + TOTAL 0.58724382 132.19567113 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 17084 +BPFP 0.0692 bits/point +EBPFP 0.1383 equivalent bits/point +MSE 132.195671 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.008s, Pack+Encode: 0.224s, Decode+Unpack: 0.310s +---------------------- -------------------------------------------------------- +💾 Converting with 132.1957 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,528B, BPFP=0.0894 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,756B, BPFP=0.1028 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,912B, BPFP=0.1119 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,624B, BPFP=0.0950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,768B, BPFP=0.1035 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,680B, BPFP=0.0983 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,740B, BPFP=0.1018 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,652B, BPFP=0.0967 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,996B, BPFP=0.1168 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,848B, BPFP=0.1081 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,372B, BPFP=0.0366 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11943251 504.90829822 + layer.0.v_cache 0.00001561 0.05212132 + layer.1.k_cache 0.49619702 35.23031879 + layer.1.v_cache 0.00000608 0.01916898 + layer.2.k_cache 0.01349212 6.31440146 + layer.2.v_cache 0.00002006 0.05764080 + layer.3.k_cache 0.01080775 22.67659688 + layer.3.v_cache 0.00002073 0.06660594 + layer.4.k_cache 0.00068886 1.85028991 + layer.4.v_cache 0.00007406 0.14029546 + layer.4.output 0.00494179 208.03644997 + ------------------------------------------------------------------------------------- + TOTAL 0.03972631 119.26887574 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 21876 +BPFP 0.0753 bits/point +EBPFP 0.1506 equivalent bits/point +MSE 119.268876 +---------------------- -------------------------------------------------------- +Time: 0.630s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 119.2689 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,184B, BPFP=0.0837 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,356B, BPFP=0.0959 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,472B, BPFP=0.1041 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,208B, BPFP=0.0854 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,384B, BPFP=0.0979 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,284B, BPFP=0.0908 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,340B, BPFP=0.0947 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,280B, BPFP=0.0905 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,588B, BPFP=0.1123 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,404B, BPFP=0.0993 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,204B, BPFP=0.0324 +⌛️ [2/4] FRONTEND: Frontend time: 0.237s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.308s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11815372 505.23352658 + layer.0.v_cache 0.00001627 0.05278906 + layer.1.k_cache 0.37623358 35.15896096 + layer.1.v_cache 0.00000616 0.01965235 + layer.2.k_cache 0.01429399 6.30394182 + layer.2.v_cache 0.00002051 0.05864141 + layer.3.k_cache 0.01320818 22.69059981 + layer.3.v_cache 0.00002004 0.06923672 + layer.4.k_cache 0.00067885 1.81351288 + layer.4.v_cache 0.00005318 0.14065169 + layer.4.output 1.38524649 245.40950226 + ------------------------------------------------------------------------------------- + TOTAL 0.60114176 134.67106054 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 16704 +BPFP 0.0695 bits/point +EBPFP 0.1389 equivalent bits/point +MSE 134.671061 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.008s, Pack+Encode: 0.237s, Decode+Unpack: 0.308s +---------------------- -------------------------------------------------------- +💾 Converting with 134.6711 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,548B, BPFP=0.0870 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,784B, BPFP=0.1003 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,896B, BPFP=0.1066 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,572B, BPFP=0.0884 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.0996 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,632B, BPFP=0.0917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,708B, BPFP=0.0960 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,604B, BPFP=0.0902 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,996B, BPFP=0.1122 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,768B, BPFP=0.0994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,132B, BPFP=0.0332 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09820673 509.32306655 + layer.0.v_cache 0.00001666 0.05176478 + layer.1.k_cache 0.50544519 35.13244365 + layer.1.v_cache 0.00000613 0.01902142 + layer.2.k_cache 0.01930715 6.25268950 + layer.2.v_cache 0.00002033 0.05646151 + layer.3.k_cache 0.01337509 22.51304484 + layer.3.v_cache 0.00001987 0.06729075 + layer.4.k_cache 0.00074230 1.77119281 + layer.4.v_cache 0.00004869 0.13469931 + layer.4.output 0.00473525 201.28565005 + ------------------------------------------------------------------------------------- + TOTAL 0.03943146 116.72477797 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 21412 +BPFP 0.0708 bits/point +EBPFP 0.1416 equivalent bits/point +MSE 116.724778 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 116.7248 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,208B, BPFP=0.0866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,420B, BPFP=0.1018 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,496B, BPFP=0.1072 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,252B, BPFP=0.0897 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,408B, BPFP=0.1009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,304B, BPFP=0.0935 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,384B, BPFP=0.0992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,320B, BPFP=0.0946 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,560B, BPFP=0.1118 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,448B, BPFP=0.1038 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,348B, BPFP=0.0343 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13853801 507.25996273 + layer.0.v_cache 0.00001803 0.05245208 + layer.1.k_cache 0.23740530 35.16315761 + layer.1.v_cache 0.00000581 0.01858523 + layer.2.k_cache 0.01896443 6.32317947 + layer.2.v_cache 0.00001977 0.05634018 + layer.3.k_cache 0.00876656 22.56106203 + layer.3.v_cache 0.00002034 0.06956475 + layer.4.k_cache 0.00073522 1.75082635 + layer.4.v_cache 0.00004914 0.13241986 + layer.4.output 1.40427464 249.59475344 + ------------------------------------------------------------------------------------- + TOTAL 0.60202618 136.50298967 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 17148 +BPFP 0.0723 bits/point +EBPFP 0.1446 equivalent bits/point +MSE 136.502990 +---------------------- -------------------------------------------------------- +Time: 0.544s Load: 0.009s, Pack+Encode: 0.223s, Decode+Unpack: 0.312s +---------------------- -------------------------------------------------------- +💾 Converting with 136.5030 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,528B, BPFP=0.0865 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,788B, BPFP=0.1012 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,904B, BPFP=0.1078 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,596B, BPFP=0.0904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,756B, BPFP=0.0994 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,632B, BPFP=0.0924 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,720B, BPFP=0.0974 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,652B, BPFP=0.0935 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,996B, BPFP=0.1130 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,828B, BPFP=0.1035 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,124B, BPFP=0.0334 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10827675 506.49626359 + layer.0.v_cache 0.00001706 0.05251413 + layer.1.k_cache 0.56919524 35.02893597 + layer.1.v_cache 0.00000592 0.01891378 + layer.2.k_cache 0.01911232 6.24871826 + layer.2.v_cache 0.00002018 0.05606964 + layer.3.k_cache 0.01342056 22.56253184 + layer.3.v_cache 0.00002333 0.07059768 + layer.4.k_cache 0.00074072 1.79486073 + layer.4.v_cache 0.00005062 0.13677702 + layer.4.output 0.00480139 201.65333851 + ------------------------------------------------------------------------------------- + TOTAL 0.04379250 116.70820895 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 21524 +BPFP 0.0717 bits/point +EBPFP 0.1434 equivalent bits/point +MSE 116.708209 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 116.7082 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,516B, BPFP=0.0887 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,732B, BPFP=0.1014 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,912B, BPFP=0.1119 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,644B, BPFP=0.0962 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,784B, BPFP=0.1044 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,660B, BPFP=0.0971 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,752B, BPFP=0.1025 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,700B, BPFP=0.0995 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,048B, BPFP=0.1199 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,832B, BPFP=0.1072 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,372B, BPFP=0.0366 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15963619 504.07794944 + layer.0.v_cache 0.00001721 0.05749997 + layer.1.k_cache 0.49929055 35.26076779 + layer.1.v_cache 0.00000624 0.02016318 + layer.2.k_cache 0.03270629 6.36958547 + layer.2.v_cache 0.00002062 0.05852037 + layer.3.k_cache 0.00946706 22.56279077 + layer.3.v_cache 0.00002056 0.07068800 + layer.4.k_cache 0.00068536 1.82704203 + layer.4.v_cache 0.00004941 0.13386080 + layer.4.output 0.00495045 208.11814473 + ------------------------------------------------------------------------------------- + TOTAL 0.04332662 119.25093417 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 21952 +BPFP 0.0756 bits/point +EBPFP 0.1511 equivalent bits/point +MSE 119.250934 +---------------------- -------------------------------------------------------- +Time: 0.624s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 119.2509 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,344B, BPFP=0.0890 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,456B, BPFP=0.0964 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,612B, BPFP=0.1067 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,336B, BPFP=0.0885 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,520B, BPFP=0.1006 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,364B, BPFP=0.0903 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,468B, BPFP=0.0972 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,408B, BPFP=0.0932 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,664B, BPFP=0.1102 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,516B, BPFP=0.1004 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,532B, BPFP=0.0334 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14874478 509.83467956 + layer.0.v_cache 0.00001714 0.05363165 + layer.1.k_cache 0.45078779 35.42894680 + layer.1.v_cache 0.00000615 0.01977239 + layer.2.k_cache 0.01724745 6.41798375 + layer.2.v_cache 0.00001987 0.05613800 + layer.3.k_cache 0.02849353 22.64987338 + layer.3.v_cache 0.00002056 0.06952081 + layer.4.k_cache 0.00066912 1.81416748 + layer.4.v_cache 0.00005068 0.13636495 + layer.4.output 1.29722614 229.76626816 + ------------------------------------------------------------------------------------- + TOTAL 0.57215530 128.52029152 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 18220 +BPFP 0.0710 bits/point +EBPFP 0.1419 equivalent bits/point +MSE 128.520292 +---------------------- -------------------------------------------------------- +Time: 0.538s Load: 0.009s, Pack+Encode: 0.220s, Decode+Unpack: 0.309s +---------------------- -------------------------------------------------------- +💾 Converting with 128.5203 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 341, 128) +Output shape: (1, 341, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.output: torch.Size([1, 341, 3584]) -> torch.Size([1, 1, 341, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,816B, BPFP=0.0832 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,092B, BPFP=0.0959 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,252B, BPFP=0.1032 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,888B, BPFP=0.0865 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,112B, BPFP=0.0968 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,956B, BPFP=0.0896 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,020B, BPFP=0.0926 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,976B, BPFP=0.0905 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,380B, BPFP=0.1091 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,212B, BPFP=0.1014 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,780B, BPFP=0.0313 +⌛️ [2/4] FRONTEND: Frontend time: 0.277s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.421s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17009405 506.23689516 + layer.0.v_cache 0.00001631 0.05407448 + layer.1.k_cache 0.79816426 34.86629112 + layer.1.v_cache 0.00000611 0.01962551 + layer.2.k_cache 0.02878924 6.26483960 + layer.2.v_cache 0.00002016 0.05766001 + layer.3.k_cache 0.01035325 22.53363986 + layer.3.v_cache 0.00002082 0.06882358 + layer.4.k_cache 0.00072160 1.83559724 + layer.4.v_cache 0.00005837 0.13226413 + layer.4.output 0.03919861 159.59494135 + ------------------------------------------------------------------------------------- + TOTAL 0.07544909 99.36672354 + (elements=2,968,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2968064 +Total Bytes 25484 +BPFP 0.0687 bits/point +EBPFP 0.1374 equivalent bits/point +MSE 99.366724 +---------------------- -------------------------------------------------------- +Time: 0.710s Load: 0.013s, Pack+Encode: 0.277s, Decode+Unpack: 0.421s +---------------------- -------------------------------------------------------- +💾 Converting with 99.3667 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 335, 128) +Output shape: (1, 335, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.output: torch.Size([1, 335, 3584]) -> torch.Size([1, 1, 335, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,784B, BPFP=0.0832 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,104B, BPFP=0.0981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,272B, BPFP=0.1060 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,900B, BPFP=0.0886 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,084B, BPFP=0.0972 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,952B, BPFP=0.0910 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,028B, BPFP=0.0946 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,984B, BPFP=0.0925 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,400B, BPFP=0.1119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,172B, BPFP=0.1013 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,896B, BPFP=0.0326 +⌛️ [2/4] FRONTEND: Frontend time: 0.297s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.448s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15195324 507.93955224 + layer.0.v_cache 0.00001666 0.05460146 + layer.1.k_cache 0.76788093 35.01794834 + layer.1.v_cache 0.00000624 0.01983758 + layer.2.k_cache 0.02686723 6.29149152 + layer.2.v_cache 0.00001960 0.05617351 + layer.3.k_cache 0.01080669 22.72676073 + layer.3.v_cache 0.00002024 0.06880080 + layer.4.k_cache 0.00071790 1.83404669 + layer.4.v_cache 0.00005106 0.13623160 + layer.4.output 3.38143948 160.46247335 + ------------------------------------------------------------------------------------- + TOTAL 1.44873036 99.84604458 + (elements=2,915,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2915840 +Total Bytes 25576 +BPFP 0.0702 bits/point +EBPFP 0.1403 equivalent bits/point +MSE 99.846045 +---------------------- -------------------------------------------------------- +Time: 0.761s Load: 0.016s, Pack+Encode: 0.297s, Decode+Unpack: 0.448s +---------------------- -------------------------------------------------------- +💾 Converting with 99.8460 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,564B, BPFP=0.0922 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,772B, BPFP=0.1045 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,892B, BPFP=0.1116 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,612B, BPFP=0.0950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,784B, BPFP=0.1052 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,656B, BPFP=0.0976 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,732B, BPFP=0.1021 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,728B, BPFP=0.1019 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,024B, BPFP=0.1193 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,844B, BPFP=0.1087 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,668B, BPFP=0.0393 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17469128 501.75619104 + layer.0.v_cache 0.00001775 0.05528021 + layer.1.k_cache 0.51265420 35.15360038 + layer.1.v_cache 0.00000620 0.01997363 + layer.2.k_cache 0.03428403 6.35622973 + layer.2.v_cache 0.00002024 0.05686918 + layer.3.k_cache 0.01443097 22.65251879 + layer.3.v_cache 0.00002182 0.07475449 + layer.4.k_cache 0.00070119 1.79617540 + layer.4.v_cache 0.00005097 0.13422829 + layer.4.output 0.00503371 209.57494946 + ------------------------------------------------------------------------------------- + TOTAL 0.04541851 119.71061573 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 22276 +BPFP 0.0773 bits/point +EBPFP 0.1545 equivalent bits/point +MSE 119.710616 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.009s, Pack+Encode: 0.252s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 119.7106 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,568B, BPFP=0.0925 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,748B, BPFP=0.1031 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,620B, BPFP=0.0955 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,780B, BPFP=0.1050 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,660B, BPFP=0.0979 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,720B, BPFP=0.1014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,684B, BPFP=0.0993 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,036B, BPFP=0.1200 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,832B, BPFP=0.1080 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,460B, BPFP=0.0376 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18043976 503.07063679 + layer.0.v_cache 0.00001814 0.05304977 + layer.1.k_cache 0.58196952 35.14826061 + layer.1.v_cache 0.00000608 0.01938009 + layer.2.k_cache 0.02409397 6.28756173 + layer.2.v_cache 0.00002020 0.05582024 + layer.3.k_cache 0.01875559 22.61868735 + layer.3.v_cache 0.00002085 0.06922124 + layer.4.k_cache 0.00069073 1.78328224 + layer.4.v_cache 0.00005389 0.13739391 + layer.4.output 0.00498411 209.57639825 + ------------------------------------------------------------------------------------- + TOTAL 0.04946809 119.78106363 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 22016 +BPFP 0.0764 bits/point +EBPFP 0.1527 equivalent bits/point +MSE 119.781064 +---------------------- -------------------------------------------------------- +Time: 0.621s Load: 0.009s, Pack+Encode: 0.246s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 119.7811 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 386, 128) +Output shape: (1, 386, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.output: torch.Size([1, 386, 3584]) -> torch.Size([1, 1, 386, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,084B, BPFP=0.0844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,256B, BPFP=0.0913 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,544B, BPFP=0.1030 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,040B, BPFP=0.0826 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,388B, BPFP=0.0967 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,136B, BPFP=0.0865 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,252B, BPFP=0.0912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,124B, BPFP=0.0860 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,664B, BPFP=0.1078 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,428B, BPFP=0.0983 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,144B, BPFP=0.0297 +⌛️ [2/4] FRONTEND: Frontend time: 0.316s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.486s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16808204 502.19009067 + layer.0.v_cache 0.00001827 0.05475845 + layer.1.k_cache 0.89383947 34.85668060 + layer.1.v_cache 0.00000605 0.01987657 + layer.2.k_cache 0.02643975 6.32929763 + layer.2.v_cache 0.00002094 0.06026395 + layer.3.k_cache 0.04368651 22.40472191 + layer.3.v_cache 0.00002149 0.07266351 + layer.4.k_cache 0.00075087 1.86969358 + layer.4.v_cache 0.00005208 0.13969454 + layer.4.output 0.03469837 142.10979136 + ------------------------------------------------------------------------------------- + TOTAL 0.08092977 91.92742829 + (elements=3,359,744) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3359744 +Total Bytes 28060 +BPFP 0.0668 bits/point +EBPFP 0.1336 equivalent bits/point +MSE 91.927428 +---------------------- -------------------------------------------------------- +Time: 0.814s Load: 0.012s, Pack+Encode: 0.316s, Decode+Unpack: 0.486s +---------------------- -------------------------------------------------------- +💾 Converting with 91.9274 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,488B, BPFP=0.0807 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,648B, BPFP=0.0894 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,832B, BPFP=0.0994 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,508B, BPFP=0.0818 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,728B, BPFP=0.0938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,548B, BPFP=0.0840 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,676B, BPFP=0.0909 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,564B, BPFP=0.0849 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,936B, BPFP=0.1050 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,732B, BPFP=0.0940 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,920B, BPFP=0.0304 +⌛️ [2/4] FRONTEND: Frontend time: 0.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14364293 506.10101997 + layer.0.v_cache 0.00001600 0.05441152 + layer.1.k_cache 0.62366660 35.13169352 + layer.1.v_cache 0.00000628 0.02035473 + layer.2.k_cache 0.01981349 6.32471042 + layer.2.v_cache 0.00002090 0.06178549 + layer.3.k_cache 0.02158341 22.62762790 + layer.3.v_cache 0.00002097 0.07085868 + layer.4.k_cache 0.00071696 1.87442440 + layer.4.v_cache 0.00005883 0.14385537 + layer.4.output 0.00463133 194.16218688 + ------------------------------------------------------------------------------------- + TOTAL 0.04952739 113.62035589 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 20580 +BPFP 0.0657 bits/point +EBPFP 0.1314 equivalent bits/point +MSE 113.620356 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.267s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 113.6204 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,524B, BPFP=0.0860 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,768B, BPFP=0.0997 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1076 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,616B, BPFP=0.0912 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,788B, BPFP=0.1009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,656B, BPFP=0.0934 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,728B, BPFP=0.0975 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,656B, BPFP=0.0934 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.1133 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,848B, BPFP=0.1042 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,216B, BPFP=0.0340 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12300906 506.22557536 + layer.0.v_cache 0.00001720 0.05214618 + layer.1.k_cache 0.58060381 34.93869162 + layer.1.v_cache 0.00000644 0.01915436 + layer.2.k_cache 0.02305070 6.17520087 + layer.2.v_cache 0.00002053 0.05932408 + layer.3.k_cache 0.05279791 22.71547834 + layer.3.v_cache 0.00002021 0.07151586 + layer.4.k_cache 0.00074809 1.84485679 + layer.4.v_cache 0.00005348 0.14191770 + layer.4.output 0.00483522 201.11073685 + ------------------------------------------------------------------------------------- + TOTAL 0.04789259 116.47170701 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 21716 +BPFP 0.0721 bits/point +EBPFP 0.1441 equivalent bits/point +MSE 116.471707 +---------------------- -------------------------------------------------------- +Time: 0.626s Load: 0.011s, Pack+Encode: 0.250s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 116.4717 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,200B, BPFP=0.0841 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,324B, BPFP=0.0928 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,496B, BPFP=0.1048 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,228B, BPFP=0.0860 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,388B, BPFP=0.0973 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,248B, BPFP=0.0874 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,336B, BPFP=0.0936 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,240B, BPFP=0.0869 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,576B, BPFP=0.1104 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,396B, BPFP=0.0978 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,152B, BPFP=0.0316 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16177035 512.08639294 + layer.0.v_cache 0.00001782 0.05526869 + layer.1.k_cache 0.31952496 35.26431124 + layer.1.v_cache 0.00000626 0.01963675 + layer.2.k_cache 0.01271096 6.28204154 + layer.2.v_cache 0.00002040 0.05951678 + layer.3.k_cache 0.01363577 22.54773332 + layer.3.v_cache 0.00002033 0.06892226 + layer.4.k_cache 0.00066992 1.82332224 + layer.4.v_cache 0.00005030 0.13474267 + layer.4.output 1.37284199 242.99073110 + ------------------------------------------------------------------------------------- + TOTAL 0.59519535 134.07511801 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 16584 +BPFP 0.0684 bits/point +EBPFP 0.1367 equivalent bits/point +MSE 134.075118 +---------------------- -------------------------------------------------------- +Time: 0.539s Load: 0.008s, Pack+Encode: 0.221s, Decode+Unpack: 0.309s +---------------------- -------------------------------------------------------- +💾 Converting with 134.0751 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,048B, BPFP=0.0895 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,124B, BPFP=0.0960 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,236B, BPFP=0.1055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,048B, BPFP=0.0895 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,188B, BPFP=0.1014 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,080B, BPFP=0.0922 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,156B, BPFP=0.0987 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,080B, BPFP=0.0922 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,256B, BPFP=0.1072 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,204B, BPFP=0.1028 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,700B, BPFP=0.0329 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.257s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09962382 514.07009904 + layer.0.v_cache 0.00001686 0.05601417 + layer.1.k_cache 0.10988800 36.21586034 + layer.1.v_cache 0.00000590 0.01983717 + layer.2.k_cache 0.01214521 6.57087616 + layer.2.v_cache 0.00002216 0.06169779 + layer.3.k_cache 0.02260447 23.30292915 + layer.3.v_cache 0.00002006 0.07405349 + layer.4.k_cache 0.00069723 1.90324410 + layer.4.v_cache 0.00005242 0.14740955 + layer.4.output 0.00882482 310.06947697 + ------------------------------------------------------------------------------------- + TOTAL 0.01804999 161.93578587 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 14120 +BPFP 0.0709 bits/point +EBPFP 0.1418 equivalent bits/point +MSE 161.935786 +---------------------- -------------------------------------------------------- +Time: 0.514s Load: 0.007s, Pack+Encode: 0.251s, Decode+Unpack: 0.257s +---------------------- -------------------------------------------------------- +💾 Converting with 161.9358 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,552B, BPFP=0.0863 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,708B, BPFP=0.0950 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,864B, BPFP=0.1036 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,520B, BPFP=0.0845 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,744B, BPFP=0.0970 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,532B, BPFP=0.0852 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,696B, BPFP=0.0943 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,596B, BPFP=0.0887 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,988B, BPFP=0.1105 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,748B, BPFP=0.0972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,068B, BPFP=0.0323 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11148910 506.15697286 + layer.0.v_cache 0.00001684 0.05487989 + layer.1.k_cache 0.60671574 35.03203890 + layer.1.v_cache 0.00000630 0.01935723 + layer.2.k_cache 0.04149003 6.34621139 + layer.2.v_cache 0.00001976 0.05526858 + layer.3.k_cache 0.02200458 22.56226020 + layer.3.v_cache 0.00001954 0.06879264 + layer.4.k_cache 0.00078895 1.80225913 + layer.4.v_cache 0.00004864 0.13080623 + layer.4.output 0.00472744 198.56537557 + ------------------------------------------------------------------------------------- + TOTAL 0.04798185 115.42273389 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 21016 +BPFP 0.0687 bits/point +EBPFP 0.1375 equivalent bits/point +MSE 115.422734 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.368s +---------------------- -------------------------------------------------------- +💾 Converting with 115.4227 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,268B, BPFP=0.0991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,392B, BPFP=0.1087 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,584B, BPFP=0.1237 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,312B, BPFP=0.1025 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,456B, BPFP=0.1138 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,272B, BPFP=0.0994 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,412B, BPFP=0.1103 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,312B, BPFP=0.1025 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,664B, BPFP=0.1300 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,492B, BPFP=0.1166 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,836B, BPFP=0.0428 +⌛️ [2/4] FRONTEND: Frontend time: 0.219s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.306s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14565937 507.81648438 + layer.0.v_cache 0.00001550 0.05067434 + layer.1.k_cache 0.22559690 35.34677246 + layer.1.v_cache 0.00000590 0.01886178 + layer.2.k_cache 0.00679684 6.38428528 + layer.2.v_cache 0.00002314 0.05791004 + layer.3.k_cache 0.06866968 22.54425781 + layer.3.v_cache 0.00002119 0.06945313 + layer.4.k_cache 0.00065507 1.78032883 + layer.4.v_cache 0.00005191 0.14060431 + layer.4.output 1.53062134 271.00497768 + ------------------------------------------------------------------------------------- + TOTAL 0.65657911 145.36732212 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 18000 +BPFP 0.0827 bits/point +EBPFP 0.1654 equivalent bits/point +MSE 145.367322 +---------------------- -------------------------------------------------------- +Time: 0.532s Load: 0.007s, Pack+Encode: 0.219s, Decode+Unpack: 0.306s +---------------------- -------------------------------------------------------- +💾 Converting with 145.3673 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,272B, BPFP=0.0989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,420B, BPFP=0.1104 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,596B, BPFP=0.1241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,296B, BPFP=0.1007 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,464B, BPFP=0.1138 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,348B, BPFP=0.1048 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,432B, BPFP=0.1113 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,348B, BPFP=0.1048 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,644B, BPFP=0.1278 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,496B, BPFP=0.1163 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,860B, BPFP=0.0429 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.305s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10663509 504.09853078 + layer.0.v_cache 0.00001613 0.05232121 + layer.1.k_cache 0.26620468 35.49290656 + layer.1.v_cache 0.00000602 0.01935518 + layer.2.k_cache 0.00649704 6.41306286 + layer.2.v_cache 0.00002002 0.05728872 + layer.3.k_cache 0.02565914 22.70131763 + layer.3.v_cache 0.00002186 0.06900567 + layer.4.k_cache 0.00070101 1.80183752 + layer.4.v_cache 0.00005421 0.14033962 + layer.4.output 1.52303164 269.56618692 + ------------------------------------------------------------------------------------- + TOTAL 0.65100216 144.57701613 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 18176 +BPFP 0.0831 bits/point +EBPFP 0.1662 equivalent bits/point +MSE 144.577016 +---------------------- -------------------------------------------------------- +Time: 0.532s Load: 0.007s, Pack+Encode: 0.220s, Decode+Unpack: 0.305s +---------------------- -------------------------------------------------------- +💾 Converting with 144.5770 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,528B, BPFP=0.0894 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,692B, BPFP=0.0990 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,916B, BPFP=0.1121 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,572B, BPFP=0.0920 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,764B, BPFP=0.1032 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,612B, BPFP=0.0943 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,736B, BPFP=0.1016 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,664B, BPFP=0.0974 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,040B, BPFP=0.1194 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,824B, BPFP=0.1067 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,304B, BPFP=0.0360 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258070 503.20745552 + layer.0.v_cache 0.00001713 0.05405808 + layer.1.k_cache 0.49703808 35.32236072 + layer.1.v_cache 0.00000670 0.02046453 + layer.2.k_cache 0.01458295 6.33719158 + layer.2.v_cache 0.00002181 0.05949468 + layer.3.k_cache 0.02940617 22.55401451 + layer.3.v_cache 0.00002120 0.07288841 + layer.4.k_cache 0.00070157 1.85695234 + layer.4.v_cache 0.00005095 0.13751686 + layer.4.output 0.00497170 207.98613898 + ------------------------------------------------------------------------------------- + TOTAL 0.04171936 119.14855118 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 21652 +BPFP 0.0745 bits/point +EBPFP 0.1491 equivalent bits/point +MSE 119.148551 +---------------------- -------------------------------------------------------- +Time: 0.621s Load: 0.009s, Pack+Encode: 0.247s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 119.1486 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,200B, BPFP=0.0830 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,348B, BPFP=0.0932 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,512B, BPFP=0.1045 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,204B, BPFP=0.0832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,392B, BPFP=0.0962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,276B, BPFP=0.0882 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,348B, BPFP=0.0932 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,280B, BPFP=0.0885 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,564B, BPFP=0.1081 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,384B, BPFP=0.0957 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,084B, BPFP=0.0305 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.306s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15433978 504.49630116 + layer.0.v_cache 0.00001575 0.05211484 + layer.1.k_cache 0.29362690 35.35477090 + layer.1.v_cache 0.00000607 0.01938242 + layer.2.k_cache 0.00940348 6.41213854 + layer.2.v_cache 0.00002159 0.05817695 + layer.3.k_cache 0.02320663 22.76362218 + layer.3.v_cache 0.00002067 0.07211528 + layer.4.k_cache 0.00067785 1.81602329 + layer.4.v_cache 0.00005016 0.13729679 + layer.4.output 1.35461534 240.50400996 + ------------------------------------------------------------------------------------- + TOTAL 0.58609860 132.63000071 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 16592 +BPFP 0.0675 bits/point +EBPFP 0.1350 equivalent bits/point +MSE 132.630001 +---------------------- -------------------------------------------------------- +Time: 0.535s Load: 0.009s, Pack+Encode: 0.220s, Decode+Unpack: 0.306s +---------------------- -------------------------------------------------------- +💾 Converting with 132.6300 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,460B, BPFP=0.0800 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,664B, BPFP=0.0912 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,828B, BPFP=0.1002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,508B, BPFP=0.0827 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,716B, BPFP=0.0941 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,556B, BPFP=0.0853 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,652B, BPFP=0.0906 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,584B, BPFP=0.0868 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,940B, BPFP=0.1064 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,744B, BPFP=0.0956 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,904B, BPFP=0.0306 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14751465 505.09797149 + layer.0.v_cache 0.00001719 0.05524145 + layer.1.k_cache 0.52245269 35.04634046 + layer.1.v_cache 0.00000618 0.01981828 + layer.2.k_cache 0.02707210 6.27190798 + layer.2.v_cache 0.00002227 0.06037707 + layer.3.k_cache 0.03364062 22.52961554 + layer.3.v_cache 0.00002083 0.07243636 + layer.4.k_cache 0.00069748 1.84243485 + layer.4.v_cache 0.00005305 0.13686076 + layer.4.output 0.00471531 194.97272870 + ------------------------------------------------------------------------------------- + TOTAL 0.04497084 113.87894736 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 20556 +BPFP 0.0663 bits/point +EBPFP 0.1326 equivalent bits/point +MSE 113.878947 +---------------------- -------------------------------------------------------- +Time: 0.627s Load: 0.012s, Pack+Encode: 0.248s, Decode+Unpack: 0.368s +---------------------- -------------------------------------------------------- +💾 Converting with 113.8789 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,136B, BPFP=0.0696 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,256B, BPFP=0.0770 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,420B, BPFP=0.0870 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,092B, BPFP=0.0669 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,304B, BPFP=0.0799 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,128B, BPFP=0.0691 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,224B, BPFP=0.0750 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,180B, BPFP=0.0723 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,468B, BPFP=0.0900 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,320B, BPFP=0.0809 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,692B, BPFP=0.0236 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17432048 499.18235294 + layer.0.v_cache 0.00001704 0.05144591 + layer.1.k_cache 0.39057378 34.21438036 + layer.1.v_cache 0.00000668 0.01946059 + layer.2.k_cache 0.01299327 6.02081323 + layer.2.v_cache 0.00002114 0.05479918 + layer.3.k_cache 0.02499930 22.18818551 + layer.3.v_cache 0.00002096 0.06718873 + layer.4.k_cache 0.00069121 1.71511925 + layer.4.v_cache 0.00005416 0.13440484 + layer.4.output 1.20070616 212.52027311 + ------------------------------------------------------------------------------------- + TOTAL 0.52992007 120.66412131 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 15220 +BPFP 0.0549 bits/point +EBPFP 0.1097 equivalent bits/point +MSE 120.664121 +---------------------- -------------------------------------------------------- +Time: 0.537s Load: 0.009s, Pack+Encode: 0.221s, Decode+Unpack: 0.307s +---------------------- -------------------------------------------------------- +💾 Converting with 120.6641 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,272B, BPFP=0.0868 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,412B, BPFP=0.0963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,536B, BPFP=0.1048 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,304B, BPFP=0.0890 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,448B, BPFP=0.0988 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,372B, BPFP=0.0936 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,388B, BPFP=0.0947 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,388B, BPFP=0.0947 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,628B, BPFP=0.1111 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,464B, BPFP=0.0999 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,496B, BPFP=0.0341 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16239141 503.78916485 + layer.0.v_cache 0.00001645 0.05278252 + layer.1.k_cache 0.33751179 35.42720771 + layer.1.v_cache 0.00000668 0.01968432 + layer.2.k_cache 0.00660169 6.28847614 + layer.2.v_cache 0.00002045 0.05768038 + layer.3.k_cache 0.02213682 22.77328824 + layer.3.v_cache 0.00002049 0.07162525 + layer.4.k_cache 0.00067824 1.81963418 + layer.4.v_cache 0.00005183 0.13606724 + layer.4.output 1.33690934 236.86644183 + ------------------------------------------------------------------------------------- + TOTAL 0.58163537 131.08827668 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 17708 +BPFP 0.0711 bits/point +EBPFP 0.1421 equivalent bits/point +MSE 131.088277 +---------------------- -------------------------------------------------------- +Time: 0.538s Load: 0.008s, Pack+Encode: 0.220s, Decode+Unpack: 0.309s +---------------------- -------------------------------------------------------- +💾 Converting with 131.0883 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,540B, BPFP=0.0898 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,700B, BPFP=0.0991 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,900B, BPFP=0.1108 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,584B, BPFP=0.0924 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,752B, BPFP=0.1021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,612B, BPFP=0.0940 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,736B, BPFP=0.1012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,620B, BPFP=0.0944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,016B, BPFP=0.1175 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,816B, BPFP=0.1059 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,288B, BPFP=0.0357 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15796844 501.04920709 + layer.0.v_cache 0.00001527 0.05137795 + layer.1.k_cache 0.50235566 35.19449991 + layer.1.v_cache 0.00000626 0.01955865 + layer.2.k_cache 0.02709456 6.18391077 + layer.2.v_cache 0.00002063 0.05439699 + layer.3.k_cache 0.00856834 22.69065379 + layer.3.v_cache 0.00001945 0.06602445 + layer.4.k_cache 0.00071483 1.80650170 + layer.4.v_cache 0.00005464 0.13620841 + layer.4.output 0.00491706 207.40769923 + ------------------------------------------------------------------------------------- + TOTAL 0.04301397 118.77095496 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 21564 +BPFP 0.0740 bits/point +EBPFP 0.1479 equivalent bits/point +MSE 118.770955 +---------------------- -------------------------------------------------------- +Time: 0.626s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.369s +---------------------- -------------------------------------------------------- +💾 Converting with 118.7710 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,324B, BPFP=0.0834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,440B, BPFP=0.0907 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,568B, BPFP=0.0988 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,308B, BPFP=0.0824 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,492B, BPFP=0.0940 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,352B, BPFP=0.0852 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,444B, BPFP=0.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,376B, BPFP=0.0867 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,612B, BPFP=0.1016 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,488B, BPFP=0.0938 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,316B, BPFP=0.0298 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.310s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13184587 511.44398942 + layer.0.v_cache 0.00001706 0.05573831 + layer.1.k_cache 0.45546043 35.60909148 + layer.1.v_cache 0.00000678 0.02128950 + layer.2.k_cache 0.02272066 6.50249949 + layer.2.v_cache 0.00001996 0.06114306 + layer.3.k_cache 0.01643563 23.36833339 + layer.3.v_cache 0.00002069 0.07447152 + layer.4.k_cache 0.00071781 1.90580405 + layer.4.v_cache 0.00005262 0.14289035 + layer.4.output 1.23455937 229.02093534 + ------------------------------------------------------------------------------------- + TOTAL 0.54524783 128.37245870 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 17720 +BPFP 0.0657 bits/point +EBPFP 0.1313 equivalent bits/point +MSE 128.372459 +---------------------- -------------------------------------------------------- +Time: 0.540s Load: 0.010s, Pack+Encode: 0.220s, Decode+Unpack: 0.310s +---------------------- -------------------------------------------------------- +💾 Converting with 128.3725 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,580B, BPFP=0.0935 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,744B, BPFP=0.1032 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,892B, BPFP=0.1120 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,556B, BPFP=0.0921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,740B, BPFP=0.1030 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,608B, BPFP=0.0952 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,716B, BPFP=0.1016 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,668B, BPFP=0.0987 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,028B, BPFP=0.1200 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,836B, BPFP=0.1087 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,476B, BPFP=0.0378 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17705467 502.10179924 + layer.0.v_cache 0.00001633 0.05104542 + layer.1.k_cache 0.50406676 35.13146603 + layer.1.v_cache 0.00000599 0.01889625 + layer.2.k_cache 0.00944760 6.25381932 + layer.2.v_cache 0.00001883 0.05462631 + layer.3.k_cache 0.04120483 22.47335538 + layer.3.v_cache 0.00002392 0.07121718 + layer.4.k_cache 0.00070100 1.79658370 + layer.4.v_cache 0.00005002 0.13320067 + layer.4.output 0.00499618 210.34926272 + ------------------------------------------------------------------------------------- + TOTAL 0.04515078 120.03122638 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 21844 +BPFP 0.0761 bits/point +EBPFP 0.1521 equivalent bits/point +MSE 120.031226 +---------------------- -------------------------------------------------------- +Time: 0.621s Load: 0.011s, Pack+Encode: 0.246s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 120.0312 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,544B, BPFP=0.0894 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,740B, BPFP=0.1007 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,904B, BPFP=0.1102 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,592B, BPFP=0.0921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,736B, BPFP=0.1005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,632B, BPFP=0.0944 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,736B, BPFP=0.1005 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,656B, BPFP=0.0958 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,992B, BPFP=0.1153 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,840B, BPFP=0.1065 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,244B, BPFP=0.0351 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20501094 510.66383102 + layer.0.v_cache 0.00001516 0.05018513 + layer.1.k_cache 0.57085758 35.31605903 + layer.1.v_cache 0.00000574 0.01854401 + layer.2.k_cache 0.01259345 6.24946696 + layer.2.v_cache 0.00002099 0.05434224 + layer.3.k_cache 0.01648921 22.51500470 + layer.3.v_cache 0.00001970 0.06702226 + layer.4.k_cache 0.00070444 1.76992380 + layer.4.v_cache 0.00005121 0.13370836 + layer.4.output 0.00487421 205.35940807 + ------------------------------------------------------------------------------------- + TOTAL 0.04940517 118.49140847 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 21616 +BPFP 0.0736 bits/point +EBPFP 0.1472 equivalent bits/point +MSE 118.491408 +---------------------- -------------------------------------------------------- +Time: 0.624s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 118.4914 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,532B, BPFP=0.0883 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,712B, BPFP=0.0987 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,912B, BPFP=0.1102 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,588B, BPFP=0.0916 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,744B, BPFP=0.1006 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,644B, BPFP=0.0948 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,720B, BPFP=0.0992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,640B, BPFP=0.0946 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.1158 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,776B, BPFP=0.1024 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,260B, BPFP=0.0351 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14466163 506.65924815 + layer.0.v_cache 0.00001631 0.05426340 + layer.1.k_cache 0.53789613 35.23693352 + layer.1.v_cache 0.00000653 0.01981016 + layer.2.k_cache 0.02095415 6.27576269 + layer.2.v_cache 0.00002169 0.05845341 + layer.3.k_cache 0.03083320 22.69303686 + layer.3.v_cache 0.00002083 0.07024503 + layer.4.k_cache 0.00069737 1.82560387 + layer.4.v_cache 0.00004943 0.13521652 + layer.4.output 0.00490546 204.90957762 + ------------------------------------------------------------------------------------- + TOTAL 0.04526444 118.08209512 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 21536 +BPFP 0.0730 bits/point +EBPFP 0.1461 equivalent bits/point +MSE 118.082095 +---------------------- -------------------------------------------------------- +Time: 0.623s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 118.0821 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,604B, BPFP=0.0838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,796B, BPFP=0.0939 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,940B, BPFP=0.1014 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,616B, BPFP=0.0844 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,824B, BPFP=0.0953 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,684B, BPFP=0.0880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,752B, BPFP=0.0916 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,704B, BPFP=0.0890 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,048B, BPFP=0.1070 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,848B, BPFP=0.0966 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,200B, BPFP=0.0314 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14240686 507.45688754 + layer.0.v_cache 0.00001673 0.05268859 + layer.1.k_cache 0.61438330 35.23869278 + layer.1.v_cache 0.00000638 0.01911337 + layer.2.k_cache 0.02027981 6.19335774 + layer.2.v_cache 0.00002012 0.05536743 + layer.3.k_cache 0.02199794 22.65094913 + layer.3.v_cache 0.00002003 0.06588517 + layer.4.k_cache 0.00078937 1.75885969 + layer.4.v_cache 0.00005038 0.13245297 + layer.4.output 0.04464151 181.94548794 + ------------------------------------------------------------------------------------- + TOTAL 0.06543891 108.66133353 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 22016 +BPFP 0.0677 bits/point +EBPFP 0.1354 equivalent bits/point +MSE 108.661334 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.012s, Pack+Encode: 0.249s, Decode+Unpack: 0.367s +---------------------- -------------------------------------------------------- +💾 Converting with 108.6613 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 399, 128) +Output shape: (1, 399, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.output: torch.Size([1, 399, 3584]) -> torch.Size([1, 1, 399, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,068B, BPFP=0.0810 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,368B, BPFP=0.0927 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,628B, BPFP=0.1029 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,164B, BPFP=0.0847 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,408B, BPFP=0.0943 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,248B, BPFP=0.0880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,332B, BPFP=0.0913 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,280B, BPFP=0.0893 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,764B, BPFP=0.1082 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,536B, BPFP=0.0993 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,460B, BPFP=0.0305 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.487s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17889012 509.29111842 + layer.0.v_cache 0.00001635 0.05201599 + layer.1.k_cache 0.94984295 34.87882059 + layer.1.v_cache 0.00000590 0.01867401 + layer.2.k_cache 0.02305104 6.17693709 + layer.2.v_cache 0.00002114 0.05895360 + layer.3.k_cache 0.01620252 22.48473968 + layer.3.v_cache 0.00002049 0.07012346 + layer.4.k_cache 0.00073175 1.81999616 + layer.4.v_cache 0.00005255 0.13489467 + layer.4.output 0.00656733 139.03036609 + ------------------------------------------------------------------------------------- + TOTAL 0.07145918 91.07051978 + (elements=3,472,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3472896 +Total Bytes 29256 +BPFP 0.0674 bits/point +EBPFP 0.1348 equivalent bits/point +MSE 91.070520 +---------------------- -------------------------------------------------------- +Time: 0.819s Load: 0.014s, Pack+Encode: 0.319s, Decode+Unpack: 0.487s +---------------------- -------------------------------------------------------- +💾 Converting with 91.0705 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,196B, BPFP=0.0827 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,420B, BPFP=0.0982 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,508B, BPFP=0.1043 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,236B, BPFP=0.0855 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,412B, BPFP=0.0976 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,284B, BPFP=0.0888 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,352B, BPFP=0.0935 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,304B, BPFP=0.0902 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,576B, BPFP=0.1090 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,400B, BPFP=0.0968 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,232B, BPFP=0.0319 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.310s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10670873 505.15780559 + layer.0.v_cache 0.00001672 0.05499064 + layer.1.k_cache 0.29393762 35.32252057 + layer.1.v_cache 0.00000627 0.01972765 + layer.2.k_cache 0.01243114 6.42612370 + layer.2.v_cache 0.00002105 0.06140523 + layer.3.k_cache 0.00919804 22.72158894 + layer.3.v_cache 0.00002002 0.07221907 + layer.4.k_cache 0.00070102 1.81900119 + layer.4.v_cache 0.00005024 0.14269226 + layer.4.output 1.35463108 240.53630689 + ------------------------------------------------------------------------------------- + TOTAL 0.58267697 132.67954253 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 16920 +BPFP 0.0688 bits/point +EBPFP 0.1376 equivalent bits/point +MSE 132.679543 +---------------------- -------------------------------------------------------- +Time: 0.544s Load: 0.009s, Pack+Encode: 0.225s, Decode+Unpack: 0.310s +---------------------- -------------------------------------------------------- +💾 Converting with 132.6795 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,164B, BPFP=0.0716 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,308B, BPFP=0.0805 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,424B, BPFP=0.0876 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,136B, BPFP=0.0699 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,324B, BPFP=0.0814 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,176B, BPFP=0.0723 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,252B, BPFP=0.0770 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,192B, BPFP=0.0733 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,472B, BPFP=0.0906 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,340B, BPFP=0.0824 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,072B, BPFP=0.0270 +⌛️ [2/4] FRONTEND: Frontend time: 0.234s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633087 503.99298720 + layer.0.v_cache 0.00001891 0.05394169 + layer.1.k_cache 0.36012592 34.37017486 + layer.1.v_cache 0.00000581 0.01848344 + layer.2.k_cache 0.01407751 6.14616334 + layer.2.v_cache 0.00002023 0.05708838 + layer.3.k_cache 0.01230825 22.11414439 + layer.3.v_cache 0.00002000 0.06926013 + layer.4.k_cache 0.00071033 1.75313382 + layer.4.v_cache 0.00004856 0.13322223 + layer.4.output 1.20534739 213.27834997 + ------------------------------------------------------------------------------------- + TOTAL 0.52535871 121.27394408 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 15860 +BPFP 0.0574 bits/point +EBPFP 0.1148 equivalent bits/point +MSE 121.273944 +---------------------- -------------------------------------------------------- +Time: 0.588s Load: 0.008s, Pack+Encode: 0.234s, Decode+Unpack: 0.346s +---------------------- -------------------------------------------------------- +💾 Converting with 121.2739 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,332B, BPFP=0.0864 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,500B, BPFP=0.0973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,604B, BPFP=0.1040 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,364B, BPFP=0.0884 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,516B, BPFP=0.0983 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,412B, BPFP=0.0915 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,480B, BPFP=0.0960 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,396B, BPFP=0.0905 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,632B, BPFP=0.1058 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,524B, BPFP=0.0988 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,528B, BPFP=0.0327 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11142051 511.43393413 + layer.0.v_cache 0.00001650 0.05159342 + layer.1.k_cache 0.42032433 35.23851627 + layer.1.v_cache 0.00000619 0.01898737 + layer.2.k_cache 0.01443043 6.25740526 + layer.2.v_cache 0.00002030 0.05769599 + layer.3.k_cache 0.02895402 22.68028924 + layer.3.v_cache 0.00001914 0.06536493 + layer.4.k_cache 0.00067703 1.79178515 + layer.4.v_cache 0.00005693 0.14015182 + layer.4.output 1.27033357 225.28343583 + ------------------------------------------------------------------------------------- + TOTAL 0.55695649 126.74822202 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 18288 +BPFP 0.0697 bits/point +EBPFP 0.1395 equivalent bits/point +MSE 126.748222 +---------------------- -------------------------------------------------------- +Time: 0.545s Load: 0.008s, Pack+Encode: 0.225s, Decode+Unpack: 0.311s +---------------------- -------------------------------------------------------- +💾 Converting with 126.7482 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 302, 128) +Output shape: (1, 302, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.output: torch.Size([1, 302, 3584]) -> torch.Size([1, 1, 302, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,592B, BPFP=0.0824 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,808B, BPFP=0.0935 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,932B, BPFP=0.1000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,616B, BPFP=0.0836 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,828B, BPFP=0.0946 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,676B, BPFP=0.0867 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,724B, BPFP=0.0892 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,644B, BPFP=0.0851 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,036B, BPFP=0.1053 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,832B, BPFP=0.0948 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,072B, BPFP=0.0301 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15012774 506.15640522 + layer.0.v_cache 0.00001504 0.05001514 + layer.1.k_cache 0.60100308 35.23614704 + layer.1.v_cache 0.00000575 0.01826546 + layer.2.k_cache 0.00814273 6.30620651 + layer.2.v_cache 0.00002032 0.05582460 + layer.3.k_cache 0.01797006 22.56955583 + layer.3.v_cache 0.00001899 0.06395637 + layer.4.k_cache 0.00070977 1.79970950 + layer.4.v_cache 0.00004868 0.13006291 + layer.4.output 0.04414053 180.15496393 + ------------------------------------------------------------------------------------- + TOTAL 0.06394387 107.85122918 + (elements=2,628,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2628608 +Total Bytes 21760 +BPFP 0.0662 bits/point +EBPFP 0.1325 equivalent bits/point +MSE 107.851229 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.011s, Pack+Encode: 0.257s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 107.8512 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,524B, BPFP=0.0889 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,760B, BPFP=0.1026 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,900B, BPFP=0.1108 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,632B, BPFP=0.0951 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,752B, BPFP=0.1021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,684B, BPFP=0.0982 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,708B, BPFP=0.0996 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,696B, BPFP=0.0989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,004B, BPFP=0.1168 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,880B, BPFP=0.1096 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,360B, BPFP=0.0363 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.412s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14350806 507.29524254 + layer.0.v_cache 0.00001655 0.05076430 + layer.1.k_cache 0.46812200 35.29023583 + layer.1.v_cache 0.00000598 0.01861673 + layer.2.k_cache 0.01096032 6.26158757 + layer.2.v_cache 0.00002019 0.05619905 + layer.3.k_cache 0.00658789 22.58357990 + layer.3.v_cache 0.00002014 0.06710484 + layer.4.k_cache 0.00067528 1.77044040 + layer.4.v_cache 0.00005785 0.14117232 + layer.4.output 0.00496594 207.35834222 + ------------------------------------------------------------------------------------- + TOTAL 0.03910211 119.12019641 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 21900 +BPFP 0.0751 bits/point +EBPFP 0.1502 equivalent bits/point +MSE 119.120196 +---------------------- -------------------------------------------------------- +Time: 0.673s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.412s +---------------------- -------------------------------------------------------- +💾 Converting with 119.1202 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 260, 128) +Output shape: (1, 260, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.output: torch.Size([1, 260, 3584]) -> torch.Size([1, 1, 260, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,552B, BPFP=0.0933 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,692B, BPFP=0.1017 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,864B, BPFP=0.1120 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,476B, BPFP=0.0887 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,740B, BPFP=0.1046 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,560B, BPFP=0.0938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,684B, BPFP=0.1012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,568B, BPFP=0.0942 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,928B, BPFP=0.1159 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,792B, BPFP=0.1077 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,396B, BPFP=0.0377 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14548651 503.72145433 + layer.0.v_cache 0.00001584 0.05126236 + layer.1.k_cache 0.47062912 35.12386944 + layer.1.v_cache 0.00000604 0.01863000 + layer.2.k_cache 0.00619981 6.27853534 + layer.2.v_cache 0.00002003 0.05580169 + layer.3.k_cache 0.01299995 22.62047401 + layer.3.v_cache 0.00001981 0.06647524 + layer.4.k_cache 0.00067777 1.76972539 + layer.4.v_cache 0.00006192 0.13582687 + layer.4.output 0.00504520 213.71342720 + ------------------------------------------------------------------------------------- + TOTAL 0.03949607 121.51976736 + (elements=2,263,040) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2263040 +Total Bytes 21252 +BPFP 0.0751 bits/point +EBPFP 0.1503 equivalent bits/point +MSE 121.519767 +---------------------- -------------------------------------------------------- +Time: 0.624s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 121.5198 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,508B, BPFP=0.0866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,784B, BPFP=0.1025 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1096 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,608B, BPFP=0.0924 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,760B, BPFP=0.1011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,668B, BPFP=0.0958 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,724B, BPFP=0.0990 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,684B, BPFP=0.0967 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,024B, BPFP=0.1163 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,884B, BPFP=0.1082 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,228B, BPFP=0.0347 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14555165 502.82732077 + layer.0.v_cache 0.00001661 0.05384193 + layer.1.k_cache 0.55468144 35.17086972 + layer.1.v_cache 0.00000621 0.01983698 + layer.2.k_cache 0.02243149 6.28500995 + layer.2.v_cache 0.00002103 0.05807287 + layer.3.k_cache 0.02028967 22.60578829 + layer.3.v_cache 0.00002036 0.07044825 + layer.4.k_cache 0.00072975 1.82685213 + layer.4.v_cache 0.00005124 0.13770990 + layer.4.output 0.00486695 204.04516807 + ------------------------------------------------------------------------------------- + TOTAL 0.04575695 117.49246631 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 21780 +BPFP 0.0736 bits/point +EBPFP 0.1472 equivalent bits/point +MSE 117.492466 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 117.4925 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 356, 128) +Output shape: (1, 356, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.output: torch.Size([1, 356, 3584]) -> torch.Size([1, 1, 356, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,792B, BPFP=0.0787 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,052B, BPFP=0.0901 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,196B, BPFP=0.0964 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,884B, BPFP=0.0827 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,072B, BPFP=0.0909 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,920B, BPFP=0.0843 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,984B, BPFP=0.0871 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,976B, BPFP=0.0867 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,372B, BPFP=0.1041 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,124B, BPFP=0.0932 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,796B, BPFP=0.0301 +⌛️ [2/4] FRONTEND: Frontend time: 0.332s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.434s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15223843 498.63303195 + layer.0.v_cache 0.00001740 0.05312700 + layer.1.k_cache 0.82972649 34.89194994 + layer.1.v_cache 0.00000631 0.01992953 + layer.2.k_cache 0.03086275 6.26128533 + layer.2.v_cache 0.00001960 0.05753738 + layer.3.k_cache 0.01100341 22.55329124 + layer.3.v_cache 0.00002062 0.06961693 + layer.4.k_cache 0.00073857 1.82848984 + layer.4.v_cache 0.00005279 0.13462582 + layer.4.output 0.03758925 152.50305979 + ------------------------------------------------------------------------------------- + TOTAL 0.07575360 96.00142962 + (elements=3,098,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3098624 +Total Bytes 25168 +BPFP 0.0650 bits/point +EBPFP 0.1300 equivalent bits/point +MSE 96.001430 +---------------------- -------------------------------------------------------- +Time: 0.779s Load: 0.013s, Pack+Encode: 0.332s, Decode+Unpack: 0.434s +---------------------- -------------------------------------------------------- +💾 Converting with 96.0014 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,532B, BPFP=0.0900 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,752B, BPFP=0.1029 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,884B, BPFP=0.1107 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,608B, BPFP=0.0945 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,756B, BPFP=0.1031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,620B, BPFP=0.0952 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,712B, BPFP=0.1006 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,648B, BPFP=0.0968 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.1180 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,828B, BPFP=0.1074 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,424B, BPFP=0.0371 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12347319 501.56396852 + layer.0.v_cache 0.00001696 0.05299477 + layer.1.k_cache 0.44284580 35.26485770 + layer.1.v_cache 0.00000660 0.02017608 + layer.2.k_cache 0.00882101 6.18146807 + layer.2.v_cache 0.00002210 0.06143423 + layer.3.k_cache 0.01145097 22.74444351 + layer.3.v_cache 0.00001984 0.06873559 + layer.4.k_cache 0.00068754 1.83095453 + layer.4.v_cache 0.00006483 0.13945205 + layer.4.output 0.00500964 208.82362715 + ------------------------------------------------------------------------------------- + TOTAL 0.03661626 119.39375736 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 21772 +BPFP 0.0752 bits/point +EBPFP 0.1505 equivalent bits/point +MSE 119.393757 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.368s +---------------------- -------------------------------------------------------- +💾 Converting with 119.3938 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,472B, BPFP=0.0807 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,676B, BPFP=0.0919 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,816B, BPFP=0.0996 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,472B, BPFP=0.0807 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,712B, BPFP=0.0939 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,528B, BPFP=0.0838 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,636B, BPFP=0.0897 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,540B, BPFP=0.0844 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,928B, BPFP=0.1057 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,724B, BPFP=0.0945 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,728B, BPFP=0.0292 +⌛️ [2/4] FRONTEND: Frontend time: 0.265s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13002123 504.90389254 + layer.0.v_cache 0.00001833 0.05449090 + layer.1.k_cache 0.50397591 34.98387473 + layer.1.v_cache 0.00000608 0.01910890 + layer.2.k_cache 0.01926375 6.23702243 + layer.2.v_cache 0.00002060 0.05827275 + layer.3.k_cache 0.04992905 22.58732525 + layer.3.v_cache 0.00002096 0.07010807 + layer.4.k_cache 0.00069540 1.83549012 + layer.4.v_cache 0.00005467 0.14125246 + layer.4.output 0.00466491 195.06679198 + ------------------------------------------------------------------------------------- + TOTAL 0.04333296 113.90343424 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 20232 +BPFP 0.0652 bits/point +EBPFP 0.1305 equivalent bits/point +MSE 113.903434 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.011s, Pack+Encode: 0.265s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 113.9034 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,552B, BPFP=0.0866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,732B, BPFP=0.0967 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,892B, BPFP=0.1056 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,592B, BPFP=0.0888 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,768B, BPFP=0.0987 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,640B, BPFP=0.0915 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,704B, BPFP=0.0951 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,672B, BPFP=0.0933 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.1107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,804B, BPFP=0.1007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,288B, BPFP=0.0342 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15974760 505.41311384 + layer.0.v_cache 0.00001739 0.05308584 + layer.1.k_cache 0.45644542 35.03846261 + layer.1.v_cache 0.00000643 0.01962512 + layer.2.k_cache 0.01631587 6.17791443 + layer.2.v_cache 0.00002005 0.05584333 + layer.3.k_cache 0.01511352 22.47641950 + layer.3.v_cache 0.00002063 0.07094615 + layer.4.k_cache 0.00073517 1.74024320 + layer.4.v_cache 0.00005091 0.13850565 + layer.4.output 0.00481773 199.15779656 + ------------------------------------------------------------------------------------- + TOTAL 0.04012924 115.60521974 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 21628 +BPFP 0.0710 bits/point +EBPFP 0.1420 equivalent bits/point +MSE 115.605220 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 115.6052 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,536B, BPFP=0.0899 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,680B, BPFP=0.0983 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1117 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,580B, BPFP=0.0925 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,744B, BPFP=0.1021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,600B, BPFP=0.0936 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,724B, BPFP=0.1009 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,604B, BPFP=0.0939 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,992B, BPFP=0.1166 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,800B, BPFP=0.1053 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,312B, BPFP=0.0360 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14148096 502.40859082 + layer.0.v_cache 0.00001588 0.05153344 + layer.1.k_cache 0.55266568 35.25855864 + layer.1.v_cache 0.00000632 0.01917415 + layer.2.k_cache 0.00687984 6.25050383 + layer.2.v_cache 0.00001885 0.05497760 + layer.3.k_cache 0.01675770 22.72714953 + layer.3.v_cache 0.00001964 0.06648689 + layer.4.k_cache 0.00068991 1.77210227 + layer.4.v_cache 0.00004859 0.13290994 + layer.4.output 0.00492290 208.12326110 + ------------------------------------------------------------------------------------- + TOTAL 0.04429669 119.15322440 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 21480 +BPFP 0.0739 bits/point +EBPFP 0.1479 equivalent bits/point +MSE 119.153224 +---------------------- -------------------------------------------------------- +Time: 0.620s Load: 0.011s, Pack+Encode: 0.246s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 119.1532 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,820B, BPFP=0.0880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,068B, BPFP=0.1000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,200B, BPFP=0.1064 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,820B, BPFP=0.0880 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,048B, BPFP=0.0991 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,888B, BPFP=0.0913 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,972B, BPFP=0.0954 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,916B, BPFP=0.0927 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,336B, BPFP=0.1130 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,100B, BPFP=0.1016 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,828B, BPFP=0.0334 +⌛️ [2/4] FRONTEND: Frontend time: 0.284s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.422s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13144862 501.16307082 + layer.0.v_cache 0.00001769 0.05498805 + layer.1.k_cache 0.64513603 34.98805752 + layer.1.v_cache 0.00000607 0.01951734 + layer.2.k_cache 0.02726801 6.39380336 + layer.2.v_cache 0.00001956 0.05648392 + layer.3.k_cache 0.02271994 22.50254571 + layer.3.v_cache 0.00002018 0.06732468 + layer.4.k_cache 0.00073057 1.81657972 + layer.4.v_cache 0.00004940 0.13007103 + layer.4.output 0.04135688 169.74296495 + ------------------------------------------------------------------------------------- + TOTAL 0.06570084 103.25842334 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 24996 +BPFP 0.0711 bits/point +EBPFP 0.1423 equivalent bits/point +MSE 103.258423 +---------------------- -------------------------------------------------------- +Time: 0.717s Load: 0.012s, Pack+Encode: 0.284s, Decode+Unpack: 0.422s +---------------------- -------------------------------------------------------- +💾 Converting with 103.2584 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,280B, BPFP=0.0930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,428B, BPFP=0.1038 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,548B, BPFP=0.1125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,288B, BPFP=0.0936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,456B, BPFP=0.1058 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,300B, BPFP=0.0945 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,392B, BPFP=0.1012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,320B, BPFP=0.0959 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,604B, BPFP=0.1166 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,484B, BPFP=0.1078 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,596B, BPFP=0.0373 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12396691 510.76293605 + layer.0.v_cache 0.00001598 0.05382881 + layer.1.k_cache 0.24854917 35.32487736 + layer.1.v_cache 0.00000584 0.01916648 + layer.2.k_cache 0.02290044 6.40927280 + layer.2.v_cache 0.00002299 0.05605458 + layer.3.k_cache 0.01294550 22.66013354 + layer.3.v_cache 0.00002077 0.06823991 + layer.4.k_cache 0.00065144 1.75378077 + layer.4.v_cache 0.00005054 0.13491222 + layer.4.output 1.42385976 253.91465947 + ------------------------------------------------------------------------------------- + TOTAL 0.61036164 138.50857758 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 17696 +BPFP 0.0756 bits/point +EBPFP 0.1513 equivalent bits/point +MSE 138.508578 +---------------------- -------------------------------------------------------- +Time: 0.544s Load: 0.008s, Pack+Encode: 0.221s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 138.5086 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,480B, BPFP=0.0800 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,688B, BPFP=0.0913 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,832B, BPFP=0.0990 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,488B, BPFP=0.0804 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,724B, BPFP=0.0932 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,560B, BPFP=0.0843 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,668B, BPFP=0.0902 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,616B, BPFP=0.0874 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,960B, BPFP=0.1060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,736B, BPFP=0.0939 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,788B, BPFP=0.0293 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14783513 502.47112889 + layer.0.v_cache 0.00001600 0.05407285 + layer.1.k_cache 0.54442752 35.12163103 + layer.1.v_cache 0.00000628 0.01946570 + layer.2.k_cache 0.01848349 6.37363568 + layer.2.v_cache 0.00002061 0.05929856 + layer.3.k_cache 0.01777397 22.64022289 + layer.3.v_cache 0.00002092 0.07239080 + layer.4.k_cache 0.00070447 1.83004074 + layer.4.v_cache 0.00005049 0.13720327 + layer.4.output 0.04615090 189.11353806 + ------------------------------------------------------------------------------------- + TOTAL 0.06190560 111.32787393 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 20540 +BPFP 0.0653 bits/point +EBPFP 0.1306 equivalent bits/point +MSE 111.327874 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 111.3279 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,620B, BPFP=0.0849 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,828B, BPFP=0.0958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,940B, BPFP=0.1017 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,648B, BPFP=0.0864 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,844B, BPFP=0.0967 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,720B, BPFP=0.0902 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,764B, BPFP=0.0925 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,740B, BPFP=0.0912 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,068B, BPFP=0.1084 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,876B, BPFP=0.0984 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,240B, BPFP=0.0318 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16274841 509.00314597 + layer.0.v_cache 0.00001794 0.05425347 + layer.1.k_cache 0.56737488 35.19659711 + layer.1.v_cache 0.00000586 0.01891285 + layer.2.k_cache 0.02027899 6.34977978 + layer.2.v_cache 0.00002117 0.05722884 + layer.3.k_cache 0.01891868 22.58139878 + layer.3.v_cache 0.00002140 0.06842347 + layer.4.k_cache 0.00069658 1.79477502 + layer.4.v_cache 0.00005005 0.13364303 + layer.4.output 0.04476916 182.54431328 + ------------------------------------------------------------------------------------- + TOTAL 0.06373636 109.00402066 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 22288 +BPFP 0.0687 bits/point +EBPFP 0.1375 equivalent bits/point +MSE 109.004021 +---------------------- -------------------------------------------------------- +Time: 0.627s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.369s +---------------------- -------------------------------------------------------- +💾 Converting with 109.0040 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,520B, BPFP=0.0870 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,756B, BPFP=0.1005 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,916B, BPFP=0.1097 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,612B, BPFP=0.0923 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,776B, BPFP=0.1016 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,620B, BPFP=0.0927 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,720B, BPFP=0.0984 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,708B, BPFP=0.0978 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,028B, BPFP=0.1161 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,864B, BPFP=0.1067 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,252B, BPFP=0.0348 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13366390 508.22355769 + layer.0.v_cache 0.00001719 0.05432977 + layer.1.k_cache 0.57618574 35.09362480 + layer.1.v_cache 0.00000658 0.01911318 + layer.2.k_cache 0.02858028 6.34596515 + layer.2.v_cache 0.00002021 0.05596684 + layer.3.k_cache 0.00873641 22.36016376 + layer.3.v_cache 0.00002023 0.06771835 + layer.4.k_cache 0.00067337 1.77004479 + layer.4.v_cache 0.00005011 0.13459929 + layer.4.output 0.00484872 203.35189364 + ------------------------------------------------------------------------------------- + TOTAL 0.04599383 117.50519642 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 21772 +BPFP 0.0733 bits/point +EBPFP 0.1466 equivalent bits/point +MSE 117.505196 +---------------------- -------------------------------------------------------- +Time: 0.622s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 117.5052 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,188B, BPFP=0.0840 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,340B, BPFP=0.0947 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,472B, BPFP=0.1041 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,212B, BPFP=0.0857 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,392B, BPFP=0.0984 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,228B, BPFP=0.0868 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,336B, BPFP=0.0945 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,256B, BPFP=0.0888 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,568B, BPFP=0.1109 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,380B, BPFP=0.0976 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,204B, BPFP=0.0324 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13057878 504.90441176 + layer.0.v_cache 0.00001797 0.05535208 + layer.1.k_cache 0.31658528 35.29309690 + layer.1.v_cache 0.00000647 0.01986545 + layer.2.k_cache 0.02849985 6.29621825 + layer.2.v_cache 0.00001991 0.05867384 + layer.3.k_cache 0.03599559 22.66833772 + layer.3.v_cache 0.00002239 0.07399110 + layer.4.k_cache 0.00078811 1.82459753 + layer.4.v_cache 0.00005540 0.14161183 + layer.4.output 1.38530015 245.41780462 + ------------------------------------------------------------------------------------- + TOTAL 0.60056887 134.66239934 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 16576 +BPFP 0.0689 bits/point +EBPFP 0.1379 equivalent bits/point +MSE 134.662399 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.010s, Pack+Encode: 0.222s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 134.6624 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,284B, BPFP=0.0925 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,428B, BPFP=0.1028 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,512B, BPFP=0.1089 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,240B, BPFP=0.0893 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,424B, BPFP=0.1025 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,292B, BPFP=0.0930 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,372B, BPFP=0.0988 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,316B, BPFP=0.0948 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,620B, BPFP=0.1166 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,484B, BPFP=0.1069 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,504B, BPFP=0.0360 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.322s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12725546 508.08021313 + layer.0.v_cache 0.00001736 0.05580057 + layer.1.k_cache 0.32262681 35.24745059 + layer.1.v_cache 0.00000608 0.01981847 + layer.2.k_cache 0.02157613 6.43931755 + layer.2.v_cache 0.00002208 0.05983759 + layer.3.k_cache 0.01760322 22.62049521 + layer.3.v_cache 0.00002192 0.07425537 + layer.4.k_cache 0.00069785 1.81191510 + layer.4.v_cache 0.00005172 0.14073937 + layer.4.output 1.41081570 250.86493993 + ------------------------------------------------------------------------------------- + TOTAL 0.60974050 137.09437779 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 17476 +BPFP 0.0740 bits/point +EBPFP 0.1480 equivalent bits/point +MSE 137.094378 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.010s, Pack+Encode: 0.222s, Decode+Unpack: 0.322s +---------------------- -------------------------------------------------------- +💾 Converting with 137.0944 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,248B, BPFP=0.0915 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,412B, BPFP=0.1036 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,572B, BPFP=0.1153 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,320B, BPFP=0.0968 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,468B, BPFP=0.1077 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,336B, BPFP=0.0980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,412B, BPFP=0.1036 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,348B, BPFP=0.0989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,612B, BPFP=0.1183 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,540B, BPFP=0.1130 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,524B, BPFP=0.0369 +⌛️ [2/4] FRONTEND: Frontend time: 0.280s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.337s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13788442 506.75586854 + layer.0.v_cache 0.00001717 0.05710365 + layer.1.k_cache 0.30723858 35.30781433 + layer.1.v_cache 0.00000664 0.02055060 + layer.2.k_cache 0.01337219 6.38899109 + layer.2.v_cache 0.00002159 0.06017980 + layer.3.k_cache 0.03522867 22.73787779 + layer.3.v_cache 0.00002126 0.07408704 + layer.4.k_cache 0.00073009 1.82306496 + layer.4.v_cache 0.00005119 0.14238088 + layer.4.output 1.43730973 255.04923290 + ------------------------------------------------------------------------------------- + TOTAL 0.62092588 138.74779700 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 17792 +BPFP 0.0768 bits/point +EBPFP 0.1535 equivalent bits/point +MSE 138.747797 +---------------------- -------------------------------------------------------- +Time: 0.625s Load: 0.007s, Pack+Encode: 0.280s, Decode+Unpack: 0.337s +---------------------- -------------------------------------------------------- +💾 Converting with 138.7478 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,244B, BPFP=0.0953 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,416B, BPFP=0.1085 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,600B, BPFP=0.1225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,352B, BPFP=0.1036 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,460B, BPFP=0.1118 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,384B, BPFP=0.1060 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,436B, BPFP=0.1100 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,356B, BPFP=0.1039 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,644B, BPFP=0.1259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,512B, BPFP=0.1158 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,772B, BPFP=0.0413 +⌛️ [2/4] FRONTEND: Frontend time: 0.281s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11546621 513.67895987 + layer.0.v_cache 0.00001823 0.05574819 + layer.1.k_cache 0.26924565 35.62092142 + layer.1.v_cache 0.00000737 0.02066434 + layer.2.k_cache 0.01761903 6.40426636 + layer.2.v_cache 0.00002145 0.06106945 + layer.3.k_cache 0.02374693 22.84868308 + layer.3.v_cache 0.00002104 0.07326839 + layer.4.k_cache 0.00067272 1.82614749 + layer.4.v_cache 0.00005153 0.13972976 + layer.4.output 1.50071327 265.72522759 + ------------------------------------------------------------------------------------- + TOTAL 0.64305077 143.57682656 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 18176 +BPFP 0.0819 bits/point +EBPFP 0.1638 equivalent bits/point +MSE 143.576827 +---------------------- -------------------------------------------------------- +Time: 0.618s Load: 0.007s, Pack+Encode: 0.281s, Decode+Unpack: 0.330s +---------------------- -------------------------------------------------------- +💾 Converting with 143.5768 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,644B, BPFP=0.0834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,752B, BPFP=0.0889 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,928B, BPFP=0.0978 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,660B, BPFP=0.0842 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,848B, BPFP=0.0938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,708B, BPFP=0.0866 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,764B, BPFP=0.0895 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,724B, BPFP=0.0875 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,016B, BPFP=0.1023 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,852B, BPFP=0.0940 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,976B, BPFP=0.0288 +⌛️ [2/4] FRONTEND: Frontend time: 0.303s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.419s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18206066 509.69389205 + layer.0.v_cache 0.00001652 0.05463949 + layer.1.k_cache 0.62405802 35.05679916 + layer.1.v_cache 0.00000616 0.01973171 + layer.2.k_cache 0.02036299 6.34948334 + layer.2.v_cache 0.00002087 0.06010062 + layer.3.k_cache 0.02546444 22.63853871 + layer.3.v_cache 0.00002064 0.06990934 + layer.4.k_cache 0.00069880 1.84944549 + layer.4.v_cache 0.00007146 0.14041311 + layer.4.output 0.04336799 177.33341547 + ------------------------------------------------------------------------------------- + TOTAL 0.06802097 106.89805066 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 21872 +BPFP 0.0653 bits/point +EBPFP 0.1305 equivalent bits/point +MSE 106.898051 +---------------------- -------------------------------------------------------- +Time: 0.732s Load: 0.009s, Pack+Encode: 0.303s, Decode+Unpack: 0.419s +---------------------- -------------------------------------------------------- +💾 Converting with 106.8981 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,264B, BPFP=0.0862 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,432B, BPFP=0.0977 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,532B, BPFP=0.1045 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,296B, BPFP=0.0884 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,456B, BPFP=0.0993 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,340B, BPFP=0.0914 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,400B, BPFP=0.0955 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,340B, BPFP=0.0914 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,616B, BPFP=0.1103 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,492B, BPFP=0.1018 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,532B, BPFP=0.0344 +⌛️ [2/4] FRONTEND: Frontend time: 0.292s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.333s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14018573 504.63949918 + layer.0.v_cache 0.00001653 0.05566483 + layer.1.k_cache 0.40671090 35.41869797 + layer.1.v_cache 0.00000662 0.02061022 + layer.2.k_cache 0.03036383 6.36421943 + layer.2.v_cache 0.00002088 0.05896373 + layer.3.k_cache 0.05940984 22.62624949 + layer.3.v_cache 0.00002025 0.07126509 + layer.4.k_cache 0.00071161 1.87031855 + layer.4.v_cache 0.00005094 0.13671586 + layer.4.output 1.33692960 236.86562305 + ------------------------------------------------------------------------------------- + TOTAL 0.58800025 131.13656269 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 17700 +BPFP 0.0710 bits/point +EBPFP 0.1421 equivalent bits/point +MSE 131.136563 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.007s, Pack+Encode: 0.292s, Decode+Unpack: 0.333s +---------------------- -------------------------------------------------------- +💾 Converting with 131.1366 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,196B, BPFP=0.0853 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,312B, BPFP=0.0936 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,492B, BPFP=0.1064 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,220B, BPFP=0.0870 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,396B, BPFP=0.0996 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,252B, BPFP=0.0893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,360B, BPFP=0.0970 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,260B, BPFP=0.0899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,552B, BPFP=0.1107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,376B, BPFP=0.0982 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,324B, BPFP=0.0339 +⌛️ [2/4] FRONTEND: Frontend time: 0.282s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351684 502.87457192 + layer.0.v_cache 0.00001872 0.05647566 + layer.1.k_cache 0.28219827 35.14023482 + layer.1.v_cache 0.00000627 0.02016323 + layer.2.k_cache 0.01639774 6.38768974 + layer.2.v_cache 0.00002172 0.06178919 + layer.3.k_cache 0.02752649 22.69478409 + layer.3.v_cache 0.00002090 0.07133844 + layer.4.k_cache 0.00067264 1.80863305 + layer.4.v_cache 0.00005058 0.13935296 + layer.4.output 1.39794763 248.09448386 + ------------------------------------------------------------------------------------- + TOTAL 0.60212139 135.64214236 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 16740 +BPFP 0.0703 bits/point +EBPFP 0.1405 equivalent bits/point +MSE 135.642142 +---------------------- -------------------------------------------------------- +Time: 0.613s Load: 0.007s, Pack+Encode: 0.282s, Decode+Unpack: 0.324s +---------------------- -------------------------------------------------------- +💾 Converting with 135.6421 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,524B, BPFP=0.0889 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,740B, BPFP=0.1014 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,916B, BPFP=0.1117 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,604B, BPFP=0.0935 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,752B, BPFP=0.1021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,676B, BPFP=0.0977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,732B, BPFP=0.1010 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,708B, BPFP=0.0996 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.1171 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,900B, BPFP=0.1108 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,460B, BPFP=0.0371 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15646637 507.71525187 + layer.0.v_cache 0.00001727 0.05320670 + layer.1.k_cache 0.53079696 35.25630029 + layer.1.v_cache 0.00000635 0.02013743 + layer.2.k_cache 0.02340859 6.35152105 + layer.2.v_cache 0.00002061 0.05860735 + layer.3.k_cache 0.01050839 22.58516135 + layer.3.v_cache 0.00002156 0.06873089 + layer.4.k_cache 0.00068596 1.81858587 + layer.4.v_cache 0.00006072 0.13666776 + layer.4.output 0.00497423 207.35787580 + ------------------------------------------------------------------------------------- + TOTAL 0.04451838 119.15113536 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 22020 +BPFP 0.0755 bits/point +EBPFP 0.1510 equivalent bits/point +MSE 119.151135 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 119.1511 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,652B, BPFP=0.0838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,752B, BPFP=0.0889 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,936B, BPFP=0.0982 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,644B, BPFP=0.0834 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,856B, BPFP=0.0942 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,672B, BPFP=0.0848 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,748B, BPFP=0.0887 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,704B, BPFP=0.0864 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.1019 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,856B, BPFP=0.0942 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,228B, BPFP=0.0306 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15292851 506.20231331 + layer.0.v_cache 0.00001614 0.05463400 + layer.1.k_cache 0.61736927 35.16436688 + layer.1.v_cache 0.00000669 0.02062165 + layer.2.k_cache 0.02003183 6.27649560 + layer.2.v_cache 0.00002344 0.06009712 + layer.3.k_cache 0.01028479 22.62280115 + layer.3.v_cache 0.00002212 0.06920776 + layer.4.k_cache 0.00070434 1.83246821 + layer.4.v_cache 0.00006490 0.13792162 + layer.4.output 0.04342050 177.37099954 + ------------------------------------------------------------------------------------- + TOTAL 0.06502327 106.70811318 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 22056 +BPFP 0.0658 bits/point +EBPFP 0.1316 equivalent bits/point +MSE 106.708113 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.369s +---------------------- -------------------------------------------------------- +💾 Converting with 106.7081 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,548B, BPFP=0.0899 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,756B, BPFP=0.1020 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,916B, BPFP=0.1113 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,612B, BPFP=0.0936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,764B, BPFP=0.1025 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,656B, BPFP=0.0962 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,720B, BPFP=0.0999 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,684B, BPFP=0.0978 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,028B, BPFP=0.1178 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,820B, BPFP=0.1057 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,336B, BPFP=0.0360 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12950333 503.78438662 + layer.0.v_cache 0.00001690 0.05438617 + layer.1.k_cache 0.52417032 35.30241854 + layer.1.v_cache 0.00000602 0.01951506 + layer.2.k_cache 0.01578268 6.31942897 + layer.2.v_cache 0.00002268 0.05875758 + layer.3.k_cache 0.03345619 22.56690724 + layer.3.v_cache 0.00002007 0.07228549 + layer.4.k_cache 0.00067301 1.77194668 + layer.4.v_cache 0.00005077 0.13836263 + layer.4.output 0.00489311 206.20799588 + ------------------------------------------------------------------------------------- + TOTAL 0.04340905 118.44378624 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 21840 +BPFP 0.0746 bits/point +EBPFP 0.1492 equivalent bits/point +MSE 118.443786 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 118.4438 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,188B, BPFP=0.0840 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,384B, BPFP=0.0979 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,472B, BPFP=0.1041 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,228B, BPFP=0.0868 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,380B, BPFP=0.0976 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,264B, BPFP=0.0894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,352B, BPFP=0.0956 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,260B, BPFP=0.0891 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,556B, BPFP=0.1100 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,436B, BPFP=0.1015 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,212B, BPFP=0.0324 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.308s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13888381 508.77078620 + layer.0.v_cache 0.00001731 0.05334754 + layer.1.k_cache 0.27925331 35.24068951 + layer.1.v_cache 0.00000589 0.01900601 + layer.2.k_cache 0.03172064 6.29978425 + layer.2.v_cache 0.00001980 0.05644842 + layer.3.k_cache 0.01738505 22.51718485 + layer.3.v_cache 0.00001961 0.06711391 + layer.4.k_cache 0.00068340 1.78870858 + layer.4.v_cache 0.00005063 0.13535884 + layer.4.output 1.38525280 245.31112637 + ------------------------------------------------------------------------------------- + TOTAL 0.59792994 134.83095957 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 16732 +BPFP 0.0696 bits/point +EBPFP 0.1392 equivalent bits/point +MSE 134.830960 +---------------------- -------------------------------------------------------- +Time: 0.540s Load: 0.008s, Pack+Encode: 0.224s, Decode+Unpack: 0.308s +---------------------- -------------------------------------------------------- +💾 Converting with 134.8310 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.0703 bits/point +Avg EBPFP 0.1406 equivalent bits/point +Avg MSE 119.795437 +Avg Time 0.634s +------------------------ ---------------------------- diff --git a/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..3f12fe083bfda3fc7ae991b26707a0cebef5f214 --- /dev/null +++ b/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 405 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other +Output output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,264B, BPFP=0.0983 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,436B, BPFP=0.1116 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,596B, BPFP=0.1241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,356B, BPFP=0.1054 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,476B, BPFP=0.1147 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,300B, BPFP=0.1011 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,452B, BPFP=0.1129 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,360B, BPFP=0.1057 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,648B, BPFP=0.1281 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,476B, BPFP=0.1147 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,852B, BPFP=0.0428 +⌛️ [2/4] FRONTEND: Frontend time: 0.460s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09475397 508.11745958 + layer.0.v_cache 0.00001927 0.05590111 + layer.1.k_cache 0.17401673 35.53711666 + layer.1.v_cache 0.00000670 0.02029670 + layer.2.k_cache 0.00814445 6.34086100 + layer.2.v_cache 0.00002157 0.06127716 + layer.3.k_cache 0.01691350 22.67425567 + layer.3.v_cache 0.00002161 0.07541533 + layer.4.k_cache 0.00069969 1.90583733 + layer.4.v_cache 0.00005477 0.14575566 + layer.4.output 1.52309445 269.57507107 + ------------------------------------------------------------------------------------- + TOTAL 0.64448903 144.82115728 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 18216 +BPFP 0.0833 bits/point +EBPFP 0.1666 equivalent bits/point +MSE 144.821157 +---------------------- -------------------------------------------------------- +Time: 0.822s Load: 0.009s, Pack+Encode: 0.460s, Decode+Unpack: 0.353s +---------------------- -------------------------------------------------------- +💾 Converting with 144.8212 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 212, 128) +Output shape: (1, 212, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.output: torch.Size([1, 212, 3584]) -> torch.Size([1, 1, 212, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,268B, BPFP=0.0935 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,456B, BPFP=0.1073 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,580B, BPFP=0.1165 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,312B, BPFP=0.0967 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,464B, BPFP=0.1079 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,328B, BPFP=0.0979 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,428B, BPFP=0.1052 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,352B, BPFP=0.0996 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,636B, BPFP=0.1206 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,604B, BPFP=0.1182 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,484B, BPFP=0.0367 +⌛️ [2/4] FRONTEND: Frontend time: 0.235s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09403566 508.45791568 + layer.0.v_cache 0.00001799 0.05618577 + layer.1.k_cache 0.24957257 35.28218050 + layer.1.v_cache 0.00000621 0.01964872 + layer.2.k_cache 0.01169950 6.42940708 + layer.2.v_cache 0.00002031 0.06021865 + layer.3.k_cache 0.03645871 22.78002930 + layer.3.v_cache 0.00002058 0.07298866 + layer.4.k_cache 0.00069161 1.87671589 + layer.4.v_cache 0.00005130 0.14044230 + layer.4.output 1.44404146 256.11731385 + ------------------------------------------------------------------------------------- + TOTAL 0.61769792 139.29393703 + (elements=1,845,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1845248 +Total Bytes 17912 +BPFP 0.0777 bits/point +EBPFP 0.1553 equivalent bits/point +MSE 139.293937 +---------------------- -------------------------------------------------------- +Time: 0.561s Load: 0.009s, Pack+Encode: 0.235s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 139.2939 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 250, 128) +Output shape: (1, 250, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.output: torch.Size([1, 250, 3584]) -> torch.Size([1, 1, 250, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,268B, BPFP=0.0793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,408B, BPFP=0.0880 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,532B, BPFP=0.0958 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,304B, BPFP=0.0815 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,436B, BPFP=0.0897 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,324B, BPFP=0.0828 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,384B, BPFP=0.0865 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,300B, BPFP=0.0813 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,572B, BPFP=0.0983 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,448B, BPFP=0.0905 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,104B, BPFP=0.0277 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09121272 509.94284375 + layer.0.v_cache 0.00001793 0.05388723 + layer.1.k_cache 0.34710925 35.06844922 + layer.1.v_cache 0.00000615 0.01939299 + layer.2.k_cache 0.02211576 6.26685059 + layer.2.v_cache 0.00002179 0.05946542 + layer.3.k_cache 0.02174057 22.57266602 + layer.3.v_cache 0.00001938 0.06667292 + layer.4.k_cache 0.00073173 1.83106689 + layer.4.v_cache 0.00005145 0.13568355 + layer.4.output 1.22463198 217.58650000 + ------------------------------------------------------------------------------------- + TOTAL 0.53267356 123.47779286 + (elements=2,176,000) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2176000 +Total Bytes 17080 +BPFP 0.0628 bits/point +EBPFP 0.1256 equivalent bits/point +MSE 123.477793 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.008s, Pack+Encode: 0.225s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 123.4778 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,216B, BPFP=0.0837 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,384B, BPFP=0.0953 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,504B, BPFP=0.1035 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,236B, BPFP=0.0851 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,412B, BPFP=0.0972 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,288B, BPFP=0.0887 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,356B, BPFP=0.0933 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,272B, BPFP=0.0876 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,572B, BPFP=0.1082 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,400B, BPFP=0.0964 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,212B, BPFP=0.0316 +⌛️ [2/4] FRONTEND: Frontend time: 0.241s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12780563 509.45574064 + layer.0.v_cache 0.00001798 0.05519765 + layer.1.k_cache 0.36413978 35.46389730 + layer.1.v_cache 0.00000619 0.01987899 + layer.2.k_cache 0.01529831 6.38635523 + layer.2.v_cache 0.00002052 0.05915960 + layer.3.k_cache 0.01269315 22.71574288 + layer.3.v_cache 0.00002023 0.06904491 + layer.4.k_cache 0.00069261 1.85863075 + layer.4.v_cache 0.00005545 0.13657757 + layer.4.output 1.34865724 239.22561359 + ------------------------------------------------------------------------------------- + TOTAL 0.58596180 132.39997180 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 16852 +BPFP 0.0682 bits/point +EBPFP 0.1365 equivalent bits/point +MSE 132.399972 +---------------------- -------------------------------------------------------- +Time: 0.573s Load: 0.007s, Pack+Encode: 0.241s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 132.4000 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 214, 128) +Output shape: (1, 214, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.output: torch.Size([1, 214, 3584]) -> torch.Size([1, 1, 214, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,272B, BPFP=0.0929 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,452B, BPFP=0.1060 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,556B, BPFP=0.1136 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,300B, BPFP=0.0949 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,444B, BPFP=0.1054 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,344B, BPFP=0.0981 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,412B, BPFP=0.1031 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,312B, BPFP=0.0958 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,620B, BPFP=0.1183 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,480B, BPFP=0.1081 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,588B, BPFP=0.0374 +⌛️ [2/4] FRONTEND: Frontend time: 0.281s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12820032 511.21225175 + layer.0.v_cache 0.00001693 0.05361606 + layer.1.k_cache 0.24875343 35.42509674 + layer.1.v_cache 0.00000587 0.01891377 + layer.2.k_cache 0.01453762 6.37029174 + layer.2.v_cache 0.00001988 0.05751827 + layer.3.k_cache 0.02276520 22.72509537 + layer.3.v_cache 0.00001932 0.06624237 + layer.4.k_cache 0.00071399 1.85377859 + layer.4.v_cache 0.00004895 0.13391213 + layer.4.output 1.43051836 255.34139269 + ------------------------------------------------------------------------------------- + TOTAL 0.61345353 139.13567445 + (elements=1,862,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1862656 +Total Bytes 17780 +BPFP 0.0764 bits/point +EBPFP 0.1527 equivalent bits/point +MSE 139.135674 +---------------------- -------------------------------------------------------- +Time: 0.618s Load: 0.007s, Pack+Encode: 0.281s, Decode+Unpack: 0.330s +---------------------- -------------------------------------------------------- +💾 Converting with 139.1357 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,516B, BPFP=0.0855 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,768B, BPFP=0.0997 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,896B, BPFP=0.1069 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,624B, BPFP=0.0916 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,788B, BPFP=0.1009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,660B, BPFP=0.0936 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,728B, BPFP=0.0975 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,652B, BPFP=0.0932 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,988B, BPFP=0.1121 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,828B, BPFP=0.1031 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,100B, BPFP=0.0330 +⌛️ [2/4] FRONTEND: Frontend time: 0.462s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.427s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11875103 507.09205776 + layer.0.v_cache 0.00001689 0.05402637 + layer.1.k_cache 0.49171178 35.02076870 + layer.1.v_cache 0.00000616 0.01925188 + layer.2.k_cache 0.02532451 6.23750784 + layer.2.v_cache 0.00002086 0.05923131 + layer.3.k_cache 0.02370731 22.47858437 + layer.3.v_cache 0.00002129 0.06964007 + layer.4.k_cache 0.00073268 1.82056160 + layer.4.v_cache 0.00004879 0.13674563 + layer.4.output 0.00480151 201.20052218 + ------------------------------------------------------------------------------------- + TOTAL 0.04082070 116.55247240 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 21548 +BPFP 0.0715 bits/point +EBPFP 0.1430 equivalent bits/point +MSE 116.552472 +---------------------- -------------------------------------------------------- +Time: 0.898s Load: 0.009s, Pack+Encode: 0.462s, Decode+Unpack: 0.427s +---------------------- -------------------------------------------------------- +💾 Converting with 116.5525 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,336B, BPFP=0.0852 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,464B, BPFP=0.0934 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,580B, BPFP=0.1008 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,336B, BPFP=0.0852 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,508B, BPFP=0.0962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,384B, BPFP=0.0883 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,468B, BPFP=0.0936 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,384B, BPFP=0.0883 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,652B, BPFP=0.1054 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,500B, BPFP=0.0957 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,252B, BPFP=0.0296 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10657067 509.18303571 + layer.0.v_cache 0.00001604 0.05372169 + layer.1.k_cache 0.39270948 35.39799904 + layer.1.v_cache 0.00000611 0.01966287 + layer.2.k_cache 0.01541229 6.37728097 + layer.2.v_cache 0.00002285 0.06133026 + layer.3.k_cache 0.01321528 22.83583785 + layer.3.v_cache 0.00001985 0.06972691 + layer.4.k_cache 0.00076706 1.88583075 + layer.4.v_cache 0.00005205 0.13716023 + layer.4.output 1.24960368 222.47866254 + ------------------------------------------------------------------------------------- + TOTAL 0.54564808 125.49248377 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 17864 +BPFP 0.0670 bits/point +EBPFP 0.1340 equivalent bits/point +MSE 125.492484 +---------------------- -------------------------------------------------------- +Time: 0.563s Load: 0.008s, Pack+Encode: 0.227s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 125.4925 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 197, 128) +Output shape: (1, 197, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.output: torch.Size([1, 197, 3584]) -> torch.Size([1, 1, 197, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,264B, BPFP=0.1003 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,396B, BPFP=0.1107 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,564B, BPFP=0.1240 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,272B, BPFP=0.1009 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,408B, BPFP=0.1117 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,324B, BPFP=0.1050 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,388B, BPFP=0.1101 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,260B, BPFP=0.0999 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,596B, BPFP=0.1266 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,448B, BPFP=0.1148 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,912B, BPFP=0.0443 +⌛️ [2/4] FRONTEND: Frontend time: 0.219s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09219642 510.52795844 + layer.0.v_cache 0.00001621 0.05273299 + layer.1.k_cache 0.17840836 35.52781964 + layer.1.v_cache 0.00000639 0.01911997 + layer.2.k_cache 0.00922235 6.38319792 + layer.2.v_cache 0.00002142 0.05890549 + layer.3.k_cache 0.01957878 22.87964487 + layer.3.v_cache 0.00002124 0.06823590 + layer.4.k_cache 0.00068177 1.83252573 + layer.4.v_cache 0.00005086 0.14009778 + layer.4.output 1.55401793 275.15373459 + ------------------------------------------------------------------------------------- + TOTAL 0.65754878 147.26861064 + (elements=1,714,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1714688 +Total Bytes 17832 +BPFP 0.0832 bits/point +EBPFP 0.1664 equivalent bits/point +MSE 147.268611 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.009s, Pack+Encode: 0.219s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 147.2686 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,612B, BPFP=0.0813 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,824B, BPFP=0.0919 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,936B, BPFP=0.0976 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,640B, BPFP=0.0827 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,832B, BPFP=0.0923 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,692B, BPFP=0.0853 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,744B, BPFP=0.0879 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,644B, BPFP=0.0829 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,004B, BPFP=0.1010 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,832B, BPFP=0.0923 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,772B, BPFP=0.0272 +⌛️ [2/4] FRONTEND: Frontend time: 0.268s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11169702 515.11169355 + layer.0.v_cache 0.00001595 0.05247928 + layer.1.k_cache 0.54955656 35.30940965 + layer.1.v_cache 0.00000603 0.01879489 + layer.2.k_cache 0.01751511 6.32488856 + layer.2.v_cache 0.00002011 0.05883399 + layer.3.k_cache 0.01650532 22.78367251 + layer.3.v_cache 0.00001967 0.06549535 + layer.4.k_cache 0.00077751 1.85464596 + layer.4.v_cache 0.00005008 0.13663566 + layer.4.output 0.04306356 177.49143145 + ------------------------------------------------------------------------------------- + TOTAL 0.05868284 107.30332762 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 21532 +BPFP 0.0638 bits/point +EBPFP 0.1277 equivalent bits/point +MSE 107.303328 +---------------------- -------------------------------------------------------- +Time: 0.691s Load: 0.012s, Pack+Encode: 0.268s, Decode+Unpack: 0.411s +---------------------- -------------------------------------------------------- +💾 Converting with 107.3033 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,204B, BPFP=0.0863 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,452B, BPFP=0.1041 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,500B, BPFP=0.1075 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,248B, BPFP=0.0894 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,424B, BPFP=0.1021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,252B, BPFP=0.0897 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,356B, BPFP=0.0972 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,276B, BPFP=0.0915 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,552B, BPFP=0.1112 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,412B, BPFP=0.1012 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,464B, BPFP=0.0355 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.308s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09832749 511.63137901 + layer.0.v_cache 0.00001658 0.05608569 + layer.1.k_cache 0.33493462 35.29501147 + layer.1.v_cache 0.00000656 0.02016571 + layer.2.k_cache 0.02762392 6.32138481 + layer.2.v_cache 0.00002012 0.05839017 + layer.3.k_cache 0.01647770 22.52408257 + layer.3.v_cache 0.00001951 0.06939839 + layer.4.k_cache 0.00067889 1.86256227 + layer.4.v_cache 0.00004874 0.13551646 + layer.4.output 1.40435175 249.62350508 + ------------------------------------------------------------------------------------- + TOTAL 0.60638920 136.78461836 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 17140 +BPFP 0.0723 bits/point +EBPFP 0.1445 equivalent bits/point +MSE 136.784618 +---------------------- -------------------------------------------------------- +Time: 0.540s Load: 0.008s, Pack+Encode: 0.224s, Decode+Unpack: 0.308s +---------------------- -------------------------------------------------------- +💾 Converting with 136.7846 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,196B, BPFP=0.0827 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,372B, BPFP=0.0949 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,504B, BPFP=0.1040 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,244B, BPFP=0.0860 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,392B, BPFP=0.0962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,300B, BPFP=0.0899 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,368B, BPFP=0.0946 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,296B, BPFP=0.0896 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,584B, BPFP=0.1095 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,404B, BPFP=0.0971 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,192B, BPFP=0.0315 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102250 507.90185979 + layer.0.v_cache 0.00001721 0.05222496 + layer.1.k_cache 0.37371607 35.38497304 + layer.1.v_cache 0.00000658 0.01925330 + layer.2.k_cache 0.01127678 6.33116271 + layer.2.v_cache 0.00001998 0.05763585 + layer.3.k_cache 0.02136800 22.53393555 + layer.3.v_cache 0.00001963 0.06763778 + layer.4.k_cache 0.00069450 1.78701809 + layer.4.v_cache 0.00005147 0.13736750 + layer.4.output 1.35464589 240.47536741 + ------------------------------------------------------------------------------------- + TOTAL 0.58945376 132.80003768 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 16852 +BPFP 0.0685 bits/point +EBPFP 0.1371 equivalent bits/point +MSE 132.800038 +---------------------- -------------------------------------------------------- +Time: 0.535s Load: 0.008s, Pack+Encode: 0.220s, Decode+Unpack: 0.307s +---------------------- -------------------------------------------------------- +💾 Converting with 132.8000 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,260B, BPFP=0.0784 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,412B, BPFP=0.0879 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,500B, BPFP=0.0934 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,252B, BPFP=0.0779 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,400B, BPFP=0.0872 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,296B, BPFP=0.0807 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,352B, BPFP=0.0842 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,288B, BPFP=0.0802 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,560B, BPFP=0.0971 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,440B, BPFP=0.0896 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,252B, BPFP=0.0289 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.310s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11871604 502.57454557 + layer.0.v_cache 0.00001755 0.05376845 + layer.1.k_cache 0.46315337 34.89447289 + layer.1.v_cache 0.00000588 0.01899526 + layer.2.k_cache 0.02202094 6.23568811 + layer.2.v_cache 0.00001989 0.05780424 + layer.3.k_cache 0.01278766 22.61537055 + layer.3.v_cache 0.00002092 0.07021653 + layer.4.k_cache 0.00074254 1.83160485 + layer.4.v_cache 0.00004925 0.13551030 + layer.4.output 1.21975157 216.18565026 + ------------------------------------------------------------------------------------- + TOTAL 0.53857618 122.45808991 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 17012 +BPFP 0.0623 bits/point +EBPFP 0.1246 equivalent bits/point +MSE 122.458090 +---------------------- -------------------------------------------------------- +Time: 0.541s Load: 0.011s, Pack+Encode: 0.220s, Decode+Unpack: 0.310s +---------------------- -------------------------------------------------------- +💾 Converting with 122.4581 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 195, 128) +Output shape: (1, 195, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.output: torch.Size([1, 195, 3584]) -> torch.Size([1, 1, 195, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,232B, BPFP=0.0987 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,420B, BPFP=0.1138 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,548B, BPFP=0.1240 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,224B, BPFP=0.0981 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,412B, BPFP=0.1131 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,260B, BPFP=0.1010 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,360B, BPFP=0.1090 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,336B, BPFP=0.1071 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,604B, BPFP=0.1285 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,540B, BPFP=0.1234 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,504B, BPFP=0.0401 +⌛️ [2/4] FRONTEND: Frontend time: 0.218s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.303s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605874 509.72500000 + layer.0.v_cache 0.00001676 0.05383568 + layer.1.k_cache 0.13872500 35.65053335 + layer.1.v_cache 0.00000593 0.01935945 + layer.2.k_cache 0.01222654 6.41917193 + layer.2.v_cache 0.00001968 0.05811192 + layer.3.k_cache 0.01415822 22.73741737 + layer.3.v_cache 0.00002127 0.07319488 + layer.4.k_cache 0.00071127 1.83177459 + layer.4.v_cache 0.00005419 0.14570682 + layer.4.output 1.56987251 280.86916209 + ------------------------------------------------------------------------------------- + TOTAL 0.66300619 149.57636709 + (elements=1,697,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1697280 +Total Bytes 17440 +BPFP 0.0822 bits/point +EBPFP 0.1644 equivalent bits/point +MSE 149.576367 +---------------------- -------------------------------------------------------- +Time: 0.528s Load: 0.007s, Pack+Encode: 0.218s, Decode+Unpack: 0.303s +---------------------- -------------------------------------------------------- +💾 Converting with 149.5764 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,060B, BPFP=0.0910 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,132B, BPFP=0.0972 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,248B, BPFP=0.1071 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,068B, BPFP=0.0917 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,188B, BPFP=0.1020 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,088B, BPFP=0.0934 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,152B, BPFP=0.0989 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,096B, BPFP=0.0941 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,272B, BPFP=0.1092 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,168B, BPFP=0.1003 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,704B, BPFP=0.0332 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.256s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11176265 519.04180975 + layer.0.v_cache 0.00001802 0.05666245 + layer.1.k_cache 0.18491925 36.06413118 + layer.1.v_cache 0.00000608 0.02046996 + layer.2.k_cache 0.01999094 6.55710007 + layer.2.v_cache 0.00002176 0.06232004 + layer.3.k_cache 0.03910406 23.32376266 + layer.3.v_cache 0.00002004 0.07280652 + layer.4.k_cache 0.00067358 1.89983242 + layer.4.v_cache 0.00005212 0.14453869 + layer.4.output 0.00886795 306.03826531 + ------------------------------------------------------------------------------------- + TOTAL 0.02462613 160.55948770 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 14176 +BPFP 0.0716 bits/point +EBPFP 0.1432 equivalent bits/point +MSE 160.559488 +---------------------- -------------------------------------------------------- +Time: 0.507s Load: 0.006s, Pack+Encode: 0.245s, Decode+Unpack: 0.256s +---------------------- -------------------------------------------------------- +💾 Converting with 160.5595 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,248B, BPFP=0.0915 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,436B, BPFP=0.1053 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,556B, BPFP=0.1141 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,336B, BPFP=0.0980 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,468B, BPFP=0.1077 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,380B, BPFP=0.1012 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,408B, BPFP=0.1033 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,368B, BPFP=0.1004 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,616B, BPFP=0.1185 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,508B, BPFP=0.1106 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,508B, BPFP=0.0368 +⌛️ [2/4] FRONTEND: Frontend time: 0.219s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15089922 515.40470951 + layer.0.v_cache 0.00001686 0.05502107 + layer.1.k_cache 0.31217133 35.25726232 + layer.1.v_cache 0.00000593 0.01950314 + layer.2.k_cache 0.01305961 6.29677585 + layer.2.v_cache 0.00001999 0.05900507 + layer.3.k_cache 0.01178671 22.70651775 + layer.3.v_cache 0.00002039 0.07120289 + layer.4.k_cache 0.00067606 1.81348373 + layer.4.v_cache 0.00005285 0.13916128 + layer.4.output 1.43724915 255.19382964 + ------------------------------------------------------------------------------------- + TOTAL 0.62055606 139.30467354 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 17832 +BPFP 0.0769 bits/point +EBPFP 0.1539 equivalent bits/point +MSE 139.304674 +---------------------- -------------------------------------------------------- +Time: 0.536s Load: 0.009s, Pack+Encode: 0.219s, Decode+Unpack: 0.307s +---------------------- -------------------------------------------------------- +💾 Converting with 139.3047 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 187, 128) +Output shape: (1, 187, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.output: torch.Size([1, 187, 3584]) -> torch.Size([1, 1, 187, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 976B, BPFP=0.0816 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,076B, BPFP=0.0899 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,160B, BPFP=0.0969 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 964B, BPFP=0.0805 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,084B, BPFP=0.0906 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,012B, BPFP=0.0846 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,048B, BPFP=0.0876 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,004B, BPFP=0.0839 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,184B, BPFP=0.0989 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,092B, BPFP=0.0912 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,676B, BPFP=0.0319 +⌛️ [2/4] FRONTEND: Frontend time: 0.188s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.252s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08258281 504.92956217 + layer.0.v_cache 0.00001733 0.05738355 + layer.1.k_cache 0.10049684 35.19855813 + layer.1.v_cache 0.00000647 0.02012152 + layer.2.k_cache 0.01472783 6.42139520 + layer.2.v_cache 0.00002140 0.06407315 + layer.3.k_cache 0.02125903 22.75032900 + layer.3.v_cache 0.00002180 0.07816862 + layer.4.k_cache 0.00070831 1.92152413 + layer.4.v_cache 0.00005280 0.14759212 + layer.4.output 0.00866398 294.18425325 + ------------------------------------------------------------------------------------- + TOTAL 0.01650250 154.75755767 + (elements=1,627,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1627648 +Total Bytes 13276 +BPFP 0.0653 bits/point +EBPFP 0.1305 equivalent bits/point +MSE 154.757558 +---------------------- -------------------------------------------------------- +Time: 0.446s Load: 0.006s, Pack+Encode: 0.188s, Decode+Unpack: 0.252s +---------------------- -------------------------------------------------------- +💾 Converting with 154.7576 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,252B, BPFP=0.0927 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,420B, BPFP=0.1052 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,576B, BPFP=0.1167 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,336B, BPFP=0.0989 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,452B, BPFP=0.1075 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,340B, BPFP=0.0992 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,408B, BPFP=0.1043 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,392B, BPFP=0.1031 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,628B, BPFP=0.1206 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,460B, BPFP=0.1081 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,544B, BPFP=0.0375 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.308s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11071491 511.36822423 + layer.0.v_cache 0.00001804 0.05501601 + layer.1.k_cache 0.26784508 35.25871964 + layer.1.v_cache 0.00000610 0.01984988 + layer.2.k_cache 0.01369454 6.22117868 + layer.2.v_cache 0.00002102 0.06015259 + layer.3.k_cache 0.01444715 22.73511552 + layer.3.v_cache 0.00002041 0.07242956 + layer.4.k_cache 0.00067564 1.83059070 + layer.4.v_cache 0.00005248 0.14240325 + layer.4.output 1.45091435 256.96398951 + ------------------------------------------------------------------------------------- + TOTAL 0.62140564 139.79480039 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 17808 +BPFP 0.0776 bits/point +EBPFP 0.1551 equivalent bits/point +MSE 139.794800 +---------------------- -------------------------------------------------------- +Time: 0.539s Load: 0.007s, Pack+Encode: 0.224s, Decode+Unpack: 0.308s +---------------------- -------------------------------------------------------- +💾 Converting with 139.7948 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,200B, BPFP=0.0841 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,376B, BPFP=0.0964 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,492B, BPFP=0.1045 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,248B, BPFP=0.0874 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,392B, BPFP=0.0975 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,284B, BPFP=0.0900 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,336B, BPFP=0.0936 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,296B, BPFP=0.0908 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,600B, BPFP=0.1121 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,452B, BPFP=0.1017 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,280B, BPFP=0.0328 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.310s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13389712 511.10468049 + layer.0.v_cache 0.00001748 0.05423585 + layer.1.k_cache 0.36644789 35.23060013 + layer.1.v_cache 0.00000587 0.01907158 + layer.2.k_cache 0.02260907 6.33738428 + layer.2.v_cache 0.00001979 0.05812423 + layer.3.k_cache 0.02550721 22.64311239 + layer.3.v_cache 0.00001968 0.06984038 + layer.4.k_cache 0.00067529 1.80094136 + layer.4.v_cache 0.00006705 0.14185662 + layer.4.output 1.37284074 242.83626281 + ------------------------------------------------------------------------------------- + TOTAL 0.59759716 133.95962865 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 16956 +BPFP 0.0699 bits/point +EBPFP 0.1398 equivalent bits/point +MSE 133.959629 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.010s, Pack+Encode: 0.222s, Decode+Unpack: 0.310s +---------------------- -------------------------------------------------------- +💾 Converting with 133.9596 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 406, 128) +Output shape: (1, 406, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.output: torch.Size([1, 406, 3584]) -> torch.Size([1, 1, 406, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,080B, BPFP=0.0800 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,400B, BPFP=0.0924 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,604B, BPFP=0.1002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,200B, BPFP=0.0847 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,464B, BPFP=0.0948 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,280B, BPFP=0.0877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,336B, BPFP=0.0899 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,268B, BPFP=0.0873 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,736B, BPFP=0.1053 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,480B, BPFP=0.0954 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,636B, BPFP=0.0310 +⌛️ [2/4] FRONTEND: Frontend time: 0.438s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.481s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10474718 506.75207820 + layer.0.v_cache 0.00001775 0.05486628 + layer.1.k_cache 0.89873824 34.86887123 + layer.1.v_cache 0.00000687 0.02028315 + layer.2.k_cache 0.02602503 6.21277878 + layer.2.v_cache 0.00002164 0.06360594 + layer.3.k_cache 0.01558245 22.57680929 + layer.3.v_cache 0.00002201 0.07184873 + layer.4.k_cache 0.00072958 1.84715609 + layer.4.v_cache 0.00005327 0.14139485 + layer.4.output 0.00651461 137.68449815 + ------------------------------------------------------------------------------------- + TOTAL 0.06420861 90.37653998 + (elements=3,533,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3533824 +Total Bytes 29484 +BPFP 0.0667 bits/point +EBPFP 0.1335 equivalent bits/point +MSE 90.376540 +---------------------- -------------------------------------------------------- +Time: 0.933s Load: 0.014s, Pack+Encode: 0.438s, Decode+Unpack: 0.481s +---------------------- -------------------------------------------------------- +💾 Converting with 90.3765 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,332B, BPFP=0.0849 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,480B, BPFP=0.0944 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,580B, BPFP=0.1008 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,376B, BPFP=0.0878 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,520B, BPFP=0.0969 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,412B, BPFP=0.0901 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,472B, BPFP=0.0939 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,404B, BPFP=0.0895 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,632B, BPFP=0.1041 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,496B, BPFP=0.0954 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,316B, BPFP=0.0302 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.322s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14414167 509.18673469 + layer.0.v_cache 0.00001717 0.05518672 + layer.1.k_cache 0.44179790 35.39900351 + layer.1.v_cache 0.00000667 0.02086448 + layer.2.k_cache 0.03016641 6.39626814 + layer.2.v_cache 0.00002043 0.06132525 + layer.3.k_cache 0.03726396 22.93553890 + layer.3.v_cache 0.00002099 0.07452955 + layer.4.k_cache 0.00070653 1.85239955 + layer.4.v_cache 0.00004910 0.14025170 + layer.4.output 1.24963187 222.45712464 + ------------------------------------------------------------------------------------- + TOTAL 0.55303611 125.48952794 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 18020 +BPFP 0.0676 bits/point +EBPFP 0.1352 equivalent bits/point +MSE 125.489528 +---------------------- -------------------------------------------------------- +Time: 0.567s Load: 0.015s, Pack+Encode: 0.230s, Decode+Unpack: 0.322s +---------------------- -------------------------------------------------------- +💾 Converting with 125.4895 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,264B, BPFP=0.0992 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,400B, BPFP=0.1099 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,584B, BPFP=0.1244 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,300B, BPFP=0.1021 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,444B, BPFP=0.1134 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,304B, BPFP=0.1024 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,424B, BPFP=0.1118 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,308B, BPFP=0.1027 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,604B, BPFP=0.1259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,552B, BPFP=0.1219 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,896B, BPFP=0.0437 +⌛️ [2/4] FRONTEND: Frontend time: 0.290s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.337s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15985105 504.22259736 + layer.0.v_cache 0.00001905 0.05588543 + layer.1.k_cache 0.22745673 35.48125638 + layer.1.v_cache 0.00000590 0.01968466 + layer.2.k_cache 0.01549998 6.41629412 + layer.2.v_cache 0.00002003 0.05927306 + layer.3.k_cache 0.01333464 22.63845841 + layer.3.v_cache 0.00002034 0.07175612 + layer.4.k_cache 0.00067076 1.85758198 + layer.4.v_cache 0.00004949 0.13842912 + layer.4.output 1.53832061 272.67989501 + ------------------------------------------------------------------------------------- + TOTAL 0.65795131 145.86591069 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 18080 +BPFP 0.0835 bits/point +EBPFP 0.1670 equivalent bits/point +MSE 145.865911 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.007s, Pack+Encode: 0.290s, Decode+Unpack: 0.337s +---------------------- -------------------------------------------------------- +💾 Converting with 145.8659 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 237, 128) +Output shape: (1, 237, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.output: torch.Size([1, 237, 3584]) -> torch.Size([1, 1, 237, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.0876 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,476B, BPFP=0.0973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,604B, BPFP=0.1057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,320B, BPFP=0.0870 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,544B, BPFP=0.1018 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,404B, BPFP=0.0926 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,460B, BPFP=0.0963 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,388B, BPFP=0.0915 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,652B, BPFP=0.1089 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,528B, BPFP=0.1007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,540B, BPFP=0.0333 +⌛️ [2/4] FRONTEND: Frontend time: 0.292s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.332s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351868 511.33072917 + layer.0.v_cache 0.00001655 0.05406072 + layer.1.k_cache 0.34375901 35.52102288 + layer.1.v_cache 0.00000619 0.01992602 + layer.2.k_cache 0.01282014 6.37453696 + layer.2.v_cache 0.00002104 0.06099229 + layer.3.k_cache 0.03241006 22.89725491 + layer.3.v_cache 0.00001985 0.06939008 + layer.4.k_cache 0.00067667 1.85584552 + layer.4.v_cache 0.00005279 0.14114892 + layer.4.output 1.29178675 228.85998719 + ------------------------------------------------------------------------------------- + TOTAL 0.56210637 128.25557752 + (elements=2,062,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2062848 +Total Bytes 18244 +BPFP 0.0708 bits/point +EBPFP 0.1415 equivalent bits/point +MSE 128.255578 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.008s, Pack+Encode: 0.292s, Decode+Unpack: 0.332s +---------------------- -------------------------------------------------------- +💾 Converting with 128.2556 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,268B, BPFP=0.0922 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,440B, BPFP=0.1047 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,544B, BPFP=0.1122 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,308B, BPFP=0.0951 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,460B, BPFP=0.1061 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,348B, BPFP=0.0980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,396B, BPFP=0.1015 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,336B, BPFP=0.0971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,612B, BPFP=0.1172 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,492B, BPFP=0.1084 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,628B, BPFP=0.0377 +⌛️ [2/4] FRONTEND: Frontend time: 0.274s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11023432 512.61224564 + layer.0.v_cache 0.00001677 0.05364892 + layer.1.k_cache 0.30291741 35.39254179 + layer.1.v_cache 0.00000597 0.01918503 + layer.2.k_cache 0.03464115 6.26251590 + layer.2.v_cache 0.00001998 0.05730512 + layer.3.k_cache 0.01330193 22.74817633 + layer.3.v_cache 0.00002112 0.06722585 + layer.4.k_cache 0.00069379 1.81078789 + layer.4.v_cache 0.00005021 0.13594896 + layer.4.output 1.42388847 253.89968854 + ------------------------------------------------------------------------------------- + TOTAL 0.61347776 138.61514125 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 17832 +BPFP 0.0762 bits/point +EBPFP 0.1525 equivalent bits/point +MSE 138.615141 +---------------------- -------------------------------------------------------- +Time: 0.626s Load: 0.007s, Pack+Encode: 0.274s, Decode+Unpack: 0.345s +---------------------- -------------------------------------------------------- +💾 Converting with 138.6151 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 231, 128) +Output shape: (1, 231, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.output: torch.Size([1, 231, 3584]) -> torch.Size([1, 1, 231, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,320B, BPFP=0.0893 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,452B, BPFP=0.0982 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,576B, BPFP=0.1066 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,376B, BPFP=0.0931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,496B, BPFP=0.1012 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,400B, BPFP=0.0947 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,448B, BPFP=0.0979 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,396B, BPFP=0.0944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,632B, BPFP=0.1104 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,468B, BPFP=0.0993 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,708B, BPFP=0.0358 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.323s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13257201 507.05350379 + layer.0.v_cache 0.00001701 0.05620176 + layer.1.k_cache 0.40462213 35.30016191 + layer.1.v_cache 0.00000658 0.02090311 + layer.2.k_cache 0.02649644 6.32277927 + layer.2.v_cache 0.00002056 0.06220677 + layer.3.k_cache 0.06147716 22.76360423 + layer.3.v_cache 0.00002105 0.07483957 + layer.4.k_cache 0.00069437 1.86898090 + layer.4.v_cache 0.00005136 0.14384830 + layer.4.output 1.32536713 235.02108457 + ------------------------------------------------------------------------------------- + TOTAL 0.58256168 130.51850715 + (elements=2,010,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2010624 +Total Bytes 18272 +BPFP 0.0727 bits/point +EBPFP 0.1454 equivalent bits/point +MSE 130.518507 +---------------------- -------------------------------------------------------- +Time: 0.593s Load: 0.008s, Pack+Encode: 0.262s, Decode+Unpack: 0.323s +---------------------- -------------------------------------------------------- +💾 Converting with 130.5185 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,136B, BPFP=0.0696 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,260B, BPFP=0.0772 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,420B, BPFP=0.0870 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,100B, BPFP=0.0674 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,312B, BPFP=0.0804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,164B, BPFP=0.0713 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,228B, BPFP=0.0752 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,140B, BPFP=0.0699 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,448B, BPFP=0.0887 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,336B, BPFP=0.0819 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,568B, BPFP=0.0225 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12841698 500.52812500 + layer.0.v_cache 0.00001718 0.05557779 + layer.1.k_cache 0.45450200 34.27527574 + layer.1.v_cache 0.00000627 0.02008840 + layer.2.k_cache 0.01434461 6.13373640 + layer.2.v_cache 0.00002215 0.06289368 + layer.3.k_cache 0.03530993 22.13776425 + layer.3.v_cache 0.00002139 0.07255941 + layer.4.k_cache 0.00068344 1.79199769 + layer.4.v_cache 0.00005073 0.13979983 + layer.4.output 1.20064817 212.50427171 + ------------------------------------------------------------------------------------- + TOTAL 0.53164187 120.74986589 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 15112 +BPFP 0.0545 bits/point +EBPFP 0.1089 equivalent bits/point +MSE 120.749866 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.009s, Pack+Encode: 0.228s, Decode+Unpack: 0.312s +---------------------- -------------------------------------------------------- +💾 Converting with 120.7499 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,500B, BPFP=0.0831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,704B, BPFP=0.0944 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,852B, BPFP=0.1026 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,524B, BPFP=0.0844 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,724B, BPFP=0.0955 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,584B, BPFP=0.0878 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,692B, BPFP=0.0938 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,612B, BPFP=0.0893 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,940B, BPFP=0.1075 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,764B, BPFP=0.0977 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,988B, BPFP=0.0316 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13731835 507.26723183 + layer.0.v_cache 0.00001709 0.05408975 + layer.1.k_cache 0.55879569 34.93124238 + layer.1.v_cache 0.00000637 0.01971171 + layer.2.k_cache 0.01348169 6.28458356 + layer.2.v_cache 0.00002375 0.06060278 + layer.3.k_cache 0.01306405 22.49157455 + layer.3.v_cache 0.00002005 0.06887129 + layer.4.k_cache 0.00072533 1.83547368 + layer.4.v_cache 0.00005384 0.14076858 + layer.4.output 0.00472113 197.55241578 + ------------------------------------------------------------------------------------- + TOTAL 0.04450318 115.06006239 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 20884 +BPFP 0.0681 bits/point +EBPFP 0.1361 equivalent bits/point +MSE 115.060062 +---------------------- -------------------------------------------------------- +Time: 0.625s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 115.0601 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,588B, BPFP=0.0824 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,744B, BPFP=0.0905 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,948B, BPFP=0.1011 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,616B, BPFP=0.0839 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,864B, BPFP=0.0968 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,680B, BPFP=0.0872 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,732B, BPFP=0.0899 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,696B, BPFP=0.0880 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,020B, BPFP=0.1049 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,828B, BPFP=0.0949 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,112B, BPFP=0.0305 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13972237 509.24818314 + layer.0.v_cache 0.00001689 0.05524093 + layer.1.k_cache 0.51719427 35.30983960 + layer.1.v_cache 0.00000613 0.02027905 + layer.2.k_cache 0.02019965 6.37704316 + layer.2.v_cache 0.00002053 0.06156048 + layer.3.k_cache 0.01759115 22.65470405 + layer.3.v_cache 0.00002125 0.07329977 + layer.4.k_cache 0.00069604 1.87642997 + layer.4.v_cache 0.00005872 0.14101222 + layer.4.output 0.04434862 180.39033579 + ------------------------------------------------------------------------------------- + TOTAL 0.05917455 108.14999664 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 21828 +BPFP 0.0667 bits/point +EBPFP 0.1333 equivalent bits/point +MSE 108.149997 +---------------------- -------------------------------------------------------- +Time: 0.625s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 108.1500 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,260B, BPFP=0.0989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,400B, BPFP=0.1099 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,580B, BPFP=0.1241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,264B, BPFP=0.0992 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,444B, BPFP=0.1134 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,348B, BPFP=0.1058 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,424B, BPFP=0.1118 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,320B, BPFP=0.1036 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,628B, BPFP=0.1278 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,472B, BPFP=0.1156 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,932B, BPFP=0.0441 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.310s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13044639 507.20991677 + layer.0.v_cache 0.00001737 0.05428815 + layer.1.k_cache 0.22565776 35.37684026 + layer.1.v_cache 0.00000620 0.01986069 + layer.2.k_cache 0.01998205 6.41963111 + layer.2.v_cache 0.00001989 0.05849625 + layer.3.k_cache 0.00629339 22.68271779 + layer.3.v_cache 0.00002050 0.07065810 + layer.4.k_cache 0.00068118 1.82381267 + layer.4.v_cache 0.00005692 0.13633246 + layer.4.output 1.53833376 272.51579325 + ------------------------------------------------------------------------------------- + TOTAL 0.65597165 145.96841806 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 18072 +BPFP 0.0835 bits/point +EBPFP 0.1669 equivalent bits/point +MSE 145.968418 +---------------------- -------------------------------------------------------- +Time: 0.537s Load: 0.007s, Pack+Encode: 0.220s, Decode+Unpack: 0.310s +---------------------- -------------------------------------------------------- +💾 Converting with 145.9684 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 246, 128) +Output shape: (1, 246, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.output: torch.Size([1, 246, 3584]) -> torch.Size([1, 1, 246, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,336B, BPFP=0.0849 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,448B, BPFP=0.0920 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,588B, BPFP=0.1009 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,340B, BPFP=0.0851 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,500B, BPFP=0.0953 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,388B, BPFP=0.0882 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,456B, BPFP=0.0925 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,400B, BPFP=0.0889 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,632B, BPFP=0.1037 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,504B, BPFP=0.0955 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,276B, BPFP=0.0297 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131485 511.53074187 + layer.0.v_cache 0.00001613 0.05436591 + layer.1.k_cache 0.49701033 35.46880955 + layer.1.v_cache 0.00000620 0.01969512 + layer.2.k_cache 0.01298508 6.46298342 + layer.2.v_cache 0.00002120 0.05961214 + layer.3.k_cache 0.02762915 22.85342432 + layer.3.v_cache 0.00002136 0.07305369 + layer.4.k_cache 0.00068392 1.83544773 + layer.4.v_cache 0.00006372 0.13661281 + layer.4.output 1.24453647 223.26782085 + ------------------------------------------------------------------------------------- + TOTAL 0.55302984 125.96291132 + (elements=2,141,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2141184 +Total Bytes 17868 +BPFP 0.0668 bits/point +EBPFP 0.1335 equivalent bits/point +MSE 125.962911 +---------------------- -------------------------------------------------------- +Time: 0.537s Load: 0.008s, Pack+Encode: 0.222s, Decode+Unpack: 0.307s +---------------------- -------------------------------------------------------- +💾 Converting with 125.9629 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,264B, BPFP=0.0988 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,420B, BPFP=0.1109 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,596B, BPFP=0.1247 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,300B, BPFP=0.1016 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,456B, BPFP=0.1138 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,332B, BPFP=0.1041 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,400B, BPFP=0.1094 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,280B, BPFP=0.1000 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,656B, BPFP=0.1294 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,516B, BPFP=0.1184 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,780B, BPFP=0.0422 +⌛️ [2/4] FRONTEND: Frontend time: 0.219s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11599884 509.14144531 + layer.0.v_cache 0.00001671 0.05163140 + layer.1.k_cache 0.22559120 35.42832031 + layer.1.v_cache 0.00000595 0.01884878 + layer.2.k_cache 0.01146694 6.39024963 + layer.2.v_cache 0.00002201 0.05807664 + layer.3.k_cache 0.03668072 22.81620605 + layer.3.v_cache 0.00002131 0.06660142 + layer.4.k_cache 0.00066762 1.77904480 + layer.4.v_cache 0.00005146 0.13713012 + layer.4.output 1.53061454 270.86875000 + ------------------------------------------------------------------------------------- + TOTAL 0.65322497 145.40992967 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 18000 +BPFP 0.0827 bits/point +EBPFP 0.1654 equivalent bits/point +MSE 145.409930 +---------------------- -------------------------------------------------------- +Time: 0.534s Load: 0.008s, Pack+Encode: 0.219s, Decode+Unpack: 0.307s +---------------------- -------------------------------------------------------- +💾 Converting with 145.4099 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,268B, BPFP=0.0865 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,376B, BPFP=0.0939 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,540B, BPFP=0.1051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,284B, BPFP=0.0876 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,452B, BPFP=0.0991 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,324B, BPFP=0.0903 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,392B, BPFP=0.0950 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,348B, BPFP=0.0920 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,588B, BPFP=0.1084 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,448B, BPFP=0.0988 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,512B, BPFP=0.0342 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.308s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605523 505.94196916 + layer.0.v_cache 0.00001866 0.05496313 + layer.1.k_cache 0.31881074 35.42511173 + layer.1.v_cache 0.00000653 0.02043372 + layer.2.k_cache 0.02091764 6.36919393 + layer.2.v_cache 0.00002113 0.06177984 + layer.3.k_cache 0.03221333 22.88940536 + layer.3.v_cache 0.00002185 0.07271116 + layer.4.k_cache 0.00068277 1.85278827 + layer.4.v_cache 0.00005596 0.14069110 + layer.4.output 1.33695214 236.86685122 + ------------------------------------------------------------------------------------- + TOTAL 0.57926287 131.22923564 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 17532 +BPFP 0.0704 bits/point +EBPFP 0.1407 equivalent bits/point +MSE 131.229236 +---------------------- -------------------------------------------------------- +Time: 0.536s Load: 0.008s, Pack+Encode: 0.220s, Decode+Unpack: 0.308s +---------------------- -------------------------------------------------------- +💾 Converting with 131.2292 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 192, 128) +Output shape: (1, 192, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.output: torch.Size([1, 192, 3584]) -> torch.Size([1, 1, 192, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 836B, BPFP=0.0680 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 916B, BPFP=0.0745 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,072B, BPFP=0.0872 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 808B, BPFP=0.0658 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 988B, BPFP=0.0804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 860B, BPFP=0.0700 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 924B, BPFP=0.0752 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 852B, BPFP=0.0693 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,096B, BPFP=0.0892 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,004B, BPFP=0.0817 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,824B, BPFP=0.0212 +⌛️ [2/4] FRONTEND: Frontend time: 0.190s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.253s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12444629 498.94222005 + layer.0.v_cache 0.00001540 0.04949565 + layer.1.k_cache 0.18411521 34.11889394 + layer.1.v_cache 0.00000585 0.01843807 + layer.2.k_cache 0.01124177 5.99246407 + layer.2.v_cache 0.00002179 0.05764349 + layer.3.k_cache 0.01616090 22.05768331 + layer.3.v_cache 0.00002117 0.06814589 + layer.4.k_cache 0.00068507 1.76230605 + layer.4.v_cache 0.00005275 0.13294306 + layer.4.output 0.00839530 286.14269438 + ------------------------------------------------------------------------------------- + TOTAL 0.02326666 150.95288790 + (elements=1,671,168) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1671168 +Total Bytes 11180 +BPFP 0.0535 bits/point +EBPFP 0.1070 equivalent bits/point +MSE 150.952888 +---------------------- -------------------------------------------------------- +Time: 0.450s Load: 0.007s, Pack+Encode: 0.190s, Decode+Unpack: 0.253s +---------------------- -------------------------------------------------------- +💾 Converting with 150.9529 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 996B, BPFP=0.0938 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,104B, BPFP=0.1039 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,204B, BPFP=0.1133 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,020B, BPFP=0.0960 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,132B, BPFP=0.1066 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,080B, BPFP=0.1017 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,124B, BPFP=0.1058 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,052B, BPFP=0.0990 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,232B, BPFP=0.1160 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,156B, BPFP=0.1088 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,828B, BPFP=0.0380 +⌛️ [2/4] FRONTEND: Frontend time: 0.195s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.262s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09999813 512.14956702 + layer.0.v_cache 0.00001899 0.05242536 + layer.1.k_cache 0.12989934 35.86571089 + layer.1.v_cache 0.00000610 0.01888279 + layer.2.k_cache 0.00966061 6.45444939 + layer.2.v_cache 0.00001994 0.05800144 + layer.3.k_cache 0.03384685 22.71265531 + layer.3.v_cache 0.00002040 0.06665887 + layer.4.k_cache 0.00066502 1.80794406 + layer.4.v_cache 0.00005200 0.13484353 + layer.4.output 0.00959846 331.53383176 + ------------------------------------------------------------------------------------- + TOTAL 0.02008098 170.59164476 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 13928 +BPFP 0.0771 bits/point +EBPFP 0.1542 equivalent bits/point +MSE 170.591645 +---------------------- -------------------------------------------------------- +Time: 0.464s Load: 0.007s, Pack+Encode: 0.195s, Decode+Unpack: 0.262s +---------------------- -------------------------------------------------------- +💾 Converting with 170.5916 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 428, 128) +Output shape: (1, 428, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.output: torch.Size([1, 428, 3584]) -> torch.Size([1, 1, 428, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,168B, BPFP=0.0791 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,472B, BPFP=0.0902 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,648B, BPFP=0.0967 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,232B, BPFP=0.0815 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,544B, BPFP=0.0929 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,328B, BPFP=0.0850 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,368B, BPFP=0.0864 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,268B, BPFP=0.0828 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,784B, BPFP=0.1016 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,536B, BPFP=0.0926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,464B, BPFP=0.0285 +⌛️ [2/4] FRONTEND: Frontend time: 0.325s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.491s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13760449 504.74368429 + layer.0.v_cache 0.00001624 0.05394513 + layer.1.k_cache 0.89392254 34.85939097 + layer.1.v_cache 0.00000615 0.01938972 + layer.2.k_cache 0.03917003 6.25361876 + layer.2.v_cache 0.00002113 0.06048543 + layer.3.k_cache 0.03749864 22.63743976 + layer.3.v_cache 0.00002059 0.06821441 + layer.4.k_cache 0.00086919 1.91401943 + layer.4.v_cache 0.00005177 0.13759226 + layer.4.output 0.00617707 129.71397280 + ------------------------------------------------------------------------------------- + TOTAL 0.06778943 86.98503469 + (elements=3,725,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3725312 +Total Bytes 29812 +BPFP 0.0640 bits/point +EBPFP 0.1280 equivalent bits/point +MSE 86.985035 +---------------------- -------------------------------------------------------- +Time: 0.829s Load: 0.013s, Pack+Encode: 0.325s, Decode+Unpack: 0.491s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9850 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,492B, BPFP=0.0827 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,764B, BPFP=0.0977 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,856B, BPFP=0.1028 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,560B, BPFP=0.0864 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,740B, BPFP=0.0964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,580B, BPFP=0.0875 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,692B, BPFP=0.0938 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,620B, BPFP=0.0898 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,960B, BPFP=0.1086 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,748B, BPFP=0.0969 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,136B, BPFP=0.0327 +⌛️ [2/4] FRONTEND: Frontend time: 0.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.361s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385377 505.53418661 + layer.0.v_cache 0.00001705 0.05398007 + layer.1.k_cache 0.49372572 34.94599125 + layer.1.v_cache 0.00000626 0.01996039 + layer.2.k_cache 0.02422354 6.27663730 + layer.2.v_cache 0.00002046 0.06048287 + layer.3.k_cache 0.04153839 22.59666757 + layer.3.v_cache 0.00002075 0.07133119 + layer.4.k_cache 0.00074884 1.88106542 + layer.4.v_cache 0.00004954 0.13731685 + layer.4.output 0.00474379 197.74417427 + ------------------------------------------------------------------------------------- + TOTAL 0.04161240 115.04628467 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 21148 +BPFP 0.0689 bits/point +EBPFP 0.1379 equivalent bits/point +MSE 115.046285 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.011s, Pack+Encode: 0.267s, Decode+Unpack: 0.361s +---------------------- -------------------------------------------------------- +💾 Converting with 115.0463 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,832B, BPFP=0.0886 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,048B, BPFP=0.0991 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,200B, BPFP=0.1064 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,804B, BPFP=0.0873 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,060B, BPFP=0.0997 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,900B, BPFP=0.0919 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,964B, BPFP=0.0950 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,916B, BPFP=0.0927 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,324B, BPFP=0.1124 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,124B, BPFP=0.1027 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,844B, BPFP=0.0335 +⌛️ [2/4] FRONTEND: Frontend time: 0.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.426s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12985682 501.20767221 + layer.0.v_cache 0.00001645 0.05312780 + layer.1.k_cache 0.61935935 35.08256035 + layer.1.v_cache 0.00000656 0.01964983 + layer.2.k_cache 0.01492825 6.24139537 + layer.2.v_cache 0.00002178 0.06000462 + layer.3.k_cache 0.02579401 22.62763642 + layer.3.v_cache 0.00002068 0.06787842 + layer.4.k_cache 0.00071613 1.87112058 + layer.4.v_cache 0.00005057 0.13656930 + layer.4.output 0.04141478 169.81027477 + ------------------------------------------------------------------------------------- + TOTAL 0.06356906 103.29644343 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 25016 +BPFP 0.0712 bits/point +EBPFP 0.1424 equivalent bits/point +MSE 103.296443 +---------------------- -------------------------------------------------------- +Time: 0.772s Load: 0.012s, Pack+Encode: 0.334s, Decode+Unpack: 0.426s +---------------------- -------------------------------------------------------- +💾 Converting with 103.2964 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,476B, BPFP=0.0806 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,680B, BPFP=0.0918 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,824B, BPFP=0.0997 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,516B, BPFP=0.0828 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,728B, BPFP=0.0944 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,568B, BPFP=0.0857 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,672B, BPFP=0.0913 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,608B, BPFP=0.0878 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,952B, BPFP=0.1066 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,752B, BPFP=0.0957 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,840B, BPFP=0.0300 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.361s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12359551 504.09167395 + layer.0.v_cache 0.00001766 0.05569889 + layer.1.k_cache 0.52692803 35.01109047 + layer.1.v_cache 0.00000637 0.02016122 + layer.2.k_cache 0.02102186 6.26544531 + layer.2.v_cache 0.00002187 0.06125868 + layer.3.k_cache 0.03868174 22.56298145 + layer.3.v_cache 0.00002097 0.07368358 + layer.4.k_cache 0.00070799 1.84761399 + layer.4.v_cache 0.00005204 0.13990194 + layer.4.output 0.00469152 194.08049763 + ------------------------------------------------------------------------------------- + TOTAL 0.04375851 113.45252899 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 20616 +BPFP 0.0663 bits/point +EBPFP 0.1325 equivalent bits/point +MSE 113.452529 +---------------------- -------------------------------------------------------- +Time: 0.617s Load: 0.010s, Pack+Encode: 0.245s, Decode+Unpack: 0.361s +---------------------- -------------------------------------------------------- +💾 Converting with 113.4525 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,260B, BPFP=0.0989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,424B, BPFP=0.1118 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,592B, BPFP=0.1250 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,344B, BPFP=0.1055 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,452B, BPFP=0.1140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,380B, BPFP=0.1084 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,420B, BPFP=0.1115 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,400B, BPFP=0.1099 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,636B, BPFP=0.1285 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,480B, BPFP=0.1162 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,968B, BPFP=0.0445 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.302s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10671347 512.30806376 + layer.0.v_cache 0.00001788 0.05667910 + layer.1.k_cache 0.15257932 35.46988851 + layer.1.v_cache 0.00000642 0.02046008 + layer.2.k_cache 0.02197224 6.44070082 + layer.2.v_cache 0.00002094 0.06079382 + layer.3.k_cache 0.03325213 22.84148280 + layer.3.v_cache 0.00002126 0.07467011 + layer.4.k_cache 0.00067229 1.84483161 + layer.4.v_cache 0.00005023 0.13742081 + layer.4.output 1.53834953 272.73714555 + ------------------------------------------------------------------------------------- + TOTAL 0.65198546 146.37735354 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 18356 +BPFP 0.0848 bits/point +EBPFP 0.1696 equivalent bits/point +MSE 146.377354 +---------------------- -------------------------------------------------------- +Time: 0.530s Load: 0.008s, Pack+Encode: 0.220s, Decode+Unpack: 0.302s +---------------------- -------------------------------------------------------- +💾 Converting with 146.3774 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,264B, BPFP=0.0787 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,368B, BPFP=0.0852 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,504B, BPFP=0.0936 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,252B, BPFP=0.0779 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,408B, BPFP=0.0876 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,308B, BPFP=0.0814 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,356B, BPFP=0.0844 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,308B, BPFP=0.0814 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,540B, BPFP=0.0959 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,436B, BPFP=0.0894 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,348B, BPFP=0.0298 +⌛️ [2/4] FRONTEND: Frontend time: 0.218s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15214830 502.59101096 + layer.0.v_cache 0.00001722 0.05621988 + layer.1.k_cache 0.49754258 34.94984126 + layer.1.v_cache 0.00000639 0.02080695 + layer.2.k_cache 0.01862443 6.28182582 + layer.2.v_cache 0.00002219 0.06286209 + layer.3.k_cache 0.01200796 22.50633598 + layer.3.v_cache 0.00002210 0.07571093 + layer.4.k_cache 0.00075226 1.87080973 + layer.4.v_cache 0.00005409 0.14468148 + layer.4.output 1.21979868 216.20665552 + ------------------------------------------------------------------------------------- + TOTAL 0.54234049 122.47098198 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 17092 +BPFP 0.0626 bits/point +EBPFP 0.1252 equivalent bits/point +MSE 122.470982 +---------------------- -------------------------------------------------------- +Time: 0.534s Load: 0.008s, Pack+Encode: 0.218s, Decode+Unpack: 0.307s +---------------------- -------------------------------------------------------- +💾 Converting with 122.4710 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,612B, BPFP=0.0826 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,804B, BPFP=0.0924 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,936B, BPFP=0.0992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,612B, BPFP=0.0826 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,848B, BPFP=0.0947 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,660B, BPFP=0.0850 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,752B, BPFP=0.0898 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,668B, BPFP=0.0855 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,992B, BPFP=0.1020 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,824B, BPFP=0.0934 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,908B, BPFP=0.0286 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18432382 506.26439549 + layer.0.v_cache 0.00001629 0.05352597 + layer.1.k_cache 0.53789903 35.07770876 + layer.1.v_cache 0.00000607 0.01964139 + layer.2.k_cache 0.01514517 6.32070072 + layer.2.v_cache 0.00002151 0.05985188 + layer.3.k_cache 0.04182221 22.67104572 + layer.3.v_cache 0.00002017 0.07097094 + layer.4.k_cache 0.00068627 1.82350554 + layer.4.v_cache 0.00005063 0.13916009 + layer.4.output 0.04375917 178.31699356 + ------------------------------------------------------------------------------------- + TOTAL 0.06390032 107.10114479 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 21616 +BPFP 0.0651 bits/point +EBPFP 0.1303 equivalent bits/point +MSE 107.101145 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.367s +---------------------- -------------------------------------------------------- +💾 Converting with 107.1011 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 371, 128) +Output shape: (1, 371, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.output: torch.Size([1, 371, 3584]) -> torch.Size([1, 1, 371, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,888B, BPFP=0.0795 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,080B, BPFP=0.0876 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,276B, BPFP=0.0959 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,944B, BPFP=0.0819 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,184B, BPFP=0.0920 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,984B, BPFP=0.0836 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,068B, BPFP=0.0871 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,996B, BPFP=0.0841 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,372B, BPFP=0.0999 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,160B, BPFP=0.0910 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,628B, BPFP=0.0278 +⌛️ [2/4] FRONTEND: Frontend time: 0.276s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.466s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16952868 506.88624495 + layer.0.v_cache 0.00001698 0.05560449 + layer.1.k_cache 0.82091874 34.98268510 + layer.1.v_cache 0.00000666 0.02043221 + layer.2.k_cache 0.02418445 6.33659424 + layer.2.v_cache 0.00002238 0.06273565 + layer.3.k_cache 0.03055777 22.57710632 + layer.3.v_cache 0.00002117 0.07240396 + layer.4.k_cache 0.00071103 1.85111535 + layer.4.v_cache 0.00005915 0.14290972 + layer.4.output 0.03610923 146.88893435 + ------------------------------------------------------------------------------------- + TOTAL 0.07639951 94.18884544 + (elements=3,229,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3229184 +Total Bytes 25580 +BPFP 0.0634 bits/point +EBPFP 0.1267 equivalent bits/point +MSE 94.188845 +---------------------- -------------------------------------------------------- +Time: 0.755s Load: 0.013s, Pack+Encode: 0.276s, Decode+Unpack: 0.466s +---------------------- -------------------------------------------------------- +💾 Converting with 94.1888 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,208B, BPFP=0.0866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,388B, BPFP=0.0995 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,504B, BPFP=0.1078 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,284B, BPFP=0.0920 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,412B, BPFP=0.1012 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,344B, BPFP=0.0963 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,384B, BPFP=0.0992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,352B, BPFP=0.0969 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,572B, BPFP=0.1127 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,464B, BPFP=0.1049 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,528B, BPFP=0.0361 +⌛️ [2/4] FRONTEND: Frontend time: 0.239s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13674593 508.10761898 + layer.0.v_cache 0.00001768 0.05671398 + layer.1.k_cache 0.26206151 35.28707354 + layer.1.v_cache 0.00000624 0.02051282 + layer.2.k_cache 0.02938290 6.41591931 + layer.2.v_cache 0.00002158 0.06212491 + layer.3.k_cache 0.05653211 22.61911823 + layer.3.v_cache 0.00002129 0.07443782 + layer.4.k_cache 0.00070776 1.84584815 + layer.4.v_cache 0.00005193 0.14206270 + layer.4.output 1.40434551 249.59127212 + ------------------------------------------------------------------------------------- + TOTAL 0.60682162 136.57472560 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 17440 +BPFP 0.0735 bits/point +EBPFP 0.1471 equivalent bits/point +MSE 136.574726 +---------------------- -------------------------------------------------------- +Time: 0.564s Load: 0.009s, Pack+Encode: 0.239s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 136.5747 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 193, 128) +Output shape: (1, 193, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.output: torch.Size([1, 193, 3584]) -> torch.Size([1, 1, 193, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,240B, BPFP=0.1004 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,356B, BPFP=0.1098 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,528B, BPFP=0.1237 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,216B, BPFP=0.0984 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,416B, BPFP=0.1146 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,252B, BPFP=0.1014 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,344B, BPFP=0.1088 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,232B, BPFP=0.0997 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,556B, BPFP=0.1260 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,448B, BPFP=0.1172 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,456B, BPFP=0.0400 +⌛️ [2/4] FRONTEND: Frontend time: 0.219s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.305s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14305212 508.35868685 + layer.0.v_cache 0.00001770 0.05579421 + layer.1.k_cache 0.14179654 35.43274369 + layer.1.v_cache 0.00000608 0.02034300 + layer.2.k_cache 0.01555968 6.43743549 + layer.2.v_cache 0.00002047 0.06306015 + layer.3.k_cache 0.02655400 22.77274429 + layer.3.v_cache 0.00002031 0.07354531 + layer.4.k_cache 0.00069483 1.90535345 + layer.4.v_cache 0.00005004 0.14248634 + layer.4.output 1.58615237 280.96794041 + ------------------------------------------------------------------------------------- + TOTAL 0.67240225 149.53163386 + (elements=1,679,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1679872 +Total Bytes 17044 +BPFP 0.0812 bits/point +EBPFP 0.1623 equivalent bits/point +MSE 149.531634 +---------------------- -------------------------------------------------------- +Time: 0.530s Load: 0.006s, Pack+Encode: 0.219s, Decode+Unpack: 0.305s +---------------------- -------------------------------------------------------- +💾 Converting with 149.5316 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,308B, BPFP=0.0821 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,452B, BPFP=0.0911 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,552B, BPFP=0.0974 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,328B, BPFP=0.0833 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,476B, BPFP=0.0926 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,364B, BPFP=0.0856 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,420B, BPFP=0.0891 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,372B, BPFP=0.0861 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,608B, BPFP=0.1009 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,500B, BPFP=0.0941 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,384B, BPFP=0.0303 +⌛️ [2/4] FRONTEND: Frontend time: 0.219s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10821181 508.83816516 + layer.0.v_cache 0.00001942 0.05575685 + layer.1.k_cache 0.30751975 35.31596699 + layer.1.v_cache 0.00000599 0.01959777 + layer.2.k_cache 0.01421910 6.37969750 + layer.2.v_cache 0.00002020 0.06145678 + layer.3.k_cache 0.01730646 22.97493097 + layer.3.v_cache 0.00002068 0.07346608 + layer.4.k_cache 0.00075739 1.91476845 + layer.4.v_cache 0.00005209 0.14445934 + layer.4.output 1.22958295 223.83955465 + ------------------------------------------------------------------------------------- + TOTAL 0.53265962 126.03853814 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 17764 +BPFP 0.0656 bits/point +EBPFP 0.1311 equivalent bits/point +MSE 126.038538 +---------------------- -------------------------------------------------------- +Time: 0.603s Load: 0.008s, Pack+Encode: 0.219s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 126.0385 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 202, 128) +Output shape: (1, 202, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.output: torch.Size([1, 202, 3584]) -> torch.Size([1, 1, 202, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,264B, BPFP=0.0978 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,404B, BPFP=0.1086 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,608B, BPFP=0.1244 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,304B, BPFP=0.1009 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,472B, BPFP=0.1139 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,320B, BPFP=0.1021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,436B, BPFP=0.1111 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,368B, BPFP=0.1058 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,676B, BPFP=0.1296 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,492B, BPFP=0.1154 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,716B, BPFP=0.0411 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15282421 513.71851795 + layer.0.v_cache 0.00001801 0.05745844 + layer.1.k_cache 0.08597872 35.68543568 + layer.1.v_cache 0.00000649 0.02078060 + layer.2.k_cache 0.01425038 6.45119303 + layer.2.v_cache 0.00002047 0.06212947 + layer.3.k_cache 0.05704349 22.84159383 + layer.3.v_cache 0.00002256 0.07662659 + layer.4.k_cache 0.00069767 1.91669638 + layer.4.v_cache 0.00005056 0.14017130 + layer.4.output 1.51553867 268.30248851 + ------------------------------------------------------------------------------------- + TOTAL 0.64233431 144.65223664 + (elements=1,758,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1758208 +Total Bytes 18060 +BPFP 0.0822 bits/point +EBPFP 0.1643 equivalent bits/point +MSE 144.652237 +---------------------- -------------------------------------------------------- +Time: 0.541s Load: 0.007s, Pack+Encode: 0.224s, Decode+Unpack: 0.309s +---------------------- -------------------------------------------------------- +💾 Converting with 144.6522 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,500B, BPFP=0.0828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,680B, BPFP=0.0928 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,832B, BPFP=0.1011 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,548B, BPFP=0.0855 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,736B, BPFP=0.0958 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,528B, BPFP=0.0844 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,656B, BPFP=0.0914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,540B, BPFP=0.0850 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,916B, BPFP=0.1058 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,768B, BPFP=0.0976 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,060B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14837397 502.60015459 + layer.0.v_cache 0.00001603 0.05089205 + layer.1.k_cache 0.60911565 34.92994631 + layer.1.v_cache 0.00000597 0.01823640 + layer.2.k_cache 0.02477714 6.21971233 + layer.2.v_cache 0.00001938 0.05494511 + layer.3.k_cache 0.01052417 22.55821934 + layer.3.v_cache 0.00002039 0.06766729 + layer.4.k_cache 0.00073933 1.81227764 + layer.4.v_cache 0.00005963 0.13761270 + layer.4.output 0.00470031 196.73215863 + ------------------------------------------------------------------------------------- + TOTAL 0.04862081 114.44557495 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 20764 +BPFP 0.0674 bits/point +EBPFP 0.1349 equivalent bits/point +MSE 114.445575 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.011s, Pack+Encode: 0.258s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 114.4456 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 210, 128) +Output shape: (1, 210, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.output: torch.Size([1, 210, 3584]) -> torch.Size([1, 1, 210, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,236B, BPFP=0.0920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,448B, BPFP=0.1077 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,580B, BPFP=0.1176 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,336B, BPFP=0.0994 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,440B, BPFP=0.1071 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,336B, BPFP=0.0994 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,428B, BPFP=0.1062 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,336B, BPFP=0.0994 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,652B, BPFP=0.1229 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,512B, BPFP=0.1125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,568B, BPFP=0.0379 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11415643 512.33653274 + layer.0.v_cache 0.00001882 0.05457989 + layer.1.k_cache 0.27145578 35.27583240 + layer.1.v_cache 0.00000623 0.01973480 + layer.2.k_cache 0.01201470 6.26530994 + layer.2.v_cache 0.00002208 0.05979470 + layer.3.k_cache 0.00887682 22.78847656 + layer.3.v_cache 0.00002165 0.07230018 + layer.4.k_cache 0.00072611 1.82829009 + layer.4.v_cache 0.00005233 0.14085250 + layer.4.output 1.45781997 258.20556973 + ------------------------------------------------------------------------------------- + TOTAL 0.62424063 140.36945247 + (elements=1,827,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1827840 +Total Bytes 17872 +BPFP 0.0782 bits/point +EBPFP 0.1564 equivalent bits/point +MSE 140.369452 +---------------------- -------------------------------------------------------- +Time: 0.548s Load: 0.007s, Pack+Encode: 0.226s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 140.3695 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 165, 128) +Output shape: (1, 165, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.output: torch.Size([1, 165, 3584]) -> torch.Size([1, 1, 165, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 992B, BPFP=0.0939 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,072B, BPFP=0.1015 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,188B, BPFP=0.1125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 996B, BPFP=0.0943 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,112B, BPFP=0.1053 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,004B, BPFP=0.0951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,100B, BPFP=0.1042 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,044B, BPFP=0.0989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,256B, BPFP=0.1189 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,092B, BPFP=0.1034 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,800B, BPFP=0.0379 +⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.262s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10245823 507.92154356 + layer.0.v_cache 0.00002070 0.05784057 + layer.1.k_cache 0.13127365 35.89106889 + layer.1.v_cache 0.00000615 0.02026673 + layer.2.k_cache 0.01458293 6.55516912 + layer.2.v_cache 0.00002074 0.06039872 + layer.3.k_cache 0.02375244 22.84293472 + layer.3.v_cache 0.00002214 0.07498338 + layer.4.k_cache 0.00067489 1.91885635 + layer.4.v_cache 0.00005196 0.14467859 + layer.4.output 0.00973386 333.31228355 + ------------------------------------------------------------------------------------- + TOTAL 0.02005887 171.09845444 + (elements=1,436,160) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1436160 +Total Bytes 13656 +BPFP 0.0761 bits/point +EBPFP 0.1521 equivalent bits/point +MSE 171.098454 +---------------------- -------------------------------------------------------- +Time: 0.466s Load: 0.005s, Pack+Encode: 0.199s, Decode+Unpack: 0.262s +---------------------- -------------------------------------------------------- +💾 Converting with 171.0985 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,052B, BPFP=0.0898 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,128B, BPFP=0.0963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,232B, BPFP=0.1052 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,060B, BPFP=0.0905 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,180B, BPFP=0.1008 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,068B, BPFP=0.0912 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,160B, BPFP=0.0990 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,084B, BPFP=0.0926 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,260B, BPFP=0.1076 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,148B, BPFP=0.0980 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,620B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.265s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12273688 518.43895150 + layer.0.v_cache 0.00001959 0.05756070 + layer.1.k_cache 0.11126613 36.21158321 + layer.1.v_cache 0.00000674 0.02143231 + layer.2.k_cache 0.01416154 6.64290098 + layer.2.v_cache 0.00002164 0.06310388 + layer.3.k_cache 0.04491856 23.24132300 + layer.3.v_cache 0.00002114 0.07603717 + layer.4.k_cache 0.00068813 1.93332910 + layer.4.v_cache 0.00006151 0.14569096 + layer.4.output 0.00884247 310.25892857 + ------------------------------------------------------------------------------------- + TOTAL 0.02092936 162.27320075 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 13992 +BPFP 0.0703 bits/point +EBPFP 0.1405 equivalent bits/point +MSE 162.273201 +---------------------- -------------------------------------------------------- +Time: 0.495s Load: 0.006s, Pack+Encode: 0.224s, Decode+Unpack: 0.265s +---------------------- -------------------------------------------------------- +💾 Converting with 162.2732 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 257, 128) +Output shape: (1, 257, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.output: torch.Size([1, 257, 3584]) -> torch.Size([1, 1, 257, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,520B, BPFP=0.0924 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,732B, BPFP=0.1053 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,836B, BPFP=0.1116 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,548B, BPFP=0.0941 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,704B, BPFP=0.1036 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,592B, BPFP=0.0968 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,660B, BPFP=0.1009 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,568B, BPFP=0.0953 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,944B, BPFP=0.1182 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,804B, BPFP=0.1097 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,096B, BPFP=0.0356 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12694686 506.60761187 + layer.0.v_cache 0.00002055 0.05680535 + layer.1.k_cache 0.54844232 35.11586515 + layer.1.v_cache 0.00000676 0.01968358 + layer.2.k_cache 0.01835482 6.28285496 + layer.2.v_cache 0.00002099 0.06005940 + layer.3.k_cache 0.02141366 22.56206112 + layer.3.v_cache 0.00002047 0.07276683 + layer.4.k_cache 0.00070767 1.85019075 + layer.4.v_cache 0.00005046 0.13629818 + layer.4.output 0.00513101 216.24958310 + ------------------------------------------------------------------------------------- + TOTAL 0.04422951 122.73595758 + (elements=2,236,928) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2236928 +Total Bytes 21004 +BPFP 0.0751 bits/point +EBPFP 0.1502 equivalent bits/point +MSE 122.735958 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 122.7360 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,056B, BPFP=0.0907 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,152B, BPFP=0.0989 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,248B, BPFP=0.1071 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,092B, BPFP=0.0938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,188B, BPFP=0.1020 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,092B, BPFP=0.0938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,160B, BPFP=0.0996 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,092B, BPFP=0.0938 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,272B, BPFP=0.1092 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,180B, BPFP=0.1013 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,684B, BPFP=0.0329 +⌛️ [2/4] FRONTEND: Frontend time: 0.190s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.253s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12731336 521.98806662 + layer.0.v_cache 0.00001750 0.05763045 + layer.1.k_cache 0.10052182 36.02265142 + layer.1.v_cache 0.00000572 0.01928830 + layer.2.k_cache 0.01141533 6.52821434 + layer.2.v_cache 0.00001926 0.05845971 + layer.3.k_cache 0.03745206 22.94320377 + layer.3.v_cache 0.00002083 0.07325848 + layer.4.k_cache 0.00066429 1.93328958 + layer.4.v_cache 0.00005101 0.14226431 + layer.4.output 0.00884527 305.81684164 + ------------------------------------------------------------------------------------- + TOTAL 0.01996459 160.61671873 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 14216 +BPFP 0.0718 bits/point +EBPFP 0.1436 equivalent bits/point +MSE 160.616719 +---------------------- -------------------------------------------------------- +Time: 0.450s Load: 0.007s, Pack+Encode: 0.190s, Decode+Unpack: 0.253s +---------------------- -------------------------------------------------------- +💾 Converting with 160.6167 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,188B, BPFP=0.0836 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,356B, BPFP=0.0954 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,484B, BPFP=0.1044 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,252B, BPFP=0.0881 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,404B, BPFP=0.0988 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,284B, BPFP=0.0904 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,344B, BPFP=0.0946 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,292B, BPFP=0.0909 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,592B, BPFP=0.1120 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,440B, BPFP=0.1014 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,328B, BPFP=0.0335 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.322s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12752426 511.37682995 + layer.0.v_cache 0.00002004 0.05802767 + layer.1.k_cache 0.30660921 35.13642402 + layer.1.v_cache 0.00000613 0.01942601 + layer.2.k_cache 0.01808885 6.31277328 + layer.2.v_cache 0.00002107 0.05961226 + layer.3.k_cache 0.01678786 22.58601008 + layer.3.v_cache 0.00002050 0.07193844 + layer.4.k_cache 0.00068290 1.87775772 + layer.4.v_cache 0.00004851 0.13794108 + layer.4.output 1.37905047 244.04094273 + ------------------------------------------------------------------------------------- + TOTAL 0.59548016 134.46607880 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 16964 +BPFP 0.0702 bits/point +EBPFP 0.1405 equivalent bits/point +MSE 134.466079 +---------------------- -------------------------------------------------------- +Time: 0.559s Load: 0.008s, Pack+Encode: 0.229s, Decode+Unpack: 0.322s +---------------------- -------------------------------------------------------- +💾 Converting with 134.4661 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,284B, BPFP=0.0925 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,392B, BPFP=0.1002 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,512B, BPFP=0.1089 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,272B, BPFP=0.0916 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,424B, BPFP=0.1025 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,288B, BPFP=0.0927 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,376B, BPFP=0.0991 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,300B, BPFP=0.0936 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,580B, BPFP=0.1138 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,448B, BPFP=0.1043 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,572B, BPFP=0.0367 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.329s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12766143 503.71313364 + layer.0.v_cache 0.00001653 0.05606326 + layer.1.k_cache 0.34631467 35.13366305 + layer.1.v_cache 0.00000659 0.02037086 + layer.2.k_cache 0.01786991 6.29899126 + layer.2.v_cache 0.00002041 0.06053487 + layer.3.k_cache 0.06874437 22.98536956 + layer.3.v_cache 0.00002276 0.07924818 + layer.4.k_cache 0.00067882 1.84903067 + layer.4.v_cache 0.00005621 0.14825362 + layer.4.output 1.41085444 250.64688940 + ------------------------------------------------------------------------------------- + TOTAL 0.61396310 136.75722852 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 17448 +BPFP 0.0739 bits/point +EBPFP 0.1478 equivalent bits/point +MSE 136.757229 +---------------------- -------------------------------------------------------- +Time: 0.563s Load: 0.007s, Pack+Encode: 0.227s, Decode+Unpack: 0.329s +---------------------- -------------------------------------------------------- +💾 Converting with 136.7572 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,188B, BPFP=0.0821 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,336B, BPFP=0.0924 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,492B, BPFP=0.1032 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,268B, BPFP=0.0877 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,432B, BPFP=0.0990 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,292B, BPFP=0.0893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,348B, BPFP=0.0932 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,304B, BPFP=0.0902 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,584B, BPFP=0.1095 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,416B, BPFP=0.0979 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,300B, BPFP=0.0326 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12070838 508.21149060 + layer.0.v_cache 0.00001763 0.05698397 + layer.1.k_cache 0.31015487 35.38625423 + layer.1.v_cache 0.00000677 0.02120839 + layer.2.k_cache 0.03544329 6.33993449 + layer.2.v_cache 0.00002212 0.06323656 + layer.3.k_cache 0.05682086 22.78161729 + layer.3.v_cache 0.00002136 0.07717724 + layer.4.k_cache 0.00073636 1.90387706 + layer.4.v_cache 0.00005031 0.14536260 + layer.4.output 1.35470685 240.50890882 + ------------------------------------------------------------------------------------- + TOTAL 0.58864294 132.85585319 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 16960 +BPFP 0.0690 bits/point +EBPFP 0.1379 equivalent bits/point +MSE 132.855853 +---------------------- -------------------------------------------------------- +Time: 0.557s Load: 0.010s, Pack+Encode: 0.229s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 132.8559 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,244B, BPFP=0.0913 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,400B, BPFP=0.1027 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,560B, BPFP=0.1144 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,328B, BPFP=0.0974 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,460B, BPFP=0.1071 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,360B, BPFP=0.0998 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,420B, BPFP=0.1042 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,368B, BPFP=0.1004 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,628B, BPFP=0.1194 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,480B, BPFP=0.1086 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,568B, BPFP=0.0374 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13227103 505.98624560 + layer.0.v_cache 0.00002050 0.05630330 + layer.1.k_cache 0.22864332 35.32558044 + layer.1.v_cache 0.00000662 0.02086406 + layer.2.k_cache 0.02950178 6.32017725 + layer.2.v_cache 0.00002095 0.06172176 + layer.3.k_cache 0.01835468 22.70222409 + layer.3.v_cache 0.00002123 0.07703264 + layer.4.k_cache 0.00070999 1.91172970 + layer.4.v_cache 0.00005457 0.14464878 + layer.4.output 1.43734575 255.31675469 + ------------------------------------------------------------------------------------- + TOTAL 0.61594264 138.81316532 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 17816 +BPFP 0.0769 bits/point +EBPFP 0.1538 equivalent bits/point +MSE 138.813165 +---------------------- -------------------------------------------------------- +Time: 0.552s Load: 0.008s, Pack+Encode: 0.226s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 138.8132 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,192B, BPFP=0.0839 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,360B, BPFP=0.0957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,476B, BPFP=0.1039 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,212B, BPFP=0.0853 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,388B, BPFP=0.0977 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,252B, BPFP=0.0881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,352B, BPFP=0.0952 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,264B, BPFP=0.0890 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,576B, BPFP=0.1109 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,384B, BPFP=0.0974 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,232B, BPFP=0.0325 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12518616 506.43676098 + layer.0.v_cache 0.00001676 0.05450887 + layer.1.k_cache 0.33007166 35.16244369 + layer.1.v_cache 0.00000630 0.02008541 + layer.2.k_cache 0.01131002 6.33195372 + layer.2.v_cache 0.00002125 0.06157348 + layer.3.k_cache 0.02232737 22.72014006 + layer.3.v_cache 0.00001983 0.06995941 + layer.4.k_cache 0.00071526 1.88702777 + layer.4.v_cache 0.00005006 0.13522689 + layer.4.output 1.37904434 244.02696670 + ------------------------------------------------------------------------------------- + TOTAL 0.59664912 134.18049689 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 16688 +BPFP 0.0691 bits/point +EBPFP 0.1382 equivalent bits/point +MSE 134.180497 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.007s, Pack+Encode: 0.225s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 134.1805 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,364B, BPFP=0.0873 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,444B, BPFP=0.0925 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,592B, BPFP=0.1019 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,364B, BPFP=0.0873 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,524B, BPFP=0.0976 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,384B, BPFP=0.0886 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,472B, BPFP=0.0943 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,384B, BPFP=0.0886 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,648B, BPFP=0.1055 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,504B, BPFP=0.0963 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,268B, BPFP=0.0299 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14655781 505.77110015 + layer.0.v_cache 0.00001709 0.05309321 + layer.1.k_cache 0.36683986 35.33018218 + layer.1.v_cache 0.00000651 0.01996426 + layer.2.k_cache 0.01755252 6.26434476 + layer.2.v_cache 0.00002060 0.05914203 + layer.3.k_cache 0.02713263 22.78990699 + layer.3.v_cache 0.00001997 0.06798102 + layer.4.k_cache 0.00071102 1.80525633 + layer.4.v_cache 0.00005162 0.13442691 + layer.4.output 1.25479176 223.77190061 + ------------------------------------------------------------------------------------- + TOTAL 0.54955600 125.80580601 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 17948 +BPFP 0.0676 bits/point +EBPFP 0.1352 equivalent bits/point +MSE 125.805806 +---------------------- -------------------------------------------------------- +Time: 0.551s Load: 0.009s, Pack+Encode: 0.225s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 125.8058 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 155, 128) +Output shape: (1, 155, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.output: torch.Size([1, 155, 3584]) -> torch.Size([1, 1, 155, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 928B, BPFP=0.0935 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,084B, BPFP=0.1093 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,156B, BPFP=0.1165 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 932B, BPFP=0.0940 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,076B, BPFP=0.1085 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 968B, BPFP=0.0976 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,080B, BPFP=0.1089 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 964B, BPFP=0.0972 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,204B, BPFP=0.1214 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,148B, BPFP=0.1157 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,736B, BPFP=0.0394 +⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.259s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102877 511.77842742 + layer.0.v_cache 0.00001822 0.05373913 + layer.1.k_cache 0.12773433 35.69685610 + layer.1.v_cache 0.00000585 0.01907927 + layer.2.k_cache 0.00931232 6.52356666 + layer.2.v_cache 0.00002110 0.05963590 + layer.3.k_cache 0.01537714 22.74010994 + layer.3.v_cache 0.00002091 0.07203519 + layer.4.k_cache 0.00066471 1.85550045 + layer.4.v_cache 0.00005180 0.14166137 + layer.4.output 0.01742802 358.57108295 + ------------------------------------------------------------------------------------- + TOTAL 0.02389596 181.70224659 + (elements=1,349,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1349120 +Total Bytes 13276 +BPFP 0.0787 bits/point +EBPFP 0.1574 equivalent bits/point +MSE 181.702247 +---------------------- -------------------------------------------------------- +Time: 0.464s Load: 0.006s, Pack+Encode: 0.200s, Decode+Unpack: 0.259s +---------------------- -------------------------------------------------------- +💾 Converting with 181.7022 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,280B, BPFP=0.0930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,400B, BPFP=0.1017 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,548B, BPFP=0.1125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,296B, BPFP=0.0942 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,460B, BPFP=0.1061 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,316B, BPFP=0.0956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,396B, BPFP=0.1015 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,352B, BPFP=0.0983 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,608B, BPFP=0.1169 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,528B, BPFP=0.1110 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,688B, BPFP=0.0383 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11659910 510.95559593 + layer.0.v_cache 0.00001673 0.05694895 + layer.1.k_cache 0.29571861 35.33709802 + layer.1.v_cache 0.00000615 0.02015675 + layer.2.k_cache 0.04707494 6.38643884 + layer.2.v_cache 0.00002002 0.05857198 + layer.3.k_cache 0.03274950 22.63400254 + layer.3.v_cache 0.00002078 0.07145992 + layer.4.k_cache 0.00068608 1.83865584 + layer.4.v_cache 0.00005136 0.14169789 + layer.4.output 1.42391751 253.86978821 + ------------------------------------------------------------------------------------- + TOTAL 0.61531564 138.50524377 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 17872 +BPFP 0.0764 bits/point +EBPFP 0.1528 equivalent bits/point +MSE 138.505244 +---------------------- -------------------------------------------------------- +Time: 0.551s Load: 0.009s, Pack+Encode: 0.226s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 138.5052 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.0837 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,460B, BPFP=0.0920 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,580B, BPFP=0.0995 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,320B, BPFP=0.0832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,512B, BPFP=0.0953 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,356B, BPFP=0.0854 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,440B, BPFP=0.0907 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,372B, BPFP=0.0864 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,636B, BPFP=0.1031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,488B, BPFP=0.0938 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,372B, BPFP=0.0303 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10631144 517.41223538 + layer.0.v_cache 0.00001780 0.05784382 + layer.1.k_cache 0.46192560 35.63769531 + layer.1.v_cache 0.00000619 0.02081141 + layer.2.k_cache 0.02890189 6.54926325 + layer.2.v_cache 0.00002115 0.06192840 + layer.3.k_cache 0.04994748 23.08706370 + layer.3.v_cache 0.00002140 0.07627194 + layer.4.k_cache 0.00068365 1.87959302 + layer.4.v_cache 0.00005040 0.14543143 + layer.4.output 1.23456514 229.57207661 + ------------------------------------------------------------------------------------- + TOTAL 0.54646135 128.93721611 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 17864 +BPFP 0.0662 bits/point +EBPFP 0.1324 equivalent bits/point +MSE 128.937216 +---------------------- -------------------------------------------------------- +Time: 0.556s Load: 0.011s, Pack+Encode: 0.226s, Decode+Unpack: 0.320s +---------------------- -------------------------------------------------------- +💾 Converting with 128.9372 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 194, 128) +Output shape: (1, 194, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.output: torch.Size([1, 194, 3584]) -> torch.Size([1, 1, 194, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,212B, BPFP=0.0976 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,456B, BPFP=0.1173 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,552B, BPFP=0.1250 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,208B, BPFP=0.0973 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,416B, BPFP=0.1140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,252B, BPFP=0.1008 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,368B, BPFP=0.1102 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,272B, BPFP=0.1024 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,568B, BPFP=0.1263 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,584B, BPFP=0.1276 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,444B, BPFP=0.0396 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15027005 509.10671714 + layer.0.v_cache 0.00001782 0.05573509 + layer.1.k_cache 0.21402168 35.69929426 + layer.1.v_cache 0.00000588 0.01962307 + layer.2.k_cache 0.02177384 6.51748877 + layer.2.v_cache 0.00002019 0.06081011 + layer.3.k_cache 0.04024272 22.58575125 + layer.3.v_cache 0.00002076 0.07112831 + layer.4.k_cache 0.00069591 1.87513874 + layer.4.v_cache 0.00005073 0.13945178 + layer.4.output 1.57794378 282.27478369 + ------------------------------------------------------------------------------------- + TOTAL 0.67486624 150.12086026 + (elements=1,688,576) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1688576 +Total Bytes 17332 +BPFP 0.0821 bits/point +EBPFP 0.1642 equivalent bits/point +MSE 150.120860 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.009s, Pack+Encode: 0.225s, Decode+Unpack: 0.314s +---------------------- -------------------------------------------------------- +💾 Converting with 150.1209 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 258, 128) +Output shape: (1, 258, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.output: torch.Size([1, 258, 3584]) -> torch.Size([1, 1, 258, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,504B, BPFP=0.0911 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,740B, BPFP=0.1054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,844B, BPFP=0.1117 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,488B, BPFP=0.0901 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,716B, BPFP=0.1039 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,500B, BPFP=0.0908 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,668B, BPFP=0.1010 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,512B, BPFP=0.0916 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,952B, BPFP=0.1182 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,772B, BPFP=0.1073 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,928B, BPFP=0.0340 +⌛️ [2/4] FRONTEND: Frontend time: 0.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09035230 509.56516473 + layer.0.v_cache 0.00001726 0.05459797 + layer.1.k_cache 0.42455262 35.22786837 + layer.1.v_cache 0.00000600 0.01868219 + layer.2.k_cache 0.01891134 6.39757780 + layer.2.v_cache 0.00001996 0.05772763 + layer.3.k_cache 0.02093011 22.83793794 + layer.3.v_cache 0.00001928 0.06752541 + layer.4.k_cache 0.00072082 1.84861365 + layer.4.v_cache 0.00004951 0.13867526 + layer.4.output 0.00505962 217.52062569 + ------------------------------------------------------------------------------------- + TOTAL 0.03476450 123.46227946 + (elements=2,245,632) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2245632 +Total Bytes 20624 +BPFP 0.0735 bits/point +EBPFP 0.1469 equivalent bits/point +MSE 123.462279 +---------------------- -------------------------------------------------------- +Time: 0.759s Load: 0.011s, Pack+Encode: 0.309s, Decode+Unpack: 0.439s +---------------------- -------------------------------------------------------- +💾 Converting with 123.4623 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,284B, BPFP=0.0872 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,440B, BPFP=0.0978 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,556B, BPFP=0.1057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,336B, BPFP=0.0908 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,472B, BPFP=0.1000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,400B, BPFP=0.0951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,416B, BPFP=0.0962 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,412B, BPFP=0.0959 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,652B, BPFP=0.1122 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,476B, BPFP=0.1003 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,708B, BPFP=0.0360 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15008528 508.27051630 + layer.0.v_cache 0.00001847 0.05601445 + layer.1.k_cache 0.25612781 35.59547385 + layer.1.v_cache 0.00000678 0.02063407 + layer.2.k_cache 0.03931687 6.35323646 + layer.2.v_cache 0.00002107 0.06083935 + layer.3.k_cache 0.03786138 22.83524117 + layer.3.v_cache 0.00002177 0.07541213 + layer.4.k_cache 0.00069674 1.89024658 + layer.4.v_cache 0.00005115 0.14025965 + layer.4.output 1.33117570 236.04508929 + ------------------------------------------------------------------------------------- + TOTAL 0.57661395 131.03608818 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 18152 +BPFP 0.0725 bits/point +EBPFP 0.1451 equivalent bits/point +MSE 131.036088 +---------------------- -------------------------------------------------------- +Time: 0.586s Load: 0.007s, Pack+Encode: 0.221s, Decode+Unpack: 0.357s +---------------------- -------------------------------------------------------- +💾 Converting with 131.0361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,064B, BPFP=0.0924 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,160B, BPFP=0.1007 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,256B, BPFP=0.1090 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,084B, BPFP=0.0941 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,212B, BPFP=0.1052 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,104B, BPFP=0.0958 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,184B, BPFP=0.1028 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,124B, BPFP=0.0976 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,288B, BPFP=0.1118 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,212B, BPFP=0.1052 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,932B, BPFP=0.0364 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15161699 512.80160590 + layer.0.v_cache 0.00001786 0.05659012 + layer.1.k_cache 0.08662784 35.85940484 + layer.1.v_cache 0.00000652 0.02085175 + layer.2.k_cache 0.02063684 6.42398817 + layer.2.v_cache 0.00002327 0.06340670 + layer.3.k_cache 0.08087930 23.03180339 + layer.3.v_cache 0.00002116 0.07545904 + layer.4.k_cache 0.00074161 1.95607554 + layer.4.v_cache 0.00006096 0.14767982 + layer.4.output 0.00898438 307.41679067 + ------------------------------------------------------------------------------------- + TOTAL 0.02373665 160.72672941 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 14620 +BPFP 0.0747 bits/point +EBPFP 0.1493 equivalent bits/point +MSE 160.726729 +---------------------- -------------------------------------------------------- +Time: 0.578s Load: 0.008s, Pack+Encode: 0.258s, Decode+Unpack: 0.313s +---------------------- -------------------------------------------------------- +💾 Converting with 160.7267 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,188B, BPFP=0.0844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,328B, BPFP=0.0943 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,480B, BPFP=0.1051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,212B, BPFP=0.0861 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,400B, BPFP=0.0994 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,240B, BPFP=0.0881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,356B, BPFP=0.0963 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,256B, BPFP=0.0892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,584B, BPFP=0.1125 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,420B, BPFP=0.1009 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,260B, BPFP=0.0331 +⌛️ [2/4] FRONTEND: Frontend time: 0.239s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14327769 509.10276989 + layer.0.v_cache 0.00001763 0.05567974 + layer.1.k_cache 0.34925704 35.11779119 + layer.1.v_cache 0.00000642 0.02039412 + layer.2.k_cache 0.02352169 6.34095348 + layer.2.v_cache 0.00002271 0.06073400 + layer.3.k_cache 0.00845094 22.57342196 + layer.3.v_cache 0.00002026 0.07240634 + layer.4.k_cache 0.00069169 1.84817130 + layer.4.v_cache 0.00005086 0.14186387 + layer.4.output 1.39158238 246.27605519 + ------------------------------------------------------------------------------------- + TOTAL 0.60390550 135.25097484 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 16724 +BPFP 0.0699 bits/point +EBPFP 0.1397 equivalent bits/point +MSE 135.250975 +---------------------- -------------------------------------------------------- +Time: 0.616s Load: 0.008s, Pack+Encode: 0.239s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 135.2510 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,532B, BPFP=0.0870 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,740B, BPFP=0.0989 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1084 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,580B, BPFP=0.0898 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,768B, BPFP=0.1005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,652B, BPFP=0.0939 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,708B, BPFP=0.0970 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,640B, BPFP=0.0932 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,004B, BPFP=0.1139 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,780B, BPFP=0.1011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,124B, BPFP=0.0335 +⌛️ [2/4] FRONTEND: Frontend time: 0.294s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.408s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12716211 501.86477273 + layer.0.v_cache 0.00001677 0.05479072 + layer.1.k_cache 0.51313377 34.90468395 + layer.1.v_cache 0.00000586 0.01960949 + layer.2.k_cache 0.02995576 6.31502397 + layer.2.v_cache 0.00002133 0.06118458 + layer.3.k_cache 0.03339291 22.58968750 + layer.3.v_cache 0.00002012 0.07121140 + layer.4.k_cache 0.00068476 1.83233088 + layer.4.v_cache 0.00005875 0.13793839 + layer.4.output 0.00481428 202.14879870 + ------------------------------------------------------------------------------------- + TOTAL 0.04342071 116.64075438 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 21436 +BPFP 0.0716 bits/point +EBPFP 0.1433 equivalent bits/point +MSE 116.640754 +---------------------- -------------------------------------------------------- +Time: 0.712s Load: 0.009s, Pack+Encode: 0.294s, Decode+Unpack: 0.408s +---------------------- -------------------------------------------------------- +💾 Converting with 116.6408 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,168B, BPFP=0.0719 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,304B, BPFP=0.0802 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,424B, BPFP=0.0876 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,164B, BPFP=0.0716 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,328B, BPFP=0.0817 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,220B, BPFP=0.0750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,252B, BPFP=0.0770 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,208B, BPFP=0.0743 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,488B, BPFP=0.0915 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,364B, BPFP=0.0839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,220B, BPFP=0.0283 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13396681 499.57018947 + layer.0.v_cache 0.00001641 0.05424061 + layer.1.k_cache 0.47111181 34.37654174 + layer.1.v_cache 0.00000608 0.01978478 + layer.2.k_cache 0.02933135 6.16915821 + layer.2.v_cache 0.00002255 0.06079581 + layer.3.k_cache 0.03744583 22.26555579 + layer.3.v_cache 0.00002260 0.07114072 + layer.4.k_cache 0.00071777 1.78187705 + layer.4.v_cache 0.00004917 0.13573088 + layer.4.output 1.20538323 213.27479963 + ------------------------------------------------------------------------------------- + TOTAL 0.53590429 121.02521250 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 16140 +BPFP 0.0584 bits/point +EBPFP 0.1168 equivalent bits/point +MSE 121.025213 +---------------------- -------------------------------------------------------- +Time: 0.560s Load: 0.008s, Pack+Encode: 0.227s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 121.0252 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,488B, BPFP=0.0807 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,696B, BPFP=0.0920 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,840B, BPFP=0.0998 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,520B, BPFP=0.0825 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,732B, BPFP=0.0940 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,556B, BPFP=0.0844 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,672B, BPFP=0.0907 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,580B, BPFP=0.0857 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,944B, BPFP=0.1055 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,728B, BPFP=0.0938 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,876B, BPFP=0.0300 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14060297 501.78374566 + layer.0.v_cache 0.00001598 0.05462579 + layer.1.k_cache 0.58610762 35.08656141 + layer.1.v_cache 0.00000669 0.01965129 + layer.2.k_cache 0.02834670 6.31415304 + layer.2.v_cache 0.00002256 0.06146689 + layer.3.k_cache 0.01962393 22.70783149 + layer.3.v_cache 0.00002003 0.07107094 + layer.4.k_cache 0.00067797 1.83701113 + layer.4.v_cache 0.00005035 0.13845290 + layer.4.output 0.00463834 194.15910218 + ------------------------------------------------------------------------------------- + TOTAL 0.04752607 113.36401681 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 20632 +BPFP 0.0658 bits/point +EBPFP 0.1317 equivalent bits/point +MSE 113.364017 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.010s, Pack+Encode: 0.258s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 113.3640 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,604B, BPFP=0.0824 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,812B, BPFP=0.0931 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,928B, BPFP=0.0991 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,640B, BPFP=0.0843 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,840B, BPFP=0.0946 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,696B, BPFP=0.0872 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,756B, BPFP=0.0903 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,720B, BPFP=0.0884 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,036B, BPFP=0.1046 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,816B, BPFP=0.0933 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,132B, BPFP=0.0303 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12404394 505.92660362 + layer.0.v_cache 0.00001645 0.05401199 + layer.1.k_cache 0.57408182 35.05183812 + layer.1.v_cache 0.00000616 0.01940466 + layer.2.k_cache 0.03522489 6.37048219 + layer.2.v_cache 0.00002326 0.06083678 + layer.3.k_cache 0.02908128 22.70494321 + layer.3.v_cache 0.00002078 0.07198493 + layer.4.k_cache 0.00072493 1.86619528 + layer.4.v_cache 0.00004967 0.13729570 + layer.4.output 0.04390477 178.97980792 + ------------------------------------------------------------------------------------- + TOTAL 0.06297686 107.36013247 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 21980 +BPFP 0.0665 bits/point +EBPFP 0.1329 equivalent bits/point +MSE 107.360132 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 107.3601 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 242, 128) +Output shape: (1, 242, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.output: torch.Size([1, 242, 3584]) -> torch.Size([1, 1, 242, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,336B, BPFP=0.0863 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,484B, BPFP=0.0958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,604B, BPFP=0.1036 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,396B, BPFP=0.0901 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,536B, BPFP=0.0992 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,420B, BPFP=0.0917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,468B, BPFP=0.0948 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,420B, BPFP=0.0917 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,656B, BPFP=0.1069 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,516B, BPFP=0.0979 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,556B, BPFP=0.0328 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11466674 508.15657283 + layer.0.v_cache 0.00001794 0.05447015 + layer.1.k_cache 0.39062487 35.17687887 + layer.1.v_cache 0.00000616 0.01934969 + layer.2.k_cache 0.02414889 6.27613919 + layer.2.v_cache 0.00002046 0.05866331 + layer.3.k_cache 0.03495786 22.60937096 + layer.3.v_cache 0.00002053 0.07120132 + layer.4.k_cache 0.00069125 1.81616009 + layer.4.v_cache 0.00005034 0.14054223 + layer.4.output 1.26514320 224.03800177 + ------------------------------------------------------------------------------------- + TOTAL 0.55418868 126.03796241 + (elements=2,106,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2106368 +Total Bytes 18392 +BPFP 0.0699 bits/point +EBPFP 0.1397 equivalent bits/point +MSE 126.037962 +---------------------- -------------------------------------------------------- +Time: 0.555s Load: 0.009s, Pack+Encode: 0.227s, Decode+Unpack: 0.320s +---------------------- -------------------------------------------------------- +💾 Converting with 126.0380 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 189, 128) +Output shape: (1, 189, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.output: torch.Size([1, 189, 3584]) -> torch.Size([1, 1, 189, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 904B, BPFP=0.0747 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 976B, BPFP=0.0807 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,104B, BPFP=0.0913 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 876B, BPFP=0.0724 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,012B, BPFP=0.0837 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 912B, BPFP=0.0754 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 976B, BPFP=0.0807 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 932B, BPFP=0.0771 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,140B, BPFP=0.0942 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,004B, BPFP=0.0830 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,492B, BPFP=0.0294 +⌛️ [2/4] FRONTEND: Frontend time: 0.196s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.263s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14361043 502.30415013 + layer.0.v_cache 0.00001648 0.05505163 + layer.1.k_cache 0.18932698 34.66760706 + layer.1.v_cache 0.00000663 0.02066447 + layer.2.k_cache 0.00768016 6.19845597 + layer.2.v_cache 0.00002453 0.06214336 + layer.3.k_cache 0.01421528 22.39721551 + layer.3.v_cache 0.00002024 0.07294590 + layer.4.k_cache 0.00072804 1.83681654 + layer.4.v_cache 0.00005086 0.13881494 + layer.4.output 0.00858909 290.70337302 + ------------------------------------------------------------------------------------- + TOTAL 0.02445902 153.09867510 + (elements=1,645,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1645056 +Total Bytes 12328 +BPFP 0.0600 bits/point +EBPFP 0.1199 equivalent bits/point +MSE 153.098675 +---------------------- -------------------------------------------------------- +Time: 0.466s Load: 0.007s, Pack+Encode: 0.196s, Decode+Unpack: 0.263s +---------------------- -------------------------------------------------------- +💾 Converting with 153.0987 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,248B, BPFP=0.0915 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,432B, BPFP=0.1050 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,560B, BPFP=0.1144 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,336B, BPFP=0.0980 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,448B, BPFP=0.1062 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,372B, BPFP=0.1006 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,416B, BPFP=0.1039 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,360B, BPFP=0.0998 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,628B, BPFP=0.1194 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,444B, BPFP=0.1059 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,516B, BPFP=0.0368 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16894778 506.07643779 + layer.0.v_cache 0.00001779 0.05556723 + layer.1.k_cache 0.37628758 35.26955188 + layer.1.v_cache 0.00000652 0.01993002 + layer.2.k_cache 0.02181128 6.31525375 + layer.2.v_cache 0.00002030 0.05895470 + layer.3.k_cache 0.01815432 22.80521475 + layer.3.v_cache 0.00002063 0.07226862 + layer.4.k_cache 0.00071780 1.82633062 + layer.4.v_cache 0.00005059 0.13963909 + layer.4.output 1.43727796 255.33532445 + ------------------------------------------------------------------------------------- + TOTAL 0.62629296 138.82273056 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 17760 +BPFP 0.0766 bits/point +EBPFP 0.1533 equivalent bits/point +MSE 138.822731 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.007s, Pack+Encode: 0.225s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 138.8227 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,308B, BPFP=0.0946 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,464B, BPFP=0.1059 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,544B, BPFP=0.1117 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,320B, BPFP=0.0955 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,452B, BPFP=0.1050 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,340B, BPFP=0.0969 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,400B, BPFP=0.1013 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,368B, BPFP=0.0990 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,640B, BPFP=0.1186 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,464B, BPFP=0.1059 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,660B, BPFP=0.0378 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09877178 508.49905961 + layer.0.v_cache 0.00001652 0.05469883 + layer.1.k_cache 0.28824160 35.19776521 + layer.1.v_cache 0.00000619 0.01970522 + layer.2.k_cache 0.02137197 6.40465348 + layer.2.v_cache 0.00001995 0.05787933 + layer.3.k_cache 0.03811082 22.55838126 + layer.3.v_cache 0.00002058 0.06943489 + layer.4.k_cache 0.00070031 1.82257702 + layer.4.v_cache 0.00004695 0.13227827 + layer.4.output 1.41729720 251.95211227 + ------------------------------------------------------------------------------------- + TOTAL 0.60990512 137.55771876 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 17960 +BPFP 0.0764 bits/point +EBPFP 0.1528 equivalent bits/point +MSE 137.557719 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.007s, Pack+Encode: 0.225s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 137.5577 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,192B, BPFP=0.0824 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,388B, BPFP=0.0960 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,504B, BPFP=0.1040 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,264B, BPFP=0.0874 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,392B, BPFP=0.0962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,316B, BPFP=0.0910 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,356B, BPFP=0.0938 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,332B, BPFP=0.0921 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,612B, BPFP=0.1114 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,448B, BPFP=0.1001 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,328B, BPFP=0.0329 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12161526 501.69648783 + layer.0.v_cache 0.00001916 0.05451800 + layer.1.k_cache 0.26028682 35.39405161 + layer.1.v_cache 0.00000628 0.01994421 + layer.2.k_cache 0.01296707 6.40418561 + layer.2.v_cache 0.00002040 0.05919138 + layer.3.k_cache 0.04595671 22.59478490 + layer.3.v_cache 0.00002073 0.07023770 + layer.4.k_cache 0.00067352 1.83463551 + layer.4.v_cache 0.00005047 0.13815566 + layer.4.output 1.35465500 240.41845765 + ------------------------------------------------------------------------------------- + TOTAL 0.58377655 132.42325858 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 17132 +BPFP 0.0697 bits/point +EBPFP 0.1393 equivalent bits/point +MSE 132.423259 +---------------------- -------------------------------------------------------- +Time: 0.560s Load: 0.008s, Pack+Encode: 0.225s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 132.4233 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,192B, BPFP=0.0824 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,348B, BPFP=0.0932 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,508B, BPFP=0.1043 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,248B, BPFP=0.0863 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,400B, BPFP=0.0968 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,288B, BPFP=0.0890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,360B, BPFP=0.0940 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,296B, BPFP=0.0896 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,596B, BPFP=0.1103 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,396B, BPFP=0.0965 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,244B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 0.277s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10229862 505.53131914 + layer.0.v_cache 0.00001727 0.05421735 + layer.1.k_cache 0.34669285 35.32800617 + layer.1.v_cache 0.00000643 0.01964098 + layer.2.k_cache 0.01639456 6.35763793 + layer.2.v_cache 0.00002164 0.06016414 + layer.3.k_cache 0.03343437 22.76505462 + layer.3.v_cache 0.00002043 0.07134337 + layer.4.k_cache 0.00071218 1.82304504 + layer.4.v_cache 0.00005088 0.13752838 + layer.4.output 1.35462105 240.49101217 + ------------------------------------------------------------------------------------- + TOTAL 0.58717627 132.68147308 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 16876 +BPFP 0.0686 bits/point +EBPFP 0.1373 equivalent bits/point +MSE 132.681473 +---------------------- -------------------------------------------------------- +Time: 0.609s Load: 0.007s, Pack+Encode: 0.277s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 132.6815 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,216B, BPFP=0.0837 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,360B, BPFP=0.0936 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,508B, BPFP=0.1038 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,260B, BPFP=0.0867 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,424B, BPFP=0.0980 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,308B, BPFP=0.0900 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,356B, BPFP=0.0933 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,280B, BPFP=0.0881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,576B, BPFP=0.1085 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,424B, BPFP=0.0980 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,184B, BPFP=0.0313 +⌛️ [2/4] FRONTEND: Frontend time: 0.277s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.336s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11800324 505.50130782 + layer.0.v_cache 0.00001819 0.05491545 + layer.1.k_cache 0.35011090 35.45163219 + layer.1.v_cache 0.00000612 0.01946316 + layer.2.k_cache 0.02017111 6.44365084 + layer.2.v_cache 0.00002006 0.05996655 + layer.3.k_cache 0.03026835 22.82000275 + layer.3.v_cache 0.00001975 0.07027815 + layer.4.k_cache 0.00074701 1.85414856 + layer.4.v_cache 0.00005118 0.13561931 + layer.4.output 1.34863711 239.28138767 + ------------------------------------------------------------------------------------- + TOTAL 0.58587504 132.19886461 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 16896 +BPFP 0.0684 bits/point +EBPFP 0.1368 equivalent bits/point +MSE 132.198865 +---------------------- -------------------------------------------------------- +Time: 0.620s Load: 0.007s, Pack+Encode: 0.277s, Decode+Unpack: 0.336s +---------------------- -------------------------------------------------------- +💾 Converting with 132.1989 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,264B, BPFP=0.0862 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,424B, BPFP=0.0972 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,540B, BPFP=0.1051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,292B, BPFP=0.0882 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,448B, BPFP=0.0988 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,324B, BPFP=0.0903 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,400B, BPFP=0.0955 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,348B, BPFP=0.0920 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,596B, BPFP=0.1089 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,472B, BPFP=0.1004 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,352B, BPFP=0.0327 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.337s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10957497 506.39294487 + layer.0.v_cache 0.00001697 0.05419195 + layer.1.k_cache 0.31262030 35.38882326 + layer.1.v_cache 0.00000629 0.02012180 + layer.2.k_cache 0.01311818 6.39037659 + layer.2.v_cache 0.00002149 0.06129895 + layer.3.k_cache 0.06227440 22.73036427 + layer.3.v_cache 0.00002139 0.07672301 + layer.4.k_cache 0.00068925 1.85762457 + layer.4.v_cache 0.00005302 0.14662492 + layer.4.output 1.33688878 236.83468497 + ------------------------------------------------------------------------------------- + TOTAL 0.57980104 131.23305229 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 17460 +BPFP 0.0701 bits/point +EBPFP 0.1402 equivalent bits/point +MSE 131.233052 +---------------------- -------------------------------------------------------- +Time: 0.605s Load: 0.007s, Pack+Encode: 0.261s, Decode+Unpack: 0.337s +---------------------- -------------------------------------------------------- +💾 Converting with 131.2331 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,252B, BPFP=0.0918 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,452B, BPFP=0.1065 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,564B, BPFP=0.1147 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,316B, BPFP=0.0965 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,476B, BPFP=0.1083 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,368B, BPFP=0.1004 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,408B, BPFP=0.1033 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,384B, BPFP=0.1015 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,640B, BPFP=0.1203 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,536B, BPFP=0.1127 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,520B, BPFP=0.0369 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13987277 511.19600939 + layer.0.v_cache 0.00001887 0.05610527 + layer.1.k_cache 0.23498858 35.24992435 + layer.1.v_cache 0.00000603 0.01949671 + layer.2.k_cache 0.02028183 6.30368457 + layer.2.v_cache 0.00001992 0.05850252 + layer.3.k_cache 0.05552547 22.79992848 + layer.3.v_cache 0.00002044 0.07125797 + layer.4.k_cache 0.00067773 1.80275329 + layer.4.v_cache 0.00004777 0.13527890 + layer.4.output 1.43727485 255.11154427 + ------------------------------------------------------------------------------------- + TOTAL 0.61837549 139.02786772 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 17916 +BPFP 0.0773 bits/point +EBPFP 0.1546 equivalent bits/point +MSE 139.027868 +---------------------- -------------------------------------------------------- +Time: 0.555s Load: 0.007s, Pack+Encode: 0.227s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 139.0279 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,252B, BPFP=0.0927 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,452B, BPFP=0.1075 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,580B, BPFP=0.1170 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,332B, BPFP=0.0986 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,444B, BPFP=0.1069 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,392B, BPFP=0.1031 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,420B, BPFP=0.1052 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,396B, BPFP=0.1034 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,648B, BPFP=0.1220 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,524B, BPFP=0.1129 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,540B, BPFP=0.0374 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17004660 509.91335900 + layer.0.v_cache 0.00001740 0.05292145 + layer.1.k_cache 0.16740718 35.27317832 + layer.1.v_cache 0.00000654 0.01981085 + layer.2.k_cache 0.01191964 6.27773252 + layer.2.v_cache 0.00001991 0.05757866 + layer.3.k_cache 0.00660476 22.59997269 + layer.3.v_cache 0.00002076 0.07012359 + layer.4.k_cache 0.00070551 1.81880745 + layer.4.v_cache 0.00005179 0.13694402 + layer.4.output 1.45088344 257.04199814 + ------------------------------------------------------------------------------------- + TOTAL 0.61841083 139.73614209 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 17980 +BPFP 0.0783 bits/point +EBPFP 0.1566 equivalent bits/point +MSE 139.736142 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.007s, Pack+Encode: 0.225s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 139.7361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,284B, BPFP=0.0872 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,404B, BPFP=0.0954 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,552B, BPFP=0.1054 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,312B, BPFP=0.0891 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,452B, BPFP=0.0986 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,364B, BPFP=0.0927 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,416B, BPFP=0.0962 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,384B, BPFP=0.0940 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,612B, BPFP=0.1095 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,464B, BPFP=0.0995 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,516B, BPFP=0.0341 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14712874 508.29130435 + layer.0.v_cache 0.00001826 0.05217493 + layer.1.k_cache 0.36788522 35.48761464 + layer.1.v_cache 0.00000598 0.01914727 + layer.2.k_cache 0.01156470 6.35944877 + layer.2.v_cache 0.00002030 0.05949857 + layer.3.k_cache 0.01315050 22.52349482 + layer.3.v_cache 0.00002017 0.06890047 + layer.4.k_cache 0.00068370 1.79740030 + layer.4.v_cache 0.00004817 0.13101987 + layer.4.output 1.33105469 236.01655668 + ------------------------------------------------------------------------------------- + TOTAL 0.57987697 130.99446475 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 17760 +BPFP 0.0710 bits/point +EBPFP 0.1419 equivalent bits/point +MSE 130.994465 +---------------------- -------------------------------------------------------- +Time: 0.552s Load: 0.008s, Pack+Encode: 0.226s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 130.9945 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 188, 128) +Output shape: (1, 188, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.output: torch.Size([1, 188, 3584]) -> torch.Size([1, 1, 188, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 944B, BPFP=0.0785 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,016B, BPFP=0.0844 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,120B, BPFP=0.0931 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 924B, BPFP=0.0768 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,048B, BPFP=0.0871 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 956B, BPFP=0.0795 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,004B, BPFP=0.0834 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 944B, BPFP=0.0785 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,140B, BPFP=0.0947 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,076B, BPFP=0.0894 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,488B, BPFP=0.0295 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.262s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005002 501.45341589 + layer.0.v_cache 0.00001663 0.05327481 + layer.1.k_cache 0.14149688 34.81095724 + layer.1.v_cache 0.00000584 0.01876245 + layer.2.k_cache 0.00881058 6.16499751 + layer.2.v_cache 0.00002285 0.05897406 + layer.3.k_cache 0.03192479 22.29448554 + layer.3.v_cache 0.00001958 0.06805626 + layer.4.k_cache 0.00066612 1.76464227 + layer.4.v_cache 0.00005162 0.13944354 + layer.4.output 0.00857985 292.16254274 + ------------------------------------------------------------------------------------- + TOTAL 0.02253670 153.64498875 + (elements=1,636,352) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1636352 +Total Bytes 12660 +BPFP 0.0619 bits/point +EBPFP 0.1238 equivalent bits/point +MSE 153.644989 +---------------------- -------------------------------------------------------- +Time: 0.490s Load: 0.006s, Pack+Encode: 0.222s, Decode+Unpack: 0.262s +---------------------- -------------------------------------------------------- +💾 Converting with 153.6450 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,252B, BPFP=0.0983 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,472B, BPFP=0.1156 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,580B, BPFP=0.1241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,300B, BPFP=0.1021 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,452B, BPFP=0.1140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,320B, BPFP=0.1036 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,432B, BPFP=0.1124 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,320B, BPFP=0.1036 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,644B, BPFP=0.1291 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,548B, BPFP=0.1215 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,944B, BPFP=0.0442 +⌛️ [2/4] FRONTEND: Frontend time: 0.219s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11435634 502.60921796 + layer.0.v_cache 0.00001764 0.05645934 + layer.1.k_cache 0.25025516 35.42483659 + layer.1.v_cache 0.00000690 0.01984810 + layer.2.k_cache 0.01528106 6.45874760 + layer.2.v_cache 0.00002105 0.06018801 + layer.3.k_cache 0.02533426 22.73900754 + layer.3.v_cache 0.00002127 0.07241439 + layer.4.k_cache 0.00070809 1.87663767 + layer.4.v_cache 0.00005072 0.13839871 + layer.4.output 1.53833902 272.73965811 + ------------------------------------------------------------------------------------- + TOTAL 0.65731915 145.80196251 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 18264 +BPFP 0.0844 bits/point +EBPFP 0.1687 equivalent bits/point +MSE 145.801963 +---------------------- -------------------------------------------------------- +Time: 0.543s Load: 0.007s, Pack+Encode: 0.219s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 145.8020 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,056B, BPFP=0.0917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,164B, BPFP=0.1010 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,248B, BPFP=0.1083 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,080B, BPFP=0.0938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,208B, BPFP=0.1049 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,120B, BPFP=0.0972 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,168B, BPFP=0.1014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,108B, BPFP=0.0962 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,280B, BPFP=0.1111 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,188B, BPFP=0.1031 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,856B, BPFP=0.0354 +⌛️ [2/4] FRONTEND: Frontend time: 0.201s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.255s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258550 512.10985243 + layer.0.v_cache 0.00001767 0.05677293 + layer.1.k_cache 0.14790291 35.78949110 + layer.1.v_cache 0.00000629 0.01966532 + layer.2.k_cache 0.00999225 6.49104886 + layer.2.v_cache 0.00002110 0.06143853 + layer.3.k_cache 0.02117847 23.08872070 + layer.3.v_cache 0.00002045 0.07515235 + layer.4.k_cache 0.00067024 1.90403256 + layer.4.v_cache 0.00006105 0.14588856 + layer.4.output 0.00897616 307.21101190 + ------------------------------------------------------------------------------------- + TOTAL 0.02207583 160.60112628 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 14476 +BPFP 0.0739 bits/point +EBPFP 0.1478 equivalent bits/point +MSE 160.601126 +---------------------- -------------------------------------------------------- +Time: 0.462s Load: 0.006s, Pack+Encode: 0.201s, Decode+Unpack: 0.255s +---------------------- -------------------------------------------------------- +💾 Converting with 160.6011 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,244B, BPFP=0.0953 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,464B, BPFP=0.1121 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,588B, BPFP=0.1216 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,332B, BPFP=0.1020 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,464B, BPFP=0.1121 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,396B, BPFP=0.1069 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,436B, BPFP=0.1100 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,420B, BPFP=0.1088 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,640B, BPFP=0.1256 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,480B, BPFP=0.1134 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,712B, BPFP=0.0406 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09364704 513.16831342 + layer.0.v_cache 0.00001853 0.05519159 + layer.1.k_cache 0.25163280 35.61601945 + layer.1.v_cache 0.00000592 0.01939661 + layer.2.k_cache 0.01595125 6.41726086 + layer.2.v_cache 0.00002050 0.05856559 + layer.3.k_cache 0.04325477 22.93643248 + layer.3.v_cache 0.00002016 0.06993809 + layer.4.k_cache 0.00069210 1.84535726 + layer.4.v_cache 0.00005026 0.13740311 + layer.4.output 1.50065788 265.90594363 + ------------------------------------------------------------------------------------- + TOTAL 0.64175873 143.62738140 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 18176 +BPFP 0.0819 bits/point +EBPFP 0.1638 equivalent bits/point +MSE 143.627381 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.007s, Pack+Encode: 0.225s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 143.6274 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,196B, BPFP=0.0827 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,408B, BPFP=0.0973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,512B, BPFP=0.1045 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,284B, BPFP=0.0888 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,404B, BPFP=0.0971 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,348B, BPFP=0.0932 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,364B, BPFP=0.0943 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,352B, BPFP=0.0935 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,572B, BPFP=0.1087 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,504B, BPFP=0.1040 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,320B, BPFP=0.0328 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510114 502.70810288 + layer.0.v_cache 0.00001717 0.05396133 + layer.1.k_cache 0.31797088 35.45664667 + layer.1.v_cache 0.00000673 0.01976044 + layer.2.k_cache 0.01711024 6.32631663 + layer.2.v_cache 0.00002022 0.05969599 + layer.3.k_cache 0.01307247 22.64344882 + layer.3.v_cache 0.00002089 0.07088687 + layer.4.k_cache 0.00069413 1.80842401 + layer.4.v_cache 0.00005461 0.13795527 + layer.4.output 1.35466395 240.43947535 + ------------------------------------------------------------------------------------- + TOTAL 0.58392448 132.49185449 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 17264 +BPFP 0.0702 bits/point +EBPFP 0.1404 equivalent bits/point +MSE 132.491854 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.007s, Pack+Encode: 0.225s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 132.4919 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,320B, BPFP=0.0832 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,480B, BPFP=0.0932 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,576B, BPFP=0.0993 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,352B, BPFP=0.0852 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,496B, BPFP=0.0943 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,412B, BPFP=0.0890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,444B, BPFP=0.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,424B, BPFP=0.0897 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,620B, BPFP=0.1021 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,488B, BPFP=0.0938 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,308B, BPFP=0.0298 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14860459 512.14931956 + layer.0.v_cache 0.00001629 0.05563807 + layer.1.k_cache 0.38266554 35.56220861 + layer.1.v_cache 0.00000669 0.02094104 + layer.2.k_cache 0.01628031 6.44268060 + layer.2.v_cache 0.00002106 0.06348884 + layer.3.k_cache 0.01842327 22.91280045 + layer.3.v_cache 0.00002289 0.07595415 + layer.4.k_cache 0.00072690 1.89944224 + layer.4.v_cache 0.00004923 0.14214330 + layer.4.output 1.23454798 229.03434620 + ------------------------------------------------------------------------------------- + TOTAL 0.54168545 128.38617884 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 17920 +BPFP 0.0664 bits/point +EBPFP 0.1328 equivalent bits/point +MSE 128.386179 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.008s, Pack+Encode: 0.226s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 128.3862 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 435, 128) +Output shape: (1, 435, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.output: torch.Size([1, 435, 3584]) -> torch.Size([1, 1, 435, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,168B, BPFP=0.0779 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,464B, BPFP=0.0885 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,632B, BPFP=0.0945 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,224B, BPFP=0.0799 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,516B, BPFP=0.0904 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,292B, BPFP=0.0823 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,360B, BPFP=0.0848 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,324B, BPFP=0.0835 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,748B, BPFP=0.0987 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,532B, BPFP=0.0909 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,324B, BPFP=0.0273 +⌛️ [2/4] FRONTEND: Frontend time: 0.326s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.494s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15864419 504.45696839 + layer.0.v_cache 0.00001607 0.05443615 + layer.1.k_cache 0.94329666 34.81685300 + layer.1.v_cache 0.00000639 0.01991704 + layer.2.k_cache 0.02505282 6.21698545 + layer.2.v_cache 0.00002131 0.06138909 + layer.3.k_cache 0.01724711 22.48749102 + layer.3.v_cache 0.00002032 0.07063951 + layer.4.k_cache 0.00074513 1.84863941 + layer.4.v_cache 0.00005173 0.13789252 + layer.4.output 0.00608059 127.99347291 + ------------------------------------------------------------------------------------- + TOTAL 0.06986270 86.24267776 + (elements=3,786,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3786240 +Total Bytes 29584 +BPFP 0.0625 bits/point +EBPFP 0.1250 equivalent bits/point +MSE 86.242678 +---------------------- -------------------------------------------------------- +Time: 0.834s Load: 0.014s, Pack+Encode: 0.326s, Decode+Unpack: 0.494s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2427 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 437, 128) +Output shape: (1, 437, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.output: torch.Size([1, 437, 3584]) -> torch.Size([1, 1, 437, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,176B, BPFP=0.0778 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,448B, BPFP=0.0875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,636B, BPFP=0.0943 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 2,224B, BPFP=0.0795 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,496B, BPFP=0.0892 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 2,320B, BPFP=0.0830 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,368B, BPFP=0.0847 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,360B, BPFP=0.0844 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,780B, BPFP=0.0994 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,592B, BPFP=0.0927 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,572B, BPFP=0.0285 +⌛️ [2/4] FRONTEND: Frontend time: 0.325s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.482s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14454122 504.80549199 + layer.0.v_cache 0.00001702 0.05405769 + layer.1.k_cache 0.94267751 34.91353958 + layer.1.v_cache 0.00000668 0.02043125 + layer.2.k_cache 0.01808156 6.23472417 + layer.2.v_cache 0.00002154 0.06225198 + layer.3.k_cache 0.03034366 22.55584507 + layer.3.v_cache 0.00002210 0.07476224 + layer.4.k_cache 0.00074160 1.88230718 + layer.4.v_cache 0.00005552 0.14534461 + layer.4.output 0.00611241 127.66738517 + ------------------------------------------------------------------------------------- + TOTAL 0.06937031 86.14237953 + (elements=3,803,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3803648 +Total Bytes 29972 +BPFP 0.0630 bits/point +EBPFP 0.1261 equivalent bits/point +MSE 86.142380 +---------------------- -------------------------------------------------------- +Time: 0.821s Load: 0.014s, Pack+Encode: 0.325s, Decode+Unpack: 0.482s +---------------------- -------------------------------------------------------- +💾 Converting with 86.1424 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,640B, BPFP=0.0832 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,812B, BPFP=0.0919 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,932B, BPFP=0.0980 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,704B, BPFP=0.0864 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,848B, BPFP=0.0938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,728B, BPFP=0.0877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,760B, BPFP=0.0893 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,696B, BPFP=0.0860 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,028B, BPFP=0.1029 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,860B, BPFP=0.0944 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,472B, BPFP=0.0324 +⌛️ [2/4] FRONTEND: Frontend time: 0.268s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12827574 506.49553571 + layer.0.v_cache 0.00001866 0.05542587 + layer.1.k_cache 0.53653341 34.98642324 + layer.1.v_cache 0.00000626 0.01959205 + layer.2.k_cache 0.01950284 6.37309721 + layer.2.v_cache 0.00002083 0.06053529 + layer.3.k_cache 0.04340698 22.61460024 + layer.3.v_cache 0.00002181 0.07339813 + layer.4.k_cache 0.00072091 1.85728851 + layer.4.v_cache 0.00005393 0.14400552 + layer.4.output 0.04342448 177.56405090 + ------------------------------------------------------------------------------------- + TOTAL 0.06073722 106.80166224 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 22480 +BPFP 0.0671 bits/point +EBPFP 0.1342 equivalent bits/point +MSE 106.801662 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.011s, Pack+Encode: 0.268s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 106.8017 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,540B, BPFP=0.0905 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,736B, BPFP=0.1020 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,896B, BPFP=0.1114 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,624B, BPFP=0.0954 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,748B, BPFP=0.1027 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,640B, BPFP=0.0963 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,724B, BPFP=0.1013 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,640B, BPFP=0.0963 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,044B, BPFP=0.1201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,816B, BPFP=0.1067 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,340B, BPFP=0.0364 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702504 501.61407425 + layer.0.v_cache 0.00001641 0.05501344 + layer.1.k_cache 0.51130361 35.25130331 + layer.1.v_cache 0.00000656 0.02060179 + layer.2.k_cache 0.01810413 6.34535802 + layer.2.v_cache 0.00002122 0.06320540 + layer.3.k_cache 0.02476560 22.84987554 + layer.3.v_cache 0.00002197 0.07345177 + layer.4.k_cache 0.00071356 1.87610173 + layer.4.v_cache 0.00005290 0.14462194 + layer.4.output 0.00501064 208.83294173 + ------------------------------------------------------------------------------------- + TOTAL 0.04100620 119.41907055 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 21748 +BPFP 0.0751 bits/point +EBPFP 0.1503 equivalent bits/point +MSE 119.419071 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.008s, Pack+Encode: 0.257s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 119.4191 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,332B, BPFP=0.0889 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,472B, BPFP=0.0983 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,608B, BPFP=0.1074 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,372B, BPFP=0.0916 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,524B, BPFP=0.1018 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,416B, BPFP=0.0946 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,468B, BPFP=0.0980 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,440B, BPFP=0.0962 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,676B, BPFP=0.1119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,528B, BPFP=0.1020 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,636B, BPFP=0.0347 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11892304 509.81727431 + layer.0.v_cache 0.00001803 0.05619236 + layer.1.k_cache 0.31174104 35.61200003 + layer.1.v_cache 0.00000657 0.02033208 + layer.2.k_cache 0.01452858 6.41523051 + layer.2.v_cache 0.00002161 0.06321990 + layer.3.k_cache 0.02263702 22.71646927 + layer.3.v_cache 0.00002141 0.07564925 + layer.4.k_cache 0.00068312 1.85296422 + layer.4.v_cache 0.00005251 0.14365046 + layer.4.output 1.30838121 232.15508623 + ------------------------------------------------------------------------------------- + TOTAL 0.56631185 129.52109330 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 18472 +BPFP 0.0726 bits/point +EBPFP 0.1451 equivalent bits/point +MSE 129.521093 +---------------------- -------------------------------------------------------- +Time: 0.560s Load: 0.008s, Pack+Encode: 0.226s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 129.5211 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 233, 128) +Output shape: (1, 233, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.output: torch.Size([1, 233, 3584]) -> torch.Size([1, 1, 233, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,336B, BPFP=0.0896 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,476B, BPFP=0.0990 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,616B, BPFP=0.1084 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,396B, BPFP=0.0936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,528B, BPFP=0.1025 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,440B, BPFP=0.0966 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,468B, BPFP=0.0984 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,436B, BPFP=0.0963 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,684B, BPFP=0.1129 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,540B, BPFP=0.1033 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,776B, BPFP=0.0362 +⌛️ [2/4] FRONTEND: Frontend time: 0.295s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13365312 505.07054721 + layer.0.v_cache 0.00001735 0.05542086 + layer.1.k_cache 0.40872042 35.29331243 + layer.1.v_cache 0.00000792 0.02113637 + layer.2.k_cache 0.01308113 6.26156944 + layer.2.v_cache 0.00002137 0.06255221 + layer.3.k_cache 0.00601081 22.75687575 + layer.3.v_cache 0.00002094 0.07153119 + layer.4.k_cache 0.00068546 1.82709565 + layer.4.v_cache 0.00005338 0.14362568 + layer.4.output 1.31399239 232.64739807 + ------------------------------------------------------------------------------------- + TOTAL 0.57413051 129.41737960 + (elements=2,028,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2028032 +Total Bytes 18696 +BPFP 0.0738 bits/point +EBPFP 0.1475 equivalent bits/point +MSE 129.417380 +---------------------- -------------------------------------------------------- +Time: 0.663s Load: 0.008s, Pack+Encode: 0.295s, Decode+Unpack: 0.359s +---------------------- -------------------------------------------------------- +💾 Converting with 129.4174 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 148, 128) +Output shape: (1, 148, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.output: torch.Size([1, 148, 3584]) -> torch.Size([1, 1, 148, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,064B, BPFP=0.1123 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,172B, BPFP=0.1237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,292B, BPFP=0.1364 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,040B, BPFP=0.1098 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,200B, BPFP=0.1267 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,072B, BPFP=0.1132 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,152B, BPFP=0.1216 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,076B, BPFP=0.1136 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,364B, BPFP=0.1440 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,252B, BPFP=0.1322 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,856B, BPFP=0.0431 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.283s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08608804 509.28135557 + layer.0.v_cache 0.00001767 0.05387218 + layer.1.k_cache 0.12276106 35.83729017 + layer.1.v_cache 0.00000594 0.01927992 + layer.2.k_cache 0.00886028 6.37213795 + layer.2.v_cache 0.00002088 0.06022782 + layer.3.k_cache 0.01513371 23.00231109 + layer.3.v_cache 0.00002085 0.07292894 + layer.4.k_cache 0.00068430 1.83526900 + layer.4.v_cache 0.00005331 0.14342365 + layer.4.output 0.08957551 367.38492399 + ------------------------------------------------------------------------------------- + TOTAL 0.05062792 185.19838613 + (elements=1,288,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1288192 +Total Bytes 14540 +BPFP 0.0903 bits/point +EBPFP 0.1806 equivalent bits/point +MSE 185.198386 +---------------------- -------------------------------------------------------- +Time: 0.539s Load: 0.006s, Pack+Encode: 0.251s, Decode+Unpack: 0.283s +---------------------- -------------------------------------------------------- +💾 Converting with 185.1984 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 170, 128) +Output shape: (1, 170, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.output: torch.Size([1, 170, 3584]) -> torch.Size([1, 1, 170, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,072B, BPFP=0.0985 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,156B, BPFP=0.1062 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,272B, BPFP=0.1169 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,092B, BPFP=0.1004 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,208B, BPFP=0.1110 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,124B, BPFP=0.1033 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,188B, BPFP=0.1092 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,152B, BPFP=0.1059 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,340B, BPFP=0.1232 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,216B, BPFP=0.1118 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,156B, BPFP=0.0414 +⌛️ [2/4] FRONTEND: Frontend time: 0.241s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.306s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13519069 520.08805147 + layer.0.v_cache 0.00001779 0.05609668 + layer.1.k_cache 0.14803200 36.09543313 + layer.1.v_cache 0.00000678 0.02039082 + layer.2.k_cache 0.02267574 6.48149055 + layer.2.v_cache 0.00002005 0.05896275 + layer.3.k_cache 0.04618364 23.01405963 + layer.3.v_cache 0.00002282 0.07608754 + layer.4.k_cache 0.00077641 1.88092041 + layer.4.v_cache 0.00005264 0.14135306 + layer.4.output 0.00945594 323.98668592 + ------------------------------------------------------------------------------------- + TOTAL 0.02465707 167.98939103 + (elements=1,479,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1479680 +Total Bytes 14976 +BPFP 0.0810 bits/point +EBPFP 0.1619 equivalent bits/point +MSE 167.989391 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.008s, Pack+Encode: 0.241s, Decode+Unpack: 0.306s +---------------------- -------------------------------------------------------- +💾 Converting with 167.9894 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 988B, BPFP=0.0930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,100B, BPFP=0.1035 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,208B, BPFP=0.1137 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,064B, BPFP=0.1002 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,144B, BPFP=0.1077 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,056B, BPFP=0.0994 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,112B, BPFP=0.1047 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,060B, BPFP=0.0998 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,248B, BPFP=0.1175 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,156B, BPFP=0.1088 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,876B, BPFP=0.0387 +⌛️ [2/4] FRONTEND: Frontend time: 0.240s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.305s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12520085 512.78628577 + layer.0.v_cache 0.00001716 0.05682905 + layer.1.k_cache 0.15152623 35.95626059 + layer.1.v_cache 0.00000633 0.02027833 + layer.2.k_cache 0.01605446 6.57049855 + layer.2.v_cache 0.00002200 0.06308494 + layer.3.k_cache 0.04332644 22.95489281 + layer.3.v_cache 0.00002349 0.07500208 + layer.4.k_cache 0.00069336 1.90929514 + layer.4.v_cache 0.00005614 0.13882747 + layer.4.output 0.00966471 331.53103485 + ------------------------------------------------------------------------------------- + TOTAL 0.02379879 170.66167639 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 14012 +BPFP 0.0776 bits/point +EBPFP 0.1552 equivalent bits/point +MSE 170.661676 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.008s, Pack+Encode: 0.240s, Decode+Unpack: 0.305s +---------------------- -------------------------------------------------------- +💾 Converting with 170.6617 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 181, 128) +Output shape: (1, 181, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.output: torch.Size([1, 181, 3584]) -> torch.Size([1, 1, 181, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,060B, BPFP=0.0915 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,128B, BPFP=0.0974 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,248B, BPFP=0.1077 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,064B, BPFP=0.0919 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,196B, BPFP=0.1032 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,124B, BPFP=0.0970 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,176B, BPFP=0.1015 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,080B, BPFP=0.0932 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,264B, BPFP=0.1091 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,204B, BPFP=0.1039 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,736B, BPFP=0.0337 +⌛️ [2/4] FRONTEND: Frontend time: 0.196s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13566811 512.75302141 + layer.0.v_cache 0.00002016 0.05582512 + layer.1.k_cache 0.14782082 35.90255147 + layer.1.v_cache 0.00000634 0.02057961 + layer.2.k_cache 0.01546242 6.56350489 + layer.2.v_cache 0.00002121 0.06068940 + layer.3.k_cache 0.02895848 23.21122615 + layer.3.v_cache 0.00002197 0.07541142 + layer.4.k_cache 0.00067433 1.90736785 + layer.4.v_cache 0.00005229 0.14025098 + layer.4.output 0.00892717 305.23658248 + ------------------------------------------------------------------------------------- + TOTAL 0.02301155 159.84391210 + (elements=1,575,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1575424 +Total Bytes 14280 +BPFP 0.0725 bits/point +EBPFP 0.1450 equivalent bits/point +MSE 159.843912 +---------------------- -------------------------------------------------------- +Time: 0.474s Load: 0.006s, Pack+Encode: 0.196s, Decode+Unpack: 0.272s +---------------------- -------------------------------------------------------- +💾 Converting with 159.8439 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 186, 128) +Output shape: (1, 186, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.output: torch.Size([1, 186, 3584]) -> torch.Size([1, 1, 186, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,012B, BPFP=0.0850 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,100B, BPFP=0.0924 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,184B, BPFP=0.0995 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,020B, BPFP=0.0857 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,104B, BPFP=0.0927 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,048B, BPFP=0.0880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,080B, BPFP=0.0907 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,048B, BPFP=0.0880 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,208B, BPFP=0.1015 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,120B, BPFP=0.0941 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,536B, BPFP=0.0304 +⌛️ [2/4] FRONTEND: Frontend time: 0.189s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.287s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09795722 507.52242944 + layer.0.v_cache 0.00001815 0.05648491 + layer.1.k_cache 0.13430934 35.57229188 + layer.1.v_cache 0.00000632 0.02023554 + layer.2.k_cache 0.01830538 6.53404712 + layer.2.v_cache 0.00002120 0.06009641 + layer.3.k_cache 0.04892764 22.78547127 + layer.3.v_cache 0.00002113 0.07425086 + layer.4.k_cache 0.00067235 1.86480762 + layer.4.v_cache 0.00008605 0.14110117 + layer.4.output 0.00869856 296.43087558 + ------------------------------------------------------------------------------------- + TOTAL 0.02124792 155.86160854 + (elements=1,618,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1618944 +Total Bytes 13460 +BPFP 0.0665 bits/point +EBPFP 0.1330 equivalent bits/point +MSE 155.861609 +---------------------- -------------------------------------------------------- +Time: 0.482s Load: 0.006s, Pack+Encode: 0.189s, Decode+Unpack: 0.287s +---------------------- -------------------------------------------------------- +💾 Converting with 155.8616 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 163, 128) +Output shape: (1, 163, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.output: torch.Size([1, 163, 3584]) -> torch.Size([1, 1, 163, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 940B, BPFP=0.0901 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,076B, BPFP=0.1031 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,156B, BPFP=0.1108 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 976B, BPFP=0.0936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,096B, BPFP=0.1051 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,020B, BPFP=0.0978 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,064B, BPFP=0.1020 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,044B, BPFP=0.1001 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,236B, BPFP=0.1185 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,168B, BPFP=0.1120 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,604B, BPFP=0.0357 +⌛️ [2/4] FRONTEND: Frontend time: 0.208s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.252s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11530171 509.17642830 + layer.0.v_cache 0.00001813 0.05603750 + layer.1.k_cache 0.14319220 36.08901996 + layer.1.v_cache 0.00000612 0.02016791 + layer.2.k_cache 0.02368733 6.56183947 + layer.2.v_cache 0.00002106 0.05907435 + layer.3.k_cache 0.02525310 22.55453323 + layer.3.v_cache 0.00002161 0.07533763 + layer.4.k_cache 0.00069984 1.88149107 + layer.4.v_cache 0.00005709 0.14211257 + layer.4.output 0.00982084 337.78429010 + ------------------------------------------------------------------------------------- + TOTAL 0.02217671 173.00623957 + (elements=1,418,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1418752 +Total Bytes 13380 +BPFP 0.0754 bits/point +EBPFP 0.1509 equivalent bits/point +MSE 173.006240 +---------------------- -------------------------------------------------------- +Time: 0.465s Load: 0.005s, Pack+Encode: 0.208s, Decode+Unpack: 0.252s +---------------------- -------------------------------------------------------- +💾 Converting with 173.0062 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 191, 128) +Output shape: (1, 191, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.output: torch.Size([1, 191, 3584]) -> torch.Size([1, 1, 191, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 856B, BPFP=0.0700 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 944B, BPFP=0.0772 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,072B, BPFP=0.0877 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 824B, BPFP=0.0674 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 988B, BPFP=0.0808 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 856B, BPFP=0.0700 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 928B, BPFP=0.0759 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 848B, BPFP=0.0694 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,116B, BPFP=0.0913 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,004B, BPFP=0.0821 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,960B, BPFP=0.0229 +⌛️ [2/4] FRONTEND: Frontend time: 0.188s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.257s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15112510 503.33920975 + layer.0.v_cache 0.00002069 0.05589615 + layer.1.k_cache 0.18491925 34.35338780 + layer.1.v_cache 0.00000629 0.01976433 + layer.2.k_cache 0.02122488 6.17019286 + layer.2.v_cache 0.00002327 0.06049251 + layer.3.k_cache 0.01259250 22.30854518 + layer.3.v_cache 0.00002182 0.07261910 + layer.4.k_cache 0.00068006 1.82502850 + layer.4.v_cache 0.00005885 0.14255569 + layer.4.output 0.00845948 287.65148186 + ------------------------------------------------------------------------------------- + TOTAL 0.02528759 151.87694499 + (elements=1,662,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1662464 +Total Bytes 11396 +BPFP 0.0548 bits/point +EBPFP 0.1097 equivalent bits/point +MSE 151.876945 +---------------------- -------------------------------------------------------- +Time: 0.452s Load: 0.006s, Pack+Encode: 0.188s, Decode+Unpack: 0.257s +---------------------- -------------------------------------------------------- +💾 Converting with 151.8769 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 174, 128) +Output shape: (1, 174, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.output: torch.Size([1, 174, 3584]) -> torch.Size([1, 1, 174, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,036B, BPFP=0.0930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,124B, BPFP=0.1009 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,256B, BPFP=0.1128 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,032B, BPFP=0.0927 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,204B, BPFP=0.1081 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,080B, BPFP=0.0970 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,168B, BPFP=0.1049 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,084B, BPFP=0.0973 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,296B, BPFP=0.1164 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,192B, BPFP=0.1070 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,908B, BPFP=0.0373 +⌛️ [2/4] FRONTEND: Frontend time: 0.189s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.251s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14146894 519.30154454 + layer.0.v_cache 0.00001664 0.05489121 + layer.1.k_cache 0.16292955 36.05694370 + layer.1.v_cache 0.00000602 0.01994443 + layer.2.k_cache 0.01834387 6.57400688 + layer.2.v_cache 0.00002001 0.06075999 + layer.3.k_cache 0.03015787 23.22799648 + layer.3.v_cache 0.00002180 0.07331116 + layer.4.k_cache 0.00067479 1.90403379 + layer.4.v_cache 0.00005130 0.14401504 + layer.4.output 0.00923280 316.49633108 + ------------------------------------------------------------------------------------- + TOTAL 0.02460708 164.87598616 + (elements=1,514,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1514496 +Total Bytes 14380 +BPFP 0.0760 bits/point +EBPFP 0.1519 equivalent bits/point +MSE 164.875986 +---------------------- -------------------------------------------------------- +Time: 0.447s Load: 0.006s, Pack+Encode: 0.189s, Decode+Unpack: 0.251s +---------------------- -------------------------------------------------------- +💾 Converting with 164.8760 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.0716 bits/point +Avg EBPFP 0.1433 equivalent bits/point +Avg MSE 134.924264 +Avg Time 0.582s +------------------------ ---------------------------- diff --git a/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..5346d4bcad571c7e7da0148ff8d90a69b980f5e5 --- /dev/null +++ b/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 405 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa +Output output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,496B, BPFP=0.0812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,796B, BPFP=0.0974 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,848B, BPFP=0.1003 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,564B, BPFP=0.0849 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,720B, BPFP=0.0933 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,648B, BPFP=0.0894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,672B, BPFP=0.0907 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,632B, BPFP=0.0885 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,936B, BPFP=0.1050 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,864B, BPFP=0.1011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,924B, BPFP=0.0304 +⌛️ [2/4] FRONTEND: Frontend time: 0.677s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.415s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14134069 505.05924479 + layer.0.v_cache 0.00001405 0.04617451 + layer.1.k_cache 0.62039534 35.10556369 + layer.1.v_cache 0.00000582 0.01693695 + layer.2.k_cache 0.00676800 6.27206082 + layer.2.v_cache 0.00001893 0.05241613 + layer.3.k_cache 0.03896382 22.64229499 + layer.3.v_cache 0.00001942 0.06336738 + layer.4.k_cache 0.00069280 1.67642678 + layer.4.v_cache 0.00005238 0.13525515 + layer.4.output 0.00804578 194.03132750 + ------------------------------------------------------------------------------------- + TOTAL 0.05085834 113.48759022 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 21100 +BPFP 0.0673 bits/point +EBPFP 0.1347 equivalent bits/point +MSE 113.487590 +---------------------- -------------------------------------------------------- +Time: 1.105s Load: 0.012s, Pack+Encode: 0.677s, Decode+Unpack: 0.415s +---------------------- -------------------------------------------------------- +💾 Converting with 113.4876 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,552B, BPFP=0.0828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,780B, BPFP=0.0949 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,892B, BPFP=0.1009 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,616B, BPFP=0.0862 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,764B, BPFP=0.0941 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,676B, BPFP=0.0894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,692B, BPFP=0.0902 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,692B, BPFP=0.0902 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.1058 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,888B, BPFP=0.1007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,108B, BPFP=0.0313 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15682776 507.45125853 + layer.0.v_cache 0.00001501 0.04734036 + layer.1.k_cache 0.54447479 35.04300875 + layer.1.v_cache 0.00000580 0.01741894 + layer.2.k_cache 0.00853899 6.26445096 + layer.2.v_cache 0.00001972 0.05391472 + layer.3.k_cache 0.02401569 22.61360288 + layer.3.v_cache 0.00001939 0.06525070 + layer.4.k_cache 0.00071059 1.71407029 + layer.4.v_cache 0.00005329 0.13919143 + layer.4.output 0.05121508 185.56384081 + ------------------------------------------------------------------------------------- + TOTAL 0.06430510 110.13861137 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 21644 +BPFP 0.0679 bits/point +EBPFP 0.1358 equivalent bits/point +MSE 110.138611 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.011s, Pack+Encode: 0.261s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 110.1386 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,568B, BPFP=0.0833 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,804B, BPFP=0.0959 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,896B, BPFP=0.1008 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,636B, BPFP=0.0869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,796B, BPFP=0.0955 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,680B, BPFP=0.0893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,720B, BPFP=0.0914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,700B, BPFP=0.0903 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,012B, BPFP=0.1069 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,884B, BPFP=0.1001 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,204B, BPFP=0.0319 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15363285 505.54474915 + layer.0.v_cache 0.00001412 0.04669566 + layer.1.k_cache 0.60852461 35.07004013 + layer.1.v_cache 0.00000580 0.01691309 + layer.2.k_cache 0.01182756 6.23954970 + layer.2.v_cache 0.00001863 0.05217359 + layer.3.k_cache 0.03467803 22.64770674 + layer.3.v_cache 0.00001919 0.06394223 + layer.4.k_cache 0.00072319 1.67208966 + layer.4.v_cache 0.00005301 0.13364967 + layer.4.output 0.05035121 185.06427660 + ------------------------------------------------------------------------------------- + TOTAL 0.06835032 109.81984975 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 21900 +BPFP 0.0685 bits/point +EBPFP 0.1369 equivalent bits/point +MSE 109.819850 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 109.8198 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,500B, BPFP=0.0828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,772B, BPFP=0.0978 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,840B, BPFP=0.1016 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,552B, BPFP=0.0857 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,728B, BPFP=0.0954 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,632B, BPFP=0.0901 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,640B, BPFP=0.0905 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,608B, BPFP=0.0888 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,940B, BPFP=0.1071 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,852B, BPFP=0.1023 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,080B, BPFP=0.0322 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12197856 509.72305654 + layer.0.v_cache 0.00001365 0.04608056 + layer.1.k_cache 0.62524759 34.81430474 + layer.1.v_cache 0.00000551 0.01647282 + layer.2.k_cache 0.01486346 6.25689460 + layer.2.v_cache 0.00001981 0.05317055 + layer.3.k_cache 0.07387851 22.40344799 + layer.3.v_cache 0.00001943 0.06249441 + layer.4.k_cache 0.00069560 1.65591075 + layer.4.v_cache 0.00005405 0.13872571 + layer.4.output 0.01098318 197.02790573 + ------------------------------------------------------------------------------------- + TOTAL 0.05374461 114.96269993 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 21144 +BPFP 0.0687 bits/point +EBPFP 0.1373 equivalent bits/point +MSE 114.962700 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.010s, Pack+Encode: 0.256s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 114.9627 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,468B, BPFP=0.0805 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,732B, BPFP=0.0950 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,832B, BPFP=0.1004 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,520B, BPFP=0.0833 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,712B, BPFP=0.0939 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,604B, BPFP=0.0879 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,644B, BPFP=0.0901 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,604B, BPFP=0.0879 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,924B, BPFP=0.1055 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,808B, BPFP=0.0991 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,952B, BPFP=0.0310 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15362121 505.19654605 + layer.0.v_cache 0.00001420 0.04671369 + layer.1.k_cache 0.52952420 34.98122259 + layer.1.v_cache 0.00000567 0.01686110 + layer.2.k_cache 0.00939641 6.17906944 + layer.2.v_cache 0.00001891 0.05137735 + layer.3.k_cache 0.04339437 22.50396107 + layer.3.v_cache 0.00002732 0.06375965 + layer.4.k_cache 0.00068326 1.65734885 + layer.4.v_cache 0.00005156 0.13435915 + layer.4.output 0.00776138 194.92236842 + ------------------------------------------------------------------------------------- + TOTAL 0.04653334 113.84045870 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 20800 +BPFP 0.0671 bits/point +EBPFP 0.1342 equivalent bits/point +MSE 113.840459 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 113.8405 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,564B, BPFP=0.0831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,788B, BPFP=0.0950 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,896B, BPFP=0.1008 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,636B, BPFP=0.0869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.0942 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,680B, BPFP=0.0893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,712B, BPFP=0.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,680B, BPFP=0.0893 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,976B, BPFP=0.1050 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,860B, BPFP=0.0989 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,216B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14921317 503.81122449 + layer.0.v_cache 0.00001395 0.04685556 + layer.1.k_cache 0.64083395 35.12791972 + layer.1.v_cache 0.00000574 0.01720768 + layer.2.k_cache 0.00578804 6.29655072 + layer.2.v_cache 0.00001912 0.05331941 + layer.3.k_cache 0.02848778 22.54247216 + layer.3.v_cache 0.00001956 0.06296152 + layer.4.k_cache 0.00070939 1.67462573 + layer.4.v_cache 0.00005206 0.13517294 + layer.4.output 0.04916091 185.09119898 + ------------------------------------------------------------------------------------- + TOTAL 0.06878054 109.72980605 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 21780 +BPFP 0.0681 bits/point +EBPFP 0.1362 equivalent bits/point +MSE 109.729806 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.011s, Pack+Encode: 0.254s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 109.7298 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,556B, BPFP=0.0830 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,800B, BPFP=0.0960 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,892B, BPFP=0.1009 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,628B, BPFP=0.0868 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,768B, BPFP=0.0943 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,704B, BPFP=0.0909 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,704B, BPFP=0.0909 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,684B, BPFP=0.0898 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.1058 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,864B, BPFP=0.0994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,196B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12806672 503.13379906 + layer.0.v_cache 0.00001388 0.04671619 + layer.1.k_cache 0.56701790 35.03234988 + layer.1.v_cache 0.00000557 0.01674508 + layer.2.k_cache 0.01020011 6.21191031 + layer.2.v_cache 0.00001881 0.05276228 + layer.3.k_cache 0.04301871 22.61228636 + layer.3.v_cache 0.00002051 0.06450787 + layer.4.k_cache 0.00071755 1.67818901 + layer.4.v_cache 0.00005284 0.13344163 + layer.4.output 0.04940383 185.55850804 + ------------------------------------------------------------------------------------- + TOTAL 0.06440938 109.87601553 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 21780 +BPFP 0.0683 bits/point +EBPFP 0.1366 equivalent bits/point +MSE 109.876016 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.012s, Pack+Encode: 0.254s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 109.8760 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,468B, BPFP=0.0805 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,760B, BPFP=0.0965 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,824B, BPFP=0.1000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,564B, BPFP=0.0857 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,720B, BPFP=0.0943 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,616B, BPFP=0.0886 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,652B, BPFP=0.0906 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,584B, BPFP=0.0868 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,936B, BPFP=0.1061 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,804B, BPFP=0.0989 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,884B, BPFP=0.0304 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12903612 508.73130482 + layer.0.v_cache 0.00001458 0.04639351 + layer.1.k_cache 0.59860808 34.96667009 + layer.1.v_cache 0.00000581 0.01685630 + layer.2.k_cache 0.00940504 6.18568479 + layer.2.v_cache 0.00001954 0.05354046 + layer.3.k_cache 0.02644065 22.47408683 + layer.3.v_cache 0.00002029 0.06382568 + layer.4.k_cache 0.00068693 1.65939684 + layer.4.v_cache 0.00005376 0.13547759 + layer.4.output 0.00862506 195.19338972 + ------------------------------------------------------------------------------------- + TOTAL 0.04850978 114.15805677 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 20812 +BPFP 0.0671 bits/point +EBPFP 0.1342 equivalent bits/point +MSE 114.158057 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.009s, Pack+Encode: 0.256s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 114.1581 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,668B, BPFP=0.0846 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,860B, BPFP=0.0944 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,940B, BPFP=0.0984 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,676B, BPFP=0.0850 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,820B, BPFP=0.0923 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,728B, BPFP=0.0877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,772B, BPFP=0.0899 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,716B, BPFP=0.0871 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,048B, BPFP=0.1039 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,916B, BPFP=0.0972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,020B, BPFP=0.0291 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16469599 506.62525365 + layer.0.v_cache 0.00001357 0.04667061 + layer.1.k_cache 0.70445866 34.96254185 + layer.1.v_cache 0.00000558 0.01688319 + layer.2.k_cache 0.01447187 6.28662030 + layer.2.v_cache 0.00001924 0.05244646 + layer.3.k_cache 0.02617162 22.62832919 + layer.3.v_cache 0.00001974 0.06389328 + layer.4.k_cache 0.00070375 1.67712184 + layer.4.v_cache 0.00005076 0.13525891 + layer.4.output 0.04814868 177.66196081 + ------------------------------------------------------------------------------------- + TOTAL 0.07339127 106.83110264 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 22164 +BPFP 0.0661 bits/point +EBPFP 0.1323 equivalent bits/point +MSE 106.831103 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.256s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 106.8311 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,544B, BPFP=0.0897 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,772B, BPFP=0.1029 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,920B, BPFP=0.1115 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,636B, BPFP=0.0950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,768B, BPFP=0.1027 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,636B, BPFP=0.0950 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,740B, BPFP=0.1011 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,620B, BPFP=0.0941 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.1152 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,904B, BPFP=0.1106 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,360B, BPFP=0.0362 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12132761 507.35612221 + layer.0.v_cache 0.00001366 0.04630668 + layer.1.k_cache 0.57241889 35.24758582 + layer.1.v_cache 0.00000556 0.01647360 + layer.2.k_cache 0.00926234 6.30536184 + layer.2.v_cache 0.00002145 0.05181209 + layer.3.k_cache 0.03109839 22.56255990 + layer.3.v_cache 0.00002257 0.06171392 + layer.4.k_cache 0.00070701 1.67436961 + layer.4.v_cache 0.00004920 0.13465345 + layer.4.output 0.01031505 206.64513078 + ------------------------------------------------------------------------------------- + TOTAL 0.04747836 118.82193380 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 21884 +BPFP 0.0748 bits/point +EBPFP 0.1495 equivalent bits/point +MSE 118.821934 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.010s, Pack+Encode: 0.256s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 118.8219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,620B, BPFP=0.0849 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,868B, BPFP=0.0979 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,948B, BPFP=0.1021 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,680B, BPFP=0.0881 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,832B, BPFP=0.0961 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,724B, BPFP=0.0904 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,760B, BPFP=0.0923 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,732B, BPFP=0.0908 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,036B, BPFP=0.1068 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,940B, BPFP=0.1017 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,288B, BPFP=0.0321 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14357947 508.17722315 + layer.0.v_cache 0.00001420 0.04713532 + layer.1.k_cache 0.63723908 35.19299890 + layer.1.v_cache 0.00000573 0.01680973 + layer.2.k_cache 0.00765126 6.28107550 + layer.2.v_cache 0.00002073 0.05205406 + layer.3.k_cache 0.02843528 22.65987278 + layer.3.v_cache 0.00002112 0.06338537 + layer.4.k_cache 0.00069171 1.66508965 + layer.4.v_cache 0.00005234 0.13389952 + layer.4.output 0.04871205 182.71042066 + ------------------------------------------------------------------------------------- + TOTAL 0.06815855 109.01544051 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 22428 +BPFP 0.0692 bits/point +EBPFP 0.1383 equivalent bits/point +MSE 109.015441 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 109.0154 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,616B, BPFP=0.0828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,852B, BPFP=0.0949 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,932B, BPFP=0.0990 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,696B, BPFP=0.0869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,820B, BPFP=0.0932 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,728B, BPFP=0.0885 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,768B, BPFP=0.0906 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,728B, BPFP=0.0885 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,032B, BPFP=0.1041 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,948B, BPFP=0.0998 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,288B, BPFP=0.0314 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16857396 507.18447746 + layer.0.v_cache 0.00001478 0.04676180 + layer.1.k_cache 0.58716421 35.11410092 + layer.1.v_cache 0.00000600 0.01712570 + layer.2.k_cache 0.00827454 6.30835001 + layer.2.v_cache 0.00002177 0.05272748 + layer.3.k_cache 0.02531170 22.41865394 + layer.3.v_cache 0.00001983 0.06213745 + layer.4.k_cache 0.00069463 1.65527984 + layer.4.v_cache 0.00005316 0.13574040 + layer.4.output 0.04836898 178.47434133 + ------------------------------------------------------------------------------------- + TOTAL 0.06639514 107.19504379 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 22408 +BPFP 0.0675 bits/point +EBPFP 0.1351 equivalent bits/point +MSE 107.195044 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.012s, Pack+Encode: 0.250s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 107.1950 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,624B, BPFP=0.0852 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,876B, BPFP=0.0984 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,944B, BPFP=0.1019 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,696B, BPFP=0.0889 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,852B, BPFP=0.0971 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,752B, BPFP=0.0919 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,776B, BPFP=0.0931 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,756B, BPFP=0.0921 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,068B, BPFP=0.1084 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,952B, BPFP=0.1023 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,296B, BPFP=0.0322 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14028978 507.43351510 + layer.0.v_cache 0.00001397 0.04645262 + layer.1.k_cache 0.59966084 35.18652999 + layer.1.v_cache 0.00000578 0.01698664 + layer.2.k_cache 0.01765286 6.25554437 + layer.2.v_cache 0.00001884 0.05276701 + layer.3.k_cache 0.02653811 22.60375321 + layer.3.v_cache 0.00001935 0.06274967 + layer.4.k_cache 0.00071527 1.67435675 + layer.4.v_cache 0.00005402 0.13556423 + layer.4.output 0.04907703 182.70604626 + ------------------------------------------------------------------------------------- + TOTAL 0.06638282 108.96532608 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 22592 +BPFP 0.0697 bits/point +EBPFP 0.1394 equivalent bits/point +MSE 108.965326 +---------------------- -------------------------------------------------------- +Time: 0.622s Load: 0.012s, Pack+Encode: 0.248s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 108.9653 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,524B, BPFP=0.0875 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,836B, BPFP=0.1055 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,928B, BPFP=0.1108 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,644B, BPFP=0.0944 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,788B, BPFP=0.1027 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,692B, BPFP=0.0972 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,728B, BPFP=0.0993 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,668B, BPFP=0.0958 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,996B, BPFP=0.1147 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,908B, BPFP=0.1096 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,308B, BPFP=0.0354 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13255837 508.44485294 + layer.0.v_cache 0.00001394 0.04645958 + layer.1.k_cache 0.54921044 35.13919965 + layer.1.v_cache 0.00000578 0.01674349 + layer.2.k_cache 0.01292971 6.24675212 + layer.2.v_cache 0.00001901 0.05316358 + layer.3.k_cache 0.07190662 22.57271262 + layer.3.v_cache 0.00001944 0.06274846 + layer.4.k_cache 0.00068159 1.61618827 + layer.4.v_cache 0.00005125 0.13472673 + layer.4.output 0.00790709 204.28180804 + ------------------------------------------------------------------------------------- + TOTAL 0.04839681 117.90036492 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 22020 +BPFP 0.0744 bits/point +EBPFP 0.1488 equivalent bits/point +MSE 117.900365 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.011s, Pack+Encode: 0.258s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 117.9004 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,500B, BPFP=0.0831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,768B, BPFP=0.0980 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,852B, BPFP=0.1026 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,580B, BPFP=0.0875 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,736B, BPFP=0.0962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,632B, BPFP=0.0904 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,676B, BPFP=0.0929 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,608B, BPFP=0.0891 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,948B, BPFP=0.1079 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,844B, BPFP=0.1022 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,148B, BPFP=0.0328 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13695534 508.58189273 + layer.0.v_cache 0.00001452 0.04728647 + layer.1.k_cache 0.52272531 34.98662248 + layer.1.v_cache 0.00000574 0.01726095 + layer.2.k_cache 0.00793556 6.27361887 + layer.2.v_cache 0.00001907 0.05346562 + layer.3.k_cache 0.04682703 22.65720405 + layer.3.v_cache 0.00002008 0.06398506 + layer.4.k_cache 0.00068945 1.70405243 + layer.4.v_cache 0.00005221 0.13719456 + layer.4.output 0.00941004 197.91248734 + ------------------------------------------------------------------------------------- + TOTAL 0.04594792 115.28882321 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 21292 +BPFP 0.0694 bits/point +EBPFP 0.1388 equivalent bits/point +MSE 115.288823 +---------------------- -------------------------------------------------------- +Time: 0.626s Load: 0.012s, Pack+Encode: 0.249s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 115.2888 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,584B, BPFP=0.0884 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,840B, BPFP=0.1027 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,888B, BPFP=0.1054 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,620B, BPFP=0.0904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,760B, BPFP=0.0982 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,684B, BPFP=0.0940 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,704B, BPFP=0.0951 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,668B, BPFP=0.0931 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.1107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,876B, BPFP=0.1047 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,288B, BPFP=0.0342 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14340984 510.59101563 + layer.0.v_cache 0.00001397 0.04668994 + layer.1.k_cache 0.51622396 34.91778390 + layer.1.v_cache 0.00000557 0.01692387 + layer.2.k_cache 0.00797994 6.24865025 + layer.2.v_cache 0.00001813 0.05266110 + layer.3.k_cache 0.02756977 22.47939104 + layer.3.v_cache 0.00001890 0.06466500 + layer.4.k_cache 0.00068956 1.66444571 + layer.4.v_cache 0.00005177 0.13575850 + layer.4.output 0.00947078 199.32951212 + ------------------------------------------------------------------------------------- + TOTAL 0.04483982 115.97203352 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 21896 +BPFP 0.0719 bits/point +EBPFP 0.1437 equivalent bits/point +MSE 115.972034 +---------------------- -------------------------------------------------------- +Time: 0.620s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 115.9720 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,564B, BPFP=0.0831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,800B, BPFP=0.0957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,896B, BPFP=0.1008 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,628B, BPFP=0.0865 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,776B, BPFP=0.0944 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,676B, BPFP=0.0891 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,716B, BPFP=0.0912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,660B, BPFP=0.0882 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,992B, BPFP=0.1059 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,876B, BPFP=0.0997 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,264B, BPFP=0.0324 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13622742 502.82599915 + layer.0.v_cache 0.00001380 0.04639258 + layer.1.k_cache 0.61520739 35.10108418 + layer.1.v_cache 0.00000584 0.01706064 + layer.2.k_cache 0.00900942 6.28266128 + layer.2.v_cache 0.00001916 0.05302589 + layer.3.k_cache 0.02667915 22.59096314 + layer.3.v_cache 0.00001881 0.06265440 + layer.4.k_cache 0.00071466 1.66927706 + layer.4.v_cache 0.00005760 0.13610515 + layer.4.output 0.04972813 185.10188897 + ------------------------------------------------------------------------------------- + TOTAL 0.06682648 109.67637919 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 21848 +BPFP 0.0683 bits/point +EBPFP 0.1366 equivalent bits/point +MSE 109.676379 +---------------------- -------------------------------------------------------- +Time: 0.621s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 109.6764 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,616B, BPFP=0.0844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,856B, BPFP=0.0970 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,952B, BPFP=0.1020 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,660B, BPFP=0.0867 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,840B, BPFP=0.0962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,704B, BPFP=0.0890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,756B, BPFP=0.0918 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,680B, BPFP=0.0878 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,036B, BPFP=0.1064 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,928B, BPFP=0.1008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,208B, BPFP=0.0314 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11478914 507.92286789 + layer.0.v_cache 0.00001377 0.04654068 + layer.1.k_cache 0.64218915 35.25207397 + layer.1.v_cache 0.00000556 0.01649915 + layer.2.k_cache 0.00898698 6.23419945 + layer.2.v_cache 0.00001980 0.05288402 + layer.3.k_cache 0.05442990 22.71296248 + layer.3.v_cache 0.00001964 0.06505328 + layer.4.k_cache 0.00070461 1.69358249 + layer.4.v_cache 0.00005015 0.13449229 + layer.4.output 0.05017195 182.09597468 + ------------------------------------------------------------------------------------- + TOTAL 0.06896543 108.75311638 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 22236 +BPFP 0.0684 bits/point +EBPFP 0.1367 equivalent bits/point +MSE 108.753116 +---------------------- -------------------------------------------------------- +Time: 0.619s Load: 0.010s, Pack+Encode: 0.246s, Decode+Unpack: 0.363s +---------------------- -------------------------------------------------------- +💾 Converting with 108.7531 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,580B, BPFP=0.0932 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,820B, BPFP=0.1073 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,672B, BPFP=0.0986 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,784B, BPFP=0.1052 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,668B, BPFP=0.0983 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,724B, BPFP=0.1017 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,676B, BPFP=0.0988 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,020B, BPFP=0.1191 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,968B, BPFP=0.1160 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,536B, BPFP=0.0382 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14755792 503.74457547 + layer.0.v_cache 0.00001375 0.04624894 + layer.1.k_cache 0.51940866 35.19246389 + layer.1.v_cache 0.00000586 0.01699763 + layer.2.k_cache 0.00816260 6.28565236 + layer.2.v_cache 0.00001900 0.05212478 + layer.3.k_cache 0.02921419 22.58165905 + layer.3.v_cache 0.00001979 0.06279579 + layer.4.k_cache 0.00071017 1.71546884 + layer.4.v_cache 0.00005507 0.13558419 + layer.4.output 0.00937099 209.85053908 + ------------------------------------------------------------------------------------- + TOTAL 0.04533906 119.92866733 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 22356 +BPFP 0.0775 bits/point +EBPFP 0.1551 equivalent bits/point +MSE 119.928667 +---------------------- -------------------------------------------------------- +Time: 0.621s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 119.9287 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,564B, BPFP=0.0915 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,836B, BPFP=0.1074 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,920B, BPFP=0.1124 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,660B, BPFP=0.0971 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,780B, BPFP=0.1042 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,680B, BPFP=0.0983 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,732B, BPFP=0.1014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,640B, BPFP=0.0960 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,044B, BPFP=0.1196 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,004B, BPFP=0.1173 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,396B, BPFP=0.0368 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12840328 504.11230103 + layer.0.v_cache 0.00001372 0.04542844 + layer.1.k_cache 0.60081110 35.23880794 + layer.1.v_cache 0.00000562 0.01647234 + layer.2.k_cache 0.00586274 6.28114850 + layer.2.v_cache 0.00001839 0.05128835 + layer.3.k_cache 0.02491910 22.58441962 + layer.3.v_cache 0.00001957 0.06152543 + layer.4.k_cache 0.00072047 1.70077766 + layer.4.v_cache 0.00005229 0.13532937 + layer.4.output 0.01050496 208.41843900 + ------------------------------------------------------------------------------------- + TOTAL 0.04908006 119.36215128 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 22256 +BPFP 0.0766 bits/point +EBPFP 0.1532 equivalent bits/point +MSE 119.362151 +---------------------- -------------------------------------------------------- +Time: 0.621s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.363s +---------------------- -------------------------------------------------------- +💾 Converting with 119.3622 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,484B, BPFP=0.0819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,760B, BPFP=0.0972 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,844B, BPFP=0.1018 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,552B, BPFP=0.0857 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,732B, BPFP=0.0956 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,608B, BPFP=0.0888 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,652B, BPFP=0.0912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,592B, BPFP=0.0879 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,948B, BPFP=0.1076 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,824B, BPFP=0.1007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,044B, BPFP=0.0319 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14595435 501.49122129 + layer.0.v_cache 0.00001364 0.04596416 + layer.1.k_cache 0.59759904 34.77637409 + layer.1.v_cache 0.00000566 0.01648309 + layer.2.k_cache 0.01409306 6.23344243 + layer.2.v_cache 0.00002047 0.05224987 + layer.3.k_cache 0.06079736 22.54116401 + layer.3.v_cache 0.00002215 0.06290696 + layer.4.k_cache 0.00070776 1.65490130 + layer.4.v_cache 0.00005221 0.13652771 + layer.4.output 0.00769303 197.02697501 + ------------------------------------------------------------------------------------- + TOTAL 0.05135982 114.48235647 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 21040 +BPFP 0.0683 bits/point +EBPFP 0.1367 equivalent bits/point +MSE 114.482356 +---------------------- -------------------------------------------------------- +Time: 0.622s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.363s +---------------------- -------------------------------------------------------- +💾 Converting with 114.4824 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 315, 128) +Output shape: (1, 315, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.output: torch.Size([1, 315, 3584]) -> torch.Size([1, 1, 315, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,536B, BPFP=0.0762 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,736B, BPFP=0.0861 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,844B, BPFP=0.0915 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,548B, BPFP=0.0768 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,716B, BPFP=0.0851 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,628B, BPFP=0.0808 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,656B, BPFP=0.0821 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,636B, BPFP=0.0812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,912B, BPFP=0.0948 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,852B, BPFP=0.0919 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,904B, BPFP=0.0277 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15500355 502.88115079 + layer.0.v_cache 0.00001366 0.04711573 + layer.1.k_cache 0.68167124 34.71525608 + layer.1.v_cache 0.00000571 0.01719576 + layer.2.k_cache 0.01007247 6.26132037 + layer.2.v_cache 0.00001936 0.05337295 + layer.3.k_cache 0.03272899 22.34131634 + layer.3.v_cache 0.00002006 0.06377844 + layer.4.k_cache 0.00071175 1.70019357 + layer.4.v_cache 0.00005427 0.13597928 + layer.4.output 0.04584261 172.57220805 + ------------------------------------------------------------------------------------- + TOTAL 0.07065879 104.48365504 + (elements=2,741,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2741760 +Total Bytes 20968 +BPFP 0.0612 bits/point +EBPFP 0.1224 equivalent bits/point +MSE 104.483655 +---------------------- -------------------------------------------------------- +Time: 0.620s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.363s +---------------------- -------------------------------------------------------- +💾 Converting with 104.4837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,524B, BPFP=0.0879 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,788B, BPFP=0.1031 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,928B, BPFP=0.1112 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,648B, BPFP=0.0950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.1022 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,664B, BPFP=0.0959 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,728B, BPFP=0.0996 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,644B, BPFP=0.0948 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.1144 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,888B, BPFP=0.1089 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,292B, BPFP=0.0354 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12889319 505.15503921 + layer.0.v_cache 0.00001350 0.04588028 + layer.1.k_cache 0.58775566 35.16133101 + layer.1.v_cache 0.00000553 0.01652347 + layer.2.k_cache 0.00723753 6.33276908 + layer.2.v_cache 0.00001924 0.05154629 + layer.3.k_cache 0.01761036 22.58333213 + layer.3.v_cache 0.00001847 0.06099535 + layer.4.k_cache 0.00071814 1.65996436 + layer.4.v_cache 0.00005270 0.13287641 + layer.4.output 0.01035140 205.24878097 + ------------------------------------------------------------------------------------- + TOTAL 0.04792848 118.11421908 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 21860 +BPFP 0.0741 bits/point +EBPFP 0.1483 equivalent bits/point +MSE 118.114219 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.011s, Pack+Encode: 0.253s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 118.1142 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,484B, BPFP=0.0816 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,740B, BPFP=0.0957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,824B, BPFP=0.1004 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,516B, BPFP=0.0834 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,708B, BPFP=0.0940 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,556B, BPFP=0.0856 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,644B, BPFP=0.0904 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,556B, BPFP=0.0856 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,924B, BPFP=0.1059 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,784B, BPFP=0.0982 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,880B, BPFP=0.0305 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13033964 506.37169894 + layer.0.v_cache 0.00001368 0.04594658 + layer.1.k_cache 0.63029007 34.90067603 + layer.1.v_cache 0.00000551 0.01648561 + layer.2.k_cache 0.00637632 6.18905812 + layer.2.v_cache 0.00001960 0.05153922 + layer.3.k_cache 0.02368569 22.54364443 + layer.3.v_cache 0.00001974 0.06086758 + layer.4.k_cache 0.00068050 1.61872380 + layer.4.v_cache 0.00005209 0.13440473 + layer.4.output 0.00765416 195.80830294 + ------------------------------------------------------------------------------------- + TOTAL 0.04970952 114.27006857 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 20616 +BPFP 0.0667 bits/point +EBPFP 0.1334 equivalent bits/point +MSE 114.270069 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.369s +---------------------- -------------------------------------------------------- +💾 Converting with 114.2701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,528B, BPFP=0.0878 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,808B, BPFP=0.1039 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,924B, BPFP=0.1105 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,628B, BPFP=0.0935 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,768B, BPFP=0.1016 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,680B, BPFP=0.0965 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,736B, BPFP=0.0997 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,672B, BPFP=0.0960 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,004B, BPFP=0.1151 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,936B, BPFP=0.1112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,296B, BPFP=0.0353 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15644295 507.05738741 + layer.0.v_cache 0.00001405 0.04549739 + layer.1.k_cache 0.56163917 35.26103659 + layer.1.v_cache 0.00000554 0.01672704 + layer.2.k_cache 0.01059573 6.29744451 + layer.2.v_cache 0.00001807 0.05201495 + layer.3.k_cache 0.04944291 22.65815107 + layer.3.v_cache 0.00001950 0.06441700 + layer.4.k_cache 0.00071267 1.68150733 + layer.4.v_cache 0.00005817 0.13687726 + layer.4.output 0.00956290 204.36843487 + ------------------------------------------------------------------------------------- + TOTAL 0.04975818 117.87353557 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 21980 +BPFP 0.0743 bits/point +EBPFP 0.1485 equivalent bits/point +MSE 117.873536 +---------------------- -------------------------------------------------------- +Time: 0.623s Load: 0.011s, Pack+Encode: 0.250s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 117.8735 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 312, 128) +Output shape: (1, 312, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.output: torch.Size([1, 312, 3584]) -> torch.Size([1, 1, 312, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,604B, BPFP=0.0803 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,840B, BPFP=0.0921 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,920B, BPFP=0.0962 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,660B, BPFP=0.0831 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,808B, BPFP=0.0905 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,720B, BPFP=0.0861 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,740B, BPFP=0.0871 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,720B, BPFP=0.0861 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,004B, BPFP=0.1004 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,924B, BPFP=0.0964 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,008B, BPFP=0.0287 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14984696 509.24824720 + layer.0.v_cache 0.00001474 0.04811291 + layer.1.k_cache 0.67859346 35.29409555 + layer.1.v_cache 0.00000581 0.01755926 + layer.2.k_cache 0.01626646 6.36140012 + layer.2.v_cache 0.00001951 0.05507907 + layer.3.k_cache 0.02914289 22.75564340 + layer.3.v_cache 0.00001919 0.06581217 + layer.4.k_cache 0.00071665 1.73191090 + layer.4.v_cache 0.00005096 0.13628051 + layer.4.output 0.04642455 181.59117445 + ------------------------------------------------------------------------------------- + TOTAL 0.07056756 108.63837425 + (elements=2,715,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2715648 +Total Bytes 21948 +BPFP 0.0647 bits/point +EBPFP 0.1293 equivalent bits/point +MSE 108.638374 +---------------------- -------------------------------------------------------- +Time: 0.623s Load: 0.012s, Pack+Encode: 0.247s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 108.6384 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,544B, BPFP=0.0897 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,768B, BPFP=0.1027 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1108 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,636B, BPFP=0.0950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,788B, BPFP=0.1039 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,676B, BPFP=0.0974 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,728B, BPFP=0.1004 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,672B, BPFP=0.0971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,028B, BPFP=0.1178 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,924B, BPFP=0.1118 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,344B, BPFP=0.0360 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15823836 509.66658922 + layer.0.v_cache 0.00001502 0.04699909 + layer.1.k_cache 0.53668020 35.18909009 + layer.1.v_cache 0.00000576 0.01697564 + layer.2.k_cache 0.01018995 6.31755435 + layer.2.v_cache 0.00002218 0.05377468 + layer.3.k_cache 0.05410197 22.58984738 + layer.3.v_cache 0.00002001 0.06381247 + layer.4.k_cache 0.00069143 1.67862608 + layer.4.v_cache 0.00005304 0.13988472 + layer.4.output 0.01041225 206.75097916 + ------------------------------------------------------------------------------------- + TOTAL 0.04899433 119.00117693 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 22016 +BPFP 0.0752 bits/point +EBPFP 0.1504 equivalent bits/point +MSE 119.001177 +---------------------- -------------------------------------------------------- +Time: 0.649s Load: 0.010s, Pack+Encode: 0.260s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 119.0012 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 329, 128) +Output shape: (1, 329, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.output: torch.Size([1, 329, 3584]) -> torch.Size([1, 1, 329, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.0878 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,140B, BPFP=0.1016 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,260B, BPFP=0.1073 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,936B, BPFP=0.0919 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,104B, BPFP=0.0999 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,996B, BPFP=0.0948 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,044B, BPFP=0.0971 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,004B, BPFP=0.0952 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,404B, BPFP=0.1142 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,260B, BPFP=0.1073 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,348B, BPFP=0.0363 +⌛️ [2/4] FRONTEND: Frontend time: 0.351s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11391156 504.05357143 + layer.0.v_cache 0.00001522 0.04788050 + layer.1.k_cache 0.63933222 34.96145398 + layer.1.v_cache 0.00000652 0.01810239 + layer.2.k_cache 0.01643764 6.15867471 + layer.2.v_cache 0.00001873 0.05360665 + layer.3.k_cache 0.02826017 22.61366713 + layer.3.v_cache 0.00001876 0.06294791 + layer.4.k_cache 0.00074607 1.69670481 + layer.4.v_cache 0.00005483 0.13846020 + layer.4.output 3.44632065 163.65501520 + ------------------------------------------------------------------------------------- + TOTAL 1.46606155 100.90530448 + (elements=2,863,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2863616 +Total Bytes 26344 +BPFP 0.0736 bits/point +EBPFP 0.1472 equivalent bits/point +MSE 100.905304 +---------------------- -------------------------------------------------------- +Time: 0.802s Load: 0.012s, Pack+Encode: 0.351s, Decode+Unpack: 0.439s +---------------------- -------------------------------------------------------- +💾 Converting with 100.9053 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,572B, BPFP=0.0880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,852B, BPFP=0.1037 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,892B, BPFP=0.1060 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,612B, BPFP=0.0903 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.0992 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,676B, BPFP=0.0939 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,700B, BPFP=0.0952 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,640B, BPFP=0.0918 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,972B, BPFP=0.1104 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,956B, BPFP=0.1095 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,276B, BPFP=0.0342 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13536828 509.14527330 + layer.0.v_cache 0.00001433 0.04621883 + layer.1.k_cache 0.57185238 35.06020385 + layer.1.v_cache 0.00000558 0.01660113 + layer.2.k_cache 0.00766427 6.30953706 + layer.2.v_cache 0.00001891 0.05254312 + layer.3.k_cache 0.04143807 22.45092511 + layer.3.v_cache 0.00001947 0.06258318 + layer.4.k_cache 0.00070404 1.66018195 + layer.4.v_cache 0.00005498 0.13601397 + layer.4.output 0.00948056 200.65037762 + ------------------------------------------------------------------------------------- + TOTAL 0.04844142 116.44074852 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 21920 +BPFP 0.0722 bits/point +EBPFP 0.1444 equivalent bits/point +MSE 116.440749 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.011s, Pack+Encode: 0.257s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 116.4407 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,548B, BPFP=0.0876 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,836B, BPFP=0.1039 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,920B, BPFP=0.1087 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,668B, BPFP=0.0944 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,792B, BPFP=0.1014 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,724B, BPFP=0.0976 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,724B, BPFP=0.0976 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,648B, BPFP=0.0933 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,012B, BPFP=0.1139 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,956B, BPFP=0.1107 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,184B, BPFP=0.0338 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15132187 504.60546875 + layer.0.v_cache 0.00001456 0.04515407 + layer.1.k_cache 0.61798344 34.92556188 + layer.1.v_cache 0.00000569 0.01639857 + layer.2.k_cache 0.01686249 6.19967076 + layer.2.v_cache 0.00001898 0.05250701 + layer.3.k_cache 0.04401700 22.37434542 + layer.3.v_cache 0.00001935 0.06074360 + layer.4.k_cache 0.00071189 1.64866804 + layer.4.v_cache 0.00005111 0.13170397 + layer.4.output 0.01058244 201.81609084 + ------------------------------------------------------------------------------------- + TOTAL 0.05324020 116.63369753 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 22012 +BPFP 0.0733 bits/point +EBPFP 0.1466 equivalent bits/point +MSE 116.633698 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 116.6337 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,540B, BPFP=0.0872 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,816B, BPFP=0.1028 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,912B, BPFP=0.1082 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,616B, BPFP=0.0915 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,776B, BPFP=0.1005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,652B, BPFP=0.0935 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,716B, BPFP=0.0971 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,652B, BPFP=0.0935 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,964B, BPFP=0.1112 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,860B, BPFP=0.1053 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,096B, BPFP=0.0331 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12581717 503.91677989 + layer.0.v_cache 0.00001375 0.04706148 + layer.1.k_cache 0.57814745 34.96144701 + layer.1.v_cache 0.00000592 0.01701336 + layer.2.k_cache 0.00733882 6.24828792 + layer.2.v_cache 0.00002058 0.05360037 + layer.3.k_cache 0.02353939 22.45359205 + layer.3.v_cache 0.00002077 0.06260631 + layer.4.k_cache 0.00069182 1.67827319 + layer.4.v_cache 0.00005027 0.13391879 + layer.4.output 0.01015823 201.76455745 + ------------------------------------------------------------------------------------- + TOTAL 0.04745609 116.58379309 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 21600 +BPFP 0.0719 bits/point +EBPFP 0.1439 equivalent bits/point +MSE 116.583793 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.010s, Pack+Encode: 0.262s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 116.5838 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,528B, BPFP=0.0871 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,772B, BPFP=0.1010 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,920B, BPFP=0.1095 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,624B, BPFP=0.0926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.1010 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,684B, BPFP=0.0960 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,732B, BPFP=0.0988 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,656B, BPFP=0.0944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,000B, BPFP=0.1141 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,888B, BPFP=0.1077 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,192B, BPFP=0.0342 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13625057 503.42558166 + layer.0.v_cache 0.00001362 0.04587037 + layer.1.k_cache 0.57512581 34.92359646 + layer.1.v_cache 0.00000584 0.01676742 + layer.2.k_cache 0.00872732 6.23130676 + layer.2.v_cache 0.00001851 0.05271182 + layer.3.k_cache 0.04362779 22.40560490 + layer.3.v_cache 0.00001925 0.06440382 + layer.4.k_cache 0.00073619 1.69012641 + layer.4.v_cache 0.00005148 0.13695981 + layer.4.output 0.00783937 203.08004758 + ------------------------------------------------------------------------------------- + TOTAL 0.04820306 117.09136838 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 21768 +BPFP 0.0730 bits/point +EBPFP 0.1460 equivalent bits/point +MSE 117.091368 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 117.0914 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,620B, BPFP=0.0833 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,872B, BPFP=0.0962 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,936B, BPFP=0.0995 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,680B, BPFP=0.0863 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,836B, BPFP=0.0944 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,740B, BPFP=0.0894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,760B, BPFP=0.0905 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,764B, BPFP=0.0907 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,044B, BPFP=0.1051 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,944B, BPFP=0.0999 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,320B, BPFP=0.0317 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16730670 509.30309416 + layer.0.v_cache 0.00001388 0.04678468 + layer.1.k_cache 0.64158324 35.13974160 + layer.1.v_cache 0.00000571 0.01682293 + layer.2.k_cache 0.02207795 6.24677558 + layer.2.v_cache 0.00001922 0.05250858 + layer.3.k_cache 0.04503629 22.58076879 + layer.3.v_cache 0.00001955 0.06493873 + layer.4.k_cache 0.00070478 1.66041083 + layer.4.v_cache 0.00005260 0.13599383 + layer.4.output 0.04763387 179.23732672 + ------------------------------------------------------------------------------------- + TOTAL 0.07119159 107.64171334 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 22516 +BPFP 0.0681 bits/point +EBPFP 0.1362 equivalent bits/point +MSE 107.641713 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.011s, Pack+Encode: 0.258s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 107.6417 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,528B, BPFP=0.0862 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,820B, BPFP=0.1027 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1076 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,656B, BPFP=0.0934 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,784B, BPFP=0.1006 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,748B, BPFP=0.0986 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,728B, BPFP=0.0975 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,720B, BPFP=0.0970 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,012B, BPFP=0.1135 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,940B, BPFP=0.1094 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,236B, BPFP=0.0341 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15225300 503.50101534 + layer.0.v_cache 0.00001427 0.04522282 + layer.1.k_cache 0.52731114 34.96158972 + layer.1.v_cache 0.00000556 0.01654243 + layer.2.k_cache 0.01483519 6.20505212 + layer.2.v_cache 0.00001864 0.05259052 + layer.3.k_cache 0.04699480 22.52870283 + layer.3.v_cache 0.00001968 0.06354755 + layer.4.k_cache 0.00070142 1.66262443 + layer.4.v_cache 0.00005177 0.13413737 + layer.4.output 0.01077793 201.45213383 + ------------------------------------------------------------------------------------- + TOTAL 0.04809712 116.43152717 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 22080 +BPFP 0.0733 bits/point +EBPFP 0.1465 equivalent bits/point +MSE 116.431527 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 116.4315 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,536B, BPFP=0.0866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,820B, BPFP=0.1027 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,912B, BPFP=0.1079 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,660B, BPFP=0.0936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,792B, BPFP=0.1011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,700B, BPFP=0.0959 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,728B, BPFP=0.0975 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,680B, BPFP=0.0948 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,032B, BPFP=0.1146 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,940B, BPFP=0.1094 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,228B, BPFP=0.0341 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15700974 507.03361913 + layer.0.v_cache 0.00001374 0.04513366 + layer.1.k_cache 0.52255304 34.93142910 + layer.1.v_cache 0.00000566 0.01646661 + layer.2.k_cache 0.01360668 6.22386109 + layer.2.v_cache 0.00001926 0.05381404 + layer.3.k_cache 0.08085050 22.52248385 + layer.3.v_cache 0.00001943 0.06346578 + layer.4.k_cache 0.00071029 1.67835927 + layer.4.v_cache 0.00005450 0.13648165 + layer.4.output 0.01014708 201.69973247 + ------------------------------------------------------------------------------------- + TOTAL 0.04975720 116.74136714 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 22028 +BPFP 0.0731 bits/point +EBPFP 0.1462 equivalent bits/point +MSE 116.741367 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 116.7414 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,480B, BPFP=0.0811 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,732B, BPFP=0.0950 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,824B, BPFP=0.1000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,516B, BPFP=0.0831 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,704B, BPFP=0.0934 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,572B, BPFP=0.0862 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,640B, BPFP=0.0899 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,572B, BPFP=0.0862 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,916B, BPFP=0.1050 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,816B, BPFP=0.0996 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,860B, BPFP=0.0302 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16436556 509.21705044 + layer.0.v_cache 0.00001441 0.04616118 + layer.1.k_cache 0.62741763 34.95562637 + layer.1.v_cache 0.00000560 0.01677961 + layer.2.k_cache 0.01487184 6.22593673 + layer.2.v_cache 0.00002167 0.05327278 + layer.3.k_cache 0.02142572 22.56089638 + layer.3.v_cache 0.00001943 0.06275668 + layer.4.k_cache 0.00070776 1.64903179 + layer.4.v_cache 0.00005251 0.13406806 + layer.4.output 0.00894435 195.04260652 + ------------------------------------------------------------------------------------- + TOTAL 0.05244192 114.13057798 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 20632 +BPFP 0.0665 bits/point +EBPFP 0.1331 equivalent bits/point +MSE 114.130578 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.011s, Pack+Encode: 0.257s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 114.1306 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,568B, BPFP=0.0872 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,824B, BPFP=0.1014 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,864B, BPFP=0.1036 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,544B, BPFP=0.0859 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,748B, BPFP=0.0972 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,608B, BPFP=0.0894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,664B, BPFP=0.0925 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,596B, BPFP=0.0887 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,936B, BPFP=0.1077 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,852B, BPFP=0.1030 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,092B, BPFP=0.0325 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11705062 501.75767349 + layer.0.v_cache 0.00001445 0.04641192 + layer.1.k_cache 0.56620099 34.89913604 + layer.1.v_cache 0.00000566 0.01649884 + layer.2.k_cache 0.01299965 6.26326570 + layer.2.v_cache 0.00001972 0.05147183 + layer.3.k_cache 0.05069733 22.48411088 + layer.3.v_cache 0.00001944 0.06229401 + layer.4.k_cache 0.00068883 1.64737492 + layer.4.v_cache 0.00005144 0.13226707 + layer.4.output 0.00838925 198.71846403 + ------------------------------------------------------------------------------------- + TOTAL 0.04743958 115.19939723 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 21296 +BPFP 0.0697 bits/point +EBPFP 0.1393 equivalent bits/point +MSE 115.199397 +---------------------- -------------------------------------------------------- +Time: 0.642s Load: 0.009s, Pack+Encode: 0.256s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 115.1994 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,500B, BPFP=0.0828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,760B, BPFP=0.0972 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,836B, BPFP=0.1014 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,536B, BPFP=0.0848 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,716B, BPFP=0.0947 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,588B, BPFP=0.0877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,664B, BPFP=0.0919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,584B, BPFP=0.0875 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,916B, BPFP=0.1058 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,820B, BPFP=0.1005 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,020B, BPFP=0.0317 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14370025 507.10865724 + layer.0.v_cache 0.00001382 0.04598617 + layer.1.k_cache 0.61448195 34.77146367 + layer.1.v_cache 0.00000570 0.01662654 + layer.2.k_cache 0.01208573 6.22828960 + layer.2.v_cache 0.00001823 0.05167797 + layer.3.k_cache 0.01319265 22.52580644 + layer.3.v_cache 0.00002242 0.06155296 + layer.4.k_cache 0.00071920 1.64462431 + layer.4.v_cache 0.00005145 0.13468750 + layer.4.output 0.00992508 196.97267794 + ------------------------------------------------------------------------------------- + TOTAL 0.05022159 114.78812459 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 20940 +BPFP 0.0680 bits/point +EBPFP 0.1360 equivalent bits/point +MSE 114.788125 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.011s, Pack+Encode: 0.258s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 114.7881 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 259, 128) +Output shape: (1, 259, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.output: torch.Size([1, 259, 3584]) -> torch.Size([1, 1, 259, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,532B, BPFP=0.0924 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,680B, BPFP=0.1014 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,844B, BPFP=0.1112 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,532B, BPFP=0.0924 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,720B, BPFP=0.1038 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,580B, BPFP=0.0953 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,640B, BPFP=0.0989 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,572B, BPFP=0.0948 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,912B, BPFP=0.1153 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,812B, BPFP=0.1093 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,156B, BPFP=0.0358 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13349269 505.34619329 + layer.0.v_cache 0.00001436 0.04673017 + layer.1.k_cache 0.50562990 35.27107716 + layer.1.v_cache 0.00000590 0.01715840 + layer.2.k_cache 0.00820730 6.29326756 + layer.2.v_cache 0.00001794 0.05148429 + layer.3.k_cache 0.03685973 22.76938609 + layer.3.v_cache 0.00002061 0.06144066 + layer.4.k_cache 0.00069539 1.67867182 + layer.4.v_cache 0.00005422 0.13498158 + layer.4.output 0.01018517 216.89802813 + ------------------------------------------------------------------------------------- + TOTAL 0.04448790 122.93862282 + (elements=2,254,336) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2254336 +Total Bytes 20980 +BPFP 0.0745 bits/point +EBPFP 0.1489 equivalent bits/point +MSE 122.938623 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 122.9386 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,556B, BPFP=0.0900 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,768B, BPFP=0.1023 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,916B, BPFP=0.1109 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,620B, BPFP=0.0938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,776B, BPFP=0.1028 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,696B, BPFP=0.0981 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,728B, BPFP=0.1000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,636B, BPFP=0.0947 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,024B, BPFP=0.1171 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,948B, BPFP=0.1127 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,424B, BPFP=0.0366 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12649573 507.15769676 + layer.0.v_cache 0.00001373 0.04635990 + layer.1.k_cache 0.51574391 35.19003183 + layer.1.v_cache 0.00000589 0.01695352 + layer.2.k_cache 0.01194659 6.31944987 + layer.2.v_cache 0.00001984 0.05415416 + layer.3.k_cache 0.04962628 22.67850297 + layer.3.v_cache 0.00002059 0.06358932 + layer.4.k_cache 0.00068541 1.69755972 + layer.4.v_cache 0.00005704 0.14090527 + layer.4.output 0.00772497 206.06701389 + ------------------------------------------------------------------------------------- + TOTAL 0.04462881 118.57848827 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 22092 +BPFP 0.0752 bits/point +EBPFP 0.1504 equivalent bits/point +MSE 118.578488 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.009s, Pack+Encode: 0.257s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 118.5785 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,552B, BPFP=0.0901 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,792B, BPFP=0.1041 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,916B, BPFP=0.1113 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,644B, BPFP=0.0955 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,784B, BPFP=0.1036 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,688B, BPFP=0.0980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,724B, BPFP=0.1001 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,640B, BPFP=0.0953 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.1166 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,936B, BPFP=0.1125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,332B, BPFP=0.0359 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12210708 504.23733736 + layer.0.v_cache 0.00001342 0.04577702 + layer.1.k_cache 0.54942129 35.21869554 + layer.1.v_cache 0.00000547 0.01650178 + layer.2.k_cache 0.01053270 6.30056842 + layer.2.v_cache 0.00001820 0.05210106 + layer.3.k_cache 0.02850239 22.48828851 + layer.3.v_cache 0.00002491 0.06310702 + layer.4.k_cache 0.00067987 1.69317525 + layer.4.v_cache 0.00005179 0.13415932 + layer.4.output 0.01126663 206.72747942 + ------------------------------------------------------------------------------------- + TOTAL 0.04648373 118.66718042 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 22016 +BPFP 0.0752 bits/point +EBPFP 0.1504 equivalent bits/point +MSE 118.667180 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.011s, Pack+Encode: 0.261s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 118.6672 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,596B, BPFP=0.0823 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,888B, BPFP=0.0974 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,944B, BPFP=0.1002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,708B, BPFP=0.0881 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,856B, BPFP=0.0957 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,756B, BPFP=0.0906 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,764B, BPFP=0.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,776B, BPFP=0.0916 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,044B, BPFP=0.1054 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,964B, BPFP=0.1013 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,344B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14443027 511.11592409 + layer.0.v_cache 0.00001375 0.04670911 + layer.1.k_cache 0.62837083 35.25268152 + layer.1.v_cache 0.00000570 0.01657198 + layer.2.k_cache 0.01557094 6.29017000 + layer.2.v_cache 0.00001871 0.05251521 + layer.3.k_cache 0.04676386 22.58684638 + layer.3.v_cache 0.00001981 0.06187153 + layer.4.k_cache 0.00070582 1.69528752 + layer.4.v_cache 0.00005437 0.13458061 + layer.4.output 0.04838977 179.72041490 + ------------------------------------------------------------------------------------- + TOTAL 0.06909897 107.95859190 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 22640 +BPFP 0.0687 bits/point +EBPFP 0.1374 equivalent bits/point +MSE 107.958592 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.367s +---------------------- -------------------------------------------------------- +💾 Converting with 107.9586 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,484B, BPFP=0.0819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,744B, BPFP=0.0963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,840B, BPFP=0.1016 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,564B, BPFP=0.0864 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,724B, BPFP=0.0952 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,620B, BPFP=0.0894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,660B, BPFP=0.0917 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,600B, BPFP=0.0883 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,944B, BPFP=0.1073 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,872B, BPFP=0.1034 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,052B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13293209 506.97007509 + layer.0.v_cache 0.00001388 0.04607463 + layer.1.k_cache 0.56267701 34.81476714 + layer.1.v_cache 0.00000566 0.01659635 + layer.2.k_cache 0.01145046 6.26105362 + layer.2.v_cache 0.00001921 0.05274850 + layer.3.k_cache 0.02674564 22.59516481 + layer.3.v_cache 0.00001926 0.06220546 + layer.4.k_cache 0.00069135 1.67836209 + layer.4.v_cache 0.00005493 0.13608122 + layer.4.output 0.00973303 196.98233216 + ------------------------------------------------------------------------------------- + TOTAL 0.04722004 114.79467376 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 21104 +BPFP 0.0685 bits/point +EBPFP 0.1371 equivalent bits/point +MSE 114.794674 +---------------------- -------------------------------------------------------- +Time: 0.621s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 114.7947 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,528B, BPFP=0.0862 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,808B, BPFP=0.1020 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,904B, BPFP=0.1074 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,612B, BPFP=0.0909 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,780B, BPFP=0.1004 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,660B, BPFP=0.0936 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,708B, BPFP=0.0963 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,636B, BPFP=0.0923 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,980B, BPFP=0.1117 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,880B, BPFP=0.1060 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,128B, BPFP=0.0333 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15182159 506.48809792 + layer.0.v_cache 0.00001420 0.04636432 + layer.1.k_cache 0.56224567 34.93212714 + layer.1.v_cache 0.00000588 0.01702473 + layer.2.k_cache 0.00652889 6.21041837 + layer.2.v_cache 0.00001844 0.05135088 + layer.3.k_cache 0.06011557 22.41840415 + layer.3.v_cache 0.00001921 0.06199227 + layer.4.k_cache 0.00068317 1.65113164 + layer.4.v_cache 0.00005096 0.13284850 + layer.4.output 0.01130176 201.30416774 + ------------------------------------------------------------------------------------- + TOTAL 0.05062446 116.53758436 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 21624 +BPFP 0.0718 bits/point +EBPFP 0.1435 equivalent bits/point +MSE 116.537584 +---------------------- -------------------------------------------------------- +Time: 0.618s Load: 0.010s, Pack+Encode: 0.245s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 116.5376 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,468B, BPFP=0.0805 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,740B, BPFP=0.0954 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,832B, BPFP=0.1004 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,548B, BPFP=0.0849 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,716B, BPFP=0.0941 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,596B, BPFP=0.0875 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,660B, BPFP=0.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,612B, BPFP=0.0884 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,924B, BPFP=0.1055 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,824B, BPFP=0.1000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,920B, BPFP=0.0307 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13053865 506.48366228 + layer.0.v_cache 0.00001414 0.04795822 + layer.1.k_cache 0.56363723 34.99023095 + layer.1.v_cache 0.00000590 0.01718502 + layer.2.k_cache 0.01033439 6.23519351 + layer.2.v_cache 0.00002015 0.05314747 + layer.3.k_cache 0.03655195 22.56999383 + layer.3.v_cache 0.00002026 0.06255540 + layer.4.k_cache 0.00068516 1.65776346 + layer.4.v_cache 0.00005378 0.13651575 + layer.4.output 0.00863618 194.99437657 + ------------------------------------------------------------------------------------- + TOTAL 0.04719499 113.95381423 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 20840 +BPFP 0.0672 bits/point +EBPFP 0.1344 equivalent bits/point +MSE 113.953814 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.011s, Pack+Encode: 0.246s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 113.9538 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,528B, BPFP=0.0862 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,784B, BPFP=0.1006 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,904B, BPFP=0.1074 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,612B, BPFP=0.0909 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,792B, BPFP=0.1011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,644B, BPFP=0.0927 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,704B, BPFP=0.0961 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,636B, BPFP=0.0923 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.1119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,876B, BPFP=0.1058 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,164B, BPFP=0.0336 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15908699 507.83444269 + layer.0.v_cache 0.00001433 0.04649304 + layer.1.k_cache 0.53822481 34.97880119 + layer.1.v_cache 0.00000570 0.01696497 + layer.2.k_cache 0.01139215 6.21738079 + layer.2.v_cache 0.00001845 0.05259523 + layer.3.k_cache 0.02907990 22.46182240 + layer.3.v_cache 0.00001963 0.06480635 + layer.4.k_cache 0.00069750 1.66835931 + layer.4.v_cache 0.00005397 0.13851115 + layer.4.output 0.01069233 201.53104048 + ------------------------------------------------------------------------------------- + TOTAL 0.04784940 116.71749768 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 21628 +BPFP 0.0718 bits/point +EBPFP 0.1435 equivalent bits/point +MSE 116.717498 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 116.7175 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,476B, BPFP=0.0806 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,716B, BPFP=0.0938 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,828B, BPFP=0.0999 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,520B, BPFP=0.0830 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,700B, BPFP=0.0929 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,576B, BPFP=0.0861 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,656B, BPFP=0.0905 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,588B, BPFP=0.0868 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,924B, BPFP=0.1051 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,788B, BPFP=0.0977 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,888B, BPFP=0.0303 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16084659 507.45137675 + layer.0.v_cache 0.00001440 0.04662923 + layer.1.k_cache 0.59028764 34.91452005 + layer.1.v_cache 0.00000559 0.01694979 + layer.2.k_cache 0.00929020 6.25353748 + layer.2.v_cache 0.00002205 0.05302515 + layer.3.k_cache 0.04382113 22.50713813 + layer.3.v_cache 0.00002103 0.06353807 + layer.4.k_cache 0.00069271 1.65820377 + layer.4.v_cache 0.00005052 0.13414892 + layer.4.output 0.00984751 194.20217283 + ------------------------------------------------------------------------------------- + TOTAL 0.05141085 113.67731042 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 20660 +BPFP 0.0664 bits/point +EBPFP 0.1328 equivalent bits/point +MSE 113.677310 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.011s, Pack+Encode: 0.253s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 113.6773 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,500B, BPFP=0.0814 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,756B, BPFP=0.0953 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,852B, BPFP=0.1005 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,576B, BPFP=0.0855 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,728B, BPFP=0.0938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,648B, BPFP=0.0894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,672B, BPFP=0.0907 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,652B, BPFP=0.0896 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,980B, BPFP=0.1074 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,844B, BPFP=0.1000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,048B, BPFP=0.0314 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14211286 508.51443142 + layer.0.v_cache 0.00001516 0.04642885 + layer.1.k_cache 0.60291523 35.06779650 + layer.1.v_cache 0.00000570 0.01709069 + layer.2.k_cache 0.01041983 6.22301483 + layer.2.v_cache 0.00001835 0.05219238 + layer.3.k_cache 0.03896446 22.55389404 + layer.3.v_cache 0.00001926 0.06317221 + layer.4.k_cache 0.00071470 1.68203100 + layer.4.v_cache 0.00005236 0.13548855 + layer.4.output 0.00754456 194.42478919 + ------------------------------------------------------------------------------------- + TOTAL 0.04988528 113.84288616 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 21256 +BPFP 0.0678 bits/point +EBPFP 0.1357 equivalent bits/point +MSE 113.842886 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 113.8429 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 296, 128) +Output shape: (1, 296, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.output: torch.Size([1, 296, 3584]) -> torch.Size([1, 1, 296, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,612B, BPFP=0.0851 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,804B, BPFP=0.0952 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,928B, BPFP=0.1018 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,660B, BPFP=0.0876 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,840B, BPFP=0.0971 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,716B, BPFP=0.0906 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,756B, BPFP=0.0927 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,716B, BPFP=0.0906 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,020B, BPFP=0.1066 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,860B, BPFP=0.0982 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,180B, BPFP=0.0315 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14616069 508.10341005 + layer.0.v_cache 0.00001366 0.04611164 + layer.1.k_cache 0.61022733 35.07364139 + layer.1.v_cache 0.00000576 0.01663595 + layer.2.k_cache 0.01378242 6.23829610 + layer.2.v_cache 0.00001907 0.05276121 + layer.3.k_cache 0.03184220 22.50127184 + layer.3.v_cache 0.00001934 0.06357250 + layer.4.k_cache 0.00072724 1.67306539 + layer.4.v_cache 0.00005080 0.13458623 + layer.4.output 0.04784788 183.52119028 + ------------------------------------------------------------------------------------- + TOTAL 0.06692845 109.32656966 + (elements=2,576,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2576384 +Total Bytes 22092 +BPFP 0.0686 bits/point +EBPFP 0.1372 equivalent bits/point +MSE 109.326570 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.254s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 109.3266 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,608B, BPFP=0.0843 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,836B, BPFP=0.0963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,944B, BPFP=0.1019 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,688B, BPFP=0.0885 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,836B, BPFP=0.0963 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,732B, BPFP=0.0908 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,768B, BPFP=0.0927 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,728B, BPFP=0.0906 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,060B, BPFP=0.1080 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,940B, BPFP=0.1017 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,340B, BPFP=0.0325 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15387250 509.74213507 + layer.0.v_cache 0.00001382 0.04680786 + layer.1.k_cache 0.63844893 35.21950045 + layer.1.v_cache 0.00000583 0.01723458 + layer.2.k_cache 0.01105453 6.29725473 + layer.2.v_cache 0.00001885 0.05353693 + layer.3.k_cache 0.03893430 22.61954043 + layer.3.v_cache 0.00001923 0.06562082 + layer.4.k_cache 0.00070257 1.69457096 + layer.4.v_cache 0.00005146 0.13605608 + layer.4.output 0.04873301 182.70986637 + ------------------------------------------------------------------------------------- + TOTAL 0.06966195 109.10948956 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 22480 +BPFP 0.0693 bits/point +EBPFP 0.1387 equivalent bits/point +MSE 109.109490 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 109.1095 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,520B, BPFP=0.0876 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,816B, BPFP=0.1047 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,928B, BPFP=0.1112 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,652B, BPFP=0.0952 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,784B, BPFP=0.1029 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,720B, BPFP=0.0992 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,720B, BPFP=0.0992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,704B, BPFP=0.0982 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,044B, BPFP=0.1179 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,012B, BPFP=0.1160 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,368B, BPFP=0.0360 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12629252 508.40590406 + layer.0.v_cache 0.00001451 0.04757815 + layer.1.k_cache 0.49097243 35.21624193 + layer.1.v_cache 0.00000577 0.01748769 + layer.2.k_cache 0.01060468 6.27922435 + layer.2.v_cache 0.00001909 0.05416760 + layer.3.k_cache 0.02748298 22.54963532 + layer.3.v_cache 0.00001876 0.06411625 + layer.4.k_cache 0.00069967 1.71342530 + layer.4.v_cache 0.00005414 0.13709014 + layer.4.output 0.01047103 205.21413745 + ------------------------------------------------------------------------------------- + TOTAL 0.04290952 118.29316665 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 22268 +BPFP 0.0755 bits/point +EBPFP 0.1510 equivalent bits/point +MSE 118.293167 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 118.2932 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,580B, BPFP=0.0885 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,788B, BPFP=0.1001 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,888B, BPFP=0.1057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,604B, BPFP=0.0898 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.0992 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,632B, BPFP=0.0914 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,692B, BPFP=0.0948 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,640B, BPFP=0.0918 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,964B, BPFP=0.1100 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,888B, BPFP=0.1057 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,300B, BPFP=0.0344 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15022702 508.49238351 + layer.0.v_cache 0.00001395 0.04655899 + layer.1.k_cache 0.54271635 35.03499874 + layer.1.v_cache 0.00000573 0.01685281 + layer.2.k_cache 0.01332666 6.19239769 + layer.2.v_cache 0.00001832 0.05221454 + layer.3.k_cache 0.02579262 22.48477753 + layer.3.v_cache 0.00001838 0.06161562 + layer.4.k_cache 0.00069325 1.67248065 + layer.4.v_cache 0.00005155 0.13219099 + layer.4.output 0.01006869 200.63368856 + ------------------------------------------------------------------------------------- + TOTAL 0.04725557 116.38954653 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 21748 +BPFP 0.0716 bits/point +EBPFP 0.1433 equivalent bits/point +MSE 116.389547 +---------------------- -------------------------------------------------------- +Time: 0.619s Load: 0.011s, Pack+Encode: 0.245s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 116.3895 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,600B, BPFP=0.0847 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,832B, BPFP=0.0970 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,920B, BPFP=0.1017 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,676B, BPFP=0.0888 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,816B, BPFP=0.0962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,740B, BPFP=0.0922 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,744B, BPFP=0.0924 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,748B, BPFP=0.0926 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,016B, BPFP=0.1068 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,908B, BPFP=0.1011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,368B, BPFP=0.0331 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12416320 508.27552966 + layer.0.v_cache 0.00001370 0.04622955 + layer.1.k_cache 0.62725913 35.11079515 + layer.1.v_cache 0.00000565 0.01690012 + layer.2.k_cache 0.01102555 6.23524646 + layer.2.v_cache 0.00002033 0.05289630 + layer.3.k_cache 0.04118381 22.65957693 + layer.3.v_cache 0.00001949 0.06349818 + layer.4.k_cache 0.00070768 1.68435928 + layer.4.v_cache 0.00005397 0.13469406 + layer.4.output 0.05105524 184.43096247 + ------------------------------------------------------------------------------------- + TOTAL 0.06834348 109.72332135 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 22368 +BPFP 0.0697 bits/point +EBPFP 0.1394 equivalent bits/point +MSE 109.723321 +---------------------- -------------------------------------------------------- +Time: 0.622s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.363s +---------------------- -------------------------------------------------------- +💾 Converting with 109.7233 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,484B, BPFP=0.0802 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,756B, BPFP=0.0949 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,848B, BPFP=0.0999 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,564B, BPFP=0.0846 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,720B, BPFP=0.0930 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,636B, BPFP=0.0885 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,676B, BPFP=0.0906 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,624B, BPFP=0.0878 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,936B, BPFP=0.1047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,848B, BPFP=0.0999 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,936B, BPFP=0.0304 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15516999 504.64484213 + layer.0.v_cache 0.00001361 0.04672008 + layer.1.k_cache 0.56212566 35.11470385 + layer.1.v_cache 0.00000565 0.01730615 + layer.2.k_cache 0.00839357 6.31499547 + layer.2.v_cache 0.00001980 0.05400323 + layer.3.k_cache 0.02402391 22.56535535 + layer.3.v_cache 0.00001987 0.06408832 + layer.4.k_cache 0.00071212 1.67167035 + layer.4.v_cache 0.00005188 0.13674878 + layer.4.output 0.05073151 189.45549617 + ------------------------------------------------------------------------------------- + TOTAL 0.06503863 111.57758276 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 21028 +BPFP 0.0669 bits/point +EBPFP 0.1338 equivalent bits/point +MSE 111.577583 +---------------------- -------------------------------------------------------- +Time: 0.622s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 111.5776 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,620B, BPFP=0.0852 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,876B, BPFP=0.0987 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,948B, BPFP=0.1025 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,688B, BPFP=0.0888 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,844B, BPFP=0.0970 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,720B, BPFP=0.0905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,764B, BPFP=0.0928 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,740B, BPFP=0.0915 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,064B, BPFP=0.1086 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,924B, BPFP=0.1012 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,328B, BPFP=0.0325 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005571 508.42666246 + layer.0.v_cache 0.00001459 0.04670806 + layer.1.k_cache 0.64038677 35.17554451 + layer.1.v_cache 0.00000545 0.01662350 + layer.2.k_cache 0.01623841 6.29436783 + layer.2.v_cache 0.00002030 0.05310651 + layer.3.k_cache 0.01774169 22.61112098 + layer.3.v_cache 0.00001948 0.06450471 + layer.4.k_cache 0.00069419 1.67789395 + layer.4.v_cache 0.00005057 0.13546560 + layer.4.output 0.05052170 182.94925445 + ------------------------------------------------------------------------------------- + TOTAL 0.06875759 109.12628113 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 22516 +BPFP 0.0697 bits/point +EBPFP 0.1394 equivalent bits/point +MSE 109.126281 +---------------------- -------------------------------------------------------- +Time: 0.627s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 109.1263 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,572B, BPFP=0.0880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,780B, BPFP=0.0997 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,884B, BPFP=0.1055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,596B, BPFP=0.0894 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,768B, BPFP=0.0990 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,652B, BPFP=0.0925 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,688B, BPFP=0.0945 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,640B, BPFP=0.0918 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,956B, BPFP=0.1095 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,884B, BPFP=0.1055 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,252B, BPFP=0.0340 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14010120 508.36626344 + layer.0.v_cache 0.00001437 0.04606119 + layer.1.k_cache 0.58692467 35.04401882 + layer.1.v_cache 0.00000569 0.01690259 + layer.2.k_cache 0.00806839 6.27741157 + layer.2.v_cache 0.00001934 0.05229504 + layer.3.k_cache 0.03832107 22.40799136 + layer.3.v_cache 0.00001992 0.06249114 + layer.4.k_cache 0.00069804 1.65085615 + layer.4.v_cache 0.00005131 0.13699904 + layer.4.output 0.00759059 200.66231439 + ------------------------------------------------------------------------------------- + TOTAL 0.04866812 116.39397006 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 21672 +BPFP 0.0714 bits/point +EBPFP 0.1428 equivalent bits/point +MSE 116.393970 +---------------------- -------------------------------------------------------- +Time: 0.624s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 116.3940 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,532B, BPFP=0.0893 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,788B, BPFP=0.1042 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,924B, BPFP=0.1122 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,672B, BPFP=0.0975 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,788B, BPFP=0.1042 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,700B, BPFP=0.0991 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,744B, BPFP=0.1017 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,652B, BPFP=0.0963 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.1171 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,960B, BPFP=0.1143 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,360B, BPFP=0.0363 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13199043 505.58092351 + layer.0.v_cache 0.00001386 0.04643862 + layer.1.k_cache 0.56989129 35.27891791 + layer.1.v_cache 0.00000577 0.01713765 + layer.2.k_cache 0.00892691 6.30015359 + layer.2.v_cache 0.00001816 0.05237591 + layer.3.k_cache 0.04626510 22.56627507 + layer.3.v_cache 0.00001927 0.06313997 + layer.4.k_cache 0.00069392 1.67377643 + layer.4.v_cache 0.00005195 0.13586626 + layer.4.output 0.00911696 207.64242404 + ------------------------------------------------------------------------------------- + TOTAL 0.04833502 119.13011607 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 22128 +BPFP 0.0759 bits/point +EBPFP 0.1518 equivalent bits/point +MSE 119.130116 +---------------------- -------------------------------------------------------- +Time: 0.623s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 119.1301 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,544B, BPFP=0.0880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,780B, BPFP=0.1015 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,916B, BPFP=0.1093 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,632B, BPFP=0.0931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.1010 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,680B, BPFP=0.0958 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,736B, BPFP=0.0990 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,684B, BPFP=0.0960 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,012B, BPFP=0.1147 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,920B, BPFP=0.1095 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,196B, BPFP=0.0342 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16500157 505.92877509 + layer.0.v_cache 0.00001391 0.04633109 + layer.1.k_cache 0.57858722 34.97258497 + layer.1.v_cache 0.00000575 0.01680909 + layer.2.k_cache 0.01018238 6.23250073 + layer.2.v_cache 0.00001877 0.05280098 + layer.3.k_cache 0.03425542 22.28699889 + layer.3.v_cache 0.00002021 0.06336741 + layer.4.k_cache 0.00069807 1.65767241 + layer.4.v_cache 0.00005194 0.13414059 + layer.4.output 0.01023613 203.10321624 + ------------------------------------------------------------------------------------- + TOTAL 0.05061695 117.24202911 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 21872 +BPFP 0.0734 bits/point +EBPFP 0.1467 equivalent bits/point +MSE 117.242029 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.369s +---------------------- -------------------------------------------------------- +💾 Converting with 117.2420 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,552B, BPFP=0.0898 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,792B, BPFP=0.1037 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,924B, BPFP=0.1113 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,656B, BPFP=0.0958 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,792B, BPFP=0.1037 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,704B, BPFP=0.0986 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,724B, BPFP=0.0998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,728B, BPFP=0.1000 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.1162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,916B, BPFP=0.1109 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,376B, BPFP=0.0362 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385483 505.37777778 + layer.0.v_cache 0.00001387 0.04589300 + layer.1.k_cache 0.55871571 35.22802373 + layer.1.v_cache 0.00000565 0.01634909 + layer.2.k_cache 0.00818202 6.27438241 + layer.2.v_cache 0.00001875 0.05205309 + layer.3.k_cache 0.04697755 22.52820095 + layer.3.v_cache 0.00001978 0.06355029 + layer.4.k_cache 0.00071636 1.66570921 + layer.4.v_cache 0.00005146 0.13486768 + layer.4.output 0.01106052 205.94619709 + ------------------------------------------------------------------------------------- + TOTAL 0.04741057 118.41236393 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 22172 +BPFP 0.0755 bits/point +EBPFP 0.1510 equivalent bits/point +MSE 118.412364 +---------------------- -------------------------------------------------------- +Time: 0.620s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 118.4124 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,540B, BPFP=0.0875 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,788B, BPFP=0.1016 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1084 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,664B, BPFP=0.0945 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.1007 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,692B, BPFP=0.0961 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,720B, BPFP=0.0977 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,680B, BPFP=0.0955 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,988B, BPFP=0.1130 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,888B, BPFP=0.1073 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,144B, BPFP=0.0336 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203608 504.64170455 + layer.0.v_cache 0.00001420 0.04563554 + layer.1.k_cache 0.56673284 34.84248224 + layer.1.v_cache 0.00000552 0.01646671 + layer.2.k_cache 0.00530624 6.24325195 + layer.2.v_cache 0.00001825 0.05193089 + layer.3.k_cache 0.02188354 22.43744141 + layer.3.v_cache 0.00001854 0.06168522 + layer.4.k_cache 0.00069888 1.66357355 + layer.4.v_cache 0.00004918 0.13169269 + layer.4.output 0.00799176 202.35555195 + ------------------------------------------------------------------------------------- + TOTAL 0.04662974 116.86027814 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 21784 +BPFP 0.0728 bits/point +EBPFP 0.1456 equivalent bits/point +MSE 116.860278 +---------------------- -------------------------------------------------------- +Time: 0.619s Load: 0.010s, Pack+Encode: 0.246s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 116.8603 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,492B, BPFP=0.0824 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,772B, BPFP=0.0978 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,840B, BPFP=0.1016 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,532B, BPFP=0.0846 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,720B, BPFP=0.0950 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,612B, BPFP=0.0890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,656B, BPFP=0.0914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,584B, BPFP=0.0875 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,936B, BPFP=0.1069 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,836B, BPFP=0.1014 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,124B, BPFP=0.0325 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12732459 506.48365724 + layer.0.v_cache 0.00001480 0.04631547 + layer.1.k_cache 0.57322790 34.79069125 + layer.1.v_cache 0.00000557 0.01659968 + layer.2.k_cache 0.00950604 6.25732379 + layer.2.v_cache 0.00002039 0.05200728 + layer.3.k_cache 0.01510038 22.68814529 + layer.3.v_cache 0.00001916 0.06281268 + layer.4.k_cache 0.00068141 1.62065507 + layer.4.v_cache 0.00005165 0.13420391 + layer.4.output 0.00936729 196.99186017 + ------------------------------------------------------------------------------------- + TOTAL 0.04656017 114.77031958 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 21104 +BPFP 0.0685 bits/point +EBPFP 0.1371 equivalent bits/point +MSE 114.770320 +---------------------- -------------------------------------------------------- +Time: 0.625s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 114.7703 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,544B, BPFP=0.0874 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,796B, BPFP=0.1017 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,912B, BPFP=0.1082 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,624B, BPFP=0.0919 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,776B, BPFP=0.1005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,684B, BPFP=0.0953 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,724B, BPFP=0.0976 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,660B, BPFP=0.0940 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,028B, BPFP=0.1148 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,972B, BPFP=0.1116 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,152B, BPFP=0.0336 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13630490 506.04019475 + layer.0.v_cache 0.00001544 0.04611255 + layer.1.k_cache 0.58776402 34.98794158 + layer.1.v_cache 0.00000559 0.01640498 + layer.2.k_cache 0.01000112 6.25870724 + layer.2.v_cache 0.00001940 0.05284004 + layer.3.k_cache 0.04424953 22.49684032 + layer.3.v_cache 0.00002139 0.06252154 + layer.4.k_cache 0.00071093 1.67283796 + layer.4.v_cache 0.00005158 0.13751945 + layer.4.output 0.01101471 201.74653856 + ------------------------------------------------------------------------------------- + TOTAL 0.05036746 116.70574649 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 21872 +BPFP 0.0728 bits/point +EBPFP 0.1457 equivalent bits/point +MSE 116.705746 +---------------------- -------------------------------------------------------- +Time: 0.627s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 116.7057 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,568B, BPFP=0.0881 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,844B, BPFP=0.1036 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,908B, BPFP=0.1072 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,636B, BPFP=0.0920 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,784B, BPFP=0.1003 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,692B, BPFP=0.0951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,712B, BPFP=0.0962 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,672B, BPFP=0.0940 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.1115 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,924B, BPFP=0.1081 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,264B, BPFP=0.0342 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12958164 508.29524505 + layer.0.v_cache 0.00001430 0.04625325 + layer.1.k_cache 0.54121377 35.02063076 + layer.1.v_cache 0.00000577 0.01665073 + layer.2.k_cache 0.01080061 6.23691916 + layer.2.v_cache 0.00001903 0.05304278 + layer.3.k_cache 0.01783211 22.43842738 + layer.3.v_cache 0.00002028 0.06150780 + layer.4.k_cache 0.00068935 1.66099735 + layer.4.v_cache 0.00005053 0.13480012 + layer.4.output 0.00833506 201.88203366 + ------------------------------------------------------------------------------------- + TOTAL 0.04462193 116.89051235 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 21988 +BPFP 0.0727 bits/point +EBPFP 0.1454 equivalent bits/point +MSE 116.890512 +---------------------- -------------------------------------------------------- +Time: 0.622s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 116.8905 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,584B, BPFP=0.0941 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,792B, BPFP=0.1065 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,864B, BPFP=0.1107 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,636B, BPFP=0.0972 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,760B, BPFP=0.1046 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,656B, BPFP=0.0984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,708B, BPFP=0.1015 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,656B, BPFP=0.0984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,976B, BPFP=0.1174 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,920B, BPFP=0.1141 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,620B, BPFP=0.0392 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12547956 505.45603612 + layer.0.v_cache 0.00001492 0.04681108 + layer.1.k_cache 0.50446328 35.20594701 + layer.1.v_cache 0.00000633 0.01703702 + layer.2.k_cache 0.00671430 6.29077427 + layer.2.v_cache 0.00001992 0.05391251 + layer.3.k_cache 0.02142675 22.64695966 + layer.3.v_cache 0.00001986 0.06395782 + layer.4.k_cache 0.00069859 1.67145889 + layer.4.v_cache 0.00005414 0.13700672 + layer.4.output 0.01038299 211.70021388 + ------------------------------------------------------------------------------------- + TOTAL 0.04303403 120.79361166 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 22172 +BPFP 0.0775 bits/point +EBPFP 0.1550 equivalent bits/point +MSE 120.793612 +---------------------- -------------------------------------------------------- +Time: 0.627s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.367s +---------------------- -------------------------------------------------------- +💾 Converting with 120.7936 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,600B, BPFP=0.0831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,856B, BPFP=0.0963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,956B, BPFP=0.1015 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,692B, BPFP=0.0878 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,832B, BPFP=0.0951 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,756B, BPFP=0.0912 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,768B, BPFP=0.0918 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,768B, BPFP=0.0918 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,056B, BPFP=0.1067 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,956B, BPFP=0.1015 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,320B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13970643 505.83077243 + layer.0.v_cache 0.00001441 0.04680184 + layer.1.k_cache 0.60015088 35.24968205 + layer.1.v_cache 0.00000584 0.01702860 + layer.2.k_cache 0.00648033 6.27587607 + layer.2.v_cache 0.00001840 0.05330753 + layer.3.k_cache 0.03288257 22.55883708 + layer.3.v_cache 0.00002024 0.06362842 + layer.4.k_cache 0.00070287 1.68960602 + layer.4.v_cache 0.00005060 0.13534761 + layer.4.output 0.04874230 180.45491220 + ------------------------------------------------------------------------------------- + TOTAL 0.06595463 107.94736900 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 22560 +BPFP 0.0689 bits/point +EBPFP 0.1378 equivalent bits/point +MSE 107.947369 +---------------------- -------------------------------------------------------- +Time: 0.625s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 107.9474 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.0889 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,092B, BPFP=0.1000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,212B, BPFP=0.1057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,860B, BPFP=0.0889 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,068B, BPFP=0.0988 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,928B, BPFP=0.0921 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,000B, BPFP=0.0956 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,876B, BPFP=0.0896 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,312B, BPFP=0.1105 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,188B, BPFP=0.1045 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,176B, BPFP=0.0353 +⌛️ [2/4] FRONTEND: Frontend time: 0.280s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.425s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13960611 503.34862385 + layer.0.v_cache 0.00001491 0.04759152 + layer.1.k_cache 0.68890666 34.96254719 + layer.1.v_cache 0.00000583 0.01712133 + layer.2.k_cache 0.00940153 6.24700685 + layer.2.v_cache 0.00001882 0.05342558 + layer.3.k_cache 0.04667800 22.56999594 + layer.3.v_cache 0.00001939 0.06411005 + layer.4.k_cache 0.00070827 1.67242301 + layer.4.v_cache 0.00005291 0.13657574 + layer.4.output 0.04427218 166.38731433 + ------------------------------------------------------------------------------------- + TOTAL 0.07031281 101.99003655 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 25572 +BPFP 0.0719 bits/point +EBPFP 0.1438 equivalent bits/point +MSE 101.990037 +---------------------- -------------------------------------------------------- +Time: 0.718s Load: 0.013s, Pack+Encode: 0.280s, Decode+Unpack: 0.425s +---------------------- -------------------------------------------------------- +💾 Converting with 101.9900 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,620B, BPFP=0.0847 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,864B, BPFP=0.0974 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,948B, BPFP=0.1018 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,696B, BPFP=0.0886 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,844B, BPFP=0.0964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,752B, BPFP=0.0916 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,760B, BPFP=0.0920 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,736B, BPFP=0.0907 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,048B, BPFP=0.1070 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,964B, BPFP=0.1026 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,260B, BPFP=0.0318 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15817952 506.87635870 + layer.0.v_cache 0.00001474 0.04618829 + layer.1.k_cache 0.61609703 35.29098296 + layer.1.v_cache 0.00000566 0.01681751 + layer.2.k_cache 0.00844678 6.22145351 + layer.2.v_cache 0.00001876 0.05217590 + layer.3.k_cache 0.04682518 22.50077080 + layer.3.v_cache 0.00001896 0.06261975 + layer.4.k_cache 0.00069859 1.66231238 + layer.4.v_cache 0.00005216 0.13369091 + layer.4.output 0.04649196 182.12823997 + ------------------------------------------------------------------------------------- + TOTAL 0.06798830 108.69182650 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 22492 +BPFP 0.0691 bits/point +EBPFP 0.1383 equivalent bits/point +MSE 108.691826 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.012s, Pack+Encode: 0.250s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 108.6918 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,484B, BPFP=0.0808 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,752B, BPFP=0.0954 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,844B, BPFP=0.1004 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,540B, BPFP=0.0838 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,720B, BPFP=0.0936 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,612B, BPFP=0.0878 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,660B, BPFP=0.0904 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,612B, BPFP=0.0878 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,936B, BPFP=0.1054 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,820B, BPFP=0.0991 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,864B, BPFP=0.0301 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14002854 507.94414199 + layer.0.v_cache 0.00001601 0.04724641 + layer.1.k_cache 0.54217949 34.97023696 + layer.1.v_cache 0.00000579 0.01722846 + layer.2.k_cache 0.01082110 6.22657526 + layer.2.v_cache 0.00001971 0.05275679 + layer.3.k_cache 0.03102738 22.49065801 + layer.3.v_cache 0.00001907 0.06395095 + layer.4.k_cache 0.00066782 1.63946321 + layer.4.v_cache 0.00005141 0.13447599 + layer.4.output 0.00856273 193.57688838 + ------------------------------------------------------------------------------------- + TOTAL 0.04616326 113.44852663 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 20844 +BPFP 0.0668 bits/point +EBPFP 0.1335 equivalent bits/point +MSE 113.448527 +---------------------- -------------------------------------------------------- +Time: 0.626s Load: 0.012s, Pack+Encode: 0.249s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 113.4485 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,616B, BPFP=0.0850 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,864B, BPFP=0.0981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,936B, BPFP=0.1019 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,676B, BPFP=0.0882 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,840B, BPFP=0.0968 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,744B, BPFP=0.0918 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,768B, BPFP=0.0930 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,740B, BPFP=0.0915 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,068B, BPFP=0.1088 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,940B, BPFP=0.1021 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,316B, BPFP=0.0324 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14705232 511.44391835 + layer.0.v_cache 0.00001374 0.04607145 + layer.1.k_cache 0.65666707 35.09525923 + layer.1.v_cache 0.00000579 0.01669782 + layer.2.k_cache 0.01062412 6.22526453 + layer.2.v_cache 0.00001918 0.05243653 + layer.3.k_cache 0.03205616 22.47863400 + layer.3.v_cache 0.00001999 0.06374318 + layer.4.k_cache 0.00071739 1.64902853 + layer.4.v_cache 0.00005111 0.13236089 + layer.4.output 0.04976816 182.93195346 + ------------------------------------------------------------------------------------- + TOTAL 0.07032965 109.27806404 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 22508 +BPFP 0.0697 bits/point +EBPFP 0.1393 equivalent bits/point +MSE 109.278064 +---------------------- -------------------------------------------------------- +Time: 0.624s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.363s +---------------------- -------------------------------------------------------- +💾 Converting with 109.2781 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,568B, BPFP=0.0872 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,836B, BPFP=0.1021 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,868B, BPFP=0.1039 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,608B, BPFP=0.0894 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,744B, BPFP=0.0970 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,668B, BPFP=0.0927 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,664B, BPFP=0.0925 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,672B, BPFP=0.0930 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,960B, BPFP=0.1090 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,868B, BPFP=0.1039 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,208B, BPFP=0.0334 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131647 505.72269795 + layer.0.v_cache 0.00001392 0.04594674 + layer.1.k_cache 0.51967005 34.93095599 + layer.1.v_cache 0.00000557 0.01653272 + layer.2.k_cache 0.01013049 6.20213888 + layer.2.v_cache 0.00001908 0.05243654 + layer.3.k_cache 0.05186962 22.51225916 + layer.3.v_cache 0.00001891 0.06349004 + layer.4.k_cache 0.00069958 1.67203443 + layer.4.v_cache 0.00005017 0.13230634 + layer.4.output 0.01073789 198.61389489 + ------------------------------------------------------------------------------------- + TOTAL 0.04758583 115.39106253 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 21664 +BPFP 0.0709 bits/point +EBPFP 0.1417 equivalent bits/point +MSE 115.391063 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.012s, Pack+Encode: 0.250s, Decode+Unpack: 0.368s +---------------------- -------------------------------------------------------- +💾 Converting with 115.3911 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,500B, BPFP=0.0814 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,732B, BPFP=0.0940 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,848B, BPFP=0.1003 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,548B, BPFP=0.0840 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,716B, BPFP=0.0931 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,620B, BPFP=0.0879 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,672B, BPFP=0.0907 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,604B, BPFP=0.0870 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,952B, BPFP=0.1059 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,816B, BPFP=0.0985 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,956B, BPFP=0.0307 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15566496 506.21321615 + layer.0.v_cache 0.00001437 0.04572762 + layer.1.k_cache 0.62643353 35.10440403 + layer.1.v_cache 0.00000573 0.01684925 + layer.2.k_cache 0.01010078 6.26455604 + layer.2.v_cache 0.00001868 0.05305424 + layer.3.k_cache 0.06271097 22.59852600 + layer.3.v_cache 0.00001958 0.06360448 + layer.4.k_cache 0.00073784 1.66270203 + layer.4.v_cache 0.00005294 0.13556422 + layer.4.output 0.00677609 194.22733755 + ------------------------------------------------------------------------------------- + TOTAL 0.05312894 113.63232746 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 20964 +BPFP 0.0669 bits/point +EBPFP 0.1338 equivalent bits/point +MSE 113.632327 +---------------------- -------------------------------------------------------- +Time: 0.627s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 113.6323 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,480B, BPFP=0.0809 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,764B, BPFP=0.0964 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,840B, BPFP=0.1005 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,564B, BPFP=0.0854 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,716B, BPFP=0.0938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,628B, BPFP=0.0889 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,656B, BPFP=0.0905 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,612B, BPFP=0.0881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,944B, BPFP=0.1062 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,820B, BPFP=0.0994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,932B, BPFP=0.0307 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13918158 503.61860795 + layer.0.v_cache 0.00001566 0.04582562 + layer.1.k_cache 0.57997750 34.86494755 + layer.1.v_cache 0.00000571 0.01671114 + layer.2.k_cache 0.01637170 6.19885211 + layer.2.v_cache 0.00001887 0.05275037 + layer.3.k_cache 0.04393929 22.37985891 + layer.3.v_cache 0.00001903 0.06134610 + layer.4.k_cache 0.00071104 1.63651708 + layer.4.v_cache 0.00005175 0.13397842 + layer.4.output 0.01065216 194.30759865 + ------------------------------------------------------------------------------------- + TOTAL 0.05028572 113.48015211 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 20956 +BPFP 0.0673 bits/point +EBPFP 0.1347 equivalent bits/point +MSE 113.480152 +---------------------- -------------------------------------------------------- +Time: 0.623s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 113.4802 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,644B, BPFP=0.0829 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,828B, BPFP=0.0921 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,936B, BPFP=0.0976 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,632B, BPFP=0.0823 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,808B, BPFP=0.0911 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,660B, BPFP=0.0837 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,740B, BPFP=0.0877 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,684B, BPFP=0.0849 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,000B, BPFP=0.1008 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,884B, BPFP=0.0950 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,968B, BPFP=0.0286 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16496690 505.42757056 + layer.0.v_cache 0.00001433 0.04737436 + layer.1.k_cache 0.65265572 35.16623299 + layer.1.v_cache 0.00000578 0.01736308 + layer.2.k_cache 0.01241827 6.34286755 + layer.2.v_cache 0.00001920 0.05360854 + layer.3.k_cache 0.08444468 22.78409621 + layer.3.v_cache 0.00001975 0.06395174 + layer.4.k_cache 0.00068928 1.68084382 + layer.4.v_cache 0.00005259 0.13770845 + layer.4.output 0.04785494 177.51392569 + ------------------------------------------------------------------------------------- + TOTAL 0.07354536 106.72465277 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 21784 +BPFP 0.0646 bits/point +EBPFP 0.1292 equivalent bits/point +MSE 106.724653 +---------------------- -------------------------------------------------------- +Time: 0.626s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 106.7247 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,564B, BPFP=0.0831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,800B, BPFP=0.0957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,900B, BPFP=0.1010 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,660B, BPFP=0.0882 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,788B, BPFP=0.0950 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,708B, BPFP=0.0908 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,716B, BPFP=0.0912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,684B, BPFP=0.0895 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,000B, BPFP=0.1063 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,884B, BPFP=0.1001 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,184B, BPFP=0.0318 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15001445 505.70121173 + layer.0.v_cache 0.00001441 0.04700403 + layer.1.k_cache 0.57475359 35.12805259 + layer.1.v_cache 0.00000584 0.01700783 + layer.2.k_cache 0.01434258 6.22496811 + layer.2.v_cache 0.00001954 0.05324837 + layer.3.k_cache 0.02649183 22.49899022 + layer.3.v_cache 0.00001945 0.06267453 + layer.4.k_cache 0.00069414 1.66720882 + layer.4.v_cache 0.00005306 0.13552511 + layer.4.output 0.05035525 185.07642432 + ------------------------------------------------------------------------------------- + TOTAL 0.06581739 109.82769774 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 21888 +BPFP 0.0684 bits/point +EBPFP 0.1369 equivalent bits/point +MSE 109.827698 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.011s, Pack+Encode: 0.250s, Decode+Unpack: 0.367s +---------------------- -------------------------------------------------------- +💾 Converting with 109.8277 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,496B, BPFP=0.0829 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,792B, BPFP=0.0993 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,852B, BPFP=0.1026 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,568B, BPFP=0.0869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,740B, BPFP=0.0964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,612B, BPFP=0.0893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,680B, BPFP=0.0931 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,608B, BPFP=0.0891 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,968B, BPFP=0.1090 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,824B, BPFP=0.1011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,096B, BPFP=0.0324 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11854591 506.07158688 + layer.0.v_cache 0.00001566 0.04624230 + layer.1.k_cache 0.56985614 34.85639337 + layer.1.v_cache 0.00000555 0.01667789 + layer.2.k_cache 0.01267573 6.26809108 + layer.2.v_cache 0.00001902 0.05219390 + layer.3.k_cache 0.04477839 22.60927111 + layer.3.v_cache 0.00001896 0.06242935 + layer.4.k_cache 0.00071754 1.68609944 + layer.4.v_cache 0.00005321 0.13580625 + layer.4.output 0.00761333 197.78791477 + ------------------------------------------------------------------------------------- + TOTAL 0.04705761 115.07765853 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 21236 +BPFP 0.0692 bits/point +EBPFP 0.1384 equivalent bits/point +MSE 115.077659 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.009s, Pack+Encode: 0.250s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 115.0777 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,560B, BPFP=0.0829 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,800B, BPFP=0.0957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,896B, BPFP=0.1008 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,624B, BPFP=0.0863 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,776B, BPFP=0.0944 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,704B, BPFP=0.0906 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,712B, BPFP=0.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,700B, BPFP=0.0903 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,996B, BPFP=0.1061 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,880B, BPFP=0.0999 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,156B, BPFP=0.0316 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582189 503.98644770 + layer.0.v_cache 0.00001509 0.04738283 + layer.1.k_cache 0.58311327 35.06242028 + layer.1.v_cache 0.00000606 0.01714200 + layer.2.k_cache 0.00657600 6.31519883 + layer.2.v_cache 0.00001909 0.05387617 + layer.3.k_cache 0.02234380 22.59369021 + layer.3.v_cache 0.00001905 0.06343653 + layer.4.k_cache 0.00071828 1.68877437 + layer.4.v_cache 0.00005370 0.13613654 + layer.4.output 0.05044473 185.10620141 + ------------------------------------------------------------------------------------- + TOTAL 0.06363526 109.74752443 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 21804 +BPFP 0.0682 bits/point +EBPFP 0.1363 equivalent bits/point +MSE 109.747524 +---------------------- -------------------------------------------------------- +Time: 0.623s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.363s +---------------------- -------------------------------------------------------- +💾 Converting with 109.7475 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 292, 128) +Output shape: (1, 292, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.output: torch.Size([1, 292, 3584]) -> torch.Size([1, 1, 292, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,532B, BPFP=0.0820 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,768B, BPFP=0.0946 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,872B, BPFP=0.1002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,572B, BPFP=0.0841 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,744B, BPFP=0.0933 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,632B, BPFP=0.0873 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,688B, BPFP=0.0903 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,624B, BPFP=0.0869 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,952B, BPFP=0.1045 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,832B, BPFP=0.0980 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,936B, BPFP=0.0301 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887579 504.03933005 + layer.0.v_cache 0.00001384 0.04679759 + layer.1.k_cache 0.57277983 35.01502970 + layer.1.v_cache 0.00000568 0.01704204 + layer.2.k_cache 0.01314942 6.31778372 + layer.2.v_cache 0.00002105 0.05385326 + layer.3.k_cache 0.03125376 22.67498863 + layer.3.v_cache 0.00001978 0.06378456 + layer.4.k_cache 0.00068035 1.67111415 + layer.4.v_cache 0.00005123 0.13591695 + layer.4.output 0.04859251 186.02025746 + ------------------------------------------------------------------------------------- + TOTAL 0.06511755 110.12808487 + (elements=2,541,568) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2541568 +Total Bytes 21152 +BPFP 0.0666 bits/point +EBPFP 0.1332 equivalent bits/point +MSE 110.128085 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.009s, Pack+Encode: 0.252s, Decode+Unpack: 0.367s +---------------------- -------------------------------------------------------- +💾 Converting with 110.1281 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 338, 128) +Output shape: (1, 338, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.output: torch.Size([1, 338, 3584]) -> torch.Size([1, 1, 338, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,820B, BPFP=0.0841 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 2,104B, BPFP=0.0973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,280B, BPFP=0.1054 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,956B, BPFP=0.0904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,104B, BPFP=0.0973 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,980B, BPFP=0.0915 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,028B, BPFP=0.0938 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 2,012B, BPFP=0.0930 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,392B, BPFP=0.1106 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 2,284B, BPFP=0.1056 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,880B, BPFP=0.0322 +⌛️ [2/4] FRONTEND: Frontend time: 0.287s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.425s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13976314 506.48511464 + layer.0.v_cache 0.00001472 0.04757456 + layer.1.k_cache 0.73750278 34.81707944 + layer.1.v_cache 0.00000589 0.01724076 + layer.2.k_cache 0.02011468 6.18447425 + layer.2.v_cache 0.00002041 0.05253284 + layer.3.k_cache 0.02722352 22.55125509 + layer.3.v_cache 0.00002029 0.06312737 + layer.4.k_cache 0.00072609 1.68614152 + layer.4.v_cache 0.00005477 0.13767526 + layer.4.output 0.04270584 160.92231086 + ------------------------------------------------------------------------------------- + TOTAL 0.07202277 99.91167010 + (elements=2,941,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2941952 +Total Bytes 25840 +BPFP 0.0703 bits/point +EBPFP 0.1405 equivalent bits/point +MSE 99.911670 +---------------------- -------------------------------------------------------- +Time: 0.722s Load: 0.011s, Pack+Encode: 0.287s, Decode+Unpack: 0.425s +---------------------- -------------------------------------------------------- +💾 Converting with 99.9117 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,500B, BPFP=0.0805 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,736B, BPFP=0.0932 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,856B, BPFP=0.0997 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,520B, BPFP=0.0816 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,720B, BPFP=0.0924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,572B, BPFP=0.0844 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,668B, BPFP=0.0896 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,572B, BPFP=0.0844 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,948B, BPFP=0.1046 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,784B, BPFP=0.0958 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,904B, BPFP=0.0299 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150811 502.15560567 + layer.0.v_cache 0.00001374 0.04685737 + layer.1.k_cache 0.66021251 35.04999597 + layer.1.v_cache 0.00000574 0.01697746 + layer.2.k_cache 0.00963001 6.27679548 + layer.2.v_cache 0.00001988 0.05224529 + layer.3.k_cache 0.08055985 22.56357556 + layer.3.v_cache 0.00001942 0.06189657 + layer.4.k_cache 0.00069026 1.70537229 + layer.4.v_cache 0.00005126 0.13538072 + layer.4.output 0.05093874 187.07248711 + ------------------------------------------------------------------------------------- + TOTAL 0.07289894 110.44541836 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 20780 +BPFP 0.0656 bits/point +EBPFP 0.1313 equivalent bits/point +MSE 110.445418 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 110.4454 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,604B, BPFP=0.0827 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,860B, BPFP=0.0959 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,936B, BPFP=0.0998 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,720B, BPFP=0.0887 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,836B, BPFP=0.0947 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,764B, BPFP=0.0910 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,768B, BPFP=0.0912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,768B, BPFP=0.0912 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,052B, BPFP=0.1058 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,932B, BPFP=0.0996 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,348B, BPFP=0.0320 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14893842 510.50969472 + layer.0.v_cache 0.00001415 0.04542518 + layer.1.k_cache 0.65110169 35.19894157 + layer.1.v_cache 0.00000575 0.01682538 + layer.2.k_cache 0.01411552 6.22955947 + layer.2.v_cache 0.00001904 0.05222580 + layer.3.k_cache 0.03939451 22.53104051 + layer.3.v_cache 0.00001876 0.06138699 + layer.4.k_cache 0.00070558 1.68039132 + layer.4.v_cache 0.00005345 0.13246696 + layer.4.output 0.04819407 179.83432049 + ------------------------------------------------------------------------------------- + TOTAL 0.07010149 107.95871772 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 22588 +BPFP 0.0685 bits/point +EBPFP 0.1370 equivalent bits/point +MSE 107.958718 +---------------------- -------------------------------------------------------- +Time: 0.626s Load: 0.012s, Pack+Encode: 0.250s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 107.9587 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,576B, BPFP=0.0936 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,804B, BPFP=0.1072 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,872B, BPFP=0.1112 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,608B, BPFP=0.0955 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,764B, BPFP=0.1048 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,636B, BPFP=0.0972 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,708B, BPFP=0.1015 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,648B, BPFP=0.0979 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,980B, BPFP=0.1176 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,896B, BPFP=0.1126 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,560B, BPFP=0.0387 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14492303 505.13331749 + layer.0.v_cache 0.00001375 0.04587666 + layer.1.k_cache 0.50395557 35.07775368 + layer.1.v_cache 0.00000564 0.01669434 + layer.2.k_cache 0.00856250 6.29184134 + layer.2.v_cache 0.00001902 0.05294720 + layer.3.k_cache 0.02692018 22.67644034 + layer.3.v_cache 0.00002118 0.06258785 + layer.4.k_cache 0.00070790 1.65566088 + layer.4.v_cache 0.00005130 0.13183688 + layer.4.output 0.01131245 211.67359791 + ------------------------------------------------------------------------------------- + TOTAL 0.04496278 120.75647894 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 22052 +BPFP 0.0771 bits/point +EBPFP 0.1541 equivalent bits/point +MSE 120.756479 +---------------------- -------------------------------------------------------- +Time: 0.626s Load: 0.012s, Pack+Encode: 0.249s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 120.7565 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,528B, BPFP=0.0865 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,828B, BPFP=0.1035 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,916B, BPFP=0.1085 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,644B, BPFP=0.0931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,800B, BPFP=0.1019 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,704B, BPFP=0.0965 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,708B, BPFP=0.0967 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,684B, BPFP=0.0953 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,016B, BPFP=0.1141 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,896B, BPFP=0.1073 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,180B, BPFP=0.0338 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14113761 508.54761096 + layer.0.v_cache 0.00001369 0.04640708 + layer.1.k_cache 0.59212339 34.95528334 + layer.1.v_cache 0.00000552 0.01644461 + layer.2.k_cache 0.01129283 6.19710994 + layer.2.v_cache 0.00001853 0.05252991 + layer.3.k_cache 0.01358872 22.64139634 + layer.3.v_cache 0.00001883 0.06250121 + layer.4.k_cache 0.00070763 1.65320576 + layer.4.v_cache 0.00005153 0.13443952 + layer.4.output 0.01019159 201.84624094 + ------------------------------------------------------------------------------------- + TOTAL 0.04884114 116.89591855 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 21904 +BPFP 0.0729 bits/point +EBPFP 0.1459 equivalent bits/point +MSE 116.895919 +---------------------- -------------------------------------------------------- +Time: 0.630s Load: 0.012s, Pack+Encode: 0.250s, Decode+Unpack: 0.368s +---------------------- -------------------------------------------------------- +💾 Converting with 116.8959 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,496B, BPFP=0.0814 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,768B, BPFP=0.0963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,848B, BPFP=0.1006 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,576B, BPFP=0.0858 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,728B, BPFP=0.0941 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,648B, BPFP=0.0897 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,668B, BPFP=0.0908 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,632B, BPFP=0.0889 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,948B, BPFP=0.1061 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,816B, BPFP=0.0989 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,920B, BPFP=0.0305 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11028925 506.14514373 + layer.0.v_cache 0.00001466 0.04696211 + layer.1.k_cache 0.56032453 34.97166267 + layer.1.v_cache 0.00000595 0.01697247 + layer.2.k_cache 0.00809042 6.22920679 + layer.2.v_cache 0.00001929 0.05315789 + layer.3.k_cache 0.03501172 22.44620570 + layer.3.v_cache 0.00001861 0.06277635 + layer.4.k_cache 0.00070074 1.68642178 + layer.4.v_cache 0.00005246 0.13411590 + layer.4.output 0.00889761 193.62425336 + ------------------------------------------------------------------------------------- + TOTAL 0.04569476 113.36249405 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 21048 +BPFP 0.0674 bits/point +EBPFP 0.1348 equivalent bits/point +MSE 113.362494 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 113.3625 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 311, 128) +Output shape: (1, 311, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.output: torch.Size([1, 311, 3584]) -> torch.Size([1, 1, 311, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,636B, BPFP=0.0822 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,860B, BPFP=0.0934 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,932B, BPFP=0.0971 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,656B, BPFP=0.0832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,812B, BPFP=0.0910 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,716B, BPFP=0.0862 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,748B, BPFP=0.0878 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,696B, BPFP=0.0852 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,984B, BPFP=0.0997 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,924B, BPFP=0.0967 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,040B, BPFP=0.0290 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14227538 508.75035169 + layer.0.v_cache 0.00001500 0.04706488 + layer.1.k_cache 0.64922478 35.28931057 + layer.1.v_cache 0.00000572 0.01709159 + layer.2.k_cache 0.00828172 6.37633846 + layer.2.v_cache 0.00001922 0.05307848 + layer.3.k_cache 0.02636586 22.81478283 + layer.3.v_cache 0.00001996 0.06570430 + layer.4.k_cache 0.00069494 1.69616032 + layer.4.v_cache 0.00005248 0.13862737 + layer.4.output 0.04686997 180.65310634 + ------------------------------------------------------------------------------------- + TOTAL 0.06794382 108.22472087 + (elements=2,706,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2706944 +Total Bytes 22004 +BPFP 0.0650 bits/point +EBPFP 0.1301 equivalent bits/point +MSE 108.224721 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.012s, Pack+Encode: 0.250s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 108.2247 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,140B, BPFP=0.0699 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,300B, BPFP=0.0797 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,416B, BPFP=0.0868 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,092B, BPFP=0.0669 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,312B, BPFP=0.0804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,180B, BPFP=0.0723 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,216B, BPFP=0.0745 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,152B, BPFP=0.0706 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,456B, BPFP=0.0892 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,352B, BPFP=0.0828 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,592B, BPFP=0.0227 +⌛️ [2/4] FRONTEND: Frontend time: 0.287s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13149661 498.55143995 + layer.0.v_cache 0.00001370 0.04547329 + layer.1.k_cache 0.50214604 34.22792969 + layer.1.v_cache 0.00000625 0.01670538 + layer.2.k_cache 0.00807812 6.07939740 + layer.2.v_cache 0.00001969 0.05223834 + layer.3.k_cache 0.06061829 22.08176317 + layer.3.v_cache 0.00001868 0.06153243 + layer.4.k_cache 0.00068906 1.58418160 + layer.4.v_cache 0.00005044 0.13205812 + layer.4.output 1.20679682 212.66859244 + ------------------------------------------------------------------------------------- + TOTAL 0.53827733 120.67722744 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 15208 +BPFP 0.0548 bits/point +EBPFP 0.1096 equivalent bits/point +MSE 120.677227 +---------------------- -------------------------------------------------------- +Time: 0.611s Load: 0.009s, Pack+Encode: 0.287s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 120.6772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 300, 128) +Output shape: (1, 300, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.output: torch.Size([1, 300, 3584]) -> torch.Size([1, 1, 300, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,616B, BPFP=0.0842 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,844B, BPFP=0.0960 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,948B, BPFP=0.1015 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,656B, BPFP=0.0862 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,852B, BPFP=0.0965 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,704B, BPFP=0.0887 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,752B, BPFP=0.0912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,708B, BPFP=0.0890 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,032B, BPFP=0.1058 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,924B, BPFP=0.1002 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,264B, BPFP=0.0317 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13491969 509.11296875 + layer.0.v_cache 0.00001558 0.04643215 + layer.1.k_cache 0.64590566 35.22377279 + layer.1.v_cache 0.00000556 0.01658740 + layer.2.k_cache 0.01395306 6.24054077 + layer.2.v_cache 0.00001972 0.05309159 + layer.3.k_cache 0.02379198 22.44383789 + layer.3.v_cache 0.00001942 0.06434862 + layer.4.k_cache 0.00073123 1.70016927 + layer.4.v_cache 0.00005274 0.13542811 + layer.4.output 0.05011183 181.16114583 + ------------------------------------------------------------------------------------- + TOTAL 0.06883514 108.42148225 + (elements=2,611,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2611200 +Total Bytes 22300 +BPFP 0.0683 bits/point +EBPFP 0.1366 equivalent bits/point +MSE 108.421482 +---------------------- -------------------------------------------------------- +Time: 0.624s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 108.4215 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,512B, BPFP=0.0812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,784B, BPFP=0.0958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,860B, BPFP=0.0999 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,584B, BPFP=0.0851 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,720B, BPFP=0.0924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,660B, BPFP=0.0891 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,672B, BPFP=0.0898 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,684B, BPFP=0.0904 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,960B, BPFP=0.1052 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,856B, BPFP=0.0997 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,976B, BPFP=0.0305 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14826349 499.50289948 + layer.0.v_cache 0.00001415 0.04636622 + layer.1.k_cache 0.61440683 35.01378933 + layer.1.v_cache 0.00000562 0.01707268 + layer.2.k_cache 0.00551110 6.35029020 + layer.2.v_cache 0.00001912 0.05217824 + layer.3.k_cache 0.01294636 22.44138276 + layer.3.v_cache 0.00002032 0.06219530 + layer.4.k_cache 0.00070633 1.68444992 + layer.4.v_cache 0.00005074 0.13405499 + layer.4.output 0.04972342 187.05021171 + ------------------------------------------------------------------------------------- + TOTAL 0.06647106 110.27389183 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 21268 +BPFP 0.0672 bits/point +EBPFP 0.1343 equivalent bits/point +MSE 110.273892 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 110.2739 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,536B, BPFP=0.0866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,820B, BPFP=0.1027 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,904B, BPFP=0.1074 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,620B, BPFP=0.0914 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,788B, BPFP=0.1009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,660B, BPFP=0.0936 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,716B, BPFP=0.0968 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,668B, BPFP=0.0941 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,976B, BPFP=0.1115 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,836B, BPFP=0.1036 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,172B, BPFP=0.0336 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13861307 507.00767148 + layer.0.v_cache 0.00001419 0.04632708 + layer.1.k_cache 0.53867822 35.05169083 + layer.1.v_cache 0.00000573 0.01664808 + layer.2.k_cache 0.00871976 6.23532204 + layer.2.v_cache 0.00001877 0.05174781 + layer.3.k_cache 0.02195022 22.58365298 + layer.3.v_cache 0.00001967 0.06463704 + layer.4.k_cache 0.00069227 1.65905013 + layer.4.v_cache 0.00004961 0.13245176 + layer.4.output 0.00954030 201.32793966 + ------------------------------------------------------------------------------------- + TOTAL 0.04562021 116.59675158 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 21696 +BPFP 0.0720 bits/point +EBPFP 0.1440 equivalent bits/point +MSE 116.596752 +---------------------- -------------------------------------------------------- +Time: 0.625s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 116.5968 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,476B, BPFP=0.0806 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,736B, BPFP=0.0948 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,836B, BPFP=0.1003 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,548B, BPFP=0.0846 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,716B, BPFP=0.0938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,608B, BPFP=0.0878 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,656B, BPFP=0.0905 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,612B, BPFP=0.0881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,952B, BPFP=0.1066 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,816B, BPFP=0.0992 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,888B, BPFP=0.0303 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13928204 507.28529283 + layer.0.v_cache 0.00001373 0.04489414 + layer.1.k_cache 0.60860016 34.85624385 + layer.1.v_cache 0.00000555 0.01628783 + layer.2.k_cache 0.01026906 6.22565764 + layer.2.v_cache 0.00001872 0.05164260 + layer.3.k_cache 0.05166252 22.41680609 + layer.3.v_cache 0.00001937 0.06247686 + layer.4.k_cache 0.00073448 1.66068273 + layer.4.v_cache 0.00005026 0.13187083 + layer.4.output 0.00697322 194.35324051 + ------------------------------------------------------------------------------------- + TOTAL 0.05055697 113.71909053 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 20844 +BPFP 0.0670 bits/point +EBPFP 0.1340 equivalent bits/point +MSE 113.719091 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 113.7191 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,600B, BPFP=0.0847 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,856B, BPFP=0.0983 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,928B, BPFP=0.1021 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,684B, BPFP=0.0892 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,816B, BPFP=0.0962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,708B, BPFP=0.0905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,740B, BPFP=0.0922 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,736B, BPFP=0.0919 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,016B, BPFP=0.1068 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,908B, BPFP=0.1011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,368B, BPFP=0.0331 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13820392 504.90815678 + layer.0.v_cache 0.00001454 0.04609080 + layer.1.k_cache 0.61684219 35.14710673 + layer.1.v_cache 0.00000570 0.01654615 + layer.2.k_cache 0.01285210 6.25819071 + layer.2.v_cache 0.00002011 0.05206322 + layer.3.k_cache 0.05980752 22.53712924 + layer.3.v_cache 0.00002263 0.06261696 + layer.4.k_cache 0.00071023 1.65753515 + layer.4.v_cache 0.00005833 0.13474152 + layer.4.output 0.05086143 184.46828087 + ------------------------------------------------------------------------------------- + TOTAL 0.06968043 109.53518490 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 22360 +BPFP 0.0697 bits/point +EBPFP 0.1393 equivalent bits/point +MSE 109.535185 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.368s +---------------------- -------------------------------------------------------- +💾 Converting with 109.5352 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,604B, BPFP=0.0833 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,856B, BPFP=0.0963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,956B, BPFP=0.1015 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,716B, BPFP=0.0891 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,844B, BPFP=0.0957 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,744B, BPFP=0.0905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,768B, BPFP=0.0918 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,756B, BPFP=0.0912 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,052B, BPFP=0.1065 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,968B, BPFP=0.1022 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,368B, BPFP=0.0324 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14825445 510.05367525 + layer.0.v_cache 0.00001396 0.04755491 + layer.1.k_cache 0.63210902 35.34511265 + layer.1.v_cache 0.00000581 0.01728143 + layer.2.k_cache 0.01184132 6.28001024 + layer.2.v_cache 0.00001890 0.05394627 + layer.3.k_cache 0.01767834 22.60570721 + layer.3.v_cache 0.00002009 0.06489474 + layer.4.k_cache 0.00068272 1.70417395 + layer.4.v_cache 0.00005247 0.13823762 + layer.4.output 0.04867053 180.70041825 + ------------------------------------------------------------------------------------- + TOTAL 0.06772769 108.30667776 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 22632 +BPFP 0.0691 bits/point +EBPFP 0.1382 equivalent bits/point +MSE 108.306678 +---------------------- -------------------------------------------------------- +Time: 0.630s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 108.3067 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,576B, BPFP=0.0933 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,796B, BPFP=0.1063 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,884B, BPFP=0.1115 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,624B, BPFP=0.0961 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,776B, BPFP=0.1051 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,668B, BPFP=0.0987 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,700B, BPFP=0.1006 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,660B, BPFP=0.0982 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,996B, BPFP=0.1181 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,880B, BPFP=0.1113 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,488B, BPFP=0.0379 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14315909 500.16986269 + layer.0.v_cache 0.00001391 0.04647600 + layer.1.k_cache 0.49711875 35.12083111 + layer.1.v_cache 0.00000560 0.01675322 + layer.2.k_cache 0.00646762 6.27359610 + layer.2.v_cache 0.00001905 0.05227609 + layer.3.k_cache 0.04707547 22.62665535 + layer.3.v_cache 0.00001924 0.06232677 + layer.4.k_cache 0.00071833 1.65882989 + layer.4.v_cache 0.00005197 0.13230262 + layer.4.output 0.00997522 210.63003923 + ------------------------------------------------------------------------------------- + TOTAL 0.04496915 120.03354026 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 22048 +BPFP 0.0768 bits/point +EBPFP 0.1535 equivalent bits/point +MSE 120.033540 +---------------------- -------------------------------------------------------- +Time: 0.625s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 120.0335 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,564B, BPFP=0.0876 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,808B, BPFP=0.1013 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,884B, BPFP=0.1055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,576B, BPFP=0.0883 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,776B, BPFP=0.0995 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,644B, BPFP=0.0921 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,692B, BPFP=0.0948 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,620B, BPFP=0.0907 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,968B, BPFP=0.1102 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,880B, BPFP=0.1053 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,252B, BPFP=0.0340 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15581964 509.28018593 + layer.0.v_cache 0.00001703 0.04676182 + layer.1.k_cache 0.56298232 35.00708095 + layer.1.v_cache 0.00000550 0.01670637 + layer.2.k_cache 0.00979712 6.26322647 + layer.2.v_cache 0.00002023 0.05448221 + layer.3.k_cache 0.04960944 22.50058454 + layer.3.v_cache 0.00001971 0.06499956 + layer.4.k_cache 0.00068561 1.65756740 + layer.4.v_cache 0.00005245 0.13763573 + layer.4.output 0.01122868 200.67994752 + ------------------------------------------------------------------------------------- + TOTAL 0.05044764 116.45816845 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 21664 +BPFP 0.0714 bits/point +EBPFP 0.1427 equivalent bits/point +MSE 116.458168 +---------------------- -------------------------------------------------------- +Time: 0.618s Load: 0.009s, Pack+Encode: 0.247s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 116.4582 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,496B, BPFP=0.0826 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,764B, BPFP=0.0974 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,836B, BPFP=0.1014 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,576B, BPFP=0.0870 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,732B, BPFP=0.0956 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,636B, BPFP=0.0903 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,660B, BPFP=0.0917 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,624B, BPFP=0.0897 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,952B, BPFP=0.1078 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,828B, BPFP=0.1009 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,116B, BPFP=0.0325 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13599485 502.30587456 + layer.0.v_cache 0.00001418 0.04645160 + layer.1.k_cache 0.56373192 34.87346787 + layer.1.v_cache 0.00000588 0.01709605 + layer.2.k_cache 0.01549433 6.21925424 + layer.2.v_cache 0.00001867 0.05231128 + layer.3.k_cache 0.03475572 22.49850928 + layer.3.v_cache 0.00001907 0.06139190 + layer.4.k_cache 0.00072111 1.66683680 + layer.4.v_cache 0.00005295 0.13230086 + layer.4.output 0.00820773 197.19947312 + ------------------------------------------------------------------------------------- + TOTAL 0.04754487 114.60410625 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 21220 +BPFP 0.0689 bits/point +EBPFP 0.1378 equivalent bits/point +MSE 114.604106 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.257s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 114.6041 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,664B, BPFP=0.0844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,868B, BPFP=0.0948 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,948B, BPFP=0.0988 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,684B, BPFP=0.0854 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,824B, BPFP=0.0925 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,728B, BPFP=0.0877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,768B, BPFP=0.0897 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,744B, BPFP=0.0885 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,036B, BPFP=0.1033 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,924B, BPFP=0.0976 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,028B, BPFP=0.0292 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13331154 503.70926339 + layer.0.v_cache 0.00001460 0.04623076 + layer.1.k_cache 0.64535681 35.00649668 + layer.1.v_cache 0.00000563 0.01688185 + layer.2.k_cache 0.00762887 6.31149490 + layer.2.v_cache 0.00001910 0.05366283 + layer.3.k_cache 0.03538683 22.56939301 + layer.3.v_cache 0.00002103 0.06493952 + layer.4.k_cache 0.00072257 1.67132965 + layer.4.v_cache 0.00005300 0.13780995 + layer.4.output 0.04522907 177.18665932 + ------------------------------------------------------------------------------------- + TOTAL 0.06700726 106.46435987 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 22216 +BPFP 0.0663 bits/point +EBPFP 0.1326 equivalent bits/point +MSE 106.464360 +---------------------- -------------------------------------------------------- +Time: 0.626s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.368s +---------------------- -------------------------------------------------------- +💾 Converting with 106.4644 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,472B, BPFP=0.0810 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,764B, BPFP=0.0971 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,824B, BPFP=0.1004 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,588B, BPFP=0.0874 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,724B, BPFP=0.0949 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,628B, BPFP=0.0896 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,652B, BPFP=0.0909 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,624B, BPFP=0.0893 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,928B, BPFP=0.1061 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,836B, BPFP=0.1010 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,032B, BPFP=0.0317 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14806804 506.71275308 + layer.0.v_cache 0.00001376 0.04604143 + layer.1.k_cache 0.60662186 34.84838523 + layer.1.v_cache 0.00000571 0.01693194 + layer.2.k_cache 0.01239752 6.19060076 + layer.2.v_cache 0.00002027 0.05210714 + layer.3.k_cache 0.02456893 22.50763885 + layer.3.v_cache 0.00001975 0.06285851 + layer.4.k_cache 0.00069718 1.64711708 + layer.4.v_cache 0.00005379 0.13392216 + layer.4.output 0.00818182 195.87735790 + ------------------------------------------------------------------------------------- + TOTAL 0.04998468 114.31528597 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 21072 +BPFP 0.0682 bits/point +EBPFP 0.1364 equivalent bits/point +MSE 114.315286 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.012s, Pack+Encode: 0.255s, Decode+Unpack: 0.367s +---------------------- -------------------------------------------------------- +💾 Converting with 114.3153 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,492B, BPFP=0.0824 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,760B, BPFP=0.0972 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,840B, BPFP=0.1016 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,556B, BPFP=0.0859 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,716B, BPFP=0.0947 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,616B, BPFP=0.0892 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,656B, BPFP=0.0914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,612B, BPFP=0.0890 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,940B, BPFP=0.1071 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,804B, BPFP=0.0996 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,016B, BPFP=0.0317 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14592592 505.28649514 + layer.0.v_cache 0.00001391 0.04693429 + layer.1.k_cache 0.53254910 34.82309726 + layer.1.v_cache 0.00000583 0.01743884 + layer.2.k_cache 0.01029040 6.23504056 + layer.2.v_cache 0.00002009 0.05471602 + layer.3.k_cache 0.03848921 22.53025618 + layer.3.v_cache 0.00001972 0.06527389 + layer.4.k_cache 0.00069417 1.69850466 + layer.4.v_cache 0.00005427 0.13715720 + layer.4.output 0.01004909 196.69076540 + ------------------------------------------------------------------------------------- + TOTAL 0.04696507 114.57236893 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 21008 +BPFP 0.0682 bits/point +EBPFP 0.1365 equivalent bits/point +MSE 114.572369 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 114.5724 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,532B, BPFP=0.0880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,804B, BPFP=0.1036 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,928B, BPFP=0.1108 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,644B, BPFP=0.0944 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.1018 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,684B, BPFP=0.0967 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,716B, BPFP=0.0986 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,700B, BPFP=0.0977 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,008B, BPFP=0.1153 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,924B, BPFP=0.1105 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,260B, BPFP=0.0350 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15384994 510.22851562 + layer.0.v_cache 0.00001404 0.04705759 + layer.1.k_cache 0.48768111 35.17868581 + layer.1.v_cache 0.00000584 0.01732356 + layer.2.k_cache 0.00538594 6.24996140 + layer.2.v_cache 0.00001948 0.05364011 + layer.3.k_cache 0.02347840 22.56995347 + layer.3.v_cache 0.00001899 0.06298808 + layer.4.k_cache 0.00068094 1.67664584 + layer.4.v_cache 0.00005116 0.13657846 + layer.4.output 0.00909345 204.30060071 + ------------------------------------------------------------------------------------- + TOTAL 0.04322588 118.01915029 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 21972 +BPFP 0.0742 bits/point +EBPFP 0.1485 equivalent bits/point +MSE 118.019150 +---------------------- -------------------------------------------------------- +Time: 0.625s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 118.0192 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,572B, BPFP=0.0835 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,812B, BPFP=0.0963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,896B, BPFP=0.1008 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,632B, BPFP=0.0867 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,792B, BPFP=0.0952 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,664B, BPFP=0.0884 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,712B, BPFP=0.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,648B, BPFP=0.0876 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 1,976B, BPFP=0.1050 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,848B, BPFP=0.0982 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,108B, BPFP=0.0312 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13453574 501.22730655 + layer.0.v_cache 0.00001386 0.04652586 + layer.1.k_cache 0.70462929 35.12308673 + layer.1.v_cache 0.00000580 0.01697260 + layer.2.k_cache 0.00896267 6.31451084 + layer.2.v_cache 0.00001881 0.05247255 + layer.3.k_cache 0.02468209 22.49805850 + layer.3.v_cache 0.00001940 0.06333111 + layer.4.k_cache 0.00068809 1.66779229 + layer.4.v_cache 0.00005209 0.13030121 + layer.4.output 0.04953035 185.11186528 + ------------------------------------------------------------------------------------- + TOTAL 0.07178355 109.58373031 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 21660 +BPFP 0.0677 bits/point +EBPFP 0.1354 equivalent bits/point +MSE 109.583730 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 109.5837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,536B, BPFP=0.0896 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 1,812B, BPFP=0.1056 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,932B, BPFP=0.1126 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 1,624B, BPFP=0.0947 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,772B, BPFP=0.1033 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 1,676B, BPFP=0.0977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,740B, BPFP=0.1014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 1,688B, BPFP=0.0984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,028B, BPFP=0.1182 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 1,984B, BPFP=0.1157 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,360B, BPFP=0.0363 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150386 507.07229478 + layer.0.v_cache 0.00001401 0.04651595 + layer.1.k_cache 0.57442873 35.25020406 + layer.1.v_cache 0.00000554 0.01678663 + layer.2.k_cache 0.00779028 6.29013836 + layer.2.v_cache 0.00001875 0.05309494 + layer.3.k_cache 0.05176373 22.64581572 + layer.3.v_cache 0.00002027 0.06406874 + layer.4.k_cache 0.00069872 1.67530823 + layer.4.v_cache 0.00005023 0.13734177 + layer.4.output 0.01044900 207.60149587 + ------------------------------------------------------------------------------------- + TOTAL 0.04937866 119.20364943 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 22152 +BPFP 0.0760 bits/point +EBPFP 0.1519 equivalent bits/point +MSE 119.203649 +---------------------- -------------------------------------------------------- +Time: 0.622s Load: 0.009s, Pack+Encode: 0.250s, Decode+Unpack: 0.363s +---------------------- -------------------------------------------------------- +💾 Converting with 119.2036 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.0700 bits/point +Avg EBPFP 0.1401 equivalent bits/point +Avg MSE 113.352911 +Avg Time 0.639s +------------------------ ---------------------------- diff --git a/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..f3ec9ce75ffc060e69ea3bb5209736329b809b11 --- /dev/null +++ b/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 405 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.001_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa +Output output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 644B, BPFP=0.1198 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1414 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1525 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 692B, BPFP=0.1287 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1362 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 684B, BPFP=0.1272 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 712B, BPFP=0.1324 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 820B, BPFP=0.1525 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 772B, BPFP=0.1436 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,284B, BPFP=0.0607 +⌛️ [2/4] FRONTEND: Frontend time: 0.400s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.245s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14073507 527.62527902 + layer.0.v_cache 0.00001426 0.04833912 + layer.1.k_cache 0.05042761 37.25770787 + layer.1.v_cache 0.00000517 0.01657190 + layer.2.k_cache 0.00244098 6.60887800 + layer.2.v_cache 0.00001715 0.04815933 + layer.3.k_cache 0.02726505 23.43162900 + layer.3.v_cache 0.00001796 0.05947283 + layer.4.k_cache 0.00069123 1.72494362 + layer.4.v_cache 0.00005084 0.13460226 + layer.4.output 0.16186065 647.18436437 + ------------------------------------------------------------------------------------- + TOTAL 0.07968764 301.60271374 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 9616 +BPFP 0.1052 bits/point +EBPFP 0.2104 equivalent bits/point +MSE 301.602714 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.005s, Pack+Encode: 0.400s, Decode+Unpack: 0.245s +---------------------- -------------------------------------------------------- +💾 Converting with 301.6027 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 800B, BPFP=0.1506 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1506 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 712B, BPFP=0.1340 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 724B, BPFP=0.1363 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 696B, BPFP=0.1310 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1303 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 732B, BPFP=0.1378 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 816B, BPFP=0.1536 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 780B, BPFP=0.1468 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,408B, BPFP=0.0648 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11392649 527.32713667 + layer.0.v_cache 0.00001670 0.04566670 + layer.1.k_cache 0.05249626 37.29998706 + layer.1.v_cache 0.00000506 0.01586184 + layer.2.k_cache 0.00416583 6.60247729 + layer.2.v_cache 0.00001711 0.04701122 + layer.3.k_cache 0.05931453 23.43140972 + layer.3.v_cache 0.00001789 0.05629708 + layer.4.k_cache 0.00069248 1.65259763 + layer.4.v_cache 0.00004729 0.12767614 + layer.4.output 0.16383441 654.24042599 + ------------------------------------------------------------------------------------- + TOTAL 0.08103179 304.48759431 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 9792 +BPFP 0.1084 bits/point +EBPFP 0.2169 equivalent bits/point +MSE 304.487594 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 304.4876 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 692B, BPFP=0.1150 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 828B, BPFP=0.1376 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 844B, BPFP=0.1403 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 768B, BPFP=0.1277 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 776B, BPFP=0.1290 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 744B, BPFP=0.1237 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 748B, BPFP=0.1243 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 744B, BPFP=0.1237 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 864B, BPFP=0.1436 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 828B, BPFP=0.1376 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,040B, BPFP=0.0484 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17268721 517.30132148 + layer.0.v_cache 0.00001453 0.04791676 + layer.1.k_cache 0.08276152 36.53047602 + layer.1.v_cache 0.00000545 0.01585309 + layer.2.k_cache 0.00739249 6.50023505 + layer.2.v_cache 0.00001625 0.04474382 + layer.3.k_cache 0.05929063 23.05225648 + layer.3.v_cache 0.00001766 0.05641342 + layer.4.k_cache 0.00068371 1.61556942 + layer.4.v_cache 0.00005055 0.12324084 + layer.4.output 0.14466690 577.00446429 + ------------------------------------------------------------------------------------- + TOTAL 0.07856402 272.01878096 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 9876 +BPFP 0.0966 bits/point +EBPFP 0.1931 equivalent bits/point +MSE 272.018781 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.005s, Pack+Encode: 0.163s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 272.0188 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 112, 128) +Output shape: (1, 112, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.output: torch.Size([1, 112, 3584]) -> torch.Size([1, 1, 112, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 760B, BPFP=0.1060 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 872B, BPFP=0.1217 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 880B, BPFP=0.1228 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 768B, BPFP=0.1071 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 812B, BPFP=0.1133 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 796B, BPFP=0.1110 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 816B, BPFP=0.1138 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 860B, BPFP=0.1200 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 904B, BPFP=0.1261 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 848B, BPFP=0.1183 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,356B, BPFP=0.0470 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17321212 523.24312919 + layer.0.v_cache 0.00001450 0.04946280 + layer.1.k_cache 0.08643453 36.92156982 + layer.1.v_cache 0.00000659 0.01768181 + layer.2.k_cache 0.00690480 6.62502071 + layer.2.v_cache 0.00001899 0.05018273 + layer.3.k_cache 0.01205242 23.28347996 + layer.3.v_cache 0.00001923 0.06372704 + layer.4.k_cache 0.00071999 1.69516209 + layer.4.v_cache 0.00004947 0.12823464 + layer.4.output 10.17892331 480.91900510 + ------------------------------------------------------------------------------------- + TOTAL 4.20775858 232.85356979 + (elements=974,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 974848 +Total Bytes 10672 +BPFP 0.0876 bits/point +EBPFP 0.1752 equivalent bits/point +MSE 232.853570 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.006s, Pack+Encode: 0.163s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 232.8536 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1186 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 804B, BPFP=0.1428 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1449 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 740B, BPFP=0.1314 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 772B, BPFP=0.1371 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 784B, BPFP=0.1392 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 736B, BPFP=0.1307 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 804B, BPFP=0.1428 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 880B, BPFP=0.1562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 812B, BPFP=0.1442 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,420B, BPFP=0.0614 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11687154 529.69193892 + layer.0.v_cache 0.00001423 0.04764365 + layer.1.k_cache 0.05324238 36.89142401 + layer.1.v_cache 0.00000539 0.01659927 + layer.2.k_cache 0.00397935 6.63882169 + layer.2.v_cache 0.00001766 0.04605336 + layer.3.k_cache 0.01210797 23.44640836 + layer.3.v_cache 0.00001803 0.05835517 + layer.4.k_cache 0.00075037 1.68783101 + layer.4.v_cache 0.00004481 0.12108943 + layer.4.output 0.15448949 618.90482955 + ------------------------------------------------------------------------------------- + TOTAL 0.07461636 290.05764539 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 10236 +BPFP 0.1069 bits/point +EBPFP 0.2138 equivalent bits/point +MSE 290.057645 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.005s, Pack+Encode: 0.164s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 290.0576 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 744B, BPFP=0.1435 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 796B, BPFP=0.1535 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 660B, BPFP=0.1273 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 708B, BPFP=0.1366 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 672B, BPFP=0.1296 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 668B, BPFP=0.1289 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 708B, BPFP=0.1366 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1520 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1458 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,416B, BPFP=0.0666 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203883 532.42491319 + layer.0.v_cache 0.00001438 0.04926980 + layer.1.k_cache 0.05633916 37.59367163 + layer.1.v_cache 0.00000602 0.01810055 + layer.2.k_cache 0.00244657 6.68224796 + layer.2.v_cache 0.00001847 0.05172042 + layer.3.k_cache 0.05492775 23.39474073 + layer.3.v_cache 0.00001883 0.06111530 + layer.4.k_cache 0.00062984 1.65308314 + layer.4.v_cache 0.00005178 0.13744943 + layer.4.output 0.18538215 670.03009259 + ------------------------------------------------------------------------------------- + TOTAL 0.09142157 311.31040943 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 9548 +BPFP 0.1083 bits/point +EBPFP 0.2167 equivalent bits/point +MSE 311.310409 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 311.3104 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 676B, BPFP=0.1174 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 844B, BPFP=0.1465 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 840B, BPFP=0.1458 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 752B, BPFP=0.1306 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 784B, BPFP=0.1361 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 796B, BPFP=0.1382 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 736B, BPFP=0.1278 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 792B, BPFP=0.1375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 868B, BPFP=0.1507 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 844B, BPFP=0.1465 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,284B, BPFP=0.0566 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13028670 517.44066840 + layer.0.v_cache 0.00001462 0.04693386 + layer.1.k_cache 0.04946813 36.78160807 + layer.1.v_cache 0.00000624 0.01612608 + layer.2.k_cache 0.00411291 6.62242839 + layer.2.v_cache 0.00001729 0.04775574 + layer.3.k_cache 0.01793825 23.42819553 + layer.3.v_cache 0.00001895 0.06222938 + layer.4.k_cache 0.00077120 1.71803623 + layer.4.v_cache 0.00004532 0.12377011 + layer.4.output 0.15109633 605.06701389 + ------------------------------------------------------------------------------------- + TOTAL 0.07413847 283.63275583 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 10216 +BPFP 0.1043 bits/point +EBPFP 0.2087 equivalent bits/point +MSE 283.632756 +---------------------- -------------------------------------------------------- +Time: 0.382s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 283.6328 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 432B, BPFP=0.1184 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 464B, BPFP=0.1272 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 472B, BPFP=0.1294 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 416B, BPFP=0.1140 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 464B, BPFP=0.1272 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 412B, BPFP=0.1129 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 436B, BPFP=0.1195 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 432B, BPFP=0.1184 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 480B, BPFP=0.1316 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 468B, BPFP=0.1283 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,452B, BPFP=0.0569 +⌛️ [2/4] FRONTEND: Frontend time: 0.281s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.150s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11683119 544.07301946 + layer.0.v_cache 0.00001514 0.05710623 + layer.1.k_cache 0.01218004 39.95992239 + layer.1.v_cache 0.00000521 0.01955268 + layer.2.k_cache 0.00491684 7.56144848 + layer.2.v_cache 0.00001706 0.05881794 + layer.3.k_cache 0.10823302 25.36230469 + layer.3.v_cache 0.00002043 0.07398451 + layer.4.k_cache 0.00062874 1.91398928 + layer.4.v_cache 0.00004794 0.15099087 + layer.4.output 0.23833110 979.17042607 + ------------------------------------------------------------------------------------- + TOTAL 0.11242431 439.61318347 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 5928 +BPFP 0.0956 bits/point +EBPFP 0.1912 equivalent bits/point +MSE 439.613183 +---------------------- -------------------------------------------------------- +Time: 0.434s Load: 0.003s, Pack+Encode: 0.281s, Decode+Unpack: 0.150s +---------------------- -------------------------------------------------------- +💾 Converting with 439.6132 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 416B, BPFP=0.1300 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 436B, BPFP=0.1363 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 484B, BPFP=0.1512 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 384B, BPFP=0.1200 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 468B, BPFP=0.1462 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 400B, BPFP=0.1250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 436B, BPFP=0.1363 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 412B, BPFP=0.1288 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 472B, BPFP=0.1475 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 444B, BPFP=0.1388 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,532B, BPFP=0.0684 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.143s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14598725 546.49320313 + layer.0.v_cache 0.00001449 0.05298679 + layer.1.k_cache 0.01347294 40.08372803 + layer.1.v_cache 0.00000521 0.01736285 + layer.2.k_cache 0.00488209 6.98775757 + layer.2.v_cache 0.00001649 0.04901268 + layer.3.k_cache 0.04421538 24.41555176 + layer.3.v_cache 0.00001905 0.06616210 + layer.4.k_cache 0.00073779 1.85576859 + layer.4.v_cache 0.00005089 0.14059450 + layer.4.output 0.27158585 1085.30687500 + ------------------------------------------------------------------------------------- + TOTAL 0.12414721 483.37119135 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 5884 +BPFP 0.1082 bits/point +EBPFP 0.2163 equivalent bits/point +MSE 483.371191 +---------------------- -------------------------------------------------------- +Time: 0.275s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.143s +---------------------- -------------------------------------------------------- +💾 Converting with 483.3712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 46, 128) +Output shape: (1, 46, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.output: torch.Size([1, 46, 3584]) -> torch.Size([1, 1, 46, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 408B, BPFP=0.1386 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 428B, BPFP=0.1454 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 484B, BPFP=0.1644 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 396B, BPFP=0.1345 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 464B, BPFP=0.1576 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 404B, BPFP=0.1372 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 428B, BPFP=0.1454 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 420B, BPFP=0.1427 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 472B, BPFP=0.1603 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 448B, BPFP=0.1522 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,508B, BPFP=0.0732 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14129612 556.93359375 + layer.0.v_cache 0.00001532 0.05577659 + layer.1.k_cache 0.01315293 42.11122728 + layer.1.v_cache 0.00000553 0.01869217 + layer.2.k_cache 0.00779452 7.61447343 + layer.2.v_cache 0.00001723 0.05197764 + layer.3.k_cache 0.01828384 25.60222593 + layer.3.v_cache 0.00001944 0.07306880 + layer.4.k_cache 0.00060559 1.92882256 + layer.4.v_cache 0.00005113 0.15064921 + layer.4.output 0.29517000 1181.53784938 + ------------------------------------------------------------------------------------- + TOTAL 0.13220186 523.84149723 + (elements=400,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 400384 +Total Bytes 5860 +BPFP 0.1171 bits/point +EBPFP 0.2342 equivalent bits/point +MSE 523.841497 +---------------------- -------------------------------------------------------- +Time: 0.272s Load: 0.002s, Pack+Encode: 0.128s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 523.8415 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 110, 128) +Output shape: (1, 110, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.output: torch.Size([1, 110, 3584]) -> torch.Size([1, 1, 110, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 764B, BPFP=0.1085 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 860B, BPFP=0.1222 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 864B, BPFP=0.1227 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 784B, BPFP=0.1114 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 832B, BPFP=0.1182 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 808B, BPFP=0.1148 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 804B, BPFP=0.1142 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 836B, BPFP=0.1187 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 896B, BPFP=0.1273 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 840B, BPFP=0.1193 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,516B, BPFP=0.0511 +⌛️ [2/4] FRONTEND: Frontend time: 0.178s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14476244 522.25774148 + layer.0.v_cache 0.00001476 0.05025144 + layer.1.k_cache 0.05665370 37.35177113 + layer.1.v_cache 0.00000525 0.01690215 + layer.2.k_cache 0.00786287 6.61802146 + layer.2.v_cache 0.00001801 0.04916671 + layer.3.k_cache 0.01410385 23.53706499 + layer.3.v_cache 0.00002027 0.06549771 + layer.4.k_cache 0.00066873 1.70883581 + layer.4.v_cache 0.00004626 0.12690913 + layer.4.output 10.36401748 489.26834416 + ------------------------------------------------------------------------------------- + TOTAL 4.28072226 236.27415124 + (elements=957,440) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 957440 +Total Bytes 10804 +BPFP 0.0903 bits/point +EBPFP 0.1805 equivalent bits/point +MSE 236.274151 +---------------------- -------------------------------------------------------- +Time: 0.385s Load: 0.004s, Pack+Encode: 0.178s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 236.2742 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 624B, BPFP=0.1204 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 716B, BPFP=0.1381 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 788B, BPFP=0.1520 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 668B, BPFP=0.1289 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 712B, BPFP=0.1373 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 672B, BPFP=0.1296 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1304 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 668B, BPFP=0.1289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 776B, BPFP=0.1497 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 720B, BPFP=0.1389 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,392B, BPFP=0.0659 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10533771 529.25597994 + layer.0.v_cache 0.00001610 0.05548246 + layer.1.k_cache 0.05553475 37.58858386 + layer.1.v_cache 0.00000748 0.02106442 + layer.2.k_cache 0.02166171 6.56842644 + layer.2.v_cache 0.00002000 0.06005131 + layer.3.k_cache 0.06386927 23.52337119 + layer.3.v_cache 0.00001944 0.07003775 + layer.4.k_cache 0.00064315 1.78974425 + layer.4.v_cache 0.00005385 0.14153941 + layer.4.output 0.16801396 669.61331570 + ------------------------------------------------------------------------------------- + TOTAL 0.08372125 310.96279358 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 9412 +BPFP 0.1068 bits/point +EBPFP 0.2136 equivalent bits/point +MSE 310.962794 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 310.9628 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 49, 128) +Output shape: (1, 49, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.output: torch.Size([1, 49, 3584]) -> torch.Size([1, 1, 49, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 408B, BPFP=0.1301 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 428B, BPFP=0.1365 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 484B, BPFP=0.1543 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 388B, BPFP=0.1237 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 460B, BPFP=0.1467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 408B, BPFP=0.1301 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 424B, BPFP=0.1352 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 404B, BPFP=0.1288 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 464B, BPFP=0.1480 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 440B, BPFP=0.1403 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,528B, BPFP=0.0696 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12410851 553.13233418 + layer.0.v_cache 0.00001740 0.05311058 + layer.1.k_cache 0.01401279 40.45063875 + layer.1.v_cache 0.00000579 0.01913195 + layer.2.k_cache 0.00545398 7.12506041 + layer.2.v_cache 0.00001955 0.05596985 + layer.3.k_cache 0.07546502 24.72871991 + layer.3.v_cache 0.00001978 0.06997606 + layer.4.k_cache 0.00063206 1.91039385 + layer.4.v_cache 0.00005205 0.14498577 + layer.4.output 0.27729562 1109.26093294 + ------------------------------------------------------------------------------------- + TOTAL 0.12710919 493.67746188 + (elements=426,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 426496 +Total Bytes 5836 +BPFP 0.1095 bits/point +EBPFP 0.2189 equivalent bits/point +MSE 493.677462 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 493.6775 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 428B, BPFP=0.1173 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 468B, BPFP=0.1283 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 468B, BPFP=0.1283 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 416B, BPFP=0.1140 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 464B, BPFP=0.1272 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 428B, BPFP=0.1173 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 436B, BPFP=0.1195 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 432B, BPFP=0.1184 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 480B, BPFP=0.1316 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 464B, BPFP=0.1272 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,440B, BPFP=0.0564 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19197876 544.61283580 + layer.0.v_cache 0.00001449 0.05567596 + layer.1.k_cache 0.01466701 40.12189899 + layer.1.v_cache 0.00000519 0.01935751 + layer.2.k_cache 0.00720649 7.58006367 + layer.2.v_cache 0.00001785 0.05749416 + layer.3.k_cache 0.10851213 25.28865560 + layer.3.v_cache 0.00001996 0.07416872 + layer.4.k_cache 0.00062085 1.95957840 + layer.4.v_cache 0.00004770 0.14699223 + layer.4.output 0.23832821 982.12672306 + ------------------------------------------------------------------------------------- + TOTAL 0.11714047 440.87081073 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 5924 +BPFP 0.0955 bits/point +EBPFP 0.1910 equivalent bits/point +MSE 440.870811 +---------------------- -------------------------------------------------------- +Time: 0.272s Load: 0.003s, Pack+Encode: 0.128s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 440.8708 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 416B, BPFP=0.1275 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 444B, BPFP=0.1360 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 480B, BPFP=0.1471 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 404B, BPFP=0.1238 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 464B, BPFP=0.1422 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 396B, BPFP=0.1213 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 428B, BPFP=0.1311 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 412B, BPFP=0.1262 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 464B, BPFP=0.1422 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 448B, BPFP=0.1373 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,464B, BPFP=0.0641 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.143s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12665685 553.55966605 + layer.0.v_cache 0.00001436 0.05386019 + layer.1.k_cache 0.01497617 40.45255055 + layer.1.v_cache 0.00000537 0.01751678 + layer.2.k_cache 0.00986420 7.32712331 + layer.2.v_cache 0.00001802 0.05145123 + layer.3.k_cache 0.09846177 24.49591184 + layer.3.v_cache 0.00001923 0.06797732 + layer.4.k_cache 0.00062397 1.85856793 + layer.4.v_cache 0.00005014 0.14266371 + layer.4.output 0.26631049 1068.60092787 + ------------------------------------------------------------------------------------- + TOTAL 0.12440374 476.95492847 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 5820 +BPFP 0.1049 bits/point +EBPFP 0.2098 equivalent bits/point +MSE 476.954928 +---------------------- -------------------------------------------------------- +Time: 0.274s Load: 0.003s, Pack+Encode: 0.128s, Decode+Unpack: 0.143s +---------------------- -------------------------------------------------------- +💾 Converting with 476.9549 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 412B, BPFP=0.1341 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 420B, BPFP=0.1367 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 484B, BPFP=0.1576 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 396B, BPFP=0.1289 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 464B, BPFP=0.1510 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 396B, BPFP=0.1289 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 440B, BPFP=0.1432 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 428B, BPFP=0.1393 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 472B, BPFP=0.1536 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 432B, BPFP=0.1406 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,496B, BPFP=0.0696 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18531116 555.97159831 + layer.0.v_cache 0.00001874 0.05707821 + layer.1.k_cache 0.01459585 41.03561656 + layer.1.v_cache 0.00000674 0.02092027 + layer.2.k_cache 0.01042949 7.31788635 + layer.2.v_cache 0.00001841 0.05851215 + layer.3.k_cache 0.15260293 25.15558879 + layer.3.v_cache 0.00002390 0.07959348 + layer.4.k_cache 0.00061422 1.90573867 + layer.4.v_cache 0.00004881 0.15145933 + layer.4.output 0.28297247 1132.08816964 + ------------------------------------------------------------------------------------- + TOTAL 0.13791044 503.31595174 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 5840 +BPFP 0.1118 bits/point +EBPFP 0.2237 equivalent bits/point +MSE 503.315952 +---------------------- -------------------------------------------------------- +Time: 0.272s Load: 0.002s, Pack+Encode: 0.128s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 503.3160 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 408B, BPFP=0.1081 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 456B, BPFP=0.1208 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 436B, BPFP=0.1155 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 396B, BPFP=0.1049 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 416B, BPFP=0.1102 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 404B, BPFP=0.1070 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 412B, BPFP=0.1091 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 404B, BPFP=0.1070 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 440B, BPFP=0.1165 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 436B, BPFP=0.1155 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,324B, BPFP=0.0501 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.143s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12052289 512.75877251 + layer.0.v_cache 0.00001546 0.05346850 + layer.1.k_cache 0.01434235 37.70302569 + layer.1.v_cache 0.00000574 0.01854827 + layer.2.k_cache 0.00254285 6.92114258 + layer.2.v_cache 0.00001710 0.05517105 + layer.3.k_cache 0.03737745 23.53149828 + layer.3.v_cache 0.00001946 0.07201761 + layer.4.k_cache 0.00063766 1.85190362 + layer.4.v_cache 0.00005039 0.14668166 + layer.4.output 0.23032015 920.57581719 + ------------------------------------------------------------------------------------- + TOTAL 0.10516308 413.36135001 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 5532 +BPFP 0.0862 bits/point +EBPFP 0.1724 equivalent bits/point +MSE 413.361350 +---------------------- -------------------------------------------------------- +Time: 0.273s Load: 0.002s, Pack+Encode: 0.128s, Decode+Unpack: 0.143s +---------------------- -------------------------------------------------------- +💾 Converting with 413.3614 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1343 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 744B, BPFP=0.1571 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 788B, BPFP=0.1664 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 700B, BPFP=0.1478 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1554 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 684B, BPFP=0.1444 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1427 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 688B, BPFP=0.1453 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 800B, BPFP=0.1689 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 736B, BPFP=0.1554 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,588B, BPFP=0.0781 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11335979 533.00422297 + layer.0.v_cache 0.00001513 0.05476561 + layer.1.k_cache 0.01655370 38.34015058 + layer.1.v_cache 0.00000667 0.01863675 + layer.2.k_cache 0.00428703 6.80038823 + layer.2.v_cache 0.00001789 0.05360310 + layer.3.k_cache 0.04938591 23.93636343 + layer.3.v_cache 0.00002111 0.07042621 + layer.4.k_cache 0.00066117 1.82898650 + layer.4.v_cache 0.00004716 0.13089851 + layer.4.output 0.18372514 733.29180743 + ------------------------------------------------------------------------------------- + TOTAL 0.08649597 337.48712317 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 9776 +BPFP 0.1214 bits/point +EBPFP 0.2428 equivalent bits/point +MSE 337.487123 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 337.4871 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 404B, BPFP=0.1435 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 432B, BPFP=0.1534 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 484B, BPFP=0.1719 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 376B, BPFP=0.1335 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 468B, BPFP=0.1662 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 392B, BPFP=0.1392 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 444B, BPFP=0.1577 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 412B, BPFP=0.1463 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 468B, BPFP=0.1662 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 436B, BPFP=0.1548 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,756B, BPFP=0.0891 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14777968 560.01651278 + layer.0.v_cache 0.00001531 0.05725022 + layer.1.k_cache 0.01728138 42.15199419 + layer.1.v_cache 0.00000547 0.01987593 + layer.2.k_cache 0.00823456 7.67224468 + layer.2.v_cache 0.00001932 0.06167182 + layer.3.k_cache 0.16770924 25.39528032 + layer.3.v_cache 0.00002046 0.07734023 + layer.4.k_cache 0.00061569 1.95366859 + layer.4.v_cache 0.00005592 0.16808489 + layer.4.output 0.30867824 1233.98701299 + ------------------------------------------------------------------------------------- + TOTAL 0.14720498 545.61664792 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 6072 +BPFP 0.1268 bits/point +EBPFP 0.2537 equivalent bits/point +MSE 545.616648 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 545.6166 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 628B, BPFP=0.1258 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 728B, BPFP=0.1458 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1611 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 692B, BPFP=0.1386 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 724B, BPFP=0.1450 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 676B, BPFP=0.1354 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1378 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 716B, BPFP=0.1434 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 792B, BPFP=0.1587 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 752B, BPFP=0.1506 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,740B, BPFP=0.0784 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14408422 538.70412660 + layer.0.v_cache 0.00002097 0.05543521 + layer.1.k_cache 0.05735342 38.03558819 + layer.1.v_cache 0.00000575 0.01920683 + layer.2.k_cache 0.01678689 6.91169230 + layer.2.v_cache 0.00001894 0.05535265 + layer.3.k_cache 0.03029843 23.95237380 + layer.3.v_cache 0.00002125 0.07391728 + layer.4.k_cache 0.00062582 1.83640661 + layer.4.v_cache 0.00005050 0.14710371 + layer.4.output 0.17433185 694.78594322 + ------------------------------------------------------------------------------------- + TOTAL 0.08644642 321.95840034 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9940 +BPFP 0.1171 bits/point +EBPFP 0.2343 equivalent bits/point +MSE 321.958400 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 321.9584 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 404B, BPFP=0.1435 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 428B, BPFP=0.1520 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 480B, BPFP=0.1705 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 388B, BPFP=0.1378 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 456B, BPFP=0.1619 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 400B, BPFP=0.1420 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 436B, BPFP=0.1548 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 404B, BPFP=0.1435 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 476B, BPFP=0.1690 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 452B, BPFP=0.1605 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,736B, BPFP=0.0881 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.151s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13243668 559.02081854 + layer.0.v_cache 0.00001576 0.05640999 + layer.1.k_cache 0.01981235 42.21400036 + layer.1.v_cache 0.00000577 0.01967055 + layer.2.k_cache 0.01756107 7.58681280 + layer.2.v_cache 0.00001728 0.05685773 + layer.3.k_cache 0.04920139 25.52294367 + layer.3.v_cache 0.00001837 0.06866617 + layer.4.k_cache 0.00061593 1.97279323 + layer.4.v_cache 0.00004907 0.15054351 + layer.4.output 0.30854983 1233.87185471 + ------------------------------------------------------------------------------------- + TOTAL 0.13997544 545.51602938 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 6060 +BPFP 0.1266 bits/point +EBPFP 0.2532 equivalent bits/point +MSE 545.516029 +---------------------- -------------------------------------------------------- +Time: 0.283s Load: 0.003s, Pack+Encode: 0.129s, Decode+Unpack: 0.151s +---------------------- -------------------------------------------------------- +💾 Converting with 545.5160 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 416B, BPFP=0.1300 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 428B, BPFP=0.1338 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 484B, BPFP=0.1512 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 392B, BPFP=0.1225 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 472B, BPFP=0.1475 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 404B, BPFP=0.1263 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 428B, BPFP=0.1338 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 408B, BPFP=0.1275 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 464B, BPFP=0.1450 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 440B, BPFP=0.1375 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,500B, BPFP=0.0670 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.143s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20666794 545.76363281 + layer.0.v_cache 0.00002565 0.05589179 + layer.1.k_cache 0.01518770 40.01627197 + layer.1.v_cache 0.00000581 0.01876274 + layer.2.k_cache 0.01031687 7.10325134 + layer.2.v_cache 0.00001873 0.05671113 + layer.3.k_cache 0.09638391 24.51259277 + layer.3.v_cache 0.00001967 0.07347186 + layer.4.k_cache 0.00062172 1.91188950 + layer.4.v_cache 0.00004827 0.15117727 + layer.4.output 0.27162739 1085.33812500 + ------------------------------------------------------------------------------------- + TOTAL 0.13121694 483.35473695 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 5836 +BPFP 0.1073 bits/point +EBPFP 0.2146 equivalent bits/point +MSE 483.354737 +---------------------- -------------------------------------------------------- +Time: 0.275s Load: 0.003s, Pack+Encode: 0.129s, Decode+Unpack: 0.143s +---------------------- -------------------------------------------------------- +💾 Converting with 483.3547 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1507 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 724B, BPFP=0.1664 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 784B, BPFP=0.1801 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 660B, BPFP=0.1517 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 744B, BPFP=0.1710 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 672B, BPFP=0.1544 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 704B, BPFP=0.1618 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 684B, BPFP=0.1572 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 776B, BPFP=0.1783 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 732B, BPFP=0.1682 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,512B, BPFP=0.0825 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13431002 516.69726562 + layer.0.v_cache 0.00001720 0.05425506 + layer.1.k_cache 0.01179634 38.35248880 + layer.1.v_cache 0.00000537 0.01781382 + layer.2.k_cache 0.00798416 6.99997666 + layer.2.v_cache 0.00001783 0.05447305 + layer.3.k_cache 0.05795463 23.82229614 + layer.3.v_cache 0.00001974 0.07240845 + layer.4.k_cache 0.00062059 1.77671275 + layer.4.v_cache 0.00005204 0.14404782 + layer.4.output 0.19989206 797.86488971 + ------------------------------------------------------------------------------------- + TOTAL 0.09482484 363.12035095 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 9648 +BPFP 0.1304 bits/point +EBPFP 0.2608 equivalent bits/point +MSE 363.120351 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.003s, Pack+Encode: 0.164s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 363.1204 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 412B, BPFP=0.1288 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 440B, BPFP=0.1375 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 488B, BPFP=0.1525 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 388B, BPFP=0.1212 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 460B, BPFP=0.1437 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 388B, BPFP=0.1212 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 440B, BPFP=0.1375 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 408B, BPFP=0.1275 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 460B, BPFP=0.1437 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 440B, BPFP=0.1375 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,508B, BPFP=0.0673 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15467889 549.16750000 + layer.0.v_cache 0.00001741 0.05387859 + layer.1.k_cache 0.01620592 39.94215820 + layer.1.v_cache 0.00000548 0.01830414 + layer.2.k_cache 0.00723796 7.05623291 + layer.2.v_cache 0.00001737 0.05312834 + layer.3.k_cache 0.06661982 24.64008301 + layer.3.v_cache 0.00002040 0.06936862 + layer.4.k_cache 0.00058776 1.83329559 + layer.4.v_cache 0.00004656 0.14654072 + layer.4.output 0.27162074 1085.27089286 + ------------------------------------------------------------------------------------- + TOTAL 0.12628134 483.52216118 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 5832 +BPFP 0.1072 bits/point +EBPFP 0.2144 equivalent bits/point +MSE 483.522161 +---------------------- -------------------------------------------------------- +Time: 0.273s Load: 0.003s, Pack+Encode: 0.128s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 483.5222 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 416B, BPFP=0.1275 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 440B, BPFP=0.1348 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 484B, BPFP=0.1483 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 400B, BPFP=0.1225 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 464B, BPFP=0.1422 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 400B, BPFP=0.1225 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 436B, BPFP=0.1336 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 416B, BPFP=0.1275 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 464B, BPFP=0.1422 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 448B, BPFP=0.1373 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,444B, BPFP=0.0632 +⌛️ [2/4] FRONTEND: Frontend time: 0.127s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11261671 551.38239124 + layer.0.v_cache 0.00001829 0.05242501 + layer.1.k_cache 0.01190703 40.42711684 + layer.1.v_cache 0.00000528 0.01737813 + layer.2.k_cache 0.00973116 7.18303546 + layer.2.v_cache 0.00001668 0.04900335 + layer.3.k_cache 0.04200072 24.72520058 + layer.3.v_cache 0.00001874 0.06749121 + layer.4.k_cache 0.00062040 1.83087861 + layer.4.v_cache 0.00004776 0.13934587 + layer.4.output 0.26626884 1068.73363095 + ------------------------------------------------------------------------------------- + TOTAL 0.12005086 476.88292253 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 5812 +BPFP 0.1047 bits/point +EBPFP 0.2095 equivalent bits/point +MSE 476.882923 +---------------------- -------------------------------------------------------- +Time: 0.272s Load: 0.002s, Pack+Encode: 0.127s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 476.8829 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 424B, BPFP=0.1274 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 428B, BPFP=0.1286 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 488B, BPFP=0.1466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 392B, BPFP=0.1178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 476B, BPFP=0.1430 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 396B, BPFP=0.1190 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 436B, BPFP=0.1310 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 400B, BPFP=0.1202 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 468B, BPFP=0.1406 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 444B, BPFP=0.1334 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,508B, BPFP=0.0647 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.140s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12932784 550.94850511 + layer.0.v_cache 0.00001491 0.05672541 + layer.1.k_cache 0.01488123 40.23507925 + layer.1.v_cache 0.00000563 0.01941167 + layer.2.k_cache 0.00939683 7.34994976 + layer.2.v_cache 0.00001981 0.05726262 + layer.3.k_cache 0.07051506 24.94631489 + layer.3.v_cache 0.00001927 0.07282568 + layer.4.k_cache 0.00064413 1.97210371 + layer.4.v_cache 0.00005939 0.15306443 + layer.4.output 0.26127321 1050.17384959 + ------------------------------------------------------------------------------------- + TOTAL 0.12081156 469.23695233 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 5860 +BPFP 0.1036 bits/point +EBPFP 0.2072 equivalent bits/point +MSE 469.236952 +---------------------- -------------------------------------------------------- +Time: 0.271s Load: 0.002s, Pack+Encode: 0.128s, Decode+Unpack: 0.140s +---------------------- -------------------------------------------------------- +💾 Converting with 469.2370 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 424B, BPFP=0.1274 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 428B, BPFP=0.1286 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 488B, BPFP=0.1466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 392B, BPFP=0.1178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 472B, BPFP=0.1418 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 392B, BPFP=0.1178 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 432B, BPFP=0.1298 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 412B, BPFP=0.1238 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 472B, BPFP=0.1418 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 444B, BPFP=0.1334 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,500B, BPFP=0.0644 +⌛️ [2/4] FRONTEND: Frontend time: 0.127s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.141s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13190371 550.71702224 + layer.0.v_cache 0.00001552 0.05766635 + layer.1.k_cache 0.01668521 40.12798603 + layer.1.v_cache 0.00000581 0.02055987 + layer.2.k_cache 0.00490881 7.36174715 + layer.2.v_cache 0.00001873 0.05997540 + layer.3.k_cache 0.11824917 24.88290640 + layer.3.v_cache 0.00002115 0.07767891 + layer.4.k_cache 0.00061206 1.91133499 + layer.4.v_cache 0.00004844 0.14662990 + layer.4.output 0.26122982 1051.14431662 + ------------------------------------------------------------------------------------- + TOTAL 0.12359279 469.61021903 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 5856 +BPFP 0.1035 bits/point +EBPFP 0.2070 equivalent bits/point +MSE 469.610219 +---------------------- -------------------------------------------------------- +Time: 0.270s Load: 0.002s, Pack+Encode: 0.127s, Decode+Unpack: 0.141s +---------------------- -------------------------------------------------------- +💾 Converting with 469.6102 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 688B, BPFP=0.1168 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 836B, BPFP=0.1420 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 844B, BPFP=0.1433 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 780B, BPFP=0.1325 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 836B, BPFP=0.1420 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 764B, BPFP=0.1298 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 860B, BPFP=0.1461 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 888B, BPFP=0.1508 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,152B, BPFP=0.0522 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10326390 508.12572181 + layer.0.v_cache 0.00001541 0.04855314 + layer.1.k_cache 0.04976616 36.79274053 + layer.1.v_cache 0.00000534 0.01615647 + layer.2.k_cache 0.01168035 6.52035058 + layer.2.v_cache 0.00001798 0.04941583 + layer.3.k_cache 0.02907468 23.28783384 + layer.3.v_cache 0.00001946 0.06436598 + layer.4.k_cache 0.00073861 1.63778836 + layer.4.v_cache 0.00005077 0.12996998 + layer.4.output 0.14788401 590.52620342 + ------------------------------------------------------------------------------------- + TOTAL 0.07234240 277.07978355 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 10152 +BPFP 0.1014 bits/point +EBPFP 0.2028 equivalent bits/point +MSE 277.079784 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.003s, Pack+Encode: 0.163s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 277.0798 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 64, 128) +Output shape: (1, 64, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.output: torch.Size([1, 64, 3584]) -> torch.Size([1, 1, 64, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 300B, BPFP=0.0732 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 336B, BPFP=0.0820 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 388B, BPFP=0.0947 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 300B, BPFP=0.0732 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 372B, BPFP=0.0908 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 308B, BPFP=0.0752 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 324B, BPFP=0.0791 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 308B, BPFP=0.0752 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 384B, BPFP=0.0938 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 364B, BPFP=0.0889 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 644B, BPFP=0.0225 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09722494 494.41949463 + layer.0.v_cache 0.00001391 0.04473723 + layer.1.k_cache 0.01806639 34.09102631 + layer.1.v_cache 0.00000557 0.01606327 + layer.2.k_cache 0.00423524 5.95113564 + layer.2.v_cache 0.00001868 0.04844291 + layer.3.k_cache 0.03759564 21.71732521 + layer.3.v_cache 0.00001825 0.05824482 + layer.4.k_cache 0.00061753 1.44283402 + layer.4.v_cache 0.00004894 0.12470674 + layer.4.output 0.21408452 847.46902902 + ------------------------------------------------------------------------------------- + TOTAL 0.09743746 381.77630670 + (elements=557,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 557056 +Total Bytes 4028 +BPFP 0.0578 bits/point +EBPFP 0.1157 equivalent bits/point +MSE 381.776307 +---------------------- -------------------------------------------------------- +Time: 0.273s Load: 0.004s, Pack+Encode: 0.128s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 381.7763 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 704B, BPFP=0.1078 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 804B, BPFP=0.1232 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 836B, BPFP=0.1281 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 740B, BPFP=0.1134 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 772B, BPFP=0.1183 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 772B, BPFP=0.1183 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 768B, BPFP=0.1176 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 768B, BPFP=0.1176 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 872B, BPFP=0.1336 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 840B, BPFP=0.1287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,252B, BPFP=0.0493 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11648470 510.95580576 + layer.0.v_cache 0.00001454 0.04792946 + layer.1.k_cache 0.02721847 37.01560106 + layer.1.v_cache 0.00000531 0.01646601 + layer.2.k_cache 0.01045542 6.52436200 + layer.2.v_cache 0.00001699 0.04690823 + layer.3.k_cache 0.01744144 23.18812950 + layer.3.v_cache 0.00001839 0.05905499 + layer.4.k_cache 0.00076342 1.67473079 + layer.4.v_cache 0.00004795 0.12696762 + layer.4.output 11.17676328 527.70308123 + ------------------------------------------------------------------------------------- + TOTAL 4.61234174 251.38691318 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 10128 +BPFP 0.0913 bits/point +EBPFP 0.1825 equivalent bits/point +MSE 251.386913 +---------------------- -------------------------------------------------------- +Time: 0.377s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 251.3869 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 628B, BPFP=0.1227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 772B, BPFP=0.1508 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 808B, BPFP=0.1578 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 680B, BPFP=0.1328 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 676B, BPFP=0.1320 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 672B, BPFP=0.1313 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 704B, BPFP=0.1375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 804B, BPFP=0.1570 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 784B, BPFP=0.1531 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,364B, BPFP=0.0660 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11433966 534.09687500 + layer.0.v_cache 0.00001437 0.04769918 + layer.1.k_cache 0.05315235 37.79661255 + layer.1.v_cache 0.00000524 0.01581829 + layer.2.k_cache 0.00678293 6.69766235 + layer.2.v_cache 0.00001704 0.04737816 + layer.3.k_cache 0.01951189 23.62744751 + layer.3.v_cache 0.00001758 0.05854605 + layer.4.k_cache 0.00076624 1.74403362 + layer.4.v_cache 0.00004856 0.12729701 + layer.4.output 0.16992235 677.53867187 + ------------------------------------------------------------------------------------- + TOTAL 0.08141837 314.53118076 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 9608 +BPFP 0.1104 bits/point +EBPFP 0.2208 equivalent bits/point +MSE 314.531181 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 314.5312 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 680B, BPFP=0.1181 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 852B, BPFP=0.1479 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 764B, BPFP=0.1326 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 772B, BPFP=0.1340 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 788B, BPFP=0.1368 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 744B, BPFP=0.1292 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 796B, BPFP=0.1382 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 844B, BPFP=0.1465 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 840B, BPFP=0.1458 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,324B, BPFP=0.0576 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11570581 521.02222222 + layer.0.v_cache 0.00001517 0.04997402 + layer.1.k_cache 0.03236232 36.84955512 + layer.1.v_cache 0.00000662 0.01650927 + layer.2.k_cache 0.00641356 6.56935357 + layer.2.v_cache 0.00001676 0.04777186 + layer.3.k_cache 0.02648412 23.42363281 + layer.3.v_cache 0.00001896 0.06282242 + layer.4.k_cache 0.00074524 1.72453071 + layer.4.v_cache 0.00004843 0.12862214 + layer.4.output 0.15113260 604.99523810 + ------------------------------------------------------------------------------------- + TOTAL 0.07292619 283.81539181 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 10224 +BPFP 0.1044 bits/point +EBPFP 0.2088 equivalent bits/point +MSE 283.815392 +---------------------- -------------------------------------------------------- +Time: 0.377s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 283.8154 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 114, 128) +Output shape: (1, 114, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.output: torch.Size([1, 114, 3584]) -> torch.Size([1, 1, 114, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 752B, BPFP=0.1031 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 852B, BPFP=0.1168 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 892B, BPFP=0.1223 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 804B, BPFP=0.1102 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 844B, BPFP=0.1157 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 800B, BPFP=0.1096 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 828B, BPFP=0.1135 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 808B, BPFP=0.1107 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 900B, BPFP=0.1234 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 844B, BPFP=0.1157 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,368B, BPFP=0.0464 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10081583 518.20278235 + layer.0.v_cache 0.00001633 0.04822261 + layer.1.k_cache 0.08473034 36.55794699 + layer.1.v_cache 0.00000563 0.01649896 + layer.2.k_cache 0.00792928 6.40477070 + layer.2.v_cache 0.00001730 0.04710083 + layer.3.k_cache 0.01019665 23.25952363 + layer.3.v_cache 0.00001810 0.05590405 + layer.4.k_cache 0.00077074 1.65987383 + layer.4.v_cache 0.00005022 0.12764807 + layer.4.output 10.00037814 471.60968828 + ------------------------------------------------------------------------------------- + TOTAL 4.12983514 228.68518176 + (elements=992,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 992256 +Total Bytes 10692 +BPFP 0.0862 bits/point +EBPFP 0.1724 equivalent bits/point +MSE 228.685182 +---------------------- -------------------------------------------------------- +Time: 0.380s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 228.6852 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 98, 128) +Output shape: (1, 98, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.output: torch.Size([1, 98, 3584]) -> torch.Size([1, 1, 98, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 688B, BPFP=0.1097 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 856B, BPFP=0.1365 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 732B, BPFP=0.1167 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 780B, BPFP=0.1244 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 744B, BPFP=0.1186 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 748B, BPFP=0.1193 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 784B, BPFP=0.1250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 848B, BPFP=0.1352 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 828B, BPFP=0.1320 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,000B, BPFP=0.0456 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11008905 520.43136161 + layer.0.v_cache 0.00001718 0.04790959 + layer.1.k_cache 0.06240322 36.90036571 + layer.1.v_cache 0.00000548 0.01624389 + layer.2.k_cache 0.00473963 6.62572246 + layer.2.v_cache 0.00001718 0.04734979 + layer.3.k_cache 0.01141277 23.20312500 + layer.3.v_cache 0.00001769 0.05585801 + layer.4.k_cache 0.00076610 1.72884976 + layer.4.v_cache 0.00004814 0.12285301 + layer.4.output 0.02426045 570.03065780 + ------------------------------------------------------------------------------------- + TOTAL 0.02113762 269.37613197 + (elements=852,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 852992 +Total Bytes 9820 +BPFP 0.0921 bits/point +EBPFP 0.1842 equivalent bits/point +MSE 269.376132 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 269.3761 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 684B, BPFP=0.1162 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 808B, BPFP=0.1372 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 848B, BPFP=0.1440 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 708B, BPFP=0.1202 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 796B, BPFP=0.1352 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 768B, BPFP=0.1304 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 748B, BPFP=0.1270 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 776B, BPFP=0.1318 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 852B, BPFP=0.1447 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 800B, BPFP=0.1359 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,136B, BPFP=0.0518 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08919133 512.45872962 + layer.0.v_cache 0.00001414 0.04725779 + layer.1.k_cache 0.08591949 36.75571608 + layer.1.v_cache 0.00000528 0.01632720 + layer.2.k_cache 0.00389855 6.57186492 + layer.2.v_cache 0.00001738 0.04929766 + layer.3.k_cache 0.07679920 23.47868546 + layer.3.v_cache 0.00002028 0.06485373 + layer.4.k_cache 0.00072407 1.70547933 + layer.4.v_cache 0.00004779 0.13192540 + layer.4.output 0.14786680 590.06871118 + ------------------------------------------------------------------------------------- + TOTAL 0.07598265 277.16241855 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 9924 +BPFP 0.0991 bits/point +EBPFP 0.1983 equivalent bits/point +MSE 277.162419 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 277.1624 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 740B, BPFP=0.1081 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 864B, BPFP=0.1262 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 868B, BPFP=0.1268 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 800B, BPFP=0.1168 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 824B, BPFP=0.1203 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 796B, BPFP=0.1162 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 824B, BPFP=0.1203 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 840B, BPFP=0.1227 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 896B, BPFP=0.1308 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 840B, BPFP=0.1227 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,472B, BPFP=0.0516 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11353527 519.06184287 + layer.0.v_cache 0.00001674 0.04981886 + layer.1.k_cache 0.04276301 37.15376752 + layer.1.v_cache 0.00000542 0.01708422 + layer.2.k_cache 0.00247095 6.64439492 + layer.2.v_cache 0.00001885 0.05082015 + layer.3.k_cache 0.02722364 23.54302807 + layer.3.v_cache 0.00001942 0.06156471 + layer.4.k_cache 0.00068256 1.73819312 + layer.4.v_cache 0.00004843 0.13445833 + layer.4.output 10.65454330 503.30511515 + ------------------------------------------------------------------------------------- + TOTAL 4.39815220 241.85828111 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 10764 +BPFP 0.0925 bits/point +EBPFP 0.1849 equivalent bits/point +MSE 241.858281 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.006s, Pack+Encode: 0.165s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 241.8583 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 644B, BPFP=0.1198 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 800B, BPFP=0.1488 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1518 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 720B, BPFP=0.1339 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1362 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 688B, BPFP=0.1280 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 700B, BPFP=0.1302 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 748B, BPFP=0.1391 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 824B, BPFP=0.1533 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 804B, BPFP=0.1496 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,324B, BPFP=0.0618 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12550225 530.67447917 + layer.0.v_cache 0.00001468 0.04774639 + layer.1.k_cache 0.05234316 37.31785366 + layer.1.v_cache 0.00000508 0.01558863 + layer.2.k_cache 0.00733058 6.58449881 + layer.2.v_cache 0.00001689 0.04794191 + layer.3.k_cache 0.01592888 23.40318952 + layer.3.v_cache 0.00001847 0.06052275 + layer.4.k_cache 0.00069460 1.65755553 + layer.4.v_cache 0.00005205 0.12877330 + layer.4.output 0.16184913 647.26403061 + ------------------------------------------------------------------------------------- + TOTAL 0.07852062 301.81096259 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 9800 +BPFP 0.1072 bits/point +EBPFP 0.2145 equivalent bits/point +MSE 301.810963 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.003s, Pack+Encode: 0.165s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 301.8110 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1274 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 780B, BPFP=0.1562 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1611 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1378 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 720B, BPFP=0.1442 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 692B, BPFP=0.1386 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1378 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 712B, BPFP=0.1426 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 796B, BPFP=0.1595 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1514 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,656B, BPFP=0.0760 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11879376 537.64738582 + layer.0.v_cache 0.00001438 0.04968809 + layer.1.k_cache 0.05518859 38.17611303 + layer.1.v_cache 0.00000514 0.01659030 + layer.2.k_cache 0.00243519 6.84577042 + layer.2.v_cache 0.00001742 0.04739804 + layer.3.k_cache 0.07959401 23.75864821 + layer.3.v_cache 0.00001812 0.06051369 + layer.4.k_cache 0.00069358 1.79044557 + layer.4.v_cache 0.00004620 0.12678194 + layer.4.output 0.17426339 694.83430632 + ------------------------------------------------------------------------------------- + TOTAL 0.08686177 321.90349878 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 9928 +BPFP 0.1170 bits/point +EBPFP 0.2340 equivalent bits/point +MSE 321.903499 +---------------------- -------------------------------------------------------- +Time: 0.377s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 321.9035 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1204 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1456 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 796B, BPFP=0.1517 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1311 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1387 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 668B, BPFP=0.1273 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 724B, BPFP=0.1380 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 808B, BPFP=0.1540 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 772B, BPFP=0.1471 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,304B, BPFP=0.0627 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11325797 531.54697027 + layer.0.v_cache 0.00001565 0.04715259 + layer.1.k_cache 0.07651102 37.15681867 + layer.1.v_cache 0.00000530 0.01656970 + layer.2.k_cache 0.00872486 6.57001923 + layer.2.v_cache 0.00001696 0.04653436 + layer.3.k_cache 0.12876377 23.26673554 + layer.3.v_cache 0.00001905 0.05899504 + layer.4.k_cache 0.00069190 1.69609200 + layer.4.v_cache 0.00004968 0.13066569 + layer.4.output 0.16578155 662.23328615 + ------------------------------------------------------------------------------------- + TOTAL 0.08756041 308.00997389 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 9576 +BPFP 0.1073 bits/point +EBPFP 0.2147 equivalent bits/point +MSE 308.009974 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.005s, Pack+Encode: 0.164s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 308.0100 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 752B, BPFP=0.1526 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1623 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 668B, BPFP=0.1356 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 720B, BPFP=0.1461 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 664B, BPFP=0.1347 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1380 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 688B, BPFP=0.1396 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 792B, BPFP=0.1607 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 764B, BPFP=0.1550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,708B, BPFP=0.0785 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12835005 539.96423498 + layer.0.v_cache 0.00001616 0.04613803 + layer.1.k_cache 0.05672838 38.26433137 + layer.1.v_cache 0.00000535 0.01653669 + layer.2.k_cache 0.00240846 6.86904551 + layer.2.v_cache 0.00001644 0.04580960 + layer.3.k_cache 0.04570918 23.62219872 + layer.3.v_cache 0.00001733 0.05732446 + layer.4.k_cache 0.00072090 1.71938136 + layer.4.v_cache 0.00004908 0.12916653 + layer.4.output 0.18444509 705.68077458 + ------------------------------------------------------------------------------------- + TOTAL 0.08971394 326.49997584 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9868 +BPFP 0.1178 bits/point +EBPFP 0.2356 equivalent bits/point +MSE 326.499976 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 326.5000 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1444 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 792B, BPFP=0.1743 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 784B, BPFP=0.1725 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 692B, BPFP=0.1523 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1620 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 716B, BPFP=0.1576 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1514 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 736B, BPFP=0.1620 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 792B, BPFP=0.1743 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 764B, BPFP=0.1681 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,632B, BPFP=0.0827 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10085608 531.33208627 + layer.0.v_cache 0.00001554 0.04498956 + layer.1.k_cache 0.06589249 38.00387186 + layer.1.v_cache 0.00000499 0.01556303 + layer.2.k_cache 0.00422055 6.80185581 + layer.2.v_cache 0.00001579 0.04337126 + layer.3.k_cache 0.05337993 23.46444831 + layer.3.v_cache 0.00001747 0.05542100 + layer.4.k_cache 0.00075412 1.73488477 + layer.4.v_cache 0.00004712 0.12413148 + layer.4.output 0.19135183 765.20472837 + ------------------------------------------------------------------------------------- + TOTAL 0.09203923 350.47374835 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 9988 +BPFP 0.1293 bits/point +EBPFP 0.2586 equivalent bits/point +MSE 350.473748 +---------------------- -------------------------------------------------------- +Time: 0.375s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 350.4737 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 644B, BPFP=0.1198 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1429 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1525 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 700B, BPFP=0.1302 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 740B, BPFP=0.1376 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 704B, BPFP=0.1310 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 704B, BPFP=0.1310 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 720B, BPFP=0.1339 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 824B, BPFP=0.1533 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 764B, BPFP=0.1421 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,188B, BPFP=0.0581 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09057366 526.18903460 + layer.0.v_cache 0.00001789 0.04482018 + layer.1.k_cache 0.03614600 37.26095145 + layer.1.v_cache 0.00000481 0.01506416 + layer.2.k_cache 0.00554357 6.68783860 + layer.2.v_cache 0.00001714 0.04510281 + layer.3.k_cache 0.02742965 23.43550328 + layer.3.v_cache 0.00001729 0.05618011 + layer.4.k_cache 0.00093027 1.72472763 + layer.4.v_cache 0.00004906 0.12633299 + layer.4.output 0.16441821 647.32658376 + ------------------------------------------------------------------------------------- + TOTAL 0.07715628 301.58068483 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 9576 +BPFP 0.1048 bits/point +EBPFP 0.2096 equivalent bits/point +MSE 301.580685 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 301.5807 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1542 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 796B, BPFP=0.1615 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1396 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 724B, BPFP=0.1469 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 688B, BPFP=0.1396 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 700B, BPFP=0.1420 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 808B, BPFP=0.1640 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 764B, BPFP=0.1550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,760B, BPFP=0.0800 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09325875 539.67431006 + layer.0.v_cache 0.00001406 0.04666631 + layer.1.k_cache 0.05775181 38.25054535 + layer.1.v_cache 0.00000524 0.01607856 + layer.2.k_cache 0.00413042 6.81782036 + layer.2.v_cache 0.00001743 0.04696469 + layer.3.k_cache 0.04387648 23.85899135 + layer.3.v_cache 0.00002497 0.06385048 + layer.4.k_cache 0.00071512 1.74765768 + layer.4.v_cache 0.00005147 0.13351277 + layer.4.output 0.17652170 704.80525278 + ------------------------------------------------------------------------------------- + TOTAL 0.08444104 326.13489218 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 10004 +BPFP 0.1194 bits/point +EBPFP 0.2388 equivalent bits/point +MSE 326.134892 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 326.1349 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 676B, BPFP=0.1200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 840B, BPFP=0.1491 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 824B, BPFP=0.1463 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 768B, BPFP=0.1364 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 784B, BPFP=0.1392 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 772B, BPFP=0.1371 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 744B, BPFP=0.1321 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 772B, BPFP=0.1371 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 844B, BPFP=0.1499 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 824B, BPFP=0.1463 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,376B, BPFP=0.0603 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212358 525.41566051 + layer.0.v_cache 0.00001682 0.04116516 + layer.1.k_cache 0.05918076 37.05373313 + layer.1.v_cache 0.00000473 0.01349164 + layer.2.k_cache 0.00260097 6.52148715 + layer.2.v_cache 0.00001544 0.03962451 + layer.3.k_cache 0.02838851 23.13432173 + layer.3.v_cache 0.00001657 0.05049480 + layer.4.k_cache 0.00106615 1.65369173 + layer.4.v_cache 0.00004456 0.11250854 + layer.4.output 0.17173813 619.16355519 + ------------------------------------------------------------------------------------- + TOTAL 0.08091912 289.89300384 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 10224 +BPFP 0.1068 bits/point +EBPFP 0.2136 equivalent bits/point +MSE 289.893004 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 289.8930 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1186 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 840B, BPFP=0.1491 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1456 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 736B, BPFP=0.1307 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 772B, BPFP=0.1371 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 736B, BPFP=0.1307 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 784B, BPFP=0.1392 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 852B, BPFP=0.1513 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 800B, BPFP=0.1420 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,400B, BPFP=0.0609 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10731998 526.61452415 + layer.0.v_cache 0.00001618 0.04747500 + layer.1.k_cache 0.07375650 37.01947576 + layer.1.v_cache 0.00000540 0.01646442 + layer.2.k_cache 0.00246326 6.64033786 + layer.2.v_cache 0.00001746 0.04760018 + layer.3.k_cache 0.01642703 23.39034757 + layer.3.v_cache 0.00001778 0.05669509 + layer.4.k_cache 0.00066607 1.65654581 + layer.4.v_cache 0.00004728 0.12409286 + layer.4.output 0.15452130 618.86749188 + ------------------------------------------------------------------------------------- + TOTAL 0.07543447 289.86388246 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 10176 +BPFP 0.1063 bits/point +EBPFP 0.2126 equivalent bits/point +MSE 289.863882 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.003s, Pack+Encode: 0.165s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 289.8639 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 99, 128) +Output shape: (1, 99, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.output: torch.Size([1, 99, 3584]) -> torch.Size([1, 1, 99, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 692B, BPFP=0.1092 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 856B, BPFP=0.1351 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 832B, BPFP=0.1313 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 800B, BPFP=0.1263 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1212 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 796B, BPFP=0.1256 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 768B, BPFP=0.1212 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 752B, BPFP=0.1187 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 864B, BPFP=0.1364 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 884B, BPFP=0.1395 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,176B, BPFP=0.0491 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09061131 520.74881629 + layer.0.v_cache 0.00001461 0.04795553 + layer.1.k_cache 0.05062923 37.19800051 + layer.1.v_cache 0.00000709 0.01694268 + layer.2.k_cache 0.00392854 6.68577005 + layer.2.v_cache 0.00001793 0.04933638 + layer.3.k_cache 0.05082532 23.50910472 + layer.3.v_cache 0.00001820 0.05989470 + layer.4.k_cache 0.00080398 1.72758129 + layer.4.v_cache 0.00005053 0.12079530 + layer.4.output 0.02403493 564.07738095 + ------------------------------------------------------------------------------------- + TOTAL 0.02147948 266.98269789 + (elements=861,696) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 861696 +Total Bytes 10188 +BPFP 0.0946 bits/point +EBPFP 0.1892 equivalent bits/point +MSE 266.982698 +---------------------- -------------------------------------------------------- +Time: 0.375s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 266.9827 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1444 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 748B, BPFP=0.1646 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 784B, BPFP=0.1725 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 692B, BPFP=0.1523 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 748B, BPFP=0.1646 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 720B, BPFP=0.1585 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1523 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 732B, BPFP=0.1611 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 780B, BPFP=0.1717 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 768B, BPFP=0.1690 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,700B, BPFP=0.0849 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11136492 525.65713028 + layer.0.v_cache 0.00001955 0.05080723 + layer.1.k_cache 0.06280328 38.18217705 + layer.1.v_cache 0.00000544 0.01793754 + layer.2.k_cache 0.00235174 6.78739908 + layer.2.v_cache 0.00001859 0.05191535 + layer.3.k_cache 0.03326307 23.66171050 + layer.3.v_cache 0.00001926 0.06529423 + layer.4.k_cache 0.00064473 1.72684167 + layer.4.v_cache 0.00005254 0.13433541 + layer.4.output 0.19494371 765.66769366 + ------------------------------------------------------------------------------------- + TOTAL 0.09265583 350.35349435 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 10020 +BPFP 0.1297 bits/point +EBPFP 0.2594 equivalent bits/point +MSE 350.353494 +---------------------- -------------------------------------------------------- +Time: 0.356s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 350.3535 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1351 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 748B, BPFP=0.1579 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 792B, BPFP=0.1672 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1453 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1554 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 692B, BPFP=0.1461 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1427 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 712B, BPFP=0.1503 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 780B, BPFP=0.1647 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 764B, BPFP=0.1613 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,644B, BPFP=0.0798 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08829598 530.83393159 + layer.0.v_cache 0.00001455 0.04919349 + layer.1.k_cache 0.05950350 38.24370183 + layer.1.v_cache 0.00000529 0.01636856 + layer.2.k_cache 0.00411868 6.70466366 + layer.2.v_cache 0.00001680 0.04734799 + layer.3.k_cache 0.04877868 23.73620111 + layer.3.v_cache 0.00001896 0.06271704 + layer.4.k_cache 0.00063741 1.72118295 + layer.4.v_cache 0.00004764 0.12948708 + layer.4.output 0.18365649 733.57028234 + ------------------------------------------------------------------------------------- + TOTAL 0.08747253 337.44333951 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 9872 +BPFP 0.1226 bits/point +EBPFP 0.2452 equivalent bits/point +MSE 337.443340 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 337.4433 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 740B, BPFP=0.1502 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1631 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 664B, BPFP=0.1347 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1453 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 680B, BPFP=0.1380 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1380 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 696B, BPFP=0.1412 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1591 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 768B, BPFP=0.1558 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,700B, BPFP=0.0783 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11195674 541.49523133 + layer.0.v_cache 0.00001346 0.04614767 + layer.1.k_cache 0.03755230 38.36866186 + layer.1.v_cache 0.00000535 0.01650867 + layer.2.k_cache 0.00239600 6.85103538 + layer.2.v_cache 0.00001753 0.04836031 + layer.3.k_cache 0.02867049 23.60393415 + layer.3.v_cache 0.00001751 0.05703076 + layer.4.k_cache 0.00071811 1.73266621 + layer.4.v_cache 0.00005486 0.13754614 + layer.4.output 0.19495771 705.69080473 + ------------------------------------------------------------------------------------- + TOTAL 0.09094743 326.59957386 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9864 +BPFP 0.1177 bits/point +EBPFP 0.2355 equivalent bits/point +MSE 326.599574 +---------------------- -------------------------------------------------------- +Time: 0.356s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 326.5996 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 672B, BPFP=0.1167 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 812B, BPFP=0.1410 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 824B, BPFP=0.1431 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 756B, BPFP=0.1313 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1333 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 768B, BPFP=0.1333 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 736B, BPFP=0.1278 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 772B, BPFP=0.1340 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 844B, BPFP=0.1465 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 828B, BPFP=0.1437 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,172B, BPFP=0.0539 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08657128 517.95412326 + layer.0.v_cache 0.00001574 0.04517962 + layer.1.k_cache 0.05618976 36.90617947 + layer.1.v_cache 0.00000506 0.01493530 + layer.2.k_cache 0.00392605 6.51790907 + layer.2.v_cache 0.00001570 0.04207154 + layer.3.k_cache 0.02849737 23.42666558 + layer.3.v_cache 0.00001769 0.05702091 + layer.4.k_cache 0.00099738 1.69466044 + layer.4.v_cache 0.00004751 0.12205325 + layer.4.output 0.15107292 604.79365079 + ------------------------------------------------------------------------------------- + TOTAL 0.07257612 283.54919729 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 9952 +BPFP 0.1016 bits/point +EBPFP 0.2033 equivalent bits/point +MSE 283.549197 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 283.5492 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 628B, BPFP=0.1227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1492 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1586 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1344 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 724B, BPFP=0.1414 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 696B, BPFP=0.1359 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 672B, BPFP=0.1313 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 696B, BPFP=0.1359 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 796B, BPFP=0.1555 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 760B, BPFP=0.1484 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,308B, BPFP=0.0644 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09254595 536.26845703 + layer.0.v_cache 0.00001776 0.04470598 + layer.1.k_cache 0.05775445 37.71629028 + layer.1.v_cache 0.00000504 0.01507079 + layer.2.k_cache 0.00891839 6.68795090 + layer.2.v_cache 0.00001732 0.04529485 + layer.3.k_cache 0.02836140 23.50653076 + layer.3.v_cache 0.00001672 0.05638015 + layer.4.k_cache 0.00083566 1.69156017 + layer.4.v_cache 0.00004876 0.12718091 + layer.4.output 0.16993865 678.34698661 + ------------------------------------------------------------------------------------- + TOTAL 0.08106423 314.97578400 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 9544 +BPFP 0.1097 bits/point +EBPFP 0.2193 equivalent bits/point +MSE 314.975784 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 314.9758 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 688B, BPFP=0.1108 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 852B, BPFP=0.1372 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 824B, BPFP=0.1327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 748B, BPFP=0.1205 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 772B, BPFP=0.1244 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 760B, BPFP=0.1224 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 760B, BPFP=0.1224 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 784B, BPFP=0.1263 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 848B, BPFP=0.1366 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 848B, BPFP=0.1366 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,080B, BPFP=0.0479 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13093194 521.77931701 + layer.0.v_cache 0.00001635 0.04646907 + layer.1.k_cache 0.05235285 37.13226381 + layer.1.v_cache 0.00000496 0.01555826 + layer.2.k_cache 0.00512381 6.62771575 + layer.2.v_cache 0.00001647 0.04495300 + layer.3.k_cache 0.01138228 23.28786948 + layer.3.v_cache 0.00001850 0.05737926 + layer.4.k_cache 0.00081880 1.66075244 + layer.4.v_cache 0.00004796 0.12057363 + layer.4.output 0.02447439 577.94868373 + ------------------------------------------------------------------------------------- + TOTAL 0.02188439 272.73021399 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 9964 +BPFP 0.0944 bits/point +EBPFP 0.1888 equivalent bits/point +MSE 272.730214 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 272.7302 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1519 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 788B, BPFP=0.1559 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 672B, BPFP=0.1329 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 708B, BPFP=0.1400 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 688B, BPFP=0.1361 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 668B, BPFP=0.1321 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 712B, BPFP=0.1408 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1551 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 768B, BPFP=0.1519 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,716B, BPFP=0.0767 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510638 536.34434335 + layer.0.v_cache 0.00001399 0.04834682 + layer.1.k_cache 0.05467150 38.01875865 + layer.1.v_cache 0.00000511 0.01614434 + layer.2.k_cache 0.00545186 6.82180516 + layer.2.v_cache 0.00001652 0.04750165 + layer.3.k_cache 0.02986256 23.72007887 + layer.3.v_cache 0.00001747 0.05520410 + layer.4.k_cache 0.00068266 1.75196665 + layer.4.v_cache 0.00005277 0.13162633 + layer.4.output 0.17642128 687.78582731 + ------------------------------------------------------------------------------------- + TOTAL 0.08357822 318.90920983 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 9904 +BPFP 0.1152 bits/point +EBPFP 0.2305 equivalent bits/point +MSE 318.909210 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.006s, Pack+Encode: 0.165s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 318.9092 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 636B, BPFP=0.1291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1550 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 792B, BPFP=0.1607 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 668B, BPFP=0.1356 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 720B, BPFP=0.1461 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 668B, BPFP=0.1356 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 712B, BPFP=0.1445 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1599 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 752B, BPFP=0.1526 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,720B, BPFP=0.0788 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08730992 537.90635146 + layer.0.v_cache 0.00001543 0.04832107 + layer.1.k_cache 0.05635426 38.52679840 + layer.1.v_cache 0.00000520 0.01669959 + layer.2.k_cache 0.00390536 6.84042477 + layer.2.v_cache 0.00001661 0.04604041 + layer.3.k_cache 0.06483487 23.54701134 + layer.3.v_cache 0.00002010 0.05808766 + layer.4.k_cache 0.00072652 1.72665762 + layer.4.v_cache 0.00004842 0.12812763 + layer.4.output 0.18845482 705.96353200 + ------------------------------------------------------------------------------------- + TOTAL 0.09014238 326.50524965 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9904 +BPFP 0.1182 bits/point +EBPFP 0.2364 equivalent bits/point +MSE 326.505250 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 326.5052 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 660B, BPFP=0.1517 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1765 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 776B, BPFP=0.1783 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1581 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 740B, BPFP=0.1700 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 692B, BPFP=0.1590 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 704B, BPFP=0.1618 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 712B, BPFP=0.1636 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 776B, BPFP=0.1783 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1737 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,524B, BPFP=0.0829 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10784986 515.38821232 + layer.0.v_cache 0.00001387 0.04628707 + layer.1.k_cache 0.01409798 38.18916590 + layer.1.v_cache 0.00000489 0.01570997 + layer.2.k_cache 0.00254294 6.95349660 + layer.2.v_cache 0.00001653 0.04673809 + layer.3.k_cache 0.05709259 23.62795482 + layer.3.v_cache 0.00001736 0.05693941 + layer.4.k_cache 0.00062715 1.69222506 + layer.4.v_cache 0.00004758 0.12934760 + layer.4.output 0.22376562 798.57549895 + ------------------------------------------------------------------------------------- + TOTAL 0.10286295 363.30438644 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 9796 +BPFP 0.1324 bits/point +EBPFP 0.2648 equivalent bits/point +MSE 363.304386 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 363.3044 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1519 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1590 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 700B, BPFP=0.1384 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 724B, BPFP=0.1432 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 680B, BPFP=0.1345 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 668B, BPFP=0.1321 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 708B, BPFP=0.1400 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 800B, BPFP=0.1582 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1495 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,708B, BPFP=0.0765 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14563359 541.76453718 + layer.0.v_cache 0.00001368 0.04379987 + layer.1.k_cache 0.05947259 38.05605963 + layer.1.v_cache 0.00000496 0.01492491 + layer.2.k_cache 0.00553216 6.83842526 + layer.2.v_cache 0.00001557 0.04352439 + layer.3.k_cache 0.04377548 23.73511978 + layer.3.v_cache 0.00001730 0.05533972 + layer.4.k_cache 0.00079035 1.71326031 + layer.4.v_cache 0.00005153 0.12888761 + layer.4.output 0.17347779 687.92241184 + ------------------------------------------------------------------------------------- + TOTAL 0.08645010 319.28533892 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 9948 +BPFP 0.1157 bits/point +EBPFP 0.2315 equivalent bits/point +MSE 319.285339 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 319.2853 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 624B, BPFP=0.1477 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1809 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 776B, BPFP=0.1837 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 692B, BPFP=0.1638 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 724B, BPFP=0.1714 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 692B, BPFP=0.1638 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1648 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 704B, BPFP=0.1667 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 764B, BPFP=0.1809 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 744B, BPFP=0.1761 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,260B, BPFP=0.0764 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12373332 524.31338778 + layer.0.v_cache 0.00001436 0.04777015 + layer.1.k_cache 0.01270404 38.78285171 + layer.1.v_cache 0.00000490 0.01602385 + layer.2.k_cache 0.00252672 7.10243826 + layer.2.v_cache 0.00001629 0.04828679 + layer.3.k_cache 0.05185594 23.97522897 + layer.3.v_cache 0.00001849 0.06078287 + layer.4.k_cache 0.00068458 1.77113539 + layer.4.v_cache 0.00004975 0.13310037 + layer.4.output 0.22283003 832.15861742 + ------------------------------------------------------------------------------------- + TOTAL 0.10302462 377.72713695 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 9440 +BPFP 0.1315 bits/point +EBPFP 0.2629 equivalent bits/point +MSE 377.727137 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 377.7271 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 644B, BPFP=0.1198 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 792B, BPFP=0.1473 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 824B, BPFP=0.1533 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 680B, BPFP=0.1265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1369 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 680B, BPFP=0.1265 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 724B, BPFP=0.1347 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 816B, BPFP=0.1518 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 780B, BPFP=0.1451 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,236B, BPFP=0.0594 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12002775 526.13276600 + layer.0.v_cache 0.00001466 0.05025096 + layer.1.k_cache 0.05271155 37.29709008 + layer.1.v_cache 0.00000524 0.01659161 + layer.2.k_cache 0.00808975 6.66579910 + layer.2.v_cache 0.00001826 0.04827119 + layer.3.k_cache 0.07501186 23.45886666 + layer.3.v_cache 0.00001895 0.06343232 + layer.4.k_cache 0.00071216 1.71559107 + layer.4.v_cache 0.00005170 0.13449683 + layer.4.output 0.17053346 647.84056122 + ------------------------------------------------------------------------------------- + TOTAL 0.08531742 301.79218143 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 9608 +BPFP 0.1051 bits/point +EBPFP 0.2103 equivalent bits/point +MSE 301.792181 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 301.7922 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 75, 128) +Output shape: (1, 75, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.output: torch.Size([1, 75, 3584]) -> torch.Size([1, 1, 75, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1333 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 772B, BPFP=0.1608 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 788B, BPFP=0.1642 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 704B, BPFP=0.1467 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1517 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 676B, BPFP=0.1408 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1450 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 716B, BPFP=0.1492 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1642 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 772B, BPFP=0.1608 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,740B, BPFP=0.0815 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08149677 530.12125000 + layer.0.v_cache 0.00001391 0.04826121 + layer.1.k_cache 0.05841089 38.29092448 + layer.1.v_cache 0.00000541 0.01640949 + layer.2.k_cache 0.00252637 6.90304606 + layer.2.v_cache 0.00001693 0.04717457 + layer.3.k_cache 0.02876982 23.50883789 + layer.3.v_cache 0.00001932 0.05832598 + layer.4.k_cache 0.00075246 1.75955526 + layer.4.v_cache 0.00005041 0.13272509 + layer.4.output 0.19266937 724.33869048 + ------------------------------------------------------------------------------------- + TOTAL 0.08945576 333.60337314 + (elements=652,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 652800 +Total Bytes 10020 +BPFP 0.1228 bits/point +EBPFP 0.2456 equivalent bits/point +MSE 333.603373 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 333.6034 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1507 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 764B, BPFP=0.1756 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 768B, BPFP=0.1765 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 688B, BPFP=0.1581 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1691 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 696B, BPFP=0.1599 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 712B, BPFP=0.1636 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 724B, BPFP=0.1664 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 800B, BPFP=0.1838 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 760B, BPFP=0.1746 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,516B, BPFP=0.0826 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10404026 520.33605239 + layer.0.v_cache 0.00001395 0.04824450 + layer.1.k_cache 0.01458383 38.41424561 + layer.1.v_cache 0.00000499 0.01539640 + layer.2.k_cache 0.00236740 6.96733273 + layer.2.v_cache 0.00001641 0.04613851 + layer.3.k_cache 0.03645969 23.59326172 + layer.3.v_cache 0.00002420 0.05695927 + layer.4.k_cache 0.00069996 1.72316832 + layer.4.v_cache 0.00005119 0.13567152 + layer.4.output 0.20980707 798.79759716 + ------------------------------------------------------------------------------------- + TOTAL 0.09570067 363.70115595 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 9820 +BPFP 0.1327 bits/point +EBPFP 0.2655 equivalent bits/point +MSE 363.701156 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 363.7012 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 652B, BPFP=0.1455 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1696 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 776B, BPFP=0.1732 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 684B, BPFP=0.1527 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1634 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 692B, BPFP=0.1545 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 700B, BPFP=0.1562 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 688B, BPFP=0.1536 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1759 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,540B, BPFP=0.0810 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11725240 529.48136161 + layer.0.v_cache 0.00001567 0.04764767 + layer.1.k_cache 0.03693308 38.34276646 + layer.1.v_cache 0.00000533 0.01666593 + layer.2.k_cache 0.00240566 6.99163034 + layer.2.v_cache 0.00001850 0.05111261 + layer.3.k_cache 0.06751672 23.63242536 + layer.3.v_cache 0.00001909 0.05877238 + layer.4.k_cache 0.00067984 1.74822627 + layer.4.v_cache 0.00004901 0.13203784 + layer.4.output 0.20094273 776.97920918 + ------------------------------------------------------------------------------------- + TOTAL 0.09597026 355.25630063 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 9768 +BPFP 0.1283 bits/point +EBPFP 0.2565 equivalent bits/point +MSE 355.256301 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 355.2563 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 652B, BPFP=0.1455 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 772B, BPFP=0.1723 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 772B, BPFP=0.1723 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 696B, BPFP=0.1554 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1634 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 700B, BPFP=0.1562 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 708B, BPFP=0.1580 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 732B, BPFP=0.1634 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 796B, BPFP=0.1777 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 772B, BPFP=0.1723 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,488B, BPFP=0.0793 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07685234 530.25290179 + layer.0.v_cache 0.00001390 0.04808702 + layer.1.k_cache 0.01187498 38.10247628 + layer.1.v_cache 0.00000934 0.01627546 + layer.2.k_cache 0.00239603 6.96721409 + layer.2.v_cache 0.00001621 0.04468389 + layer.3.k_cache 0.07241479 23.49944894 + layer.3.v_cache 0.00001716 0.05716070 + layer.4.k_cache 0.00072067 1.71885725 + layer.4.v_cache 0.00004958 0.12448087 + layer.4.output 0.21417375 776.68144133 + ------------------------------------------------------------------------------------- + TOTAL 0.09785772 355.15303974 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 9820 +BPFP 0.1289 bits/point +EBPFP 0.2579 equivalent bits/point +MSE 355.153040 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 355.1530 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1316 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 752B, BPFP=0.1546 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 800B, BPFP=0.1645 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 660B, BPFP=0.1357 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1497 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 672B, BPFP=0.1382 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1406 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 680B, BPFP=0.1398 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 788B, BPFP=0.1620 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 744B, BPFP=0.1530 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,652B, BPFP=0.0779 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09264422 531.23797286 + layer.0.v_cache 0.00001680 0.04811831 + layer.1.k_cache 0.05606840 38.30415425 + layer.1.v_cache 0.00000542 0.01673474 + layer.2.k_cache 0.00241068 6.89789300 + layer.2.v_cache 0.00001702 0.04652522 + layer.3.k_cache 0.15270803 23.68269910 + layer.3.v_cache 0.00001849 0.06065432 + layer.4.k_cache 0.00070970 1.74555648 + layer.4.v_cache 0.00005539 0.13301040 + layer.4.output 0.18838861 715.38868656 + ------------------------------------------------------------------------------------- + TOTAL 0.09549261 329.99377204 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 9800 +BPFP 0.1185 bits/point +EBPFP 0.2370 equivalent bits/point +MSE 329.993772 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 329.9938 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1299 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 728B, BPFP=0.1477 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 788B, BPFP=0.1599 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 672B, BPFP=0.1364 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1477 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 688B, BPFP=0.1396 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 688B, BPFP=0.1396 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1591 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1534 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,632B, BPFP=0.0763 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10144231 534.99913758 + layer.0.v_cache 0.00001330 0.04586132 + layer.1.k_cache 0.09893422 38.25193727 + layer.1.v_cache 0.00000513 0.01556940 + layer.2.k_cache 0.00368564 6.89058061 + layer.2.v_cache 0.00001777 0.04656501 + layer.3.k_cache 0.06341467 23.36755371 + layer.3.v_cache 0.00001739 0.05954424 + layer.4.k_cache 0.00073693 1.75289957 + layer.4.v_cache 0.00004994 0.13155883 + layer.4.output 0.18261438 705.34792440 + ------------------------------------------------------------------------------------- + TOTAL 0.09097753 326.05862814 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9788 +BPFP 0.1168 bits/point +EBPFP 0.2337 equivalent bits/point +MSE 326.058628 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 326.0586 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 672B, BPFP=0.1207 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 820B, BPFP=0.1473 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1473 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 752B, BPFP=0.1351 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 772B, BPFP=0.1386 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 756B, BPFP=0.1358 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 732B, BPFP=0.1315 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 768B, BPFP=0.1379 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 840B, BPFP=0.1509 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 800B, BPFP=0.1437 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,496B, BPFP=0.0640 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08947531 527.83072917 + layer.0.v_cache 0.00001432 0.04300256 + layer.1.k_cache 0.05320675 37.26836386 + layer.1.v_cache 0.00000497 0.01465953 + layer.2.k_cache 0.00562141 6.57422296 + layer.2.v_cache 0.00001574 0.04168253 + layer.3.k_cache 0.02742601 23.60682696 + layer.3.v_cache 0.00001759 0.05424421 + layer.4.k_cache 0.00114861 1.72453185 + layer.4.v_cache 0.00004734 0.11849532 + layer.4.output 0.15628038 628.11227422 + ------------------------------------------------------------------------------------- + TOTAL 0.07476122 293.76839285 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 10228 +BPFP 0.1081 bits/point +EBPFP 0.2161 equivalent bits/point +MSE 293.768393 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 293.7684 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 676B, BPFP=0.1148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 836B, BPFP=0.1420 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 840B, BPFP=0.1427 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 732B, BPFP=0.1243 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1304 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 792B, BPFP=0.1345 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 756B, BPFP=0.1284 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 780B, BPFP=0.1325 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 856B, BPFP=0.1454 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 832B, BPFP=0.1413 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,060B, BPFP=0.0500 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12311893 512.50212296 + layer.0.v_cache 0.00001357 0.04264581 + layer.1.k_cache 0.03743907 36.64955206 + layer.1.v_cache 0.00000515 0.01462343 + layer.2.k_cache 0.00511221 6.43866564 + layer.2.v_cache 0.00001545 0.04166181 + layer.3.k_cache 0.04225797 23.11406409 + layer.3.v_cache 0.00001799 0.05345094 + layer.4.k_cache 0.00119426 1.65638733 + layer.4.v_cache 0.00004621 0.11159623 + layer.4.output 0.14783020 589.99631211 + ------------------------------------------------------------------------------------- + TOTAL 0.07317837 277.09405618 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 9928 +BPFP 0.0992 bits/point +EBPFP 0.1984 equivalent bits/point +MSE 277.094056 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 277.0941 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 700B, BPFP=0.1072 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 888B, BPFP=0.1360 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 836B, BPFP=0.1281 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 780B, BPFP=0.1195 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 776B, BPFP=0.1189 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 772B, BPFP=0.1183 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 768B, BPFP=0.1176 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 820B, BPFP=0.1256 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 884B, BPFP=0.1354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 860B, BPFP=0.1317 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,384B, BPFP=0.0522 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09293266 515.72349877 + layer.0.v_cache 0.00001503 0.04842854 + layer.1.k_cache 0.07854857 37.10027478 + layer.1.v_cache 0.00000559 0.01629323 + layer.2.k_cache 0.00650115 6.56422454 + layer.2.v_cache 0.00001763 0.04775591 + layer.3.k_cache 0.03715762 23.29500326 + layer.3.v_cache 0.00001788 0.05777862 + layer.4.k_cache 0.00073365 1.60005188 + layer.4.v_cache 0.00004630 0.11993376 + layer.4.output 11.17680568 527.64990371 + ------------------------------------------------------------------------------------- + TOTAL 4.61491858 251.65426878 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 10468 +BPFP 0.0943 bits/point +EBPFP 0.1887 equivalent bits/point +MSE 251.654269 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.006s, Pack+Encode: 0.156s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 251.6543 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 96, 128) +Output shape: (1, 96, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.output: torch.Size([1, 96, 3584]) -> torch.Size([1, 1, 96, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 688B, BPFP=0.1120 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 856B, BPFP=0.1393 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 812B, BPFP=0.1322 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 740B, BPFP=0.1204 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 796B, BPFP=0.1296 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 776B, BPFP=0.1263 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 764B, BPFP=0.1243 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 792B, BPFP=0.1289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 864B, BPFP=0.1406 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 824B, BPFP=0.1341 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,168B, BPFP=0.0504 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10158994 519.46191406 + layer.0.v_cache 0.00001467 0.04971003 + layer.1.k_cache 0.06416908 37.20018514 + layer.1.v_cache 0.00000542 0.01697497 + layer.2.k_cache 0.00400108 6.69987615 + layer.2.v_cache 0.00001872 0.05048951 + layer.3.k_cache 0.01491791 23.42224630 + layer.3.v_cache 0.00001877 0.06495642 + layer.4.k_cache 0.00070988 1.73138237 + layer.4.v_cache 0.00004832 0.12964187 + layer.4.output 0.14169503 569.18019903 + ------------------------------------------------------------------------------------- + TOTAL 0.06925641 269.00522177 + (elements=835,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 835584 +Total Bytes 10080 +BPFP 0.0965 bits/point +EBPFP 0.1930 equivalent bits/point +MSE 269.005222 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 269.0052 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 652B, BPFP=0.1415 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1667 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 780B, BPFP=0.1693 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 712B, BPFP=0.1545 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1597 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 696B, BPFP=0.1510 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1493 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 696B, BPFP=0.1510 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 804B, BPFP=0.1745 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 772B, BPFP=0.1675 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,716B, BPFP=0.0842 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10321087 530.15115017 + layer.0.v_cache 0.00001557 0.04772559 + layer.1.k_cache 0.06175087 37.87112088 + layer.1.v_cache 0.00000571 0.01613750 + layer.2.k_cache 0.01329466 6.73962657 + layer.2.v_cache 0.00001784 0.04875596 + layer.3.k_cache 0.05039416 23.36130778 + layer.3.v_cache 0.00001745 0.05810291 + layer.4.k_cache 0.00079786 1.70339712 + layer.4.v_cache 0.00005071 0.13040123 + layer.4.output 0.18870974 753.55791171 + ------------------------------------------------------------------------------------- + TOTAL 0.09120729 345.59018280 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 10020 +BPFP 0.1279 bits/point +EBPFP 0.2558 equivalent bits/point +MSE 345.590183 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 345.5902 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 652B, BPFP=0.1415 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 776B, BPFP=0.1684 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 776B, BPFP=0.1684 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 700B, BPFP=0.1519 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1597 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 704B, BPFP=0.1528 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1510 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 712B, BPFP=0.1545 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 800B, BPFP=0.1736 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 780B, BPFP=0.1693 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,652B, BPFP=0.0822 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09666174 526.07866753 + layer.0.v_cache 0.00001715 0.04516016 + layer.1.k_cache 0.08476152 37.99026150 + layer.1.v_cache 0.00000506 0.01517094 + layer.2.k_cache 0.00238475 6.63141039 + layer.2.v_cache 0.00001567 0.04376604 + layer.3.k_cache 0.07332428 23.33880785 + layer.3.v_cache 0.00001796 0.05741778 + layer.4.k_cache 0.00072290 1.65860314 + layer.4.v_cache 0.00005092 0.12129180 + layer.4.output 0.18875189 753.63740079 + ------------------------------------------------------------------------------------- + TOTAL 0.09289560 345.37896251 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 9984 +BPFP 0.1275 bits/point +EBPFP 0.2549 equivalent bits/point +MSE 345.378963 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 345.3790 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 648B, BPFP=0.1489 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 772B, BPFP=0.1774 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 772B, BPFP=0.1774 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 692B, BPFP=0.1590 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1691 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 704B, BPFP=0.1618 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 708B, BPFP=0.1627 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 720B, BPFP=0.1654 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 780B, BPFP=0.1792 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 764B, BPFP=0.1756 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,540B, BPFP=0.0834 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10601801 522.66256893 + layer.0.v_cache 0.00001817 0.04809410 + layer.1.k_cache 0.01696859 38.24356259 + layer.1.v_cache 0.00000501 0.01676279 + layer.2.k_cache 0.00274148 6.93209076 + layer.2.v_cache 0.00001757 0.04965682 + layer.3.k_cache 0.01706722 23.76971436 + layer.3.v_cache 0.00001843 0.06027436 + layer.4.k_cache 0.00080906 1.79744720 + layer.4.v_cache 0.00005076 0.13114590 + layer.4.output 0.19983917 797.85136555 + ------------------------------------------------------------------------------------- + TOTAL 0.09074050 363.45122804 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 9836 +BPFP 0.1329 bits/point +EBPFP 0.2659 equivalent bits/point +MSE 363.451228 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 363.4512 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 65, 128) +Output shape: (1, 65, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.output: torch.Size([1, 65, 3584]) -> torch.Size([1, 1, 65, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 620B, BPFP=0.1490 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 712B, BPFP=0.1712 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 764B, BPFP=0.1837 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 676B, BPFP=0.1625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 712B, BPFP=0.1712 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 676B, BPFP=0.1625 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1654 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 676B, BPFP=0.1625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 748B, BPFP=0.1798 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 732B, BPFP=0.1760 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,208B, BPFP=0.0758 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.196s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09579474 526.96947115 + layer.0.v_cache 0.00001312 0.04575128 + layer.1.k_cache 0.01648773 38.31337139 + layer.1.v_cache 0.00000712 0.01535896 + layer.2.k_cache 0.00431219 6.92408166 + layer.2.v_cache 0.00001803 0.04773267 + layer.3.k_cache 0.03460605 23.30217849 + layer.3.v_cache 0.00001721 0.05852320 + layer.4.k_cache 0.00068162 1.71793424 + layer.4.v_cache 0.00004727 0.13318226 + layer.4.output 0.20898957 835.49436813 + ------------------------------------------------------------------------------------- + TOTAL 0.09499483 379.17577425 + (elements=565,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 565760 +Total Bytes 9212 +BPFP 0.1303 bits/point +EBPFP 0.2605 equivalent bits/point +MSE 379.175774 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.196s +---------------------- -------------------------------------------------------- +💾 Converting with 379.1758 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 41, 128) +Output shape: (1, 41, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.output: torch.Size([1, 41, 3584]) -> torch.Size([1, 1, 41, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 392B, BPFP=0.1494 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 428B, BPFP=0.1631 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 472B, BPFP=0.1799 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 388B, BPFP=0.1479 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 460B, BPFP=0.1753 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 392B, BPFP=0.1494 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 420B, BPFP=0.1601 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 388B, BPFP=0.1479 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 480B, BPFP=0.1829 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 440B, BPFP=0.1677 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,696B, BPFP=0.0923 +⌛️ [2/4] FRONTEND: Frontend time: 0.126s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.141s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10098721 562.84379764 + layer.0.v_cache 0.00001362 0.05219413 + layer.1.k_cache 0.01708067 41.60630836 + layer.1.v_cache 0.00000533 0.01799610 + layer.2.k_cache 0.00250472 7.47077979 + layer.2.v_cache 0.00001755 0.05577155 + layer.3.k_cache 0.08199174 25.22231088 + layer.3.v_cache 0.00001815 0.06283539 + layer.4.k_cache 0.00059437 1.86850273 + layer.4.v_cache 0.00005232 0.15280375 + layer.4.output 0.35770382 1324.43205575 + ------------------------------------------------------------------------------------- + TOTAL 0.15924661 582.96339356 + (elements=356,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 356864 +Total Bytes 5956 +BPFP 0.1335 bits/point +EBPFP 0.2670 equivalent bits/point +MSE 582.963394 +---------------------- -------------------------------------------------------- +Time: 0.269s Load: 0.003s, Pack+Encode: 0.126s, Decode+Unpack: 0.141s +---------------------- -------------------------------------------------------- +💾 Converting with 582.9634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 424B, BPFP=0.1274 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 436B, BPFP=0.1310 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 488B, BPFP=0.1466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 404B, BPFP=0.1214 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 464B, BPFP=0.1394 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 400B, BPFP=0.1202 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 428B, BPFP=0.1286 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 408B, BPFP=0.1226 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 480B, BPFP=0.1442 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 448B, BPFP=0.1346 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,524B, BPFP=0.0654 +⌛️ [2/4] FRONTEND: Frontend time: 0.127s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09210638 548.99308894 + layer.0.v_cache 0.00001369 0.04884452 + layer.1.k_cache 0.01682195 40.25744629 + layer.1.v_cache 0.00000520 0.01791459 + layer.2.k_cache 0.00487203 7.30406835 + layer.2.v_cache 0.00001755 0.05456549 + layer.3.k_cache 0.04965206 24.68587787 + layer.3.v_cache 0.00001944 0.06563762 + layer.4.k_cache 0.00060442 1.80026876 + layer.4.v_cache 0.00005207 0.14080800 + layer.4.output 0.26182200 1052.01734203 + ------------------------------------------------------------------------------------- + TOTAL 0.11746581 469.85234792 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 5904 +BPFP 0.1044 bits/point +EBPFP 0.2087 equivalent bits/point +MSE 469.852348 +---------------------- -------------------------------------------------------- +Time: 0.271s Load: 0.002s, Pack+Encode: 0.127s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 469.8523 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 47, 128) +Output shape: (1, 47, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.output: torch.Size([1, 47, 3584]) -> torch.Size([1, 1, 47, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 408B, BPFP=0.1356 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 428B, BPFP=0.1423 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 484B, BPFP=0.1609 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 404B, BPFP=0.1343 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 448B, BPFP=0.1489 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 408B, BPFP=0.1356 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 436B, BPFP=0.1449 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 412B, BPFP=0.1370 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 460B, BPFP=0.1529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 440B, BPFP=0.1463 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,540B, BPFP=0.0731 +⌛️ [2/4] FRONTEND: Frontend time: 0.126s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08461995 563.51117852 + layer.0.v_cache 0.00001366 0.05054650 + layer.1.k_cache 0.01838745 41.77278040 + layer.1.v_cache 0.00000609 0.01908153 + layer.2.k_cache 0.00780689 7.48454577 + layer.2.v_cache 0.00001932 0.05499954 + layer.3.k_cache 0.04292001 25.18231071 + layer.3.v_cache 0.00001942 0.06665540 + layer.4.k_cache 0.00059704 1.85923166 + layer.4.v_cache 0.00004850 0.14904988 + layer.4.output 0.29441516 1159.18056611 + ------------------------------------------------------------------------------------- + TOTAL 0.13031438 514.96554957 + (elements=409,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 409088 +Total Bytes 5868 +BPFP 0.1148 bits/point +EBPFP 0.2295 equivalent bits/point +MSE 514.965550 +---------------------- -------------------------------------------------------- +Time: 0.270s Load: 0.002s, Pack+Encode: 0.126s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 514.9655 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 432B, BPFP=0.1184 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 460B, BPFP=0.1261 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 472B, BPFP=0.1294 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 404B, BPFP=0.1107 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 468B, BPFP=0.1283 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 428B, BPFP=0.1173 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 436B, BPFP=0.1195 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 424B, BPFP=0.1162 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 468B, BPFP=0.1283 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 464B, BPFP=0.1272 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,436B, BPFP=0.0562 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.140s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10653571 544.51075932 + layer.0.v_cache 0.00001396 0.05184699 + layer.1.k_cache 0.01513677 40.07973975 + layer.1.v_cache 0.00000574 0.01951213 + layer.2.k_cache 0.00241650 7.43967853 + layer.2.v_cache 0.00001997 0.06115340 + layer.3.k_cache 0.04250543 25.10060307 + layer.3.v_cache 0.00001889 0.06943307 + layer.4.k_cache 0.00060099 1.91864121 + layer.4.v_cache 0.00005317 0.15571599 + layer.4.output 0.25650722 980.40671992 + ------------------------------------------------------------------------------------- + TOTAL 0.11546222 440.13259547 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 5892 +BPFP 0.0950 bits/point +EBPFP 0.1900 equivalent bits/point +MSE 440.132595 +---------------------- -------------------------------------------------------- +Time: 0.271s Load: 0.003s, Pack+Encode: 0.128s, Decode+Unpack: 0.140s +---------------------- -------------------------------------------------------- +💾 Converting with 440.1326 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 412B, BPFP=0.1341 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 432B, BPFP=0.1406 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 484B, BPFP=0.1576 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 392B, BPFP=0.1276 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 460B, BPFP=0.1497 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 412B, BPFP=0.1341 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 432B, BPFP=0.1406 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 404B, BPFP=0.1315 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 460B, BPFP=0.1497 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 444B, BPFP=0.1445 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,500B, BPFP=0.0698 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10088900 564.34977214 + layer.0.v_cache 0.00001321 0.04798734 + layer.1.k_cache 0.01601078 41.03089142 + layer.1.v_cache 0.00000529 0.01773517 + layer.2.k_cache 0.00240732 7.37995911 + layer.2.v_cache 0.00001858 0.05604496 + layer.3.k_cache 0.04236705 25.11397044 + layer.3.v_cache 0.00001968 0.06710851 + layer.4.k_cache 0.00061965 1.94418399 + layer.4.v_cache 0.00005235 0.15704257 + layer.4.output 0.30498679 1134.03869048 + ------------------------------------------------------------------------------------- + TOTAL 0.13513591 504.61385465 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 5832 +BPFP 0.1117 bits/point +EBPFP 0.2233 equivalent bits/point +MSE 504.613855 +---------------------- -------------------------------------------------------- +Time: 0.272s Load: 0.002s, Pack+Encode: 0.128s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 504.6139 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 53, 128) +Output shape: (1, 53, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.output: torch.Size([1, 53, 3584]) -> torch.Size([1, 1, 53, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 424B, BPFP=0.1250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 424B, BPFP=0.1250 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 488B, BPFP=0.1439 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 400B, BPFP=0.1179 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 460B, BPFP=0.1356 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 408B, BPFP=0.1203 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 432B, BPFP=0.1274 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 408B, BPFP=0.1203 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 472B, BPFP=0.1392 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 456B, BPFP=0.1344 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,468B, BPFP=0.0618 +⌛️ [2/4] FRONTEND: Frontend time: 0.127s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10150213 550.04481132 + layer.0.v_cache 0.00001370 0.05342224 + layer.1.k_cache 0.01753690 40.18176039 + layer.1.v_cache 0.00000536 0.01962621 + layer.2.k_cache 0.00234856 7.34895641 + layer.2.v_cache 0.00001946 0.06056100 + layer.3.k_cache 0.03772523 25.40287349 + layer.3.v_cache 0.00001792 0.06830806 + layer.4.k_cache 0.00061365 1.90332247 + layer.4.v_cache 0.00006478 0.15904151 + layer.4.output 0.25806696 1030.10545822 + ------------------------------------------------------------------------------------- + TOTAL 0.11566567 460.94005239 + (elements=461,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 461312 +Total Bytes 5840 +BPFP 0.1013 bits/point +EBPFP 0.2026 equivalent bits/point +MSE 460.940052 +---------------------- -------------------------------------------------------- +Time: 0.271s Load: 0.002s, Pack+Encode: 0.127s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 460.9401 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 380B, BPFP=0.0990 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 388B, BPFP=0.1010 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 412B, BPFP=0.1073 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 364B, BPFP=0.0948 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 396B, BPFP=0.1031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 380B, BPFP=0.0990 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 388B, BPFP=0.1010 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 380B, BPFP=0.0990 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 408B, BPFP=0.1062 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 404B, BPFP=0.1052 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,308B, BPFP=0.0487 +⌛️ [2/4] FRONTEND: Frontend time: 0.126s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11850098 511.90120443 + layer.0.v_cache 0.00001385 0.04681279 + layer.1.k_cache 0.01468890 36.65203857 + layer.1.v_cache 0.00000523 0.01672727 + layer.2.k_cache 0.01477869 6.54785665 + layer.2.v_cache 0.00001915 0.05287379 + layer.3.k_cache 0.14661819 22.63347982 + layer.3.v_cache 0.00001925 0.06769396 + layer.4.k_cache 0.00060183 1.61046054 + layer.4.v_cache 0.00005752 0.14883399 + layer.4.output 0.22966646 904.47202381 + ------------------------------------------------------------------------------------- + TOTAL 0.11193934 406.52836167 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 5208 +BPFP 0.0798 bits/point +EBPFP 0.1596 equivalent bits/point +MSE 406.528362 +---------------------- -------------------------------------------------------- +Time: 0.271s Load: 0.002s, Pack+Encode: 0.126s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 406.5284 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 684B, BPFP=0.1162 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 828B, BPFP=0.1406 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 848B, BPFP=0.1440 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 736B, BPFP=0.1250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1304 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 752B, BPFP=0.1277 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 784B, BPFP=0.1332 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 864B, BPFP=0.1467 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 824B, BPFP=0.1399 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,024B, BPFP=0.0491 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10836371 518.62831182 + layer.0.v_cache 0.00001445 0.04730704 + layer.1.k_cache 0.06846340 36.44446332 + layer.1.v_cache 0.00000523 0.01591231 + layer.2.k_cache 0.00506382 6.48714746 + layer.2.v_cache 0.00001740 0.04621402 + layer.3.k_cache 0.03973729 23.26394255 + layer.3.v_cache 0.00001777 0.05792712 + layer.4.k_cache 0.00066797 1.60633435 + layer.4.v_cache 0.00005099 0.12785569 + layer.4.output 0.14786712 590.37902756 + ------------------------------------------------------------------------------------- + TOTAL 0.07396893 277.61050639 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 9864 +BPFP 0.0985 bits/point +EBPFP 0.1971 equivalent bits/point +MSE 277.610506 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 277.6105 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 624B, BPFP=0.1477 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 724B, BPFP=0.1714 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 776B, BPFP=0.1837 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 672B, BPFP=0.1591 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 724B, BPFP=0.1714 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 692B, BPFP=0.1638 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 692B, BPFP=0.1638 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 696B, BPFP=0.1648 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 756B, BPFP=0.1790 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 736B, BPFP=0.1742 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,236B, BPFP=0.0756 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08728886 524.96809896 + layer.0.v_cache 0.00001734 0.05195268 + layer.1.k_cache 0.01499850 38.78292199 + layer.1.v_cache 0.00000528 0.01756942 + layer.2.k_cache 0.00240521 7.11569306 + layer.2.v_cache 0.00001739 0.05192927 + layer.3.k_cache 0.03126823 24.16336892 + layer.3.v_cache 0.00001843 0.06457200 + layer.4.k_cache 0.00061395 1.78472114 + layer.4.v_cache 0.00004830 0.13597997 + layer.4.output 0.21953388 831.84246483 + ------------------------------------------------------------------------------------- + TOTAL 0.09843639 377.64906242 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 9328 +BPFP 0.1299 bits/point +EBPFP 0.2598 equivalent bits/point +MSE 377.649062 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 377.6491 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 620B, BPFP=0.1468 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 728B, BPFP=0.1723 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 784B, BPFP=0.1856 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 664B, BPFP=0.1572 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 716B, BPFP=0.1695 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 676B, BPFP=0.1600 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1648 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 684B, BPFP=0.1619 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 752B, BPFP=0.1780 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 740B, BPFP=0.1752 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,216B, BPFP=0.0749 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08223097 532.40648674 + layer.0.v_cache 0.00001442 0.05154438 + layer.1.k_cache 0.01480860 38.67957653 + layer.1.v_cache 0.00000502 0.01693766 + layer.2.k_cache 0.00576455 7.16505293 + layer.2.v_cache 0.00002304 0.05084130 + layer.3.k_cache 0.03215047 24.04711729 + layer.3.v_cache 0.00001933 0.06702900 + layer.4.k_cache 0.00061510 1.75164471 + layer.4.v_cache 0.00004916 0.13918884 + layer.4.output 0.22077460 831.56838474 + ------------------------------------------------------------------------------------- + TOTAL 0.09888840 377.96200662 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 9276 +BPFP 0.1292 bits/point +EBPFP 0.2584 equivalent bits/point +MSE 377.962007 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.003s, Pack+Encode: 0.155s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 377.9620 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 652B, BPFP=0.1396 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 788B, BPFP=0.1687 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 780B, BPFP=0.1670 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 700B, BPFP=0.1498 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1567 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 712B, BPFP=0.1524 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1473 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 744B, BPFP=0.1592 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 800B, BPFP=0.1712 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 772B, BPFP=0.1652 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,728B, BPFP=0.0834 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10114760 524.10862586 + layer.0.v_cache 0.00001420 0.04632294 + layer.1.k_cache 0.03683303 37.99278949 + layer.1.v_cache 0.00000544 0.01603207 + layer.2.k_cache 0.00248627 6.70627124 + layer.2.v_cache 0.00001713 0.04623039 + layer.3.k_cache 0.02909129 23.45927868 + layer.3.v_cache 0.00001766 0.05879083 + layer.4.k_cache 0.00072613 1.72207318 + layer.4.v_cache 0.00005154 0.13160016 + layer.4.output 0.19731793 744.18529843 + ------------------------------------------------------------------------------------- + TOTAL 0.09127152 341.38735905 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 10096 +BPFP 0.1271 bits/point +EBPFP 0.2542 equivalent bits/point +MSE 341.387359 +---------------------- -------------------------------------------------------- +Time: 0.356s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 341.3874 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 644B, BPFP=0.1324 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 756B, BPFP=0.1554 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 796B, BPFP=0.1637 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 692B, BPFP=0.1423 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1497 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 688B, BPFP=0.1414 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1390 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 712B, BPFP=0.1464 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 772B, BPFP=0.1587 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 756B, BPFP=0.1554 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,668B, BPFP=0.0784 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10230959 538.73221628 + layer.0.v_cache 0.00001409 0.04932510 + layer.1.k_cache 0.07973289 38.25510125 + layer.1.v_cache 0.00000506 0.01584655 + layer.2.k_cache 0.00553133 6.87562641 + layer.2.v_cache 0.00001928 0.04969035 + layer.3.k_cache 0.03946657 23.74363146 + layer.3.v_cache 0.00001980 0.06376785 + layer.4.k_cache 0.00066525 1.74672579 + layer.4.v_cache 0.00005119 0.13493119 + layer.4.output 0.17985767 714.81144267 + ------------------------------------------------------------------------------------- + TOTAL 0.08745993 330.19688005 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 9888 +BPFP 0.1196 bits/point +EBPFP 0.2392 equivalent bits/point +MSE 330.196880 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 330.1969 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1299 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 784B, BPFP=0.1591 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 788B, BPFP=0.1599 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 684B, BPFP=0.1388 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 720B, BPFP=0.1461 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 676B, BPFP=0.1372 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 684B, BPFP=0.1388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 728B, BPFP=0.1477 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1591 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 768B, BPFP=0.1558 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,736B, BPFP=0.0793 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633976 542.32863231 + layer.0.v_cache 0.00001461 0.04826570 + layer.1.k_cache 0.07748085 38.36784066 + layer.1.v_cache 0.00000515 0.01520171 + layer.2.k_cache 0.00505460 6.80916922 + layer.2.v_cache 0.00001651 0.04389906 + layer.3.k_cache 0.02793909 23.81157892 + layer.3.v_cache 0.00001809 0.05422941 + layer.4.k_cache 0.00070819 1.69538344 + layer.4.v_cache 0.00004716 0.12001872 + layer.4.output 0.18765952 705.49918831 + ------------------------------------------------------------------------------------- + TOTAL 0.09007298 326.57579631 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 9992 +BPFP 0.1193 bits/point +EBPFP 0.2385 equivalent bits/point +MSE 326.575796 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 326.5758 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 652B, BPFP=0.1455 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 748B, BPFP=0.1670 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 772B, BPFP=0.1723 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 692B, BPFP=0.1545 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 736B, BPFP=0.1643 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 708B, BPFP=0.1580 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 700B, BPFP=0.1562 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 700B, BPFP=0.1562 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 784B, BPFP=0.1750 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 772B, BPFP=0.1723 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,496B, BPFP=0.0796 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08399124 528.96997768 + layer.0.v_cache 0.00001671 0.04802493 + layer.1.k_cache 0.03849935 38.28431920 + layer.1.v_cache 0.00000503 0.01593958 + layer.2.k_cache 0.00738276 6.91338501 + layer.2.v_cache 0.00001692 0.04747448 + layer.3.k_cache 0.04914829 23.63072510 + layer.3.v_cache 0.00001713 0.05853500 + layer.4.k_cache 0.00063731 1.70865032 + layer.4.v_cache 0.00004711 0.12580956 + layer.4.output 0.21328995 776.65637755 + ------------------------------------------------------------------------------------- + TOTAL 0.09839950 355.08220492 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 9760 +BPFP 0.1282 bits/point +EBPFP 0.2563 equivalent bits/point +MSE 355.082205 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 355.0822 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 56, 128) +Output shape: (1, 56, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.output: torch.Size([1, 56, 3584]) -> torch.Size([1, 1, 56, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 428B, BPFP=0.1194 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 468B, BPFP=0.1306 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 492B, BPFP=0.1373 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 408B, BPFP=0.1138 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 476B, BPFP=0.1328 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 420B, BPFP=0.1172 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 448B, BPFP=0.1250 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 436B, BPFP=0.1217 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 476B, BPFP=0.1328 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 468B, BPFP=0.1306 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,456B, BPFP=0.0580 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.140s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13768739 564.12377930 + layer.0.v_cache 0.00001432 0.05451903 + layer.1.k_cache 0.01638686 41.14870344 + layer.1.v_cache 0.00000533 0.01953186 + layer.2.k_cache 0.00251956 7.79597255 + layer.2.v_cache 0.00001745 0.05984925 + layer.3.k_cache 0.08599430 25.94219535 + layer.3.v_cache 0.00002012 0.07493381 + layer.4.k_cache 0.00061995 2.06519154 + layer.4.v_cache 0.00004798 0.14715534 + layer.4.output 0.24256272 1017.22759885 + ------------------------------------------------------------------------------------- + TOTAL 0.11419131 456.58970726 + (elements=487,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 487424 +Total Bytes 5976 +BPFP 0.0981 bits/point +EBPFP 0.1962 equivalent bits/point +MSE 456.589707 +---------------------- -------------------------------------------------------- +Time: 0.272s Load: 0.003s, Pack+Encode: 0.128s, Decode+Unpack: 0.140s +---------------------- -------------------------------------------------------- +💾 Converting with 456.5897 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 632B, BPFP=0.1234 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 768B, BPFP=0.1500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 792B, BPFP=0.1547 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 668B, BPFP=0.1305 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 712B, BPFP=0.1391 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 684B, BPFP=0.1336 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 676B, BPFP=0.1320 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 700B, BPFP=0.1367 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 804B, BPFP=0.1570 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 768B, BPFP=0.1500 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,392B, BPFP=0.0667 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09278157 536.73212891 + layer.0.v_cache 0.00001504 0.04797013 + layer.1.k_cache 0.07523044 37.66615295 + layer.1.v_cache 0.00000517 0.01608807 + layer.2.k_cache 0.00381728 6.66987457 + layer.2.v_cache 0.00001789 0.04647681 + layer.3.k_cache 0.03327554 23.53988953 + layer.3.v_cache 0.00001939 0.05777473 + layer.4.k_cache 0.00067710 1.75220928 + layer.4.v_cache 0.00005192 0.12862072 + layer.4.output 0.16994183 677.54849330 + ------------------------------------------------------------------------------------- + TOTAL 0.08208730 314.67627287 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 9596 +BPFP 0.1102 bits/point +EBPFP 0.2205 equivalent bits/point +MSE 314.676273 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 314.6763 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 404B, BPFP=0.1070 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 436B, BPFP=0.1155 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 440B, BPFP=0.1165 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 392B, BPFP=0.1038 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 420B, BPFP=0.1112 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 404B, BPFP=0.1070 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 408B, BPFP=0.1081 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 404B, BPFP=0.1070 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 436B, BPFP=0.1155 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 440B, BPFP=0.1165 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,340B, BPFP=0.0507 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12468526 520.36407574 + layer.0.v_cache 0.00001451 0.05312255 + layer.1.k_cache 0.01313243 37.73221084 + layer.1.v_cache 0.00000580 0.01873284 + layer.2.k_cache 0.00906488 6.85278165 + layer.2.v_cache 0.00002099 0.05781592 + layer.3.k_cache 0.01569995 23.67013705 + layer.3.v_cache 0.00002002 0.06959871 + layer.4.k_cache 0.00064964 1.77533321 + layer.4.v_cache 0.00005478 0.15159327 + layer.4.output 0.23031524 920.74220642 + ------------------------------------------------------------------------------------- + TOTAL 0.10444441 413.87887334 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 5524 +BPFP 0.0861 bits/point +EBPFP 0.1721 equivalent bits/point +MSE 413.878873 +---------------------- -------------------------------------------------------- +Time: 0.272s Load: 0.003s, Pack+Encode: 0.128s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 413.8789 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 384B, BPFP=0.1000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 400B, BPFP=0.1042 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 412B, BPFP=0.1073 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 380B, BPFP=0.0990 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 400B, BPFP=0.1042 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 384B, BPFP=0.1000 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 384B, BPFP=0.1000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 392B, BPFP=0.1021 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 412B, BPFP=0.1073 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 400B, BPFP=0.1042 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 1,368B, BPFP=0.0509 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15164433 504.98834635 + layer.0.v_cache 0.00001435 0.05103664 + layer.1.k_cache 0.01704682 36.63837891 + layer.1.v_cache 0.00000573 0.01835612 + layer.2.k_cache 0.00240009 6.51820170 + layer.2.v_cache 0.00001845 0.05353671 + layer.3.k_cache 0.02161660 23.04301554 + layer.3.v_cache 0.00001985 0.06642773 + layer.4.k_cache 0.00061459 1.71165517 + layer.4.v_cache 0.00005275 0.14510156 + layer.4.output 0.22653077 903.61599702 + ------------------------------------------------------------------------------------- + TOTAL 0.10465582 405.79682562 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 5316 +BPFP 0.0814 bits/point +EBPFP 0.1629 equivalent bits/point +MSE 405.796826 +---------------------- -------------------------------------------------------- +Time: 0.273s Load: 0.003s, Pack+Encode: 0.128s, Decode+Unpack: 0.142s +---------------------- -------------------------------------------------------- +💾 Converting with 405.7968 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 648B, BPFP=0.1387 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 784B, BPFP=0.1678 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 780B, BPFP=0.1670 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 704B, BPFP=0.1507 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 740B, BPFP=0.1584 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 708B, BPFP=0.1515 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 688B, BPFP=0.1473 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 732B, BPFP=0.1567 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 800B, BPFP=0.1712 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 780B, BPFP=0.1670 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,948B, BPFP=0.0901 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.196s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08869729 529.71243579 + layer.0.v_cache 0.00001594 0.04545069 + layer.1.k_cache 0.08105310 37.87809022 + layer.1.v_cache 0.00000494 0.01546950 + layer.2.k_cache 0.00256802 6.69950219 + layer.2.v_cache 0.00001661 0.04402080 + layer.3.k_cache 0.01835501 23.42703707 + layer.3.v_cache 0.00001861 0.05749120 + layer.4.k_cache 0.00078324 1.74742189 + layer.4.v_cache 0.00004769 0.12064034 + layer.4.output 0.18611241 743.06170499 + ------------------------------------------------------------------------------------- + TOTAL 0.08790278 341.24585262 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 10312 +BPFP 0.1298 bits/point +EBPFP 0.2597 equivalent bits/point +MSE 341.245853 +---------------------- -------------------------------------------------------- +Time: 0.356s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.196s +---------------------- -------------------------------------------------------- +💾 Converting with 341.2459 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 696B, BPFP=0.1121 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 812B, BPFP=0.1308 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 816B, BPFP=0.1314 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 764B, BPFP=0.1231 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 792B, BPFP=0.1276 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 776B, BPFP=0.1250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 748B, BPFP=0.1205 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 780B, BPFP=0.1256 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 856B, BPFP=0.1379 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 796B, BPFP=0.1282 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,072B, BPFP=0.0477 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17228427 515.10204575 + layer.0.v_cache 0.00001543 0.04916706 + layer.1.k_cache 0.04621154 37.15412321 + layer.1.v_cache 0.00000545 0.01687020 + layer.2.k_cache 0.00739691 6.57041947 + layer.2.v_cache 0.00001797 0.05050435 + layer.3.k_cache 0.03096496 23.44832776 + layer.3.v_cache 0.00001890 0.06263244 + layer.4.k_cache 0.00071091 1.77429388 + layer.4.v_cache 0.00004900 0.12379288 + layer.4.output 0.02450962 578.54896907 + ------------------------------------------------------------------------------------- + TOTAL 0.02524957 272.59970356 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 9908 +BPFP 0.0939 bits/point +EBPFP 0.1878 equivalent bits/point +MSE 272.599704 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 272.5997 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 688B, BPFP=0.1156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 820B, BPFP=0.1378 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 852B, BPFP=0.1431 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 712B, BPFP=0.1196 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 776B, BPFP=0.1304 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 768B, BPFP=0.1290 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 732B, BPFP=0.1230 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 788B, BPFP=0.1324 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 852B, BPFP=0.1431 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 860B, BPFP=0.1445 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,076B, BPFP=0.0498 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16174046 512.46299563 + layer.0.v_cache 0.00001612 0.04899947 + layer.1.k_cache 0.03352992 36.70098811 + layer.1.v_cache 0.00000563 0.01713897 + layer.2.k_cache 0.00496730 6.56650207 + layer.2.v_cache 0.00001740 0.04880651 + layer.3.k_cache 0.01521110 23.14741368 + layer.3.v_cache 0.00001939 0.06227843 + layer.4.k_cache 0.00066431 1.64746684 + layer.4.v_cache 0.00005067 0.12806082 + layer.4.output 0.14624557 583.97254224 + ------------------------------------------------------------------------------------- + TOTAL 0.07293772 274.62579096 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 9924 +BPFP 0.0981 bits/point +EBPFP 0.1962 equivalent bits/point +MSE 274.625791 +---------------------- -------------------------------------------------------- +Time: 0.357s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 274.6258 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 652B, BPFP=0.1185 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 808B, BPFP=0.1468 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 804B, BPFP=0.1461 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 732B, BPFP=0.1330 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1395 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 736B, BPFP=0.1337 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 728B, BPFP=0.1323 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 740B, BPFP=0.1344 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 844B, BPFP=0.1533 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 808B, BPFP=0.1468 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,496B, BPFP=0.0648 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07121617 527.68404797 + layer.0.v_cache 0.00001681 0.04833441 + layer.1.k_cache 0.06877367 37.34679892 + layer.1.v_cache 0.00000530 0.01628818 + layer.2.k_cache 0.00486504 6.63434406 + layer.2.v_cache 0.00001749 0.04611903 + layer.3.k_cache 0.08833314 23.64286627 + layer.3.v_cache 0.00001723 0.05767114 + layer.4.k_cache 0.00077278 1.67642159 + layer.4.v_cache 0.00004595 0.12444434 + layer.4.output 0.15812202 636.34271179 + ------------------------------------------------------------------------------------- + TOTAL 0.07887751 297.15743050 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 10116 +BPFP 0.1081 bits/point +EBPFP 0.2162 equivalent bits/point +MSE 297.157430 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 297.1574 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 700B, BPFP=0.1083 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 892B, BPFP=0.1380 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 828B, BPFP=0.1281 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 784B, BPFP=0.1213 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 768B, BPFP=0.1188 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 776B, BPFP=0.1200 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 764B, BPFP=0.1182 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 804B, BPFP=0.1244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 880B, BPFP=0.1361 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 840B, BPFP=0.1300 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,252B, BPFP=0.0498 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13964315 512.00355817 + layer.0.v_cache 0.00001432 0.04929923 + layer.1.k_cache 0.04810369 36.99160736 + layer.1.v_cache 0.00000572 0.01738518 + layer.2.k_cache 0.00492308 6.64394326 + layer.2.v_cache 0.00001777 0.04825214 + layer.3.k_cache 0.04255116 23.22504931 + layer.3.v_cache 0.00001880 0.05966839 + layer.4.k_cache 0.00073971 1.68851554 + layer.4.v_cache 0.00004573 0.12687340 + layer.4.output 11.28739116 532.46799859 + ------------------------------------------------------------------------------------- + TOTAL 4.66163537 253.41942012 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 10288 +BPFP 0.0936 bits/point +EBPFP 0.1872 equivalent bits/point +MSE 253.419420 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 253.4194 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 668B, BPFP=0.1186 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 828B, BPFP=0.1470 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 820B, BPFP=0.1456 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 748B, BPFP=0.1328 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 780B, BPFP=0.1385 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 776B, BPFP=0.1378 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 740B, BPFP=0.1314 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 804B, BPFP=0.1428 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 864B, BPFP=0.1534 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 820B, BPFP=0.1456 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,440B, BPFP=0.0619 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14145413 529.70845170 + layer.0.v_cache 0.00001516 0.04937457 + layer.1.k_cache 0.10647225 36.98365922 + layer.1.v_cache 0.00000560 0.01666593 + layer.2.k_cache 0.00244946 6.58317150 + layer.2.v_cache 0.00001685 0.04681359 + layer.3.k_cache 0.01581429 23.18190141 + layer.3.v_cache 0.00001841 0.05778968 + layer.4.k_cache 0.00072023 1.67398401 + layer.4.v_cache 0.00004652 0.11924130 + layer.4.output 0.15451367 619.09750406 + ------------------------------------------------------------------------------------- + TOTAL 0.07932992 290.12374008 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 10288 +BPFP 0.1075 bits/point +EBPFP 0.2149 equivalent bits/point +MSE 290.123740 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1237 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 640B, BPFP=0.1351 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 760B, BPFP=0.1605 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 788B, BPFP=0.1664 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 680B, BPFP=0.1436 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 728B, BPFP=0.1537 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 692B, BPFP=0.1461 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 680B, BPFP=0.1436 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 704B, BPFP=0.1486 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 792B, BPFP=0.1672 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 788B, BPFP=0.1664 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,628B, BPFP=0.0793 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.197s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10237078 532.61840160 + layer.0.v_cache 0.00001391 0.04687464 + layer.1.k_cache 0.05953315 38.33656765 + layer.1.v_cache 0.00000517 0.01585480 + layer.2.k_cache 0.00728044 6.81187068 + layer.2.v_cache 0.00001648 0.04483541 + layer.3.k_cache 0.03101032 23.72049857 + layer.3.v_cache 0.00001666 0.05489744 + layer.4.k_cache 0.00072555 1.66677176 + layer.4.v_cache 0.00004870 0.12377356 + layer.4.output 0.18364459 733.02925917 + ------------------------------------------------------------------------------------- + TOTAL 0.08744314 337.33206825 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 9880 +BPFP 0.1227 bits/point +EBPFP 0.2454 equivalent bits/point +MSE 337.332068 +---------------------- -------------------------------------------------------- +Time: 0.356s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.197s +---------------------- -------------------------------------------------------- +💾 Converting with 337.3321 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 111, 128) +Output shape: (1, 111, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.output: torch.Size([1, 111, 3584]) -> torch.Size([1, 1, 111, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 764B, BPFP=0.1075 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 848B, BPFP=0.1194 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 880B, BPFP=0.1239 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 764B, BPFP=0.1075 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 800B, BPFP=0.1126 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 784B, BPFP=0.1104 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 808B, BPFP=0.1137 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 804B, BPFP=0.1132 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 888B, BPFP=0.1250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 864B, BPFP=0.1216 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,364B, BPFP=0.0475 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11107565 525.88949887 + layer.0.v_cache 0.00001597 0.04790534 + layer.1.k_cache 0.04333284 37.18999859 + layer.1.v_cache 0.00000558 0.01665729 + layer.2.k_cache 0.00477623 6.58741334 + layer.2.v_cache 0.00001758 0.04664574 + layer.3.k_cache 0.01020195 23.49360396 + layer.3.v_cache 0.00001870 0.05667361 + layer.4.k_cache 0.00079488 1.69326507 + layer.4.v_cache 0.00005041 0.13116574 + layer.4.output 10.27057757 485.73934202 + ------------------------------------------------------------------------------------- + TOTAL 4.23907840 235.01930716 + (elements=966,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 966144 +Total Bytes 10568 +BPFP 0.0875 bits/point +EBPFP 0.1750 equivalent bits/point +MSE 235.019307 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.006s, Pack+Encode: 0.157s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 235.0193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 656B, BPFP=0.1444 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 772B, BPFP=0.1699 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 788B, BPFP=0.1734 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 684B, BPFP=0.1505 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 732B, BPFP=0.1611 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 704B, BPFP=0.1549 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 696B, BPFP=0.1532 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 728B, BPFP=0.1602 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 796B, BPFP=0.1752 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 764B, BPFP=0.1681 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,644B, BPFP=0.0831 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09375787 527.13407790 + layer.0.v_cache 0.00001622 0.04664076 + layer.1.k_cache 0.05984993 38.12248638 + layer.1.v_cache 0.00000512 0.01597733 + layer.2.k_cache 0.00243208 6.85684247 + layer.2.v_cache 0.00001865 0.04948250 + layer.3.k_cache 0.04944374 23.43622084 + layer.3.v_cache 0.00001908 0.06131211 + layer.4.k_cache 0.00084877 1.75343237 + layer.4.v_cache 0.00005139 0.13202057 + layer.4.output 0.19371969 765.60368461 + ------------------------------------------------------------------------------------- + TOTAL 0.09191063 350.40201679 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 9964 +BPFP 0.1290 bits/point +EBPFP 0.2580 equivalent bits/point +MSE 350.402017 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 350.4020 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 109, 128) +Output shape: (1, 109, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.output: torch.Size([1, 109, 3584]) -> torch.Size([1, 1, 109, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 752B, BPFP=0.1078 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 880B, BPFP=0.1261 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 868B, BPFP=0.1244 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 792B, BPFP=0.1135 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 836B, BPFP=0.1198 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 792B, BPFP=0.1135 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 804B, BPFP=0.1153 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 840B, BPFP=0.1204 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 908B, BPFP=0.1302 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 860B, BPFP=0.1233 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,516B, BPFP=0.0515 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860953 529.65320384 + layer.0.v_cache 0.00001500 0.04746254 + layer.1.k_cache 0.05917679 37.38574443 + layer.1.v_cache 0.00000528 0.01676900 + layer.2.k_cache 0.01253135 6.64273883 + layer.2.v_cache 0.00001869 0.04905081 + layer.3.k_cache 0.07406004 23.57519755 + layer.3.v_cache 0.00001854 0.06066650 + layer.4.k_cache 0.00071234 1.71320826 + layer.4.v_cache 0.00004916 0.13037384 + layer.4.output 10.45909253 492.96977392 + ------------------------------------------------------------------------------------- + TOTAL 4.32287320 238.23899018 + (elements=948,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 948736 +Total Bytes 10848 +BPFP 0.0915 bits/point +EBPFP 0.1829 equivalent bits/point +MSE 238.238990 +---------------------- -------------------------------------------------------- +Time: 0.377s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 238.2390 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst + to output-fixed/kimiaudio/lambda0.001/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.1093 bits/point +Avg EBPFP 0.2185 equivalent bits/point +Avg MSE 351.529627 +Avg Time 0.347s +------------------------ ---------------------------- diff --git a/lambda0.004/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.004/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..d178d145fa6729db5a6c3f9c4477f904b80be81c --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 255 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- -------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench +Output output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench +---------------- -------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.3588 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,320B, BPFP=2.3765 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,056B, BPFP=0.5895 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,896B, BPFP=2.2948 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,016B, BPFP=0.9676 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,632B, BPFP=2.2438 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,860B, BPFP=0.7446 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,756B, BPFP=2.2677 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,116B, BPFP=2.1443 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,256B, BPFP=2.1713 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,972B, BPFP=0.1095 +⌛️ [2/4] FRONTEND: Frontend time: 2.011s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.053s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12089630 52.74350164 + layer.0.v_cache 0.00001394 0.01269992 + layer.1.k_cache 0.01174029 6.27670514 + layer.1.v_cache 0.00000579 0.00516914 + layer.2.k_cache 0.00835781 1.13821308 + layer.2.v_cache 0.00001852 0.01557265 + layer.3.k_cache 0.02896961 4.41758219 + layer.3.v_cache 0.00001872 0.01833137 + layer.4.k_cache 0.00063445 0.40759593 + layer.4.v_cache 0.00005063 0.03517684 + layer.4.output 0.17246709 668.58989198 + ------------------------------------------------------------------------------------- + TOTAL 0.08105739 279.12939951 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 87740 +BPFP 0.9956 bits/point +EBPFP 1.9912 equivalent bits/point +MSE 279.129400 +---------------------- -------------------------------------------------------- +Time: 3.067s Load: 0.004s, Pack+Encode: 2.011s, Decode+Unpack: 1.053s +---------------------- -------------------------------------------------------- +💾 Converting with 279.1294 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3625 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,380B, BPFP=2.4180 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,048B, BPFP=0.5953 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,232B, BPFP=2.3891 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,696B, BPFP=0.7219 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,892B, BPFP=2.3227 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,192B, BPFP=0.6234 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,040B, BPFP=2.3516 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,560B, BPFP=1.8672 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,456B, BPFP=2.2375 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,608B, BPFP=0.1007 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.026s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08029544 56.47017822 + layer.0.v_cache 0.00001376 0.01281122 + layer.1.k_cache 0.03632395 6.64041595 + layer.1.v_cache 0.00000563 0.00535662 + layer.2.k_cache 0.00522719 1.13559246 + layer.2.v_cache 0.00001997 0.01673213 + layer.3.k_cache 0.02707962 4.29460411 + layer.3.v_cache 0.00001843 0.01877384 + layer.4.k_cache 0.00063057 0.39691586 + layer.4.v_cache 0.00005085 0.03723103 + layer.4.output 0.18451144 675.63186384 + ------------------------------------------------------------------------------------- + TOTAL 0.08477915 282.26186225 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 84960 +BPFP 0.9761 bits/point +EBPFP 1.9522 equivalent bits/point +MSE 282.261862 +---------------------- -------------------------------------------------------- +Time: 2.528s Load: 0.003s, Pack+Encode: 1.498s, Decode+Unpack: 1.026s +---------------------- -------------------------------------------------------- +💾 Converting with 282.2619 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,852B, BPFP=0.3365 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,408B, BPFP=2.2544 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,144B, BPFP=0.5712 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,044B, BPFP=2.1882 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,068B, BPFP=0.9208 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,720B, BPFP=2.1294 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,748B, BPFP=0.8626 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,788B, BPFP=2.1417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,236B, BPFP=1.8597 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,448B, BPFP=2.0799 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,020B, BPFP=0.1043 +⌛️ [2/4] FRONTEND: Frontend time: 1.514s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10874708 50.84092251 + layer.0.v_cache 0.00001608 0.01293832 + layer.1.k_cache 0.03270756 6.24055942 + layer.1.v_cache 0.00000572 0.00536918 + layer.2.k_cache 0.00777350 1.17187589 + layer.2.v_cache 0.00002080 0.01524295 + layer.3.k_cache 0.05920834 5.43856386 + layer.3.v_cache 0.00001846 0.01797741 + layer.4.k_cache 0.00062493 0.39337770 + layer.4.v_cache 0.00005036 0.03368196 + layer.4.output 0.17137358 629.63237126 + ------------------------------------------------------------------------------------- + TOTAL 0.08286988 263.03512400 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 88476 +BPFP 0.9456 bits/point +EBPFP 1.8912 equivalent bits/point +MSE 263.035124 +---------------------- -------------------------------------------------------- +Time: 2.538s Load: 0.004s, Pack+Encode: 1.514s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 263.0351 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3606 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,072B, BPFP=2.4183 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,996B, BPFP=0.6002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,044B, BPFP=2.4127 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,424B, BPFP=0.6859 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,388B, BPFP=2.2812 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,172B, BPFP=0.6354 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,808B, BPFP=2.3654 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,272B, BPFP=1.6571 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,832B, BPFP=2.1699 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,708B, BPFP=0.1061 +⌛️ [2/4] FRONTEND: Frontend time: 1.501s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.022s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12073612 51.87618940 + layer.0.v_cache 0.00001348 0.01248669 + layer.1.k_cache 0.03279715 6.58799470 + layer.1.v_cache 0.00000568 0.00525046 + layer.2.k_cache 0.00696802 1.00866210 + layer.2.v_cache 0.00001833 0.01548475 + layer.3.k_cache 0.03227850 4.23727652 + layer.3.v_cache 0.00001834 0.01832237 + layer.4.k_cache 0.00060115 0.39887179 + layer.4.v_cache 0.00005191 0.03528062 + layer.4.output 0.18320683 691.07566392 + ------------------------------------------------------------------------------------- + TOTAL 0.08681979 288.33679217 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 81516 +BPFP 0.9605 bits/point +EBPFP 1.9211 equivalent bits/point +MSE 288.336792 +---------------------- -------------------------------------------------------- +Time: 2.527s Load: 0.003s, Pack+Encode: 1.501s, Decode+Unpack: 1.022s +---------------------- -------------------------------------------------------- +💾 Converting with 288.3368 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,796B, BPFP=0.3598 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,120B, BPFP=2.4279 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,996B, BPFP=0.6002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,912B, BPFP=2.3862 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,396B, BPFP=0.6803 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,084B, BPFP=2.2204 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,304B, BPFP=0.6619 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,636B, BPFP=2.3309 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,176B, BPFP=1.6378 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,952B, BPFP=2.1939 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,728B, BPFP=0.1067 +⌛️ [2/4] FRONTEND: Frontend time: 1.515s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.064s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10595027 52.63249324 + layer.0.v_cache 0.00001376 0.01281243 + layer.1.k_cache 0.03307601 6.10416432 + layer.1.v_cache 0.00000550 0.00508875 + layer.2.k_cache 0.00695119 1.12336516 + layer.2.v_cache 0.00001929 0.01522827 + layer.3.k_cache 0.03450640 4.70832629 + layer.3.v_cache 0.00001844 0.01855861 + layer.4.k_cache 0.00061739 0.37465878 + layer.4.v_cache 0.00005297 0.03529232 + layer.4.output 0.18589628 691.13221154 + ------------------------------------------------------------------------------------- + TOTAL 0.08720501 288.40914523 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 81100 +BPFP 0.9556 bits/point +EBPFP 1.9113 equivalent bits/point +MSE 288.409145 +---------------------- -------------------------------------------------------- +Time: 2.583s Load: 0.004s, Pack+Encode: 1.515s, Decode+Unpack: 1.064s +---------------------- -------------------------------------------------------- +💾 Converting with 288.4091 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3639 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,376B, BPFP=2.4478 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,008B, BPFP=0.5949 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,000B, BPFP=2.3734 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,276B, BPFP=0.8457 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,632B, BPFP=2.3006 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,088B, BPFP=0.6108 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,948B, BPFP=2.3631 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,232B, BPFP=1.8259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,320B, BPFP=2.2389 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,408B, BPFP=0.0963 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07772056 53.55680133 + layer.0.v_cache 0.00001351 0.01334550 + layer.1.k_cache 0.03518496 6.06440291 + layer.1.v_cache 0.00000556 0.00532227 + layer.2.k_cache 0.00230807 1.09863204 + layer.2.v_cache 0.00001833 0.01637265 + layer.3.k_cache 0.04514034 4.94788911 + layer.3.v_cache 0.00001901 0.01881415 + layer.4.k_cache 0.00062178 0.40666715 + layer.4.v_cache 0.00004990 0.03748794 + layer.4.output 0.18426369 682.90545886 + ------------------------------------------------------------------------------------- + TOTAL 0.08534870 285.08846748 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 84128 +BPFP 0.9788 bits/point +EBPFP 1.9576 equivalent bits/point +MSE 285.088467 +---------------------- -------------------------------------------------------- +Time: 2.508s Load: 0.004s, Pack+Encode: 1.498s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 285.0885 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,780B, BPFP=0.3612 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,296B, BPFP=2.4951 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,956B, BPFP=0.5998 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,856B, BPFP=2.4058 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,604B, BPFP=0.7313 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,612B, BPFP=2.3563 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,152B, BPFP=0.6396 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,876B, BPFP=2.4099 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,192B, BPFP=1.8653 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,940B, BPFP=2.2200 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,692B, BPFP=0.1070 +⌛️ [2/4] FRONTEND: Frontend time: 1.516s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14568528 54.47117238 + layer.0.v_cache 0.00001322 0.01261854 + layer.1.k_cache 0.01379078 6.19312158 + layer.1.v_cache 0.00000548 0.00516117 + layer.2.k_cache 0.00727733 1.28947865 + layer.2.v_cache 0.00001786 0.01613841 + layer.3.k_cache 0.03082054 5.12502933 + layer.3.v_cache 0.00001879 0.01925729 + layer.4.k_cache 0.00061820 0.40497133 + layer.4.v_cache 0.00005377 0.03714617 + layer.4.output 0.19662610 701.20025510 + ------------------------------------------------------------------------------------- + TOTAL 0.09262847 292.70446356 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 82956 +BPFP 0.9902 bits/point +EBPFP 1.9804 equivalent bits/point +MSE 292.704464 +---------------------- -------------------------------------------------------- +Time: 2.535s Load: 0.003s, Pack+Encode: 1.516s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 292.7045 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,824B, BPFP=0.3563 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,308B, BPFP=2.4039 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,084B, BPFP=0.6023 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,024B, BPFP=2.3484 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,668B, BPFP=0.7164 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,644B, BPFP=2.2742 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,568B, BPFP=0.6969 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,768B, BPFP=2.2984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,768B, BPFP=1.9078 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,372B, BPFP=2.2211 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,980B, BPFP=0.1110 +⌛️ [2/4] FRONTEND: Frontend time: 1.487s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10149734 51.29268188 + layer.0.v_cache 0.00001421 0.01363844 + layer.1.k_cache 0.01324784 5.53975983 + layer.1.v_cache 0.00000586 0.00559430 + layer.2.k_cache 0.00706157 1.59800148 + layer.2.v_cache 0.00001822 0.01569878 + layer.3.k_cache 0.02712160 5.23167610 + layer.3.v_cache 0.00002016 0.01873830 + layer.4.k_cache 0.00063687 0.38120153 + layer.4.v_cache 0.00005451 0.03652930 + layer.4.output 0.17858309 671.93281250 + ------------------------------------------------------------------------------------- + TOTAL 0.08233881 280.45077691 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 85008 +BPFP 0.9767 bits/point +EBPFP 1.9533 equivalent bits/point +MSE 280.450777 +---------------------- -------------------------------------------------------- +Time: 2.496s Load: 0.004s, Pack+Encode: 1.487s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 280.4508 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,880B, BPFP=0.3376 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,468B, BPFP=2.2392 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,136B, BPFP=0.5632 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,136B, BPFP=2.1796 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,476B, BPFP=0.9835 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,756B, BPFP=2.1114 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,780B, BPFP=1.2177 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,956B, BPFP=2.1473 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,136B, BPFP=2.0000 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,600B, BPFP=2.0833 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,708B, BPFP=0.0951 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10772817 55.04314835 + layer.0.v_cache 0.00001348 0.01252405 + layer.1.k_cache 0.03276526 6.45123501 + layer.1.v_cache 0.00000562 0.00535485 + layer.2.k_cache 0.01020141 1.11343840 + layer.2.v_cache 0.00001934 0.01576335 + layer.3.k_cache 0.04694761 7.29320728 + layer.3.v_cache 0.00001910 0.01992259 + layer.4.k_cache 0.00062564 0.39311692 + layer.4.v_cache 0.00005323 0.03539097 + layer.4.output 0.17166949 622.67939245 + ------------------------------------------------------------------------------------- + TOTAL 0.08235678 260.53757935 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 92032 +BPFP 0.9723 bits/point +EBPFP 1.9446 equivalent bits/point +MSE 260.537579 +---------------------- -------------------------------------------------------- +Time: 2.518s Load: 0.003s, Pack+Encode: 1.500s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 260.5376 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.3502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,468B, BPFP=2.3471 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,044B, BPFP=0.5730 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,124B, BPFP=2.2824 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,400B, BPFP=1.0166 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,768B, BPFP=2.2154 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,584B, BPFP=0.6747 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,848B, BPFP=2.2304 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,340B, BPFP=1.7583 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,472B, BPFP=2.1596 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,484B, BPFP=0.0937 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12335577 52.62249388 + layer.0.v_cache 0.00001400 0.01274721 + layer.1.k_cache 0.01519604 5.34803312 + layer.1.v_cache 0.00000547 0.00512537 + layer.2.k_cache 0.01098718 1.11133566 + layer.2.v_cache 0.00001779 0.01569833 + layer.3.k_cache 0.07416311 4.96758776 + layer.3.v_cache 0.00001944 0.01890184 + layer.4.k_cache 0.00061719 0.40448623 + layer.4.v_cache 0.00006444 0.03598346 + layer.4.output 0.17504946 651.26425344 + ------------------------------------------------------------------------------------- + TOTAL 0.08528157 271.96424512 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 86392 +BPFP 0.9567 bits/point +EBPFP 1.9134 equivalent bits/point +MSE 271.964245 +---------------------- -------------------------------------------------------- +Time: 2.502s Load: 0.004s, Pack+Encode: 1.497s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 271.9642 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,776B, BPFP=0.3604 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,140B, BPFP=2.4635 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,984B, BPFP=0.6055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,764B, BPFP=2.3872 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,372B, BPFP=0.6843 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,344B, BPFP=2.3019 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,012B, BPFP=0.6112 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,700B, BPFP=2.3742 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,428B, BPFP=1.7102 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,336B, BPFP=2.3003 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,564B, BPFP=0.1033 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13955514 56.97982828 + layer.0.v_cache 0.00001392 0.01309958 + layer.1.k_cache 0.03454666 6.56130308 + layer.1.v_cache 0.00000526 0.00538263 + layer.2.k_cache 0.00236268 1.08102853 + layer.2.v_cache 0.00001660 0.01522747 + layer.3.k_cache 0.02874016 3.99799852 + layer.3.v_cache 0.00001899 0.01916002 + layer.4.k_cache 0.00062191 0.40364853 + layer.4.v_cache 0.00005265 0.03773721 + layer.4.output 0.18728454 701.01107375 + ------------------------------------------------------------------------------------- + TOTAL 0.08923092 292.71717236 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 81420 +BPFP 0.9719 bits/point +EBPFP 1.9438 equivalent bits/point +MSE 292.717172 +---------------------- -------------------------------------------------------- +Time: 2.506s Load: 0.004s, Pack+Encode: 1.496s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 292.7172 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,784B, BPFP=0.3620 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,356B, BPFP=2.5073 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,996B, BPFP=0.6080 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,172B, BPFP=2.4700 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,540B, BPFP=0.7183 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,524B, BPFP=2.3385 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,072B, BPFP=0.6234 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,792B, BPFP=2.3929 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,972B, BPFP=1.8206 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,324B, BPFP=2.2979 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,688B, BPFP=0.1069 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11454884 51.98205408 + layer.0.v_cache 0.00001351 0.01353175 + layer.1.k_cache 0.01665594 6.81167325 + layer.1.v_cache 0.00000570 0.00580988 + layer.2.k_cache 0.00836552 1.32867293 + layer.2.v_cache 0.00001974 0.01672034 + layer.3.k_cache 0.03124422 4.85949350 + layer.3.v_cache 0.00001876 0.01906481 + layer.4.k_cache 0.00062231 0.40715465 + layer.4.v_cache 0.00005487 0.03710329 + layer.4.output 0.19663499 699.25185529 + ------------------------------------------------------------------------------------- + TOTAL 0.09105849 291.77907444 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 83220 +BPFP 0.9934 bits/point +EBPFP 1.9867 equivalent bits/point +MSE 291.779074 +---------------------- -------------------------------------------------------- +Time: 2.509s Load: 0.004s, Pack+Encode: 1.498s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 291.7791 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.3438 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,320B, BPFP=2.2917 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,088B, BPFP=0.5744 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,064B, BPFP=2.2440 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,840B, BPFP=0.9003 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,732B, BPFP=2.1823 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,296B, BPFP=0.6131 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,904B, BPFP=2.2143 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,868B, BPFP=1.6496 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,432B, BPFP=2.1265 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,740B, BPFP=0.0994 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12290487 52.98843238 + layer.0.v_cache 0.00001422 0.01292302 + layer.1.k_cache 0.01283375 5.89707002 + layer.1.v_cache 0.00000579 0.00549666 + layer.2.k_cache 0.01100613 1.21143623 + layer.2.v_cache 0.00001929 0.01564045 + layer.3.k_cache 0.03011658 4.50145758 + layer.3.v_cache 0.00001858 0.01866945 + layer.4.k_cache 0.00063020 0.41109703 + layer.4.v_cache 0.00005223 0.03627384 + layer.4.output 0.17536174 644.54979804 + ------------------------------------------------------------------------------------- + TOTAL 0.08265493 269.23218135 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 85132 +BPFP 0.9315 bits/point +EBPFP 1.8630 equivalent bits/point +MSE 269.232181 +---------------------- -------------------------------------------------------- +Time: 2.500s Load: 0.004s, Pack+Encode: 1.491s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 269.2322 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3412 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,516B, BPFP=2.3007 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,128B, BPFP=0.5750 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,064B, BPFP=2.2176 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,784B, BPFP=1.0632 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,784B, BPFP=2.1662 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,080B, BPFP=1.3015 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,904B, BPFP=2.1882 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,848B, BPFP=1.9941 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,440B, BPFP=2.1029 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,740B, BPFP=0.0982 +⌛️ [2/4] FRONTEND: Frontend time: 1.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11218034 52.71287914 + layer.0.v_cache 0.00001438 0.01277183 + layer.1.k_cache 0.03157643 6.21805061 + layer.1.v_cache 0.00000545 0.00526557 + layer.2.k_cache 0.00802605 1.10646712 + layer.2.v_cache 0.00001826 0.01533846 + layer.3.k_cache 0.04519741 7.29943704 + layer.3.v_cache 0.00001794 0.01778168 + layer.4.k_cache 0.00061137 0.38808091 + layer.4.v_cache 0.00005913 0.03603825 + layer.4.output 0.17010492 632.76024160 + ------------------------------------------------------------------------------------- + TOTAL 0.08167301 264.53728246 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 92144 +BPFP 0.9964 bits/point +EBPFP 1.9927 equivalent bits/point +MSE 264.537282 +---------------------- -------------------------------------------------------- +Time: 2.498s Load: 0.004s, Pack+Encode: 1.488s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 264.5373 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,864B, BPFP=0.3387 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,424B, BPFP=2.2573 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,148B, BPFP=0.5719 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,056B, BPFP=2.1904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,172B, BPFP=0.9397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,780B, BPFP=2.1403 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,808B, BPFP=1.2369 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,912B, BPFP=2.1642 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,668B, BPFP=1.7565 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,496B, BPFP=2.0887 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,844B, BPFP=0.0998 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10146660 52.38980174 + layer.0.v_cache 0.00001344 0.01245233 + layer.1.k_cache 0.03090366 6.60546023 + layer.1.v_cache 0.00000566 0.00521708 + layer.2.k_cache 0.01188999 0.94654190 + layer.2.v_cache 0.00001845 0.01506002 + layer.3.k_cache 0.02542871 6.83374520 + layer.3.v_cache 0.00001884 0.01881227 + layer.4.k_cache 0.00060243 0.41540115 + layer.4.v_cache 0.00005224 0.03498670 + layer.4.output 0.16839122 629.01541736 + ------------------------------------------------------------------------------------- + TOTAL 0.07936109 262.96384707 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 90172 +BPFP 0.9637 bits/point +EBPFP 1.9274 equivalent bits/point +MSE 262.963847 +---------------------- -------------------------------------------------------- +Time: 2.506s Load: 0.004s, Pack+Encode: 1.497s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 262.9638 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,928B, BPFP=0.3274 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,404B, BPFP=2.1067 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,240B, BPFP=0.5503 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,020B, BPFP=2.0414 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,692B, BPFP=0.9667 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,628B, BPFP=1.9749 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,388B, BPFP=1.2548 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,852B, BPFP=2.0129 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,364B, BPFP=1.9300 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,420B, BPFP=1.9395 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,816B, BPFP=0.0926 +⌛️ [2/4] FRONTEND: Frontend time: 1.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12981709 57.07852836 + layer.0.v_cache 0.00001628 0.01297663 + layer.1.k_cache 0.06761242 5.76210354 + layer.1.v_cache 0.00000547 0.00509611 + layer.2.k_cache 0.00384689 1.10442949 + layer.2.v_cache 0.00001812 0.01617373 + layer.3.k_cache 0.05327560 6.52595321 + layer.3.v_cache 0.00001896 0.01940136 + layer.4.k_cache 0.00060800 0.39149085 + layer.4.v_cache 0.00005116 0.03539921 + layer.4.output 0.15518735 587.87718362 + ------------------------------------------------------------------------------------- + TOTAL 0.07891655 246.24069634 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 92752 +BPFP 0.9266 bits/point +EBPFP 1.8533 equivalent bits/point +MSE 246.240696 +---------------------- -------------------------------------------------------- +Time: 2.499s Load: 0.004s, Pack+Encode: 1.488s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 246.2407 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3549 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,524B, BPFP=2.4159 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,032B, BPFP=0.5849 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,076B, BPFP=2.3295 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,004B, BPFP=0.7724 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,684B, BPFP=2.2539 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,608B, BPFP=0.6960 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,956B, BPFP=2.3063 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,580B, BPFP=2.0409 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,404B, BPFP=2.1998 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,484B, BPFP=0.0960 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887507 54.56002845 + layer.0.v_cache 0.00001345 0.01299427 + layer.1.k_cache 0.03259810 5.93295627 + layer.1.v_cache 0.00000523 0.00515704 + layer.2.k_cache 0.00244350 1.11079049 + layer.2.v_cache 0.00001834 0.01581270 + layer.3.k_cache 0.10864470 5.19389061 + layer.3.v_cache 0.00001904 0.01879439 + layer.4.k_cache 0.00060864 0.38135491 + layer.4.v_cache 0.00004940 0.03493806 + layer.4.output 0.18628448 666.80820106 + ------------------------------------------------------------------------------------- + TOTAL 0.09395687 278.52494851 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 86192 +BPFP 0.9780 bits/point +EBPFP 1.9561 equivalent bits/point +MSE 278.524949 +---------------------- -------------------------------------------------------- +Time: 2.503s Load: 0.003s, Pack+Encode: 1.493s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 278.5249 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,928B, BPFP=0.3385 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,456B, BPFP=2.1868 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,208B, BPFP=0.5632 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,012B, BPFP=2.1088 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,660B, BPFP=0.8181 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,632B, BPFP=2.0421 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,320B, BPFP=1.1096 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,944B, BPFP=2.0969 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,704B, BPFP=1.8792 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,536B, BPFP=2.0253 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,072B, BPFP=0.1021 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.032s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422350 53.01920207 + layer.0.v_cache 0.00001493 0.01337418 + layer.1.k_cache 0.01422887 6.36133841 + layer.1.v_cache 0.00000595 0.00549189 + layer.2.k_cache 0.00668612 1.12389845 + layer.2.v_cache 0.00001902 0.01604673 + layer.3.k_cache 0.02497406 6.43889489 + layer.3.v_cache 0.00001869 0.01922476 + layer.4.k_cache 0.00061935 0.38959117 + layer.4.v_cache 0.00005212 0.03524953 + layer.4.output 0.16899524 608.42902287 + ------------------------------------------------------------------------------------- + TOTAL 0.07904761 254.49561601 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 90472 +BPFP 0.9343 bits/point +EBPFP 1.8686 equivalent bits/point +MSE 254.495616 +---------------------- -------------------------------------------------------- +Time: 2.530s Load: 0.003s, Pack+Encode: 1.495s, Decode+Unpack: 1.032s +---------------------- -------------------------------------------------------- +💾 Converting with 254.4956 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 113, 128) +Output shape: (1, 113, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.output: torch.Size([1, 113, 3584]) -> torch.Size([1, 1, 113, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,360B, BPFP=0.3263 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,384B, BPFP=1.7124 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,992B, BPFP=0.5520 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,156B, BPFP=1.6809 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,964B, BPFP=0.8247 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,684B, BPFP=1.6156 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,636B, BPFP=0.6410 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,780B, BPFP=1.6289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,896B, BPFP=1.5066 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,388B, BPFP=1.5747 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,156B, BPFP=0.0821 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10214825 55.67048361 + layer.0.v_cache 0.00001637 0.01247713 + layer.1.k_cache 0.04284562 6.51276823 + layer.1.v_cache 0.00000578 0.00511969 + layer.2.k_cache 0.00913706 1.03510676 + layer.2.v_cache 0.00001928 0.01492245 + layer.3.k_cache 0.05585717 5.21748447 + layer.3.v_cache 0.00001870 0.01717432 + layer.4.k_cache 0.00065968 0.42223956 + layer.4.v_cache 0.00005753 0.03490188 + layer.4.output 10.09706179 472.74727402 + ------------------------------------------------------------------------------------- + TOTAL 4.17001165 198.71609390 + (elements=983,552) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 983552 +Total Bytes 91396 +BPFP 0.7434 bits/point +EBPFP 1.4868 equivalent bits/point +MSE 198.716094 +---------------------- -------------------------------------------------------- +Time: 2.507s Load: 0.006s, Pack+Encode: 1.492s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 198.7161 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,908B, BPFP=0.3276 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,380B, BPFP=2.1257 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,196B, BPFP=0.5488 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,948B, BPFP=2.0515 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,292B, BPFP=0.7370 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,688B, BPFP=2.0069 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,536B, BPFP=0.9505 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,796B, BPFP=2.0254 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,056B, BPFP=1.7266 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,404B, BPFP=1.9581 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,200B, BPFP=0.1030 +⌛️ [2/4] FRONTEND: Frontend time: 1.505s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11481400 56.31433508 + layer.0.v_cache 0.00001396 0.01351635 + layer.1.k_cache 0.03135788 6.03321050 + layer.1.v_cache 0.00000545 0.00538388 + layer.2.k_cache 0.00370715 1.02931213 + layer.2.v_cache 0.00001965 0.01656839 + layer.3.k_cache 0.03930106 6.71140231 + layer.3.v_cache 0.00001886 0.01985210 + layer.4.k_cache 0.00060819 0.40775027 + layer.4.v_cache 0.00005437 0.03762359 + layer.4.output 0.16591480 593.78002355 + ------------------------------------------------------------------------------------- + TOTAL 0.07948848 248.64994820 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 88404 +BPFP 0.8929 bits/point +EBPFP 1.7858 equivalent bits/point +MSE 248.649948 +---------------------- -------------------------------------------------------- +Time: 2.515s Load: 0.004s, Pack+Encode: 1.505s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 248.6499 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,892B, BPFP=0.3322 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,428B, BPFP=2.1819 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,208B, BPFP=0.5632 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,288B, BPFP=2.1573 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,316B, BPFP=0.7577 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,592B, BPFP=2.0351 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,220B, BPFP=0.5653 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,796B, BPFP=2.0709 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,248B, BPFP=1.9747 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,548B, BPFP=2.0274 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,548B, BPFP=0.0890 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09788724 55.56409103 + layer.0.v_cache 0.00001308 0.01294912 + layer.1.k_cache 0.03299916 6.04512555 + layer.1.v_cache 0.00000544 0.00529584 + layer.2.k_cache 0.00367894 1.06037406 + layer.2.v_cache 0.00002012 0.01650887 + layer.3.k_cache 0.03939044 4.98044715 + layer.3.v_cache 0.00001875 0.01866553 + layer.4.k_cache 0.00061788 0.39850368 + layer.4.v_cache 0.00005010 0.03719593 + layer.4.output 0.17021053 608.30658106 + ------------------------------------------------------------------------------------- + TOTAL 0.08036205 254.48736613 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 87084 +BPFP 0.8993 bits/point +EBPFP 1.7987 equivalent bits/point +MSE 254.487366 +---------------------- -------------------------------------------------------- +Time: 2.501s Load: 0.004s, Pack+Encode: 1.493s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 254.4874 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,272B, BPFP=0.3287 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,420B, BPFP=1.7969 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,920B, BPFP=0.5671 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,048B, BPFP=1.7431 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,400B, BPFP=0.7812 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,676B, BPFP=1.6892 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,116B, BPFP=1.0295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,976B, BPFP=1.7326 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,428B, BPFP=1.6534 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,388B, BPFP=1.6476 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,576B, BPFP=0.1152 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12672405 53.64374457 + layer.0.v_cache 0.00001401 0.01252894 + layer.1.k_cache 0.12044146 6.73812075 + layer.1.v_cache 0.00000568 0.00525466 + layer.2.k_cache 0.00570256 1.18672816 + layer.2.v_cache 0.00001866 0.01616808 + layer.3.k_cache 0.03611955 5.83215389 + layer.3.v_cache 0.00001983 0.01962569 + layer.4.k_cache 0.00066233 0.42424940 + layer.4.v_cache 0.00005331 0.03746427 + layer.4.output 10.56037249 495.79691634 + ------------------------------------------------------------------------------------- + TOTAL 4.36543346 208.14673252 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 95220 +BPFP 0.8104 bits/point +EBPFP 1.6207 equivalent bits/point +MSE 208.146733 +---------------------- -------------------------------------------------------- +Time: 2.501s Load: 0.005s, Pack+Encode: 1.492s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 208.1467 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,900B, BPFP=0.3374 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,520B, BPFP=2.2230 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,144B, BPFP=0.5582 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,160B, BPFP=2.1591 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,880B, BPFP=0.8665 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,740B, BPFP=2.0845 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,324B, BPFP=0.7678 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,812B, BPFP=2.0973 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,380B, BPFP=2.0206 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,376B, BPFP=2.0199 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,624B, BPFP=0.0919 +⌛️ [2/4] FRONTEND: Frontend time: 1.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13733016 53.67148105 + layer.0.v_cache 0.00001373 0.01322337 + layer.1.k_cache 0.03155638 6.04841891 + layer.1.v_cache 0.00000617 0.00564793 + layer.2.k_cache 0.00751131 1.00331185 + layer.2.v_cache 0.00001754 0.01596292 + layer.3.k_cache 0.02548880 5.26341802 + layer.3.v_cache 0.00001765 0.01927696 + layer.4.k_cache 0.00061011 0.41413433 + layer.4.v_cache 0.00005045 0.03702028 + layer.4.output 0.16239834 614.71966315 + ------------------------------------------------------------------------------------- + TOTAL 0.07878769 257.03114927 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 88860 +BPFP 0.9281 bits/point +EBPFP 1.8562 equivalent bits/point +MSE 257.031149 +---------------------- -------------------------------------------------------- +Time: 2.498s Load: 0.005s, Pack+Encode: 1.488s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 257.0311 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3639 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,360B, BPFP=2.4446 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,968B, BPFP=0.5870 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,904B, BPFP=2.3544 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,400B, BPFP=0.6725 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,584B, BPFP=2.2911 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,404B, BPFP=0.6733 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,732B, BPFP=2.3204 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,732B, BPFP=1.9248 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,292B, BPFP=2.2334 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,680B, BPFP=0.1040 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10979201 52.19081908 + layer.0.v_cache 0.00001339 0.01278442 + layer.1.k_cache 0.01160004 6.27935250 + layer.1.v_cache 0.00000567 0.00535543 + layer.2.k_cache 0.00714037 1.18460450 + layer.2.v_cache 0.00001829 0.01542214 + layer.3.k_cache 0.04622446 4.32605521 + layer.3.v_cache 0.00001896 0.01893567 + layer.4.k_cache 0.00064378 0.39384007 + layer.4.v_cache 0.00005251 0.03614778 + layer.4.output 0.18415536 683.44416817 + ------------------------------------------------------------------------------------- + TOTAL 0.08615276 285.21014671 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 83896 +BPFP 0.9761 bits/point +EBPFP 1.9522 equivalent bits/point +MSE 285.210147 +---------------------- -------------------------------------------------------- +Time: 2.507s Load: 0.003s, Pack+Encode: 1.493s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 285.2101 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,792B, BPFP=0.3636 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,236B, BPFP=2.4830 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,960B, BPFP=0.6006 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,736B, BPFP=2.3815 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,408B, BPFP=0.6916 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,792B, BPFP=2.1899 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,164B, BPFP=0.6420 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,684B, BPFP=2.3709 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,456B, BPFP=1.7159 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,336B, BPFP=2.3003 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,920B, BPFP=0.1136 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11290005 53.65632610 + layer.0.v_cache 0.00001570 0.01329851 + layer.1.k_cache 0.03434501 5.90914362 + layer.1.v_cache 0.00000534 0.00535960 + layer.2.k_cache 0.00238675 1.09144315 + layer.2.v_cache 0.00001838 0.01599039 + layer.3.k_cache 0.02838124 4.82152052 + layer.3.v_cache 0.00002047 0.02006920 + layer.4.k_cache 0.00062757 0.37640416 + layer.4.v_cache 0.00005455 0.03773788 + layer.4.output 0.18420308 700.04812152 + ------------------------------------------------------------------------------------- + TOTAL 0.08636333 292.13436140 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 81484 +BPFP 0.9726 bits/point +EBPFP 1.9453 equivalent bits/point +MSE 292.134361 +---------------------- -------------------------------------------------------- +Time: 2.508s Load: 0.004s, Pack+Encode: 1.491s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 292.1344 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,816B, BPFP=0.3638 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,304B, BPFP=2.4647 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,008B, BPFP=0.6026 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,004B, BPFP=2.4046 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,448B, BPFP=0.6907 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,576B, BPFP=2.3189 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,276B, BPFP=0.6562 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,688B, BPFP=2.3413 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,468B, BPFP=1.6963 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,264B, BPFP=2.2564 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,728B, BPFP=0.1067 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09752406 50.12809245 + layer.0.v_cache 0.00001393 0.01313993 + layer.1.k_cache 0.01493649 5.76259593 + layer.1.v_cache 0.00000588 0.00526859 + layer.2.k_cache 0.01003601 1.16966590 + layer.2.v_cache 0.00001805 0.01604616 + layer.3.k_cache 0.04451986 4.36953227 + layer.3.v_cache 0.00001914 0.02031794 + layer.4.k_cache 0.00061869 0.38255692 + layer.4.v_cache 0.00005251 0.03834035 + layer.4.output 0.18409089 690.08041438 + ------------------------------------------------------------------------------------- + TOTAL 0.08566946 287.79226218 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 82580 +BPFP 0.9731 bits/point +EBPFP 1.9462 equivalent bits/point +MSE 287.792262 +---------------------- -------------------------------------------------------- +Time: 2.514s Load: 0.004s, Pack+Encode: 1.494s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 287.7923 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3464 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,476B, BPFP=2.3486 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,044B, BPFP=0.5730 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,008B, BPFP=2.2605 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,336B, BPFP=0.8163 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,672B, BPFP=2.1973 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,748B, BPFP=0.7056 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,936B, BPFP=2.2470 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,040B, BPFP=2.0783 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,464B, BPFP=2.1581 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,736B, BPFP=0.1005 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13658243 50.48845185 + layer.0.v_cache 0.00001379 0.01276121 + layer.1.k_cache 0.01234693 5.57941777 + layer.1.v_cache 0.00000566 0.00526767 + layer.2.k_cache 0.00375735 1.12279612 + layer.2.v_cache 0.00001852 0.01493870 + layer.3.k_cache 0.02700986 5.05205159 + layer.3.v_cache 0.00001930 0.01812442 + layer.4.k_cache 0.00062007 0.40630658 + layer.4.v_cache 0.00005146 0.03609110 + layer.4.output 0.17614663 652.00059165 + ------------------------------------------------------------------------------------- + TOTAL 0.08314423 272.16119697 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 87300 +BPFP 0.9667 bits/point +EBPFP 1.9335 equivalent bits/point +MSE 272.161197 +---------------------- -------------------------------------------------------- +Time: 2.514s Load: 0.004s, Pack+Encode: 1.494s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 272.1612 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,728B, BPFP=0.3699 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,072B, BPFP=2.5839 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,816B, BPFP=0.6027 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,816B, BPFP=2.5291 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,300B, BPFP=0.7063 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,412B, BPFP=2.4426 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,904B, BPFP=0.6216 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,600B, BPFP=2.4829 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,596B, BPFP=1.8399 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,216B, BPFP=2.4007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,384B, BPFP=0.1035 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08738674 57.68224930 + layer.0.v_cache 0.00001368 0.01367594 + layer.1.k_cache 0.01286346 5.78772433 + layer.1.v_cache 0.00000597 0.00547763 + layer.2.k_cache 0.00251921 1.00879136 + layer.2.v_cache 0.00001769 0.01705289 + layer.3.k_cache 0.08324167 4.82667228 + layer.3.v_cache 0.00002021 0.02034082 + layer.4.k_cache 0.00062021 0.40586477 + layer.4.v_cache 0.00005375 0.03977927 + layer.4.output 0.19567458 739.37726272 + ------------------------------------------------------------------------------------- + TOTAL 0.09155674 308.55579221 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 80844 +BPFP 1.0179 bits/point +EBPFP 2.0358 equivalent bits/point +MSE 308.555792 +---------------------- -------------------------------------------------------- +Time: 2.510s Load: 0.004s, Pack+Encode: 1.492s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 308.5558 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.3502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,364B, BPFP=2.3276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,044B, BPFP=0.5730 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,980B, BPFP=2.2553 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,584B, BPFP=0.8630 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,632B, BPFP=2.1898 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,212B, BPFP=0.6047 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,816B, BPFP=2.2244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,568B, BPFP=1.6130 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,420B, BPFP=2.1498 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,928B, BPFP=0.1056 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.019s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12176956 55.05720538 + layer.0.v_cache 0.00001396 0.01283589 + layer.1.k_cache 0.01120206 5.57255903 + layer.1.v_cache 0.00000562 0.00535070 + layer.2.k_cache 0.00678912 1.08055354 + layer.2.v_cache 0.00001849 0.01551363 + layer.3.k_cache 0.02622183 4.03537281 + layer.3.v_cache 0.00001965 0.01889520 + layer.4.k_cache 0.00062413 0.39905291 + layer.4.v_cache 0.00005248 0.03602699 + layer.4.output 0.17940697 651.59853701 + ------------------------------------------------------------------------------------- + TOTAL 0.08368034 272.20136030 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 84408 +BPFP 0.9347 bits/point +EBPFP 1.8694 equivalent bits/point +MSE 272.201360 +---------------------- -------------------------------------------------------- +Time: 2.512s Load: 0.004s, Pack+Encode: 1.489s, Decode+Unpack: 1.019s +---------------------- -------------------------------------------------------- +💾 Converting with 272.2014 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,224B, BPFP=0.3310 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,452B, BPFP=1.8530 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,756B, BPFP=0.5589 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,008B, BPFP=1.7869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,096B, BPFP=0.9071 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,684B, BPFP=1.7387 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,072B, BPFP=1.2012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,960B, BPFP=1.7798 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,512B, BPFP=1.7131 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,428B, BPFP=1.7006 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,168B, BPFP=0.1099 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120816 57.18468192 + layer.0.v_cache 0.00001607 0.01310048 + layer.1.k_cache 0.10678518 6.22033924 + layer.1.v_cache 0.00000585 0.00522104 + layer.2.k_cache 0.00813519 1.18121599 + layer.2.v_cache 0.00001919 0.01600379 + layer.3.k_cache 0.06560528 6.91013590 + layer.3.v_cache 0.00001973 0.01939753 + layer.4.k_cache 0.00062300 0.40876232 + layer.4.v_cache 0.00005287 0.03651016 + layer.4.output 10.86395687 500.74982993 + ------------------------------------------------------------------------------------- + TOTAL 4.49059815 210.42612811 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 96360 +BPFP 0.8435 bits/point +EBPFP 1.6870 equivalent bits/point +MSE 210.426128 +---------------------- -------------------------------------------------------- +Time: 2.511s Load: 0.004s, Pack+Encode: 1.492s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 210.4261 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,888B, BPFP=0.3391 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,352B, BPFP=2.2184 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,132B, BPFP=0.5625 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,004B, BPFP=2.1559 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,112B, BPFP=0.9181 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,588B, BPFP=2.0812 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,340B, BPFP=0.9591 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,944B, BPFP=2.1451 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,760B, BPFP=1.9325 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,464B, BPFP=2.0589 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,156B, BPFP=0.1066 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.025s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12848519 50.61364045 + layer.0.v_cache 0.00001371 0.01228565 + layer.1.k_cache 0.03007136 6.67366396 + layer.1.v_cache 0.00000570 0.00521644 + layer.2.k_cache 0.00936044 1.17806147 + layer.2.v_cache 0.00001842 0.01545934 + layer.3.k_cache 0.04065019 6.28160709 + layer.3.v_cache 0.00001877 0.01857203 + layer.4.k_cache 0.00062204 0.39531581 + layer.4.v_cache 0.00010837 0.03466209 + layer.4.output 0.16426703 621.70689655 + ------------------------------------------------------------------------------------- + TOTAL 0.07995432 259.83392707 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 89740 +BPFP 0.9481 bits/point +EBPFP 1.8961 equivalent bits/point +MSE 259.833927 +---------------------- -------------------------------------------------------- +Time: 2.523s Load: 0.005s, Pack+Encode: 1.493s, Decode+Unpack: 1.025s +---------------------- -------------------------------------------------------- +💾 Converting with 259.8339 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3594 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,284B, BPFP=2.3992 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,992B, BPFP=0.5844 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,896B, BPFP=2.3234 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,092B, BPFP=0.7992 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,408B, BPFP=2.2281 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,052B, BPFP=0.7914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,732B, BPFP=2.2914 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,508B, BPFP=1.8570 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,840B, BPFP=2.1172 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,884B, BPFP=0.1084 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09257430 54.19656372 + layer.0.v_cache 0.00001329 0.01291608 + layer.1.k_cache 0.01202901 5.43609161 + layer.1.v_cache 0.00000544 0.00512389 + layer.2.k_cache 0.00549439 1.23278532 + layer.2.v_cache 0.00001849 0.01603083 + layer.3.k_cache 0.11014032 6.13407440 + layer.3.v_cache 0.00001916 0.01941097 + layer.4.k_cache 0.00063444 0.38662860 + layer.4.v_cache 0.00005494 0.03642788 + layer.4.output 0.19143520 675.09308036 + ------------------------------------------------------------------------------------- + TOTAL 0.09182531 281.94868328 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 84528 +BPFP 0.9711 bits/point +EBPFP 1.9423 equivalent bits/point +MSE 281.948683 +---------------------- -------------------------------------------------------- +Time: 2.513s Load: 0.005s, Pack+Encode: 1.494s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 281.9487 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3423 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,528B, BPFP=2.3304 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,076B, BPFP=0.5722 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,072B, BPFP=2.2455 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,140B, BPFP=0.9561 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,684B, BPFP=2.1734 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,360B, BPFP=0.6250 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,884B, BPFP=2.2106 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,788B, BPFP=2.0067 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,324B, BPFP=2.1064 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,612B, BPFP=0.0960 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.015s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12365316 53.28659784 + layer.0.v_cache 0.00001332 0.01288510 + layer.1.k_cache 0.03511774 4.89485277 + layer.1.v_cache 0.00000569 0.00537666 + layer.2.k_cache 0.01236608 1.14307876 + layer.2.v_cache 0.00001835 0.01540678 + layer.3.k_cache 0.02851222 5.35672469 + layer.3.v_cache 0.00001844 0.01868219 + layer.4.k_cache 0.00063207 0.39777447 + layer.4.v_cache 0.00005066 0.03612158 + layer.4.output 0.17964420 642.49372874 + ------------------------------------------------------------------------------------- + TOTAL 0.08575865 268.38962365 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 87308 +BPFP 0.9553 bits/point +EBPFP 1.9106 equivalent bits/point +MSE 268.389624 +---------------------- -------------------------------------------------------- +Time: 2.513s Load: 0.004s, Pack+Encode: 1.494s, Decode+Unpack: 1.015s +---------------------- -------------------------------------------------------- +💾 Converting with 268.3896 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,804B, BPFP=0.3614 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,276B, BPFP=2.4591 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,984B, BPFP=0.5978 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,676B, BPFP=2.3389 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,776B, BPFP=0.7564 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,052B, BPFP=2.2139 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,412B, BPFP=0.6835 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,672B, BPFP=2.1378 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,404B, BPFP=1.6835 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,220B, BPFP=2.0473 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,764B, BPFP=0.1077 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12096892 53.45370092 + layer.0.v_cache 0.00001393 0.01309839 + layer.1.k_cache 0.03247837 6.10795241 + layer.1.v_cache 0.00000613 0.00517861 + layer.2.k_cache 0.00394396 1.10075652 + layer.2.v_cache 0.00001920 0.01596300 + layer.3.k_cache 0.02751121 4.59419329 + layer.3.v_cache 0.00001919 0.01938454 + layer.4.k_cache 0.00062481 0.38763985 + layer.4.v_cache 0.00005293 0.03533698 + layer.4.output 0.18564256 691.62562958 + ------------------------------------------------------------------------------------- + TOTAL 0.08736098 288.65368303 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 80040 +BPFP 0.9432 bits/point +EBPFP 1.8863 equivalent bits/point +MSE 288.653683 +---------------------- -------------------------------------------------------- +Time: 2.511s Load: 0.004s, Pack+Encode: 1.494s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 288.6537 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,784B, BPFP=0.3620 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,088B, BPFP=2.4529 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,972B, BPFP=0.6031 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,936B, BPFP=2.4221 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,716B, BPFP=0.7541 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,576B, BPFP=2.3490 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,252B, BPFP=0.6599 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,664B, BPFP=2.3669 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,780B, BPFP=1.7817 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,400B, BPFP=2.3133 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,860B, BPFP=0.1119 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08564486 54.86363636 + layer.0.v_cache 0.00001408 0.01346929 + layer.1.k_cache 0.03507188 6.17944772 + layer.1.v_cache 0.00000566 0.00545729 + layer.2.k_cache 0.00239554 1.13821490 + layer.2.v_cache 0.00001938 0.01640875 + layer.3.k_cache 0.03313774 5.20739112 + layer.3.v_cache 0.00001993 0.02065505 + layer.4.k_cache 0.00061714 0.38609577 + layer.4.v_cache 0.00005134 0.03661847 + layer.4.output 0.18226704 700.48411410 + ------------------------------------------------------------------------------------- + TOTAL 0.08428511 292.42683491 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 83028 +BPFP 0.9911 bits/point +EBPFP 1.9821 equivalent bits/point +MSE 292.426835 +---------------------- -------------------------------------------------------- +Time: 2.510s Load: 0.004s, Pack+Encode: 1.490s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 292.4268 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 117, 128) +Output shape: (1, 117, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.output: torch.Size([1, 117, 3584]) -> torch.Size([1, 1, 117, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,396B, BPFP=0.3200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,644B, BPFP=1.6886 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,676B, BPFP=0.6245 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,272B, BPFP=1.6389 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,364B, BPFP=0.7163 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,844B, BPFP=1.5817 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,364B, BPFP=0.8499 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,052B, BPFP=1.6095 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,996B, BPFP=1.4685 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,528B, BPFP=1.5395 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,380B, BPFP=0.1026 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11900647 52.95496127 + layer.0.v_cache 0.00001428 0.01326172 + layer.1.k_cache 0.12535267 5.89142420 + layer.1.v_cache 0.00000579 0.00561695 + layer.2.k_cache 0.00517510 1.06243988 + layer.2.v_cache 0.00001992 0.01619470 + layer.3.k_cache 0.09315677 6.62284812 + layer.3.v_cache 0.00001978 0.01910054 + layer.4.k_cache 0.00061896 0.40406287 + layer.4.v_cache 0.00005133 0.03383855 + layer.4.output 9.75369281 457.17723596 + ------------------------------------------------------------------------------------- + TOTAL 4.03642769 192.19202356 + (elements=1,018,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1018368 +Total Bytes 95516 +BPFP 0.7503 bits/point +EBPFP 1.5007 equivalent bits/point +MSE 192.192024 +---------------------- -------------------------------------------------------- +Time: 2.512s Load: 0.005s, Pack+Encode: 1.500s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 192.1920 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,224B, BPFP=0.3310 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,396B, BPFP=1.8446 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,800B, BPFP=0.5655 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,056B, BPFP=1.7940 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,976B, BPFP=0.8893 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,660B, BPFP=1.7351 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,192B, BPFP=1.2190 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,892B, BPFP=1.7696 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,372B, BPFP=1.6923 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,396B, BPFP=1.6958 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,032B, BPFP=0.0857 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14131550 57.05398996 + layer.0.v_cache 0.00001383 0.01286708 + layer.1.k_cache 0.02978072 6.86613653 + layer.1.v_cache 0.00000537 0.00515777 + layer.2.k_cache 0.00578868 1.20008458 + layer.2.v_cache 0.00001872 0.01572452 + layer.3.k_cache 0.04020810 8.12955845 + layer.3.v_cache 0.00001850 0.01866291 + layer.4.k_cache 0.00063358 0.41364768 + layer.4.v_cache 0.00005294 0.03725597 + layer.4.output 10.86925024 509.88550170 + ------------------------------------------------------------------------------------- + TOTAL 4.48838751 214.29127043 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 94996 +BPFP 0.8315 bits/point +EBPFP 1.6631 equivalent bits/point +MSE 214.291270 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.006s, Pack+Encode: 1.492s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 214.2913 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,844B, BPFP=0.3430 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,488B, BPFP=2.3229 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,092B, BPFP=0.5751 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,132B, BPFP=2.2567 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,536B, BPFP=1.0298 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,644B, BPFP=2.1659 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,144B, BPFP=0.7708 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,956B, BPFP=2.2240 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,304B, BPFP=1.9167 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,504B, BPFP=2.1399 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,568B, BPFP=0.0948 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09514355 49.80558268 + layer.0.v_cache 0.00001423 0.01296831 + layer.1.k_cache 0.03283414 5.71252151 + layer.1.v_cache 0.00000553 0.00502912 + layer.2.k_cache 0.00489674 1.19445855 + layer.2.v_cache 0.00001883 0.01552352 + layer.3.k_cache 0.07199075 5.66544742 + layer.3.v_cache 0.00001932 0.01858963 + layer.4.k_cache 0.00061592 0.39037623 + layer.4.v_cache 0.00006816 0.03571333 + layer.4.output 0.17253765 642.08604379 + ------------------------------------------------------------------------------------- + TOTAL 0.08313946 268.08579511 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 88212 +BPFP 0.9652 bits/point +EBPFP 1.9304 equivalent bits/point +MSE 268.085795 +---------------------- -------------------------------------------------------- +Time: 2.502s Load: 0.003s, Pack+Encode: 1.495s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 268.0858 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.3341 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,368B, BPFP=2.2213 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,132B, BPFP=0.5625 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,128B, BPFP=2.1782 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,780B, BPFP=0.6789 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,816B, BPFP=2.1221 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,456B, BPFP=0.8003 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,924B, BPFP=2.1415 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,044B, BPFP=1.8039 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,616B, BPFP=2.0862 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,588B, BPFP=0.0921 +⌛️ [2/4] FRONTEND: Frontend time: 1.487s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.002s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11660385 54.28832166 + layer.0.v_cache 0.00001381 0.01295076 + layer.1.k_cache 0.05175667 6.73879619 + layer.1.v_cache 0.00000569 0.00533440 + layer.2.k_cache 0.00227456 1.00457036 + layer.2.v_cache 0.00001743 0.01495384 + layer.3.k_cache 0.02781237 4.93321912 + layer.3.v_cache 0.00001883 0.01773311 + layer.4.k_cache 0.00061846 0.41357027 + layer.4.v_cache 0.00005344 0.03503913 + layer.4.output 0.17410791 622.74338054 + ------------------------------------------------------------------------------------- + TOTAL 0.08340767 260.39224427 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 86712 +BPFP 0.9161 bits/point +EBPFP 1.8322 equivalent bits/point +MSE 260.392244 +---------------------- -------------------------------------------------------- +Time: 2.493s Load: 0.004s, Pack+Encode: 1.487s, Decode+Unpack: 1.002s +---------------------- -------------------------------------------------------- +💾 Converting with 260.3922 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,852B, BPFP=0.3486 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,428B, BPFP=2.3396 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,012B, BPFP=0.5670 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,984B, BPFP=2.2560 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,024B, BPFP=0.9458 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,716B, BPFP=2.2056 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,324B, BPFP=0.6258 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,868B, BPFP=2.2342 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,100B, BPFP=2.0896 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,412B, BPFP=2.1483 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,760B, BPFP=0.1011 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212080 52.72675075 + layer.0.v_cache 0.00001360 0.01256128 + layer.1.k_cache 0.03288655 5.86359966 + layer.1.v_cache 0.00000536 0.00503382 + layer.2.k_cache 0.00973156 1.14469064 + layer.2.v_cache 0.00001762 0.01491470 + layer.3.k_cache 0.02906220 4.50705949 + layer.3.v_cache 0.00001891 0.01825663 + layer.4.k_cache 0.00063934 0.40231222 + layer.4.v_cache 0.00004866 0.03467285 + layer.4.output 0.18425958 652.34364243 + ------------------------------------------------------------------------------------- + TOTAL 0.08496245 272.41972641 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 87480 +BPFP 0.9687 bits/point +EBPFP 1.9375 equivalent bits/point +MSE 272.419726 +---------------------- -------------------------------------------------------- +Time: 2.490s Load: 0.004s, Pack+Encode: 1.484s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 272.4197 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,888B, BPFP=0.3471 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,452B, BPFP=2.2890 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,128B, BPFP=0.5750 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,076B, BPFP=2.2199 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,912B, BPFP=0.9029 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,588B, BPFP=2.1301 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,528B, BPFP=1.0162 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,868B, BPFP=2.1816 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,808B, BPFP=1.6191 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,520B, BPFP=2.1176 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,680B, BPFP=0.0966 +⌛️ [2/4] FRONTEND: Frontend time: 1.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10758395 53.06964614 + layer.0.v_cache 0.00001397 0.01342400 + layer.1.k_cache 0.03293324 5.88310619 + layer.1.v_cache 0.00000575 0.00514049 + layer.2.k_cache 0.00823421 0.96858251 + layer.2.v_cache 0.00001745 0.01529868 + layer.3.k_cache 0.04116619 6.45267119 + layer.3.v_cache 0.00001985 0.01937608 + layer.4.k_cache 0.00062779 0.37816364 + layer.4.v_cache 0.00005042 0.03572341 + layer.4.output 0.16954060 635.00246849 + ------------------------------------------------------------------------------------- + TOTAL 0.08102571 265.40343598 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 87448 +BPFP 0.9456 bits/point +EBPFP 1.8912 equivalent bits/point +MSE 265.403436 +---------------------- -------------------------------------------------------- +Time: 2.494s Load: 0.005s, Pack+Encode: 1.488s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 265.4034 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,900B, BPFP=0.3299 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,492B, BPFP=2.1688 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,280B, BPFP=0.5694 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,972B, BPFP=2.0785 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,976B, BPFP=0.6903 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,640B, BPFP=2.0208 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,156B, BPFP=1.2424 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,824B, BPFP=2.0528 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,360B, BPFP=1.7986 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,448B, BPFP=1.9875 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,800B, BPFP=0.0942 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12307767 52.52813585 + layer.0.v_cache 0.00001410 0.01348450 + layer.1.k_cache 0.03088827 6.20206434 + layer.1.v_cache 0.00000584 0.00551829 + layer.2.k_cache 0.00371391 1.20984294 + layer.2.v_cache 0.00001950 0.01640879 + layer.3.k_cache 0.03079298 6.15620253 + layer.3.v_cache 0.00001950 0.01979531 + layer.4.k_cache 0.00063014 0.41290800 + layer.4.v_cache 0.00005282 0.03654718 + layer.4.output 0.15767360 601.47703373 + ------------------------------------------------------------------------------------- + TOTAL 0.07605470 251.58471434 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 89848 +BPFP 0.9176 bits/point +EBPFP 1.8351 equivalent bits/point +MSE 251.584714 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.005s, Pack+Encode: 1.496s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 251.5847 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,820B, BPFP=0.3468 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,228B, BPFP=2.3300 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,980B, BPFP=0.5678 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,020B, BPFP=2.2904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,592B, BPFP=0.6845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,496B, BPFP=2.1905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,516B, BPFP=0.6700 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,780B, BPFP=2.2447 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,400B, BPFP=1.6006 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,164B, BPFP=2.1273 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,900B, BPFP=0.1062 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10295349 55.04289134 + layer.0.v_cache 0.00001348 0.01221591 + layer.1.k_cache 0.03334363 5.37199588 + layer.1.v_cache 0.00000550 0.00526393 + layer.2.k_cache 0.01126997 1.10153226 + layer.2.v_cache 0.00001933 0.01569034 + layer.3.k_cache 0.09131574 5.24517115 + layer.3.v_cache 0.00001769 0.01791771 + layer.4.k_cache 0.00062275 0.40545096 + layer.4.v_cache 0.00006541 0.03545168 + layer.4.output 0.17512874 656.60452962 + ------------------------------------------------------------------------------------- + TOTAL 0.08620754 274.32266403 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 82896 +BPFP 0.9292 bits/point +EBPFP 1.8583 equivalent bits/point +MSE 274.322664 +---------------------- -------------------------------------------------------- +Time: 2.524s Load: 0.003s, Pack+Encode: 1.503s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 274.3227 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,888B, BPFP=0.3391 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,376B, BPFP=2.2227 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,176B, BPFP=0.5704 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,144B, BPFP=2.1810 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,448B, BPFP=0.7989 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,724B, BPFP=2.1056 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,028B, BPFP=0.9030 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,904B, BPFP=2.1379 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,796B, BPFP=1.7593 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,440B, BPFP=2.0546 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,696B, BPFP=0.0948 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.023s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12314748 54.02386404 + layer.0.v_cache 0.00001636 0.01289786 + layer.1.k_cache 0.01190477 5.80790412 + layer.1.v_cache 0.00000588 0.00529259 + layer.2.k_cache 0.01614088 1.04386314 + layer.2.v_cache 0.00001834 0.01522628 + layer.3.k_cache 0.05604688 6.56191420 + layer.3.v_cache 0.00001882 0.01934492 + layer.4.k_cache 0.00063774 0.39928568 + layer.4.v_cache 0.00005162 0.03509311 + layer.4.output 0.17348555 622.20258621 + ------------------------------------------------------------------------------------- + TOTAL 0.08366986 260.19663467 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 87620 +BPFP 0.9257 bits/point +EBPFP 1.8513 equivalent bits/point +MSE 260.196635 +---------------------- -------------------------------------------------------- +Time: 2.520s Load: 0.003s, Pack+Encode: 1.493s, Decode+Unpack: 1.023s +---------------------- -------------------------------------------------------- +💾 Converting with 260.1966 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3549 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,536B, BPFP=2.4182 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,020B, BPFP=0.5826 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,040B, BPFP=2.3225 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,152B, BPFP=0.8009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,756B, BPFP=2.2677 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,444B, BPFP=0.6644 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,816B, BPFP=2.2793 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,952B, BPFP=1.9198 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,512B, BPFP=2.2207 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,976B, BPFP=0.1096 +⌛️ [2/4] FRONTEND: Frontend time: 1.499s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11560643 51.82551481 + layer.0.v_cache 0.00001402 0.01313937 + layer.1.k_cache 0.03272506 5.91137808 + layer.1.v_cache 0.00000563 0.00551853 + layer.2.k_cache 0.00645547 1.11972573 + layer.2.v_cache 0.00001830 0.01598052 + layer.3.k_cache 0.02731901 5.38515143 + layer.3.v_cache 0.00001960 0.01859889 + layer.4.k_cache 0.00061302 0.39327014 + layer.4.v_cache 0.00004978 0.03500407 + layer.4.output 0.17251868 666.42559524 + ------------------------------------------------------------------------------------- + TOTAL 0.08179159 278.21779107 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 86044 +BPFP 0.9764 bits/point +EBPFP 1.9527 equivalent bits/point +MSE 278.217791 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.004s, Pack+Encode: 1.499s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 278.2178 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,896B, BPFP=0.3366 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,488B, BPFP=2.2173 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,164B, BPFP=0.5618 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,012B, BPFP=2.1328 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,368B, BPFP=0.9531 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,748B, BPFP=2.0859 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,884B, BPFP=1.0447 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,908B, BPFP=2.1143 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,532B, BPFP=1.8700 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,460B, BPFP=2.0348 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,088B, BPFP=0.1037 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.019s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11690623 57.11138361 + layer.0.v_cache 0.00001418 0.01304561 + layer.1.k_cache 0.03305150 6.31968203 + layer.1.v_cache 0.00000545 0.00534514 + layer.2.k_cache 0.01095151 1.13032428 + layer.2.v_cache 0.00001854 0.01535412 + layer.3.k_cache 0.04293454 6.52633390 + layer.3.v_cache 0.00002007 0.01958256 + layer.4.k_cache 0.00062489 0.39204441 + layer.4.v_cache 0.00005451 0.03615630 + layer.4.output 0.16672201 614.35049716 + ------------------------------------------------------------------------------------- + TOTAL 0.08068444 257.17780777 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 90548 +BPFP 0.9457 bits/point +EBPFP 1.8915 equivalent bits/point +MSE 257.177808 +---------------------- -------------------------------------------------------- +Time: 2.521s Load: 0.005s, Pack+Encode: 1.496s, Decode+Unpack: 1.019s +---------------------- -------------------------------------------------------- +💾 Converting with 257.1778 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,916B, BPFP=0.3290 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,384B, BPFP=2.1264 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,196B, BPFP=0.5488 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,016B, BPFP=2.0632 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,920B, BPFP=0.6731 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,596B, BPFP=1.9911 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,932B, BPFP=1.1902 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,764B, BPFP=2.0199 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,396B, BPFP=1.9567 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,292B, BPFP=1.9389 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,732B, BPFP=0.0915 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11067300 54.83371609 + layer.0.v_cache 0.00001409 0.01305111 + layer.1.k_cache 0.05152405 6.00655960 + layer.1.v_cache 0.00000569 0.00543203 + layer.2.k_cache 0.01362713 1.02182543 + layer.2.v_cache 0.00001889 0.01642985 + layer.3.k_cache 0.08227781 7.04454627 + layer.3.v_cache 0.00001889 0.01953202 + layer.4.k_cache 0.00062414 0.41511708 + layer.4.v_cache 0.00005257 0.03583032 + layer.4.output 0.15746453 593.45638736 + ------------------------------------------------------------------------------------- + TOTAL 0.08006400 248.44745596 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 90144 +BPFP 0.9105 bits/point +EBPFP 1.8209 equivalent bits/point +MSE 248.447456 +---------------------- -------------------------------------------------------- +Time: 2.523s Load: 0.004s, Pack+Encode: 1.502s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 248.4475 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 116, 128) +Output shape: (1, 116, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.output: torch.Size([1, 116, 3584]) -> torch.Size([1, 1, 116, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,392B, BPFP=0.3222 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,528B, BPFP=1.6875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,100B, BPFP=0.5523 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,120B, BPFP=1.6325 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,128B, BPFP=0.6907 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,796B, BPFP=1.5889 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,848B, BPFP=0.6530 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,920B, BPFP=1.6056 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,724B, BPFP=1.4445 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,468B, BPFP=1.5447 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,508B, BPFP=0.0867 +⌛️ [2/4] FRONTEND: Frontend time: 1.506s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16314862 51.98664803 + layer.0.v_cache 0.00001373 0.01226600 + layer.1.k_cache 0.05481785 5.79830038 + layer.1.v_cache 0.00000557 0.00507964 + layer.2.k_cache 0.00968372 1.13302823 + layer.2.v_cache 0.00001954 0.01591212 + layer.3.k_cache 0.02418142 5.31497613 + layer.3.v_cache 0.00001860 0.01813245 + layer.4.k_cache 0.00062333 0.39932928 + layer.4.v_cache 0.00005467 0.03512422 + layer.4.output 9.83417319 462.33243534 + ------------------------------------------------------------------------------------- + TOTAL 4.06422232 194.17916729 + (elements=1,009,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1009664 +Total Bytes 91532 +BPFP 0.7252 bits/point +EBPFP 1.4505 equivalent bits/point +MSE 194.179167 +---------------------- -------------------------------------------------------- +Time: 2.530s Load: 0.004s, Pack+Encode: 1.506s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 194.1792 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,780B, BPFP=0.3566 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,864B, BPFP=2.3766 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,968B, BPFP=0.5946 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,876B, BPFP=2.3790 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,348B, BPFP=0.6707 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,480B, BPFP=2.2997 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,992B, BPFP=0.5994 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,368B, BPFP=2.2772 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,132B, BPFP=1.4287 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,896B, BPFP=2.1827 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,960B, BPFP=0.1133 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.022s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11749315 52.68515875 + layer.0.v_cache 0.00001389 0.01313664 + layer.1.k_cache 0.03399578 6.34446364 + layer.1.v_cache 0.00000555 0.00542548 + layer.2.k_cache 0.00854440 1.16468498 + layer.2.v_cache 0.00001777 0.01538781 + layer.3.k_cache 0.02740619 4.83935156 + layer.3.v_cache 0.00001935 0.01959525 + layer.4.k_cache 0.00063531 0.38418765 + layer.4.v_cache 0.00005166 0.03725469 + layer.4.output 0.18490290 690.41025641 + ------------------------------------------------------------------------------------- + TOTAL 0.08720608 288.14002596 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 79664 +BPFP 0.9387 bits/point +EBPFP 1.8775 equivalent bits/point +MSE 288.140026 +---------------------- -------------------------------------------------------- +Time: 2.524s Load: 0.004s, Pack+Encode: 1.497s, Decode+Unpack: 1.022s +---------------------- -------------------------------------------------------- +💾 Converting with 288.1400 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,932B, BPFP=0.3211 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,248B, BPFP=2.0359 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,224B, BPFP=0.5359 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,808B, BPFP=1.9628 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,236B, BPFP=0.8703 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,520B, BPFP=1.9149 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,028B, BPFP=1.1682 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,756B, BPFP=1.9541 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,144B, BPFP=1.8524 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,196B, BPFP=1.8610 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,000B, BPFP=0.0950 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11492289 57.80063061 + layer.0.v_cache 0.00001440 0.01307621 + layer.1.k_cache 0.01112092 5.59162156 + layer.1.v_cache 0.00000555 0.00525184 + layer.2.k_cache 0.01096075 1.20743074 + layer.2.v_cache 0.00001882 0.01580627 + layer.3.k_cache 0.03781782 7.32918078 + layer.3.v_cache 0.00001905 0.01917715 + layer.4.k_cache 0.00062716 0.41741521 + layer.4.v_cache 0.00006343 0.03473012 + layer.4.output 0.15224552 572.52840046 + ------------------------------------------------------------------------------------- + TOTAL 0.07301703 240.00783080 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 91092 +BPFP 0.8907 bits/point +EBPFP 1.7814 equivalent bits/point +MSE 240.007831 +---------------------- -------------------------------------------------------- +Time: 2.520s Load: 0.005s, Pack+Encode: 1.498s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 240.0078 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,784B, BPFP=0.3620 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,144B, BPFP=2.4643 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,928B, BPFP=0.5942 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,904B, BPFP=2.4156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,460B, BPFP=0.7021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,432B, BPFP=2.3198 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,392B, BPFP=0.6883 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,708B, BPFP=2.3758 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,960B, BPFP=1.6153 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,000B, BPFP=2.2321 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,568B, BPFP=0.1034 +⌛️ [2/4] FRONTEND: Frontend time: 1.504s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10405910 54.79700183 + layer.0.v_cache 0.00001453 0.01296651 + layer.1.k_cache 0.03397077 6.46040166 + layer.1.v_cache 0.00000547 0.00538228 + layer.2.k_cache 0.00568739 1.26252449 + layer.2.v_cache 0.00001916 0.01632312 + layer.3.k_cache 0.09581309 5.31189322 + layer.3.v_cache 0.00001862 0.01917176 + layer.4.k_cache 0.00061896 0.40310585 + layer.4.v_cache 0.00005653 0.03570614 + layer.4.output 0.19123089 700.53345315 + ------------------------------------------------------------------------------------- + TOTAL 0.09287528 292.47403817 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 81280 +BPFP 0.9702 bits/point +EBPFP 1.9404 equivalent bits/point +MSE 292.474038 +---------------------- -------------------------------------------------------- +Time: 2.529s Load: 0.005s, Pack+Encode: 1.504s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 292.4740 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,852B, BPFP=0.3404 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,464B, BPFP=2.2912 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,148B, BPFP=0.5787 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,164B, BPFP=2.2360 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,544B, BPFP=0.8353 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,716B, BPFP=2.1537 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,148B, BPFP=0.9463 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,964B, BPFP=2.1993 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,380B, BPFP=1.9081 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,600B, BPFP=2.1324 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,016B, BPFP=0.1055 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09384689 53.33417969 + layer.0.v_cache 0.00001364 0.01294625 + layer.1.k_cache 0.01313462 5.87191090 + layer.1.v_cache 0.00000592 0.00553306 + layer.2.k_cache 0.00813576 1.18346199 + layer.2.v_cache 0.00001762 0.01508656 + layer.3.k_cache 0.02628646 5.23367489 + layer.3.v_cache 0.00002044 0.01904720 + layer.4.k_cache 0.00061596 0.39341350 + layer.4.v_cache 0.00005334 0.03608596 + layer.4.output 0.17343949 632.51743697 + ------------------------------------------------------------------------------------- + TOTAL 0.07977689 264.33690581 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 88996 +BPFP 0.9623 bits/point +EBPFP 1.9247 equivalent bits/point +MSE 264.336906 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.005s, Pack+Encode: 1.494s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 264.3369 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 126, 128) +Output shape: (1, 126, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.output: torch.Size([1, 126, 3584]) -> torch.Size([1, 1, 126, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,828B, BPFP=0.3507 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,476B, BPFP=1.5471 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,292B, BPFP=0.5322 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,040B, BPFP=1.4931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,952B, BPFP=0.7381 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,504B, BPFP=1.4266 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,100B, BPFP=0.7564 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,668B, BPFP=1.4469 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,352B, BPFP=1.2837 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,176B, BPFP=1.3859 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,336B, BPFP=0.0945 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.008s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13141363 59.35777840 + layer.0.v_cache 0.00001429 0.01241875 + layer.1.k_cache 0.06496696 7.09264798 + layer.1.v_cache 0.00000583 0.00540551 + layer.2.k_cache 0.01057190 1.24159761 + layer.2.v_cache 0.00002020 0.01679666 + layer.3.k_cache 0.05250616 8.32881867 + layer.3.v_cache 0.00001853 0.01967296 + layer.4.k_cache 0.00062861 0.45277556 + layer.4.v_cache 0.00004930 0.03565047 + layer.4.output 9.05470577 424.26314484 + ------------------------------------------------------------------------------------- + TOTAL 3.74371387 179.20032803 + (elements=1,096,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1096704 +Total Bytes 93724 +BPFP 0.6837 bits/point +EBPFP 1.3674 equivalent bits/point +MSE 179.200328 +---------------------- -------------------------------------------------------- +Time: 2.508s Load: 0.005s, Pack+Encode: 1.494s, Decode+Unpack: 1.008s +---------------------- -------------------------------------------------------- +💾 Converting with 179.2003 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,884B, BPFP=0.3384 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,444B, BPFP=2.2349 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,124B, BPFP=0.5611 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,108B, BPFP=2.1746 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,336B, BPFP=0.9583 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,744B, BPFP=2.1092 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,584B, BPFP=0.8233 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,956B, BPFP=2.1473 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,960B, BPFP=1.9684 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,468B, BPFP=2.0596 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,676B, BPFP=0.0943 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13059476 53.18068651 + layer.0.v_cache 0.00001380 0.01257314 + layer.1.k_cache 0.03034586 6.54295753 + layer.1.v_cache 0.00000531 0.00510274 + layer.2.k_cache 0.00374181 1.07569105 + layer.2.v_cache 0.00001886 0.01567273 + layer.3.k_cache 0.04454902 5.34820416 + layer.3.v_cache 0.00001968 0.01914341 + layer.4.k_cache 0.00060938 0.41444599 + layer.4.v_cache 0.00007316 0.03664826 + layer.4.output 0.16485390 620.68103448 + ------------------------------------------------------------------------------------- + TOTAL 0.08023229 259.49519805 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 89284 +BPFP 0.9432 bits/point +EBPFP 1.8865 equivalent bits/point +MSE 259.495198 +---------------------- -------------------------------------------------------- +Time: 2.501s Load: 0.004s, Pack+Encode: 1.490s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 259.4952 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,912B, BPFP=0.3247 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,408B, BPFP=2.1073 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,276B, BPFP=0.5564 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,956B, BPFP=2.0306 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,040B, BPFP=1.0258 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,528B, BPFP=1.9579 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,608B, BPFP=0.9524 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,780B, BPFP=2.0007 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,216B, BPFP=1.9049 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,372B, BPFP=1.9314 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,356B, BPFP=0.1057 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.003s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15436328 53.10695482 + layer.0.v_cache 0.00001586 0.01314633 + layer.1.k_cache 0.01080767 6.06908052 + layer.1.v_cache 0.00000575 0.00552031 + layer.2.k_cache 0.00786338 1.02682114 + layer.2.v_cache 0.00001961 0.01547866 + layer.3.k_cache 0.05652578 7.04153840 + layer.3.v_cache 0.00002042 0.01998856 + layer.4.k_cache 0.00062828 0.40902656 + layer.4.v_cache 0.00005674 0.03897368 + layer.4.output 0.15192759 586.99966033 + ------------------------------------------------------------------------------------- + TOTAL 0.07610588 245.69083243 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 91452 +BPFP 0.9136 bits/point +EBPFP 1.8273 equivalent bits/point +MSE 245.690832 +---------------------- -------------------------------------------------------- +Time: 2.497s Load: 0.005s, Pack+Encode: 1.489s, Decode+Unpack: 1.003s +---------------------- -------------------------------------------------------- +💾 Converting with 245.6908 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,796B, BPFP=0.3598 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,280B, BPFP=2.4599 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,992B, BPFP=0.5994 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,780B, BPFP=2.3598 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,720B, BPFP=0.7452 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,584B, BPFP=2.3205 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,368B, BPFP=0.6747 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,744B, BPFP=2.3526 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,328B, BPFP=1.8686 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,808B, BPFP=2.1651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,496B, BPFP=0.1000 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10254851 52.15394318 + layer.0.v_cache 0.00001399 0.01305606 + layer.1.k_cache 0.05476840 5.84575907 + layer.1.v_cache 0.00000542 0.00518601 + layer.2.k_cache 0.00244555 1.13151179 + layer.2.v_cache 0.00001844 0.01629013 + layer.3.k_cache 0.02811479 4.84165759 + layer.3.v_cache 0.00001905 0.01864699 + layer.4.k_cache 0.00060111 0.38218598 + layer.4.v_cache 0.00005313 0.03610343 + layer.4.output 0.19635826 692.74118590 + ------------------------------------------------------------------------------------- + TOTAL 0.09194683 289.03721421 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 82896 +BPFP 0.9768 bits/point +EBPFP 1.9536 equivalent bits/point +MSE 289.037214 +---------------------- -------------------------------------------------------- +Time: 2.501s Load: 0.004s, Pack+Encode: 1.491s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 289.0372 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,816B, BPFP=0.3460 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,176B, BPFP=2.3201 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,028B, BPFP=0.5770 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,988B, BPFP=2.2843 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,616B, BPFP=0.6890 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,540B, BPFP=2.1989 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,404B, BPFP=0.6486 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,876B, BPFP=2.2630 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,744B, BPFP=1.6662 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,340B, BPFP=2.1608 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,680B, BPFP=0.1002 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11987135 55.20494712 + layer.0.v_cache 0.00001390 0.01251759 + layer.1.k_cache 0.03139273 6.31535153 + layer.1.v_cache 0.00000562 0.00524301 + layer.2.k_cache 0.00816743 1.18947983 + layer.2.v_cache 0.00001857 0.01522245 + layer.3.k_cache 0.02640573 4.22311625 + layer.3.v_cache 0.00001968 0.01882826 + layer.4.k_cache 0.00062091 0.39423966 + layer.4.v_cache 0.00004972 0.03471232 + layer.4.output 0.16983549 655.91109538 + ------------------------------------------------------------------------------------- + TOTAL 0.08090671 274.04654857 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 83208 +BPFP 0.9327 bits/point +EBPFP 1.8653 equivalent bits/point +MSE 274.046549 +---------------------- -------------------------------------------------------- +Time: 2.506s Load: 0.004s, Pack+Encode: 1.497s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 274.0465 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,876B, BPFP=0.3369 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,416B, BPFP=2.2299 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,208B, BPFP=0.5761 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,100B, BPFP=2.1731 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,568B, BPFP=0.8204 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,644B, BPFP=2.0912 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,836B, BPFP=1.0481 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,896B, BPFP=2.1365 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,356B, BPFP=2.0395 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,484B, BPFP=2.0625 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,692B, BPFP=0.0947 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.008s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14089697 54.01821233 + layer.0.v_cache 0.00001393 0.01288480 + layer.1.k_cache 0.01618495 6.47121456 + layer.1.v_cache 0.00000582 0.00540305 + layer.2.k_cache 0.00661465 1.14018013 + layer.2.v_cache 0.00002083 0.01551269 + layer.3.k_cache 0.07087009 6.19447905 + layer.3.v_cache 0.00001822 0.01921131 + layer.4.k_cache 0.00061891 0.41332604 + layer.4.v_cache 0.00005311 0.03612898 + layer.4.output 0.16425455 620.62310140 + ------------------------------------------------------------------------------------- + TOTAL 0.08147525 259.56989781 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 90076 +BPFP 0.9516 bits/point +EBPFP 1.9032 equivalent bits/point +MSE 259.569898 +---------------------- -------------------------------------------------------- +Time: 2.506s Load: 0.005s, Pack+Encode: 1.494s, Decode+Unpack: 1.008s +---------------------- -------------------------------------------------------- +💾 Converting with 259.5699 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,916B, BPFP=0.3364 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,508B, BPFP=2.1959 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,144B, BPFP=0.5520 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,168B, BPFP=2.1362 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,224B, BPFP=0.9171 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,828B, BPFP=2.0765 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,788B, BPFP=1.0162 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,096B, BPFP=2.1236 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,756B, BPFP=1.7128 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,592B, BPFP=2.0351 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,144B, BPFP=0.1039 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12786581 55.08710060 + layer.0.v_cache 0.00001426 0.01342342 + layer.1.k_cache 0.03183105 5.46749501 + layer.1.v_cache 0.00000593 0.00562834 + layer.2.k_cache 0.00241907 1.11867900 + layer.2.v_cache 0.00001785 0.01566755 + layer.3.k_cache 0.02946699 6.33852018 + layer.3.v_cache 0.00001957 0.02026381 + layer.4.k_cache 0.00063747 0.41096462 + layer.4.v_cache 0.00005003 0.03542634 + layer.4.output 0.16568538 608.00285915 + ------------------------------------------------------------------------------------- + TOTAL 0.07953681 254.38430488 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 90164 +BPFP 0.9311 bits/point +EBPFP 1.8623 equivalent bits/point +MSE 254.384305 +---------------------- -------------------------------------------------------- +Time: 2.502s Load: 0.005s, Pack+Encode: 1.493s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 254.3843 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,792B, BPFP=0.3590 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,288B, BPFP=2.4615 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,980B, BPFP=0.5970 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,032B, BPFP=2.4103 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,744B, BPFP=0.7500 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,440B, BPFP=2.2917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,688B, BPFP=0.7388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,904B, BPFP=2.3846 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,596B, BPFP=1.9223 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,564B, BPFP=2.1162 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,792B, BPFP=0.1085 +⌛️ [2/4] FRONTEND: Frontend time: 1.486s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12225728 54.46619591 + layer.0.v_cache 0.00001337 0.01323100 + layer.1.k_cache 0.03480693 5.18540602 + layer.1.v_cache 0.00000545 0.00525159 + layer.2.k_cache 0.00853881 1.21025085 + layer.2.v_cache 0.00001924 0.01652650 + layer.3.k_cache 0.06130744 4.84174914 + layer.3.v_cache 0.00001928 0.02026674 + layer.4.k_cache 0.00058936 0.39302855 + layer.4.v_cache 0.00007197 0.03826479 + layer.4.output 0.18657821 688.70215201 + ------------------------------------------------------------------------------------- + TOTAL 0.09021627 287.47677854 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 83820 +BPFP 0.9877 bits/point +EBPFP 1.9754 equivalent bits/point +MSE 287.476779 +---------------------- -------------------------------------------------------- +Time: 2.495s Load: 0.004s, Pack+Encode: 1.486s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 287.4768 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,788B, BPFP=0.3628 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,360B, BPFP=2.5081 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,944B, BPFP=0.5974 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,916B, BPFP=2.4180 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,308B, BPFP=0.6713 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,612B, BPFP=2.3563 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,004B, BPFP=0.6096 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,772B, BPFP=2.3888 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,024B, BPFP=1.8312 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,972B, BPFP=2.2265 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,584B, BPFP=0.1039 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.002s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582910 55.89435369 + layer.0.v_cache 0.00001386 0.01316157 + layer.1.k_cache 0.03408930 5.89008797 + layer.1.v_cache 0.00000538 0.00533422 + layer.2.k_cache 0.00228146 1.07455048 + layer.2.v_cache 0.00001859 0.01605464 + layer.3.k_cache 0.06276476 4.91532838 + layer.3.v_cache 0.00001913 0.01964451 + layer.4.k_cache 0.00060520 0.40474215 + layer.4.v_cache 0.00005125 0.03790735 + layer.4.output 0.19356344 697.92248377 + ------------------------------------------------------------------------------------- + TOTAL 0.09238954 291.39579714 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 82284 +BPFP 0.9822 bits/point +EBPFP 1.9644 equivalent bits/point +MSE 291.395797 +---------------------- -------------------------------------------------------- +Time: 2.499s Load: 0.004s, Pack+Encode: 1.492s, Decode+Unpack: 1.002s +---------------------- -------------------------------------------------------- +💾 Converting with 291.3958 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.3502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,444B, BPFP=2.3426 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,012B, BPFP=0.5670 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,140B, BPFP=2.2854 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,468B, BPFP=0.8411 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,676B, BPFP=2.1980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,396B, BPFP=0.8276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,116B, BPFP=2.2809 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,796B, BPFP=2.0324 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,368B, BPFP=2.1401 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,836B, BPFP=0.1032 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09700135 53.58204302 + layer.0.v_cache 0.00001355 0.01297075 + layer.1.k_cache 0.01103399 5.97395122 + layer.1.v_cache 0.00000598 0.00519290 + layer.2.k_cache 0.00794396 1.17723911 + layer.2.v_cache 0.00001935 0.01634628 + layer.3.k_cache 0.04207947 5.93371214 + layer.3.v_cache 0.00002010 0.01978333 + layer.4.k_cache 0.00062246 0.40537703 + layer.4.v_cache 0.00005226 0.03573789 + layer.4.output 0.17535226 651.77527969 + ------------------------------------------------------------------------------------- + TOTAL 0.08154460 272.32878303 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 88112 +BPFP 0.9757 bits/point +EBPFP 1.9515 equivalent bits/point +MSE 272.328783 +---------------------- -------------------------------------------------------- +Time: 2.498s Load: 0.005s, Pack+Encode: 1.489s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 272.3288 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,876B, BPFP=0.3449 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,240B, BPFP=2.2500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,144B, BPFP=0.5779 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,020B, BPFP=2.2096 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,860B, BPFP=0.7096 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,740B, BPFP=2.1581 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,584B, BPFP=0.8426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,952B, BPFP=2.1971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,820B, BPFP=1.6213 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,568B, BPFP=2.1265 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,724B, BPFP=0.0978 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11944316 51.25643382 + layer.0.v_cache 0.00001525 0.01280368 + layer.1.k_cache 0.01376067 5.63681353 + layer.1.v_cache 0.00000561 0.00523437 + layer.2.k_cache 0.00522648 1.20930894 + layer.2.v_cache 0.00001966 0.01590450 + layer.3.k_cache 0.04600189 5.19142851 + layer.3.v_cache 0.00001912 0.01909743 + layer.4.k_cache 0.00062079 0.36793760 + layer.4.v_cache 0.00005114 0.03571577 + layer.4.output 0.17175110 636.20283613 + ------------------------------------------------------------------------------------- + TOTAL 0.08161303 265.71591359 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 85528 +BPFP 0.9248 bits/point +EBPFP 1.8497 equivalent bits/point +MSE 265.715914 +---------------------- -------------------------------------------------------- +Time: 2.501s Load: 0.003s, Pack+Encode: 1.493s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 265.7159 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,804B, BPFP=0.3614 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,064B, BPFP=2.4167 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,996B, BPFP=0.6002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,864B, BPFP=2.3766 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,064B, BPFP=0.8141 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,328B, BPFP=2.2692 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,188B, BPFP=0.6386 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,764B, BPFP=2.3566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,816B, BPFP=1.5657 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,384B, BPFP=2.0801 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,868B, BPFP=0.1107 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.008s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120220 56.65792768 + layer.0.v_cache 0.00001368 0.01288002 + layer.1.k_cache 0.03273370 5.74424587 + layer.1.v_cache 0.00000573 0.00528372 + layer.2.k_cache 0.00892877 1.26026789 + layer.2.v_cache 0.00001953 0.01558571 + layer.3.k_cache 0.02974504 4.69399633 + layer.3.v_cache 0.00001879 0.01905393 + layer.4.k_cache 0.00063874 0.39910338 + layer.4.v_cache 0.00005200 0.03732147 + layer.4.output 0.17853031 692.49404762 + ------------------------------------------------------------------------------------- + TOTAL 0.08429825 289.19435290 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 81140 +BPFP 0.9561 bits/point +EBPFP 1.9122 equivalent bits/point +MSE 289.194353 +---------------------- -------------------------------------------------------- +Time: 2.503s Load: 0.004s, Pack+Encode: 1.492s, Decode+Unpack: 1.008s +---------------------- -------------------------------------------------------- +💾 Converting with 289.1944 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,832B, BPFP=0.3449 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,532B, BPFP=2.3592 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,092B, BPFP=0.5821 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,168B, BPFP=2.2907 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,492B, BPFP=0.8456 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,716B, BPFP=2.2056 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,008B, BPFP=0.9428 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,028B, BPFP=2.2643 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,080B, BPFP=1.8976 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,556B, BPFP=2.1755 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,864B, BPFP=0.1039 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09802409 53.90109657 + layer.0.v_cache 0.00001402 0.01279936 + layer.1.k_cache 0.03453328 6.62452569 + layer.1.v_cache 0.00000594 0.00546752 + layer.2.k_cache 0.01027090 1.12103740 + layer.2.v_cache 0.00001933 0.01534192 + layer.3.k_cache 0.02613793 4.93512294 + layer.3.v_cache 0.00001931 0.01928795 + layer.4.k_cache 0.00063818 0.39179859 + layer.4.v_cache 0.00005393 0.03481721 + layer.4.output 0.18084440 652.36725473 + ------------------------------------------------------------------------------------- + TOTAL 0.08444869 272.56659284 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 88368 +BPFP 0.9786 bits/point +EBPFP 1.9571 equivalent bits/point +MSE 272.566593 +---------------------- -------------------------------------------------------- +Time: 2.497s Load: 0.004s, Pack+Encode: 1.489s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 272.5666 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,196B, BPFP=0.3331 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,420B, BPFP=1.8841 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,748B, BPFP=0.5686 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,156B, BPFP=1.8441 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,792B, BPFP=1.0303 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,660B, BPFP=1.7688 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,196B, BPFP=1.0916 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,908B, BPFP=1.8064 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,340B, BPFP=1.7203 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,396B, BPFP=1.7288 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,308B, BPFP=0.0934 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13030120 53.95269835 + layer.0.v_cache 0.00001385 0.01248967 + layer.1.k_cache 0.02890817 6.33102447 + layer.1.v_cache 0.00000611 0.00539570 + layer.2.k_cache 0.00721959 1.24586146 + layer.2.v_cache 0.00001830 0.01527131 + layer.3.k_cache 0.02195997 6.51813277 + layer.3.v_cache 0.00001903 0.01827255 + layer.4.k_cache 0.00063766 0.41546068 + layer.4.v_cache 0.00005061 0.03620588 + layer.4.output 11.07892277 519.67662101 + ------------------------------------------------------------------------------------- + TOTAL 4.57303494 218.01689176 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 95120 +BPFP 0.8488 bits/point +EBPFP 1.6976 equivalent bits/point +MSE 218.016892 +---------------------- -------------------------------------------------------- +Time: 2.502s Load: 0.005s, Pack+Encode: 1.490s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 218.0169 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.3502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,504B, BPFP=2.3539 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,044B, BPFP=0.5730 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,148B, BPFP=2.2869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,540B, BPFP=0.8547 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,824B, BPFP=2.2259 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,344B, BPFP=0.8178 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,080B, BPFP=2.2741 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,708B, BPFP=2.0158 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,552B, BPFP=2.1747 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,548B, BPFP=0.0954 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12633393 53.63897779 + layer.0.v_cache 0.00001372 0.01276824 + layer.1.k_cache 0.01411889 5.83928561 + layer.1.v_cache 0.00000549 0.00526495 + layer.2.k_cache 0.00833277 0.97353777 + layer.2.v_cache 0.00001851 0.01536314 + layer.3.k_cache 0.05797521 5.67932386 + layer.3.v_cache 0.00001879 0.01823610 + layer.4.k_cache 0.00064672 0.41657652 + layer.4.v_cache 0.00005349 0.03614699 + layer.4.output 0.17447430 649.93965146 + ------------------------------------------------------------------------------------- + TOTAL 0.08404927 271.54194360 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 88152 +BPFP 0.9762 bits/point +EBPFP 1.9523 equivalent bits/point +MSE 271.541944 +---------------------- -------------------------------------------------------- +Time: 2.505s Load: 0.004s, Pack+Encode: 1.495s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 271.5419 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,672B, BPFP=0.3732 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,920B, BPFP=2.6607 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,656B, BPFP=0.5929 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,612B, BPFP=2.5920 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,092B, BPFP=0.6902 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,432B, BPFP=2.5518 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,592B, BPFP=0.5786 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,604B, BPFP=2.5902 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,596B, BPFP=1.6955 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,096B, BPFP=2.4768 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,460B, BPFP=0.1103 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.003s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08086312 53.96932199 + layer.0.v_cache 0.00001331 0.01383736 + layer.1.k_cache 0.01551799 6.15891244 + layer.1.v_cache 0.00000513 0.00524494 + layer.2.k_cache 0.00232868 1.22524763 + layer.2.v_cache 0.00002262 0.01779096 + layer.3.k_cache 0.10623371 4.56787240 + layer.3.v_cache 0.00001836 0.02007095 + layer.4.k_cache 0.00061897 0.42388652 + layer.4.v_cache 0.00005922 0.03851412 + layer.4.output 0.20964142 773.51141582 + ------------------------------------------------------------------------------------- + TOTAL 0.09842183 322.41297706 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 78732 +BPFP 1.0338 bits/point +EBPFP 2.0675 equivalent bits/point +MSE 322.412977 +---------------------- -------------------------------------------------------- +Time: 2.500s Load: 0.003s, Pack+Encode: 1.494s, Decode+Unpack: 1.003s +---------------------- -------------------------------------------------------- +💾 Converting with 322.4130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,912B, BPFP=0.3247 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,360B, BPFP=2.0992 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,548B, BPFP=0.6026 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,856B, BPFP=2.0136 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,628B, BPFP=0.7860 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,556B, BPFP=1.9626 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,004B, BPFP=0.6800 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,552B, BPFP=1.9620 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,820B, BPFP=1.6678 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,268B, BPFP=1.9137 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,296B, BPFP=0.1042 +⌛️ [2/4] FRONTEND: Frontend time: 1.486s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13670269 55.64325217 + layer.0.v_cache 0.00001568 0.01296292 + layer.1.k_cache 0.03139175 5.07816248 + layer.1.v_cache 0.00000602 0.00555208 + layer.2.k_cache 0.00629326 1.20785970 + layer.2.v_cache 0.00001871 0.01532094 + layer.3.k_cache 0.03837063 5.27051047 + layer.3.v_cache 0.00001860 0.01840807 + layer.4.k_cache 0.00062597 0.40079871 + layer.4.v_cache 0.00005354 0.03728147 + layer.4.output 0.15974679 588.68793672 + ------------------------------------------------------------------------------------- + TOTAL 0.07833673 246.38268624 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 86800 +BPFP 0.8672 bits/point +EBPFP 1.7343 equivalent bits/point +MSE 246.382686 +---------------------- -------------------------------------------------------- +Time: 2.494s Load: 0.003s, Pack+Encode: 1.486s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 246.3827 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,240B, BPFP=0.3271 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,536B, BPFP=1.8306 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,848B, BPFP=0.5619 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,180B, BPFP=1.7786 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,480B, BPFP=0.8002 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,880B, BPFP=1.7348 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,316B, BPFP=1.0683 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,980B, BPFP=1.7494 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,120B, BPFP=1.6238 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,384B, BPFP=1.6624 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,092B, BPFP=0.0854 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12281527 55.17963274 + layer.0.v_cache 0.00001417 0.01268040 + layer.1.k_cache 0.04442539 6.01161779 + layer.1.v_cache 0.00000572 0.00528145 + layer.2.k_cache 0.01033308 1.07089618 + layer.2.v_cache 0.00001923 0.01605602 + layer.3.k_cache 0.04555132 7.06707707 + layer.3.v_cache 0.00001841 0.01925097 + layer.4.k_cache 0.00060580 0.40111627 + layer.4.v_cache 0.00007068 0.03683567 + layer.4.output 10.66341841 498.03125000 + ------------------------------------------------------------------------------------- + TOTAL 4.40398753 209.17877615 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 94056 +BPFP 0.8079 bits/point +EBPFP 1.6159 equivalent bits/point +MSE 209.178776 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.005s, Pack+Encode: 1.493s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 209.1788 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3343 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,444B, BPFP=2.2609 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,188B, BPFP=0.5792 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,188B, BPFP=2.2144 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,604B, BPFP=0.8365 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,804B, BPFP=2.1446 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,340B, BPFP=1.1519 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,848B, BPFP=2.1526 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,488B, BPFP=1.9055 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,552B, BPFP=2.0988 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,780B, BPFP=0.0981 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11362342 54.88641215 + layer.0.v_cache 0.00001418 0.01309574 + layer.1.k_cache 0.05429294 6.11084836 + layer.1.v_cache 0.00000580 0.00540912 + layer.2.k_cache 0.00513919 1.11724321 + layer.2.v_cache 0.00001816 0.01583173 + layer.3.k_cache 0.05885953 6.00528167 + layer.3.v_cache 0.00001841 0.01857091 + layer.4.k_cache 0.00061763 0.40157225 + layer.4.v_cache 0.00005080 0.03550142 + layer.4.output 0.16727475 628.13584925 + ------------------------------------------------------------------------------------- + TOTAL 0.08256255 262.68004184 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 90076 +BPFP 0.9627 bits/point +EBPFP 1.9254 equivalent bits/point +MSE 262.680042 +---------------------- -------------------------------------------------------- +Time: 2.501s Load: 0.003s, Pack+Encode: 1.492s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 262.6800 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,920B, BPFP=0.3226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,356B, BPFP=2.0759 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,324B, BPFP=0.5585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,936B, BPFP=2.0054 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,212B, BPFP=0.8757 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,580B, BPFP=1.9456 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,236B, BPFP=1.3837 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,696B, BPFP=1.9651 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,264B, BPFP=1.8925 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,348B, BPFP=1.9066 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,944B, BPFP=0.0947 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702597 57.51833942 + layer.0.v_cache 0.00001402 0.01344440 + layer.1.k_cache 0.01332537 5.82748249 + layer.1.v_cache 0.00000581 0.00557006 + layer.2.k_cache 0.00507878 1.28132416 + layer.2.v_cache 0.00001944 0.01631751 + layer.3.k_cache 0.04051446 6.61014680 + layer.3.v_cache 0.00001931 0.01966526 + layer.4.k_cache 0.00062870 0.40679981 + layer.4.v_cache 0.00006514 0.03761641 + layer.4.output 0.16549673 580.73343894 + ------------------------------------------------------------------------------------- + TOTAL 0.07795142 243.34533994 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 92816 +BPFP 0.9173 bits/point +EBPFP 1.8346 equivalent bits/point +MSE 243.345340 +---------------------- -------------------------------------------------------- +Time: 2.505s Load: 0.004s, Pack+Encode: 1.494s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 243.3453 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3412 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,404B, BPFP=2.2801 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,088B, BPFP=0.5676 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,964B, BPFP=2.1993 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,184B, BPFP=0.9529 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,648B, BPFP=2.1412 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,016B, BPFP=0.7382 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,656B, BPFP=2.1426 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,988B, BPFP=1.6522 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,464B, BPFP=2.1074 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,980B, BPFP=0.1045 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.008s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13925676 54.66869256 + layer.0.v_cache 0.00001387 0.01289764 + layer.1.k_cache 0.03290265 6.01129114 + layer.1.v_cache 0.00000589 0.00533539 + layer.2.k_cache 0.00800494 1.09151755 + layer.2.v_cache 0.00001921 0.01513241 + layer.3.k_cache 0.08301921 5.09086052 + layer.3.v_cache 0.00001931 0.01786948 + layer.4.k_cache 0.00065893 0.38251818 + layer.4.v_cache 0.00004891 0.03395879 + layer.4.output 0.17367665 632.35089286 + ------------------------------------------------------------------------------------- + TOTAL 0.08704037 264.34037198 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 86248 +BPFP 0.9326 bits/point +EBPFP 1.8652 equivalent bits/point +MSE 264.340372 +---------------------- -------------------------------------------------------- +Time: 2.503s Load: 0.005s, Pack+Encode: 1.490s, Decode+Unpack: 1.008s +---------------------- -------------------------------------------------------- +💾 Converting with 264.3404 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,188B, BPFP=0.3319 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,440B, BPFP=1.8871 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,744B, BPFP=0.5680 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,136B, BPFP=1.8410 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,860B, BPFP=1.0407 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,696B, BPFP=1.7743 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,880B, BPFP=1.0437 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,916B, BPFP=1.8076 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,532B, BPFP=1.7494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,324B, BPFP=1.7178 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,900B, BPFP=0.0845 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10583916 52.86882775 + layer.0.v_cache 0.00001406 0.01250949 + layer.1.k_cache 0.01351924 6.93359256 + layer.1.v_cache 0.00000536 0.00504166 + layer.2.k_cache 0.00567923 1.23678419 + layer.2.v_cache 0.00001856 0.01564193 + layer.3.k_cache 0.05293757 6.48502328 + layer.3.v_cache 0.00001841 0.01832272 + layer.4.k_cache 0.00064880 0.42897063 + layer.4.v_cache 0.00004836 0.03608236 + layer.4.output 11.07886250 504.94577843 + ------------------------------------------------------------------------------------- + TOTAL 4.57239802 211.92124974 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 94616 +BPFP 0.8443 bits/point +EBPFP 1.6886 equivalent bits/point +MSE 211.921250 +---------------------- -------------------------------------------------------- +Time: 2.510s Load: 0.004s, Pack+Encode: 1.495s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 211.9212 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.3502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,468B, BPFP=2.3471 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,056B, BPFP=0.5753 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,232B, BPFP=2.3027 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,108B, BPFP=0.7733 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,844B, BPFP=2.2297 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,560B, BPFP=0.6702 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,936B, BPFP=2.2470 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,916B, BPFP=1.8667 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,428B, BPFP=2.1514 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,728B, BPFP=0.1003 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860243 51.99254047 + layer.0.v_cache 0.00001403 0.01285318 + layer.1.k_cache 0.01400426 5.07172403 + layer.1.v_cache 0.00000584 0.00529408 + layer.2.k_cache 0.01277778 1.04612015 + layer.2.v_cache 0.00001980 0.01633198 + layer.3.k_cache 0.07376247 5.80445623 + layer.3.v_cache 0.00001877 0.01931833 + layer.4.k_cache 0.00062152 0.41070534 + layer.4.v_cache 0.00005140 0.03734862 + layer.4.output 0.18179212 651.15926205 + ------------------------------------------------------------------------------------- + TOTAL 0.08837783 271.91361922 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 86136 +BPFP 0.9538 bits/point +EBPFP 1.9077 equivalent bits/point +MSE 271.913619 +---------------------- -------------------------------------------------------- +Time: 2.514s Load: 0.004s, Pack+Encode: 1.494s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 271.9136 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,896B, BPFP=0.3292 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,472B, BPFP=2.1653 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,204B, BPFP=0.5563 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,072B, BPFP=2.0958 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,864B, BPFP=0.8444 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,788B, BPFP=2.0465 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,928B, BPFP=0.8556 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,044B, BPFP=2.0910 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,404B, BPFP=1.9799 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,552B, BPFP=2.0056 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,788B, BPFP=0.0939 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.025s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13781073 56.15878906 + layer.0.v_cache 0.00001414 0.01344473 + layer.1.k_cache 0.01252146 6.19989827 + layer.1.v_cache 0.00000561 0.00534500 + layer.2.k_cache 0.00238173 1.07931824 + layer.2.v_cache 0.00001849 0.01642966 + layer.3.k_cache 0.04360520 5.56554565 + layer.3.v_cache 0.00001973 0.02024542 + layer.4.k_cache 0.00061668 0.40352512 + layer.4.v_cache 0.00005448 0.03838905 + layer.4.output 0.16772948 601.57316468 + ------------------------------------------------------------------------------------- + TOTAL 0.08065615 251.79488723 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 90012 +BPFP 0.9192 bits/point +EBPFP 1.8385 equivalent bits/point +MSE 251.794887 +---------------------- -------------------------------------------------------- +Time: 2.525s Load: 0.004s, Pack+Encode: 1.496s, Decode+Unpack: 1.025s +---------------------- -------------------------------------------------------- +💾 Converting with 251.7949 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,844B, BPFP=0.3647 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,352B, BPFP=2.4430 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,924B, BPFP=0.5783 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,888B, BPFP=2.3513 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,416B, BPFP=0.6756 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,712B, BPFP=2.3165 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,176B, BPFP=0.6282 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,748B, BPFP=2.3236 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,264B, BPFP=1.8323 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,424B, BPFP=2.2595 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,916B, BPFP=0.1106 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07550411 52.85597557 + layer.0.v_cache 0.00001381 0.01309584 + layer.1.k_cache 0.01271453 5.74315325 + layer.1.v_cache 0.00000531 0.00523234 + layer.2.k_cache 0.00681068 1.07219609 + layer.2.v_cache 0.00001904 0.01619271 + layer.3.k_cache 0.02844424 4.95302804 + layer.3.v_cache 0.00001798 0.01946003 + layer.4.k_cache 0.00061127 0.40286361 + layer.4.v_cache 0.00009323 0.03830610 + layer.4.output 0.18255964 683.70162749 + ------------------------------------------------------------------------------------- + TOTAL 0.08247951 285.35475859 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 83664 +BPFP 0.9734 bits/point +EBPFP 1.9468 equivalent bits/point +MSE 285.354759 +---------------------- -------------------------------------------------------- +Time: 2.520s Load: 0.005s, Pack+Encode: 1.500s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 285.3548 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,864B, BPFP=0.3641 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,308B, BPFP=2.4039 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,012B, BPFP=0.5883 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,812B, BPFP=2.3070 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,504B, BPFP=0.8797 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,548B, BPFP=2.2555 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,504B, BPFP=0.6844 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,804B, BPFP=2.3055 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,292B, BPFP=1.8148 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,248B, BPFP=2.1969 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,988B, BPFP=0.1113 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08590809 48.42576904 + layer.0.v_cache 0.00001388 0.01296300 + layer.1.k_cache 0.01594085 6.45068665 + layer.1.v_cache 0.00000552 0.00537964 + layer.2.k_cache 0.00791870 1.11735001 + layer.2.v_cache 0.00001801 0.01581604 + layer.3.k_cache 0.02757127 4.93420067 + layer.3.v_cache 0.00001862 0.01946979 + layer.4.k_cache 0.00063773 0.40379810 + layer.4.v_cache 0.00005492 0.03731033 + layer.4.output 0.18408036 674.38180804 + ------------------------------------------------------------------------------------- + TOTAL 0.08392059 281.29972938 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 84884 +BPFP 0.9752 bits/point +EBPFP 1.9505 equivalent bits/point +MSE 281.299729 +---------------------- -------------------------------------------------------- +Time: 2.511s Load: 0.003s, Pack+Encode: 1.494s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 281.2997 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,768B, BPFP=0.3635 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,084B, BPFP=2.4844 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,896B, BPFP=0.5954 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,520B, BPFP=2.3684 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,336B, BPFP=0.6859 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,404B, BPFP=2.3446 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,976B, BPFP=0.6118 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,500B, BPFP=2.3643 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,596B, BPFP=1.5617 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,964B, BPFP=2.2541 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,576B, BPFP=0.1050 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08991440 53.89999872 + layer.0.v_cache 0.00001347 0.01271871 + layer.1.k_cache 0.03345199 6.26512708 + layer.1.v_cache 0.00000518 0.00520916 + layer.2.k_cache 0.00398020 1.06561500 + layer.2.v_cache 0.00001815 0.01586180 + layer.3.k_cache 0.04607732 4.43554125 + layer.3.v_cache 0.00001896 0.01863946 + layer.4.k_cache 0.00062119 0.39098208 + layer.4.v_cache 0.00005040 0.03584028 + layer.4.output 0.18612516 709.98731203 + ------------------------------------------------------------------------------------- + TOTAL 0.08688396 296.23863046 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 79620 +BPFP 0.9629 bits/point +EBPFP 1.9258 equivalent bits/point +MSE 296.238630 +---------------------- -------------------------------------------------------- +Time: 2.514s Load: 0.005s, Pack+Encode: 1.495s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 296.2386 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,932B, BPFP=0.3317 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,312B, BPFP=2.1140 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,192B, BPFP=0.5481 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,000B, BPFP=2.0604 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,456B, BPFP=0.9368 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,668B, BPFP=2.0034 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,460B, BPFP=1.2809 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,732B, BPFP=2.0144 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,968B, BPFP=1.8832 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,340B, BPFP=1.9471 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,952B, BPFP=0.0969 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.021s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11186004 55.66706731 + layer.0.v_cache 0.00001364 0.01311161 + layer.1.k_cache 0.01386243 6.14071890 + layer.1.v_cache 0.00000608 0.00528119 + layer.2.k_cache 0.00506424 1.00462509 + layer.2.v_cache 0.00001872 0.01577806 + layer.3.k_cache 0.02496567 5.98034064 + layer.3.v_cache 0.00001876 0.02009828 + layer.4.k_cache 0.00064518 0.42381941 + layer.4.v_cache 0.00005436 0.03719622 + layer.4.output 0.16682192 594.08202512 + ------------------------------------------------------------------------------------- + TOTAL 0.07789780 248.69895368 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 92012 +BPFP 0.9293 bits/point +EBPFP 1.8587 equivalent bits/point +MSE 248.698954 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.004s, Pack+Encode: 1.494s, Decode+Unpack: 1.021s +---------------------- -------------------------------------------------------- +💾 Converting with 248.6990 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3494 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,388B, BPFP=2.3321 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,992B, BPFP=0.5633 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,024B, BPFP=2.2636 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,316B, BPFP=0.8125 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,736B, BPFP=2.2093 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,412B, BPFP=0.6423 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,912B, BPFP=2.2425 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,980B, BPFP=2.0670 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,392B, BPFP=2.1446 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,660B, BPFP=0.0984 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.029s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08542032 53.22080902 + layer.0.v_cache 0.00001402 0.01298635 + layer.1.k_cache 0.01083013 5.69878341 + layer.1.v_cache 0.00000556 0.00524284 + layer.2.k_cache 0.00382075 1.05301381 + layer.2.v_cache 0.00001860 0.01575017 + layer.3.k_cache 0.02610962 4.60466647 + layer.3.v_cache 0.00002051 0.01902877 + layer.4.k_cache 0.00063312 0.38816034 + layer.4.v_cache 0.00005614 0.03808947 + layer.4.output 0.16811641 646.68351979 + ------------------------------------------------------------------------------------- + TOTAL 0.07669080 270.10830407 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 86668 +BPFP 0.9597 bits/point +EBPFP 1.9195 equivalent bits/point +MSE 270.108304 +---------------------- -------------------------------------------------------- +Time: 2.526s Load: 0.004s, Pack+Encode: 1.493s, Decode+Unpack: 1.029s +---------------------- -------------------------------------------------------- +💾 Converting with 270.1083 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,836B, BPFP=0.3542 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,368B, BPFP=2.3858 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,068B, BPFP=0.5918 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,096B, BPFP=2.3333 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,672B, BPFP=0.7083 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,784B, BPFP=2.2731 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,768B, BPFP=0.7269 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,992B, BPFP=2.3133 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,660B, BPFP=1.8634 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,572B, BPFP=2.2323 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,856B, BPFP=0.1063 +⌛️ [2/4] FRONTEND: Frontend time: 1.505s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10540552 50.68627025 + layer.0.v_cache 0.00001541 0.01339267 + layer.1.k_cache 0.03402772 6.07434270 + layer.1.v_cache 0.00000555 0.00528152 + layer.2.k_cache 0.00394259 1.01011149 + layer.2.v_cache 0.00001868 0.01656290 + layer.3.k_cache 0.02873192 5.65822799 + layer.3.v_cache 0.00001983 0.02003388 + layer.4.k_cache 0.00063741 0.39875808 + layer.4.v_cache 0.00005806 0.03832417 + layer.4.output 0.17961645 665.89203042 + ------------------------------------------------------------------------------------- + TOTAL 0.08412811 277.95091286 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 85672 +BPFP 0.9721 bits/point +EBPFP 1.9443 equivalent bits/point +MSE 277.950913 +---------------------- -------------------------------------------------------- +Time: 2.523s Load: 0.004s, Pack+Encode: 1.505s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 277.9509 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3625 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,380B, BPFP=2.4180 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,040B, BPFP=0.5938 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,048B, BPFP=2.3531 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,068B, BPFP=0.9898 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,608B, BPFP=2.2672 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,492B, BPFP=0.6820 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,936B, BPFP=2.3312 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,548B, BPFP=1.8648 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,308B, BPFP=2.2086 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,976B, BPFP=0.1109 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10022384 55.97363281 + layer.0.v_cache 0.00001350 0.01317728 + layer.1.k_cache 0.03445883 5.43388710 + layer.1.v_cache 0.00000556 0.00543503 + layer.2.k_cache 0.00550650 1.35805645 + layer.2.v_cache 0.00001848 0.01607915 + layer.3.k_cache 0.03016714 4.55925369 + layer.3.v_cache 0.00001937 0.02071009 + layer.4.k_cache 0.00063435 0.40084124 + layer.4.v_cache 0.00006510 0.03864052 + layer.4.output 0.18593751 676.24860491 + ------------------------------------------------------------------------------------- + TOTAL 0.08662796 282.44470281 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 86260 +BPFP 0.9910 bits/point +EBPFP 1.9821 equivalent bits/point +MSE 282.444703 +---------------------- -------------------------------------------------------- +Time: 2.510s Load: 0.004s, Pack+Encode: 1.495s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 282.4447 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,932B, BPFP=0.3112 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,476B, BPFP=2.0097 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,276B, BPFP=0.5277 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,028B, BPFP=1.9375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,264B, BPFP=0.8479 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,668B, BPFP=1.8795 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,768B, BPFP=1.0902 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,932B, BPFP=1.9220 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,224B, BPFP=1.8080 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,304B, BPFP=1.8209 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,916B, BPFP=0.0901 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11898830 56.64913720 + layer.0.v_cache 0.00001403 0.01286368 + layer.1.k_cache 0.01243949 6.26286395 + layer.1.v_cache 0.00000551 0.00520248 + layer.2.k_cache 0.00346032 1.11528141 + layer.2.v_cache 0.00001901 0.01627083 + layer.3.k_cache 0.02403709 6.19001251 + layer.3.v_cache 0.00001994 0.01915440 + layer.4.k_cache 0.00063949 0.41694602 + layer.4.v_cache 0.00006591 0.03723588 + layer.4.output 0.03773703 572.82561672 + ------------------------------------------------------------------------------------- + TOTAL 0.02493225 240.02966385 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 91788 +BPFP 0.8697 bits/point +EBPFP 1.7395 equivalent bits/point +MSE 240.029664 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.004s, Pack+Encode: 1.497s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 240.0297 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,824B, BPFP=0.3476 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,236B, BPFP=2.3316 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,992B, BPFP=0.5701 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,096B, BPFP=2.3049 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,432B, BPFP=0.8445 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,724B, BPFP=2.2340 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,396B, BPFP=0.6471 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,968B, BPFP=2.2805 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,424B, BPFP=1.9863 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,544B, BPFP=2.1997 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,812B, BPFP=0.1038 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10693085 53.87162967 + layer.0.v_cache 0.00001341 0.01225017 + layer.1.k_cache 0.01172146 5.56450225 + layer.1.v_cache 0.00000553 0.00526524 + layer.2.k_cache 0.00236719 1.10535068 + layer.2.v_cache 0.00001835 0.01518925 + layer.3.k_cache 0.02622769 5.30297479 + layer.3.v_cache 0.00001864 0.01813439 + layer.4.k_cache 0.00061713 0.40031824 + layer.4.v_cache 0.00004943 0.03483095 + layer.4.output 0.18273007 658.87543554 + ------------------------------------------------------------------------------------- + TOTAL 0.08394589 275.20344085 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 86448 +BPFP 0.9690 bits/point +EBPFP 1.9379 equivalent bits/point +MSE 275.203441 +---------------------- -------------------------------------------------------- +Time: 2.517s Load: 0.004s, Pack+Encode: 1.496s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 275.2034 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,136B, BPFP=0.3304 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,464B, BPFP=1.9282 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,632B, BPFP=0.5619 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,152B, BPFP=1.8800 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,136B, BPFP=0.7946 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,728B, BPFP=1.8144 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,768B, BPFP=1.2017 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,788B, BPFP=1.8236 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,636B, BPFP=1.8001 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,428B, BPFP=1.7679 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,012B, BPFP=0.0887 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13156438 56.77203550 + layer.0.v_cache 0.00001398 0.01289524 + layer.1.k_cache 0.04507210 5.64586556 + layer.1.v_cache 0.00000603 0.00551081 + layer.2.k_cache 0.00848899 0.94614592 + layer.2.v_cache 0.00001983 0.01616855 + layer.3.k_cache 0.06093804 7.55217720 + layer.3.v_cache 0.00001891 0.01837160 + layer.4.k_cache 0.00062755 0.43087089 + layer.4.v_cache 0.00004781 0.03393644 + layer.4.output 11.29576791 530.43148868 + ------------------------------------------------------------------------------------- + TOTAL 4.66571606 222.61496462 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 93880 +BPFP 0.8543 bits/point +EBPFP 1.7086 equivalent bits/point +MSE 222.614965 +---------------------- -------------------------------------------------------- +Time: 2.528s Load: 0.005s, Pack+Encode: 1.503s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 222.6150 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,852B, BPFP=0.3486 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,548B, BPFP=2.3622 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,036B, BPFP=0.5715 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,252B, BPFP=2.3065 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,100B, BPFP=0.9601 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,848B, BPFP=2.2304 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,532B, BPFP=0.8532 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,180B, BPFP=2.2929 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,136B, BPFP=1.7199 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,536B, BPFP=2.1717 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,612B, BPFP=0.0971 +⌛️ [2/4] FRONTEND: Frontend time: 1.499s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.015s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12340720 51.41866293 + layer.0.v_cache 0.00001379 0.01294280 + layer.1.k_cache 0.01365945 5.34605941 + layer.1.v_cache 0.00000586 0.00540773 + layer.2.k_cache 0.00515996 1.20381450 + layer.2.v_cache 0.00002118 0.01637016 + layer.3.k_cache 0.04682968 6.03269462 + layer.3.v_cache 0.00001899 0.01923234 + layer.4.k_cache 0.00062596 0.40828875 + layer.4.v_cache 0.00004975 0.03542353 + layer.4.output 0.18155291 652.49230852 + ------------------------------------------------------------------------------------- + TOTAL 0.08592130 272.46735626 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 87632 +BPFP 0.9704 bits/point +EBPFP 1.9408 equivalent bits/point +MSE 272.467356 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.005s, Pack+Encode: 1.499s, Decode+Unpack: 1.015s +---------------------- -------------------------------------------------------- +💾 Converting with 272.4674 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3423 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,500B, BPFP=2.3251 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,084B, BPFP=0.5737 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,116B, BPFP=2.2537 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,408B, BPFP=0.8199 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,728B, BPFP=2.1815 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,320B, BPFP=0.8036 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,976B, BPFP=2.2277 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,844B, BPFP=2.0171 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,380B, BPFP=2.1168 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,944B, BPFP=0.1048 +⌛️ [2/4] FRONTEND: Frontend time: 1.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10400932 56.60405622 + layer.0.v_cache 0.00001381 0.01275653 + layer.1.k_cache 0.01300498 5.73692867 + layer.1.v_cache 0.00000559 0.00517849 + layer.2.k_cache 0.00535372 1.20122855 + layer.2.v_cache 0.00001862 0.01540654 + layer.3.k_cache 0.04415199 5.78913371 + layer.3.v_cache 0.00001841 0.01870412 + layer.4.k_cache 0.00060249 0.38013499 + layer.4.v_cache 0.00005274 0.03583778 + layer.4.output 0.16694923 643.70907738 + ------------------------------------------------------------------------------------- + TOTAL 0.07858096 269.16252396 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 88140 +BPFP 0.9644 bits/point +EBPFP 1.9288 equivalent bits/point +MSE 269.162524 +---------------------- -------------------------------------------------------- +Time: 2.514s Load: 0.004s, Pack+Encode: 1.494s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 269.1625 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3594 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,488B, BPFP=2.4391 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,948B, BPFP=0.5758 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,088B, BPFP=2.3609 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,732B, BPFP=0.9242 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,684B, BPFP=2.2820 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,300B, BPFP=0.6445 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,984B, BPFP=2.3406 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,336B, BPFP=1.8234 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,388B, BPFP=2.2242 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,744B, BPFP=0.1045 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.015s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09321334 56.63485718 + layer.0.v_cache 0.00001862 0.01357518 + layer.1.k_cache 0.03257937 6.25235023 + layer.1.v_cache 0.00000532 0.00506879 + layer.2.k_cache 0.01442847 1.16823578 + layer.2.v_cache 0.00001865 0.01596343 + layer.3.k_cache 0.04385926 4.80168037 + layer.3.v_cache 0.00001828 0.01848393 + layer.4.k_cache 0.00061064 0.39095984 + layer.4.v_cache 0.00005353 0.03698569 + layer.4.output 0.18500509 674.97477679 + ------------------------------------------------------------------------------------- + TOTAL 0.08704948 282.00950576 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 85532 +BPFP 0.9827 bits/point +EBPFP 1.9653 equivalent bits/point +MSE 282.009506 +---------------------- -------------------------------------------------------- +Time: 2.515s Load: 0.005s, Pack+Encode: 1.496s, Decode+Unpack: 1.015s +---------------------- -------------------------------------------------------- +💾 Converting with 282.0095 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,220B, BPFP=0.3368 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,424B, BPFP=1.8847 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,704B, BPFP=0.5619 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,188B, BPFP=1.8489 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,116B, BPFP=0.9278 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,668B, BPFP=1.7700 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,240B, BPFP=1.0983 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,880B, BPFP=1.8022 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,532B, BPFP=1.7494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,444B, BPFP=1.7360 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,088B, BPFP=0.0886 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11527186 51.70614002 + layer.0.v_cache 0.00001510 0.01257216 + layer.1.k_cache 0.02920971 6.83862305 + layer.1.v_cache 0.00000551 0.00521450 + layer.2.k_cache 0.00857845 1.14878238 + layer.2.v_cache 0.00001837 0.01585833 + layer.3.k_cache 0.03550095 6.76503480 + layer.3.v_cache 0.00001950 0.01919658 + layer.4.k_cache 0.00063242 0.41230707 + layer.4.v_cache 0.00005150 0.03738868 + layer.4.output 11.08123415 518.93290569 + ------------------------------------------------------------------------------------- + TOTAL 4.57399661 217.61714455 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 94504 +BPFP 0.8433 bits/point +EBPFP 1.6866 equivalent bits/point +MSE 217.617145 +---------------------- -------------------------------------------------------- +Time: 2.511s Load: 0.005s, Pack+Encode: 1.489s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 217.6171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,920B, BPFP=0.3261 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,280B, BPFP=2.0856 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,296B, BPFP=0.5598 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,948B, BPFP=2.0292 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,660B, BPFP=0.7914 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,652B, BPFP=1.9789 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,004B, BPFP=1.0197 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,720B, BPFP=1.9905 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,288B, BPFP=1.9171 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,500B, BPFP=1.9531 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,536B, BPFP=0.0858 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.022s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09673090 57.05609927 + layer.0.v_cache 0.00001344 0.01304485 + layer.1.k_cache 0.04918879 5.83355315 + layer.1.v_cache 0.00000572 0.00545178 + layer.2.k_cache 0.00880511 1.04011768 + layer.2.v_cache 0.00001944 0.01600767 + layer.3.k_cache 0.06586690 6.29088825 + layer.3.v_cache 0.00001823 0.01879281 + layer.4.k_cache 0.00062484 0.42104862 + layer.4.v_cache 0.00006325 0.03599745 + layer.4.output 0.15294319 587.99374030 + ------------------------------------------------------------------------------------- + TOTAL 0.07599641 246.27571668 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 89804 +BPFP 0.8972 bits/point +EBPFP 1.7944 equivalent bits/point +MSE 246.275717 +---------------------- -------------------------------------------------------- +Time: 2.530s Load: 0.005s, Pack+Encode: 1.503s, Decode+Unpack: 1.022s +---------------------- -------------------------------------------------------- +💾 Converting with 246.2757 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,932B, BPFP=0.3281 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,332B, BPFP=2.0944 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,312B, BPFP=0.5625 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,968B, BPFP=2.0326 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,868B, BPFP=0.9966 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,340B, BPFP=1.9260 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,520B, BPFP=0.9375 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,800B, BPFP=2.0041 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,848B, BPFP=1.8424 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,344B, BPFP=1.9266 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,088B, BPFP=0.0992 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12676482 54.22768767 + layer.0.v_cache 0.00001339 0.01287633 + layer.1.k_cache 0.03133728 6.19639786 + layer.1.v_cache 0.00000590 0.00534705 + layer.2.k_cache 0.00372887 1.08178230 + layer.2.v_cache 0.00001820 0.01590338 + layer.3.k_cache 0.02418065 6.80939384 + layer.3.v_cache 0.00002034 0.02062456 + layer.4.k_cache 0.00062996 0.41108795 + layer.4.v_cache 0.00006053 0.03750365 + layer.4.output 0.15389095 587.56317935 + ------------------------------------------------------------------------------------- + TOTAL 0.07435274 245.98593294 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 90352 +BPFP 0.9027 bits/point +EBPFP 1.8053 equivalent bits/point +MSE 245.985933 +---------------------- -------------------------------------------------------- +Time: 2.521s Load: 0.005s, Pack+Encode: 1.498s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 245.9859 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,908B, BPFP=0.3312 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,408B, BPFP=2.1542 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,236B, BPFP=0.5618 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,056B, BPFP=2.0931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,544B, BPFP=0.7889 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,480B, BPFP=1.9931 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,436B, BPFP=1.1174 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,628B, BPFP=2.0187 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,012B, BPFP=1.9118 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,372B, BPFP=1.9743 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,084B, BPFP=0.1013 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12863920 54.49441189 + layer.0.v_cache 0.00001425 0.01314723 + layer.1.k_cache 0.01389826 6.30484280 + layer.1.v_cache 0.00000571 0.00548302 + layer.2.k_cache 0.00659207 1.22289937 + layer.2.v_cache 0.00001883 0.01561649 + layer.3.k_cache 0.05540012 6.26949666 + layer.3.v_cache 0.00001922 0.01877717 + layer.4.k_cache 0.00063279 0.41181484 + layer.4.v_cache 0.00005416 0.03530536 + layer.4.output 0.16196846 601.11453373 + ------------------------------------------------------------------------------------- + TOTAL 0.07876787 251.56432535 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 90164 +BPFP 0.9208 bits/point +EBPFP 1.8416 equivalent bits/point +MSE 251.564325 +---------------------- -------------------------------------------------------- +Time: 2.512s Load: 0.005s, Pack+Encode: 1.493s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 251.5643 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,276B, BPFP=0.3293 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,532B, BPFP=1.8131 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,912B, BPFP=0.5660 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,164B, BPFP=1.7598 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,976B, BPFP=0.8646 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,740B, BPFP=1.6985 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,756B, BPFP=0.9774 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,976B, BPFP=1.7326 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,992B, BPFP=1.5903 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,444B, BPFP=1.6557 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,116B, BPFP=0.0851 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13160179 52.06736021 + layer.0.v_cache 0.00001577 0.01237117 + layer.1.k_cache 0.05775581 6.03301267 + layer.1.v_cache 0.00000549 0.00504520 + layer.2.k_cache 0.01027601 1.05117614 + layer.2.v_cache 0.00002197 0.01563982 + layer.3.k_cache 0.07329086 6.80206186 + layer.3.v_cache 0.00001983 0.01919435 + layer.4.k_cache 0.00064477 0.41470125 + layer.4.v_cache 0.00005260 0.03626283 + layer.4.output 10.56448284 495.15042163 + ------------------------------------------------------------------------------------- + TOTAL 4.36618028 207.79469276 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 93884 +BPFP 0.7990 bits/point +EBPFP 1.5980 equivalent bits/point +MSE 207.794693 +---------------------- -------------------------------------------------------- +Time: 2.525s Load: 0.005s, Pack+Encode: 1.503s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 207.7947 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,844B, BPFP=0.3471 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,364B, BPFP=2.3276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,028B, BPFP=0.5700 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,992B, BPFP=2.2575 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,796B, BPFP=0.9029 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,652B, BPFP=2.1935 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,232B, BPFP=0.6084 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,312B, BPFP=2.1295 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,544B, BPFP=1.7967 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,488B, BPFP=2.1627 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,628B, BPFP=0.0976 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10139442 52.16537439 + layer.0.v_cache 0.00001346 0.01275717 + layer.1.k_cache 0.03266963 5.64405179 + layer.1.v_cache 0.00000526 0.00511777 + layer.2.k_cache 0.01283502 1.02436038 + layer.2.v_cache 0.00001884 0.01563774 + layer.3.k_cache 0.02666297 4.29173481 + layer.3.v_cache 0.00001780 0.01753388 + layer.4.k_cache 0.00061512 0.38753932 + layer.4.v_cache 0.00004947 0.03429065 + layer.4.output 0.18393325 652.24397590 + ------------------------------------------------------------------------------------- + TOTAL 0.08598910 272.31213113 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 84880 +BPFP 0.9399 bits/point +EBPFP 1.8799 equivalent bits/point +MSE 272.312131 +---------------------- -------------------------------------------------------- +Time: 2.517s Load: 0.005s, Pack+Encode: 1.497s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 272.3121 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,132B, BPFP=0.3298 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,412B, BPFP=1.9202 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,592B, BPFP=0.5557 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,048B, BPFP=1.8639 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,020B, BPFP=0.9313 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,780B, BPFP=1.8224 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,668B, BPFP=1.1863 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,960B, BPFP=1.8502 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,480B, BPFP=1.7760 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,412B, BPFP=1.7655 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,080B, BPFP=0.0902 +⌛️ [2/4] FRONTEND: Frontend time: 1.507s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11149573 58.52981416 + layer.0.v_cache 0.00001357 0.01318531 + layer.1.k_cache 0.08092193 6.04745695 + layer.1.v_cache 0.00000571 0.00524533 + layer.2.k_cache 0.01109533 1.24205289 + layer.2.v_cache 0.00001920 0.01644507 + layer.3.k_cache 0.03617269 7.53027404 + layer.3.v_cache 0.00001858 0.01876882 + layer.4.k_cache 0.00062559 0.42104272 + layer.4.v_cache 0.00005684 0.03736912 + layer.4.output 11.30219499 530.16451556 + ------------------------------------------------------------------------------------- + TOTAL 4.66798765 222.64783902 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 94584 +BPFP 0.8607 bits/point +EBPFP 1.7215 equivalent bits/point +MSE 222.647839 +---------------------- -------------------------------------------------------- +Time: 2.529s Load: 0.005s, Pack+Encode: 1.507s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 222.6478 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3606 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,468B, BPFP=2.4976 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,996B, BPFP=0.6002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,136B, BPFP=2.4311 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,616B, BPFP=0.7244 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,440B, BPFP=2.2917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,040B, BPFP=0.6090 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,672B, BPFP=2.3381 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,496B, BPFP=1.7019 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,596B, BPFP=2.1226 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,676B, BPFP=0.1052 +⌛️ [2/4] FRONTEND: Frontend time: 1.499s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10167113 54.00773738 + layer.0.v_cache 0.00001386 0.01349451 + layer.1.k_cache 0.03545215 6.43183312 + layer.1.v_cache 0.00000569 0.00531427 + layer.2.k_cache 0.01047806 1.20629198 + layer.2.v_cache 0.00001755 0.01627706 + layer.3.k_cache 0.06687863 5.33766213 + layer.3.v_cache 0.00001960 0.01942827 + layer.4.k_cache 0.00059181 0.39595325 + layer.4.v_cache 0.00004882 0.03762159 + layer.4.output 0.19633365 691.15710852 + ------------------------------------------------------------------------------------- + TOTAL 0.09350076 288.56302195 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 81936 +BPFP 0.9655 bits/point +EBPFP 1.9310 equivalent bits/point +MSE 288.563022 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.004s, Pack+Encode: 1.499s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 288.5630 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,880B, BPFP=0.3376 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,492B, BPFP=2.2435 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,092B, BPFP=0.5553 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,144B, BPFP=2.1810 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,404B, BPFP=0.9705 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,824B, BPFP=2.1236 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,820B, BPFP=0.6861 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,912B, BPFP=2.1394 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,880B, BPFP=1.9540 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,596B, BPFP=2.0826 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,972B, BPFP=0.1019 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.021s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11445767 56.04177330 + layer.0.v_cache 0.00001583 0.01271492 + layer.1.k_cache 0.03092440 5.74646101 + layer.1.v_cache 0.00000551 0.00509843 + layer.2.k_cache 0.00639992 1.11285856 + layer.2.v_cache 0.00001897 0.01559456 + layer.3.k_cache 0.05613507 5.71180322 + layer.3.v_cache 0.00002026 0.01889040 + layer.4.k_cache 0.00063356 0.40198762 + layer.4.v_cache 0.00005123 0.03490867 + layer.4.output 0.16055210 621.29469417 + ------------------------------------------------------------------------------------- + TOTAL 0.07838395 259.89205588 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 89016 +BPFP 0.9404 bits/point +EBPFP 1.8808 equivalent bits/point +MSE 259.892056 +---------------------- -------------------------------------------------------- +Time: 2.522s Load: 0.003s, Pack+Encode: 1.497s, Decode+Unpack: 1.021s +---------------------- -------------------------------------------------------- +💾 Converting with 259.8921 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3452 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,576B, BPFP=2.3393 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,080B, BPFP=0.5729 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,232B, BPFP=2.2753 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,664B, BPFP=0.8676 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,832B, BPFP=2.2009 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,124B, BPFP=0.5811 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,936B, BPFP=2.2202 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,120B, BPFP=2.0685 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,696B, BPFP=2.1756 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,580B, BPFP=0.0951 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13956541 52.64099702 + layer.0.v_cache 0.00001349 0.01296785 + layer.1.k_cache 0.03336124 5.13885135 + layer.1.v_cache 0.00000558 0.00523528 + layer.2.k_cache 0.00698305 1.13437371 + layer.2.v_cache 0.00001831 0.01508847 + layer.3.k_cache 0.07289400 4.76380194 + layer.3.v_cache 0.00001855 0.01817766 + layer.4.k_cache 0.00061006 0.40198090 + layer.4.v_cache 0.00005429 0.03641862 + layer.4.output 0.17891866 642.52970876 + ------------------------------------------------------------------------------------- + TOTAL 0.08858556 268.34563848 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 87696 +BPFP 0.9596 bits/point +EBPFP 1.9191 equivalent bits/point +MSE 268.345638 +---------------------- -------------------------------------------------------- +Time: 2.503s Load: 0.005s, Pack+Encode: 1.492s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 268.3456 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,864B, BPFP=0.3596 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,484B, BPFP=2.4082 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,000B, BPFP=0.5787 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,100B, BPFP=2.3341 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,500B, BPFP=0.6752 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,420B, BPFP=2.2029 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,944B, BPFP=0.5679 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,800B, BPFP=2.2762 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,556B, BPFP=1.6505 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,300B, BPFP=2.1798 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,964B, BPFP=0.1092 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.021s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10837562 53.40441744 + layer.0.v_cache 0.00001644 0.01269077 + layer.1.k_cache 0.01241452 6.43447537 + layer.1.v_cache 0.00000532 0.00510984 + layer.2.k_cache 0.01724591 0.99530397 + layer.2.v_cache 0.00001958 0.01527380 + layer.3.k_cache 0.04641535 4.67164141 + layer.3.v_cache 0.00001847 0.01754300 + layer.4.k_cache 0.00063678 0.39877541 + layer.4.v_cache 0.00004974 0.03466855 + layer.4.output 0.18014439 667.15222663 + ------------------------------------------------------------------------------------- + TOTAL 0.08507109 278.59149917 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 82932 +BPFP 0.9410 bits/point +EBPFP 1.8821 equivalent bits/point +MSE 278.591499 +---------------------- -------------------------------------------------------- +Time: 2.515s Load: 0.003s, Pack+Encode: 1.490s, Decode+Unpack: 1.021s +---------------------- -------------------------------------------------------- +💾 Converting with 278.5915 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.9319 bits/point +Avg EBPFP 1.8637 equivalent bits/point +Avg MSE 262.378483 +Avg Time 2.518s +------------------------ ---------------------------- diff --git a/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..ce14c3c0a34a2e02ed4511b99da95dedffaa2669 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 255 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean +Output output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean +---------------- ------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,008B, BPFP=0.2860 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,464B, BPFP=1.8168 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,412B, BPFP=0.5288 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,108B, BPFP=1.7200 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,292B, BPFP=0.8770 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,644B, BPFP=1.6869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,044B, BPFP=0.8593 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,096B, BPFP=1.7192 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,796B, BPFP=1.4124 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,960B, BPFP=1.6381 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,884B, BPFP=0.0702 +⌛️ [2/4] FRONTEND: Frontend time: 2.066s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12261864 53.28967341 + layer.0.v_cache 0.00001806 0.01404509 + layer.1.k_cache 0.30198812 6.39565942 + layer.1.v_cache 0.00000607 0.00541308 + layer.2.k_cache 0.01966049 1.30450607 + layer.2.v_cache 0.00002021 0.01607369 + layer.3.k_cache 0.01413035 7.97340651 + layer.3.v_cache 0.00002054 0.01889822 + layer.4.k_cache 0.00067868 0.43836811 + layer.4.v_cache 0.00004987 0.03363149 + layer.4.output 1.39792533 247.16711513 + ------------------------------------------------------------------------------------- + TOTAL 0.60262755 105.86232241 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 182708 +BPFP 0.7668 bits/point +EBPFP 1.5336 equivalent bits/point +MSE 105.862322 +---------------------- -------------------------------------------------------- +Time: 3.346s Load: 0.008s, Pack+Encode: 2.066s, Decode+Unpack: 1.272s +---------------------- -------------------------------------------------------- +💾 Converting with 105.8623 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,020B, BPFP=0.2908 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,212B, BPFP=1.8238 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,236B, BPFP=0.5234 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,360B, BPFP=1.7622 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,936B, BPFP=0.9358 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,800B, BPFP=1.7216 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,932B, BPFP=0.8631 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,244B, BPFP=1.7538 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,116B, BPFP=1.4552 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,124B, BPFP=1.6727 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,732B, BPFP=0.0592 +⌛️ [2/4] FRONTEND: Frontend time: 1.707s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11787021 48.88542571 + layer.0.v_cache 0.00001738 0.01335305 + layer.1.k_cache 0.25265321 6.15088286 + layer.1.v_cache 0.00000586 0.00531589 + layer.2.k_cache 0.01010228 1.31675282 + layer.2.v_cache 0.00002017 0.01613335 + layer.3.k_cache 0.01159736 8.59421511 + layer.3.v_cache 0.00002197 0.02046956 + layer.4.k_cache 0.00067244 0.44797830 + layer.4.v_cache 0.00004845 0.03336524 + layer.4.output 1.41728832 250.50694444 + ------------------------------------------------------------------------------------- + TOTAL 0.60670750 107.00191194 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 182712 +BPFP 0.7775 bits/point +EBPFP 1.5549 equivalent bits/point +MSE 107.001912 +---------------------- -------------------------------------------------------- +Time: 2.935s Load: 0.008s, Pack+Encode: 1.707s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 107.0019 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,980B, BPFP=0.2789 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,952B, BPFP=1.7483 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,316B, BPFP=0.5126 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,980B, BPFP=1.6802 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,372B, BPFP=0.7968 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,308B, BPFP=1.6331 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,876B, BPFP=0.7621 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,908B, BPFP=1.6752 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,236B, BPFP=1.4179 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,700B, BPFP=1.5905 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,364B, BPFP=0.0537 +⌛️ [2/4] FRONTEND: Frontend time: 1.717s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13311632 55.67035980 + layer.0.v_cache 0.00001526 0.01289922 + layer.1.k_cache 0.36474110 6.34350422 + layer.1.v_cache 0.00000573 0.00513575 + layer.2.k_cache 0.01296863 1.33559236 + layer.2.v_cache 0.00001984 0.01588200 + layer.3.k_cache 0.01490937 8.74239770 + layer.3.v_cache 0.00002017 0.01969160 + layer.4.k_cache 0.00067370 0.45764232 + layer.4.v_cache 0.00005228 0.03348945 + layer.4.output 1.37278975 242.68597854 + ------------------------------------------------------------------------------------- + TOTAL 0.59623828 104.20226142 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 177992 +BPFP 0.7336 bits/point +EBPFP 1.4672 equivalent bits/point +MSE 104.202261 +---------------------- -------------------------------------------------------- +Time: 2.946s Load: 0.008s, Pack+Encode: 1.717s, Decode+Unpack: 1.222s +---------------------- -------------------------------------------------------- +💾 Converting with 104.2023 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,320B, BPFP=0.2766 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,400B, BPFP=1.6265 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,292B, BPFP=0.5310 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,516B, BPFP=1.5699 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,856B, BPFP=0.7592 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,772B, BPFP=1.5223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,376B, BPFP=0.7285 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,360B, BPFP=1.5599 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,588B, BPFP=1.3184 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,080B, BPFP=1.4780 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,708B, BPFP=0.0614 +⌛️ [2/4] FRONTEND: Frontend time: 1.721s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.228s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11263794 52.94849033 + layer.0.v_cache 0.00001639 0.01346779 + layer.1.k_cache 0.42101050 6.02159244 + layer.1.v_cache 0.00000577 0.00526593 + layer.2.k_cache 0.00499437 1.25367262 + layer.2.v_cache 0.00001973 0.01597622 + layer.3.k_cache 0.03029772 9.00148085 + layer.3.v_cache 0.00002073 0.01976958 + layer.4.k_cache 0.00068836 0.43247908 + layer.4.v_cache 0.00005112 0.03357326 + layer.4.output 1.25471843 221.72775176 + ------------------------------------------------------------------------------------- + TOTAL 0.55016304 95.40235473 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 184268 +BPFP 0.6941 bits/point +EBPFP 1.3882 equivalent bits/point +MSE 95.402355 +---------------------- -------------------------------------------------------- +Time: 2.957s Load: 0.008s, Pack+Encode: 1.721s, Decode+Unpack: 1.228s +---------------------- -------------------------------------------------------- +💾 Converting with 95.4024 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,072B, BPFP=0.2919 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,512B, BPFP=1.8286 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,456B, BPFP=0.5344 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,416B, BPFP=1.7500 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,460B, BPFP=0.8931 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,748B, BPFP=1.7021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,808B, BPFP=0.7747 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,244B, BPFP=1.7377 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,520B, BPFP=1.3991 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,224B, BPFP=1.6646 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,156B, BPFP=0.0630 +⌛️ [2/4] FRONTEND: Frontend time: 1.713s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.226s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10233516 52.36398007 + layer.0.v_cache 0.00001822 0.01387616 + layer.1.k_cache 0.31275415 6.06899324 + layer.1.v_cache 0.00000587 0.00531203 + layer.2.k_cache 0.01225604 1.40720220 + layer.2.v_cache 0.00002018 0.01635956 + layer.3.k_cache 0.02847941 7.58020915 + layer.3.v_cache 0.00002101 0.01987023 + layer.4.k_cache 0.00067306 0.45989850 + layer.4.v_cache 0.00005947 0.03468213 + layer.4.output 1.40430263 248.08901950 + ------------------------------------------------------------------------------------- + TOTAL 0.60510241 106.15255999 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 181616 +BPFP 0.7657 bits/point +EBPFP 1.5314 equivalent bits/point +MSE 106.152560 +---------------------- -------------------------------------------------------- +Time: 2.949s Load: 0.010s, Pack+Encode: 1.713s, Decode+Unpack: 1.226s +---------------------- -------------------------------------------------------- +💾 Converting with 106.1526 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,212B, BPFP=0.2770 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,908B, BPFP=1.6958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,932B, BPFP=0.5278 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,360B, BPFP=1.6135 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,548B, BPFP=0.7732 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,280B, BPFP=1.5561 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,536B, BPFP=0.7725 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,268B, BPFP=1.6086 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,296B, BPFP=1.3444 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,368B, BPFP=1.5077 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,100B, BPFP=0.0615 +⌛️ [2/4] FRONTEND: Frontend time: 2.078s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.409s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16298727 53.67255926 + layer.0.v_cache 0.00001770 0.01333197 + layer.1.k_cache 0.61736183 6.33945678 + layer.1.v_cache 0.00000612 0.00538738 + layer.2.k_cache 0.02593859 1.28943452 + layer.2.v_cache 0.00002024 0.01579081 + layer.3.k_cache 0.01326758 9.05847749 + layer.3.v_cache 0.00002270 0.01987458 + layer.4.k_cache 0.00075427 0.45103190 + layer.4.v_cache 0.00005043 0.03334381 + layer.4.output 0.04538776 184.22309281 + ------------------------------------------------------------------------------------- + TOTAL 0.06694947 80.02707872 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 227808 +BPFP 0.7122 bits/point +EBPFP 1.4244 equivalent bits/point +MSE 80.027079 +---------------------- -------------------------------------------------------- +Time: 3.498s Load: 0.011s, Pack+Encode: 2.078s, Decode+Unpack: 1.409s +---------------------- -------------------------------------------------------- +💾 Converting with 80.0271 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 239, 128) +Output shape: (1, 239, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.output: torch.Size([1, 239, 3584]) -> torch.Size([1, 1, 239, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,312B, BPFP=0.2819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,700B, BPFP=1.6802 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,056B, BPFP=0.5267 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,600B, BPFP=1.6083 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,300B, BPFP=0.8041 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,756B, BPFP=1.5531 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,764B, BPFP=0.7691 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,476B, BPFP=1.6002 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,456B, BPFP=1.3373 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,992B, BPFP=1.5031 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,728B, BPFP=0.0722 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.229s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11863919 53.68906087 + layer.0.v_cache 0.00001702 0.01372115 + layer.1.k_cache 0.43747612 6.74944481 + layer.1.v_cache 0.00000797 0.00570112 + layer.2.k_cache 0.02286501 1.25442684 + layer.2.v_cache 0.00002174 0.01568689 + layer.3.k_cache 0.01903599 8.29471298 + layer.3.v_cache 0.00002123 0.01918222 + layer.4.k_cache 0.00068597 0.45080359 + layer.4.v_cache 0.00006921 0.03240107 + layer.4.output 1.28101462 226.20576808 + ------------------------------------------------------------------------------------- + TOTAL 0.56270246 97.29208930 + (elements=2,080,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2080256 +Total Bytes 186140 +BPFP 0.7158 bits/point +EBPFP 1.4317 equivalent bits/point +MSE 97.292089 +---------------------- -------------------------------------------------------- +Time: 2.967s Load: 0.008s, Pack+Encode: 1.730s, Decode+Unpack: 1.229s +---------------------- -------------------------------------------------------- +💾 Converting with 97.2921 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,748B, BPFP=0.2821 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,016B, BPFP=1.8427 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,020B, BPFP=0.5359 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,044B, BPFP=1.7849 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,124B, BPFP=0.8985 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,744B, BPFP=1.7077 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,924B, BPFP=0.8272 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,700B, BPFP=1.7645 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,608B, BPFP=1.4620 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,352B, BPFP=1.6844 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,208B, BPFP=0.0697 +⌛️ [2/4] FRONTEND: Frontend time: 1.836s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15151563 54.10004753 + layer.0.v_cache 0.00001631 0.01376325 + layer.1.k_cache 0.46755480 6.06946590 + layer.1.v_cache 0.00000667 0.00550974 + layer.2.k_cache 0.02171374 1.41553210 + layer.2.v_cache 0.00002115 0.01606919 + layer.3.k_cache 0.01385514 9.19107949 + layer.3.v_cache 0.00002107 0.01974891 + layer.4.k_cache 0.00069081 0.44129599 + layer.4.v_cache 0.00005638 0.03445464 + layer.4.output 0.00504555 211.03773425 + ------------------------------------------------------------------------------------- + TOTAL 0.04063356 91.09241803 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 223488 +BPFP 0.7810 bits/point +EBPFP 1.5621 equivalent bits/point +MSE 91.092418 +---------------------- -------------------------------------------------------- +Time: 3.202s Load: 0.010s, Pack+Encode: 1.836s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 91.0924 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,996B, BPFP=0.2775 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,224B, BPFP=1.7517 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,384B, BPFP=0.5128 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,312B, BPFP=1.6883 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,080B, BPFP=0.8389 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,484B, BPFP=1.6308 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,860B, BPFP=0.8236 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,040B, BPFP=1.6694 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,916B, BPFP=1.3831 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,860B, BPFP=1.5875 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,860B, BPFP=0.0681 +⌛️ [2/4] FRONTEND: Frontend time: 1.732s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.229s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14529029 56.30499566 + layer.0.v_cache 0.00001764 0.01340892 + layer.1.k_cache 0.31089128 6.48249891 + layer.1.v_cache 0.00000630 0.00546434 + layer.2.k_cache 0.01355181 1.25892727 + layer.2.v_cache 0.00002032 0.01587494 + layer.3.k_cache 0.02426839 9.27373155 + layer.3.v_cache 0.00002043 0.01901722 + layer.4.k_cache 0.00068925 0.45036879 + layer.4.v_cache 0.00005165 0.03284784 + layer.4.output 1.36066019 240.55250000 + ------------------------------------------------------------------------------------- + TOTAL 0.58937816 103.39556679 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 182016 +BPFP 0.7435 bits/point +EBPFP 1.4871 equivalent bits/point +MSE 103.395567 +---------------------- -------------------------------------------------------- +Time: 2.969s Load: 0.008s, Pack+Encode: 1.732s, Decode+Unpack: 1.229s +---------------------- -------------------------------------------------------- +💾 Converting with 103.3956 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 262, 128) +Output shape: (1, 262, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.output: torch.Size([1, 262, 3584]) -> torch.Size([1, 1, 262, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,796B, BPFP=0.2860 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,360B, BPFP=1.8702 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,932B, BPFP=0.5327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,896B, BPFP=1.7829 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,692B, BPFP=0.8762 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,924B, BPFP=1.7250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,072B, BPFP=0.8989 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,684B, BPFP=1.7703 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,152B, BPFP=1.4404 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,276B, BPFP=1.6863 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,944B, BPFP=0.0677 +⌛️ [2/4] FRONTEND: Frontend time: 1.844s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12917162 51.30288347 + layer.0.v_cache 0.00001830 0.01348944 + layer.1.k_cache 0.40634074 6.54275536 + layer.1.v_cache 0.00000633 0.00539989 + layer.2.k_cache 0.02363680 1.17418280 + layer.2.v_cache 0.00002248 0.01607071 + layer.3.k_cache 0.02343226 9.54686475 + layer.3.v_cache 0.00002042 0.01941602 + layer.4.k_cache 0.00071170 0.45009604 + layer.4.v_cache 0.00005093 0.03325200 + layer.4.output 0.00505940 211.77131611 + ------------------------------------------------------------------------------------- + TOTAL 0.03640161 91.26491902 + (elements=2,280,448) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2280448 +Total Bytes 223728 +BPFP 0.7849 bits/point +EBPFP 1.5697 equivalent bits/point +MSE 91.264919 +---------------------- -------------------------------------------------------- +Time: 3.206s Load: 0.009s, Pack+Encode: 1.844s, Decode+Unpack: 1.353s +---------------------- -------------------------------------------------------- +💾 Converting with 91.2649 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,008B, BPFP=0.2783 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,396B, BPFP=1.7636 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,608B, BPFP=0.5283 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,380B, BPFP=1.6931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,524B, BPFP=0.8697 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,516B, BPFP=1.6331 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,544B, BPFP=0.8017 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,272B, BPFP=1.6856 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,256B, BPFP=1.4067 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,896B, BPFP=1.5900 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,976B, BPFP=0.0692 +⌛️ [2/4] FRONTEND: Frontend time: 1.717s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.230s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18074451 52.58641493 + layer.0.v_cache 0.00001764 0.01357071 + layer.1.k_cache 0.33421082 6.50509332 + layer.1.v_cache 0.00000655 0.00544796 + layer.2.k_cache 0.01487562 1.19390693 + layer.2.v_cache 0.00002082 0.01630205 + layer.3.k_cache 0.01319740 9.04590441 + layer.3.v_cache 0.00002096 0.01958313 + layer.4.k_cache 0.00067848 0.45243446 + layer.4.v_cache 0.00005239 0.03429037 + layer.4.output 1.36066546 240.44057540 + ------------------------------------------------------------------------------------- + TOTAL 0.59226373 103.11511624 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 183376 +BPFP 0.7491 bits/point +EBPFP 1.4982 equivalent bits/point +MSE 103.115116 +---------------------- -------------------------------------------------------- +Time: 2.956s Load: 0.009s, Pack+Encode: 1.717s, Decode+Unpack: 1.230s +---------------------- -------------------------------------------------------- +💾 Converting with 103.1151 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 232, 128) +Output shape: (1, 232, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.output: torch.Size([1, 232, 3584]) -> torch.Size([1, 1, 232, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,208B, BPFP=0.2834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,384B, BPFP=1.7096 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,848B, BPFP=0.5286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,360B, BPFP=1.6406 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,548B, BPFP=0.8451 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,700B, BPFP=1.5962 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,144B, BPFP=0.8179 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,232B, BPFP=1.6320 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,668B, BPFP=1.3920 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,016B, BPFP=1.5501 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,736B, BPFP=0.0744 +⌛️ [2/4] FRONTEND: Frontend time: 1.727s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11978169 53.20831930 + layer.0.v_cache 0.00001746 0.01357607 + layer.1.k_cache 0.32299256 6.31467938 + layer.1.v_cache 0.00000596 0.00519525 + layer.2.k_cache 0.02091339 1.16839363 + layer.2.v_cache 0.00002092 0.01635064 + layer.3.k_cache 0.01769928 9.37125791 + layer.3.v_cache 0.00002180 0.01991281 + layer.4.k_cache 0.00071208 0.45020748 + layer.4.v_cache 0.00005763 0.03446176 + layer.4.output 1.31963615 233.45056573 + ------------------------------------------------------------------------------------- + TOTAL 0.57174564 100.27978320 + (elements=2,019,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2019328 +Total Bytes 185844 +BPFP 0.7363 bits/point +EBPFP 1.4725 equivalent bits/point +MSE 100.279783 +---------------------- -------------------------------------------------------- +Time: 2.953s Load: 0.008s, Pack+Encode: 1.727s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 100.2798 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,968B, BPFP=0.2818 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,332B, BPFP=1.7991 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,388B, BPFP=0.5247 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,536B, BPFP=1.7426 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,796B, BPFP=0.9798 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,604B, BPFP=1.6764 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,364B, BPFP=0.8781 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,240B, BPFP=1.7216 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,756B, BPFP=1.4031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,140B, BPFP=1.6435 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,888B, BPFP=0.0800 +⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.216s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15639258 54.66369851 + layer.0.v_cache 0.00001700 0.01420497 + layer.1.k_cache 0.33150226 6.37005338 + layer.1.v_cache 0.00000613 0.00551168 + layer.2.k_cache 0.01694729 1.21307082 + layer.2.v_cache 0.00002208 0.01589540 + layer.3.k_cache 0.03858858 8.18521507 + layer.3.v_cache 0.00002090 0.01990459 + layer.4.k_cache 0.00067984 0.44575098 + layer.4.v_cache 0.00005217 0.03440921 + layer.4.output 1.39161012 246.04748377 + ------------------------------------------------------------------------------------- + TOTAL 0.60502939 105.48824123 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 186012 +BPFP 0.7771 bits/point +EBPFP 1.5542 equivalent bits/point +MSE 105.488241 +---------------------- -------------------------------------------------------- +Time: 2.938s Load: 0.007s, Pack+Encode: 1.714s, Decode+Unpack: 1.216s +---------------------- -------------------------------------------------------- +💾 Converting with 105.4882 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,904B, BPFP=0.2807 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,624B, BPFP=1.8100 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,156B, BPFP=0.5240 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,416B, BPFP=1.7408 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,184B, BPFP=0.8690 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,448B, BPFP=1.6854 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,144B, BPFP=0.8095 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,976B, BPFP=1.7157 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,448B, BPFP=1.3993 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,716B, BPFP=1.6435 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,428B, BPFP=0.0771 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13605902 54.33821972 + layer.0.v_cache 0.00001757 0.01326252 + layer.1.k_cache 0.58683352 6.13396881 + layer.1.v_cache 0.00000623 0.00544694 + layer.2.k_cache 0.01500385 1.17257702 + layer.2.v_cache 0.00002229 0.01605629 + layer.3.k_cache 0.00978726 8.82000308 + layer.3.v_cache 0.00001998 0.01822248 + layer.4.k_cache 0.00067570 0.43417439 + layer.4.v_cache 0.00005472 0.03285471 + layer.4.output 0.00490606 203.03401361 + ------------------------------------------------------------------------------------- + TOTAL 0.04604839 87.77781654 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 227444 +BPFP 0.7657 bits/point +EBPFP 1.5315 equivalent bits/point +MSE 87.777817 +---------------------- -------------------------------------------------------- +Time: 3.198s Load: 0.011s, Pack+Encode: 1.840s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 87.7778 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,364B, BPFP=0.2674 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,484B, BPFP=1.5615 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,324B, BPFP=0.5100 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,452B, BPFP=1.4983 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,208B, BPFP=0.6868 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,512B, BPFP=1.4407 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,464B, BPFP=0.6412 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,136B, BPFP=1.4789 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,508B, BPFP=1.2566 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,692B, BPFP=1.3904 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,128B, BPFP=0.0624 +⌛️ [2/4] FRONTEND: Frontend time: 1.711s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12967140 54.74704733 + layer.0.v_cache 0.00001601 0.01307664 + layer.1.k_cache 0.49353925 6.42763959 + layer.1.v_cache 0.00000620 0.00521144 + layer.2.k_cache 0.03189284 1.36197498 + layer.2.v_cache 0.00002040 0.01531150 + layer.3.k_cache 0.04285583 9.30194164 + layer.3.v_cache 0.00002221 0.01872846 + layer.4.k_cache 0.00071132 0.44523106 + layer.4.v_cache 0.00005107 0.03189291 + layer.4.output 1.20063613 211.72081583 + ------------------------------------------------------------------------------------- + TOTAL 0.53548467 91.43610390 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 182272 +BPFP 0.6570 bits/point +EBPFP 1.3140 equivalent bits/point +MSE 91.436104 +---------------------- -------------------------------------------------------- +Time: 2.939s Load: 0.010s, Pack+Encode: 1.711s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,420B, BPFP=0.2774 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,684B, BPFP=1.6117 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,424B, BPFP=0.5286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,588B, BPFP=1.5429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,180B, BPFP=0.7643 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,620B, BPFP=1.4822 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,320B, BPFP=0.7103 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,396B, BPFP=1.5309 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,800B, BPFP=1.3052 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,920B, BPFP=1.4383 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,528B, BPFP=0.0585 +⌛️ [2/4] FRONTEND: Frontend time: 1.718s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16529650 52.34111054 + layer.0.v_cache 0.00001613 0.01375962 + layer.1.k_cache 0.54577502 6.06130430 + layer.1.v_cache 0.00000662 0.00565307 + layer.2.k_cache 0.01453520 1.17675475 + layer.2.v_cache 0.00002096 0.01679070 + layer.3.k_cache 0.02409829 7.93433892 + layer.3.v_cache 0.00002071 0.01988772 + layer.4.k_cache 0.00068699 0.44074476 + layer.4.v_cache 0.00004971 0.03334399 + layer.4.output 1.22953407 217.36637622 + ------------------------------------------------------------------------------------- + TOTAL 0.55042615 93.50637188 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 184880 +BPFP 0.6824 bits/point +EBPFP 1.3649 equivalent bits/point +MSE 93.506372 +---------------------- -------------------------------------------------------- +Time: 2.940s Load: 0.010s, Pack+Encode: 1.718s, Decode+Unpack: 1.212s +---------------------- -------------------------------------------------------- +💾 Converting with 93.5064 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,800B, BPFP=0.2771 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,356B, BPFP=1.7850 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,904B, BPFP=0.5210 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,168B, BPFP=1.7282 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,016B, BPFP=0.8609 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,896B, BPFP=1.6674 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,496B, BPFP=0.7882 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,708B, BPFP=1.7062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,180B, BPFP=1.3943 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,888B, BPFP=1.6193 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,628B, BPFP=0.0657 +⌛️ [2/4] FRONTEND: Frontend time: 2.178s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19043710 54.19767775 + layer.0.v_cache 0.00001603 0.01345195 + layer.1.k_cache 0.64205214 6.31670863 + layer.1.v_cache 0.00000630 0.00553056 + layer.2.k_cache 0.01513053 1.30053188 + layer.2.v_cache 0.00002191 0.01637056 + layer.3.k_cache 0.02775778 9.24161859 + layer.3.v_cache 0.00002125 0.01927655 + layer.4.k_cache 0.00071623 0.44586350 + layer.4.v_cache 0.00005146 0.03232982 + layer.4.output 0.04089398 165.87546418 + ------------------------------------------------------------------------------------- + TOTAL 0.06838051 72.51280053 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 268040 +BPFP 0.7534 bits/point +EBPFP 1.5068 equivalent bits/point +MSE 72.512801 +---------------------- -------------------------------------------------------- +Time: 3.660s Load: 0.012s, Pack+Encode: 2.178s, Decode+Unpack: 1.470s +---------------------- -------------------------------------------------------- +💾 Converting with 72.5128 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,024B, BPFP=0.2774 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,592B, BPFP=1.7443 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,528B, BPFP=0.5261 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,324B, BPFP=1.6742 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,336B, BPFP=0.8467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,384B, BPFP=1.6223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,148B, BPFP=0.7811 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,048B, BPFP=1.6590 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,336B, BPFP=1.3436 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,484B, BPFP=1.5727 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,144B, BPFP=0.0642 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14392723 54.85067221 + layer.0.v_cache 0.00001682 0.01359266 + layer.1.k_cache 0.53349789 6.51522989 + layer.1.v_cache 0.00000638 0.00557957 + layer.2.k_cache 0.01523476 1.31253149 + layer.2.v_cache 0.00002109 0.01589070 + layer.3.k_cache 0.03867486 9.02789932 + layer.3.v_cache 0.00001994 0.01892143 + layer.4.k_cache 0.00070525 0.44241994 + layer.4.v_cache 0.00005078 0.03285647 + layer.4.output 0.00472849 196.12214475 + ------------------------------------------------------------------------------------- + TOTAL 0.04501497 85.00532982 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 226348 +BPFP 0.7351 bits/point +EBPFP 1.4703 equivalent bits/point +MSE 85.005330 +---------------------- -------------------------------------------------------- +Time: 3.184s Load: 0.011s, Pack+Encode: 1.830s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 85.0053 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,996B, BPFP=0.2775 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,340B, BPFP=1.7597 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,388B, BPFP=0.5131 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,256B, BPFP=1.6844 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,016B, BPFP=0.8344 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,376B, BPFP=1.6233 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,144B, BPFP=0.7739 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,152B, BPFP=1.6772 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,096B, BPFP=1.3956 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,732B, BPFP=1.5786 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,156B, BPFP=0.0611 +⌛️ [2/4] FRONTEND: Frontend time: 1.707s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14673459 53.94287760 + layer.0.v_cache 0.00001655 0.01346139 + layer.1.k_cache 0.37581726 6.14107042 + layer.1.v_cache 0.00000603 0.00530095 + layer.2.k_cache 0.01062915 1.24886217 + layer.2.v_cache 0.00002006 0.01640850 + layer.3.k_cache 0.04027582 8.37632270 + layer.3.v_cache 0.00002066 0.01937309 + layer.4.k_cache 0.00066373 0.43947001 + layer.4.v_cache 0.00004861 0.03331306 + layer.4.output 1.36064577 240.48873016 + ------------------------------------------------------------------------------------- + TOTAL 0.59404429 103.15632771 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 180652 +BPFP 0.7380 bits/point +EBPFP 1.4759 equivalent bits/point +MSE 103.156328 +---------------------- -------------------------------------------------------- +Time: 2.926s Load: 0.007s, Pack+Encode: 1.707s, Decode+Unpack: 1.212s +---------------------- -------------------------------------------------------- +💾 Converting with 103.1563 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,228B, BPFP=0.2823 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,884B, BPFP=1.7284 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,896B, BPFP=0.5272 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,760B, BPFP=1.6533 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,208B, BPFP=0.8152 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,812B, BPFP=1.5900 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,600B, BPFP=0.7746 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,716B, BPFP=1.6504 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,808B, BPFP=1.3894 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,092B, BPFP=1.5419 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,092B, BPFP=0.0677 +⌛️ [2/4] FRONTEND: Frontend time: 1.712s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12385076 54.36025975 + layer.0.v_cache 0.00002004 0.01446201 + layer.1.k_cache 0.38583452 6.28727944 + layer.1.v_cache 0.00000646 0.00567710 + layer.2.k_cache 0.02500911 1.26012636 + layer.2.v_cache 0.00002014 0.01588727 + layer.3.k_cache 0.03425702 8.76822082 + layer.3.v_cache 0.00002057 0.01979753 + layer.4.k_cache 0.00068928 0.45260307 + layer.4.v_cache 0.00004882 0.03316729 + layer.4.output 1.30836928 231.04073184 + ------------------------------------------------------------------------------------- + TOTAL 0.57225539 99.32368256 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 186096 +BPFP 0.7310 bits/point +EBPFP 1.4619 equivalent bits/point +MSE 99.323683 +---------------------- -------------------------------------------------------- +Time: 2.940s Load: 0.008s, Pack+Encode: 1.712s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 99.3237 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 247, 128) +Output shape: (1, 247, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.output: torch.Size([1, 247, 3584]) -> torch.Size([1, 1, 247, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,400B, BPFP=0.2783 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,696B, BPFP=1.6255 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,184B, BPFP=0.5177 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,408B, BPFP=1.5440 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,728B, BPFP=0.7419 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,468B, BPFP=1.4846 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,252B, BPFP=0.7118 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,104B, BPFP=1.5248 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,412B, BPFP=1.2912 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,760B, BPFP=1.4398 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,304B, BPFP=0.0660 +⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13907220 51.92649291 + layer.0.v_cache 0.00001735 0.01368488 + layer.1.k_cache 0.47208408 6.52387735 + layer.1.v_cache 0.00000611 0.00537296 + layer.2.k_cache 0.03408500 1.23525272 + layer.2.v_cache 0.00001989 0.01570406 + layer.3.k_cache 0.02387486 8.32398845 + layer.3.v_cache 0.00002031 0.01906197 + layer.4.k_cache 0.00069623 0.42893173 + layer.4.v_cache 0.00004922 0.03266222 + layer.4.output 1.23951819 219.17376374 + ------------------------------------------------------------------------------------- + TOTAL 0.54979721 94.27890444 + (elements=2,149,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2149888 +Total Bytes 183716 +BPFP 0.6836 bits/point +EBPFP 1.3673 equivalent bits/point +MSE 94.278904 +---------------------- -------------------------------------------------------- +Time: 2.935s Load: 0.008s, Pack+Encode: 1.714s, Decode+Unpack: 1.213s +---------------------- -------------------------------------------------------- +💾 Converting with 94.2789 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 427, 128) +Output shape: (1, 427, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.output: torch.Size([1, 427, 3584]) -> torch.Size([1, 1, 427, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,412B, BPFP=0.2712 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,304B, BPFP=1.6212 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,172B, BPFP=0.5186 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,416B, BPFP=1.5521 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,636B, BPFP=0.7185 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,172B, BPFP=1.4700 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,680B, BPFP=0.7201 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,256B, BPFP=1.5097 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,400B, BPFP=1.2588 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,260B, BPFP=1.4366 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,708B, BPFP=0.0717 +⌛️ [2/4] FRONTEND: Frontend time: 2.300s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.567s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15738957 52.00024243 + layer.0.v_cache 0.00001693 0.01264262 + layer.1.k_cache 1.00627419 6.36805829 + layer.1.v_cache 0.00000663 0.00529196 + layer.2.k_cache 0.03590691 1.26796935 + layer.2.v_cache 0.00002108 0.01484788 + layer.3.k_cache 0.02176077 8.87919899 + layer.3.v_cache 0.00002004 0.01730593 + layer.4.k_cache 0.00078997 0.44782384 + layer.4.v_cache 0.00005212 0.03097423 + layer.4.output 0.00621442 129.79074314 + ------------------------------------------------------------------------------------- + TOTAL 0.07445524 57.50467985 + (elements=3,716,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3716608 +Total Bytes 316416 +BPFP 0.6811 bits/point +EBPFP 1.3622 equivalent bits/point +MSE 57.504680 +---------------------- -------------------------------------------------------- +Time: 3.883s Load: 0.016s, Pack+Encode: 2.300s, Decode+Unpack: 1.567s +---------------------- -------------------------------------------------------- +💾 Converting with 57.5047 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,936B, BPFP=0.2846 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,276B, BPFP=1.8033 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,084B, BPFP=0.5238 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,012B, BPFP=1.7304 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,008B, BPFP=0.8653 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,880B, BPFP=1.6651 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,800B, BPFP=0.8533 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,628B, BPFP=1.7083 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,740B, BPFP=1.3688 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,092B, BPFP=1.6197 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,428B, BPFP=0.0694 +⌛️ [2/4] FRONTEND: Frontend time: 1.837s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.337s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15504150 51.48157504 + layer.0.v_cache 0.00001776 0.01295094 + layer.1.k_cache 0.59434689 6.29242777 + layer.1.v_cache 0.00000612 0.00516637 + layer.2.k_cache 0.01687181 1.26237347 + layer.2.v_cache 0.00002031 0.01533385 + layer.3.k_cache 0.01346795 8.41010670 + layer.3.v_cache 0.00002023 0.01859188 + layer.4.k_cache 0.00071492 0.42572596 + layer.4.v_cache 0.00006004 0.03168372 + layer.4.output 0.00488892 204.49481089 + ------------------------------------------------------------------------------------- + TOTAL 0.04792882 88.20115364 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 223884 +BPFP 0.7593 bits/point +EBPFP 1.5186 equivalent bits/point +MSE 88.201154 +---------------------- -------------------------------------------------------- +Time: 3.184s Load: 0.010s, Pack+Encode: 1.837s, Decode+Unpack: 1.337s +---------------------- -------------------------------------------------------- +💾 Converting with 88.2012 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 361, 128) +Output shape: (1, 361, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.output: torch.Size([1, 361, 3584]) -> torch.Size([1, 1, 361, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,288B, BPFP=0.2722 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,336B, BPFP=1.6160 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,844B, BPFP=0.5126 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,668B, BPFP=1.5438 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,024B, BPFP=0.7801 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,632B, BPFP=1.4990 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,400B, BPFP=0.7531 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,096B, BPFP=1.5190 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,484B, BPFP=1.2761 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,352B, BPFP=1.4436 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,544B, BPFP=0.0590 +⌛️ [2/4] FRONTEND: Frontend time: 1.963s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.461s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21057716 53.13468988 + layer.0.v_cache 0.00001554 0.01206299 + layer.1.k_cache 0.82271198 6.05886832 + layer.1.v_cache 0.00000592 0.00491177 + layer.2.k_cache 0.01184860 1.24312450 + layer.2.v_cache 0.00001987 0.01552433 + layer.3.k_cache 0.01345186 8.53183702 + layer.3.v_cache 0.00002111 0.01777959 + layer.4.k_cache 0.00076976 0.43698162 + layer.4.v_cache 0.00005077 0.03160430 + layer.4.output 0.03703259 149.90877275 + ------------------------------------------------------------------------------------- + TOTAL 0.07757063 65.81463492 + (elements=3,142,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3142144 +Total Bytes 268668 +BPFP 0.6840 bits/point +EBPFP 1.3681 equivalent bits/point +MSE 65.814635 +---------------------- -------------------------------------------------------- +Time: 3.438s Load: 0.014s, Pack+Encode: 1.963s, Decode+Unpack: 1.461s +---------------------- -------------------------------------------------------- +💾 Converting with 65.8146 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,684B, BPFP=0.3149 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,808B, BPFP=1.7070 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,624B, BPFP=0.5332 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,608B, BPFP=1.6405 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,884B, BPFP=0.8247 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,656B, BPFP=1.5878 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,816B, BPFP=0.7655 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,176B, BPFP=1.6166 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,956B, BPFP=1.3273 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,972B, BPFP=1.5499 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,420B, BPFP=0.0666 +⌛️ [2/4] FRONTEND: Frontend time: 1.826s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.335s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21757852 54.91588403 + layer.0.v_cache 0.00001544 0.01173678 + layer.1.k_cache 0.59705023 6.27845017 + layer.1.v_cache 0.00000603 0.00496873 + layer.2.k_cache 0.00787334 1.26058159 + layer.2.v_cache 0.00001970 0.01503328 + layer.3.k_cache 0.01914395 9.09120730 + layer.3.v_cache 0.00001916 0.01724990 + layer.4.k_cache 0.00071861 0.42040891 + layer.4.v_cache 0.00005003 0.03114229 + layer.4.output 0.00471288 196.57796669 + ------------------------------------------------------------------------------------- + TOTAL 0.05149795 85.18190764 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 222604 +BPFP 0.7255 bits/point +EBPFP 1.4511 equivalent bits/point +MSE 85.181908 +---------------------- -------------------------------------------------------- +Time: 3.173s Load: 0.011s, Pack+Encode: 1.826s, Decode+Unpack: 1.335s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1819 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 372, 128) +Output shape: (1, 372, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.output: torch.Size([1, 372, 3584]) -> torch.Size([1, 1, 372, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,432B, BPFP=0.2702 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,924B, BPFP=1.5929 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,196B, BPFP=0.5123 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,288B, BPFP=1.5242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,736B, BPFP=0.7450 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,980B, BPFP=1.4693 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,572B, BPFP=0.6961 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,760B, BPFP=1.5020 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,260B, BPFP=1.2290 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,824B, BPFP=1.4207 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,824B, BPFP=0.0829 +⌛️ [2/4] FRONTEND: Frontend time: 1.971s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.473s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18192377 52.57926432 + layer.0.v_cache 0.00001786 0.01266800 + layer.1.k_cache 0.82695811 6.28042242 + layer.1.v_cache 0.00000687 0.00526908 + layer.2.k_cache 0.01216373 1.08906161 + layer.2.v_cache 0.00002073 0.01534427 + layer.3.k_cache 0.01743993 9.34400070 + layer.3.v_cache 0.00002103 0.01783151 + layer.4.k_cache 0.00069676 0.41496306 + layer.4.v_cache 0.00005544 0.03114012 + layer.4.output 0.03606073 145.64564132 + ------------------------------------------------------------------------------------- + TOTAL 0.07598408 64.07702673 + (elements=3,237,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3237888 +Total Bytes 274796 +BPFP 0.6790 bits/point +EBPFP 1.3579 equivalent bits/point +MSE 64.077027 +---------------------- -------------------------------------------------------- +Time: 3.455s Load: 0.012s, Pack+Encode: 1.971s, Decode+Unpack: 1.473s +---------------------- -------------------------------------------------------- +💾 Converting with 64.0770 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,740B, BPFP=0.2777 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,908B, BPFP=1.7854 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,752B, BPFP=0.5201 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,500B, BPFP=1.7173 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,008B, BPFP=0.7260 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,944B, BPFP=1.6420 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,164B, BPFP=0.7819 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,900B, BPFP=1.6883 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,932B, BPFP=1.3028 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,160B, BPFP=1.6041 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,348B, BPFP=0.0784 +⌛️ [2/4] FRONTEND: Frontend time: 1.958s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.474s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16470559 53.11643769 + layer.0.v_cache 0.00001558 0.01278769 + layer.1.k_cache 0.70577214 6.31182313 + layer.1.v_cache 0.00000663 0.00519268 + layer.2.k_cache 0.01487649 1.15617876 + layer.2.v_cache 0.00002061 0.01543698 + layer.3.k_cache 0.03012228 8.35877939 + layer.3.v_cache 0.00002030 0.01817369 + layer.4.k_cache 0.00074575 0.43565638 + layer.4.v_cache 0.00005065 0.03116740 + layer.4.output 0.04139163 167.86938854 + ------------------------------------------------------------------------------------- + TOTAL 0.07094573 73.20866786 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 260356 +BPFP 0.7409 bits/point +EBPFP 1.4817 equivalent bits/point +MSE 73.208668 +---------------------- -------------------------------------------------------- +Time: 3.444s Load: 0.012s, Pack+Encode: 1.958s, Decode+Unpack: 1.474s +---------------------- -------------------------------------------------------- +💾 Converting with 73.2087 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,780B, BPFP=0.2818 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,776B, BPFP=1.8146 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,988B, BPFP=0.5300 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,668B, BPFP=1.7493 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,836B, BPFP=0.8748 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,856B, BPFP=1.7014 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,012B, BPFP=0.8851 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,428B, BPFP=1.7351 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,396B, BPFP=1.3795 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,136B, BPFP=1.6590 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,624B, BPFP=0.0726 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16585089 52.41062058 + layer.0.v_cache 0.00001549 0.01288699 + layer.1.k_cache 0.52552974 5.84385595 + layer.1.v_cache 0.00000676 0.00526844 + layer.2.k_cache 0.01473994 1.29568263 + layer.2.v_cache 0.00002188 0.01614376 + layer.3.k_cache 0.05129043 8.17094819 + layer.3.v_cache 0.00002012 0.01884922 + layer.4.k_cache 0.00075129 0.44528129 + layer.4.v_cache 0.00005717 0.03238963 + layer.4.output 0.00501330 209.12035040 + ------------------------------------------------------------------------------------- + TOTAL 0.04666922 90.12319879 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 222500 +BPFP 0.7717 bits/point +EBPFP 1.5434 equivalent bits/point +MSE 90.123199 +---------------------- -------------------------------------------------------- +Time: 3.193s Load: 0.009s, Pack+Encode: 1.833s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 90.1232 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 368, 128) +Output shape: (1, 368, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.output: torch.Size([1, 368, 3584]) -> torch.Size([1, 1, 368, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,400B, BPFP=0.2717 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,740B, BPFP=1.6024 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,188B, BPFP=0.5175 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,028B, BPFP=1.5297 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,436B, BPFP=0.7403 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,776B, BPFP=1.4766 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,544B, BPFP=0.7024 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,724B, BPFP=1.5168 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,484B, BPFP=1.2519 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,848B, BPFP=1.4372 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,672B, BPFP=0.0708 +⌛️ [2/4] FRONTEND: Frontend time: 1.961s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.469s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14147143 54.58423913 + layer.0.v_cache 0.00001611 0.01264313 + layer.1.k_cache 0.89062865 6.49686133 + layer.1.v_cache 0.00000637 0.00512405 + layer.2.k_cache 0.02201829 1.20482354 + layer.2.v_cache 0.00002088 0.01559105 + layer.3.k_cache 0.00965744 8.85594708 + layer.3.v_cache 0.00002101 0.01863134 + layer.4.k_cache 0.00077426 0.43532384 + layer.4.v_cache 0.00005366 0.03362716 + layer.4.output 0.03638816 147.18225932 + ------------------------------------------------------------------------------------- + TOTAL 0.07761090 64.81991923 + (elements=3,203,072) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3203072 +Total Bytes 271840 +BPFP 0.6789 bits/point +EBPFP 1.3579 equivalent bits/point +MSE 64.819919 +---------------------- -------------------------------------------------------- +Time: 3.445s Load: 0.014s, Pack+Encode: 1.961s, Decode+Unpack: 1.469s +---------------------- -------------------------------------------------------- +💾 Converting with 64.8199 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.020s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 453, 128) +Output shape: (1, 453, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.output: torch.Size([1, 453, 3584]) -> torch.Size([1, 1, 453, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,856B, BPFP=0.2710 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 50,040B, BPFP=1.7260 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,836B, BPFP=0.5117 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,516B, BPFP=1.6389 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,936B, BPFP=0.7566 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,868B, BPFP=1.5821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,132B, BPFP=0.7979 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,260B, BPFP=1.6301 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,776B, BPFP=1.3375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,920B, BPFP=1.5494 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,576B, BPFP=0.0620 +⌛️ [2/4] FRONTEND: Frontend time: 2.424s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.783s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13665763 52.12140418 + layer.0.v_cache 0.00001897 0.01296489 + layer.1.k_cache 1.07282356 6.74064667 + layer.1.v_cache 0.00000630 0.00508551 + layer.2.k_cache 0.02241544 1.17131544 + layer.2.v_cache 0.00001979 0.01497807 + layer.3.k_cache 0.01806638 8.21241258 + layer.3.v_cache 0.00002079 0.01828334 + layer.4.k_cache 0.00076567 0.43815060 + layer.4.v_cache 0.00005104 0.03156189 + layer.4.output 0.00584071 122.23391675 + ------------------------------------------------------------------------------------- + TOTAL 0.07598415 54.37671885 + (elements=3,942,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3942912 +Total Bytes 354716 +BPFP 0.7197 bits/point +EBPFP 1.4394 equivalent bits/point +MSE 54.376719 +---------------------- -------------------------------------------------------- +Time: 4.227s Load: 0.020s, Pack+Encode: 2.424s, Decode+Unpack: 1.783s +---------------------- -------------------------------------------------------- +💾 Converting with 54.3767 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.017s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 402, 128) +Output shape: (1, 402, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.output: torch.Size([1, 402, 3584]) -> torch.Size([1, 1, 402, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,040B, BPFP=0.2736 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,900B, BPFP=1.7063 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,076B, BPFP=0.5082 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,420B, BPFP=1.6099 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,440B, BPFP=0.7167 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,232B, BPFP=1.5637 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,924B, BPFP=0.6578 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,252B, BPFP=1.6034 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,780B, BPFP=1.2741 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,928B, BPFP=1.5131 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,132B, BPFP=0.0618 +⌛️ [2/4] FRONTEND: Frontend time: 2.088s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.575s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18062694 52.95540850 + layer.0.v_cache 0.00001752 0.01256762 + layer.1.k_cache 0.89169403 6.21076381 + layer.1.v_cache 0.00000593 0.00487294 + layer.2.k_cache 0.03854401 1.11309017 + layer.2.v_cache 0.00001967 0.01480712 + layer.3.k_cache 0.02336847 7.77731217 + layer.3.v_cache 0.00001924 0.01804258 + layer.4.k_cache 0.00085529 0.43457221 + layer.4.v_cache 0.00004959 0.03297577 + layer.4.output 0.00648942 137.48431947 + ------------------------------------------------------------------------------------- + TOTAL 0.06944863 60.64497937 + (elements=3,499,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3499008 +Total Bytes 305124 +BPFP 0.6976 bits/point +EBPFP 1.3952 equivalent bits/point +MSE 60.644979 +---------------------- -------------------------------------------------------- +Time: 3.680s Load: 0.017s, Pack+Encode: 2.088s, Decode+Unpack: 1.575s +---------------------- -------------------------------------------------------- +💾 Converting with 60.6450 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,976B, BPFP=0.2787 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,764B, BPFP=1.7789 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,824B, BPFP=0.5502 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,628B, BPFP=1.7153 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,608B, BPFP=0.8741 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,680B, BPFP=1.6622 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,292B, BPFP=0.8004 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,152B, BPFP=1.6886 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,960B, BPFP=1.3978 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,840B, BPFP=1.6151 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,004B, BPFP=0.0640 +⌛️ [2/4] FRONTEND: Frontend time: 1.850s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13958113 51.92088444 + layer.0.v_cache 0.00001688 0.01359819 + layer.1.k_cache 0.56864935 6.17537260 + layer.1.v_cache 0.00000621 0.00557818 + layer.2.k_cache 0.01018704 1.26474065 + layer.2.v_cache 0.00002046 0.01597126 + layer.3.k_cache 0.04374025 8.82384248 + layer.3.v_cache 0.00002178 0.01974520 + layer.4.k_cache 0.00068741 0.45142323 + layer.4.v_cache 0.00005272 0.03346190 + layer.4.output 0.00479460 198.90122568 + ------------------------------------------------------------------------------------- + TOTAL 0.04685444 85.94312929 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 228728 +BPFP 0.7535 bits/point +EBPFP 1.5070 equivalent bits/point +MSE 85.943129 +---------------------- -------------------------------------------------------- +Time: 3.204s Load: 0.010s, Pack+Encode: 1.850s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 85.9431 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 381, 128) +Output shape: (1, 381, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.output: torch.Size([1, 381, 3584]) -> torch.Size([1, 1, 381, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,516B, BPFP=0.2672 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,256B, BPFP=1.5689 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,460B, BPFP=0.5110 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,424B, BPFP=1.4938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,044B, BPFP=0.6990 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,964B, BPFP=1.4339 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,940B, BPFP=0.6537 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,108B, BPFP=1.4808 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,328B, BPFP=1.2438 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,956B, BPFP=1.3926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,448B, BPFP=0.0729 +⌛️ [2/4] FRONTEND: Frontend time: 1.962s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.467s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15713709 53.83034469 + layer.0.v_cache 0.00001719 0.01317339 + layer.1.k_cache 0.86767426 6.83058883 + layer.1.v_cache 0.00000622 0.00535369 + layer.2.k_cache 0.01635739 1.22252793 + layer.2.v_cache 0.00002098 0.01594128 + layer.3.k_cache 0.01709203 9.46485272 + layer.3.v_cache 0.00002073 0.01869591 + layer.4.k_cache 0.00073405 0.46434241 + layer.4.v_cache 0.00005083 0.03278824 + layer.4.output 0.03519400 142.16376078 + ------------------------------------------------------------------------------------- + TOTAL 0.07679228 62.76734909 + (elements=3,316,224) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3316224 +Total Bytes 274444 +BPFP 0.6621 bits/point +EBPFP 1.3241 equivalent bits/point +MSE 62.767349 +---------------------- -------------------------------------------------------- +Time: 3.443s Load: 0.014s, Pack+Encode: 1.962s, Decode+Unpack: 1.467s +---------------------- -------------------------------------------------------- +💾 Converting with 62.7673 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,228B, BPFP=0.2823 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,648B, BPFP=1.7126 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,988B, BPFP=0.5334 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,540B, BPFP=1.6386 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,280B, BPFP=0.8200 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,696B, BPFP=1.5823 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,140B, BPFP=0.8106 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,336B, BPFP=1.6250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,320B, BPFP=1.3568 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,188B, BPFP=1.5483 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,512B, BPFP=0.0717 +⌛️ [2/4] FRONTEND: Frontend time: 1.716s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422791 53.13917268 + layer.0.v_cache 0.00001930 0.01356552 + layer.1.k_cache 0.38928963 6.35301378 + layer.1.v_cache 0.00000631 0.00543423 + layer.2.k_cache 0.01623082 1.32946921 + layer.2.v_cache 0.00002095 0.01605270 + layer.3.k_cache 0.01313608 8.94693802 + layer.3.v_cache 0.00002100 0.01911828 + layer.4.k_cache 0.00066994 0.44627276 + layer.4.v_cache 0.00005661 0.03353736 + layer.4.output 1.30836332 231.28468407 + ------------------------------------------------------------------------------------- + TOTAL 0.57013069 99.37031547 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 185876 +BPFP 0.7301 bits/point +EBPFP 1.4602 equivalent bits/point +MSE 99.370315 +---------------------- -------------------------------------------------------- +Time: 2.943s Load: 0.012s, Pack+Encode: 1.716s, Decode+Unpack: 1.215s +---------------------- -------------------------------------------------------- +💾 Converting with 99.3703 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,292B, BPFP=0.2842 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,428B, BPFP=1.6835 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,052B, BPFP=0.5331 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,412B, BPFP=1.6163 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,900B, BPFP=0.7879 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,436B, BPFP=1.5516 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,872B, BPFP=0.7860 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,400B, BPFP=1.6155 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,432B, BPFP=1.3528 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,892B, BPFP=1.5156 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,716B, BPFP=0.0635 +⌛️ [2/4] FRONTEND: Frontend time: 1.725s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13591524 53.44896220 + layer.0.v_cache 0.00001656 0.01357808 + layer.1.k_cache 0.34520291 6.11970339 + layer.1.v_cache 0.00000598 0.00538674 + layer.2.k_cache 0.01250870 1.24403200 + layer.2.v_cache 0.00001982 0.01536231 + layer.3.k_cache 0.02633334 8.94528043 + layer.3.v_cache 0.00002177 0.01990076 + layer.4.k_cache 0.00067446 0.44505853 + layer.4.v_cache 0.00005295 0.03294238 + layer.4.output 1.29725540 229.27487515 + ------------------------------------------------------------------------------------- + TOTAL 0.56479644 98.54201958 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 183832 +BPFP 0.7159 bits/point +EBPFP 1.4319 equivalent bits/point +MSE 98.542020 +---------------------- -------------------------------------------------------- +Time: 2.957s Load: 0.010s, Pack+Encode: 1.725s, Decode+Unpack: 1.222s +---------------------- -------------------------------------------------------- +💾 Converting with 98.5420 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,432B, BPFP=0.3191 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,296B, BPFP=1.8214 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,348B, BPFP=0.5291 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,072B, BPFP=1.7333 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,380B, BPFP=0.8194 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,564B, BPFP=1.6967 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,352B, BPFP=0.8174 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,124B, BPFP=1.7370 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,824B, BPFP=1.4274 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,984B, BPFP=1.6550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,316B, BPFP=0.0650 +⌛️ [2/4] FRONTEND: Frontend time: 1.722s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150146 53.77047181 + layer.0.v_cache 0.00001564 0.01338080 + layer.1.k_cache 0.36060091 6.35874271 + layer.1.v_cache 0.00000598 0.00533886 + layer.2.k_cache 0.01419523 1.04432446 + layer.2.v_cache 0.00002081 0.01596883 + layer.3.k_cache 0.03680107 7.90532125 + layer.3.v_cache 0.00002092 0.01892883 + layer.4.k_cache 0.00067164 0.42597438 + layer.4.v_cache 0.00004838 0.03212978 + layer.4.output 1.41076765 249.41431452 + ------------------------------------------------------------------------------------- + TOTAL 0.61289739 106.79357549 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 180692 +BPFP 0.7653 bits/point +EBPFP 1.5307 equivalent bits/point +MSE 106.793575 +---------------------- -------------------------------------------------------- +Time: 2.953s Load: 0.007s, Pack+Encode: 1.722s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 106.7936 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 343, 128) +Output shape: (1, 343, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.output: torch.Size([1, 343, 3584]) -> torch.Size([1, 1, 343, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,992B, BPFP=0.2730 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,060B, BPFP=1.7338 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,672B, BPFP=0.5317 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,776B, BPFP=1.6753 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,640B, BPFP=0.8036 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,276B, BPFP=1.6070 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,312B, BPFP=0.7886 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,232B, BPFP=1.6505 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,508B, BPFP=1.3442 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,324B, BPFP=1.5636 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,472B, BPFP=0.0747 +⌛️ [2/4] FRONTEND: Frontend time: 1.977s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.467s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16677478 52.93906022 + layer.0.v_cache 0.00001698 0.01310420 + layer.1.k_cache 0.72144124 6.55376005 + layer.1.v_cache 0.00000659 0.00553997 + layer.2.k_cache 0.01351720 1.29574354 + layer.2.v_cache 0.00002117 0.01566283 + layer.3.k_cache 0.02347133 8.93526287 + layer.3.v_cache 0.00002006 0.01863567 + layer.4.k_cache 0.00069292 0.43912633 + layer.4.v_cache 0.00005144 0.03143804 + layer.4.output 0.03905215 158.07395356 + ------------------------------------------------------------------------------------- + TOTAL 0.07055169 69.22147110 + (elements=2,985,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2985472 +Total Bytes 274264 +BPFP 0.7349 bits/point +EBPFP 1.4699 equivalent bits/point +MSE 69.221471 +---------------------- -------------------------------------------------------- +Time: 3.456s Load: 0.013s, Pack+Encode: 1.977s, Decode+Unpack: 1.467s +---------------------- -------------------------------------------------------- +💾 Converting with 69.2215 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,864B, BPFP=0.2764 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,504B, BPFP=1.7900 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,136B, BPFP=0.5191 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,348B, BPFP=1.7243 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,080B, BPFP=0.8000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,400B, BPFP=1.6705 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,960B, BPFP=0.9068 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,764B, BPFP=1.6911 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,496B, BPFP=1.3350 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,600B, BPFP=1.6250 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,040B, BPFP=0.0734 +⌛️ [2/4] FRONTEND: Frontend time: 1.863s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13317820 52.84698864 + layer.0.v_cache 0.00001784 0.01316937 + layer.1.k_cache 0.51782770 6.01769798 + layer.1.v_cache 0.00000612 0.00527874 + layer.2.k_cache 0.01694398 1.26877064 + layer.2.v_cache 0.00002016 0.01560547 + layer.3.k_cache 0.01920085 7.72783913 + layer.3.v_cache 0.00002016 0.01879169 + layer.4.k_cache 0.00071488 0.44061490 + layer.4.v_cache 0.00005349 0.03313447 + layer.4.output 0.00484358 201.65048701 + ------------------------------------------------------------------------------------- + TOTAL 0.04246403 87.05537060 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 226192 +BPFP 0.7560 bits/point +EBPFP 1.5120 equivalent bits/point +MSE 87.055371 +---------------------- -------------------------------------------------------- +Time: 3.221s Load: 0.011s, Pack+Encode: 1.863s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 87.0554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 374, 128) +Output shape: (1, 374, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.output: torch.Size([1, 374, 3584]) -> torch.Size([1, 1, 374, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,436B, BPFP=0.2689 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,164B, BPFP=1.5944 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,152B, BPFP=0.5077 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,404B, BPFP=1.5209 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,708B, BPFP=0.6980 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,564B, BPFP=1.4440 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,788B, BPFP=0.7014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,520B, BPFP=1.4840 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,776B, BPFP=1.2440 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,652B, BPFP=1.4059 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,784B, BPFP=0.0584 +⌛️ [2/4] FRONTEND: Frontend time: 1.966s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18781370 54.28567325 + layer.0.v_cache 0.00001670 0.01280416 + layer.1.k_cache 0.85217587 6.41625781 + layer.1.v_cache 0.00000605 0.00516944 + layer.2.k_cache 0.02919085 1.10147111 + layer.2.v_cache 0.00002159 0.01497612 + layer.3.k_cache 0.02976600 8.39217666 + layer.3.v_cache 0.00002022 0.01847318 + layer.4.k_cache 0.00072409 0.43614340 + layer.4.v_cache 0.00004949 0.03319996 + layer.4.output 0.03578218 144.78695569 + ------------------------------------------------------------------------------------- + TOTAL 0.07942705 63.77794323 + (elements=3,255,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3255296 +Total Bytes 269948 +BPFP 0.6634 bits/point +EBPFP 1.3268 equivalent bits/point +MSE 63.777943 +---------------------- -------------------------------------------------------- +Time: 3.450s Load: 0.014s, Pack+Encode: 1.966s, Decode+Unpack: 1.470s +---------------------- -------------------------------------------------------- +💾 Converting with 63.7779 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,308B, BPFP=0.2793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,296B, BPFP=1.6400 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,228B, BPFP=0.5335 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,312B, BPFP=1.5762 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,916B, BPFP=0.7726 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,484B, BPFP=1.5226 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,552B, BPFP=0.7490 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,096B, BPFP=1.5622 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,300B, BPFP=1.3161 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,952B, BPFP=1.4881 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,880B, BPFP=0.0637 +⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13979621 52.83264312 + layer.0.v_cache 0.00001743 0.01289639 + layer.1.k_cache 0.46959303 6.87593908 + layer.1.v_cache 0.00000602 0.00536712 + layer.2.k_cache 0.01642122 1.19948809 + layer.2.v_cache 0.00002053 0.01596527 + layer.3.k_cache 0.03876302 8.73028742 + layer.3.v_cache 0.00002036 0.01926594 + layer.4.k_cache 0.00069112 0.43971639 + layer.4.v_cache 0.00005399 0.03435292 + layer.4.output 1.27033806 224.23544013 + ------------------------------------------------------------------------------------- + TOTAL 0.56222055 96.45964722 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 183324 +BPFP 0.6992 bits/point +EBPFP 1.3983 equivalent bits/point +MSE 96.459647 +---------------------- -------------------------------------------------------- +Time: 2.966s Load: 0.009s, Pack+Encode: 1.735s, Decode+Unpack: 1.222s +---------------------- -------------------------------------------------------- +💾 Converting with 96.4596 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,040B, BPFP=0.2781 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,384B, BPFP=1.7472 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,780B, BPFP=0.5355 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,432B, BPFP=1.6817 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,584B, BPFP=0.7974 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,612B, BPFP=1.6253 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,372B, BPFP=0.8516 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,320B, BPFP=1.6740 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,308B, BPFP=1.3979 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,980B, BPFP=1.5818 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,788B, BPFP=0.0667 +⌛️ [2/4] FRONTEND: Frontend time: 1.724s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13893990 51.08535672 + layer.0.v_cache 0.00001711 0.01357969 + layer.1.k_cache 0.36820191 6.23061717 + layer.1.v_cache 0.00000637 0.00553709 + layer.2.k_cache 0.01566778 1.27905852 + layer.2.v_cache 0.00002161 0.01654349 + layer.3.k_cache 0.01884081 8.69352178 + layer.3.v_cache 0.00002076 0.01994106 + layer.4.k_cache 0.00070054 0.45895016 + layer.4.v_cache 0.00005210 0.03440710 + layer.4.output 1.34866800 238.47506293 + ------------------------------------------------------------------------------------- + TOTAL 0.58724382 102.18605608 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 183600 +BPFP 0.7434 bits/point +EBPFP 1.4868 equivalent bits/point +MSE 102.186056 +---------------------- -------------------------------------------------------- +Time: 2.959s Load: 0.010s, Pack+Encode: 1.724s, Decode+Unpack: 1.225s +---------------------- -------------------------------------------------------- +💾 Converting with 102.1861 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,836B, BPFP=0.2830 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,036B, BPFP=1.8162 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,132B, BPFP=0.5344 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,864B, BPFP=1.7477 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,836B, BPFP=0.8682 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,888B, BPFP=1.6905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,920B, BPFP=0.8146 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,612B, BPFP=1.7329 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,640B, BPFP=1.4419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,568B, BPFP=1.6718 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,492B, BPFP=0.0626 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11943251 54.96276627 + layer.0.v_cache 0.00001561 0.01296117 + layer.1.k_cache 0.49619702 6.38017679 + layer.1.v_cache 0.00000608 0.00521629 + layer.2.k_cache 0.01349212 1.31305869 + layer.2.v_cache 0.00002006 0.01568095 + layer.3.k_cache 0.01080775 8.34891993 + layer.3.v_cache 0.00002073 0.01858155 + layer.4.k_cache 0.00068886 0.44172260 + layer.4.v_cache 0.00007406 0.03346214 + layer.4.output 0.00494179 207.79736825 + ------------------------------------------------------------------------------------- + TOTAL 0.03972631 89.77141906 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 222824 +BPFP 0.7670 bits/point +EBPFP 1.5341 equivalent bits/point +MSE 89.771419 +---------------------- -------------------------------------------------------- +Time: 3.199s Load: 0.009s, Pack+Encode: 1.841s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 89.7714 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,984B, BPFP=0.2817 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,248B, BPFP=1.7851 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,328B, BPFP=0.5181 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,204B, BPFP=1.7113 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,536B, BPFP=0.8863 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,600B, BPFP=1.6686 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,332B, BPFP=0.8012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,268B, BPFP=1.7158 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,084B, BPFP=1.4200 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,080B, BPFP=1.6318 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,464B, BPFP=0.0653 +⌛️ [2/4] FRONTEND: Frontend time: 1.727s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.226s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11815372 55.76227110 + layer.0.v_cache 0.00001627 0.01345619 + layer.1.k_cache 0.37623358 6.50778599 + layer.1.v_cache 0.00000616 0.00548155 + layer.2.k_cache 0.01429399 1.32227764 + layer.2.v_cache 0.00002051 0.01608811 + layer.3.k_cache 0.01320818 8.32700759 + layer.3.v_cache 0.00002004 0.01988576 + layer.4.k_cache 0.00067885 0.45047408 + layer.4.v_cache 0.00005318 0.03499671 + layer.4.output 1.38524649 244.63497899 + ------------------------------------------------------------------------------------- + TOTAL 0.60114176 104.99438692 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 182128 +BPFP 0.7575 bits/point +EBPFP 1.5149 equivalent bits/point +MSE 104.994387 +---------------------- -------------------------------------------------------- +Time: 2.961s Load: 0.008s, Pack+Encode: 1.727s, Decode+Unpack: 1.226s +---------------------- -------------------------------------------------------- +💾 Converting with 104.9944 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,960B, BPFP=0.2788 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,280B, BPFP=1.7581 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,404B, BPFP=0.5286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,384B, BPFP=1.7077 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,536B, BPFP=0.8732 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,172B, BPFP=1.6396 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,552B, BPFP=0.8179 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,932B, BPFP=1.6823 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,116B, BPFP=1.3554 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,740B, BPFP=1.6153 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,120B, BPFP=0.0572 +⌛️ [2/4] FRONTEND: Frontend time: 1.852s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09820673 51.84890681 + layer.0.v_cache 0.00001666 0.01274610 + layer.1.k_cache 0.50544519 6.54817299 + layer.1.v_cache 0.00000613 0.00530184 + layer.2.k_cache 0.01930715 1.21816599 + layer.2.v_cache 0.00002033 0.01588134 + layer.3.k_cache 0.01337509 8.06500903 + layer.3.v_cache 0.00001987 0.01889171 + layer.4.k_cache 0.00074230 0.43370707 + layer.4.v_cache 0.00004869 0.03176655 + layer.4.output 0.00473525 199.17181077 + ------------------------------------------------------------------------------------- + TOTAL 0.03943146 86.02360146 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 225196 +BPFP 0.7445 bits/point +EBPFP 1.4891 equivalent bits/point +MSE 86.023601 +---------------------- -------------------------------------------------------- +Time: 3.217s Load: 0.012s, Pack+Encode: 1.852s, Decode+Unpack: 1.354s +---------------------- -------------------------------------------------------- +💾 Converting with 86.0236 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,356B, BPFP=0.3122 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,288B, BPFP=1.8125 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,352B, BPFP=0.5269 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,148B, BPFP=1.7308 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,088B, BPFP=0.8664 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,572B, BPFP=1.6895 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,072B, BPFP=0.8653 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,328B, BPFP=1.7437 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,744B, BPFP=1.4151 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,124B, BPFP=1.6574 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,688B, BPFP=0.0582 +⌛️ [2/4] FRONTEND: Frontend time: 1.724s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13853801 54.25628942 + layer.0.v_cache 0.00001803 0.01332862 + layer.1.k_cache 0.23740530 6.15691355 + layer.1.v_cache 0.00000581 0.00524420 + layer.2.k_cache 0.01896443 1.28717937 + layer.2.v_cache 0.00001977 0.01623768 + layer.3.k_cache 0.00876656 7.85054282 + layer.3.v_cache 0.00002034 0.01978806 + layer.4.k_cache 0.00073522 0.45627038 + layer.4.v_cache 0.00004914 0.03345714 + layer.4.output 1.40427464 248.28340023 + ------------------------------------------------------------------------------------- + TOTAL 0.60202618 106.35759134 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 181760 +BPFP 0.7663 bits/point +EBPFP 1.5326 equivalent bits/point +MSE 106.357591 +---------------------- -------------------------------------------------------- +Time: 2.956s Load: 0.009s, Pack+Encode: 1.724s, Decode+Unpack: 1.222s +---------------------- -------------------------------------------------------- +💾 Converting with 106.3576 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,964B, BPFP=0.2810 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,280B, BPFP=1.7708 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,124B, BPFP=0.5165 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,164B, BPFP=1.7077 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,300B, BPFP=0.8662 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,156B, BPFP=1.6506 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,576B, BPFP=0.7686 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,236B, BPFP=1.7117 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,772B, BPFP=1.4024 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,728B, BPFP=1.6264 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,120B, BPFP=0.0657 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10827675 51.14870287 + layer.0.v_cache 0.00001706 0.01282624 + layer.1.k_cache 0.56919524 5.69915816 + layer.1.v_cache 0.00000592 0.00512078 + layer.2.k_cache 0.01911232 1.32764048 + layer.2.v_cache 0.00002018 0.01562842 + layer.3.k_cache 0.01342056 8.40096117 + layer.3.v_cache 0.00002333 0.01976270 + layer.4.k_cache 0.00074072 0.44302006 + layer.4.v_cache 0.00005062 0.03281088 + layer.4.output 0.00480139 200.76740424 + ------------------------------------------------------------------------------------- + TOTAL 0.04379250 86.61632126 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 225420 +BPFP 0.7507 bits/point +EBPFP 1.5014 equivalent bits/point +MSE 86.616321 +---------------------- -------------------------------------------------------- +Time: 3.216s Load: 0.009s, Pack+Encode: 1.856s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 86.6163 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,820B, BPFP=0.2821 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,564B, BPFP=1.8471 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,544B, BPFP=0.5585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,152B, BPFP=1.7645 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,616B, BPFP=0.9139 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,056B, BPFP=1.7004 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,888B, BPFP=0.8127 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,836B, BPFP=1.7460 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,832B, BPFP=1.3947 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,376B, BPFP=1.6606 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,628B, BPFP=0.0638 +⌛️ [2/4] FRONTEND: Frontend time: 1.849s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.361s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15963619 55.36418466 + layer.0.v_cache 0.00001721 0.01422594 + layer.1.k_cache 0.49929055 6.38477751 + layer.1.v_cache 0.00000624 0.00540880 + layer.2.k_cache 0.03270629 1.38626270 + layer.2.v_cache 0.00002062 0.01570542 + layer.3.k_cache 0.00946706 8.88718509 + layer.3.v_cache 0.00002056 0.01895847 + layer.4.k_cache 0.00068536 0.43908246 + layer.4.v_cache 0.00004941 0.03253413 + layer.4.output 0.00495045 207.73724251 + ------------------------------------------------------------------------------------- + TOTAL 0.04332662 89.80641310 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 224312 +BPFP 0.7722 bits/point +EBPFP 1.5443 equivalent bits/point +MSE 89.806413 +---------------------- -------------------------------------------------------- +Time: 3.220s Load: 0.009s, Pack+Encode: 1.849s, Decode+Unpack: 1.361s +---------------------- -------------------------------------------------------- +💾 Converting with 89.8064 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,284B, BPFP=0.2836 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,528B, BPFP=1.6901 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,960B, BPFP=0.5270 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,464B, BPFP=1.6197 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,228B, BPFP=0.8096 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,344B, BPFP=1.5456 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,904B, BPFP=0.7881 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,084B, BPFP=1.5945 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,596B, BPFP=1.3636 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,912B, BPFP=1.5169 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,508B, BPFP=0.0616 +⌛️ [2/4] FRONTEND: Frontend time: 1.731s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.227s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14874478 52.40370514 + layer.0.v_cache 0.00001714 0.01328847 + layer.1.k_cache 0.45078779 6.13747186 + layer.1.v_cache 0.00000615 0.00531276 + layer.2.k_cache 0.01724745 1.32253589 + layer.2.v_cache 0.00001987 0.01512382 + layer.3.k_cache 0.02849353 8.24497132 + layer.3.v_cache 0.00002056 0.01882261 + layer.4.k_cache 0.00066912 0.43798498 + layer.4.v_cache 0.00005068 0.03223056 + layer.4.output 1.29722614 229.31581038 + ------------------------------------------------------------------------------------- + TOTAL 0.57215530 98.46130118 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 183812 +BPFP 0.7159 bits/point +EBPFP 1.4317 equivalent bits/point +MSE 98.461301 +---------------------- -------------------------------------------------------- +Time: 2.966s Load: 0.008s, Pack+Encode: 1.731s, Decode+Unpack: 1.227s +---------------------- -------------------------------------------------------- +💾 Converting with 98.4613 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 341, 128) +Output shape: (1, 341, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.output: torch.Size([1, 341, 3584]) -> torch.Size([1, 1, 341, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,024B, BPFP=0.2760 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,836B, BPFP=1.7337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,736B, BPFP=0.5378 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,508B, BPFP=1.6728 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,008B, BPFP=0.7793 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,912B, BPFP=1.5997 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,228B, BPFP=0.7894 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,956B, BPFP=1.6475 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,272B, BPFP=1.2955 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,116B, BPFP=1.5632 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,672B, BPFP=0.0633 +⌛️ [2/4] FRONTEND: Frontend time: 1.961s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.462s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17009405 51.92382812 + layer.0.v_cache 0.00001631 0.01316376 + layer.1.k_cache 0.79816426 6.19042253 + layer.1.v_cache 0.00000611 0.00516547 + layer.2.k_cache 0.02878924 1.26103483 + layer.2.v_cache 0.00002016 0.01520942 + layer.3.k_cache 0.01035325 8.97092865 + layer.3.v_cache 0.00002082 0.01817949 + layer.4.k_cache 0.00072160 0.43225684 + layer.4.v_cache 0.00005837 0.03128094 + layer.4.output 0.03919861 158.95485966 + ------------------------------------------------------------------------------------- + TOTAL 0.07544909 69.50267574 + (elements=2,968,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2968064 +Total Bytes 269268 +BPFP 0.7258 bits/point +EBPFP 1.4515 equivalent bits/point +MSE 69.502676 +---------------------- -------------------------------------------------------- +Time: 3.435s Load: 0.012s, Pack+Encode: 1.961s, Decode+Unpack: 1.462s +---------------------- -------------------------------------------------------- +💾 Converting with 69.5027 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 335, 128) +Output shape: (1, 335, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.output: torch.Size([1, 335, 3584]) -> torch.Size([1, 1, 335, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,956B, BPFP=0.2778 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,832B, BPFP=1.7646 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,220B, BPFP=0.5233 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,088B, BPFP=1.6832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,332B, BPFP=0.8084 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,524B, BPFP=1.6103 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,488B, BPFP=0.8157 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,728B, BPFP=1.6664 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,372B, BPFP=1.3233 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,748B, BPFP=1.5741 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,944B, BPFP=0.0663 +⌛️ [2/4] FRONTEND: Frontend time: 1.974s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.474s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15195324 53.38887593 + layer.0.v_cache 0.00001666 0.01340307 + layer.1.k_cache 0.76788093 6.16454276 + layer.1.v_cache 0.00000624 0.00542209 + layer.2.k_cache 0.02686723 1.22780470 + layer.2.v_cache 0.00001960 0.01489780 + layer.3.k_cache 0.01080669 8.74803084 + layer.3.v_cache 0.00002024 0.01833201 + layer.4.k_cache 0.00071790 0.44205655 + layer.4.v_cache 0.00005106 0.03247896 + layer.4.output 3.38143948 160.08909915 + ------------------------------------------------------------------------------------- + TOTAL 1.44873036 70.03997287 + (elements=2,915,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2915840 +Total Bytes 268232 +BPFP 0.7359 bits/point +EBPFP 1.4719 equivalent bits/point +MSE 70.039973 +---------------------- -------------------------------------------------------- +Time: 3.460s Load: 0.012s, Pack+Encode: 1.974s, Decode+Unpack: 1.474s +---------------------- -------------------------------------------------------- +💾 Converting with 70.0400 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,768B, BPFP=0.2811 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,484B, BPFP=1.8564 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,364B, BPFP=0.5521 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,884B, BPFP=1.7620 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,812B, BPFP=0.9323 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,576B, BPFP=1.6849 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,532B, BPFP=0.9158 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,672B, BPFP=1.7495 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,456B, BPFP=1.4420 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,036B, BPFP=1.6531 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,296B, BPFP=0.0783 +⌛️ [2/4] FRONTEND: Frontend time: 1.846s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17469128 56.55456589 + layer.0.v_cache 0.00001775 0.01377080 + layer.1.k_cache 0.51265420 5.86403164 + layer.1.v_cache 0.00000620 0.00547732 + layer.2.k_cache 0.03428403 1.33076805 + layer.2.v_cache 0.00002024 0.01534511 + layer.3.k_cache 0.01443097 8.63855027 + layer.3.v_cache 0.00002182 0.02025227 + layer.4.k_cache 0.00070119 0.43409165 + layer.4.v_cache 0.00005097 0.03310368 + layer.4.output 0.00503371 208.97710580 + ------------------------------------------------------------------------------------- + TOTAL 0.04541851 90.33821748 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 226880 +BPFP 0.7869 bits/point +EBPFP 1.5738 equivalent bits/point +MSE 90.338217 +---------------------- -------------------------------------------------------- +Time: 3.211s Load: 0.011s, Pack+Encode: 1.846s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 90.3382 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,792B, BPFP=0.2825 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,996B, BPFP=1.8276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,124B, BPFP=0.5380 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,776B, BPFP=1.7557 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,592B, BPFP=0.9193 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,552B, BPFP=1.6835 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,580B, BPFP=0.8597 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,512B, BPFP=1.7401 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,080B, BPFP=1.4198 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,140B, BPFP=1.6592 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,700B, BPFP=0.0649 +⌛️ [2/4] FRONTEND: Frontend time: 1.844s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18043976 57.06047317 + layer.0.v_cache 0.00001814 0.01304246 + layer.1.k_cache 0.58196952 6.47155808 + layer.1.v_cache 0.00000608 0.00529998 + layer.2.k_cache 0.02409397 1.26626990 + layer.2.v_cache 0.00002020 0.01518581 + layer.3.k_cache 0.01875559 8.53771005 + layer.3.v_cache 0.00002085 0.01886589 + layer.4.k_cache 0.00069073 0.43924713 + layer.4.v_cache 0.00005389 0.03348402 + layer.4.output 0.00498411 209.15055593 + ------------------------------------------------------------------------------------- + TOTAL 0.04946809 90.46558988 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 222844 +BPFP 0.7729 bits/point +EBPFP 1.5458 equivalent bits/point +MSE 90.465590 +---------------------- -------------------------------------------------------- +Time: 3.201s Load: 0.010s, Pack+Encode: 1.844s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 90.4656 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 386, 128) +Output shape: (1, 386, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.output: torch.Size([1, 386, 3584]) -> torch.Size([1, 1, 386, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,660B, BPFP=0.2696 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,940B, BPFP=1.7787 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,884B, BPFP=0.5215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,916B, BPFP=1.6967 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,896B, BPFP=0.7649 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,060B, BPFP=1.6621 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,716B, BPFP=0.7576 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,920B, BPFP=1.6969 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,084B, BPFP=1.3797 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,872B, BPFP=1.6140 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,024B, BPFP=0.0580 +⌛️ [2/4] FRONTEND: Frontend time: 2.079s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.574s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16808204 53.24425194 + layer.0.v_cache 0.00001827 0.01358549 + layer.1.k_cache 0.89383947 6.65343226 + layer.1.v_cache 0.00000605 0.00536549 + layer.2.k_cache 0.02643975 1.31790683 + layer.2.v_cache 0.00002094 0.01610265 + layer.3.k_cache 0.04368651 9.10416709 + layer.3.v_cache 0.00002149 0.01979364 + layer.4.k_cache 0.00075087 0.46577564 + layer.4.v_cache 0.00005208 0.03284171 + layer.4.output 0.03469837 140.40195920 + ------------------------------------------------------------------------------------- + TOTAL 0.08092977 61.98158454 + (elements=3,359,744) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3359744 +Total Bytes 309972 +BPFP 0.7381 bits/point +EBPFP 1.4762 equivalent bits/point +MSE 61.981585 +---------------------- -------------------------------------------------------- +Time: 3.665s Load: 0.012s, Pack+Encode: 2.079s, Decode+Unpack: 1.574s +---------------------- -------------------------------------------------------- +💾 Converting with 61.9816 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,032B, BPFP=0.2730 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,624B, BPFP=1.7157 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,524B, BPFP=0.5167 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,492B, BPFP=1.6543 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,184B, BPFP=0.7695 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,408B, BPFP=1.5955 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,616B, BPFP=0.7387 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,084B, BPFP=1.6322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,296B, BPFP=1.3724 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,332B, BPFP=1.5371 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,856B, BPFP=0.0609 +⌛️ [2/4] FRONTEND: Frontend time: 1.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14364293 53.14444987 + layer.0.v_cache 0.00001600 0.01369544 + layer.1.k_cache 0.62366660 6.34645208 + layer.1.v_cache 0.00000628 0.00566678 + layer.2.k_cache 0.01981349 1.20635690 + layer.2.v_cache 0.00002090 0.01628117 + layer.3.k_cache 0.02158341 9.33163622 + layer.3.v_cache 0.00002097 0.01965684 + layer.4.k_cache 0.00071696 0.45315234 + layer.4.v_cache 0.00005883 0.03376757 + layer.4.output 0.00463133 192.62501550 + ------------------------------------------------------------------------------------- + TOTAL 0.04952739 83.46742492 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 225448 +BPFP 0.7195 bits/point +EBPFP 1.4390 equivalent bits/point +MSE 83.467425 +---------------------- -------------------------------------------------------- +Time: 3.207s Load: 0.012s, Pack+Encode: 1.847s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 83.4674 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,928B, BPFP=0.2780 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,408B, BPFP=1.7717 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,364B, BPFP=0.5282 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,996B, BPFP=1.6920 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,036B, BPFP=0.8481 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,180B, BPFP=1.6460 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,720B, BPFP=0.8303 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,044B, BPFP=1.6947 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,300B, BPFP=1.3143 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,836B, BPFP=1.6266 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,376B, BPFP=0.0836 +⌛️ [2/4] FRONTEND: Frontend time: 1.852s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12300906 51.45684087 + layer.0.v_cache 0.00001720 0.01270821 + layer.1.k_cache 0.58060381 6.20379044 + layer.1.v_cache 0.00000644 0.00507936 + layer.2.k_cache 0.02305070 1.27816034 + layer.2.v_cache 0.00002053 0.01575953 + layer.3.k_cache 0.05279791 8.37947562 + layer.3.v_cache 0.00002021 0.01889713 + layer.4.k_cache 0.00074809 0.41771332 + layer.4.v_cache 0.00005348 0.03308627 + layer.4.output 0.00483522 200.22853275 + ------------------------------------------------------------------------------------- + TOTAL 0.04789259 86.43654355 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 227188 +BPFP 0.7538 bits/point +EBPFP 1.5077 equivalent bits/point +MSE 86.436544 +---------------------- -------------------------------------------------------- +Time: 3.213s Load: 0.010s, Pack+Encode: 1.852s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 86.4365 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,936B, BPFP=0.2758 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,400B, BPFP=1.7797 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,556B, BPFP=0.5294 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,360B, BPFP=1.7068 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,656B, BPFP=0.8167 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,612B, BPFP=1.6544 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,096B, BPFP=0.7775 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,236B, BPFP=1.6982 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,544B, BPFP=1.4395 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,892B, BPFP=1.6040 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,880B, BPFP=0.0589 +⌛️ [2/4] FRONTEND: Frontend time: 1.717s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16177035 53.53926394 + layer.0.v_cache 0.00001782 0.01411698 + layer.1.k_cache 0.31952496 6.23827961 + layer.1.v_cache 0.00000626 0.00551999 + layer.2.k_cache 0.01271096 1.41675954 + layer.2.v_cache 0.00002040 0.01652919 + layer.3.k_cache 0.01363577 8.79838528 + layer.3.v_cache 0.00002033 0.01951507 + layer.4.k_cache 0.00066992 0.45159153 + layer.4.v_cache 0.00005030 0.03322181 + layer.4.output 1.37284199 242.77250160 + ------------------------------------------------------------------------------------- + TOTAL 0.59519535 104.11415848 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 181168 +BPFP 0.7467 bits/point +EBPFP 1.4934 equivalent bits/point +MSE 104.114158 +---------------------- -------------------------------------------------------- +Time: 2.944s Load: 0.008s, Pack+Encode: 1.717s, Decode+Unpack: 1.219s +---------------------- -------------------------------------------------------- +💾 Converting with 104.1142 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,384B, BPFP=0.2889 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,472B, BPFP=1.6626 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,136B, BPFP=0.5239 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,500B, BPFP=1.5796 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,532B, BPFP=0.8139 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,108B, BPFP=1.5461 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,396B, BPFP=0.8023 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,740B, BPFP=1.6001 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,056B, BPFP=1.3709 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,684B, BPFP=1.5099 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,780B, BPFP=0.0705 +⌛️ [2/4] FRONTEND: Frontend time: 1.848s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.120s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09962382 52.45799180 + layer.0.v_cache 0.00001686 0.01425502 + layer.1.k_cache 0.10988800 6.60562684 + layer.1.v_cache 0.00000590 0.00557014 + layer.2.k_cache 0.01214521 1.31745144 + layer.2.v_cache 0.00002216 0.01685677 + layer.3.k_cache 0.02260447 8.92584128 + layer.3.v_cache 0.00002006 0.02131635 + layer.4.k_cache 0.00069723 0.46238325 + layer.4.v_cache 0.00005242 0.03601724 + layer.4.output 0.00882482 299.99146175 + ------------------------------------------------------------------------------------- + TOTAL 0.01804999 127.63550249 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 142788 +BPFP 0.7172 bits/point +EBPFP 1.4343 equivalent bits/point +MSE 127.635502 +---------------------- -------------------------------------------------------- +Time: 2.976s Load: 0.008s, Pack+Encode: 1.848s, Decode+Unpack: 1.120s +---------------------- -------------------------------------------------------- +💾 Converting with 127.6355 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,056B, BPFP=0.2811 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,828B, BPFP=1.7698 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,072B, BPFP=0.5601 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,464B, BPFP=1.6940 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,280B, BPFP=0.7940 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,180B, BPFP=1.6226 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,248B, BPFP=0.8479 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,172B, BPFP=1.6777 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,416B, BPFP=1.3577 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,412B, BPFP=1.5798 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,812B, BPFP=0.0621 +⌛️ [2/4] FRONTEND: Frontend time: 1.843s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11148910 52.83828334 + layer.0.v_cache 0.00001684 0.01320215 + layer.1.k_cache 0.60671574 6.65148709 + layer.1.v_cache 0.00000630 0.00524722 + layer.2.k_cache 0.04149003 1.17422518 + layer.2.v_cache 0.00001976 0.01517221 + layer.3.k_cache 0.02200458 8.90927005 + layer.3.v_cache 0.00001954 0.01864886 + layer.4.k_cache 0.00078895 0.44640663 + layer.4.v_cache 0.00004864 0.03199046 + layer.4.output 0.00472744 197.31097483 + ------------------------------------------------------------------------------------- + TOTAL 0.04798185 85.36945630 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 226940 +BPFP 0.7423 bits/point +EBPFP 1.4846 equivalent bits/point +MSE 85.369456 +---------------------- -------------------------------------------------------- +Time: 3.216s Load: 0.009s, Pack+Encode: 1.843s, Decode+Unpack: 1.364s +---------------------- -------------------------------------------------------- +💾 Converting with 85.3695 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,768B, BPFP=0.2944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,528B, BPFP=1.9163 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,912B, BPFP=0.5400 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,176B, BPFP=1.8887 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,816B, BPFP=1.0012 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,220B, BPFP=1.8141 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,296B, BPFP=1.0388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,572B, BPFP=1.8416 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,456B, BPFP=1.5200 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,840B, BPFP=1.7844 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,660B, BPFP=0.0632 +⌛️ [2/4] FRONTEND: Frontend time: 1.715s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14565937 55.75337402 + layer.0.v_cache 0.00001550 0.01333680 + layer.1.k_cache 0.22559690 6.67132935 + layer.1.v_cache 0.00000590 0.00540322 + layer.2.k_cache 0.00679684 1.31255783 + layer.2.v_cache 0.00002314 0.01662104 + layer.3.k_cache 0.06866968 7.75803650 + layer.3.v_cache 0.00002119 0.02049482 + layer.4.k_cache 0.00065507 0.43779301 + layer.4.v_cache 0.00005191 0.03490817 + layer.4.output 1.53062134 270.74004464 + ------------------------------------------------------------------------------------- + TOTAL 0.65657911 115.71789219 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 180244 +BPFP 0.8283 bits/point +EBPFP 1.6567 equivalent bits/point +MSE 115.717892 +---------------------- -------------------------------------------------------- +Time: 2.940s Load: 0.007s, Pack+Encode: 1.715s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 115.7179 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,748B, BPFP=0.2914 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,704B, BPFP=1.9204 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,380B, BPFP=0.5737 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,804B, BPFP=1.8504 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,064B, BPFP=0.9378 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,108B, BPFP=1.7963 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,060B, BPFP=1.0152 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,616B, BPFP=1.8358 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,104B, BPFP=1.5628 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,804B, BPFP=1.7727 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,792B, BPFP=0.0643 +⌛️ [2/4] FRONTEND: Frontend time: 1.722s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.216s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10663509 52.96477573 + layer.0.v_cache 0.00001613 0.01354918 + layer.1.k_cache 0.26620468 6.16805742 + layer.1.v_cache 0.00000602 0.00547005 + layer.2.k_cache 0.00649704 1.25458796 + layer.2.v_cache 0.00002002 0.01610892 + layer.3.k_cache 0.02565914 8.63955370 + layer.3.v_cache 0.00002186 0.02002202 + layer.4.k_cache 0.00070101 0.45532610 + layer.4.v_cache 0.00005421 0.03512401 + layer.4.output 1.52303164 269.18014837 + ------------------------------------------------------------------------------------- + TOTAL 0.65100216 114.93138904 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 180184 +BPFP 0.8239 bits/point +EBPFP 1.6479 equivalent bits/point +MSE 114.931389 +---------------------- -------------------------------------------------------- +Time: 2.945s Load: 0.007s, Pack+Encode: 1.722s, Decode+Unpack: 1.216s +---------------------- -------------------------------------------------------- +💾 Converting with 114.9314 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,808B, BPFP=0.2814 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,552B, BPFP=1.8464 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,096B, BPFP=0.5323 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,540B, BPFP=1.7872 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,456B, BPFP=0.9045 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,116B, BPFP=1.7039 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,508B, BPFP=0.8490 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,088B, BPFP=1.7608 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,600B, BPFP=1.4396 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,548B, BPFP=1.6706 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,404B, BPFP=0.0619 +⌛️ [2/4] FRONTEND: Frontend time: 1.842s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258070 53.14259276 + layer.0.v_cache 0.00001713 0.01384563 + layer.1.k_cache 0.49703808 6.56449473 + layer.1.v_cache 0.00000670 0.00569278 + layer.2.k_cache 0.01458295 1.23580773 + layer.2.v_cache 0.00002181 0.01592707 + layer.3.k_cache 0.02940617 8.58399169 + layer.3.v_cache 0.00002120 0.01975208 + layer.4.k_cache 0.00070157 0.45693570 + layer.4.v_cache 0.00005095 0.03305634 + layer.4.output 0.00497170 207.78770399 + ------------------------------------------------------------------------------------- + TOTAL 0.04171936 89.68153085 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 225716 +BPFP 0.7770 bits/point +EBPFP 1.5540 equivalent bits/point +MSE 89.681531 +---------------------- -------------------------------------------------------- +Time: 3.200s Load: 0.010s, Pack+Encode: 1.842s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 89.6815 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,000B, BPFP=0.2765 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,196B, BPFP=1.7420 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,404B, BPFP=0.5119 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,392B, BPFP=1.6864 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,408B, BPFP=0.8579 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,644B, BPFP=1.6347 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,352B, BPFP=0.7848 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,604B, BPFP=1.7011 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,644B, BPFP=1.4273 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,036B, BPFP=1.5926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,744B, BPFP=0.0567 +⌛️ [2/4] FRONTEND: Frontend time: 1.725s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.226s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15433978 54.73567132 + layer.0.v_cache 0.00001575 0.01337164 + layer.1.k_cache 0.29362690 6.34214040 + layer.1.v_cache 0.00000607 0.00535227 + layer.2.k_cache 0.00940348 1.40845982 + layer.2.v_cache 0.00002159 0.01643463 + layer.3.k_cache 0.02320663 8.29518006 + layer.3.v_cache 0.00002067 0.02060160 + layer.4.k_cache 0.00067785 0.46079673 + layer.4.v_cache 0.00005016 0.03384711 + layer.4.output 1.35461534 239.43623578 + ------------------------------------------------------------------------------------- + TOTAL 0.58609860 102.78738271 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 182424 +BPFP 0.7419 bits/point +EBPFP 1.4838 equivalent bits/point +MSE 102.787383 +---------------------- -------------------------------------------------------- +Time: 2.958s Load: 0.007s, Pack+Encode: 1.725s, Decode+Unpack: 1.226s +---------------------- -------------------------------------------------------- +💾 Converting with 102.7874 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,028B, BPFP=0.2757 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,840B, BPFP=1.7456 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,396B, BPFP=0.5151 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,312B, BPFP=1.6618 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,488B, BPFP=0.8491 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,312B, BPFP=1.6070 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,108B, BPFP=0.7735 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,216B, BPFP=1.6566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,428B, BPFP=1.3393 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,456B, BPFP=1.5601 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,132B, BPFP=0.0715 +⌛️ [2/4] FRONTEND: Frontend time: 1.863s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.360s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14751465 51.46728002 + layer.0.v_cache 0.00001719 0.01361982 + layer.1.k_cache 0.52245269 6.21094007 + layer.1.v_cache 0.00000618 0.00546757 + layer.2.k_cache 0.02707210 1.36660842 + layer.2.v_cache 0.00002227 0.01613332 + layer.3.k_cache 0.03364062 9.03864703 + layer.3.v_cache 0.00002083 0.01950896 + layer.4.k_cache 0.00069748 0.43342266 + layer.4.v_cache 0.00005305 0.03201421 + layer.4.output 0.00471531 194.23701441 + ------------------------------------------------------------------------------------- + TOTAL 0.04497084 84.01545547 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 227716 +BPFP 0.7344 bits/point +EBPFP 1.4688 equivalent bits/point +MSE 84.015455 +---------------------- -------------------------------------------------------- +Time: 3.232s Load: 0.009s, Pack+Encode: 1.863s, Decode+Unpack: 1.360s +---------------------- -------------------------------------------------------- +💾 Converting with 84.0155 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,548B, BPFP=0.2787 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,072B, BPFP=1.5363 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,440B, BPFP=0.5172 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,176B, BPFP=1.4814 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,400B, BPFP=0.6985 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,084B, BPFP=1.4145 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,748B, BPFP=0.6586 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,816B, BPFP=1.4593 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,060B, BPFP=1.2292 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,428B, BPFP=1.3743 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,332B, BPFP=0.0817 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.227s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17432048 54.70435049 + layer.0.v_cache 0.00001704 0.01222875 + layer.1.k_cache 0.39057378 6.68401118 + layer.1.v_cache 0.00000668 0.00527269 + layer.2.k_cache 0.01299327 1.22524438 + layer.2.v_cache 0.00002114 0.01428933 + layer.3.k_cache 0.02499930 8.80222120 + layer.3.v_cache 0.00002096 0.01775118 + layer.4.k_cache 0.00069121 0.44195976 + layer.4.v_cache 0.00005416 0.03093234 + layer.4.output 1.20070616 211.80322129 + ------------------------------------------------------------------------------------- + TOTAL 0.52992007 91.44475355 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 183104 +BPFP 0.6600 bits/point +EBPFP 1.3200 equivalent bits/point +MSE 91.444754 +---------------------- -------------------------------------------------------- +Time: 2.965s Load: 0.008s, Pack+Encode: 1.730s, Decode+Unpack: 1.227s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4448 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,168B, BPFP=0.2844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,240B, BPFP=1.7222 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,636B, BPFP=0.5210 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,268B, BPFP=1.6558 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,236B, BPFP=0.9031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,340B, BPFP=1.5925 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,024B, BPFP=0.8204 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,228B, BPFP=1.6531 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,200B, BPFP=1.3783 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,732B, BPFP=1.5510 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,236B, BPFP=0.0705 +⌛️ [2/4] FRONTEND: Frontend time: 1.720s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16239141 51.86924724 + layer.0.v_cache 0.00001645 0.01336449 + layer.1.k_cache 0.33751179 6.20402800 + layer.1.v_cache 0.00000668 0.00551332 + layer.2.k_cache 0.00660169 1.32142539 + layer.2.v_cache 0.00002045 0.01562688 + layer.3.k_cache 0.02213682 8.58414696 + layer.3.v_cache 0.00002049 0.01961314 + layer.4.k_cache 0.00067824 0.45214717 + layer.4.v_cache 0.00005183 0.03301291 + layer.4.output 1.33690934 236.42024719 + ------------------------------------------------------------------------------------- + TOTAL 0.58163537 101.37999152 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 184308 +BPFP 0.7397 bits/point +EBPFP 1.4795 equivalent bits/point +MSE 101.379992 +---------------------- -------------------------------------------------------- +Time: 2.948s Load: 0.008s, Pack+Encode: 1.720s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 101.3800 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,892B, BPFP=0.2852 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,688B, BPFP=1.7892 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,064B, BPFP=0.5285 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,956B, BPFP=1.7465 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,316B, BPFP=0.9513 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,540B, BPFP=1.6639 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,640B, BPFP=0.7952 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,080B, BPFP=1.6954 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,272B, BPFP=1.4151 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,072B, BPFP=1.6367 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,136B, BPFP=0.0594 +⌛️ [2/4] FRONTEND: Frontend time: 1.867s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15796844 54.44059731 + layer.0.v_cache 0.00001527 0.01283331 + layer.1.k_cache 0.50235566 6.02688690 + layer.1.v_cache 0.00000626 0.00533114 + layer.2.k_cache 0.02709456 1.17564392 + layer.2.v_cache 0.00002063 0.01460012 + layer.3.k_cache 0.00856834 7.77489870 + layer.3.v_cache 0.00001945 0.01798660 + layer.4.k_cache 0.00071483 0.44875176 + layer.4.v_cache 0.00005464 0.03290282 + layer.4.output 0.00491706 207.01325959 + ------------------------------------------------------------------------------------- + TOTAL 0.04301397 89.35548528 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 221656 +BPFP 0.7602 bits/point +EBPFP 1.5204 equivalent bits/point +MSE 89.355485 +---------------------- -------------------------------------------------------- +Time: 3.228s Load: 0.010s, Pack+Encode: 1.867s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 89.3555 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,408B, BPFP=0.2777 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,828B, BPFP=1.6273 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,472B, BPFP=0.5338 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,708B, BPFP=1.5567 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,416B, BPFP=0.7823 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,684B, BPFP=1.4922 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,796B, BPFP=0.7432 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,580B, BPFP=1.5486 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,604B, BPFP=1.2981 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,996B, BPFP=1.4488 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,456B, BPFP=0.0761 +⌛️ [2/4] FRONTEND: Frontend time: 1.725s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13184587 53.42309964 + layer.0.v_cache 0.00001706 0.01392030 + layer.1.k_cache 0.45546043 6.68755119 + layer.1.v_cache 0.00000678 0.00588114 + layer.2.k_cache 0.02272066 1.27228079 + layer.2.v_cache 0.00001996 0.01619871 + layer.3.k_cache 0.01643563 8.27880367 + layer.3.v_cache 0.00002069 0.02040315 + layer.4.k_cache 0.00071781 0.44466419 + layer.4.v_cache 0.00005262 0.03381285 + layer.4.output 1.23455937 218.24103543 + ------------------------------------------------------------------------------------- + TOTAL 0.54524783 93.99316845 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 187948 +BPFP 0.6966 bits/point +EBPFP 1.3931 equivalent bits/point +MSE 93.993168 +---------------------- -------------------------------------------------------- +Time: 2.958s Load: 0.010s, Pack+Encode: 1.725s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 93.9932 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,792B, BPFP=0.2836 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,876B, BPFP=1.8274 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,080B, BPFP=0.5374 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,684B, BPFP=1.7569 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,352B, BPFP=0.8494 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,508B, BPFP=1.6873 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,104B, BPFP=0.8348 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,384B, BPFP=1.7391 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,716B, BPFP=1.4036 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,040B, BPFP=1.6596 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,212B, BPFP=0.0694 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17705467 53.78701320 + layer.0.v_cache 0.00001633 0.01300522 + layer.1.k_cache 0.50406676 6.30889661 + layer.1.v_cache 0.00000599 0.00518302 + layer.2.k_cache 0.00944760 1.17450818 + layer.2.v_cache 0.00001883 0.01475345 + layer.3.k_cache 0.04120483 8.20822421 + layer.3.v_cache 0.00002392 0.01949924 + layer.4.k_cache 0.00070100 0.42896487 + layer.4.v_cache 0.00005002 0.03232579 + layer.4.output 0.00499618 210.22793222 + ------------------------------------------------------------------------------------- + TOTAL 0.04515078 90.68164114 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 220748 +BPFP 0.7685 bits/point +EBPFP 1.5371 equivalent bits/point +MSE 90.681641 +---------------------- -------------------------------------------------------- +Time: 3.213s Load: 0.011s, Pack+Encode: 1.855s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 90.6816 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,908B, BPFP=0.2840 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,984B, BPFP=1.7931 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,064B, BPFP=0.5245 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,980B, BPFP=1.7350 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,068B, BPFP=0.7562 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,660B, BPFP=1.6586 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,544B, BPFP=0.7838 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,528B, BPFP=1.7088 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,328B, BPFP=1.4079 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,184B, BPFP=1.6310 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,720B, BPFP=0.0638 +⌛️ [2/4] FRONTEND: Frontend time: 1.862s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20501094 53.78710938 + layer.0.v_cache 0.00001516 0.01251063 + layer.1.k_cache 0.57085758 6.03239384 + layer.1.v_cache 0.00000574 0.00499122 + layer.2.k_cache 0.01259345 1.19368783 + layer.2.v_cache 0.00002099 0.01485097 + layer.3.k_cache 0.01648921 8.55895815 + layer.3.v_cache 0.00001970 0.01824247 + layer.4.k_cache 0.00070444 0.41718035 + layer.4.v_cache 0.00005121 0.03262412 + layer.4.output 0.00487421 205.31195437 + ------------------------------------------------------------------------------------- + TOTAL 0.04940517 88.66213115 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 219968 +BPFP 0.7488 bits/point +EBPFP 1.4976 equivalent bits/point +MSE 88.662131 +---------------------- -------------------------------------------------------- +Time: 3.221s Load: 0.010s, Pack+Encode: 1.862s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6621 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,928B, BPFP=0.2841 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,432B, BPFP=1.8123 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,076B, BPFP=0.5233 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,324B, BPFP=1.7484 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,404B, BPFP=0.8305 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,332B, BPFP=1.6912 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,772B, BPFP=0.7940 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,124B, BPFP=1.7369 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,592B, BPFP=1.4179 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,396B, BPFP=1.6372 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,900B, BPFP=0.0651 +⌛️ [2/4] FRONTEND: Frontend time: 1.888s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14466163 50.84688869 + layer.0.v_cache 0.00001631 0.01339210 + layer.1.k_cache 0.53789613 6.76943305 + layer.1.v_cache 0.00000653 0.00548861 + layer.2.k_cache 0.02095415 1.21840305 + layer.2.v_cache 0.00002169 0.01564370 + layer.3.k_cache 0.03083320 8.56013157 + layer.3.v_cache 0.00002083 0.01895623 + layer.4.k_cache 0.00069737 0.44375613 + layer.4.v_cache 0.00004943 0.03216982 + layer.4.output 0.00490546 204.46015090 + ------------------------------------------------------------------------------------- + TOTAL 0.04526444 88.18501878 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 224280 +BPFP 0.7607 bits/point +EBPFP 1.5213 equivalent bits/point +MSE 88.185019 +---------------------- -------------------------------------------------------- +Time: 3.269s Load: 0.011s, Pack+Encode: 1.888s, Decode+Unpack: 1.370s +---------------------- -------------------------------------------------------- +💾 Converting with 88.1850 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,340B, BPFP=0.2791 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,384B, BPFP=1.6401 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,856B, BPFP=0.5151 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,036B, BPFP=1.5696 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,848B, BPFP=0.8282 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,664B, BPFP=1.4979 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,928B, BPFP=0.7801 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,600B, BPFP=1.5468 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,672B, BPFP=1.2893 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,024B, BPFP=1.4645 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,580B, BPFP=0.0641 +⌛️ [2/4] FRONTEND: Frontend time: 1.866s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14240686 52.86025685 + layer.0.v_cache 0.00001673 0.01262099 + layer.1.k_cache 0.61438330 6.54877178 + layer.1.v_cache 0.00000638 0.00522686 + layer.2.k_cache 0.02027981 1.16323373 + layer.2.v_cache 0.00002012 0.01520752 + layer.3.k_cache 0.02199794 8.49408184 + layer.3.v_cache 0.00002003 0.01822323 + layer.4.k_cache 0.00078937 0.43769872 + layer.4.v_cache 0.00005038 0.03207735 + layer.4.output 0.04464151 181.32272754 + ------------------------------------------------------------------------------------- + TOTAL 0.06543891 78.75567598 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 226932 +BPFP 0.6976 bits/point +EBPFP 1.3952 equivalent bits/point +MSE 78.755676 +---------------------- -------------------------------------------------------- +Time: 3.222s Load: 0.010s, Pack+Encode: 1.866s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 78.7557 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 399, 128) +Output shape: (1, 399, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.output: torch.Size([1, 399, 3584]) -> torch.Size([1, 1, 399, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,984B, BPFP=0.2735 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,612B, BPFP=1.7079 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,128B, BPFP=0.5141 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,668B, BPFP=1.6317 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,240B, BPFP=0.7926 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,716B, BPFP=1.5945 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,096B, BPFP=0.7086 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,792B, BPFP=1.6366 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,672B, BPFP=1.3578 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,568B, BPFP=1.5495 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,104B, BPFP=0.0621 +⌛️ [2/4] FRONTEND: Frontend time: 2.127s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.598s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17889012 53.36585605 + layer.0.v_cache 0.00001635 0.01256723 + layer.1.k_cache 0.94984295 6.37787278 + layer.1.v_cache 0.00000590 0.00500657 + layer.2.k_cache 0.02305104 1.26617638 + layer.2.v_cache 0.00002114 0.01588569 + layer.3.k_cache 0.01620252 8.64804222 + layer.3.v_cache 0.00002049 0.01908649 + layer.4.k_cache 0.00073175 0.43108137 + layer.4.v_cache 0.00005255 0.03189582 + layer.4.output 0.00656733 138.71130952 + ------------------------------------------------------------------------------------- + TOTAL 0.07145918 61.24427278 + (elements=3,472,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3472896 +Total Bytes 311580 +BPFP 0.7177 bits/point +EBPFP 1.4355 equivalent bits/point +MSE 61.244273 +---------------------- -------------------------------------------------------- +Time: 3.739s Load: 0.015s, Pack+Encode: 2.127s, Decode+Unpack: 1.598s +---------------------- -------------------------------------------------------- +💾 Converting with 61.2443 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,012B, BPFP=0.2774 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,420B, BPFP=1.7575 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,372B, BPFP=0.5097 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,296B, BPFP=1.6798 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,336B, BPFP=0.8529 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,888B, BPFP=1.6515 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,248B, BPFP=0.7777 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,472B, BPFP=1.6919 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,316B, BPFP=1.4046 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,164B, BPFP=1.6015 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,276B, BPFP=0.0620 +⌛️ [2/4] FRONTEND: Frontend time: 1.712s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10670873 54.59191354 + layer.0.v_cache 0.00001672 0.01405788 + layer.1.k_cache 0.29393762 6.54000881 + layer.1.v_cache 0.00000627 0.00553905 + layer.2.k_cache 0.01243114 1.35355411 + layer.2.v_cache 0.00002105 0.01663320 + layer.3.k_cache 0.00919804 9.24859241 + layer.3.v_cache 0.00002002 0.02014667 + layer.4.k_cache 0.00070102 0.46224415 + layer.4.v_cache 0.00005024 0.03460299 + layer.4.output 1.35463108 239.49539744 + ------------------------------------------------------------------------------------- + TOTAL 0.58267697 102.86794558 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 182800 +BPFP 0.7434 bits/point +EBPFP 1.4869 equivalent bits/point +MSE 102.867946 +---------------------- -------------------------------------------------------- +Time: 2.946s Load: 0.008s, Pack+Encode: 1.712s, Decode+Unpack: 1.225s +---------------------- -------------------------------------------------------- +💾 Converting with 102.8679 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,476B, BPFP=0.2753 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,504B, BPFP=1.5689 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,336B, BPFP=0.5128 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,124B, BPFP=1.4840 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,584B, BPFP=0.7126 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,188B, BPFP=1.4264 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,284B, BPFP=0.6941 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,076B, BPFP=1.4811 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,332B, BPFP=1.2507 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,680B, BPFP=1.3952 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,664B, BPFP=0.0674 +⌛️ [2/4] FRONTEND: Frontend time: 1.727s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633087 53.97700080 + layer.0.v_cache 0.00001891 0.01339039 + layer.1.k_cache 0.36012592 6.71629177 + layer.1.v_cache 0.00000581 0.00512597 + layer.2.k_cache 0.01407751 1.38188424 + layer.2.v_cache 0.00002023 0.01608027 + layer.3.k_cache 0.01230825 9.47401921 + layer.3.v_cache 0.00002000 0.01944876 + layer.4.k_cache 0.00071033 0.46412689 + layer.4.v_cache 0.00004856 0.03227682 + layer.4.output 1.20534739 212.94813344 + ------------------------------------------------------------------------------------- + TOTAL 0.52535871 91.92568113 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 183248 +BPFP 0.6631 bits/point +EBPFP 1.3262 equivalent bits/point +MSE 91.925681 +---------------------- -------------------------------------------------------- +Time: 2.963s Load: 0.011s, Pack+Encode: 1.727s, Decode+Unpack: 1.225s +---------------------- -------------------------------------------------------- +💾 Converting with 91.9257 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,340B, BPFP=0.2814 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,268B, BPFP=1.6382 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,112B, BPFP=0.5259 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,352B, BPFP=1.5788 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,880B, BPFP=0.7702 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,520B, BPFP=1.5249 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,600B, BPFP=0.7521 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,924B, BPFP=1.5511 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,736B, BPFP=1.3444 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,028B, BPFP=1.4930 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,764B, BPFP=0.0626 +⌛️ [2/4] FRONTEND: Frontend time: 1.723s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.229s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11142051 53.55579373 + layer.0.v_cache 0.00001650 0.01287112 + layer.1.k_cache 0.42032433 6.26068748 + layer.1.v_cache 0.00000619 0.00522302 + layer.2.k_cache 0.01443043 1.23933050 + layer.2.v_cache 0.00002030 0.01560296 + layer.3.k_cache 0.02895402 9.01947249 + layer.3.v_cache 0.00001914 0.01811285 + layer.4.k_cache 0.00067703 0.43324862 + layer.4.v_cache 0.00005693 0.03351124 + layer.4.output 1.27033357 224.33350622 + ------------------------------------------------------------------------------------- + TOTAL 0.55695649 96.52519986 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 183524 +BPFP 0.6999 bits/point +EBPFP 1.3998 equivalent bits/point +MSE 96.525200 +---------------------- -------------------------------------------------------- +Time: 2.963s Load: 0.010s, Pack+Encode: 1.723s, Decode+Unpack: 1.229s +---------------------- -------------------------------------------------------- +💾 Converting with 96.5252 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 302, 128) +Output shape: (1, 302, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.output: torch.Size([1, 302, 3584]) -> torch.Size([1, 1, 302, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,340B, BPFP=0.2763 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,248B, BPFP=1.6167 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,900B, BPFP=0.5122 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,960B, BPFP=1.5501 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,640B, BPFP=0.7575 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,008B, BPFP=1.5008 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,152B, BPFP=0.7322 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,760B, BPFP=1.5397 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,120B, BPFP=1.2997 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,276B, BPFP=1.4630 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,876B, BPFP=0.0508 +⌛️ [2/4] FRONTEND: Frontend time: 1.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15012774 51.64829845 + layer.0.v_cache 0.00001504 0.01219024 + layer.1.k_cache 0.60100308 6.23891909 + layer.1.v_cache 0.00000575 0.00491524 + layer.2.k_cache 0.00814273 1.42399163 + layer.2.v_cache 0.00002032 0.01543493 + layer.3.k_cache 0.01797006 8.82901587 + layer.3.v_cache 0.00001899 0.01800661 + layer.4.k_cache 0.00070977 0.43632831 + layer.4.v_cache 0.00004868 0.03120000 + layer.4.output 0.04414053 179.41619856 + ------------------------------------------------------------------------------------- + TOTAL 0.06394387 77.91598178 + (elements=2,628,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2628608 +Total Bytes 224280 +BPFP 0.6826 bits/point +EBPFP 1.3652 equivalent bits/point +MSE 77.915982 +---------------------- -------------------------------------------------------- +Time: 3.210s Load: 0.012s, Pack+Encode: 1.847s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 77.9160 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,848B, BPFP=0.2826 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,664B, BPFP=1.7878 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,352B, BPFP=0.5452 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,516B, BPFP=1.7208 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,096B, BPFP=0.8801 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,748B, BPFP=1.6761 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,212B, BPFP=0.8869 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,416B, BPFP=1.7150 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,880B, BPFP=1.3923 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,416B, BPFP=1.6567 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,936B, BPFP=0.0744 +⌛️ [2/4] FRONTEND: Frontend time: 1.845s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14350806 54.66628527 + layer.0.v_cache 0.00001655 0.01268957 + layer.1.k_cache 0.46812200 6.30278858 + layer.1.v_cache 0.00000598 0.00515307 + layer.2.k_cache 0.01096032 1.19912868 + layer.2.v_cache 0.00002019 0.01550865 + layer.3.k_cache 0.00658789 8.47143646 + layer.3.v_cache 0.00002014 0.01849020 + layer.4.k_cache 0.00067528 0.43039470 + layer.4.v_cache 0.00005785 0.03373167 + layer.4.output 0.00496594 206.92420709 + ------------------------------------------------------------------------------------- + TOTAL 0.03910211 89.38970920 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 224084 +BPFP 0.7685 bits/point +EBPFP 1.5370 equivalent bits/point +MSE 89.389709 +---------------------- -------------------------------------------------------- +Time: 3.205s Load: 0.010s, Pack+Encode: 1.845s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 89.3897 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 260, 128) +Output shape: (1, 260, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.output: torch.Size([1, 260, 3584]) -> torch.Size([1, 1, 260, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,648B, BPFP=0.2793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,684B, BPFP=1.8440 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,760B, BPFP=0.5264 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,792B, BPFP=1.7904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,168B, BPFP=0.9115 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,552B, BPFP=1.7159 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,852B, BPFP=0.8925 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,200B, BPFP=1.7548 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,364B, BPFP=1.4642 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,204B, BPFP=1.6950 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,780B, BPFP=0.0582 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14548651 53.11329252 + layer.0.v_cache 0.00001584 0.01307587 + layer.1.k_cache 0.47062912 6.25877592 + layer.1.v_cache 0.00000604 0.00521544 + layer.2.k_cache 0.00619981 1.36130442 + layer.2.v_cache 0.00002003 0.01588860 + layer.3.k_cache 0.01299995 8.63893010 + layer.3.v_cache 0.00001981 0.01865228 + layer.4.k_cache 0.00067777 0.44772685 + layer.4.v_cache 0.00006192 0.03335735 + layer.4.output 0.00504520 213.42005495 + ------------------------------------------------------------------------------------- + TOTAL 0.03949607 91.99097670 + (elements=2,263,040) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2263040 +Total Bytes 221004 +BPFP 0.7813 bits/point +EBPFP 1.5625 equivalent bits/point +MSE 91.990977 +---------------------- -------------------------------------------------------- +Time: 3.215s Load: 0.011s, Pack+Encode: 1.855s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 91.9910 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,908B, BPFP=0.2819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,488B, BPFP=1.8088 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,236B, BPFP=0.5306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,276B, BPFP=1.7392 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,912B, BPFP=0.8566 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,168B, BPFP=1.6756 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,848B, BPFP=0.8529 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,028B, BPFP=1.7250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,956B, BPFP=1.3761 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,448B, BPFP=1.6342 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,188B, BPFP=0.0672 +⌛️ [2/4] FRONTEND: Frontend time: 1.873s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14555165 51.65691421 + layer.0.v_cache 0.00001661 0.01337332 + layer.1.k_cache 0.55468144 5.92580907 + layer.1.v_cache 0.00000621 0.00530251 + layer.2.k_cache 0.02243149 1.32795480 + layer.2.v_cache 0.00002103 0.01547397 + layer.3.k_cache 0.02028967 8.66458040 + layer.3.v_cache 0.00002036 0.01901591 + layer.4.k_cache 0.00072975 0.44488985 + layer.4.v_cache 0.00005124 0.03289077 + layer.4.output 0.00486695 203.88606224 + ------------------------------------------------------------------------------------- + TOTAL 0.04575695 87.95933179 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 225456 +BPFP 0.7618 bits/point +EBPFP 1.5237 equivalent bits/point +MSE 87.959332 +---------------------- -------------------------------------------------------- +Time: 3.233s Load: 0.009s, Pack+Encode: 1.873s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 87.9593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 356, 128) +Output shape: (1, 356, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.output: torch.Size([1, 356, 3584]) -> torch.Size([1, 1, 356, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,180B, BPFP=0.2712 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,800B, BPFP=1.6591 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,668B, BPFP=0.5121 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,236B, BPFP=1.5904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,340B, BPFP=0.8050 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,596B, BPFP=1.5184 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,216B, BPFP=0.7117 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,452B, BPFP=1.5560 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,308B, BPFP=1.2863 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,388B, BPFP=1.4654 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,700B, BPFP=0.0734 +⌛️ [2/4] FRONTEND: Frontend time: 1.971s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.471s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15223843 52.26374868 + layer.0.v_cache 0.00001740 0.01270551 + layer.1.k_cache 0.82972649 6.20476266 + layer.1.v_cache 0.00000631 0.00522958 + layer.2.k_cache 0.03086275 1.18329740 + layer.2.v_cache 0.00001960 0.01503291 + layer.3.k_cache 0.01100341 8.39414789 + layer.3.v_cache 0.00002062 0.01824603 + layer.4.k_cache 0.00073857 0.43559634 + layer.4.v_cache 0.00005279 0.03145019 + layer.4.output 0.03758925 152.34787570 + ------------------------------------------------------------------------------------- + TOTAL 0.07575360 66.76466748 + (elements=3,098,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3098624 +Total Bytes 270884 +BPFP 0.6994 bits/point +EBPFP 1.3987 equivalent bits/point +MSE 66.764667 +---------------------- -------------------------------------------------------- +Time: 3.455s Load: 0.012s, Pack+Encode: 1.971s, Decode+Unpack: 1.471s +---------------------- -------------------------------------------------------- +💾 Converting with 66.7647 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,852B, BPFP=0.2850 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,120B, BPFP=1.8280 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,504B, BPFP=0.5583 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,924B, BPFP=1.7578 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,456B, BPFP=0.8492 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,912B, BPFP=1.6983 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,580B, BPFP=0.8564 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,692B, BPFP=1.7441 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,436B, BPFP=1.4354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,368B, BPFP=1.6664 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,824B, BPFP=0.0740 +⌛️ [2/4] FRONTEND: Frontend time: 1.880s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12347319 53.36495903 + layer.0.v_cache 0.00001696 0.01316924 + layer.1.k_cache 0.44284580 6.46161485 + layer.1.v_cache 0.00000660 0.00551489 + layer.2.k_cache 0.00882101 1.34085668 + layer.2.v_cache 0.00002210 0.01641861 + layer.3.k_cache 0.01145097 8.69196245 + layer.3.v_cache 0.00001984 0.01881971 + layer.4.k_cache 0.00068754 0.43652272 + layer.4.v_cache 0.00006483 0.03252660 + layer.4.output 0.00500964 208.49642521 + ------------------------------------------------------------------------------------- + TOTAL 0.03661626 89.99160831 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 224668 +BPFP 0.7763 bits/point +EBPFP 1.5526 equivalent bits/point +MSE 89.991608 +---------------------- -------------------------------------------------------- +Time: 3.262s Load: 0.010s, Pack+Encode: 1.880s, Decode+Unpack: 1.371s +---------------------- -------------------------------------------------------- +💾 Converting with 89.9916 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,020B, BPFP=0.2752 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,640B, BPFP=1.7346 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,408B, BPFP=0.5158 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,240B, BPFP=1.6579 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,932B, BPFP=0.8735 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,428B, BPFP=1.6134 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,996B, BPFP=0.7673 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,228B, BPFP=1.6572 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,348B, BPFP=1.3897 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,732B, BPFP=1.5752 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,420B, BPFP=0.0581 +⌛️ [2/4] FRONTEND: Frontend time: 1.865s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.352s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13002123 53.26580318 + layer.0.v_cache 0.00001833 0.01367253 + layer.1.k_cache 0.50397591 6.22092199 + layer.1.v_cache 0.00000608 0.00529897 + layer.2.k_cache 0.01926375 1.38059542 + layer.2.v_cache 0.00002060 0.01563719 + layer.3.k_cache 0.04992905 8.72707048 + layer.3.v_cache 0.00002096 0.01980819 + layer.4.k_cache 0.00069540 0.44539452 + layer.4.v_cache 0.00005467 0.03452504 + layer.4.output 0.00466491 194.44072682 + ------------------------------------------------------------------------------------- + TOTAL 0.04333296 84.18904796 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 227392 +BPFP 0.7333 bits/point +EBPFP 1.4667 equivalent bits/point +MSE 84.189048 +---------------------- -------------------------------------------------------- +Time: 3.228s Load: 0.011s, Pack+Encode: 1.865s, Decode+Unpack: 1.352s +---------------------- -------------------------------------------------------- +💾 Converting with 84.1890 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,988B, BPFP=0.2783 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,380B, BPFP=1.7511 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,364B, BPFP=0.5225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,072B, BPFP=1.6781 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,984B, BPFP=0.8362 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,124B, BPFP=1.6252 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,776B, BPFP=0.8246 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,012B, BPFP=1.6748 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,612B, BPFP=1.3176 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,588B, BPFP=1.5953 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,420B, BPFP=0.0831 +⌛️ [2/4] FRONTEND: Frontend time: 1.886s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15974760 51.60551409 + layer.0.v_cache 0.00001739 0.01295365 + layer.1.k_cache 0.45644542 6.08224923 + layer.1.v_cache 0.00000643 0.00527784 + layer.2.k_cache 0.01631587 1.16185041 + layer.2.v_cache 0.00002005 0.01494414 + layer.3.k_cache 0.01511352 8.46050328 + layer.3.v_cache 0.00002063 0.01896448 + layer.4.k_cache 0.00073517 0.43285452 + layer.4.v_cache 0.00005091 0.03294744 + layer.4.output 0.00481773 198.10184949 + ------------------------------------------------------------------------------------- + TOTAL 0.04012924 85.56123562 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 227320 +BPFP 0.7462 bits/point +EBPFP 1.4924 equivalent bits/point +MSE 85.561236 +---------------------- -------------------------------------------------------- +Time: 3.272s Load: 0.010s, Pack+Encode: 1.886s, Decode+Unpack: 1.375s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5612 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,860B, BPFP=0.2844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,876B, BPFP=1.8069 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,128B, BPFP=0.5342 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,716B, BPFP=1.7390 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,120B, BPFP=0.9434 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,552B, BPFP=1.6709 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,320B, BPFP=0.7210 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,448B, BPFP=1.7233 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,064B, BPFP=1.4082 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,196B, BPFP=1.6500 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,936B, BPFP=0.0580 +⌛️ [2/4] FRONTEND: Frontend time: 1.876s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14148096 53.85016532 + layer.0.v_cache 0.00001588 0.01278504 + layer.1.k_cache 0.55266568 6.29865668 + layer.1.v_cache 0.00000632 0.00526932 + layer.2.k_cache 0.00687984 1.27357054 + layer.2.v_cache 0.00001885 0.01498619 + layer.3.k_cache 0.01675770 7.91408627 + layer.3.v_cache 0.00001964 0.01846128 + layer.4.k_cache 0.00068991 0.42551388 + layer.4.v_cache 0.00004859 0.03248628 + layer.4.output 0.00492290 207.82198034 + ------------------------------------------------------------------------------------- + TOTAL 0.04429669 89.68234372 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 220216 +BPFP 0.7581 bits/point +EBPFP 1.5161 equivalent bits/point +MSE 89.682344 +---------------------- -------------------------------------------------------- +Time: 3.234s Load: 0.009s, Pack+Encode: 1.876s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 89.6823 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,740B, BPFP=0.2777 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,332B, BPFP=1.8059 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,740B, BPFP=0.5195 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,516B, BPFP=1.7181 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,836B, BPFP=0.8144 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,224B, BPFP=1.6556 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,152B, BPFP=0.7330 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,040B, BPFP=1.6950 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,140B, BPFP=1.3613 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,228B, BPFP=1.6074 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,224B, BPFP=0.0637 +⌛️ [2/4] FRONTEND: Frontend time: 1.969s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.467s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13144862 52.42081076 + layer.0.v_cache 0.00001769 0.01344475 + layer.1.k_cache 0.64513603 6.57035936 + layer.1.v_cache 0.00000607 0.00521146 + layer.2.k_cache 0.02726801 1.33364519 + layer.2.v_cache 0.00001956 0.01516945 + layer.3.k_cache 0.02271994 7.94798973 + layer.3.v_cache 0.00002018 0.01810606 + layer.4.k_cache 0.00073057 0.44159956 + layer.4.v_cache 0.00004940 0.03077182 + layer.4.output 0.04135688 167.86590557 + ------------------------------------------------------------------------------------- + TOTAL 0.06570084 73.16814395 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 261172 +BPFP 0.7432 bits/point +EBPFP 1.4864 equivalent bits/point +MSE 73.168144 +---------------------- -------------------------------------------------------- +Time: 3.447s Load: 0.011s, Pack+Encode: 1.969s, Decode+Unpack: 1.467s +---------------------- -------------------------------------------------------- +💾 Converting with 73.1681 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,072B, BPFP=0.2959 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,204B, BPFP=1.8317 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,316B, BPFP=0.5317 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,228B, BPFP=1.7608 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,396B, BPFP=0.9009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,556B, BPFP=1.7119 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,460B, BPFP=0.8328 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,996B, BPFP=1.7439 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,128B, BPFP=1.4628 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,032B, BPFP=1.6738 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,512B, BPFP=0.0572 +⌛️ [2/4] FRONTEND: Frontend time: 1.720s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12396691 52.71569767 + layer.0.v_cache 0.00001598 0.01360092 + layer.1.k_cache 0.24854917 6.10467501 + layer.1.v_cache 0.00000584 0.00528019 + layer.2.k_cache 0.02290044 1.33483731 + layer.2.v_cache 0.00002299 0.01572900 + layer.3.k_cache 0.01294550 8.26202535 + layer.3.v_cache 0.00002077 0.01927722 + layer.4.k_cache 0.00065144 0.42817163 + layer.4.v_cache 0.00005054 0.03273229 + layer.4.output 1.42385976 251.72634967 + ------------------------------------------------------------------------------------- + TOTAL 0.61036164 107.70685143 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 180900 +BPFP 0.7733 bits/point +EBPFP 1.5467 equivalent bits/point +MSE 107.706851 +---------------------- -------------------------------------------------------- +Time: 2.950s Load: 0.008s, Pack+Encode: 1.720s, Decode+Unpack: 1.222s +---------------------- -------------------------------------------------------- +💾 Converting with 107.7069 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,040B, BPFP=0.2725 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,704B, BPFP=1.7141 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,512B, BPFP=0.5143 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,340B, BPFP=1.6404 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,436B, BPFP=0.7805 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,312B, BPFP=1.5848 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,840B, BPFP=0.7483 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,060B, BPFP=1.6252 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,976B, BPFP=1.3503 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,420B, BPFP=1.5365 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,576B, BPFP=0.0585 +⌛️ [2/4] FRONTEND: Frontend time: 1.854s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14783513 53.44790090 + layer.0.v_cache 0.00001600 0.01335938 + layer.1.k_cache 0.54442752 5.93801743 + layer.1.v_cache 0.00000628 0.00529718 + layer.2.k_cache 0.01848349 1.36305538 + layer.2.v_cache 0.00002061 0.01598946 + layer.3.k_cache 0.01777397 8.92281017 + layer.3.v_cache 0.00002092 0.01991151 + layer.4.k_cache 0.00070447 0.44859045 + layer.4.v_cache 0.00005049 0.03316407 + layer.4.output 0.04615090 187.49105598 + ------------------------------------------------------------------------------------- + TOTAL 0.06190560 81.33208752 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 225216 +BPFP 0.7163 bits/point +EBPFP 1.4325 equivalent bits/point +MSE 81.332088 +---------------------- -------------------------------------------------------- +Time: 3.216s Load: 0.011s, Pack+Encode: 1.854s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 81.3321 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,260B, BPFP=0.2758 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,740B, BPFP=1.6642 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,948B, BPFP=0.5216 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,236B, BPFP=1.5854 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,464B, BPFP=0.7584 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,980B, BPFP=1.5195 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,420B, BPFP=0.7561 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,996B, BPFP=1.5728 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,572B, BPFP=1.2884 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,172B, BPFP=1.4771 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,636B, BPFP=0.0647 +⌛️ [2/4] FRONTEND: Frontend time: 1.846s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16274841 52.66766071 + layer.0.v_cache 0.00001794 0.01312919 + layer.1.k_cache 0.56737488 6.39474098 + layer.1.v_cache 0.00000586 0.00501629 + layer.2.k_cache 0.02027899 1.32445019 + layer.2.v_cache 0.00002117 0.01543310 + layer.3.k_cache 0.01891868 9.34433332 + layer.3.v_cache 0.00002140 0.01853003 + layer.4.k_cache 0.00069658 0.43425049 + layer.4.v_cache 0.00005005 0.03115580 + layer.4.output 0.04476916 181.71250300 + ------------------------------------------------------------------------------------- + TOTAL 0.06373636 78.95507183 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 226424 +BPFP 0.6984 bits/point +EBPFP 1.3967 equivalent bits/point +MSE 78.955072 +---------------------- -------------------------------------------------------- +Time: 3.203s Load: 0.010s, Pack+Encode: 1.846s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 78.9551 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,948B, BPFP=0.2832 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,544B, BPFP=1.8054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,076B, BPFP=0.5195 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,068B, BPFP=1.7209 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,748B, BPFP=0.8441 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,912B, BPFP=1.6548 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,428B, BPFP=0.7685 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,664B, BPFP=1.6978 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,124B, BPFP=1.3807 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,336B, BPFP=1.6218 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,044B, BPFP=0.0658 +⌛️ [2/4] FRONTEND: Frontend time: 1.861s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.352s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13366390 52.89757183 + layer.0.v_cache 0.00001719 0.01340294 + layer.1.k_cache 0.57618574 6.81029916 + layer.1.v_cache 0.00000658 0.00519232 + layer.2.k_cache 0.02858028 1.32735010 + layer.2.v_cache 0.00002021 0.01509389 + layer.3.k_cache 0.00873641 7.48594984 + layer.3.v_cache 0.00002023 0.01855505 + layer.4.k_cache 0.00067337 0.42846096 + layer.4.v_cache 0.00005011 0.03174165 + layer.4.output 0.00484872 202.78975340 + ------------------------------------------------------------------------------------- + TOTAL 0.04599383 87.56246421 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 222892 +BPFP 0.7504 bits/point +EBPFP 1.5008 equivalent bits/point +MSE 87.562464 +---------------------- -------------------------------------------------------- +Time: 3.224s Load: 0.011s, Pack+Encode: 1.861s, Decode+Unpack: 1.352s +---------------------- -------------------------------------------------------- +💾 Converting with 87.5625 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,972B, BPFP=0.2808 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,520B, BPFP=1.8043 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,408B, BPFP=0.5238 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,328B, BPFP=1.7200 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,896B, BPFP=0.9118 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,708B, BPFP=1.6762 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,696B, BPFP=0.8269 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,444B, BPFP=1.7282 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,372B, BPFP=1.4403 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,284B, BPFP=1.6462 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,900B, BPFP=0.0697 +⌛️ [2/4] FRONTEND: Frontend time: 1.719s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13057878 53.67667827 + layer.0.v_cache 0.00001797 0.01451365 + layer.1.k_cache 0.31658528 6.97811793 + layer.1.v_cache 0.00000647 0.00556036 + layer.2.k_cache 0.02849985 1.41737221 + layer.2.v_cache 0.00001991 0.01607262 + layer.3.k_cache 0.03599559 8.66999479 + layer.3.v_cache 0.00002239 0.02089998 + layer.4.k_cache 0.00078811 0.46753765 + layer.4.v_cache 0.00005540 0.03483733 + layer.4.output 1.38530015 244.63760504 + ------------------------------------------------------------------------------------- + TOTAL 0.60056887 104.92734236 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 184528 +BPFP 0.7674 bits/point +EBPFP 1.5349 equivalent bits/point +MSE 104.927342 +---------------------- -------------------------------------------------------- +Time: 2.948s Load: 0.008s, Pack+Encode: 1.719s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 104.9273 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,972B, BPFP=0.2860 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,488B, BPFP=1.8353 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,712B, BPFP=0.5553 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,408B, BPFP=1.7575 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,092B, BPFP=0.8707 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,628B, BPFP=1.7013 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,012B, BPFP=0.8649 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,532B, BPFP=1.7664 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,564B, BPFP=1.4807 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,324B, BPFP=1.6794 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,972B, BPFP=0.0717 +⌛️ [2/4] FRONTEND: Frontend time: 1.719s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12725546 51.52138087 + layer.0.v_cache 0.00001736 0.01424242 + layer.1.k_cache 0.32262681 6.30001170 + layer.1.v_cache 0.00000608 0.00555790 + layer.2.k_cache 0.02157613 1.38371495 + layer.2.v_cache 0.00002208 0.01650800 + layer.3.k_cache 0.01760322 8.40403248 + layer.3.v_cache 0.00002192 0.02075265 + layer.4.k_cache 0.00069785 0.45476753 + layer.4.v_cache 0.00005172 0.03393868 + layer.4.output 1.41081570 249.33893598 + ------------------------------------------------------------------------------------- + TOTAL 0.60974050 106.67808583 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 184704 +BPFP 0.7823 bits/point +EBPFP 1.5647 equivalent bits/point +MSE 106.678086 +---------------------- -------------------------------------------------------- +Time: 2.946s Load: 0.008s, Pack+Encode: 1.719s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 106.6781 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,900B, BPFP=0.2861 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,596B, BPFP=1.8776 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,636B, BPFP=0.5602 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,728B, BPFP=1.8140 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,932B, BPFP=0.9487 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,888B, BPFP=1.7523 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,688B, BPFP=0.8574 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,720B, BPFP=1.8134 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,780B, BPFP=1.5244 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,508B, BPFP=1.7245 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,560B, BPFP=0.0687 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13788442 53.22523749 + layer.0.v_cache 0.00001717 0.01479530 + layer.1.k_cache 0.30723858 6.03128668 + layer.1.v_cache 0.00000664 0.00572812 + layer.2.k_cache 0.01337219 1.41361348 + layer.2.v_cache 0.00002159 0.01681754 + layer.3.k_cache 0.03522867 8.63468310 + layer.3.v_cache 0.00002126 0.02051237 + layer.4.k_cache 0.00073009 0.44859790 + layer.4.v_cache 0.00005119 0.03503444 + layer.4.output 1.43730973 254.05361335 + ------------------------------------------------------------------------------------- + TOTAL 0.62092588 108.71891764 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 185936 +BPFP 0.8023 bits/point +EBPFP 1.6047 equivalent bits/point +MSE 108.718918 +---------------------- -------------------------------------------------------- +Time: 2.956s Load: 0.008s, Pack+Encode: 1.730s, Decode+Unpack: 1.217s +---------------------- -------------------------------------------------------- +💾 Converting with 108.7189 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,832B, BPFP=0.2935 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,192B, BPFP=1.9295 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,160B, BPFP=0.5484 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,216B, BPFP=1.8548 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,520B, BPFP=0.9589 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,472B, BPFP=1.7978 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,356B, BPFP=0.8698 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,240B, BPFP=1.8566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,444B, BPFP=1.4893 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,048B, BPFP=1.7653 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,920B, BPFP=0.0757 +⌛️ [2/4] FRONTEND: Frontend time: 1.752s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11546621 52.12266391 + layer.0.v_cache 0.00001823 0.01450925 + layer.1.k_cache 0.26924565 5.91172521 + layer.1.v_cache 0.00000737 0.00590491 + layer.2.k_cache 0.01761903 1.31587593 + layer.2.v_cache 0.00002145 0.01663184 + layer.3.k_cache 0.02374693 8.77382047 + layer.3.v_cache 0.00002104 0.02033144 + layer.4.k_cache 0.00067272 0.43004092 + layer.4.v_cache 0.00005153 0.03394497 + layer.4.output 1.50071327 265.08996411 + ------------------------------------------------------------------------------------- + TOTAL 0.64305077 113.19265868 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 181400 +BPFP 0.8173 bits/point +EBPFP 1.6346 equivalent bits/point +MSE 113.192659 +---------------------- -------------------------------------------------------- +Time: 2.984s Load: 0.007s, Pack+Encode: 1.752s, Decode+Unpack: 1.225s +---------------------- -------------------------------------------------------- +💾 Converting with 113.1927 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,352B, BPFP=0.2715 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,036B, BPFP=1.6252 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,220B, BPFP=0.5185 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,772B, BPFP=1.5611 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,576B, BPFP=0.7394 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,608B, BPFP=1.5020 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,636B, BPFP=0.6918 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,288B, BPFP=1.5365 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,808B, BPFP=1.3093 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,824B, BPFP=1.4623 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,628B, BPFP=0.0625 +⌛️ [2/4] FRONTEND: Frontend time: 1.864s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18206066 53.49987952 + layer.0.v_cache 0.00001652 0.01347391 + layer.1.k_cache 0.62405802 6.51275199 + layer.1.v_cache 0.00000616 0.00527672 + layer.2.k_cache 0.02036299 1.28190652 + layer.2.v_cache 0.00002087 0.01599412 + layer.3.k_cache 0.02546444 9.25534256 + layer.3.v_cache 0.00002064 0.01879142 + layer.4.k_cache 0.00069880 0.44678235 + layer.4.v_cache 0.00007146 0.03282945 + layer.4.output 0.04336799 175.95019712 + ------------------------------------------------------------------------------------- + TOTAL 0.06802097 76.63143579 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 229748 +BPFP 0.6856 bits/point +EBPFP 1.3712 equivalent bits/point +MSE 76.631436 +---------------------- -------------------------------------------------------- +Time: 3.234s Load: 0.011s, Pack+Encode: 1.864s, Decode+Unpack: 1.359s +---------------------- -------------------------------------------------------- +💾 Converting with 76.6314 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,128B, BPFP=0.2817 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,564B, BPFP=1.7443 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,736B, BPFP=0.5278 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,440B, BPFP=1.6676 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,096B, BPFP=0.8253 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,496B, BPFP=1.6032 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,544B, BPFP=0.7877 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,216B, BPFP=1.6523 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,972B, BPFP=1.3627 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,736B, BPFP=1.5513 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,608B, BPFP=0.0742 +⌛️ [2/4] FRONTEND: Frontend time: 1.739s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.229s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14018573 53.28500273 + layer.0.v_cache 0.00001653 0.01387405 + layer.1.k_cache 0.40671090 6.01175180 + layer.1.v_cache 0.00000662 0.00571897 + layer.2.k_cache 0.03036383 1.22772163 + layer.2.v_cache 0.00002088 0.01558421 + layer.3.k_cache 0.05940984 8.91684080 + layer.3.v_cache 0.00002025 0.01897810 + layer.4.k_cache 0.00071161 0.45605419 + layer.4.v_cache 0.00005094 0.03249594 + layer.4.output 1.33692960 236.25023394 + ------------------------------------------------------------------------------------- + TOTAL 0.58800025 101.39621529 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 183536 +BPFP 0.7366 bits/point +EBPFP 1.4733 equivalent bits/point +MSE 101.396215 +---------------------- -------------------------------------------------------- +Time: 2.975s Load: 0.008s, Pack+Encode: 1.739s, Decode+Unpack: 1.229s +---------------------- -------------------------------------------------------- +💾 Converting with 101.3962 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,992B, BPFP=0.2848 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,584B, BPFP=1.8253 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,596B, BPFP=0.5420 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,540B, BPFP=1.7509 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,764B, BPFP=0.9820 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,972B, BPFP=1.7103 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,248B, BPFP=0.8739 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,436B, BPFP=1.7434 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,488B, BPFP=1.4618 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,324B, BPFP=1.6641 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,472B, BPFP=0.0660 +⌛️ [2/4] FRONTEND: Frontend time: 1.736s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351684 52.98004959 + layer.0.v_cache 0.00001872 0.01470535 + layer.1.k_cache 0.28219827 6.56469838 + layer.1.v_cache 0.00000627 0.00568452 + layer.2.k_cache 0.01639774 1.40148842 + layer.2.v_cache 0.00002172 0.01729028 + layer.3.k_cache 0.02752649 8.21778737 + layer.3.v_cache 0.00002090 0.02020896 + layer.4.k_cache 0.00067264 0.46089284 + layer.4.v_cache 0.00005058 0.03523471 + layer.4.output 1.39794763 247.24898076 + ------------------------------------------------------------------------------------- + TOTAL 0.60212139 105.90946504 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 186416 +BPFP 0.7824 bits/point +EBPFP 1.5647 equivalent bits/point +MSE 105.909465 +---------------------- -------------------------------------------------------- +Time: 2.966s Load: 0.010s, Pack+Encode: 1.736s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 105.9095 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,848B, BPFP=0.2826 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,368B, BPFP=1.8288 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,096B, BPFP=0.5303 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,188B, BPFP=1.7600 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,904B, BPFP=0.8689 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,008B, BPFP=1.6912 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,380B, BPFP=0.8384 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,608B, BPFP=1.7262 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,152B, BPFP=1.4081 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,268B, BPFP=1.6481 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,788B, BPFP=0.0732 +⌛️ [2/4] FRONTEND: Frontend time: 1.877s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15646637 53.44926248 + layer.0.v_cache 0.00001727 0.01331268 + layer.1.k_cache 0.53079696 5.96977393 + layer.1.v_cache 0.00000635 0.00548152 + layer.2.k_cache 0.02340859 1.37032831 + layer.2.v_cache 0.00002061 0.01556605 + layer.3.k_cache 0.01050839 8.02036388 + layer.3.v_cache 0.00002156 0.01856935 + layer.4.k_cache 0.00068596 0.43149772 + layer.4.v_cache 0.00006072 0.03339675 + layer.4.output 0.00497423 206.99142124 + ------------------------------------------------------------------------------------- + TOTAL 0.04451838 89.30985302 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 224608 +BPFP 0.7703 bits/point +EBPFP 1.5406 equivalent bits/point +MSE 89.309853 +---------------------- -------------------------------------------------------- +Time: 3.237s Load: 0.010s, Pack+Encode: 1.877s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 89.3099 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,348B, BPFP=0.2713 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,968B, BPFP=1.6218 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,380B, BPFP=0.5266 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,808B, BPFP=1.5629 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,676B, BPFP=0.7445 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,448B, BPFP=1.4939 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,416B, BPFP=0.7313 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,128B, BPFP=1.5284 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,220B, BPFP=1.2794 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,400B, BPFP=1.4407 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,812B, BPFP=0.0784 +⌛️ [2/4] FRONTEND: Frontend time: 1.844s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15292851 54.31998909 + layer.0.v_cache 0.00001614 0.01372390 + layer.1.k_cache 0.61736927 6.04441259 + layer.1.v_cache 0.00000669 0.00569140 + layer.2.k_cache 0.02003183 1.29893315 + layer.2.v_cache 0.00002344 0.01592526 + layer.3.k_cache 0.01028479 9.04753450 + layer.3.v_cache 0.00002212 0.01868942 + layer.4.k_cache 0.00070434 0.44764145 + layer.4.v_cache 0.00006490 0.03281346 + layer.4.output 0.04342050 175.85547600 + ------------------------------------------------------------------------------------- + TOTAL 0.06502327 76.60198154 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 231604 +BPFP 0.6911 bits/point +EBPFP 1.3823 equivalent bits/point +MSE 76.601982 +---------------------- -------------------------------------------------------- +Time: 3.205s Load: 0.012s, Pack+Encode: 1.844s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 76.6020 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,848B, BPFP=0.2816 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,620B, BPFP=1.8367 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,128B, BPFP=0.5302 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,404B, BPFP=1.7660 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,516B, BPFP=0.8432 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,552B, BPFP=1.7165 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,668B, BPFP=0.7939 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,480B, BPFP=1.7704 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,544B, BPFP=1.4257 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,640B, BPFP=1.6636 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,500B, BPFP=0.0622 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12950333 53.01020490 + layer.0.v_cache 0.00001690 0.01345380 + layer.1.k_cache 0.52417032 6.17439627 + layer.1.v_cache 0.00000602 0.00529490 + layer.2.k_cache 0.01578268 1.31176701 + layer.2.v_cache 0.00002268 0.01577030 + layer.3.k_cache 0.03345619 7.32444746 + layer.3.v_cache 0.00002007 0.01941417 + layer.4.k_cache 0.00067301 0.43211047 + layer.4.v_cache 0.00005077 0.03251371 + layer.4.output 0.00489311 206.10447092 + ------------------------------------------------------------------------------------- + TOTAL 0.04340905 88.88650997 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 224900 +BPFP 0.7684 bits/point +EBPFP 1.5369 equivalent bits/point +MSE 88.886510 +---------------------- -------------------------------------------------------- +Time: 3.198s Load: 0.011s, Pack+Encode: 1.841s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 88.8865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,996B, BPFP=0.2825 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,332B, BPFP=1.7910 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,380B, BPFP=0.5218 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,936B, BPFP=1.6923 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,472B, BPFP=0.8818 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,284B, BPFP=1.6462 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,184B, BPFP=0.8614 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,956B, BPFP=1.6937 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,752B, BPFP=1.3965 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,836B, BPFP=1.6145 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,432B, BPFP=0.0650 +⌛️ [2/4] FRONTEND: Frontend time: 1.749s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.241s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13888381 54.06703814 + layer.0.v_cache 0.00001731 0.01349031 + layer.1.k_cache 0.27925331 6.81895095 + layer.1.v_cache 0.00000589 0.00532959 + layer.2.k_cache 0.03172064 1.45427114 + layer.2.v_cache 0.00001980 0.01567143 + layer.3.k_cache 0.01738505 8.44368305 + layer.3.v_cache 0.00001961 0.01896826 + layer.4.k_cache 0.00068340 0.44452778 + layer.4.v_cache 0.00005063 0.03357987 + layer.4.output 1.38525280 244.81759050 + ------------------------------------------------------------------------------------- + TOTAL 0.59792994 105.00227318 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 181560 +BPFP 0.7551 bits/point +EBPFP 1.5102 equivalent bits/point +MSE 105.002273 +---------------------- -------------------------------------------------------- +Time: 2.998s Load: 0.009s, Pack+Encode: 1.749s, Decode+Unpack: 1.241s +---------------------- -------------------------------------------------------- +💾 Converting with 105.0023 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.7387 bits/point +Avg EBPFP 1.4774 equivalent bits/point +Avg MSE 89.766048 +Avg Time 3.174s +------------------------ ---------------------------- diff --git a/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..0b0745bd1dd88ab118dd019015873b4047c23298 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 255 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other +Output output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other +---------------- ------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,764B, BPFP=0.2926 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,984B, BPFP=1.9422 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,004B, BPFP=0.5445 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,276B, BPFP=1.8871 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,524B, BPFP=1.0513 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,684B, BPFP=1.8411 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,532B, BPFP=0.9742 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,236B, BPFP=1.8840 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,140B, BPFP=1.4879 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,120B, BPFP=1.7973 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,548B, BPFP=0.0727 +⌛️ [2/4] FRONTEND: Frontend time: 2.088s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.281s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09475397 54.58137535 + layer.0.v_cache 0.00001927 0.01430668 + layer.1.k_cache 0.17401673 5.56005252 + layer.1.v_cache 0.00000670 0.00570279 + layer.2.k_cache 0.00814445 1.27758971 + layer.2.v_cache 0.00002157 0.01650044 + layer.3.k_cache 0.01691350 8.99494228 + layer.3.v_cache 0.00002161 0.02129694 + layer.4.k_cache 0.00069969 0.46443806 + layer.4.v_cache 0.00005477 0.03583330 + layer.4.output 1.52309445 268.93796642 + ------------------------------------------------------------------------------------- + TOTAL 0.64448903 114.91398841 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 182812 +BPFP 0.8359 bits/point +EBPFP 1.6719 equivalent bits/point +MSE 114.913988 +---------------------- -------------------------------------------------------- +Time: 3.377s Load: 0.007s, Pack+Encode: 2.088s, Decode+Unpack: 1.281s +---------------------- -------------------------------------------------------- +💾 Converting with 114.9140 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 212, 128) +Output shape: (1, 212, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.output: torch.Size([1, 212, 3584]) -> torch.Size([1, 1, 212, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,876B, BPFP=0.2857 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,652B, BPFP=1.8906 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,060B, BPFP=0.5203 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,552B, BPFP=1.8096 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,124B, BPFP=0.9673 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,244B, BPFP=1.7869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,544B, BPFP=0.9245 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,768B, BPFP=1.8255 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,724B, BPFP=1.5274 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,600B, BPFP=1.7394 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,792B, BPFP=0.0610 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.232s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09403566 51.53660267 + layer.0.v_cache 0.00001799 0.01440172 + layer.1.k_cache 0.24957257 5.91674805 + layer.1.v_cache 0.00000621 0.00550206 + layer.2.k_cache 0.01169950 1.40261726 + layer.2.v_cache 0.00002031 0.01654451 + layer.3.k_cache 0.03645871 8.63137069 + layer.3.v_cache 0.00002058 0.02054338 + layer.4.k_cache 0.00069161 0.45491269 + layer.4.v_cache 0.00005130 0.03423272 + layer.4.output 1.44404146 255.15292284 + ------------------------------------------------------------------------------------- + TOTAL 0.61769792 109.06493739 + (elements=1,845,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1845248 +Total Bytes 185936 +BPFP 0.8061 bits/point +EBPFP 1.6122 equivalent bits/point +MSE 109.064937 +---------------------- -------------------------------------------------------- +Time: 2.965s Load: 0.007s, Pack+Encode: 1.726s, Decode+Unpack: 1.232s +---------------------- -------------------------------------------------------- +💾 Converting with 109.0649 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 250, 128) +Output shape: (1, 250, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.output: torch.Size([1, 250, 3584]) -> torch.Size([1, 1, 250, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,420B, BPFP=0.2762 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,552B, BPFP=1.5970 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,204B, BPFP=0.5128 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,300B, BPFP=1.5188 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,992B, BPFP=0.7495 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,408B, BPFP=1.4630 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,624B, BPFP=0.7890 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,908B, BPFP=1.4943 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,348B, BPFP=1.2717 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,812B, BPFP=1.4258 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,264B, BPFP=0.0649 +⌛️ [2/4] FRONTEND: Frontend time: 1.707s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.224s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09121272 53.80653125 + layer.0.v_cache 0.00001793 0.01346887 + layer.1.k_cache 0.34710925 6.59807715 + layer.1.v_cache 0.00000615 0.00555976 + layer.2.k_cache 0.02211576 1.10281836 + layer.2.v_cache 0.00002179 0.01701654 + layer.3.k_cache 0.02174057 8.52383691 + layer.3.v_cache 0.00001938 0.01916413 + layer.4.k_cache 0.00073173 0.43744632 + layer.4.v_cache 0.00005145 0.03393967 + layer.4.output 1.22463198 216.17891071 + ------------------------------------------------------------------------------------- + TOTAL 0.53267356 93.16530788 + (elements=2,176,000) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2176000 +Total Bytes 184832 +BPFP 0.6795 bits/point +EBPFP 1.3591 equivalent bits/point +MSE 93.165308 +---------------------- -------------------------------------------------------- +Time: 2.940s Load: 0.009s, Pack+Encode: 1.707s, Decode+Unpack: 1.224s +---------------------- -------------------------------------------------------- +💾 Converting with 93.1653 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,028B, BPFP=0.2773 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,364B, BPFP=1.7459 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,568B, BPFP=0.5209 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,416B, BPFP=1.6806 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,680B, BPFP=0.8728 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,464B, BPFP=1.6151 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,964B, BPFP=0.8235 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,120B, BPFP=1.6602 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,336B, BPFP=1.3998 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,836B, BPFP=1.5719 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,272B, BPFP=0.0617 +⌛️ [2/4] FRONTEND: Frontend time: 1.716s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12780563 54.20931133 + layer.0.v_cache 0.00001798 0.01381172 + layer.1.k_cache 0.36413978 6.26960815 + layer.1.v_cache 0.00000619 0.00544569 + layer.2.k_cache 0.01529831 1.31552124 + layer.2.v_cache 0.00002052 0.01605425 + layer.3.k_cache 0.01269315 8.82936883 + layer.3.v_cache 0.00002023 0.01915441 + layer.4.k_cache 0.00069261 0.45459287 + layer.4.v_cache 0.00005545 0.03322809 + layer.4.output 1.34865724 238.48007788 + ------------------------------------------------------------------------------------- + TOTAL 0.58596180 102.38392010 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 183048 +BPFP 0.7412 bits/point +EBPFP 1.4823 equivalent bits/point +MSE 102.383920 +---------------------- -------------------------------------------------------- +Time: 2.948s Load: 0.008s, Pack+Encode: 1.716s, Decode+Unpack: 1.225s +---------------------- -------------------------------------------------------- +💾 Converting with 102.3839 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 214, 128) +Output shape: (1, 214, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.output: torch.Size([1, 214, 3584]) -> torch.Size([1, 1, 214, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,928B, BPFP=0.2868 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,308B, BPFP=1.8478 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,160B, BPFP=0.5228 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,256B, BPFP=1.7710 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,816B, BPFP=0.9357 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,504B, BPFP=1.7161 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,576B, BPFP=0.9182 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,060B, BPFP=1.7567 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,600B, BPFP=1.4311 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,944B, BPFP=1.6752 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,944B, BPFP=0.0620 +⌛️ [2/4] FRONTEND: Frontend time: 1.723s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12820032 51.39128213 + layer.0.v_cache 0.00001693 0.01332824 + layer.1.k_cache 0.24875343 5.30336184 + layer.1.v_cache 0.00000587 0.00512602 + layer.2.k_cache 0.01453762 1.18885297 + layer.2.v_cache 0.00001988 0.01573566 + layer.3.k_cache 0.02276520 7.22135512 + layer.3.v_cache 0.00001932 0.01860028 + layer.4.k_cache 0.00071399 0.44125930 + layer.4.v_cache 0.00004895 0.03313401 + layer.4.output 1.43051836 252.85843625 + ------------------------------------------------------------------------------------- + TOTAL 0.61345353 107.97888761 + (elements=1,862,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1862656 +Total Bytes 182096 +BPFP 0.7821 bits/point +EBPFP 1.5642 equivalent bits/point +MSE 107.978888 +---------------------- -------------------------------------------------------- +Time: 2.953s Load: 0.008s, Pack+Encode: 1.723s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 107.9789 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,976B, BPFP=0.2807 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,620B, BPFP=1.7836 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,264B, BPFP=0.5226 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,288B, BPFP=1.7085 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,292B, BPFP=0.9190 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,648B, BPFP=1.6724 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,300B, BPFP=0.8630 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,164B, BPFP=1.7015 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,628B, BPFP=1.3892 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,824B, BPFP=1.6259 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,060B, BPFP=0.0649 +⌛️ [2/4] FRONTEND: Frontend time: 2.072s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.356s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11875103 51.94692718 + layer.0.v_cache 0.00001689 0.01353175 + layer.1.k_cache 0.49171178 6.31861674 + layer.1.v_cache 0.00000616 0.00540274 + layer.2.k_cache 0.02532451 1.29944255 + layer.2.v_cache 0.00002086 0.01636402 + layer.3.k_cache 0.02370731 8.11500698 + layer.3.v_cache 0.00002129 0.01991557 + layer.4.k_cache 0.00073268 0.43911264 + layer.4.v_cache 0.00004879 0.03307230 + layer.4.output 0.00480151 200.19007865 + ------------------------------------------------------------------------------------- + TOTAL 0.04082070 86.44340841 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 229064 +BPFP 0.7601 bits/point +EBPFP 1.5201 equivalent bits/point +MSE 86.443408 +---------------------- -------------------------------------------------------- +Time: 3.437s Load: 0.009s, Pack+Encode: 2.072s, Decode+Unpack: 1.356s +---------------------- -------------------------------------------------------- +💾 Converting with 86.4434 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,344B, BPFP=0.2770 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,608B, BPFP=1.6332 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,288B, BPFP=0.5286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,612B, BPFP=1.5696 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,752B, BPFP=0.7495 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,888B, BPFP=1.5235 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,212B, BPFP=0.7151 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,392B, BPFP=1.5556 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,608B, BPFP=1.3143 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,128B, BPFP=1.4750 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,636B, BPFP=0.0605 +⌛️ [2/4] FRONTEND: Frontend time: 1.737s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.236s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10657067 52.37338170 + layer.0.v_cache 0.00001604 0.01342106 + layer.1.k_cache 0.39270948 6.12598503 + layer.1.v_cache 0.00000611 0.00539282 + layer.2.k_cache 0.01541229 1.23916439 + layer.2.v_cache 0.00002285 0.01667973 + layer.3.k_cache 0.01321528 8.99957350 + layer.3.v_cache 0.00001985 0.01934968 + layer.4.k_cache 0.00076706 0.45728174 + layer.4.v_cache 0.00005205 0.03325799 + layer.4.output 1.24960368 220.59846939 + ------------------------------------------------------------------------------------- + TOTAL 0.54564808 94.91016314 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 184468 +BPFP 0.6920 bits/point +EBPFP 1.3841 equivalent bits/point +MSE 94.910163 +---------------------- -------------------------------------------------------- +Time: 2.981s Load: 0.008s, Pack+Encode: 1.737s, Decode+Unpack: 1.236s +---------------------- -------------------------------------------------------- +💾 Converting with 94.9102 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 197, 128) +Output shape: (1, 197, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.output: torch.Size([1, 197, 3584]) -> torch.Size([1, 1, 197, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,692B, BPFP=0.2928 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,604B, BPFP=1.9515 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,632B, BPFP=0.5260 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,508B, BPFP=1.8645 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,016B, BPFP=0.9530 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,176B, BPFP=1.8382 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,664B, BPFP=0.8458 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,780B, BPFP=1.8861 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,952B, BPFP=1.5032 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,836B, BPFP=1.8112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,436B, BPFP=0.0729 +⌛️ [2/4] FRONTEND: Frontend time: 1.757s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09219642 52.07244408 + layer.0.v_cache 0.00001621 0.01358630 + layer.1.k_cache 0.17840836 5.94967977 + layer.1.v_cache 0.00000639 0.00538584 + layer.2.k_cache 0.00922235 1.25242653 + layer.2.v_cache 0.00002142 0.01650181 + layer.3.k_cache 0.01957878 7.78897761 + layer.3.v_cache 0.00002124 0.01949081 + layer.4.k_cache 0.00068177 0.44486341 + layer.4.v_cache 0.00005086 0.03488943 + layer.4.output 1.55401793 274.80427393 + ------------------------------------------------------------------------------------- + TOTAL 0.65754878 117.13106842 + (elements=1,714,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1714688 +Total Bytes 176296 +BPFP 0.8225 bits/point +EBPFP 1.6450 equivalent bits/point +MSE 117.131068 +---------------------- -------------------------------------------------------- +Time: 2.985s Load: 0.009s, Pack+Encode: 1.757s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 117.1311 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,416B, BPFP=0.2730 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,744B, BPFP=1.6000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,968B, BPFP=0.5024 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,120B, BPFP=1.5181 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,800B, BPFP=0.7460 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,364B, BPFP=1.4800 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,936B, BPFP=0.6520 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,776B, BPFP=1.5008 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,628B, BPFP=1.2413 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,596B, BPFP=1.4413 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,288B, BPFP=0.0597 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.356s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11169702 53.69565272 + layer.0.v_cache 0.00001595 0.01298766 + layer.1.k_cache 0.54955656 6.85917575 + layer.1.v_cache 0.00000603 0.00528209 + layer.2.k_cache 0.01751511 1.26578320 + layer.2.v_cache 0.00002011 0.01637334 + layer.3.k_cache 0.01650532 8.45079149 + layer.3.v_cache 0.00001967 0.01863081 + layer.4.k_cache 0.00077751 0.44828767 + layer.4.v_cache 0.00005008 0.03382850 + layer.4.output 0.04306356 174.61602823 + ------------------------------------------------------------------------------------- + TOTAL 0.05868284 76.06582299 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 225636 +BPFP 0.6690 bits/point +EBPFP 1.3380 equivalent bits/point +MSE 76.065823 +---------------------- -------------------------------------------------------- +Time: 3.225s Load: 0.011s, Pack+Encode: 1.857s, Decode+Unpack: 1.356s +---------------------- -------------------------------------------------------- +💾 Converting with 76.0658 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,012B, BPFP=0.2876 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,508B, BPFP=1.8283 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,412B, BPFP=0.5312 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,260B, BPFP=1.7388 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,828B, BPFP=0.9194 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,420B, BPFP=1.6786 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,380B, BPFP=0.8873 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,112B, BPFP=1.7282 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,808B, BPFP=1.3481 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,968B, BPFP=1.6462 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,908B, BPFP=0.0707 +⌛️ [2/4] FRONTEND: Frontend time: 1.725s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.224s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09832749 52.49453483 + layer.0.v_cache 0.00001658 0.01384121 + layer.1.k_cache 0.33493462 6.21476984 + layer.1.v_cache 0.00000656 0.00565026 + layer.2.k_cache 0.02762392 1.36699907 + layer.2.v_cache 0.00002012 0.01621203 + layer.3.k_cache 0.01647770 8.14028847 + layer.3.v_cache 0.00001951 0.01972381 + layer.4.k_cache 0.00067889 0.43365360 + layer.4.v_cache 0.00004874 0.03366019 + layer.4.output 1.40435175 248.23341252 + ------------------------------------------------------------------------------------- + TOTAL 0.60638920 106.25724829 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 182616 +BPFP 0.7699 bits/point +EBPFP 1.5399 equivalent bits/point +MSE 106.257248 +---------------------- -------------------------------------------------------- +Time: 2.958s Load: 0.009s, Pack+Encode: 1.725s, Decode+Unpack: 1.224s +---------------------- -------------------------------------------------------- +💾 Converting with 106.2572 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,020B, BPFP=0.2779 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,080B, BPFP=1.7340 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,452B, BPFP=0.5152 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,240B, BPFP=1.6759 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,640B, BPFP=0.8739 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,468B, BPFP=1.6225 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,908B, BPFP=0.8233 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,936B, BPFP=1.6549 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,792B, BPFP=1.3684 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,848B, BPFP=1.5796 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,580B, BPFP=0.0650 +⌛️ [2/4] FRONTEND: Frontend time: 1.719s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102250 54.10043038 + layer.0.v_cache 0.00001721 0.01305337 + layer.1.k_cache 0.37371607 6.26260241 + layer.1.v_cache 0.00000658 0.00536991 + layer.2.k_cache 0.01127678 1.20520533 + layer.2.v_cache 0.00001998 0.01627191 + layer.3.k_cache 0.02136800 8.12960140 + layer.3.v_cache 0.00001963 0.01910723 + layer.4.k_cache 0.00069450 0.44017552 + layer.4.v_cache 0.00005147 0.03412012 + layer.4.output 1.35464589 239.39965234 + ------------------------------------------------------------------------------------- + TOTAL 0.58945376 102.70726494 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 181964 +BPFP 0.7400 bits/point +EBPFP 1.4801 equivalent bits/point +MSE 102.707265 +---------------------- -------------------------------------------------------- +Time: 2.948s Load: 0.009s, Pack+Encode: 1.719s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 102.7073 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,440B, BPFP=0.2764 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,300B, BPFP=1.5750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,164B, BPFP=0.5082 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,144B, BPFP=1.5030 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,956B, BPFP=0.7443 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,144B, BPFP=1.4407 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,340B, BPFP=0.7059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,936B, BPFP=1.4900 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,852B, BPFP=1.2358 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,676B, BPFP=1.4116 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,068B, BPFP=0.0717 +⌛️ [2/4] FRONTEND: Frontend time: 1.713s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.224s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11871604 52.14217349 + layer.0.v_cache 0.00001755 0.01329151 + layer.1.k_cache 0.46315337 6.43109301 + layer.1.v_cache 0.00000588 0.00520228 + layer.2.k_cache 0.02202094 1.24856008 + layer.2.v_cache 0.00001989 0.01629161 + layer.3.k_cache 0.01278766 9.51745070 + layer.3.v_cache 0.00002092 0.01976516 + layer.4.k_cache 0.00074254 0.43153457 + layer.4.v_cache 0.00004925 0.03211209 + layer.4.output 1.21975157 215.65680137 + ------------------------------------------------------------------------------------- + TOTAL 0.53857618 92.90912259 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 183020 +BPFP 0.6702 bits/point +EBPFP 1.3404 equivalent bits/point +MSE 92.909123 +---------------------- -------------------------------------------------------- +Time: 2.945s Load: 0.009s, Pack+Encode: 1.713s, Decode+Unpack: 1.224s +---------------------- -------------------------------------------------------- +💾 Converting with 92.9091 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 195, 128) +Output shape: (1, 195, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.output: torch.Size([1, 195, 3584]) -> torch.Size([1, 1, 195, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,620B, BPFP=0.2901 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,668B, BPFP=1.9766 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,936B, BPFP=0.5558 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,908B, BPFP=1.9157 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,760B, BPFP=0.8622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,056B, BPFP=1.8474 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,356B, BPFP=0.9901 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,676B, BPFP=1.8971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,116B, BPFP=1.5317 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,740B, BPFP=1.8221 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,300B, BPFP=0.0721 +⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605874 53.64728566 + layer.0.v_cache 0.00001676 0.01395431 + layer.1.k_cache 0.13872500 6.47770996 + layer.1.v_cache 0.00000593 0.00545835 + layer.2.k_cache 0.01222654 1.18519663 + layer.2.v_cache 0.00001968 0.01588381 + layer.3.k_cache 0.01415822 8.35706944 + layer.3.v_cache 0.00002127 0.02015454 + layer.4.k_cache 0.00071127 0.45761961 + layer.4.v_cache 0.00005419 0.03525336 + layer.4.output 1.56987251 277.62692308 + ------------------------------------------------------------------------------------- + TOTAL 0.66300619 118.44729689 + (elements=1,697,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1697280 +Total Bytes 177136 +BPFP 0.8349 bits/point +EBPFP 1.6698 equivalent bits/point +MSE 118.447297 +---------------------- -------------------------------------------------------- +Time: 2.941s Load: 0.008s, Pack+Encode: 1.714s, Decode+Unpack: 1.219s +---------------------- -------------------------------------------------------- +💾 Converting with 118.4473 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,368B, BPFP=0.2891 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,576B, BPFP=1.6806 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,364B, BPFP=0.5464 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,660B, BPFP=1.6020 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,020B, BPFP=0.7744 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,052B, BPFP=1.5498 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,248B, BPFP=0.7940 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,532B, BPFP=1.5910 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,772B, BPFP=1.3541 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,548B, BPFP=1.5065 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,848B, BPFP=0.0717 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.108s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11176265 52.44873047 + layer.0.v_cache 0.00001802 0.01434132 + layer.1.k_cache 0.18491925 5.93400356 + layer.1.v_cache 0.00000608 0.00570140 + layer.2.k_cache 0.01999094 1.39530618 + layer.2.v_cache 0.00002176 0.01727060 + layer.3.k_cache 0.03910406 8.46969839 + layer.3.v_cache 0.00002004 0.02095094 + layer.4.k_cache 0.00067358 0.44108481 + layer.4.v_cache 0.00005212 0.03495098 + layer.4.output 0.00886795 301.17231652 + ------------------------------------------------------------------------------------- + TOTAL 0.02462613 128.05813261 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 141988 +BPFP 0.7171 bits/point +EBPFP 1.4341 equivalent bits/point +MSE 128.058133 +---------------------- -------------------------------------------------------- +Time: 2.949s Load: 0.006s, Pack+Encode: 1.834s, Decode+Unpack: 1.108s +---------------------- -------------------------------------------------------- +💾 Converting with 128.0581 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,924B, BPFP=0.2879 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,424B, BPFP=1.8650 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,176B, BPFP=0.5264 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,352B, BPFP=1.7864 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,876B, BPFP=0.8712 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,668B, BPFP=1.7362 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,680B, BPFP=0.8568 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,340B, BPFP=1.7855 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,804B, BPFP=1.4528 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,268B, BPFP=1.7069 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,232B, BPFP=0.0653 +⌛️ [2/4] FRONTEND: Frontend time: 1.703s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15089922 52.70770980 + layer.0.v_cache 0.00001686 0.01405107 + layer.1.k_cache 0.31217133 6.19772124 + layer.1.v_cache 0.00000593 0.00547316 + layer.2.k_cache 0.01305961 1.35308709 + layer.2.v_cache 0.00001999 0.01635761 + layer.3.k_cache 0.01178671 8.49389075 + layer.3.v_cache 0.00002039 0.01983293 + layer.4.k_cache 0.00067606 0.42904362 + layer.4.v_cache 0.00005285 0.03377274 + layer.4.output 1.43724915 253.96501928 + ------------------------------------------------------------------------------------- + TOTAL 0.62055606 108.64859265 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 181744 +BPFP 0.7842 bits/point +EBPFP 1.5685 equivalent bits/point +MSE 108.648593 +---------------------- -------------------------------------------------------- +Time: 2.928s Load: 0.007s, Pack+Encode: 1.703s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 108.6486 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 187, 128) +Output shape: (1, 187, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.output: torch.Size([1, 187, 3584]) -> torch.Size([1, 1, 187, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,376B, BPFP=0.2821 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,500B, BPFP=1.6293 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,184B, BPFP=0.5167 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,412B, BPFP=1.5384 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,344B, BPFP=0.7807 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,048B, BPFP=1.5080 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,948B, BPFP=0.7477 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,684B, BPFP=1.5612 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,816B, BPFP=1.3215 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,636B, BPFP=1.4736 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,416B, BPFP=0.0766 +⌛️ [2/4] FRONTEND: Frontend time: 1.590s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.104s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08258281 53.03901028 + layer.0.v_cache 0.00001733 0.01486788 + layer.1.k_cache 0.10049684 5.75259220 + layer.1.v_cache 0.00000647 0.00581472 + layer.2.k_cache 0.01472783 1.30211643 + layer.2.v_cache 0.00002140 0.01799490 + layer.3.k_cache 0.02125903 8.96147017 + layer.3.v_cache 0.00002180 0.02218718 + layer.4.k_cache 0.00070831 0.47595509 + layer.4.v_cache 0.00005280 0.03697001 + layer.4.output 0.00866398 293.28561879 + ------------------------------------------------------------------------------------- + TOTAL 0.01650250 124.86048885 + (elements=1,627,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1627648 +Total Bytes 142364 +BPFP 0.6997 bits/point +EBPFP 1.3995 equivalent bits/point +MSE 124.860489 +---------------------- -------------------------------------------------------- +Time: 2.700s Load: 0.007s, Pack+Encode: 1.590s, Decode+Unpack: 1.104s +---------------------- -------------------------------------------------------- +💾 Converting with 124.8605 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,844B, BPFP=0.2847 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,156B, BPFP=1.8629 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,148B, BPFP=0.5293 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,452B, BPFP=1.8107 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,784B, BPFP=0.9467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,780B, BPFP=1.7610 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,800B, BPFP=0.9479 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,384B, BPFP=1.8057 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,120B, BPFP=1.4159 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,216B, BPFP=1.7192 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,972B, BPFP=0.0738 +⌛️ [2/4] FRONTEND: Frontend time: 1.711s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11071491 52.70050541 + layer.0.v_cache 0.00001804 0.01401152 + layer.1.k_cache 0.26784508 5.62308043 + layer.1.v_cache 0.00000610 0.00551638 + layer.2.k_cache 0.01369454 1.34783849 + layer.2.v_cache 0.00002102 0.01617829 + layer.3.k_cache 0.01444715 8.45637288 + layer.3.v_cache 0.00002041 0.01985889 + layer.4.k_cache 0.00067564 0.43675239 + layer.4.v_cache 0.00005248 0.03380505 + layer.4.output 1.45091435 255.93919262 + ------------------------------------------------------------------------------------- + TOTAL 0.62140564 109.42519224 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 183656 +BPFP 0.8000 bits/point +EBPFP 1.6000 equivalent bits/point +MSE 109.425192 +---------------------- -------------------------------------------------------- +Time: 2.935s Load: 0.007s, Pack+Encode: 1.711s, Decode+Unpack: 1.217s +---------------------- -------------------------------------------------------- +💾 Converting with 109.4252 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,988B, BPFP=0.2794 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,276B, BPFP=1.7710 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,416B, BPFP=0.5196 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,116B, BPFP=1.6897 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,108B, BPFP=0.8484 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,336B, BPFP=1.6351 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,984B, BPFP=0.7696 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,028B, BPFP=1.6836 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,736B, BPFP=1.3828 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,872B, BPFP=1.6026 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,920B, BPFP=0.0693 +⌛️ [2/4] FRONTEND: Frontend time: 1.704s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13389712 54.11607080 + layer.0.v_cache 0.00001748 0.01346479 + layer.1.k_cache 0.36644789 6.02855132 + layer.1.v_cache 0.00000587 0.00526541 + layer.2.k_cache 0.02260907 1.24831346 + layer.2.v_cache 0.00001979 0.01561984 + layer.3.k_cache 0.02550721 9.00509082 + layer.3.v_cache 0.00001968 0.01937480 + layer.4.k_cache 0.00067529 0.44680092 + layer.4.v_cache 0.00006705 0.03453723 + layer.4.output 1.37284074 242.72241352 + ------------------------------------------------------------------------------------- + TOTAL 0.59759716 104.11705788 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 180780 +BPFP 0.7451 bits/point +EBPFP 1.4902 equivalent bits/point +MSE 104.117058 +---------------------- -------------------------------------------------------- +Time: 2.927s Load: 0.008s, Pack+Encode: 1.704s, Decode+Unpack: 1.215s +---------------------- -------------------------------------------------------- +💾 Converting with 104.1171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 406, 128) +Output shape: (1, 406, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.output: torch.Size([1, 406, 3584]) -> torch.Size([1, 1, 406, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,072B, BPFP=0.2722 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,544B, BPFP=1.7143 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,312B, BPFP=0.5123 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,688B, BPFP=1.6429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,084B, BPFP=0.8114 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,808B, BPFP=1.6090 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,996B, BPFP=0.7696 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,404B, BPFP=1.6319 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,072B, BPFP=1.3113 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 40,152B, BPFP=1.5453 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,164B, BPFP=0.0779 +⌛️ [2/4] FRONTEND: Frontend time: 2.298s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.572s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10474718 51.88670432 + layer.0.v_cache 0.00001775 0.01344480 + layer.1.k_cache 0.89873824 6.00212030 + layer.1.v_cache 0.00000687 0.00550159 + layer.2.k_cache 0.02602503 1.29506138 + layer.2.v_cache 0.00002164 0.01639172 + layer.3.k_cache 0.01558245 9.04096511 + layer.3.v_cache 0.00002201 0.01912835 + layer.4.k_cache 0.00072958 0.44679005 + layer.4.v_cache 0.00005327 0.03292982 + layer.4.output 0.00651461 136.32998329 + ------------------------------------------------------------------------------------- + TOTAL 0.06420861 60.18052473 + (elements=3,533,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3533824 +Total Bytes 321296 +BPFP 0.7274 bits/point +EBPFP 1.4547 equivalent bits/point +MSE 60.180525 +---------------------- -------------------------------------------------------- +Time: 3.883s Load: 0.013s, Pack+Encode: 2.298s, Decode+Unpack: 1.572s +---------------------- -------------------------------------------------------- +💾 Converting with 60.1805 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,380B, BPFP=0.2793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,724B, BPFP=1.6406 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,468B, BPFP=0.5401 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,756B, BPFP=1.5788 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,148B, BPFP=0.7747 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,916B, BPFP=1.5253 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,928B, BPFP=0.7607 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,628B, BPFP=1.5707 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,720B, BPFP=1.3214 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,260B, BPFP=1.4834 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,408B, BPFP=0.0675 +⌛️ [2/4] FRONTEND: Frontend time: 1.715s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14414167 53.46249203 + layer.0.v_cache 0.00001717 0.01368694 + layer.1.k_cache 0.44179790 5.88289919 + layer.1.v_cache 0.00000667 0.00575429 + layer.2.k_cache 0.03016641 1.25319314 + layer.2.v_cache 0.00002043 0.01663942 + layer.3.k_cache 0.03726396 8.48192662 + layer.3.v_cache 0.00002099 0.02034306 + layer.4.k_cache 0.00070653 0.43789439 + layer.4.v_cache 0.00004910 0.03336554 + layer.4.output 1.24963187 220.74139942 + ------------------------------------------------------------------------------------- + TOTAL 0.55303611 94.98811709 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 187336 +BPFP 0.7028 bits/point +EBPFP 1.4056 equivalent bits/point +MSE 94.988117 +---------------------- -------------------------------------------------------- +Time: 2.943s Load: 0.010s, Pack+Encode: 1.715s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 94.9881 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,756B, BPFP=0.2949 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,868B, BPFP=1.9526 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,848B, BPFP=0.5377 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,152B, BPFP=1.8964 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,600B, BPFP=0.9893 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,388B, BPFP=1.8364 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,936B, BPFP=0.9372 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,948B, BPFP=1.8803 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,344B, BPFP=1.5188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,812B, BPFP=1.7911 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,488B, BPFP=0.0616 +⌛️ [2/4] FRONTEND: Frontend time: 1.705s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15985105 52.38295481 + layer.0.v_cache 0.00001905 0.01476884 + layer.1.k_cache 0.22745673 5.86463875 + layer.1.v_cache 0.00000590 0.00558374 + layer.2.k_cache 0.01549998 1.21517400 + layer.2.v_cache 0.00002003 0.01641346 + layer.3.k_cache 0.01333464 8.26387860 + layer.3.v_cache 0.00002034 0.02040072 + layer.4.k_cache 0.00067076 0.45252634 + layer.4.v_cache 0.00004949 0.03440024 + layer.4.output 1.53832061 272.08612258 + ------------------------------------------------------------------------------------- + TOTAL 0.65795131 116.05138809 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 179140 +BPFP 0.8274 bits/point +EBPFP 1.6548 equivalent bits/point +MSE 116.051388 +---------------------- -------------------------------------------------------- +Time: 2.928s Load: 0.008s, Pack+Encode: 1.705s, Decode+Unpack: 1.214s +---------------------- -------------------------------------------------------- +💾 Converting with 116.0514 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 237, 128) +Output shape: (1, 237, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.output: torch.Size([1, 237, 3584]) -> torch.Size([1, 1, 237, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,304B, BPFP=0.2838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,584B, BPFP=1.6867 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,048B, BPFP=0.5306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,492B, BPFP=1.6147 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,048B, BPFP=0.8602 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,824B, BPFP=1.5707 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,480B, BPFP=0.7569 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,340B, BPFP=1.6047 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,180B, BPFP=1.3304 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,056B, BPFP=1.5200 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,936B, BPFP=0.0653 +⌛️ [2/4] FRONTEND: Frontend time: 1.713s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351868 52.23922485 + layer.0.v_cache 0.00001655 0.01342922 + layer.1.k_cache 0.34375901 6.10034077 + layer.1.v_cache 0.00000619 0.00539378 + layer.2.k_cache 0.01282014 1.21521215 + layer.2.v_cache 0.00002104 0.01625561 + layer.3.k_cache 0.03241006 8.33214044 + layer.3.v_cache 0.00001985 0.01883529 + layer.4.k_cache 0.00067667 0.44680857 + layer.4.v_cache 0.00005279 0.03327081 + layer.4.output 1.29178675 228.17678195 + ------------------------------------------------------------------------------------- + TOTAL 0.56210637 97.97990501 + (elements=2,062,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2062848 +Total Bytes 185292 +BPFP 0.7186 bits/point +EBPFP 1.4372 equivalent bits/point +MSE 97.979905 +---------------------- -------------------------------------------------------- +Time: 2.944s Load: 0.008s, Pack+Encode: 1.713s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 97.9799 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,932B, BPFP=0.2858 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,220B, BPFP=1.8328 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,180B, BPFP=0.5218 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,104B, BPFP=1.7517 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,576B, BPFP=0.9866 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,512B, BPFP=1.7087 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,300B, BPFP=0.8212 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,000B, BPFP=1.7442 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,204B, BPFP=1.3956 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,992B, BPFP=1.6709 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,200B, BPFP=0.0644 +⌛️ [2/4] FRONTEND: Frontend time: 1.715s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11023432 51.60277980 + layer.0.v_cache 0.00001677 0.01354998 + layer.1.k_cache 0.30291741 6.18493823 + layer.1.v_cache 0.00000597 0.00530189 + layer.2.k_cache 0.03464115 1.21346847 + layer.2.v_cache 0.00001998 0.01580373 + layer.3.k_cache 0.01330193 7.71449628 + layer.3.v_cache 0.00002112 0.01894299 + layer.4.k_cache 0.00069379 0.43252205 + layer.4.v_cache 0.00005021 0.03274433 + layer.4.output 1.42388847 251.80795266 + ------------------------------------------------------------------------------------- + TOTAL 0.61347776 107.64060096 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 181220 +BPFP 0.7747 bits/point +EBPFP 1.5494 equivalent bits/point +MSE 107.640601 +---------------------- -------------------------------------------------------- +Time: 2.940s Load: 0.007s, Pack+Encode: 1.715s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 107.6406 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 231, 128) +Output shape: (1, 231, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.output: torch.Size([1, 231, 3584]) -> torch.Size([1, 1, 231, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,212B, BPFP=0.2849 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,604B, BPFP=1.7319 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,088B, BPFP=0.5471 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,576B, BPFP=1.6623 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,376B, BPFP=0.8371 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,792B, BPFP=1.6093 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,896B, BPFP=0.8047 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,480B, BPFP=1.6558 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,480B, BPFP=1.3853 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,024B, BPFP=1.5574 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,012B, BPFP=0.0774 +⌛️ [2/4] FRONTEND: Frontend time: 1.717s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13257201 52.23273049 + layer.0.v_cache 0.00001701 0.01417571 + layer.1.k_cache 0.40462213 6.12770880 + layer.1.v_cache 0.00000658 0.00576188 + layer.2.k_cache 0.02649644 1.27942766 + layer.2.v_cache 0.00002056 0.01625665 + layer.3.k_cache 0.06147716 9.20460253 + layer.3.v_cache 0.00002105 0.02001803 + layer.4.k_cache 0.00069437 0.44864341 + layer.4.v_cache 0.00005136 0.03367258 + layer.4.output 1.32536713 234.41709184 + ------------------------------------------------------------------------------------- + TOTAL 0.58256168 100.60603768 + (elements=2,010,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2010624 +Total Bytes 186540 +BPFP 0.7422 bits/point +EBPFP 1.4844 equivalent bits/point +MSE 100.606038 +---------------------- -------------------------------------------------------- +Time: 2.946s Load: 0.008s, Pack+Encode: 1.717s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 100.6060 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,484B, BPFP=0.2748 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,808B, BPFP=1.5814 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,348B, BPFP=0.5115 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,572B, BPFP=1.5056 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,844B, BPFP=0.6645 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,072B, BPFP=1.4750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,288B, BPFP=0.6304 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,596B, BPFP=1.5071 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,268B, BPFP=1.3032 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,224B, BPFP=1.4230 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,032B, BPFP=0.0616 +⌛️ [2/4] FRONTEND: Frontend time: 1.728s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12841698 57.02842371 + layer.0.v_cache 0.00001718 0.01355161 + layer.1.k_cache 0.45450200 6.39221766 + layer.1.v_cache 0.00000627 0.00548280 + layer.2.k_cache 0.01434461 1.45906205 + layer.2.v_cache 0.00002215 0.01686023 + layer.3.k_cache 0.03530993 8.82734566 + layer.3.v_cache 0.00002139 0.01939168 + layer.4.k_cache 0.00068344 0.45442948 + layer.4.v_cache 0.00005073 0.03268092 + layer.4.output 1.20064817 211.64508053 + ------------------------------------------------------------------------------------- + TOTAL 0.53164187 91.51558880 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 184536 +BPFP 0.6651 bits/point +EBPFP 1.3303 equivalent bits/point +MSE 91.515589 +---------------------- -------------------------------------------------------- +Time: 2.959s Load: 0.009s, Pack+Encode: 1.728s, Decode+Unpack: 1.222s +---------------------- -------------------------------------------------------- +💾 Converting with 91.5156 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,204B, BPFP=0.2883 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,708B, BPFP=1.7569 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,568B, BPFP=0.5301 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,288B, BPFP=1.6782 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,212B, BPFP=0.8429 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,564B, BPFP=1.6381 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,288B, BPFP=0.7917 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,024B, BPFP=1.6636 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,820B, BPFP=1.3752 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,764B, BPFP=1.5938 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,048B, BPFP=0.0637 +⌛️ [2/4] FRONTEND: Frontend time: 1.826s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13731835 55.13392758 + layer.0.v_cache 0.00001709 0.01341509 + layer.1.k_cache 0.55879569 6.25564380 + layer.1.v_cache 0.00000637 0.00535624 + layer.2.k_cache 0.01348169 1.31762078 + layer.2.v_cache 0.00002375 0.01634554 + layer.3.k_cache 0.01306405 8.90028760 + layer.3.v_cache 0.00002005 0.01868234 + layer.4.k_cache 0.00072533 0.44565904 + layer.4.v_cache 0.00005384 0.03410204 + layer.4.output 0.00472113 196.65625000 + ------------------------------------------------------------------------------------- + TOTAL 0.04450318 85.21969353 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 227488 +BPFP 0.7414 bits/point +EBPFP 1.4829 equivalent bits/point +MSE 85.219694 +---------------------- -------------------------------------------------------- +Time: 3.184s Load: 0.009s, Pack+Encode: 1.826s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2197 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,304B, BPFP=0.2753 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,216B, BPFP=1.6723 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,044B, BPFP=0.5214 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,884B, BPFP=1.6032 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,760B, BPFP=0.7662 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,896B, BPFP=1.5519 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,192B, BPFP=0.7367 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,744B, BPFP=1.5959 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,636B, BPFP=1.3308 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,856B, BPFP=1.4979 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,144B, BPFP=0.0604 +⌛️ [2/4] FRONTEND: Frontend time: 1.832s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13972237 53.67153434 + layer.0.v_cache 0.00001689 0.01348894 + layer.1.k_cache 0.51719427 6.38188339 + layer.1.v_cache 0.00000613 0.00543040 + layer.2.k_cache 0.02019965 1.34207691 + layer.2.v_cache 0.00002053 0.01594295 + layer.3.k_cache 0.01759115 8.74784248 + layer.3.v_cache 0.00002125 0.01930901 + layer.4.k_cache 0.00069604 0.45452638 + layer.4.v_cache 0.00005872 0.03280993 + layer.4.output 0.04434862 179.94675486 + ------------------------------------------------------------------------------------- + TOTAL 0.05917455 78.25365463 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 230676 +BPFP 0.7044 bits/point +EBPFP 1.4088 equivalent bits/point +MSE 78.253655 +---------------------- -------------------------------------------------------- +Time: 3.193s Load: 0.012s, Pack+Encode: 1.832s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 78.2537 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,724B, BPFP=0.2924 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,716B, BPFP=1.9406 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,824B, BPFP=0.5358 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,172B, BPFP=1.8979 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,396B, BPFP=1.0518 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,776B, BPFP=1.8668 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,616B, BPFP=0.9906 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,900B, BPFP=1.8766 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,696B, BPFP=1.5465 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,720B, BPFP=1.7839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,968B, BPFP=0.0669 +⌛️ [2/4] FRONTEND: Frontend time: 1.711s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13044639 51.78878278 + layer.0.v_cache 0.00001737 0.01392829 + layer.1.k_cache 0.22565776 5.62570911 + layer.1.v_cache 0.00000620 0.00558332 + layer.2.k_cache 0.01998205 1.34787309 + layer.2.v_cache 0.00001989 0.01602866 + layer.3.k_cache 0.00629339 7.60102453 + layer.3.v_cache 0.00002050 0.02006608 + layer.4.k_cache 0.00068118 0.45435506 + layer.4.v_cache 0.00005692 0.03384762 + layer.4.output 1.53833376 272.11333453 + ------------------------------------------------------------------------------------- + TOTAL 0.65597165 115.98238472 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 181508 +BPFP 0.8383 bits/point +EBPFP 1.6767 equivalent bits/point +MSE 115.982385 +---------------------- -------------------------------------------------------- +Time: 2.937s Load: 0.007s, Pack+Encode: 1.711s, Decode+Unpack: 1.219s +---------------------- -------------------------------------------------------- +💾 Converting with 115.9824 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 246, 128) +Output shape: (1, 246, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.output: torch.Size([1, 246, 3584]) -> torch.Size([1, 1, 246, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,360B, BPFP=0.2769 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,748B, BPFP=1.6354 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,136B, BPFP=0.5168 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,552B, BPFP=1.5595 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,924B, BPFP=0.7574 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,796B, BPFP=1.5114 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,224B, BPFP=0.7129 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,480B, BPFP=1.5549 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,948B, BPFP=1.3305 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,024B, BPFP=1.4624 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,952B, BPFP=0.0631 +⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131485 54.55557673 + layer.0.v_cache 0.00001613 0.01379201 + layer.1.k_cache 0.49701033 6.55970975 + layer.1.v_cache 0.00000620 0.00547868 + layer.2.k_cache 0.01298508 1.32573024 + layer.2.v_cache 0.00002120 0.01591970 + layer.3.k_cache 0.02762915 8.86957730 + layer.3.v_cache 0.00002136 0.02000372 + layer.4.k_cache 0.00068392 0.43685228 + layer.4.v_cache 0.00006372 0.03240239 + layer.4.output 1.24453647 219.87006388 + ------------------------------------------------------------------------------------- + TOTAL 0.55302984 94.76032294 + (elements=2,141,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2141184 +Total Bytes 185144 +BPFP 0.6917 bits/point +EBPFP 1.3835 equivalent bits/point +MSE 94.760323 +---------------------- -------------------------------------------------------- +Time: 2.941s Load: 0.009s, Pack+Encode: 1.714s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 94.7603 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,740B, BPFP=0.2922 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,552B, BPFP=1.9181 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,860B, BPFP=0.5359 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,800B, BPFP=1.8594 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,624B, BPFP=0.9862 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,284B, BPFP=1.8191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,464B, BPFP=0.8956 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,520B, BPFP=1.8375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,936B, BPFP=1.5575 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,596B, BPFP=1.7653 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,416B, BPFP=0.0604 +⌛️ [2/4] FRONTEND: Frontend time: 1.777s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11599884 53.23968750 + layer.0.v_cache 0.00001671 0.01346979 + layer.1.k_cache 0.22559120 6.68170410 + layer.1.v_cache 0.00000595 0.00541384 + layer.2.k_cache 0.01146694 1.30540466 + layer.2.v_cache 0.00002201 0.01661623 + layer.3.k_cache 0.03668072 8.60181641 + layer.3.v_cache 0.00002131 0.01945292 + layer.4.k_cache 0.00066762 0.44705791 + layer.4.v_cache 0.00005146 0.03400537 + layer.4.output 1.53061454 270.72933036 + ------------------------------------------------------------------------------------- + TOTAL 0.65322497 115.61587890 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 177792 +BPFP 0.8171 bits/point +EBPFP 1.6341 equivalent bits/point +MSE 115.615879 +---------------------- -------------------------------------------------------- +Time: 3.004s Load: 0.009s, Pack+Encode: 1.777s, Decode+Unpack: 1.217s +---------------------- -------------------------------------------------------- +💾 Converting with 115.6159 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,132B, BPFP=0.2819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,512B, BPFP=1.7407 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,612B, BPFP=0.5194 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,300B, BPFP=1.6580 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,732B, BPFP=0.8687 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,776B, BPFP=1.6223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,136B, BPFP=0.8281 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,412B, BPFP=1.6657 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,588B, BPFP=1.4047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,028B, BPFP=1.5712 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,804B, BPFP=0.0761 +⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605523 52.51542457 + layer.0.v_cache 0.00001866 0.01393713 + layer.1.k_cache 0.31881074 6.32605887 + layer.1.v_cache 0.00000653 0.00567199 + layer.2.k_cache 0.02091764 1.26421093 + layer.2.v_cache 0.00002113 0.01666127 + layer.3.k_cache 0.03221333 8.05802924 + layer.3.v_cache 0.00002185 0.02002498 + layer.4.k_cache 0.00068277 0.46326550 + layer.4.v_cache 0.00005596 0.03447295 + layer.4.output 1.33695214 236.18479024 + ------------------------------------------------------------------------------------- + TOTAL 0.57926287 101.29478171 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 186032 +BPFP 0.7467 bits/point +EBPFP 1.4933 equivalent bits/point +MSE 101.294782 +---------------------- -------------------------------------------------------- +Time: 2.938s Load: 0.008s, Pack+Encode: 1.714s, Decode+Unpack: 1.217s +---------------------- -------------------------------------------------------- +💾 Converting with 101.2948 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 192, 128) +Output shape: (1, 192, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.output: torch.Size([1, 192, 3584]) -> torch.Size([1, 1, 192, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,456B, BPFP=0.2812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,732B, BPFP=1.5244 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,840B, BPFP=0.5566 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,980B, BPFP=1.4632 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,188B, BPFP=0.6663 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,416B, BPFP=1.4173 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,476B, BPFP=0.6084 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,920B, BPFP=1.4583 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,260B, BPFP=1.2419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,976B, BPFP=1.3815 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,584B, BPFP=0.0533 +⌛️ [2/4] FRONTEND: Frontend time: 1.596s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.106s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12444629 59.58028158 + layer.0.v_cache 0.00001540 0.01188266 + layer.1.k_cache 0.18411521 6.45990372 + layer.1.v_cache 0.00000585 0.00483784 + layer.2.k_cache 0.01124177 1.23701255 + layer.2.v_cache 0.00002179 0.01561665 + layer.3.k_cache 0.01616090 9.65200615 + layer.3.v_cache 0.00002117 0.01838228 + layer.4.k_cache 0.00068507 0.45835833 + layer.4.v_cache 0.00005275 0.03101285 + layer.4.output 0.00839530 284.99172247 + ------------------------------------------------------------------------------------- + TOTAL 0.02326666 121.90655011 + (elements=1,671,168) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1671168 +Total Bytes 134828 +BPFP 0.6454 bits/point +EBPFP 1.2909 equivalent bits/point +MSE 121.906550 +---------------------- -------------------------------------------------------- +Time: 2.711s Load: 0.009s, Pack+Encode: 1.596s, Decode+Unpack: 1.106s +---------------------- -------------------------------------------------------- +💾 Converting with 121.9066 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,128B, BPFP=0.2944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,024B, BPFP=1.7907 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,672B, BPFP=0.5339 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,252B, BPFP=1.7180 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,092B, BPFP=1.0441 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,724B, BPFP=1.6683 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,316B, BPFP=0.8769 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,236B, BPFP=1.7165 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,876B, BPFP=1.4944 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,284B, BPFP=1.6269 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,472B, BPFP=0.0601 +⌛️ [2/4] FRONTEND: Frontend time: 1.599s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.103s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09999813 52.71955008 + layer.0.v_cache 0.00001899 0.01373617 + layer.1.k_cache 0.12989934 6.06893149 + layer.1.v_cache 0.00000610 0.00539997 + layer.2.k_cache 0.00966061 1.49049892 + layer.2.v_cache 0.00001994 0.01663584 + layer.3.k_cache 0.03384685 8.67664163 + layer.3.v_cache 0.00002040 0.01936345 + layer.4.k_cache 0.00066502 0.44222903 + layer.4.v_cache 0.00005200 0.03471046 + layer.4.output 0.00959846 330.56459768 + ------------------------------------------------------------------------------------- + TOTAL 0.02008098 140.20234593 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 140076 +BPFP 0.7756 bits/point +EBPFP 1.5512 equivalent bits/point +MSE 140.202346 +---------------------- -------------------------------------------------------- +Time: 2.708s Load: 0.006s, Pack+Encode: 1.599s, Decode+Unpack: 1.103s +---------------------- -------------------------------------------------------- +💾 Converting with 140.2023 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 428, 128) +Output shape: (1, 428, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.output: torch.Size([1, 428, 3584]) -> torch.Size([1, 1, 428, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,388B, BPFP=0.2697 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,540B, BPFP=1.6260 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,064B, BPFP=0.5134 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,360B, BPFP=1.5464 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,848B, BPFP=0.7246 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,968B, BPFP=1.4956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,392B, BPFP=0.7079 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,976B, BPFP=1.5324 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,396B, BPFP=1.2922 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,956B, BPFP=1.4587 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,812B, BPFP=0.0616 +⌛️ [2/4] FRONTEND: Frontend time: 2.068s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.576s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13760449 51.38331447 + layer.0.v_cache 0.00001624 0.01298248 + layer.1.k_cache 0.89392254 6.39331169 + layer.1.v_cache 0.00000615 0.00513578 + layer.2.k_cache 0.03917003 1.32118325 + layer.2.v_cache 0.00002113 0.01619400 + layer.3.k_cache 0.03749864 9.17528669 + layer.3.v_cache 0.00002059 0.01802916 + layer.4.k_cache 0.00086919 0.45797109 + layer.4.v_cache 0.00005177 0.03163327 + layer.4.output 0.00617707 129.43225342 + ------------------------------------------------------------------------------------- + TOTAL 0.06778943 57.34357740 + (elements=3,725,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3725312 +Total Bytes 317700 +BPFP 0.6823 bits/point +EBPFP 1.3645 equivalent bits/point +MSE 57.343577 +---------------------- -------------------------------------------------------- +Time: 3.657s Load: 0.014s, Pack+Encode: 2.068s, Decode+Unpack: 1.576s +---------------------- -------------------------------------------------------- +💾 Converting with 57.3436 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,088B, BPFP=0.2819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,616B, BPFP=1.7518 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,548B, BPFP=0.5290 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,456B, BPFP=1.6875 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,452B, BPFP=0.8562 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,584B, BPFP=1.6392 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,024B, BPFP=0.8324 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,212B, BPFP=1.6740 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,840B, BPFP=1.3763 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,696B, BPFP=1.5900 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,156B, BPFP=0.0725 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385377 53.59548842 + layer.0.v_cache 0.00001705 0.01330263 + layer.1.k_cache 0.49372572 6.10559818 + layer.1.v_cache 0.00000626 0.00548906 + layer.2.k_cache 0.02422354 1.32628464 + layer.2.v_cache 0.00002046 0.01621531 + layer.3.k_cache 0.04153839 8.86248909 + layer.3.v_cache 0.00002075 0.01924252 + layer.4.k_cache 0.00074884 0.44151874 + layer.4.v_cache 0.00004954 0.03230088 + layer.4.output 0.00474379 196.74205294 + ------------------------------------------------------------------------------------- + TOTAL 0.04161240 85.15366471 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 229672 +BPFP 0.7486 bits/point +EBPFP 1.4971 equivalent bits/point +MSE 85.153665 +---------------------- -------------------------------------------------------- +Time: 3.195s Load: 0.011s, Pack+Encode: 1.834s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1537 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,700B, BPFP=0.2757 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,868B, BPFP=1.7835 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,688B, BPFP=0.5170 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,508B, BPFP=1.7177 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,312B, BPFP=0.7891 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,492B, BPFP=1.6685 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,152B, BPFP=0.6846 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,844B, BPFP=1.6856 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,436B, BPFP=1.3272 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,688B, BPFP=1.6296 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,048B, BPFP=0.0763 +⌛️ [2/4] FRONTEND: Frontend time: 2.180s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.474s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12985682 52.87318595 + layer.0.v_cache 0.00001645 0.01292054 + layer.1.k_cache 0.61935935 6.05172229 + layer.1.v_cache 0.00000656 0.00540522 + layer.2.k_cache 0.01492825 1.19990762 + layer.2.v_cache 0.00002178 0.01591441 + layer.3.k_cache 0.02579401 8.17663083 + layer.3.v_cache 0.00002068 0.01820387 + layer.4.k_cache 0.00071613 0.45575768 + layer.4.v_cache 0.00005057 0.03259191 + layer.4.output 0.04141478 167.80661212 + ------------------------------------------------------------------------------------- + TOTAL 0.06356906 73.14638383 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 260736 +BPFP 0.7419 bits/point +EBPFP 1.4839 equivalent bits/point +MSE 73.146384 +---------------------- -------------------------------------------------------- +Time: 3.666s Load: 0.012s, Pack+Encode: 2.180s, Decode+Unpack: 1.474s +---------------------- -------------------------------------------------------- +💾 Converting with 73.1464 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,000B, BPFP=0.2732 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,760B, BPFP=1.7351 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,524B, BPFP=0.5203 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,324B, BPFP=1.6567 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,900B, BPFP=0.8140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,404B, BPFP=1.6064 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,860B, BPFP=0.7572 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,524B, BPFP=1.6676 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,812B, BPFP=1.3556 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,784B, BPFP=1.5726 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,844B, BPFP=0.0690 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12359551 52.43496640 + layer.0.v_cache 0.00001766 0.01382524 + layer.1.k_cache 0.52692803 6.37251463 + layer.1.v_cache 0.00000637 0.00551676 + layer.2.k_cache 0.02102186 1.40690858 + layer.2.v_cache 0.00002187 0.01645000 + layer.3.k_cache 0.03868174 9.54741365 + layer.3.v_cache 0.00002097 0.01994032 + layer.4.k_cache 0.00070799 0.44832216 + layer.4.v_cache 0.00005204 0.03316757 + layer.4.output 0.00469152 193.82757867 + ------------------------------------------------------------------------------------- + TOTAL 0.04375851 83.94659271 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 227736 +BPFP 0.7319 bits/point +EBPFP 1.4637 equivalent bits/point +MSE 83.946593 +---------------------- -------------------------------------------------------- +Time: 3.193s Load: 0.011s, Pack+Encode: 1.833s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9466 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,732B, BPFP=0.2930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,900B, BPFP=1.9551 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,820B, BPFP=0.5355 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,048B, BPFP=1.8882 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,220B, BPFP=1.1165 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,292B, BPFP=1.8288 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,416B, BPFP=0.9749 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,996B, BPFP=1.8841 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,624B, BPFP=1.5408 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,764B, BPFP=1.7874 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,416B, BPFP=0.0720 +⌛️ [2/4] FRONTEND: Frontend time: 1.710s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10671347 52.74129436 + layer.0.v_cache 0.00001788 0.01437387 + layer.1.k_cache 0.15257932 5.75983310 + layer.1.v_cache 0.00000642 0.00552199 + layer.2.k_cache 0.02197224 1.42689046 + layer.2.v_cache 0.00002094 0.01624232 + layer.3.k_cache 0.03325213 8.35727834 + layer.3.v_cache 0.00002126 0.02073544 + layer.4.k_cache 0.00067229 0.46187465 + layer.4.v_cache 0.00005023 0.03268872 + layer.4.output 1.53834953 272.15874013 + ------------------------------------------------------------------------------------- + TOTAL 0.65198546 116.11458319 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 182228 +BPFP 0.8417 bits/point +EBPFP 1.6833 equivalent bits/point +MSE 116.114583 +---------------------- -------------------------------------------------------- +Time: 2.934s Load: 0.007s, Pack+Encode: 1.710s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 116.1146 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,392B, BPFP=0.2734 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,832B, BPFP=1.6081 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,524B, BPFP=0.5306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,692B, BPFP=1.5371 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,716B, BPFP=0.7293 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,120B, BPFP=1.5015 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,328B, BPFP=0.7052 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,524B, BPFP=1.5266 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,124B, BPFP=1.3150 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,208B, BPFP=1.4447 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,052B, BPFP=0.0716 +⌛️ [2/4] FRONTEND: Frontend time: 1.718s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15214830 54.63557100 + layer.0.v_cache 0.00001722 0.01428042 + layer.1.k_cache 0.49754258 6.30199806 + layer.1.v_cache 0.00000639 0.00576953 + layer.2.k_cache 0.01862443 1.49100789 + layer.2.v_cache 0.00002219 0.01727641 + layer.3.k_cache 0.01200796 9.44847271 + layer.3.v_cache 0.00002210 0.02094974 + layer.4.k_cache 0.00075226 0.46645364 + layer.4.v_cache 0.00005409 0.03437989 + layer.4.output 1.21979868 215.56602163 + ------------------------------------------------------------------------------------- + TOTAL 0.54234049 93.02343004 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 187512 +BPFP 0.6866 bits/point +EBPFP 1.3733 equivalent bits/point +MSE 93.023430 +---------------------- -------------------------------------------------------- +Time: 2.945s Load: 0.010s, Pack+Encode: 1.718s, Decode+Unpack: 1.217s +---------------------- -------------------------------------------------------- +💾 Converting with 93.0234 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,328B, BPFP=0.2730 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,948B, BPFP=1.6367 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,064B, BPFP=0.5156 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,700B, BPFP=1.5727 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,552B, BPFP=0.7455 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,644B, BPFP=1.5186 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,552B, BPFP=0.6943 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,464B, BPFP=1.5607 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,004B, BPFP=1.3322 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,860B, BPFP=1.4785 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,324B, BPFP=0.0536 +⌛️ [2/4] FRONTEND: Frontend time: 1.835s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18432382 53.09706071 + layer.0.v_cache 0.00001629 0.01333715 + layer.1.k_cache 0.53789903 6.51323362 + layer.1.v_cache 0.00000607 0.00531056 + layer.2.k_cache 0.01514517 1.42093276 + layer.2.v_cache 0.00002151 0.01606584 + layer.3.k_cache 0.04182221 9.26712026 + layer.3.v_cache 0.00002017 0.01946296 + layer.4.k_cache 0.00068627 0.43601570 + layer.4.v_cache 0.00005063 0.03277025 + layer.4.output 0.04375917 177.41794496 + ------------------------------------------------------------------------------------- + TOTAL 0.06390032 77.22040733 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 228440 +BPFP 0.6884 bits/point +EBPFP 1.3768 equivalent bits/point +MSE 77.220407 +---------------------- -------------------------------------------------------- +Time: 3.192s Load: 0.010s, Pack+Encode: 1.835s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 77.2204 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 371, 128) +Output shape: (1, 371, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.output: torch.Size([1, 371, 3584]) -> torch.Size([1, 1, 371, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,360B, BPFP=0.2679 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,392B, BPFP=1.6169 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,232B, BPFP=0.5152 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,692B, BPFP=1.5453 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,960B, BPFP=0.7143 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,612B, BPFP=1.4998 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,144B, BPFP=0.6799 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,388B, BPFP=1.5325 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,656B, BPFP=1.2911 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,460B, BPFP=1.4513 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,672B, BPFP=0.0642 +⌛️ [2/4] FRONTEND: Frontend time: 1.969s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.464s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16952868 53.38328630 + layer.0.v_cache 0.00001698 0.01366518 + layer.1.k_cache 0.82091874 6.60180072 + layer.1.v_cache 0.00000666 0.00552187 + layer.2.k_cache 0.02418445 1.31926618 + layer.2.v_cache 0.00002238 0.01645274 + layer.3.k_cache 0.03055777 9.51416476 + layer.3.v_cache 0.00002117 0.01945236 + layer.4.k_cache 0.00071103 0.45543006 + layer.4.v_cache 0.00005915 0.03323097 + layer.4.output 0.03610923 146.08020071 + ------------------------------------------------------------------------------------- + TOTAL 0.07639951 64.34845154 + (elements=3,229,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3229184 +Total Bytes 274568 +BPFP 0.6802 bits/point +EBPFP 1.3604 equivalent bits/point +MSE 64.348452 +---------------------- -------------------------------------------------------- +Time: 3.445s Load: 0.012s, Pack+Encode: 1.969s, Decode+Unpack: 1.464s +---------------------- -------------------------------------------------------- +💾 Converting with 64.3485 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,376B, BPFP=0.3136 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,584B, BPFP=1.8337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,548B, BPFP=0.5410 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,520B, BPFP=1.7575 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,460B, BPFP=0.8214 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,032B, BPFP=1.7225 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,452B, BPFP=0.8208 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,544B, BPFP=1.7592 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,128B, BPFP=1.4427 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,240B, BPFP=1.6657 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,044B, BPFP=0.0721 +⌛️ [2/4] FRONTEND: Frontend time: 1.715s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13674593 52.50273258 + layer.0.v_cache 0.00001768 0.01437373 + layer.1.k_cache 0.26206151 6.76730711 + layer.1.v_cache 0.00000624 0.00558196 + layer.2.k_cache 0.02938290 1.32272815 + layer.2.v_cache 0.00002158 0.01634997 + layer.3.k_cache 0.05653211 9.02795018 + layer.3.v_cache 0.00002129 0.02010378 + layer.4.k_cache 0.00070776 0.45883823 + layer.4.v_cache 0.00005193 0.03485725 + layer.4.output 1.40434551 248.24031373 + ------------------------------------------------------------------------------------- + TOTAL 0.60682162 106.34429524 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 183928 +BPFP 0.7755 bits/point +EBPFP 1.5509 equivalent bits/point +MSE 106.344295 +---------------------- -------------------------------------------------------- +Time: 2.943s Load: 0.009s, Pack+Encode: 1.715s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 106.3443 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 193, 128) +Output shape: (1, 193, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.output: torch.Size([1, 193, 3584]) -> torch.Size([1, 1, 193, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,540B, BPFP=0.2866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,120B, BPFP=2.0337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,536B, BPFP=0.5291 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,848B, BPFP=1.9307 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,392B, BPFP=1.0842 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,440B, BPFP=1.8977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,352B, BPFP=0.9190 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,796B, BPFP=1.9265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,824B, BPFP=1.5240 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,772B, BPFP=1.8436 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,920B, BPFP=0.0685 +⌛️ [2/4] FRONTEND: Frontend time: 1.710s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14305212 53.43750000 + layer.0.v_cache 0.00001770 0.01442177 + layer.1.k_cache 0.14179654 6.44195335 + layer.1.v_cache 0.00000608 0.00569806 + layer.2.k_cache 0.01555968 1.32990166 + layer.2.v_cache 0.00002047 0.01707316 + layer.3.k_cache 0.02655400 8.82162555 + layer.3.v_cache 0.00002031 0.02090974 + layer.4.k_cache 0.00069483 0.44811917 + layer.4.v_cache 0.00005004 0.03514959 + layer.4.output 1.58615237 280.21814859 + ------------------------------------------------------------------------------------- + TOTAL 0.67240225 119.53525836 + (elements=1,679,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1679872 +Total Bytes 178540 +BPFP 0.8503 bits/point +EBPFP 1.7005 equivalent bits/point +MSE 119.535258 +---------------------- -------------------------------------------------------- +Time: 2.932s Load: 0.009s, Pack+Encode: 1.710s, Decode+Unpack: 1.213s +---------------------- -------------------------------------------------------- +💾 Converting with 119.5353 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,432B, BPFP=0.2781 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,480B, BPFP=1.5989 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,288B, BPFP=0.5201 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,204B, BPFP=1.5188 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,164B, BPFP=0.7633 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,784B, BPFP=1.4925 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,684B, BPFP=0.7332 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,456B, BPFP=1.5346 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,452B, BPFP=1.2834 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,048B, BPFP=1.4463 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,448B, BPFP=0.0757 +⌛️ [2/4] FRONTEND: Frontend time: 1.704s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10821181 54.39522935 + layer.0.v_cache 0.00001942 0.01369344 + layer.1.k_cache 0.30751975 6.45269408 + layer.1.v_cache 0.00000599 0.00533769 + layer.2.k_cache 0.01421910 1.23115264 + layer.2.v_cache 0.00002020 0.01656278 + layer.3.k_cache 0.01730646 9.14938465 + layer.3.v_cache 0.00002068 0.02052838 + layer.4.k_cache 0.00075739 0.43006201 + layer.4.v_cache 0.00005209 0.03531929 + layer.4.output 1.22958295 217.34554289 + ------------------------------------------------------------------------------------- + TOTAL 0.53265962 93.71580968 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 186440 +BPFP 0.6882 bits/point +EBPFP 1.3764 equivalent bits/point +MSE 93.715810 +---------------------- -------------------------------------------------------- +Time: 2.929s Load: 0.008s, Pack+Encode: 1.704s, Decode+Unpack: 1.217s +---------------------- -------------------------------------------------------- +💾 Converting with 93.7158 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 202, 128) +Output shape: (1, 202, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.output: torch.Size([1, 202, 3584]) -> torch.Size([1, 1, 202, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,796B, BPFP=0.2936 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,028B, BPFP=1.9360 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,968B, BPFP=0.5390 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,208B, BPFP=1.8725 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,052B, BPFP=0.8549 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,528B, BPFP=1.8199 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,292B, BPFP=1.0282 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,324B, BPFP=1.8815 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,896B, BPFP=1.5390 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,944B, BPFP=1.7748 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,764B, BPFP=0.0747 +⌛️ [2/4] FRONTEND: Frontend time: 1.708s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15282421 51.81881381 + layer.0.v_cache 0.00001801 0.01482871 + layer.1.k_cache 0.08597872 6.23724426 + layer.1.v_cache 0.00000649 0.00581275 + layer.2.k_cache 0.01425038 1.20492599 + layer.2.v_cache 0.00002047 0.01629115 + layer.3.k_cache 0.05704349 9.21618229 + layer.3.v_cache 0.00002256 0.02103953 + layer.4.k_cache 0.00069767 0.45670175 + layer.4.v_cache 0.00005056 0.03403480 + layer.4.output 1.51553867 267.69419643 + ------------------------------------------------------------------------------------- + TOTAL 0.64233431 114.28736765 + (elements=1,758,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1758208 +Total Bytes 181800 +BPFP 0.8272 bits/point +EBPFP 1.6544 equivalent bits/point +MSE 114.287368 +---------------------- -------------------------------------------------------- +Time: 2.934s Load: 0.007s, Pack+Encode: 1.708s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 114.2874 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,048B, BPFP=0.2787 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,888B, BPFP=1.7054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,288B, BPFP=0.5128 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,616B, BPFP=1.6352 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,900B, BPFP=0.7674 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,772B, BPFP=1.5886 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,364B, BPFP=0.7931 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,560B, BPFP=1.6321 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,012B, BPFP=1.3258 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,108B, BPFP=1.5519 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,668B, BPFP=0.0684 +⌛️ [2/4] FRONTEND: Frontend time: 1.831s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14837397 50.91048062 + layer.0.v_cache 0.00001603 0.01247663 + layer.1.k_cache 0.60911565 5.75005176 + layer.1.v_cache 0.00000597 0.00489316 + layer.2.k_cache 0.02477714 1.23057152 + layer.2.v_cache 0.00001938 0.01517862 + layer.3.k_cache 0.01052417 9.39173959 + layer.3.v_cache 0.00002039 0.01935514 + layer.4.k_cache 0.00073933 0.43055005 + layer.4.v_cache 0.00005963 0.03564477 + layer.4.output 0.00470031 195.95305401 + ------------------------------------------------------------------------------------- + TOTAL 0.04862081 84.67484235 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 222224 +BPFP 0.7217 bits/point +EBPFP 1.4435 equivalent bits/point +MSE 84.674842 +---------------------- -------------------------------------------------------- +Time: 3.193s Load: 0.012s, Pack+Encode: 1.831s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 84.6748 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 210, 128) +Output shape: (1, 210, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.output: torch.Size([1, 210, 3584]) -> torch.Size([1, 1, 210, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,896B, BPFP=0.2899 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,320B, BPFP=1.8839 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,092B, BPFP=0.5277 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,348B, BPFP=1.8116 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,524B, BPFP=0.8574 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,764B, BPFP=1.7682 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,640B, BPFP=0.8661 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,332B, BPFP=1.8104 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,472B, BPFP=1.5232 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,280B, BPFP=1.7321 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,544B, BPFP=0.0696 +⌛️ [2/4] FRONTEND: Frontend time: 1.712s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11415643 52.73151972 + layer.0.v_cache 0.00001882 0.01367350 + layer.1.k_cache 0.27145578 6.16991490 + layer.1.v_cache 0.00000623 0.00549865 + layer.2.k_cache 0.01201470 1.36402559 + layer.2.v_cache 0.00002208 0.01618341 + layer.3.k_cache 0.00887682 8.92565685 + layer.3.v_cache 0.00002165 0.02035874 + layer.4.k_cache 0.00072611 0.45398251 + layer.4.v_cache 0.00005233 0.03441336 + layer.4.output 1.45781997 257.39638605 + ------------------------------------------------------------------------------------- + TOTAL 0.62424063 110.08881939 + (elements=1,827,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1827840 +Total Bytes 182212 +BPFP 0.7975 bits/point +EBPFP 1.5950 equivalent bits/point +MSE 110.088819 +---------------------- -------------------------------------------------------- +Time: 2.944s Load: 0.010s, Pack+Encode: 1.712s, Decode+Unpack: 1.222s +---------------------- -------------------------------------------------------- +💾 Converting with 110.0888 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 165, 128) +Output shape: (1, 165, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.output: torch.Size([1, 165, 3584]) -> torch.Size([1, 1, 165, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,100B, BPFP=0.2936 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,348B, BPFP=1.8322 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,628B, BPFP=0.5330 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,600B, BPFP=1.7614 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,044B, BPFP=0.9511 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,096B, BPFP=1.7136 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,004B, BPFP=0.8527 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,608B, BPFP=1.7621 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,180B, BPFP=1.5322 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,508B, BPFP=1.6580 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,508B, BPFP=0.0745 +⌛️ [2/4] FRONTEND: Frontend time: 1.597s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.108s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10245823 53.91696851 + layer.0.v_cache 0.00002070 0.01507257 + layer.1.k_cache 0.13127365 5.86213157 + layer.1.v_cache 0.00000615 0.00568583 + layer.2.k_cache 0.01458293 1.66516576 + layer.2.v_cache 0.00002074 0.01696840 + layer.3.k_cache 0.02375244 8.93882576 + layer.3.v_cache 0.00002214 0.02177451 + layer.4.k_cache 0.00067489 0.46247461 + layer.4.v_cache 0.00005196 0.03669300 + layer.4.output 0.00973386 332.49921537 + ------------------------------------------------------------------------------------- + TOTAL 0.02005887 141.08448636 + (elements=1,436,160) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1436160 +Total Bytes 141624 +BPFP 0.7889 bits/point +EBPFP 1.5778 equivalent bits/point +MSE 141.084486 +---------------------- -------------------------------------------------------- +Time: 2.710s Load: 0.006s, Pack+Encode: 1.597s, Decode+Unpack: 1.108s +---------------------- -------------------------------------------------------- +💾 Converting with 141.0845 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,328B, BPFP=0.2842 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,636B, BPFP=1.6766 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,188B, BPFP=0.5283 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,892B, BPFP=1.6130 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,816B, BPFP=0.8381 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,244B, BPFP=1.5577 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,036B, BPFP=0.7715 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,760B, BPFP=1.6018 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,424B, BPFP=1.4023 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,756B, BPFP=1.5161 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,536B, BPFP=0.0675 +⌛️ [2/4] FRONTEND: Frontend time: 1.601s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.109s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12273688 54.19686006 + layer.0.v_cache 0.00001959 0.01503339 + layer.1.k_cache 0.11126613 6.12456642 + layer.1.v_cache 0.00000674 0.00611377 + layer.2.k_cache 0.01416154 1.36277871 + layer.2.v_cache 0.00002164 0.01770181 + layer.3.k_cache 0.04491856 8.56543502 + layer.3.v_cache 0.00002114 0.02180392 + layer.4.k_cache 0.00068813 0.47784170 + layer.4.v_cache 0.00006151 0.03606887 + layer.4.output 0.00884247 299.87407299 + ------------------------------------------------------------------------------------- + TOTAL 0.02092936 127.64368909 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 143616 +BPFP 0.7213 bits/point +EBPFP 1.4426 equivalent bits/point +MSE 127.643689 +---------------------- -------------------------------------------------------- +Time: 2.718s Load: 0.009s, Pack+Encode: 1.601s, Decode+Unpack: 1.109s +---------------------- -------------------------------------------------------- +💾 Converting with 127.6437 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 257, 128) +Output shape: (1, 257, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.output: torch.Size([1, 257, 3584]) -> torch.Size([1, 1, 257, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,564B, BPFP=0.2775 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,388B, BPFP=1.9083 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,452B, BPFP=0.5139 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,656B, BPFP=1.8030 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,488B, BPFP=0.7592 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,940B, BPFP=1.7595 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,156B, BPFP=0.7999 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,932B, BPFP=1.8198 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,440B, BPFP=1.4251 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,124B, BPFP=1.7099 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,272B, BPFP=0.0718 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.342s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12694686 53.64560281 + layer.0.v_cache 0.00002055 0.01402901 + layer.1.k_cache 0.54844232 6.60579079 + layer.1.v_cache 0.00000676 0.00537125 + layer.2.k_cache 0.01835482 1.29947458 + layer.2.v_cache 0.00002099 0.01642048 + layer.3.k_cache 0.02141366 7.91261333 + layer.3.v_cache 0.00002047 0.01976311 + layer.4.k_cache 0.00070767 0.45408443 + layer.4.v_cache 0.00005046 0.03204794 + layer.4.output 0.00513101 215.79697054 + ------------------------------------------------------------------------------------- + TOTAL 0.04422951 92.97552891 + (elements=2,236,928) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2236928 +Total Bytes 218412 +BPFP 0.7811 bits/point +EBPFP 1.5622 equivalent bits/point +MSE 92.975529 +---------------------- -------------------------------------------------------- +Time: 3.185s Load: 0.009s, Pack+Encode: 1.834s, Decode+Unpack: 1.342s +---------------------- -------------------------------------------------------- +💾 Converting with 92.9755 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,352B, BPFP=0.2878 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,540B, BPFP=1.6775 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,036B, BPFP=0.5182 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,468B, BPFP=1.5855 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,968B, BPFP=0.9416 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,912B, BPFP=1.5378 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,388B, BPFP=0.8060 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,644B, BPFP=1.6006 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,756B, BPFP=1.3527 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,560B, BPFP=1.5076 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,420B, BPFP=0.0665 +⌛️ [2/4] FRONTEND: Frontend time: 1.607s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.109s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12731336 51.76192801 + layer.0.v_cache 0.00001750 0.01467303 + layer.1.k_cache 0.10052182 6.48866758 + layer.1.v_cache 0.00000572 0.00536288 + layer.2.k_cache 0.01141533 1.32745127 + layer.2.v_cache 0.00001926 0.01619458 + layer.3.k_cache 0.03745206 9.17253163 + layer.3.v_cache 0.00002083 0.02088348 + layer.4.k_cache 0.00066429 0.44161535 + layer.4.v_cache 0.00005101 0.03509402 + layer.4.output 0.00884527 301.23098999 + ------------------------------------------------------------------------------------- + TOTAL 0.01996459 128.11184305 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 143044 +BPFP 0.7224 bits/point +EBPFP 1.4448 equivalent bits/point +MSE 128.111843 +---------------------- -------------------------------------------------------- +Time: 2.723s Load: 0.006s, Pack+Encode: 1.607s, Decode+Unpack: 1.109s +---------------------- -------------------------------------------------------- +💾 Converting with 128.1118 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,952B, BPFP=0.2782 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,736B, BPFP=1.8114 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,324B, BPFP=0.5155 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,140B, BPFP=1.6990 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,840B, BPFP=0.9037 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,460B, BPFP=1.6512 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,024B, BPFP=0.8463 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,372B, BPFP=1.7154 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,580B, BPFP=1.3781 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,892B, BPFP=1.6112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,268B, BPFP=0.0731 +⌛️ [2/4] FRONTEND: Frontend time: 1.715s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.236s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12752426 55.54211096 + layer.0.v_cache 0.00002004 0.01447230 + layer.1.k_cache 0.30660921 5.98207051 + layer.1.v_cache 0.00000613 0.00538541 + layer.2.k_cache 0.01808885 1.25434408 + layer.2.v_cache 0.00002107 0.01600894 + layer.3.k_cache 0.01678786 8.03302827 + layer.3.v_cache 0.00002050 0.01955053 + layer.4.k_cache 0.00068290 0.43750179 + layer.4.v_cache 0.00004851 0.03306238 + layer.4.output 1.37905047 243.72480293 + ------------------------------------------------------------------------------------- + TOTAL 0.59548016 104.55359739 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 183588 +BPFP 0.7601 bits/point +EBPFP 1.5202 equivalent bits/point +MSE 104.553597 +---------------------- -------------------------------------------------------- +Time: 2.960s Load: 0.009s, Pack+Encode: 1.715s, Decode+Unpack: 1.236s +---------------------- -------------------------------------------------------- +💾 Converting with 104.5536 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,020B, BPFP=0.2895 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,328B, BPFP=1.8237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,616B, BPFP=0.5484 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,464B, BPFP=1.7615 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,740B, BPFP=0.9173 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,688B, BPFP=1.7056 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,336B, BPFP=0.8882 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,804B, BPFP=1.7860 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,704B, BPFP=1.4188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,328B, BPFP=1.6797 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,148B, BPFP=0.0838 +⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12766143 51.79972818 + layer.0.v_cache 0.00001653 0.01399973 + layer.1.k_cache 0.34631467 5.79596482 + layer.1.v_cache 0.00000659 0.00561765 + layer.2.k_cache 0.01786991 1.21443901 + layer.2.v_cache 0.00002041 0.01622403 + layer.3.k_cache 0.06874437 8.80978183 + layer.3.v_cache 0.00002276 0.02154483 + layer.4.k_cache 0.00067882 0.45362949 + layer.4.v_cache 0.00005621 0.03548731 + layer.4.output 1.41085444 249.45412278 + ------------------------------------------------------------------------------------- + TOTAL 0.61396310 106.72619272 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 186176 +BPFP 0.7886 bits/point +EBPFP 1.5771 equivalent bits/point +MSE 106.726193 +---------------------- -------------------------------------------------------- +Time: 2.944s Load: 0.008s, Pack+Encode: 1.714s, Decode+Unpack: 1.222s +---------------------- -------------------------------------------------------- +💾 Converting with 106.7262 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,004B, BPFP=0.2768 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,624B, BPFP=1.7716 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,976B, BPFP=0.5514 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,624B, BPFP=1.7024 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,188B, BPFP=0.9118 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,928B, BPFP=1.6543 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,732B, BPFP=0.8111 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,756B, BPFP=1.7116 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,268B, BPFP=1.4013 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,184B, BPFP=1.6029 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,796B, BPFP=0.0770 +⌛️ [2/4] FRONTEND: Frontend time: 1.720s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.228s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12070838 55.23065888 + layer.0.v_cache 0.00001763 0.01405055 + layer.1.k_cache 0.31015487 6.46051133 + layer.1.v_cache 0.00000677 0.00588269 + layer.2.k_cache 0.03544329 1.36959609 + layer.2.v_cache 0.00002212 0.01680487 + layer.3.k_cache 0.05682086 9.45903515 + layer.3.v_cache 0.00002136 0.02104721 + layer.4.k_cache 0.00073636 0.44674335 + layer.4.v_cache 0.00005031 0.03478919 + layer.4.output 1.35470685 238.97072535 + ------------------------------------------------------------------------------------- + TOTAL 0.58864294 102.69730569 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 187080 +BPFP 0.7608 bits/point +EBPFP 1.5217 equivalent bits/point +MSE 102.697306 +---------------------- -------------------------------------------------------- +Time: 2.958s Load: 0.010s, Pack+Encode: 1.720s, Decode+Unpack: 1.228s +---------------------- -------------------------------------------------------- +💾 Converting with 102.6973 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,936B, BPFP=0.2887 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,488B, BPFP=1.8697 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,300B, BPFP=0.5355 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,640B, BPFP=1.8075 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,460B, BPFP=0.9874 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,980B, BPFP=1.7591 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,780B, BPFP=0.8641 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,708B, BPFP=1.8125 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,964B, BPFP=1.4645 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,432B, BPFP=1.7189 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,892B, BPFP=0.0827 +⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13227103 51.45551368 + layer.0.v_cache 0.00002050 0.01420120 + layer.1.k_cache 0.22864332 6.64638752 + layer.1.v_cache 0.00000662 0.00579400 + layer.2.k_cache 0.02950178 1.33866259 + layer.2.v_cache 0.00002095 0.01643357 + layer.3.k_cache 0.01835468 8.47145904 + layer.3.v_cache 0.00002123 0.02104380 + layer.4.k_cache 0.00070999 0.45748138 + layer.4.v_cache 0.00005457 0.03381280 + layer.4.output 1.43734575 253.84077381 + ------------------------------------------------------------------------------------- + TOTAL 0.61594264 108.54977684 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 186580 +BPFP 0.8051 bits/point +EBPFP 1.6102 equivalent bits/point +MSE 108.549777 +---------------------- -------------------------------------------------------- +Time: 2.946s Load: 0.008s, Pack+Encode: 1.714s, Decode+Unpack: 1.225s +---------------------- -------------------------------------------------------- +💾 Converting with 108.5498 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,984B, BPFP=0.2804 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,268B, BPFP=1.7784 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,416B, BPFP=0.5220 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,196B, BPFP=1.7030 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,436B, BPFP=0.8753 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,524B, BPFP=1.6557 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,708B, BPFP=0.8944 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,028B, BPFP=1.6912 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,692B, BPFP=1.3860 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,772B, BPFP=1.6028 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,804B, BPFP=0.0684 +⌛️ [2/4] FRONTEND: Frontend time: 1.715s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12518616 53.40046541 + layer.0.v_cache 0.00001676 0.01370975 + layer.1.k_cache 0.33007166 6.21274326 + layer.1.v_cache 0.00000630 0.00562230 + layer.2.k_cache 0.01131002 1.29756646 + layer.2.v_cache 0.00002125 0.01713748 + layer.3.k_cache 0.02232737 9.49416702 + layer.3.v_cache 0.00001983 0.01981915 + layer.4.k_cache 0.00071526 0.47494472 + layer.4.v_cache 0.00005006 0.03261862 + layer.4.output 1.37904434 243.47033864 + ------------------------------------------------------------------------------------- + TOTAL 0.59664912 104.42712733 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 182828 +BPFP 0.7569 bits/point +EBPFP 1.5139 equivalent bits/point +MSE 104.427127 +---------------------- -------------------------------------------------------- +Time: 2.943s Load: 0.007s, Pack+Encode: 1.715s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 104.4271 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,368B, BPFP=0.2797 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,348B, BPFP=1.6232 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,008B, BPFP=0.5128 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,576B, BPFP=1.5738 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,096B, BPFP=0.7746 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,688B, BPFP=1.5169 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,604B, BPFP=0.7431 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,080B, BPFP=1.5420 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,244B, BPFP=1.2964 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,960B, BPFP=1.4703 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,048B, BPFP=0.0736 +⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14655781 51.11853628 + layer.0.v_cache 0.00001709 0.01346769 + layer.1.k_cache 0.36683986 6.10671297 + layer.1.v_cache 0.00000651 0.00561314 + layer.2.k_cache 0.01755252 1.14181531 + layer.2.v_cache 0.00002060 0.01639091 + layer.3.k_cache 0.02713263 8.50212422 + layer.3.v_cache 0.00001997 0.01904818 + layer.4.k_cache 0.00071102 0.44181020 + layer.4.v_cache 0.00005162 0.03240553 + layer.4.output 1.25479176 221.60335553 + ------------------------------------------------------------------------------------- + TOTAL 0.54955600 95.21302430 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 185020 +BPFP 0.6969 bits/point +EBPFP 1.3939 equivalent bits/point +MSE 95.213024 +---------------------- -------------------------------------------------------- +Time: 2.945s Load: 0.011s, Pack+Encode: 1.714s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 95.2130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 155, 128) +Output shape: (1, 155, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.output: torch.Size([1, 155, 3584]) -> torch.Size([1, 1, 155, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,964B, BPFP=0.2988 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,024B, BPFP=1.9177 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,344B, BPFP=0.5387 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,176B, BPFP=1.8323 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,828B, BPFP=1.0915 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,836B, BPFP=1.7980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,028B, BPFP=0.9101 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,500B, BPFP=1.8649 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,232B, BPFP=1.6363 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,492B, BPFP=1.7633 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,596B, BPFP=0.0662 +⌛️ [2/4] FRONTEND: Frontend time: 1.607s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.121s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102877 51.49036353 + layer.0.v_cache 0.00001822 0.01440545 + layer.1.k_cache 0.12773433 6.63238250 + layer.1.v_cache 0.00000585 0.00562222 + layer.2.k_cache 0.00931232 1.49623000 + layer.2.v_cache 0.00002110 0.01746757 + layer.3.k_cache 0.01537714 9.48303459 + layer.3.v_cache 0.00002091 0.02188617 + layer.4.k_cache 0.00066471 0.46160278 + layer.4.v_cache 0.00005180 0.03736227 + layer.4.output 0.01742802 357.01699309 + ------------------------------------------------------------------------------------- + TOTAL 0.02389596 151.10466522 + (elements=1,349,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1349120 +Total Bytes 140020 +BPFP 0.8303 bits/point +EBPFP 1.6606 equivalent bits/point +MSE 151.104665 +---------------------- -------------------------------------------------------- +Time: 2.736s Load: 0.008s, Pack+Encode: 1.607s, Decode+Unpack: 1.121s +---------------------- -------------------------------------------------------- +💾 Converting with 151.1047 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,960B, BPFP=0.2878 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,320B, BPFP=1.8401 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,760B, BPFP=0.5640 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,440B, BPFP=1.7762 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,816B, BPFP=0.8587 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,556B, BPFP=1.7119 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,072B, BPFP=0.8773 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,332B, BPFP=1.7683 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,452B, BPFP=1.4863 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,140B, BPFP=1.6817 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,688B, BPFP=0.0694 +⌛️ [2/4] FRONTEND: Frontend time: 1.722s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.233s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11659910 50.15492369 + layer.0.v_cache 0.00001673 0.01439245 + layer.1.k_cache 0.29571861 6.44953613 + layer.1.v_cache 0.00000615 0.00568234 + layer.2.k_cache 0.04707494 1.30925662 + layer.2.v_cache 0.00002002 0.01601486 + layer.3.k_cache 0.03274950 8.72091888 + layer.3.v_cache 0.00002078 0.02029336 + layer.4.k_cache 0.00068608 0.43236389 + layer.4.v_cache 0.00005136 0.03407486 + layer.4.output 1.42391751 251.71939369 + ------------------------------------------------------------------------------------- + TOTAL 0.61531564 107.59960076 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 183536 +BPFP 0.7846 bits/point +EBPFP 1.5692 equivalent bits/point +MSE 107.599601 +---------------------- -------------------------------------------------------- +Time: 2.962s Load: 0.007s, Pack+Encode: 1.722s, Decode+Unpack: 1.233s +---------------------- -------------------------------------------------------- +💾 Converting with 107.5996 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,380B, BPFP=0.2760 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,980B, BPFP=1.6368 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,664B, BPFP=0.5459 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,820B, BPFP=1.5638 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,324B, BPFP=0.7765 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,980B, BPFP=1.5108 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,428B, BPFP=0.7200 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,812B, BPFP=1.5633 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,184B, BPFP=1.3347 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,376B, BPFP=1.4728 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,204B, BPFP=0.0738 +⌛️ [2/4] FRONTEND: Frontend time: 1.721s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.232s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10631144 54.38253686 + layer.0.v_cache 0.00001780 0.01451916 + layer.1.k_cache 0.46192560 6.10183716 + layer.1.v_cache 0.00000619 0.00587016 + layer.2.k_cache 0.02890189 1.29328266 + layer.2.v_cache 0.00002115 0.01664336 + layer.3.k_cache 0.04994748 8.75866404 + layer.3.v_cache 0.00002140 0.02116185 + layer.4.k_cache 0.00068365 0.45993282 + layer.4.v_cache 0.00005040 0.03472836 + layer.4.output 1.23456514 218.27221342 + ------------------------------------------------------------------------------------- + TOTAL 0.54646135 94.05851002 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 189152 +BPFP 0.7010 bits/point +EBPFP 1.4020 equivalent bits/point +MSE 94.058510 +---------------------- -------------------------------------------------------- +Time: 2.962s Load: 0.009s, Pack+Encode: 1.721s, Decode+Unpack: 1.232s +---------------------- -------------------------------------------------------- +💾 Converting with 94.0585 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 194, 128) +Output shape: (1, 194, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.output: torch.Size([1, 194, 3584]) -> torch.Size([1, 1, 194, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,556B, BPFP=0.2864 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,928B, BPFP=2.0077 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,556B, BPFP=0.5280 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,668B, BPFP=1.9062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,348B, BPFP=0.9140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,296B, BPFP=1.8763 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,764B, BPFP=0.8669 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,764B, BPFP=1.9140 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,036B, BPFP=1.5332 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,572B, BPFP=1.8180 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,252B, BPFP=0.0604 +⌛️ [2/4] FRONTEND: Frontend time: 1.727s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.239s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15027005 52.30229140 + layer.0.v_cache 0.00001782 0.01458332 + layer.1.k_cache 0.21402168 6.53560678 + layer.1.v_cache 0.00000588 0.00553162 + layer.2.k_cache 0.02177384 1.40025550 + layer.2.v_cache 0.00002019 0.01710368 + layer.3.k_cache 0.04024272 8.81043802 + layer.3.v_cache 0.00002076 0.02090430 + layer.4.k_cache 0.00069591 0.45532596 + layer.4.v_cache 0.00005073 0.03520581 + layer.4.output 1.57794378 279.02386322 + ------------------------------------------------------------------------------------- + TOTAL 0.67486624 118.98613464 + (elements=1,688,576) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1688576 +Total Bytes 174740 +BPFP 0.8279 bits/point +EBPFP 1.6557 equivalent bits/point +MSE 118.986135 +---------------------- -------------------------------------------------------- +Time: 2.974s Load: 0.008s, Pack+Encode: 1.727s, Decode+Unpack: 1.239s +---------------------- -------------------------------------------------------- +💾 Converting with 118.9861 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 258, 128) +Output shape: (1, 258, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.output: torch.Size([1, 258, 3584]) -> torch.Size([1, 1, 258, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,596B, BPFP=0.2783 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,868B, BPFP=1.8694 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,484B, BPFP=0.5138 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,616B, BPFP=1.7936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,780B, BPFP=0.7740 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,704B, BPFP=1.7384 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,264B, BPFP=0.8639 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,352B, BPFP=1.7776 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,608B, BPFP=1.4297 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,240B, BPFP=1.7103 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,488B, BPFP=0.0561 +⌛️ [2/4] FRONTEND: Frontend time: 1.839s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09035230 53.13108648 + layer.0.v_cache 0.00001726 0.01342705 + layer.1.k_cache 0.42455262 6.02213069 + layer.1.v_cache 0.00000600 0.00515154 + layer.2.k_cache 0.01891134 1.28776633 + layer.2.v_cache 0.00001996 0.01580407 + layer.3.k_cache 0.02093011 7.88465775 + layer.3.v_cache 0.00001928 0.01902181 + layer.4.k_cache 0.00072082 0.45020137 + layer.4.v_cache 0.00004951 0.03411891 + layer.4.output 0.00505962 214.94366002 + ------------------------------------------------------------------------------------- + TOTAL 0.03476450 92.55699919 + (elements=2,245,632) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2245632 +Total Bytes 217000 +BPFP 0.7731 bits/point +EBPFP 1.5461 equivalent bits/point +MSE 92.556999 +---------------------- -------------------------------------------------------- +Time: 3.215s Load: 0.012s, Pack+Encode: 1.839s, Decode+Unpack: 1.365s +---------------------- -------------------------------------------------------- +💾 Converting with 92.5570 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,176B, BPFP=0.2837 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,416B, BPFP=1.7266 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,676B, BPFP=0.5215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,328B, BPFP=1.6527 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,836B, BPFP=0.8720 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,536B, BPFP=1.5989 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,836B, BPFP=0.8041 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,404B, BPFP=1.6579 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,472B, BPFP=1.3228 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,788B, BPFP=1.5481 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,360B, BPFP=0.0908 +⌛️ [2/4] FRONTEND: Frontend time: 1.724s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.230s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15008528 53.31964164 + layer.0.v_cache 0.00001847 0.01387775 + layer.1.k_cache 0.25612781 6.09269330 + layer.1.v_cache 0.00000678 0.00560147 + layer.2.k_cache 0.03931687 1.14547079 + layer.2.v_cache 0.00002107 0.01602277 + layer.3.k_cache 0.03786138 8.46713177 + layer.3.v_cache 0.00002177 0.02024193 + layer.4.k_cache 0.00069674 0.44839275 + layer.4.v_cache 0.00005115 0.03397393 + layer.4.output 1.33117570 235.36937112 + ------------------------------------------------------------------------------------- + TOTAL 0.57661395 101.00874388 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 185828 +BPFP 0.7426 bits/point +EBPFP 1.4852 equivalent bits/point +MSE 101.008744 +---------------------- -------------------------------------------------------- +Time: 2.963s Load: 0.009s, Pack+Encode: 1.724s, Decode+Unpack: 1.230s +---------------------- -------------------------------------------------------- +💾 Converting with 101.0087 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,364B, BPFP=0.2920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,472B, BPFP=1.6903 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,204B, BPFP=0.5385 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,660B, BPFP=1.6198 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,560B, BPFP=0.8299 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,160B, BPFP=1.5764 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,588B, BPFP=0.8323 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,700B, BPFP=1.6233 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,112B, BPFP=1.3986 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,720B, BPFP=1.5382 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,028B, BPFP=0.0748 +⌛️ [2/4] FRONTEND: Frontend time: 1.612s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.111s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15161699 50.88803168 + layer.0.v_cache 0.00001786 0.01494464 + layer.1.k_cache 0.08662784 6.74674615 + layer.1.v_cache 0.00000652 0.00585551 + layer.2.k_cache 0.02063684 1.18627887 + layer.2.v_cache 0.00002327 0.01709492 + layer.3.k_cache 0.08087930 9.01609836 + layer.3.v_cache 0.00002116 0.02108237 + layer.4.k_cache 0.00074161 0.46600333 + layer.4.v_cache 0.00006096 0.03613887 + layer.4.output 0.00898438 304.48390377 + ------------------------------------------------------------------------------------- + TOTAL 0.02373665 129.39915301 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 143568 +BPFP 0.7331 bits/point +EBPFP 1.4662 equivalent bits/point +MSE 129.399153 +---------------------- -------------------------------------------------------- +Time: 2.732s Load: 0.009s, Pack+Encode: 1.612s, Decode+Unpack: 1.111s +---------------------- -------------------------------------------------------- +💾 Converting with 129.3992 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,992B, BPFP=0.2835 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,460B, BPFP=1.8082 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,572B, BPFP=0.5378 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,580B, BPFP=1.7457 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,320B, BPFP=0.9460 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,884B, BPFP=1.6963 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,180B, BPFP=0.8651 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,404B, BPFP=1.7332 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,624B, BPFP=1.4648 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,272B, BPFP=1.6528 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,724B, BPFP=0.0682 +⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14327769 53.24201438 + layer.0.v_cache 0.00001763 0.01416141 + layer.1.k_cache 0.34925704 6.65342130 + layer.1.v_cache 0.00000642 0.00568322 + layer.2.k_cache 0.02352169 1.43257127 + layer.2.v_cache 0.00002271 0.01665244 + layer.3.k_cache 0.00845094 8.42356401 + layer.3.v_cache 0.00002026 0.02017767 + layer.4.k_cache 0.00069169 0.45360128 + layer.4.v_cache 0.00005086 0.03449108 + layer.4.output 1.39158238 246.06237825 + ------------------------------------------------------------------------------------- + TOTAL 0.60390550 105.45488152 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 186012 +BPFP 0.7771 bits/point +EBPFP 1.5542 equivalent bits/point +MSE 105.454882 +---------------------- -------------------------------------------------------- +Time: 2.947s Load: 0.010s, Pack+Encode: 1.714s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 105.4549 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,916B, BPFP=0.2793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,652B, BPFP=1.7984 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,344B, BPFP=0.5309 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,352B, BPFP=1.7245 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,376B, BPFP=0.8736 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,576B, BPFP=1.6805 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,536B, BPFP=0.7691 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,216B, BPFP=1.7168 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,476B, BPFP=1.3907 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,796B, BPFP=1.6361 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,744B, BPFP=0.0629 +⌛️ [2/4] FRONTEND: Frontend time: 1.843s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12716211 51.27470170 + layer.0.v_cache 0.00001677 0.01362425 + layer.1.k_cache 0.51313377 6.19245783 + layer.1.v_cache 0.00000586 0.00527198 + layer.2.k_cache 0.02995576 1.48560924 + layer.2.v_cache 0.00002133 0.01639123 + layer.3.k_cache 0.03339291 9.08393910 + layer.3.v_cache 0.00002012 0.01962263 + layer.4.k_cache 0.00068476 0.43115565 + layer.4.v_cache 0.00005875 0.03282320 + layer.4.output 0.00481428 201.50334416 + ------------------------------------------------------------------------------------- + TOTAL 0.04342071 87.00464741 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 225984 +BPFP 0.7553 bits/point +EBPFP 1.5106 equivalent bits/point +MSE 87.004647 +---------------------- -------------------------------------------------------- +Time: 3.205s Load: 0.012s, Pack+Encode: 1.843s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 87.0046 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,468B, BPFP=0.2749 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,756B, BPFP=1.5844 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,552B, BPFP=0.5261 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,512B, BPFP=1.5079 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,376B, BPFP=0.6998 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,568B, BPFP=1.4498 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,700B, BPFP=0.6582 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,308B, BPFP=1.4953 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,460B, BPFP=1.2586 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,952B, BPFP=1.4119 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,824B, BPFP=0.0775 +⌛️ [2/4] FRONTEND: Frontend time: 1.717s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13396681 55.68156373 + layer.0.v_cache 0.00001641 0.01331699 + layer.1.k_cache 0.47111181 6.11581637 + layer.1.v_cache 0.00000608 0.00538850 + layer.2.k_cache 0.02933135 1.34065907 + layer.2.v_cache 0.00002255 0.01676235 + layer.3.k_cache 0.03744583 9.67621186 + layer.3.v_cache 0.00002260 0.01958951 + layer.4.k_cache 0.00071777 0.46424698 + layer.4.v_cache 0.00004917 0.03220097 + layer.4.output 1.20538323 212.97061305 + ------------------------------------------------------------------------------------- + TOTAL 0.53590429 92.00941457 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 185476 +BPFP 0.6712 bits/point +EBPFP 1.3423 equivalent bits/point +MSE 92.009415 +---------------------- -------------------------------------------------------- +Time: 2.950s Load: 0.009s, Pack+Encode: 1.717s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 92.0094 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,020B, BPFP=0.2724 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,428B, BPFP=1.7051 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,544B, BPFP=0.5178 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,128B, BPFP=1.6345 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,376B, BPFP=0.7799 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,348B, BPFP=1.5922 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,716B, BPFP=0.7441 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,036B, BPFP=1.6296 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,116B, BPFP=1.3626 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,444B, BPFP=1.5432 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,192B, BPFP=0.0635 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14060297 53.35778130 + layer.0.v_cache 0.00001598 0.01344747 + layer.1.k_cache 0.58610762 6.65594228 + layer.1.v_cache 0.00000669 0.00530554 + layer.2.k_cache 0.02834670 1.37649451 + layer.2.v_cache 0.00002256 0.01616114 + layer.3.k_cache 0.01962393 9.29163106 + layer.3.v_cache 0.00002003 0.01932406 + layer.4.k_cache 0.00067797 0.44472906 + layer.4.v_cache 0.00005035 0.03271386 + layer.4.output 0.00463834 192.68449281 + ------------------------------------------------------------------------------------- + TOTAL 0.04752607 83.52970470 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 225348 +BPFP 0.7192 bits/point +EBPFP 1.4383 equivalent bits/point +MSE 83.529705 +---------------------- -------------------------------------------------------- +Time: 3.198s Load: 0.012s, Pack+Encode: 1.838s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 83.5297 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,332B, BPFP=0.2741 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,752B, BPFP=1.6320 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,068B, BPFP=0.5175 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,204B, BPFP=1.5524 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,864B, BPFP=0.7640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,376B, BPFP=1.5099 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,260B, BPFP=0.7329 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,208B, BPFP=1.5526 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,360B, BPFP=1.2521 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,400B, BPFP=1.4597 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,248B, BPFP=0.0606 +⌛️ [2/4] FRONTEND: Frontend time: 1.854s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.352s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12404394 51.34552002 + layer.0.v_cache 0.00001645 0.01339109 + layer.1.k_cache 0.57408182 6.49644751 + layer.1.v_cache 0.00000616 0.00522317 + layer.2.k_cache 0.03522489 1.30605597 + layer.2.v_cache 0.00002326 0.01609913 + layer.3.k_cache 0.02908128 9.39354184 + layer.3.v_cache 0.00002078 0.01944587 + layer.4.k_cache 0.00072493 0.43418101 + layer.4.v_cache 0.00004967 0.03220735 + layer.4.output 0.04390477 178.18330005 + ------------------------------------------------------------------------------------- + TOTAL 0.06297686 77.43207137 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 227072 +BPFP 0.6865 bits/point +EBPFP 1.3731 equivalent bits/point +MSE 77.432071 +---------------------- -------------------------------------------------------- +Time: 3.216s Load: 0.010s, Pack+Encode: 1.854s, Decode+Unpack: 1.352s +---------------------- -------------------------------------------------------- +💾 Converting with 77.4321 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 242, 128) +Output shape: (1, 242, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.output: torch.Size([1, 242, 3584]) -> torch.Size([1, 1, 242, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,304B, BPFP=0.2779 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,660B, BPFP=1.6568 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,996B, BPFP=0.5163 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,476B, BPFP=1.5803 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,184B, BPFP=0.7867 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,808B, BPFP=1.5372 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,452B, BPFP=0.7394 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,416B, BPFP=1.5764 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,380B, BPFP=1.3159 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,120B, BPFP=1.4928 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,852B, BPFP=0.0724 +⌛️ [2/4] FRONTEND: Frontend time: 1.714s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11466674 53.81956192 + layer.0.v_cache 0.00001794 0.01349637 + layer.1.k_cache 0.39062487 6.45289700 + layer.1.v_cache 0.00000616 0.00531214 + layer.2.k_cache 0.02414889 1.31683224 + layer.2.v_cache 0.00002046 0.01579888 + layer.3.k_cache 0.03495786 9.45403296 + layer.3.v_cache 0.00002053 0.01966700 + layer.4.k_cache 0.00069125 0.43627845 + layer.4.v_cache 0.00005034 0.03403580 + layer.4.output 1.26514320 223.37258338 + ------------------------------------------------------------------------------------- + TOTAL 0.55418868 96.18682332 + (elements=2,106,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2106368 +Total Bytes 185648 +BPFP 0.7051 bits/point +EBPFP 1.4102 equivalent bits/point +MSE 96.186823 +---------------------- -------------------------------------------------------- +Time: 2.942s Load: 0.008s, Pack+Encode: 1.714s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 96.1868 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 189, 128) +Output shape: (1, 189, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.output: torch.Size([1, 189, 3584]) -> torch.Size([1, 1, 189, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,428B, BPFP=0.2834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,476B, BPFP=1.6101 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,240B, BPFP=0.5159 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,576B, BPFP=1.5357 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,204B, BPFP=0.7609 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,012B, BPFP=1.4891 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,860B, BPFP=0.7325 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,612B, BPFP=1.5387 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,824B, BPFP=1.3082 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,460B, BPFP=1.4435 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,236B, BPFP=0.0736 +⌛️ [2/4] FRONTEND: Frontend time: 1.619s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.107s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14361043 54.17960483 + layer.0.v_cache 0.00001648 0.01410453 + layer.1.k_cache 0.18932698 6.08608670 + layer.1.v_cache 0.00000663 0.00601249 + layer.2.k_cache 0.00768016 1.45053004 + layer.2.v_cache 0.00002453 0.01764017 + layer.3.k_cache 0.01421528 9.17862052 + layer.3.v_cache 0.00002024 0.02071209 + layer.4.k_cache 0.00072804 0.49605960 + layer.4.v_cache 0.00005086 0.03476123 + layer.4.output 0.00858909 289.95214475 + ------------------------------------------------------------------------------------- + TOTAL 0.02445902 123.59700855 + (elements=1,645,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1645056 +Total Bytes 141928 +BPFP 0.6902 bits/point +EBPFP 1.3804 equivalent bits/point +MSE 123.597009 +---------------------- -------------------------------------------------------- +Time: 2.735s Load: 0.009s, Pack+Encode: 1.619s, Decode+Unpack: 1.107s +---------------------- -------------------------------------------------------- +💾 Converting with 123.5970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,932B, BPFP=0.2884 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,532B, BPFP=1.8729 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,340B, BPFP=0.5384 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,660B, BPFP=1.8090 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,492B, BPFP=0.9164 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,756B, BPFP=1.7427 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,480B, BPFP=0.9888 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,568B, BPFP=1.8022 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,076B, BPFP=1.4727 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,308B, BPFP=1.7098 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,396B, BPFP=0.0670 +⌛️ [2/4] FRONTEND: Frontend time: 1.710s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16894778 50.92194377 + layer.0.v_cache 0.00001779 0.01399501 + layer.1.k_cache 0.37628758 6.13444698 + layer.1.v_cache 0.00000652 0.00542063 + layer.2.k_cache 0.02181128 1.19999050 + layer.2.v_cache 0.00002030 0.01634549 + layer.3.k_cache 0.01815432 8.27341859 + layer.3.v_cache 0.00002063 0.02048524 + layer.4.k_cache 0.00071780 0.43867278 + layer.4.v_cache 0.00005059 0.03375987 + layer.4.output 1.43727796 253.95701291 + ------------------------------------------------------------------------------------- + TOTAL 0.62629296 108.51515113 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 185540 +BPFP 0.8006 bits/point +EBPFP 1.6012 equivalent bits/point +MSE 108.515151 +---------------------- -------------------------------------------------------- +Time: 2.938s Load: 0.009s, Pack+Encode: 1.710s, Decode+Unpack: 1.219s +---------------------- -------------------------------------------------------- +💾 Converting with 108.5152 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,008B, BPFP=0.2899 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,348B, BPFP=1.8336 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,388B, BPFP=0.5344 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,324B, BPFP=1.7595 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,000B, BPFP=0.8681 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,488B, BPFP=1.6991 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,636B, BPFP=0.9141 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,128B, BPFP=1.7454 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,752B, BPFP=1.4288 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,740B, BPFP=1.6450 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,564B, BPFP=0.0678 +⌛️ [2/4] FRONTEND: Frontend time: 1.711s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.228s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09877178 51.13784903 + layer.0.v_cache 0.00001652 0.01350312 + layer.1.k_cache 0.28824160 5.82624308 + layer.1.v_cache 0.00000619 0.00557421 + layer.2.k_cache 0.02137197 1.26461453 + layer.2.v_cache 0.00001995 0.01616616 + layer.3.k_cache 0.03811082 8.55686894 + layer.3.v_cache 0.00002058 0.01989877 + layer.4.k_cache 0.00070031 0.44279699 + layer.4.v_cache 0.00004695 0.03226218 + layer.4.output 1.41729720 250.56084656 + ------------------------------------------------------------------------------------- + TOTAL 0.60990512 107.13186488 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 182376 +BPFP 0.7760 bits/point +EBPFP 1.5521 equivalent bits/point +MSE 107.131865 +---------------------- -------------------------------------------------------- +Time: 2.946s Load: 0.007s, Pack+Encode: 1.711s, Decode+Unpack: 1.228s +---------------------- -------------------------------------------------------- +💾 Converting with 107.1319 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,020B, BPFP=0.2779 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,440B, BPFP=1.7588 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,648B, BPFP=0.5288 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,364B, BPFP=1.6845 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,076B, BPFP=0.8349 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,564B, BPFP=1.6291 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,088B, BPFP=0.8357 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,244B, BPFP=1.6762 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,948B, BPFP=1.3791 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,936B, BPFP=1.5857 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,548B, BPFP=0.0745 +⌛️ [2/4] FRONTEND: Frontend time: 1.721s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.224s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12161526 54.22588841 + layer.0.v_cache 0.00001916 0.01358510 + layer.1.k_cache 0.26028682 5.94005916 + layer.1.v_cache 0.00000628 0.00537694 + layer.2.k_cache 0.01296707 1.21371109 + layer.2.v_cache 0.00002040 0.01578770 + layer.3.k_cache 0.04595671 8.38015990 + layer.3.v_cache 0.00002073 0.01884666 + layer.4.k_cache 0.00067352 0.45238934 + layer.4.v_cache 0.00005047 0.03315558 + layer.4.output 1.35465500 239.46254741 + ------------------------------------------------------------------------------------- + TOTAL 0.58377655 102.73745834 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 183876 +BPFP 0.7478 bits/point +EBPFP 1.4956 equivalent bits/point +MSE 102.737458 +---------------------- -------------------------------------------------------- +Time: 2.952s Load: 0.008s, Pack+Encode: 1.721s, Decode+Unpack: 1.224s +---------------------- -------------------------------------------------------- +💾 Converting with 102.7375 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,024B, BPFP=0.2782 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,388B, BPFP=1.7553 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,584B, BPFP=0.5243 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,296B, BPFP=1.6798 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,324B, BPFP=0.8520 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,656B, BPFP=1.6355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,396B, BPFP=0.8570 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,244B, BPFP=1.6762 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,120B, BPFP=1.3219 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,908B, BPFP=1.5838 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,864B, BPFP=0.0678 +⌛️ [2/4] FRONTEND: Frontend time: 1.709s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10229862 55.88674468 + layer.0.v_cache 0.00001727 0.01336967 + layer.1.k_cache 0.34669285 5.85086627 + layer.1.v_cache 0.00000643 0.00540853 + layer.2.k_cache 0.01639456 1.22131699 + layer.2.v_cache 0.00002164 0.01613884 + layer.3.k_cache 0.03343437 8.59183360 + layer.3.v_cache 0.00002043 0.01927541 + layer.4.k_cache 0.00071218 0.43723817 + layer.4.v_cache 0.00005088 0.03298949 + layer.4.output 1.35462105 239.45583123 + ------------------------------------------------------------------------------------- + TOTAL 0.58717627 102.83917648 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 182804 +BPFP 0.7434 bits/point +EBPFP 1.4869 equivalent bits/point +MSE 102.839176 +---------------------- -------------------------------------------------------- +Time: 2.939s Load: 0.010s, Pack+Encode: 1.709s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 102.8392 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,040B, BPFP=0.2781 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,424B, BPFP=1.7500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,452B, BPFP=0.5129 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,208B, BPFP=1.6663 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,540B, BPFP=0.8632 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,584B, BPFP=1.6233 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,168B, BPFP=0.8376 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,324B, BPFP=1.6743 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,592B, BPFP=1.3486 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,868B, BPFP=1.5741 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,304B, BPFP=0.0620 +⌛️ [2/4] FRONTEND: Frontend time: 1.712s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11800324 53.53310848 + layer.0.v_cache 0.00001819 0.01368738 + layer.1.k_cache 0.35011090 6.48622541 + layer.1.v_cache 0.00000612 0.00530900 + layer.2.k_cache 0.02017111 1.24858624 + layer.2.v_cache 0.00002006 0.01597498 + layer.3.k_cache 0.03026835 9.08732491 + layer.3.v_cache 0.00001975 0.01917205 + layer.4.k_cache 0.00074701 0.44878949 + layer.4.v_cache 0.00005118 0.03245544 + layer.4.output 1.34863711 238.55848804 + ------------------------------------------------------------------------------------- + TOTAL 0.58587504 102.40000292 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 182504 +BPFP 0.7390 bits/point +EBPFP 1.4779 equivalent bits/point +MSE 102.400003 +---------------------- -------------------------------------------------------- +Time: 2.944s Load: 0.010s, Pack+Encode: 1.712s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 102.4000 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,136B, BPFP=0.2822 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,432B, BPFP=1.7353 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,792B, BPFP=0.5317 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,524B, BPFP=1.6733 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,804B, BPFP=0.8736 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,920B, BPFP=1.6321 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,932B, BPFP=0.8141 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,716B, BPFP=1.6864 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,868B, BPFP=1.4239 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,412B, BPFP=1.5974 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,116B, BPFP=0.0596 +⌛️ [2/4] FRONTEND: Frontend time: 1.707s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10957497 53.52681922 + layer.0.v_cache 0.00001697 0.01385042 + layer.1.k_cache 0.31262030 5.81398777 + layer.1.v_cache 0.00000629 0.00561300 + layer.2.k_cache 0.01311818 1.46215794 + layer.2.v_cache 0.00002149 0.01695590 + layer.3.k_cache 0.06227440 8.94530184 + layer.3.v_cache 0.00002139 0.02177316 + layer.4.k_cache 0.00068925 0.46039018 + layer.4.v_cache 0.00005302 0.03640815 + layer.4.output 1.33688878 236.37654008 + ------------------------------------------------------------------------------------- + TOTAL 0.57980104 101.46700224 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 185652 +BPFP 0.7451 bits/point +EBPFP 1.4903 equivalent bits/point +MSE 101.467002 +---------------------- -------------------------------------------------------- +Time: 2.935s Load: 0.010s, Pack+Encode: 1.707s, Decode+Unpack: 1.219s +---------------------- -------------------------------------------------------- +💾 Converting with 101.4670 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,896B, BPFP=0.2858 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,488B, BPFP=1.8697 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,692B, BPFP=0.5643 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,300B, BPFP=1.7826 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,444B, BPFP=0.9129 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,712B, BPFP=1.7394 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,120B, BPFP=0.8891 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,332B, BPFP=1.7849 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,240B, BPFP=1.4114 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,144B, BPFP=1.6978 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,732B, BPFP=0.0705 +⌛️ [2/4] FRONTEND: Frontend time: 1.712s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13987277 51.92329170 + layer.0.v_cache 0.00001887 0.01409135 + layer.1.k_cache 0.23498858 5.99405466 + layer.1.v_cache 0.00000603 0.00532377 + layer.2.k_cache 0.02028183 1.38441546 + layer.2.v_cache 0.00001992 0.01594678 + layer.3.k_cache 0.05552547 8.94644838 + layer.3.v_cache 0.00002044 0.01973978 + layer.4.k_cache 0.00067773 0.43169518 + layer.4.v_cache 0.00004777 0.03308033 + layer.4.output 1.43727485 254.07478203 + ------------------------------------------------------------------------------------- + TOTAL 0.61837549 108.66420950 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 183100 +BPFP 0.7901 bits/point +EBPFP 1.5802 equivalent bits/point +MSE 108.664210 +---------------------- -------------------------------------------------------- +Time: 2.940s Load: 0.007s, Pack+Encode: 1.712s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 108.6642 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,868B, BPFP=0.2864 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,136B, BPFP=1.8614 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,132B, BPFP=0.5281 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,404B, BPFP=1.8072 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,596B, BPFP=0.8587 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,584B, BPFP=1.7464 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,616B, BPFP=0.9342 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,972B, BPFP=1.7752 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,500B, BPFP=1.4440 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,988B, BPFP=1.7023 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,416B, BPFP=0.0679 +⌛️ [2/4] FRONTEND: Frontend time: 1.723s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17004660 48.03889125 + layer.0.v_cache 0.00001740 0.01348949 + layer.1.k_cache 0.16740718 6.17037602 + layer.1.v_cache 0.00000654 0.00552469 + layer.2.k_cache 0.01191964 1.16382205 + layer.2.v_cache 0.00001991 0.01573859 + layer.3.k_cache 0.00660476 8.96470548 + layer.3.v_cache 0.00002076 0.01926983 + layer.4.k_cache 0.00070551 0.46294190 + layer.4.v_cache 0.00005179 0.03308875 + layer.4.output 1.45088344 256.32246530 + ------------------------------------------------------------------------------------- + TOTAL 0.61841083 109.36147677 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 181212 +BPFP 0.7894 bits/point +EBPFP 1.5787 equivalent bits/point +MSE 109.361477 +---------------------- -------------------------------------------------------- +Time: 2.952s Load: 0.007s, Pack+Encode: 1.723s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 109.3615 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,200B, BPFP=0.2853 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,092B, BPFP=1.7046 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,656B, BPFP=0.5201 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,172B, BPFP=1.6421 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,080B, BPFP=0.8207 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,572B, BPFP=1.6014 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,852B, BPFP=0.8052 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,004B, BPFP=1.6307 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,164B, BPFP=1.3698 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,676B, BPFP=1.5405 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,716B, BPFP=0.0652 +⌛️ [2/4] FRONTEND: Frontend time: 1.717s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14712874 52.87806131 + layer.0.v_cache 0.00001826 0.01313249 + layer.1.k_cache 0.36788522 6.52095470 + layer.1.v_cache 0.00000598 0.00531125 + layer.2.k_cache 0.01156470 1.44024287 + layer.2.v_cache 0.00002030 0.01599372 + layer.3.k_cache 0.01315050 8.44767164 + layer.3.v_cache 0.00002017 0.01923716 + layer.4.k_cache 0.00068370 0.43975090 + layer.4.v_cache 0.00004817 0.03214010 + layer.4.output 1.33105469 235.20419255 + ------------------------------------------------------------------------------------- + TOTAL 0.57987697 100.95540259 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 182184 +BPFP 0.7280 bits/point +EBPFP 1.4561 equivalent bits/point +MSE 100.955403 +---------------------- -------------------------------------------------------- +Time: 2.944s Load: 0.008s, Pack+Encode: 1.717s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 100.9554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 188, 128) +Output shape: (1, 188, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.output: torch.Size([1, 188, 3584]) -> torch.Size([1, 1, 188, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,396B, BPFP=0.2822 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,148B, BPFP=1.5914 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,148B, BPFP=0.5110 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,288B, BPFP=1.5199 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,304B, BPFP=0.7733 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,768B, BPFP=1.4767 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,712B, BPFP=0.7241 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,196B, BPFP=1.5123 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,808B, BPFP=1.3138 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,304B, BPFP=1.4382 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,844B, BPFP=0.0694 +⌛️ [2/4] FRONTEND: Frontend time: 1.600s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.111s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005002 54.18003034 + layer.0.v_cache 0.00001663 0.01402822 + layer.1.k_cache 0.14149688 5.85344680 + layer.1.v_cache 0.00000584 0.00531257 + layer.2.k_cache 0.00881058 1.41434268 + layer.2.v_cache 0.00002285 0.01713162 + layer.3.k_cache 0.03192479 8.48779622 + layer.3.v_cache 0.00001958 0.01994718 + layer.4.k_cache 0.00066612 0.45560041 + layer.4.v_cache 0.00005162 0.03588858 + layer.4.output 0.00857985 291.73166793 + ------------------------------------------------------------------------------------- + TOTAL 0.02253670 124.27089413 + (elements=1,636,352) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1636352 +Total Bytes 139916 +BPFP 0.6840 bits/point +EBPFP 1.3681 equivalent bits/point +MSE 124.270894 +---------------------- -------------------------------------------------------- +Time: 2.718s Load: 0.006s, Pack+Encode: 1.600s, Decode+Unpack: 1.111s +---------------------- -------------------------------------------------------- +💾 Converting with 124.2709 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,728B, BPFP=0.2927 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,872B, BPFP=1.9529 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,872B, BPFP=0.5396 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,128B, BPFP=1.8945 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,596B, BPFP=0.9890 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,376B, BPFP=1.8354 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,876B, BPFP=0.9325 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,824B, BPFP=1.8706 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,880B, BPFP=1.5609 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,804B, BPFP=1.7905 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,140B, BPFP=0.0689 +⌛️ [2/4] FRONTEND: Frontend time: 1.715s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11435634 54.05362751 + layer.0.v_cache 0.00001764 0.01468345 + layer.1.k_cache 0.25025516 6.76053485 + layer.1.v_cache 0.00000690 0.00557560 + layer.2.k_cache 0.01528106 1.32394731 + layer.2.v_cache 0.00002105 0.01683116 + layer.3.k_cache 0.02533426 8.22859292 + layer.3.v_cache 0.00002127 0.02043501 + layer.4.k_cache 0.00070809 0.44537434 + layer.4.v_cache 0.00005072 0.03417259 + layer.4.output 1.53833902 271.98745962 + ------------------------------------------------------------------------------------- + TOTAL 0.65731915 116.16564659 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 180096 +BPFP 0.8318 bits/point +EBPFP 1.6636 equivalent bits/point +MSE 116.165647 +---------------------- -------------------------------------------------------- +Time: 2.941s Load: 0.007s, Pack+Encode: 1.715s, Decode+Unpack: 1.219s +---------------------- -------------------------------------------------------- +💾 Converting with 116.1656 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,344B, BPFP=0.2903 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,524B, BPFP=1.6948 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,980B, BPFP=0.5191 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,592B, BPFP=1.6139 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,336B, BPFP=0.8104 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,236B, BPFP=1.5830 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,396B, BPFP=0.8156 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,744B, BPFP=1.6271 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,944B, BPFP=1.3840 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,772B, BPFP=1.5427 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,676B, BPFP=0.0704 +⌛️ [2/4] FRONTEND: Frontend time: 1.614s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.110s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258550 53.29735243 + layer.0.v_cache 0.00001767 0.01464928 + layer.1.k_cache 0.14790291 6.49765761 + layer.1.v_cache 0.00000629 0.00561711 + layer.2.k_cache 0.00999225 1.36842838 + layer.2.v_cache 0.00002110 0.01680319 + layer.3.k_cache 0.02117847 8.78760444 + layer.3.v_cache 0.00002045 0.02147260 + layer.4.k_cache 0.00067024 0.45627675 + layer.4.v_cache 0.00006105 0.03537784 + layer.4.output 0.00897616 304.60141369 + ------------------------------------------------------------------------------------- + TOTAL 0.02207583 129.57124326 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 142544 +BPFP 0.7279 bits/point +EBPFP 1.4557 equivalent bits/point +MSE 129.571243 +---------------------- -------------------------------------------------------- +Time: 2.730s Load: 0.006s, Pack+Encode: 1.614s, Decode+Unpack: 1.110s +---------------------- -------------------------------------------------------- +💾 Converting with 129.5712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,808B, BPFP=0.2917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,232B, BPFP=1.9326 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,984B, BPFP=0.5349 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,148B, BPFP=1.8496 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,096B, BPFP=0.9265 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,476B, BPFP=1.7981 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,376B, BPFP=1.0245 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,148B, BPFP=1.8496 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,856B, BPFP=1.4442 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,992B, BPFP=1.7610 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,184B, BPFP=0.0677 +⌛️ [2/4] FRONTEND: Frontend time: 1.727s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.238s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09364704 52.62048579 + layer.0.v_cache 0.00001853 0.01398784 + layer.1.k_cache 0.25163280 6.31674015 + layer.1.v_cache 0.00000592 0.00548261 + layer.2.k_cache 0.01595125 1.06159764 + layer.2.v_cache 0.00002050 0.01639347 + layer.3.k_cache 0.04325477 8.16136140 + layer.3.v_cache 0.00002016 0.01985253 + layer.4.k_cache 0.00069210 0.44643739 + layer.4.v_cache 0.00005026 0.03447524 + layer.4.output 1.50065788 265.28400735 + ------------------------------------------------------------------------------------- + TOTAL 0.64175873 113.27558032 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 181300 +BPFP 0.8168 bits/point +EBPFP 1.6337 equivalent bits/point +MSE 113.275580 +---------------------- -------------------------------------------------------- +Time: 2.972s Load: 0.007s, Pack+Encode: 1.727s, Decode+Unpack: 1.238s +---------------------- -------------------------------------------------------- +💾 Converting with 113.2756 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,008B, BPFP=0.2771 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,456B, BPFP=1.7600 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,480B, BPFP=0.5171 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,208B, BPFP=1.6737 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,196B, BPFP=0.7741 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,592B, BPFP=1.6311 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,288B, BPFP=0.7804 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,140B, BPFP=1.6690 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,372B, BPFP=1.3393 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,800B, BPFP=1.5763 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,912B, BPFP=0.0781 +⌛️ [2/4] FRONTEND: Frontend time: 1.722s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.231s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510114 54.96671910 + layer.0.v_cache 0.00001717 0.01331110 + layer.1.k_cache 0.31797088 6.79965615 + layer.1.v_cache 0.00000673 0.00545950 + layer.2.k_cache 0.01711024 1.03016122 + layer.2.v_cache 0.00002022 0.01642379 + layer.3.k_cache 0.01307247 7.48173158 + layer.3.v_cache 0.00002089 0.01958964 + layer.4.k_cache 0.00069413 0.45609456 + layer.4.v_cache 0.00005461 0.03455802 + layer.4.output 1.35466395 239.55080594 + ------------------------------------------------------------------------------------- + TOTAL 0.58392448 102.80466743 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 181452 +BPFP 0.7379 bits/point +EBPFP 1.4759 equivalent bits/point +MSE 102.804667 +---------------------- -------------------------------------------------------- +Time: 2.960s Load: 0.008s, Pack+Encode: 1.722s, Decode+Unpack: 1.231s +---------------------- -------------------------------------------------------- +💾 Converting with 102.8047 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,440B, BPFP=0.2797 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,648B, BPFP=1.6159 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,164B, BPFP=0.5144 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,720B, BPFP=1.5575 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,624B, BPFP=0.7954 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,996B, BPFP=1.5118 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,748B, BPFP=0.7402 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,556B, BPFP=1.5471 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,476B, BPFP=1.2901 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,188B, BPFP=1.4609 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,516B, BPFP=0.0766 +⌛️ [2/4] FRONTEND: Frontend time: 1.718s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.229s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14860459 52.38276525 + layer.0.v_cache 0.00001629 0.01365463 + layer.1.k_cache 0.38266554 6.52323766 + layer.1.v_cache 0.00000669 0.00574419 + layer.2.k_cache 0.01628031 1.30206274 + layer.2.v_cache 0.00002106 0.01683562 + layer.3.k_cache 0.01842327 9.19788779 + layer.3.v_cache 0.00002289 0.02096145 + layer.4.k_cache 0.00072690 0.44280757 + layer.4.v_cache 0.00004923 0.03398981 + layer.4.output 1.23454798 218.19943476 + ------------------------------------------------------------------------------------- + TOTAL 0.54168545 93.96094059 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 188076 +BPFP 0.6970 bits/point +EBPFP 1.3941 equivalent bits/point +MSE 93.960941 +---------------------- -------------------------------------------------------- +Time: 2.956s Load: 0.008s, Pack+Encode: 1.718s, Decode+Unpack: 1.229s +---------------------- -------------------------------------------------------- +💾 Converting with 93.9609 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 435, 128) +Output shape: (1, 435, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.output: torch.Size([1, 435, 3584]) -> torch.Size([1, 1, 435, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,408B, BPFP=0.2661 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,600B, BPFP=1.6020 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,256B, BPFP=0.5121 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,588B, BPFP=1.5297 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,780B, BPFP=0.7105 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,300B, BPFP=1.4835 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,124B, BPFP=0.6869 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,920B, BPFP=1.5057 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,392B, BPFP=1.2353 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,696B, BPFP=1.4259 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,276B, BPFP=0.0681 +⌛️ [2/4] FRONTEND: Frontend time: 2.119s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.603s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15864419 53.15616020 + layer.0.v_cache 0.00001607 0.01314892 + layer.1.k_cache 0.94329666 6.17429396 + layer.1.v_cache 0.00000639 0.00530503 + layer.2.k_cache 0.02505282 1.20384283 + layer.2.v_cache 0.00002131 0.01568735 + layer.3.k_cache 0.01724711 9.09496621 + layer.3.v_cache 0.00002032 0.01824785 + layer.4.k_cache 0.00074513 0.44357693 + layer.4.v_cache 0.00005173 0.03104904 + layer.4.output 0.00608059 127.29926108 + ------------------------------------------------------------------------------------- + TOTAL 0.06986270 56.54418270 + (elements=3,786,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3786240 +Total Bytes 318340 +BPFP 0.6726 bits/point +EBPFP 1.3453 equivalent bits/point +MSE 56.544183 +---------------------- -------------------------------------------------------- +Time: 3.736s Load: 0.014s, Pack+Encode: 2.119s, Decode+Unpack: 1.603s +---------------------- -------------------------------------------------------- +💾 Converting with 56.5442 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 437, 128) +Output shape: (1, 437, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.output: torch.Size([1, 437, 3584]) -> torch.Size([1, 1, 437, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,452B, BPFP=0.2664 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,680B, BPFP=1.5975 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,264B, BPFP=0.5100 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,040B, BPFP=1.5389 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,508B, BPFP=0.7333 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,424B, BPFP=1.4811 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,236B, BPFP=0.6878 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,660B, BPFP=1.5253 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,072B, BPFP=1.2540 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 40,236B, BPFP=1.4386 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,736B, BPFP=0.0855 +⌛️ [2/4] FRONTEND: Frontend time: 2.097s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.588s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14454122 53.26357355 + layer.0.v_cache 0.00001702 0.01287381 + layer.1.k_cache 0.94267751 6.43961123 + layer.1.v_cache 0.00000668 0.00542517 + layer.2.k_cache 0.01808156 1.21171406 + layer.2.v_cache 0.00002154 0.01584381 + layer.3.k_cache 0.03034366 9.08811186 + layer.3.v_cache 0.00002210 0.01949477 + layer.4.k_cache 0.00074160 0.45140659 + layer.4.v_cache 0.00005552 0.03344059 + layer.4.output 0.00611241 126.69017653 + ------------------------------------------------------------------------------------- + TOTAL 0.06937031 56.31604301 + (elements=3,803,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3803648 +Total Bytes 325308 +BPFP 0.6842 bits/point +EBPFP 1.3684 equivalent bits/point +MSE 56.316043 +---------------------- -------------------------------------------------------- +Time: 3.700s Load: 0.014s, Pack+Encode: 2.097s, Decode+Unpack: 1.588s +---------------------- -------------------------------------------------------- +💾 Converting with 56.3160 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,396B, BPFP=0.2737 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,064B, BPFP=1.6266 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,088B, BPFP=0.5118 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,352B, BPFP=1.5398 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,512B, BPFP=0.7362 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,576B, BPFP=1.5004 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,080B, BPFP=0.7143 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,624B, BPFP=1.5536 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,972B, BPFP=1.2668 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,664B, BPFP=1.4541 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,736B, BPFP=0.0851 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.358s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12827574 53.34302075 + layer.0.v_cache 0.00001866 0.01364862 + layer.1.k_cache 0.53653341 6.10678101 + layer.1.v_cache 0.00000626 0.00531232 + layer.2.k_cache 0.01950284 1.10883064 + layer.2.v_cache 0.00002083 0.01572589 + layer.3.k_cache 0.04340698 8.73002961 + layer.3.v_cache 0.00002181 0.01990688 + layer.4.k_cache 0.00072091 0.44255071 + layer.4.v_cache 0.00005393 0.03313768 + layer.4.output 0.04342448 175.93850012 + ------------------------------------------------------------------------------------- + TOTAL 0.06073722 76.55226147 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 232064 +BPFP 0.6925 bits/point +EBPFP 1.3850 equivalent bits/point +MSE 76.552261 +---------------------- -------------------------------------------------------- +Time: 3.208s Load: 0.012s, Pack+Encode: 1.838s, Decode+Unpack: 1.358s +---------------------- -------------------------------------------------------- +💾 Converting with 76.5523 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,808B, BPFP=0.2824 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,376B, BPFP=1.8430 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,204B, BPFP=0.5406 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,336B, BPFP=1.7820 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,588B, BPFP=0.9156 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,412B, BPFP=1.7277 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,776B, BPFP=0.8680 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,968B, BPFP=1.7603 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,968B, BPFP=1.4079 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,716B, BPFP=1.6868 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,812B, BPFP=0.0739 +⌛️ [2/4] FRONTEND: Frontend time: 1.839s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702504 53.95882284 + layer.0.v_cache 0.00001641 0.01371852 + layer.1.k_cache 0.51130361 6.30443693 + layer.1.v_cache 0.00000656 0.00564526 + layer.2.k_cache 0.01810413 1.33295699 + layer.2.v_cache 0.00002122 0.01646556 + layer.3.k_cache 0.02476560 8.93471349 + layer.3.v_cache 0.00002197 0.02005550 + layer.4.k_cache 0.00071356 0.44827159 + layer.4.v_cache 0.00005290 0.03424073 + layer.4.output 0.00501064 208.20122852 + ------------------------------------------------------------------------------------- + TOTAL 0.04100620 89.91046630 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 226964 +BPFP 0.7842 bits/point +EBPFP 1.5685 equivalent bits/point +MSE 89.910466 +---------------------- -------------------------------------------------------- +Time: 3.202s Load: 0.012s, Pack+Encode: 1.839s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 89.9105 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,240B, BPFP=0.2831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,576B, BPFP=1.7078 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,816B, BPFP=0.5219 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,636B, BPFP=1.6450 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,432B, BPFP=0.8301 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,944B, BPFP=1.5988 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,856B, BPFP=0.7917 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,604B, BPFP=1.6429 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,008B, BPFP=1.3360 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,148B, BPFP=1.5457 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,672B, BPFP=0.0827 +⌛️ [2/4] FRONTEND: Frontend time: 1.729s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.226s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11892304 53.29614884 + layer.0.v_cache 0.00001803 0.01398308 + layer.1.k_cache 0.31174104 6.06580738 + layer.1.v_cache 0.00000657 0.00553845 + layer.2.k_cache 0.01452858 1.23513155 + layer.2.v_cache 0.00002161 0.01662262 + layer.3.k_cache 0.02263702 8.67052701 + layer.3.v_cache 0.00002141 0.02075252 + layer.4.k_cache 0.00068312 0.45462427 + layer.4.v_cache 0.00005251 0.03464224 + layer.4.output 1.30838121 231.20503281 + ------------------------------------------------------------------------------------- + TOTAL 0.56631185 99.30876516 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 186932 +BPFP 0.7342 bits/point +EBPFP 1.4685 equivalent bits/point +MSE 99.308765 +---------------------- -------------------------------------------------------- +Time: 2.965s Load: 0.010s, Pack+Encode: 1.729s, Decode+Unpack: 1.226s +---------------------- -------------------------------------------------------- +💾 Converting with 99.3088 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 233, 128) +Output shape: (1, 233, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.output: torch.Size([1, 233, 3584]) -> torch.Size([1, 1, 233, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,184B, BPFP=0.2806 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,564B, BPFP=1.7143 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,808B, BPFP=0.5236 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,576B, BPFP=1.6481 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,524B, BPFP=0.8399 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,720B, BPFP=1.5907 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,056B, BPFP=0.8085 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,304B, BPFP=1.6298 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,080B, BPFP=1.3466 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,060B, BPFP=1.5464 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,192B, BPFP=0.0785 +⌛️ [2/4] FRONTEND: Frontend time: 1.725s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13365312 53.74811394 + layer.0.v_cache 0.00001735 0.01396512 + layer.1.k_cache 0.40872042 6.24822238 + layer.1.v_cache 0.00000792 0.00581553 + layer.2.k_cache 0.01308113 1.34464418 + layer.2.v_cache 0.00002137 0.01653686 + layer.3.k_cache 0.00601081 8.61837048 + layer.3.v_cache 0.00002094 0.01932395 + layer.4.k_cache 0.00068546 0.45657859 + layer.4.v_cache 0.00005338 0.03473568 + layer.4.output 1.31399239 232.02542535 + ------------------------------------------------------------------------------------- + TOTAL 0.57413051 99.68731083 + (elements=2,028,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2028032 +Total Bytes 186068 +BPFP 0.7340 bits/point +EBPFP 1.4680 equivalent bits/point +MSE 99.687311 +---------------------- -------------------------------------------------------- +Time: 2.958s Load: 0.010s, Pack+Encode: 1.725s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 99.6873 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 148, 128) +Output shape: (1, 148, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.output: torch.Size([1, 148, 3584]) -> torch.Size([1, 1, 148, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,892B, BPFP=0.3053 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,180B, BPFP=2.0249 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,148B, BPFP=0.5435 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,768B, BPFP=1.9814 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,848B, BPFP=1.0397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,268B, BPFP=1.9286 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,144B, BPFP=0.9654 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,360B, BPFP=1.9383 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,916B, BPFP=1.6803 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,688B, BPFP=1.8674 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,708B, BPFP=0.0710 +⌛️ [2/4] FRONTEND: Frontend time: 1.614s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.115s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08608804 52.57627085 + layer.0.v_cache 0.00001767 0.01439079 + layer.1.k_cache 0.12276106 6.01946032 + layer.1.v_cache 0.00000594 0.00552565 + layer.2.k_cache 0.00886028 1.04566079 + layer.2.v_cache 0.00002088 0.01684442 + layer.3.k_cache 0.01513371 7.31689536 + layer.3.v_cache 0.00002085 0.02119784 + layer.4.k_cache 0.00068430 0.44640206 + layer.4.v_cache 0.00005331 0.03505216 + layer.4.output 0.08957551 365.99496260 + ------------------------------------------------------------------------------------- + TOTAL 0.05062792 154.67426108 + (elements=1,288,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1288192 +Total Bytes 139920 +BPFP 0.8689 bits/point +EBPFP 1.7379 equivalent bits/point +MSE 154.674261 +---------------------- -------------------------------------------------------- +Time: 2.736s Load: 0.006s, Pack+Encode: 1.614s, Decode+Unpack: 1.115s +---------------------- -------------------------------------------------------- +💾 Converting with 154.6743 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 170, 128) +Output shape: (1, 170, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.output: torch.Size([1, 170, 3584]) -> torch.Size([1, 1, 170, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,252B, BPFP=0.2989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,312B, BPFP=1.7750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,828B, BPFP=0.5357 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,636B, BPFP=1.7129 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,952B, BPFP=0.9147 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,904B, BPFP=1.6456 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,308B, BPFP=0.8555 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,728B, BPFP=1.7213 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,444B, BPFP=1.4195 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,384B, BPFP=1.5978 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,948B, BPFP=0.0781 +⌛️ [2/4] FRONTEND: Frontend time: 1.626s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.115s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13519069 51.71663028 + layer.0.v_cache 0.00001779 0.01463228 + layer.1.k_cache 0.14803200 6.92481905 + layer.1.v_cache 0.00000678 0.00564957 + layer.2.k_cache 0.02267574 1.40643095 + layer.2.v_cache 0.00002005 0.01609102 + layer.3.k_cache 0.04618364 8.61988813 + layer.3.v_cache 0.00002282 0.02045659 + layer.4.k_cache 0.00077641 0.46242824 + layer.4.v_cache 0.00005264 0.03516958 + layer.4.output 0.00945594 322.61959034 + ------------------------------------------------------------------------------------- + TOTAL 0.02465707 136.91525459 + (elements=1,479,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1479680 +Total Bytes 141696 +BPFP 0.7661 bits/point +EBPFP 1.5322 equivalent bits/point +MSE 136.915255 +---------------------- -------------------------------------------------------- +Time: 2.750s Load: 0.008s, Pack+Encode: 1.626s, Decode+Unpack: 1.115s +---------------------- -------------------------------------------------------- +💾 Converting with 136.9153 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,136B, BPFP=0.2952 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,380B, BPFP=1.8242 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,720B, BPFP=0.5384 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,524B, BPFP=1.7436 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,012B, BPFP=0.9424 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,080B, BPFP=1.7018 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,000B, BPFP=0.8471 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,668B, BPFP=1.7572 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,524B, BPFP=1.4612 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,488B, BPFP=1.6461 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,108B, BPFP=0.0687 +⌛️ [2/4] FRONTEND: Frontend time: 1.628s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.115s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12520085 55.23358081 + layer.0.v_cache 0.00001716 0.01500802 + layer.1.k_cache 0.15152623 6.48943504 + layer.1.v_cache 0.00000633 0.00570339 + layer.2.k_cache 0.01605446 1.42811823 + layer.2.v_cache 0.00002200 0.01763354 + layer.3.k_cache 0.04332644 8.31097265 + layer.3.v_cache 0.00002349 0.02137884 + layer.4.k_cache 0.00069336 0.47903566 + layer.4.v_cache 0.00005614 0.03553758 + layer.4.output 0.00966471 330.41429109 + ------------------------------------------------------------------------------------- + TOTAL 0.02379879 140.29037891 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 140640 +BPFP 0.7787 bits/point +EBPFP 1.5574 equivalent bits/point +MSE 140.290379 +---------------------- -------------------------------------------------------- +Time: 2.749s Load: 0.006s, Pack+Encode: 1.628s, Decode+Unpack: 1.115s +---------------------- -------------------------------------------------------- +💾 Converting with 140.2904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 181, 128) +Output shape: (1, 181, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.output: torch.Size([1, 181, 3584]) -> torch.Size([1, 1, 181, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,356B, BPFP=0.2897 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,448B, BPFP=1.6789 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,116B, BPFP=0.5280 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,680B, BPFP=1.6126 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,056B, BPFP=0.7818 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,124B, BPFP=1.5646 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,884B, BPFP=0.7669 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,700B, BPFP=1.6143 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,000B, BPFP=1.3812 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,628B, BPFP=1.5218 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,112B, BPFP=0.0754 +⌛️ [2/4] FRONTEND: Frontend time: 1.633s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.118s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13566811 51.47030387 + layer.0.v_cache 0.00002016 0.01472500 + layer.1.k_cache 0.14782082 6.25170966 + layer.1.v_cache 0.00000634 0.00582806 + layer.2.k_cache 0.01546242 1.35623970 + layer.2.v_cache 0.00002121 0.01677981 + layer.3.k_cache 0.02895848 8.30693939 + layer.3.v_cache 0.00002197 0.02146877 + layer.4.k_cache 0.00067433 0.46835163 + layer.4.v_cache 0.00005229 0.03520918 + layer.4.output 0.00892717 302.72565608 + ------------------------------------------------------------------------------------- + TOTAL 0.02301155 128.64865574 + (elements=1,575,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1575424 +Total Bytes 142104 +BPFP 0.7216 bits/point +EBPFP 1.4432 equivalent bits/point +MSE 128.648656 +---------------------- -------------------------------------------------------- +Time: 2.758s Load: 0.006s, Pack+Encode: 1.633s, Decode+Unpack: 1.118s +---------------------- -------------------------------------------------------- +💾 Converting with 128.6487 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 186, 128) +Output shape: (1, 186, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.output: torch.Size([1, 186, 3584]) -> torch.Size([1, 1, 186, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,380B, BPFP=0.2839 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,628B, BPFP=1.6489 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,232B, BPFP=0.5235 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,540B, BPFP=1.5575 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,396B, BPFP=0.8733 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,828B, BPFP=1.4976 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,140B, BPFP=0.7678 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,592B, BPFP=1.5618 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,912B, BPFP=1.3367 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,440B, BPFP=1.4651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,084B, BPFP=0.0730 +⌛️ [2/4] FRONTEND: Frontend time: 1.620s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.112s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09795722 53.42891045 + layer.0.v_cache 0.00001815 0.01452473 + layer.1.k_cache 0.13430934 5.59581535 + layer.1.v_cache 0.00000632 0.00586899 + layer.2.k_cache 0.01830538 1.29048600 + layer.2.v_cache 0.00002120 0.01711878 + layer.3.k_cache 0.04892764 6.13215227 + layer.3.v_cache 0.00002113 0.02117180 + layer.4.k_cache 0.00067235 0.45454854 + layer.4.v_cache 0.00008605 0.03533755 + layer.4.output 0.00869856 295.11791955 + ------------------------------------------------------------------------------------- + TOTAL 0.02124792 125.46008066 + (elements=1,618,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1618944 +Total Bytes 143172 +BPFP 0.7075 bits/point +EBPFP 1.4150 equivalent bits/point +MSE 125.460081 +---------------------- -------------------------------------------------------- +Time: 2.739s Load: 0.006s, Pack+Encode: 1.620s, Decode+Unpack: 1.112s +---------------------- -------------------------------------------------------- +💾 Converting with 125.4601 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 163, 128) +Output shape: (1, 163, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.output: torch.Size([1, 163, 3584]) -> torch.Size([1, 1, 163, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,028B, BPFP=0.2903 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,176B, BPFP=1.8382 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,564B, BPFP=0.5334 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,356B, BPFP=1.7596 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,528B, BPFP=1.0092 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,784B, BPFP=1.7048 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,044B, BPFP=0.8669 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,416B, BPFP=1.7653 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,476B, BPFP=1.4835 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,328B, BPFP=1.6610 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,076B, BPFP=0.0695 +⌛️ [2/4] FRONTEND: Frontend time: 1.639s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.114s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11530171 55.07553681 + layer.0.v_cache 0.00001813 0.01488548 + layer.1.k_cache 0.14319220 6.29339038 + layer.1.v_cache 0.00000612 0.00574024 + layer.2.k_cache 0.02368733 1.50321109 + layer.2.v_cache 0.00002106 0.01637406 + layer.3.k_cache 0.02525310 9.06091758 + layer.3.v_cache 0.00002161 0.02166427 + layer.4.k_cache 0.00069984 0.45849544 + layer.4.v_cache 0.00005709 0.03660077 + layer.4.output 0.00982084 336.78286591 + ------------------------------------------------------------------------------------- + TOTAL 0.02217671 142.93922809 + (elements=1,418,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1418752 +Total Bytes 139776 +BPFP 0.7882 bits/point +EBPFP 1.5763 equivalent bits/point +MSE 142.939228 +---------------------- -------------------------------------------------------- +Time: 2.759s Load: 0.006s, Pack+Encode: 1.639s, Decode+Unpack: 1.114s +---------------------- -------------------------------------------------------- +💾 Converting with 142.9392 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 191, 128) +Output shape: (1, 191, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.output: torch.Size([1, 191, 3584]) -> torch.Size([1, 1, 191, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,388B, BPFP=0.2772 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,508B, BPFP=1.5959 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,484B, BPFP=0.5304 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,724B, BPFP=1.5317 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,400B, BPFP=0.6872 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,040B, BPFP=1.4758 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,988B, BPFP=0.6535 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,608B, BPFP=1.5223 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,436B, BPFP=1.3446 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,560B, BPFP=1.4365 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,324B, BPFP=0.0622 +⌛️ [2/4] FRONTEND: Frontend time: 1.647s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.138s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15112510 56.28911363 + layer.0.v_cache 0.00002069 0.01394047 + layer.1.k_cache 0.18491925 6.51806193 + layer.1.v_cache 0.00000629 0.00543204 + layer.2.k_cache 0.02122488 1.48244840 + layer.2.v_cache 0.00002327 0.01668823 + layer.3.k_cache 0.01259250 9.23408868 + layer.3.v_cache 0.00002182 0.02016980 + layer.4.k_cache 0.00068006 0.47333994 + layer.4.v_cache 0.00005885 0.03435423 + layer.4.output 0.00845948 286.35758695 + ------------------------------------------------------------------------------------- + TOTAL 0.02528759 122.27004388 + (elements=1,662,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1662464 +Total Bytes 140460 +BPFP 0.6759 bits/point +EBPFP 1.3518 equivalent bits/point +MSE 122.270044 +---------------------- -------------------------------------------------------- +Time: 2.792s Load: 0.007s, Pack+Encode: 1.647s, Decode+Unpack: 1.138s +---------------------- -------------------------------------------------------- +💾 Converting with 122.2700 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 174, 128) +Output shape: (1, 174, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.output: torch.Size([1, 174, 3584]) -> torch.Size([1, 1, 174, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,312B, BPFP=0.2974 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,316B, BPFP=1.7346 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,932B, BPFP=0.5327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,536B, BPFP=1.6645 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,996B, BPFP=0.8976 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,060B, BPFP=1.6218 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,388B, BPFP=0.8430 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,620B, BPFP=1.6721 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,048B, BPFP=1.4411 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,580B, BPFP=1.5787 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,868B, BPFP=0.0624 +⌛️ [2/4] FRONTEND: Frontend time: 1.626s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.119s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14146894 55.07742906 + layer.0.v_cache 0.00001664 0.01424973 + layer.1.k_cache 0.16292955 6.71313406 + layer.1.v_cache 0.00000602 0.00550031 + layer.2.k_cache 0.01834387 1.37284281 + layer.2.v_cache 0.00002001 0.01672866 + layer.3.k_cache 0.03015787 8.15666953 + layer.3.v_cache 0.00002180 0.02085276 + layer.4.k_cache 0.00067479 0.45468197 + layer.4.v_cache 0.00005130 0.03524839 + layer.4.output 0.00923280 315.16561474 + ------------------------------------------------------------------------------------- + TOTAL 0.02460708 134.00156709 + (elements=1,514,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1514496 +Total Bytes 141656 +BPFP 0.7483 bits/point +EBPFP 1.4965 equivalent bits/point +MSE 134.001567 +---------------------- -------------------------------------------------------- +Time: 2.751s Load: 0.006s, Pack+Encode: 1.626s, Decode+Unpack: 1.119s +---------------------- -------------------------------------------------------- +💾 Converting with 134.0016 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.7482 bits/point +Avg EBPFP 1.4964 equivalent bits/point +Avg MSE 104.581793 +Avg Time 2.996s +------------------------ ---------------------------- diff --git a/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..be11c95c5f763e9beea78fffa32e739dee0c11e1 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 255 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- -------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa +Output output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa +---------------- -------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,040B, BPFP=0.2734 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,104B, BPFP=1.6332 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,524B, BPFP=0.5167 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,740B, BPFP=1.5592 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,028B, BPFP=0.7611 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,068B, BPFP=1.5228 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,448B, BPFP=0.7296 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,760B, BPFP=1.5603 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,636B, BPFP=1.2823 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,636B, BPFP=1.4993 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,120B, BPFP=0.0629 +⌛️ [2/4] FRONTEND: Frontend time: 2.265s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.403s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14134069 52.79901801 + layer.0.v_cache 0.00001405 0.01116861 + layer.1.k_cache 0.62039534 6.19490729 + layer.1.v_cache 0.00000582 0.00466262 + layer.2.k_cache 0.00676800 1.16286861 + layer.2.v_cache 0.00001893 0.01483716 + layer.3.k_cache 0.03896382 8.11230808 + layer.3.v_cache 0.00001942 0.01782552 + layer.4.k_cache 0.00069280 0.40546205 + layer.4.v_cache 0.00005238 0.03362450 + layer.4.output 0.00804578 192.31550719 + ------------------------------------------------------------------------------------- + TOTAL 0.05085834 83.23324899 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 217104 +BPFP 0.6929 bits/point +EBPFP 1.3857 equivalent bits/point +MSE 83.233249 +---------------------- -------------------------------------------------------- +Time: 3.678s Load: 0.009s, Pack+Encode: 2.265s, Decode+Unpack: 1.403s +---------------------- -------------------------------------------------------- +💾 Converting with 83.2332 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,176B, BPFP=0.2760 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,484B, BPFP=1.6256 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,672B, BPFP=0.5158 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,248B, BPFP=1.5597 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,928B, BPFP=0.7961 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,404B, BPFP=1.5147 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,624B, BPFP=0.7265 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,120B, BPFP=1.5529 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,196B, BPFP=1.2903 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,004B, BPFP=1.4934 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,664B, BPFP=0.0660 +⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.352s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15682776 55.81999587 + layer.0.v_cache 0.00001501 0.01140746 + layer.1.k_cache 0.54447479 6.50682636 + layer.1.v_cache 0.00000580 0.00482937 + layer.2.k_cache 0.00853899 1.23460529 + layer.2.v_cache 0.00001972 0.01504195 + layer.3.k_cache 0.02401569 8.20950265 + layer.3.v_cache 0.00001939 0.01791061 + layer.4.k_cache 0.00071059 0.41664748 + layer.4.v_cache 0.00005329 0.03347391 + layer.4.output 0.05121508 184.45224890 + ------------------------------------------------------------------------------------- + TOTAL 0.06430510 80.20211666 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 221520 +BPFP 0.6949 bits/point +EBPFP 1.3898 equivalent bits/point +MSE 80.202117 +---------------------- -------------------------------------------------------- +Time: 3.192s Load: 0.013s, Pack+Encode: 1.828s, Decode+Unpack: 1.352s +---------------------- -------------------------------------------------------- +💾 Converting with 80.2021 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,216B, BPFP=0.2772 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,152B, BPFP=1.6025 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,684B, BPFP=0.5147 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,856B, BPFP=1.5336 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,620B, BPFP=0.8301 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,176B, BPFP=1.4974 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,152B, BPFP=0.7521 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,860B, BPFP=1.5338 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,004B, BPFP=1.2757 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,720B, BPFP=1.4732 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,296B, BPFP=0.0706 +⌛️ [2/4] FRONTEND: Frontend time: 1.831s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15363285 53.38208174 + layer.0.v_cache 0.00001412 0.01102704 + layer.1.k_cache 0.60852461 6.82262353 + layer.1.v_cache 0.00000580 0.00468961 + layer.2.k_cache 0.01182756 1.21213921 + layer.2.v_cache 0.00001863 0.01461654 + layer.3.k_cache 0.03467803 8.37669570 + layer.3.v_cache 0.00001919 0.01784445 + layer.4.k_cache 0.00072319 0.39221482 + layer.4.v_cache 0.00005301 0.03251467 + layer.4.output 0.05035121 183.89205236 + ------------------------------------------------------------------------------------- + TOTAL 0.06835032 79.85357728 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 221736 +BPFP 0.6932 bits/point +EBPFP 1.3864 equivalent bits/point +MSE 79.853577 +---------------------- -------------------------------------------------------- +Time: 3.190s Load: 0.010s, Pack+Encode: 1.831s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 79.8536 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,068B, BPFP=0.2798 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,348B, BPFP=1.6756 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,600B, BPFP=0.5300 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,068B, BPFP=1.6049 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,132B, BPFP=0.8907 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,512B, BPFP=1.5742 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,448B, BPFP=0.7977 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,020B, BPFP=1.6023 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,124B, BPFP=1.3319 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,128B, BPFP=1.5530 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,160B, BPFP=0.0644 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12197856 51.11590037 + layer.0.v_cache 0.00001365 0.01114653 + layer.1.k_cache 0.62524759 6.05904968 + layer.1.v_cache 0.00000551 0.00447778 + layer.2.k_cache 0.01486346 1.26433183 + layer.2.v_cache 0.00001981 0.01521637 + layer.3.k_cache 0.07387851 8.12223767 + layer.3.v_cache 0.00001943 0.01787649 + layer.4.k_cache 0.00069560 0.40266046 + layer.4.v_cache 0.00005405 0.03354194 + layer.4.output 0.01098318 195.87171883 + ------------------------------------------------------------------------------------- + TOTAL 0.05374461 84.59696888 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 222608 +BPFP 0.7230 bits/point +EBPFP 1.4460 equivalent bits/point +MSE 84.596969 +---------------------- -------------------------------------------------------- +Time: 3.196s Load: 0.009s, Pack+Encode: 1.830s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 84.5970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,068B, BPFP=0.2779 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,024B, BPFP=1.6461 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,344B, BPFP=0.5123 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,716B, BPFP=1.5743 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,892B, BPFP=0.7616 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,100B, BPFP=1.5406 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,740B, BPFP=0.7533 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,616B, BPFP=1.5689 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,648B, BPFP=1.2965 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,292B, BPFP=1.4963 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,796B, BPFP=0.0689 +⌛️ [2/4] FRONTEND: Frontend time: 1.839s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15362121 53.15691132 + layer.0.v_cache 0.00001420 0.01129832 + layer.1.k_cache 0.52952420 6.73717576 + layer.1.v_cache 0.00000567 0.00470970 + layer.2.k_cache 0.00939641 1.20166872 + layer.2.v_cache 0.00001891 0.01442764 + layer.3.k_cache 0.04339437 7.51475552 + layer.3.v_cache 0.00002732 0.01807770 + layer.4.k_cache 0.00068326 0.40839881 + layer.4.v_cache 0.00005156 0.03212816 + layer.4.output 0.00776138 193.92142857 + ------------------------------------------------------------------------------------- + TOTAL 0.04653334 83.91467951 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 217236 +BPFP 0.7006 bits/point +EBPFP 1.4012 equivalent bits/point +MSE 83.914680 +---------------------- -------------------------------------------------------- +Time: 3.200s Load: 0.010s, Pack+Encode: 1.839s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9147 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,196B, BPFP=0.2761 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,428B, BPFP=1.6171 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,644B, BPFP=0.5125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,180B, BPFP=1.5508 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,324B, BPFP=0.8144 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,420B, BPFP=1.5104 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,876B, BPFP=0.7375 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,012B, BPFP=1.5419 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,948B, BPFP=1.2727 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,980B, BPFP=1.4870 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,372B, BPFP=0.0636 +⌛️ [2/4] FRONTEND: Frontend time: 1.869s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14921317 53.76646870 + layer.0.v_cache 0.00001395 0.01118423 + layer.1.k_cache 0.64083395 6.41927000 + layer.1.v_cache 0.00000574 0.00474591 + layer.2.k_cache 0.00578804 1.28180907 + layer.2.v_cache 0.00001912 0.01498940 + layer.3.k_cache 0.02848778 8.44998356 + layer.3.v_cache 0.00001956 0.01784005 + layer.4.k_cache 0.00070939 0.41117457 + layer.4.v_cache 0.00005206 0.03332305 + layer.4.output 0.04916091 184.17334791 + ------------------------------------------------------------------------------------- + TOTAL 0.06878054 79.97789552 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 221380 +BPFP 0.6921 bits/point +EBPFP 1.3842 equivalent bits/point +MSE 79.977896 +---------------------- -------------------------------------------------------- +Time: 3.223s Load: 0.009s, Pack+Encode: 1.869s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 79.9779 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,168B, BPFP=0.2756 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,228B, BPFP=1.6120 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,648B, BPFP=0.5145 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,904B, BPFP=1.5414 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,148B, BPFP=0.8078 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,288B, BPFP=1.5085 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,476B, BPFP=0.7186 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,852B, BPFP=1.5386 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,676B, BPFP=1.2626 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,720B, BPFP=1.4782 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,244B, BPFP=0.0704 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12806672 53.27233762 + layer.0.v_cache 0.00001388 0.01132229 + layer.1.k_cache 0.56701790 6.08372232 + layer.1.v_cache 0.00000557 0.00464800 + layer.2.k_cache 0.01020011 1.20046424 + layer.2.v_cache 0.00001881 0.01491780 + layer.3.k_cache 0.04301871 8.44432344 + layer.3.v_cache 0.00002051 0.01821848 + layer.4.k_cache 0.00071755 0.39633851 + layer.4.v_cache 0.00005284 0.03358704 + layer.4.output 0.04940383 184.27023403 + ------------------------------------------------------------------------------------- + TOTAL 0.06440938 79.96303047 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 220352 +BPFP 0.6912 bits/point +EBPFP 1.3825 equivalent bits/point +MSE 79.963030 +---------------------- -------------------------------------------------------- +Time: 3.196s Load: 0.012s, Pack+Encode: 1.833s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 79.9630 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,072B, BPFP=0.2781 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,080B, BPFP=1.6491 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,472B, BPFP=0.5193 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,764B, BPFP=1.5770 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,520B, BPFP=0.8509 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,232B, BPFP=1.5478 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,492B, BPFP=0.7397 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,888B, BPFP=1.5838 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,980B, BPFP=1.3147 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,944B, BPFP=1.5320 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,236B, BPFP=0.0645 +⌛️ [2/4] FRONTEND: Frontend time: 1.825s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12903612 55.20477316 + layer.0.v_cache 0.00001458 0.01129933 + layer.1.k_cache 0.59860808 6.50700855 + layer.1.v_cache 0.00000581 0.00473983 + layer.2.k_cache 0.00940504 1.25418048 + layer.2.v_cache 0.00001954 0.01538398 + layer.3.k_cache 0.02644065 8.42775151 + layer.3.v_cache 0.00002029 0.01792409 + layer.4.k_cache 0.00068693 0.40791773 + layer.4.v_cache 0.00005376 0.03293282 + layer.4.output 0.00862506 193.00422932 + ------------------------------------------------------------------------------------- + TOTAL 0.04850978 83.70079510 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 219680 +BPFP 0.7085 bits/point +EBPFP 1.4169 equivalent bits/point +MSE 83.700795 +---------------------- -------------------------------------------------------- +Time: 3.181s Load: 0.012s, Pack+Encode: 1.825s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 83.7008 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,444B, BPFP=0.2762 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,512B, BPFP=1.5479 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,188B, BPFP=0.5168 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,048B, BPFP=1.4736 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,208B, BPFP=0.7208 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,316B, BPFP=1.4365 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,704B, BPFP=0.6952 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,008B, BPFP=1.4716 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,108B, BPFP=1.2230 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,852B, BPFP=1.4129 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,564B, BPFP=0.0621 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16469599 52.75543451 + layer.0.v_cache 0.00001357 0.01095903 + layer.1.k_cache 0.70445866 6.49414221 + layer.1.v_cache 0.00000558 0.00457375 + layer.2.k_cache 0.01447187 1.07845168 + layer.2.v_cache 0.00001924 0.01448139 + layer.3.k_cache 0.02617162 8.29261839 + layer.3.v_cache 0.00001974 0.01754503 + layer.4.k_cache 0.00070375 0.38804319 + layer.4.v_cache 0.00005076 0.03327729 + layer.4.output 0.04814868 175.34093808 + ------------------------------------------------------------------------------------- + TOTAL 0.07339127 76.26329959 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 220952 +BPFP 0.6594 bits/point +EBPFP 1.3187 equivalent bits/point +MSE 76.263300 +---------------------- -------------------------------------------------------- +Time: 3.189s Load: 0.011s, Pack+Encode: 1.834s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 76.2633 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,948B, BPFP=0.2874 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,944B, BPFP=1.7393 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,176B, BPFP=0.5330 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,692B, BPFP=1.6666 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,196B, BPFP=0.8827 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,976B, BPFP=1.6250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,540B, BPFP=0.7865 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,404B, BPFP=1.6499 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,936B, BPFP=1.3322 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,432B, BPFP=1.5934 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,392B, BPFP=0.0613 +⌛️ [2/4] FRONTEND: Frontend time: 1.829s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12132761 50.85600604 + layer.0.v_cache 0.00001366 0.01113729 + layer.1.k_cache 0.57241889 6.63979002 + layer.1.v_cache 0.00000556 0.00452326 + layer.2.k_cache 0.00926234 1.25507352 + layer.2.v_cache 0.00002145 0.01465528 + layer.3.k_cache 0.03109839 8.51145010 + layer.3.v_cache 0.00002257 0.01752244 + layer.4.k_cache 0.00070701 0.39547483 + layer.4.v_cache 0.00004920 0.03266527 + layer.4.output 0.01031505 205.58251460 + ------------------------------------------------------------------------------------- + TOTAL 0.04747836 88.63622943 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 215636 +BPFP 0.7368 bits/point +EBPFP 1.4736 equivalent bits/point +MSE 88.636229 +---------------------- -------------------------------------------------------- +Time: 3.188s Load: 0.010s, Pack+Encode: 1.829s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6362 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,284B, BPFP=0.2771 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,404B, BPFP=1.5942 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,888B, BPFP=0.5185 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,084B, BPFP=1.5250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,748B, BPFP=0.7733 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,416B, BPFP=1.4899 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,796B, BPFP=0.7234 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,000B, BPFP=1.5206 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,096B, BPFP=1.2634 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,672B, BPFP=1.4509 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,480B, BPFP=0.0635 +⌛️ [2/4] FRONTEND: Frontend time: 1.835s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14357947 52.67019715 + layer.0.v_cache 0.00001420 0.01142056 + layer.1.k_cache 0.63723908 6.38009746 + layer.1.v_cache 0.00000573 0.00465057 + layer.2.k_cache 0.00765126 1.28436894 + layer.2.v_cache 0.00002073 0.01480550 + layer.3.k_cache 0.02843528 8.55453348 + layer.3.v_cache 0.00002112 0.01770290 + layer.4.k_cache 0.00069171 0.39804569 + layer.4.v_cache 0.00005234 0.03327959 + layer.4.output 0.04871205 178.39845997 + ------------------------------------------------------------------------------------- + TOTAL 0.06815855 77.53872480 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 220868 +BPFP 0.6812 bits/point +EBPFP 1.3624 equivalent bits/point +MSE 77.538725 +---------------------- -------------------------------------------------------- +Time: 3.202s Load: 0.010s, Pack+Encode: 1.835s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 77.5387 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,380B, BPFP=0.2756 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,504B, BPFP=1.5627 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,464B, BPFP=0.5361 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,160B, BPFP=1.4939 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,972B, BPFP=0.7158 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,540B, BPFP=1.4621 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,800B, BPFP=0.7070 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,892B, BPFP=1.4801 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,412B, BPFP=1.2506 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,844B, BPFP=1.4264 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,560B, BPFP=0.0626 +⌛️ [2/4] FRONTEND: Frontend time: 1.824s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16857396 54.19514600 + layer.0.v_cache 0.00001478 0.01142446 + layer.1.k_cache 0.58716421 6.11285020 + layer.1.v_cache 0.00000600 0.00476883 + layer.2.k_cache 0.00827454 1.21242105 + layer.2.v_cache 0.00002177 0.01481488 + layer.3.k_cache 0.02531170 8.02264744 + layer.3.v_cache 0.00001983 0.01758416 + layer.4.k_cache 0.00069463 0.39972654 + layer.4.v_cache 0.00005316 0.03310624 + layer.4.output 0.04836898 177.33217213 + ------------------------------------------------------------------------------------- + TOTAL 0.06639514 77.13821734 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 221528 +BPFP 0.6676 bits/point +EBPFP 1.3351 equivalent bits/point +MSE 77.138217 +---------------------- -------------------------------------------------------- +Time: 3.178s Load: 0.011s, Pack+Encode: 1.824s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 77.1382 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,296B, BPFP=0.2777 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,384B, BPFP=1.5931 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,736B, BPFP=0.5105 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,080B, BPFP=1.5247 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,432B, BPFP=0.8091 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,224B, BPFP=1.4799 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,888B, BPFP=0.7282 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,904B, BPFP=1.5155 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,260B, BPFP=1.2720 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,848B, BPFP=1.4602 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,864B, BPFP=0.0664 +⌛️ [2/4] FRONTEND: Frontend time: 1.839s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.361s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14028978 53.09242607 + layer.0.v_cache 0.00001397 0.01120523 + layer.1.k_cache 0.59966084 6.68406483 + layer.1.v_cache 0.00000578 0.00471008 + layer.2.k_cache 0.01765286 1.22563202 + layer.2.v_cache 0.00001884 0.01477878 + layer.3.k_cache 0.02653811 8.63466404 + layer.3.v_cache 0.00001935 0.01743703 + layer.4.k_cache 0.00071527 0.40746691 + layer.4.v_cache 0.00005402 0.03287469 + layer.4.output 0.04907703 180.66691635 + ------------------------------------------------------------------------------------- + TOTAL 0.06638282 78.51727495 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 221916 +BPFP 0.6845 bits/point +EBPFP 1.3689 equivalent bits/point +MSE 78.517275 +---------------------- -------------------------------------------------------- +Time: 3.210s Load: 0.010s, Pack+Encode: 1.839s, Decode+Unpack: 1.361s +---------------------- -------------------------------------------------------- +💾 Converting with 78.5173 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,976B, BPFP=0.2858 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,136B, BPFP=1.7312 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,256B, BPFP=0.5317 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,960B, BPFP=1.6636 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,960B, BPFP=0.8019 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,268B, BPFP=1.6239 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,048B, BPFP=0.7495 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,932B, BPFP=1.6620 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,208B, BPFP=1.2757 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,860B, BPFP=1.6004 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,916B, BPFP=0.0732 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13255837 50.93877097 + layer.0.v_cache 0.00001394 0.01115173 + layer.1.k_cache 0.54921044 6.85554684 + layer.1.v_cache 0.00000578 0.00456298 + layer.2.k_cache 0.01292971 1.25577983 + layer.2.v_cache 0.00001901 0.01525064 + layer.3.k_cache 0.07190662 8.04021050 + layer.3.v_cache 0.00001944 0.01805678 + layer.4.k_cache 0.00068159 0.38203736 + layer.4.v_cache 0.00005125 0.03278663 + layer.4.output 0.00790709 203.33196560 + ------------------------------------------------------------------------------------- + TOTAL 0.04839681 87.69870079 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 216520 +BPFP 0.7316 bits/point +EBPFP 1.4633 equivalent bits/point +MSE 87.698701 +---------------------- -------------------------------------------------------- +Time: 3.201s Load: 0.009s, Pack+Encode: 1.833s, Decode+Unpack: 1.359s +---------------------- -------------------------------------------------------- +💾 Converting with 87.6987 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,392B, BPFP=0.2988 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,472B, BPFP=1.6884 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,576B, BPFP=0.5306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,240B, BPFP=1.6201 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,024B, BPFP=0.8324 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,508B, BPFP=1.5796 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,252B, BPFP=0.8451 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,256B, BPFP=1.6210 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,936B, BPFP=1.3262 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,232B, BPFP=1.5643 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,532B, BPFP=0.0675 +⌛️ [2/4] FRONTEND: Frontend time: 1.853s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13695534 52.11756843 + layer.0.v_cache 0.00001452 0.01142597 + layer.1.k_cache 0.52272531 6.60726691 + layer.1.v_cache 0.00000574 0.00482237 + layer.2.k_cache 0.00793556 1.11306492 + layer.2.v_cache 0.00001907 0.01499800 + layer.3.k_cache 0.04682703 7.98766224 + layer.3.v_cache 0.00002008 0.01781201 + layer.4.k_cache 0.00068945 0.40850008 + layer.4.v_cache 0.00005221 0.03260801 + layer.4.output 0.00941004 196.49821112 + ------------------------------------------------------------------------------------- + TOTAL 0.04594792 84.92960040 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 223420 +BPFP 0.7282 bits/point +EBPFP 1.4564 equivalent bits/point +MSE 84.929600 +---------------------- -------------------------------------------------------- +Time: 3.219s Load: 0.013s, Pack+Encode: 1.853s, Decode+Unpack: 1.354s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9296 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,024B, BPFP=0.2804 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,388B, BPFP=1.6958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,720B, BPFP=0.5424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,168B, BPFP=1.6277 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,092B, BPFP=0.7864 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,352B, BPFP=1.5821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,312B, BPFP=0.7987 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,036B, BPFP=1.6203 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,268B, BPFP=1.2984 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,068B, BPFP=1.5663 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,556B, BPFP=0.0682 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14340984 52.69296177 + layer.0.v_cache 0.00001397 0.01121477 + layer.1.k_cache 0.51622396 6.37555716 + layer.1.v_cache 0.00000557 0.00455381 + layer.2.k_cache 0.00797994 1.19575947 + layer.2.v_cache 0.00001813 0.01453591 + layer.3.k_cache 0.02756977 7.56308419 + layer.3.v_cache 0.00001890 0.01761395 + layer.4.k_cache 0.00068956 0.38754809 + layer.4.v_cache 0.00005177 0.03212314 + layer.4.output 0.00947078 198.00978954 + ------------------------------------------------------------------------------------- + TOTAL 0.04483982 85.55079289 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 219984 +BPFP 0.7221 bits/point +EBPFP 1.4442 equivalent bits/point +MSE 85.550793 +---------------------- -------------------------------------------------------- +Time: 3.211s Load: 0.012s, Pack+Encode: 1.840s, Decode+Unpack: 1.359s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5508 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,240B, BPFP=0.2785 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,208B, BPFP=1.6054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,348B, BPFP=0.5500 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,936B, BPFP=1.5378 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,424B, BPFP=0.7134 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,128B, BPFP=1.4949 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,812B, BPFP=0.7341 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,804B, BPFP=1.5308 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,164B, BPFP=1.2842 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,756B, BPFP=1.4751 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,120B, BPFP=0.0692 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.356s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13622742 51.74344308 + layer.0.v_cache 0.00001380 0.01122796 + layer.1.k_cache 0.61520739 6.34644261 + layer.1.v_cache 0.00000584 0.00481329 + layer.2.k_cache 0.00900942 1.06482380 + layer.2.v_cache 0.00001916 0.01511813 + layer.3.k_cache 0.02667915 8.07011320 + layer.3.v_cache 0.00001881 0.01745076 + layer.4.k_cache 0.00071466 0.40180102 + layer.4.v_cache 0.00005760 0.03263566 + layer.4.output 0.04972813 183.22163508 + ------------------------------------------------------------------------------------- + TOTAL 0.06682648 79.42701853 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 219940 +BPFP 0.6876 bits/point +EBPFP 1.3752 equivalent bits/point +MSE 79.427019 +---------------------- -------------------------------------------------------- +Time: 3.203s Load: 0.013s, Pack+Encode: 1.833s, Decode+Unpack: 1.356s +---------------------- -------------------------------------------------------- +💾 Converting with 79.4270 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,312B, BPFP=0.2776 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,440B, BPFP=1.5907 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,056B, BPFP=0.5255 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,972B, BPFP=1.5140 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,660B, BPFP=0.8184 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,284B, BPFP=1.4781 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,720B, BPFP=0.7170 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,104B, BPFP=1.5209 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,884B, BPFP=1.2481 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,832B, BPFP=1.4544 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,268B, BPFP=0.0617 +⌛️ [2/4] FRONTEND: Frontend time: 1.836s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11478914 53.99447377 + layer.0.v_cache 0.00001377 0.01109509 + layer.1.k_cache 0.64218915 5.76779205 + layer.1.v_cache 0.00000556 0.00447709 + layer.2.k_cache 0.00898698 1.29416251 + layer.2.v_cache 0.00001980 0.01497506 + layer.3.k_cache 0.05442990 8.01469253 + layer.3.v_cache 0.00001964 0.01810267 + layer.4.k_cache 0.00070461 0.39624728 + layer.4.v_cache 0.00005015 0.03236508 + layer.4.output 0.05017195 180.60027473 + ------------------------------------------------------------------------------------- + TOTAL 0.06896543 78.45590036 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 221532 +BPFP 0.6810 bits/point +EBPFP 1.3620 equivalent bits/point +MSE 78.455900 +---------------------- -------------------------------------------------------- +Time: 3.210s Load: 0.011s, Pack+Encode: 1.836s, Decode+Unpack: 1.362s +---------------------- -------------------------------------------------------- +💾 Converting with 78.4559 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,860B, BPFP=0.2866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,856B, BPFP=1.7604 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,976B, BPFP=0.5292 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,740B, BPFP=1.6946 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,428B, BPFP=0.9097 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,956B, BPFP=1.6483 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,064B, BPFP=0.8882 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,676B, BPFP=1.6908 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,124B, BPFP=1.3634 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,544B, BPFP=1.6241 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,204B, BPFP=0.0607 +⌛️ [2/4] FRONTEND: Frontend time: 1.831s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14755792 56.48985849 + layer.0.v_cache 0.00001375 0.01110424 + layer.1.k_cache 0.51940866 6.25361236 + layer.1.v_cache 0.00000586 0.00476342 + layer.2.k_cache 0.00816260 1.18337771 + layer.2.v_cache 0.00001900 0.01483472 + layer.3.k_cache 0.02921419 7.80870707 + layer.3.v_cache 0.00001979 0.01737336 + layer.4.k_cache 0.00071017 0.40170968 + layer.4.v_cache 0.00005507 0.03349499 + layer.4.output 0.00937099 207.14309299 + ------------------------------------------------------------------------------------- + TOTAL 0.04533906 89.54238159 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 217428 +BPFP 0.7541 bits/point +EBPFP 1.5082 equivalent bits/point +MSE 89.542382 +---------------------- -------------------------------------------------------- +Time: 3.198s Load: 0.009s, Pack+Encode: 1.831s, Decode+Unpack: 1.359s +---------------------- -------------------------------------------------------- +💾 Converting with 89.5424 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,860B, BPFP=0.2844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,800B, BPFP=1.7439 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,368B, BPFP=0.5482 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,660B, BPFP=1.6772 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,140B, BPFP=0.7690 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,072B, BPFP=1.6428 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,236B, BPFP=0.7746 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,556B, BPFP=1.6711 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,580B, BPFP=1.3214 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,652B, BPFP=1.6182 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,256B, BPFP=0.0607 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12840328 54.24054527 + layer.0.v_cache 0.00001372 0.01091108 + layer.1.k_cache 0.60081110 6.22410098 + layer.1.v_cache 0.00000562 0.00455053 + layer.2.k_cache 0.00586274 1.16714146 + layer.2.v_cache 0.00001839 0.01465193 + layer.3.k_cache 0.02491910 7.84938901 + layer.3.v_cache 0.00001957 0.01742741 + layer.4.k_cache 0.00072047 0.40858599 + layer.4.v_cache 0.00005229 0.03441295 + layer.4.output 0.01050496 207.26456327 + ------------------------------------------------------------------------------------- + TOTAL 0.04908006 89.46021526 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 213180 +BPFP 0.7338 bits/point +EBPFP 1.4677 equivalent bits/point +MSE 89.460215 +---------------------- -------------------------------------------------------- +Time: 3.223s Load: 0.012s, Pack+Encode: 1.840s, Decode+Unpack: 1.371s +---------------------- -------------------------------------------------------- +💾 Converting with 89.4602 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,068B, BPFP=0.2798 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,924B, BPFP=1.6522 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,796B, BPFP=0.5409 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,608B, BPFP=1.5795 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,236B, BPFP=0.8412 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,200B, BPFP=1.5570 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,952B, BPFP=0.7703 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,784B, BPFP=1.5892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,488B, BPFP=1.2416 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,532B, BPFP=1.5201 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,136B, BPFP=0.0642 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14595435 53.04224754 + layer.0.v_cache 0.00001364 0.01097149 + layer.1.k_cache 0.59759904 6.20518424 + layer.1.v_cache 0.00000566 0.00456549 + layer.2.k_cache 0.01409306 1.22853169 + layer.2.v_cache 0.00002047 0.01508693 + layer.3.k_cache 0.06079736 7.67934760 + layer.3.v_cache 0.00002215 0.01789125 + layer.4.k_cache 0.00070776 0.40213495 + layer.4.v_cache 0.00005221 0.03330833 + layer.4.output 0.00769303 195.50730376 + ------------------------------------------------------------------------------------- + TOTAL 0.05135982 84.54061152 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 217724 +BPFP 0.7071 bits/point +EBPFP 1.4142 equivalent bits/point +MSE 84.540612 +---------------------- -------------------------------------------------------- +Time: 3.203s Load: 0.010s, Pack+Encode: 1.834s, Decode+Unpack: 1.359s +---------------------- -------------------------------------------------------- +💾 Converting with 84.5406 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 315, 128) +Output shape: (1, 315, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.output: torch.Size([1, 315, 3584]) -> torch.Size([1, 1, 315, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,512B, BPFP=0.2734 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,584B, BPFP=1.5171 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,208B, BPFP=0.5063 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,156B, BPFP=1.4462 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,912B, BPFP=0.6901 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,184B, BPFP=1.3980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,280B, BPFP=0.7083 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,036B, BPFP=1.4403 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,344B, BPFP=1.2075 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,832B, BPFP=1.3806 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,276B, BPFP=0.0657 +⌛️ [2/4] FRONTEND: Frontend time: 1.846s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.358s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15500355 53.93207465 + layer.0.v_cache 0.00001366 0.01118163 + layer.1.k_cache 0.68167124 6.00878209 + layer.1.v_cache 0.00000571 0.00473469 + layer.2.k_cache 0.01007247 1.20003216 + layer.2.v_cache 0.00001936 0.01526690 + layer.3.k_cache 0.03272899 8.84350818 + layer.3.v_cache 0.00002006 0.01769001 + layer.4.k_cache 0.00071175 0.40799420 + layer.4.v_cache 0.00005427 0.03359713 + layer.4.output 0.04584261 171.64842687 + ------------------------------------------------------------------------------------- + TOTAL 0.07065879 74.82434410 + (elements=2,741,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2741760 +Total Bytes 222324 +BPFP 0.6487 bits/point +EBPFP 1.2974 equivalent bits/point +MSE 74.824344 +---------------------- -------------------------------------------------------- +Time: 3.215s Load: 0.011s, Pack+Encode: 1.846s, Decode+Unpack: 1.358s +---------------------- -------------------------------------------------------- +💾 Converting with 74.8243 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,924B, BPFP=0.2839 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,980B, BPFP=1.7286 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,336B, BPFP=0.5383 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,696B, BPFP=1.6545 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,356B, BPFP=0.8277 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,284B, BPFP=1.6308 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,444B, BPFP=0.7751 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,772B, BPFP=1.6589 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,100B, BPFP=1.3319 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,684B, BPFP=1.5962 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,476B, BPFP=0.0616 +⌛️ [2/4] FRONTEND: Frontend time: 1.835s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.360s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12889319 52.32755204 + layer.0.v_cache 0.00001350 0.01099815 + layer.1.k_cache 0.58775566 6.22971561 + layer.1.v_cache 0.00000553 0.00448158 + layer.2.k_cache 0.00723753 1.23031537 + layer.2.v_cache 0.00001924 0.01477006 + layer.3.k_cache 0.01761036 7.82503766 + layer.3.v_cache 0.00001847 0.01753499 + layer.4.k_cache 0.00071814 0.40056723 + layer.4.v_cache 0.00005270 0.03332177 + layer.4.output 0.01035140 204.47300013 + ------------------------------------------------------------------------------------- + TOTAL 0.04792848 88.20031149 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 216052 +BPFP 0.7328 bits/point +EBPFP 1.4655 equivalent bits/point +MSE 88.200311 +---------------------- -------------------------------------------------------- +Time: 3.206s Load: 0.011s, Pack+Encode: 1.835s, Decode+Unpack: 1.360s +---------------------- -------------------------------------------------------- +💾 Converting with 88.2003 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,052B, BPFP=0.2779 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,072B, BPFP=1.6545 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,736B, BPFP=0.5357 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,836B, BPFP=1.5865 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,540B, BPFP=0.7449 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,092B, BPFP=1.5456 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,232B, BPFP=0.7280 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,484B, BPFP=1.5671 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,660B, BPFP=1.3017 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,468B, BPFP=1.5112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,988B, BPFP=0.0628 +⌛️ [2/4] FRONTEND: Frontend time: 1.829s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.339s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13033964 51.94430155 + layer.0.v_cache 0.00001368 0.01110007 + layer.1.k_cache 0.63029007 6.13402805 + layer.1.v_cache 0.00000551 0.00450880 + layer.2.k_cache 0.00637632 1.27930483 + layer.2.v_cache 0.00001960 0.01442727 + layer.3.k_cache 0.02368569 8.08647220 + layer.3.v_cache 0.00001974 0.01725039 + layer.4.k_cache 0.00068050 0.39075760 + layer.4.v_cache 0.00005209 0.03289941 + layer.4.output 0.00765416 194.83600038 + ------------------------------------------------------------------------------------- + TOTAL 0.04970952 84.22159134 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 216160 +BPFP 0.6996 bits/point +EBPFP 1.3991 equivalent bits/point +MSE 84.221591 +---------------------- -------------------------------------------------------- +Time: 3.177s Load: 0.010s, Pack+Encode: 1.829s, Decode+Unpack: 1.339s +---------------------- -------------------------------------------------------- +💾 Converting with 84.2216 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,976B, BPFP=0.2858 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,140B, BPFP=1.7314 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,036B, BPFP=0.5191 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,028B, BPFP=1.6675 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,740B, BPFP=0.8467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,384B, BPFP=1.6305 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,632B, BPFP=0.7256 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,900B, BPFP=1.6602 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,620B, BPFP=1.3568 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,756B, BPFP=1.5944 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,088B, BPFP=0.0664 +⌛️ [2/4] FRONTEND: Frontend time: 1.835s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15644295 52.15809183 + layer.0.v_cache 0.00001405 0.01088937 + layer.1.k_cache 0.56163917 5.92288118 + layer.1.v_cache 0.00000554 0.00458452 + layer.2.k_cache 0.01059573 1.28286317 + layer.2.v_cache 0.00001807 0.01481048 + layer.3.k_cache 0.04944291 7.87509694 + layer.3.v_cache 0.00001950 0.01786642 + layer.4.k_cache 0.00071267 0.40596480 + layer.4.v_cache 0.00005817 0.03314728 + layer.4.output 0.00956290 203.38686647 + ------------------------------------------------------------------------------------- + TOTAL 0.04975818 87.73142713 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 217300 +BPFP 0.7343 bits/point +EBPFP 1.4686 equivalent bits/point +MSE 87.731427 +---------------------- -------------------------------------------------------- +Time: 3.185s Load: 0.009s, Pack+Encode: 1.835s, Decode+Unpack: 1.341s +---------------------- -------------------------------------------------------- +💾 Converting with 87.7314 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 312, 128) +Output shape: (1, 312, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.output: torch.Size([1, 312, 3584]) -> torch.Size([1, 1, 312, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,464B, BPFP=0.2736 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,756B, BPFP=1.5403 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,380B, BPFP=0.5198 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,360B, BPFP=1.4704 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,588B, BPFP=0.7306 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,516B, BPFP=1.4281 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,536B, BPFP=0.6779 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,232B, BPFP=1.4639 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,620B, BPFP=1.2330 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,924B, BPFP=1.3984 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,460B, BPFP=0.0677 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14984696 53.84589093 + layer.0.v_cache 0.00001474 0.01149652 + layer.1.k_cache 0.67859346 6.17082136 + layer.1.v_cache 0.00000581 0.00480959 + layer.2.k_cache 0.01626646 1.30342102 + layer.2.v_cache 0.00001951 0.01519633 + layer.3.k_cache 0.02914289 7.98089365 + layer.3.v_cache 0.00001919 0.01807325 + layer.4.k_cache 0.00071665 0.40037967 + layer.4.v_cache 0.00005096 0.03249874 + layer.4.output 0.04642455 173.57344609 + ------------------------------------------------------------------------------------- + TOTAL 0.07056756 75.57632963 + (elements=2,715,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2715648 +Total Bytes 223836 +BPFP 0.6594 bits/point +EBPFP 1.3188 equivalent bits/point +MSE 75.576330 +---------------------- -------------------------------------------------------- +Time: 3.198s Load: 0.010s, Pack+Encode: 1.841s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 75.5763 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,944B, BPFP=0.2872 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,124B, BPFP=1.7498 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,048B, BPFP=0.5256 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,732B, BPFP=1.6689 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,732B, BPFP=0.8557 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,156B, BPFP=1.6355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,888B, BPFP=0.7486 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,772B, BPFP=1.6712 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,716B, BPFP=1.3195 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,736B, BPFP=1.6111 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,668B, BPFP=0.0636 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15823836 52.09170612 + layer.0.v_cache 0.00001502 0.01115670 + layer.1.k_cache 0.53668020 6.43387147 + layer.1.v_cache 0.00000576 0.00467807 + layer.2.k_cache 0.01018995 1.19698042 + layer.2.v_cache 0.00002218 0.01507402 + layer.3.k_cache 0.05410197 7.39295704 + layer.3.v_cache 0.00002001 0.01805800 + layer.4.k_cache 0.00069143 0.39967695 + layer.4.v_cache 0.00005304 0.03277517 + layer.4.output 0.01041225 205.75977496 + ------------------------------------------------------------------------------------- + TOTAL 0.04899433 88.70090345 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 215516 +BPFP 0.7364 bits/point +EBPFP 1.4727 equivalent bits/point +MSE 88.700903 +---------------------- -------------------------------------------------------- +Time: 3.193s Load: 0.009s, Pack+Encode: 1.838s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 88.7009 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 329, 128) +Output shape: (1, 329, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.output: torch.Size([1, 329, 3584]) -> torch.Size([1, 1, 329, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,888B, BPFP=0.2796 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,064B, BPFP=1.7128 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,104B, BPFP=0.5274 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,760B, BPFP=1.6508 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,180B, BPFP=0.8159 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,560B, BPFP=1.5938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,040B, BPFP=0.8093 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,380B, BPFP=1.6328 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,968B, BPFP=1.2808 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,084B, BPFP=1.5712 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,112B, BPFP=0.0754 +⌛️ [2/4] FRONTEND: Frontend time: 2.181s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.471s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11391156 52.89359328 + layer.0.v_cache 0.00001522 0.01133450 + layer.1.k_cache 0.63933222 6.32776881 + layer.1.v_cache 0.00000652 0.00506418 + layer.2.k_cache 0.01643764 1.18023422 + layer.2.v_cache 0.00001873 0.01475242 + layer.3.k_cache 0.02826017 7.96622326 + layer.3.v_cache 0.00001876 0.01714767 + layer.4.k_cache 0.00074607 0.40939122 + layer.4.v_cache 0.00005483 0.03367466 + layer.4.output 3.44632065 162.74824957 + ------------------------------------------------------------------------------------- + TOTAL 1.46606155 71.06452536 + (elements=2,863,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2863616 +Total Bytes 261140 +BPFP 0.7295 bits/point +EBPFP 1.4591 equivalent bits/point +MSE 71.064525 +---------------------- -------------------------------------------------------- +Time: 3.666s Load: 0.014s, Pack+Encode: 2.181s, Decode+Unpack: 1.471s +---------------------- -------------------------------------------------------- +💾 Converting with 71.0645 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,012B, BPFP=0.2807 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,212B, BPFP=1.6920 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,252B, BPFP=0.5181 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,112B, BPFP=1.6304 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,660B, BPFP=0.8770 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,356B, BPFP=1.5880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,240B, BPFP=0.7975 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,000B, BPFP=1.6241 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,652B, BPFP=1.3246 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,988B, BPFP=1.5674 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,756B, BPFP=0.0621 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13536828 54.06407860 + layer.0.v_cache 0.00001433 0.01097367 + layer.1.k_cache 0.57185238 6.43714473 + layer.1.v_cache 0.00000558 0.00448480 + layer.2.k_cache 0.00766427 1.23658036 + layer.2.v_cache 0.00001891 0.01473789 + layer.3.k_cache 0.04143807 7.99739496 + layer.3.v_cache 0.00001947 0.01737823 + layer.4.k_cache 0.00070404 0.38815515 + layer.4.v_cache 0.00005498 0.03269508 + layer.4.output 0.00948056 198.49716782 + ------------------------------------------------------------------------------------- + TOTAL 0.04844142 85.86375284 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 220240 +BPFP 0.7255 bits/point +EBPFP 1.4511 equivalent bits/point +MSE 85.863753 +---------------------- -------------------------------------------------------- +Time: 3.192s Load: 0.012s, Pack+Encode: 1.833s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8638 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,964B, BPFP=0.2810 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,116B, BPFP=1.7049 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,188B, BPFP=0.5202 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,932B, BPFP=1.6379 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,292B, BPFP=0.8657 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,280B, BPFP=1.6010 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,820B, BPFP=0.7824 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,740B, BPFP=1.6270 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,324B, BPFP=1.3204 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,812B, BPFP=1.5745 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,960B, BPFP=0.0644 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15132187 53.09592957 + layer.0.v_cache 0.00001456 0.01078997 + layer.1.k_cache 0.61798344 6.09674957 + layer.1.v_cache 0.00000569 0.00446976 + layer.2.k_cache 0.01686249 1.27167854 + layer.2.v_cache 0.00001898 0.01496175 + layer.3.k_cache 0.04401700 7.54497406 + layer.3.v_cache 0.00001935 0.01723077 + layer.4.k_cache 0.00071189 0.38970041 + layer.4.v_cache 0.00005111 0.03230295 + layer.4.output 0.01058244 200.37554995 + ------------------------------------------------------------------------------------- + TOTAL 0.05324020 86.53574335 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 218428 +BPFP 0.7274 bits/point +EBPFP 1.4548 equivalent bits/point +MSE 86.535743 +---------------------- -------------------------------------------------------- +Time: 3.190s Load: 0.010s, Pack+Encode: 1.833s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 86.5357 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,948B, BPFP=0.2801 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,508B, BPFP=1.7271 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,212B, BPFP=0.5215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,232B, BPFP=1.6549 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,416B, BPFP=0.8161 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,512B, BPFP=1.6141 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,944B, BPFP=0.7894 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,060B, BPFP=1.6452 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,368B, BPFP=1.3229 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,948B, BPFP=1.5822 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,436B, BPFP=0.0601 +⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12581717 53.43419526 + layer.0.v_cache 0.00001375 0.01128279 + layer.1.k_cache 0.57814745 6.38583507 + layer.1.v_cache 0.00000592 0.00464645 + layer.2.k_cache 0.00733882 1.26845164 + layer.2.v_cache 0.00002058 0.01516280 + layer.3.k_cache 0.02353939 8.65743355 + layer.3.v_cache 0.00002077 0.01784300 + layer.4.k_cache 0.00069182 0.41690782 + layer.4.v_cache 0.00005027 0.03224505 + layer.4.output 0.01015823 200.43289014 + ------------------------------------------------------------------------------------- + TOTAL 0.04745609 86.66319026 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 218584 +BPFP 0.7279 bits/point +EBPFP 1.4558 equivalent bits/point +MSE 86.663190 +---------------------- -------------------------------------------------------- +Time: 3.182s Load: 0.010s, Pack+Encode: 1.828s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 86.6632 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,944B, BPFP=0.2819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,172B, BPFP=1.7206 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,092B, BPFP=0.5185 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,052B, BPFP=1.6567 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,640B, BPFP=0.8349 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,428B, BPFP=1.6211 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,532B, BPFP=0.7717 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,108B, BPFP=1.6599 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,932B, BPFP=1.3077 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,980B, BPFP=1.5956 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,152B, BPFP=0.0664 +⌛️ [2/4] FRONTEND: Frontend time: 1.826s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13625057 52.82419737 + layer.0.v_cache 0.00001362 0.01092378 + layer.1.k_cache 0.57512581 6.36376508 + layer.1.v_cache 0.00000584 0.00461883 + layer.2.k_cache 0.00872732 1.18241181 + layer.2.v_cache 0.00001851 0.01501029 + layer.3.k_cache 0.04362779 8.39028106 + layer.3.v_cache 0.00001925 0.01855613 + layer.4.k_cache 0.00073619 0.39431083 + layer.4.v_cache 0.00005148 0.03375485 + layer.4.output 0.00783937 201.72598410 + ------------------------------------------------------------------------------------- + TOTAL 0.04820306 87.13645404 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 218032 +BPFP 0.7314 bits/point +EBPFP 1.4628 equivalent bits/point +MSE 87.136454 +---------------------- -------------------------------------------------------- +Time: 3.180s Load: 0.011s, Pack+Encode: 1.826s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 87.1365 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,384B, BPFP=0.2767 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,228B, BPFP=1.5537 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,028B, BPFP=0.5154 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,772B, BPFP=1.4788 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,516B, BPFP=0.7461 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,084B, BPFP=1.4435 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,408B, BPFP=0.6891 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,912B, BPFP=1.4860 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,388B, BPFP=1.2021 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,772B, BPFP=1.4274 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,772B, BPFP=0.0718 +⌛️ [2/4] FRONTEND: Frontend time: 1.826s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16730670 53.78323525 + layer.0.v_cache 0.00001388 0.01110642 + layer.1.k_cache 0.64158324 6.35395331 + layer.1.v_cache 0.00000571 0.00467606 + layer.2.k_cache 0.02207795 1.19791422 + layer.2.v_cache 0.00001922 0.01463127 + layer.3.k_cache 0.04503629 8.03216633 + layer.3.v_cache 0.00001955 0.01795416 + layer.4.k_cache 0.00070478 0.39529635 + layer.4.v_cache 0.00005260 0.03241057 + layer.4.output 0.04763387 177.59756814 + ------------------------------------------------------------------------------------- + TOTAL 0.07119159 77.23684241 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 220264 +BPFP 0.6659 bits/point +EBPFP 1.3319 equivalent bits/point +MSE 77.236842 +---------------------- -------------------------------------------------------- +Time: 3.187s Load: 0.011s, Pack+Encode: 1.826s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 77.2368 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,984B, BPFP=0.2811 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,124B, BPFP=1.6992 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,292B, BPFP=0.5241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,908B, BPFP=1.6306 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,612B, BPFP=0.8242 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,276B, BPFP=1.5950 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,212B, BPFP=0.8017 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,912B, BPFP=1.6309 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,272B, BPFP=1.3127 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,072B, BPFP=1.5835 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,808B, BPFP=0.0710 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15225300 51.15867554 + layer.0.v_cache 0.00001427 0.01083675 + layer.1.k_cache 0.52731114 5.92872284 + layer.1.v_cache 0.00000556 0.00456936 + layer.2.k_cache 0.01483519 1.24730586 + layer.2.v_cache 0.00001864 0.01485435 + layer.3.k_cache 0.04699480 7.98572967 + layer.3.v_cache 0.00001968 0.01742877 + layer.4.k_cache 0.00070142 0.38576508 + layer.4.v_cache 0.00005177 0.03237227 + layer.4.output 0.01077793 198.26444043 + ------------------------------------------------------------------------------------- + TOTAL 0.04809712 85.56690256 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 219472 +BPFP 0.7282 bits/point +EBPFP 1.4565 equivalent bits/point +MSE 85.566903 +---------------------- -------------------------------------------------------- +Time: 3.196s Load: 0.011s, Pack+Encode: 1.830s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5669 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,996B, BPFP=0.2818 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,196B, BPFP=1.7033 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,200B, BPFP=0.5190 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,956B, BPFP=1.6333 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,704B, BPFP=0.8294 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,416B, BPFP=1.6029 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,500B, BPFP=0.8743 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,860B, BPFP=1.6279 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,848B, BPFP=1.2888 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,004B, BPFP=1.5796 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,412B, BPFP=0.0678 +⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15700974 52.40861561 + layer.0.v_cache 0.00001374 0.01080747 + layer.1.k_cache 0.52255304 6.02998523 + layer.1.v_cache 0.00000566 0.00451676 + layer.2.k_cache 0.01360668 1.24316583 + layer.2.v_cache 0.00001926 0.01539640 + layer.3.k_cache 0.08085050 7.59322029 + layer.3.v_cache 0.00001943 0.01817812 + layer.4.k_cache 0.00071029 0.38159951 + layer.4.v_cache 0.00005450 0.03356029 + layer.4.output 0.01014708 198.75610817 + ------------------------------------------------------------------------------------- + TOTAL 0.04975720 85.82540016 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 220092 +BPFP 0.7303 bits/point +EBPFP 1.4606 equivalent bits/point +MSE 85.825400 +---------------------- -------------------------------------------------------- +Time: 3.188s Load: 0.010s, Pack+Encode: 1.828s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8254 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,068B, BPFP=0.2779 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,156B, BPFP=1.6533 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,764B, BPFP=0.5353 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,024B, BPFP=1.5912 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,988B, BPFP=0.8217 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,216B, BPFP=1.5469 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,164B, BPFP=0.7765 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,836B, BPFP=1.5809 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,296B, BPFP=1.2772 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,616B, BPFP=1.5140 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,408B, BPFP=0.0659 +⌛️ [2/4] FRONTEND: Frontend time: 1.835s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16436556 52.11868832 + layer.0.v_cache 0.00001441 0.01114783 + layer.1.k_cache 0.62741763 5.90886445 + layer.1.v_cache 0.00000560 0.00469077 + layer.2.k_cache 0.01487184 1.36169209 + layer.2.v_cache 0.00002167 0.01523863 + layer.3.k_cache 0.02142572 8.03019377 + layer.3.v_cache 0.00001943 0.01816340 + layer.4.k_cache 0.00070776 0.39837676 + layer.4.v_cache 0.00005251 0.03339697 + layer.4.output 0.00894435 193.84953008 + ------------------------------------------------------------------------------------- + TOTAL 0.05244192 83.81453903 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 219536 +BPFP 0.7080 bits/point +EBPFP 1.4160 equivalent bits/point +MSE 83.814539 +---------------------- -------------------------------------------------------- +Time: 3.196s Load: 0.011s, Pack+Encode: 1.835s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8145 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,036B, BPFP=0.2800 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,096B, BPFP=1.6735 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,316B, BPFP=0.5180 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,728B, BPFP=1.5974 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,552B, BPFP=0.8092 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,040B, BPFP=1.5592 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,916B, BPFP=0.7182 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,708B, BPFP=1.5963 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,956B, BPFP=1.2765 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,828B, BPFP=1.5474 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,524B, BPFP=0.0598 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.342s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11705062 52.23247053 + layer.0.v_cache 0.00001445 0.01123152 + layer.1.k_cache 0.56620099 6.29169157 + layer.1.v_cache 0.00000566 0.00454306 + layer.2.k_cache 0.01299965 1.08603915 + layer.2.v_cache 0.00001972 0.01481030 + layer.3.k_cache 0.05069733 7.75624948 + layer.3.v_cache 0.00001944 0.01765712 + layer.4.k_cache 0.00068883 0.38378876 + layer.4.v_cache 0.00005144 0.03242959 + layer.4.output 0.00838925 197.36570602 + ------------------------------------------------------------------------------------- + TOTAL 0.04743958 85.25828549 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 215700 +BPFP 0.7055 bits/point +EBPFP 1.4111 equivalent bits/point +MSE 85.258285 +---------------------- -------------------------------------------------------- +Time: 3.190s Load: 0.010s, Pack+Encode: 1.838s, Decode+Unpack: 1.342s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2583 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,060B, BPFP=0.2794 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,092B, BPFP=1.6614 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,712B, BPFP=0.5362 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,068B, BPFP=1.6049 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,012B, BPFP=0.8288 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,220B, BPFP=1.5581 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,052B, BPFP=0.7758 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,660B, BPFP=1.5824 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,836B, BPFP=1.3160 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,816B, BPFP=1.5358 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,192B, BPFP=0.0646 +⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.342s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14370025 53.40326855 + layer.0.v_cache 0.00001382 0.01106841 + layer.1.k_cache 0.61448195 6.39169473 + layer.1.v_cache 0.00000570 0.00459051 + layer.2.k_cache 0.01208573 1.15747760 + layer.2.v_cache 0.00001823 0.01481847 + layer.3.k_cache 0.01319265 8.12416837 + layer.3.v_cache 0.00002242 0.01757213 + layer.4.k_cache 0.00071920 0.40194125 + layer.4.v_cache 0.00005145 0.03270552 + layer.4.output 0.00992508 195.76266721 + ------------------------------------------------------------------------------------- + TOTAL 0.05022159 84.69988094 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 219720 +BPFP 0.7136 bits/point +EBPFP 1.4272 equivalent bits/point +MSE 84.699881 +---------------------- -------------------------------------------------------- +Time: 3.180s Load: 0.009s, Pack+Encode: 1.828s, Decode+Unpack: 1.342s +---------------------- -------------------------------------------------------- +💾 Converting with 84.6999 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 259, 128) +Output shape: (1, 259, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.output: torch.Size([1, 259, 3584]) -> torch.Size([1, 1, 259, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,652B, BPFP=0.2806 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,664B, BPFP=1.7896 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,788B, BPFP=0.5302 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,552B, BPFP=1.7225 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,832B, BPFP=0.7138 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,624B, BPFP=1.6665 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,932B, BPFP=0.7802 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,428B, BPFP=1.7150 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,968B, BPFP=1.3253 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,360B, BPFP=1.6506 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,716B, BPFP=0.0665 +⌛️ [2/4] FRONTEND: Frontend time: 1.826s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13349269 53.59164605 + layer.0.v_cache 0.00001436 0.01140545 + layer.1.k_cache 0.50562990 6.26749140 + layer.1.v_cache 0.00000590 0.00486981 + layer.2.k_cache 0.00820730 1.03739370 + layer.2.v_cache 0.00001794 0.01456556 + layer.3.k_cache 0.03685973 7.28081733 + layer.3.v_cache 0.00002061 0.01757055 + layer.4.k_cache 0.00069539 0.40487164 + layer.4.v_cache 0.00005422 0.03313390 + layer.4.output 0.01018517 213.36662300 + ------------------------------------------------------------------------------------- + TOTAL 0.04448790 91.89588979 + (elements=2,254,336) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2254336 +Total Bytes 209516 +BPFP 0.7435 bits/point +EBPFP 1.4870 equivalent bits/point +MSE 91.895890 +---------------------- -------------------------------------------------------- +Time: 3.182s Load: 0.010s, Pack+Encode: 1.826s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8959 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,940B, BPFP=0.2859 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,348B, BPFP=1.7563 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,112B, BPFP=0.5273 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,840B, BPFP=1.6690 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,256B, BPFP=0.7671 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,204B, BPFP=1.6322 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,168B, BPFP=0.8199 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,824B, BPFP=1.6681 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,524B, BPFP=1.3613 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,012B, BPFP=1.6211 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,232B, BPFP=0.0681 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12649573 53.05215929 + layer.0.v_cache 0.00001373 0.01113675 + layer.1.k_cache 0.51574391 6.05878725 + layer.1.v_cache 0.00000589 0.00463024 + layer.2.k_cache 0.01194659 1.10210006 + layer.2.v_cache 0.00001984 0.01505520 + layer.3.k_cache 0.04962628 8.17289225 + layer.3.v_cache 0.00002059 0.01754636 + layer.4.k_cache 0.00068541 0.39141993 + layer.4.v_cache 0.00005704 0.03337073 + layer.4.output 0.00772497 204.58649140 + ------------------------------------------------------------------------------------- + TOTAL 0.04462881 88.29203164 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 217460 +BPFP 0.7403 bits/point +EBPFP 1.4805 equivalent bits/point +MSE 88.292032 +---------------------- -------------------------------------------------------- +Time: 3.186s Load: 0.009s, Pack+Encode: 1.834s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 88.2920 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,916B, BPFP=0.2855 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,112B, BPFP=1.7491 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,028B, BPFP=0.5244 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,788B, BPFP=1.6722 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,928B, BPFP=0.8090 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,244B, BPFP=1.6406 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,328B, BPFP=0.7742 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,800B, BPFP=1.6729 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,652B, BPFP=1.3158 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,788B, BPFP=1.6141 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,344B, BPFP=0.0609 +⌛️ [2/4] FRONTEND: Frontend time: 1.835s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.342s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12210708 51.74491389 + layer.0.v_cache 0.00001342 0.01089613 + layer.1.k_cache 0.54942129 6.36351405 + layer.1.v_cache 0.00000547 0.00451703 + layer.2.k_cache 0.01053270 1.17655985 + layer.2.v_cache 0.00001820 0.01465069 + layer.3.k_cache 0.02850239 7.84936614 + layer.3.v_cache 0.00002491 0.01809898 + layer.4.k_cache 0.00067987 0.40431939 + layer.4.v_cache 0.00005179 0.03162327 + layer.4.output 0.01126663 205.58540228 + ------------------------------------------------------------------------------------- + TOTAL 0.04648373 88.63036914 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 214928 +BPFP 0.7344 bits/point +EBPFP 1.4687 equivalent bits/point +MSE 88.630369 +---------------------- -------------------------------------------------------- +Time: 3.187s Load: 0.009s, Pack+Encode: 1.835s, Decode+Unpack: 1.342s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6304 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,376B, BPFP=0.2772 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,476B, BPFP=1.5716 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,840B, BPFP=0.5074 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,072B, BPFP=1.4992 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,732B, BPFP=0.8113 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,412B, BPFP=1.4651 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,652B, BPFP=0.7040 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,888B, BPFP=1.4897 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,136B, BPFP=1.2446 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,748B, BPFP=1.4309 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,436B, BPFP=0.0621 +⌛️ [2/4] FRONTEND: Frontend time: 1.825s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14443027 53.61513446 + layer.0.v_cache 0.00001375 0.01095440 + layer.1.k_cache 0.62837083 6.58554430 + layer.1.v_cache 0.00000570 0.00451431 + layer.2.k_cache 0.01557094 1.23503450 + layer.2.v_cache 0.00001871 0.01459394 + layer.3.k_cache 0.04676386 8.35872638 + layer.3.v_cache 0.00001981 0.01700442 + layer.4.k_cache 0.00070582 0.39905674 + layer.4.v_cache 0.00005437 0.03187263 + layer.4.output 0.04838977 178.24846770 + ------------------------------------------------------------------------------------- + TOTAL 0.06909897 77.53010059 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 221768 +BPFP 0.6727 bits/point +EBPFP 1.3454 equivalent bits/point +MSE 77.530101 +---------------------- -------------------------------------------------------- +Time: 3.183s Load: 0.010s, Pack+Encode: 1.825s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 77.5301 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,024B, BPFP=0.2774 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,112B, BPFP=1.6625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,508B, BPFP=0.5250 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,052B, BPFP=1.6040 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,420B, BPFP=0.8514 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,440B, BPFP=1.5702 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,908B, BPFP=0.8231 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,980B, BPFP=1.6000 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,104B, BPFP=1.2756 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,968B, BPFP=1.5442 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,796B, BPFP=0.0615 +⌛️ [2/4] FRONTEND: Frontend time: 1.851s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13293209 53.31804881 + layer.0.v_cache 0.00001388 0.01112111 + layer.1.k_cache 0.56267701 6.18646348 + layer.1.v_cache 0.00000566 0.00454084 + layer.2.k_cache 0.01145046 1.28979330 + layer.2.v_cache 0.00001921 0.01498105 + layer.3.k_cache 0.02674564 7.67006336 + layer.3.v_cache 0.00001926 0.01731033 + layer.4.k_cache 0.00069135 0.41475074 + layer.4.v_cache 0.00005493 0.03384492 + layer.4.output 0.00973303 195.37676678 + ------------------------------------------------------------------------------------- + TOTAL 0.04722004 84.50578150 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 220312 +BPFP 0.7155 bits/point +EBPFP 1.4310 equivalent bits/point +MSE 84.505781 +---------------------- -------------------------------------------------------- +Time: 3.208s Load: 0.011s, Pack+Encode: 1.851s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 84.5058 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,992B, BPFP=0.2816 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,336B, BPFP=1.7112 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,404B, BPFP=0.5305 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,212B, BPFP=1.6478 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,796B, BPFP=0.8346 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,412B, BPFP=1.6027 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,404B, BPFP=0.8125 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,992B, BPFP=1.6354 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,088B, BPFP=1.3588 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,960B, BPFP=1.5772 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,544B, BPFP=0.0608 +⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15182159 54.04569748 + layer.0.v_cache 0.00001420 0.01116499 + layer.1.k_cache 0.56224567 6.40182726 + layer.1.v_cache 0.00000588 0.00468707 + layer.2.k_cache 0.00652889 1.28751029 + layer.2.v_cache 0.00001844 0.01449538 + layer.3.k_cache 0.06011557 7.83094011 + layer.3.v_cache 0.00001921 0.01770936 + layer.4.k_cache 0.00068317 0.40203637 + layer.4.v_cache 0.00005096 0.03135989 + layer.4.output 0.01130176 200.04074265 + ------------------------------------------------------------------------------------- + TOTAL 0.05062446 86.49015452 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 220140 +BPFP 0.7304 bits/point +EBPFP 1.4609 equivalent bits/point +MSE 86.490155 +---------------------- -------------------------------------------------------- +Time: 3.193s Load: 0.010s, Pack+Encode: 1.828s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 86.4902 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,080B, BPFP=0.2785 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,360B, BPFP=1.6645 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,648B, BPFP=0.5289 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,308B, BPFP=1.6068 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,632B, BPFP=0.8570 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,404B, BPFP=1.5572 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,844B, BPFP=0.7590 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,856B, BPFP=1.5820 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,936B, BPFP=1.3123 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,028B, BPFP=1.5366 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,176B, BPFP=0.0640 +⌛️ [2/4] FRONTEND: Frontend time: 1.842s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13053865 52.73756168 + layer.0.v_cache 0.00001414 0.01170137 + layer.1.k_cache 0.56363723 6.16224644 + layer.1.v_cache 0.00000590 0.00481933 + layer.2.k_cache 0.01033439 1.27945375 + layer.2.v_cache 0.00002015 0.01490843 + layer.3.k_cache 0.03655195 8.47349061 + layer.3.v_cache 0.00002026 0.01774491 + layer.4.k_cache 0.00068516 0.40111791 + layer.4.v_cache 0.00005378 0.03303993 + layer.4.output 0.00863618 193.53137531 + ------------------------------------------------------------------------------------- + TOTAL 0.04719499 83.75621833 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 221272 +BPFP 0.7136 bits/point +EBPFP 1.4272 equivalent bits/point +MSE 83.756218 +---------------------- -------------------------------------------------------- +Time: 3.199s Load: 0.010s, Pack+Encode: 1.842s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 83.7562 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,976B, BPFP=0.2807 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,296B, BPFP=1.7089 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,204B, BPFP=0.5192 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,080B, BPFP=1.6403 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,980B, BPFP=0.8450 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,316B, BPFP=1.5972 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,224B, BPFP=0.8023 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,904B, BPFP=1.6304 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,156B, BPFP=1.3062 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,928B, BPFP=1.5754 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,048B, BPFP=0.0649 +⌛️ [2/4] FRONTEND: Frontend time: 1.842s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15908699 51.77691011 + layer.0.v_cache 0.00001433 0.01102095 + layer.1.k_cache 0.53822481 6.61565303 + layer.1.v_cache 0.00000570 0.00464001 + layer.2.k_cache 0.01139215 1.22653507 + layer.2.v_cache 0.00001845 0.01470229 + layer.3.k_cache 0.02907990 7.77758172 + layer.3.v_cache 0.00001963 0.01782661 + layer.4.k_cache 0.00069750 0.39702953 + layer.4.v_cache 0.00005397 0.03354553 + layer.4.output 0.01069233 198.73091800 + ------------------------------------------------------------------------------------- + TOTAL 0.04784940 85.82305123 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 219112 +BPFP 0.7270 bits/point +EBPFP 1.4541 equivalent bits/point +MSE 85.823051 +---------------------- -------------------------------------------------------- +Time: 3.202s Load: 0.009s, Pack+Encode: 1.842s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8231 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,036B, BPFP=0.2751 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,208B, BPFP=1.6503 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,696B, BPFP=0.5297 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,828B, BPFP=1.5750 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,048B, BPFP=0.8767 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,236B, BPFP=1.5426 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,824B, BPFP=0.7552 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,848B, BPFP=1.5760 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,060B, BPFP=1.3145 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,808B, BPFP=1.5192 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,360B, BPFP=0.0652 +⌛️ [2/4] FRONTEND: Frontend time: 1.844s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16084659 52.42655635 + layer.0.v_cache 0.00001440 0.01110836 + layer.1.k_cache 0.59028764 6.25666948 + layer.1.v_cache 0.00000559 0.00468799 + layer.2.k_cache 0.00929020 1.43353239 + layer.2.v_cache 0.00002205 0.01489307 + layer.3.k_cache 0.04382113 8.29801631 + layer.3.v_cache 0.00002103 0.01787587 + layer.4.k_cache 0.00069271 0.39415896 + layer.4.v_cache 0.00005052 0.03222864 + layer.4.output 0.00984751 193.02168144 + ------------------------------------------------------------------------------------- + TOTAL 0.05141085 83.53185280 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 220952 +BPFP 0.7101 bits/point +EBPFP 1.4201 equivalent bits/point +MSE 83.531853 +---------------------- -------------------------------------------------------- +Time: 3.201s Load: 0.009s, Pack+Encode: 1.844s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 83.5319 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,048B, BPFP=0.2739 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,200B, BPFP=1.6385 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,600B, BPFP=0.5208 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,768B, BPFP=1.5608 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,100B, BPFP=0.7650 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,216B, BPFP=1.5308 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,912B, BPFP=0.7548 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,800B, BPFP=1.5625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,400B, BPFP=1.2695 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,580B, BPFP=1.4963 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,100B, BPFP=0.0705 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14211286 54.65640259 + layer.0.v_cache 0.00001516 0.01105373 + layer.1.k_cache 0.60291523 6.18056997 + layer.1.v_cache 0.00000570 0.00472296 + layer.2.k_cache 0.01041983 1.25204150 + layer.2.v_cache 0.00001835 0.01472367 + layer.3.k_cache 0.03896446 8.19982232 + layer.3.v_cache 0.00001926 0.01764336 + layer.4.k_cache 0.00071470 0.40545461 + layer.4.v_cache 0.00005236 0.03325067 + layer.4.output 0.00754456 192.45086186 + ------------------------------------------------------------------------------------- + TOTAL 0.04988528 83.40774814 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 218724 +BPFP 0.6980 bits/point +EBPFP 1.3961 equivalent bits/point +MSE 83.407748 +---------------------- -------------------------------------------------------- +Time: 3.204s Load: 0.012s, Pack+Encode: 1.838s, Decode+Unpack: 1.354s +---------------------- -------------------------------------------------------- +💾 Converting with 83.4077 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 296, 128) +Output shape: (1, 296, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.output: torch.Size([1, 296, 3584]) -> torch.Size([1, 1, 296, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,236B, BPFP=0.2764 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,132B, BPFP=1.5906 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,856B, BPFP=0.5203 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,892B, BPFP=1.5251 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,212B, BPFP=0.6974 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,320B, BPFP=1.4949 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,732B, BPFP=0.7777 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,908B, BPFP=1.5260 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,384B, BPFP=1.2872 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,740B, BPFP=1.4643 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,156B, BPFP=0.0615 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14616069 53.17009013 + layer.0.v_cache 0.00001366 0.01105895 + layer.1.k_cache 0.61022733 6.62305430 + layer.1.v_cache 0.00000576 0.00463676 + layer.2.k_cache 0.01378242 1.20364751 + layer.2.v_cache 0.00001907 0.01507669 + layer.3.k_cache 0.03184220 8.09775605 + layer.3.v_cache 0.00001934 0.01835467 + layer.4.k_cache 0.00072724 0.41196841 + layer.4.v_cache 0.00005080 0.03296847 + layer.4.output 0.04784788 182.26932010 + ------------------------------------------------------------------------------------- + TOTAL 0.06692845 79.14552074 + (elements=2,576,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2576384 +Total Bytes 219568 +BPFP 0.6818 bits/point +EBPFP 1.3636 equivalent bits/point +MSE 79.145521 +---------------------- -------------------------------------------------------- +Time: 3.187s Load: 0.009s, Pack+Encode: 1.830s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 79.1455 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,308B, BPFP=0.2783 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,528B, BPFP=1.6007 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,780B, BPFP=0.5128 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,164B, BPFP=1.5292 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,312B, BPFP=0.8029 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,408B, BPFP=1.4895 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,976B, BPFP=0.7328 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,148B, BPFP=1.5283 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,992B, BPFP=1.2580 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,012B, BPFP=1.4688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,464B, BPFP=0.0634 +⌛️ [2/4] FRONTEND: Frontend time: 1.836s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15387250 52.37966325 + layer.0.v_cache 0.00001382 0.01099544 + layer.1.k_cache 0.63844893 6.74259754 + layer.1.v_cache 0.00000583 0.00477265 + layer.2.k_cache 0.01105453 1.22580284 + layer.2.v_cache 0.00001885 0.01509292 + layer.3.k_cache 0.03893430 8.32075900 + layer.3.v_cache 0.00001923 0.01806579 + layer.4.k_cache 0.00070257 0.39806893 + layer.4.v_cache 0.00005146 0.03254197 + layer.4.output 0.04873301 180.35918025 + ------------------------------------------------------------------------------------- + TOTAL 0.06966195 78.33309542 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 222092 +BPFP 0.6850 bits/point +EBPFP 1.3700 equivalent bits/point +MSE 78.333095 +---------------------- -------------------------------------------------------- +Time: 3.195s Load: 0.010s, Pack+Encode: 1.836s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 78.3331 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,952B, BPFP=0.2855 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,476B, BPFP=1.7571 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,088B, BPFP=0.5240 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,252B, BPFP=1.6866 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,780B, BPFP=0.7945 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,436B, BPFP=1.6395 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,932B, BPFP=0.8033 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,100B, BPFP=1.6778 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,692B, BPFP=1.3660 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,696B, BPFP=1.5969 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,664B, BPFP=0.0714 +⌛️ [2/4] FRONTEND: Frontend time: 1.827s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12629252 55.06502249 + layer.0.v_cache 0.00001451 0.01129972 + layer.1.k_cache 0.49097243 6.43201900 + layer.1.v_cache 0.00000577 0.00474425 + layer.2.k_cache 0.01060468 1.09018552 + layer.2.v_cache 0.00001909 0.01496790 + layer.3.k_cache 0.02748298 8.18253250 + layer.3.v_cache 0.00001876 0.01730535 + layer.4.k_cache 0.00069967 0.39484118 + layer.4.v_cache 0.00005414 0.03243954 + layer.4.output 0.01047103 203.12386334 + ------------------------------------------------------------------------------------- + TOTAL 0.04290952 87.83014122 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 219068 +BPFP 0.7430 bits/point +EBPFP 1.4860 equivalent bits/point +MSE 87.830141 +---------------------- -------------------------------------------------------- +Time: 3.184s Load: 0.010s, Pack+Encode: 1.827s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 87.8301 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,164B, BPFP=0.2892 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,972B, BPFP=1.6785 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,520B, BPFP=0.5332 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,884B, BPFP=1.6176 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,212B, BPFP=0.7959 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,212B, BPFP=1.5800 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,908B, BPFP=0.7789 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,780B, BPFP=1.6118 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,504B, BPFP=1.3163 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,944B, BPFP=1.5650 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,328B, BPFP=0.0666 +⌛️ [2/4] FRONTEND: Frontend time: 1.829s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15022702 52.10870646 + layer.0.v_cache 0.00001395 0.01094908 + layer.1.k_cache 0.54271635 6.20971855 + layer.1.v_cache 0.00000573 0.00462363 + layer.2.k_cache 0.01332666 1.17887064 + layer.2.v_cache 0.00001832 0.01457739 + layer.3.k_cache 0.02579262 7.95183429 + layer.3.v_cache 0.00001838 0.01665405 + layer.4.k_cache 0.00069325 0.39202815 + layer.4.v_cache 0.00005155 0.03227918 + layer.4.output 0.01006869 198.58083717 + ------------------------------------------------------------------------------------- + TOTAL 0.04725557 85.76388833 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 218428 +BPFP 0.7196 bits/point +EBPFP 1.4391 equivalent bits/point +MSE 85.763888 +---------------------- -------------------------------------------------------- +Time: 3.185s Load: 0.009s, Pack+Encode: 1.829s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7639 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,216B, BPFP=0.2763 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,288B, BPFP=1.6042 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,780B, BPFP=0.5180 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,888B, BPFP=1.5301 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,460B, BPFP=0.7129 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,172B, BPFP=1.4922 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,992B, BPFP=0.7411 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,904B, BPFP=1.5309 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,076B, BPFP=1.2752 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,792B, BPFP=1.4720 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,940B, BPFP=0.0676 +⌛️ [2/4] FRONTEND: Frontend time: 1.832s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12416320 52.81327463 + layer.0.v_cache 0.00001370 0.01087077 + layer.1.k_cache 0.62725913 6.72515476 + layer.1.v_cache 0.00000565 0.00461268 + layer.2.k_cache 0.01102555 1.08944785 + layer.2.v_cache 0.00002033 0.01485113 + layer.3.k_cache 0.04118381 8.18421527 + layer.3.v_cache 0.00001949 0.01771923 + layer.4.k_cache 0.00070768 0.40433895 + layer.4.v_cache 0.00005397 0.03273697 + layer.4.output 0.05105524 183.26174334 + ------------------------------------------------------------------------------------- + TOTAL 0.06834348 79.53702504 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 219508 +BPFP 0.6839 bits/point +EBPFP 1.3678 equivalent bits/point +MSE 79.537025 +---------------------- -------------------------------------------------------- +Time: 3.187s Load: 0.009s, Pack+Encode: 1.832s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 79.5370 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,032B, BPFP=0.2721 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,352B, BPFP=1.6410 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,528B, BPFP=0.5151 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,208B, BPFP=1.5792 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,520B, BPFP=0.7310 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,516B, BPFP=1.5417 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,584B, BPFP=0.7344 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,028B, BPFP=1.5694 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,792B, BPFP=1.2863 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,848B, BPFP=1.5056 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,184B, BPFP=0.0632 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15516999 53.57278601 + layer.0.v_cache 0.00001361 0.01125732 + layer.1.k_cache 0.56212566 6.29511364 + layer.1.v_cache 0.00000565 0.00474357 + layer.2.k_cache 0.00839357 1.28365318 + layer.2.v_cache 0.00001980 0.01500247 + layer.3.k_cache 0.02402391 7.99619174 + layer.3.v_cache 0.00001987 0.01743745 + layer.4.k_cache 0.00071212 0.41125211 + layer.4.v_cache 0.00005188 0.03297317 + layer.4.output 0.05073151 187.16298505 + ------------------------------------------------------------------------------------- + TOTAL 0.06503863 81.16360623 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 218592 +BPFP 0.6952 bits/point +EBPFP 1.3904 equivalent bits/point +MSE 81.163606 +---------------------- -------------------------------------------------------- +Time: 3.195s Load: 0.012s, Pack+Encode: 1.838s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 81.1636 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,272B, BPFP=0.2774 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,264B, BPFP=1.5922 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,864B, BPFP=0.5189 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,860B, BPFP=1.5183 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,904B, BPFP=0.7315 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,180B, BPFP=1.4825 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,468B, BPFP=0.7085 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,912B, BPFP=1.5210 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,336B, BPFP=1.2803 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,696B, BPFP=1.4571 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,348B, BPFP=0.0627 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005571 52.74804030 + layer.0.v_cache 0.00001459 0.01113700 + layer.1.k_cache 0.64038677 6.58331854 + layer.1.v_cache 0.00000545 0.00457706 + layer.2.k_cache 0.01623841 1.19862093 + layer.2.v_cache 0.00002030 0.01537084 + layer.3.k_cache 0.01774169 8.37293590 + layer.3.v_cache 0.00001948 0.01862275 + layer.4.k_cache 0.00069419 0.40504001 + layer.4.v_cache 0.00005057 0.03312583 + layer.4.output 0.05052170 181.49619709 + ------------------------------------------------------------------------------------- + TOTAL 0.06875759 78.81553934 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 219104 +BPFP 0.6781 bits/point +EBPFP 1.3561 equivalent bits/point +MSE 78.815539 +---------------------- -------------------------------------------------------- +Time: 3.199s Load: 0.010s, Pack+Encode: 1.838s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 78.8155 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,004B, BPFP=0.2802 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,284B, BPFP=1.6960 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,328B, BPFP=0.5224 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,176B, BPFP=1.6340 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,784B, BPFP=0.7720 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,372B, BPFP=1.5889 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,488B, BPFP=0.8114 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,944B, BPFP=1.6210 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,604B, BPFP=1.3219 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,052B, BPFP=1.5710 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,880B, BPFP=0.0630 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.360s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14010120 51.68433930 + layer.0.v_cache 0.00001437 0.01099937 + layer.1.k_cache 0.58692467 6.12341834 + layer.1.v_cache 0.00000569 0.00462171 + layer.2.k_cache 0.00806839 1.17141374 + layer.2.v_cache 0.00001934 0.01490541 + layer.3.k_cache 0.03832107 7.80023171 + layer.3.v_cache 0.00001992 0.01796352 + layer.4.k_cache 0.00069804 0.38740020 + layer.4.v_cache 0.00005131 0.03227162 + layer.4.output 0.00759059 198.47489439 + ------------------------------------------------------------------------------------- + TOTAL 0.04866812 85.68069563 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 218916 +BPFP 0.7212 bits/point +EBPFP 1.4424 equivalent bits/point +MSE 85.680696 +---------------------- -------------------------------------------------------- +Time: 3.212s Load: 0.011s, Pack+Encode: 1.841s, Decode+Unpack: 1.360s +---------------------- -------------------------------------------------------- +💾 Converting with 85.6807 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,896B, BPFP=0.2854 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,820B, BPFP=1.7386 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,180B, BPFP=0.5352 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,836B, BPFP=1.6812 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,296B, BPFP=0.8335 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,008B, BPFP=1.6329 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,652B, BPFP=0.7959 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,584B, BPFP=1.6665 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,572B, BPFP=1.3743 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,688B, BPFP=1.6143 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,556B, BPFP=0.0629 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13199043 55.04058565 + layer.0.v_cache 0.00001386 0.01119168 + layer.1.k_cache 0.56989129 6.46770785 + layer.1.v_cache 0.00000577 0.00474082 + layer.2.k_cache 0.00892691 1.38077181 + layer.2.v_cache 0.00001816 0.01447033 + layer.3.k_cache 0.04626510 7.26722376 + layer.3.v_cache 0.00001927 0.01738341 + layer.4.k_cache 0.00069392 0.39724845 + layer.4.v_cache 0.00005195 0.03270081 + layer.4.output 0.00911696 206.18201959 + ------------------------------------------------------------------------------------- + TOTAL 0.04833502 89.05342128 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 216088 +BPFP 0.7411 bits/point +EBPFP 1.4822 equivalent bits/point +MSE 89.053421 +---------------------- -------------------------------------------------------- +Time: 3.200s Load: 0.009s, Pack+Encode: 1.834s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 89.0534 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,916B, BPFP=0.2803 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,192B, BPFP=1.7217 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,184B, BPFP=0.5237 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,060B, BPFP=1.6572 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,828B, BPFP=0.7885 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,232B, BPFP=1.6099 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,116B, BPFP=0.8050 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,148B, BPFP=1.6622 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,948B, BPFP=1.3656 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,840B, BPFP=1.5876 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,816B, BPFP=0.0637 +⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16500157 51.41084768 + layer.0.v_cache 0.00001391 0.01106752 + layer.1.k_cache 0.57858722 6.18768088 + layer.1.v_cache 0.00000575 0.00461105 + layer.2.k_cache 0.01018238 1.28613627 + layer.2.v_cache 0.00001877 0.01511200 + layer.3.k_cache 0.03425542 7.51562946 + layer.3.v_cache 0.00002021 0.01777191 + layer.4.k_cache 0.00069807 0.39704650 + layer.4.v_cache 0.00005194 0.03167161 + layer.4.output 0.01023613 201.90748827 + ------------------------------------------------------------------------------------- + TOTAL 0.05061695 87.07235251 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 218280 +BPFP 0.7322 bits/point +EBPFP 1.4644 equivalent bits/point +MSE 87.072353 +---------------------- -------------------------------------------------------- +Time: 3.185s Load: 0.009s, Pack+Encode: 1.828s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 87.0724 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,920B, BPFP=0.2847 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,152B, BPFP=1.7449 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,108B, BPFP=0.5271 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,576B, BPFP=1.6537 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,600B, BPFP=0.7870 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,828B, BPFP=1.6104 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,188B, BPFP=0.7632 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,900B, BPFP=1.6725 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,596B, BPFP=1.3076 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,820B, BPFP=1.6100 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,456B, BPFP=0.0699 +⌛️ [2/4] FRONTEND: Frontend time: 1.836s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.342s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385483 53.74815538 + layer.0.v_cache 0.00001387 0.01106620 + layer.1.k_cache 0.55871571 6.70909153 + layer.1.v_cache 0.00000565 0.00452077 + layer.2.k_cache 0.00818202 1.05670369 + layer.2.v_cache 0.00001875 0.01472291 + layer.3.k_cache 0.04697755 8.04081670 + layer.3.v_cache 0.00001978 0.01736839 + layer.4.k_cache 0.00071636 0.39005619 + layer.4.v_cache 0.00005146 0.03174986 + layer.4.output 0.01106052 204.73032407 + ------------------------------------------------------------------------------------- + TOTAL 0.04741057 88.41979530 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 215144 +BPFP 0.7324 bits/point +EBPFP 1.4648 equivalent bits/point +MSE 88.419795 +---------------------- -------------------------------------------------------- +Time: 3.186s Load: 0.009s, Pack+Encode: 1.836s, Decode+Unpack: 1.342s +---------------------- -------------------------------------------------------- +💾 Converting with 88.4198 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,968B, BPFP=0.2823 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,940B, BPFP=1.7011 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,060B, BPFP=0.5148 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,796B, BPFP=1.6361 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,888B, BPFP=0.7891 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,320B, BPFP=1.6091 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,736B, BPFP=0.7805 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,964B, BPFP=1.6457 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,872B, BPFP=1.2995 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,908B, BPFP=1.5857 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,708B, BPFP=0.0626 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203608 51.33450639 + layer.0.v_cache 0.00001420 0.01097828 + layer.1.k_cache 0.56673284 6.03674627 + layer.1.v_cache 0.00000552 0.00458780 + layer.2.k_cache 0.00530624 1.07844050 + layer.2.v_cache 0.00001825 0.01483273 + layer.3.k_cache 0.02188354 7.85099343 + layer.3.v_cache 0.00001854 0.01719727 + layer.4.k_cache 0.00069888 0.38796725 + layer.4.v_cache 0.00004918 0.03128896 + layer.4.output 0.00799176 200.92448052 + ------------------------------------------------------------------------------------- + TOTAL 0.04662974 86.66111191 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 216160 +BPFP 0.7225 bits/point +EBPFP 1.4449 equivalent bits/point +MSE 86.661112 +---------------------- -------------------------------------------------------- +Time: 3.205s Load: 0.010s, Pack+Encode: 1.838s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 86.6611 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,060B, BPFP=0.2794 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,132B, BPFP=1.6636 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,428B, BPFP=0.5205 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,964B, BPFP=1.5992 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,780B, BPFP=0.8160 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,380B, BPFP=1.5669 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,180B, BPFP=0.7829 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,076B, BPFP=1.6053 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,452B, BPFP=1.2396 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,856B, BPFP=1.5380 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,788B, BPFP=0.0693 +⌛️ [2/4] FRONTEND: Frontend time: 1.835s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12732459 54.37792279 + layer.0.v_cache 0.00001480 0.01086980 + layer.1.k_cache 0.57322790 6.52269257 + layer.1.v_cache 0.00000557 0.00460712 + layer.2.k_cache 0.00950604 1.24367003 + layer.2.v_cache 0.00002039 0.01487569 + layer.3.k_cache 0.01510038 7.85393662 + layer.3.v_cache 0.00001916 0.01765977 + layer.4.k_cache 0.00068141 0.39211653 + layer.4.v_cache 0.00005165 0.03234294 + layer.4.output 0.00936729 194.59444409 + ------------------------------------------------------------------------------------- + TOTAL 0.04656017 84.27245897 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 219096 +BPFP 0.7116 bits/point +EBPFP 1.4231 equivalent bits/point +MSE 84.272459 +---------------------- -------------------------------------------------------- +Time: 3.196s Load: 0.011s, Pack+Encode: 1.835s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 84.2725 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,948B, BPFP=0.2801 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,296B, BPFP=1.7151 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,148B, BPFP=0.5179 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,936B, BPFP=1.6381 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,052B, BPFP=0.8521 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,416B, BPFP=1.6087 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,924B, BPFP=0.7883 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,928B, BPFP=1.6377 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,596B, BPFP=1.3358 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,252B, BPFP=1.5994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,992B, BPFP=0.0646 +⌛️ [2/4] FRONTEND: Frontend time: 1.837s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13630490 52.93268795 + layer.0.v_cache 0.00001544 0.01096036 + layer.1.k_cache 0.58776402 6.08991982 + layer.1.v_cache 0.00000559 0.00448797 + layer.2.k_cache 0.01000112 1.32167650 + layer.2.v_cache 0.00001940 0.01504157 + layer.3.k_cache 0.04424953 7.93374943 + layer.3.v_cache 0.00002139 0.01793015 + layer.4.k_cache 0.00071093 0.39187578 + layer.4.v_cache 0.00005158 0.03259454 + layer.4.output 0.01101471 200.41249353 + ------------------------------------------------------------------------------------- + TOTAL 0.05036746 86.56696346 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 219488 +BPFP 0.7309 bits/point +EBPFP 1.4618 equivalent bits/point +MSE 86.566963 +---------------------- -------------------------------------------------------- +Time: 3.192s Load: 0.010s, Pack+Encode: 1.837s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 86.5670 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,952B, BPFP=0.2783 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,156B, BPFP=1.6949 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,336B, BPFP=0.5247 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,044B, BPFP=1.6324 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,956B, BPFP=0.7844 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,472B, BPFP=1.6003 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,504B, BPFP=0.8152 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,924B, BPFP=1.6257 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,732B, BPFP=1.3339 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,088B, BPFP=1.5787 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,384B, BPFP=0.0673 +⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12958164 51.12541451 + layer.0.v_cache 0.00001430 0.01108796 + layer.1.k_cache 0.54121377 6.41269426 + layer.1.v_cache 0.00000577 0.00457951 + layer.2.k_cache 0.01080061 1.22438280 + layer.2.v_cache 0.00001903 0.01458643 + layer.3.k_cache 0.01783211 8.07052415 + layer.3.v_cache 0.00002028 0.01699384 + layer.4.k_cache 0.00068935 0.39837690 + layer.4.v_cache 0.00005053 0.03244957 + layer.4.output 0.00833506 198.64725719 + ------------------------------------------------------------------------------------- + TOTAL 0.04462193 85.75540531 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 219548 +BPFP 0.7259 bits/point +EBPFP 1.4517 equivalent bits/point +MSE 85.755405 +---------------------- -------------------------------------------------------- +Time: 3.189s Load: 0.010s, Pack+Encode: 1.828s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,768B, BPFP=0.2833 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,800B, BPFP=1.7704 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,040B, BPFP=0.5371 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,744B, BPFP=1.7077 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,232B, BPFP=0.7861 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,160B, BPFP=1.6730 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,680B, BPFP=0.8127 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,612B, BPFP=1.6999 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,480B, BPFP=1.3356 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,536B, BPFP=1.6359 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,568B, BPFP=0.0727 +⌛️ [2/4] FRONTEND: Frontend time: 1.827s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12547956 54.71005377 + layer.0.v_cache 0.00001492 0.01148285 + layer.1.k_cache 0.50446328 5.90552222 + layer.1.v_cache 0.00000633 0.00472456 + layer.2.k_cache 0.00671430 1.09722668 + layer.2.v_cache 0.00001992 0.01533989 + layer.3.k_cache 0.02142675 8.32151278 + layer.3.v_cache 0.00001986 0.01824188 + layer.4.k_cache 0.00069859 0.38633200 + layer.4.v_cache 0.00005414 0.03328062 + layer.4.output 0.01038299 210.38151820 + ------------------------------------------------------------------------------------- + TOTAL 0.04303403 90.77496145 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 214620 +BPFP 0.7500 bits/point +EBPFP 1.5001 equivalent bits/point +MSE 90.774961 +---------------------- -------------------------------------------------------- +Time: 3.182s Load: 0.009s, Pack+Encode: 1.827s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 90.7750 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,360B, BPFP=0.2782 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,396B, BPFP=1.5779 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,924B, BPFP=0.5152 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,120B, BPFP=1.5116 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,708B, BPFP=0.7635 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,252B, BPFP=1.4666 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,404B, BPFP=0.6958 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,012B, BPFP=1.5060 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,056B, BPFP=1.2488 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,896B, BPFP=1.4481 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,116B, BPFP=0.0676 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13970643 51.65207122 + layer.0.v_cache 0.00001441 0.01107537 + layer.1.k_cache 0.60015088 6.24489495 + layer.1.v_cache 0.00000584 0.00462993 + layer.2.k_cache 0.00648033 1.15224358 + layer.2.v_cache 0.00001840 0.01488512 + layer.3.k_cache 0.03288257 7.80392182 + layer.3.v_cache 0.00002024 0.01772683 + layer.4.k_cache 0.00070287 0.40327185 + layer.4.v_cache 0.00005060 0.03260463 + layer.4.output 0.04874230 179.24823505 + ------------------------------------------------------------------------------------- + TOTAL 0.06595463 77.76911592 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 221244 +BPFP 0.6756 bits/point +EBPFP 1.3512 equivalent bits/point +MSE 77.769116 +---------------------- -------------------------------------------------------- +Time: 3.190s Load: 0.012s, Pack+Encode: 1.830s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 77.7691 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,816B, BPFP=0.2779 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,820B, BPFP=1.7116 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,964B, BPFP=0.5239 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,344B, BPFP=1.6411 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,276B, BPFP=0.8255 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,528B, BPFP=1.6021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,376B, BPFP=0.7825 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,408B, BPFP=1.6441 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,956B, BPFP=1.2880 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,840B, BPFP=1.5692 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,040B, BPFP=0.0617 +⌛️ [2/4] FRONTEND: Frontend time: 1.945s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.459s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13960611 53.20264120 + layer.0.v_cache 0.00001491 0.01126504 + layer.1.k_cache 0.68890666 6.24706733 + layer.1.v_cache 0.00000583 0.00466500 + layer.2.k_cache 0.00940153 1.10144407 + layer.2.v_cache 0.00001882 0.01485111 + layer.3.k_cache 0.04667800 8.59970867 + layer.3.v_cache 0.00001939 0.01752877 + layer.4.k_cache 0.00070827 0.40084438 + layer.4.v_cache 0.00005291 0.03325043 + layer.4.output 0.04427218 165.53762560 + ------------------------------------------------------------------------------------- + TOTAL 0.07031281 72.25862619 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 257368 +BPFP 0.7234 bits/point +EBPFP 1.4468 equivalent bits/point +MSE 72.258626 +---------------------- -------------------------------------------------------- +Time: 3.417s Load: 0.013s, Pack+Encode: 1.945s, Decode+Unpack: 1.459s +---------------------- -------------------------------------------------------- +💾 Converting with 72.2586 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,320B, BPFP=0.2780 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,224B, BPFP=1.5794 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,940B, BPFP=0.5194 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,992B, BPFP=1.5151 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,980B, BPFP=0.7828 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,180B, BPFP=1.4726 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,852B, BPFP=0.7239 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,816B, BPFP=1.5059 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,816B, BPFP=1.2446 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,816B, BPFP=1.4536 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,216B, BPFP=0.0688 +⌛️ [2/4] FRONTEND: Frontend time: 1.832s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15817952 51.25756428 + layer.0.v_cache 0.00001474 0.01103155 + layer.1.k_cache 0.61609703 6.39973006 + layer.1.v_cache 0.00000566 0.00465643 + layer.2.k_cache 0.00844678 1.22549418 + layer.2.v_cache 0.00001876 0.01490312 + layer.3.k_cache 0.04682518 7.99852699 + layer.3.v_cache 0.00001896 0.01728456 + layer.4.k_cache 0.00069859 0.39576629 + layer.4.v_cache 0.00005216 0.03314434 + layer.4.output 0.04649196 180.69082955 + ------------------------------------------------------------------------------------- + TOTAL 0.06798830 78.36434757 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 221152 +BPFP 0.6798 bits/point +EBPFP 1.3596 equivalent bits/point +MSE 78.364348 +---------------------- -------------------------------------------------------- +Time: 3.193s Load: 0.011s, Pack+Encode: 1.832s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 78.3643 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,040B, BPFP=0.2744 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,316B, BPFP=1.6505 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,088B, BPFP=0.5492 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,960B, BPFP=1.5767 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,976B, BPFP=0.8153 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,264B, BPFP=1.5388 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,848B, BPFP=0.7539 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,836B, BPFP=1.5699 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,644B, BPFP=1.2872 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,684B, BPFP=1.5072 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,240B, BPFP=0.0641 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14002854 53.46175414 + layer.0.v_cache 0.00001601 0.01138899 + layer.1.k_cache 0.54217949 6.41927580 + layer.1.v_cache 0.00000579 0.00471785 + layer.2.k_cache 0.01082110 1.21886909 + layer.2.v_cache 0.00001971 0.01476563 + layer.3.k_cache 0.03102738 8.52675849 + layer.3.v_cache 0.00001907 0.01771443 + layer.4.k_cache 0.00066782 0.40098843 + layer.4.v_cache 0.00005141 0.03291351 + layer.4.output 0.00856273 192.69177763 + ------------------------------------------------------------------------------------- + TOTAL 0.04616326 83.46774057 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 219896 +BPFP 0.7042 bits/point +EBPFP 1.4084 equivalent bits/point +MSE 83.467741 +---------------------- -------------------------------------------------------- +Time: 3.189s Load: 0.011s, Pack+Encode: 1.834s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 83.4677 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,284B, BPFP=0.2780 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,084B, BPFP=1.5827 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,808B, BPFP=0.5160 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,856B, BPFP=1.5181 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,652B, BPFP=0.7708 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,056B, BPFP=1.4760 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,200B, BPFP=0.7471 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,812B, BPFP=1.5158 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,268B, BPFP=1.2241 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,564B, BPFP=1.4501 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,840B, BPFP=0.0664 +⌛️ [2/4] FRONTEND: Frontend time: 1.825s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14705232 53.18625710 + layer.0.v_cache 0.00001374 0.01112462 + layer.1.k_cache 0.65666707 6.30107142 + layer.1.v_cache 0.00000579 0.00462127 + layer.2.k_cache 0.01062412 1.25800896 + layer.2.v_cache 0.00001918 0.01498972 + layer.3.k_cache 0.03205616 8.27129860 + layer.3.v_cache 0.00001999 0.01846887 + layer.4.k_cache 0.00071739 0.39446277 + layer.4.v_cache 0.00005111 0.03204225 + layer.4.output 0.04976816 179.63206469 + ------------------------------------------------------------------------------------- + TOTAL 0.07032965 78.05392932 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 219424 +BPFP 0.6790 bits/point +EBPFP 1.3581 equivalent bits/point +MSE 78.053929 +---------------------- -------------------------------------------------------- +Time: 3.180s Load: 0.011s, Pack+Encode: 1.825s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 78.0539 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,332B, BPFP=0.2965 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,292B, BPFP=1.6844 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,524B, BPFP=0.5296 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,176B, BPFP=1.6223 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,560B, BPFP=0.8652 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,388B, BPFP=1.5785 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,460B, BPFP=0.8040 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,072B, BPFP=1.6165 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,352B, BPFP=1.2429 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,040B, BPFP=1.5592 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,580B, BPFP=0.0682 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131647 53.16308941 + layer.0.v_cache 0.00001392 0.01114598 + layer.1.k_cache 0.51967005 6.31166506 + layer.1.v_cache 0.00000557 0.00458990 + layer.2.k_cache 0.01013049 1.28503201 + layer.2.v_cache 0.00001908 0.01483455 + layer.3.k_cache 0.05186962 7.80772927 + layer.3.v_cache 0.00001891 0.01796022 + layer.4.k_cache 0.00069958 0.39323778 + layer.4.v_cache 0.00005017 0.03171285 + layer.4.output 0.01073789 196.84214540 + ------------------------------------------------------------------------------------- + TOTAL 0.04758583 85.11388322 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 220776 +BPFP 0.7221 bits/point +EBPFP 1.4443 equivalent bits/point +MSE 85.113883 +---------------------- -------------------------------------------------------- +Time: 3.190s Load: 0.009s, Pack+Encode: 1.834s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1139 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,060B, BPFP=0.2745 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,812B, BPFP=1.6174 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,460B, BPFP=0.5132 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,592B, BPFP=1.5512 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,852B, BPFP=0.8058 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,036B, BPFP=1.5211 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,432B, BPFP=0.7287 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,636B, BPFP=1.5536 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,656B, BPFP=1.2834 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,540B, BPFP=1.4941 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,784B, BPFP=0.0681 +⌛️ [2/4] FRONTEND: Frontend time: 1.825s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15566496 52.48521593 + layer.0.v_cache 0.00001437 0.01084065 + layer.1.k_cache 0.62643353 6.26618618 + layer.1.v_cache 0.00000573 0.00463256 + layer.2.k_cache 0.01010078 1.25345972 + layer.2.v_cache 0.00001868 0.01484659 + layer.3.k_cache 0.06271097 8.46314155 + layer.3.v_cache 0.00001958 0.01821129 + layer.4.k_cache 0.00073784 0.39455409 + layer.4.v_cache 0.00005294 0.03306772 + layer.4.output 0.00677609 192.02491009 + ------------------------------------------------------------------------------------- + TOTAL 0.05312894 83.12461923 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 217860 +BPFP 0.6953 bits/point +EBPFP 1.3905 equivalent bits/point +MSE 83.124619 +---------------------- -------------------------------------------------------- +Time: 3.182s Load: 0.012s, Pack+Encode: 1.825s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 83.1246 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,076B, BPFP=0.2773 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,972B, BPFP=1.6375 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,440B, BPFP=0.5157 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,668B, BPFP=1.5662 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,064B, BPFP=0.7684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,068B, BPFP=1.5334 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,124B, BPFP=0.7170 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,656B, BPFP=1.5656 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,644B, BPFP=1.2917 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,696B, BPFP=1.5131 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,452B, BPFP=0.0660 +⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13918158 55.36154666 + layer.0.v_cache 0.00001566 0.01117716 + layer.1.k_cache 0.57997750 6.34090312 + layer.1.v_cache 0.00000571 0.00466614 + layer.2.k_cache 0.01637170 1.17508127 + layer.2.v_cache 0.00001887 0.01515340 + layer.3.k_cache 0.04393929 7.85069350 + layer.3.v_cache 0.00001903 0.01748760 + layer.4.k_cache 0.00071104 0.40238705 + layer.4.v_cache 0.00005175 0.03283578 + layer.4.output 0.01065216 192.99338162 + ------------------------------------------------------------------------------------- + TOTAL 0.05028572 83.65680018 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 216860 +BPFP 0.6969 bits/point +EBPFP 1.3938 equivalent bits/point +MSE 83.656800 +---------------------- -------------------------------------------------------- +Time: 3.183s Load: 0.012s, Pack+Encode: 1.828s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 83.6568 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,432B, BPFP=0.2738 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,704B, BPFP=1.5476 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,216B, BPFP=0.5149 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,336B, BPFP=1.4786 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,724B, BPFP=0.6917 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,396B, BPFP=1.4312 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,304B, BPFP=0.6706 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,088B, BPFP=1.4661 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,620B, BPFP=1.2409 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,024B, BPFP=1.4125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,996B, BPFP=0.0648 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16496690 53.19034148 + layer.0.v_cache 0.00001433 0.01130131 + layer.1.k_cache 0.65265572 6.33182334 + layer.1.v_cache 0.00000578 0.00469532 + layer.2.k_cache 0.01241827 1.14281341 + layer.2.v_cache 0.00001920 0.01435993 + layer.3.k_cache 0.08444468 8.60010632 + layer.3.v_cache 0.00001975 0.01730327 + layer.4.k_cache 0.00068928 0.39074247 + layer.4.v_cache 0.00005259 0.03251229 + layer.4.output 0.04785494 173.08780242 + ------------------------------------------------------------------------------------- + TOTAL 0.07354536 75.37356565 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 221840 +BPFP 0.6577 bits/point +EBPFP 1.3155 equivalent bits/point +MSE 75.373566 +---------------------- -------------------------------------------------------- +Time: 3.218s Load: 0.010s, Pack+Encode: 1.841s, Decode+Unpack: 1.367s +---------------------- -------------------------------------------------------- +💾 Converting with 75.3736 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,220B, BPFP=0.2774 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,356B, BPFP=1.6133 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,640B, BPFP=0.5123 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,108B, BPFP=1.5470 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,292B, BPFP=0.8127 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,172B, BPFP=1.4972 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,916B, BPFP=0.7396 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,852B, BPFP=1.5334 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,968B, BPFP=1.2738 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,776B, BPFP=1.4762 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,572B, BPFP=0.0651 +⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15001445 53.94950773 + layer.0.v_cache 0.00001441 0.01124350 + layer.1.k_cache 0.57475359 6.42608933 + layer.1.v_cache 0.00000584 0.00472399 + layer.2.k_cache 0.01434258 1.24177458 + layer.2.v_cache 0.00001954 0.01513993 + layer.3.k_cache 0.02649183 7.91318476 + layer.3.v_cache 0.00001945 0.01738980 + layer.4.k_cache 0.00069414 0.40260694 + layer.4.v_cache 0.00005306 0.03308659 + layer.4.output 0.05035525 184.07848943 + ------------------------------------------------------------------------------------- + TOTAL 0.06581739 79.91553960 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 220872 +BPFP 0.6905 bits/point +EBPFP 1.3810 equivalent bits/point +MSE 79.915540 +---------------------- -------------------------------------------------------- +Time: 3.184s Load: 0.012s, Pack+Encode: 1.828s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 79.9155 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,040B, BPFP=0.2793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,320B, BPFP=1.6800 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,676B, BPFP=0.5361 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,048B, BPFP=1.6095 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,992B, BPFP=0.8307 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,348B, BPFP=1.5707 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,280B, BPFP=0.7912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,984B, BPFP=1.6059 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,448B, BPFP=1.2992 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,040B, BPFP=1.5536 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,980B, BPFP=0.0632 +⌛️ [2/4] FRONTEND: Frontend time: 1.826s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11854591 53.90694260 + layer.0.v_cache 0.00001566 0.01116324 + layer.1.k_cache 0.56985614 6.72562966 + layer.1.v_cache 0.00000555 0.00453049 + layer.2.k_cache 0.01267573 1.14152949 + layer.2.v_cache 0.00001902 0.01521888 + layer.3.k_cache 0.04477839 7.78086301 + layer.3.v_cache 0.00001896 0.01762108 + layer.4.k_cache 0.00071754 0.40799986 + layer.4.v_cache 0.00005321 0.03293378 + layer.4.output 0.00761333 195.98185790 + ------------------------------------------------------------------------------------- + TOTAL 0.04705761 84.81867279 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 220156 +BPFP 0.7176 bits/point +EBPFP 1.4351 equivalent bits/point +MSE 84.818673 +---------------------- -------------------------------------------------------- +Time: 3.179s Load: 0.010s, Pack+Encode: 1.826s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 84.8187 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,200B, BPFP=0.2764 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,668B, BPFP=1.6299 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,688B, BPFP=0.5149 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,192B, BPFP=1.5514 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,804B, BPFP=0.8399 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,504B, BPFP=1.5149 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,048B, BPFP=0.7466 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,076B, BPFP=1.5453 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,476B, BPFP=1.3008 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,012B, BPFP=1.4887 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,416B, BPFP=0.0639 +⌛️ [2/4] FRONTEND: Frontend time: 1.828s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582189 52.31904031 + layer.0.v_cache 0.00001509 0.01139507 + layer.1.k_cache 0.58311327 6.34374419 + layer.1.v_cache 0.00000606 0.00473072 + layer.2.k_cache 0.00657600 1.23999023 + layer.2.v_cache 0.00001909 0.01501436 + layer.3.k_cache 0.02234380 8.39842172 + layer.3.v_cache 0.00001905 0.01758464 + layer.4.k_cache 0.00071828 0.41236857 + layer.4.v_cache 0.00005370 0.03271314 + layer.4.output 0.05044473 182.81593173 + ------------------------------------------------------------------------------------- + TOTAL 0.06363526 79.32391324 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 223084 +BPFP 0.6974 bits/point +EBPFP 1.3948 equivalent bits/point +MSE 79.323913 +---------------------- -------------------------------------------------------- +Time: 3.189s Load: 0.011s, Pack+Encode: 1.828s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 79.3239 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 292, 128) +Output shape: (1, 292, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.output: torch.Size([1, 292, 3584]) -> torch.Size([1, 1, 292, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,152B, BPFP=0.2757 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,324B, BPFP=1.6226 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,580B, BPFP=0.5126 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,368B, BPFP=1.5715 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,944B, BPFP=0.7997 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,392B, BPFP=1.5193 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,504B, BPFP=0.7761 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,976B, BPFP=1.5505 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,000B, BPFP=1.2842 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,856B, BPFP=1.4906 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,844B, BPFP=0.0600 +⌛️ [2/4] FRONTEND: Frontend time: 1.824s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887579 52.75656170 + layer.0.v_cache 0.00001384 0.01119852 + layer.1.k_cache 0.57277983 6.51114979 + layer.1.v_cache 0.00000568 0.00465670 + layer.2.k_cache 0.01314942 1.33794539 + layer.2.v_cache 0.00002105 0.01523111 + layer.3.k_cache 0.03125376 8.69713269 + layer.3.v_cache 0.00001978 0.01797976 + layer.4.k_cache 0.00068035 0.40731508 + layer.4.v_cache 0.00005123 0.03362574 + layer.4.output 0.04859251 185.40300881 + ------------------------------------------------------------------------------------- + TOTAL 0.06511755 80.44787401 + (elements=2,541,568) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2541568 +Total Bytes 220940 +BPFP 0.6954 bits/point +EBPFP 1.3909 equivalent bits/point +MSE 80.447874 +---------------------- -------------------------------------------------------- +Time: 3.179s Load: 0.011s, Pack+Encode: 1.824s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 80.4479 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 338, 128) +Output shape: (1, 338, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.output: torch.Size([1, 338, 3584]) -> torch.Size([1, 1, 338, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,996B, BPFP=0.2772 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,460B, BPFP=1.6855 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,140B, BPFP=0.5150 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,964B, BPFP=1.6163 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,140B, BPFP=0.7461 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,848B, BPFP=1.5647 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,928B, BPFP=0.7363 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,716B, BPFP=1.6048 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,852B, BPFP=1.2875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,360B, BPFP=1.5422 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,204B, BPFP=0.0674 +⌛️ [2/4] FRONTEND: Frontend time: 1.962s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.472s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13976314 54.25755247 + layer.0.v_cache 0.00001472 0.01124522 + layer.1.k_cache 0.73750278 5.65759061 + layer.1.v_cache 0.00000589 0.00474599 + layer.2.k_cache 0.02011468 1.11955677 + layer.2.v_cache 0.00002041 0.01444355 + layer.3.k_cache 0.02722352 8.39737065 + layer.3.v_cache 0.00002029 0.01739165 + layer.4.k_cache 0.00072609 0.39971337 + layer.4.v_cache 0.00005477 0.03331265 + layer.4.output 0.04270584 159.93164888 + ------------------------------------------------------------------------------------- + TOTAL 0.07202277 69.96673324 + (elements=2,941,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2941952 +Total Bytes 260608 +BPFP 0.7087 bits/point +EBPFP 1.4173 equivalent bits/point +MSE 69.966733 +---------------------- -------------------------------------------------------- +Time: 3.448s Load: 0.014s, Pack+Encode: 1.962s, Decode+Unpack: 1.472s +---------------------- -------------------------------------------------------- +💾 Converting with 69.9667 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,100B, BPFP=0.2738 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,324B, BPFP=1.6282 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,660B, BPFP=0.5187 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,172B, BPFP=1.5664 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,796B, BPFP=0.7945 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,352B, BPFP=1.5223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,424B, BPFP=0.7208 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,936B, BPFP=1.5537 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,408B, BPFP=1.3106 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,724B, BPFP=1.4886 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,252B, BPFP=0.0633 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150811 54.13802083 + layer.0.v_cache 0.00001374 0.01106703 + layer.1.k_cache 0.66021251 6.69222844 + layer.1.v_cache 0.00000574 0.00474005 + layer.2.k_cache 0.00963001 1.21401610 + layer.2.v_cache 0.00001988 0.01469734 + layer.3.k_cache 0.08055985 7.92864382 + layer.3.v_cache 0.00001942 0.01753005 + layer.4.k_cache 0.00069026 0.40784772 + layer.4.v_cache 0.00005126 0.03373839 + layer.4.output 0.05093874 185.79364568 + ------------------------------------------------------------------------------------- + TOTAL 0.07289894 80.64812056 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 220148 +BPFP 0.6953 bits/point +EBPFP 1.3907 equivalent bits/point +MSE 80.648121 +---------------------- -------------------------------------------------------- +Time: 3.194s Load: 0.010s, Pack+Encode: 1.830s, Decode+Unpack: 1.354s +---------------------- -------------------------------------------------------- +💾 Converting with 80.6481 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,392B, BPFP=0.2781 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,244B, BPFP=1.5596 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,924B, BPFP=0.5118 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,020B, BPFP=1.4965 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,048B, BPFP=0.7244 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,260B, BPFP=1.4573 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,756B, BPFP=0.7094 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,776B, BPFP=1.4839 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,996B, BPFP=1.2374 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,756B, BPFP=1.4313 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,088B, BPFP=0.0669 +⌛️ [2/4] FRONTEND: Frontend time: 1.843s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14893842 51.54664939 + layer.0.v_cache 0.00001415 0.01087166 + layer.1.k_cache 0.65110169 6.49517963 + layer.1.v_cache 0.00000575 0.00460953 + layer.2.k_cache 0.01411552 1.17081202 + layer.2.v_cache 0.00001904 0.01490817 + layer.3.k_cache 0.03939451 7.97679617 + layer.3.v_cache 0.00001876 0.01682800 + layer.4.k_cache 0.00070558 0.40390755 + layer.4.v_cache 0.00005345 0.03274573 + layer.4.output 0.04819407 178.22184406 + ------------------------------------------------------------------------------------- + TOTAL 0.07010149 77.36624802 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 220260 +BPFP 0.6681 bits/point +EBPFP 1.3363 equivalent bits/point +MSE 77.366248 +---------------------- -------------------------------------------------------- +Time: 3.209s Load: 0.011s, Pack+Encode: 1.843s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 77.3662 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,760B, BPFP=0.2828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,656B, BPFP=1.7619 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,892B, BPFP=0.5283 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,520B, BPFP=1.6944 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,836B, BPFP=0.8220 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,088B, BPFP=1.6687 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,520B, BPFP=0.7438 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,560B, BPFP=1.6968 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,296B, BPFP=1.3246 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,280B, BPFP=1.6207 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,480B, BPFP=0.0635 +⌛️ [2/4] FRONTEND: Frontend time: 1.831s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14492303 55.25907126 + layer.0.v_cache 0.00001375 0.01105387 + layer.1.k_cache 0.50395557 5.90291233 + layer.1.v_cache 0.00000564 0.00461315 + layer.2.k_cache 0.00856250 1.25229079 + layer.2.v_cache 0.00001902 0.01531370 + layer.3.k_cache 0.02692018 7.39037529 + layer.3.v_cache 0.00002118 0.01742678 + layer.4.k_cache 0.00070790 0.39665599 + layer.4.v_cache 0.00005130 0.03195598 + layer.4.output 0.01131245 209.40424701 + ------------------------------------------------------------------------------------- + TOTAL 0.04496278 90.35949401 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 211888 +BPFP 0.7405 bits/point +EBPFP 1.4810 equivalent bits/point +MSE 90.359494 +---------------------- -------------------------------------------------------- +Time: 3.196s Load: 0.011s, Pack+Encode: 1.831s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 90.3595 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,984B, BPFP=0.2822 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,140B, BPFP=1.7063 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,144B, BPFP=0.5177 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,948B, BPFP=1.6388 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,516B, BPFP=0.8784 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,312B, BPFP=1.6028 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,896B, BPFP=0.7867 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,848B, BPFP=1.6332 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,240B, BPFP=1.3157 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,040B, BPFP=1.5874 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,232B, BPFP=0.0666 +⌛️ [2/4] FRONTEND: Frontend time: 1.863s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14113761 53.80842391 + layer.0.v_cache 0.00001369 0.01103574 + layer.1.k_cache 0.59212339 5.94089276 + layer.1.v_cache 0.00000552 0.00450263 + layer.2.k_cache 0.01129283 1.20516028 + layer.2.v_cache 0.00001853 0.01489676 + layer.3.k_cache 0.01358872 7.93133810 + layer.3.v_cache 0.00001883 0.01759267 + layer.4.k_cache 0.00070763 0.40351862 + layer.4.v_cache 0.00005153 0.03195996 + layer.4.output 0.01019159 200.39728908 + ------------------------------------------------------------------------------------- + TOTAL 0.04884114 86.59707912 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 219300 +BPFP 0.7303 bits/point +EBPFP 1.4606 equivalent bits/point +MSE 86.597079 +---------------------- -------------------------------------------------------- +Time: 3.259s Load: 0.011s, Pack+Encode: 1.863s, Decode+Unpack: 1.384s +---------------------- -------------------------------------------------------- +💾 Converting with 86.5971 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,032B, BPFP=0.2740 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,232B, BPFP=1.6459 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,400B, BPFP=0.5118 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,768B, BPFP=1.5662 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,288B, BPFP=0.8323 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,204B, BPFP=1.5355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,836B, BPFP=0.7533 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,860B, BPFP=1.5712 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,968B, BPFP=1.2504 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,916B, BPFP=1.5198 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,836B, BPFP=0.0687 +⌛️ [2/4] FRONTEND: Frontend time: 1.829s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11028925 51.83641387 + layer.0.v_cache 0.00001466 0.01113530 + layer.1.k_cache 0.56032453 6.56663848 + layer.1.v_cache 0.00000595 0.00468642 + layer.2.k_cache 0.00809042 1.35034913 + layer.2.v_cache 0.00001929 0.01489615 + layer.3.k_cache 0.03501172 7.98599381 + layer.3.v_cache 0.00001861 0.01721027 + layer.4.k_cache 0.00070074 0.40999064 + layer.4.v_cache 0.00005246 0.03205915 + layer.4.output 0.00889761 192.66013875 + ------------------------------------------------------------------------------------- + TOTAL 0.04569476 83.34413791 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 219340 +BPFP 0.7024 bits/point +EBPFP 1.4049 equivalent bits/point +MSE 83.344138 +---------------------- -------------------------------------------------------- +Time: 3.194s Load: 0.011s, Pack+Encode: 1.829s, Decode+Unpack: 1.354s +---------------------- -------------------------------------------------------- +💾 Converting with 83.3441 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 311, 128) +Output shape: (1, 311, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.output: torch.Size([1, 311, 3584]) -> torch.Size([1, 1, 311, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,464B, BPFP=0.2745 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,436B, BPFP=1.5291 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,304B, BPFP=0.5177 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,128B, BPFP=1.4634 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,352B, BPFP=0.7211 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,292B, BPFP=1.4214 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,336B, BPFP=0.6700 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,248B, BPFP=1.4695 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,600B, BPFP=1.2359 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,872B, BPFP=1.4003 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,236B, BPFP=0.0663 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14227538 52.90403939 + layer.0.v_cache 0.00001500 0.01098924 + layer.1.k_cache 0.64922478 6.06315667 + layer.1.v_cache 0.00000572 0.00463020 + layer.2.k_cache 0.00828172 1.23408749 + layer.2.v_cache 0.00001922 0.01488234 + layer.3.k_cache 0.02636586 8.06981166 + layer.3.v_cache 0.00001996 0.01819027 + layer.4.k_cache 0.00069494 0.38536800 + layer.4.v_cache 0.00005248 0.03261006 + layer.4.output 0.04686997 174.06996440 + ------------------------------------------------------------------------------------- + TOTAL 0.06794382 75.71926565 + (elements=2,706,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2706944 +Total Bytes 222268 +BPFP 0.6569 bits/point +EBPFP 1.3138 equivalent bits/point +MSE 75.719266 +---------------------- -------------------------------------------------------- +Time: 3.203s Load: 0.010s, Pack+Encode: 1.840s, Decode+Unpack: 1.353s +---------------------- -------------------------------------------------------- +💾 Converting with 75.7193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,616B, BPFP=0.2828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,252B, BPFP=1.4860 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,436B, BPFP=0.5169 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,996B, BPFP=1.4091 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,596B, BPFP=0.7105 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,360B, BPFP=1.3701 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,196B, BPFP=0.6248 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,888B, BPFP=1.4025 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,384B, BPFP=1.1877 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,152B, BPFP=1.3574 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,964B, BPFP=0.0697 +⌛️ [2/4] FRONTEND: Frontend time: 1.943s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13149661 56.04759881 + layer.0.v_cache 0.00001370 0.01070573 + layer.1.k_cache 0.50214604 6.64027507 + layer.1.v_cache 0.00000625 0.00468022 + layer.2.k_cache 0.00807812 1.24715301 + layer.2.v_cache 0.00001969 0.01534388 + layer.3.k_cache 0.06061829 8.90258119 + layer.3.v_cache 0.00001868 0.01757378 + layer.4.k_cache 0.00068906 0.40481546 + layer.4.v_cache 0.00005044 0.03241554 + layer.4.output 1.20679682 210.95845588 + ------------------------------------------------------------------------------------- + TOTAL 0.53827733 91.17837258 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 176840 +BPFP 0.6374 bits/point +EBPFP 1.2748 equivalent bits/point +MSE 91.178373 +---------------------- -------------------------------------------------------- +Time: 3.176s Load: 0.010s, Pack+Encode: 1.943s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 91.1784 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 300, 128) +Output shape: (1, 300, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.output: torch.Size([1, 300, 3584]) -> torch.Size([1, 1, 300, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,368B, BPFP=0.2796 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,252B, BPFP=1.5756 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,916B, BPFP=0.5165 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,728B, BPFP=1.4963 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,976B, BPFP=0.7800 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,044B, BPFP=1.4606 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,196B, BPFP=0.7394 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,872B, BPFP=1.5037 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,004B, BPFP=1.2502 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,696B, BPFP=1.4425 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,416B, BPFP=0.0626 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13491969 53.49034505 + layer.0.v_cache 0.00001558 0.01102093 + layer.1.k_cache 0.64590566 6.33527059 + layer.1.v_cache 0.00000556 0.00457506 + layer.2.k_cache 0.01395306 1.18827291 + layer.2.v_cache 0.00001972 0.01497086 + layer.3.k_cache 0.02379198 8.28741618 + layer.3.v_cache 0.00001942 0.01818515 + layer.4.k_cache 0.00073123 0.40028214 + layer.4.v_cache 0.00005274 0.03291372 + layer.4.output 0.05011183 180.20610119 + ------------------------------------------------------------------------------------- + TOTAL 0.06883514 78.30740947 + (elements=2,611,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2611200 +Total Bytes 220468 +BPFP 0.6755 bits/point +EBPFP 1.3509 equivalent bits/point +MSE 78.307409 +---------------------- -------------------------------------------------------- +Time: 3.206s Load: 0.010s, Pack+Encode: 1.833s, Decode+Unpack: 1.363s +---------------------- -------------------------------------------------------- +💾 Converting with 78.3074 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,136B, BPFP=0.2758 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,228B, BPFP=1.6231 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,472B, BPFP=0.5086 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,124B, BPFP=1.5638 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,240B, BPFP=0.7646 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,264B, BPFP=1.5176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,756B, BPFP=0.7386 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,936B, BPFP=1.5537 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,736B, BPFP=1.2745 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,772B, BPFP=1.4912 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,684B, BPFP=0.0666 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.360s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14826349 54.11868422 + layer.0.v_cache 0.00001415 0.01093273 + layer.1.k_cache 0.61440683 6.33825390 + layer.1.v_cache 0.00000562 0.00463608 + layer.2.k_cache 0.00551110 1.29406392 + layer.2.v_cache 0.00001912 0.01473860 + layer.3.k_cache 0.01294636 7.91129641 + layer.3.v_cache 0.00002032 0.01696483 + layer.4.k_cache 0.00070633 0.41367665 + layer.4.v_cache 0.00005074 0.03275734 + layer.4.output 0.04972342 185.82713549 + ------------------------------------------------------------------------------------- + TOTAL 0.06647106 80.64387960 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 219348 +BPFP 0.6928 bits/point +EBPFP 1.3856 equivalent bits/point +MSE 80.643880 +---------------------- -------------------------------------------------------- +Time: 3.225s Load: 0.010s, Pack+Encode: 1.855s, Decode+Unpack: 1.360s +---------------------- -------------------------------------------------------- +💾 Converting with 80.6439 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,980B, BPFP=0.2809 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,128B, BPFP=1.6995 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,324B, BPFP=0.5259 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,112B, BPFP=1.6421 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,176B, BPFP=0.8560 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,196B, BPFP=1.5905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,120B, BPFP=0.7965 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,868B, BPFP=1.6284 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,992B, BPFP=1.2969 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,800B, BPFP=1.5681 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,160B, BPFP=0.0658 +⌛️ [2/4] FRONTEND: Frontend time: 1.860s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.352s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13861307 52.17050358 + layer.0.v_cache 0.00001419 0.01114391 + layer.1.k_cache 0.53867822 6.11950595 + layer.1.v_cache 0.00000573 0.00462304 + layer.2.k_cache 0.00871976 1.33082443 + layer.2.v_cache 0.00001877 0.01444587 + layer.3.k_cache 0.02195022 7.51962291 + layer.3.v_cache 0.00001967 0.01831533 + layer.4.k_cache 0.00069227 0.38638603 + layer.4.v_cache 0.00004961 0.03161579 + layer.4.output 0.00954030 199.21534941 + ------------------------------------------------------------------------------------- + TOTAL 0.04562021 86.00673133 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 218856 +BPFP 0.7262 bits/point +EBPFP 1.4524 equivalent bits/point +MSE 86.006731 +---------------------- -------------------------------------------------------- +Time: 3.223s Load: 0.011s, Pack+Encode: 1.860s, Decode+Unpack: 1.352s +---------------------- -------------------------------------------------------- +💾 Converting with 86.0067 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,052B, BPFP=0.2760 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,824B, BPFP=1.6294 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,256B, BPFP=0.5057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,464B, BPFP=1.5551 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,428B, BPFP=0.8429 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,012B, BPFP=1.5304 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,008B, BPFP=0.7107 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,732B, BPFP=1.5697 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,672B, BPFP=1.2933 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,684B, BPFP=1.5125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,340B, BPFP=0.0651 +⌛️ [2/4] FRONTEND: Frontend time: 1.846s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.358s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13928204 54.60304783 + layer.0.v_cache 0.00001373 0.01106480 + layer.1.k_cache 0.60860016 5.93789950 + layer.1.v_cache 0.00000555 0.00457102 + layer.2.k_cache 0.01026906 1.20858647 + layer.2.v_cache 0.00001872 0.01513848 + layer.3.k_cache 0.05166252 7.28806289 + layer.3.v_cache 0.00001937 0.01841977 + layer.4.k_cache 0.00073448 0.40207171 + layer.4.v_cache 0.00005026 0.03282919 + layer.4.output 0.00697322 190.85589411 + ------------------------------------------------------------------------------------- + TOTAL 0.05055697 82.67723238 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 217472 +BPFP 0.6989 bits/point +EBPFP 1.3978 equivalent bits/point +MSE 82.677232 +---------------------- -------------------------------------------------------- +Time: 3.214s Load: 0.011s, Pack+Encode: 1.846s, Decode+Unpack: 1.358s +---------------------- -------------------------------------------------------- +💾 Converting with 82.6772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,196B, BPFP=0.2752 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,256B, BPFP=1.6025 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,684B, BPFP=0.5129 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,816B, BPFP=1.5263 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,456B, BPFP=0.8186 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,092B, BPFP=1.4879 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,748B, BPFP=0.7282 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,780B, BPFP=1.5244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,400B, BPFP=1.2394 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,816B, BPFP=1.4733 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,324B, BPFP=0.0630 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13820392 53.99666645 + layer.0.v_cache 0.00001454 0.01098988 + layer.1.k_cache 0.61684219 6.52375281 + layer.1.v_cache 0.00000570 0.00450685 + layer.2.k_cache 0.01285210 1.19157642 + layer.2.v_cache 0.00002011 0.01496660 + layer.3.k_cache 0.05980752 8.32655836 + layer.3.v_cache 0.00002263 0.01758098 + layer.4.k_cache 0.00071023 0.40179521 + layer.4.v_cache 0.00005833 0.03263883 + layer.4.output 0.05086143 182.93638015 + ------------------------------------------------------------------------------------- + TOTAL 0.06968043 79.47504079 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 219568 +BPFP 0.6841 bits/point +EBPFP 1.3682 equivalent bits/point +MSE 79.475041 +---------------------- -------------------------------------------------------- +Time: 3.190s Load: 0.011s, Pack+Encode: 1.830s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 79.4750 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,372B, BPFP=0.2789 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,740B, BPFP=1.5957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,996B, BPFP=0.5189 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,260B, BPFP=1.5189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,892B, BPFP=0.7730 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,492B, BPFP=1.4790 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,560B, BPFP=0.7039 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,200B, BPFP=1.5158 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,352B, BPFP=1.2641 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,160B, BPFP=1.4618 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,772B, BPFP=0.0651 +⌛️ [2/4] FRONTEND: Frontend time: 1.848s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14825445 52.35934904 + layer.0.v_cache 0.00001396 0.01126910 + layer.1.k_cache 0.63210902 6.11286881 + layer.1.v_cache 0.00000581 0.00477967 + layer.2.k_cache 0.01184132 1.32333273 + layer.2.v_cache 0.00001890 0.01477884 + layer.3.k_cache 0.01767834 8.34552874 + layer.3.v_cache 0.00002009 0.01775849 + layer.4.k_cache 0.00068272 0.40399396 + layer.4.v_cache 0.00005247 0.03283130 + layer.4.output 0.04867053 179.40447022 + ------------------------------------------------------------------------------------- + TOTAL 0.06772769 77.90928131 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 222796 +BPFP 0.6803 bits/point +EBPFP 1.3606 equivalent bits/point +MSE 77.909281 +---------------------- -------------------------------------------------------- +Time: 3.208s Load: 0.010s, Pack+Encode: 1.848s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 77.9093 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,764B, BPFP=0.2820 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,648B, BPFP=1.7547 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,956B, BPFP=0.5301 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,428B, BPFP=1.6825 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,072B, BPFP=0.8329 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,936B, BPFP=1.6534 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,976B, BPFP=0.8864 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,440B, BPFP=1.6832 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,132B, BPFP=1.3691 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,284B, BPFP=1.6148 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,120B, BPFP=0.0602 +⌛️ [2/4] FRONTEND: Frontend time: 1.842s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14315909 53.60648600 + layer.0.v_cache 0.00001391 0.01136655 + layer.1.k_cache 0.49711875 6.43231895 + layer.1.v_cache 0.00000560 0.00471614 + layer.2.k_cache 0.00646762 1.30516538 + layer.2.v_cache 0.00001905 0.01491461 + layer.3.k_cache 0.04707547 8.50113008 + layer.3.v_cache 0.00001924 0.01788276 + layer.4.k_cache 0.00071833 0.40497040 + layer.4.v_cache 0.00005197 0.03242299 + layer.4.output 0.00997522 209.52592330 + ------------------------------------------------------------------------------------- + TOTAL 0.04496915 90.41251982 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 214756 +BPFP 0.7477 bits/point +EBPFP 1.4953 equivalent bits/point +MSE 90.412520 +---------------------- -------------------------------------------------------- +Time: 3.198s Load: 0.010s, Pack+Encode: 1.842s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 90.4125 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,004B, BPFP=0.2802 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,164B, BPFP=1.6893 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,264B, BPFP=0.5188 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,892B, BPFP=1.6181 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,844B, BPFP=0.7753 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,408B, BPFP=1.5909 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,492B, BPFP=0.7556 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,988B, BPFP=1.6234 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,896B, BPFP=1.2823 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,004B, BPFP=1.5683 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,052B, BPFP=0.0644 +⌛️ [2/4] FRONTEND: Frontend time: 1.880s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15581964 52.49376960 + layer.0.v_cache 0.00001703 0.01100323 + layer.1.k_cache 0.56298232 6.39186845 + layer.1.v_cache 0.00000550 0.00453071 + layer.2.k_cache 0.00979712 1.20879208 + layer.2.v_cache 0.00002023 0.01545736 + layer.3.k_cache 0.04960944 8.00697857 + layer.3.v_cache 0.00001971 0.01808696 + layer.4.k_cache 0.00068561 0.38934786 + layer.4.v_cache 0.00005245 0.03193020 + layer.4.output 0.01122868 198.60658282 + ------------------------------------------------------------------------------------- + TOTAL 0.05044764 85.81281440 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 217008 +BPFP 0.7149 bits/point +EBPFP 1.4298 equivalent bits/point +MSE 85.812814 +---------------------- -------------------------------------------------------- +Time: 3.264s Load: 0.009s, Pack+Encode: 1.880s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8128 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,088B, BPFP=0.2809 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,196B, BPFP=1.6672 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,368B, BPFP=0.5172 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,760B, BPFP=1.5879 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,848B, BPFP=0.8198 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,184B, BPFP=1.5561 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,692B, BPFP=0.7560 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,796B, BPFP=1.5899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,640B, BPFP=1.2500 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,752B, BPFP=1.5322 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,016B, BPFP=0.0711 +⌛️ [2/4] FRONTEND: Frontend time: 1.871s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13599485 52.44891163 + layer.0.v_cache 0.00001418 0.01109241 + layer.1.k_cache 0.56373192 6.10716652 + layer.1.v_cache 0.00000588 0.00480024 + layer.2.k_cache 0.01549433 1.19419661 + layer.2.v_cache 0.00001867 0.01486160 + layer.3.k_cache 0.03475572 7.85043238 + layer.3.v_cache 0.00001907 0.01691498 + layer.4.k_cache 0.00072111 0.41459340 + layer.4.v_cache 0.00005295 0.03409034 + layer.4.output 0.00820773 195.83094081 + ------------------------------------------------------------------------------------- + TOTAL 0.04754487 84.64197916 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 218340 +BPFP 0.7091 bits/point +EBPFP 1.4182 equivalent bits/point +MSE 84.641979 +---------------------- -------------------------------------------------------- +Time: 3.261s Load: 0.011s, Pack+Encode: 1.871s, Decode+Unpack: 1.380s +---------------------- -------------------------------------------------------- +💾 Converting with 84.6420 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,432B, BPFP=0.2756 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,364B, BPFP=1.5404 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,080B, BPFP=0.5114 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,992B, BPFP=1.4708 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,648B, BPFP=0.7431 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,456B, BPFP=1.4436 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,148B, BPFP=0.6670 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,252B, BPFP=1.4840 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,480B, BPFP=1.2419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,956B, BPFP=1.4182 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,320B, BPFP=0.0603 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13331154 54.62318004 + layer.0.v_cache 0.00001460 0.01095290 + layer.1.k_cache 0.64535681 6.77275739 + layer.1.v_cache 0.00000563 0.00462815 + layer.2.k_cache 0.00762887 1.19256552 + layer.2.v_cache 0.00001910 0.01499583 + layer.3.k_cache 0.03538683 8.38299798 + layer.3.v_cache 0.00002103 0.01815956 + layer.4.k_cache 0.00072257 0.40099409 + layer.4.v_cache 0.00005300 0.03357749 + layer.4.output 0.04522907 175.43029627 + ------------------------------------------------------------------------------------- + TOTAL 0.06700726 76.43922840 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 221128 +BPFP 0.6599 bits/point +EBPFP 1.3198 equivalent bits/point +MSE 76.439228 +---------------------- -------------------------------------------------------- +Time: 3.203s Load: 0.011s, Pack+Encode: 1.838s, Decode+Unpack: 1.353s +---------------------- -------------------------------------------------------- +💾 Converting with 76.4392 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,064B, BPFP=0.2786 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,176B, BPFP=1.6602 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,940B, BPFP=0.5469 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,052B, BPFP=1.5984 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,928B, BPFP=0.8213 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,296B, BPFP=1.5568 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,304B, BPFP=0.7320 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,904B, BPFP=1.5902 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,768B, BPFP=1.3077 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,916B, BPFP=1.5359 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,996B, BPFP=0.0707 +⌛️ [2/4] FRONTEND: Frontend time: 1.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14806804 49.83302844 + layer.0.v_cache 0.00001376 0.01112498 + layer.1.k_cache 0.60662186 5.95712882 + layer.1.v_cache 0.00000571 0.00463478 + layer.2.k_cache 0.01239752 1.20184982 + layer.2.v_cache 0.00002027 0.01458231 + layer.3.k_cache 0.02456893 7.99989684 + layer.3.v_cache 0.00001975 0.01749457 + layer.4.k_cache 0.00069718 0.40582676 + layer.4.v_cache 0.00005379 0.03211908 + layer.4.output 0.00818182 194.77034079 + ------------------------------------------------------------------------------------- + TOTAL 0.04998468 84.05118070 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 220344 +BPFP 0.7131 bits/point +EBPFP 1.4262 equivalent bits/point +MSE 84.051181 +---------------------- -------------------------------------------------------- +Time: 3.223s Load: 0.010s, Pack+Encode: 1.847s, Decode+Unpack: 1.366s +---------------------- -------------------------------------------------------- +💾 Converting with 84.0512 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,052B, BPFP=0.2789 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,228B, BPFP=1.6689 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,420B, BPFP=0.5201 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,076B, BPFP=1.6053 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,320B, BPFP=0.8458 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,340B, BPFP=1.5647 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,500B, BPFP=0.8006 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,088B, BPFP=1.6060 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,276B, BPFP=1.2851 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,056B, BPFP=1.5490 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,656B, BPFP=0.0683 +⌛️ [2/4] FRONTEND: Frontend time: 1.871s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14592592 53.97799801 + layer.0.v_cache 0.00001391 0.01114971 + layer.1.k_cache 0.53254910 6.51908653 + layer.1.v_cache 0.00000583 0.00481992 + layer.2.k_cache 0.01029040 1.23370307 + layer.2.v_cache 0.00002009 0.01535474 + layer.3.k_cache 0.03848921 7.74530353 + layer.3.v_cache 0.00001972 0.01792327 + layer.4.k_cache 0.00069417 0.41256611 + layer.4.v_cache 0.00005427 0.03419955 + layer.4.output 0.01004909 195.82376325 + ------------------------------------------------------------------------------------- + TOTAL 0.04696507 84.74932043 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 221012 +BPFP 0.7178 bits/point +EBPFP 1.4356 equivalent bits/point +MSE 84.749320 +---------------------- -------------------------------------------------------- +Time: 3.259s Load: 0.011s, Pack+Encode: 1.871s, Decode+Unpack: 1.377s +---------------------- -------------------------------------------------------- +💾 Converting with 84.7493 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,960B, BPFP=0.2849 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,324B, BPFP=1.7420 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,176B, BPFP=0.5271 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,064B, BPFP=1.6696 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,652B, BPFP=0.7842 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,296B, BPFP=1.6255 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,136B, BPFP=0.7546 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,864B, BPFP=1.6581 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,820B, BPFP=1.3109 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,708B, BPFP=1.5917 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,296B, BPFP=0.0681 +⌛️ [2/4] FRONTEND: Frontend time: 1.875s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15384994 50.87145996 + layer.0.v_cache 0.00001404 0.01124667 + layer.1.k_cache 0.48768111 5.88763832 + layer.1.v_cache 0.00000584 0.00475506 + layer.2.k_cache 0.00538594 1.21222058 + layer.2.v_cache 0.00001948 0.01466919 + layer.3.k_cache 0.02347840 8.04275872 + layer.3.v_cache 0.00001899 0.01682696 + layer.4.k_cache 0.00068094 0.38978033 + layer.4.v_cache 0.00005116 0.03222165 + layer.4.output 0.00909345 202.66938025 + ------------------------------------------------------------------------------------- + TOTAL 0.04322588 87.36289642 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 216296 +BPFP 0.7309 bits/point +EBPFP 1.4618 equivalent bits/point +MSE 87.362896 +---------------------- -------------------------------------------------------- +Time: 3.267s Load: 0.009s, Pack+Encode: 1.875s, Decode+Unpack: 1.384s +---------------------- -------------------------------------------------------- +💾 Converting with 87.3629 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,236B, BPFP=0.2783 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,084B, BPFP=1.5989 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,820B, BPFP=0.5219 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,944B, BPFP=1.5383 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,100B, BPFP=0.7494 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,180B, BPFP=1.4977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,444B, BPFP=0.7145 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,004B, BPFP=1.5415 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,016B, BPFP=1.2764 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,592B, BPFP=1.4664 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,288B, BPFP=0.0629 +⌛️ [2/4] FRONTEND: Frontend time: 1.839s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13453574 52.99124415 + layer.0.v_cache 0.00001386 0.01107559 + layer.1.k_cache 0.70462929 5.89777649 + layer.1.v_cache 0.00000580 0.00463158 + layer.2.k_cache 0.00896267 1.12659107 + layer.2.v_cache 0.00001881 0.01466471 + layer.3.k_cache 0.02468209 8.03056740 + layer.3.v_cache 0.00001940 0.01773906 + layer.4.k_cache 0.00068809 0.40252885 + layer.4.v_cache 0.00005209 0.03331252 + layer.4.output 0.04953035 183.57261297 + ------------------------------------------------------------------------------------- + TOTAL 0.07178355 79.61990719 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 218708 +BPFP 0.6837 bits/point +EBPFP 1.3675 equivalent bits/point +MSE 79.619907 +---------------------- -------------------------------------------------------- +Time: 3.204s Load: 0.010s, Pack+Encode: 1.839s, Decode+Unpack: 1.354s +---------------------- -------------------------------------------------------- +💾 Converting with 79.6199 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,872B, BPFP=0.2840 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,072B, BPFP=1.7533 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,456B, BPFP=0.5513 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,864B, BPFP=1.6828 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,008B, BPFP=0.8167 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,100B, BPFP=1.6383 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,148B, BPFP=0.8249 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,652B, BPFP=1.6705 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,724B, BPFP=1.3249 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,700B, BPFP=1.6150 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,736B, BPFP=0.0644 +⌛️ [2/4] FRONTEND: Frontend time: 1.854s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150386 52.00545855 + layer.0.v_cache 0.00001401 0.01107904 + layer.1.k_cache 0.57442873 6.55456953 + layer.1.v_cache 0.00000554 0.00461028 + layer.2.k_cache 0.00779028 1.30196654 + layer.2.v_cache 0.00001875 0.01519673 + layer.3.k_cache 0.05176373 8.53451857 + layer.3.v_cache 0.00002027 0.01813849 + layer.4.k_cache 0.00069872 0.38040434 + layer.4.v_cache 0.00005023 0.03346006 + layer.4.output 0.01044900 206.17455690 + ------------------------------------------------------------------------------------- + TOTAL 0.04937866 88.94595885 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 216332 +BPFP 0.7419 bits/point +EBPFP 1.4838 equivalent bits/point +MSE 88.945959 +---------------------- -------------------------------------------------------- +Time: 3.222s Load: 0.010s, Pack+Encode: 1.854s, Decode+Unpack: 1.359s +---------------------- -------------------------------------------------------- +💾 Converting with 88.9460 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.7067 bits/point +Avg EBPFP 1.4133 equivalent bits/point +Avg MSE 83.046943 +Avg Time 3.214s +------------------------ ---------------------------- diff --git a/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..7bc5c53ae1a1f7c519938348845e2f2eab2b6e09 --- /dev/null +++ b/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 255 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- -------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.004_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa +Output output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa +---------------- -------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,868B, BPFP=0.3475 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,180B, BPFP=2.2656 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,980B, BPFP=0.5543 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,392B, BPFP=2.1190 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,540B, BPFP=0.8445 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,928B, BPFP=2.0327 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,724B, BPFP=0.6927 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,564B, BPFP=2.1510 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,948B, BPFP=1.6644 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,968B, BPFP=2.0402 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,944B, BPFP=0.1048 +⌛️ [2/4] FRONTEND: Frontend time: 1.968s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.043s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14073507 53.51905459 + layer.0.v_cache 0.00001426 0.01296053 + layer.1.k_cache 0.05042761 5.07507579 + layer.1.v_cache 0.00000517 0.00472245 + layer.2.k_cache 0.00244098 1.30694489 + layer.2.v_cache 0.00001715 0.01353485 + layer.3.k_cache 0.02726505 5.94055648 + layer.3.v_cache 0.00001796 0.01715841 + layer.4.k_cache 0.00069123 0.38463234 + layer.4.v_cache 0.00005084 0.03584781 + layer.4.output 0.16186065 644.13275935 + ------------------------------------------------------------------------------------- + TOTAL 0.07968764 269.13175315 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 83036 +BPFP 0.9086 bits/point +EBPFP 1.8171 equivalent bits/point +MSE 269.131753 +---------------------- -------------------------------------------------------- +Time: 3.016s Load: 0.004s, Pack+Encode: 1.968s, Decode+Unpack: 1.043s +---------------------- -------------------------------------------------------- +💾 Converting with 269.1318 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.3479 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,364B, BPFP=2.1393 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,012B, BPFP=0.5670 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,852B, BPFP=2.0429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,636B, BPFP=0.6845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,808B, BPFP=1.8464 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,544B, BPFP=0.6672 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,508B, BPFP=1.9782 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,064B, BPFP=1.5181 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,764B, BPFP=1.8381 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,648B, BPFP=0.1250 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11392649 54.38260072 + layer.0.v_cache 0.00001670 0.01303877 + layer.1.k_cache 0.05249626 5.72398799 + layer.1.v_cache 0.00000506 0.00457937 + layer.2.k_cache 0.00416583 1.17619195 + layer.2.v_cache 0.00001711 0.01309025 + layer.3.k_cache 0.05931453 4.95931382 + layer.3.v_cache 0.00001789 0.01569795 + layer.4.k_cache 0.00069248 0.36232215 + layer.4.v_cache 0.00004729 0.03416418 + layer.4.output 0.16383441 651.98628442 + ------------------------------------------------------------------------------------- + TOTAL 0.08103179 272.38758695 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 77048 +BPFP 0.8532 bits/point +EBPFP 1.7064 equivalent bits/point +MSE 272.387587 +---------------------- -------------------------------------------------------- +Time: 2.500s Load: 0.005s, Pack+Encode: 1.482s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 272.3876 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,936B, BPFP=0.3218 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,100B, BPFP=2.0113 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,180B, BPFP=0.5286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,308B, BPFP=1.8797 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,616B, BPFP=0.7673 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,800B, BPFP=1.7952 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,196B, BPFP=0.8637 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,444B, BPFP=1.9023 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,028B, BPFP=1.5007 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,908B, BPFP=1.8132 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,104B, BPFP=0.0975 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17268721 55.84712641 + layer.0.v_cache 0.00001453 0.01343351 + layer.1.k_cache 0.08276152 6.29988260 + layer.1.v_cache 0.00000545 0.00473696 + layer.2.k_cache 0.00739249 1.02311536 + layer.2.v_cache 0.00001625 0.01292705 + layer.3.k_cache 0.05929063 6.65746031 + layer.3.v_cache 0.00001766 0.01708727 + layer.4.k_cache 0.00068371 0.39089572 + layer.4.v_cache 0.00005055 0.03524998 + layer.4.output 0.14466690 574.29526026 + ------------------------------------------------------------------------------------- + TOTAL 0.07856402 240.60992571 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 84620 +BPFP 0.8274 bits/point +EBPFP 1.6548 equivalent bits/point +MSE 240.609926 +---------------------- -------------------------------------------------------- +Time: 2.505s Load: 0.005s, Pack+Encode: 1.489s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 240.6099 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 112, 128) +Output shape: (1, 112, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.output: torch.Size([1, 112, 3584]) -> torch.Size([1, 1, 112, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,324B, BPFP=0.3242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,364B, BPFP=1.7249 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,000B, BPFP=0.5580 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,816B, BPFP=1.6484 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,448B, BPFP=0.7600 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,480B, BPFP=1.6016 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,496B, BPFP=0.7667 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,884B, BPFP=1.6579 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,800B, BPFP=1.3672 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,224B, BPFP=1.5658 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,560B, BPFP=0.1108 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.015s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17321212 52.75509644 + layer.0.v_cache 0.00001450 0.01285143 + layer.1.k_cache 0.08643453 6.85437502 + layer.1.v_cache 0.00000659 0.00506233 + layer.2.k_cache 0.00690480 0.99089582 + layer.2.v_cache 0.00001899 0.01404595 + layer.3.k_cache 0.01205242 6.77350507 + layer.3.v_cache 0.00001923 0.01695250 + layer.4.k_cache 0.00071999 0.38901298 + layer.4.v_cache 0.00004947 0.03420720 + layer.4.output 10.17892331 478.60833865 + ------------------------------------------------------------------------------------- + TOTAL 4.20775858 201.06496325 + (elements=974,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 974848 +Total Bytes 91396 +BPFP 0.7500 bits/point +EBPFP 1.5001 equivalent bits/point +MSE 201.064963 +---------------------- -------------------------------------------------------- +Time: 2.505s Load: 0.006s, Pack+Encode: 1.484s, Decode+Unpack: 1.015s +---------------------- -------------------------------------------------------- +💾 Converting with 201.0650 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,900B, BPFP=0.3374 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,152B, BPFP=2.1577 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,168B, BPFP=0.5625 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,656B, BPFP=2.0696 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,404B, BPFP=0.7820 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,804B, BPFP=1.9183 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,960B, BPFP=0.7031 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,492B, BPFP=2.0405 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,308B, BPFP=1.8303 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,060B, BPFP=1.9638 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,720B, BPFP=0.0944 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11687154 57.10507480 + layer.0.v_cache 0.00001423 0.01328974 + layer.1.k_cache 0.05324238 6.01900343 + layer.1.v_cache 0.00000539 0.00480914 + layer.2.k_cache 0.00397935 1.20911962 + layer.2.v_cache 0.00001766 0.01312537 + layer.3.k_cache 0.01210797 5.23750201 + layer.3.v_cache 0.00001803 0.01738435 + layer.4.k_cache 0.00075037 0.39481098 + layer.4.v_cache 0.00004481 0.03378159 + layer.4.output 0.15448949 615.19795049 + ------------------------------------------------------------------------------------- + TOTAL 0.07461636 257.43726791 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 84624 +BPFP 0.8839 bits/point +EBPFP 1.7677 equivalent bits/point +MSE 257.437268 +---------------------- -------------------------------------------------------- +Time: 2.512s Load: 0.004s, Pack+Encode: 1.495s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 257.4373 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3580 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,348B, BPFP=2.3819 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,024B, BPFP=0.5833 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,852B, BPFP=2.2863 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,524B, BPFP=0.8727 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,440B, BPFP=2.2068 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,920B, BPFP=0.7562 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,576B, BPFP=2.2330 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,176B, BPFP=1.7701 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,392B, BPFP=2.1975 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,168B, BPFP=0.1149 +⌛️ [2/4] FRONTEND: Frontend time: 1.516s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.027s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203883 53.76904900 + layer.0.v_cache 0.00001438 0.01309598 + layer.1.k_cache 0.05633916 5.99622788 + layer.1.v_cache 0.00000602 0.00534305 + layer.2.k_cache 0.00244657 1.02213890 + layer.2.v_cache 0.00001847 0.01484533 + layer.3.k_cache 0.05492775 5.28220170 + layer.3.v_cache 0.00001883 0.01765543 + layer.4.k_cache 0.00062984 0.38718647 + layer.4.v_cache 0.00005178 0.03572181 + layer.4.output 0.18538215 667.18628748 + ------------------------------------------------------------------------------------- + TOTAL 0.09142157 278.63808693 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 85276 +BPFP 0.9676 bits/point +EBPFP 1.9353 equivalent bits/point +MSE 278.638087 +---------------------- -------------------------------------------------------- +Time: 2.547s Load: 0.003s, Pack+Encode: 1.516s, Decode+Unpack: 1.027s +---------------------- -------------------------------------------------------- +💾 Converting with 278.6381 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,908B, BPFP=0.3312 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,724B, BPFP=2.0354 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,144B, BPFP=0.5458 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,468B, BPFP=1.9910 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,580B, BPFP=0.7951 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,052B, BPFP=1.9187 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,472B, BPFP=0.6028 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,492B, BPFP=1.9951 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,564B, BPFP=1.6604 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,248B, BPFP=1.7792 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,444B, BPFP=0.1102 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13028670 55.34609918 + layer.0.v_cache 0.00001462 0.01357424 + layer.1.k_cache 0.04946813 5.81545003 + layer.1.v_cache 0.00000624 0.00473968 + layer.2.k_cache 0.00411291 0.97185067 + layer.2.v_cache 0.00001729 0.01320732 + layer.3.k_cache 0.01793825 6.31653510 + layer.3.v_cache 0.00001895 0.01686747 + layer.4.k_cache 0.00077120 0.40116853 + layer.4.v_cache 0.00004532 0.03310650 + layer.4.output 0.15109633 600.70213294 + ------------------------------------------------------------------------------------- + TOTAL 0.07413847 251.40279584 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 83096 +BPFP 0.8486 bits/point +EBPFP 1.6972 equivalent bits/point +MSE 251.402796 +---------------------- -------------------------------------------------------- +Time: 2.498s Load: 0.005s, Pack+Encode: 1.481s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 251.4028 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,696B, BPFP=0.4649 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,564B, BPFP=1.7993 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,336B, BPFP=0.6404 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,304B, BPFP=1.7281 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,868B, BPFP=0.7862 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,100B, BPFP=1.6721 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,508B, BPFP=0.6875 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,356B, BPFP=1.7423 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,680B, BPFP=1.5570 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,008B, BPFP=1.6469 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,096B, BPFP=0.1212 +⌛️ [2/4] FRONTEND: Frontend time: 1.740s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.978s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11683119 66.96363590 + layer.0.v_cache 0.00001514 0.01650501 + layer.1.k_cache 0.01218004 6.38180007 + layer.1.v_cache 0.00000521 0.00610004 + layer.2.k_cache 0.00491684 1.19306732 + layer.2.v_cache 0.00001706 0.01719241 + layer.3.k_cache 0.10823302 5.67060611 + layer.3.v_cache 0.00002043 0.02268369 + layer.4.k_cache 0.00062874 0.39280908 + layer.4.v_cache 0.00004794 0.04068831 + layer.4.output 0.23833110 949.46193609 + ------------------------------------------------------------------------------------- + TOTAL 0.11242431 395.70227297 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 49516 +BPFP 0.7984 bits/point +EBPFP 1.5969 equivalent bits/point +MSE 395.702273 +---------------------- -------------------------------------------------------- +Time: 2.720s Load: 0.002s, Pack+Encode: 1.740s, Decode+Unpack: 0.978s +---------------------- -------------------------------------------------------- +💾 Converting with 395.7023 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,596B, BPFP=0.4988 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,400B, BPFP=2.0000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,704B, BPFP=0.8450 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,112B, BPFP=1.9100 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,520B, BPFP=0.7875 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,840B, BPFP=1.8250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,368B, BPFP=0.7400 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,196B, BPFP=1.9363 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,208B, BPFP=1.6275 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,788B, BPFP=1.8088 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,792B, BPFP=0.1246 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.969s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14598725 55.89195313 + layer.0.v_cache 0.00001449 0.01462481 + layer.1.k_cache 0.01347294 9.44897705 + layer.1.v_cache 0.00000521 0.00515172 + layer.2.k_cache 0.00488209 0.95456650 + layer.2.v_cache 0.00001649 0.01468128 + layer.3.k_cache 0.04421538 4.71205902 + layer.3.v_cache 0.00001905 0.01924296 + layer.4.k_cache 0.00073779 0.42088093 + layer.4.v_cache 0.00005089 0.03427374 + layer.4.output 0.27158585 1082.17517857 + ------------------------------------------------------------------------------------- + TOTAL 0.12414721 449.80839183 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 47524 +BPFP 0.8736 bits/point +EBPFP 1.7472 equivalent bits/point +MSE 449.808392 +---------------------- -------------------------------------------------------- +Time: 2.450s Load: 0.003s, Pack+Encode: 1.478s, Decode+Unpack: 0.969s +---------------------- -------------------------------------------------------- +💾 Converting with 449.8084 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 46, 128) +Output shape: (1, 46, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.output: torch.Size([1, 46, 3584]) -> torch.Size([1, 1, 46, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,624B, BPFP=0.5516 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,276B, BPFP=2.1318 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,840B, BPFP=1.3043 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,896B, BPFP=2.0027 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,960B, BPFP=1.0054 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,608B, BPFP=1.9049 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,056B, BPFP=1.0380 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,896B, BPFP=2.0027 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,192B, BPFP=1.7636 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,704B, BPFP=1.9375 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,720B, BPFP=0.1320 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.969s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14129612 63.12709642 + layer.0.v_cache 0.00001532 0.01449564 + layer.1.k_cache 0.01315293 8.67800174 + layer.1.v_cache 0.00000553 0.00534807 + layer.2.k_cache 0.00779452 1.01933537 + layer.2.v_cache 0.00001723 0.01444022 + layer.3.k_cache 0.01828384 5.63595581 + layer.3.v_cache 0.00001944 0.02056940 + layer.4.k_cache 0.00060559 0.39688492 + layer.4.v_cache 0.00005113 0.03767470 + layer.4.output 0.29517000 1172.75805512 + ------------------------------------------------------------------------------------- + TOTAL 0.13220186 487.54448166 + (elements=400,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 400384 +Total Bytes 48772 +BPFP 0.9745 bits/point +EBPFP 1.9490 equivalent bits/point +MSE 487.544482 +---------------------- -------------------------------------------------------- +Time: 2.449s Load: 0.002s, Pack+Encode: 1.478s, Decode+Unpack: 0.969s +---------------------- -------------------------------------------------------- +💾 Converting with 487.5445 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 110, 128) +Output shape: (1, 110, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.output: torch.Size([1, 110, 3584]) -> torch.Size([1, 1, 110, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,292B, BPFP=0.3256 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,284B, BPFP=1.7449 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,852B, BPFP=0.5472 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,704B, BPFP=1.6625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,072B, BPFP=0.7205 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,436B, BPFP=1.6244 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,336B, BPFP=1.1841 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,928B, BPFP=1.6943 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,824B, BPFP=1.5375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,212B, BPFP=1.5926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,052B, BPFP=0.1228 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14476244 51.81620650 + layer.0.v_cache 0.00001476 0.01338319 + layer.1.k_cache 0.05665370 6.45332586 + layer.1.v_cache 0.00000525 0.00469040 + layer.2.k_cache 0.00786287 1.21279602 + layer.2.v_cache 0.00001801 0.01342299 + layer.3.k_cache 0.01410385 7.18625211 + layer.3.v_cache 0.00002027 0.01751375 + layer.4.k_cache 0.00066873 0.39486563 + layer.4.v_cache 0.00004626 0.03246421 + layer.4.output 10.36401748 486.75726461 + ------------------------------------------------------------------------------------- + TOTAL 4.28072226 204.37916311 + (elements=957,440) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 957440 +Total Bytes 94992 +BPFP 0.7937 bits/point +EBPFP 1.5874 equivalent bits/point +MSE 204.379163 +---------------------- -------------------------------------------------------- +Time: 2.499s Load: 0.004s, Pack+Encode: 1.482s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 204.3792 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.3565 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,668B, BPFP=2.4437 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,004B, BPFP=0.5795 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,068B, BPFP=2.3279 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,788B, BPFP=0.7307 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,888B, BPFP=2.2932 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,700B, BPFP=0.7137 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,276B, BPFP=2.3681 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,072B, BPFP=1.5571 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,288B, BPFP=2.1775 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,668B, BPFP=0.1286 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.015s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10533771 51.27771750 + layer.0.v_cache 0.00001610 0.01491279 + layer.1.k_cache 0.05553475 5.85210277 + layer.1.v_cache 0.00000748 0.00639166 + layer.2.k_cache 0.02166171 1.32103079 + layer.2.v_cache 0.00002000 0.01659807 + layer.3.k_cache 0.06386927 4.60026758 + layer.3.v_cache 0.00001944 0.02013230 + layer.4.k_cache 0.00064315 0.40806278 + layer.4.v_cache 0.00005385 0.03714783 + layer.4.output 0.16801396 668.27584877 + ------------------------------------------------------------------------------------- + TOTAL 0.08372125 278.91090032 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 85268 +BPFP 0.9675 bits/point +EBPFP 1.9351 equivalent bits/point +MSE 278.910900 +---------------------- -------------------------------------------------------- +Time: 2.500s Load: 0.005s, Pack+Encode: 1.481s, Decode+Unpack: 1.015s +---------------------- -------------------------------------------------------- +💾 Converting with 278.9109 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 49, 128) +Output shape: (1, 49, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.output: torch.Size([1, 49, 3584]) -> torch.Size([1, 1, 49, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,556B, BPFP=0.4962 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,400B, BPFP=2.0408 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,172B, BPFP=1.0115 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,112B, BPFP=1.9490 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,588B, BPFP=0.8253 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,952B, BPFP=1.8980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,400B, BPFP=0.7653 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,080B, BPFP=1.9388 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,244B, BPFP=1.6722 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,912B, BPFP=1.8852 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,732B, BPFP=0.1700 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.970s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12410851 62.36318160 + layer.0.v_cache 0.00001740 0.01510808 + layer.1.k_cache 0.01401279 8.52751253 + layer.1.v_cache 0.00000579 0.00585505 + layer.2.k_cache 0.00545398 0.99115286 + layer.2.v_cache 0.00001955 0.01626270 + layer.3.k_cache 0.07546502 4.58274187 + layer.3.v_cache 0.00001978 0.02058280 + layer.4.k_cache 0.00063206 0.43807446 + layer.4.v_cache 0.00005205 0.03537678 + layer.4.output 0.27729562 1105.37153790 + ------------------------------------------------------------------------------------- + TOTAL 0.12710919 459.68215377 + (elements=426,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 426496 +Total Bytes 49148 +BPFP 0.9219 bits/point +EBPFP 1.8438 equivalent bits/point +MSE 459.682154 +---------------------- -------------------------------------------------------- +Time: 2.451s Load: 0.003s, Pack+Encode: 1.478s, Decode+Unpack: 0.970s +---------------------- -------------------------------------------------------- +💾 Converting with 459.6822 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,700B, BPFP=0.4660 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,508B, BPFP=1.7840 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,484B, BPFP=0.6809 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,252B, BPFP=1.7138 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,708B, BPFP=0.7423 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,112B, BPFP=1.6754 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,596B, BPFP=0.7116 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,280B, BPFP=1.7215 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,736B, BPFP=1.5724 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,000B, BPFP=1.6447 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,268B, BPFP=0.1280 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.972s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19197876 65.59214809 + layer.0.v_cache 0.00001449 0.01587580 + layer.1.k_cache 0.01466701 6.29116286 + layer.1.v_cache 0.00000519 0.00561221 + layer.2.k_cache 0.00720649 1.10455751 + layer.2.v_cache 0.00001785 0.01557823 + layer.3.k_cache 0.10851213 5.51909919 + layer.3.v_cache 0.00001996 0.02167974 + layer.4.k_cache 0.00062085 0.41277183 + layer.4.v_cache 0.00004770 0.03999119 + layer.4.output 0.23832821 950.29519110 + ------------------------------------------------------------------------------------- + TOTAL 0.11714047 395.94616555 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 49644 +BPFP 0.8005 bits/point +EBPFP 1.6010 equivalent bits/point +MSE 395.946166 +---------------------- -------------------------------------------------------- +Time: 2.456s Load: 0.003s, Pack+Encode: 1.480s, Decode+Unpack: 0.972s +---------------------- -------------------------------------------------------- +💾 Converting with 395.9462 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,412B, BPFP=0.4326 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,344B, BPFP=1.9436 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,752B, BPFP=0.8431 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,092B, BPFP=1.8664 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,608B, BPFP=0.7990 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,972B, BPFP=1.8297 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,460B, BPFP=0.7537 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,152B, BPFP=1.8848 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,328B, BPFP=1.6324 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,836B, BPFP=1.7880 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,948B, BPFP=0.1290 +⌛️ [2/4] FRONTEND: Frontend time: 1.479s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.976s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12665685 65.50769761 + layer.0.v_cache 0.00001436 0.01463788 + layer.1.k_cache 0.01497617 8.97947124 + layer.1.v_cache 0.00000537 0.00509612 + layer.2.k_cache 0.00986420 1.01783708 + layer.2.v_cache 0.00001802 0.01423974 + layer.3.k_cache 0.09846177 5.88233559 + layer.3.v_cache 0.00001923 0.01953471 + layer.4.k_cache 0.00062397 0.38878583 + layer.4.v_cache 0.00005014 0.03506135 + layer.4.output 0.26631049 1060.92489496 + ------------------------------------------------------------------------------------- + TOTAL 0.12440374 441.66699776 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 47904 +BPFP 0.8633 bits/point +EBPFP 1.7266 equivalent bits/point +MSE 441.666998 +---------------------- -------------------------------------------------------- +Time: 2.457s Load: 0.002s, Pack+Encode: 1.479s, Decode+Unpack: 0.976s +---------------------- -------------------------------------------------------- +💾 Converting with 441.6670 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,400B, BPFP=0.4557 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,540B, BPFP=2.1289 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,736B, BPFP=0.8906 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,300B, BPFP=2.0508 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,220B, BPFP=1.0482 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,116B, BPFP=1.9909 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,140B, BPFP=0.6966 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,428B, BPFP=2.0924 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,336B, BPFP=1.7370 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,056B, BPFP=1.9714 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,768B, BPFP=0.1287 +⌛️ [2/4] FRONTEND: Frontend time: 1.487s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.973s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18531116 59.30748494 + layer.0.v_cache 0.00001874 0.01567261 + layer.1.k_cache 0.01459585 9.57946078 + layer.1.v_cache 0.00000674 0.00635834 + layer.2.k_cache 0.01042949 1.08773518 + layer.2.v_cache 0.00001841 0.01667557 + layer.3.k_cache 0.15260293 4.39334297 + layer.3.v_cache 0.00002390 0.02293593 + layer.4.k_cache 0.00061422 0.41554534 + layer.4.v_cache 0.00004881 0.03819526 + layer.4.output 0.28297247 1125.94019717 + ------------------------------------------------------------------------------------- + TOTAL 0.13791044 468.02734042 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 49040 +BPFP 0.9390 bits/point +EBPFP 1.8781 equivalent bits/point +MSE 468.027340 +---------------------- -------------------------------------------------------- +Time: 2.463s Load: 0.003s, Pack+Encode: 1.487s, Decode+Unpack: 0.973s +---------------------- -------------------------------------------------------- +💾 Converting with 468.0273 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,792B, BPFP=0.4746 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,448B, BPFP=1.7076 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,356B, BPFP=0.6239 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,128B, BPFP=1.6229 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,716B, BPFP=0.9841 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,932B, BPFP=1.5710 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,512B, BPFP=0.6653 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,260B, BPFP=1.6578 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,692B, BPFP=1.5074 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,960B, BPFP=1.5784 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,684B, BPFP=0.1394 +⌛️ [2/4] FRONTEND: Frontend time: 1.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.977s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12052289 66.93576205 + layer.0.v_cache 0.00001546 0.01570586 + layer.1.k_cache 0.01434235 5.32235666 + layer.1.v_cache 0.00000574 0.00613649 + layer.2.k_cache 0.00254285 1.19586583 + layer.2.v_cache 0.00001710 0.01821665 + layer.3.k_cache 0.03737745 4.53815525 + layer.3.v_cache 0.00001946 0.02362295 + layer.4.k_cache 0.00063766 0.44171104 + layer.4.v_cache 0.00005039 0.04263948 + layer.4.output 0.23032015 917.62605932 + ------------------------------------------------------------------------------------- + TOTAL 0.10516308 382.46603456 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 50480 +BPFP 0.7864 bits/point +EBPFP 1.5728 equivalent bits/point +MSE 382.466035 +---------------------- -------------------------------------------------------- +Time: 2.468s Load: 0.002s, Pack+Encode: 1.488s, Decode+Unpack: 0.977s +---------------------- -------------------------------------------------------- +💾 Converting with 382.4660 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,772B, BPFP=0.3742 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,044B, BPFP=2.5431 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,800B, BPFP=0.5912 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,160B, BPFP=2.3564 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,380B, BPFP=0.7137 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,176B, BPFP=2.3598 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,976B, BPFP=0.6284 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,652B, BPFP=2.4603 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,868B, BPFP=1.6613 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,960B, BPFP=2.3142 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,760B, BPFP=0.1134 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11335979 52.59863941 + layer.0.v_cache 0.00001513 0.01452145 + layer.1.k_cache 0.01655370 5.08467308 + layer.1.v_cache 0.00000667 0.00560815 + layer.2.k_cache 0.00428703 1.11927672 + layer.2.v_cache 0.00001789 0.01575525 + layer.3.k_cache 0.04938591 4.22629485 + layer.3.v_cache 0.00002111 0.02112004 + layer.4.k_cache 0.00066117 0.42720274 + layer.4.v_cache 0.00004716 0.03573068 + layer.4.output 0.18372514 730.49565637 + ------------------------------------------------------------------------------------- + TOTAL 0.08649597 304.53049512 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 79548 +BPFP 0.9880 bits/point +EBPFP 1.9761 equivalent bits/point +MSE 304.530495 +---------------------- -------------------------------------------------------- +Time: 2.508s Load: 0.003s, Pack+Encode: 1.491s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 304.5305 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,052B, BPFP=0.7287 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,424B, BPFP=2.2812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,596B, BPFP=1.2770 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,060B, BPFP=2.1520 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,792B, BPFP=1.3466 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,896B, BPFP=2.0938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,104B, BPFP=1.1023 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,120B, BPFP=2.1733 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,192B, BPFP=1.8438 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,796B, BPFP=2.0582 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,780B, BPFP=0.1410 +⌛️ [2/4] FRONTEND: Frontend time: 1.509s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.969s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14777968 61.52329324 + layer.0.v_cache 0.00001531 0.01554903 + layer.1.k_cache 0.01728138 9.68485884 + layer.1.v_cache 0.00000547 0.00583937 + layer.2.k_cache 0.00823456 1.33616838 + layer.2.v_cache 0.00001932 0.01800127 + layer.3.k_cache 0.16770924 5.69985095 + layer.3.v_cache 0.00002046 0.02284478 + layer.4.k_cache 0.00061569 0.42362746 + layer.4.v_cache 0.00005592 0.03983225 + layer.4.output 0.30867824 1229.97828734 + ------------------------------------------------------------------------------------- + TOTAL 0.14720498 511.09516923 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 50812 +BPFP 1.0614 bits/point +EBPFP 2.1228 equivalent bits/point +MSE 511.095169 +---------------------- -------------------------------------------------------- +Time: 2.481s Load: 0.003s, Pack+Encode: 1.509s, Decode+Unpack: 0.969s +---------------------- -------------------------------------------------------- +💾 Converting with 511.0952 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3606 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,692B, BPFP=2.5425 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,008B, BPFP=0.6026 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,024B, BPFP=2.4087 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,656B, BPFP=0.7324 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,612B, BPFP=2.3261 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,516B, BPFP=0.7043 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,092B, BPFP=2.4223 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,232B, BPFP=2.0497 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,308B, BPFP=2.2652 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,052B, BPFP=0.1160 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14408422 54.58608774 + layer.0.v_cache 0.00002097 0.01495907 + layer.1.k_cache 0.05735342 5.86815858 + layer.1.v_cache 0.00000575 0.00556433 + layer.2.k_cache 0.01678689 1.03098356 + layer.2.v_cache 0.00001894 0.01600140 + layer.3.k_cache 0.03029843 4.34865238 + layer.3.v_cache 0.00002125 0.02090272 + layer.4.k_cache 0.00062582 0.41692289 + layer.4.v_cache 0.00005050 0.03739378 + layer.4.output 0.17433185 692.82881181 + ------------------------------------------------------------------------------------- + TOTAL 0.08644642 289.18513583 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 85992 +BPFP 1.0133 bits/point +EBPFP 2.0266 equivalent bits/point +MSE 289.185136 +---------------------- -------------------------------------------------------- +Time: 2.501s Load: 0.003s, Pack+Encode: 1.480s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 289.1851 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,108B, BPFP=0.7486 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,256B, BPFP=2.2216 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,656B, BPFP=1.2983 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,012B, BPFP=2.1349 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,688B, BPFP=1.3097 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,844B, BPFP=2.0753 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,160B, BPFP=1.1222 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,132B, BPFP=2.1776 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,208B, BPFP=1.8494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,660B, BPFP=2.0099 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,360B, BPFP=0.1197 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.973s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13243668 62.68727805 + layer.0.v_cache 0.00001576 0.01483751 + layer.1.k_cache 0.01981235 9.42839466 + layer.1.v_cache 0.00000577 0.00554074 + layer.2.k_cache 0.01756107 1.06724557 + layer.2.v_cache 0.00001728 0.01601122 + layer.3.k_cache 0.04920139 5.65070863 + layer.3.v_cache 0.00001837 0.02016945 + layer.4.k_cache 0.00061593 0.40449996 + layer.4.v_cache 0.00004907 0.03716665 + layer.4.output 0.30854983 1228.55225244 + ------------------------------------------------------------------------------------- + TOTAL 0.13997544 510.54103644 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 50084 +BPFP 1.0462 bits/point +EBPFP 2.0924 equivalent bits/point +MSE 510.541036 +---------------------- -------------------------------------------------------- +Time: 2.467s Load: 0.002s, Pack+Encode: 1.492s, Decode+Unpack: 0.973s +---------------------- -------------------------------------------------------- +💾 Converting with 510.5410 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,432B, BPFP=0.4475 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,472B, BPFP=2.0225 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,748B, BPFP=0.8588 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,212B, BPFP=1.9412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,680B, BPFP=0.8375 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,052B, BPFP=1.8913 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,524B, BPFP=0.7887 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,300B, BPFP=1.9688 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,304B, BPFP=1.6575 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,956B, BPFP=1.8613 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,812B, BPFP=0.1255 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.974s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20666794 63.73823242 + layer.0.v_cache 0.00002565 0.01539046 + layer.1.k_cache 0.01518770 8.60250000 + layer.1.v_cache 0.00000581 0.00546145 + layer.2.k_cache 0.01031687 1.09742203 + layer.2.v_cache 0.00001873 0.01608348 + layer.3.k_cache 0.09638391 4.72381683 + layer.3.v_cache 0.00001967 0.02049155 + layer.4.k_cache 0.00062172 0.40842392 + layer.4.v_cache 0.00004827 0.03590430 + layer.4.output 0.27162739 1080.44071429 + ------------------------------------------------------------------------------------- + TOTAL 0.13121694 449.51463097 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 48492 +BPFP 0.8914 bits/point +EBPFP 1.7828 equivalent bits/point +MSE 449.514631 +---------------------- -------------------------------------------------------- +Time: 2.458s Load: 0.002s, Pack+Encode: 1.481s, Decode+Unpack: 0.974s +---------------------- -------------------------------------------------------- +💾 Converting with 449.5146 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,580B, BPFP=0.3631 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,148B, BPFP=2.7914 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,548B, BPFP=0.5855 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,616B, BPFP=2.6691 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,268B, BPFP=0.7509 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,836B, BPFP=2.4899 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,276B, BPFP=0.9825 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,780B, BPFP=2.7068 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,820B, BPFP=1.3373 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,320B, BPFP=2.6011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,532B, BPFP=0.1159 +⌛️ [2/4] FRONTEND: Frontend time: 1.486s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.015s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13431002 58.10455681 + layer.0.v_cache 0.00001720 0.01534336 + layer.1.k_cache 0.01179634 5.85588343 + layer.1.v_cache 0.00000537 0.00535401 + layer.2.k_cache 0.00798416 1.23624757 + layer.2.v_cache 0.00001783 0.01631879 + layer.3.k_cache 0.05795463 5.87093847 + layer.3.v_cache 0.00001974 0.02110300 + layer.4.k_cache 0.00062059 0.41223408 + layer.4.v_cache 0.00005204 0.03754222 + layer.4.output 0.19989206 796.32497374 + ------------------------------------------------------------------------------------- + TOTAL 0.09482484 332.10884341 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 78724 +BPFP 1.0641 bits/point +EBPFP 2.1281 equivalent bits/point +MSE 332.108843 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.003s, Pack+Encode: 1.486s, Decode+Unpack: 1.015s +---------------------- -------------------------------------------------------- +💾 Converting with 332.1088 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,436B, BPFP=0.4487 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,440B, BPFP=2.0125 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,736B, BPFP=0.8550 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,208B, BPFP=1.9400 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,348B, BPFP=1.0462 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,052B, BPFP=1.8913 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,504B, BPFP=0.7825 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,236B, BPFP=1.9487 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,300B, BPFP=1.6562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,028B, BPFP=1.8838 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,776B, BPFP=0.1239 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.971s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15467889 62.36973633 + layer.0.v_cache 0.00001741 0.01525779 + layer.1.k_cache 0.01620592 8.42490356 + layer.1.v_cache 0.00000548 0.00552267 + layer.2.k_cache 0.00723796 1.16506279 + layer.2.v_cache 0.00001737 0.01550114 + layer.3.k_cache 0.06661982 4.81865204 + layer.3.v_cache 0.00002040 0.02087519 + layer.4.k_cache 0.00058776 0.40266270 + layer.4.v_cache 0.00004656 0.03581648 + layer.4.output 0.27162074 1081.33705357 + ------------------------------------------------------------------------------------- + TOTAL 0.12628134 449.80196269 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 49064 +BPFP 0.9019 bits/point +EBPFP 1.8038 equivalent bits/point +MSE 449.801963 +---------------------- -------------------------------------------------------- +Time: 2.454s Load: 0.002s, Pack+Encode: 1.481s, Decode+Unpack: 0.971s +---------------------- -------------------------------------------------------- +💾 Converting with 449.8020 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,428B, BPFP=0.4375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,304B, BPFP=1.9314 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,608B, BPFP=0.7990 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,084B, BPFP=1.8640 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,344B, BPFP=0.7181 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,896B, BPFP=1.8064 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,488B, BPFP=0.7623 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,136B, BPFP=1.8799 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,220B, BPFP=1.5993 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,848B, BPFP=1.7917 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,748B, BPFP=0.1203 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.970s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11261671 62.37943283 + layer.0.v_cache 0.00001829 0.01483473 + layer.1.k_cache 0.01190703 8.67442232 + layer.1.v_cache 0.00000528 0.00506546 + layer.2.k_cache 0.00973116 1.14300148 + layer.2.v_cache 0.00001668 0.01356608 + layer.3.k_cache 0.04200072 5.27647370 + layer.3.v_cache 0.00001874 0.01906361 + layer.4.k_cache 0.00062040 0.39928182 + layer.4.v_cache 0.00004776 0.03443530 + layer.4.output 0.26626884 1061.87640056 + ------------------------------------------------------------------------------------- + TOTAL 0.12005086 441.82908125 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 47104 +BPFP 0.8489 bits/point +EBPFP 1.6978 equivalent bits/point +MSE 441.829081 +---------------------- -------------------------------------------------------- +Time: 2.456s Load: 0.002s, Pack+Encode: 1.485s, Decode+Unpack: 0.970s +---------------------- -------------------------------------------------------- +💾 Converting with 441.8291 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,428B, BPFP=0.4291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,544B, BPFP=1.9663 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,624B, BPFP=0.7885 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,308B, BPFP=1.8954 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,564B, BPFP=0.7704 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,160B, BPFP=1.8510 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,476B, BPFP=0.7440 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,332B, BPFP=1.9026 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,528B, BPFP=1.6611 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,096B, BPFP=1.8317 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,096B, BPFP=0.1329 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.977s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12932784 58.54678110 + layer.0.v_cache 0.00001491 0.01613193 + layer.1.k_cache 0.01488123 7.32953116 + layer.1.v_cache 0.00000563 0.00568697 + layer.2.k_cache 0.00939683 1.04632715 + layer.2.v_cache 0.00001981 0.01627642 + layer.3.k_cache 0.07051506 4.42489155 + layer.3.v_cache 0.00001927 0.02134484 + layer.4.k_cache 0.00064413 0.44124552 + layer.4.v_cache 0.00005939 0.03882766 + layer.4.output 0.26127321 1039.83945742 + ------------------------------------------------------------------------------------- + TOTAL 0.12081156 432.39783801 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 49156 +BPFP 0.8688 bits/point +EBPFP 1.7377 equivalent bits/point +MSE 432.397838 +---------------------- -------------------------------------------------------- +Time: 2.471s Load: 0.002s, Pack+Encode: 1.492s, Decode+Unpack: 0.977s +---------------------- -------------------------------------------------------- +💾 Converting with 432.3978 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,432B, BPFP=0.4303 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,628B, BPFP=1.9916 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,604B, BPFP=0.7825 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,404B, BPFP=1.9243 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,036B, BPFP=0.9123 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,240B, BPFP=1.8750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,180B, BPFP=0.6550 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,488B, BPFP=1.9495 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,424B, BPFP=1.6298 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,120B, BPFP=1.8389 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,048B, BPFP=0.1308 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.975s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13190371 60.34262789 + layer.0.v_cache 0.00001552 0.01672312 + layer.1.k_cache 0.01668521 7.54391949 + layer.1.v_cache 0.00000581 0.00604777 + layer.2.k_cache 0.00490881 1.24276220 + layer.2.v_cache 0.00001873 0.01664554 + layer.3.k_cache 0.11824917 4.94556838 + layer.3.v_cache 0.00002115 0.02197078 + layer.4.k_cache 0.00061206 0.41191776 + layer.4.v_cache 0.00004844 0.03682326 + layer.4.output 0.26122982 1041.35044643 + ------------------------------------------------------------------------------------- + TOTAL 0.12359279 433.17871360 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 49604 +BPFP 0.8768 bits/point +EBPFP 1.7535 equivalent bits/point +MSE 433.178714 +---------------------- -------------------------------------------------------- +Time: 2.458s Load: 0.002s, Pack+Encode: 1.480s, Decode+Unpack: 0.975s +---------------------- -------------------------------------------------------- +💾 Converting with 433.1787 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,924B, BPFP=0.3268 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,180B, BPFP=2.0686 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,220B, BPFP=0.5469 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,472B, BPFP=1.9484 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,604B, BPFP=0.7819 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,944B, BPFP=1.8587 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,796B, BPFP=0.8145 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,432B, BPFP=1.9416 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,440B, BPFP=1.4334 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,904B, BPFP=1.8519 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,892B, BPFP=0.1187 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10326390 56.21358590 + layer.0.v_cache 0.00001541 0.01342324 + layer.1.k_cache 0.04976616 6.01305489 + layer.1.v_cache 0.00000534 0.00484103 + layer.2.k_cache 0.01168035 1.02824659 + layer.2.v_cache 0.00001798 0.01447618 + layer.3.k_cache 0.02907468 5.47551495 + layer.3.v_cache 0.00001946 0.01914082 + layer.4.k_cache 0.00073861 0.38218125 + layer.4.v_cache 0.00005077 0.03593661 + layer.4.output 0.14788401 588.04309006 + ------------------------------------------------------------------------------------- + TOTAL 0.07234240 246.20600188 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 84808 +BPFP 0.8473 bits/point +EBPFP 1.6945 equivalent bits/point +MSE 246.206002 +---------------------- -------------------------------------------------------- +Time: 2.513s Load: 0.004s, Pack+Encode: 1.493s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 246.2060 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 64, 128) +Output shape: (1, 64, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.output: torch.Size([1, 64, 3584]) -> torch.Size([1, 1, 64, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,296B, BPFP=0.3164 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,024B, BPFP=1.4707 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,344B, BPFP=0.5723 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,676B, BPFP=1.3857 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,732B, BPFP=0.6670 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,440B, BPFP=1.3281 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,412B, BPFP=0.5889 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,616B, BPFP=1.3711 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,452B, BPFP=1.0869 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,444B, BPFP=1.3291 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,768B, BPFP=0.0965 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.966s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09722494 72.30528259 + layer.0.v_cache 0.00001391 0.01077929 + layer.1.k_cache 0.01806639 9.12917233 + layer.1.v_cache 0.00000557 0.00434836 + layer.2.k_cache 0.00423524 1.40542567 + layer.2.v_cache 0.00001868 0.01395631 + layer.3.k_cache 0.03759564 8.20583439 + layer.3.v_cache 0.00001825 0.01598062 + layer.4.k_cache 0.00061753 0.39309105 + layer.4.v_cache 0.00004894 0.02923811 + layer.4.output 0.21408452 843.14320592 + ------------------------------------------------------------------------------------- + TOTAL 0.09743746 352.55973824 + (elements=557,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 557056 +Total Bytes 44204 +BPFP 0.6348 bits/point +EBPFP 1.2696 equivalent bits/point +MSE 352.559738 +---------------------- -------------------------------------------------------- +Time: 2.465s Load: 0.003s, Pack+Encode: 1.496s, Decode+Unpack: 0.966s +---------------------- -------------------------------------------------------- +💾 Converting with 352.5597 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,180B, BPFP=0.3339 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,128B, BPFP=1.8578 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,572B, BPFP=0.5472 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,576B, BPFP=1.7733 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,056B, BPFP=0.7745 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,452B, BPFP=1.7543 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,160B, BPFP=0.7904 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,764B, BPFP=1.8021 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,136B, BPFP=1.5527 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,160B, BPFP=1.7096 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,808B, BPFP=0.1052 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.015s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11648470 56.87307560 + layer.0.v_cache 0.00001454 0.01310674 + layer.1.k_cache 0.02721847 7.25604368 + layer.1.v_cache 0.00000531 0.00476064 + layer.2.k_cache 0.01045542 0.92975377 + layer.2.v_cache 0.00001699 0.01306544 + layer.3.k_cache 0.01744144 6.93531470 + layer.3.v_cache 0.00001839 0.01703579 + layer.4.k_cache 0.00076342 0.39874163 + layer.4.v_cache 0.00004795 0.03477776 + layer.4.output 11.17676328 525.77586660 + ------------------------------------------------------------------------------------- + TOTAL 4.61234174 220.75922011 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 88992 +BPFP 0.8019 bits/point +EBPFP 1.6038 equivalent bits/point +MSE 220.759220 +---------------------- -------------------------------------------------------- +Time: 2.505s Load: 0.005s, Pack+Encode: 1.485s, Decode+Unpack: 1.015s +---------------------- -------------------------------------------------------- +💾 Converting with 220.7592 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,880B, BPFP=0.3672 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,968B, BPFP=2.1422 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,956B, BPFP=0.5773 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,208B, BPFP=2.1891 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,324B, BPFP=0.6492 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,608B, BPFP=2.0719 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,256B, BPFP=0.6359 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,336B, BPFP=2.2141 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,804B, BPFP=1.5242 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,752B, BPFP=1.9047 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,124B, BPFP=0.1151 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11433966 53.12902832 + layer.0.v_cache 0.00001437 0.01292745 + layer.1.k_cache 0.05315235 6.01879196 + layer.1.v_cache 0.00000524 0.00448005 + layer.2.k_cache 0.00678293 1.05004549 + layer.2.v_cache 0.00001704 0.01317996 + layer.3.k_cache 0.01951189 3.98854828 + layer.3.v_cache 0.00001758 0.01635268 + layer.4.k_cache 0.00076624 0.36633983 + layer.4.v_cache 0.00004856 0.03379250 + layer.4.output 0.16992235 676.45619420 + ------------------------------------------------------------------------------------- + TOTAL 0.08141837 282.34275564 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 77216 +BPFP 0.8871 bits/point +EBPFP 1.7743 equivalent bits/point +MSE 282.342756 +---------------------- -------------------------------------------------------- +Time: 2.490s Load: 0.003s, Pack+Encode: 1.476s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 282.3428 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,896B, BPFP=0.3292 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,112B, BPFP=2.1028 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,132B, BPFP=0.5437 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,612B, BPFP=2.0160 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,804B, BPFP=0.6604 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,796B, BPFP=1.8743 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,412B, BPFP=0.5924 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,616B, BPFP=2.0167 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,960B, BPFP=1.5556 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,916B, BPFP=1.8951 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,836B, BPFP=0.1199 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11570581 56.08879666 + layer.0.v_cache 0.00001517 0.01391309 + layer.1.k_cache 0.03236232 5.66031223 + layer.1.v_cache 0.00000662 0.00484210 + layer.2.k_cache 0.00641356 1.03362054 + layer.2.v_cache 0.00001676 0.01390840 + layer.3.k_cache 0.02648412 4.80281169 + layer.3.v_cache 0.00001896 0.01819608 + layer.4.k_cache 0.00074524 0.39378259 + layer.4.v_cache 0.00004843 0.03460129 + layer.4.output 0.15113260 600.41542659 + ------------------------------------------------------------------------------------- + TOTAL 0.07292619 251.23369240 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 83092 +BPFP 0.8486 bits/point +EBPFP 1.6971 equivalent bits/point +MSE 251.233692 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.004s, Pack+Encode: 1.483s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 251.2337 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 114, 128) +Output shape: (1, 114, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.output: torch.Size([1, 114, 3584]) -> torch.Size([1, 1, 114, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,380B, BPFP=0.3262 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,172B, BPFP=1.6683 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,920B, BPFP=0.5373 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,520B, BPFP=1.5789 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,524B, BPFP=0.6201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,496B, BPFP=1.5757 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,712B, BPFP=0.6458 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,620B, BPFP=1.5927 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,380B, BPFP=1.4227 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,180B, BPFP=1.5323 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,072B, BPFP=0.1189 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10081583 50.55598102 + layer.0.v_cache 0.00001633 0.01321122 + layer.1.k_cache 0.08473034 6.08579616 + layer.1.v_cache 0.00000563 0.00494119 + layer.2.k_cache 0.00792928 0.98007570 + layer.2.v_cache 0.00001730 0.01321596 + layer.3.k_cache 0.01019665 5.55156721 + layer.3.v_cache 0.00001810 0.01546742 + layer.4.k_cache 0.00077074 0.40858941 + layer.4.v_cache 0.00005022 0.03301662 + layer.4.output 10.00037814 470.52279135 + ------------------------------------------------------------------------------------- + TOTAL 4.12983514 197.48949420 + (elements=992,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 992256 +Total Bytes 89976 +BPFP 0.7254 bits/point +EBPFP 1.4509 equivalent bits/point +MSE 197.489494 +---------------------- -------------------------------------------------------- +Time: 2.509s Load: 0.004s, Pack+Encode: 1.491s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 197.4895 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 98, 128) +Output shape: (1, 98, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.output: torch.Size([1, 98, 3584]) -> torch.Size([1, 1, 98, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,964B, BPFP=0.3131 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,032B, BPFP=1.9184 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,280B, BPFP=0.5230 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,548B, BPFP=1.8412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,088B, BPFP=0.8112 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,504B, BPFP=1.8342 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,452B, BPFP=0.7098 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,440B, BPFP=1.8240 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,424B, BPFP=1.5026 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,084B, BPFP=1.7672 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,136B, BPFP=0.0942 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11008905 57.89211974 + layer.0.v_cache 0.00001718 0.01305553 + layer.1.k_cache 0.06240322 6.16303300 + layer.1.v_cache 0.00000548 0.00475214 + layer.2.k_cache 0.00473963 1.17554645 + layer.2.v_cache 0.00001718 0.01369311 + layer.3.k_cache 0.01141277 6.25644295 + layer.3.v_cache 0.00001769 0.01613358 + layer.4.k_cache 0.00076610 0.40402447 + layer.4.v_cache 0.00004814 0.03361589 + layer.4.output 0.02426045 567.33345481 + ------------------------------------------------------------------------------------- + TOTAL 0.02113762 237.84156474 + (elements=852,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 852992 +Total Bytes 85952 +BPFP 0.8061 bits/point +EBPFP 1.6122 equivalent bits/point +MSE 237.841565 +---------------------- -------------------------------------------------------- +Time: 2.512s Load: 0.005s, Pack+Encode: 1.495s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 237.8416 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,908B, BPFP=0.3240 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,104B, BPFP=2.0557 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,296B, BPFP=0.5598 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,488B, BPFP=1.9511 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,964B, BPFP=0.8431 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,340B, BPFP=1.9260 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,120B, BPFP=1.3791 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,812B, BPFP=2.0061 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,036B, BPFP=1.5346 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,080B, BPFP=1.8818 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,900B, BPFP=0.1189 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08919133 53.98899244 + layer.0.v_cache 0.00001414 0.01303210 + layer.1.k_cache 0.08591949 6.30080911 + layer.1.v_cache 0.00000528 0.00490716 + layer.2.k_cache 0.00389855 1.06442841 + layer.2.v_cache 0.00001738 0.01395664 + layer.3.k_cache 0.07679920 7.64749809 + layer.3.v_cache 0.00002028 0.01926329 + layer.4.k_cache 0.00072407 0.36606946 + layer.4.v_cache 0.00004779 0.03467163 + layer.4.output 0.14786680 587.15576475 + ------------------------------------------------------------------------------------- + TOTAL 0.07598265 245.85552833 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 90048 +BPFP 0.8996 bits/point +EBPFP 1.7992 equivalent bits/point +MSE 245.855528 +---------------------- -------------------------------------------------------- +Time: 2.510s Load: 0.005s, Pack+Encode: 1.485s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 245.8555 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,268B, BPFP=0.3312 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,288B, BPFP=1.7944 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,804B, BPFP=0.5555 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,728B, BPFP=1.7126 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,744B, BPFP=0.6928 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,612B, BPFP=1.6957 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,216B, BPFP=1.0537 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,836B, BPFP=1.7284 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,720B, BPFP=1.4194 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,328B, BPFP=1.6542 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,136B, BPFP=0.1071 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11353527 54.54955827 + layer.0.v_cache 0.00001674 0.01285768 + layer.1.k_cache 0.04276301 6.22068829 + layer.1.v_cache 0.00000542 0.00477585 + layer.2.k_cache 0.00247095 1.03331621 + layer.2.v_cache 0.00001885 0.01470312 + layer.3.k_cache 0.02722364 8.59813997 + layer.3.v_cache 0.00001942 0.01780945 + layer.4.k_cache 0.00068256 0.41130419 + layer.4.v_cache 0.00004843 0.03440973 + layer.4.output 10.65454330 500.64360814 + ------------------------------------------------------------------------------------- + TOTAL 4.39815220 210.31781293 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 91680 +BPFP 0.7875 bits/point +EBPFP 1.5750 equivalent bits/point +MSE 210.317813 +---------------------- -------------------------------------------------------- +Time: 2.509s Load: 0.004s, Pack+Encode: 1.489s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 210.3178 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3452 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,012B, BPFP=2.2344 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,980B, BPFP=0.5543 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,152B, BPFP=2.0744 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,884B, BPFP=0.9085 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,004B, BPFP=1.8609 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,160B, BPFP=0.5878 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,076B, BPFP=2.0603 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,636B, BPFP=1.6064 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,124B, BPFP=1.8832 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,264B, BPFP=0.1133 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12550225 52.37761579 + layer.0.v_cache 0.00001468 0.01295666 + layer.1.k_cache 0.05234316 5.24247596 + layer.1.v_cache 0.00000508 0.00448695 + layer.2.k_cache 0.00733058 0.97976512 + layer.2.v_cache 0.00001689 0.01317193 + layer.3.k_cache 0.01592888 4.55232021 + layer.3.v_cache 0.00001847 0.01712292 + layer.4.k_cache 0.00069460 0.38217231 + layer.4.v_cache 0.00005205 0.03362525 + layer.4.output 0.16184913 644.74627976 + ------------------------------------------------------------------------------------- + TOTAL 0.07852062 269.22586303 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 80148 +BPFP 0.8770 bits/point +EBPFP 1.7539 equivalent bits/point +MSE 269.225863 +---------------------- -------------------------------------------------------- +Time: 2.492s Load: 0.003s, Pack+Encode: 1.480s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 269.2259 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,832B, BPFP=0.3670 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,936B, BPFP=2.3910 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,968B, BPFP=0.5946 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,444B, BPFP=2.0921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,144B, BPFP=0.6298 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,852B, BPFP=2.1739 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,884B, BPFP=0.5777 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,816B, BPFP=2.1667 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,808B, BPFP=1.5641 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,756B, BPFP=1.9543 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,944B, BPFP=0.1129 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11879376 56.07322967 + layer.0.v_cache 0.00001438 0.01413360 + layer.1.k_cache 0.05518859 5.69427255 + layer.1.v_cache 0.00000514 0.00502355 + layer.2.k_cache 0.00243519 0.97074391 + layer.2.v_cache 0.00001742 0.01427947 + layer.3.k_cache 0.07959401 4.47953131 + layer.3.v_cache 0.00001812 0.01771623 + layer.4.k_cache 0.00069358 0.40538920 + layer.4.v_cache 0.00004620 0.03546389 + layer.4.output 0.17426339 693.68984661 + ------------------------------------------------------------------------------------- + TOTAL 0.08686177 289.61992410 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 76384 +BPFP 0.9001 bits/point +EBPFP 1.8002 equivalent bits/point +MSE 289.619924 +---------------------- -------------------------------------------------------- +Time: 2.498s Load: 0.003s, Pack+Encode: 1.482s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 289.6199 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,832B, BPFP=0.3491 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,804B, BPFP=2.2492 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,012B, BPFP=0.5739 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,440B, BPFP=2.1799 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,464B, BPFP=0.6601 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,184B, BPFP=2.1311 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,164B, BPFP=0.6029 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,448B, BPFP=2.1814 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,328B, BPFP=1.3963 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,184B, BPFP=1.9405 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,788B, BPFP=0.1031 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11325797 52.25893197 + layer.0.v_cache 0.00001565 0.01244574 + layer.1.k_cache 0.07651102 5.78121055 + layer.1.v_cache 0.00000530 0.00498545 + layer.2.k_cache 0.00872486 1.17214603 + layer.2.v_cache 0.00001696 0.01368238 + layer.3.k_cache 0.12876377 4.95185926 + layer.3.v_cache 0.00001905 0.01680107 + layer.4.k_cache 0.00069190 0.37000158 + layer.4.v_cache 0.00004968 0.03295427 + layer.4.output 0.16578155 659.87793990 + ------------------------------------------------------------------------------------- + TOTAL 0.08756041 275.51532927 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 78648 +BPFP 0.8815 bits/point +EBPFP 1.7631 equivalent bits/point +MSE 275.515329 +---------------------- -------------------------------------------------------- +Time: 2.501s Load: 0.004s, Pack+Encode: 1.485s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 275.5153 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,796B, BPFP=0.3644 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,820B, BPFP=2.3985 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,940B, BPFP=0.5966 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,372B, BPFP=2.3076 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,236B, BPFP=0.6567 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,872B, BPFP=2.2062 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,800B, BPFP=0.5682 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,308B, BPFP=2.0917 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,300B, BPFP=1.4813 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,740B, BPFP=1.7735 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,600B, BPFP=0.1044 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12835005 52.84053812 + layer.0.v_cache 0.00001616 0.01334281 + layer.1.k_cache 0.05672838 5.31701739 + layer.1.v_cache 0.00000535 0.00518346 + layer.2.k_cache 0.00240846 0.97582176 + layer.2.v_cache 0.00001644 0.01441391 + layer.3.k_cache 0.04570918 4.48338140 + layer.3.v_cache 0.00001733 0.01743440 + layer.4.k_cache 0.00072090 0.38901044 + layer.4.v_cache 0.00004908 0.03579744 + layer.4.output 0.18444509 700.88589981 + ------------------------------------------------------------------------------------- + TOTAL 0.08971394 292.37019058 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 74784 +BPFP 0.8927 bits/point +EBPFP 1.7853 equivalent bits/point +MSE 292.370191 +---------------------- -------------------------------------------------------- +Time: 2.493s Load: 0.003s, Pack+Encode: 1.480s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 292.3702 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,736B, BPFP=0.3820 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,448B, BPFP=2.5194 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,680B, BPFP=0.5898 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,024B, BPFP=2.4261 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,848B, BPFP=0.6268 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,688B, BPFP=2.3521 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,468B, BPFP=0.5431 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,204B, BPFP=2.4657 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,972B, BPFP=1.3143 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,328B, BPFP=2.2729 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,216B, BPFP=0.1011 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10085608 54.41108467 + layer.0.v_cache 0.00001554 0.01314453 + layer.1.k_cache 0.06589249 6.05627441 + layer.1.v_cache 0.00000499 0.00489046 + layer.2.k_cache 0.00422055 0.86813000 + layer.2.v_cache 0.00001579 0.01362622 + layer.3.k_cache 0.05337993 4.38841570 + layer.3.v_cache 0.00001747 0.01718673 + layer.4.k_cache 0.00075412 0.36304565 + layer.4.v_cache 0.00004712 0.03554282 + layer.4.output 0.19135183 763.31482646 + ------------------------------------------------------------------------------------- + TOTAL 0.09203923 318.19853685 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 73612 +BPFP 0.9529 bits/point +EBPFP 1.9059 equivalent bits/point +MSE 318.198537 +---------------------- -------------------------------------------------------- +Time: 2.493s Load: 0.003s, Pack+Encode: 1.480s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 318.1985 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,888B, BPFP=0.3512 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,880B, BPFP=2.2098 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,032B, BPFP=0.5640 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,184B, BPFP=2.0804 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,336B, BPFP=0.6205 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,964B, BPFP=2.0394 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,156B, BPFP=0.5871 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,384B, BPFP=2.1176 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,532B, BPFP=1.4010 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,308B, BPFP=1.9174 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,496B, BPFP=0.0929 +⌛️ [2/4] FRONTEND: Frontend time: 1.479s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09057366 52.14525205 + layer.0.v_cache 0.00001789 0.01226994 + layer.1.k_cache 0.03614600 6.36833627 + layer.1.v_cache 0.00000481 0.00450009 + layer.2.k_cache 0.00554357 0.93526495 + layer.2.v_cache 0.00001714 0.01335279 + layer.3.k_cache 0.02742965 4.64866130 + layer.3.v_cache 0.00001729 0.01690622 + layer.4.k_cache 0.00093027 0.36327516 + layer.4.v_cache 0.00004906 0.03401779 + layer.4.output 0.16441821 641.83960459 + ------------------------------------------------------------------------------------- + TOTAL 0.07715628 268.08347463 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 78160 +BPFP 0.8552 bits/point +EBPFP 1.7104 equivalent bits/point +MSE 268.083475 +---------------------- -------------------------------------------------------- +Time: 2.494s Load: 0.005s, Pack+Encode: 1.479s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 268.0835 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,804B, BPFP=0.3661 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,836B, BPFP=2.4018 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,932B, BPFP=0.5950 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,240B, BPFP=2.0779 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,140B, BPFP=0.6372 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,576B, BPFP=1.9432 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,952B, BPFP=0.5990 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,796B, BPFP=1.9878 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,520B, BPFP=1.3231 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,124B, BPFP=1.8515 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,852B, BPFP=0.1117 +⌛️ [2/4] FRONTEND: Frontend time: 1.477s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09325875 53.12089717 + layer.0.v_cache 0.00001406 0.01321374 + layer.1.k_cache 0.05775181 6.38683923 + layer.1.v_cache 0.00000524 0.00488002 + layer.2.k_cache 0.00413042 1.05262875 + layer.2.v_cache 0.00001743 0.01392991 + layer.3.k_cache 0.04387648 4.88557652 + layer.3.v_cache 0.00002497 0.01880460 + layer.4.k_cache 0.00071512 0.37433416 + layer.4.v_cache 0.00005147 0.03631638 + layer.4.output 0.17652170 702.58290816 + ------------------------------------------------------------------------------------- + TOTAL 0.08444104 293.17575163 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 71772 +BPFP 0.8567 bits/point +EBPFP 1.7134 equivalent bits/point +MSE 293.175752 +---------------------- -------------------------------------------------------- +Time: 2.490s Load: 0.004s, Pack+Encode: 1.477s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 293.1758 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,912B, BPFP=0.3395 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,280B, BPFP=2.0028 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,044B, BPFP=0.5405 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,080B, BPFP=1.9673 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,436B, BPFP=0.6101 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,892B, BPFP=1.9339 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,992B, BPFP=0.5312 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,356B, BPFP=2.0163 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,300B, BPFP=1.2962 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,036B, BPFP=1.7820 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,508B, BPFP=0.0890 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212358 54.28454590 + layer.0.v_cache 0.00001682 0.01131052 + layer.1.k_cache 0.05918076 5.35475193 + layer.1.v_cache 0.00000473 0.00405320 + layer.2.k_cache 0.00260097 1.03851769 + layer.2.v_cache 0.00001544 0.01223719 + layer.3.k_cache 0.02838851 4.54079090 + layer.3.v_cache 0.00001657 0.01540894 + layer.4.k_cache 0.00106615 0.33557703 + layer.4.v_cache 0.00004456 0.03197626 + layer.4.output 0.17173813 615.24939123 + ------------------------------------------------------------------------------------- + TOTAL 0.08091912 257.19852401 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 76836 +BPFP 0.8025 bits/point +EBPFP 1.6050 equivalent bits/point +MSE 257.198524 +---------------------- -------------------------------------------------------- +Time: 2.495s Load: 0.005s, Pack+Encode: 1.481s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 257.1985 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,904B, BPFP=0.3381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,096B, BPFP=2.1477 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,132B, BPFP=0.5561 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,632B, BPFP=2.0653 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,716B, BPFP=0.6598 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,076B, BPFP=1.9666 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,608B, BPFP=0.6406 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,748B, BPFP=2.0859 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,156B, BPFP=1.4482 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,248B, BPFP=1.9972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,960B, BPFP=0.1004 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10731998 56.32589444 + layer.0.v_cache 0.00001618 0.01285056 + layer.1.k_cache 0.07375650 6.13148915 + layer.1.v_cache 0.00000540 0.00489955 + layer.2.k_cache 0.00246326 0.98785695 + layer.2.v_cache 0.00001746 0.01321789 + layer.3.k_cache 0.01642703 5.09113242 + layer.3.v_cache 0.00001778 0.01590170 + layer.4.k_cache 0.00066607 0.36558281 + layer.4.v_cache 0.00004728 0.03299689 + layer.4.output 0.15452130 615.55352070 + ------------------------------------------------------------------------------------- + TOTAL 0.07543447 257.52096866 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 82276 +BPFP 0.8593 bits/point +EBPFP 1.7187 equivalent bits/point +MSE 257.520969 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.004s, Pack+Encode: 1.482s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 257.5210 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 99, 128) +Output shape: (1, 99, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.output: torch.Size([1, 99, 3584]) -> torch.Size([1, 1, 99, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,036B, BPFP=0.3213 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,896B, BPFP=1.8775 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,348B, BPFP=0.5284 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,352B, BPFP=1.7917 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,560B, BPFP=0.7197 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,636B, BPFP=1.8365 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,216B, BPFP=0.6654 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,604B, BPFP=1.8314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,420B, BPFP=1.4867 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,032B, BPFP=1.7412 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,660B, BPFP=0.1051 +⌛️ [2/4] FRONTEND: Frontend time: 1.487s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09061131 59.46354167 + layer.0.v_cache 0.00001461 0.01277636 + layer.1.k_cache 0.05062923 5.62874596 + layer.1.v_cache 0.00000709 0.00520519 + layer.2.k_cache 0.00392854 1.07928960 + layer.2.v_cache 0.00001793 0.01389144 + layer.3.k_cache 0.05082532 6.40104352 + layer.3.v_cache 0.00001820 0.01710298 + layer.4.k_cache 0.00080398 0.41897949 + layer.4.v_cache 0.00005053 0.03546535 + layer.4.output 0.02403493 561.79247835 + ------------------------------------------------------------------------------------- + TOTAL 0.02147948 235.62490530 + (elements=861,696) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 861696 +Total Bytes 85760 +BPFP 0.7962 bits/point +EBPFP 1.5924 equivalent bits/point +MSE 235.624905 +---------------------- -------------------------------------------------------- +Time: 2.505s Load: 0.004s, Pack+Encode: 1.487s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 235.6249 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,712B, BPFP=0.3768 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,932B, BPFP=2.6259 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,672B, BPFP=0.5880 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,324B, BPFP=2.4921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,012B, BPFP=0.6629 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,868B, BPFP=2.1717 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,708B, BPFP=0.5960 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,120B, BPFP=2.4472 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,812B, BPFP=1.2790 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,204B, BPFP=2.0255 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,132B, BPFP=0.1299 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11136492 55.54265240 + layer.0.v_cache 0.00001955 0.01359362 + layer.1.k_cache 0.06280328 6.14530536 + layer.1.v_cache 0.00000544 0.00513068 + layer.2.k_cache 0.00235174 0.92279354 + layer.2.v_cache 0.00001859 0.01487654 + layer.3.k_cache 0.03326307 4.71696537 + layer.3.v_cache 0.00001926 0.01859713 + layer.4.k_cache 0.00064473 0.37654514 + layer.4.v_cache 0.00005254 0.03501857 + layer.4.output 0.19494371 762.91907696 + ------------------------------------------------------------------------------------- + TOTAL 0.09265583 318.13088336 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 73496 +BPFP 0.9514 bits/point +EBPFP 1.9029 equivalent bits/point +MSE 318.130883 +---------------------- -------------------------------------------------------- +Time: 2.496s Load: 0.004s, Pack+Encode: 1.481s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 318.1309 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,748B, BPFP=0.3691 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,532B, BPFP=2.4350 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,792B, BPFP=0.5895 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,484B, BPFP=2.4248 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,040B, BPFP=0.6419 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,808B, BPFP=2.2821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,692B, BPFP=0.5684 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,408B, BPFP=2.4088 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,856B, BPFP=1.6588 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,928B, BPFP=2.3074 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,592B, BPFP=0.1083 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08829598 53.83311339 + layer.0.v_cache 0.00001455 0.01352603 + layer.1.k_cache 0.05950350 6.16699549 + layer.1.v_cache 0.00000529 0.00491180 + layer.2.k_cache 0.00411868 1.29056477 + layer.2.v_cache 0.00001680 0.01425785 + layer.3.k_cache 0.04877868 4.67457127 + layer.3.v_cache 0.00001896 0.01908435 + layer.4.k_cache 0.00063741 0.41547257 + layer.4.v_cache 0.00004764 0.03631844 + layer.4.output 0.18365649 730.65884411 + ------------------------------------------------------------------------------------- + TOTAL 0.08747253 304.76945440 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 77880 +BPFP 0.9673 bits/point +EBPFP 1.9346 equivalent bits/point +MSE 304.769454 +---------------------- -------------------------------------------------------- +Time: 2.495s Load: 0.003s, Pack+Encode: 1.478s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 304.7695 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,788B, BPFP=0.3628 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,152B, BPFP=2.2630 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,924B, BPFP=0.5933 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,084B, BPFP=2.2492 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,204B, BPFP=0.6502 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,116B, BPFP=2.0528 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,884B, BPFP=0.5852 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,440B, BPFP=2.3214 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,020B, BPFP=1.6274 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,436B, BPFP=2.1177 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,548B, BPFP=0.1029 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11195674 50.41737267 + layer.0.v_cache 0.00001346 0.01219792 + layer.1.k_cache 0.03755230 6.12282572 + layer.1.v_cache 0.00000535 0.00486527 + layer.2.k_cache 0.00239600 1.23121167 + layer.2.v_cache 0.00001753 0.01466686 + layer.3.k_cache 0.02867049 4.86424672 + layer.3.v_cache 0.00001751 0.01691668 + layer.4.k_cache 0.00071811 0.36205884 + layer.4.v_cache 0.00005486 0.03578202 + layer.4.output 0.19495771 701.06760204 + ------------------------------------------------------------------------------------- + TOTAL 0.09094743 292.38560933 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 76596 +BPFP 0.9143 bits/point +EBPFP 1.8286 equivalent bits/point +MSE 292.385609 +---------------------- -------------------------------------------------------- +Time: 2.496s Load: 0.004s, Pack+Encode: 1.480s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 292.3856 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,908B, BPFP=0.3312 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,836B, BPFP=2.0549 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,112B, BPFP=0.5403 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,268B, BPFP=1.9563 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,448B, BPFP=0.5986 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,196B, BPFP=1.9438 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,132B, BPFP=0.5437 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,752B, BPFP=2.0403 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,452B, BPFP=1.4674 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,612B, BPFP=1.8424 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,772B, BPFP=0.0936 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08657128 54.93554145 + layer.0.v_cache 0.00001574 0.01214557 + layer.1.k_cache 0.05618976 5.53236016 + layer.1.v_cache 0.00000506 0.00452797 + layer.2.k_cache 0.00392605 1.00446370 + layer.2.v_cache 0.00001570 0.01263221 + layer.3.k_cache 0.02849737 5.04796278 + layer.3.v_cache 0.00001769 0.01768509 + layer.4.k_cache 0.00099738 0.38868001 + layer.4.v_cache 0.00004751 0.03459602 + layer.4.output 0.15107292 601.53010913 + ------------------------------------------------------------------------------------- + TOTAL 0.07257612 251.62949170 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 80488 +BPFP 0.8220 bits/point +EBPFP 1.6440 equivalent bits/point +MSE 251.629492 +---------------------- -------------------------------------------------------- +Time: 2.500s Load: 0.004s, Pack+Encode: 1.484s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 251.6295 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,884B, BPFP=0.3680 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,732B, BPFP=2.2914 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,948B, BPFP=0.5758 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,472B, BPFP=2.0453 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,264B, BPFP=0.6375 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,276B, BPFP=2.2023 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,820B, BPFP=0.5508 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,624B, BPFP=2.2703 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,548B, BPFP=1.4742 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,988B, BPFP=1.9508 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,632B, BPFP=0.1013 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09254595 51.51166992 + layer.0.v_cache 0.00001776 0.01199843 + layer.1.k_cache 0.05775445 5.78045807 + layer.1.v_cache 0.00000504 0.00466186 + layer.2.k_cache 0.00891839 1.04105377 + layer.2.v_cache 0.00001732 0.01364775 + layer.3.k_cache 0.02836140 4.06871109 + layer.3.v_cache 0.00001672 0.01634093 + layer.4.k_cache 0.00083566 0.35726752 + layer.4.v_cache 0.00004876 0.03447590 + layer.4.output 0.16993865 676.52466518 + ------------------------------------------------------------------------------------- + TOTAL 0.08106423 282.26546715 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 77188 +BPFP 0.8868 bits/point +EBPFP 1.7736 equivalent bits/point +MSE 282.265467 +---------------------- -------------------------------------------------------- +Time: 2.491s Load: 0.003s, Pack+Encode: 1.478s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 282.2655 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,944B, BPFP=0.3131 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,772B, BPFP=1.8963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,248B, BPFP=0.5232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,380B, BPFP=1.8331 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,692B, BPFP=0.7558 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,228B, BPFP=1.8086 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,556B, BPFP=0.5728 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,312B, BPFP=1.8222 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,532B, BPFP=1.3744 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,668B, BPFP=1.7184 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,344B, BPFP=0.1000 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13093194 59.21357523 + layer.0.v_cache 0.00001635 0.01325938 + layer.1.k_cache 0.05235285 6.57165779 + layer.1.v_cache 0.00000496 0.00472277 + layer.2.k_cache 0.00512381 1.08566882 + layer.2.v_cache 0.00001647 0.01303838 + layer.3.k_cache 0.01138228 4.89512210 + layer.3.v_cache 0.00001850 0.01578512 + layer.4.k_cache 0.00081880 0.39045692 + layer.4.v_cache 0.00004796 0.03515671 + layer.4.output 0.02447439 573.36924705 + ------------------------------------------------------------------------------------- + TOTAL 0.02188439 240.34253956 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 82676 +BPFP 0.7834 bits/point +EBPFP 1.5668 equivalent bits/point +MSE 240.342540 +---------------------- -------------------------------------------------------- +Time: 2.496s Load: 0.005s, Pack+Encode: 1.480s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 240.3425 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,832B, BPFP=0.3623 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,852B, BPFP=2.3441 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,952B, BPFP=0.5839 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,416B, BPFP=2.2579 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,220B, BPFP=0.6369 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,188B, BPFP=2.2128 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,772B, BPFP=0.5483 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,324B, BPFP=2.2397 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,972B, BPFP=1.5767 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,804B, BPFP=2.1369 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,324B, BPFP=0.1222 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510638 49.24310535 + layer.0.v_cache 0.00001399 0.01282226 + layer.1.k_cache 0.05467150 6.04442432 + layer.1.v_cache 0.00000511 0.00476759 + layer.2.k_cache 0.00545186 1.07922315 + layer.2.v_cache 0.00001652 0.01436423 + layer.3.k_cache 0.02986256 4.34036139 + layer.3.v_cache 0.00001747 0.01617701 + layer.4.k_cache 0.00068266 0.35992567 + layer.4.v_cache 0.00005277 0.03657686 + layer.4.output 0.17642128 685.56945072 + ------------------------------------------------------------------------------------- + TOTAL 0.08357822 285.89046488 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 79656 +BPFP 0.9267 bits/point +EBPFP 1.8535 equivalent bits/point +MSE 285.890465 +---------------------- -------------------------------------------------------- +Time: 2.499s Load: 0.004s, Pack+Encode: 1.483s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 285.8905 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,804B, BPFP=0.3661 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,860B, BPFP=2.4067 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,904B, BPFP=0.5893 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,040B, BPFP=2.2403 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,232B, BPFP=0.6558 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,420B, BPFP=2.1144 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,772B, BPFP=0.5625 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,716B, BPFP=2.1745 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,988B, BPFP=1.4180 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,784B, BPFP=1.7825 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,808B, BPFP=0.1104 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08730992 51.80945934 + layer.0.v_cache 0.00001543 0.01386157 + layer.1.k_cache 0.05635426 6.02145544 + layer.1.v_cache 0.00000520 0.00514661 + layer.2.k_cache 0.00390536 1.16624926 + layer.2.v_cache 0.00001661 0.01394331 + layer.3.k_cache 0.06483487 4.94832978 + layer.3.v_cache 0.00002010 0.01771505 + layer.4.k_cache 0.00072652 0.36918677 + layer.4.v_cache 0.00004842 0.03230568 + layer.4.output 0.18845482 701.54522263 + ------------------------------------------------------------------------------------- + TOTAL 0.09014238 292.65965949 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 74328 +BPFP 0.8872 bits/point +EBPFP 1.7744 equivalent bits/point +MSE 292.659659 +---------------------- -------------------------------------------------------- +Time: 2.490s Load: 0.005s, Pack+Encode: 1.476s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 292.6597 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,604B, BPFP=0.3686 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,284B, BPFP=2.5928 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,532B, BPFP=0.5818 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,196B, BPFP=2.3428 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,608B, BPFP=0.5993 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,428B, BPFP=1.7068 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,324B, BPFP=0.5340 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,736B, BPFP=2.0074 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,840B, BPFP=1.3419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,576B, BPFP=1.9706 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,952B, BPFP=0.1297 +⌛️ [2/4] FRONTEND: Frontend time: 1.479s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10784986 58.19301471 + layer.0.v_cache 0.00001387 0.01341984 + layer.1.k_cache 0.01409798 5.71234400 + layer.1.v_cache 0.00000489 0.00465540 + layer.2.k_cache 0.00254294 1.03484558 + layer.2.v_cache 0.00001653 0.01396808 + layer.3.k_cache 0.05709259 4.90478201 + layer.3.v_cache 0.00001736 0.01651502 + layer.4.k_cache 0.00062715 0.34580550 + layer.4.v_cache 0.00004758 0.03292159 + layer.4.output 0.22376562 796.84598214 + ------------------------------------------------------------------------------------- + TOTAL 0.10286295 332.24671451 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 65080 +BPFP 0.8796 bits/point +EBPFP 1.7593 equivalent bits/point +MSE 332.246715 +---------------------- -------------------------------------------------------- +Time: 2.494s Load: 0.004s, Pack+Encode: 1.479s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 332.2467 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.3655 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,656B, BPFP=2.3054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,920B, BPFP=0.5775 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,060B, BPFP=1.7919 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,168B, BPFP=0.6266 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,400B, BPFP=2.0570 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,004B, BPFP=0.5941 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,384B, BPFP=2.2516 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,668B, BPFP=1.3188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,164B, BPFP=1.6147 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,280B, BPFP=0.1209 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14563359 54.30012609 + layer.0.v_cache 0.00001368 0.01225683 + layer.1.k_cache 0.05947259 5.91318174 + layer.1.v_cache 0.00000496 0.00441667 + layer.2.k_cache 0.00553216 0.93203909 + layer.2.v_cache 0.00001557 0.01261305 + layer.3.k_cache 0.04377548 4.36525910 + layer.3.v_cache 0.00001730 0.01659270 + layer.4.k_cache 0.00079035 0.33392986 + layer.4.v_cache 0.00005153 0.03421577 + layer.4.output 0.17347779 684.98598553 + ------------------------------------------------------------------------------------- + TOTAL 0.08645010 285.93097233 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 72552 +BPFP 0.8441 bits/point +EBPFP 1.6882 equivalent bits/point +MSE 285.930972 +---------------------- -------------------------------------------------------- +Time: 2.491s Load: 0.004s, Pack+Encode: 1.478s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 285.9310 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,480B, BPFP=0.3504 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,368B, BPFP=2.6913 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,360B, BPFP=0.5587 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,536B, BPFP=2.2576 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,632B, BPFP=0.6231 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,764B, BPFP=2.0748 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,268B, BPFP=0.5369 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,544B, BPFP=2.2595 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,668B, BPFP=1.1051 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,116B, BPFP=1.6847 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,496B, BPFP=0.1182 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12373332 65.47049598 + layer.0.v_cache 0.00001436 0.01383460 + layer.1.k_cache 0.01270404 6.72720522 + layer.1.v_cache 0.00000490 0.00482041 + layer.2.k_cache 0.00252672 1.05695389 + layer.2.v_cache 0.00001629 0.01425101 + layer.3.k_cache 0.05185594 4.10069367 + layer.3.v_cache 0.00001849 0.01726924 + layer.4.k_cache 0.00068458 0.33529628 + layer.4.v_cache 0.00004975 0.03310286 + layer.4.output 0.22283003 819.60930736 + ------------------------------------------------------------------------------------- + TOTAL 0.10302462 342.06112204 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 63232 +BPFP 0.8806 bits/point +EBPFP 1.7611 equivalent bits/point +MSE 342.061122 +---------------------- -------------------------------------------------------- +Time: 2.501s Load: 0.004s, Pack+Encode: 1.485s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 342.0611 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,852B, BPFP=0.3445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,440B, BPFP=2.3140 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,052B, BPFP=0.5677 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,776B, BPFP=2.1905 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,732B, BPFP=0.8802 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,912B, BPFP=2.0298 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,296B, BPFP=0.7991 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,936B, BPFP=2.2202 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,444B, BPFP=1.5707 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,792B, BPFP=2.0074 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,048B, BPFP=0.1076 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12002775 53.29117839 + layer.0.v_cache 0.00001466 0.01310321 + layer.1.k_cache 0.05271155 5.48770287 + layer.1.v_cache 0.00000524 0.00480181 + layer.2.k_cache 0.00808975 0.98468890 + layer.2.v_cache 0.00001826 0.01425269 + layer.3.k_cache 0.07501186 6.02788289 + layer.3.v_cache 0.00001895 0.01870326 + layer.4.k_cache 0.00071216 0.39920925 + layer.4.v_cache 0.00005170 0.03470155 + layer.4.output 0.17053346 644.43516156 + ------------------------------------------------------------------------------------- + TOTAL 0.08531742 269.25425622 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 84280 +BPFP 0.9222 bits/point +EBPFP 1.8444 equivalent bits/point +MSE 269.254256 +---------------------- -------------------------------------------------------- +Time: 2.499s Load: 0.004s, Pack+Encode: 1.476s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 269.2543 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 75, 128) +Output shape: (1, 75, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.output: torch.Size([1, 75, 3584]) -> torch.Size([1, 1, 75, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,768B, BPFP=0.3683 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,868B, BPFP=2.4725 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,868B, BPFP=0.5975 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,152B, BPFP=2.3233 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,200B, BPFP=0.6667 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,820B, BPFP=2.2542 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,652B, BPFP=0.5525 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,892B, BPFP=2.2692 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,976B, BPFP=1.6617 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,352B, BPFP=1.9483 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,776B, BPFP=0.1124 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08149677 55.63179687 + layer.0.v_cache 0.00001391 0.01404399 + layer.1.k_cache 0.05841089 6.69578084 + layer.1.v_cache 0.00000541 0.00524306 + layer.2.k_cache 0.00252637 1.02011495 + layer.2.v_cache 0.00001693 0.01438361 + layer.3.k_cache 0.02876982 4.65072591 + layer.3.v_cache 0.00001932 0.01764166 + layer.4.k_cache 0.00075246 0.41349309 + layer.4.v_cache 0.00005041 0.03620049 + layer.4.output 0.19266937 719.61779762 + ------------------------------------------------------------------------------------- + TOTAL 0.08945576 300.34258869 + (elements=652,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 652800 +Total Bytes 76324 +BPFP 0.9353 bits/point +EBPFP 1.8707 equivalent bits/point +MSE 300.342589 +---------------------- -------------------------------------------------------- +Time: 2.496s Load: 0.003s, Pack+Encode: 1.480s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 300.3426 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,612B, BPFP=0.3704 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,360B, BPFP=2.6103 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,528B, BPFP=0.5809 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,052B, BPFP=2.3097 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,832B, BPFP=0.6507 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,000B, BPFP=2.2978 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,372B, BPFP=0.5450 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,640B, BPFP=2.4449 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,764B, BPFP=1.3244 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,172B, BPFP=2.1075 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,284B, BPFP=0.1078 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10404026 55.76130227 + layer.0.v_cache 0.00001395 0.01458245 + layer.1.k_cache 0.01458383 5.82722787 + layer.1.v_cache 0.00000499 0.00483400 + layer.2.k_cache 0.00236740 1.07652215 + layer.2.v_cache 0.00001641 0.01434202 + layer.3.k_cache 0.03645969 3.94706502 + layer.3.v_cache 0.00002420 0.01795399 + layer.4.k_cache 0.00069996 0.36570392 + layer.4.v_cache 0.00005119 0.03663940 + layer.4.output 0.20980707 795.10976891 + ------------------------------------------------------------------------------------- + TOTAL 0.09570067 331.34320914 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 69616 +BPFP 0.9410 bits/point +EBPFP 1.8819 equivalent bits/point +MSE 331.343209 +---------------------- -------------------------------------------------------- +Time: 2.492s Load: 0.004s, Pack+Encode: 1.478s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 331.3432 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,672B, BPFP=0.3732 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,792B, BPFP=2.6321 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,652B, BPFP=0.5920 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,260B, BPFP=2.5134 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,932B, BPFP=0.6545 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,012B, BPFP=2.4580 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,456B, BPFP=0.5482 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,220B, BPFP=2.5045 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,760B, BPFP=1.5089 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,304B, BPFP=2.0768 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,864B, BPFP=0.1232 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11725240 55.67732980 + layer.0.v_cache 0.00001567 0.01408610 + layer.1.k_cache 0.03693308 5.52793230 + layer.1.v_cache 0.00000533 0.00529923 + layer.2.k_cache 0.00240566 1.10196795 + layer.2.v_cache 0.00001850 0.01581403 + layer.3.k_cache 0.06751672 4.59383894 + layer.3.v_cache 0.00001909 0.01845376 + layer.4.k_cache 0.00067984 0.39867434 + layer.4.v_cache 0.00004901 0.03732773 + layer.4.output 0.20094273 774.00580357 + ------------------------------------------------------------------------------------- + TOTAL 0.09597026 322.67243230 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 74924 +BPFP 0.9838 bits/point +EBPFP 1.9675 equivalent bits/point +MSE 322.672432 +---------------------- -------------------------------------------------------- +Time: 2.500s Load: 0.004s, Pack+Encode: 1.485s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 322.6724 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,668B, BPFP=0.3723 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,440B, BPFP=2.5536 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,624B, BPFP=0.5857 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,916B, BPFP=2.4366 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,772B, BPFP=0.6188 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,636B, BPFP=2.3741 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,436B, BPFP=0.5437 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,668B, BPFP=2.3813 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,936B, BPFP=1.1018 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,140B, BPFP=2.0402 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,320B, BPFP=0.1059 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07685234 51.86113979 + layer.0.v_cache 0.00001390 0.01464049 + layer.1.k_cache 0.01187498 5.69575457 + layer.1.v_cache 0.00000934 0.00527438 + layer.2.k_cache 0.00239603 1.08636693 + layer.2.v_cache 0.00001621 0.01390978 + layer.3.k_cache 0.07241479 4.56339373 + layer.3.v_cache 0.00001716 0.01748805 + layer.4.k_cache 0.00072067 0.34460234 + layer.4.v_cache 0.00004958 0.03481868 + layer.4.output 0.21417375 771.20994898 + ------------------------------------------------------------------------------------- + TOTAL 0.09785772 321.30041362 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 70556 +BPFP 0.9264 bits/point +EBPFP 1.8528 equivalent bits/point +MSE 321.300414 +---------------------- -------------------------------------------------------- +Time: 2.498s Load: 0.003s, Pack+Encode: 1.483s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 321.3004 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,772B, BPFP=0.3643 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,076B, BPFP=2.4827 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,884B, BPFP=0.5929 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,508B, BPFP=2.3660 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,216B, BPFP=0.6612 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,156B, BPFP=2.2936 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,916B, BPFP=0.5995 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,500B, BPFP=2.3643 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,328B, BPFP=1.7122 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,348B, BPFP=2.1275 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,412B, BPFP=0.1002 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09264422 51.64446379 + layer.0.v_cache 0.00001680 0.01375029 + layer.1.k_cache 0.05606840 6.10952839 + layer.1.v_cache 0.00000542 0.00504468 + layer.2.k_cache 0.00241068 1.06412817 + layer.2.v_cache 0.00001702 0.01437610 + layer.3.k_cache 0.15270803 5.17354825 + layer.3.v_cache 0.00001849 0.01893006 + layer.4.k_cache 0.00070970 0.41400741 + layer.4.v_cache 0.00005539 0.03672858 + layer.4.output 0.18838861 710.41917293 + ------------------------------------------------------------------------------------- + TOTAL 0.09549261 296.31933625 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 79116 +BPFP 0.9568 bits/point +EBPFP 1.9136 equivalent bits/point +MSE 296.319336 +---------------------- -------------------------------------------------------- +Time: 2.496s Load: 0.004s, Pack+Encode: 1.480s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 296.3193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,804B, BPFP=0.3661 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,484B, BPFP=2.3304 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,904B, BPFP=0.5893 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,820B, BPFP=2.1956 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,168B, BPFP=0.6429 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,340B, BPFP=1.8953 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,836B, BPFP=0.5755 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,260B, BPFP=2.0820 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,160B, BPFP=1.4529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,436B, BPFP=1.7119 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,444B, BPFP=0.0998 +⌛️ [2/4] FRONTEND: Frontend time: 1.479s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10144231 51.40563806 + layer.0.v_cache 0.00001330 0.01328515 + layer.1.k_cache 0.09893422 6.45179293 + layer.1.v_cache 0.00000513 0.00484196 + layer.2.k_cache 0.00368564 0.97536201 + layer.2.v_cache 0.00001777 0.01378352 + layer.3.k_cache 0.06341467 4.82122208 + layer.3.v_cache 0.00001739 0.01790025 + layer.4.k_cache 0.00073693 0.37406887 + layer.4.v_cache 0.00004994 0.03451100 + layer.4.output 0.18261438 700.81534091 + ------------------------------------------------------------------------------------- + TOTAL 0.09097753 292.34234072 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 71656 +BPFP 0.8553 bits/point +EBPFP 1.7107 equivalent bits/point +MSE 292.342341 +---------------------- -------------------------------------------------------- +Time: 2.496s Load: 0.004s, Pack+Encode: 1.479s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 292.3423 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,912B, BPFP=0.3434 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,416B, BPFP=2.0503 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,060B, BPFP=0.5496 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,352B, BPFP=2.0388 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,268B, BPFP=0.5869 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,524B, BPFP=1.8901 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,852B, BPFP=0.5122 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,224B, BPFP=2.0158 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,920B, BPFP=1.4224 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,892B, BPFP=1.5970 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,136B, BPFP=0.1061 +⌛️ [2/4] FRONTEND: Frontend time: 1.475s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.008s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08947531 51.34958131 + layer.0.v_cache 0.00001432 0.01224030 + layer.1.k_cache 0.05320675 5.98010955 + layer.1.v_cache 0.00000497 0.00464410 + layer.2.k_cache 0.00562141 0.96063057 + layer.2.v_cache 0.00001574 0.01250432 + layer.3.k_cache 0.02742601 5.23274810 + layer.3.v_cache 0.00001759 0.01589035 + layer.4.k_cache 0.00114861 0.38624744 + layer.4.v_cache 0.00004734 0.03392482 + layer.4.output 0.15628038 622.64244663 + ------------------------------------------------------------------------------------- + TOTAL 0.07476122 260.14621455 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 76556 +BPFP 0.8088 bits/point +EBPFP 1.6176 equivalent bits/point +MSE 260.146215 +---------------------- -------------------------------------------------------- +Time: 2.487s Load: 0.004s, Pack+Encode: 1.475s, Decode+Unpack: 1.008s +---------------------- -------------------------------------------------------- +💾 Converting with 260.1462 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,932B, BPFP=0.3281 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,764B, BPFP=1.9980 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,168B, BPFP=0.5380 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,208B, BPFP=1.9035 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,496B, BPFP=0.7636 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,100B, BPFP=1.8852 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,200B, BPFP=0.5435 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,224B, BPFP=1.9062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,324B, BPFP=1.4137 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,420B, BPFP=1.7697 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,872B, BPFP=0.0939 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12311893 56.44256326 + layer.0.v_cache 0.00001357 0.01268318 + layer.1.k_cache 0.03743907 5.52798296 + layer.1.v_cache 0.00000515 0.00449798 + layer.2.k_cache 0.00511221 0.97402307 + layer.2.v_cache 0.00001545 0.01215727 + layer.3.k_cache 0.04225797 4.30857053 + layer.3.v_cache 0.00001799 0.01569424 + layer.4.k_cache 0.00119426 0.38952089 + layer.4.v_cache 0.00004621 0.03274068 + layer.4.output 0.14783020 588.15644410 + ------------------------------------------------------------------------------------- + TOTAL 0.07317837 246.16562016 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 80708 +BPFP 0.8063 bits/point +EBPFP 1.6126 equivalent bits/point +MSE 246.165620 +---------------------- -------------------------------------------------------- +Time: 2.493s Load: 0.005s, Pack+Encode: 1.476s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 246.1656 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,284B, BPFP=0.3499 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,228B, BPFP=1.8732 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,580B, BPFP=0.5484 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,436B, BPFP=1.7518 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,800B, BPFP=0.8885 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,444B, BPFP=1.7531 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,112B, BPFP=0.7831 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,720B, BPFP=1.7953 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,904B, BPFP=1.3640 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,156B, BPFP=1.7089 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,288B, BPFP=0.1157 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09293266 55.48590208 + layer.0.v_cache 0.00001503 0.01324988 + layer.1.k_cache 0.07854857 6.62415508 + layer.1.v_cache 0.00000559 0.00501116 + layer.2.k_cache 0.00650115 1.04919987 + layer.2.v_cache 0.00001763 0.01367216 + layer.3.k_cache 0.03715762 6.08140654 + layer.3.v_cache 0.00001788 0.01690158 + layer.4.k_cache 0.00073365 0.39582137 + layer.4.v_cache 0.00004630 0.03424201 + layer.4.output 11.17680568 525.84554447 + ------------------------------------------------------------------------------------- + TOTAL 4.61491858 220.62578665 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 88952 +BPFP 0.8015 bits/point +EBPFP 1.6031 equivalent bits/point +MSE 220.625787 +---------------------- -------------------------------------------------------- +Time: 2.495s Load: 0.004s, Pack+Encode: 1.478s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 220.6258 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 96, 128) +Output shape: (1, 96, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.output: torch.Size([1, 96, 3584]) -> torch.Size([1, 1, 96, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,912B, BPFP=0.3112 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,208B, BPFP=1.9870 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,164B, BPFP=0.5150 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,492B, BPFP=1.8704 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,960B, BPFP=0.8073 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,352B, BPFP=1.8477 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,924B, BPFP=0.9642 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,688B, BPFP=1.9023 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,476B, BPFP=1.7051 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,112B, BPFP=1.8086 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,156B, BPFP=0.0966 +⌛️ [2/4] FRONTEND: Frontend time: 1.486s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10158994 56.15913900 + layer.0.v_cache 0.00001467 0.01305772 + layer.1.k_cache 0.06416908 6.19909223 + layer.1.v_cache 0.00000542 0.00500825 + layer.2.k_cache 0.00400108 0.97583405 + layer.2.v_cache 0.00001872 0.01407209 + layer.3.k_cache 0.01491791 6.59954198 + layer.3.v_cache 0.00001877 0.01857243 + layer.4.k_cache 0.00070988 0.42727542 + layer.4.v_cache 0.00004832 0.03598420 + layer.4.output 0.14169503 564.30333891 + ------------------------------------------------------------------------------------- + TOTAL 0.06925641 236.50417352 + (elements=835,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 835584 +Total Bytes 88444 +BPFP 0.8468 bits/point +EBPFP 1.6936 equivalent bits/point +MSE 236.504174 +---------------------- -------------------------------------------------------- +Time: 2.503s Load: 0.006s, Pack+Encode: 1.486s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 236.5042 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,712B, BPFP=0.3715 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,472B, BPFP=2.4896 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,708B, BPFP=0.5877 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,780B, BPFP=2.3394 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,040B, BPFP=0.6597 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,716B, BPFP=2.3255 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,940B, BPFP=0.6380 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,544B, BPFP=2.2882 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,524B, BPFP=1.6328 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,516B, BPFP=2.0651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,248B, BPFP=0.1007 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10321087 59.17384847 + layer.0.v_cache 0.00001557 0.01440939 + layer.1.k_cache 0.06175087 6.61150360 + layer.1.v_cache 0.00000571 0.00508479 + layer.2.k_cache 0.01329466 1.03336292 + layer.2.v_cache 0.00001784 0.01491447 + layer.3.k_cache 0.05039416 5.20357259 + layer.3.v_cache 0.00001745 0.01814305 + layer.4.k_cache 0.00079786 0.39948967 + layer.4.v_cache 0.00005071 0.03750231 + layer.4.output 0.18870974 752.95504712 + ------------------------------------------------------------------------------------- + TOTAL 0.09120729 314.30571536 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 74200 +BPFP 0.9472 bits/point +EBPFP 1.8944 equivalent bits/point +MSE 314.305715 +---------------------- -------------------------------------------------------- +Time: 2.495s Load: 0.004s, Pack+Encode: 1.478s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 314.3057 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,716B, BPFP=0.3724 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,868B, BPFP=2.3585 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,728B, BPFP=0.5920 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,944B, BPFP=2.1580 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,816B, BPFP=0.6111 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,664B, BPFP=2.3142 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,772B, BPFP=0.6016 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,500B, BPFP=2.2786 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,504B, BPFP=1.1944 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,000B, BPFP=1.7361 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,876B, BPFP=0.1202 +⌛️ [2/4] FRONTEND: Frontend time: 1.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09666174 57.54089355 + layer.0.v_cache 0.00001715 0.01421725 + layer.1.k_cache 0.08476152 5.34891722 + layer.1.v_cache 0.00000506 0.00452573 + layer.2.k_cache 0.00238475 0.96061113 + layer.2.v_cache 0.00001567 0.01315878 + layer.3.k_cache 0.07332428 5.23043696 + layer.3.v_cache 0.00001796 0.01661450 + layer.4.k_cache 0.00072290 0.33894740 + layer.4.v_cache 0.00005092 0.03228536 + layer.4.output 0.18875189 752.63764881 + ------------------------------------------------------------------------------------- + TOTAL 0.09289560 313.99789115 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 69388 +BPFP 0.8858 bits/point +EBPFP 1.7715 equivalent bits/point +MSE 313.997891 +---------------------- -------------------------------------------------------- +Time: 2.497s Load: 0.004s, Pack+Encode: 1.482s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 313.9979 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,616B, BPFP=0.3713 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,624B, BPFP=2.4412 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,528B, BPFP=0.5809 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,472B, BPFP=2.4062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,928B, BPFP=0.6728 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,228B, BPFP=2.3502 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,480B, BPFP=0.5699 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,932B, BPFP=2.5119 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,744B, BPFP=1.0901 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,720B, BPFP=2.0037 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,728B, BPFP=0.1224 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10601801 54.87395163 + layer.0.v_cache 0.00001817 0.01389918 + layer.1.k_cache 0.01696859 6.44181913 + layer.1.v_cache 0.00000501 0.00501936 + layer.2.k_cache 0.00274148 1.18459544 + layer.2.v_cache 0.00001757 0.01448022 + layer.3.k_cache 0.01706722 5.96719001 + layer.3.v_cache 0.00001843 0.01796588 + layer.4.k_cache 0.00080906 0.32487847 + layer.4.v_cache 0.00005076 0.03609454 + layer.4.output 0.19983917 796.94327731 + ------------------------------------------------------------------------------------- + TOTAL 0.09074050 332.20487265 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 69000 +BPFP 0.9326 bits/point +EBPFP 1.8653 equivalent bits/point +MSE 332.204873 +---------------------- -------------------------------------------------------- +Time: 2.498s Load: 0.004s, Pack+Encode: 1.481s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 332.2049 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 65, 128) +Output shape: (1, 65, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.output: torch.Size([1, 65, 3584]) -> torch.Size([1, 1, 65, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,476B, BPFP=0.3548 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,136B, BPFP=2.6769 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,300B, BPFP=0.5529 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,380B, BPFP=2.4952 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,568B, BPFP=0.6173 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,516B, BPFP=2.5279 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,252B, BPFP=0.5413 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,048B, BPFP=2.6558 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,680B, BPFP=1.3654 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,296B, BPFP=2.2346 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,056B, BPFP=0.1049 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09579474 56.91049805 + layer.0.v_cache 0.00001312 0.01280129 + layer.1.k_cache 0.01648773 5.61685979 + layer.1.v_cache 0.00000712 0.00481523 + layer.2.k_cache 0.00431219 1.20574118 + layer.2.v_cache 0.00001803 0.01452609 + layer.3.k_cache 0.03460605 4.82427133 + layer.3.v_cache 0.00001721 0.01859197 + layer.4.k_cache 0.00068162 0.35724593 + layer.4.v_cache 0.00004727 0.03676925 + layer.4.output 0.20898957 833.57273352 + ------------------------------------------------------------------------------------- + TOTAL 0.09499483 347.29477969 + (elements=565,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 565760 +Total Bytes 69708 +BPFP 0.9857 bits/point +EBPFP 1.9714 equivalent bits/point +MSE 347.294780 +---------------------- -------------------------------------------------------- +Time: 2.497s Load: 0.003s, Pack+Encode: 1.489s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 347.2948 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 41, 128) +Output shape: (1, 41, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.output: torch.Size([1, 41, 3584]) -> torch.Size([1, 1, 41, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,616B, BPFP=0.6159 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,108B, BPFP=2.3277 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,024B, BPFP=1.5335 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,800B, BPFP=2.2104 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,892B, BPFP=1.4832 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,592B, BPFP=2.1311 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,016B, BPFP=1.5305 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,708B, BPFP=2.1753 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,008B, BPFP=1.9085 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,596B, BPFP=2.1326 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 2,640B, BPFP=0.1437 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.975s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10098721 79.08666992 + layer.0.v_cache 0.00001362 0.01475497 + layer.1.k_cache 0.01708067 9.28549641 + layer.1.v_cache 0.00000533 0.00555412 + layer.2.k_cache 0.00250472 1.16488229 + layer.2.v_cache 0.00001755 0.01709232 + layer.3.k_cache 0.08199174 6.60929461 + layer.3.v_cache 0.00001815 0.01870447 + layer.4.k_cache 0.00059437 0.38656798 + layer.4.v_cache 0.00005232 0.03989179 + layer.4.output 0.35770382 1314.74052700 + ------------------------------------------------------------------------------------- + TOTAL 0.15924661 547.04779988 + (elements=356,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 356864 +Total Bytes 50000 +BPFP 1.1209 bits/point +EBPFP 2.2418 equivalent bits/point +MSE 547.047800 +---------------------- -------------------------------------------------------- +Time: 2.472s Load: 0.002s, Pack+Encode: 1.496s, Decode+Unpack: 0.975s +---------------------- -------------------------------------------------------- +💾 Converting with 547.0478 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,432B, BPFP=0.4303 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,420B, BPFP=1.9291 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,720B, BPFP=0.8173 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,216B, BPFP=1.8678 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,880B, BPFP=0.8654 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,120B, BPFP=1.8389 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,008B, BPFP=0.6034 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,240B, BPFP=1.8750 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,364B, BPFP=1.6118 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,932B, BPFP=1.7825 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,460B, BPFP=0.1485 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.981s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09210638 57.97940768 + layer.0.v_cache 0.00001369 0.01392457 + layer.1.k_cache 0.01682195 7.39933249 + layer.1.v_cache 0.00000520 0.00523076 + layer.2.k_cache 0.00487203 1.17670345 + layer.2.v_cache 0.00001755 0.01577327 + layer.3.k_cache 0.04965206 4.92780069 + layer.3.v_cache 0.00001944 0.01884122 + layer.4.k_cache 0.00060442 0.39638431 + layer.4.v_cache 0.00005207 0.03551216 + layer.4.output 0.26182200 1040.15564904 + ------------------------------------------------------------------------------------- + TOTAL 0.11746581 432.53285023 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 48792 +BPFP 0.8624 bits/point +EBPFP 1.7248 equivalent bits/point +MSE 432.532850 +---------------------- -------------------------------------------------------- +Time: 2.486s Load: 0.003s, Pack+Encode: 1.503s, Decode+Unpack: 0.981s +---------------------- -------------------------------------------------------- +💾 Converting with 432.5329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 47, 128) +Output shape: (1, 47, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.output: torch.Size([1, 47, 3584]) -> torch.Size([1, 1, 47, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,520B, BPFP=0.5053 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,364B, BPFP=2.1157 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,640B, BPFP=0.8777 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,264B, BPFP=2.0824 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,968B, BPFP=0.9867 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,980B, BPFP=1.9880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,116B, BPFP=0.7035 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,224B, BPFP=2.0691 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,216B, BPFP=1.7340 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,952B, BPFP=1.9787 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,036B, BPFP=0.1442 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.979s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08461995 59.41965696 + layer.0.v_cache 0.00001366 0.01436050 + layer.1.k_cache 0.01838745 8.70059724 + layer.1.v_cache 0.00000609 0.00574921 + layer.2.k_cache 0.00780689 1.12537165 + layer.2.v_cache 0.00001932 0.01640854 + layer.3.k_cache 0.04292001 4.82289935 + layer.3.v_cache 0.00001942 0.01868767 + layer.4.k_cache 0.00059704 0.39113292 + layer.4.v_cache 0.00004850 0.03706125 + layer.4.output 0.29441516 1151.23290274 + ------------------------------------------------------------------------------------- + TOTAL 0.13031438 478.42248497 + (elements=409,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 409088 +Total Bytes 48280 +BPFP 0.9441 bits/point +EBPFP 1.8883 equivalent bits/point +MSE 478.422485 +---------------------- -------------------------------------------------------- +Time: 2.480s Load: 0.002s, Pack+Encode: 1.498s, Decode+Unpack: 0.979s +---------------------- -------------------------------------------------------- +💾 Converting with 478.4225 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,748B, BPFP=0.4792 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,516B, BPFP=1.7862 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,648B, BPFP=0.7259 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,316B, BPFP=1.7314 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,740B, BPFP=0.7511 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,208B, BPFP=1.7018 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,408B, BPFP=0.6601 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,312B, BPFP=1.7303 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,744B, BPFP=1.5746 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,136B, BPFP=1.6820 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,392B, BPFP=0.1328 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.984s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10653571 64.15475089 + layer.0.v_cache 0.00001396 0.01496610 + layer.1.k_cache 0.01513677 7.93426514 + layer.1.v_cache 0.00000574 0.00600255 + layer.2.k_cache 0.00241650 1.23016090 + layer.2.v_cache 0.00001997 0.01760374 + layer.3.k_cache 0.04250543 5.03727535 + layer.3.v_cache 0.00001889 0.02096658 + layer.4.k_cache 0.00060099 0.39050748 + layer.4.v_cache 0.00005317 0.04213371 + layer.4.output 0.25650722 947.39692982 + ------------------------------------------------------------------------------------- + TOTAL 0.11546222 394.74277301 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 50168 +BPFP 0.8090 bits/point +EBPFP 1.6179 equivalent bits/point +MSE 394.742773 +---------------------- -------------------------------------------------------- +Time: 2.482s Load: 0.002s, Pack+Encode: 1.497s, Decode+Unpack: 0.984s +---------------------- -------------------------------------------------------- +💾 Converting with 394.7428 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,768B, BPFP=0.5755 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,252B, BPFP=2.0352 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,980B, BPFP=0.9701 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,088B, BPFP=1.9818 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,924B, BPFP=0.9518 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,016B, BPFP=1.9583 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,460B, BPFP=0.8008 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,068B, BPFP=1.9753 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,224B, BPFP=1.7005 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,924B, BPFP=1.9284 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,056B, BPFP=0.1421 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.978s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10088900 68.91607157 + layer.0.v_cache 0.00001321 0.01362387 + layer.1.k_cache 0.01601078 8.27796936 + layer.1.v_cache 0.00000529 0.00506729 + layer.2.k_cache 0.00240732 1.04556378 + layer.2.v_cache 0.00001858 0.01616887 + layer.3.k_cache 0.04236705 5.16486994 + layer.3.v_cache 0.00001968 0.01875881 + layer.4.k_cache 0.00061965 0.39632336 + layer.4.v_cache 0.00005235 0.03789787 + layer.4.output 0.30498679 1122.89229911 + ------------------------------------------------------------------------------------- + TOTAL 0.13513591 467.30225932 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 48760 +BPFP 0.9337 bits/point +EBPFP 1.8673 equivalent bits/point +MSE 467.302259 +---------------------- -------------------------------------------------------- +Time: 2.476s Load: 0.002s, Pack+Encode: 1.496s, Decode+Unpack: 0.978s +---------------------- -------------------------------------------------------- +💾 Converting with 467.3023 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 53, 128) +Output shape: (1, 53, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.output: torch.Size([1, 53, 3584]) -> torch.Size([1, 1, 53, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,448B, BPFP=0.4269 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,592B, BPFP=1.9434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,800B, BPFP=0.8255 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,464B, BPFP=1.9057 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,540B, BPFP=0.7488 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,368B, BPFP=1.8774 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,224B, BPFP=0.6557 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,372B, BPFP=1.8785 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,204B, BPFP=1.5342 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,232B, BPFP=1.8373 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,468B, BPFP=0.1461 +⌛️ [2/4] FRONTEND: Frontend time: 1.501s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.978s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10150213 63.09802476 + layer.0.v_cache 0.00001370 0.01540916 + layer.1.k_cache 0.01753690 7.90337732 + layer.1.v_cache 0.00000536 0.00611561 + layer.2.k_cache 0.00234856 1.18330599 + layer.2.v_cache 0.00001946 0.01773130 + layer.3.k_cache 0.03772523 4.58412573 + layer.3.v_cache 0.00001792 0.01989687 + layer.4.k_cache 0.00061365 0.40375666 + layer.4.v_cache 0.00006478 0.04076671 + layer.4.output 0.25806696 1002.98795485 + ------------------------------------------------------------------------------------- + TOTAL 0.11566567 417.54048200 + (elements=461,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 461312 +Total Bytes 49712 +BPFP 0.8621 bits/point +EBPFP 1.7242 equivalent bits/point +MSE 417.540482 +---------------------- -------------------------------------------------------- +Time: 2.481s Load: 0.002s, Pack+Encode: 1.501s, Decode+Unpack: 0.978s +---------------------- -------------------------------------------------------- +💾 Converting with 417.5405 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,888B, BPFP=0.4917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,312B, BPFP=1.6438 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,264B, BPFP=0.5896 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,012B, BPFP=1.5656 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,524B, BPFP=0.9177 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,808B, BPFP=1.5125 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,604B, BPFP=0.6781 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,060B, BPFP=1.5781 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,540B, BPFP=1.4427 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,868B, BPFP=1.5281 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,368B, BPFP=0.1625 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.979s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11850098 73.93541667 + layer.0.v_cache 0.00001385 0.01398400 + layer.1.k_cache 0.01468890 6.71817983 + layer.1.v_cache 0.00000523 0.00552939 + layer.2.k_cache 0.01477869 1.31482786 + layer.2.v_cache 0.00001915 0.01892957 + layer.3.k_cache 0.14661819 5.31297913 + layer.3.v_cache 0.00001925 0.02291277 + layer.4.k_cache 0.00060183 0.45383476 + layer.4.v_cache 0.00005752 0.04337618 + layer.4.output 0.22966646 900.57946429 + ------------------------------------------------------------------------------------- + TOTAL 0.11193934 375.99389530 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 50248 +BPFP 0.7697 bits/point +EBPFP 1.5395 equivalent bits/point +MSE 375.993895 +---------------------- -------------------------------------------------------- +Time: 2.484s Load: 0.003s, Pack+Encode: 1.502s, Decode+Unpack: 0.979s +---------------------- -------------------------------------------------------- +💾 Converting with 375.9939 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,920B, BPFP=0.3261 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,124B, BPFP=2.0591 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,232B, BPFP=0.5489 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,428B, BPFP=1.9409 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,184B, BPFP=0.8804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,172B, BPFP=1.8974 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,768B, BPFP=1.1495 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,676B, BPFP=1.9830 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,752B, BPFP=1.4864 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,348B, BPFP=1.7575 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,268B, BPFP=0.1036 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10836371 56.15113366 + layer.0.v_cache 0.00001445 0.01289983 + layer.1.k_cache 0.06846340 5.52512293 + layer.1.v_cache 0.00000523 0.00474881 + layer.2.k_cache 0.00506382 1.02341030 + layer.2.v_cache 0.00001740 0.01416312 + layer.3.k_cache 0.03973729 5.88546289 + layer.3.v_cache 0.00001777 0.01729692 + layer.4.k_cache 0.00066797 0.39506721 + layer.4.v_cache 0.00005099 0.03540967 + layer.4.output 0.14786712 588.22918284 + ------------------------------------------------------------------------------------- + TOTAL 0.07396893 246.27464678 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 86872 +BPFP 0.8679 bits/point +EBPFP 1.7358 equivalent bits/point +MSE 246.274647 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.003s, Pack+Encode: 1.489s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 246.2746 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,480B, BPFP=0.3504 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,720B, BPFP=2.7746 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,368B, BPFP=0.5606 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,012B, BPFP=2.6070 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,624B, BPFP=0.6212 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,400B, BPFP=2.4621 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,348B, BPFP=0.5559 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,832B, BPFP=2.5644 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,992B, BPFP=1.1818 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,204B, BPFP=2.4157 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,412B, BPFP=0.1154 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08728886 57.78136837 + layer.0.v_cache 0.00001734 0.01456951 + layer.1.k_cache 0.01499850 6.15693341 + layer.1.v_cache 0.00000528 0.00511222 + layer.2.k_cache 0.00240521 1.11664581 + layer.2.v_cache 0.00001739 0.01470150 + layer.3.k_cache 0.03126823 4.88199222 + layer.3.v_cache 0.00001843 0.01909595 + layer.4.k_cache 0.00061395 0.35249386 + layer.4.v_cache 0.00004830 0.03518514 + layer.4.output 0.21953388 818.81256764 + ------------------------------------------------------------------------------------- + TOTAL 0.09843639 341.29800420 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 71392 +BPFP 0.9942 bits/point +EBPFP 1.9884 equivalent bits/point +MSE 341.298004 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.004s, Pack+Encode: 1.490s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 341.2980 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,464B, BPFP=0.3466 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,728B, BPFP=2.7765 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,388B, BPFP=0.5653 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,136B, BPFP=2.6364 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,736B, BPFP=0.6477 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,972B, BPFP=2.5975 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,576B, BPFP=0.6098 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,168B, BPFP=2.6439 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,088B, BPFP=1.4413 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,412B, BPFP=2.4650 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,196B, BPFP=0.1081 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08223097 59.32359360 + layer.0.v_cache 0.00001442 0.01456231 + layer.1.k_cache 0.01480860 6.74552594 + layer.1.v_cache 0.00000502 0.00498648 + layer.2.k_cache 0.00576455 1.37308907 + layer.2.v_cache 0.00002304 0.01530174 + layer.3.k_cache 0.03215047 4.65069025 + layer.3.v_cache 0.00001933 0.02079249 + layer.4.k_cache 0.00061510 0.35519493 + layer.4.v_cache 0.00004916 0.03592004 + layer.4.output 0.22077460 819.22436418 + ------------------------------------------------------------------------------------- + TOTAL 0.09888840 341.59471801 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 73864 +BPFP 1.0286 bits/point +EBPFP 2.0573 equivalent bits/point +MSE 341.594718 +---------------------- -------------------------------------------------------- +Time: 2.495s Load: 0.003s, Pack+Encode: 1.484s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 341.5947 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,712B, BPFP=0.3664 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,800B, BPFP=2.5257 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,864B, BPFP=0.6130 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,000B, BPFP=2.3545 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,012B, BPFP=0.8587 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,912B, BPFP=2.3356 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,948B, BPFP=0.6310 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,224B, BPFP=2.4024 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,052B, BPFP=1.5094 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,984B, BPFP=2.1370 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,620B, BPFP=0.1107 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.008s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10114760 54.67327964 + layer.0.v_cache 0.00001420 0.01362854 + layer.1.k_cache 0.03683303 5.94031598 + layer.1.v_cache 0.00000544 0.00510410 + layer.2.k_cache 0.00248627 1.06168480 + layer.2.v_cache 0.00001713 0.01454799 + layer.3.k_cache 0.02909129 5.22646342 + layer.3.v_cache 0.00001766 0.01824552 + layer.4.k_cache 0.00072613 0.38368173 + layer.4.v_cache 0.00005154 0.03658688 + layer.4.output 0.19731793 736.68144569 + ------------------------------------------------------------------------------------- + TOTAL 0.09127152 307.30256814 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 77128 +BPFP 0.9711 bits/point +EBPFP 1.9422 equivalent bits/point +MSE 307.302568 +---------------------- -------------------------------------------------------- +Time: 2.502s Load: 0.003s, Pack+Encode: 1.491s, Decode+Unpack: 1.008s +---------------------- -------------------------------------------------------- +💾 Converting with 307.3026 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,780B, BPFP=0.3660 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,920B, BPFP=2.4507 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,908B, BPFP=0.5979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,276B, BPFP=2.3183 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,308B, BPFP=0.6801 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,764B, BPFP=2.2130 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,096B, BPFP=0.6365 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,436B, BPFP=2.3512 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,796B, BPFP=1.6028 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,900B, BPFP=2.2410 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,928B, BPFP=0.1154 +⌛️ [2/4] FRONTEND: Frontend time: 1.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10230959 51.77657599 + layer.0.v_cache 0.00001409 0.01358143 + layer.1.k_cache 0.07973289 5.42176739 + layer.1.v_cache 0.00000506 0.00483421 + layer.2.k_cache 0.00553133 1.15591963 + layer.2.v_cache 0.00001928 0.01475417 + layer.3.k_cache 0.03946657 5.77002194 + layer.3.v_cache 0.00001980 0.01878702 + layer.4.k_cache 0.00066525 0.38333572 + layer.4.v_cache 0.00005119 0.03636454 + layer.4.output 0.17985767 711.57277961 + ------------------------------------------------------------------------------------- + TOTAL 0.08745993 296.80031760 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 79112 +BPFP 0.9568 bits/point +EBPFP 1.9135 equivalent bits/point +MSE 296.800318 +---------------------- -------------------------------------------------------- +Time: 2.506s Load: 0.004s, Pack+Encode: 1.488s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 296.8003 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,816B, BPFP=0.3685 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,908B, BPFP=2.4164 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,896B, BPFP=0.5877 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,716B, BPFP=2.1745 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,324B, BPFP=0.6745 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,068B, BPFP=1.8401 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,764B, BPFP=0.5609 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,712B, BPFP=2.1737 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,068B, BPFP=1.4343 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,864B, BPFP=1.7987 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,124B, BPFP=0.1196 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633976 50.49621423 + layer.0.v_cache 0.00001461 0.01427681 + layer.1.k_cache 0.07748085 6.34872397 + layer.1.v_cache 0.00000515 0.00452711 + layer.2.k_cache 0.00505460 0.98285606 + layer.2.v_cache 0.00001651 0.01282578 + layer.3.k_cache 0.02793909 5.11482457 + layer.3.v_cache 0.00001809 0.01685570 + layer.4.k_cache 0.00070819 0.36319584 + layer.4.v_cache 0.00004716 0.03351853 + layer.4.output 0.18765952 697.73173701 + ------------------------------------------------------------------------------------- + TOTAL 0.09007298 291.02999869 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 73260 +BPFP 0.8745 bits/point +EBPFP 1.7489 equivalent bits/point +MSE 291.029999 +---------------------- -------------------------------------------------------- +Time: 2.505s Load: 0.004s, Pack+Encode: 1.489s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 291.0300 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,660B, BPFP=0.3705 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,768B, BPFP=2.6268 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,636B, BPFP=0.5884 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,224B, BPFP=2.5054 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,892B, BPFP=0.6455 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,604B, BPFP=2.3670 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,456B, BPFP=0.5482 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,248B, BPFP=2.5107 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,476B, BPFP=1.4455 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,716B, BPFP=2.1688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,228B, BPFP=0.1029 +⌛️ [2/4] FRONTEND: Frontend time: 1.486s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08399124 56.49621582 + layer.0.v_cache 0.00001671 0.01410936 + layer.1.k_cache 0.03849935 6.46477748 + layer.1.v_cache 0.00000503 0.00492324 + layer.2.k_cache 0.00738276 0.97515651 + layer.2.v_cache 0.00001692 0.01413672 + layer.3.k_cache 0.04914829 4.77760359 + layer.3.v_cache 0.00001713 0.01812920 + layer.4.k_cache 0.00063731 0.34721345 + layer.4.v_cache 0.00004711 0.03506402 + layer.4.output 0.21328995 774.02519133 + ------------------------------------------------------------------------------------- + TOTAL 0.09839950 322.78374522 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 73908 +BPFP 0.9704 bits/point +EBPFP 1.9409 equivalent bits/point +MSE 322.783745 +---------------------- -------------------------------------------------------- +Time: 2.499s Load: 0.003s, Pack+Encode: 1.486s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 322.7837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 56, 128) +Output shape: (1, 56, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.output: torch.Size([1, 56, 3584]) -> torch.Size([1, 1, 56, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,512B, BPFP=0.4219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,436B, BPFP=1.7958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,684B, BPFP=0.7489 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,232B, BPFP=1.7388 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,544B, BPFP=0.7098 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,108B, BPFP=1.7042 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,404B, BPFP=0.6708 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,392B, BPFP=1.7835 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,508B, BPFP=1.5368 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,968B, BPFP=1.6652 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,320B, BPFP=0.1323 +⌛️ [2/4] FRONTEND: Frontend time: 1.507s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.974s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13768739 66.91672189 + layer.0.v_cache 0.00001432 0.01488546 + layer.1.k_cache 0.01638686 8.35392979 + layer.1.v_cache 0.00000533 0.00540154 + layer.2.k_cache 0.00251956 1.10221658 + layer.2.v_cache 0.00001745 0.01530219 + layer.3.k_cache 0.08599430 5.49292428 + layer.3.v_cache 0.00002012 0.02070564 + layer.4.k_cache 0.00061995 0.41043408 + layer.4.v_cache 0.00004798 0.03855870 + layer.4.output 0.24256272 966.91254783 + ------------------------------------------------------------------------------------- + TOTAL 0.11419131 402.98581853 + (elements=487,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 487424 +Total Bytes 49108 +BPFP 0.8060 bits/point +EBPFP 1.6120 equivalent bits/point +MSE 402.985819 +---------------------- -------------------------------------------------------- +Time: 2.484s Load: 0.003s, Pack+Encode: 1.507s, Decode+Unpack: 0.974s +---------------------- -------------------------------------------------------- +💾 Converting with 402.9858 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.3633 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,048B, BPFP=2.3531 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,960B, BPFP=0.5781 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,864B, BPFP=2.1219 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,256B, BPFP=0.6359 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,132B, BPFP=1.9789 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,028B, BPFP=0.5914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,176B, BPFP=2.1828 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,568B, BPFP=1.6734 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,032B, BPFP=1.9594 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,900B, BPFP=0.1088 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09278157 54.54735718 + layer.0.v_cache 0.00001504 0.01347903 + layer.1.k_cache 0.07523044 5.49166870 + layer.1.v_cache 0.00000517 0.00486393 + layer.2.k_cache 0.00381728 1.05088043 + layer.2.v_cache 0.00001789 0.01352905 + layer.3.k_cache 0.03327554 4.10453644 + layer.3.v_cache 0.00001939 0.01715637 + layer.4.k_cache 0.00067710 0.39766519 + layer.4.v_cache 0.00005192 0.03644916 + layer.4.output 0.16994183 676.95792411 + ------------------------------------------------------------------------------------- + TOTAL 0.08208730 282.61076790 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 77824 +BPFP 0.8941 bits/point +EBPFP 1.7882 equivalent bits/point +MSE 282.610768 +---------------------- -------------------------------------------------------- +Time: 2.491s Load: 0.004s, Pack+Encode: 1.483s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 282.6108 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,792B, BPFP=0.4746 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,400B, BPFP=1.6949 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,396B, BPFP=0.6345 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,048B, BPFP=1.6017 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,564B, BPFP=0.9439 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,948B, BPFP=1.5752 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,496B, BPFP=0.6610 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,160B, BPFP=1.6314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,644B, BPFP=1.4947 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,912B, BPFP=1.5657 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,004B, BPFP=0.1515 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.968s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12468526 68.58541611 + layer.0.v_cache 0.00001451 0.01636923 + layer.1.k_cache 0.01313243 6.21283890 + layer.1.v_cache 0.00000580 0.00615477 + layer.2.k_cache 0.00906488 1.56240871 + layer.2.v_cache 0.00002099 0.01843878 + layer.3.k_cache 0.01569995 5.00007604 + layer.3.v_cache 0.00002002 0.02157429 + layer.4.k_cache 0.00064964 0.43929633 + layer.4.v_cache 0.00005478 0.04110900 + layer.4.output 0.23031524 917.48872579 + ------------------------------------------------------------------------------------- + TOTAL 0.10444441 382.60733898 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 50364 +BPFP 0.7846 bits/point +EBPFP 1.5692 equivalent bits/point +MSE 382.607339 +---------------------- -------------------------------------------------------- +Time: 2.459s Load: 0.003s, Pack+Encode: 1.489s, Decode+Unpack: 0.968s +---------------------- -------------------------------------------------------- +💾 Converting with 382.6073 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,868B, BPFP=0.4865 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,404B, BPFP=1.6677 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,376B, BPFP=0.6188 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,020B, BPFP=1.5677 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,172B, BPFP=0.8260 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,888B, BPFP=1.5333 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,492B, BPFP=0.6490 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,044B, BPFP=1.5740 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,624B, BPFP=1.4646 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,932B, BPFP=1.5448 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,496B, BPFP=0.1673 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.968s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15164433 78.18001302 + layer.0.v_cache 0.00001435 0.01557362 + layer.1.k_cache 0.01704682 6.60751292 + layer.1.v_cache 0.00000573 0.00609738 + layer.2.k_cache 0.00240009 1.27841517 + layer.2.v_cache 0.00001845 0.01801522 + layer.3.k_cache 0.02161660 5.77453817 + layer.3.v_cache 0.00001985 0.02207854 + layer.4.k_cache 0.00061459 0.43752391 + layer.4.v_cache 0.00005275 0.04244104 + layer.4.output 0.22653077 902.84144345 + ------------------------------------------------------------------------------------- + TOTAL 0.10465582 377.19248901 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 50316 +BPFP 0.7708 bits/point +EBPFP 1.5415 equivalent bits/point +MSE 377.192489 +---------------------- -------------------------------------------------------- +Time: 2.455s Load: 0.003s, Pack+Encode: 1.484s, Decode+Unpack: 0.968s +---------------------- -------------------------------------------------------- +💾 Converting with 377.1925 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,716B, BPFP=0.3673 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,700B, BPFP=2.2902 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,732B, BPFP=0.5848 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,092B, BPFP=1.9461 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,104B, BPFP=0.6644 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,508B, BPFP=2.0351 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,408B, BPFP=0.7295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,980B, BPFP=1.9221 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,784B, BPFP=1.4521 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,888B, BPFP=1.6884 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,528B, BPFP=0.1079 +⌛️ [2/4] FRONTEND: Frontend time: 1.479s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.002s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08869729 54.57700128 + layer.0.v_cache 0.00001594 0.01388640 + layer.1.k_cache 0.08105310 5.09436704 + layer.1.v_cache 0.00000494 0.00466148 + layer.2.k_cache 0.00256802 0.85125607 + layer.2.v_cache 0.00001661 0.01357049 + layer.3.k_cache 0.01835501 5.36295350 + layer.3.v_cache 0.00001861 0.01650672 + layer.4.k_cache 0.00078324 0.37531239 + layer.4.v_cache 0.00004769 0.03244766 + layer.4.output 0.18611241 741.85377935 + ------------------------------------------------------------------------------------- + TOTAL 0.08790278 309.37167168 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 67440 +BPFP 0.8491 bits/point +EBPFP 1.6982 equivalent bits/point +MSE 309.371672 +---------------------- -------------------------------------------------------- +Time: 2.484s Load: 0.003s, Pack+Encode: 1.479s, Decode+Unpack: 1.002s +---------------------- -------------------------------------------------------- +💾 Converting with 309.3717 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,924B, BPFP=0.3099 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,048B, BPFP=1.9407 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,256B, BPFP=0.5245 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,512B, BPFP=1.8544 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,240B, BPFP=0.8441 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,476B, BPFP=1.8486 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,652B, BPFP=0.9104 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,700B, BPFP=1.8847 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,996B, BPFP=1.4491 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,968B, BPFP=1.7668 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,496B, BPFP=0.1035 +⌛️ [2/4] FRONTEND: Frontend time: 1.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17228427 58.73809500 + layer.0.v_cache 0.00001543 0.01340437 + layer.1.k_cache 0.04621154 6.25784019 + layer.1.v_cache 0.00000545 0.00494865 + layer.2.k_cache 0.00739691 1.18360429 + layer.2.v_cache 0.00001797 0.01397912 + layer.3.k_cache 0.03096496 7.05784937 + layer.3.v_cache 0.00001890 0.01735261 + layer.4.k_cache 0.00071091 0.41633134 + layer.4.v_cache 0.00004900 0.03235852 + layer.4.output 0.02450962 573.57096834 + ------------------------------------------------------------------------------------- + TOTAL 0.02524957 240.51367893 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 87268 +BPFP 0.8269 bits/point +EBPFP 1.6538 equivalent bits/point +MSE 240.513679 +---------------------- -------------------------------------------------------- +Time: 2.481s Load: 0.004s, Pack+Encode: 1.476s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 240.5137 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,920B, BPFP=0.3226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,248B, BPFP=2.0578 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,512B, BPFP=0.5901 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,696B, BPFP=1.9651 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,232B, BPFP=0.8790 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,368B, BPFP=1.9099 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,584B, BPFP=0.7702 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,696B, BPFP=1.9651 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,896B, BPFP=1.6626 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,004B, BPFP=1.8488 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,172B, BPFP=0.1001 +⌛️ [2/4] FRONTEND: Frontend time: 1.474s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.002s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16174046 58.72676726 + layer.0.v_cache 0.00001612 0.01361496 + layer.1.k_cache 0.03352992 6.02129536 + layer.1.v_cache 0.00000563 0.00503030 + layer.2.k_cache 0.00496730 1.17244171 + layer.2.v_cache 0.00001740 0.01420621 + layer.3.k_cache 0.01521110 6.09401711 + layer.3.v_cache 0.00001939 0.01851853 + layer.4.k_cache 0.00066431 0.39937887 + layer.4.v_cache 0.00005067 0.03501480 + layer.4.output 0.14624557 582.59331797 + ------------------------------------------------------------------------------------- + TOTAL 0.07293772 244.15608888 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 87328 +BPFP 0.8631 bits/point +EBPFP 1.7261 equivalent bits/point +MSE 244.156089 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.004s, Pack+Encode: 1.474s, Decode+Unpack: 1.002s +---------------------- -------------------------------------------------------- +💾 Converting with 244.1561 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.3379 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,012B, BPFP=2.1824 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,048B, BPFP=0.5538 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,040B, BPFP=2.0058 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,088B, BPFP=0.7427 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,976B, BPFP=1.9942 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,304B, BPFP=0.6003 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,564B, BPFP=2.1010 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,076B, BPFP=1.4673 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,756B, BPFP=1.9542 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,196B, BPFP=0.1089 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.001s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07121617 54.91685592 + layer.0.v_cache 0.00001681 0.01276269 + layer.1.k_cache 0.06877367 6.03235431 + layer.1.v_cache 0.00000530 0.00480169 + layer.2.k_cache 0.00486504 1.05907281 + layer.2.v_cache 0.00001749 0.01277257 + layer.3.k_cache 0.08833314 4.58870892 + layer.3.v_cache 0.00001723 0.01618326 + layer.4.k_cache 0.00077278 0.36284385 + layer.4.v_cache 0.00004595 0.03630753 + layer.4.output 0.15812202 629.99127907 + ------------------------------------------------------------------------------------- + TOTAL 0.07887751 263.35185983 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 80920 +BPFP 0.8648 bits/point +EBPFP 1.7297 equivalent bits/point +MSE 263.351860 +---------------------- -------------------------------------------------------- +Time: 2.490s Load: 0.005s, Pack+Encode: 1.484s, Decode+Unpack: 1.001s +---------------------- -------------------------------------------------------- +💾 Converting with 263.3519 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,144B, BPFP=0.3317 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,280B, BPFP=1.8998 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,544B, BPFP=0.5483 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,712B, BPFP=1.8119 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,828B, BPFP=0.9016 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,480B, BPFP=1.7760 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,444B, BPFP=1.1516 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,724B, BPFP=1.8137 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,384B, BPFP=1.7611 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,204B, BPFP=1.7333 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,352B, BPFP=0.0962 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13964315 55.19835821 + layer.0.v_cache 0.00001432 0.01352339 + layer.1.k_cache 0.04810369 5.66454353 + layer.1.v_cache 0.00000572 0.00526741 + layer.2.k_cache 0.00492308 1.25022390 + layer.2.v_cache 0.00001777 0.01413360 + layer.3.k_cache 0.04255116 6.72967952 + layer.3.v_cache 0.00001880 0.01764505 + layer.4.k_cache 0.00073971 0.41985457 + layer.4.v_cache 0.00004573 0.03614869 + layer.4.output 11.28739116 530.96631895 + ------------------------------------------------------------------------------------- + TOTAL 4.66163537 222.71256533 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 93096 +BPFP 0.8472 bits/point +EBPFP 1.6944 equivalent bits/point +MSE 222.712565 +---------------------- -------------------------------------------------------- +Time: 2.492s Load: 0.005s, Pack+Encode: 1.480s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 222.7126 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,896B, BPFP=0.3366 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,156B, BPFP=2.1584 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,156B, BPFP=0.5604 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,116B, BPFP=1.9737 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,352B, BPFP=0.7727 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,292B, BPFP=2.0050 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,184B, BPFP=0.7429 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,268B, BPFP=2.0007 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,136B, BPFP=1.6222 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,068B, BPFP=1.9652 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,120B, BPFP=0.1045 +⌛️ [2/4] FRONTEND: Frontend time: 1.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.002s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14145413 53.22909268 + layer.0.v_cache 0.00001516 0.01319957 + layer.1.k_cache 0.10647225 6.14593645 + layer.1.v_cache 0.00000560 0.00476722 + layer.2.k_cache 0.00244946 1.09183918 + layer.2.v_cache 0.00001685 0.01293066 + layer.3.k_cache 0.01581429 5.72734278 + layer.3.v_cache 0.00001841 0.01632396 + layer.4.k_cache 0.00072023 0.38069703 + layer.4.v_cache 0.00004652 0.03409933 + layer.4.output 0.15451367 615.55357143 + ------------------------------------------------------------------------------------- + TOTAL 0.07932992 257.38418993 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 83744 +BPFP 0.8747 bits/point +EBPFP 1.7493 equivalent bits/point +MSE 257.384190 +---------------------- -------------------------------------------------------- +Time: 2.487s Load: 0.004s, Pack+Encode: 1.480s, Decode+Unpack: 1.002s +---------------------- -------------------------------------------------------- +💾 Converting with 257.3842 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,756B, BPFP=0.3708 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,708B, BPFP=2.4721 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,816B, BPFP=0.5946 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,020B, BPFP=2.3269 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,876B, BPFP=0.6073 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,452B, BPFP=1.7846 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,580B, BPFP=0.5448 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,768B, BPFP=2.2736 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,572B, BPFP=1.3877 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,032B, BPFP=1.6959 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,340B, BPFP=0.1007 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10237078 55.54926362 + layer.0.v_cache 0.00001391 0.01319854 + layer.1.k_cache 0.05953315 5.49067234 + layer.1.v_cache 0.00000517 0.00503117 + layer.2.k_cache 0.00728044 1.08529890 + layer.2.v_cache 0.00001648 0.01377557 + layer.3.k_cache 0.03101032 4.45896623 + layer.3.v_cache 0.00001666 0.01698803 + layer.4.k_cache 0.00072555 0.35389184 + layer.4.v_cache 0.00004870 0.03182009 + layer.4.output 0.18364459 725.01345319 + ------------------------------------------------------------------------------------- + TOTAL 0.08744314 302.47723992 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 69920 +BPFP 0.8684 bits/point +EBPFP 1.7369 equivalent bits/point +MSE 302.477240 +---------------------- -------------------------------------------------------- +Time: 2.507s Load: 0.003s, Pack+Encode: 1.495s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 302.4772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 111, 128) +Output shape: (1, 111, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.output: torch.Size([1, 111, 3584]) -> torch.Size([1, 1, 111, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,328B, BPFP=0.3277 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,272B, BPFP=1.7275 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,928B, BPFP=0.5529 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,680B, BPFP=1.6441 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,632B, BPFP=0.6520 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,480B, BPFP=1.6160 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,336B, BPFP=0.8919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,768B, BPFP=1.6565 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,256B, BPFP=1.4437 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,200B, BPFP=1.5766 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,880B, BPFP=0.0981 +⌛️ [2/4] FRONTEND: Frontend time: 1.479s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11107565 51.95220562 + layer.0.v_cache 0.00001597 0.01218073 + layer.1.k_cache 0.04333284 6.35675654 + layer.1.v_cache 0.00000558 0.00470472 + layer.2.k_cache 0.00477623 0.92785122 + layer.2.v_cache 0.00001758 0.01288381 + layer.3.k_cache 0.01020195 6.48994075 + layer.3.v_cache 0.00001870 0.01552014 + layer.4.k_cache 0.00079488 0.40437111 + layer.4.v_cache 0.00005041 0.03359907 + layer.4.output 10.27057757 482.41493726 + ------------------------------------------------------------------------------------- + TOTAL 4.23907840 202.53615144 + (elements=966,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 966144 +Total Bytes 90760 +BPFP 0.7515 bits/point +EBPFP 1.5030 equivalent bits/point +MSE 202.536151 +---------------------- -------------------------------------------------------- +Time: 2.488s Load: 0.005s, Pack+Encode: 1.479s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 202.5362 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,692B, BPFP=0.3724 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,528B, BPFP=2.5370 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,692B, BPFP=0.5924 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,084B, BPFP=2.4393 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,012B, BPFP=0.6629 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,904B, BPFP=2.3996 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,636B, BPFP=0.5801 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,020B, BPFP=2.4252 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,820B, BPFP=1.2808 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,312B, BPFP=2.2694 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,776B, BPFP=0.1187 +⌛️ [2/4] FRONTEND: Frontend time: 1.486s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09375787 55.65735379 + layer.0.v_cache 0.00001622 0.01337316 + layer.1.k_cache 0.05984993 5.61245771 + layer.1.v_cache 0.00000512 0.00479787 + layer.2.k_cache 0.00243208 1.02995848 + layer.2.v_cache 0.00001865 0.01475839 + layer.3.k_cache 0.04944374 5.13197176 + layer.3.v_cache 0.00001908 0.01800219 + layer.4.k_cache 0.00084877 0.36567739 + layer.4.v_cache 0.00005139 0.03416236 + layer.4.output 0.19371969 761.58695926 + ------------------------------------------------------------------------------------- + TOTAL 0.09191063 317.58771929 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 74476 +BPFP 0.9641 bits/point +EBPFP 1.9282 equivalent bits/point +MSE 317.587719 +---------------------- -------------------------------------------------------- +Time: 2.494s Load: 0.004s, Pack+Encode: 1.486s, Decode+Unpack: 1.004s +---------------------- -------------------------------------------------------- +💾 Converting with 317.5877 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 109, 128) +Output shape: (1, 109, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.output: torch.Size([1, 109, 3584]) -> torch.Size([1, 1, 109, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,316B, BPFP=0.3320 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,272B, BPFP=1.7592 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,912B, BPFP=0.5608 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,748B, BPFP=1.6841 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,764B, BPFP=0.8263 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,408B, BPFP=1.6353 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,648B, BPFP=1.0963 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,808B, BPFP=1.6927 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,316B, BPFP=1.4788 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,232B, BPFP=1.6101 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,360B, BPFP=0.1098 +⌛️ [2/4] FRONTEND: Frontend time: 1.478s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.003s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860953 53.33703053 + layer.0.v_cache 0.00001500 0.01284550 + layer.1.k_cache 0.05917679 6.39069891 + layer.1.v_cache 0.00000528 0.00489254 + layer.2.k_cache 0.01253135 1.24757063 + layer.2.v_cache 0.00001869 0.01421268 + layer.3.k_cache 0.07406004 7.21802094 + layer.3.v_cache 0.00001854 0.01722607 + layer.4.k_cache 0.00071234 0.40178253 + layer.4.v_cache 0.00004916 0.03651149 + layer.4.output 10.45909253 490.15522608 + ------------------------------------------------------------------------------------- + TOTAL 4.32287320 205.86866908 + (elements=948,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 948736 +Total Bytes 93784 +BPFP 0.7908 bits/point +EBPFP 1.5816 equivalent bits/point +MSE 205.868669 +---------------------- -------------------------------------------------------- +Time: 2.486s Load: 0.005s, Pack+Encode: 1.478s, Decode+Unpack: 1.003s +---------------------- -------------------------------------------------------- +💾 Converting with 205.8687 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst + to output-fixed/kimiaudio/lambda0.004/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.8823 bits/point +Avg EBPFP 1.7646 equivalent bits/point +Avg MSE 317.952833 +Avg Time 2.498s +------------------------ ---------------------------- diff --git a/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..a14203f57394818805098e56a0b3a600398d2325 --- /dev/null +++ b/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 559 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench +Output output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,788B, BPFP=0.3449 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,124B, BPFP=1.7600 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,344B, BPFP=0.6451 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,628B, BPFP=1.6644 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,676B, BPFP=0.7091 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,720B, BPFP=1.4892 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,268B, BPFP=0.6304 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,640B, BPFP=1.4738 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,824B, BPFP=1.1235 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,796B, BPFP=1.5039 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,856B, BPFP=0.1889 +⌛️ [2/4] FRONTEND: Frontend time: 0.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.247s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12089630 166.23648486 + layer.0.v_cache 0.00001394 0.01434611 + layer.1.k_cache 0.01174029 16.73025927 + layer.1.v_cache 0.00000579 0.00581185 + layer.2.k_cache 0.00835781 1.92379083 + layer.2.v_cache 0.00001852 0.01585332 + layer.3.k_cache 0.02896961 7.37861539 + layer.3.v_cache 0.00001872 0.01678492 + layer.4.k_cache 0.00063445 0.39766858 + layer.4.v_cache 0.00005063 0.03756030 + layer.4.output 0.17246709 669.42857143 + ------------------------------------------------------------------------------------- + TOTAL 0.08105739 286.98571620 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 65664 +BPFP 0.7451 bits/point +EBPFP 1.4902 equivalent bits/point +MSE 286.985716 +---------------------- -------------------------------------------------------- +Time: 0.735s Load: 0.005s, Pack+Encode: 0.483s, Decode+Unpack: 0.247s +---------------------- -------------------------------------------------------- +💾 Converting with 286.9857 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,788B, BPFP=0.3492 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,252B, BPFP=1.8070 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,356B, BPFP=0.6555 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,892B, BPFP=1.7367 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,828B, BPFP=0.7477 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,960B, BPFP=1.5547 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,180B, BPFP=0.6211 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,360B, BPFP=1.6328 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,260B, BPFP=1.2227 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,984B, BPFP=1.5594 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,536B, BPFP=0.1824 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08029544 162.18452148 + layer.0.v_cache 0.00001376 0.01513198 + layer.1.k_cache 0.03632395 16.68359375 + layer.1.v_cache 0.00000563 0.00631564 + layer.2.k_cache 0.00522719 1.99500103 + layer.2.v_cache 0.00001997 0.01690002 + layer.3.k_cache 0.02707962 8.23684235 + layer.3.v_cache 0.00001843 0.01751847 + layer.4.k_cache 0.00063057 0.41491914 + layer.4.v_cache 0.00005085 0.04042881 + layer.4.output 0.18451144 678.07739955 + ------------------------------------------------------------------------------------- + TOTAL 0.08477915 290.36193938 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 67396 +BPFP 0.7743 bits/point +EBPFP 1.5486 equivalent bits/point +MSE 290.361939 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.003s, Pack+Encode: 0.166s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 290.3619 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,892B, BPFP=0.3438 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,472B, BPFP=1.7209 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,444B, BPFP=0.6257 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,012B, BPFP=1.6374 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,820B, BPFP=0.6940 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,300B, BPFP=1.5080 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,376B, BPFP=0.6134 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,504B, BPFP=1.5451 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,588B, BPFP=1.1969 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,484B, BPFP=1.5414 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,868B, BPFP=0.1783 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10874708 165.58848110 + layer.0.v_cache 0.00001608 0.01480039 + layer.1.k_cache 0.03270756 16.12196528 + layer.1.v_cache 0.00000572 0.00544375 + layer.2.k_cache 0.00777350 1.88263010 + layer.2.v_cache 0.00002080 0.01558545 + layer.3.k_cache 0.05920834 7.94062486 + layer.3.v_cache 0.00001846 0.01653428 + layer.4.k_cache 0.00062493 0.37338510 + layer.4.v_cache 0.00005036 0.03524736 + layer.4.output 0.17137358 630.66190822 + ------------------------------------------------------------------------------------- + TOTAL 0.08286988 270.97812090 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 69760 +BPFP 0.7456 bits/point +EBPFP 1.4911 equivalent bits/point +MSE 270.978121 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 270.9781 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,780B, BPFP=0.3566 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,876B, BPFP=1.7780 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,268B, BPFP=0.6546 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,436B, BPFP=1.6899 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,432B, BPFP=0.6875 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,428B, BPFP=1.4880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,108B, BPFP=0.6226 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,704B, BPFP=1.5433 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,912B, BPFP=1.1843 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,696B, BPFP=1.5417 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,744B, BPFP=0.1930 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12073612 163.55222105 + layer.0.v_cache 0.00001348 0.01416282 + layer.1.k_cache 0.03279715 16.75624437 + layer.1.v_cache 0.00000568 0.00639563 + layer.2.k_cache 0.00696802 1.83993628 + layer.2.v_cache 0.00001833 0.01664433 + layer.3.k_cache 0.03227850 7.58987114 + layer.3.v_cache 0.00001834 0.01844704 + layer.4.k_cache 0.00060115 0.41213476 + layer.4.v_cache 0.00005191 0.04041498 + layer.4.output 0.18320683 694.96697573 + ------------------------------------------------------------------------------------- + TOTAL 0.08681979 297.35384133 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 64384 +BPFP 0.7587 bits/point +EBPFP 1.5173 equivalent bits/point +MSE 297.353841 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.003s, Pack+Encode: 0.165s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 297.3538 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,768B, BPFP=0.3542 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,096B, BPFP=1.8221 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,276B, BPFP=0.6562 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,152B, BPFP=1.6330 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,628B, BPFP=0.7268 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,196B, BPFP=1.4415 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,204B, BPFP=0.6418 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,856B, BPFP=1.5737 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,976B, BPFP=1.1971 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,964B, BPFP=1.5954 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,720B, BPFP=0.1923 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10595027 167.76807893 + layer.0.v_cache 0.00001376 0.01414487 + layer.1.k_cache 0.03307601 16.44654885 + layer.1.v_cache 0.00000550 0.00581739 + layer.2.k_cache 0.00695119 2.07577593 + layer.2.v_cache 0.00001929 0.01542387 + layer.3.k_cache 0.03450640 7.55618912 + layer.3.v_cache 0.00001844 0.01675944 + layer.4.k_cache 0.00061739 0.38500884 + layer.4.v_cache 0.00005297 0.03685998 + layer.4.output 0.18589628 694.80471612 + ------------------------------------------------------------------------------------- + TOTAL 0.08720501 297.52668353 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 64836 +BPFP 0.7640 bits/point +EBPFP 1.5280 equivalent bits/point +MSE 297.526684 +---------------------- -------------------------------------------------------- +Time: 0.380s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 297.5267 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,792B, BPFP=0.3544 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,256B, BPFP=1.8307 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,340B, BPFP=0.6606 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,892B, BPFP=1.7587 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,780B, BPFP=0.7476 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,008B, BPFP=1.5839 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,220B, BPFP=0.6369 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,024B, BPFP=1.5870 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,192B, BPFP=1.2247 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,228B, BPFP=1.6274 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,272B, BPFP=0.1772 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07772056 161.53874110 + layer.0.v_cache 0.00001351 0.01440087 + layer.1.k_cache 0.03518496 16.87351352 + layer.1.v_cache 0.00000556 0.00562094 + layer.2.k_cache 0.00230807 1.88387308 + layer.2.v_cache 0.00001833 0.01557586 + layer.3.k_cache 0.04514034 8.42198316 + layer.3.v_cache 0.00001901 0.01713284 + layer.4.k_cache 0.00062178 0.40783759 + layer.4.v_cache 0.00004990 0.03852666 + layer.4.output 0.18426369 686.62596067 + ------------------------------------------------------------------------------------- + TOTAL 0.08534870 293.85876061 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 67004 +BPFP 0.7796 bits/point +EBPFP 1.5591 equivalent bits/point +MSE 293.858761 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 293.8588 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,772B, BPFP=0.3596 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,192B, BPFP=1.8653 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,280B, BPFP=0.6656 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,432B, BPFP=1.7110 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,612B, BPFP=0.7330 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,676B, BPFP=1.5576 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,088B, BPFP=0.6266 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,016B, BPFP=1.6266 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,980B, BPFP=1.2135 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,716B, BPFP=1.5657 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,720B, BPFP=0.1948 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14568528 171.28140219 + layer.0.v_cache 0.00001322 0.01429060 + layer.1.k_cache 0.01379078 16.59532265 + layer.1.v_cache 0.00000548 0.00559345 + layer.2.k_cache 0.00727733 1.78373639 + layer.2.v_cache 0.00001786 0.01617329 + layer.3.k_cache 0.03082054 7.06244768 + layer.3.v_cache 0.00001879 0.01863574 + layer.4.k_cache 0.00061820 0.40305789 + layer.4.v_cache 0.00005377 0.03894777 + layer.4.output 0.19662610 703.35505566 + ------------------------------------------------------------------------------------- + TOTAL 0.09262847 301.21794102 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 65484 +BPFP 0.7817 bits/point +EBPFP 1.5633 equivalent bits/point +MSE 301.217941 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 301.2179 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3516 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,512B, BPFP=1.8578 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,400B, BPFP=0.6641 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,956B, BPFP=1.7492 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,884B, BPFP=0.7586 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,384B, BPFP=1.6375 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,348B, BPFP=0.6539 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,148B, BPFP=1.5914 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,284B, BPFP=1.2273 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,176B, BPFP=1.5969 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,884B, BPFP=0.1921 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10149734 160.77661133 + layer.0.v_cache 0.00001421 0.01533714 + layer.1.k_cache 0.01324784 16.27150574 + layer.1.v_cache 0.00000586 0.00605325 + layer.2.k_cache 0.00706157 1.84704399 + layer.2.v_cache 0.00001822 0.01592588 + layer.3.k_cache 0.02712160 7.49992905 + layer.3.v_cache 0.00002016 0.01718642 + layer.4.k_cache 0.00063687 0.41192937 + layer.4.v_cache 0.00005451 0.03853097 + layer.4.output 0.17858309 677.83889509 + ------------------------------------------------------------------------------------- + TOTAL 0.08233881 290.10425404 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 68776 +BPFP 0.7902 bits/point +EBPFP 1.5803 equivalent bits/point +MSE 290.104254 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1043 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,896B, BPFP=0.3405 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,880B, BPFP=1.7744 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,460B, BPFP=0.6214 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,196B, BPFP=1.6516 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,276B, BPFP=0.7680 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,472B, BPFP=1.5216 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,632B, BPFP=0.6523 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,832B, BPFP=1.5862 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,876B, BPFP=1.2349 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,884B, BPFP=1.5955 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,768B, BPFP=0.1736 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10772817 167.81993085 + layer.0.v_cache 0.00001348 0.01432632 + layer.1.k_cache 0.03276526 16.09541128 + layer.1.v_cache 0.00000562 0.00557621 + layer.2.k_cache 0.01020141 2.03039656 + layer.2.v_cache 0.00001934 0.01487235 + layer.3.k_cache 0.04694761 7.05366183 + layer.3.v_cache 0.00001910 0.01665701 + layer.4.k_cache 0.00062564 0.37643297 + layer.4.v_cache 0.00005323 0.03600971 + layer.4.output 0.17166949 623.59180008 + ------------------------------------------------------------------------------------- + TOTAL 0.08235678 268.15328680 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 72172 +BPFP 0.7625 bits/point +EBPFP 1.5249 equivalent bits/point +MSE 268.153287 +---------------------- -------------------------------------------------------- +Time: 0.382s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 268.1533 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3494 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,192B, BPFP=1.7304 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,360B, BPFP=0.6325 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,712B, BPFP=1.6401 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,680B, BPFP=0.6928 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,100B, BPFP=1.5248 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,260B, BPFP=0.6137 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,328B, BPFP=1.5678 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,024B, BPFP=1.1340 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,980B, BPFP=1.5023 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,568B, BPFP=0.1766 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12335577 162.14361352 + layer.0.v_cache 0.00001400 0.01481432 + layer.1.k_cache 0.01519604 16.08021637 + layer.1.v_cache 0.00000547 0.00590334 + layer.2.k_cache 0.01098718 1.96195782 + layer.2.v_cache 0.00001779 0.01626686 + layer.3.k_cache 0.07416311 8.61495641 + layer.3.v_cache 0.00001944 0.01819891 + layer.4.k_cache 0.00061719 0.40869655 + layer.4.v_cache 0.00006444 0.03848225 + layer.4.output 0.17504946 652.77307444 + ------------------------------------------------------------------------------------- + TOTAL 0.08528157 279.92438985 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 67060 +BPFP 0.7426 bits/point +EBPFP 1.4852 equivalent bits/point +MSE 279.924390 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.005s, Pack+Encode: 0.168s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 279.9244 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,756B, BPFP=0.3563 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,112B, BPFP=1.8490 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,204B, BPFP=0.6502 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,300B, BPFP=1.6843 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,432B, BPFP=0.6964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,448B, BPFP=1.5114 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,016B, BPFP=0.6120 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,648B, BPFP=1.5519 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,652B, BPFP=1.1469 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,556B, BPFP=1.5333 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,552B, BPFP=0.1899 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13955514 173.72162896 + layer.0.v_cache 0.00001392 0.01477893 + layer.1.k_cache 0.03454666 17.39443613 + layer.1.v_cache 0.00000526 0.00615834 + layer.2.k_cache 0.00236268 1.90705535 + layer.2.v_cache 0.00001660 0.01562698 + layer.3.k_cache 0.02874016 7.26891773 + layer.3.v_cache 0.00001899 0.01741648 + layer.4.k_cache 0.00062191 0.41881314 + layer.4.v_cache 0.00005265 0.03910118 + layer.4.output 0.18728454 703.30264378 + ------------------------------------------------------------------------------------- + TOTAL 0.08923092 301.40720234 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 63676 +BPFP 0.7601 bits/point +EBPFP 1.5201 equivalent bits/point +MSE 301.407202 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 301.4072 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,768B, BPFP=0.3588 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,844B, BPFP=1.7946 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,256B, BPFP=0.6607 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,364B, BPFP=1.6972 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,476B, BPFP=0.7054 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,660B, BPFP=1.5544 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,064B, BPFP=0.6218 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,660B, BPFP=1.5544 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,028B, BPFP=1.2232 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,536B, BPFP=1.5292 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,676B, BPFP=0.1935 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11454884 174.27435065 + layer.0.v_cache 0.00001351 0.01569355 + layer.1.k_cache 0.01665594 16.56874302 + layer.1.v_cache 0.00000570 0.00647326 + layer.2.k_cache 0.00836552 1.92216313 + layer.2.v_cache 0.00001974 0.01755690 + layer.3.k_cache 0.03124422 7.50440087 + layer.3.v_cache 0.00001876 0.01810029 + layer.4.k_cache 0.00062231 0.41837182 + layer.4.v_cache 0.00005487 0.04105764 + layer.4.output 0.19663499 703.78525046 + ------------------------------------------------------------------------------------- + TOTAL 0.09105849 301.60492143 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 64332 +BPFP 0.7679 bits/point +EBPFP 1.5358 equivalent bits/point +MSE 301.604921 +---------------------- -------------------------------------------------------- +Time: 0.382s Load: 0.004s, Pack+Encode: 0.168s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 301.6049 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,880B, BPFP=0.3497 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,116B, BPFP=1.6957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,408B, BPFP=0.6339 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,544B, BPFP=1.5893 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,668B, BPFP=0.6823 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,800B, BPFP=1.4509 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,228B, BPFP=0.6004 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,100B, BPFP=1.5067 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,200B, BPFP=1.1533 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,896B, BPFP=1.4688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,904B, BPFP=0.1835 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12290487 159.96128627 + layer.0.v_cache 0.00001422 0.01513587 + layer.1.k_cache 0.01283375 16.16368176 + layer.1.v_cache 0.00000579 0.00646862 + layer.2.k_cache 0.01100613 1.87576603 + layer.2.v_cache 0.00001929 0.01643028 + layer.3.k_cache 0.03011658 8.09386480 + layer.3.v_cache 0.00001858 0.01786173 + layer.4.k_cache 0.00063020 0.41739704 + layer.4.v_cache 0.00005223 0.04005766 + layer.4.output 0.17536174 645.31212798 + ------------------------------------------------------------------------------------- + TOTAL 0.08265493 276.69369682 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 66744 +BPFP 0.7303 bits/point +EBPFP 1.4606 equivalent bits/point +MSE 276.693697 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 276.6937 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,880B, BPFP=0.3456 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,444B, BPFP=1.7360 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,588B, BPFP=0.6596 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,712B, BPFP=1.6015 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,992B, BPFP=0.7338 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,220B, BPFP=1.5110 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,460B, BPFP=0.6360 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,468B, BPFP=1.5566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,532B, BPFP=1.2007 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,376B, BPFP=1.5397 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,796B, BPFP=0.1785 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11218034 164.40799632 + layer.0.v_cache 0.00001438 0.01475445 + layer.1.k_cache 0.03157643 16.20251608 + layer.1.v_cache 0.00000545 0.00586344 + layer.2.k_cache 0.00802605 2.03575045 + layer.2.v_cache 0.00001826 0.01625041 + layer.3.k_cache 0.04519741 8.66556899 + layer.3.v_cache 0.00001794 0.01699568 + layer.4.k_cache 0.00061137 0.39510624 + layer.4.v_cache 0.00005913 0.03914746 + layer.4.output 0.17010492 638.02830882 + ------------------------------------------------------------------------------------- + TOTAL 0.08167301 273.99988890 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 69468 +BPFP 0.7512 bits/point +EBPFP 1.5023 equivalent bits/point +MSE 273.999889 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.003s, Pack+Encode: 0.165s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 273.9999 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,912B, BPFP=0.3474 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,356B, BPFP=1.6999 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,528B, BPFP=0.6410 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,892B, BPFP=1.6156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,992B, BPFP=0.7253 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,260B, BPFP=1.5007 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,476B, BPFP=0.6315 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,396B, BPFP=1.5254 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,616B, BPFP=1.2020 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,400B, BPFP=1.5262 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,892B, BPFP=0.1789 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10146660 162.04413835 + layer.0.v_cache 0.00001344 0.01425005 + layer.1.k_cache 0.03090366 16.35558798 + layer.1.v_cache 0.00000566 0.00576822 + layer.2.k_cache 0.01188999 1.76408848 + layer.2.v_cache 0.00001845 0.01593103 + layer.3.k_cache 0.02542871 8.21101841 + layer.3.v_cache 0.00001884 0.01758702 + layer.4.k_cache 0.00060243 0.39484636 + layer.4.v_cache 0.00005224 0.03915831 + layer.4.output 0.16839122 630.62878945 + ------------------------------------------------------------------------------------- + TOTAL 0.07936109 270.78022943 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 69720 +BPFP 0.7451 bits/point +EBPFP 1.4903 equivalent bits/point +MSE 270.780229 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 270.7802 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,972B, BPFP=0.3349 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,020B, BPFP=1.7018 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,952B, BPFP=0.6712 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,728B, BPFP=1.6522 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,616B, BPFP=0.7840 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,368B, BPFP=1.5910 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,020B, BPFP=0.6827 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,752B, BPFP=1.6562 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,476B, BPFP=1.2697 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,868B, BPFP=1.5061 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,152B, BPFP=0.1735 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12981709 169.39672852 + layer.0.v_cache 0.00001628 0.01492722 + layer.1.k_cache 0.06761242 16.10616800 + layer.1.v_cache 0.00000547 0.00530898 + layer.2.k_cache 0.00384689 1.83688255 + layer.2.v_cache 0.00001812 0.01533931 + layer.3.k_cache 0.05327560 7.30186661 + layer.3.v_cache 0.00001896 0.01700432 + layer.4.k_cache 0.00060800 0.38268268 + layer.4.v_cache 0.00005116 0.03716358 + layer.4.output 0.15518735 589.72107919 + ------------------------------------------------------------------------------------- + TOTAL 0.07891655 254.30362507 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 76924 +BPFP 0.7685 bits/point +EBPFP 1.5370 equivalent bits/point +MSE 254.303625 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 254.3036 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,812B, BPFP=0.3495 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,348B, BPFP=1.8032 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,272B, BPFP=0.6312 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,176B, BPFP=1.7701 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,756B, BPFP=0.7245 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,144B, BPFP=1.5710 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,180B, BPFP=0.6134 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,556B, BPFP=1.6505 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,540B, BPFP=1.2616 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,180B, BPFP=1.5779 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,516B, BPFP=0.1796 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887507 170.00421971 + layer.0.v_cache 0.00001345 0.01505202 + layer.1.k_cache 0.03259810 17.18093834 + layer.1.v_cache 0.00000523 0.00565896 + layer.2.k_cache 0.00244350 1.98153800 + layer.2.v_cache 0.00001834 0.01535936 + layer.3.k_cache 0.10864470 7.72814866 + layer.3.v_cache 0.00001904 0.01656980 + layer.4.k_cache 0.00060864 0.40560339 + layer.4.v_cache 0.00004940 0.03810144 + layer.4.output 0.18628448 669.71208113 + ------------------------------------------------------------------------------------- + TOTAL 0.09395687 287.37504456 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 68480 +BPFP 0.7771 bits/point +EBPFP 1.5541 equivalent bits/point +MSE 287.375045 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 287.3750 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,928B, BPFP=0.3385 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,004B, BPFP=1.7563 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,572B, BPFP=0.6271 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,440B, BPFP=1.6573 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,204B, BPFP=0.7381 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,500B, BPFP=1.4923 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,668B, BPFP=0.6440 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,580B, BPFP=1.5063 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,748B, BPFP=1.1847 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,440B, BPFP=1.4817 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,140B, BPFP=0.1791 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422350 165.85518346 + layer.0.v_cache 0.00001493 0.01489982 + layer.1.k_cache 0.01422887 16.36784586 + layer.1.v_cache 0.00000595 0.00539095 + layer.2.k_cache 0.00668612 1.94072509 + layer.2.v_cache 0.00001902 0.01492886 + layer.3.k_cache 0.02497406 7.68386395 + layer.3.v_cache 0.00001869 0.01671153 + layer.4.k_cache 0.00061935 0.37239700 + layer.4.v_cache 0.00005212 0.03643210 + layer.4.output 0.16899524 609.25642055 + ------------------------------------------------------------------------------------- + TOTAL 0.07904761 262.18254838 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 72224 +BPFP 0.7459 bits/point +EBPFP 1.4917 equivalent bits/point +MSE 262.182548 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 262.1825 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 113, 128) +Output shape: (1, 113, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.output: torch.Size([1, 113, 3584]) -> torch.Size([1, 1, 113, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,136B, BPFP=0.2954 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,584B, BPFP=1.4635 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,136B, BPFP=0.5719 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,236B, BPFP=1.4154 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,372B, BPFP=0.6045 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,952B, BPFP=1.3761 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,168B, BPFP=0.5763 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,884B, BPFP=1.3667 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,476B, BPFP=1.0337 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,064B, BPFP=1.2533 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,604B, BPFP=0.1502 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10214825 161.00739768 + layer.0.v_cache 0.00001637 0.01425497 + layer.1.k_cache 0.04284562 16.44069327 + layer.1.v_cache 0.00000578 0.00590114 + layer.2.k_cache 0.00913706 1.87929110 + layer.2.v_cache 0.00001928 0.01569405 + layer.3.k_cache 0.05585717 7.37327265 + layer.3.v_cache 0.00001870 0.01705014 + layer.4.k_cache 0.00065968 0.41298405 + layer.4.v_cache 0.00005753 0.03793703 + layer.4.output 10.09706179 475.44346555 + ------------------------------------------------------------------------------------- + TOTAL 4.17001165 206.78286676 + (elements=983,552) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 983552 +Total Bytes 79612 +BPFP 0.6475 bits/point +EBPFP 1.2951 equivalent bits/point +MSE 206.782867 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 206.7829 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,964B, BPFP=0.3372 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,960B, BPFP=1.7102 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,704B, BPFP=0.6360 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,436B, BPFP=1.6202 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,504B, BPFP=0.7734 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,912B, BPFP=1.5302 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,748B, BPFP=0.6435 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,140B, BPFP=1.5694 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,104B, BPFP=1.2198 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,584B, BPFP=1.4739 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,352B, BPFP=0.1803 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11481400 172.05500944 + layer.0.v_cache 0.00001396 0.01504814 + layer.1.k_cache 0.03135788 15.93377484 + layer.1.v_cache 0.00000545 0.00576499 + layer.2.k_cache 0.00370715 1.85303506 + layer.2.v_cache 0.00001965 0.01563453 + layer.3.k_cache 0.03930106 8.13148113 + layer.3.v_cache 0.00001886 0.01772089 + layer.4.k_cache 0.00060819 0.38598499 + layer.4.v_cache 0.00005437 0.03778697 + layer.4.output 0.16591480 595.80945840 + ------------------------------------------------------------------------------------- + TOTAL 0.07948848 257.00690881 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 74408 +BPFP 0.7515 bits/point +EBPFP 1.5031 equivalent bits/point +MSE 257.006909 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 257.0069 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,924B, BPFP=0.3378 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,728B, BPFP=1.7079 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,480B, BPFP=0.6110 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,044B, BPFP=1.5878 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,892B, BPFP=0.6833 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,360B, BPFP=1.4677 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,332B, BPFP=0.5850 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,568B, BPFP=1.5042 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,540B, BPFP=1.1482 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,096B, BPFP=1.4213 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,684B, BPFP=0.1676 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09788724 166.95307014 + layer.0.v_cache 0.00001308 0.01444872 + layer.1.k_cache 0.03299916 16.74938416 + layer.1.v_cache 0.00000544 0.00585344 + layer.2.k_cache 0.00367894 1.82417829 + layer.2.v_cache 0.00002012 0.01666615 + layer.3.k_cache 0.03939044 7.94181292 + layer.3.v_cache 0.00001875 0.01708817 + layer.4.k_cache 0.00061788 0.39527666 + layer.4.v_cache 0.00005010 0.03861331 + layer.4.output 0.17021053 609.23299559 + ------------------------------------------------------------------------------------- + TOTAL 0.08036205 262.26984477 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 69648 +BPFP 0.7193 bits/point +EBPFP 1.4385 equivalent bits/point +MSE 262.269845 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 262.2698 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,112B, BPFP=0.3056 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,536B, BPFP=1.5243 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,252B, BPFP=0.6152 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,352B, BPFP=1.4977 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,324B, BPFP=0.7703 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,036B, BPFP=1.4520 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,684B, BPFP=0.6777 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,640B, BPFP=1.5394 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,096B, BPFP=1.1713 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,584B, BPFP=1.3866 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,128B, BPFP=0.1887 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12672405 162.04649523 + layer.0.v_cache 0.00001401 0.01443270 + layer.1.k_cache 0.12044146 16.60778583 + layer.1.v_cache 0.00000568 0.00552914 + layer.2.k_cache 0.00570256 1.92598527 + layer.2.v_cache 0.00001866 0.01544018 + layer.3.k_cache 0.03611955 7.46528399 + layer.3.v_cache 0.00001983 0.01768122 + layer.4.k_cache 0.00066233 0.37494921 + layer.4.v_cache 0.00005331 0.03770330 + layer.4.output 10.56037249 497.18654927 + ------------------------------------------------------------------------------------- + TOTAL 4.36543346 215.81277241 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 84744 +BPFP 0.7212 bits/point +EBPFP 1.4424 equivalent bits/point +MSE 215.812772 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 215.8128 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,932B, BPFP=0.3430 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,736B, BPFP=1.7287 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,464B, BPFP=0.6151 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,456B, BPFP=1.6790 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,792B, BPFP=0.6733 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,648B, BPFP=1.5355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,208B, BPFP=0.5696 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,844B, BPFP=1.5703 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,512B, BPFP=1.1562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,536B, BPFP=1.5156 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,644B, BPFP=0.1685 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13733016 163.25113192 + layer.0.v_cache 0.00001373 0.01535540 + layer.1.k_cache 0.03155638 16.51162026 + layer.1.v_cache 0.00000617 0.00647060 + layer.2.k_cache 0.00751131 1.66279377 + layer.2.v_cache 0.00001754 0.01692508 + layer.3.k_cache 0.02548880 7.87553544 + layer.3.v_cache 0.00001765 0.01839862 + layer.4.k_cache 0.00061011 0.38999610 + layer.4.v_cache 0.00005045 0.03988394 + layer.4.output 0.16239834 616.55585430 + ------------------------------------------------------------------------------------- + TOTAL 0.07878769 265.03994654 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 70772 +BPFP 0.7392 bits/point +EBPFP 1.4784 equivalent bits/point +MSE 265.039947 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 265.0399 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,804B, BPFP=0.3568 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,300B, BPFP=1.8394 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,376B, BPFP=0.6677 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,732B, BPFP=1.7271 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,720B, BPFP=0.7358 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,668B, BPFP=1.5166 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,316B, BPFP=0.6559 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,124B, BPFP=1.6068 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,260B, BPFP=1.2381 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,968B, BPFP=1.5759 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,676B, BPFP=0.1886 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10979201 159.78424150 + layer.0.v_cache 0.00001339 0.01454900 + layer.1.k_cache 0.01160004 16.85130909 + layer.1.v_cache 0.00000567 0.00573082 + layer.2.k_cache 0.00714037 1.84204720 + layer.2.v_cache 0.00001829 0.01513684 + layer.3.k_cache 0.04622446 7.56836865 + layer.3.v_cache 0.00001896 0.01671171 + layer.4.k_cache 0.00064378 0.39428240 + layer.4.v_cache 0.00005251 0.03742963 + layer.4.output 0.18415536 686.13455018 + ------------------------------------------------------------------------------------- + TOTAL 0.08615276 293.49833283 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 66944 +BPFP 0.7789 bits/point +EBPFP 1.5577 equivalent bits/point +MSE 293.498333 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 293.4983 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,748B, BPFP=0.3547 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,040B, BPFP=1.8344 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,212B, BPFP=0.6518 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,064B, BPFP=1.6364 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,680B, BPFP=0.7468 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,348B, BPFP=1.4911 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,136B, BPFP=0.6364 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,580B, BPFP=1.5381 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,748B, BPFP=1.1664 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,520B, BPFP=1.5260 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,908B, BPFP=0.2003 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11290005 172.38045353 + layer.0.v_cache 0.00001570 0.01449939 + layer.1.k_cache 0.03434501 16.96443156 + layer.1.v_cache 0.00000534 0.00532308 + layer.2.k_cache 0.00238675 1.90111789 + layer.2.v_cache 0.00001838 0.01559759 + layer.3.k_cache 0.02838124 7.65950547 + layer.3.v_cache 0.00002047 0.01690189 + layer.4.k_cache 0.00062757 0.40416943 + layer.4.v_cache 0.00005455 0.03924373 + layer.4.output 0.18420308 704.09537338 + ------------------------------------------------------------------------------------- + TOTAL 0.08636333 301.65110925 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 63984 +BPFP 0.7638 bits/point +EBPFP 1.5275 equivalent bits/point +MSE 301.651109 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 301.6511 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,768B, BPFP=0.3542 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,120B, BPFP=1.8269 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,312B, BPFP=0.6635 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,636B, BPFP=1.7300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,576B, BPFP=0.7163 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,756B, BPFP=1.5537 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,204B, BPFP=0.6418 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,908B, BPFP=1.5841 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,848B, BPFP=1.1715 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,844B, BPFP=1.5713 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,804B, BPFP=0.1947 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09752406 165.16684195 + layer.0.v_cache 0.00001393 0.01511998 + layer.1.k_cache 0.01493649 16.69805752 + layer.1.v_cache 0.00000588 0.00609642 + layer.2.k_cache 0.01003601 1.91462864 + layer.2.v_cache 0.00001805 0.01574931 + layer.3.k_cache 0.04451986 6.55181181 + layer.3.v_cache 0.00001914 0.01765416 + layer.4.k_cache 0.00061869 0.41198197 + layer.4.v_cache 0.00005251 0.03927476 + layer.4.output 0.18409089 694.59128892 + ------------------------------------------------------------------------------------- + TOTAL 0.08566946 297.23389641 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 65776 +BPFP 0.7751 bits/point +EBPFP 1.5502 equivalent bits/point +MSE 297.233896 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 297.2339 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,868B, BPFP=0.3517 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,304B, BPFP=1.7515 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,376B, BPFP=0.6355 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,488B, BPFP=1.5979 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,724B, BPFP=0.7011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,764B, BPFP=1.4616 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,376B, BPFP=0.6355 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,884B, BPFP=1.4842 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,736B, BPFP=1.0798 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,588B, BPFP=1.4285 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,816B, BPFP=0.1833 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13658243 162.29875753 + layer.0.v_cache 0.00001379 0.01527581 + layer.1.k_cache 0.01234693 16.61468138 + layer.1.v_cache 0.00000566 0.00599587 + layer.2.k_cache 0.00375735 1.89873716 + layer.2.v_cache 0.00001852 0.01568543 + layer.3.k_cache 0.02700986 8.80075404 + layer.3.v_cache 0.00001930 0.01777921 + layer.4.k_cache 0.00062007 0.40192662 + layer.4.v_cache 0.00005146 0.03799439 + layer.4.output 0.17614663 652.88333692 + ------------------------------------------------------------------------------------- + TOTAL 0.08314423 280.01711446 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 65924 +BPFP 0.7300 bits/point +EBPFP 1.4600 equivalent bits/point +MSE 280.017114 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 280.0171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,720B, BPFP=0.3682 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,312B, BPFP=1.7791 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,072B, BPFP=0.6575 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,568B, BPFP=1.6199 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,348B, BPFP=0.7166 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,948B, BPFP=1.4872 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,784B, BPFP=0.5959 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,984B, BPFP=1.4949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,532B, BPFP=1.1841 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,124B, BPFP=1.5248 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,728B, BPFP=0.2057 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08738674 169.12239137 + layer.0.v_cache 0.00001368 0.01540136 + layer.1.k_cache 0.01286346 16.40344071 + layer.1.v_cache 0.00000597 0.00597913 + layer.2.k_cache 0.00251921 1.85765933 + layer.2.v_cache 0.00001769 0.01605603 + layer.3.k_cache 0.08324167 7.23116355 + layer.3.v_cache 0.00002021 0.01721783 + layer.4.k_cache 0.00062021 0.40691522 + layer.4.v_cache 0.00005375 0.04043020 + layer.4.output 0.19567458 743.23104207 + ------------------------------------------------------------------------------------- + TOTAL 0.09155674 317.51376172 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 60120 +BPFP 0.7570 bits/point +EBPFP 1.5139 equivalent bits/point +MSE 317.513762 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 317.5138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,872B, BPFP=0.3524 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,304B, BPFP=1.7515 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,312B, BPFP=0.6235 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,176B, BPFP=1.7274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,692B, BPFP=0.6950 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,152B, BPFP=1.5346 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,268B, BPFP=0.6152 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,180B, BPFP=1.5399 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,904B, BPFP=1.1114 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,972B, BPFP=1.5008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,996B, BPFP=0.1881 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12176956 155.91295651 + layer.0.v_cache 0.00001396 0.01498957 + layer.1.k_cache 0.01120206 16.07863387 + layer.1.v_cache 0.00000562 0.00567423 + layer.2.k_cache 0.00678912 1.96652167 + layer.2.v_cache 0.00001849 0.01564855 + layer.3.k_cache 0.02622183 8.33621400 + layer.3.v_cache 0.00001965 0.01757330 + layer.4.k_cache 0.00062413 0.40732556 + layer.4.v_cache 0.00005248 0.03796465 + layer.4.output 0.17940697 653.14936532 + ------------------------------------------------------------------------------------- + TOTAL 0.08368034 279.69641524 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 67828 +BPFP 0.7511 bits/point +EBPFP 1.5022 equivalent bits/point +MSE 279.696415 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 279.6964 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,092B, BPFP=0.3113 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,132B, BPFP=1.6565 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,188B, BPFP=0.6232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,772B, BPFP=1.6030 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,288B, BPFP=0.7869 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,440B, BPFP=1.5536 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,488B, BPFP=0.6679 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,628B, BPFP=1.5815 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,220B, BPFP=1.2232 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,564B, BPFP=1.4232 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,632B, BPFP=0.1835 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120816 155.29676339 + layer.0.v_cache 0.00001607 0.01438695 + layer.1.k_cache 0.10678518 16.62713565 + layer.1.v_cache 0.00000585 0.00545720 + layer.2.k_cache 0.00813519 1.83494030 + layer.2.v_cache 0.00001919 0.01575771 + layer.3.k_cache 0.06560528 7.52299514 + layer.3.v_cache 0.00001973 0.01769048 + layer.4.k_cache 0.00062300 0.38483694 + layer.4.v_cache 0.00005287 0.03771145 + layer.4.output 10.86395687 511.62551020 + ------------------------------------------------------------------------------------- + TOTAL 4.49059815 221.36095568 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 85444 +BPFP 0.7479 bits/point +EBPFP 1.4959 equivalent bits/point +MSE 221.360956 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.006s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 221.3610 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,908B, BPFP=0.3427 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,476B, BPFP=1.7019 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,508B, BPFP=0.6300 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,960B, BPFP=1.6092 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,932B, BPFP=0.7062 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,120B, BPFP=1.4583 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,396B, BPFP=0.6099 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,232B, BPFP=1.4784 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,284B, BPFP=1.1286 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,980B, BPFP=1.4332 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,060B, BPFP=0.1811 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12848519 171.14801096 + layer.0.v_cache 0.00001371 0.01441577 + layer.1.k_cache 0.03007136 16.40754647 + layer.1.v_cache 0.00000570 0.00597115 + layer.2.k_cache 0.00936044 1.82180120 + layer.2.v_cache 0.00001842 0.01587462 + layer.3.k_cache 0.04065019 7.22790668 + layer.3.v_cache 0.00001877 0.01815916 + layer.4.k_cache 0.00062204 0.38766401 + layer.4.v_cache 0.00010837 0.03766444 + layer.4.output 0.16426703 623.74184113 + ------------------------------------------------------------------------------------- + TOTAL 0.07995432 268.42811190 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 68856 +BPFP 0.7274 bits/point +EBPFP 1.4549 equivalent bits/point +MSE 268.428112 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 268.4281 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,796B, BPFP=0.3508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,592B, BPFP=1.8734 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,420B, BPFP=0.6680 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,080B, BPFP=1.7734 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,844B, BPFP=0.7508 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,396B, BPFP=1.6398 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,360B, BPFP=0.6562 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,684B, BPFP=1.6961 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,752B, BPFP=1.3188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,444B, BPFP=1.6492 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,816B, BPFP=0.1902 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09257430 157.96778564 + layer.0.v_cache 0.00001329 0.01386579 + layer.1.k_cache 0.01202901 16.83247681 + layer.1.v_cache 0.00000544 0.00518077 + layer.2.k_cache 0.00549439 2.12454033 + layer.2.v_cache 0.00001849 0.01529317 + layer.3.k_cache 0.11014032 9.06759338 + layer.3.v_cache 0.00001916 0.01626897 + layer.4.k_cache 0.00063444 0.37457414 + layer.4.v_cache 0.00005494 0.03513484 + layer.4.output 0.19143520 677.67857143 + ------------------------------------------------------------------------------------- + TOTAL 0.09182531 290.01192434 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 70184 +BPFP 0.8063 bits/point +EBPFP 1.6127 equivalent bits/point +MSE 290.011924 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 290.0119 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,864B, BPFP=0.3467 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,048B, BPFP=1.6830 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,348B, BPFP=0.6228 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,668B, BPFP=1.6124 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,588B, BPFP=0.6674 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,808B, BPFP=1.4524 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,256B, BPFP=0.6057 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,220B, BPFP=1.5290 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,128B, BPFP=1.1399 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,780B, BPFP=1.4472 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,568B, BPFP=0.1745 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12365316 165.35021391 + layer.0.v_cache 0.00001332 0.01496309 + layer.1.k_cache 0.03511774 16.48085531 + layer.1.v_cache 0.00000569 0.00615391 + layer.2.k_cache 0.01236608 2.11082313 + layer.2.v_cache 0.00001835 0.01562627 + layer.3.k_cache 0.02851222 7.49425398 + layer.3.v_cache 0.00001844 0.01758162 + layer.4.k_cache 0.00063207 0.39649982 + layer.4.v_cache 0.00005066 0.03941189 + layer.4.output 0.17964420 645.26333971 + ------------------------------------------------------------------------------------- + TOTAL 0.08575865 276.98645652 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 66276 +BPFP 0.7252 bits/point +EBPFP 1.4504 equivalent bits/point +MSE 276.986457 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 276.9865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,764B, BPFP=0.3534 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,940B, BPFP=1.7909 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,288B, BPFP=0.6587 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,464B, BPFP=1.6955 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,476B, BPFP=0.6963 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,448B, BPFP=1.4920 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,132B, BPFP=0.6274 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,656B, BPFP=1.5337 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,948B, BPFP=1.1915 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,440B, BPFP=1.4904 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,748B, BPFP=0.1931 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12096892 170.36033904 + layer.0.v_cache 0.00001393 0.01488512 + layer.1.k_cache 0.03247837 16.80906638 + layer.1.v_cache 0.00000613 0.00592877 + layer.2.k_cache 0.00394396 1.79508659 + layer.2.v_cache 0.00001920 0.01631257 + layer.3.k_cache 0.02751121 6.81388112 + layer.3.v_cache 0.00001919 0.01800765 + layer.4.k_cache 0.00062481 0.40199480 + layer.4.v_cache 0.00005293 0.03834511 + layer.4.output 0.18564256 694.81501832 + ------------------------------------------------------------------------------------- + TOTAL 0.08736098 297.64582208 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 64304 +BPFP 0.7577 bits/point +EBPFP 1.5155 equivalent bits/point +MSE 297.645822 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 297.6458 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,776B, BPFP=0.3604 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,412B, BPFP=1.9099 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,244B, BPFP=0.6583 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,264B, BPFP=1.6769 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,508B, BPFP=0.7119 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,584B, BPFP=1.5390 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,076B, BPFP=0.6242 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,952B, BPFP=1.6136 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,084B, BPFP=1.2346 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,768B, BPFP=1.5763 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,976B, BPFP=0.2022 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08564486 173.72876928 + layer.0.v_cache 0.00001408 0.01542837 + layer.1.k_cache 0.03507188 16.78759449 + layer.1.v_cache 0.00000566 0.00596272 + layer.2.k_cache 0.00239554 1.85310126 + layer.2.v_cache 0.00001938 0.01665436 + layer.3.k_cache 0.03313774 6.91781022 + layer.3.v_cache 0.00001993 0.01903637 + layer.4.k_cache 0.00061714 0.42587033 + layer.4.v_cache 0.00005134 0.04137631 + layer.4.output 0.18226704 703.93233998 + ------------------------------------------------------------------------------------- + TOTAL 0.08428511 301.60811668 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 65644 +BPFP 0.7836 bits/point +EBPFP 1.5671 equivalent bits/point +MSE 301.608117 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 301.6081 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 117, 128) +Output shape: (1, 117, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.output: torch.Size([1, 117, 3584]) -> torch.Size([1, 1, 117, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,184B, BPFP=0.2917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,864B, BPFP=1.4509 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,224B, BPFP=0.5641 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,752B, BPFP=1.4359 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,872B, BPFP=0.6506 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,092B, BPFP=1.3478 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,548B, BPFP=0.6074 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,268B, BPFP=1.3713 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,752B, BPFP=1.0353 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,432B, BPFP=1.2596 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,272B, BPFP=0.1578 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11900647 150.62730369 + layer.0.v_cache 0.00001428 0.01485758 + layer.1.k_cache 0.12535267 15.90103019 + layer.1.v_cache 0.00000579 0.00573740 + layer.2.k_cache 0.00517510 1.81890126 + layer.2.v_cache 0.00001992 0.01536583 + layer.3.k_cache 0.09315677 7.79394270 + layer.3.v_cache 0.00001978 0.01674786 + layer.4.k_cache 0.00061896 0.37396576 + layer.4.v_cache 0.00005133 0.03487862 + layer.4.output 9.75369281 458.99793956 + ------------------------------------------------------------------------------------- + TOTAL 4.03642769 199.38754752 + (elements=1,018,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1018368 +Total Bytes 83260 +BPFP 0.6541 bits/point +EBPFP 1.3081 equivalent bits/point +MSE 199.387548 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.006s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 199.3875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,088B, BPFP=0.3107 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,624B, BPFP=1.5810 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,228B, BPFP=0.6292 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,608B, BPFP=1.5786 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,664B, BPFP=0.8429 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,384B, BPFP=1.5452 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,524B, BPFP=0.6732 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,000B, BPFP=1.4881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,288B, BPFP=1.2333 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,488B, BPFP=1.4119 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,592B, BPFP=0.1614 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14131550 158.20100446 + layer.0.v_cache 0.00001383 0.01458800 + layer.1.k_cache 0.02978072 16.31773391 + layer.1.v_cache 0.00000537 0.00537188 + layer.2.k_cache 0.00578868 1.71387474 + layer.2.v_cache 0.00001872 0.01475781 + layer.3.k_cache 0.04020810 8.00209322 + layer.3.v_cache 0.00001850 0.01614717 + layer.4.k_cache 0.00063358 0.38669543 + layer.4.v_cache 0.00005294 0.03787027 + layer.4.output 10.86925024 511.84204932 + ------------------------------------------------------------------------------------- + TOTAL 4.48838751 221.62379307 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 83488 +BPFP 0.7308 bits/point +EBPFP 1.4616 equivalent bits/point +MSE 221.623793 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 221.6238 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,876B, BPFP=0.3490 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,452B, BPFP=1.7582 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,392B, BPFP=0.6310 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,880B, BPFP=1.6518 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,840B, BPFP=0.7143 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,364B, BPFP=1.5558 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,512B, BPFP=0.6533 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,588B, BPFP=1.5975 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,320B, BPFP=1.1756 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,304B, BPFP=1.5446 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,616B, BPFP=0.1758 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09514355 164.87064035 + layer.0.v_cache 0.00001423 0.01514170 + layer.1.k_cache 0.03283414 16.38508969 + layer.1.v_cache 0.00000553 0.00569332 + layer.2.k_cache 0.00489674 1.90418752 + layer.2.v_cache 0.00001883 0.01551198 + layer.3.k_cache 0.07199075 7.60082862 + layer.3.v_cache 0.00001932 0.01692634 + layer.4.k_cache 0.00061592 0.39194130 + layer.4.v_cache 0.00006816 0.03755408 + layer.4.output 0.17253765 644.66198980 + ------------------------------------------------------------------------------------- + TOTAL 0.08313946 276.69867314 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 69144 +BPFP 0.7566 bits/point +EBPFP 1.5131 equivalent bits/point +MSE 276.698673 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 276.6987 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,900B, BPFP=0.3412 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,568B, BPFP=1.7184 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,480B, BPFP=0.6250 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,992B, BPFP=1.6149 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,052B, BPFP=0.7277 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,320B, BPFP=1.4943 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,424B, BPFP=0.6149 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,544B, BPFP=1.5345 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,624B, BPFP=1.1897 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,324B, BPFP=1.4950 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,656B, BPFP=0.1708 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11660385 171.17257094 + layer.0.v_cache 0.00001381 0.01511625 + layer.1.k_cache 0.05175667 15.91985312 + layer.1.v_cache 0.00000569 0.00588532 + layer.2.k_cache 0.00227456 1.77106537 + layer.2.v_cache 0.00001743 0.01513664 + layer.3.k_cache 0.02781237 7.74072616 + layer.3.v_cache 0.00001883 0.01684398 + layer.4.k_cache 0.00061846 0.38030032 + layer.4.v_cache 0.00005344 0.03720133 + layer.4.output 0.17410791 623.25148810 + ------------------------------------------------------------------------------------- + TOTAL 0.08340767 268.22559507 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 69884 +BPFP 0.7383 bits/point +EBPFP 1.4766 equivalent bits/point +MSE 268.225595 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 268.2256 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,828B, BPFP=0.3441 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,108B, BPFP=1.7146 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,296B, BPFP=0.6205 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,640B, BPFP=1.6265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,724B, BPFP=0.7011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,780B, BPFP=1.4646 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,304B, BPFP=0.6220 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,900B, BPFP=1.4872 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,976B, BPFP=1.1250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,748B, BPFP=1.4586 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,744B, BPFP=0.1814 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212080 166.33607869 + layer.0.v_cache 0.00001360 0.01455595 + layer.1.k_cache 0.03288655 16.52899685 + layer.1.v_cache 0.00000536 0.00576683 + layer.2.k_cache 0.00973156 1.98957880 + layer.2.v_cache 0.00001762 0.01533212 + layer.3.k_cache 0.02906220 7.96690718 + layer.3.v_cache 0.00001891 0.01712203 + layer.4.k_cache 0.00063934 0.40946335 + layer.4.v_cache 0.00004866 0.03830392 + layer.4.output 0.18425958 652.94739673 + ------------------------------------------------------------------------------------- + TOTAL 0.08496245 280.23258134 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 66048 +BPFP 0.7314 bits/point +EBPFP 1.4628 equivalent bits/point +MSE 280.232581 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 280.2326 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,888B, BPFP=0.3471 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,704B, BPFP=1.7838 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,572B, BPFP=0.6566 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,100B, BPFP=1.6728 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,012B, BPFP=0.7375 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,412B, BPFP=1.5463 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,564B, BPFP=0.6551 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,868B, BPFP=1.6301 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,580B, BPFP=1.2096 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,360B, BPFP=1.5368 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,704B, BPFP=0.1761 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10758395 162.51246553 + layer.0.v_cache 0.00001397 0.01494083 + layer.1.k_cache 0.03293324 16.17419577 + layer.1.v_cache 0.00000575 0.00569772 + layer.2.k_cache 0.00823421 1.99419574 + layer.2.v_cache 0.00001745 0.01508375 + layer.3.k_cache 0.04116619 8.10522246 + layer.3.v_cache 0.00001985 0.01756604 + layer.4.k_cache 0.00062779 0.39298033 + layer.4.v_cache 0.00005042 0.03716705 + layer.4.output 0.16954060 637.83755252 + ------------------------------------------------------------------------------------- + TOTAL 0.08102571 273.77249311 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 70764 +BPFP 0.7652 bits/point +EBPFP 1.5304 equivalent bits/point +MSE 273.772493 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 273.7725 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,916B, BPFP=0.3326 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,764B, BPFP=1.6951 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,620B, BPFP=0.6285 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,508B, BPFP=1.6507 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,228B, BPFP=0.7340 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,744B, BPFP=1.5181 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,732B, BPFP=0.6479 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,228B, BPFP=1.6021 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,772B, BPFP=1.1757 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,580B, BPFP=1.4896 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,984B, BPFP=0.1732 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12307767 173.07616102 + layer.0.v_cache 0.00001410 0.01494293 + layer.1.k_cache 0.03088827 16.78237305 + layer.1.v_cache 0.00000584 0.00604570 + layer.2.k_cache 0.00371391 1.98358748 + layer.2.v_cache 0.00001950 0.01615175 + layer.3.k_cache 0.03079298 7.51094699 + layer.3.v_cache 0.00001950 0.01744108 + layer.4.k_cache 0.00063014 0.38387977 + layer.4.v_cache 0.00005282 0.03770549 + layer.4.output 0.15767360 602.68100198 + ------------------------------------------------------------------------------------- + TOTAL 0.07605470 259.91742642 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 73076 +BPFP 0.7463 bits/point +EBPFP 1.4926 equivalent bits/point +MSE 259.917426 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 259.9174 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3430 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,028B, BPFP=1.7203 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,288B, BPFP=0.6265 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,680B, BPFP=1.6540 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,696B, BPFP=0.7043 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,800B, BPFP=1.4863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,376B, BPFP=0.6433 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,796B, BPFP=1.4855 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,020B, BPFP=1.1471 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,932B, BPFP=1.5114 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,844B, BPFP=0.1863 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10295349 164.75242950 + layer.0.v_cache 0.00001348 0.01402438 + layer.1.k_cache 0.03334363 16.64663622 + layer.1.v_cache 0.00000550 0.00568962 + layer.2.k_cache 0.01126997 2.20533901 + layer.2.v_cache 0.00001933 0.01618274 + layer.3.k_cache 0.09131574 7.91734202 + layer.3.v_cache 0.00001769 0.01699678 + layer.4.k_cache 0.00062275 0.39987359 + layer.4.v_cache 0.00006541 0.03751879 + layer.4.output 0.17512874 661.35246080 + ------------------------------------------------------------------------------------- + TOTAL 0.08620754 283.61642696 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 66260 +BPFP 0.7427 bits/point +EBPFP 1.4854 equivalent bits/point +MSE 283.616427 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 283.6164 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,916B, BPFP=0.3441 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,616B, BPFP=1.7270 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,496B, BPFP=0.6279 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,212B, BPFP=1.6545 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,976B, BPFP=0.7141 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,428B, BPFP=1.5136 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,516B, BPFP=0.6315 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,660B, BPFP=1.5553 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,880B, BPFP=1.2356 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,524B, BPFP=1.5309 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,680B, BPFP=0.1714 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12314748 168.35535650 + layer.0.v_cache 0.00001636 0.01465429 + layer.1.k_cache 0.01190477 16.52037592 + layer.1.v_cache 0.00000588 0.00595914 + layer.2.k_cache 0.01614088 1.94181701 + layer.2.v_cache 0.00001834 0.01480610 + layer.3.k_cache 0.05604688 7.40872368 + layer.3.v_cache 0.00001882 0.01737348 + layer.4.k_cache 0.00063774 0.36378082 + layer.4.v_cache 0.00005162 0.03413249 + layer.4.output 0.17348555 623.56660509 + ------------------------------------------------------------------------------------- + TOTAL 0.08366986 268.21430677 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 70904 +BPFP 0.7491 bits/point +EBPFP 1.4981 equivalent bits/point +MSE 268.214307 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 268.2143 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,812B, BPFP=0.3495 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,348B, BPFP=1.8032 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,348B, BPFP=0.6458 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,744B, BPFP=1.6867 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,736B, BPFP=0.7207 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,924B, BPFP=1.5285 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,276B, BPFP=0.6319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,924B, BPFP=1.5285 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,136B, BPFP=1.1836 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,972B, BPFP=1.5378 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,904B, BPFP=0.1903 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11560643 166.54861111 + layer.0.v_cache 0.00001402 0.01518789 + layer.1.k_cache 0.03272506 16.92646545 + layer.1.v_cache 0.00000563 0.00600652 + layer.2.k_cache 0.00645547 1.87263790 + layer.2.v_cache 0.00001830 0.01687566 + layer.3.k_cache 0.02731901 8.50651946 + layer.3.v_cache 0.00001960 0.01769087 + layer.4.k_cache 0.00061302 0.40042364 + layer.4.v_cache 0.00004978 0.03857456 + layer.4.output 0.17251868 669.45601852 + ------------------------------------------------------------------------------------- + TOTAL 0.08179159 287.09065428 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 67124 +BPFP 0.7617 bits/point +EBPFP 1.5233 equivalent bits/point +MSE 287.090654 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 287.0907 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,944B, BPFP=0.3452 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,708B, BPFP=1.7237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,476B, BPFP=0.6172 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,020B, BPFP=1.6016 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,896B, BPFP=0.6918 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,244B, BPFP=1.4638 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,592B, BPFP=0.6378 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,440B, BPFP=1.4986 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,436B, BPFP=1.1428 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,112B, BPFP=1.4403 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,220B, BPFP=0.1831 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11690623 163.82690430 + layer.0.v_cache 0.00001418 0.01455447 + layer.1.k_cache 0.03305150 16.08714433 + layer.1.v_cache 0.00000545 0.00573377 + layer.2.k_cache 0.01095151 1.90052709 + layer.2.v_cache 0.00001854 0.01521347 + layer.3.k_cache 0.04293454 7.17404175 + layer.3.v_cache 0.00002007 0.01782805 + layer.4.k_cache 0.00062489 0.38370145 + layer.4.v_cache 0.00005451 0.03630530 + layer.4.output 0.16672201 616.43293425 + ------------------------------------------------------------------------------------- + TOTAL 0.08068444 264.97014669 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 70088 +BPFP 0.7320 bits/point +EBPFP 1.4641 equivalent bits/point +MSE 264.970147 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 264.9701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,944B, BPFP=0.3338 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,772B, BPFP=1.6779 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,700B, BPFP=0.6353 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,180B, BPFP=1.5762 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,276B, BPFP=0.7342 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,628B, BPFP=1.4815 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,688B, BPFP=0.6332 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,920B, BPFP=1.5316 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,004B, BPFP=1.2026 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,580B, BPFP=1.4732 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,884B, BPFP=0.1689 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11067300 171.43147965 + layer.0.v_cache 0.00001409 0.01485699 + layer.1.k_cache 0.05152405 15.97896366 + layer.1.v_cache 0.00000569 0.00589261 + layer.2.k_cache 0.01362713 1.90727133 + layer.2.v_cache 0.00001889 0.01631025 + layer.3.k_cache 0.08227781 7.81721379 + layer.3.v_cache 0.00001889 0.01733574 + layer.4.k_cache 0.00062414 0.40260210 + layer.4.v_cache 0.00005257 0.03624892 + layer.4.output 0.15746453 595.67979788 + ------------------------------------------------------------------------------------- + TOTAL 0.08006400 256.90510354 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 72576 +BPFP 0.7330 bits/point +EBPFP 1.4661 equivalent bits/point +MSE 256.905104 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 256.9051 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 116, 128) +Output shape: (1, 116, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.output: torch.Size([1, 116, 3584]) -> torch.Size([1, 1, 116, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,156B, BPFP=0.2904 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,676B, BPFP=1.4380 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,176B, BPFP=0.5625 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,468B, BPFP=1.4100 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,556B, BPFP=0.6137 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,908B, BPFP=1.3346 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,612B, BPFP=0.6212 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,988B, BPFP=1.3454 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,516B, BPFP=1.0124 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,248B, BPFP=1.2457 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,688B, BPFP=0.1479 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16314862 163.91450027 + layer.0.v_cache 0.00001373 0.01437017 + layer.1.k_cache 0.05481785 16.67619587 + layer.1.v_cache 0.00000557 0.00596259 + layer.2.k_cache 0.00968372 1.87610521 + layer.2.v_cache 0.00001954 0.01677301 + layer.3.k_cache 0.02418142 7.51867676 + layer.3.v_cache 0.00001860 0.01740617 + layer.4.k_cache 0.00062333 0.38816695 + layer.4.v_cache 0.00005467 0.03928316 + layer.4.output 9.83417319 462.96524784 + ------------------------------------------------------------------------------------- + TOTAL 4.06422232 201.83671618 + (elements=1,009,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1009664 +Total Bytes 80992 +BPFP 0.6417 bits/point +EBPFP 1.2835 equivalent bits/point +MSE 201.836716 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 201.8367 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,764B, BPFP=0.3534 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,296B, BPFP=1.8622 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,280B, BPFP=0.6571 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,704B, BPFP=1.7436 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,464B, BPFP=0.6939 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,328B, BPFP=1.4679 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,164B, BPFP=0.6338 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,504B, BPFP=1.5032 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,604B, BPFP=1.1226 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,692B, BPFP=1.5409 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,940B, BPFP=0.1986 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11749315 166.43915264 + layer.0.v_cache 0.00001389 0.01509067 + layer.1.k_cache 0.03399578 16.72265155 + layer.1.v_cache 0.00000555 0.00582255 + layer.2.k_cache 0.00854440 1.94015190 + layer.2.v_cache 0.00001777 0.01571931 + layer.3.k_cache 0.02740619 6.28091705 + layer.3.v_cache 0.00001935 0.01787360 + layer.4.k_cache 0.00063531 0.41092540 + layer.4.v_cache 0.00005166 0.03750600 + layer.4.output 0.18490290 695.02506868 + ------------------------------------------------------------------------------------- + TOTAL 0.08720608 297.47419361 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 64740 +BPFP 0.7629 bits/point +EBPFP 1.5257 equivalent bits/point +MSE 297.474194 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 297.4742 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,876B, BPFP=0.3118 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,424B, BPFP=1.5665 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,868B, BPFP=0.6430 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,012B, BPFP=1.4980 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,700B, BPFP=0.7812 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,688B, BPFP=1.4441 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,244B, BPFP=0.7055 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,748B, BPFP=1.4541 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,000B, BPFP=1.1636 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,160B, BPFP=1.3564 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,288B, BPFP=0.1731 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11492289 167.02856965 + layer.0.v_cache 0.00001440 0.01491027 + layer.1.k_cache 0.01112092 16.79838919 + layer.1.v_cache 0.00000555 0.00559119 + layer.2.k_cache 0.01096075 1.92684255 + layer.2.v_cache 0.00001882 0.01586710 + layer.3.k_cache 0.03781782 7.48597490 + layer.3.v_cache 0.00001905 0.01740232 + layer.4.k_cache 0.00062716 0.37624586 + layer.4.v_cache 0.00006343 0.03593487 + layer.4.output 0.15224552 576.63905775 + ------------------------------------------------------------------------------------- + TOTAL 0.07301703 248.83406660 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 73008 +BPFP 0.7139 bits/point +EBPFP 1.4277 equivalent bits/point +MSE 248.834067 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 248.8341 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,760B, BPFP=0.3571 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,092B, BPFP=1.8450 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,280B, BPFP=0.6656 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,500B, BPFP=1.7248 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,592B, BPFP=0.7289 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,412B, BPFP=1.5041 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,088B, BPFP=0.6266 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,868B, BPFP=1.5966 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,372B, BPFP=1.2930 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,064B, BPFP=1.6364 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,728B, BPFP=0.1950 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10405910 172.03505479 + layer.0.v_cache 0.00001453 0.01438638 + layer.1.k_cache 0.03397077 17.20125431 + layer.1.v_cache 0.00000547 0.00567826 + layer.2.k_cache 0.00568739 1.94190206 + layer.2.v_cache 0.00001916 0.01576912 + layer.3.k_cache 0.09581309 8.01446137 + layer.3.v_cache 0.00001862 0.01693147 + layer.4.k_cache 0.00061896 0.38993365 + layer.4.v_cache 0.00005653 0.03699852 + layer.4.output 0.19123089 703.63555195 + ------------------------------------------------------------------------------------- + TOTAL 0.09287528 301.47771962 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 65756 +BPFP 0.7849 bits/point +EBPFP 1.5698 equivalent bits/point +MSE 301.477720 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 301.4777 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,888B, BPFP=0.3471 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,324B, BPFP=1.7140 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,484B, BPFP=0.6404 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,880B, BPFP=1.6324 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,808B, BPFP=0.7000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,168B, BPFP=1.5015 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,508B, BPFP=0.6449 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,244B, BPFP=1.5154 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,080B, BPFP=1.1176 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,760B, BPFP=1.4265 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,976B, BPFP=0.1832 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09384689 157.78687960 + layer.0.v_cache 0.00001364 0.01484895 + layer.1.k_cache 0.01313462 16.11598834 + layer.1.v_cache 0.00000592 0.00642169 + layer.2.k_cache 0.00813576 1.90870649 + layer.2.v_cache 0.00001762 0.01653263 + layer.3.k_cache 0.02628646 7.42414623 + layer.3.v_cache 0.00002044 0.01896360 + layer.4.k_cache 0.00061596 0.40084143 + layer.4.v_cache 0.00005334 0.03985282 + layer.4.output 0.17343949 638.01407563 + ------------------------------------------------------------------------------------- + TOTAL 0.07977689 273.51951242 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 68120 +BPFP 0.7366 bits/point +EBPFP 1.4732 equivalent bits/point +MSE 273.519512 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 273.5195 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 126, 128) +Output shape: (1, 126, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.output: torch.Size([1, 126, 3584]) -> torch.Size([1, 1, 126, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,404B, BPFP=0.2981 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,912B, BPFP=1.3532 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,580B, BPFP=0.5680 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,544B, BPFP=1.3075 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,644B, BPFP=0.5759 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,008B, BPFP=1.2411 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,320B, BPFP=0.5357 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,176B, BPFP=1.2619 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,656B, BPFP=0.9494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,568B, BPFP=1.1865 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,560B, BPFP=0.1516 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13141363 153.78504774 + layer.0.v_cache 0.00001429 0.01441522 + layer.1.k_cache 0.06496696 15.85148209 + layer.1.v_cache 0.00000583 0.00619574 + layer.2.k_cache 0.01057190 1.74232338 + layer.2.v_cache 0.00002020 0.01631543 + layer.3.k_cache 0.05250616 7.19058276 + layer.3.v_cache 0.00001853 0.01753241 + layer.4.k_cache 0.00062861 0.39026575 + layer.4.v_cache 0.00004930 0.03644108 + layer.4.output 9.05470577 426.04425312 + ------------------------------------------------------------------------------------- + TOTAL 3.74371387 185.96237491 + (elements=1,096,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1096704 +Total Bytes 83372 +BPFP 0.6082 bits/point +EBPFP 1.2163 equivalent bits/point +MSE 185.962375 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 185.9624 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,892B, BPFP=0.3398 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,536B, BPFP=1.7126 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,504B, BPFP=0.6293 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,308B, BPFP=1.6717 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,044B, BPFP=0.7263 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,476B, BPFP=1.5223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,488B, BPFP=0.6264 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,736B, BPFP=1.5690 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,956B, BPFP=1.2493 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,504B, BPFP=1.5273 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,740B, BPFP=0.1729 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13059476 166.69563802 + layer.0.v_cache 0.00001380 0.01432872 + layer.1.k_cache 0.03034586 16.17032316 + layer.1.v_cache 0.00000531 0.00552415 + layer.2.k_cache 0.00374181 1.74635753 + layer.2.v_cache 0.00001886 0.01553431 + layer.3.k_cache 0.04454902 7.56147573 + layer.3.v_cache 0.00001968 0.01745923 + layer.4.k_cache 0.00060938 0.38988705 + layer.4.v_cache 0.00007316 0.03703086 + layer.4.output 0.16485390 623.49502258 + ------------------------------------------------------------------------------------- + TOTAL 0.08023229 268.06580687 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 71184 +BPFP 0.7520 bits/point +EBPFP 1.5041 equivalent bits/point +MSE 268.065807 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 268.0658 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,964B, BPFP=0.3336 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,048B, BPFP=1.7065 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,796B, BPFP=0.6447 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,536B, BPFP=1.6196 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,420B, BPFP=0.7507 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,968B, BPFP=1.5231 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,936B, BPFP=0.6685 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,464B, BPFP=1.6073 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,040B, BPFP=1.1957 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,236B, BPFP=1.5686 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,688B, BPFP=0.1865 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15436328 171.59943954 + layer.0.v_cache 0.00001586 0.01525466 + layer.1.k_cache 0.01080767 16.00721409 + layer.1.v_cache 0.00000575 0.00602716 + layer.2.k_cache 0.00786338 1.86731156 + layer.2.v_cache 0.00001961 0.01616274 + layer.3.k_cache 0.05652578 8.44211877 + layer.3.v_cache 0.00002042 0.01842646 + layer.4.k_cache 0.00062828 0.39937044 + layer.4.v_cache 0.00005674 0.04059793 + layer.4.output 0.15192759 589.66236413 + ------------------------------------------------------------------------------------- + TOTAL 0.07610588 254.47343954 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 76096 +BPFP 0.7602 bits/point +EBPFP 1.5205 equivalent bits/point +MSE 254.473440 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 254.4734 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,752B, BPFP=0.3510 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,900B, BPFP=1.7829 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,260B, BPFP=0.6530 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,368B, BPFP=1.6763 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,724B, BPFP=0.7460 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,616B, BPFP=1.5256 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,072B, BPFP=0.6154 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,864B, BPFP=1.5753 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,932B, BPFP=1.1883 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,944B, BPFP=1.5913 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,488B, BPFP=0.1857 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10254851 170.58931791 + layer.0.v_cache 0.00001399 0.01508732 + layer.1.k_cache 0.05476840 16.92876728 + layer.1.v_cache 0.00000542 0.00571705 + layer.2.k_cache 0.00244555 1.92311096 + layer.2.v_cache 0.00001844 0.01591145 + layer.3.k_cache 0.02811479 8.70878484 + layer.3.v_cache 0.00001905 0.01709597 + layer.4.k_cache 0.00060111 0.39574349 + layer.4.v_cache 0.00005313 0.03813804 + layer.4.output 0.19635826 695.13673306 + ------------------------------------------------------------------------------------- + TOTAL 0.09194683 297.91734151 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 64920 +BPFP 0.7650 bits/point +EBPFP 1.5300 equivalent bits/point +MSE 297.917342 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 297.9173 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,796B, BPFP=0.3422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,336B, BPFP=1.7790 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,316B, BPFP=0.6319 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,044B, BPFP=1.7233 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,768B, BPFP=0.7180 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,160B, BPFP=1.5549 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,368B, BPFP=0.6418 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,064B, BPFP=1.5366 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,028B, BPFP=1.1486 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,936B, BPFP=1.5122 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,800B, BPFP=0.1851 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11987135 163.93247428 + layer.0.v_cache 0.00001390 0.01390794 + layer.1.k_cache 0.03139273 16.84748654 + layer.1.v_cache 0.00000562 0.00559128 + layer.2.k_cache 0.00816743 1.97401838 + layer.2.v_cache 0.00001857 0.01548000 + layer.3.k_cache 0.02640573 8.12569744 + layer.3.v_cache 0.00001968 0.01679571 + layer.4.k_cache 0.00062091 0.39246550 + layer.4.v_cache 0.00004972 0.03586441 + layer.4.output 0.16983549 661.36944686 + ------------------------------------------------------------------------------------- + TOTAL 0.08090671 283.58505350 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 67616 +BPFP 0.7579 bits/point +EBPFP 1.5158 equivalent bits/point +MSE 283.585054 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.164s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 283.5851 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,904B, BPFP=0.3420 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,768B, BPFP=1.7543 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,548B, BPFP=0.6372 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,608B, BPFP=1.7256 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,064B, BPFP=0.7299 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,652B, BPFP=1.5539 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,564B, BPFP=0.6401 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,572B, BPFP=1.5395 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,704B, BPFP=1.2040 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,440B, BPFP=1.5158 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,808B, BPFP=0.1747 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14089697 161.50927173 + layer.0.v_cache 0.00001393 0.01437261 + layer.1.k_cache 0.01618495 16.04541437 + layer.1.v_cache 0.00000582 0.00564168 + layer.2.k_cache 0.00661465 1.77562249 + layer.2.v_cache 0.00002083 0.01488131 + layer.3.k_cache 0.07087009 7.44108283 + layer.3.v_cache 0.00001822 0.01697027 + layer.4.k_cache 0.00061891 0.39046566 + layer.4.v_cache 0.00005311 0.03794012 + layer.4.output 0.16425455 623.25738916 + ------------------------------------------------------------------------------------- + TOTAL 0.08147525 267.65019925 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 71632 +BPFP 0.7568 bits/point +EBPFP 1.5135 equivalent bits/point +MSE 267.650199 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 267.6502 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,928B, BPFP=0.3385 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,972B, BPFP=1.7507 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,548B, BPFP=0.6229 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,588B, BPFP=1.6833 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,120B, BPFP=0.7233 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,748B, BPFP=1.5358 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,448B, BPFP=0.6053 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,476B, BPFP=1.6636 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,116B, BPFP=1.2493 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,040B, BPFP=1.5871 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,252B, BPFP=0.1819 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12786581 171.21975948 + layer.0.v_cache 0.00001426 0.01526173 + layer.1.k_cache 0.03183105 16.27992231 + layer.1.v_cache 0.00000593 0.00603682 + layer.2.k_cache 0.00241907 1.82128392 + layer.2.v_cache 0.00001785 0.01535803 + layer.3.k_cache 0.02946699 8.54431701 + layer.3.v_cache 0.00001957 0.01881351 + layer.4.k_cache 0.00063747 0.39193070 + layer.4.v_cache 0.00005003 0.03763996 + layer.4.output 0.16568538 609.10062199 + ------------------------------------------------------------------------------------- + TOTAL 0.07953681 262.47380455 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 74236 +BPFP 0.7666 bits/point +EBPFP 1.5333 equivalent bits/point +MSE 262.473805 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 262.4738 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,756B, BPFP=0.3518 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,756B, BPFP=1.7540 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,288B, BPFP=0.6587 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,540B, BPFP=1.7107 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,644B, BPFP=0.7300 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,616B, BPFP=1.5256 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,200B, BPFP=0.6410 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,852B, BPFP=1.5729 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,772B, BPFP=1.1562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,512B, BPFP=1.5048 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,924B, BPFP=0.1981 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12225728 166.32414363 + layer.0.v_cache 0.00001337 0.01462978 + layer.1.k_cache 0.03480693 16.98149852 + layer.1.v_cache 0.00000545 0.00558949 + layer.2.k_cache 0.00853881 1.93928626 + layer.2.v_cache 0.00001924 0.01596787 + layer.3.k_cache 0.06130744 6.77593916 + layer.3.v_cache 0.00001928 0.01840780 + layer.4.k_cache 0.00058936 0.38310853 + layer.4.v_cache 0.00007197 0.03821798 + layer.4.output 0.18657821 694.96634615 + ------------------------------------------------------------------------------------- + TOTAL 0.09021627 297.48595365 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 64860 +BPFP 0.7643 bits/point +EBPFP 1.5286 equivalent bits/point +MSE 297.485954 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 297.4860 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,760B, BPFP=0.3571 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,100B, BPFP=1.8466 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,264B, BPFP=0.6623 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,412B, BPFP=1.7070 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,408B, BPFP=0.6916 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,664B, BPFP=1.5552 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,992B, BPFP=0.6071 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,028B, BPFP=1.6291 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,124B, BPFP=1.2427 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,044B, BPFP=1.6323 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,592B, BPFP=0.1911 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582910 170.46765929 + layer.0.v_cache 0.00001386 0.01519094 + layer.1.k_cache 0.03408930 16.99562449 + layer.1.v_cache 0.00000538 0.00615364 + layer.2.k_cache 0.00228146 1.73013940 + layer.2.v_cache 0.00001859 0.01672320 + layer.3.k_cache 0.06276476 6.72857587 + layer.3.v_cache 0.00001913 0.01865388 + layer.4.k_cache 0.00060520 0.41401038 + layer.4.v_cache 0.00005125 0.04049267 + layer.4.output 0.19356344 703.14407468 + ------------------------------------------------------------------------------------- + TOTAL 0.09238954 301.08480862 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 65388 +BPFP 0.7805 bits/point +EBPFP 1.5610 equivalent bits/point +MSE 301.084809 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 301.0848 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,884B, BPFP=0.3547 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,732B, BPFP=1.6438 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,372B, BPFP=0.6348 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,288B, BPFP=1.5602 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,664B, BPFP=0.6898 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,728B, BPFP=1.4548 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,256B, BPFP=0.6130 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,028B, BPFP=1.5113 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,844B, BPFP=1.1002 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,552B, BPFP=1.4217 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,904B, BPFP=0.1857 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09700135 159.90721480 + layer.0.v_cache 0.00001355 0.01505028 + layer.1.k_cache 0.01103399 16.15088773 + layer.1.v_cache 0.00000598 0.00606808 + layer.2.k_cache 0.00794396 1.94314097 + layer.2.v_cache 0.00001935 0.01742351 + layer.3.k_cache 0.04207947 8.78982507 + layer.3.v_cache 0.00002010 0.01967798 + layer.4.k_cache 0.00062246 0.41333762 + layer.4.v_cache 0.00005226 0.04054439 + layer.4.output 0.17535226 652.14516997 + ------------------------------------------------------------------------------------- + TOTAL 0.08154460 279.54819766 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 65252 +BPFP 0.7226 bits/point +EBPFP 1.4452 equivalent bits/point +MSE 279.548198 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 279.5482 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,892B, BPFP=0.3478 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,572B, BPFP=1.7596 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,576B, BPFP=0.6574 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,084B, BPFP=1.6699 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,980B, BPFP=0.7316 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,212B, BPFP=1.5096 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,608B, BPFP=0.6632 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,524B, BPFP=1.5669 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,644B, BPFP=1.2213 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,376B, BPFP=1.5397 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,812B, BPFP=0.1789 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11944316 159.82364430 + layer.0.v_cache 0.00001525 0.01437722 + layer.1.k_cache 0.01376067 15.95525620 + layer.1.v_cache 0.00000561 0.00564334 + layer.2.k_cache 0.00522648 1.95383552 + layer.2.v_cache 0.00001966 0.01604320 + layer.3.k_cache 0.04600189 8.55885297 + layer.3.v_cache 0.00001912 0.01735477 + layer.4.k_cache 0.00062079 0.37721962 + layer.4.v_cache 0.00005114 0.03622489 + layer.4.output 0.17175110 637.86265756 + ------------------------------------------------------------------------------------- + TOTAL 0.08161303 273.63512088 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 70280 +BPFP 0.7599 bits/point +EBPFP 1.5199 equivalent bits/point +MSE 273.635121 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 273.6351 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,764B, BPFP=0.3534 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,692B, BPFP=1.7412 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,296B, BPFP=0.6603 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,396B, BPFP=1.6819 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,556B, BPFP=0.7123 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,592B, BPFP=1.5208 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,184B, BPFP=0.6378 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,508B, BPFP=1.5040 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,712B, BPFP=1.1442 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,500B, BPFP=1.5024 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,860B, BPFP=0.1963 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120220 167.48950821 + layer.0.v_cache 0.00001368 0.01464857 + layer.1.k_cache 0.03273370 16.73670686 + layer.1.v_cache 0.00000573 0.00575895 + layer.2.k_cache 0.00892877 1.95569845 + layer.2.v_cache 0.00001953 0.01624224 + layer.3.k_cache 0.02974504 6.47020154 + layer.3.v_cache 0.00001879 0.01774776 + layer.4.k_cache 0.00063874 0.40859225 + layer.4.v_cache 0.00005200 0.04006891 + layer.4.output 0.17853031 695.09312042 + ------------------------------------------------------------------------------------- + TOTAL 0.08429825 297.57688333 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 64060 +BPFP 0.7549 bits/point +EBPFP 1.5097 equivalent bits/point +MSE 297.576883 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 297.5769 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,872B, BPFP=0.3524 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,232B, BPFP=1.7380 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,368B, BPFP=0.6340 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,916B, BPFP=1.6785 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,840B, BPFP=0.7229 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,988B, BPFP=1.5038 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,544B, BPFP=0.6672 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,088B, BPFP=1.5226 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,244B, BPFP=1.1755 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,000B, BPFP=1.5060 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,976B, BPFP=0.1876 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09802409 159.24089326 + layer.0.v_cache 0.00001402 0.01452375 + layer.1.k_cache 0.03453328 16.41897767 + layer.1.v_cache 0.00000594 0.00573140 + layer.2.k_cache 0.01027090 2.05685811 + layer.2.v_cache 0.00001933 0.01594722 + layer.3.k_cache 0.02613793 7.46863749 + layer.3.v_cache 0.00001931 0.01732884 + layer.4.k_cache 0.00063818 0.40250029 + layer.4.v_cache 0.00005393 0.03851743 + layer.4.output 0.18084440 652.50510972 + ------------------------------------------------------------------------------------- + TOTAL 0.08444869 279.60092256 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 68068 +BPFP 0.7538 bits/point +EBPFP 1.5075 equivalent bits/point +MSE 279.600923 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 279.6009 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,084B, BPFP=0.3161 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,532B, BPFP=1.5977 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,336B, BPFP=0.6578 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,624B, BPFP=1.6117 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,224B, BPFP=0.7925 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,476B, BPFP=1.5892 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,332B, BPFP=0.6572 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,172B, BPFP=1.5431 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,356B, BPFP=1.2676 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,792B, BPFP=1.4854 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,876B, BPFP=0.1707 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13030120 148.79530112 + layer.0.v_cache 0.00001385 0.01499759 + layer.1.k_cache 0.02890817 16.14056811 + layer.1.v_cache 0.00000611 0.00627942 + layer.2.k_cache 0.00721959 1.69570967 + layer.2.v_cache 0.00001830 0.01594251 + layer.3.k_cache 0.02195997 6.17842354 + layer.3.v_cache 0.00001903 0.01818607 + layer.4.k_cache 0.00063766 0.38868617 + layer.4.v_cache 0.00005061 0.03718041 + layer.4.output 11.07892277 521.71493585 + ------------------------------------------------------------------------------------- + TOTAL 4.57303494 225.01740151 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 83804 +BPFP 0.7478 bits/point +EBPFP 1.4956 equivalent bits/point +MSE 225.017402 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 225.0174 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,864B, BPFP=0.3509 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,204B, BPFP=1.7327 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,312B, BPFP=0.6235 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,588B, BPFP=1.6167 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,724B, BPFP=0.7011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,020B, BPFP=1.5098 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,252B, BPFP=0.6122 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,292B, BPFP=1.5610 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,156B, BPFP=1.1589 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,928B, BPFP=1.4925 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,628B, BPFP=0.1782 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12633393 164.95826666 + layer.0.v_cache 0.00001372 0.01448533 + layer.1.k_cache 0.01411889 16.47780379 + layer.1.v_cache 0.00000549 0.00585852 + layer.2.k_cache 0.00833277 1.84441219 + layer.2.v_cache 0.00001851 0.01592359 + layer.3.k_cache 0.05797521 8.13969697 + layer.3.v_cache 0.00001879 0.01816732 + layer.4.k_cache 0.00064672 0.42124296 + layer.4.v_cache 0.00005349 0.04103619 + layer.4.output 0.17447430 653.02995912 + ------------------------------------------------------------------------------------- + TOTAL 0.08404927 280.18509455 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 66968 +BPFP 0.7416 bits/point +EBPFP 1.4832 equivalent bits/point +MSE 280.185095 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 280.1851 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,636B, BPFP=0.3652 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,992B, BPFP=1.7839 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,036B, BPFP=0.6777 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,448B, BPFP=1.6625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,284B, BPFP=0.7330 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,588B, BPFP=1.4705 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,828B, BPFP=0.6312 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,032B, BPFP=1.5696 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,264B, BPFP=1.1750 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,204B, BPFP=1.6080 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,720B, BPFP=0.2143 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08086312 165.32339565 + layer.0.v_cache 0.00001331 0.01544352 + layer.1.k_cache 0.01551799 16.00196010 + layer.1.v_cache 0.00000513 0.00559407 + layer.2.k_cache 0.00232868 1.84749146 + layer.2.v_cache 0.00002262 0.01623513 + layer.3.k_cache 0.10623371 8.44260777 + layer.3.v_cache 0.00001836 0.01769301 + layer.4.k_cache 0.00061897 0.39431253 + layer.4.v_cache 0.00005922 0.03739960 + layer.4.output 0.20964142 774.83463010 + ------------------------------------------------------------------------------------- + TOTAL 0.09842183 330.34967903 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 59032 +BPFP 0.7751 bits/point +EBPFP 1.5502 equivalent bits/point +MSE 330.349679 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 330.3497 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,952B, BPFP=0.3315 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,924B, BPFP=1.6855 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,820B, BPFP=0.6488 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,172B, BPFP=1.5577 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,304B, BPFP=0.7310 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,584B, BPFP=1.4579 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,848B, BPFP=0.6535 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,720B, BPFP=1.4810 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,832B, BPFP=1.1603 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,404B, BPFP=1.4273 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,532B, BPFP=0.1827 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13670269 171.22112772 + layer.0.v_cache 0.00001568 0.01541701 + layer.1.k_cache 0.03139175 16.59140147 + layer.1.v_cache 0.00000602 0.00608155 + layer.2.k_cache 0.00629326 1.87130472 + layer.2.v_cache 0.00001871 0.01550629 + layer.3.k_cache 0.03837063 7.28426660 + layer.3.v_cache 0.00001860 0.01706298 + layer.4.k_cache 0.00062597 0.39451052 + layer.4.v_cache 0.00005354 0.03781222 + layer.4.output 0.15974679 589.49568129 + ------------------------------------------------------------------------------------- + TOTAL 0.07833673 254.34848589 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 73092 +BPFP 0.7302 bits/point +EBPFP 1.4604 equivalent bits/point +MSE 254.348486 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 254.3485 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,128B, BPFP=0.3107 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,684B, BPFP=1.5602 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,316B, BPFP=0.6303 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,644B, BPFP=1.5543 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,444B, BPFP=0.7950 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,212B, BPFP=1.4912 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,452B, BPFP=0.6501 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,468B, BPFP=1.5286 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,748B, BPFP=1.2775 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,680B, BPFP=1.4136 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,748B, BPFP=0.1616 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12281527 163.21104702 + layer.0.v_cache 0.00001417 0.01475006 + layer.1.k_cache 0.04442539 16.25377163 + layer.1.v_cache 0.00000572 0.00603752 + layer.2.k_cache 0.01033308 1.85111971 + layer.2.v_cache 0.00001923 0.01625417 + layer.3.k_cache 0.04555132 7.61945835 + layer.3.v_cache 0.00001841 0.01808089 + layer.4.k_cache 0.00060580 0.37972951 + layer.4.v_cache 0.00007068 0.03770740 + layer.4.output 10.66341841 502.01009680 + ------------------------------------------------------------------------------------- + TOTAL 4.40398753 217.85168434 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 84524 +BPFP 0.7261 bits/point +EBPFP 1.4521 equivalent bits/point +MSE 217.851684 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 217.8517 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,884B, BPFP=0.3423 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,632B, BPFP=1.7500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,512B, BPFP=0.6381 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,280B, BPFP=1.6860 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,992B, BPFP=0.7253 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,720B, BPFP=1.5843 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,404B, BPFP=0.6185 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,944B, BPFP=1.6250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,528B, BPFP=1.1860 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,728B, BPFP=1.5858 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,680B, BPFP=0.1734 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11362342 166.74219886 + layer.0.v_cache 0.00001418 0.01407565 + layer.1.k_cache 0.05429294 16.53169854 + layer.1.v_cache 0.00000580 0.00572571 + layer.2.k_cache 0.00513919 1.79235343 + layer.2.v_cache 0.00001816 0.01549574 + layer.3.k_cache 0.05885953 8.09240226 + layer.3.v_cache 0.00001841 0.01668533 + layer.4.k_cache 0.00061763 0.38464861 + layer.4.v_cache 0.00005080 0.03626542 + layer.4.output 0.16727475 630.73193522 + ------------------------------------------------------------------------------------- + TOTAL 0.08256255 271.10324094 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 71304 +BPFP 0.7621 bits/point +EBPFP 1.5241 equivalent bits/point +MSE 271.103241 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 271.1032 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,972B, BPFP=0.3313 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,944B, BPFP=1.6707 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,144B, BPFP=0.6962 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,516B, BPFP=1.5988 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,912B, BPFP=0.8253 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,220B, BPFP=1.5491 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,248B, BPFP=0.7137 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,588B, BPFP=1.6109 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,996B, BPFP=1.3434 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,692B, BPFP=1.4603 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,644B, BPFP=0.1835 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702597 169.84965138 + layer.0.v_cache 0.00001402 0.01448108 + layer.1.k_cache 0.01332537 16.65130681 + layer.1.v_cache 0.00000581 0.00587405 + layer.2.k_cache 0.00507878 1.86664557 + layer.2.v_cache 0.00001944 0.01523946 + layer.3.k_cache 0.04051446 7.72318522 + layer.3.v_cache 0.00001931 0.01668232 + layer.4.k_cache 0.00062870 0.37953662 + layer.4.v_cache 0.00006514 0.03541609 + layer.4.output 0.16549673 583.34235791 + ------------------------------------------------------------------------------------- + TOTAL 0.07795142 251.76203082 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 77876 +BPFP 0.7696 bits/point +EBPFP 1.5393 equivalent bits/point +MSE 251.762031 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.005s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 251.7620 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,884B, BPFP=0.3463 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,288B, BPFP=1.7074 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,520B, BPFP=0.6471 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,744B, BPFP=1.6074 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,888B, BPFP=0.7147 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,672B, BPFP=1.4103 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,444B, BPFP=0.6331 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,880B, BPFP=1.4485 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,396B, BPFP=1.1757 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,756B, BPFP=1.4257 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,920B, BPFP=0.1817 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13925676 161.96017923 + layer.0.v_cache 0.00001387 0.01485280 + layer.1.k_cache 0.03290265 16.32913890 + layer.1.v_cache 0.00000589 0.00603078 + layer.2.k_cache 0.00800494 1.98424862 + layer.2.v_cache 0.00001921 0.01541931 + layer.3.k_cache 0.08301921 7.63272633 + layer.3.v_cache 0.00001931 0.01760082 + layer.4.k_cache 0.00065893 0.37794364 + layer.4.v_cache 0.00004891 0.03578911 + layer.4.output 0.17367665 638.11339286 + ------------------------------------------------------------------------------------- + TOTAL 0.08704037 273.83339291 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 67392 +BPFP 0.7287 bits/point +EBPFP 1.4574 equivalent bits/point +MSE 273.833393 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 273.8334 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,068B, BPFP=0.3137 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,460B, BPFP=1.5868 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,268B, BPFP=0.6475 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,064B, BPFP=1.5267 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,108B, BPFP=0.7749 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,964B, BPFP=1.5115 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,248B, BPFP=0.6444 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,720B, BPFP=1.4745 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,148B, BPFP=1.2360 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,996B, BPFP=1.3647 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,336B, BPFP=0.1590 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10583916 152.09490671 + layer.0.v_cache 0.00001406 0.01493664 + layer.1.k_cache 0.01351924 15.94849344 + layer.1.v_cache 0.00000536 0.00590770 + layer.2.k_cache 0.00567923 1.72890435 + layer.2.v_cache 0.00001856 0.01601036 + layer.3.k_cache 0.05293757 7.19198875 + layer.3.v_cache 0.00001841 0.01709090 + layer.4.k_cache 0.00064880 0.38389076 + layer.4.v_cache 0.00004836 0.03587727 + layer.4.output 11.07886250 521.55170770 + ------------------------------------------------------------------------------------- + TOTAL 4.57239802 225.19411534 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 80380 +BPFP 0.7173 bits/point +EBPFP 1.4345 equivalent bits/point +MSE 225.194115 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.005s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 225.1941 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,872B, BPFP=0.3524 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,272B, BPFP=1.7455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,412B, BPFP=0.6423 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,820B, BPFP=1.6604 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,612B, BPFP=0.6800 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,020B, BPFP=1.5098 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,328B, BPFP=0.6265 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,260B, BPFP=1.5550 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,032B, BPFP=1.1355 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,108B, BPFP=1.5264 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,648B, BPFP=0.1788 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860243 162.53670934 + layer.0.v_cache 0.00001403 0.01476194 + layer.1.k_cache 0.01400426 16.36835526 + layer.1.v_cache 0.00000584 0.00613968 + layer.2.k_cache 0.01277778 1.90402369 + layer.2.v_cache 0.00001980 0.01689203 + layer.3.k_cache 0.07376247 7.62701195 + layer.3.v_cache 0.00001877 0.01774343 + layer.4.k_cache 0.00062152 0.40671277 + layer.4.v_cache 0.00005140 0.03942147 + layer.4.output 0.18179212 652.84385757 + ------------------------------------------------------------------------------------- + TOTAL 0.08837783 279.93204556 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 67384 +BPFP 0.7462 bits/point +EBPFP 1.4924 equivalent bits/point +MSE 279.932046 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 279.9320 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,924B, BPFP=0.3340 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,924B, BPFP=1.7229 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,612B, BPFP=0.6271 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,240B, BPFP=1.6042 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,180B, BPFP=0.7257 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,756B, BPFP=1.5201 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,636B, BPFP=0.6312 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,064B, BPFP=1.5736 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,788B, BPFP=1.1785 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,868B, BPFP=1.5396 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,016B, BPFP=0.1740 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13781073 175.94294705 + layer.0.v_cache 0.00001414 0.01518310 + layer.1.k_cache 0.01252146 16.55884060 + layer.1.v_cache 0.00000561 0.00571589 + layer.2.k_cache 0.00238173 1.86681892 + layer.2.v_cache 0.00001849 0.01592135 + layer.3.k_cache 0.04360520 8.45893894 + layer.3.v_cache 0.00001973 0.01748148 + layer.4.k_cache 0.00061668 0.37655254 + layer.4.v_cache 0.00005448 0.03808466 + layer.4.output 0.16772948 602.61448413 + ------------------------------------------------------------------------------------- + TOTAL 0.08065615 260.09399255 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 73008 +BPFP 0.7456 bits/point +EBPFP 1.4912 equivalent bits/point +MSE 260.093993 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.005s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 260.0940 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,784B, BPFP=0.3528 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,560B, BPFP=1.8908 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,308B, BPFP=0.6543 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,604B, BPFP=1.7017 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,712B, BPFP=0.7342 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,804B, BPFP=1.5435 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,280B, BPFP=0.6487 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,140B, BPFP=1.6100 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,084B, BPFP=1.2033 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,200B, BPFP=1.6218 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,836B, BPFP=0.1932 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07550411 161.79680083 + layer.0.v_cache 0.00001381 0.01472387 + layer.1.k_cache 0.01271453 16.69679589 + layer.1.v_cache 0.00000531 0.00566719 + layer.2.k_cache 0.00681068 1.98850723 + layer.2.v_cache 0.00001904 0.01631937 + layer.3.k_cache 0.02844424 7.71964158 + layer.3.v_cache 0.00001798 0.01740427 + layer.4.k_cache 0.00061127 0.41803162 + layer.4.v_cache 0.00009323 0.03981721 + layer.4.output 0.18255964 686.22202758 + ------------------------------------------------------------------------------------- + TOTAL 0.08247951 293.66281777 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 67312 +BPFP 0.7831 bits/point +EBPFP 1.5663 equivalent bits/point +MSE 293.662818 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 293.6628 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,796B, BPFP=0.3508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,404B, BPFP=1.8367 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,312B, BPFP=0.6469 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,116B, BPFP=1.7805 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,688B, BPFP=0.7203 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,072B, BPFP=1.5766 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,308B, BPFP=0.6461 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,376B, BPFP=1.6359 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,456B, BPFP=1.2609 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,344B, BPFP=1.6297 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,876B, BPFP=0.1919 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08590809 158.72929687 + layer.0.v_cache 0.00001388 0.01403057 + layer.1.k_cache 0.01594085 16.58738708 + layer.1.v_cache 0.00000552 0.00532539 + layer.2.k_cache 0.00791870 1.89236755 + layer.2.v_cache 0.00001801 0.01456605 + layer.3.k_cache 0.02757127 7.73371582 + layer.3.v_cache 0.00001862 0.01615960 + layer.4.k_cache 0.00063773 0.39134731 + layer.4.v_cache 0.00005492 0.03802974 + layer.4.output 0.18408036 677.62678571 + ------------------------------------------------------------------------------------- + TOTAL 0.08392059 289.92998388 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 68748 +BPFP 0.7898 bits/point +EBPFP 1.5797 equivalent bits/point +MSE 289.929984 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 289.9300 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,748B, BPFP=0.3594 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,664B, BPFP=1.7812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,144B, BPFP=0.6464 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,956B, BPFP=1.6357 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,348B, BPFP=0.6883 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,996B, BPFP=1.4383 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,964B, BPFP=0.6094 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,420B, BPFP=1.5255 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,552B, BPFP=1.1414 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,548B, BPFP=1.5518 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,884B, BPFP=0.2022 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08991440 177.31350226 + layer.0.v_cache 0.00001347 0.01505948 + layer.1.k_cache 0.03345199 16.44574938 + layer.1.v_cache 0.00000518 0.00589588 + layer.2.k_cache 0.00398020 1.88969883 + layer.2.v_cache 0.00001815 0.01576968 + layer.3.k_cache 0.04607732 8.17153288 + layer.3.v_cache 0.00001896 0.01706561 + layer.4.k_cache 0.00062119 0.40919038 + layer.4.v_cache 0.00005040 0.03840957 + layer.4.output 0.18612516 713.50117481 + ------------------------------------------------------------------------------------- + TOTAL 0.08688396 305.81353516 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 62224 +BPFP 0.7525 bits/point +EBPFP 1.5050 equivalent bits/point +MSE 305.813535 +---------------------- -------------------------------------------------------- +Time: 0.375s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 305.8135 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,952B, BPFP=0.3352 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,956B, BPFP=1.7095 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,708B, BPFP=0.6367 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,264B, BPFP=1.5907 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,472B, BPFP=0.7679 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,676B, BPFP=1.4897 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,876B, BPFP=0.6655 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,196B, BPFP=1.5790 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,024B, BPFP=1.2060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,664B, BPFP=1.4876 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,196B, BPFP=0.1765 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11186004 172.79349459 + layer.0.v_cache 0.00001364 0.01487016 + layer.1.k_cache 0.01386243 16.29210755 + layer.1.v_cache 0.00000608 0.00601206 + layer.2.k_cache 0.00506424 1.84718038 + layer.2.v_cache 0.00001872 0.01612868 + layer.3.k_cache 0.02496567 6.94444593 + layer.3.v_cache 0.00001876 0.01796983 + layer.4.k_cache 0.00064518 0.39731061 + layer.4.v_cache 0.00005436 0.03857723 + layer.4.output 0.16682192 596.17930730 + ------------------------------------------------------------------------------------- + TOTAL 0.07789780 257.15430871 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 73984 +BPFP 0.7473 bits/point +EBPFP 1.4945 equivalent bits/point +MSE 257.154309 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 257.1543 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.3479 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,308B, BPFP=1.7523 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,352B, BPFP=0.6310 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,932B, BPFP=1.6815 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,756B, BPFP=0.7071 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,304B, BPFP=1.5633 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,240B, BPFP=0.6099 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,332B, BPFP=1.5685 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,024B, BPFP=1.1340 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,884B, BPFP=1.4842 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,792B, BPFP=0.1827 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08542032 166.01268355 + layer.0.v_cache 0.00001402 0.01488630 + layer.1.k_cache 0.01083013 16.19529044 + layer.1.v_cache 0.00000556 0.00590638 + layer.2.k_cache 0.00382075 1.96622072 + layer.2.v_cache 0.00001860 0.01621205 + layer.3.k_cache 0.02610962 9.29360852 + layer.3.v_cache 0.00002051 0.01843306 + layer.4.k_cache 0.00063312 0.39994021 + layer.4.v_cache 0.00005614 0.04100916 + layer.4.output 0.16811641 653.36483434 + ------------------------------------------------------------------------------------- + TOTAL 0.07669080 280.44223710 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 67772 +BPFP 0.7505 bits/point +EBPFP 1.5010 equivalent bits/point +MSE 280.442237 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 280.4422 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,812B, BPFP=0.3495 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,092B, BPFP=1.7539 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,296B, BPFP=0.6358 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,392B, BPFP=1.6188 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,808B, BPFP=0.7346 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,640B, BPFP=1.4738 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,320B, BPFP=0.6404 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,852B, BPFP=1.5147 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,880B, BPFP=1.1343 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,464B, BPFP=1.4398 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,792B, BPFP=0.1872 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10540552 172.05713493 + layer.0.v_cache 0.00001541 0.01466440 + layer.1.k_cache 0.03402772 16.79469582 + layer.1.v_cache 0.00000555 0.00618173 + layer.2.k_cache 0.00394259 1.99368964 + layer.2.v_cache 0.00001868 0.01698877 + layer.3.k_cache 0.02873192 7.95361780 + layer.3.v_cache 0.00001983 0.01822996 + layer.4.k_cache 0.00063741 0.41825330 + layer.4.v_cache 0.00005806 0.04062828 + layer.4.output 0.17961645 669.22360009 + ------------------------------------------------------------------------------------- + TOTAL 0.08412811 287.28701678 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 65348 +BPFP 0.7415 bits/point +EBPFP 1.4830 equivalent bits/point +MSE 287.287017 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 287.2870 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3516 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,648B, BPFP=1.8844 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,364B, BPFP=0.6570 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,988B, BPFP=1.7555 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,876B, BPFP=0.7570 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,328B, BPFP=1.6266 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,212B, BPFP=0.6273 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,288B, BPFP=1.6187 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,288B, BPFP=1.2281 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,160B, BPFP=1.5938 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,976B, BPFP=0.1946 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10022384 159.03897705 + layer.0.v_cache 0.00001350 0.01459510 + layer.1.k_cache 0.03445883 16.33270569 + layer.1.v_cache 0.00000556 0.00575475 + layer.2.k_cache 0.00550650 1.99117928 + layer.2.v_cache 0.00001848 0.01592626 + layer.3.k_cache 0.03016714 7.42081451 + layer.3.v_cache 0.00001937 0.01854682 + layer.4.k_cache 0.00063435 0.41688318 + layer.4.v_cache 0.00006510 0.03855597 + layer.4.output 0.18593751 677.73197545 + ------------------------------------------------------------------------------------- + TOTAL 0.08662796 289.96575098 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 68928 +BPFP 0.7919 bits/point +EBPFP 1.5838 equivalent bits/point +MSE 289.965751 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 289.9658 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,760B, BPFP=0.2835 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,676B, BPFP=1.5586 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,576B, BPFP=0.5760 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,392B, BPFP=1.5129 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,236B, BPFP=0.6823 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,024B, BPFP=1.4536 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,912B, BPFP=0.6302 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,112B, BPFP=1.4678 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,012B, BPFP=1.1295 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,620B, BPFP=1.3885 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,644B, BPFP=0.1529 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11898830 169.39525612 + layer.0.v_cache 0.00001403 0.01473254 + layer.1.k_cache 0.01243949 16.56414858 + layer.1.v_cache 0.00000551 0.00566909 + layer.2.k_cache 0.00346032 1.90529184 + layer.2.v_cache 0.00001901 0.01562665 + layer.3.k_cache 0.02403709 7.50599151 + layer.3.v_cache 0.00001994 0.01719824 + layer.4.k_cache 0.00063949 0.39563570 + layer.4.v_cache 0.00006591 0.03760573 + layer.4.output 0.03773703 574.49967784 + ------------------------------------------------------------------------------------- + TOTAL 0.02493225 248.07970005 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 72964 +BPFP 0.6914 bits/point +EBPFP 1.3827 equivalent bits/point +MSE 248.079700 +---------------------- -------------------------------------------------------- +Time: 0.382s Load: 0.004s, Pack+Encode: 0.167s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 248.0797 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3430 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,116B, BPFP=1.7370 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,304B, BPFP=0.6296 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,740B, BPFP=1.6654 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,640B, BPFP=0.6936 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,040B, BPFP=1.5320 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,200B, BPFP=0.6098 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,836B, BPFP=1.4931 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,996B, BPFP=1.1425 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,080B, BPFP=1.5396 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,828B, BPFP=0.1859 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10693085 158.92717464 + layer.0.v_cache 0.00001341 0.01484076 + layer.1.k_cache 0.01172146 16.96669565 + layer.1.v_cache 0.00000553 0.00600821 + layer.2.k_cache 0.00236719 2.00614948 + layer.2.v_cache 0.00001835 0.01642302 + layer.3.k_cache 0.02622769 8.43742185 + layer.3.v_cache 0.00001864 0.01772325 + layer.4.k_cache 0.00061713 0.41748652 + layer.4.v_cache 0.00004943 0.03960705 + layer.4.output 0.18273007 661.39552483 + ------------------------------------------------------------------------------------- + TOTAL 0.08394589 283.33048260 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 66580 +BPFP 0.7463 bits/point +EBPFP 1.4926 equivalent bits/point +MSE 283.330483 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 283.3305 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,952B, BPFP=0.3020 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,260B, BPFP=1.5873 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,204B, BPFP=0.6504 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,288B, BPFP=1.5916 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,772B, BPFP=0.7382 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,600B, BPFP=1.4851 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,988B, BPFP=0.6170 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,668B, BPFP=1.4957 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,632B, BPFP=1.1807 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,864B, BPFP=1.3713 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,504B, BPFP=0.1658 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13156438 156.78329981 + layer.0.v_cache 0.00001398 0.01521993 + layer.1.k_cache 0.04507210 16.18679417 + layer.1.v_cache 0.00000603 0.00628662 + layer.2.k_cache 0.00848899 1.83837619 + layer.2.v_cache 0.00001983 0.01689689 + layer.3.k_cache 0.06093804 6.62452985 + layer.3.v_cache 0.00001891 0.01780646 + layer.4.k_cache 0.00062755 0.39640706 + layer.4.v_cache 0.00004781 0.03614913 + layer.4.output 11.29576791 532.17830622 + ------------------------------------------------------------------------------------- + TOTAL 4.66571606 229.83352410 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 78732 +BPFP 0.7165 bits/point +EBPFP 1.4329 equivalent bits/point +MSE 229.833524 +---------------------- -------------------------------------------------------- +Time: 0.377s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 229.8335 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,840B, BPFP=0.3464 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,484B, BPFP=1.7854 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,356B, BPFP=0.6318 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,924B, BPFP=1.6800 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,012B, BPFP=0.7553 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,508B, BPFP=1.6017 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,472B, BPFP=0.6536 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,900B, BPFP=1.6755 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,516B, BPFP=1.2267 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,684B, BPFP=1.6348 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,788B, BPFP=0.1826 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12340720 160.35484516 + layer.0.v_cache 0.00001379 0.01504226 + layer.1.k_cache 0.01365945 16.37301158 + layer.1.v_cache 0.00000586 0.00605851 + layer.2.k_cache 0.00515996 1.88103402 + layer.2.v_cache 0.00002118 0.01641711 + layer.3.k_cache 0.04682968 8.10426496 + layer.3.v_cache 0.00001899 0.01783804 + layer.4.k_cache 0.00062596 0.40556873 + layer.4.v_cache 0.00004975 0.03936472 + layer.4.output 0.18155291 653.15345310 + ------------------------------------------------------------------------------------- + TOTAL 0.08592130 279.95809510 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 70484 +BPFP 0.7805 bits/point +EBPFP 1.5610 equivalent bits/point +MSE 279.958095 +---------------------- -------------------------------------------------------- +Time: 0.380s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 279.9581 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,844B, BPFP=0.3430 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,384B, BPFP=1.7455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,396B, BPFP=0.6317 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,044B, BPFP=1.6823 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,800B, BPFP=0.7068 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,104B, BPFP=1.5074 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,520B, BPFP=0.6548 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,220B, BPFP=1.5290 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,408B, BPFP=1.1920 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,388B, BPFP=1.5603 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,228B, BPFP=0.1921 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10400932 165.10187640 + layer.0.v_cache 0.00001381 0.01445049 + layer.1.k_cache 0.01300498 16.03567360 + layer.1.v_cache 0.00000559 0.00593190 + layer.2.k_cache 0.00535372 1.91844977 + layer.2.v_cache 0.00001862 0.01533760 + layer.3.k_cache 0.04415199 7.24799383 + layer.3.v_cache 0.00001841 0.01719866 + layer.4.k_cache 0.00060249 0.38648487 + layer.4.v_cache 0.00005274 0.03951375 + layer.4.output 0.16694923 644.65593112 + ------------------------------------------------------------------------------------- + TOTAL 0.07858096 276.66908404 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 69336 +BPFP 0.7587 bits/point +EBPFP 1.5173 equivalent bits/point +MSE 276.669084 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.005s, Pack+Encode: 0.164s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 276.6691 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3516 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,228B, BPFP=1.8023 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,316B, BPFP=0.6477 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,892B, BPFP=1.7367 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,764B, BPFP=0.7352 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,376B, BPFP=1.6359 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,200B, BPFP=0.6250 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,360B, BPFP=1.6328 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,376B, BPFP=1.2453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,296B, BPFP=1.6203 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,588B, BPFP=0.1838 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09321334 158.96184082 + layer.0.v_cache 0.00001862 0.01565094 + layer.1.k_cache 0.03257937 16.44676819 + layer.1.v_cache 0.00000532 0.00573330 + layer.2.k_cache 0.01442847 1.82757530 + layer.2.v_cache 0.00001865 0.01653403 + layer.3.k_cache 0.04385926 8.50195465 + layer.3.v_cache 0.00001828 0.01764493 + layer.4.k_cache 0.00061064 0.40038157 + layer.4.v_cache 0.00005353 0.04019069 + layer.4.output 0.18500509 677.97410714 + ------------------------------------------------------------------------------------- + TOTAL 0.08704948 290.12076614 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 68196 +BPFP 0.7835 bits/point +EBPFP 1.5670 equivalent bits/point +MSE 290.120766 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1208 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,084B, BPFP=0.3161 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,636B, BPFP=1.6135 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,328B, BPFP=0.6566 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,752B, BPFP=1.6311 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,164B, BPFP=0.7834 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,500B, BPFP=1.5928 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,312B, BPFP=0.6541 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,532B, BPFP=1.5977 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,540B, BPFP=1.2955 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,708B, BPFP=1.4727 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,684B, BPFP=0.1665 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11527186 154.41487788 + layer.0.v_cache 0.00001510 0.01464080 + layer.1.k_cache 0.02920971 15.75178957 + layer.1.v_cache 0.00000551 0.00584475 + layer.2.k_cache 0.00857845 1.90993522 + layer.2.v_cache 0.00001837 0.01580825 + layer.3.k_cache 0.03550095 7.32344426 + layer.3.v_cache 0.00001950 0.01755591 + layer.4.k_cache 0.00063242 0.38753384 + layer.4.v_cache 0.00005150 0.03906255 + layer.4.output 11.08123415 521.60341540 + ------------------------------------------------------------------------------------- + TOTAL 4.57399661 225.35908240 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 84240 +BPFP 0.7517 bits/point +EBPFP 1.5034 equivalent bits/point +MSE 225.359082 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.006s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 225.3591 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,964B, BPFP=0.3336 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,944B, BPFP=1.6889 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,896B, BPFP=0.6617 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,580B, BPFP=1.6270 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,476B, BPFP=0.7602 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,784B, BPFP=1.4918 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,036B, BPFP=0.6855 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,244B, BPFP=1.5700 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,144B, BPFP=1.2133 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,872B, BPFP=1.5068 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,052B, BPFP=0.1711 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09673090 169.60511846 + layer.0.v_cache 0.00001344 0.01403423 + layer.1.k_cache 0.04918879 16.06882377 + layer.1.v_cache 0.00000572 0.00580610 + layer.2.k_cache 0.00880511 1.92862204 + layer.2.v_cache 0.00001944 0.01602427 + layer.3.k_cache 0.06586690 7.64597818 + layer.3.v_cache 0.00001823 0.01744024 + layer.4.k_cache 0.00062484 0.40756864 + layer.4.v_cache 0.00006325 0.03768435 + layer.4.output 0.15294319 589.59539984 + ------------------------------------------------------------------------------------- + TOTAL 0.07599641 254.28911172 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 74992 +BPFP 0.7492 bits/point +EBPFP 1.4984 equivalent bits/point +MSE 254.289112 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 254.2891 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,956B, BPFP=0.3322 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,684B, BPFP=1.6447 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,924B, BPFP=0.6664 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,624B, BPFP=1.6345 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,528B, BPFP=0.7690 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,240B, BPFP=1.5693 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,932B, BPFP=0.6678 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,456B, BPFP=1.6060 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,120B, BPFP=1.2092 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,584B, BPFP=1.4579 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,496B, BPFP=0.1819 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12676482 167.32553499 + layer.0.v_cache 0.00001339 0.01521623 + layer.1.k_cache 0.03133728 16.46129840 + layer.1.v_cache 0.00000590 0.00561846 + layer.2.k_cache 0.00372887 1.82409369 + layer.2.v_cache 0.00001820 0.01569888 + layer.3.k_cache 0.02418065 7.69160859 + layer.3.v_cache 0.00002034 0.01797245 + layer.4.k_cache 0.00062996 0.39850011 + layer.4.v_cache 0.00006053 0.03782097 + layer.4.output 0.15389095 589.46457686 + ------------------------------------------------------------------------------------- + TOTAL 0.07435274 254.12031770 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 75544 +BPFP 0.7547 bits/point +EBPFP 1.5094 equivalent bits/point +MSE 254.120318 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 254.1203 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,924B, BPFP=0.3340 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,168B, BPFP=1.7653 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,608B, BPFP=0.6264 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,444B, BPFP=1.6396 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,084B, BPFP=0.7090 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,752B, BPFP=1.5194 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,544B, BPFP=0.6153 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,252B, BPFP=1.6062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,936B, BPFP=1.2042 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,524B, BPFP=1.4799 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,148B, BPFP=0.1773 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12863920 178.62594401 + layer.0.v_cache 0.00001425 0.01460098 + layer.1.k_cache 0.01389826 16.97346191 + layer.1.v_cache 0.00000571 0.00590223 + layer.2.k_cache 0.00659207 1.85747986 + layer.2.v_cache 0.00001883 0.01605707 + layer.3.k_cache 0.05540012 7.75301921 + layer.3.v_cache 0.00001922 0.01633037 + layer.4.k_cache 0.00063279 0.39045122 + layer.4.v_cache 0.00005416 0.03653984 + layer.4.output 0.16196846 602.18958333 + ------------------------------------------------------------------------------------- + TOTAL 0.07876787 260.05981588 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 73384 +BPFP 0.7494 bits/point +EBPFP 1.4989 equivalent bits/point +MSE 260.059816 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 260.0598 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,096B, BPFP=0.3032 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,500B, BPFP=1.5191 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,140B, BPFP=0.5990 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,352B, BPFP=1.4977 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,276B, BPFP=0.7633 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,028B, BPFP=1.4508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,384B, BPFP=0.6343 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,996B, BPFP=1.4462 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,848B, BPFP=1.1354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,544B, BPFP=1.3808 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,684B, BPFP=0.1588 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13160179 165.07127098 + layer.0.v_cache 0.00001577 0.01483489 + layer.1.k_cache 0.05775581 16.77331543 + layer.1.v_cache 0.00000549 0.00574547 + layer.2.k_cache 0.01027601 1.90441414 + layer.2.v_cache 0.00002197 0.01665076 + layer.3.k_cache 0.07329086 7.53695679 + layer.3.v_cache 0.00001983 0.01878785 + layer.4.k_cache 0.00064477 0.40168133 + layer.4.v_cache 0.00005260 0.03775363 + layer.4.output 10.56448284 497.20457176 + ------------------------------------------------------------------------------------- + TOTAL 4.36618028 216.01255374 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 81848 +BPFP 0.6966 bits/point +EBPFP 1.3931 equivalent bits/point +MSE 216.012554 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 216.0126 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,868B, BPFP=0.3517 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,036B, BPFP=1.7011 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,364B, BPFP=0.6333 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,812B, BPFP=1.6589 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,724B, BPFP=0.7011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,912B, BPFP=1.4895 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,244B, BPFP=0.6107 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,092B, BPFP=1.5233 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,308B, BPFP=1.1875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,020B, BPFP=1.5098 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,720B, BPFP=0.1807 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10139442 159.36391660 + layer.0.v_cache 0.00001346 0.01473404 + layer.1.k_cache 0.03266963 16.33401673 + layer.1.v_cache 0.00000526 0.00574557 + layer.2.k_cache 0.01283502 1.98432720 + layer.2.v_cache 0.00001884 0.01549760 + layer.3.k_cache 0.02666297 8.54060162 + layer.3.v_cache 0.00001780 0.01630153 + layer.4.k_cache 0.00061512 0.39569910 + layer.4.v_cache 0.00004947 0.03691647 + layer.4.output 0.18393325 652.31599613 + ------------------------------------------------------------------------------------- + TOTAL 0.08598910 279.58351349 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 67100 +BPFP 0.7430 bits/point +EBPFP 1.4861 equivalent bits/point +MSE 279.583513 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 279.5835 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,940B, BPFP=0.3001 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,500B, BPFP=1.6244 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,320B, BPFP=0.6683 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,644B, BPFP=1.6467 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,260B, BPFP=0.8137 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,372B, BPFP=1.6046 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,276B, BPFP=0.6615 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,128B, BPFP=1.5668 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,256B, BPFP=1.2772 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,696B, BPFP=1.5000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,796B, BPFP=0.1723 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11149573 159.77624149 + layer.0.v_cache 0.00001357 0.01456228 + layer.1.k_cache 0.08092193 16.67118609 + layer.1.v_cache 0.00000571 0.00558965 + layer.2.k_cache 0.01109533 2.03396274 + layer.2.v_cache 0.00001920 0.01556064 + layer.3.k_cache 0.03617269 8.07456683 + layer.3.v_cache 0.00001858 0.01671764 + layer.4.k_cache 0.00062559 0.38845576 + layer.4.v_cache 0.00005684 0.03690888 + layer.4.output 11.30219499 532.06696429 + ------------------------------------------------------------------------------------- + TOTAL 4.66798765 230.08838247 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 83188 +BPFP 0.7570 bits/point +EBPFP 1.5141 equivalent bits/point +MSE 230.088382 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 230.0884 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,768B, BPFP=0.3542 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,996B, BPFP=1.8021 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,308B, BPFP=0.6627 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,544B, BPFP=1.7115 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,608B, BPFP=0.7228 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,680B, BPFP=1.5385 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,208B, BPFP=0.6426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,736B, BPFP=1.5497 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,988B, BPFP=1.1995 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,648B, BPFP=1.5321 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,700B, BPFP=0.1917 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10167113 169.48804337 + layer.0.v_cache 0.00001386 0.01529739 + layer.1.k_cache 0.03545215 16.99662429 + layer.1.v_cache 0.00000569 0.00574334 + layer.2.k_cache 0.01047806 2.05951026 + layer.2.v_cache 0.00001755 0.01535178 + layer.3.k_cache 0.06687863 7.95616737 + layer.3.v_cache 0.00001960 0.01755327 + layer.4.k_cache 0.00059181 0.38684410 + layer.4.v_cache 0.00004882 0.03794470 + layer.4.output 0.19633365 694.41935668 + ------------------------------------------------------------------------------------- + TOTAL 0.09350076 297.52438686 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 65184 +BPFP 0.7681 bits/point +EBPFP 1.5362 equivalent bits/point +MSE 297.524387 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 297.5244 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,900B, BPFP=0.3412 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,924B, BPFP=1.7823 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,484B, BPFP=0.6257 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,276B, BPFP=1.6659 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,888B, BPFP=0.6983 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,260B, BPFP=1.4835 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,380B, BPFP=0.6070 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,636B, BPFP=1.5510 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,456B, BPFP=1.1595 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,288B, BPFP=1.4885 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,948B, BPFP=0.1783 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11445767 165.62279993 + layer.0.v_cache 0.00001583 0.01453578 + layer.1.k_cache 0.03092440 16.48598857 + layer.1.v_cache 0.00000551 0.00571919 + layer.2.k_cache 0.00639992 1.80963801 + layer.2.v_cache 0.00001897 0.01613772 + layer.3.k_cache 0.05613507 7.06499683 + layer.3.v_cache 0.00002026 0.01802092 + layer.4.k_cache 0.00063356 0.37914430 + layer.4.v_cache 0.00005123 0.03819658 + layer.4.output 0.16055210 623.58497537 + ------------------------------------------------------------------------------------- + TOTAL 0.07838395 268.03235326 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 70440 +BPFP 0.7442 bits/point +EBPFP 1.4883 equivalent bits/point +MSE 268.032353 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 268.0324 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3452 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,388B, BPFP=1.7463 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,384B, BPFP=0.6295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,952B, BPFP=1.6652 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,760B, BPFP=0.6994 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,404B, BPFP=1.5632 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,328B, BPFP=0.6190 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,364B, BPFP=1.5558 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,412B, BPFP=1.1927 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,216B, BPFP=1.5283 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,588B, BPFP=0.1751 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13956541 164.93454706 + layer.0.v_cache 0.00001349 0.01546186 + layer.1.k_cache 0.03336124 16.42577253 + layer.1.v_cache 0.00000558 0.00630034 + layer.2.k_cache 0.00698305 2.06826455 + layer.2.v_cache 0.00001831 0.01617439 + layer.3.k_cache 0.07289400 7.94385420 + layer.3.v_cache 0.00001855 0.01692141 + layer.4.k_cache 0.00061006 0.41102932 + layer.4.v_cache 0.00005429 0.04138853 + layer.4.output 0.17891866 644.46933461 + ------------------------------------------------------------------------------------- + TOTAL 0.08858556 276.65676803 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 68652 +BPFP 0.7512 bits/point +EBPFP 1.5024 equivalent bits/point +MSE 276.656768 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 276.6568 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,804B, BPFP=0.3480 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,096B, BPFP=1.7546 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,292B, BPFP=0.6350 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,568B, BPFP=1.6528 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,684B, BPFP=0.7106 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,888B, BPFP=1.5216 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,136B, BPFP=0.6049 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,684B, BPFP=1.4823 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,840B, BPFP=1.1265 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,636B, BPFP=1.4730 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,836B, BPFP=0.1884 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10837562 167.69598765 + layer.0.v_cache 0.00001644 0.01442528 + layer.1.k_cache 0.01241452 16.88271304 + layer.1.v_cache 0.00000532 0.00570352 + layer.2.k_cache 0.01724591 1.96916086 + layer.2.v_cache 0.00001958 0.01532444 + layer.3.k_cache 0.04641535 7.91353353 + layer.3.v_cache 0.00001847 0.01600778 + layer.4.k_cache 0.00063678 0.39671085 + layer.4.v_cache 0.00004974 0.03632758 + layer.4.output 0.18014439 669.12516534 + ------------------------------------------------------------------------------------- + TOTAL 0.08507109 286.98953247 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 65464 +BPFP 0.7428 bits/point +EBPFP 1.4857 equivalent bits/point +MSE 286.989532 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 286.9895 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.7479 bits/point +Avg EBPFP 1.4958 equivalent bits/point +Avg MSE 270.808676 +Avg Time 0.370s +------------------------ ---------------------------- diff --git a/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..531333ac488784611e7988ed0af5bd4f73b07d29 --- /dev/null +++ b/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 559 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean +Output output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,532B, BPFP=0.3233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,304B, BPFP=1.5200 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,772B, BPFP=0.5545 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,200B, BPFP=1.4412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,536B, BPFP=0.6804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,716B, BPFP=1.4067 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,792B, BPFP=0.6273 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,236B, BPFP=1.4438 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,500B, BPFP=1.0345 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,808B, BPFP=1.3419 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,588B, BPFP=0.1385 +⌛️ [2/4] FRONTEND: Frontend time: 0.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.441s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12261864 157.75255066 + layer.0.v_cache 0.00001806 0.01706406 + layer.1.k_cache 0.30198812 16.07976710 + layer.1.v_cache 0.00000607 0.00593172 + layer.2.k_cache 0.01966049 1.71981937 + layer.2.v_cache 0.00002021 0.01646118 + layer.3.k_cache 0.01413035 7.32810047 + layer.3.v_cache 0.00002054 0.01787268 + layer.4.k_cache 0.00067868 0.40869594 + layer.4.v_cache 0.00004987 0.03595074 + layer.4.output 1.39792533 247.21579419 + ------------------------------------------------------------------------------------- + TOTAL 0.60262755 112.58192784 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 158984 +BPFP 0.6672 bits/point +EBPFP 1.3345 equivalent bits/point +MSE 112.581928 +---------------------- -------------------------------------------------------- +Time: 0.931s Load: 0.008s, Pack+Encode: 0.482s, Decode+Unpack: 0.441s +---------------------- -------------------------------------------------------- +💾 Converting with 112.5819 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,464B, BPFP=0.3229 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,412B, BPFP=1.5489 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,776B, BPFP=0.5625 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,684B, BPFP=1.4962 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,540B, BPFP=0.6178 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,700B, BPFP=1.4251 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,948B, BPFP=0.5749 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,896B, BPFP=1.4392 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,580B, BPFP=1.0547 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,172B, BPFP=1.3869 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,272B, BPFP=0.1268 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11787021 153.50593171 + layer.0.v_cache 0.00001738 0.01599867 + layer.1.k_cache 0.25265321 15.96426166 + layer.1.v_cache 0.00000586 0.00618250 + layer.2.k_cache 0.01010228 1.78126710 + layer.2.v_cache 0.00002017 0.01702380 + layer.3.k_cache 0.01159736 7.52301421 + layer.3.v_cache 0.00002197 0.01881231 + layer.4.k_cache 0.00067244 0.42608254 + layer.4.v_cache 0.00004845 0.03628283 + layer.4.output 1.41728832 250.54654431 + ------------------------------------------------------------------------------------- + TOTAL 0.60670750 113.71298044 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 156444 +BPFP 0.6657 bits/point +EBPFP 1.3314 equivalent bits/point +MSE 113.712980 +---------------------- -------------------------------------------------------- +Time: 0.593s Load: 0.010s, Pack+Encode: 0.256s, Decode+Unpack: 0.327s +---------------------- -------------------------------------------------------- +💾 Converting with 113.7130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,368B, BPFP=0.3061 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,820B, BPFP=1.4588 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,500B, BPFP=0.5255 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,928B, BPFP=1.3963 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,320B, BPFP=0.5830 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,312B, BPFP=1.3531 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,048B, BPFP=0.5639 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,516B, BPFP=1.4375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,496B, BPFP=1.0157 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,512B, BPFP=1.2971 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,108B, BPFP=0.1212 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13311632 157.94786996 + layer.0.v_cache 0.00001526 0.01589036 + layer.1.k_cache 0.36474110 15.75026275 + layer.1.v_cache 0.00000573 0.00599964 + layer.2.k_cache 0.01296863 1.80129444 + layer.2.v_cache 0.00001984 0.01626706 + layer.3.k_cache 0.01490937 7.45451047 + layer.3.v_cache 0.00002017 0.01771936 + layer.4.k_cache 0.00067370 0.40739550 + layer.4.v_cache 0.00005228 0.03485788 + layer.4.output 1.37278975 242.79740551 + ------------------------------------------------------------------------------------- + TOTAL 0.59623828 110.76670035 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 153928 +BPFP 0.6344 bits/point +EBPFP 1.2689 equivalent bits/point +MSE 110.766700 +---------------------- -------------------------------------------------------- +Time: 0.592s Load: 0.008s, Pack+Encode: 0.260s, Decode+Unpack: 0.324s +---------------------- -------------------------------------------------------- +💾 Converting with 110.7667 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,560B, BPFP=0.2920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,040B, BPFP=1.4114 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,808B, BPFP=0.5000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,444B, BPFP=1.3732 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,724B, BPFP=0.5587 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,516B, BPFP=1.3138 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,000B, BPFP=0.5123 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,972B, BPFP=1.3430 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,072B, BPFP=0.9652 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,520B, BPFP=1.2500 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,664B, BPFP=0.1250 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11263794 152.56131532 + layer.0.v_cache 0.00001639 0.01664261 + layer.1.k_cache 0.42101050 16.28225258 + layer.1.v_cache 0.00000577 0.00608018 + layer.2.k_cache 0.00499437 1.76079172 + layer.2.v_cache 0.00001973 0.01657853 + layer.3.k_cache 0.03029772 7.55657259 + layer.3.v_cache 0.00002073 0.01812709 + layer.4.k_cache 0.00068836 0.40969276 + layer.4.v_cache 0.00005112 0.03638107 + layer.4.output 1.25471843 221.87192623 + ------------------------------------------------------------------------------------- + TOTAL 0.55016304 101.86870106 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 162320 +BPFP 0.6114 bits/point +EBPFP 1.2229 equivalent bits/point +MSE 101.868701 +---------------------- -------------------------------------------------------- +Time: 0.561s Load: 0.010s, Pack+Encode: 0.225s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 101.8687 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,464B, BPFP=0.3200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,500B, BPFP=1.5410 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,652B, BPFP=0.5485 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,648B, BPFP=1.4799 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,832B, BPFP=0.6330 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,844B, BPFP=1.4223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,900B, BPFP=0.5662 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,444B, BPFP=1.4653 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,712B, BPFP=1.0545 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,144B, BPFP=1.3721 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,960B, BPFP=0.1327 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.323s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10233516 160.33507741 + layer.0.v_cache 0.00001822 0.01693566 + layer.1.k_cache 0.31275415 16.17027353 + layer.1.v_cache 0.00000587 0.00584066 + layer.2.k_cache 0.01225604 1.81030147 + layer.2.v_cache 0.00002018 0.01610142 + layer.3.k_cache 0.02847941 8.08489710 + layer.3.v_cache 0.00002101 0.01800817 + layer.4.k_cache 0.00067306 0.41257379 + layer.4.v_cache 0.00005947 0.03583355 + layer.4.output 1.40430263 248.22489351 + ------------------------------------------------------------------------------------- + TOTAL 0.60510241 113.20471161 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 158100 +BPFP 0.6666 bits/point +EBPFP 1.3331 equivalent bits/point +MSE 113.204712 +---------------------- -------------------------------------------------------- +Time: 0.560s Load: 0.008s, Pack+Encode: 0.228s, Decode+Unpack: 0.323s +---------------------- -------------------------------------------------------- +💾 Converting with 113.2047 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,512B, BPFP=0.2929 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,504B, BPFP=1.4617 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,852B, BPFP=0.5236 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,860B, BPFP=1.4275 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,468B, BPFP=0.6095 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,572B, BPFP=1.3591 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,720B, BPFP=0.5697 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,012B, BPFP=1.3824 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,644B, BPFP=0.9909 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,924B, BPFP=1.2715 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,432B, BPFP=0.1323 +⌛️ [2/4] FRONTEND: Frontend time: 0.333s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16298727 152.33692071 + layer.0.v_cache 0.00001770 0.01640227 + layer.1.k_cache 0.61736183 15.63832809 + layer.1.v_cache 0.00000612 0.00583801 + layer.2.k_cache 0.02593859 1.66471479 + layer.2.v_cache 0.00002024 0.01599748 + layer.3.k_cache 0.01326758 7.69319661 + layer.3.v_cache 0.00002270 0.01870518 + layer.4.k_cache 0.00075427 0.40891175 + layer.4.v_cache 0.00005043 0.03471286 + layer.4.output 0.04538776 184.50959670 + ------------------------------------------------------------------------------------- + TOTAL 0.06694947 86.43534733 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 203500 +BPFP 0.6362 bits/point +EBPFP 1.2724 equivalent bits/point +MSE 86.435347 +---------------------- -------------------------------------------------------- +Time: 0.732s Load: 0.012s, Pack+Encode: 0.333s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 86.4353 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 239, 128) +Output shape: (1, 239, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.output: torch.Size([1, 239, 3584]) -> torch.Size([1, 1, 239, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,488B, BPFP=0.2934 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,092B, BPFP=1.4443 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,780B, BPFP=0.5086 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,200B, BPFP=1.3860 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,064B, BPFP=0.5926 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,220B, BPFP=1.3219 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,304B, BPFP=0.5429 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,648B, BPFP=1.3499 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,804B, BPFP=0.9678 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,208B, BPFP=1.2558 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,804B, BPFP=0.1383 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11863919 152.34732937 + layer.0.v_cache 0.00001702 0.01702376 + layer.1.k_cache 0.43747612 15.65707129 + layer.1.v_cache 0.00000797 0.00637515 + layer.2.k_cache 0.02286501 1.67308728 + layer.2.v_cache 0.00002174 0.01594970 + layer.3.k_cache 0.01903599 7.27206510 + layer.3.v_cache 0.00002123 0.01821803 + layer.4.k_cache 0.00068597 0.41671019 + layer.4.v_cache 0.00006921 0.03594063 + layer.4.output 1.28101462 226.51651225 + ------------------------------------------------------------------------------------- + TOTAL 0.56270246 103.71031507 + (elements=2,080,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2080256 +Total Bytes 162612 +BPFP 0.6254 bits/point +EBPFP 1.2507 equivalent bits/point +MSE 103.710315 +---------------------- -------------------------------------------------------- +Time: 0.565s Load: 0.009s, Pack+Encode: 0.230s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 103.7103 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,492B, BPFP=0.3263 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,972B, BPFP=1.5430 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,440B, BPFP=0.5608 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,948B, BPFP=1.4822 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,200B, BPFP=0.6654 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,680B, BPFP=1.4068 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,544B, BPFP=0.6264 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,096B, BPFP=1.4316 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,856B, BPFP=1.0608 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,076B, BPFP=1.3710 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,808B, BPFP=0.1427 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15151563 151.63382248 + layer.0.v_cache 0.00001631 0.01649517 + layer.1.k_cache 0.46755480 15.88106360 + layer.1.v_cache 0.00000667 0.00606272 + layer.2.k_cache 0.02171374 1.71060633 + layer.2.v_cache 0.00002115 0.01669078 + layer.3.k_cache 0.01385514 7.26282899 + layer.3.v_cache 0.00002107 0.01874682 + layer.4.k_cache 0.00069081 0.40959478 + layer.4.v_cache 0.00005638 0.03819481 + layer.4.output 0.00504555 211.08052689 + ------------------------------------------------------------------------------------- + TOTAL 0.04063356 97.32692910 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 193112 +BPFP 0.6749 bits/point +EBPFP 1.3498 equivalent bits/point +MSE 97.326929 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.009s, Pack+Encode: 0.258s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 97.3269 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,240B, BPFP=0.2944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,768B, BPFP=1.4422 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,416B, BPFP=0.5150 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,168B, BPFP=1.4006 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,420B, BPFP=0.5847 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,452B, BPFP=1.3508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,224B, BPFP=0.5711 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,928B, BPFP=1.3839 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,288B, BPFP=0.9922 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,332B, BPFP=1.2731 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,332B, BPFP=0.1323 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14529029 155.18711806 + layer.0.v_cache 0.00001764 0.01659648 + layer.1.k_cache 0.31089128 16.27670139 + layer.1.v_cache 0.00000630 0.00623050 + layer.2.k_cache 0.01355181 1.74750502 + layer.2.v_cache 0.00002032 0.01667167 + layer.3.k_cache 0.02426839 7.35160916 + layer.3.v_cache 0.00002043 0.01855501 + layer.4.k_cache 0.00068925 0.39966047 + layer.4.v_cache 0.00005165 0.03541610 + layer.4.output 1.36066019 240.76263889 + ------------------------------------------------------------------------------------- + TOTAL 0.58937816 109.78791389 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 154568 +BPFP 0.6314 bits/point +EBPFP 1.2628 equivalent bits/point +MSE 109.787914 +---------------------- -------------------------------------------------------- +Time: 0.562s Load: 0.009s, Pack+Encode: 0.229s, Decode+Unpack: 0.324s +---------------------- -------------------------------------------------------- +💾 Converting with 109.7879 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 262, 128) +Output shape: (1, 262, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.output: torch.Size([1, 262, 3584]) -> torch.Size([1, 1, 262, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,424B, BPFP=0.3235 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,648B, BPFP=1.5296 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,516B, BPFP=0.5675 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,632B, BPFP=1.4690 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,368B, BPFP=0.6780 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,316B, BPFP=1.3905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,580B, BPFP=0.6310 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,992B, BPFP=1.4308 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,784B, BPFP=1.0606 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,004B, BPFP=1.3719 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,500B, BPFP=0.1406 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12917162 152.26124165 + layer.0.v_cache 0.00001830 0.01613539 + layer.1.k_cache 0.40634074 15.71042872 + layer.1.v_cache 0.00000633 0.00609763 + layer.2.k_cache 0.02363680 1.73063951 + layer.2.v_cache 0.00002248 0.01654114 + layer.3.k_cache 0.02343226 7.19689848 + layer.3.v_cache 0.00002042 0.01846661 + layer.4.k_cache 0.00071170 0.41846160 + layer.4.v_cache 0.00005093 0.03597962 + layer.4.output 0.00505940 211.88322996 + ------------------------------------------------------------------------------------- + TOTAL 0.03640161 97.68197059 + (elements=2,280,448) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2280448 +Total Bytes 191764 +BPFP 0.6727 bits/point +EBPFP 1.3454 equivalent bits/point +MSE 97.681971 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 97.6820 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,268B, BPFP=0.2964 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,076B, BPFP=1.4636 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,336B, BPFP=0.5094 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,340B, BPFP=1.4125 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,412B, BPFP=0.5842 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,632B, BPFP=1.3633 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,196B, BPFP=0.5692 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,656B, BPFP=1.3650 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,524B, BPFP=1.0086 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,052B, BPFP=1.2536 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,572B, BPFP=0.1346 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18074451 152.90875000 + layer.0.v_cache 0.00001764 0.01655591 + layer.1.k_cache 0.33421082 15.73503906 + layer.1.v_cache 0.00000655 0.00617660 + layer.2.k_cache 0.01487562 1.78891249 + layer.2.v_cache 0.00002082 0.01651136 + layer.3.k_cache 0.01319740 7.39551758 + layer.3.v_cache 0.00002096 0.01832455 + layer.4.k_cache 0.00067848 0.41934570 + layer.4.v_cache 0.00005239 0.03571862 + layer.4.output 1.36066546 240.75452381 + ------------------------------------------------------------------------------------- + TOTAL 0.59226373 109.62485403 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 155064 +BPFP 0.6334 bits/point +EBPFP 1.2669 equivalent bits/point +MSE 109.624854 +---------------------- -------------------------------------------------------- +Time: 0.561s Load: 0.009s, Pack+Encode: 0.227s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 109.6249 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 232, 128) +Output shape: (1, 232, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.output: torch.Size([1, 232, 3584]) -> torch.Size([1, 1, 232, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,520B, BPFP=0.3044 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,004B, BPFP=1.4820 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,024B, BPFP=0.5404 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,260B, BPFP=1.4318 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,456B, BPFP=0.6369 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,424B, BPFP=1.3755 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,156B, BPFP=0.6166 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,856B, BPFP=1.4046 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,024B, BPFP=1.0119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,344B, BPFP=1.3028 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,284B, BPFP=0.1471 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11978169 155.96694841 + layer.0.v_cache 0.00001746 0.01646129 + layer.1.k_cache 0.32299256 16.02605570 + layer.1.v_cache 0.00000596 0.00565658 + layer.2.k_cache 0.02091339 1.76982788 + layer.2.v_cache 0.00002092 0.01619683 + layer.3.k_cache 0.01769928 7.01681045 + layer.3.v_cache 0.00002180 0.01858093 + layer.4.k_cache 0.00071208 0.41812699 + layer.4.v_cache 0.00005763 0.03672384 + layer.4.output 1.31963615 233.46066810 + ------------------------------------------------------------------------------------- + TOTAL 0.57174564 106.79506268 + (elements=2,019,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2019328 +Total Bytes 165352 +BPFP 0.6551 bits/point +EBPFP 1.3102 equivalent bits/point +MSE 106.795063 +---------------------- -------------------------------------------------------- +Time: 0.561s Load: 0.008s, Pack+Encode: 0.227s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 106.7951 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,600B, BPFP=0.3267 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,956B, BPFP=1.4884 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,072B, BPFP=0.5733 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,408B, BPFP=1.4494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,288B, BPFP=0.6597 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,804B, BPFP=1.4065 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,964B, BPFP=0.6366 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,228B, BPFP=1.4366 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,544B, BPFP=1.0330 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,584B, BPFP=1.3199 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,772B, BPFP=0.1499 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15639258 158.09399858 + layer.0.v_cache 0.00001700 0.01707140 + layer.1.k_cache 0.33150226 16.06076216 + layer.1.v_cache 0.00000613 0.00592256 + layer.2.k_cache 0.01694729 1.61891618 + layer.2.v_cache 0.00002208 0.01655505 + layer.3.k_cache 0.03858858 7.67123247 + layer.3.v_cache 0.00002090 0.01862965 + layer.4.k_cache 0.00067984 0.41631071 + layer.4.v_cache 0.00005217 0.03771228 + layer.4.output 1.39161012 246.10385552 + ------------------------------------------------------------------------------------- + TOTAL 0.60502939 112.15788822 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 160220 +BPFP 0.6694 bits/point +EBPFP 1.3387 equivalent bits/point +MSE 112.157888 +---------------------- -------------------------------------------------------- +Time: 0.558s Load: 0.007s, Pack+Encode: 0.226s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 112.1579 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,560B, BPFP=0.3182 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,768B, BPFP=1.5321 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,356B, BPFP=0.5355 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,216B, BPFP=1.4432 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,816B, BPFP=0.6190 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,100B, BPFP=1.3793 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,064B, BPFP=0.5760 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,556B, BPFP=1.4054 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,848B, BPFP=1.0215 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,556B, BPFP=1.3482 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,840B, BPFP=0.1459 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13605902 154.96605998 + layer.0.v_cache 0.00001757 0.01619395 + layer.1.k_cache 0.58683352 16.01976734 + layer.1.v_cache 0.00000623 0.00620871 + layer.2.k_cache 0.01500385 1.74574051 + layer.2.v_cache 0.00002229 0.01639885 + layer.3.k_cache 0.00978726 7.45766619 + layer.3.v_cache 0.00001998 0.01727467 + layer.4.k_cache 0.00067570 0.39971723 + layer.4.v_cache 0.00005472 0.03518681 + layer.4.output 0.00490606 203.22373757 + ------------------------------------------------------------------------------------- + TOTAL 0.04604839 94.30861043 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 195680 +BPFP 0.6588 bits/point +EBPFP 1.3176 equivalent bits/point +MSE 94.308610 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.009s, Pack+Encode: 0.257s, Decode+Unpack: 0.388s +---------------------- -------------------------------------------------------- +💾 Converting with 94.3086 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,264B, BPFP=0.2613 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,180B, BPFP=1.3591 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,540B, BPFP=0.4620 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,272B, BPFP=1.3034 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,448B, BPFP=0.5176 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,340B, BPFP=1.2463 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,764B, BPFP=0.4757 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,792B, BPFP=1.2740 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,836B, BPFP=0.9091 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,480B, BPFP=1.1936 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,168B, BPFP=0.1240 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12967140 145.98529412 + layer.0.v_cache 0.00001601 0.01658248 + layer.1.k_cache 0.49353925 15.67825712 + layer.1.v_cache 0.00000620 0.00595455 + layer.2.k_cache 0.03189284 1.69526415 + layer.2.v_cache 0.00002040 0.01589207 + layer.3.k_cache 0.04285583 7.53976045 + layer.3.v_cache 0.00002221 0.01804782 + layer.4.k_cache 0.00071132 0.39973154 + layer.4.v_cache 0.00005107 0.03431148 + layer.4.output 1.20063613 212.30588235 + ------------------------------------------------------------------------------------- + TOTAL 0.53548467 97.50178072 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 161084 +BPFP 0.5806 bits/point +EBPFP 1.1612 equivalent bits/point +MSE 97.501781 +---------------------- -------------------------------------------------------- +Time: 0.551s Load: 0.010s, Pack+Encode: 0.228s, Decode+Unpack: 0.313s +---------------------- -------------------------------------------------------- +💾 Converting with 97.5018 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,612B, BPFP=0.2894 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,192B, BPFP=1.3926 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,960B, BPFP=0.4995 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,400B, BPFP=1.3429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,564B, BPFP=0.5374 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,436B, BPFP=1.2824 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,080B, BPFP=0.5070 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,816B, BPFP=1.3062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,044B, BPFP=0.9440 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,424B, BPFP=1.2189 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,712B, BPFP=0.1229 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16529650 152.68178966 + layer.0.v_cache 0.00001613 0.01696870 + layer.1.k_cache 0.54577502 15.86678746 + layer.1.v_cache 0.00000662 0.00627571 + layer.2.k_cache 0.01453520 1.66067444 + layer.2.v_cache 0.00002096 0.01736441 + layer.3.k_cache 0.02409829 7.21613652 + layer.3.v_cache 0.00002071 0.01872159 + layer.4.k_cache 0.00068699 0.42111191 + layer.4.v_cache 0.00004971 0.03589174 + layer.4.output 1.22953407 217.41223824 + ------------------------------------------------------------------------------------- + TOTAL 0.55042615 99.98984646 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 162240 +BPFP 0.5989 bits/point +EBPFP 1.1977 equivalent bits/point +MSE 99.989846 +---------------------- -------------------------------------------------------- +Time: 0.547s Load: 0.008s, Pack+Encode: 0.221s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 99.9898 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,564B, BPFP=0.3136 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,212B, BPFP=1.4914 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,200B, BPFP=0.5352 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,944B, BPFP=1.4308 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,344B, BPFP=0.6376 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,536B, BPFP=1.3635 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,304B, BPFP=0.5879 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,156B, BPFP=1.3932 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,052B, BPFP=1.0059 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,544B, BPFP=1.3161 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,796B, BPFP=0.1351 +⌛️ [2/4] FRONTEND: Frontend time: 0.352s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.437s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19043710 150.63750717 + layer.0.v_cache 0.00001603 0.01704196 + layer.1.k_cache 0.64205214 15.63718911 + layer.1.v_cache 0.00000630 0.00628023 + layer.2.k_cache 0.01513053 1.73074966 + layer.2.v_cache 0.00002191 0.01728660 + layer.3.k_cache 0.02775778 7.19011089 + layer.3.v_cache 0.00002125 0.01864669 + layer.4.k_cache 0.00071623 0.40972723 + layer.4.v_cache 0.00005146 0.03634264 + layer.4.output 0.04089398 165.90646844 + ------------------------------------------------------------------------------------- + TOTAL 0.06838051 78.64977419 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 230652 +BPFP 0.6483 bits/point +EBPFP 1.2966 equivalent bits/point +MSE 78.649774 +---------------------- -------------------------------------------------------- +Time: 0.800s Load: 0.011s, Pack+Encode: 0.352s, Decode+Unpack: 0.437s +---------------------- -------------------------------------------------------- +💾 Converting with 78.6498 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,504B, BPFP=0.3039 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,388B, BPFP=1.5121 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,524B, BPFP=0.5258 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,932B, BPFP=1.4318 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,144B, BPFP=0.6153 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,940B, BPFP=1.3770 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,492B, BPFP=0.5793 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,332B, BPFP=1.3986 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,240B, BPFP=1.0071 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,832B, BPFP=1.3158 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,784B, BPFP=0.1324 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14392723 160.89388251 + layer.0.v_cache 0.00001682 0.01672086 + layer.1.k_cache 0.53349789 16.12437714 + layer.1.v_cache 0.00000638 0.00606020 + layer.2.k_cache 0.01523476 1.79038576 + layer.2.v_cache 0.00002109 0.01579277 + layer.3.k_cache 0.03867486 7.90014131 + layer.3.v_cache 0.00001994 0.01758408 + layer.4.k_cache 0.00070525 0.41200453 + layer.4.v_cache 0.00005078 0.03441209 + layer.4.output 0.00472849 196.07254859 + ------------------------------------------------------------------------------------- + TOTAL 0.04501497 91.74818831 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 199112 +BPFP 0.6467 bits/point +EBPFP 1.2933 equivalent bits/point +MSE 91.748188 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.009s, Pack+Encode: 0.250s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 91.7482 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,272B, BPFP=0.2967 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,088B, BPFP=1.4644 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,428B, BPFP=0.5158 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,704B, BPFP=1.4378 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,532B, BPFP=0.5925 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,888B, BPFP=1.3811 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,036B, BPFP=0.5581 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,424B, BPFP=1.4183 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,688B, BPFP=1.0200 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,616B, BPFP=1.2928 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,860B, BPFP=0.1276 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14673459 158.35934028 + layer.0.v_cache 0.00001655 0.01649595 + layer.1.k_cache 0.37581726 15.98918945 + layer.1.v_cache 0.00000603 0.00603184 + layer.2.k_cache 0.01062915 1.82169922 + layer.2.v_cache 0.00002006 0.01672533 + layer.3.k_cache 0.04027582 7.52070475 + layer.3.v_cache 0.00002066 0.01843087 + layer.4.k_cache 0.00066373 0.40765520 + layer.4.v_cache 0.00004861 0.03601712 + layer.4.output 1.36064577 240.72791667 + ------------------------------------------------------------------------------------- + TOTAL 0.59404429 109.95810039 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 156536 +BPFP 0.6394 bits/point +EBPFP 1.2789 equivalent bits/point +MSE 109.958100 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.009s, Pack+Encode: 0.222s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 109.9581 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,480B, BPFP=0.2991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,544B, BPFP=1.5053 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,988B, BPFP=0.5334 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,332B, BPFP=1.4244 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,348B, BPFP=0.6242 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,932B, BPFP=1.3977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,216B, BPFP=0.5486 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,272B, BPFP=1.4204 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,312B, BPFP=1.0224 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,736B, BPFP=1.3178 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,840B, BPFP=0.1416 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12385076 158.17487981 + layer.0.v_cache 0.00002004 0.01875635 + layer.1.k_cache 0.38583452 15.80348140 + layer.1.v_cache 0.00000646 0.00663067 + layer.2.k_cache 0.02500911 1.76125199 + layer.2.v_cache 0.00002014 0.01688232 + layer.3.k_cache 0.03425702 7.27276768 + layer.3.v_cache 0.00002057 0.01916421 + layer.4.k_cache 0.00068928 0.43530599 + layer.4.v_cache 0.00004882 0.03620571 + layer.4.output 1.30836928 231.47514118 + ------------------------------------------------------------------------------------- + TOTAL 0.57225539 106.11007732 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 166000 +BPFP 0.6520 bits/point +EBPFP 1.3040 equivalent bits/point +MSE 106.110077 +---------------------- -------------------------------------------------------- +Time: 0.545s Load: 0.008s, Pack+Encode: 0.222s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 106.1101 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 247, 128) +Output shape: (1, 247, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.output: torch.Size([1, 247, 3584]) -> torch.Size([1, 1, 247, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,528B, BPFP=0.2864 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,380B, BPFP=1.4157 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,892B, BPFP=0.4992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,516B, BPFP=1.3611 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,688B, BPFP=0.5496 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,340B, BPFP=1.2867 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,296B, BPFP=0.5248 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,848B, BPFP=1.3188 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,972B, BPFP=0.9471 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,320B, BPFP=1.2222 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,240B, BPFP=0.1287 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13907220 147.58951164 + layer.0.v_cache 0.00001735 0.01596612 + layer.1.k_cache 0.47208408 15.88115788 + layer.1.v_cache 0.00000611 0.00567060 + layer.2.k_cache 0.03408500 1.73754463 + layer.2.v_cache 0.00001989 0.01497894 + layer.3.k_cache 0.02387486 7.11300845 + layer.3.v_cache 0.00002031 0.01691773 + layer.4.k_cache 0.00069623 0.39020251 + layer.4.v_cache 0.00004922 0.03213413 + layer.4.output 1.23951819 219.06161437 + ------------------------------------------------------------------------------------- + TOTAL 0.54979721 100.36637607 + (elements=2,149,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2149888 +Total Bytes 163020 +BPFP 0.6066 bits/point +EBPFP 1.2132 equivalent bits/point +MSE 100.366376 +---------------------- -------------------------------------------------------- +Time: 0.544s Load: 0.009s, Pack+Encode: 0.222s, Decode+Unpack: 0.313s +---------------------- -------------------------------------------------------- +💾 Converting with 100.3664 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 427, 128) +Output shape: (1, 427, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.output: torch.Size([1, 427, 3584]) -> torch.Size([1, 1, 427, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,628B, BPFP=0.2791 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,504B, BPFP=1.4090 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,544B, BPFP=0.4956 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,760B, BPFP=1.3451 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,096B, BPFP=0.5524 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,460B, BPFP=1.2610 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,760B, BPFP=0.5401 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,116B, BPFP=1.2850 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,288B, BPFP=0.9254 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,352B, BPFP=1.2204 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,224B, BPFP=0.1319 +⌛️ [2/4] FRONTEND: Frontend time: 0.473s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.503s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15738957 155.05466920 + layer.0.v_cache 0.00001693 0.01543223 + layer.1.k_cache 1.00627419 15.85429322 + layer.1.v_cache 0.00000663 0.00565115 + layer.2.k_cache 0.03590691 1.69175102 + layer.2.v_cache 0.00002108 0.01453126 + layer.3.k_cache 0.02176077 7.08537832 + layer.3.v_cache 0.00002004 0.01624158 + layer.4.k_cache 0.00078997 0.39164863 + layer.4.v_cache 0.00005212 0.03243883 + layer.4.output 0.00621442 129.82199314 + ------------------------------------------------------------------------------------- + TOTAL 0.07445524 64.05388161 + (elements=3,716,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3716608 +Total Bytes 279732 +BPFP 0.6021 bits/point +EBPFP 1.2042 equivalent bits/point +MSE 64.053882 +---------------------- -------------------------------------------------------- +Time: 0.990s Load: 0.014s, Pack+Encode: 0.473s, Decode+Unpack: 0.503s +---------------------- -------------------------------------------------------- +💾 Converting with 64.0539 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,536B, BPFP=0.3192 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,248B, BPFP=1.5134 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,336B, BPFP=0.5383 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,180B, BPFP=1.4518 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,732B, BPFP=0.6188 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,716B, BPFP=1.3674 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,152B, BPFP=0.5853 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,896B, BPFP=1.3778 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,424B, BPFP=1.0046 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,616B, BPFP=1.3040 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,264B, BPFP=0.1340 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15504150 155.34530673 + layer.0.v_cache 0.00001776 0.01544738 + layer.1.k_cache 0.59434689 16.13960339 + layer.1.v_cache 0.00000612 0.00543631 + layer.2.k_cache 0.01687181 1.63302624 + layer.2.v_cache 0.00002031 0.01481186 + layer.3.k_cache 0.01346795 7.17776861 + layer.3.v_cache 0.00002023 0.01651039 + layer.4.k_cache 0.00071492 0.38970339 + layer.4.v_cache 0.00006004 0.03225495 + layer.4.output 0.00488892 204.70800936 + ------------------------------------------------------------------------------------- + TOTAL 0.04792882 94.92505499 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 191100 +BPFP 0.6481 bits/point +EBPFP 1.2963 equivalent bits/point +MSE 94.925055 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.012s, Pack+Encode: 0.257s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 94.9251 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 361, 128) +Output shape: (1, 361, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.output: torch.Size([1, 361, 3584]) -> torch.Size([1, 1, 361, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,596B, BPFP=0.2855 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,188B, BPFP=1.3932 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,488B, BPFP=0.4972 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,028B, BPFP=1.3430 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,468B, BPFP=0.5829 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,464B, BPFP=1.2753 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,036B, BPFP=0.5209 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,692B, BPFP=1.2851 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,624B, BPFP=0.9359 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,968B, BPFP=1.2105 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,824B, BPFP=0.1226 +⌛️ [2/4] FRONTEND: Frontend time: 0.307s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.442s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21057716 152.66473338 + layer.0.v_cache 0.00001554 0.01449275 + layer.1.k_cache 0.82271198 15.86647469 + layer.1.v_cache 0.00000592 0.00547762 + layer.2.k_cache 0.01184860 1.67251866 + layer.2.v_cache 0.00001987 0.01492223 + layer.3.k_cache 0.01345186 7.28005973 + layer.3.v_cache 0.00002111 0.01631685 + layer.4.k_cache 0.00076976 0.39302219 + layer.4.v_cache 0.00005077 0.03208880 + layer.4.output 0.03703259 150.26780768 + ------------------------------------------------------------------------------------- + TOTAL 0.07757063 72.34322121 + (elements=3,142,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3142144 +Total Bytes 235376 +BPFP 0.5993 bits/point +EBPFP 1.1985 equivalent bits/point +MSE 72.343221 +---------------------- -------------------------------------------------------- +Time: 0.764s Load: 0.015s, Pack+Encode: 0.307s, Decode+Unpack: 0.442s +---------------------- -------------------------------------------------------- +💾 Converting with 72.3432 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,500B, BPFP=0.3047 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,204B, BPFP=1.4519 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,384B, BPFP=0.5199 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,832B, BPFP=1.4313 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,416B, BPFP=0.5771 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,412B, BPFP=1.3526 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,796B, BPFP=0.5428 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,928B, BPFP=1.3812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,932B, BPFP=0.9936 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,960B, BPFP=1.2722 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,608B, BPFP=0.1315 +⌛️ [2/4] FRONTEND: Frontend time: 0.272s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21757852 160.76699634 + layer.0.v_cache 0.00001544 0.01388390 + layer.1.k_cache 0.59705023 16.29346049 + layer.1.v_cache 0.00000603 0.00520068 + layer.2.k_cache 0.00787334 1.68727620 + layer.2.v_cache 0.00001970 0.01432702 + layer.3.k_cache 0.01914395 7.93600832 + layer.3.v_cache 0.00001916 0.01515964 + layer.4.k_cache 0.00071861 0.37866260 + layer.4.v_cache 0.00005003 0.03172701 + layer.4.output 0.00471288 196.68699341 + ------------------------------------------------------------------------------------- + TOTAL 0.05149795 91.99715624 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 193972 +BPFP 0.6322 bits/point +EBPFP 1.2644 equivalent bits/point +MSE 91.997156 +---------------------- -------------------------------------------------------- +Time: 0.668s Load: 0.011s, Pack+Encode: 0.272s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 91.9972 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.018s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 372, 128) +Output shape: (1, 372, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.output: torch.Size([1, 372, 3584]) -> torch.Size([1, 1, 372, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,644B, BPFP=0.2791 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,632B, BPFP=1.3706 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,452B, BPFP=0.4810 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,436B, BPFP=1.3204 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,548B, BPFP=0.5270 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,612B, BPFP=1.2438 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,920B, BPFP=0.5007 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,904B, BPFP=1.2560 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,540B, BPFP=0.9047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,288B, BPFP=1.1882 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,312B, BPFP=0.1339 +⌛️ [2/4] FRONTEND: Frontend time: 0.302s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.444s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18192377 150.88085938 + layer.0.v_cache 0.00001786 0.01555921 + layer.1.k_cache 0.82695811 15.98645151 + layer.1.v_cache 0.00000687 0.00581620 + layer.2.k_cache 0.01216373 1.75568086 + layer.2.v_cache 0.00002073 0.01511944 + layer.3.k_cache 0.01743993 6.74973486 + layer.3.v_cache 0.00002103 0.01640481 + layer.4.k_cache 0.00069676 0.38488798 + layer.4.v_cache 0.00005544 0.03293395 + layer.4.output 0.03606073 145.77722974 + ------------------------------------------------------------------------------------- + TOTAL 0.07598408 70.36965038 + (elements=3,237,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3237888 +Total Bytes 238288 +BPFP 0.5887 bits/point +EBPFP 1.1775 equivalent bits/point +MSE 70.369650 +---------------------- -------------------------------------------------------- +Time: 0.764s Load: 0.018s, Pack+Encode: 0.302s, Decode+Unpack: 0.444s +---------------------- -------------------------------------------------------- +💾 Converting with 70.3697 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,316B, BPFP=0.3055 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,080B, BPFP=1.4551 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,080B, BPFP=0.5360 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,988B, BPFP=1.4023 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,960B, BPFP=0.6269 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,380B, BPFP=1.3245 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,980B, BPFP=0.5795 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,632B, BPFP=1.3367 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,256B, BPFP=0.9799 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,412B, BPFP=1.2777 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,168B, BPFP=0.1463 +⌛️ [2/4] FRONTEND: Frontend time: 0.291s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.441s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16470559 154.26774139 + layer.0.v_cache 0.00001558 0.01558886 + layer.1.k_cache 0.70577214 15.97929113 + layer.1.v_cache 0.00000663 0.00558010 + layer.2.k_cache 0.01487649 1.71821656 + layer.2.v_cache 0.00002061 0.01484820 + layer.3.k_cache 0.03012228 7.09947483 + layer.3.v_cache 0.00002030 0.01645001 + layer.4.k_cache 0.00074575 0.39878396 + layer.4.v_cache 0.00005065 0.03182069 + layer.4.output 0.04139163 167.91636720 + ------------------------------------------------------------------------------------- + TOTAL 0.07094573 79.70366860 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 224252 +BPFP 0.6381 bits/point +EBPFP 1.2762 equivalent bits/point +MSE 79.703669 +---------------------- -------------------------------------------------------- +Time: 0.743s Load: 0.011s, Pack+Encode: 0.291s, Decode+Unpack: 0.441s +---------------------- -------------------------------------------------------- +💾 Converting with 79.7037 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,452B, BPFP=0.3215 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,548B, BPFP=1.5064 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,316B, BPFP=0.5493 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,636B, BPFP=1.4526 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,212B, BPFP=0.6611 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,248B, BPFP=1.3708 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,048B, BPFP=0.5925 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,648B, BPFP=1.3943 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,428B, BPFP=1.0276 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,608B, BPFP=1.3330 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,608B, BPFP=0.1399 +⌛️ [2/4] FRONTEND: Frontend time: 0.269s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16585089 152.54535672 + layer.0.v_cache 0.00001549 0.01511702 + layer.1.k_cache 0.52552974 16.00632370 + layer.1.v_cache 0.00000676 0.00565358 + layer.2.k_cache 0.01473994 1.74657524 + layer.2.v_cache 0.00002188 0.01568602 + layer.3.k_cache 0.05129043 7.59939379 + layer.3.v_cache 0.00002012 0.01739055 + layer.4.k_cache 0.00075129 0.40540607 + layer.4.v_cache 0.00005717 0.03443361 + layer.4.output 0.00501330 209.50192049 + ------------------------------------------------------------------------------------- + TOTAL 0.04666922 96.75910469 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 189752 +BPFP 0.6581 bits/point +EBPFP 1.3163 equivalent bits/point +MSE 96.759105 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.012s, Pack+Encode: 0.269s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 96.7591 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 368, 128) +Output shape: (1, 368, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.output: torch.Size([1, 368, 3584]) -> torch.Size([1, 1, 368, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,560B, BPFP=0.2785 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,556B, BPFP=1.3823 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,332B, BPFP=0.4811 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,460B, BPFP=1.3358 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,752B, BPFP=0.5414 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,492B, BPFP=1.2522 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,544B, BPFP=0.5326 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,176B, BPFP=1.2812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,616B, BPFP=0.9178 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,516B, BPFP=1.2108 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,768B, BPFP=0.1320 +⌛️ [2/4] FRONTEND: Frontend time: 0.289s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.432s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14147143 151.96259341 + layer.0.v_cache 0.00001611 0.01485819 + layer.1.k_cache 0.89062865 15.91227125 + layer.1.v_cache 0.00000637 0.00539656 + layer.2.k_cache 0.02201829 1.63845411 + layer.2.v_cache 0.00002088 0.01478013 + layer.3.k_cache 0.00965744 7.11117554 + layer.3.v_cache 0.00002101 0.01649060 + layer.4.k_cache 0.00077426 0.39460394 + layer.4.v_cache 0.00005366 0.03375501 + layer.4.output 0.03638816 147.33813325 + ------------------------------------------------------------------------------------- + TOTAL 0.07761090 71.08654773 + (elements=3,203,072) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3203072 +Total Bytes 238772 +BPFP 0.5964 bits/point +EBPFP 1.1927 equivalent bits/point +MSE 71.086548 +---------------------- -------------------------------------------------------- +Time: 0.733s Load: 0.012s, Pack+Encode: 0.289s, Decode+Unpack: 0.432s +---------------------- -------------------------------------------------------- +💾 Converting with 71.0865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 453, 128) +Output shape: (1, 453, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.output: torch.Size([1, 453, 3584]) -> torch.Size([1, 1, 453, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,604B, BPFP=0.2968 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,784B, BPFP=1.4412 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,840B, BPFP=0.5119 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,740B, BPFP=1.3707 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,552B, BPFP=0.5709 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,764B, BPFP=1.3026 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,540B, BPFP=0.5360 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,624B, BPFP=1.3322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,120B, BPFP=0.9699 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,812B, BPFP=1.2697 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,824B, BPFP=0.1272 +⌛️ [2/4] FRONTEND: Frontend time: 0.548s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.577s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13665763 150.47178532 + layer.0.v_cache 0.00001897 0.01613903 + layer.1.k_cache 1.07282356 15.97275542 + layer.1.v_cache 0.00000630 0.00561446 + layer.2.k_cache 0.02241544 1.76104575 + layer.2.v_cache 0.00001979 0.01505584 + layer.3.k_cache 0.01806638 7.22113805 + layer.3.v_cache 0.00002079 0.01715590 + layer.4.k_cache 0.00076567 0.40267554 + layer.4.v_cache 0.00005104 0.03365758 + layer.4.output 0.00584071 122.38336684 + ------------------------------------------------------------------------------------- + TOTAL 0.07598415 60.74121122 + (elements=3,942,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3942912 +Total Bytes 304204 +BPFP 0.6172 bits/point +EBPFP 1.2344 equivalent bits/point +MSE 60.741211 +---------------------- -------------------------------------------------------- +Time: 1.140s Load: 0.016s, Pack+Encode: 0.548s, Decode+Unpack: 0.577s +---------------------- -------------------------------------------------------- +💾 Converting with 60.7412 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 402, 128) +Output shape: (1, 402, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.output: torch.Size([1, 402, 3584]) -> torch.Size([1, 1, 402, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,612B, BPFP=0.2959 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,012B, BPFP=1.4386 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,976B, BPFP=0.5044 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,808B, BPFP=1.3529 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,876B, BPFP=0.5782 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,820B, BPFP=1.2757 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,792B, BPFP=0.5361 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,592B, BPFP=1.3057 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,940B, BPFP=0.9305 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,768B, BPFP=1.2348 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,928B, BPFP=0.1218 +⌛️ [2/4] FRONTEND: Frontend time: 0.346s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.505s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18062694 150.75335238 + layer.0.v_cache 0.00001752 0.01520989 + layer.1.k_cache 0.89169403 15.75583508 + layer.1.v_cache 0.00000593 0.00510936 + layer.2.k_cache 0.03854401 1.68527419 + layer.2.v_cache 0.00001967 0.01419704 + layer.3.k_cache 0.02336847 7.69984732 + layer.3.v_cache 0.00001924 0.01604570 + layer.4.k_cache 0.00085529 0.39022751 + layer.4.v_cache 0.00004959 0.03079772 + layer.4.output 0.00648942 137.80146144 + ------------------------------------------------------------------------------------- + TOTAL 0.06944863 67.11624272 + (elements=3,499,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3499008 +Total Bytes 265124 +BPFP 0.6062 bits/point +EBPFP 1.2123 equivalent bits/point +MSE 67.116243 +---------------------- -------------------------------------------------------- +Time: 0.867s Load: 0.015s, Pack+Encode: 0.346s, Decode+Unpack: 0.505s +---------------------- -------------------------------------------------------- +💾 Converting with 67.1162 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,456B, BPFP=0.3056 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,216B, BPFP=1.5242 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,504B, BPFP=0.5323 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,856B, BPFP=1.4480 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,064B, BPFP=0.6196 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,480B, BPFP=1.3710 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,568B, BPFP=0.5918 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,004B, BPFP=1.4003 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,200B, BPFP=1.0193 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,184B, BPFP=1.3544 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,496B, BPFP=0.1320 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13958113 149.35157650 + layer.0.v_cache 0.00001688 0.01687515 + layer.1.k_cache 0.56864935 15.90332731 + layer.1.v_cache 0.00000621 0.00649038 + layer.2.k_cache 0.01018704 1.82125986 + layer.2.v_cache 0.00002046 0.01707604 + layer.3.k_cache 0.04374025 7.79851661 + layer.3.v_cache 0.00002178 0.01925223 + layer.4.k_cache 0.00068741 0.41874181 + layer.4.v_cache 0.00005272 0.03731154 + layer.4.output 0.00479460 198.76344086 + ------------------------------------------------------------------------------------- + TOTAL 0.04685444 92.16085373 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 198028 +BPFP 0.6524 bits/point +EBPFP 1.3047 equivalent bits/point +MSE 92.160854 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.256s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 92.1609 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 381, 128) +Output shape: (1, 381, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.output: torch.Size([1, 381, 3584]) -> torch.Size([1, 1, 381, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,800B, BPFP=0.2789 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,288B, BPFP=1.3652 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,756B, BPFP=0.4821 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,964B, BPFP=1.3109 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,024B, BPFP=0.5341 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,416B, BPFP=1.2474 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,000B, BPFP=0.4921 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,124B, BPFP=1.2764 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,292B, BPFP=0.9142 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,236B, BPFP=1.1990 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,444B, BPFP=0.1315 +⌛️ [2/4] FRONTEND: Frontend time: 0.297s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.433s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15713709 148.74527354 + layer.0.v_cache 0.00001719 0.01634375 + layer.1.k_cache 0.86767426 15.46591259 + layer.1.v_cache 0.00000622 0.00590775 + layer.2.k_cache 0.01635739 1.69574631 + layer.2.v_cache 0.00002098 0.01613980 + layer.3.k_cache 0.01709203 7.18475870 + layer.3.v_cache 0.00002073 0.01759508 + layer.4.k_cache 0.00073405 0.40515009 + layer.4.v_cache 0.00005083 0.03460336 + layer.4.output 0.03519400 142.26961474 + ------------------------------------------------------------------------------------- + TOTAL 0.07679228 68.79263142 + (elements=3,316,224) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3316224 +Total Bytes 244344 +BPFP 0.5895 bits/point +EBPFP 1.1789 equivalent bits/point +MSE 68.792631 +---------------------- -------------------------------------------------------- +Time: 0.746s Load: 0.016s, Pack+Encode: 0.297s, Decode+Unpack: 0.433s +---------------------- -------------------------------------------------------- +💾 Converting with 68.7926 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,516B, BPFP=0.3015 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,248B, BPFP=1.4856 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,064B, BPFP=0.5385 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,368B, BPFP=1.4268 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,016B, BPFP=0.6020 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,596B, BPFP=1.3753 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,520B, BPFP=0.5689 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,984B, BPFP=1.4012 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,268B, BPFP=1.0195 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,700B, BPFP=1.3154 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,020B, BPFP=0.1433 +⌛️ [2/4] FRONTEND: Frontend time: 0.239s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422791 157.66326122 + layer.0.v_cache 0.00001930 0.01699563 + layer.1.k_cache 0.38928963 15.92644168 + layer.1.v_cache 0.00000631 0.00603570 + layer.2.k_cache 0.01623082 1.73223316 + layer.2.v_cache 0.00002095 0.01680659 + layer.3.k_cache 0.01313608 7.17114727 + layer.3.v_cache 0.00002100 0.01862254 + layer.4.k_cache 0.00066994 0.41469124 + layer.4.v_cache 0.00005661 0.03703780 + layer.4.output 1.30836332 231.47556090 + ------------------------------------------------------------------------------------- + TOTAL 0.57013069 106.07836465 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 165300 +BPFP 0.6493 bits/point +EBPFP 1.2985 equivalent bits/point +MSE 106.078365 +---------------------- -------------------------------------------------------- +Time: 0.563s Load: 0.009s, Pack+Encode: 0.239s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 106.0784 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,520B, BPFP=0.2993 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,012B, BPFP=1.4574 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,860B, BPFP=0.5204 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,168B, BPFP=1.4015 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,896B, BPFP=0.5890 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,176B, BPFP=1.3358 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,440B, BPFP=0.5588 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,816B, BPFP=1.3782 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,948B, BPFP=0.9897 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,164B, BPFP=1.2688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,512B, BPFP=0.1373 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13591524 158.23659296 + layer.0.v_cache 0.00001656 0.01692188 + layer.1.k_cache 0.34520291 15.92849990 + layer.1.v_cache 0.00000598 0.00623169 + layer.2.k_cache 0.01250870 1.70625021 + layer.2.v_cache 0.00001982 0.01647999 + layer.3.k_cache 0.02633334 7.09770384 + layer.3.v_cache 0.00002177 0.01915110 + layer.4.k_cache 0.00067446 0.41505060 + layer.4.v_cache 0.00005295 0.03663424 + layer.4.output 1.29725540 229.48149970 + ------------------------------------------------------------------------------------- + TOTAL 0.56479644 105.28529496 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 162512 +BPFP 0.6329 bits/point +EBPFP 1.2658 equivalent bits/point +MSE 105.285295 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.009s, Pack+Encode: 0.223s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 105.2853 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,400B, BPFP=0.3168 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,560B, BPFP=1.5524 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,572B, BPFP=0.5452 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,836B, BPFP=1.5003 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,824B, BPFP=0.6354 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,004B, BPFP=1.4404 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,524B, BPFP=0.6138 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,344B, BPFP=1.4649 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,652B, BPFP=1.0550 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,088B, BPFP=1.3744 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,844B, BPFP=0.1321 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150146 155.94453845 + layer.0.v_cache 0.00001564 0.01609297 + layer.1.k_cache 0.36060091 16.08150134 + layer.1.v_cache 0.00000598 0.00598335 + layer.2.k_cache 0.01419523 1.63812706 + layer.2.v_cache 0.00002081 0.01620750 + layer.3.k_cache 0.03680107 7.46697070 + layer.3.v_cache 0.00002092 0.01754468 + layer.4.k_cache 0.00067164 0.40878929 + layer.4.v_cache 0.00004838 0.03459482 + layer.4.output 1.41076765 249.36773782 + ------------------------------------------------------------------------------------- + TOTAL 0.61289739 113.36497147 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 158648 +BPFP 0.6720 bits/point +EBPFP 1.3439 equivalent bits/point +MSE 113.364971 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.008s, Pack+Encode: 0.222s, Decode+Unpack: 0.324s +---------------------- -------------------------------------------------------- +💾 Converting with 113.3650 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 343, 128) +Output shape: (1, 343, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.output: torch.Size([1, 343, 3584]) -> torch.Size([1, 1, 343, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,464B, BPFP=0.2945 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,708B, BPFP=1.4900 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,180B, BPFP=0.5093 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,336B, BPFP=1.4275 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,768B, BPFP=0.5816 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,680B, BPFP=1.3520 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,136B, BPFP=0.5528 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,484B, BPFP=1.3887 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,780B, BPFP=0.9922 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,796B, BPFP=1.3118 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,240B, BPFP=0.1382 +⌛️ [2/4] FRONTEND: Frontend time: 0.294s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.432s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16677478 148.64357462 + layer.0.v_cache 0.00001698 0.01624560 + layer.1.k_cache 0.72144124 16.08083688 + layer.1.v_cache 0.00000659 0.00610119 + layer.2.k_cache 0.01351720 1.73099689 + layer.2.v_cache 0.00002117 0.01638615 + layer.3.k_cache 0.02347133 7.18723451 + layer.3.v_cache 0.00002006 0.01802325 + layer.4.k_cache 0.00069292 0.39878632 + layer.4.v_cache 0.00005144 0.03500555 + layer.4.output 0.03905215 157.99642076 + ------------------------------------------------------------------------------------- + TOTAL 0.07055169 75.30047860 + (elements=2,985,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2985472 +Total Bytes 238572 +BPFP 0.6393 bits/point +EBPFP 1.2786 equivalent bits/point +MSE 75.300479 +---------------------- -------------------------------------------------------- +Time: 0.740s Load: 0.013s, Pack+Encode: 0.294s, Decode+Unpack: 0.432s +---------------------- -------------------------------------------------------- +💾 Converting with 75.3005 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,468B, BPFP=0.3107 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,292B, BPFP=1.4939 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,420B, BPFP=0.5352 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,004B, BPFP=1.4207 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,836B, BPFP=0.6157 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,476B, BPFP=1.3339 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,440B, BPFP=0.5932 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,192B, BPFP=1.3745 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,540B, BPFP=0.9966 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,928B, BPFP=1.3027 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,060B, BPFP=0.1385 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13317820 152.33384943 + layer.0.v_cache 0.00001784 0.01615277 + layer.1.k_cache 0.51782770 16.32082031 + layer.1.v_cache 0.00000612 0.00593287 + layer.2.k_cache 0.01694398 1.73227805 + layer.2.v_cache 0.00002016 0.01575243 + layer.3.k_cache 0.01920085 7.32230247 + layer.3.v_cache 0.00002016 0.01721987 + layer.4.k_cache 0.00071488 0.40026123 + layer.4.v_cache 0.00005349 0.03477297 + layer.4.output 0.00484358 201.73610390 + ------------------------------------------------------------------------------------- + TOTAL 0.04246403 93.55012175 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 192656 +BPFP 0.6439 bits/point +EBPFP 1.2878 equivalent bits/point +MSE 93.550122 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.009s, Pack+Encode: 0.258s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 93.5501 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 374, 128) +Output shape: (1, 374, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.output: torch.Size([1, 374, 3584]) -> torch.Size([1, 1, 374, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,732B, BPFP=0.2812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,948B, BPFP=1.3765 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,400B, BPFP=0.4763 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,724B, BPFP=1.3254 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,424B, BPFP=0.5191 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,676B, BPFP=1.2398 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,948B, BPFP=0.4992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,260B, BPFP=1.2642 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,452B, BPFP=0.8962 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,216B, BPFP=1.1788 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,888B, BPFP=0.1187 +⌛️ [2/4] FRONTEND: Frontend time: 0.279s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.432s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18781370 149.62747535 + layer.0.v_cache 0.00001670 0.01577256 + layer.1.k_cache 0.85217587 15.69838972 + layer.1.v_cache 0.00000605 0.00576985 + layer.2.k_cache 0.02919085 1.58459424 + layer.2.v_cache 0.00002159 0.01460787 + layer.3.k_cache 0.02976600 7.28908535 + layer.3.v_cache 0.00002022 0.01665704 + layer.4.k_cache 0.00072409 0.38996100 + layer.4.v_cache 0.00004949 0.03262690 + layer.4.output 0.03578218 144.95668210 + ------------------------------------------------------------------------------------- + TOTAL 0.07942705 69.96304204 + (elements=3,255,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3255296 +Total Bytes 236668 +BPFP 0.5816 bits/point +EBPFP 1.1632 equivalent bits/point +MSE 69.963042 +---------------------- -------------------------------------------------------- +Time: 0.724s Load: 0.013s, Pack+Encode: 0.279s, Decode+Unpack: 0.432s +---------------------- -------------------------------------------------------- +💾 Converting with 69.9630 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,460B, BPFP=0.2892 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,876B, BPFP=1.4183 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,752B, BPFP=0.5026 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,912B, BPFP=1.3558 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,612B, BPFP=0.5584 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,104B, BPFP=1.3034 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,080B, BPFP=0.5239 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,464B, BPFP=1.3268 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,660B, BPFP=0.9505 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,092B, BPFP=1.2378 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,236B, BPFP=0.1319 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13979621 152.25792596 + layer.0.v_cache 0.00001743 0.01565443 + layer.1.k_cache 0.46959303 15.94476648 + layer.1.v_cache 0.00000602 0.00588428 + layer.2.k_cache 0.01642122 1.80702203 + layer.2.v_cache 0.00002053 0.01588035 + layer.3.k_cache 0.03876302 7.56396383 + layer.3.v_cache 0.00002036 0.01813096 + layer.4.k_cache 0.00069112 0.41429277 + layer.4.v_cache 0.00005399 0.03571686 + layer.4.output 1.27033806 224.58372851 + ------------------------------------------------------------------------------------- + TOTAL 0.56222055 102.95090221 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 160248 +BPFP 0.6111 bits/point +EBPFP 1.2223 equivalent bits/point +MSE 102.950902 +---------------------- -------------------------------------------------------- +Time: 0.543s Load: 0.010s, Pack+Encode: 0.220s, Decode+Unpack: 0.312s +---------------------- -------------------------------------------------------- +💾 Converting with 102.9509 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,468B, BPFP=0.3075 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,296B, BPFP=1.4659 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,860B, BPFP=0.5410 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,408B, BPFP=1.4047 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,864B, BPFP=0.6101 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,780B, BPFP=1.3615 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,048B, BPFP=0.5540 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,256B, BPFP=1.3943 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,548B, BPFP=1.0014 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,840B, BPFP=1.2968 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,512B, BPFP=0.1427 +⌛️ [2/4] FRONTEND: Frontend time: 0.237s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13893990 153.26228662 + layer.0.v_cache 0.00001711 0.01649981 + layer.1.k_cache 0.36820191 15.89897310 + layer.1.v_cache 0.00000637 0.00627196 + layer.2.k_cache 0.01566778 1.76804826 + layer.2.v_cache 0.00002161 0.01710226 + layer.3.k_cache 0.01884081 6.97425768 + layer.3.v_cache 0.00002076 0.01890632 + layer.4.k_cache 0.00070054 0.42494306 + layer.4.v_cache 0.00005210 0.03641368 + layer.4.output 1.34866800 238.57380821 + ------------------------------------------------------------------------------------- + TOTAL 0.58724382 108.73178590 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 158880 +BPFP 0.6433 bits/point +EBPFP 1.2866 equivalent bits/point +MSE 108.731786 +---------------------- -------------------------------------------------------- +Time: 0.557s Load: 0.007s, Pack+Encode: 0.237s, Decode+Unpack: 0.313s +---------------------- -------------------------------------------------------- +💾 Converting with 108.7318 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,484B, BPFP=0.3209 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,800B, BPFP=1.5098 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,176B, BPFP=0.5370 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,632B, BPFP=1.4415 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,900B, BPFP=0.6379 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,468B, BPFP=1.3734 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,064B, BPFP=0.5890 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,116B, BPFP=1.4113 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,516B, BPFP=1.0250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,140B, BPFP=1.3542 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,880B, BPFP=0.1328 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11943251 161.51226007 + layer.0.v_cache 0.00001561 0.01625490 + layer.1.k_cache 0.49619702 16.09596647 + layer.1.v_cache 0.00000608 0.00600622 + layer.2.k_cache 0.01349212 1.70674436 + layer.2.v_cache 0.00002006 0.01638411 + layer.3.k_cache 0.01080775 7.15465440 + layer.3.v_cache 0.00002073 0.01737662 + layer.4.k_cache 0.00068886 0.42103337 + layer.4.v_cache 0.00007406 0.03665161 + layer.4.output 0.00494179 207.85771134 + ------------------------------------------------------------------------------------- + TOTAL 0.03972631 96.58748891 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 190176 +BPFP 0.6547 bits/point +EBPFP 1.3093 equivalent bits/point +MSE 96.587489 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 96.5875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,408B, BPFP=0.3117 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,988B, BPFP=1.4839 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,672B, BPFP=0.5424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,228B, BPFP=1.4301 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,664B, BPFP=0.6126 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,592B, BPFP=1.3852 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,912B, BPFP=0.5594 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,324B, BPFP=1.4369 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,564B, BPFP=1.0297 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,408B, BPFP=1.3015 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,804B, BPFP=0.1394 +⌛️ [2/4] FRONTEND: Frontend time: 0.218s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11815372 160.76551895 + layer.0.v_cache 0.00001627 0.01653504 + layer.1.k_cache 0.37623358 16.25731538 + layer.1.v_cache 0.00000616 0.00625204 + layer.2.k_cache 0.01429399 1.77976030 + layer.2.v_cache 0.00002051 0.01691003 + layer.3.k_cache 0.01320818 7.13729320 + layer.3.v_cache 0.00002004 0.01832427 + layer.4.k_cache 0.00067885 0.41184393 + layer.4.v_cache 0.00005318 0.03730214 + layer.4.output 1.38524649 244.86908129 + ------------------------------------------------------------------------------------- + TOTAL 0.60114176 111.79591907 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 156564 +BPFP 0.6511 bits/point +EBPFP 1.3023 equivalent bits/point +MSE 111.795919 +---------------------- -------------------------------------------------------- +Time: 0.537s Load: 0.007s, Pack+Encode: 0.218s, Decode+Unpack: 0.312s +---------------------- -------------------------------------------------------- +💾 Converting with 111.7959 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,540B, BPFP=0.3114 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,628B, BPFP=1.4966 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,468B, BPFP=0.5321 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,612B, BPFP=1.4395 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,596B, BPFP=0.5955 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,308B, BPFP=1.3662 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,744B, BPFP=0.5477 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,104B, BPFP=1.4110 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,080B, BPFP=1.0162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,736B, BPFP=1.3341 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,400B, BPFP=0.1237 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09820673 153.67909735 + layer.0.v_cache 0.00001666 0.01513810 + layer.1.k_cache 0.50544519 16.45004426 + layer.1.v_cache 0.00000613 0.00568125 + layer.2.k_cache 0.01930715 1.78295053 + layer.2.v_cache 0.00002033 0.01541320 + layer.3.k_cache 0.01337509 7.49740403 + layer.3.v_cache 0.00001987 0.01742422 + layer.4.k_cache 0.00074230 0.40435045 + layer.4.v_cache 0.00004869 0.03387697 + layer.4.output 0.00473525 199.59859969 + ------------------------------------------------------------------------------------- + TOTAL 0.03943146 92.77009284 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 194216 +BPFP 0.6421 bits/point +EBPFP 1.2842 equivalent bits/point +MSE 92.770093 +---------------------- -------------------------------------------------------- +Time: 0.627s Load: 0.010s, Pack+Encode: 0.246s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 92.7701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,424B, BPFP=0.3171 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,928B, BPFP=1.5000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,644B, BPFP=0.5479 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,184B, BPFP=1.4467 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,652B, BPFP=0.6201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,172B, BPFP=1.3741 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,392B, BPFP=0.6015 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,128B, BPFP=1.4427 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,648B, BPFP=1.0499 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,004B, BPFP=1.3621 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,220B, BPFP=0.1251 +⌛️ [2/4] FRONTEND: Frontend time: 0.218s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13853801 158.70839306 + layer.0.v_cache 0.00001803 0.01648967 + layer.1.k_cache 0.23740530 16.39421320 + layer.1.v_cache 0.00000581 0.00607370 + layer.2.k_cache 0.01896443 1.79436052 + layer.2.v_cache 0.00001977 0.01661891 + layer.3.k_cache 0.00876656 7.85428948 + layer.3.v_cache 0.00002034 0.01889110 + layer.4.k_cache 0.00073522 0.41411045 + layer.4.v_cache 0.00004914 0.03668925 + layer.4.output 1.40427464 248.21690695 + ------------------------------------------------------------------------------------- + TOTAL 0.60202618 113.10461635 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 155396 +BPFP 0.6552 bits/point +EBPFP 1.3103 equivalent bits/point +MSE 113.104616 +---------------------- -------------------------------------------------------- +Time: 0.544s Load: 0.009s, Pack+Encode: 0.218s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 113.1046 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,556B, BPFP=0.3145 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,156B, BPFP=1.4808 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,372B, BPFP=0.5306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,908B, BPFP=1.4101 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,216B, BPFP=0.5784 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,012B, BPFP=1.3594 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,800B, BPFP=0.5548 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,552B, BPFP=1.3899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,692B, BPFP=1.0016 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,572B, BPFP=1.3345 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,960B, BPFP=0.1291 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10827675 154.47412817 + layer.0.v_cache 0.00001706 0.01550517 + layer.1.k_cache 0.56919524 15.79601520 + layer.1.v_cache 0.00000592 0.00579677 + layer.2.k_cache 0.01911232 1.83223680 + layer.2.v_cache 0.00002018 0.01631274 + layer.3.k_cache 0.01342056 7.91340792 + layer.3.v_cache 0.00002333 0.01857301 + layer.4.k_cache 0.00074072 0.40439661 + layer.4.v_cache 0.00005062 0.03489884 + layer.4.output 0.00480139 201.00656703 + ------------------------------------------------------------------------------------- + TOTAL 0.04379250 93.38572003 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 191796 +BPFP 0.6387 bits/point +EBPFP 1.2774 equivalent bits/point +MSE 93.385720 +---------------------- -------------------------------------------------------- +Time: 0.627s Load: 0.009s, Pack+Encode: 0.245s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 93.3857 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,464B, BPFP=0.3198 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,560B, BPFP=1.5543 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,216B, BPFP=0.5393 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,260B, BPFP=1.4782 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,020B, BPFP=0.6449 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,648B, BPFP=1.3839 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,464B, BPFP=0.6124 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,280B, BPFP=1.4209 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,528B, BPFP=1.0257 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,216B, BPFP=1.3586 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,972B, BPFP=0.1335 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15963619 156.20085733 + layer.0.v_cache 0.00001721 0.01779108 + layer.1.k_cache 0.49929055 15.87093830 + layer.1.v_cache 0.00000624 0.00613150 + layer.2.k_cache 0.03270629 1.70516156 + layer.2.v_cache 0.00002062 0.01569806 + layer.3.k_cache 0.00946706 7.38857861 + layer.3.v_cache 0.00002056 0.01771254 + layer.4.k_cache 0.00068536 0.40512496 + layer.4.v_cache 0.00004941 0.03373960 + layer.4.output 0.00495045 207.85782838 + ------------------------------------------------------------------------------------- + TOTAL 0.04332662 96.27450190 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 192628 +BPFP 0.6631 bits/point +EBPFP 1.3262 equivalent bits/point +MSE 96.274502 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 96.2745 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,488B, BPFP=0.2971 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,328B, BPFP=1.4783 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,744B, BPFP=0.5127 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,444B, BPFP=1.4198 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,788B, BPFP=0.5818 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,180B, BPFP=1.3361 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,552B, BPFP=0.5662 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,840B, BPFP=1.3798 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,080B, BPFP=0.9984 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,372B, BPFP=1.2826 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,044B, BPFP=0.1328 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14874478 155.30429025 + layer.0.v_cache 0.00001714 0.01567064 + layer.1.k_cache 0.45078779 15.99004920 + layer.1.v_cache 0.00000615 0.00576109 + layer.2.k_cache 0.01724745 1.71809284 + layer.2.v_cache 0.00001987 0.01521005 + layer.3.k_cache 0.02849353 6.95146257 + layer.3.v_cache 0.00002056 0.01747391 + layer.4.k_cache 0.00066912 0.40439101 + layer.4.v_cache 0.00005068 0.03480010 + layer.4.output 1.29722614 229.46922291 + ------------------------------------------------------------------------------------- + TOTAL 0.57215530 105.10245659 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 162860 +BPFP 0.6343 bits/point +EBPFP 1.2685 equivalent bits/point +MSE 105.102457 +---------------------- -------------------------------------------------------- +Time: 0.543s Load: 0.008s, Pack+Encode: 0.220s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 105.1025 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 341, 128) +Output shape: (1, 341, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.output: torch.Size([1, 341, 3584]) -> torch.Size([1, 1, 341, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,444B, BPFP=0.2953 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,608B, BPFP=1.4941 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,096B, BPFP=0.5084 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,148B, BPFP=1.4272 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,404B, BPFP=0.5684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,236B, BPFP=1.3396 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,896B, BPFP=0.5451 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,832B, BPFP=1.3669 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,192B, BPFP=0.9710 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,344B, BPFP=1.2988 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,268B, BPFP=0.1261 +⌛️ [2/4] FRONTEND: Frontend time: 0.286s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.431s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17009405 150.92809522 + layer.0.v_cache 0.00001631 0.01606545 + layer.1.k_cache 0.79816426 15.81761908 + layer.1.v_cache 0.00000611 0.00597135 + layer.2.k_cache 0.02878924 1.81579661 + layer.2.v_cache 0.00002016 0.01543503 + layer.3.k_cache 0.01035325 7.17472164 + layer.3.v_cache 0.00002082 0.01734641 + layer.4.k_cache 0.00072160 0.40820572 + layer.4.v_cache 0.00005837 0.03383338 + layer.4.output 0.03919861 159.03411709 + ------------------------------------------------------------------------------------- + TOTAL 0.07544909 75.85128880 + (elements=2,968,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2968064 +Total Bytes 233468 +BPFP 0.6293 bits/point +EBPFP 1.2586 equivalent bits/point +MSE 75.851289 +---------------------- -------------------------------------------------------- +Time: 0.728s Load: 0.012s, Pack+Encode: 0.286s, Decode+Unpack: 0.431s +---------------------- -------------------------------------------------------- +💾 Converting with 75.8513 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 335, 128) +Output shape: (1, 335, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.output: torch.Size([1, 335, 3584]) -> torch.Size([1, 1, 335, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,608B, BPFP=0.3082 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,208B, BPFP=1.5022 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,220B, BPFP=0.5233 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,440B, BPFP=1.4198 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,292B, BPFP=0.6200 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,404B, BPFP=1.3248 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,056B, BPFP=0.5623 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,068B, BPFP=1.3558 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,124B, BPFP=0.9853 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,692B, BPFP=1.2916 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,692B, BPFP=0.1312 +⌛️ [2/4] FRONTEND: Frontend time: 0.277s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.431s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15195324 155.64890392 + layer.0.v_cache 0.00001666 0.01638155 + layer.1.k_cache 0.76788093 15.84961521 + layer.1.v_cache 0.00000624 0.00589570 + layer.2.k_cache 0.02686723 1.69313491 + layer.2.v_cache 0.00001960 0.01506997 + layer.3.k_cache 0.01080669 7.04069059 + layer.3.v_cache 0.00002024 0.01731277 + layer.4.k_cache 0.00071790 0.40841247 + layer.4.v_cache 0.00005106 0.03436821 + layer.4.output 3.38143948 160.28296908 + ------------------------------------------------------------------------------------- + TOTAL 1.44873036 76.63003346 + (elements=2,915,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2915840 +Total Bytes 231804 +BPFP 0.6360 bits/point +EBPFP 1.2720 equivalent bits/point +MSE 76.630033 +---------------------- -------------------------------------------------------- +Time: 0.721s Load: 0.012s, Pack+Encode: 0.277s, Decode+Unpack: 0.431s +---------------------- -------------------------------------------------------- +💾 Converting with 76.6300 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,456B, BPFP=0.3217 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,908B, BPFP=1.5276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,284B, BPFP=0.5474 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,676B, BPFP=1.4550 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,572B, BPFP=0.6823 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,244B, BPFP=1.3705 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,360B, BPFP=0.6108 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,548B, BPFP=1.3884 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,312B, BPFP=1.0208 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,328B, BPFP=1.3165 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,608B, BPFP=0.1483 +⌛️ [2/4] FRONTEND: Frontend time: 0.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17469128 155.05691333 + layer.0.v_cache 0.00001775 0.01674579 + layer.1.k_cache 0.51265420 16.13937205 + layer.1.v_cache 0.00000620 0.00599615 + layer.2.k_cache 0.03428403 1.71418434 + layer.2.v_cache 0.00002024 0.01547925 + layer.3.k_cache 0.01443097 7.15777335 + layer.3.v_cache 0.00002182 0.01917555 + layer.4.k_cache 0.00070119 0.39750231 + layer.4.v_cache 0.00005097 0.03333708 + layer.4.output 0.00503371 209.48423181 + ------------------------------------------------------------------------------------- + TOTAL 0.04541851 96.87918246 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 191296 +BPFP 0.6635 bits/point +EBPFP 1.3270 equivalent bits/point +MSE 96.879182 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.009s, Pack+Encode: 0.267s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 96.8792 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,380B, BPFP=0.3172 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,612B, BPFP=1.5101 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,236B, BPFP=0.5446 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,216B, BPFP=1.4278 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,808B, BPFP=0.6373 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,872B, BPFP=1.3486 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,260B, BPFP=0.6050 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,392B, BPFP=1.3792 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,556B, BPFP=1.0351 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,524B, BPFP=1.3281 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,052B, BPFP=0.1352 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18043976 154.97458726 + layer.0.v_cache 0.00001814 0.01613811 + layer.1.k_cache 0.58196952 15.95256117 + layer.1.v_cache 0.00000608 0.00615243 + layer.2.k_cache 0.02409397 1.78442245 + layer.2.v_cache 0.00002020 0.01564984 + layer.3.k_cache 0.01875559 6.83069769 + layer.3.v_cache 0.00002085 0.01797718 + layer.4.k_cache 0.00069073 0.41002134 + layer.4.v_cache 0.00005389 0.03672606 + layer.4.output 0.00498411 209.48114892 + ------------------------------------------------------------------------------------- + TOTAL 0.04946809 96.84782212 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 187908 +BPFP 0.6517 bits/point +EBPFP 1.3035 equivalent bits/point +MSE 96.847822 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 96.8478 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 386, 128) +Output shape: (1, 386, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.output: torch.Size([1, 386, 3584]) -> torch.Size([1, 1, 386, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,404B, BPFP=0.2997 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,440B, BPFP=1.4751 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,480B, BPFP=0.5052 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,024B, BPFP=1.4177 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,548B, BPFP=0.5889 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,420B, BPFP=1.3528 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,128B, BPFP=0.5719 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,164B, BPFP=1.3829 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,008B, BPFP=1.0123 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,528B, BPFP=1.3167 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,300B, BPFP=0.1232 +⌛️ [2/4] FRONTEND: Frontend time: 0.320s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.500s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16808204 153.83864961 + layer.0.v_cache 0.00001827 0.01689721 + layer.1.k_cache 0.89383947 15.81086565 + layer.1.v_cache 0.00000605 0.00601031 + layer.2.k_cache 0.02643975 1.77893430 + layer.2.v_cache 0.00002094 0.01654341 + layer.3.k_cache 0.04368651 7.19212729 + layer.3.v_cache 0.00002149 0.01866748 + layer.4.k_cache 0.00075087 0.42085610 + layer.4.v_cache 0.00005208 0.03619810 + layer.4.output 0.03469837 140.54591506 + ------------------------------------------------------------------------------------- + TOTAL 0.08092977 68.40924441 + (elements=3,359,744) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3359744 +Total Bytes 266444 +BPFP 0.6344 bits/point +EBPFP 1.2689 equivalent bits/point +MSE 68.409244 +---------------------- -------------------------------------------------------- +Time: 0.832s Load: 0.013s, Pack+Encode: 0.320s, Decode+Unpack: 0.500s +---------------------- -------------------------------------------------------- +💾 Converting with 68.4092 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,312B, BPFP=0.2882 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,624B, BPFP=1.4444 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,028B, BPFP=0.4898 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,744B, BPFP=1.3967 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,500B, BPFP=0.5697 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,588B, BPFP=1.3340 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,992B, BPFP=0.5421 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,804B, BPFP=1.3457 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,292B, BPFP=0.9924 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,340B, BPFP=1.2663 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,616B, BPFP=0.1288 +⌛️ [2/4] FRONTEND: Frontend time: 0.269s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14364293 153.63458930 + layer.0.v_cache 0.00001600 0.01715801 + layer.1.k_cache 0.62366660 15.78418477 + layer.1.v_cache 0.00000628 0.00637399 + layer.2.k_cache 0.01981349 1.75146347 + layer.2.v_cache 0.00002090 0.01709426 + layer.3.k_cache 0.02158341 6.95787048 + layer.3.v_cache 0.00002097 0.01878670 + layer.4.k_cache 0.00071696 0.41573434 + layer.4.v_cache 0.00005883 0.03639248 + layer.4.output 0.00463133 192.71662636 + ------------------------------------------------------------------------------------- + TOTAL 0.04952739 89.86211955 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 194840 +BPFP 0.6218 bits/point +EBPFP 1.2436 equivalent bits/point +MSE 89.862120 +---------------------- -------------------------------------------------------- +Time: 0.654s Load: 0.010s, Pack+Encode: 0.269s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 89.8621 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,444B, BPFP=0.3071 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,644B, BPFP=1.5029 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,520B, BPFP=0.5370 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,248B, BPFP=1.4242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,016B, BPFP=0.6214 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,868B, BPFP=1.3463 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,424B, BPFP=0.5880 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,556B, BPFP=1.3852 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,576B, BPFP=0.9914 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,492B, BPFP=1.3251 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,604B, BPFP=0.1499 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12300906 151.56602550 + layer.0.v_cache 0.00001720 0.01504179 + layer.1.k_cache 0.58060381 15.93526836 + layer.1.v_cache 0.00000644 0.00533469 + layer.2.k_cache 0.02305070 1.77133686 + layer.2.v_cache 0.00002053 0.01520165 + layer.3.k_cache 0.05279791 7.75111229 + layer.3.v_cache 0.00002021 0.01740557 + layer.4.k_cache 0.00074809 0.40203882 + layer.4.v_cache 0.00005348 0.03411974 + layer.4.output 0.00483522 200.33646532 + ------------------------------------------------------------------------------------- + TOTAL 0.04789259 92.93342015 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 196392 +BPFP 0.6517 bits/point +EBPFP 1.3033 equivalent bits/point +MSE 92.933420 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 92.9334 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,284B, BPFP=0.3002 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,396B, BPFP=1.4992 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,476B, BPFP=0.5238 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,164B, BPFP=1.4128 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,700B, BPFP=0.6096 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,912B, BPFP=1.3952 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,468B, BPFP=0.5933 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,540B, BPFP=1.4392 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,692B, BPFP=1.0294 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,704B, BPFP=1.3105 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,112B, BPFP=0.1312 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16177035 156.90654779 + layer.0.v_cache 0.00001782 0.01728966 + layer.1.k_cache 0.31952496 15.84341061 + layer.1.v_cache 0.00000626 0.00602242 + layer.2.k_cache 0.01271096 1.70250028 + layer.2.v_cache 0.00002040 0.01636103 + layer.3.k_cache 0.01363577 7.04204255 + layer.3.v_cache 0.00002033 0.01788323 + layer.4.k_cache 0.00066992 0.41604532 + layer.4.v_cache 0.00005030 0.03456049 + layer.4.output 1.37284199 242.81143898 + ------------------------------------------------------------------------------------- + TOTAL 0.59519535 110.68721978 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 157448 +BPFP 0.6489 bits/point +EBPFP 1.2979 equivalent bits/point +MSE 110.687220 +---------------------- -------------------------------------------------------- +Time: 0.543s Load: 0.008s, Pack+Encode: 0.220s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 110.6872 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,580B, BPFP=0.3057 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,784B, BPFP=1.4331 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,160B, BPFP=0.5260 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,264B, BPFP=1.3887 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,768B, BPFP=0.5779 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,904B, BPFP=1.3579 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,376B, BPFP=0.5444 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,296B, BPFP=1.3914 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,772B, BPFP=1.0051 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,084B, BPFP=1.2879 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,588B, BPFP=0.1413 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.261s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09962382 155.02296790 + layer.0.v_cache 0.00001686 0.01774216 + layer.1.k_cache 0.10988800 15.87865944 + layer.1.v_cache 0.00000590 0.00617842 + layer.2.k_cache 0.01214521 1.79600983 + layer.2.v_cache 0.00002216 0.01737993 + layer.3.k_cache 0.02260447 7.95159145 + layer.3.v_cache 0.00002006 0.01965840 + layer.4.k_cache 0.00069723 0.41613878 + layer.4.v_cache 0.00005242 0.03842195 + layer.4.output 0.00882482 299.80918228 + ------------------------------------------------------------------------------------- + TOTAL 0.01804999 134.10758966 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 126576 +BPFP 0.6357 bits/point +EBPFP 1.2715 equivalent bits/point +MSE 134.107590 +---------------------- -------------------------------------------------------- +Time: 0.518s Load: 0.007s, Pack+Encode: 0.250s, Decode+Unpack: 0.261s +---------------------- -------------------------------------------------------- +💾 Converting with 134.1076 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,484B, BPFP=0.3049 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,068B, BPFP=1.5051 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,356B, BPFP=0.5202 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,772B, BPFP=1.4331 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,592B, BPFP=0.5890 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,196B, BPFP=1.3454 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,784B, BPFP=0.5996 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,044B, BPFP=1.3926 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,096B, BPFP=1.0062 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,620B, BPFP=1.3134 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,348B, BPFP=0.1299 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11148910 154.20910254 + layer.0.v_cache 0.00001684 0.01621158 + layer.1.k_cache 0.60671574 16.07324219 + layer.1.v_cache 0.00000630 0.00573877 + layer.2.k_cache 0.04149003 1.73950977 + layer.2.v_cache 0.00001976 0.01480417 + layer.3.k_cache 0.02200458 7.50154912 + layer.3.v_cache 0.00001954 0.01718628 + layer.4.k_cache 0.00078895 0.40823171 + layer.4.v_cache 0.00004864 0.03331189 + layer.4.output 0.00472744 197.43141523 + ------------------------------------------------------------------------------------- + TOTAL 0.04798185 91.88463498 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 196360 +BPFP 0.6423 bits/point +EBPFP 1.2845 equivalent bits/point +MSE 91.884635 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8846 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,568B, BPFP=0.3569 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,016B, BPFP=1.5637 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,632B, BPFP=0.5962 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,504B, BPFP=1.5237 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,808B, BPFP=0.6881 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,576B, BPFP=1.4512 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,152B, BPFP=0.6369 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,920B, BPFP=1.4781 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,428B, BPFP=1.1272 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,388B, BPFP=1.4366 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,496B, BPFP=0.1395 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14565937 151.93904297 + layer.0.v_cache 0.00001550 0.01634724 + layer.1.k_cache 0.22559690 15.99971191 + layer.1.v_cache 0.00000590 0.00603171 + layer.2.k_cache 0.00679684 1.77019989 + layer.2.v_cache 0.00002314 0.01691164 + layer.3.k_cache 0.06866968 6.95229858 + layer.3.v_cache 0.00002119 0.01902119 + layer.4.k_cache 0.00065507 0.40752495 + layer.4.v_cache 0.00005191 0.03724225 + layer.4.output 1.53062134 270.81906250 + ------------------------------------------------------------------------------------- + TOTAL 0.65657911 121.93516293 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 151488 +BPFP 0.6962 bits/point +EBPFP 1.3924 equivalent bits/point +MSE 121.935163 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.007s, Pack+Encode: 0.222s, Decode+Unpack: 0.313s +---------------------- -------------------------------------------------------- +💾 Converting with 121.9352 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,412B, BPFP=0.3430 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,732B, BPFP=1.6116 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,476B, BPFP=0.5812 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,752B, BPFP=1.5354 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,272B, BPFP=0.6430 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,724B, BPFP=1.4555 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,008B, BPFP=0.6225 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,148B, BPFP=1.4885 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,136B, BPFP=1.0989 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,392B, BPFP=1.4297 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,440B, BPFP=0.1381 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10663509 157.02899565 + layer.0.v_cache 0.00001613 0.01717661 + layer.1.k_cache 0.26620468 15.93071629 + layer.1.v_cache 0.00000602 0.00656726 + layer.2.k_cache 0.00649704 1.82295614 + layer.2.v_cache 0.00002002 0.01753871 + layer.3.k_cache 0.02565914 6.39815085 + layer.3.v_cache 0.00002186 0.01948183 + layer.4.k_cache 0.00070101 0.43115447 + layer.4.v_cache 0.00005421 0.03916111 + layer.4.output 1.52303164 269.45535714 + ------------------------------------------------------------------------------------- + TOTAL 0.65100216 121.64114111 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 151492 +BPFP 0.6927 bits/point +EBPFP 1.3855 equivalent bits/point +MSE 121.641141 +---------------------- -------------------------------------------------------- +Time: 0.539s Load: 0.008s, Pack+Encode: 0.220s, Decode+Unpack: 0.311s +---------------------- -------------------------------------------------------- +💾 Converting with 121.6411 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,516B, BPFP=0.3228 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,912B, BPFP=1.5164 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,244B, BPFP=0.5410 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,992B, BPFP=1.4625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,380B, BPFP=0.6660 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,776B, BPFP=1.3914 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,232B, BPFP=0.5988 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,388B, BPFP=1.4272 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,888B, BPFP=1.0468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,252B, BPFP=1.3607 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,160B, BPFP=0.1351 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258070 158.35055302 + layer.0.v_cache 0.00001713 0.01741282 + layer.1.k_cache 0.49703808 15.80353903 + layer.1.v_cache 0.00000670 0.00659716 + layer.2.k_cache 0.01458295 1.75490327 + layer.2.v_cache 0.00002181 0.01702676 + layer.3.k_cache 0.02940617 6.84528662 + layer.3.v_cache 0.00002120 0.01940998 + layer.4.k_cache 0.00070157 0.42812933 + layer.4.v_cache 0.00005095 0.03616472 + layer.4.output 0.00497170 207.86528558 + ------------------------------------------------------------------------------------- + TOTAL 0.04171936 96.37270716 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 192740 +BPFP 0.6635 bits/point +EBPFP 1.3270 equivalent bits/point +MSE 96.372707 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.009s, Pack+Encode: 0.247s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 96.3727 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,328B, BPFP=0.2992 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,340B, BPFP=1.4754 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,556B, BPFP=0.5224 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,684B, BPFP=1.4300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,676B, BPFP=0.5998 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,120B, BPFP=1.3910 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,856B, BPFP=0.5431 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,568B, BPFP=1.4220 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,908B, BPFP=1.0307 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,780B, BPFP=1.2984 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,140B, BPFP=0.1298 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15433978 152.96318446 + layer.0.v_cache 0.00001575 0.01647117 + layer.1.k_cache 0.29362690 16.00566493 + layer.1.v_cache 0.00000607 0.00634413 + layer.2.k_cache 0.00940348 1.77660417 + layer.2.v_cache 0.00002159 0.01735827 + layer.3.k_cache 0.02320663 7.99123902 + layer.3.v_cache 0.00002067 0.01992189 + layer.4.k_cache 0.00067785 0.42945241 + layer.4.v_cache 0.00005016 0.03694456 + layer.4.output 1.35461534 239.67493679 + ------------------------------------------------------------------------------------- + TOTAL 0.58609860 109.23457309 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 157956 +BPFP 0.6424 bits/point +EBPFP 1.2848 equivalent bits/point +MSE 109.234573 +---------------------- -------------------------------------------------------- +Time: 0.545s Load: 0.009s, Pack+Encode: 0.222s, Decode+Unpack: 0.313s +---------------------- -------------------------------------------------------- +💾 Converting with 109.2346 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,348B, BPFP=0.2932 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,476B, BPFP=1.5064 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,540B, BPFP=0.5230 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,368B, BPFP=1.4456 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,248B, BPFP=0.6167 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,080B, BPFP=1.3750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,748B, BPFP=0.5893 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,516B, BPFP=1.3989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,360B, BPFP=1.0066 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,968B, BPFP=1.3140 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,216B, BPFP=0.1427 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14751465 158.99602522 + layer.0.v_cache 0.00001719 0.01614764 + layer.1.k_cache 0.52245269 15.89772649 + layer.1.v_cache 0.00000618 0.00565657 + layer.2.k_cache 0.02707210 1.73300096 + layer.2.v_cache 0.00002227 0.01604078 + layer.3.k_cache 0.03364062 7.46965461 + layer.3.v_cache 0.00002083 0.01787161 + layer.4.k_cache 0.00069748 0.40029567 + layer.4.v_cache 0.00005305 0.03362143 + layer.4.output 0.00471531 194.61389411 + ------------------------------------------------------------------------------------- + TOTAL 0.04497084 90.99313528 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 201868 +BPFP 0.6510 bits/point +EBPFP 1.3020 equivalent bits/point +MSE 90.993135 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 90.9931 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,380B, BPFP=0.2684 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,916B, BPFP=1.3429 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,576B, BPFP=0.4642 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,108B, BPFP=1.2934 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,340B, BPFP=0.5110 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,732B, BPFP=1.2091 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,840B, BPFP=0.4804 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,096B, BPFP=1.2314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,624B, BPFP=0.8961 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,044B, BPFP=1.1669 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,712B, BPFP=0.1375 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17432048 144.51767770 + layer.0.v_cache 0.00001704 0.01516121 + layer.1.k_cache 0.39057378 15.36976773 + layer.1.v_cache 0.00000668 0.00574470 + layer.2.k_cache 0.01299327 1.64728023 + layer.2.v_cache 0.00002114 0.01454970 + layer.3.k_cache 0.02499930 6.29011757 + layer.3.v_cache 0.00002096 0.01716487 + layer.4.k_cache 0.00069121 0.38585630 + layer.4.v_cache 0.00005416 0.03475135 + layer.4.output 1.20070616 212.29903711 + ------------------------------------------------------------------------------------- + TOTAL 0.52992007 97.31713713 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 160368 +BPFP 0.5780 bits/point +EBPFP 1.1561 equivalent bits/point +MSE 97.317137 +---------------------- -------------------------------------------------------- +Time: 0.547s Load: 0.008s, Pack+Encode: 0.222s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 97.3171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,460B, BPFP=0.3043 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,840B, BPFP=1.4902 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,120B, BPFP=0.5540 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,000B, BPFP=1.4329 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,064B, BPFP=0.6184 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,112B, BPFP=1.3723 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,636B, BPFP=0.5892 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,620B, BPFP=1.4069 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,112B, BPFP=1.0311 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,076B, BPFP=1.3016 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,880B, BPFP=0.1450 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16239141 154.58825737 + layer.0.v_cache 0.00001645 0.01626284 + layer.1.k_cache 0.33751179 16.31802888 + layer.1.v_cache 0.00000668 0.00606617 + layer.2.k_cache 0.00660169 1.74138871 + layer.2.v_cache 0.00002045 0.01589727 + layer.3.k_cache 0.02213682 7.59469884 + layer.3.v_cache 0.00002049 0.01933611 + layer.4.k_cache 0.00067824 0.41990458 + layer.4.v_cache 0.00005183 0.03612286 + layer.4.output 1.33690934 236.52384201 + ------------------------------------------------------------------------------------- + TOTAL 0.58163537 108.02487398 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 162920 +BPFP 0.6539 bits/point +EBPFP 1.3078 equivalent bits/point +MSE 108.024874 +---------------------- -------------------------------------------------------- +Time: 0.545s Load: 0.008s, Pack+Encode: 0.222s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 108.0249 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,492B, BPFP=0.3202 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,480B, BPFP=1.4855 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,256B, BPFP=0.5396 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,408B, BPFP=1.4230 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,656B, BPFP=0.6213 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,688B, BPFP=1.3228 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,000B, BPFP=0.5830 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,320B, BPFP=1.3596 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,304B, BPFP=1.0089 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,176B, BPFP=1.2929 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,576B, BPFP=0.1297 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15796844 158.17280784 + layer.0.v_cache 0.00001527 0.01555090 + layer.1.k_cache 0.50235566 16.27773547 + layer.1.v_cache 0.00000626 0.00618765 + layer.2.k_cache 0.02709456 1.78203070 + layer.2.v_cache 0.00002063 0.01523847 + layer.3.k_cache 0.00856834 6.46253307 + layer.3.v_cache 0.00001945 0.01733021 + layer.4.k_cache 0.00071483 0.41000138 + layer.4.v_cache 0.00005464 0.03513295 + layer.4.output 0.00491706 207.06838020 + ------------------------------------------------------------------------------------- + TOTAL 0.04301397 96.03960059 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 186356 +BPFP 0.6391 bits/point +EBPFP 1.2782 equivalent bits/point +MSE 96.039601 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.009s, Pack+Encode: 0.248s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 96.0396 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,636B, BPFP=0.2921 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,432B, BPFP=1.4133 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,996B, BPFP=0.5038 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,684B, BPFP=1.3662 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,728B, BPFP=0.5499 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,456B, BPFP=1.2888 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,876B, BPFP=0.5592 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,040B, BPFP=1.3256 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,036B, BPFP=0.9473 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,648B, BPFP=1.2379 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,440B, BPFP=0.1390 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13184587 150.61811681 + layer.0.v_cache 0.00001706 0.01619601 + layer.1.k_cache 0.45546043 15.77534042 + layer.1.v_cache 0.00000678 0.00621735 + layer.2.k_cache 0.02272066 1.81174161 + layer.2.v_cache 0.00001996 0.01543901 + layer.3.k_cache 0.01643563 6.92470181 + layer.3.v_cache 0.00002069 0.01826687 + layer.4.k_cache 0.00071781 0.40177370 + layer.4.v_cache 0.00005262 0.03419372 + layer.4.output 1.23455937 218.14510729 + ------------------------------------------------------------------------------------- + TOTAL 0.54524783 100.15516108 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 165972 +BPFP 0.6151 bits/point +EBPFP 1.2302 equivalent bits/point +MSE 100.155161 +---------------------- -------------------------------------------------------- +Time: 0.546s Load: 0.008s, Pack+Encode: 0.221s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 100.1552 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,540B, BPFP=0.3279 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,744B, BPFP=1.5237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,372B, BPFP=0.5547 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,624B, BPFP=1.4574 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,668B, BPFP=0.6314 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,148B, BPFP=1.3700 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,124B, BPFP=0.5992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,560B, BPFP=1.3944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,696B, BPFP=1.0473 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,688B, BPFP=1.3428 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,488B, BPFP=0.1394 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17705467 152.30545691 + layer.0.v_cache 0.00001633 0.01501526 + layer.1.k_cache 0.50406676 16.32726681 + layer.1.v_cache 0.00000599 0.00568570 + layer.2.k_cache 0.00944760 1.69260233 + layer.2.v_cache 0.00001883 0.01488246 + layer.3.k_cache 0.04120483 7.29390832 + layer.3.v_cache 0.00002392 0.01787653 + layer.4.k_cache 0.00070100 0.40072938 + layer.4.v_cache 0.00005002 0.03433572 + layer.4.output 0.00499618 210.27687027 + ------------------------------------------------------------------------------------- + TOTAL 0.04515078 97.06152066 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 189652 +BPFP 0.6603 bits/point +EBPFP 1.3205 equivalent bits/point +MSE 97.061521 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.009s, Pack+Encode: 0.247s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 97.0615 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,360B, BPFP=0.3102 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,528B, BPFP=1.5352 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,236B, BPFP=0.5345 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,320B, BPFP=1.4653 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,808B, BPFP=0.6255 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,804B, BPFP=1.3775 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,620B, BPFP=0.5567 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,244B, BPFP=1.4030 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,916B, BPFP=1.0368 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,464B, BPFP=1.3579 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,504B, BPFP=0.1282 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20501094 153.21166088 + layer.0.v_cache 0.00001516 0.01499059 + layer.1.k_cache 0.57085758 16.23953812 + layer.1.v_cache 0.00000574 0.00578611 + layer.2.k_cache 0.01259345 1.83411537 + layer.2.v_cache 0.00002099 0.01558971 + layer.3.k_cache 0.01648921 7.05158420 + layer.3.v_cache 0.00001970 0.01753953 + layer.4.k_cache 0.00070444 0.39580084 + layer.4.v_cache 0.00005121 0.03560568 + layer.4.output 0.00487421 205.50882937 + ------------------------------------------------------------------------------------- + TOTAL 0.04940517 95.14023627 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 191804 +BPFP 0.6529 bits/point +EBPFP 1.3059 equivalent bits/point +MSE 95.140236 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 95.1402 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,504B, BPFP=0.3173 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,364B, BPFP=1.5201 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,304B, BPFP=0.5364 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,068B, BPFP=1.4453 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,912B, BPFP=0.6292 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,716B, BPFP=1.3674 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,244B, BPFP=0.5906 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,264B, BPFP=1.3990 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,836B, BPFP=1.0284 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,000B, BPFP=1.3261 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,020B, BPFP=0.1320 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14466163 154.29782634 + layer.0.v_cache 0.00001631 0.01616280 + layer.1.k_cache 0.53789613 16.09178246 + layer.1.v_cache 0.00000653 0.00606907 + layer.2.k_cache 0.02095415 1.71142285 + layer.2.v_cache 0.00002169 0.01620809 + layer.3.k_cache 0.03083320 7.07924931 + layer.3.v_cache 0.00002083 0.01813154 + layer.4.k_cache 0.00069737 0.41460309 + layer.4.v_cache 0.00004943 0.03497366 + layer.4.output 0.00490546 204.66325119 + ------------------------------------------------------------------------------------- + TOTAL 0.04526444 94.84289338 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 192232 +BPFP 0.6520 bits/point +EBPFP 1.3039 equivalent bits/point +MSE 94.842893 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 94.8429 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,588B, BPFP=0.2920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,128B, BPFP=1.4176 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,896B, BPFP=0.5171 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,072B, BPFP=1.3625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,012B, BPFP=0.5755 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,736B, BPFP=1.2926 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,288B, BPFP=0.5376 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,044B, BPFP=1.3087 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,880B, BPFP=0.9344 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,432B, BPFP=1.2245 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,268B, BPFP=0.1289 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14240686 152.97269544 + layer.0.v_cache 0.00001673 0.01477039 + layer.1.k_cache 0.61438330 16.03447200 + layer.1.v_cache 0.00000638 0.00551723 + layer.2.k_cache 0.02027981 1.72312024 + layer.2.v_cache 0.00002012 0.01476296 + layer.3.k_cache 0.02199794 7.08785960 + layer.3.v_cache 0.00002003 0.01652559 + layer.4.k_cache 0.00078937 0.39113836 + layer.4.v_cache 0.00005038 0.03263677 + layer.4.output 0.04464151 181.38670569 + ------------------------------------------------------------------------------------- + TOTAL 0.06543891 85.17649638 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 198344 +BPFP 0.6097 bits/point +EBPFP 1.2194 equivalent bits/point +MSE 85.176496 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.012s, Pack+Encode: 0.249s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1765 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 399, 128) +Output shape: (1, 399, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.output: torch.Size([1, 399, 3584]) -> torch.Size([1, 1, 399, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,652B, BPFP=0.2997 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,264B, BPFP=1.4593 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,880B, BPFP=0.5044 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,716B, BPFP=1.3987 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,484B, BPFP=0.6064 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,964B, BPFP=1.3300 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,736B, BPFP=0.5771 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,640B, BPFP=1.3565 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,108B, BPFP=0.9832 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,924B, BPFP=1.2893 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,544B, BPFP=0.1261 +⌛️ [2/4] FRONTEND: Frontend time: 0.331s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.508s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17889012 151.37028117 + layer.0.v_cache 0.00001635 0.01518745 + layer.1.k_cache 0.94984295 15.59860711 + layer.1.v_cache 0.00000590 0.00550211 + layer.2.k_cache 0.02305104 1.67309402 + layer.2.v_cache 0.00002114 0.01578250 + layer.3.k_cache 0.01620252 7.62579851 + layer.3.v_cache 0.00002049 0.01775910 + layer.4.k_cache 0.00073175 0.40194985 + layer.4.v_cache 0.00005255 0.03381504 + layer.4.output 0.00656733 138.84970238 + ------------------------------------------------------------------------------------- + TOTAL 0.07145918 67.57092315 + (elements=3,472,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3472896 +Total Bytes 272912 +BPFP 0.6287 bits/point +EBPFP 1.2573 equivalent bits/point +MSE 67.570923 +---------------------- -------------------------------------------------------- +Time: 0.853s Load: 0.014s, Pack+Encode: 0.331s, Decode+Unpack: 0.508s +---------------------- -------------------------------------------------------- +💾 Converting with 67.5709 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,408B, BPFP=0.3048 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,436B, BPFP=1.4820 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,604B, BPFP=0.5257 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,308B, BPFP=1.4732 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,860B, BPFP=0.6126 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,160B, BPFP=1.3938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,500B, BPFP=0.5877 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,968B, BPFP=1.4497 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,740B, BPFP=1.0191 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,032B, BPFP=1.3158 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,760B, BPFP=0.1359 +⌛️ [2/4] FRONTEND: Frontend time: 0.236s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10670873 157.95504356 + layer.0.v_cache 0.00001672 0.01705409 + layer.1.k_cache 0.29393762 15.87577131 + layer.1.v_cache 0.00000627 0.00632206 + layer.2.k_cache 0.01243114 1.84700296 + layer.2.v_cache 0.00002105 0.01752061 + layer.3.k_cache 0.00919804 7.36825805 + layer.3.v_cache 0.00002002 0.01958634 + layer.4.k_cache 0.00070102 0.41310501 + layer.4.v_cache 0.00005024 0.03773261 + layer.4.output 1.35463108 239.69018647 + ------------------------------------------------------------------------------------- + TOTAL 0.58267697 109.49345305 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 160776 +BPFP 0.6539 bits/point +EBPFP 1.3077 equivalent bits/point +MSE 109.493453 +---------------------- -------------------------------------------------------- +Time: 0.572s Load: 0.009s, Pack+Encode: 0.236s, Decode+Unpack: 0.327s +---------------------- -------------------------------------------------------- +💾 Converting with 109.4935 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,872B, BPFP=0.2997 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,088B, BPFP=1.3588 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,116B, BPFP=0.4993 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,248B, BPFP=1.3071 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,000B, BPFP=0.5536 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,360B, BPFP=1.2525 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,276B, BPFP=0.5091 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,812B, BPFP=1.2803 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,048B, BPFP=0.9257 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,588B, BPFP=1.2050 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,312B, BPFP=0.1346 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633087 155.89736097 + layer.0.v_cache 0.00001891 0.01631129 + layer.1.k_cache 0.36012592 15.70191775 + layer.1.v_cache 0.00000581 0.00566914 + layer.2.k_cache 0.01407751 1.63237204 + layer.2.v_cache 0.00002023 0.01571573 + layer.3.k_cache 0.01230825 7.34779178 + layer.3.v_cache 0.00002000 0.01785819 + layer.4.k_cache 0.00071033 0.39553259 + layer.4.v_cache 0.00004856 0.03410076 + layer.4.output 1.20534739 213.05531145 + ------------------------------------------------------------------------------------- + TOTAL 0.52535871 98.37951826 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 164720 +BPFP 0.5961 bits/point +EBPFP 1.1921 equivalent bits/point +MSE 98.379518 +---------------------- -------------------------------------------------------- +Time: 0.566s Load: 0.010s, Pack+Encode: 0.230s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 98.3795 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,520B, BPFP=0.2930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,668B, BPFP=1.4048 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,756B, BPFP=0.5029 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,088B, BPFP=1.3672 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,548B, BPFP=0.5542 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,080B, BPFP=1.3019 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,480B, BPFP=0.5498 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,472B, BPFP=1.3273 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,860B, BPFP=0.9634 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,168B, BPFP=1.2427 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,792B, BPFP=0.1277 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11142051 155.60720306 + layer.0.v_cache 0.00001650 0.01587582 + layer.1.k_cache 0.42032433 16.09480355 + layer.1.v_cache 0.00000619 0.00591530 + layer.2.k_cache 0.01443043 1.73921995 + layer.2.v_cache 0.00002030 0.01607359 + layer.3.k_cache 0.02895402 7.97926274 + layer.3.v_cache 0.00001914 0.01685042 + layer.4.k_cache 0.00067703 0.40479674 + layer.4.v_cache 0.00005693 0.03584315 + layer.4.output 1.27033357 224.57850474 + ------------------------------------------------------------------------------------- + TOTAL 0.55695649 103.17443397 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 160432 +BPFP 0.6119 bits/point +EBPFP 1.2237 equivalent bits/point +MSE 103.174434 +---------------------- -------------------------------------------------------- +Time: 0.565s Load: 0.010s, Pack+Encode: 0.228s, Decode+Unpack: 0.327s +---------------------- -------------------------------------------------------- +💾 Converting with 103.1744 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 302, 128) +Output shape: (1, 302, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.output: torch.Size([1, 302, 3584]) -> torch.Size([1, 1, 302, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,516B, BPFP=0.2854 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,008B, BPFP=1.3974 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,408B, BPFP=0.4868 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,188B, BPFP=1.3549 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,592B, BPFP=0.5480 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,096B, BPFP=1.2984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,828B, BPFP=0.5085 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,632B, BPFP=1.3262 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,492B, BPFP=0.9567 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,036B, BPFP=1.2436 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,736B, BPFP=0.1163 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15012774 155.21542581 + layer.0.v_cache 0.00001504 0.01509344 + layer.1.k_cache 0.60100308 16.27857092 + layer.1.v_cache 0.00000575 0.00558237 + layer.2.k_cache 0.00814273 1.72649257 + layer.2.v_cache 0.00002032 0.01554262 + layer.3.k_cache 0.01797006 7.28565019 + layer.3.v_cache 0.00001899 0.01667741 + layer.4.k_cache 0.00070977 0.40061493 + layer.4.v_cache 0.00004868 0.03322064 + layer.4.output 0.04414053 179.55407403 + ------------------------------------------------------------------------------------- + TOTAL 0.06394387 84.58066995 + (elements=2,628,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2628608 +Total Bytes 197532 +BPFP 0.6012 bits/point +EBPFP 1.2024 equivalent bits/point +MSE 84.580670 +---------------------- -------------------------------------------------------- +Time: 0.660s Load: 0.011s, Pack+Encode: 0.262s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 84.5807 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,508B, BPFP=0.3211 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,900B, BPFP=1.5100 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,116B, BPFP=0.5315 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,628B, BPFP=1.4359 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,516B, BPFP=0.6131 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,432B, BPFP=1.3661 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,280B, BPFP=0.5993 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,924B, BPFP=1.3948 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,576B, BPFP=1.0247 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,764B, BPFP=1.3272 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,132B, BPFP=0.1427 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14350806 154.05783582 + layer.0.v_cache 0.00001655 0.01517588 + layer.1.k_cache 0.46812200 16.15164412 + layer.1.v_cache 0.00000598 0.00556099 + layer.2.k_cache 0.01096032 1.84464082 + layer.2.v_cache 0.00002019 0.01527344 + layer.3.k_cache 0.00658789 6.61793359 + layer.3.v_cache 0.00002014 0.01686431 + layer.4.k_cache 0.00067528 0.38902103 + layer.4.v_cache 0.00005785 0.03548768 + layer.4.output 0.00496594 207.08685368 + ------------------------------------------------------------------------------------- + TOTAL 0.03910211 95.80925961 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 190776 +BPFP 0.6543 bits/point +EBPFP 1.3085 equivalent bits/point +MSE 95.809260 +---------------------- -------------------------------------------------------- +Time: 0.654s Load: 0.011s, Pack+Encode: 0.260s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 95.8093 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 260, 128) +Output shape: (1, 260, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.output: torch.Size([1, 260, 3584]) -> torch.Size([1, 1, 260, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,260B, BPFP=0.3161 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,868B, BPFP=1.4945 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,028B, BPFP=0.5425 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,192B, BPFP=1.4538 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,888B, BPFP=0.6543 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,020B, BPFP=1.3834 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,764B, BPFP=0.5868 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,228B, BPFP=1.3959 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,440B, BPFP=1.0481 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,512B, BPFP=1.3529 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,552B, BPFP=0.1335 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14548651 152.76661659 + layer.0.v_cache 0.00001584 0.01577381 + layer.1.k_cache 0.47062912 15.73663800 + layer.1.v_cache 0.00000604 0.00604301 + layer.2.k_cache 0.00619981 1.83201165 + layer.2.v_cache 0.00002003 0.01671504 + layer.3.k_cache 0.01299995 7.28343975 + layer.3.v_cache 0.00001981 0.01803294 + layer.4.k_cache 0.00067777 0.41443675 + layer.4.v_cache 0.00006192 0.03710278 + layer.4.output 0.00504520 213.42582418 + ------------------------------------------------------------------------------------- + TOTAL 0.03949607 98.35926939 + (elements=2,263,040) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2263040 +Total Bytes 185752 +BPFP 0.6566 bits/point +EBPFP 1.3133 equivalent bits/point +MSE 98.359269 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 98.3593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,500B, BPFP=0.3159 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,716B, BPFP=1.5347 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,436B, BPFP=0.5420 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,512B, BPFP=1.4655 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,712B, BPFP=0.6153 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,140B, BPFP=1.3867 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,464B, BPFP=0.6011 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,388B, BPFP=1.4010 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,688B, BPFP=1.0161 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,224B, BPFP=1.3341 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,828B, BPFP=0.1381 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14555165 150.12083525 + layer.0.v_cache 0.00001661 0.01639673 + layer.1.k_cache 0.55468144 15.73942656 + layer.1.v_cache 0.00000621 0.00604197 + layer.2.k_cache 0.02243149 1.74399320 + layer.2.v_cache 0.00002103 0.01553650 + layer.3.k_cache 0.02028967 7.39845814 + layer.3.v_cache 0.00002036 0.01732085 + layer.4.k_cache 0.00072975 0.40167607 + layer.4.v_cache 0.00005124 0.03472384 + layer.4.output 0.00486695 203.89343159 + ------------------------------------------------------------------------------------- + TOTAL 0.04575695 94.27931943 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 194608 +BPFP 0.6576 bits/point +EBPFP 1.3152 equivalent bits/point +MSE 94.279319 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 94.2793 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 356, 128) +Output shape: (1, 356, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.output: torch.Size([1, 356, 3584]) -> torch.Size([1, 1, 356, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,620B, BPFP=0.2906 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,364B, BPFP=1.4205 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,624B, BPFP=0.5102 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,308B, BPFP=1.3741 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,952B, BPFP=0.5685 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,308B, BPFP=1.2863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,308B, BPFP=0.5402 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,496B, BPFP=1.2946 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,308B, BPFP=0.9352 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,524B, BPFP=1.2080 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,724B, BPFP=0.1362 +⌛️ [2/4] FRONTEND: Frontend time: 0.287s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.448s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15223843 153.12309077 + layer.0.v_cache 0.00001740 0.01548065 + layer.1.k_cache 0.82972649 15.92092861 + layer.1.v_cache 0.00000631 0.00570548 + layer.2.k_cache 0.03086275 1.79532906 + layer.2.v_cache 0.00001960 0.01465980 + layer.3.k_cache 0.01100341 6.87474077 + layer.3.v_cache 0.00002062 0.01654862 + layer.4.k_cache 0.00073857 0.39630393 + layer.4.v_cache 0.00005279 0.03267736 + layer.4.output 0.03758925 152.36372642 + ------------------------------------------------------------------------------------- + TOTAL 0.07575360 73.22009118 + (elements=3,098,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3098624 +Total Bytes 236536 +BPFP 0.6107 bits/point +EBPFP 1.2214 equivalent bits/point +MSE 73.220091 +---------------------- -------------------------------------------------------- +Time: 0.748s Load: 0.013s, Pack+Encode: 0.287s, Decode+Unpack: 0.448s +---------------------- -------------------------------------------------------- +💾 Converting with 73.2201 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,520B, BPFP=0.3242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,768B, BPFP=1.5136 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,268B, BPFP=0.5444 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,636B, BPFP=1.4471 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,288B, BPFP=0.6631 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,668B, BPFP=1.3903 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,248B, BPFP=0.6020 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,172B, BPFP=1.4199 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,572B, BPFP=1.0322 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,180B, BPFP=1.3616 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,088B, BPFP=0.1434 +⌛️ [2/4] FRONTEND: Frontend time: 0.270s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12347319 157.83417528 + layer.0.v_cache 0.00001696 0.01585014 + layer.1.k_cache 0.44284580 16.14444130 + layer.1.v_cache 0.00000660 0.00605139 + layer.2.k_cache 0.00882101 1.73839770 + layer.2.v_cache 0.00002210 0.01676423 + layer.3.k_cache 0.01145097 7.03897531 + layer.3.v_cache 0.00001984 0.01766882 + layer.4.k_cache 0.00068754 0.42142971 + layer.4.v_cache 0.00006483 0.03589060 + layer.4.output 0.00500964 208.70881445 + ------------------------------------------------------------------------------------- + TOTAL 0.03661626 96.71949092 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 192408 +BPFP 0.6648 bits/point +EBPFP 1.3297 equivalent bits/point +MSE 96.719491 +---------------------- -------------------------------------------------------- +Time: 0.663s Load: 0.012s, Pack+Encode: 0.270s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 96.7195 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,380B, BPFP=0.2950 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,020B, BPFP=1.4814 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,460B, BPFP=0.5186 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,908B, BPFP=1.4204 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,220B, BPFP=0.6151 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,052B, BPFP=1.3735 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,332B, BPFP=0.5664 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,580B, BPFP=1.4024 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,444B, BPFP=1.0112 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,680B, BPFP=1.2982 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,716B, BPFP=0.1309 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13002123 154.76768092 + layer.0.v_cache 0.00001833 0.01658158 + layer.1.k_cache 0.50397591 16.04139083 + layer.1.v_cache 0.00000608 0.00596483 + layer.2.k_cache 0.01926375 1.70483591 + layer.2.v_cache 0.00002060 0.01596729 + layer.3.k_cache 0.04992905 7.23525562 + layer.3.v_cache 0.00002096 0.01833058 + layer.4.k_cache 0.00069540 0.41825503 + layer.4.v_cache 0.00005467 0.03660625 + layer.4.output 0.00466491 194.58663847 + ------------------------------------------------------------------------------------- + TOTAL 0.04333296 90.72749048 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 198792 +BPFP 0.6411 bits/point +EBPFP 1.2822 equivalent bits/point +MSE 90.727490 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.012s, Pack+Encode: 0.257s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 90.7275 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,524B, BPFP=0.3083 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,584B, BPFP=1.4835 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,600B, BPFP=0.5357 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,560B, BPFP=1.4263 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,384B, BPFP=0.6353 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,844B, BPFP=1.3306 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,232B, BPFP=0.5710 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,316B, BPFP=1.3569 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,428B, BPFP=0.9725 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,412B, BPFP=1.3065 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,772B, BPFP=0.1496 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15974760 152.92476283 + layer.0.v_cache 0.00001739 0.01554886 + layer.1.k_cache 0.45644542 15.70650112 + layer.1.v_cache 0.00000643 0.00558345 + layer.2.k_cache 0.01631587 1.73098330 + layer.2.v_cache 0.00002005 0.01461678 + layer.3.k_cache 0.01511352 7.19789777 + layer.3.v_cache 0.00002063 0.01731799 + layer.4.k_cache 0.00073517 0.38550153 + layer.4.v_cache 0.00005091 0.03332567 + layer.4.output 0.00481773 198.04296875 + ------------------------------------------------------------------------------------- + TOTAL 0.04012924 92.01957768 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 196656 +BPFP 0.6455 bits/point +EBPFP 1.2911 equivalent bits/point +MSE 92.019578 +---------------------- -------------------------------------------------------- +Time: 0.646s Load: 0.010s, Pack+Encode: 0.259s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 92.0196 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,492B, BPFP=0.3214 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,672B, BPFP=1.5023 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,480B, BPFP=0.5548 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,892B, BPFP=1.4567 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,964B, BPFP=0.6416 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,348B, BPFP=1.3663 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,312B, BPFP=0.6035 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,044B, BPFP=1.4071 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,932B, BPFP=1.0494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,936B, BPFP=1.3422 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,448B, BPFP=0.1291 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14148096 157.46330758 + layer.0.v_cache 0.00001588 0.01506696 + layer.1.k_cache 0.55266568 15.99278734 + layer.1.v_cache 0.00000632 0.00572742 + layer.2.k_cache 0.00687984 1.77445807 + layer.2.v_cache 0.00001885 0.01509594 + layer.3.k_cache 0.01675770 7.25071139 + layer.3.v_cache 0.00001964 0.01712741 + layer.4.k_cache 0.00068991 0.40112885 + layer.4.v_cache 0.00004859 0.03372955 + layer.4.output 0.00492290 207.84714419 + ------------------------------------------------------------------------------------- + TOTAL 0.04429669 96.34700882 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 190520 +BPFP 0.6558 bits/point +EBPFP 1.3117 equivalent bits/point +MSE 96.347009 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 96.3470 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,272B, BPFP=0.3034 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,364B, BPFP=1.4688 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,956B, BPFP=0.5300 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,792B, BPFP=1.3928 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,752B, BPFP=0.6169 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,296B, BPFP=1.3204 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,552B, BPFP=0.5588 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,044B, BPFP=1.3566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,020B, BPFP=0.9685 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,388B, BPFP=1.2765 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,464B, BPFP=0.1345 +⌛️ [2/4] FRONTEND: Frontend time: 0.282s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.430s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13144862 153.36435759 + layer.0.v_cache 0.00001769 0.01668325 + layer.1.k_cache 0.64513603 15.98460478 + layer.1.v_cache 0.00000607 0.00593834 + layer.2.k_cache 0.02726801 1.76273367 + layer.2.v_cache 0.00001956 0.01531673 + layer.3.k_cache 0.02271994 7.12804155 + layer.3.v_cache 0.00002018 0.01713680 + layer.4.k_cache 0.00073057 0.39958158 + layer.4.v_cache 0.00004940 0.03303973 + layer.4.output 0.04135688 167.89393521 + ------------------------------------------------------------------------------------- + TOTAL 0.06570084 79.64617532 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 221900 +BPFP 0.6314 bits/point +EBPFP 1.2629 equivalent bits/point +MSE 79.646175 +---------------------- -------------------------------------------------------- +Time: 0.724s Load: 0.011s, Pack+Encode: 0.282s, Decode+Unpack: 0.430s +---------------------- -------------------------------------------------------- +💾 Converting with 79.6462 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,364B, BPFP=0.3172 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,420B, BPFP=1.5567 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,784B, BPFP=0.5657 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,344B, BPFP=1.4785 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,412B, BPFP=0.6113 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,268B, BPFP=1.4003 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,176B, BPFP=0.5942 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,572B, BPFP=1.4224 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,696B, BPFP=1.0680 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,108B, BPFP=1.3887 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,364B, BPFP=0.1284 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12396691 155.12605378 + layer.0.v_cache 0.00001598 0.01629561 + layer.1.k_cache 0.24854917 16.27034657 + layer.1.v_cache 0.00000584 0.00573879 + layer.2.k_cache 0.02290044 1.73715380 + layer.2.v_cache 0.00002299 0.01536269 + layer.3.k_cache 0.01294550 6.83814953 + layer.3.v_cache 0.00002077 0.01766789 + layer.4.k_cache 0.00065144 0.40246767 + layer.4.v_cache 0.00005054 0.03446627 + layer.4.output 1.42385976 251.58359635 + ------------------------------------------------------------------------------------- + TOTAL 0.61036164 114.20875747 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 155508 +BPFP 0.6648 bits/point +EBPFP 1.3296 equivalent bits/point +MSE 114.208757 +---------------------- -------------------------------------------------------- +Time: 0.552s Load: 0.008s, Pack+Encode: 0.230s, Decode+Unpack: 0.314s +---------------------- -------------------------------------------------------- +💾 Converting with 114.2088 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,356B, BPFP=0.2896 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,416B, BPFP=1.4282 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,152B, BPFP=0.4948 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,360B, BPFP=1.3711 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,580B, BPFP=0.5720 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,484B, BPFP=1.3237 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,268B, BPFP=0.5551 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,788B, BPFP=1.3402 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,812B, BPFP=0.9630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,020B, BPFP=1.2446 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,228B, BPFP=0.1253 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14783513 155.49987835 + layer.0.v_cache 0.00001600 0.01687407 + layer.1.k_cache 0.54442752 15.99455457 + layer.1.v_cache 0.00000628 0.00591000 + layer.2.k_cache 0.01848349 1.67265600 + layer.2.v_cache 0.00002061 0.01612096 + layer.3.k_cache 0.01777397 7.46748959 + layer.3.v_cache 0.00002092 0.01842765 + layer.4.k_cache 0.00070447 0.41276521 + layer.4.v_cache 0.00005049 0.03558533 + layer.4.output 0.04615090 187.71306537 + ------------------------------------------------------------------------------------- + TOTAL 0.06190560 87.94892467 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 193464 +BPFP 0.6153 bits/point +EBPFP 1.2306 equivalent bits/point +MSE 87.948925 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 87.9489 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,480B, BPFP=0.2873 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,348B, BPFP=1.4339 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,712B, BPFP=0.5092 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,240B, BPFP=1.3758 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,136B, BPFP=0.5839 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,736B, BPFP=1.2970 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,424B, BPFP=0.5466 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,456B, BPFP=1.3347 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,168B, BPFP=0.9526 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,724B, BPFP=1.2439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,440B, BPFP=0.1306 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16274841 155.15143928 + layer.0.v_cache 0.00001794 0.01672308 + layer.1.k_cache 0.56737488 16.00772074 + layer.1.v_cache 0.00000586 0.00566650 + layer.2.k_cache 0.02027899 1.68659779 + layer.2.v_cache 0.00002117 0.01546322 + layer.3.k_cache 0.01891868 7.41233795 + layer.3.v_cache 0.00002140 0.01744156 + layer.4.k_cache 0.00069658 0.39743411 + layer.4.v_cache 0.00005005 0.03322934 + layer.4.output 0.04476916 182.02731004 + ------------------------------------------------------------------------------------- + TOTAL 0.06373636 85.58442493 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 199864 +BPFP 0.6164 bits/point +EBPFP 1.2329 equivalent bits/point +MSE 85.584425 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.012s, Pack+Encode: 0.251s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5844 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,512B, BPFP=0.3155 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,480B, BPFP=1.5156 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,368B, BPFP=0.5362 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,968B, BPFP=1.4290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,784B, BPFP=0.6172 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,536B, BPFP=1.3471 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,192B, BPFP=0.5833 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,068B, BPFP=1.3775 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,452B, BPFP=0.9989 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,988B, BPFP=1.3157 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,000B, BPFP=0.1308 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13366390 154.84904418 + layer.0.v_cache 0.00001719 0.01628652 + layer.1.k_cache 0.57618574 16.27731334 + layer.1.v_cache 0.00000658 0.00587929 + layer.2.k_cache 0.02858028 1.74246764 + layer.2.v_cache 0.00002021 0.01554414 + layer.3.k_cache 0.00873641 7.28454143 + layer.3.v_cache 0.00002023 0.01780050 + layer.4.k_cache 0.00067337 0.40737582 + layer.4.v_cache 0.00005011 0.03406841 + layer.4.output 0.00484872 203.21552852 + ------------------------------------------------------------------------------------- + TOTAL 0.04599383 94.30347182 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 191348 +BPFP 0.6442 bits/point +EBPFP 1.2884 equivalent bits/point +MSE 94.303472 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 94.3035 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,380B, BPFP=0.3097 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,544B, BPFP=1.5232 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,716B, BPFP=0.5455 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,800B, BPFP=1.4706 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,100B, BPFP=0.6434 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,112B, BPFP=1.4219 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,264B, BPFP=0.5843 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,528B, BPFP=1.4514 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,864B, BPFP=1.0509 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,368B, BPFP=1.3693 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,396B, BPFP=0.1454 +⌛️ [2/4] FRONTEND: Frontend time: 0.232s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13057878 161.63898119 + layer.0.v_cache 0.00001797 0.01786458 + layer.1.k_cache 0.31658528 16.25589694 + layer.1.v_cache 0.00000647 0.00617715 + layer.2.k_cache 0.02849985 1.73340451 + layer.2.v_cache 0.00001991 0.01632708 + layer.3.k_cache 0.03599559 7.60025894 + layer.3.v_cache 0.00002239 0.01958998 + layer.4.k_cache 0.00078811 0.42867894 + layer.4.v_cache 0.00005540 0.03793598 + layer.4.output 1.38530015 244.88524160 + ------------------------------------------------------------------------------------- + TOTAL 0.60056887 111.87951803 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 161072 +BPFP 0.6699 bits/point +EBPFP 1.3398 equivalent bits/point +MSE 111.879518 +---------------------- -------------------------------------------------------- +Time: 0.559s Load: 0.010s, Pack+Encode: 0.232s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 111.8795 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,456B, BPFP=0.3209 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,144B, BPFP=1.5945 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,580B, BPFP=0.5458 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,736B, BPFP=1.4931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,928B, BPFP=0.6429 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,932B, BPFP=1.4352 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,572B, BPFP=0.6172 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,096B, BPFP=1.4470 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,864B, BPFP=1.0703 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,140B, BPFP=1.3782 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,540B, BPFP=0.1393 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12725546 160.85530674 + layer.0.v_cache 0.00001736 0.01793746 + layer.1.k_cache 0.32262681 15.95502637 + layer.1.v_cache 0.00000608 0.00623186 + layer.2.k_cache 0.02157613 1.74449671 + layer.2.v_cache 0.00002208 0.01686596 + layer.3.k_cache 0.01760322 7.67349454 + layer.3.v_cache 0.00002192 0.01967798 + layer.4.k_cache 0.00069785 0.42056777 + layer.4.v_cache 0.00005172 0.03741402 + layer.4.output 1.41081570 249.42750165 + ------------------------------------------------------------------------------------- + TOTAL 0.60974050 113.69056064 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 159988 +BPFP 0.6776 bits/point +EBPFP 1.3553 equivalent bits/point +MSE 113.690561 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.008s, Pack+Encode: 0.221s, Decode+Unpack: 0.312s +---------------------- -------------------------------------------------------- +💾 Converting with 113.6906 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,404B, BPFP=0.3231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,884B, BPFP=1.6053 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,644B, BPFP=0.5607 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,856B, BPFP=1.5299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,052B, BPFP=0.6640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,896B, BPFP=1.4595 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,672B, BPFP=0.6362 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,384B, BPFP=1.4953 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,824B, BPFP=1.0874 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,380B, BPFP=1.4217 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,636B, BPFP=0.1429 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13788442 156.19802670 + layer.0.v_cache 0.00001717 0.01827129 + layer.1.k_cache 0.30723858 15.85954693 + layer.1.v_cache 0.00000664 0.00662833 + layer.2.k_cache 0.01337219 1.78141906 + layer.2.v_cache 0.00002159 0.01743606 + layer.3.k_cache 0.03522867 6.99809559 + layer.3.v_cache 0.00002126 0.01995014 + layer.4.k_cache 0.00073009 0.42133188 + layer.4.v_cache 0.00005119 0.03770489 + layer.4.output 1.43730973 254.21885480 + ------------------------------------------------------------------------------------- + TOTAL 0.62092588 115.34649379 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 160632 +BPFP 0.6931 bits/point +EBPFP 1.3863 equivalent bits/point +MSE 115.346494 +---------------------- -------------------------------------------------------- +Time: 0.544s Load: 0.007s, Pack+Encode: 0.220s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 115.3465 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,408B, BPFP=0.3376 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,072B, BPFP=1.6140 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,400B, BPFP=0.5668 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,748B, BPFP=1.5126 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,216B, BPFP=0.7059 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,956B, BPFP=1.4519 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,952B, BPFP=0.6857 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,764B, BPFP=1.5138 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,204B, BPFP=1.0879 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,488B, BPFP=1.4161 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,040B, BPFP=0.1536 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11546621 156.57753140 + layer.0.v_cache 0.00001823 0.01760460 + layer.1.k_cache 0.26924565 16.00122310 + layer.1.v_cache 0.00000737 0.00652815 + layer.2.k_cache 0.01761903 1.72641395 + layer.2.v_cache 0.00002145 0.01714395 + layer.3.k_cache 0.02374693 7.78052536 + layer.3.v_cache 0.00002104 0.01928719 + layer.4.k_cache 0.00067272 0.41635334 + layer.4.v_cache 0.00005153 0.03644896 + layer.4.output 1.50071327 265.43452381 + ------------------------------------------------------------------------------------- + TOTAL 0.64305077 120.03768980 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 156248 +BPFP 0.7040 bits/point +EBPFP 1.4079 equivalent bits/point +MSE 120.037690 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.008s, Pack+Encode: 0.221s, Decode+Unpack: 0.314s +---------------------- -------------------------------------------------------- +💾 Converting with 120.0377 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,556B, BPFP=0.2819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,784B, BPFP=1.4095 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,560B, BPFP=0.4850 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,792B, BPFP=1.3592 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,676B, BPFP=0.5416 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,428B, BPFP=1.2900 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,328B, BPFP=0.5239 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,032B, BPFP=1.3206 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,752B, BPFP=0.9513 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,340B, BPFP=1.2348 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,652B, BPFP=0.1279 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18206066 152.39397321 + layer.0.v_cache 0.00001652 0.01661958 + layer.1.k_cache 0.62405802 16.16759144 + layer.1.v_cache 0.00000616 0.00592275 + layer.2.k_cache 0.02036299 1.76369883 + layer.2.v_cache 0.00002087 0.01611433 + layer.3.k_cache 0.02546444 7.43209561 + layer.3.v_cache 0.00002064 0.01747569 + layer.4.k_cache 0.00069880 0.40865309 + layer.4.v_cache 0.00007146 0.03662329 + layer.4.output 0.04336799 176.02758291 + ------------------------------------------------------------------------------------- + TOTAL 0.06802097 82.96775578 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 202900 +BPFP 0.6055 bits/point +EBPFP 1.2110 equivalent bits/point +MSE 82.967756 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.012s, Pack+Encode: 0.249s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 82.9678 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,416B, BPFP=0.3013 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,252B, BPFP=1.5183 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,120B, BPFP=0.5540 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,476B, BPFP=1.4653 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,188B, BPFP=0.6269 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,580B, BPFP=1.4042 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,468B, BPFP=0.5778 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,740B, BPFP=1.4151 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,904B, BPFP=1.0169 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,104B, BPFP=1.3035 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,252B, BPFP=0.1487 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14018573 157.58781386 + layer.0.v_cache 0.00001653 0.01708108 + layer.1.k_cache 0.40671090 16.23231313 + layer.1.v_cache 0.00000662 0.00622645 + layer.2.k_cache 0.03036383 1.79229576 + layer.2.v_cache 0.00002088 0.01598986 + layer.3.k_cache 0.05940984 7.68198818 + layer.3.v_cache 0.00002025 0.01798313 + layer.4.k_cache 0.00071161 0.41643661 + layer.4.v_cache 0.00005094 0.03355568 + layer.4.output 1.33692960 236.53175686 + ------------------------------------------------------------------------------------- + TOTAL 0.58800025 108.20729305 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 164500 +BPFP 0.6602 bits/point +EBPFP 1.3205 equivalent bits/point +MSE 108.207293 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.010s, Pack+Encode: 0.225s, Decode+Unpack: 0.314s +---------------------- -------------------------------------------------------- +💾 Converting with 108.2073 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,464B, BPFP=0.3185 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,932B, BPFP=1.5648 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,832B, BPFP=0.5588 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,676B, BPFP=1.4752 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,096B, BPFP=0.6490 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,896B, BPFP=1.4195 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,448B, BPFP=0.6027 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,488B, BPFP=1.4618 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,732B, BPFP=1.0511 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,500B, BPFP=1.3913 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,500B, BPFP=0.1376 +⌛️ [2/4] FRONTEND: Frontend time: 0.236s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351684 155.54878353 + layer.0.v_cache 0.00001872 0.01786088 + layer.1.k_cache 0.28219827 16.32371709 + layer.1.v_cache 0.00000627 0.00664911 + layer.2.k_cache 0.01639774 1.86241143 + layer.2.v_cache 0.00002172 0.01799873 + layer.3.k_cache 0.02752649 7.97882331 + layer.3.v_cache 0.00002090 0.01927950 + layer.4.k_cache 0.00067264 0.42613879 + layer.4.v_cache 0.00005058 0.03872513 + layer.4.output 1.39794763 247.11992417 + ------------------------------------------------------------------------------------- + TOTAL 0.60212139 112.47528569 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 160564 +BPFP 0.6739 bits/point +EBPFP 1.3477 equivalent bits/point +MSE 112.475286 +---------------------- -------------------------------------------------------- +Time: 0.558s Load: 0.008s, Pack+Encode: 0.236s, Decode+Unpack: 0.314s +---------------------- -------------------------------------------------------- +💾 Converting with 112.4753 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,424B, BPFP=0.3162 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,820B, BPFP=1.5054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,236B, BPFP=0.5385 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,780B, BPFP=1.4447 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,516B, BPFP=0.6131 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,276B, BPFP=1.3570 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,040B, BPFP=0.5854 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,652B, BPFP=1.3790 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,244B, BPFP=1.0054 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,712B, BPFP=1.3242 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,240B, BPFP=0.1436 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15646637 156.91050606 + layer.0.v_cache 0.00001727 0.01612523 + layer.1.k_cache 0.53079696 15.87926882 + layer.1.v_cache 0.00000635 0.00633425 + layer.2.k_cache 0.02340859 1.79099433 + layer.2.v_cache 0.00002061 0.01613237 + layer.3.k_cache 0.01050839 7.29195119 + layer.3.v_cache 0.00002156 0.01777047 + layer.4.k_cache 0.00068596 0.40224858 + layer.4.v_cache 0.00006072 0.03547401 + layer.4.output 0.00497423 207.07019590 + ------------------------------------------------------------------------------------- + TOTAL 0.04451838 95.99165745 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 189940 +BPFP 0.6514 bits/point +EBPFP 1.3028 equivalent bits/point +MSE 95.991657 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.388s +---------------------- -------------------------------------------------------- +💾 Converting with 95.9917 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,508B, BPFP=0.2794 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,624B, BPFP=1.4014 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,712B, BPFP=0.4927 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,732B, BPFP=1.3561 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,876B, BPFP=0.5517 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,180B, BPFP=1.2774 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,220B, BPFP=0.5185 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,592B, BPFP=1.2983 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,320B, BPFP=0.9294 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,768B, BPFP=1.2058 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,960B, BPFP=0.1374 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.389s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15292851 147.92053064 + layer.0.v_cache 0.00001614 0.01695784 + layer.1.k_cache 0.61736927 15.97700164 + layer.1.v_cache 0.00000669 0.00657795 + layer.2.k_cache 0.02003183 1.77926715 + layer.2.v_cache 0.00002344 0.01695902 + layer.3.k_cache 0.01028479 6.86548010 + layer.3.v_cache 0.00002212 0.01793871 + layer.4.k_cache 0.00070434 0.41621998 + layer.4.v_cache 0.00006490 0.03550277 + layer.4.output 0.04342050 176.03625058 + ------------------------------------------------------------------------------------- + TOTAL 0.06502327 82.66506999 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 202492 +BPFP 0.6043 bits/point +EBPFP 1.2085 equivalent bits/point +MSE 82.665070 +---------------------- -------------------------------------------------------- +Time: 0.657s Load: 0.010s, Pack+Encode: 0.258s, Decode+Unpack: 0.389s +---------------------- -------------------------------------------------------- +💾 Converting with 82.6651 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,376B, BPFP=0.3123 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,432B, BPFP=1.5353 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,044B, BPFP=0.5253 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,288B, BPFP=1.4689 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,832B, BPFP=0.6292 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,916B, BPFP=1.3892 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,948B, BPFP=0.5778 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,048B, BPFP=1.4549 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,788B, BPFP=1.0332 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,704B, BPFP=1.3769 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,364B, BPFP=0.1275 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12950333 155.36335386 + layer.0.v_cache 0.00001690 0.01707189 + layer.1.k_cache 0.52417032 16.08173901 + layer.1.v_cache 0.00000602 0.00619806 + layer.2.k_cache 0.01578268 1.74935403 + layer.2.v_cache 0.00002268 0.01678989 + layer.3.k_cache 0.03345619 7.11069690 + layer.3.v_cache 0.00002007 0.01893356 + layer.4.k_cache 0.00067301 0.40337928 + layer.4.v_cache 0.00005077 0.03558103 + layer.4.output 0.00489311 206.26694437 + ------------------------------------------------------------------------------------- + TOTAL 0.04340905 95.56892401 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 192740 +BPFP 0.6586 bits/point +EBPFP 1.3171 equivalent bits/point +MSE 95.568924 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 95.5689 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,424B, BPFP=0.3128 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,052B, BPFP=1.4884 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,828B, BPFP=0.5535 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,212B, BPFP=1.4290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,032B, BPFP=0.6386 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,848B, BPFP=1.4033 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,764B, BPFP=0.6196 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,524B, BPFP=1.4511 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,608B, BPFP=1.0328 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,880B, BPFP=1.3348 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,572B, BPFP=0.1371 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13888381 161.17993495 + layer.0.v_cache 0.00001731 0.01628734 + layer.1.k_cache 0.27925331 15.96037078 + layer.1.v_cache 0.00000589 0.00591258 + layer.2.k_cache 0.03172064 1.72915332 + layer.2.v_cache 0.00001980 0.01602190 + layer.3.k_cache 0.01738505 7.38765665 + layer.3.v_cache 0.00001961 0.01767010 + layer.4.k_cache 0.00068340 0.40079419 + layer.4.v_cache 0.00005063 0.03509705 + layer.4.output 1.38525280 244.88445378 + ------------------------------------------------------------------------------------- + TOTAL 0.59792994 111.82000443 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 158744 +BPFP 0.6602 bits/point +EBPFP 1.3204 equivalent bits/point +MSE 111.820004 +---------------------- -------------------------------------------------------- +Time: 0.563s Load: 0.009s, Pack+Encode: 0.227s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 111.8200 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.6405 bits/point +Avg EBPFP 1.2809 equivalent bits/point +Avg MSE 96.296327 +Avg Time 0.636s +------------------------ ---------------------------- diff --git a/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..5cc3326820cb8b8ab5b50b3a94afcb6a43e42acd --- /dev/null +++ b/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 559 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other +Output output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,400B, BPFP=0.3420 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,500B, BPFP=1.5936 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,392B, BPFP=0.5746 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,684B, BPFP=1.5302 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,664B, BPFP=0.6735 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,788B, BPFP=1.4605 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,600B, BPFP=0.6685 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,288B, BPFP=1.4994 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,432B, BPFP=1.1219 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,576B, BPFP=1.4440 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,520B, BPFP=0.1501 +⌛️ [2/4] FRONTEND: Frontend time: 0.471s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09475397 156.81158660 + layer.0.v_cache 0.00001927 0.01758911 + layer.1.k_cache 0.17401673 15.95421938 + layer.1.v_cache 0.00000670 0.00662555 + layer.2.k_cache 0.00814445 1.78827193 + layer.2.v_cache 0.00002157 0.01792737 + layer.3.k_cache 0.01691350 7.09295685 + layer.3.v_cache 0.00002161 0.02097015 + layer.4.k_cache 0.00069969 0.44364200 + layer.4.v_cache 0.00005477 0.03930042 + layer.4.output 1.52309445 269.48240938 + ------------------------------------------------------------------------------------- + TOTAL 0.64448903 121.68058559 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 153844 +BPFP 0.7035 bits/point +EBPFP 1.4070 equivalent bits/point +MSE 121.680586 +---------------------- -------------------------------------------------------- +Time: 0.839s Load: 0.009s, Pack+Encode: 0.471s, Decode+Unpack: 0.359s +---------------------- -------------------------------------------------------- +💾 Converting with 121.6806 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 212, 128) +Output shape: (1, 212, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.output: torch.Size([1, 212, 3584]) -> torch.Size([1, 1, 212, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,364B, BPFP=0.3216 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,856B, BPFP=1.6108 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,404B, BPFP=0.5457 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,536B, BPFP=1.5136 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,108B, BPFP=0.6713 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,940B, BPFP=1.4696 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,708B, BPFP=0.6418 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,412B, BPFP=1.5044 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,876B, BPFP=1.0964 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,320B, BPFP=1.4239 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,760B, BPFP=0.1343 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09403566 153.37385761 + layer.0.v_cache 0.00001799 0.01843705 + layer.1.k_cache 0.24957257 16.27323942 + layer.1.v_cache 0.00000621 0.00635123 + layer.2.k_cache 0.01169950 1.80292684 + layer.2.v_cache 0.00002031 0.01732689 + layer.3.k_cache 0.03645871 7.76151376 + layer.3.v_cache 0.00002058 0.02007968 + layer.4.k_cache 0.00069161 0.44525370 + layer.4.v_cache 0.00005130 0.03818380 + layer.4.output 1.44404146 255.31239471 + ------------------------------------------------------------------------------------- + TOTAL 0.61769792 115.70258429 + (elements=1,845,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1845248 +Total Bytes 159284 +BPFP 0.6906 bits/point +EBPFP 1.3811 equivalent bits/point +MSE 115.702584 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.008s, Pack+Encode: 0.229s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 115.7026 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 250, 128) +Output shape: (1, 250, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.output: torch.Size([1, 250, 3584]) -> torch.Size([1, 1, 250, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,608B, BPFP=0.2880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,276B, BPFP=1.3922 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,956B, BPFP=0.4973 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,304B, BPFP=1.3315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,676B, BPFP=0.5423 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,456B, BPFP=1.2785 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,472B, BPFP=0.5295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,944B, BPFP=1.3090 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,160B, BPFP=0.9475 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,652B, BPFP=1.2283 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,472B, BPFP=0.1292 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09121272 153.61523438 + layer.0.v_cache 0.00001793 0.01651678 + layer.1.k_cache 0.34710925 15.56650879 + layer.1.v_cache 0.00000615 0.00601210 + layer.2.k_cache 0.02211576 1.72583301 + layer.2.v_cache 0.00002179 0.01610094 + layer.3.k_cache 0.02174057 7.03752832 + layer.3.v_cache 0.00001938 0.01710960 + layer.4.k_cache 0.00073173 0.39972748 + layer.4.v_cache 0.00005145 0.03513203 + layer.4.output 1.22463198 216.57153571 + ------------------------------------------------------------------------------------- + TOTAL 0.53267356 99.67273255 + (elements=2,176,000) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2176000 +Total Bytes 163976 +BPFP 0.6029 bits/point +EBPFP 1.2057 equivalent bits/point +MSE 99.672733 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.009s, Pack+Encode: 0.228s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 99.6727 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,440B, BPFP=0.3056 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,436B, BPFP=1.4755 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,848B, BPFP=0.5402 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,544B, BPFP=1.4141 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,152B, BPFP=0.6300 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,704B, BPFP=1.3563 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,684B, BPFP=0.5977 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,476B, BPFP=1.4094 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,636B, BPFP=1.0074 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,596B, BPFP=1.2800 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,660B, BPFP=0.1343 +⌛️ [2/4] FRONTEND: Frontend time: 0.219s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12780563 152.21875000 + layer.0.v_cache 0.00001798 0.01715694 + layer.1.k_cache 0.36413978 16.03802786 + layer.1.v_cache 0.00000619 0.00631531 + layer.2.k_cache 0.01529831 1.76969419 + layer.2.v_cache 0.00002052 0.01729963 + layer.3.k_cache 0.01269315 7.49894224 + layer.3.v_cache 0.00002023 0.01856361 + layer.4.k_cache 0.00069261 0.42501687 + layer.4.v_cache 0.00005545 0.03691605 + layer.4.output 1.34865724 238.56743628 + ------------------------------------------------------------------------------------- + TOTAL 0.58596180 108.70698451 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 159176 +BPFP 0.6445 bits/point +EBPFP 1.2890 equivalent bits/point +MSE 108.706985 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.009s, Pack+Encode: 0.219s, Decode+Unpack: 0.314s +---------------------- -------------------------------------------------------- +💾 Converting with 108.7070 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 214, 128) +Output shape: (1, 214, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.output: torch.Size([1, 214, 3584]) -> torch.Size([1, 1, 214, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,440B, BPFP=0.3242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,316B, BPFP=1.5564 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,632B, BPFP=0.5572 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,456B, BPFP=1.4936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,752B, BPFP=0.6390 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,436B, BPFP=1.4191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,404B, BPFP=0.6136 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,728B, BPFP=1.4404 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,492B, BPFP=1.0581 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,136B, BPFP=1.3972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,252B, BPFP=0.1278 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12820032 157.65216121 + layer.0.v_cache 0.00001693 0.01599768 + layer.1.k_cache 0.24875343 16.15480569 + layer.1.v_cache 0.00000587 0.00568124 + layer.2.k_cache 0.01453762 1.76904753 + layer.2.v_cache 0.00001988 0.01550905 + layer.3.k_cache 0.02276520 7.34715428 + layer.3.v_cache 0.00001932 0.01700027 + layer.4.k_cache 0.00071399 0.42072524 + layer.4.v_cache 0.00004895 0.03433753 + layer.4.output 1.43051836 252.98942340 + ------------------------------------------------------------------------------------- + TOTAL 0.61345353 114.96225785 + (elements=1,862,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1862656 +Total Bytes 156044 +BPFP 0.6702 bits/point +EBPFP 1.3404 equivalent bits/point +MSE 114.962258 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.008s, Pack+Encode: 0.221s, Decode+Unpack: 0.313s +---------------------- -------------------------------------------------------- +💾 Converting with 114.9623 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,456B, BPFP=0.3078 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,584B, BPFP=1.4995 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,384B, BPFP=0.5293 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,504B, BPFP=1.4386 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,852B, BPFP=0.6121 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,276B, BPFP=1.3694 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,016B, BPFP=0.5650 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,768B, BPFP=1.3971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,952B, BPFP=1.0126 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,876B, BPFP=1.3468 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,420B, BPFP=0.1323 +⌛️ [2/4] FRONTEND: Frontend time: 0.315s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11875103 149.61094032 + layer.0.v_cache 0.00001689 0.01646178 + layer.1.k_cache 0.49171178 16.10177403 + layer.1.v_cache 0.00000616 0.00582558 + layer.2.k_cache 0.02532451 1.73775882 + layer.2.v_cache 0.00002086 0.01625042 + layer.3.k_cache 0.02370731 7.45964717 + layer.3.v_cache 0.00002129 0.01815416 + layer.4.k_cache 0.00073268 0.41127235 + layer.4.v_cache 0.00004879 0.03523922 + layer.4.output 0.00480151 200.34581292 + ------------------------------------------------------------------------------------- + TOTAL 0.04082070 92.81376555 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 195088 +BPFP 0.6473 bits/point +EBPFP 1.2946 equivalent bits/point +MSE 92.813766 +---------------------- -------------------------------------------------------- +Time: 0.702s Load: 0.011s, Pack+Encode: 0.315s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 92.8138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,544B, BPFP=0.2898 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,000B, BPFP=1.4031 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,764B, BPFP=0.4952 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,424B, BPFP=1.3663 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,704B, BPFP=0.5551 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,476B, BPFP=1.3059 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,428B, BPFP=0.5375 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,020B, BPFP=1.3406 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,188B, BPFP=0.9686 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,640B, BPFP=1.2526 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,908B, BPFP=0.1267 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.331s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10657067 150.81208546 + layer.0.v_cache 0.00001604 0.01669499 + layer.1.k_cache 0.39270948 16.22645089 + layer.1.v_cache 0.00000611 0.00604451 + layer.2.k_cache 0.01541229 1.78749776 + layer.2.v_cache 0.00002285 0.01698760 + layer.3.k_cache 0.01321528 7.26766930 + layer.3.v_cache 0.00001985 0.01837719 + layer.4.k_cache 0.00076706 0.41954006 + layer.4.v_cache 0.00005205 0.03461977 + layer.4.output 1.24960368 220.90942055 + ------------------------------------------------------------------------------------- + TOTAL 0.54564808 101.35128891 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 163096 +BPFP 0.6119 bits/point +EBPFP 1.2237 equivalent bits/point +MSE 101.351289 +---------------------- -------------------------------------------------------- +Time: 0.569s Load: 0.008s, Pack+Encode: 0.230s, Decode+Unpack: 0.331s +---------------------- -------------------------------------------------------- +💾 Converting with 101.3513 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 197, 128) +Output shape: (1, 197, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.output: torch.Size([1, 197, 3584]) -> torch.Size([1, 1, 197, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,464B, BPFP=0.3541 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,884B, BPFP=1.5771 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,476B, BPFP=0.5930 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,988B, BPFP=1.5060 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,084B, BPFP=0.7205 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,408B, BPFP=1.4600 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,420B, BPFP=0.6678 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,540B, BPFP=1.4705 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,072B, BPFP=1.1161 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,004B, BPFP=1.4280 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,308B, BPFP=0.1508 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09219642 157.04891735 + layer.0.v_cache 0.00001621 0.01665511 + layer.1.k_cache 0.17840836 15.91694386 + layer.1.v_cache 0.00000639 0.00595156 + layer.2.k_cache 0.00922235 1.86983494 + layer.2.v_cache 0.00002142 0.01678498 + layer.3.k_cache 0.01957878 7.55079302 + layer.3.v_cache 0.00002124 0.01845362 + layer.4.k_cache 0.00068177 0.41173592 + layer.4.v_cache 0.00005086 0.03731657 + layer.4.output 1.55401793 274.91547317 + ------------------------------------------------------------------------------------- + TOTAL 0.65754878 123.95892348 + (elements=1,714,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1714688 +Total Bytes 150648 +BPFP 0.7029 bits/point +EBPFP 1.4057 equivalent bits/point +MSE 123.958923 +---------------------- -------------------------------------------------------- +Time: 0.560s Load: 0.009s, Pack+Encode: 0.227s, Decode+Unpack: 0.324s +---------------------- -------------------------------------------------------- +💾 Converting with 123.9589 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,692B, BPFP=0.2869 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,164B, BPFP=1.3692 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,584B, BPFP=0.4831 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,908B, BPFP=1.3058 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,612B, BPFP=0.5349 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,944B, BPFP=1.2573 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,176B, BPFP=0.5129 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,556B, BPFP=1.2881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,236B, BPFP=0.9192 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,104B, BPFP=1.2149 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,932B, BPFP=0.1219 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11169702 150.25723286 + layer.0.v_cache 0.00001595 0.01598304 + layer.1.k_cache 0.54955656 15.72692950 + layer.1.v_cache 0.00000603 0.00578563 + layer.2.k_cache 0.01751511 1.75305826 + layer.2.v_cache 0.00002011 0.01655553 + layer.3.k_cache 0.01650532 6.93615723 + layer.3.v_cache 0.00001967 0.01688522 + layer.4.k_cache 0.00077751 0.40244800 + layer.4.v_cache 0.00005008 0.03486049 + layer.4.output 0.04306356 174.81601382 + ------------------------------------------------------------------------------------- + TOTAL 0.05868284 82.28694074 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 198908 +BPFP 0.5897 bits/point +EBPFP 1.1795 equivalent bits/point +MSE 82.286941 +---------------------- -------------------------------------------------------- +Time: 0.653s Load: 0.010s, Pack+Encode: 0.261s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 82.2869 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,424B, BPFP=0.3171 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,732B, BPFP=1.5576 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,732B, BPFP=0.5542 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,624B, BPFP=1.4782 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,976B, BPFP=0.6433 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,800B, BPFP=1.4192 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,720B, BPFP=0.6250 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,356B, BPFP=1.4590 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,332B, BPFP=1.0272 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,700B, BPFP=1.3403 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,676B, BPFP=0.1400 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09832749 160.08514908 + layer.0.v_cache 0.00001658 0.01718915 + layer.1.k_cache 0.33493462 16.11198887 + layer.1.v_cache 0.00000656 0.00629794 + layer.2.k_cache 0.02762392 1.81038309 + layer.2.v_cache 0.00002012 0.01613339 + layer.3.k_cache 0.01647770 7.80429105 + layer.3.v_cache 0.00001951 0.01778744 + layer.4.k_cache 0.00067889 0.41088496 + layer.4.v_cache 0.00004874 0.03494438 + layer.4.output 1.40435175 248.32624099 + ------------------------------------------------------------------------------------- + TOTAL 0.60638920 113.21169037 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 159072 +BPFP 0.6707 bits/point +EBPFP 1.3413 equivalent bits/point +MSE 113.211690 +---------------------- -------------------------------------------------------- +Time: 0.562s Load: 0.010s, Pack+Encode: 0.228s, Decode+Unpack: 0.324s +---------------------- -------------------------------------------------------- +💾 Converting with 113.2117 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,400B, BPFP=0.3042 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,896B, BPFP=1.4447 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,640B, BPFP=0.5282 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,340B, BPFP=1.4062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,028B, BPFP=0.6242 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,340B, BPFP=1.3371 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,300B, BPFP=0.5738 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,960B, BPFP=1.3800 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,704B, BPFP=1.0166 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,456B, BPFP=1.2760 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,960B, BPFP=0.1379 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102250 156.32257674 + layer.0.v_cache 0.00001721 0.01575035 + layer.1.k_cache 0.37371607 16.09200969 + layer.1.v_cache 0.00000658 0.00574382 + layer.2.k_cache 0.01127678 1.75819032 + layer.2.v_cache 0.00001998 0.01636298 + layer.3.k_cache 0.02136800 7.19184700 + layer.3.v_cache 0.00001963 0.01758322 + layer.4.k_cache 0.00069450 0.40392371 + layer.4.v_cache 0.00005147 0.03591228 + layer.4.output 1.35464589 239.69708044 + ------------------------------------------------------------------------------------- + TOTAL 0.58945376 109.39643901 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 157024 +BPFP 0.6386 bits/point +EBPFP 1.2772 equivalent bits/point +MSE 109.396439 +---------------------- -------------------------------------------------------- +Time: 0.555s Load: 0.010s, Pack+Encode: 0.226s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 109.3964 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,600B, BPFP=0.2864 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,068B, BPFP=1.3738 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,936B, BPFP=0.4940 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,256B, BPFP=1.3232 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,692B, BPFP=0.5411 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,280B, BPFP=1.2625 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,792B, BPFP=0.5473 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,716B, BPFP=1.2896 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,844B, BPFP=0.9241 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,356B, BPFP=1.2049 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,764B, BPFP=0.1313 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11871604 155.47307644 + layer.0.v_cache 0.00001755 0.01573436 + layer.1.k_cache 0.46315337 15.85028344 + layer.1.v_cache 0.00000588 0.00551301 + layer.2.k_cache 0.02202094 1.75204504 + layer.2.v_cache 0.00001989 0.01482334 + layer.3.k_cache 0.01278766 7.15275714 + layer.3.v_cache 0.00002092 0.01721119 + layer.4.k_cache 0.00074254 0.39611905 + layer.4.v_cache 0.00004925 0.03270144 + layer.4.output 1.21975157 215.67453401 + ------------------------------------------------------------------------------------- + TOTAL 0.53857618 99.43717662 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 163304 +BPFP 0.5980 bits/point +EBPFP 1.1960 equivalent bits/point +MSE 99.437177 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.011s, Pack+Encode: 0.222s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 99.4372 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 195, 128) +Output shape: (1, 195, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.output: torch.Size([1, 195, 3584]) -> torch.Size([1, 1, 195, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,164B, BPFP=0.3337 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,704B, BPFP=1.5788 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,352B, BPFP=0.5891 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,904B, BPFP=1.5147 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,072B, BPFP=0.7269 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,940B, BPFP=1.4375 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,948B, BPFP=0.6369 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,456B, BPFP=1.4788 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,948B, BPFP=1.1176 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,804B, BPFP=1.4266 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,860B, BPFP=0.1472 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605874 156.09509215 + layer.0.v_cache 0.00001676 0.01676966 + layer.1.k_cache 0.13872500 15.88066657 + layer.1.v_cache 0.00000593 0.00590572 + layer.2.k_cache 0.01222654 1.70644688 + layer.2.v_cache 0.00001968 0.01582777 + layer.3.k_cache 0.01415822 7.43918394 + layer.3.v_cache 0.00002127 0.01906448 + layer.4.k_cache 0.00071127 0.41097995 + layer.4.v_cache 0.00005419 0.03664481 + layer.4.output 1.56987251 277.60966117 + ------------------------------------------------------------------------------------- + TOTAL 0.66300619 124.99377707 + (elements=1,697,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1697280 +Total Bytes 148152 +BPFP 0.6983 bits/point +EBPFP 1.3966 equivalent bits/point +MSE 124.993777 +---------------------- -------------------------------------------------------- +Time: 0.545s Load: 0.008s, Pack+Encode: 0.223s, Decode+Unpack: 0.313s +---------------------- -------------------------------------------------------- +💾 Converting with 124.9938 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,428B, BPFP=0.2943 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,948B, BPFP=1.4550 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,132B, BPFP=0.5264 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,264B, BPFP=1.3963 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,860B, BPFP=0.5889 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,544B, BPFP=1.3345 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,392B, BPFP=0.5488 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,900B, BPFP=1.3650 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,488B, BPFP=0.9863 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,804B, BPFP=1.2709 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,376B, BPFP=0.1395 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.261s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11176265 151.58902816 + layer.0.v_cache 0.00001802 0.01728428 + layer.1.k_cache 0.18491925 15.81244366 + layer.1.v_cache 0.00000608 0.00622679 + layer.2.k_cache 0.01999094 1.67033017 + layer.2.v_cache 0.00002176 0.01689516 + layer.3.k_cache 0.03910406 6.93688562 + layer.3.v_cache 0.00002004 0.01823645 + layer.4.k_cache 0.00067358 0.41185182 + layer.4.v_cache 0.00005212 0.03671852 + layer.4.output 0.00886795 301.49754710 + ------------------------------------------------------------------------------------- + TOTAL 0.02462613 134.52933708 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 125136 +BPFP 0.6319 bits/point +EBPFP 1.2639 equivalent bits/point +MSE 134.529337 +---------------------- -------------------------------------------------------- +Time: 0.519s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.261s +---------------------- -------------------------------------------------------- +💾 Converting with 134.5293 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,344B, BPFP=0.3187 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,456B, BPFP=1.5739 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,452B, BPFP=0.5467 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,352B, BPFP=1.4930 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,796B, BPFP=0.6452 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,600B, BPFP=1.4378 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,268B, BPFP=0.6065 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,004B, BPFP=1.4674 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,468B, BPFP=1.0613 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,816B, BPFP=1.3803 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,672B, BPFP=0.1328 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15089922 151.07811583 + layer.0.v_cache 0.00001686 0.01714917 + layer.1.k_cache 0.31217133 16.12620809 + layer.1.v_cache 0.00000593 0.00637386 + layer.2.k_cache 0.01305961 1.79182341 + layer.2.v_cache 0.00001999 0.01719025 + layer.3.k_cache 0.01178671 7.13047196 + layer.3.v_cache 0.00002039 0.01931672 + layer.4.k_cache 0.00067606 0.42532918 + layer.4.v_cache 0.00005285 0.03690898 + layer.4.output 1.43724915 254.17513414 + ------------------------------------------------------------------------------------- + TOTAL 0.62055606 115.05146038 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 156228 +BPFP 0.6741 bits/point +EBPFP 1.3483 equivalent bits/point +MSE 115.051460 +---------------------- -------------------------------------------------------- +Time: 0.543s Load: 0.007s, Pack+Encode: 0.222s, Decode+Unpack: 0.314s +---------------------- -------------------------------------------------------- +💾 Converting with 115.0515 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 187, 128) +Output shape: (1, 187, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.output: torch.Size([1, 187, 3584]) -> torch.Size([1, 1, 187, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,528B, BPFP=0.2948 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,068B, BPFP=1.4261 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,300B, BPFP=0.5264 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,340B, BPFP=1.3653 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,064B, BPFP=0.5902 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,852B, BPFP=1.3245 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,448B, BPFP=0.5388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,588B, BPFP=1.3860 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,744B, BPFP=0.9813 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,324B, BPFP=1.2804 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,928B, BPFP=0.1424 +⌛️ [2/4] FRONTEND: Frontend time: 0.194s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.258s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08258281 156.63601897 + layer.0.v_cache 0.00001733 0.01817300 + layer.1.k_cache 0.10049684 15.48300546 + layer.1.v_cache 0.00000647 0.00629070 + layer.2.k_cache 0.01472783 1.79022511 + layer.2.v_cache 0.00002140 0.01793179 + layer.3.k_cache 0.02125903 7.54886664 + layer.3.v_cache 0.00002180 0.01999782 + layer.4.k_cache 0.00070831 0.42573392 + layer.4.v_cache 0.00005280 0.03826745 + layer.4.output 0.00866398 293.49417494 + ------------------------------------------------------------------------------------- + TOTAL 0.01650250 131.55551385 + (elements=1,627,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1627648 +Total Bytes 128184 +BPFP 0.6300 bits/point +EBPFP 1.2601 equivalent bits/point +MSE 131.555514 +---------------------- -------------------------------------------------------- +Time: 0.459s Load: 0.007s, Pack+Encode: 0.194s, Decode+Unpack: 0.258s +---------------------- -------------------------------------------------------- +💾 Converting with 131.5555 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,364B, BPFP=0.3232 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,512B, BPFP=1.5930 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,440B, BPFP=0.5509 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,428B, BPFP=1.5127 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,824B, BPFP=0.6534 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,420B, BPFP=1.4381 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,720B, BPFP=0.6457 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,944B, BPFP=1.4769 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,460B, BPFP=1.0708 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,112B, BPFP=1.4153 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,720B, BPFP=0.1451 +⌛️ [2/4] FRONTEND: Frontend time: 0.243s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11071491 152.91626555 + layer.0.v_cache 0.00001804 0.01728896 + layer.1.k_cache 0.26784508 16.16958864 + layer.1.v_cache 0.00000610 0.00625510 + layer.2.k_cache 0.01369454 1.76636407 + layer.2.v_cache 0.00002102 0.01653822 + layer.3.k_cache 0.01444715 7.90038600 + layer.3.v_cache 0.00002041 0.01901599 + layer.4.k_cache 0.00067564 0.42028845 + layer.4.v_cache 0.00005248 0.03617215 + layer.4.output 1.45091435 256.54559496 + ------------------------------------------------------------------------------------- + TOTAL 0.62140564 116.18160752 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 157944 +BPFP 0.6880 bits/point +EBPFP 1.3760 equivalent bits/point +MSE 116.181608 +---------------------- -------------------------------------------------------- +Time: 0.571s Load: 0.010s, Pack+Encode: 0.243s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 116.1816 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,332B, BPFP=0.3035 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,140B, BPFP=1.4812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,596B, BPFP=0.5322 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,504B, BPFP=1.4367 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,532B, BPFP=0.5978 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,356B, BPFP=1.3562 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,460B, BPFP=0.5928 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,624B, BPFP=1.3750 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,468B, BPFP=1.0137 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,328B, BPFP=1.2842 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,604B, BPFP=0.1362 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13389712 154.29915219 + layer.0.v_cache 0.00001748 0.01621418 + layer.1.k_cache 0.36644789 15.65900233 + layer.1.v_cache 0.00000587 0.00588257 + layer.2.k_cache 0.02260907 1.70186926 + layer.2.v_cache 0.00001979 0.01582498 + layer.3.k_cache 0.02550721 6.77056857 + layer.3.v_cache 0.00001968 0.01776257 + layer.4.k_cache 0.00067529 0.40473671 + layer.4.v_cache 0.00006705 0.03573673 + layer.4.output 1.37284074 242.77480381 + ------------------------------------------------------------------------------------- + TOTAL 0.59759716 110.49119863 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 155944 +BPFP 0.6427 bits/point +EBPFP 1.2855 equivalent bits/point +MSE 110.491199 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.009s, Pack+Encode: 0.224s, Decode+Unpack: 0.320s +---------------------- -------------------------------------------------------- +💾 Converting with 110.4912 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 406, 128) +Output shape: (1, 406, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.output: torch.Size([1, 406, 3584]) -> torch.Size([1, 1, 406, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,584B, BPFP=0.2919 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,104B, BPFP=1.4664 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,932B, BPFP=0.4977 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,044B, BPFP=1.3872 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,108B, BPFP=0.5814 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,352B, BPFP=1.3220 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,960B, BPFP=0.5373 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,160B, BPFP=1.3531 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,728B, BPFP=0.9517 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,176B, BPFP=1.2768 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,596B, BPFP=0.1352 +⌛️ [2/4] FRONTEND: Frontend time: 0.587s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.492s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10474718 151.73135391 + layer.0.v_cache 0.00001775 0.01682282 + layer.1.k_cache 0.89873824 15.77650309 + layer.1.v_cache 0.00000687 0.00622039 + layer.2.k_cache 0.02602503 1.65305302 + layer.2.v_cache 0.00002164 0.01700118 + layer.3.k_cache 0.01558245 6.65377131 + layer.3.v_cache 0.00002201 0.01834166 + layer.4.k_cache 0.00072958 0.40725952 + layer.4.v_cache 0.00005327 0.03523592 + layer.4.output 0.00651461 136.51229328 + ------------------------------------------------------------------------------------- + TOTAL 0.06420861 66.58244799 + (elements=3,533,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3533824 +Total Bytes 275744 +BPFP 0.6242 bits/point +EBPFP 1.2485 equivalent bits/point +MSE 66.582448 +---------------------- -------------------------------------------------------- +Time: 1.094s Load: 0.015s, Pack+Encode: 0.587s, Decode+Unpack: 0.492s +---------------------- -------------------------------------------------------- +💾 Converting with 66.5824 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,528B, BPFP=0.2888 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,180B, BPFP=1.4145 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,888B, BPFP=0.5031 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,548B, BPFP=1.3742 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,908B, BPFP=0.5681 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,448B, BPFP=1.3041 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,020B, BPFP=0.5753 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,936B, BPFP=1.3352 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,988B, BPFP=0.9559 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,604B, BPFP=1.2503 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,604B, BPFP=0.1331 +⌛️ [2/4] FRONTEND: Frontend time: 0.241s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.358s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14414167 150.99692283 + layer.0.v_cache 0.00001717 0.01710262 + layer.1.k_cache 0.44179790 15.86836934 + layer.1.v_cache 0.00000667 0.00615981 + layer.2.k_cache 0.03016641 1.79324553 + layer.2.v_cache 0.00002043 0.01695537 + layer.3.k_cache 0.03726396 7.26522939 + layer.3.v_cache 0.00002099 0.01917226 + layer.4.k_cache 0.00070653 0.41331734 + layer.4.v_cache 0.00004910 0.03519312 + layer.4.output 1.24963187 220.88438411 + ------------------------------------------------------------------------------------- + TOTAL 0.55303611 101.33072685 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 164652 +BPFP 0.6177 bits/point +EBPFP 1.2354 equivalent bits/point +MSE 101.330727 +---------------------- -------------------------------------------------------- +Time: 0.609s Load: 0.010s, Pack+Encode: 0.241s, Decode+Unpack: 0.358s +---------------------- -------------------------------------------------------- +💾 Converting with 101.3307 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,396B, BPFP=0.3452 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,236B, BPFP=1.5889 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,664B, BPFP=0.6018 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,292B, BPFP=1.5148 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,708B, BPFP=0.7622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,540B, BPFP=1.4557 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,504B, BPFP=0.6677 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,144B, BPFP=1.5031 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,432B, BPFP=1.1332 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,356B, BPFP=1.4413 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,508B, BPFP=0.1403 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.340s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15985105 155.91915633 + layer.0.v_cache 0.00001905 0.01783731 + layer.1.k_cache 0.22745673 15.85448973 + layer.1.v_cache 0.00000590 0.00627997 + layer.2.k_cache 0.01549998 1.72967575 + layer.2.v_cache 0.00002003 0.01666675 + layer.3.k_cache 0.01333464 7.72431309 + layer.3.v_cache 0.00002034 0.01940494 + layer.4.k_cache 0.00067076 0.42691446 + layer.4.v_cache 0.00004949 0.03759523 + layer.4.output 1.53832061 272.18112886 + ------------------------------------------------------------------------------------- + TOTAL 0.65795131 122.76589621 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 152780 +BPFP 0.7056 bits/point +EBPFP 1.4113 equivalent bits/point +MSE 122.765896 +---------------------- -------------------------------------------------------- +Time: 0.575s Load: 0.008s, Pack+Encode: 0.228s, Decode+Unpack: 0.340s +---------------------- -------------------------------------------------------- +💾 Converting with 122.7659 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 237, 128) +Output shape: (1, 237, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.output: torch.Size([1, 237, 3584]) -> torch.Size([1, 1, 237, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,484B, BPFP=0.2956 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,968B, BPFP=1.4483 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,868B, BPFP=0.5187 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,068B, BPFP=1.3890 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,100B, BPFP=0.5999 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,096B, BPFP=1.3249 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,092B, BPFP=0.5335 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,440B, BPFP=1.3476 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,908B, BPFP=0.9829 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,316B, BPFP=1.2735 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,672B, BPFP=0.1382 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351868 154.85350738 + layer.0.v_cache 0.00001655 0.01667462 + layer.1.k_cache 0.34375901 16.23170181 + layer.1.v_cache 0.00000619 0.00598448 + layer.2.k_cache 0.01282014 1.81277067 + layer.2.v_cache 0.00002104 0.01657840 + layer.3.k_cache 0.03241006 6.97489060 + layer.3.v_cache 0.00001985 0.01811670 + layer.4.k_cache 0.00067667 0.41316970 + layer.4.v_cache 0.00005279 0.03633095 + layer.4.output 1.29178675 228.46100814 + ------------------------------------------------------------------------------------- + TOTAL 0.56210637 104.68275190 + (elements=2,062,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2062848 +Total Bytes 162012 +BPFP 0.6283 bits/point +EBPFP 1.2566 equivalent bits/point +MSE 104.682752 +---------------------- -------------------------------------------------------- +Time: 0.552s Load: 0.008s, Pack+Encode: 0.225s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 104.6828 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,448B, BPFP=0.3233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,604B, BPFP=1.5701 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,872B, BPFP=0.5721 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,448B, BPFP=1.4860 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,532B, BPFP=0.6201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,496B, BPFP=1.4169 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,052B, BPFP=0.5852 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,812B, BPFP=1.4398 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,272B, BPFP=1.0372 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,904B, BPFP=1.3738 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,284B, BPFP=0.1379 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11023432 151.82529070 + layer.0.v_cache 0.00001677 0.01569653 + layer.1.k_cache 0.30291741 16.00291152 + layer.1.v_cache 0.00000597 0.00571002 + layer.2.k_cache 0.03464115 1.66262974 + layer.2.v_cache 0.00001998 0.01581090 + layer.3.k_cache 0.01330193 6.83442042 + layer.3.v_cache 0.00002112 0.01742034 + layer.4.k_cache 0.00069379 0.40793606 + layer.4.v_cache 0.00005021 0.03412889 + layer.4.output 1.42388847 251.61459718 + ------------------------------------------------------------------------------------- + TOTAL 0.61347776 114.00730208 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 156724 +BPFP 0.6700 bits/point +EBPFP 1.3400 equivalent bits/point +MSE 114.007302 +---------------------- -------------------------------------------------------- +Time: 0.585s Load: 0.007s, Pack+Encode: 0.250s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 114.0073 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 231, 128) +Output shape: (1, 231, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.output: torch.Size([1, 231, 3584]) -> torch.Size([1, 1, 231, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,448B, BPFP=0.3009 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,104B, BPFP=1.4951 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,164B, BPFP=0.5522 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,576B, BPFP=1.4594 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,400B, BPFP=0.6358 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,676B, BPFP=1.3985 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,628B, BPFP=0.5836 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,116B, BPFP=1.4283 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,092B, BPFP=1.0208 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,436B, BPFP=1.3147 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,968B, BPFP=0.1543 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13257201 152.25618912 + layer.0.v_cache 0.00001701 0.01736745 + layer.1.k_cache 0.40462213 15.86363742 + layer.1.v_cache 0.00000658 0.00610956 + layer.2.k_cache 0.02649644 1.79311790 + layer.2.v_cache 0.00002056 0.01678977 + layer.3.k_cache 0.06147716 6.80019891 + layer.3.v_cache 0.00002105 0.01941303 + layer.4.k_cache 0.00069437 0.41719990 + layer.4.v_cache 0.00005136 0.03663106 + layer.4.output 1.32536713 234.48538961 + ------------------------------------------------------------------------------------- + TOTAL 0.58256168 106.97790479 + (elements=2,010,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2010624 +Total Bytes 166608 +BPFP 0.6629 bits/point +EBPFP 1.3258 equivalent bits/point +MSE 106.977905 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.008s, Pack+Encode: 0.225s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 106.9779 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,324B, BPFP=0.2650 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,428B, BPFP=1.3743 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,568B, BPFP=0.4637 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,504B, BPFP=1.3176 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,808B, BPFP=0.5397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,820B, BPFP=1.2757 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,916B, BPFP=0.4850 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,256B, BPFP=1.3025 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,484B, BPFP=0.9488 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,136B, BPFP=1.2338 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,432B, BPFP=0.1263 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12841698 148.15677083 + layer.0.v_cache 0.00001718 0.01730789 + layer.1.k_cache 0.45450200 15.45000766 + layer.1.v_cache 0.00000627 0.00654724 + layer.2.k_cache 0.01434461 1.78304360 + layer.2.v_cache 0.00002215 0.01769259 + layer.3.k_cache 0.03530993 6.92917672 + layer.3.v_cache 0.00002139 0.01933368 + layer.4.k_cache 0.00068344 0.41885825 + layer.4.v_cache 0.00005073 0.03813964 + layer.4.output 1.20064817 212.31596639 + ------------------------------------------------------------------------------------- + TOTAL 0.53164187 97.59109664 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 164676 +BPFP 0.5936 bits/point +EBPFP 1.1871 equivalent bits/point +MSE 97.591097 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.008s, Pack+Encode: 0.226s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 97.5911 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,468B, BPFP=0.3030 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,044B, BPFP=1.4984 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,448B, BPFP=0.5235 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,848B, BPFP=1.4322 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,760B, BPFP=0.5962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,792B, BPFP=1.3737 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,396B, BPFP=0.5760 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,160B, BPFP=1.3941 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,508B, BPFP=1.0255 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,220B, BPFP=1.3420 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,436B, BPFP=0.1301 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13731835 157.96503768 + layer.0.v_cache 0.00001709 0.01623231 + layer.1.k_cache 0.55879569 16.09428330 + layer.1.v_cache 0.00000637 0.00588340 + layer.2.k_cache 0.01348169 1.83715431 + layer.2.v_cache 0.00002375 0.01631685 + layer.3.k_cache 0.01306405 7.65460681 + layer.3.v_cache 0.00002005 0.01751008 + layer.4.k_cache 0.00072533 0.40693399 + layer.4.v_cache 0.00005384 0.03601989 + layer.4.output 0.00472113 196.82149189 + ------------------------------------------------------------------------------------- + TOTAL 0.04450318 91.87061305 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 198080 +BPFP 0.6456 bits/point +EBPFP 1.2912 equivalent bits/point +MSE 91.870613 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.011s, Pack+Encode: 0.258s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8706 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,460B, BPFP=0.2834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,788B, BPFP=1.4425 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,584B, BPFP=0.4975 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,552B, BPFP=1.3783 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,152B, BPFP=0.5789 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,412B, BPFP=1.3191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,744B, BPFP=0.5577 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,172B, BPFP=1.3586 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,588B, BPFP=0.9649 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,252B, BPFP=1.2589 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,368B, BPFP=0.1288 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13972237 155.16700841 + layer.0.v_cache 0.00001689 0.01736045 + layer.1.k_cache 0.51719427 15.82116415 + layer.1.v_cache 0.00000613 0.00651837 + layer.2.k_cache 0.02019965 1.83340241 + layer.2.v_cache 0.00002053 0.01747357 + layer.3.k_cache 0.01759115 7.04837391 + layer.3.v_cache 0.00002125 0.01912271 + layer.4.k_cache 0.00069604 0.42161661 + layer.4.v_cache 0.00005872 0.03635249 + layer.4.output 0.04434862 180.15805944 + ------------------------------------------------------------------------------------- + TOTAL 0.05917455 84.79381231 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 203072 +BPFP 0.6201 bits/point +EBPFP 1.2402 equivalent bits/point +MSE 84.793812 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 84.7938 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,344B, BPFP=0.3411 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,624B, BPFP=1.6193 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,660B, BPFP=0.6014 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,652B, BPFP=1.5430 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,040B, BPFP=0.7098 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,772B, BPFP=1.4739 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,420B, BPFP=0.6611 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,064B, BPFP=1.4969 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,280B, BPFP=1.1212 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,448B, BPFP=1.4485 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,844B, BPFP=0.1441 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13044639 150.78241795 + layer.0.v_cache 0.00001737 0.01691411 + layer.1.k_cache 0.22565776 15.63812839 + layer.1.v_cache 0.00000620 0.00610960 + layer.2.k_cache 0.01998205 1.76095535 + layer.2.v_cache 0.00001989 0.01661007 + layer.3.k_cache 0.00629339 7.16132052 + layer.3.v_cache 0.00002050 0.01888550 + layer.4.k_cache 0.00068118 0.41391988 + layer.4.v_cache 0.00005692 0.03696897 + layer.4.output 1.53833376 272.19438711 + ------------------------------------------------------------------------------------- + TOTAL 0.65597165 122.42429060 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 153148 +BPFP 0.7073 bits/point +EBPFP 1.4147 equivalent bits/point +MSE 122.424291 +---------------------- -------------------------------------------------------- +Time: 0.547s Load: 0.007s, Pack+Encode: 0.225s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 122.4243 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 246, 128) +Output shape: (1, 246, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.output: torch.Size([1, 246, 3584]) -> torch.Size([1, 1, 246, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,584B, BPFP=0.2912 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,164B, BPFP=1.4078 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,832B, BPFP=0.4975 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,228B, BPFP=1.3483 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,636B, BPFP=0.5485 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,336B, BPFP=1.2917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,624B, BPFP=0.5478 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,888B, BPFP=1.3267 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,192B, BPFP=0.9649 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,432B, BPFP=1.2342 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,300B, BPFP=0.1298 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131485 149.90642467 + layer.0.v_cache 0.00001613 0.01695859 + layer.1.k_cache 0.49701033 15.70147755 + layer.1.v_cache 0.00000620 0.00610366 + layer.2.k_cache 0.01298508 1.69977985 + layer.2.v_cache 0.00002120 0.01645870 + layer.3.k_cache 0.02762915 7.09548280 + layer.3.v_cache 0.00002136 0.01852028 + layer.4.k_cache 0.00068392 0.40718463 + layer.4.v_cache 0.00006372 0.03464511 + layer.4.output 1.24453647 220.00671458 + ------------------------------------------------------------------------------------- + TOTAL 0.55302984 100.87941399 + (elements=2,141,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2141184 +Total Bytes 163216 +BPFP 0.6098 bits/point +EBPFP 1.2196 equivalent bits/point +MSE 100.879414 +---------------------- -------------------------------------------------------- +Time: 0.543s Load: 0.008s, Pack+Encode: 0.220s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 100.8794 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,508B, BPFP=0.3522 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,248B, BPFP=1.5819 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,460B, BPFP=0.5828 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,644B, BPFP=1.5347 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,052B, BPFP=0.7072 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,740B, BPFP=1.4641 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,644B, BPFP=0.6753 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,860B, BPFP=1.4734 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,564B, BPFP=1.1378 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,316B, BPFP=1.4309 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,100B, BPFP=0.1350 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11599884 154.97798828 + layer.0.v_cache 0.00001671 0.01600617 + layer.1.k_cache 0.22559120 15.79369751 + layer.1.v_cache 0.00000595 0.00615922 + layer.2.k_cache 0.01146694 1.86475006 + layer.2.v_cache 0.00002201 0.01719274 + layer.3.k_cache 0.03668072 7.42958313 + layer.3.v_cache 0.00002131 0.01838090 + layer.4.k_cache 0.00066762 0.41748550 + layer.4.v_cache 0.00005146 0.03695099 + layer.4.output 1.53061454 270.82174107 + ------------------------------------------------------------------------------------- + TOTAL 0.65322497 122.13708129 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 152136 +BPFP 0.6992 bits/point +EBPFP 1.3983 equivalent bits/point +MSE 122.137081 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.008s, Pack+Encode: 0.220s, Decode+Unpack: 0.314s +---------------------- -------------------------------------------------------- +💾 Converting with 122.1371 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,464B, BPFP=0.3046 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,044B, BPFP=1.5041 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,160B, BPFP=0.5568 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,444B, BPFP=1.4632 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,412B, BPFP=0.6422 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,772B, BPFP=1.4173 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,276B, BPFP=0.6329 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,916B, BPFP=1.4271 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,440B, BPFP=1.0535 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,372B, BPFP=1.3218 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,660B, BPFP=0.1526 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605523 151.89326897 + layer.0.v_cache 0.00001866 0.01699227 + layer.1.k_cache 0.31881074 16.01067182 + layer.1.v_cache 0.00000653 0.00626060 + layer.2.k_cache 0.02091764 1.78086313 + layer.2.v_cache 0.00002113 0.01721229 + layer.3.k_cache 0.03221333 6.79538883 + layer.3.v_cache 0.00002185 0.01939769 + layer.4.k_cache 0.00068277 0.41895857 + layer.4.v_cache 0.00005596 0.03678833 + layer.4.output 1.33695214 236.53138646 + ------------------------------------------------------------------------------------- + TOTAL 0.57926287 107.80679457 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 166960 +BPFP 0.6701 bits/point +EBPFP 1.3402 equivalent bits/point +MSE 107.806795 +---------------------- -------------------------------------------------------- +Time: 0.546s Load: 0.008s, Pack+Encode: 0.223s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 107.8068 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 192, 128) +Output shape: (1, 192, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.output: torch.Size([1, 192, 3584]) -> torch.Size([1, 1, 192, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,004B, BPFP=0.2445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,236B, BPFP=1.3213 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,248B, BPFP=0.4271 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,700B, BPFP=1.2777 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,012B, BPFP=0.4893 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,148B, BPFP=1.2327 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,408B, BPFP=0.4401 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,432B, BPFP=1.2559 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,376B, BPFP=0.9258 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,700B, BPFP=1.1963 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,652B, BPFP=0.1122 +⌛️ [2/4] FRONTEND: Frontend time: 0.194s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.259s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12444629 140.24559530 + layer.0.v_cache 0.00001540 0.01567606 + layer.1.k_cache 0.18411521 14.97667440 + layer.1.v_cache 0.00000585 0.00585303 + layer.2.k_cache 0.01124177 1.60063521 + layer.2.v_cache 0.00002179 0.01715701 + layer.3.k_cache 0.01616090 6.70864677 + layer.3.v_cache 0.00002117 0.01857487 + layer.4.k_cache 0.00068507 0.40823007 + layer.4.v_cache 0.00005275 0.03587463 + layer.4.output 0.00839530 285.85858445 + ------------------------------------------------------------------------------------- + TOTAL 0.02326666 127.35547109 + (elements=1,671,168) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1671168 +Total Bytes 117916 +BPFP 0.5645 bits/point +EBPFP 1.1289 equivalent bits/point +MSE 127.355471 +---------------------- -------------------------------------------------------- +Time: 0.460s Load: 0.006s, Pack+Encode: 0.194s, Decode+Unpack: 0.259s +---------------------- -------------------------------------------------------- +💾 Converting with 127.3555 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,420B, BPFP=0.3219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,536B, BPFP=1.5565 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,208B, BPFP=0.5843 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,392B, BPFP=1.5429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,196B, BPFP=0.6773 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,616B, BPFP=1.4699 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,336B, BPFP=0.6905 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,956B, BPFP=1.5019 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,876B, BPFP=1.1178 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,700B, BPFP=1.3837 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,268B, BPFP=0.1381 +⌛️ [2/4] FRONTEND: Frontend time: 0.192s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.256s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09999813 157.31497082 + layer.0.v_cache 0.00001899 0.01690391 + layer.1.k_cache 0.12989934 16.03392966 + layer.1.v_cache 0.00000610 0.00600034 + layer.2.k_cache 0.00966061 1.79142963 + layer.2.v_cache 0.00001994 0.01713039 + layer.3.k_cache 0.03384685 7.10251112 + layer.3.v_cache 0.00002040 0.01843502 + layer.4.k_cache 0.00066502 0.41857005 + layer.4.v_cache 0.00005200 0.03736552 + layer.4.output 0.00959846 330.84103378 + ------------------------------------------------------------------------------------- + TOTAL 0.02008098 146.97908723 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 125504 +BPFP 0.6949 bits/point +EBPFP 1.3898 equivalent bits/point +MSE 146.979087 +---------------------- -------------------------------------------------------- +Time: 0.454s Load: 0.006s, Pack+Encode: 0.192s, Decode+Unpack: 0.256s +---------------------- -------------------------------------------------------- +💾 Converting with 146.9791 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 428, 128) +Output shape: (1, 428, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.output: torch.Size([1, 428, 3584]) -> torch.Size([1, 1, 428, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,648B, BPFP=0.2792 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,548B, BPFP=1.4073 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,096B, BPFP=0.4781 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,660B, BPFP=1.3383 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,468B, BPFP=0.5647 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,152B, BPFP=1.2833 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,760B, BPFP=0.5023 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,204B, BPFP=1.3217 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,556B, BPFP=0.9330 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,832B, BPFP=1.2351 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,680B, BPFP=0.1287 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.500s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13760449 156.83977074 + layer.0.v_cache 0.00001624 0.01620120 + layer.1.k_cache 0.89392254 15.96030547 + layer.1.v_cache 0.00000615 0.00585927 + layer.2.k_cache 0.03917003 1.77713683 + layer.2.v_cache 0.00002113 0.01643610 + layer.3.k_cache 0.03749864 7.27600212 + layer.3.v_cache 0.00002059 0.01731695 + layer.4.k_cache 0.00086919 0.42473948 + layer.4.v_cache 0.00005177 0.03601933 + layer.4.output 0.00617707 129.52275951 + ------------------------------------------------------------------------------------- + TOTAL 0.06778943 64.06053553 + (elements=3,725,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3725312 +Total Bytes 280604 +BPFP 0.6026 bits/point +EBPFP 1.2052 equivalent bits/point +MSE 64.060536 +---------------------- -------------------------------------------------------- +Time: 0.833s Load: 0.013s, Pack+Encode: 0.319s, Decode+Unpack: 0.500s +---------------------- -------------------------------------------------------- +💾 Converting with 64.0605 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,444B, BPFP=0.3016 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,768B, BPFP=1.4832 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,532B, BPFP=0.5281 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,780B, BPFP=1.4284 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,948B, BPFP=0.6066 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,900B, BPFP=1.3797 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,556B, BPFP=0.5849 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,244B, BPFP=1.3987 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,964B, BPFP=0.9953 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,740B, BPFP=1.3154 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,596B, BPFP=0.1393 +⌛️ [2/4] FRONTEND: Frontend time: 0.276s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385377 159.51471077 + layer.0.v_cache 0.00001705 0.01609349 + layer.1.k_cache 0.49372572 16.24934723 + layer.1.v_cache 0.00000626 0.00599473 + layer.2.k_cache 0.02422354 1.72521778 + layer.2.v_cache 0.00002046 0.01594962 + layer.3.k_cache 0.04153839 7.70155800 + layer.3.v_cache 0.00002075 0.01796783 + layer.4.k_cache 0.00074884 0.41892031 + layer.4.v_cache 0.00004954 0.03471816 + layer.4.output 0.00474379 196.73388425 + ------------------------------------------------------------------------------------- + TOTAL 0.04161240 91.93162751 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 198472 +BPFP 0.6469 bits/point +EBPFP 1.2938 equivalent bits/point +MSE 91.931628 +---------------------- -------------------------------------------------------- +Time: 0.676s Load: 0.014s, Pack+Encode: 0.276s, Decode+Unpack: 0.386s +---------------------- -------------------------------------------------------- +💾 Converting with 91.9316 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,252B, BPFP=0.3024 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,676B, BPFP=1.4839 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,012B, BPFP=0.5327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,340B, BPFP=1.4193 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,204B, BPFP=0.6387 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,568B, BPFP=1.3336 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,096B, BPFP=0.5851 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,396B, BPFP=1.3736 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,936B, BPFP=1.0128 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,132B, BPFP=1.3125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,932B, BPFP=0.1447 +⌛️ [2/4] FRONTEND: Frontend time: 0.362s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.436s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12985682 154.67499516 + layer.0.v_cache 0.00001645 0.01580445 + layer.1.k_cache 0.61935935 15.90536868 + layer.1.v_cache 0.00000656 0.00572018 + layer.2.k_cache 0.01492825 1.71414931 + layer.2.v_cache 0.00002178 0.01594059 + layer.3.k_cache 0.02579401 7.31694895 + layer.3.v_cache 0.00002068 0.01701579 + layer.4.k_cache 0.00071613 0.41045670 + layer.4.v_cache 0.00005057 0.03452346 + layer.4.output 0.04141478 167.91123950 + ------------------------------------------------------------------------------------- + TOTAL 0.06356906 79.73468234 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 227544 +BPFP 0.6475 bits/point +EBPFP 1.2950 equivalent bits/point +MSE 79.734682 +---------------------- -------------------------------------------------------- +Time: 0.809s Load: 0.011s, Pack+Encode: 0.362s, Decode+Unpack: 0.436s +---------------------- -------------------------------------------------------- +💾 Converting with 79.7347 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,412B, BPFP=0.2957 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,568B, BPFP=1.4515 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,380B, BPFP=0.5125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,448B, BPFP=1.3903 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,904B, BPFP=0.5957 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,412B, BPFP=1.3337 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,856B, BPFP=0.5931 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,032B, BPFP=1.3676 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,272B, BPFP=0.9983 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,460B, BPFP=1.2817 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,712B, BPFP=0.1382 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12359551 155.32845280 + layer.0.v_cache 0.00001766 0.01709727 + layer.1.k_cache 0.52692803 15.99569937 + layer.1.v_cache 0.00000637 0.00616007 + layer.2.k_cache 0.02102186 1.79695951 + layer.2.v_cache 0.00002187 0.01673044 + layer.3.k_cache 0.03868174 7.62408917 + layer.3.v_cache 0.00002097 0.01888786 + layer.4.k_cache 0.00070799 0.41940465 + layer.4.v_cache 0.00005204 0.03630801 + layer.4.output 0.00469152 193.91302448 + ------------------------------------------------------------------------------------- + TOTAL 0.04375851 90.50888003 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 197456 +BPFP 0.6346 bits/point +EBPFP 1.2691 equivalent bits/point +MSE 90.508880 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 90.5089 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,348B, BPFP=0.3414 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,288B, BPFP=1.5930 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,640B, BPFP=0.5999 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,384B, BPFP=1.5220 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,416B, BPFP=0.7393 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,408B, BPFP=1.4454 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,668B, BPFP=0.6806 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,912B, BPFP=1.4849 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,480B, BPFP=1.1369 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,976B, BPFP=1.4114 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,292B, BPFP=0.1491 +⌛️ [2/4] FRONTEND: Frontend time: 0.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.361s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10671347 153.36399772 + layer.0.v_cache 0.00001788 0.01787564 + layer.1.k_cache 0.15257932 15.97072275 + layer.1.v_cache 0.00000642 0.00609995 + layer.2.k_cache 0.02197224 1.75702763 + layer.2.v_cache 0.00002094 0.01685407 + layer.3.k_cache 0.03325213 7.50234694 + layer.3.v_cache 0.00002126 0.01988419 + layer.4.k_cache 0.00067229 0.41787965 + layer.4.v_cache 0.00005023 0.03642568 + layer.4.output 1.53834953 272.18377602 + ------------------------------------------------------------------------------------- + TOTAL 0.65198546 122.61150273 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 152812 +BPFP 0.7058 bits/point +EBPFP 1.4116 equivalent bits/point +MSE 122.611503 +---------------------- -------------------------------------------------------- +Time: 0.600s Load: 0.008s, Pack+Encode: 0.231s, Decode+Unpack: 0.361s +---------------------- -------------------------------------------------------- +💾 Converting with 122.6115 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,576B, BPFP=0.2849 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,576B, BPFP=1.4054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,004B, BPFP=0.4983 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,688B, BPFP=1.3501 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,968B, BPFP=0.5583 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,988B, BPFP=1.3065 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,368B, BPFP=0.5209 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,376B, BPFP=1.3307 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,404B, BPFP=0.9589 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,948B, BPFP=1.2418 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,536B, BPFP=0.1382 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15214830 158.83905005 + layer.0.v_cache 0.00001722 0.01734743 + layer.1.k_cache 0.49754258 15.43634738 + layer.1.v_cache 0.00000639 0.00656366 + layer.2.k_cache 0.01862443 1.76534268 + layer.2.v_cache 0.00002219 0.01750777 + layer.3.k_cache 0.01200796 7.55567476 + layer.3.v_cache 0.00002210 0.01951954 + layer.4.k_cache 0.00075226 0.42682283 + layer.4.v_cache 0.00005409 0.03764854 + layer.4.output 1.21979868 215.67549445 + ------------------------------------------------------------------------------------- + TOTAL 0.54234049 99.63825211 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 167432 +BPFP 0.6131 bits/point +EBPFP 1.2262 equivalent bits/point +MSE 99.638252 +---------------------- -------------------------------------------------------- +Time: 0.601s Load: 0.008s, Pack+Encode: 0.225s, Decode+Unpack: 0.368s +---------------------- -------------------------------------------------------- +💾 Converting with 99.6383 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,508B, BPFP=0.2822 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,652B, BPFP=1.4166 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,524B, BPFP=0.4879 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,652B, BPFP=1.3654 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,768B, BPFP=0.5516 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,604B, BPFP=1.3117 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,908B, BPFP=0.5076 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,188B, BPFP=1.3416 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,952B, BPFP=0.9709 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,464B, BPFP=1.2533 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,768B, BPFP=0.1227 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18432382 153.59483863 + layer.0.v_cache 0.00001629 0.01686815 + layer.1.k_cache 0.53789903 15.77795370 + layer.1.v_cache 0.00000607 0.00627964 + layer.2.k_cache 0.01514517 1.72738317 + layer.2.v_cache 0.00002151 0.01696566 + layer.3.k_cache 0.04182221 7.07830430 + layer.3.v_cache 0.00002017 0.01847716 + layer.4.k_cache 0.00068627 0.41127970 + layer.4.v_cache 0.00005063 0.03692314 + layer.4.output 0.04375917 177.72561475 + ------------------------------------------------------------------------------------- + TOTAL 0.06390032 83.69203391 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 201988 +BPFP 0.6087 bits/point +EBPFP 1.2174 equivalent bits/point +MSE 83.692034 +---------------------- -------------------------------------------------------- +Time: 0.642s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 83.6920 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 371, 128) +Output shape: (1, 371, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.output: torch.Size([1, 371, 3584]) -> torch.Size([1, 1, 371, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,592B, BPFP=0.2776 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,316B, BPFP=1.4031 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,432B, BPFP=0.4815 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,960B, BPFP=1.3460 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,136B, BPFP=0.5532 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,452B, BPFP=1.2825 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,124B, BPFP=0.5106 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,108B, BPFP=1.3101 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,316B, BPFP=0.9399 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,228B, BPFP=1.2310 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,048B, BPFP=0.1266 +⌛️ [2/4] FRONTEND: Frontend time: 0.290s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.444s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16952868 150.39115145 + layer.0.v_cache 0.00001698 0.01730631 + layer.1.k_cache 0.82091874 15.96435810 + layer.1.v_cache 0.00000666 0.00632148 + layer.2.k_cache 0.02418445 1.74987974 + layer.2.v_cache 0.00002238 0.01734882 + layer.3.k_cache 0.03055777 7.55245470 + layer.3.v_cache 0.00002117 0.01866492 + layer.4.k_cache 0.00071103 0.41795086 + layer.4.v_cache 0.00005915 0.03792987 + layer.4.output 0.03610923 146.16733250 + ------------------------------------------------------------------------------------- + TOTAL 0.07639951 70.54968787 + (elements=3,229,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3229184 +Total Bytes 242712 +BPFP 0.6013 bits/point +EBPFP 1.2026 equivalent bits/point +MSE 70.549688 +---------------------- -------------------------------------------------------- +Time: 0.746s Load: 0.011s, Pack+Encode: 0.290s, Decode+Unpack: 0.444s +---------------------- -------------------------------------------------------- +💾 Converting with 70.5497 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,440B, BPFP=0.3182 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,764B, BPFP=1.5599 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,648B, BPFP=0.5482 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,760B, BPFP=1.4880 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,676B, BPFP=0.6218 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,760B, BPFP=1.4163 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,504B, BPFP=0.6095 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,464B, BPFP=1.4667 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,552B, BPFP=1.0430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,004B, BPFP=1.3621 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,840B, BPFP=0.1417 +⌛️ [2/4] FRONTEND: Frontend time: 0.243s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.332s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13674593 161.63019639 + layer.0.v_cache 0.00001768 0.01828747 + layer.1.k_cache 0.26206151 16.16519025 + layer.1.v_cache 0.00000624 0.00656228 + layer.2.k_cache 0.02938290 1.89049565 + layer.2.v_cache 0.00002158 0.01740673 + layer.3.k_cache 0.05653211 7.65210353 + layer.3.v_cache 0.00002129 0.01940553 + layer.4.k_cache 0.00070776 0.41987617 + layer.4.v_cache 0.00005193 0.03788538 + layer.4.output 1.40434551 248.35611894 + ------------------------------------------------------------------------------------- + TOTAL 0.60682162 113.31472011 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 159412 +BPFP 0.6721 bits/point +EBPFP 1.3442 equivalent bits/point +MSE 113.314720 +---------------------- -------------------------------------------------------- +Time: 0.583s Load: 0.008s, Pack+Encode: 0.243s, Decode+Unpack: 0.332s +---------------------- -------------------------------------------------------- +💾 Converting with 113.3147 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 193, 128) +Output shape: (1, 193, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.output: torch.Size([1, 193, 3584]) -> torch.Size([1, 1, 193, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,112B, BPFP=0.3329 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,676B, BPFP=1.5929 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,148B, BPFP=0.5787 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,900B, BPFP=1.5301 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,120B, BPFP=0.7383 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,344B, BPFP=1.4851 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,936B, BPFP=0.7234 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,676B, BPFP=1.5120 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,024B, BPFP=1.1354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,892B, BPFP=1.4485 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,272B, BPFP=0.1419 +⌛️ [2/4] FRONTEND: Frontend time: 0.234s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14305212 154.94314686 + layer.0.v_cache 0.00001770 0.01805087 + layer.1.k_cache 0.14179654 16.15730246 + layer.1.v_cache 0.00000608 0.00629981 + layer.2.k_cache 0.01555968 1.85751643 + layer.2.v_cache 0.00002047 0.01804530 + layer.3.k_cache 0.02655400 6.97789768 + layer.3.v_cache 0.00002031 0.01965292 + layer.4.k_cache 0.00069483 0.44053116 + layer.4.v_cache 0.00005004 0.03768202 + layer.4.output 1.58615237 280.62613342 + ------------------------------------------------------------------------------------- + TOTAL 0.67240225 126.16817997 + (elements=1,679,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1679872 +Total Bytes 149100 +BPFP 0.7101 bits/point +EBPFP 1.4201 equivalent bits/point +MSE 126.168180 +---------------------- -------------------------------------------------------- +Time: 0.558s Load: 0.007s, Pack+Encode: 0.234s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 126.1682 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,620B, BPFP=0.2899 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,108B, BPFP=1.3873 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,984B, BPFP=0.5010 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,152B, BPFP=1.3273 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,120B, BPFP=0.5723 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,508B, BPFP=1.2869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,596B, BPFP=0.5394 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,772B, BPFP=1.3035 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,824B, BPFP=0.9302 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,564B, BPFP=1.2277 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,384B, BPFP=0.1379 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10821181 153.79841240 + layer.0.v_cache 0.00001942 0.01563605 + layer.1.k_cache 0.30751975 15.97344162 + layer.1.v_cache 0.00000599 0.00543365 + layer.2.k_cache 0.01421910 1.66023126 + layer.2.v_cache 0.00002020 0.01556402 + layer.3.k_cache 0.01730646 7.91772412 + layer.3.v_cache 0.00002068 0.01786475 + layer.4.k_cache 0.00075739 0.40283038 + layer.4.v_cache 0.00005209 0.03438396 + layer.4.output 1.22958295 217.42801563 + ------------------------------------------------------------------------------------- + TOTAL 0.53265962 100.10809598 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 164632 +BPFP 0.6077 bits/point +EBPFP 1.2154 equivalent bits/point +MSE 100.108096 +---------------------- -------------------------------------------------------- +Time: 0.563s Load: 0.009s, Pack+Encode: 0.226s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 100.1081 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 202, 128) +Output shape: (1, 202, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.output: torch.Size([1, 202, 3584]) -> torch.Size([1, 1, 202, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,364B, BPFP=0.3376 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,912B, BPFP=1.6176 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,552B, BPFP=0.5842 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,952B, BPFP=1.5433 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,012B, BPFP=0.6971 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,860B, BPFP=1.4588 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,752B, BPFP=0.6770 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,636B, BPFP=1.5189 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,296B, BPFP=1.1058 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,688B, BPFP=1.4455 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,480B, BPFP=0.1490 +⌛️ [2/4] FRONTEND: Frontend time: 0.240s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.405s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15282421 157.00426400 + layer.0.v_cache 0.00001801 0.01831430 + layer.1.k_cache 0.08597872 16.05780815 + layer.1.v_cache 0.00000649 0.00673325 + layer.2.k_cache 0.01425038 1.78866592 + layer.2.v_cache 0.00002047 0.01759495 + layer.3.k_cache 0.05704349 7.25446947 + layer.3.v_cache 0.00002256 0.02061930 + layer.4.k_cache 0.00069767 0.43816878 + layer.4.v_cache 0.00005056 0.03726856 + layer.4.output 1.51553867 268.14047030 + ------------------------------------------------------------------------------------- + TOTAL 0.64233431 121.15454110 + (elements=1,758,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1758208 +Total Bytes 155504 +BPFP 0.7076 bits/point +EBPFP 1.4151 equivalent bits/point +MSE 121.154541 +---------------------- -------------------------------------------------------- +Time: 0.653s Load: 0.008s, Pack+Encode: 0.240s, Decode+Unpack: 0.405s +---------------------- -------------------------------------------------------- +💾 Converting with 121.1545 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,584B, BPFP=0.3083 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,104B, BPFP=1.4413 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,432B, BPFP=0.5208 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,172B, BPFP=1.3898 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,756B, BPFP=0.5939 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,852B, BPFP=1.3169 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,404B, BPFP=0.5744 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,520B, BPFP=1.3538 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,616B, BPFP=0.9726 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,160B, BPFP=1.2787 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,672B, BPFP=0.1315 +⌛️ [2/4] FRONTEND: Frontend time: 0.379s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.495s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14837397 156.80032575 + layer.0.v_cache 0.00001603 0.01484278 + layer.1.k_cache 0.60911565 16.06596110 + layer.1.v_cache 0.00000597 0.00518321 + layer.2.k_cache 0.02477714 1.79132598 + layer.2.v_cache 0.00001938 0.01456441 + layer.3.k_cache 0.01052417 7.45856683 + layer.3.v_cache 0.00002039 0.01667099 + layer.4.k_cache 0.00073933 0.37686610 + layer.4.v_cache 0.00005963 0.03378630 + layer.4.output 0.00470031 196.03790699 + ------------------------------------------------------------------------------------- + TOTAL 0.04862081 91.46137896 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 193272 +BPFP 0.6277 bits/point +EBPFP 1.2554 equivalent bits/point +MSE 91.461379 +---------------------- -------------------------------------------------------- +Time: 0.883s Load: 0.010s, Pack+Encode: 0.379s, Decode+Unpack: 0.495s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4614 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 210, 128) +Output shape: (1, 210, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.output: torch.Size([1, 210, 3584]) -> torch.Size([1, 1, 210, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,452B, BPFP=0.3312 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,388B, BPFP=1.5914 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,444B, BPFP=0.5539 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,300B, BPFP=1.5104 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,964B, BPFP=0.6670 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,616B, BPFP=1.4595 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,584B, BPFP=0.6387 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,836B, BPFP=1.4759 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,368B, BPFP=1.0690 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,064B, BPFP=1.4185 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,388B, BPFP=0.1423 +⌛️ [2/4] FRONTEND: Frontend time: 0.320s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.410s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11415643 153.89438244 + layer.0.v_cache 0.00001882 0.01721780 + layer.1.k_cache 0.27145578 16.09650065 + layer.1.v_cache 0.00000623 0.00615971 + layer.2.k_cache 0.01201470 1.75923520 + layer.2.v_cache 0.00002208 0.01678464 + layer.3.k_cache 0.00887682 7.07083915 + layer.3.v_cache 0.00002165 0.01951681 + layer.4.k_cache 0.00072611 0.43734378 + layer.4.v_cache 0.00005233 0.03731002 + layer.4.output 1.45781997 257.74141156 + ------------------------------------------------------------------------------------- + TOTAL 0.62424063 116.67912771 + (elements=1,827,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1827840 +Total Bytes 157404 +BPFP 0.6889 bits/point +EBPFP 1.3778 equivalent bits/point +MSE 116.679128 +---------------------- -------------------------------------------------------- +Time: 0.738s Load: 0.007s, Pack+Encode: 0.320s, Decode+Unpack: 0.410s +---------------------- -------------------------------------------------------- +💾 Converting with 116.6791 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 165, 128) +Output shape: (1, 165, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.output: torch.Size([1, 165, 3584]) -> torch.Size([1, 1, 165, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,320B, BPFP=0.3144 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,000B, BPFP=1.6098 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,376B, BPFP=0.6038 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,268B, BPFP=1.5405 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,560B, BPFP=0.7159 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,780B, BPFP=1.4943 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,236B, BPFP=0.6852 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,112B, BPFP=1.5258 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,880B, BPFP=1.1250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,872B, BPFP=1.4083 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,360B, BPFP=0.1537 +⌛️ [2/4] FRONTEND: Frontend time: 0.289s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10245823 155.35999053 + layer.0.v_cache 0.00002070 0.01888618 + layer.1.k_cache 0.13127365 15.85136719 + layer.1.v_cache 0.00000615 0.00628632 + layer.2.k_cache 0.01458293 1.86523438 + layer.2.v_cache 0.00002074 0.01744213 + layer.3.k_cache 0.02375244 7.48129513 + layer.3.v_cache 0.00002214 0.02065017 + layer.4.k_cache 0.00067489 0.44887275 + layer.4.v_cache 0.00005196 0.04027434 + layer.4.output 0.00973386 332.85865801 + ------------------------------------------------------------------------------------- + TOTAL 0.02005887 147.71299442 + (elements=1,436,160) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1436160 +Total Bytes 127764 +BPFP 0.7117 bits/point +EBPFP 1.4234 equivalent bits/point +MSE 147.712994 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.007s, Pack+Encode: 0.289s, Decode+Unpack: 0.355s +---------------------- -------------------------------------------------------- +💾 Converting with 147.7130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,520B, BPFP=0.3005 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,072B, BPFP=1.4577 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,128B, BPFP=0.5232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,528B, BPFP=1.4112 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,916B, BPFP=0.5905 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,828B, BPFP=1.3514 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,408B, BPFP=0.5471 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,248B, BPFP=1.3873 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,896B, BPFP=1.0157 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,112B, BPFP=1.2903 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,348B, BPFP=0.1384 +⌛️ [2/4] FRONTEND: Frontend time: 0.271s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.299s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12273688 149.63596098 + layer.0.v_cache 0.00001959 0.01815397 + layer.1.k_cache 0.11126613 15.42454454 + layer.1.v_cache 0.00000674 0.00698034 + layer.2.k_cache 0.01416154 1.71790501 + layer.2.v_cache 0.00002164 0.01820463 + layer.3.k_cache 0.04491856 7.85880867 + layer.3.v_cache 0.00002114 0.02066467 + layer.4.k_cache 0.00068813 0.42887241 + layer.4.v_cache 0.00006151 0.03963992 + layer.4.output 0.00884247 299.75968482 + ------------------------------------------------------------------------------------- + TOTAL 0.02092936 133.73456052 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 127004 +BPFP 0.6379 bits/point +EBPFP 1.2758 equivalent bits/point +MSE 133.734561 +---------------------- -------------------------------------------------------- +Time: 0.577s Load: 0.007s, Pack+Encode: 0.271s, Decode+Unpack: 0.299s +---------------------- -------------------------------------------------------- +💾 Converting with 133.7346 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 257, 128) +Output shape: (1, 257, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.output: torch.Size([1, 257, 3584]) -> torch.Size([1, 1, 257, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,144B, BPFP=0.3127 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,264B, BPFP=1.5360 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,976B, BPFP=0.5457 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,068B, BPFP=1.4633 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,416B, BPFP=0.6333 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,908B, BPFP=1.3928 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,996B, BPFP=0.6077 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,432B, BPFP=1.4246 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,776B, BPFP=1.0199 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,060B, BPFP=1.3412 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,304B, BPFP=0.1416 +⌛️ [2/4] FRONTEND: Frontend time: 0.359s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.466s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12694686 156.27659898 + layer.0.v_cache 0.00002055 0.01712388 + layer.1.k_cache 0.54844232 16.10530156 + layer.1.v_cache 0.00000676 0.00596011 + layer.2.k_cache 0.01835482 1.83094449 + layer.2.v_cache 0.00002099 0.01644077 + layer.3.k_cache 0.02141366 7.68558345 + layer.3.v_cache 0.00002047 0.01863050 + layer.4.k_cache 0.00070767 0.41702514 + layer.4.v_cache 0.00005046 0.03422582 + layer.4.output 0.00513101 216.01306281 + ------------------------------------------------------------------------------------- + TOTAL 0.04422951 99.67642791 + (elements=2,236,928) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2236928 +Total Bytes 185344 +BPFP 0.6629 bits/point +EBPFP 1.3257 equivalent bits/point +MSE 99.676428 +---------------------- -------------------------------------------------------- +Time: 0.835s Load: 0.010s, Pack+Encode: 0.359s, Decode+Unpack: 0.466s +---------------------- -------------------------------------------------------- +💾 Converting with 99.6764 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,432B, BPFP=0.2946 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,868B, BPFP=1.4481 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,112B, BPFP=0.5247 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,108B, BPFP=1.3829 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,732B, BPFP=0.5780 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,568B, BPFP=1.3365 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,304B, BPFP=0.5412 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,036B, BPFP=1.3767 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,592B, BPFP=0.9952 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,736B, BPFP=1.2651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,892B, BPFP=0.1336 +⌛️ [2/4] FRONTEND: Frontend time: 0.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12731336 151.59990986 + layer.0.v_cache 0.00001750 0.01780725 + layer.1.k_cache 0.10052182 15.57467484 + layer.1.v_cache 0.00000572 0.00579649 + layer.2.k_cache 0.01141533 1.79215886 + layer.2.v_cache 0.00001926 0.01579369 + layer.3.k_cache 0.03745206 7.73528315 + layer.3.v_cache 0.00002083 0.01884432 + layer.4.k_cache 0.00066429 0.40960974 + layer.4.v_cache 0.00005101 0.03647328 + layer.4.output 0.00884527 301.49026197 + ------------------------------------------------------------------------------------- + TOTAL 0.01996459 134.56695207 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 124380 +BPFP 0.6281 bits/point +EBPFP 1.2563 equivalent bits/point +MSE 134.566952 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.006s, Pack+Encode: 0.267s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 134.5670 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,352B, BPFP=0.3063 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,392B, BPFP=1.5056 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,564B, BPFP=0.5324 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,352B, BPFP=1.4324 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,828B, BPFP=0.6213 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,992B, BPFP=1.4071 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,220B, BPFP=0.5785 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,340B, BPFP=1.4316 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,420B, BPFP=1.0149 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,504B, BPFP=1.3024 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,236B, BPFP=0.1431 +⌛️ [2/4] FRONTEND: Frontend time: 0.310s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.413s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12752426 157.68121833 + layer.0.v_cache 0.00002004 0.01753620 + layer.1.k_cache 0.30660921 16.41033771 + layer.1.v_cache 0.00000613 0.00598906 + layer.2.k_cache 0.01808885 1.79291541 + layer.2.v_cache 0.00002107 0.01604685 + layer.3.k_cache 0.01678786 8.13785147 + layer.3.v_cache 0.00002050 0.01842374 + layer.4.k_cache 0.00068290 0.42586139 + layer.4.v_cache 0.00004851 0.03396306 + layer.4.output 1.37905047 243.77095399 + ------------------------------------------------------------------------------------- + TOTAL 0.59548016 111.23157771 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 158200 +BPFP 0.6550 bits/point +EBPFP 1.3099 equivalent bits/point +MSE 111.231578 +---------------------- -------------------------------------------------------- +Time: 0.731s Load: 0.007s, Pack+Encode: 0.310s, Decode+Unpack: 0.413s +---------------------- -------------------------------------------------------- +💾 Converting with 111.2316 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,432B, BPFP=0.3191 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,876B, BPFP=1.5752 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,772B, BPFP=0.5596 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,864B, BPFP=1.5023 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,392B, BPFP=0.6763 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,580B, BPFP=1.4099 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,036B, BPFP=0.6506 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,076B, BPFP=1.4456 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,512B, BPFP=1.0449 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,160B, BPFP=1.3796 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,776B, BPFP=0.1520 +⌛️ [2/4] FRONTEND: Frontend time: 0.283s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.401s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12766143 157.48786722 + layer.0.v_cache 0.00001653 0.01696455 + layer.1.k_cache 0.34631467 16.06657839 + layer.1.v_cache 0.00000659 0.00610523 + layer.2.k_cache 0.01786991 1.77617564 + layer.2.v_cache 0.00002041 0.01599432 + layer.3.k_cache 0.06874437 8.16848516 + layer.3.v_cache 0.00002276 0.02004521 + layer.4.k_cache 0.00067882 0.41227392 + layer.4.v_cache 0.00005621 0.03714922 + layer.4.output 1.41085444 249.44904131 + ------------------------------------------------------------------------------------- + TOTAL 0.61396310 113.53828988 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 161476 +BPFP 0.6839 bits/point +EBPFP 1.3679 equivalent bits/point +MSE 113.538290 +---------------------- -------------------------------------------------------- +Time: 0.692s Load: 0.008s, Pack+Encode: 0.283s, Decode+Unpack: 0.401s +---------------------- -------------------------------------------------------- +💾 Converting with 113.5383 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,376B, BPFP=0.3025 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,256B, BPFP=1.4696 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,760B, BPFP=0.5365 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,564B, BPFP=1.4217 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,988B, BPFP=0.6214 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,932B, BPFP=1.3780 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,788B, BPFP=0.6076 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,740B, BPFP=1.4339 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,456B, BPFP=0.9994 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,740B, BPFP=1.2956 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,388B, BPFP=0.1520 +⌛️ [2/4] FRONTEND: Frontend time: 0.278s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12070838 156.95791275 + layer.0.v_cache 0.00001763 0.01753461 + layer.1.k_cache 0.31015487 15.96872191 + layer.1.v_cache 0.00000677 0.00649391 + layer.2.k_cache 0.03544329 1.72784478 + layer.2.v_cache 0.00002212 0.01706826 + layer.3.k_cache 0.05682086 6.85776756 + layer.3.v_cache 0.00002136 0.02043768 + layer.4.k_cache 0.00073636 0.42792670 + layer.4.v_cache 0.00005031 0.03669286 + layer.4.output 1.35470685 239.70184102 + ------------------------------------------------------------------------------------- + TOTAL 0.58864294 109.40889930 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 160988 +BPFP 0.6547 bits/point +EBPFP 1.3094 equivalent bits/point +MSE 109.408899 +---------------------- -------------------------------------------------------- +Time: 0.673s Load: 0.008s, Pack+Encode: 0.278s, Decode+Unpack: 0.388s +---------------------- -------------------------------------------------------- +💾 Converting with 109.4089 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,404B, BPFP=0.3231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,916B, BPFP=1.6077 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,584B, BPFP=0.5563 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,744B, BPFP=1.5217 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,592B, BPFP=0.7036 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,880B, BPFP=1.4583 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,980B, BPFP=0.6587 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,216B, BPFP=1.4830 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,676B, BPFP=1.0766 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,440B, BPFP=1.4261 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,696B, BPFP=0.1540 +⌛️ [2/4] FRONTEND: Frontend time: 0.329s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.456s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13227103 153.93423562 + layer.0.v_cache 0.00002050 0.01730072 + layer.1.k_cache 0.22864332 16.05786591 + layer.1.v_cache 0.00000662 0.00633039 + layer.2.k_cache 0.02950178 1.80760272 + layer.2.v_cache 0.00002095 0.01730024 + layer.3.k_cache 0.01835468 7.05338312 + layer.3.v_cache 0.00002123 0.02045047 + layer.4.k_cache 0.00070999 0.43647773 + layer.4.v_cache 0.00005457 0.03706889 + layer.4.output 1.43734575 254.20948608 + ------------------------------------------------------------------------------------- + TOTAL 0.61594264 115.22673049 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 162128 +BPFP 0.6996 bits/point +EBPFP 1.3992 equivalent bits/point +MSE 115.226730 +---------------------- -------------------------------------------------------- +Time: 0.792s Load: 0.008s, Pack+Encode: 0.329s, Decode+Unpack: 0.456s +---------------------- -------------------------------------------------------- +💾 Converting with 115.2267 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,308B, BPFP=0.3032 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,312B, BPFP=1.5000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,612B, BPFP=0.5358 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,856B, BPFP=1.4679 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,832B, BPFP=0.6216 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,120B, BPFP=1.4161 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,476B, BPFP=0.5966 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,288B, BPFP=1.4279 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,576B, BPFP=1.0259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,448B, BPFP=1.2984 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,856B, BPFP=0.1393 +⌛️ [2/4] FRONTEND: Frontend time: 0.320s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.423s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12518616 160.97469735 + layer.0.v_cache 0.00001676 0.01657037 + layer.1.k_cache 0.33007166 16.49736504 + layer.1.v_cache 0.00000630 0.00619092 + layer.2.k_cache 0.01131002 1.81418369 + layer.2.v_cache 0.00002125 0.01689851 + layer.3.k_cache 0.02232737 7.62161667 + layer.3.v_cache 0.00001983 0.01777089 + layer.4.k_cache 0.00071526 0.42617117 + layer.4.v_cache 0.00005006 0.03509149 + layer.4.output 1.37904434 243.77433237 + ------------------------------------------------------------------------------------- + TOTAL 0.59664912 111.40275781 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 158684 +BPFP 0.6570 bits/point +EBPFP 1.3140 equivalent bits/point +MSE 111.402758 +---------------------- -------------------------------------------------------- +Time: 0.753s Load: 0.010s, Pack+Encode: 0.320s, Decode+Unpack: 0.423s +---------------------- -------------------------------------------------------- +💾 Converting with 111.4028 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,528B, BPFP=0.2900 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,948B, BPFP=1.4055 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,920B, BPFP=0.5072 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,028B, BPFP=1.3466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,916B, BPFP=0.5710 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,068B, BPFP=1.2851 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,668B, BPFP=0.5551 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,560B, BPFP=1.3166 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,652B, BPFP=0.9383 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,156B, BPFP=1.2267 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,132B, BPFP=0.1384 +⌛️ [2/4] FRONTEND: Frontend time: 0.323s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.465s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14655781 153.53214652 + layer.0.v_cache 0.00001709 0.01597988 + layer.1.k_cache 0.36683986 16.30739806 + layer.1.v_cache 0.00000651 0.00611252 + layer.2.k_cache 0.01755252 1.83001146 + layer.2.v_cache 0.00002060 0.01649302 + layer.3.k_cache 0.02713263 7.63076632 + layer.3.v_cache 0.00001997 0.01759984 + layer.4.k_cache 0.00071102 0.40823955 + layer.4.v_cache 0.00005162 0.03390302 + layer.4.output 1.25479176 221.89567477 + ------------------------------------------------------------------------------------- + TOTAL 0.54955600 101.94519844 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 162576 +BPFP 0.6124 bits/point +EBPFP 1.2248 equivalent bits/point +MSE 101.945198 +---------------------- -------------------------------------------------------- +Time: 0.797s Load: 0.008s, Pack+Encode: 0.323s, Decode+Unpack: 0.465s +---------------------- -------------------------------------------------------- +💾 Converting with 101.9452 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 155, 128) +Output shape: (1, 155, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.output: torch.Size([1, 155, 3584]) -> torch.Size([1, 1, 155, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,320B, BPFP=0.3347 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,892B, BPFP=1.6020 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,028B, BPFP=0.6077 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,652B, BPFP=1.5778 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,996B, BPFP=0.7052 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,292B, BPFP=1.5415 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,932B, BPFP=0.6988 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,924B, BPFP=1.6052 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,384B, BPFP=1.1476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,320B, BPFP=1.4435 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,948B, BPFP=0.1433 +⌛️ [2/4] FRONTEND: Frontend time: 0.310s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102877 167.73828125 + layer.0.v_cache 0.00001822 0.01720588 + layer.1.k_cache 0.12773433 16.34526210 + layer.1.v_cache 0.00000585 0.00632546 + layer.2.k_cache 0.00931232 1.87618723 + layer.2.v_cache 0.00002110 0.01841028 + layer.3.k_cache 0.01537714 7.16026414 + layer.3.v_cache 0.00002091 0.02008286 + layer.4.k_cache 0.00066471 0.44307866 + layer.4.v_cache 0.00005180 0.04013102 + layer.4.output 0.01742802 357.44288594 + ------------------------------------------------------------------------------------- + TOTAL 0.02389596 158.57443709 + (elements=1,349,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1349120 +Total Bytes 121688 +BPFP 0.7216 bits/point +EBPFP 1.4432 equivalent bits/point +MSE 158.574437 +---------------------- -------------------------------------------------------- +Time: 0.698s Load: 0.006s, Pack+Encode: 0.310s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 158.5744 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,260B, BPFP=0.3096 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,192B, BPFP=1.6128 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,824B, BPFP=0.5686 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,168B, BPFP=1.5384 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,168B, BPFP=0.6663 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,784B, BPFP=1.4378 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,532B, BPFP=0.6201 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,088B, BPFP=1.4599 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,664B, BPFP=1.0657 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,188B, BPFP=1.3945 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,636B, BPFP=0.1416 +⌛️ [2/4] FRONTEND: Frontend time: 0.365s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.432s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11659910 154.42937863 + layer.0.v_cache 0.00001673 0.01776294 + layer.1.k_cache 0.29571861 15.58841524 + layer.1.v_cache 0.00000615 0.00614381 + layer.2.k_cache 0.04707494 1.84201348 + layer.2.v_cache 0.00002002 0.01589474 + layer.3.k_cache 0.03274950 7.26386605 + layer.3.v_cache 0.00002078 0.01839841 + layer.4.k_cache 0.00068608 0.42470710 + layer.4.v_cache 0.00005136 0.03572052 + layer.4.output 1.42391751 251.73091777 + ------------------------------------------------------------------------------------- + TOTAL 0.61531564 114.22110149 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 160504 +BPFP 0.6861 bits/point +EBPFP 1.3723 equivalent bits/point +MSE 114.221101 +---------------------- -------------------------------------------------------- +Time: 0.804s Load: 0.008s, Pack+Encode: 0.365s, Decode+Unpack: 0.432s +---------------------- -------------------------------------------------------- +💾 Converting with 114.2211 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,572B, BPFP=0.2881 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,696B, BPFP=1.4299 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,912B, BPFP=0.4985 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,744B, BPFP=1.3700 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,960B, BPFP=0.5645 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,836B, BPFP=1.3128 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,360B, BPFP=0.5267 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,364B, BPFP=1.3460 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,484B, BPFP=0.9756 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,012B, BPFP=1.2608 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,468B, BPFP=0.1392 +⌛️ [2/4] FRONTEND: Frontend time: 0.326s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.448s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10631144 151.08080267 + layer.0.v_cache 0.00001780 0.01715144 + layer.1.k_cache 0.46192560 15.73553565 + layer.1.v_cache 0.00000619 0.00627355 + layer.2.k_cache 0.02890189 1.67546635 + layer.2.v_cache 0.00002115 0.01654061 + layer.3.k_cache 0.04994748 7.19004969 + layer.3.v_cache 0.00002140 0.01899444 + layer.4.k_cache 0.00068365 0.41661884 + layer.4.v_cache 0.00005040 0.03623293 + layer.4.output 1.23456514 218.11594542 + ------------------------------------------------------------------------------------- + TOTAL 0.54646135 100.17678142 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 167408 +BPFP 0.6204 bits/point +EBPFP 1.2409 equivalent bits/point +MSE 100.176781 +---------------------- -------------------------------------------------------- +Time: 0.782s Load: 0.009s, Pack+Encode: 0.326s, Decode+Unpack: 0.448s +---------------------- -------------------------------------------------------- +💾 Converting with 100.1768 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 194, 128) +Output shape: (1, 194, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.output: torch.Size([1, 194, 3584]) -> torch.Size([1, 1, 194, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,160B, BPFP=0.3351 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,944B, BPFP=1.6063 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,268B, BPFP=0.5854 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,180B, BPFP=1.5448 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,796B, BPFP=0.7084 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,568B, BPFP=1.4955 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,280B, BPFP=0.6669 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,920B, BPFP=1.5238 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,940B, BPFP=1.1227 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,164B, BPFP=1.4630 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,976B, BPFP=0.1378 +⌛️ [2/4] FRONTEND: Frontend time: 0.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.424s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15027005 154.42124074 + layer.0.v_cache 0.00001782 0.01815465 + layer.1.k_cache 0.21402168 16.05123807 + layer.1.v_cache 0.00000588 0.00650293 + layer.2.k_cache 0.02177384 1.77650861 + layer.2.v_cache 0.00002019 0.01831505 + layer.3.k_cache 0.04024272 8.00407362 + layer.3.v_cache 0.00002076 0.02008644 + layer.4.k_cache 0.00069591 0.44586728 + layer.4.v_cache 0.00005073 0.03910953 + layer.4.output 1.57794378 279.13671300 + ------------------------------------------------------------------------------------- + TOTAL 0.67486624 125.57400517 + (elements=1,688,576) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1688576 +Total Bytes 149196 +BPFP 0.7068 bits/point +EBPFP 1.4137 equivalent bits/point +MSE 125.574005 +---------------------- -------------------------------------------------------- +Time: 0.741s Load: 0.007s, Pack+Encode: 0.309s, Decode+Unpack: 0.424s +---------------------- -------------------------------------------------------- +💾 Converting with 125.5740 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 258, 128) +Output shape: (1, 258, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.output: torch.Size([1, 258, 3584]) -> torch.Size([1, 1, 258, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,272B, BPFP=0.3193 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,368B, BPFP=1.5363 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,860B, BPFP=0.5366 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,152B, BPFP=1.4627 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,652B, BPFP=0.6451 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,484B, BPFP=1.4222 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,760B, BPFP=0.5911 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,796B, BPFP=1.4411 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,184B, BPFP=1.0407 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,700B, BPFP=1.3748 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,760B, BPFP=0.1277 +⌛️ [2/4] FRONTEND: Frontend time: 0.395s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.528s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09035230 157.26574612 + layer.0.v_cache 0.00001726 0.01713129 + layer.1.k_cache 0.42455262 16.14363985 + layer.1.v_cache 0.00000600 0.00585021 + layer.2.k_cache 0.01891134 1.79791792 + layer.2.v_cache 0.00001996 0.01600986 + layer.3.k_cache 0.02093011 7.42346759 + layer.3.v_cache 0.00001928 0.01741378 + layer.4.k_cache 0.00072082 0.41288663 + layer.4.v_cache 0.00004951 0.03583538 + layer.4.output 0.00505962 215.12105482 + ------------------------------------------------------------------------------------- + TOTAL 0.03476450 99.35195779 + (elements=2,245,632) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2245632 +Total Bytes 185988 +BPFP 0.6626 bits/point +EBPFP 1.3252 equivalent bits/point +MSE 99.351958 +---------------------- -------------------------------------------------------- +Time: 0.932s Load: 0.009s, Pack+Encode: 0.395s, Decode+Unpack: 0.528s +---------------------- -------------------------------------------------------- +💾 Converting with 99.3520 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,488B, BPFP=0.3049 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,936B, BPFP=1.4902 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,112B, BPFP=0.5511 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,308B, BPFP=1.4476 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,736B, BPFP=0.6614 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,332B, BPFP=1.3813 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,944B, BPFP=0.6076 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,012B, BPFP=1.4274 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,328B, BPFP=1.0413 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,208B, BPFP=1.3049 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,436B, BPFP=0.1595 +⌛️ [2/4] FRONTEND: Frontend time: 0.302s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.415s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15008528 150.74023438 + layer.0.v_cache 0.00001847 0.01678177 + layer.1.k_cache 0.25612781 15.90811184 + layer.1.v_cache 0.00000678 0.00600572 + layer.2.k_cache 0.03931687 1.69793754 + layer.2.v_cache 0.00002107 0.01623407 + layer.3.k_cache 0.03786138 7.29919327 + layer.3.v_cache 0.00002177 0.01964266 + layer.4.k_cache 0.00069674 0.41374810 + layer.4.v_cache 0.00005115 0.03482172 + layer.4.output 1.33117570 235.51853649 + ------------------------------------------------------------------------------------- + TOTAL 0.57661395 107.34014509 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 166840 +BPFP 0.6667 bits/point +EBPFP 1.3334 equivalent bits/point +MSE 107.340145 +---------------------- -------------------------------------------------------- +Time: 0.725s Load: 0.008s, Pack+Encode: 0.302s, Decode+Unpack: 0.415s +---------------------- -------------------------------------------------------- +💾 Converting with 107.3401 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,488B, BPFP=0.3028 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,908B, BPFP=1.4677 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,248B, BPFP=0.5424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,236B, BPFP=1.4094 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,992B, BPFP=0.6069 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,632B, BPFP=1.3569 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,512B, BPFP=0.5653 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,008B, BPFP=1.3896 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,472B, BPFP=0.9958 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,828B, BPFP=1.2872 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,948B, BPFP=0.1482 +⌛️ [2/4] FRONTEND: Frontend time: 0.303s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15161699 154.67878689 + layer.0.v_cache 0.00001786 0.01802833 + layer.1.k_cache 0.08662784 16.01342773 + layer.1.v_cache 0.00000652 0.00635061 + layer.2.k_cache 0.02063684 1.76421967 + layer.2.v_cache 0.00002327 0.01775277 + layer.3.k_cache 0.08087930 6.97548760 + layer.3.v_cache 0.00002116 0.02039126 + layer.4.k_cache 0.00074161 0.43854137 + layer.4.v_cache 0.00006096 0.03871926 + layer.4.output 0.00898438 304.97442956 + ------------------------------------------------------------------------------------- + TOTAL 0.02373665 136.16427720 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 126272 +BPFP 0.6448 bits/point +EBPFP 1.2895 equivalent bits/point +MSE 136.164277 +---------------------- -------------------------------------------------------- +Time: 0.677s Load: 0.006s, Pack+Encode: 0.303s, Decode+Unpack: 0.367s +---------------------- -------------------------------------------------------- +💾 Converting with 136.1643 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,480B, BPFP=0.3182 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,096B, BPFP=1.5693 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,084B, BPFP=0.5741 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,880B, BPFP=1.4830 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,144B, BPFP=0.6494 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,220B, BPFP=1.4361 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,608B, BPFP=0.6114 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,956B, BPFP=1.4884 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,872B, BPFP=1.0562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,128B, BPFP=1.3585 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,664B, BPFP=0.1386 +⌛️ [2/4] FRONTEND: Frontend time: 0.281s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.405s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14327769 157.86443537 + layer.0.v_cache 0.00001763 0.01749665 + layer.1.k_cache 0.34925704 16.04299649 + layer.1.v_cache 0.00000642 0.00650502 + layer.2.k_cache 0.02352169 1.77154957 + layer.2.v_cache 0.00002271 0.01773388 + layer.3.k_cache 0.00845094 7.65548373 + layer.3.v_cache 0.00002026 0.01943034 + layer.4.k_cache 0.00069169 0.43340610 + layer.4.v_cache 0.00005086 0.03789007 + layer.4.output 1.39158238 246.10531656 + ------------------------------------------------------------------------------------- + TOTAL 0.60390550 112.15318489 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 162132 +BPFP 0.6774 bits/point +EBPFP 1.3547 equivalent bits/point +MSE 112.153185 +---------------------- -------------------------------------------------------- +Time: 0.693s Load: 0.007s, Pack+Encode: 0.281s, Decode+Unpack: 0.405s +---------------------- -------------------------------------------------------- +💾 Converting with 112.1532 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,504B, BPFP=0.3127 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,012B, BPFP=1.5348 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,388B, BPFP=0.5334 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,804B, BPFP=1.4661 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,360B, BPFP=0.6455 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,324B, BPFP=1.3820 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,908B, BPFP=0.5630 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,864B, BPFP=1.4127 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,076B, BPFP=1.0270 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,884B, BPFP=1.3570 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,396B, BPFP=0.1331 +⌛️ [2/4] FRONTEND: Frontend time: 0.358s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.484s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12716211 151.54609375 + layer.0.v_cache 0.00001677 0.01676174 + layer.1.k_cache 0.51313377 16.11536044 + layer.1.v_cache 0.00000586 0.00605656 + layer.2.k_cache 0.02995576 1.81152299 + layer.2.v_cache 0.00002133 0.01697138 + layer.3.k_cache 0.03339291 7.38669478 + layer.3.v_cache 0.00002012 0.01804901 + layer.4.k_cache 0.00068476 0.41905900 + layer.4.v_cache 0.00005875 0.03579986 + layer.4.output 0.00481428 201.72689935 + ------------------------------------------------------------------------------------- + TOTAL 0.04342071 93.49768617 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 196520 +BPFP 0.6568 bits/point +EBPFP 1.3136 equivalent bits/point +MSE 93.497686 +---------------------- -------------------------------------------------------- +Time: 0.852s Load: 0.010s, Pack+Encode: 0.358s, Decode+Unpack: 0.484s +---------------------- -------------------------------------------------------- +💾 Converting with 93.4977 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,796B, BPFP=0.2950 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,232B, BPFP=1.3676 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,096B, BPFP=0.4980 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,492B, BPFP=1.3221 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,048B, BPFP=0.5566 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,432B, BPFP=1.2569 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,332B, BPFP=0.5125 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,904B, BPFP=1.2859 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,204B, BPFP=0.9353 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,780B, BPFP=1.2168 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,964B, BPFP=0.1403 +⌛️ [2/4] FRONTEND: Frontend time: 0.305s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.400s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13396681 155.72582431 + layer.0.v_cache 0.00001641 0.01647474 + layer.1.k_cache 0.47111181 15.85486993 + layer.1.v_cache 0.00000608 0.00597651 + layer.2.k_cache 0.02933135 1.63728357 + layer.2.v_cache 0.00002255 0.01647627 + layer.3.k_cache 0.03744583 7.02738604 + layer.3.v_cache 0.00002260 0.01762792 + layer.4.k_cache 0.00071777 0.40722080 + layer.4.v_cache 0.00004917 0.03506615 + layer.4.output 1.20538323 213.11176533 + ------------------------------------------------------------------------------------- + TOTAL 0.53590429 98.38391550 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 166280 +BPFP 0.6017 bits/point +EBPFP 1.2034 equivalent bits/point +MSE 98.383916 +---------------------- -------------------------------------------------------- +Time: 0.714s Load: 0.009s, Pack+Encode: 0.305s, Decode+Unpack: 0.400s +---------------------- -------------------------------------------------------- +💾 Converting with 98.3839 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,192B, BPFP=0.2817 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,416B, BPFP=1.4332 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,984B, BPFP=0.4874 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,728B, BPFP=1.3958 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,764B, BPFP=0.5840 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,800B, BPFP=1.3455 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,472B, BPFP=0.5681 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,216B, BPFP=1.3681 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,280B, BPFP=0.9918 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,248B, BPFP=1.2613 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,660B, BPFP=0.1291 +⌛️ [2/4] FRONTEND: Frontend time: 0.365s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.477s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14060297 157.64496528 + layer.0.v_cache 0.00001598 0.01611978 + layer.1.k_cache 0.58610762 15.70328098 + layer.1.v_cache 0.00000669 0.00583643 + layer.2.k_cache 0.02834670 1.80941688 + layer.2.v_cache 0.00002256 0.01604177 + layer.3.k_cache 0.01962393 7.19247437 + layer.3.v_cache 0.00002003 0.01755293 + layer.4.k_cache 0.00067797 0.40910572 + layer.4.v_cache 0.00005035 0.03500187 + layer.4.output 0.00463834 192.69597904 + ------------------------------------------------------------------------------------- + TOTAL 0.04752607 90.10127349 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 195760 +BPFP 0.6247 bits/point +EBPFP 1.2495 equivalent bits/point +MSE 90.101273 +---------------------- -------------------------------------------------------- +Time: 0.851s Load: 0.009s, Pack+Encode: 0.365s, Decode+Unpack: 0.477s +---------------------- -------------------------------------------------------- +💾 Converting with 90.1013 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,508B, BPFP=0.2831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,700B, BPFP=1.4237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,532B, BPFP=0.4899 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,588B, BPFP=1.3666 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,996B, BPFP=0.5652 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,272B, BPFP=1.2989 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,496B, BPFP=0.5395 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,788B, BPFP=1.3255 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,492B, BPFP=0.9505 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,156B, BPFP=1.2416 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,088B, BPFP=0.1255 +⌛️ [2/4] FRONTEND: Frontend time: 0.374s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.501s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12404394 156.37862356 + layer.0.v_cache 0.00001645 0.01600116 + layer.1.k_cache 0.57408182 15.82733957 + layer.1.v_cache 0.00000616 0.00561731 + layer.2.k_cache 0.03522489 1.79029927 + layer.2.v_cache 0.00002326 0.01595740 + layer.3.k_cache 0.02908128 7.56320190 + layer.3.v_cache 0.00002078 0.01765592 + layer.4.k_cache 0.00072493 0.40392449 + layer.4.v_cache 0.00004967 0.03412016 + layer.4.output 0.04390477 178.33244243 + ------------------------------------------------------------------------------------- + TOTAL 0.06297686 84.13999046 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 201616 +BPFP 0.6096 bits/point +EBPFP 1.2191 equivalent bits/point +MSE 84.139990 +---------------------- -------------------------------------------------------- +Time: 0.885s Load: 0.010s, Pack+Encode: 0.374s, Decode+Unpack: 0.501s +---------------------- -------------------------------------------------------- +💾 Converting with 84.1400 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 242, 128) +Output shape: (1, 242, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.output: torch.Size([1, 242, 3584]) -> torch.Size([1, 1, 242, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,504B, BPFP=0.2908 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,244B, BPFP=1.4362 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,716B, BPFP=0.4982 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,272B, BPFP=1.3735 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,524B, BPFP=0.5504 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,456B, BPFP=1.3208 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,216B, BPFP=0.5305 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,944B, BPFP=1.3523 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,916B, BPFP=0.9631 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,456B, BPFP=1.2562 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,892B, BPFP=0.1374 +⌛️ [2/4] FRONTEND: Frontend time: 0.322s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.471s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11466674 150.74641658 + layer.0.v_cache 0.00001794 0.01657135 + layer.1.k_cache 0.39062487 16.09970723 + layer.1.v_cache 0.00000616 0.00587760 + layer.2.k_cache 0.02414889 1.84453640 + layer.2.v_cache 0.00002046 0.01623663 + layer.3.k_cache 0.03495786 7.08415739 + layer.3.v_cache 0.00002053 0.01818944 + layer.4.k_cache 0.00069125 0.40974105 + layer.4.v_cache 0.00005034 0.03680522 + layer.4.output 1.26514320 223.67737234 + ------------------------------------------------------------------------------------- + TOTAL 0.55418868 102.47175561 + (elements=2,106,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2106368 +Total Bytes 163140 +BPFP 0.6196 bits/point +EBPFP 1.2392 equivalent bits/point +MSE 102.471756 +---------------------- -------------------------------------------------------- +Time: 0.802s Load: 0.009s, Pack+Encode: 0.322s, Decode+Unpack: 0.471s +---------------------- -------------------------------------------------------- +💾 Converting with 102.4718 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 189, 128) +Output shape: (1, 189, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.output: torch.Size([1, 189, 3584]) -> torch.Size([1, 1, 189, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,620B, BPFP=0.2993 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,188B, BPFP=1.4210 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,432B, BPFP=0.5317 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,432B, BPFP=1.3585 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,144B, BPFP=0.5906 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,012B, BPFP=1.3237 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,660B, BPFP=0.5506 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,388B, BPFP=1.3548 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,908B, BPFP=0.9845 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,344B, BPFP=1.2685 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,560B, BPFP=0.1483 +⌛️ [2/4] FRONTEND: Frontend time: 0.318s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14361043 154.24092675 + layer.0.v_cache 0.00001648 0.01734113 + layer.1.k_cache 0.18932698 15.61181641 + layer.1.v_cache 0.00000663 0.00659783 + layer.2.k_cache 0.00768016 1.75336549 + layer.2.v_cache 0.00002453 0.01750755 + layer.3.k_cache 0.01421528 7.15624160 + layer.3.v_cache 0.00002024 0.01933098 + layer.4.k_cache 0.00072804 0.41845594 + layer.4.v_cache 0.00005086 0.03668987 + layer.4.output 0.00858909 290.40459656 + ------------------------------------------------------------------------------------- + TOTAL 0.02445902 130.12414409 + (elements=1,645,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1645056 +Total Bytes 129688 +BPFP 0.6307 bits/point +EBPFP 1.2614 equivalent bits/point +MSE 130.124144 +---------------------- -------------------------------------------------------- +Time: 0.697s Load: 0.008s, Pack+Encode: 0.318s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 130.1241 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,372B, BPFP=0.3207 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,300B, BPFP=1.5625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,728B, BPFP=0.5669 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,384B, BPFP=1.4953 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,644B, BPFP=0.7075 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,476B, BPFP=1.4287 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,828B, BPFP=0.6476 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,936B, BPFP=1.4624 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,260B, BPFP=1.0461 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,756B, BPFP=1.3759 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,012B, BPFP=0.1364 +⌛️ [2/4] FRONTEND: Frontend time: 0.347s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16894778 153.84554724 + layer.0.v_cache 0.00001779 0.01719509 + layer.1.k_cache 0.37628758 15.91582077 + layer.1.v_cache 0.00000652 0.00626836 + layer.2.k_cache 0.02181128 1.76693110 + layer.2.v_cache 0.00002030 0.01659020 + layer.3.k_cache 0.01815432 7.80316019 + layer.3.v_cache 0.00002063 0.01904473 + layer.4.k_cache 0.00071780 0.41801711 + layer.4.v_cache 0.00005059 0.03677116 + layer.4.output 1.43727796 254.21478873 + ------------------------------------------------------------------------------------- + TOTAL 0.62629296 115.25581571 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 157696 +BPFP 0.6805 bits/point +EBPFP 1.3610 equivalent bits/point +MSE 115.255816 +---------------------- -------------------------------------------------------- +Time: 0.794s Load: 0.008s, Pack+Encode: 0.347s, Decode+Unpack: 0.439s +---------------------- -------------------------------------------------------- +💾 Converting with 115.2558 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,456B, BPFP=0.3223 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,128B, BPFP=1.5284 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,932B, BPFP=0.5738 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,576B, BPFP=1.4884 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,248B, BPFP=0.6690 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,600B, BPFP=1.4178 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,004B, BPFP=0.6513 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,088B, BPFP=1.4531 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,364B, BPFP=1.0391 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,580B, BPFP=1.3440 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,020B, BPFP=0.1345 +⌛️ [2/4] FRONTEND: Frontend time: 0.312s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.436s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09877178 161.20572917 + layer.0.v_cache 0.00001652 0.01651440 + layer.1.k_cache 0.28824160 16.00238150 + layer.1.v_cache 0.00000619 0.00602771 + layer.2.k_cache 0.02137197 1.82301105 + layer.2.v_cache 0.00001995 0.01578736 + layer.3.k_cache 0.03811082 7.87555497 + layer.3.v_cache 0.00002058 0.01773335 + layer.4.k_cache 0.00070031 0.40932369 + layer.4.v_cache 0.00004695 0.03331639 + layer.4.output 1.41729720 250.54551091 + ------------------------------------------------------------------------------------- + TOTAL 0.60990512 114.18964447 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 157996 +BPFP 0.6723 bits/point +EBPFP 1.3446 equivalent bits/point +MSE 114.189644 +---------------------- -------------------------------------------------------- +Time: 0.757s Load: 0.008s, Pack+Encode: 0.312s, Decode+Unpack: 0.436s +---------------------- -------------------------------------------------------- +💾 Converting with 114.1896 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,324B, BPFP=0.2989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,516B, BPFP=1.4876 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,788B, BPFP=0.5384 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,748B, BPFP=1.4345 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,908B, BPFP=0.6159 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,496B, BPFP=1.3479 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,240B, BPFP=0.5697 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,120B, BPFP=1.3910 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,520B, BPFP=1.0039 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,520B, BPFP=1.2804 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,676B, BPFP=0.1450 +⌛️ [2/4] FRONTEND: Frontend time: 0.286s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.436s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12161526 154.34808836 + layer.0.v_cache 0.00001916 0.01691849 + layer.1.k_cache 0.26028682 16.25802639 + layer.1.v_cache 0.00000628 0.00604687 + layer.2.k_cache 0.01296707 1.79855185 + layer.2.v_cache 0.00002040 0.01595453 + layer.3.k_cache 0.04595671 7.62444636 + layer.3.v_cache 0.00002073 0.01798005 + layer.4.k_cache 0.00067352 0.40583565 + layer.4.v_cache 0.00005047 0.03569992 + layer.4.output 1.35465500 239.67722819 + ------------------------------------------------------------------------------------- + TOTAL 0.58377655 109.30989093 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 158856 +BPFP 0.6461 bits/point +EBPFP 1.2921 equivalent bits/point +MSE 109.309891 +---------------------- -------------------------------------------------------- +Time: 0.730s Load: 0.008s, Pack+Encode: 0.286s, Decode+Unpack: 0.436s +---------------------- -------------------------------------------------------- +💾 Converting with 109.3099 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,396B, BPFP=0.3039 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,068B, BPFP=1.4566 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,784B, BPFP=0.5382 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,712B, BPFP=1.4320 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,836B, BPFP=0.6109 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,060B, BPFP=1.3869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,588B, BPFP=0.5938 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,456B, BPFP=1.4143 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,464B, BPFP=1.0000 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,872B, BPFP=1.3048 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,832B, BPFP=0.1366 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.447s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10229862 157.24168626 + layer.0.v_cache 0.00001727 0.01609169 + layer.1.k_cache 0.34669285 16.00895002 + layer.1.v_cache 0.00000643 0.00584371 + layer.2.k_cache 0.01639456 1.82092042 + layer.2.v_cache 0.00002164 0.01619623 + layer.3.k_cache 0.03343437 7.45482534 + layer.3.v_cache 0.00002043 0.01783695 + layer.4.k_cache 0.00071218 0.39695189 + layer.4.v_cache 0.00005088 0.03366960 + layer.4.output 1.35462105 239.65862042 + ------------------------------------------------------------------------------------- + TOTAL 0.58717627 109.44843030 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 159068 +BPFP 0.6469 bits/point +EBPFP 1.2938 equivalent bits/point +MSE 109.448430 +---------------------- -------------------------------------------------------- +Time: 0.775s Load: 0.009s, Pack+Encode: 0.319s, Decode+Unpack: 0.447s +---------------------- -------------------------------------------------------- +💾 Converting with 109.4484 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,460B, BPFP=0.3070 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,748B, BPFP=1.4970 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,804B, BPFP=0.5372 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,984B, BPFP=1.4444 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,300B, BPFP=0.6401 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,192B, BPFP=1.3899 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,056B, BPFP=0.6233 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,096B, BPFP=1.4521 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,692B, BPFP=1.0113 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,004B, BPFP=1.3081 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,120B, BPFP=0.1388 +⌛️ [2/4] FRONTEND: Frontend time: 0.354s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.428s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11800324 153.97838656 + layer.0.v_cache 0.00001819 0.01654043 + layer.1.k_cache 0.35011090 16.27369992 + layer.1.v_cache 0.00000612 0.00566834 + layer.2.k_cache 0.02017111 1.72254735 + layer.2.v_cache 0.00002006 0.01593855 + layer.3.k_cache 0.03026835 7.81010000 + layer.3.v_cache 0.00001975 0.01824882 + layer.4.k_cache 0.00074701 0.40849580 + layer.4.v_cache 0.00005118 0.03430158 + layer.4.output 1.34863711 238.56039569 + ------------------------------------------------------------------------------------- + TOTAL 0.58587504 108.83568807 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 162456 +BPFP 0.6578 bits/point +EBPFP 1.3156 equivalent bits/point +MSE 108.835688 +---------------------- -------------------------------------------------------- +Time: 0.793s Load: 0.010s, Pack+Encode: 0.354s, Decode+Unpack: 0.428s +---------------------- -------------------------------------------------------- +💾 Converting with 108.8357 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,452B, BPFP=0.3038 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,656B, BPFP=1.4776 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,104B, BPFP=0.5529 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,220B, BPFP=1.4479 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,424B, BPFP=0.6430 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,016B, BPFP=1.4340 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,828B, BPFP=0.6023 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,476B, BPFP=1.4653 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,288B, BPFP=1.0431 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,440B, BPFP=1.3264 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,132B, BPFP=0.1377 +⌛️ [2/4] FRONTEND: Frontend time: 0.323s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.405s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10957497 154.33975846 + layer.0.v_cache 0.00001697 0.01698342 + layer.1.k_cache 0.31262030 15.91354757 + layer.1.v_cache 0.00000629 0.00635898 + layer.2.k_cache 0.01311818 1.79852255 + layer.2.v_cache 0.00002149 0.01723817 + layer.3.k_cache 0.06227440 7.00258640 + layer.3.v_cache 0.00002139 0.02011348 + layer.4.k_cache 0.00068925 0.42945802 + layer.4.v_cache 0.00005302 0.03948930 + layer.4.output 1.33688878 236.51618060 + ------------------------------------------------------------------------------------- + TOTAL 0.57980104 107.95278356 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 165036 +BPFP 0.6624 bits/point +EBPFP 1.3248 equivalent bits/point +MSE 107.952784 +---------------------- -------------------------------------------------------- +Time: 0.736s Load: 0.008s, Pack+Encode: 0.323s, Decode+Unpack: 0.405s +---------------------- -------------------------------------------------------- +💾 Converting with 107.9528 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,372B, BPFP=0.3207 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,224B, BPFP=1.5569 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,436B, BPFP=0.5455 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,404B, BPFP=1.4968 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,436B, BPFP=0.6188 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,416B, BPFP=1.4243 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,084B, BPFP=0.5930 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,536B, BPFP=1.4331 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,264B, BPFP=1.0464 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,932B, BPFP=1.3888 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,244B, BPFP=0.1388 +⌛️ [2/4] FRONTEND: Frontend time: 0.293s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.429s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13987277 150.03550469 + layer.0.v_cache 0.00001887 0.01677887 + layer.1.k_cache 0.23498858 16.11755772 + layer.1.v_cache 0.00000603 0.00598241 + layer.2.k_cache 0.02028183 1.81863991 + layer.2.v_cache 0.00001992 0.01615319 + layer.3.k_cache 0.05552547 7.94658536 + layer.3.v_cache 0.00002044 0.01916733 + layer.4.k_cache 0.00067773 0.41852767 + layer.4.v_cache 0.00004777 0.03546925 + layer.4.output 1.43727485 254.20334507 + ------------------------------------------------------------------------------------- + TOTAL 0.61837549 115.05022246 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 155348 +BPFP 0.6703 bits/point +EBPFP 1.3407 equivalent bits/point +MSE 115.050222 +---------------------- -------------------------------------------------------- +Time: 0.729s Load: 0.008s, Pack+Encode: 0.293s, Decode+Unpack: 0.429s +---------------------- -------------------------------------------------------- +💾 Converting with 115.0502 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,448B, BPFP=0.3294 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,976B, BPFP=1.5533 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,680B, BPFP=0.5687 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,084B, BPFP=1.4873 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,104B, BPFP=0.6742 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,824B, BPFP=1.3940 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,276B, BPFP=0.6129 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,160B, BPFP=1.4188 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,052B, BPFP=1.0406 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,324B, BPFP=1.3569 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,256B, BPFP=0.1402 +⌛️ [2/4] FRONTEND: Frontend time: 0.303s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.409s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17004660 151.68264959 + layer.0.v_cache 0.00001740 0.01679798 + layer.1.k_cache 0.16740718 16.07711604 + layer.1.v_cache 0.00000654 0.00637371 + layer.2.k_cache 0.01191964 1.83688745 + layer.2.v_cache 0.00001991 0.01683038 + layer.3.k_cache 0.00660476 7.19022662 + layer.3.v_cache 0.00002076 0.01923541 + layer.4.k_cache 0.00070551 0.43359151 + layer.4.v_cache 0.00005179 0.03701649 + layer.4.output 1.45088344 256.51449306 + ------------------------------------------------------------------------------------- + TOTAL 0.61841083 116.05401039 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 154184 +BPFP 0.6716 bits/point +EBPFP 1.3433 equivalent bits/point +MSE 116.054010 +---------------------- -------------------------------------------------------- +Time: 0.719s Load: 0.007s, Pack+Encode: 0.303s, Decode+Unpack: 0.409s +---------------------- -------------------------------------------------------- +💾 Converting with 116.0540 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,500B, BPFP=0.3057 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,568B, BPFP=1.4652 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,976B, BPFP=0.5418 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,076B, BPFP=1.4318 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,108B, BPFP=0.6188 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,200B, BPFP=1.3723 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,636B, BPFP=0.5867 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,372B, BPFP=1.3840 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,888B, BPFP=1.0114 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,956B, BPFP=1.2878 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,692B, BPFP=0.1329 +⌛️ [2/4] FRONTEND: Frontend time: 0.273s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.428s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14712874 150.97715693 + layer.0.v_cache 0.00001826 0.01608033 + layer.1.k_cache 0.36788522 15.99071629 + layer.1.v_cache 0.00000598 0.00622850 + layer.2.k_cache 0.01156470 1.75814673 + layer.2.v_cache 0.00002030 0.01701247 + layer.3.k_cache 0.01315050 7.23685674 + layer.3.v_cache 0.00002017 0.01801563 + layer.4.k_cache 0.00068370 0.41027898 + layer.4.v_cache 0.00004817 0.03442621 + layer.4.output 1.33105469 235.49745730 + ------------------------------------------------------------------------------------- + TOTAL 0.57987697 107.34983058 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 160972 +BPFP 0.6433 bits/point +EBPFP 1.2865 equivalent bits/point +MSE 107.349831 +---------------------- -------------------------------------------------------- +Time: 0.709s Load: 0.008s, Pack+Encode: 0.273s, Decode+Unpack: 0.428s +---------------------- -------------------------------------------------------- +💾 Converting with 107.3498 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 188, 128) +Output shape: (1, 188, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.output: torch.Size([1, 188, 3584]) -> torch.Size([1, 1, 188, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,520B, BPFP=0.2926 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,980B, BPFP=1.4112 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,292B, BPFP=0.5229 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,340B, BPFP=1.3580 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,936B, BPFP=0.5765 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,796B, BPFP=1.3128 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,352B, BPFP=0.5279 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,164B, BPFP=1.3434 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,876B, BPFP=0.9870 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,012B, BPFP=1.2477 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,344B, BPFP=0.1347 +⌛️ [2/4] FRONTEND: Frontend time: 0.269s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005002 148.50280502 + layer.0.v_cache 0.00001663 0.01594007 + layer.1.k_cache 0.14149688 15.83675433 + layer.1.v_cache 0.00000584 0.00568653 + layer.2.k_cache 0.00881058 1.70446274 + layer.2.v_cache 0.00002285 0.01625787 + layer.3.k_cache 0.03192479 7.05617312 + layer.3.v_cache 0.00001958 0.01768910 + layer.4.k_cache 0.00066612 0.40602246 + layer.4.v_cache 0.00005162 0.03617166 + layer.4.output 0.00857985 291.94082447 + ------------------------------------------------------------------------------------- + TOTAL 0.02253670 130.42257260 + (elements=1,636,352) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1636352 +Total Bytes 126612 +BPFP 0.6190 bits/point +EBPFP 1.2380 equivalent bits/point +MSE 130.422573 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.008s, Pack+Encode: 0.269s, Decode+Unpack: 0.355s +---------------------- -------------------------------------------------------- +💾 Converting with 130.4226 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,420B, BPFP=0.3470 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,504B, BPFP=1.6099 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,572B, BPFP=0.5945 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,440B, BPFP=1.5264 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,448B, BPFP=0.7418 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,616B, BPFP=1.4617 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,304B, BPFP=0.6520 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,128B, BPFP=1.5019 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,092B, BPFP=1.1065 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,388B, BPFP=1.4438 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,736B, BPFP=0.1429 +⌛️ [2/4] FRONTEND: Frontend time: 0.343s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.436s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11435634 152.47915358 + layer.0.v_cache 0.00001764 0.01755660 + layer.1.k_cache 0.25025516 15.75155931 + layer.1.v_cache 0.00000690 0.00622166 + layer.2.k_cache 0.01528106 1.75226674 + layer.2.v_cache 0.00002105 0.01713896 + layer.3.k_cache 0.02533426 7.42581192 + layer.3.v_cache 0.00002127 0.02006201 + layer.4.k_cache 0.00070809 0.42344286 + layer.4.v_cache 0.00005072 0.03668381 + layer.4.output 1.53833902 272.17217785 + ------------------------------------------------------------------------------------- + TOTAL 0.65731915 122.53736132 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 152648 +BPFP 0.7050 bits/point +EBPFP 1.4101 equivalent bits/point +MSE 122.537361 +---------------------- -------------------------------------------------------- +Time: 0.787s Load: 0.008s, Pack+Encode: 0.343s, Decode+Unpack: 0.436s +---------------------- -------------------------------------------------------- +💾 Converting with 122.5374 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,504B, BPFP=0.3042 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,988B, BPFP=1.4747 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,104B, BPFP=0.5299 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,276B, BPFP=1.4128 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,832B, BPFP=0.5931 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,540B, BPFP=1.3490 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,632B, BPFP=0.5757 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,112B, BPFP=1.3986 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,564B, BPFP=1.0038 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,852B, BPFP=1.2892 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,372B, BPFP=0.1410 +⌛️ [2/4] FRONTEND: Frontend time: 0.314s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.408s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258550 157.89083116 + layer.0.v_cache 0.00001767 0.01777822 + layer.1.k_cache 0.14790291 15.91998020 + layer.1.v_cache 0.00000629 0.00621632 + layer.2.k_cache 0.00999225 1.82000003 + layer.2.v_cache 0.00002110 0.01764152 + layer.3.k_cache 0.02117847 7.12761909 + layer.3.v_cache 0.00002045 0.02052387 + layer.4.k_cache 0.00067024 0.43298789 + layer.4.v_cache 0.00006105 0.03880930 + layer.4.output 0.00897616 304.97363591 + ------------------------------------------------------------------------------------- + TOTAL 0.02207583 136.35928465 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 125776 +BPFP 0.6422 bits/point +EBPFP 1.2845 equivalent bits/point +MSE 136.359285 +---------------------- -------------------------------------------------------- +Time: 0.730s Load: 0.008s, Pack+Encode: 0.314s, Decode+Unpack: 0.408s +---------------------- -------------------------------------------------------- +💾 Converting with 136.3593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,404B, BPFP=0.3373 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,548B, BPFP=1.5738 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,596B, BPFP=0.5818 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,568B, BPFP=1.4988 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,448B, BPFP=0.7237 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,532B, BPFP=1.4194 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,644B, BPFP=0.6621 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,132B, BPFP=1.4654 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,176B, BPFP=1.0858 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,620B, BPFP=1.4262 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,056B, BPFP=0.1429 +⌛️ [2/4] FRONTEND: Frontend time: 0.350s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.426s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09364704 159.53285846 + layer.0.v_cache 0.00001853 0.01723597 + layer.1.k_cache 0.25163280 16.08871041 + layer.1.v_cache 0.00000592 0.00608699 + layer.2.k_cache 0.01595125 1.91052336 + layer.2.v_cache 0.00002050 0.01645829 + layer.3.k_cache 0.04325477 7.83053409 + layer.3.v_cache 0.00002016 0.01901972 + layer.4.k_cache 0.00069210 0.41655099 + layer.4.v_cache 0.00005026 0.03584176 + layer.4.output 1.50065788 265.41808911 + ------------------------------------------------------------------------------------- + TOTAL 0.64175873 120.22355552 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 153724 +BPFP 0.6926 bits/point +EBPFP 1.3852 equivalent bits/point +MSE 120.223556 +---------------------- -------------------------------------------------------- +Time: 0.785s Load: 0.008s, Pack+Encode: 0.350s, Decode+Unpack: 0.426s +---------------------- -------------------------------------------------------- +💾 Converting with 120.2236 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,300B, BPFP=0.2973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,808B, BPFP=1.4386 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,572B, BPFP=0.5235 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,400B, BPFP=1.4104 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,504B, BPFP=0.5879 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,436B, BPFP=1.3438 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,164B, BPFP=0.5644 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,688B, BPFP=1.3612 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,248B, BPFP=0.9851 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,072B, BPFP=1.2494 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,844B, BPFP=0.1466 +⌛️ [2/4] FRONTEND: Frontend time: 0.321s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.395s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510114 156.76590155 + layer.0.v_cache 0.00001717 0.01634064 + layer.1.k_cache 0.31797088 15.82524933 + layer.1.v_cache 0.00000673 0.00599985 + layer.2.k_cache 0.01711024 1.66726550 + layer.2.v_cache 0.00002022 0.01616760 + layer.3.k_cache 0.01307247 7.42905338 + layer.3.v_cache 0.00002089 0.01840314 + layer.4.k_cache 0.00069413 0.40980776 + layer.4.v_cache 0.00005461 0.03567232 + layer.4.output 1.35466395 239.69467051 + ------------------------------------------------------------------------------------- + TOTAL 0.58392448 109.41485616 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 156036 +BPFP 0.6346 bits/point +EBPFP 1.2692 equivalent bits/point +MSE 109.414856 +---------------------- -------------------------------------------------------- +Time: 0.724s Load: 0.008s, Pack+Encode: 0.321s, Decode+Unpack: 0.395s +---------------------- -------------------------------------------------------- +💾 Converting with 109.4149 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,572B, BPFP=0.2881 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,128B, BPFP=1.3942 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,956B, BPFP=0.5013 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,500B, BPFP=1.3546 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,824B, BPFP=0.5559 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,416B, BPFP=1.2863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,468B, BPFP=0.5335 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,884B, BPFP=1.3158 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,980B, BPFP=0.9438 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,760B, BPFP=1.2450 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,372B, BPFP=0.1384 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14860459 151.07056452 + layer.0.v_cache 0.00001629 0.01589984 + layer.1.k_cache 0.38266554 16.06408790 + layer.1.v_cache 0.00000669 0.00581460 + layer.2.k_cache 0.01628031 1.72418496 + layer.2.v_cache 0.00002106 0.01606914 + layer.3.k_cache 0.01842327 7.56935760 + layer.3.v_cache 0.00002289 0.01850012 + layer.4.k_cache 0.00072690 0.40649876 + layer.4.v_cache 0.00004923 0.03460619 + layer.4.output 1.23454798 218.10170651 + ------------------------------------------------------------------------------------- + TOTAL 0.54168545 100.21397230 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 164860 +BPFP 0.6110 bits/point +EBPFP 1.2220 equivalent bits/point +MSE 100.213972 +---------------------- -------------------------------------------------------- +Time: 0.562s Load: 0.008s, Pack+Encode: 0.230s, Decode+Unpack: 0.324s +---------------------- -------------------------------------------------------- +💾 Converting with 100.2140 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 435, 128) +Output shape: (1, 435, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.output: torch.Size([1, 435, 3584]) -> torch.Size([1, 1, 435, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,720B, BPFP=0.2773 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,640B, BPFP=1.3879 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,180B, BPFP=0.4734 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,924B, BPFP=1.3263 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,956B, BPFP=0.5372 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,968B, BPFP=1.2560 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,380B, BPFP=0.5165 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,572B, BPFP=1.2777 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,844B, BPFP=0.8924 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,464B, BPFP=1.2020 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,640B, BPFP=0.1264 +⌛️ [2/4] FRONTEND: Frontend time: 0.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.496s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15864419 149.63073815 + layer.0.v_cache 0.00001607 0.01600718 + layer.1.k_cache 0.94329666 15.92588115 + layer.1.v_cache 0.00000639 0.00584763 + layer.2.k_cache 0.02505282 1.69844339 + layer.2.v_cache 0.00002131 0.01614051 + layer.3.k_cache 0.01724711 6.84554373 + layer.3.v_cache 0.00002032 0.01741275 + layer.4.k_cache 0.00074513 0.40496672 + layer.4.v_cache 0.00005173 0.03472625 + layer.4.output 0.00608059 127.37847906 + ------------------------------------------------------------------------------------- + TOTAL 0.06986270 62.72029770 + (elements=3,786,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3786240 +Total Bytes 279288 +BPFP 0.5901 bits/point +EBPFP 1.1802 equivalent bits/point +MSE 62.720298 +---------------------- -------------------------------------------------------- +Time: 0.844s Load: 0.014s, Pack+Encode: 0.334s, Decode+Unpack: 0.496s +---------------------- -------------------------------------------------------- +💾 Converting with 62.7203 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 437, 128) +Output shape: (1, 437, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.output: torch.Size([1, 437, 3584]) -> torch.Size([1, 1, 437, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,652B, BPFP=0.2736 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,456B, BPFP=1.3750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,260B, BPFP=0.4741 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,828B, BPFP=1.3168 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,080B, BPFP=0.5392 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,088B, BPFP=1.2546 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,676B, BPFP=0.5247 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,832B, BPFP=1.2812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,180B, BPFP=0.9003 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,676B, BPFP=1.2041 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,672B, BPFP=0.1413 +⌛️ [2/4] FRONTEND: Frontend time: 0.325s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.496s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14454122 150.32240418 + layer.0.v_cache 0.00001702 0.01601107 + layer.1.k_cache 0.94267751 15.74315065 + layer.1.v_cache 0.00000668 0.00605329 + layer.2.k_cache 0.01808156 1.70691965 + layer.2.v_cache 0.00002154 0.01628941 + layer.3.k_cache 0.03034366 6.89752630 + layer.3.v_cache 0.00002210 0.01843304 + layer.4.k_cache 0.00074160 0.41830703 + layer.4.v_cache 0.00005552 0.03595402 + layer.4.output 0.00611241 126.77001267 + ------------------------------------------------------------------------------------- + TOTAL 0.06937031 62.50418455 + (elements=3,803,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3803648 +Total Bytes 283400 +BPFP 0.5961 bits/point +EBPFP 1.1921 equivalent bits/point +MSE 62.504185 +---------------------- -------------------------------------------------------- +Time: 0.835s Load: 0.015s, Pack+Encode: 0.325s, Decode+Unpack: 0.496s +---------------------- -------------------------------------------------------- +💾 Converting with 62.5042 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,600B, BPFP=0.2841 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,636B, BPFP=1.4020 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,688B, BPFP=0.4915 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,328B, BPFP=1.3356 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,816B, BPFP=0.5487 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,184B, BPFP=1.2776 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,752B, BPFP=0.5455 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,792B, BPFP=1.3084 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,052B, BPFP=0.9158 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,948B, BPFP=1.2149 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,720B, BPFP=0.1429 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12827574 151.42897727 + layer.0.v_cache 0.00001866 0.01690124 + layer.1.k_cache 0.53653341 15.94252232 + layer.1.v_cache 0.00000626 0.00573247 + layer.2.k_cache 0.01950284 1.69031277 + layer.2.v_cache 0.00002083 0.01616274 + layer.3.k_cache 0.04340698 7.26403650 + layer.3.v_cache 0.00002181 0.01836660 + layer.4.k_cache 0.00072091 0.40772862 + layer.4.v_cache 0.00005393 0.03647215 + layer.4.output 0.04342448 176.04704893 + ------------------------------------------------------------------------------------- + TOTAL 0.06073722 82.89156207 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 203516 +BPFP 0.6073 bits/point +EBPFP 1.2146 equivalent bits/point +MSE 82.891562 +---------------------- -------------------------------------------------------- +Time: 0.642s Load: 0.012s, Pack+Encode: 0.259s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 82.8916 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,500B, BPFP=0.3231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,324B, BPFP=1.5463 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,312B, BPFP=0.5470 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,120B, BPFP=1.4756 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,368B, BPFP=0.6678 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,976B, BPFP=1.4084 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,320B, BPFP=0.6062 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,456B, BPFP=1.4366 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,680B, BPFP=1.0385 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,432B, BPFP=1.3764 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,584B, BPFP=0.1476 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702504 157.91382754 + layer.0.v_cache 0.00001641 0.01720400 + layer.1.k_cache 0.51130361 15.87344337 + layer.1.v_cache 0.00000656 0.00640468 + layer.2.k_cache 0.01810413 1.69114467 + layer.2.v_cache 0.00002122 0.01726682 + layer.3.k_cache 0.02476560 7.92202828 + layer.3.v_cache 0.00002197 0.01874921 + layer.4.k_cache 0.00071356 0.42581248 + layer.4.v_cache 0.00005290 0.03782795 + layer.4.output 0.00501064 208.70233620 + ------------------------------------------------------------------------------------- + TOTAL 0.04100620 96.75529779 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 195072 +BPFP 0.6740 bits/point +EBPFP 1.3481 equivalent bits/point +MSE 96.755298 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 96.7553 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,480B, BPFP=0.2991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,484B, BPFP=1.5013 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,972B, BPFP=0.5323 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,220B, BPFP=1.4169 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,328B, BPFP=0.6229 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,696B, BPFP=1.3819 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,732B, BPFP=0.5831 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,044B, BPFP=1.4052 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,092B, BPFP=1.0077 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,644B, BPFP=1.3117 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,684B, BPFP=0.1496 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11892304 158.45270767 + layer.0.v_cache 0.00001803 0.01776665 + layer.1.k_cache 0.31174104 15.90087473 + layer.1.v_cache 0.00000657 0.00622261 + layer.2.k_cache 0.01452858 1.72590298 + layer.2.v_cache 0.00002161 0.01726241 + layer.3.k_cache 0.02263702 7.31724822 + layer.3.v_cache 0.00002141 0.01923759 + layer.4.k_cache 0.00068312 0.40926374 + layer.4.v_cache 0.00005251 0.03737033 + layer.4.output 1.30838121 231.48939255 + ------------------------------------------------------------------------------------- + TOTAL 0.56631185 106.13703558 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 166376 +BPFP 0.6535 bits/point +EBPFP 1.3070 equivalent bits/point +MSE 106.137036 +---------------------- -------------------------------------------------------- +Time: 0.546s Load: 0.009s, Pack+Encode: 0.225s, Decode+Unpack: 0.312s +---------------------- -------------------------------------------------------- +💾 Converting with 106.1370 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 233, 128) +Output shape: (1, 233, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.output: torch.Size([1, 233, 3584]) -> torch.Size([1, 1, 233, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,476B, BPFP=0.3002 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,972B, BPFP=1.4734 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,936B, BPFP=0.5322 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,256B, BPFP=1.4254 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,000B, BPFP=0.6035 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,244B, BPFP=1.3576 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,848B, BPFP=0.5933 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,620B, BPFP=1.3828 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,876B, BPFP=0.9976 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,324B, BPFP=1.2959 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,188B, BPFP=0.1455 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13365312 154.97689780 + layer.0.v_cache 0.00001735 0.01719261 + layer.1.k_cache 0.40872042 15.83518831 + layer.1.v_cache 0.00000792 0.00681518 + layer.2.k_cache 0.01308113 1.74059718 + layer.2.v_cache 0.00002137 0.01738225 + layer.3.k_cache 0.00601081 7.38694613 + layer.3.v_cache 0.00002094 0.01871941 + layer.4.k_cache 0.00068546 0.42455030 + layer.4.v_cache 0.00005338 0.03745214 + layer.4.output 1.31399239 232.45623850 + ------------------------------------------------------------------------------------- + TOTAL 0.57413051 106.33267123 + (elements=2,028,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2028032 +Total Bytes 163740 +BPFP 0.6459 bits/point +EBPFP 1.2918 equivalent bits/point +MSE 106.332671 +---------------------- -------------------------------------------------------- +Time: 0.541s Load: 0.008s, Pack+Encode: 0.222s, Decode+Unpack: 0.311s +---------------------- -------------------------------------------------------- +💾 Converting with 106.3327 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 148, 128) +Output shape: (1, 148, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.output: torch.Size([1, 148, 3584]) -> torch.Size([1, 1, 148, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,328B, BPFP=0.3514 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,340B, BPFP=1.7251 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,004B, BPFP=0.6339 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,524B, BPFP=1.6389 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,324B, BPFP=0.7732 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,832B, BPFP=1.5659 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,964B, BPFP=0.7352 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,064B, BPFP=1.5904 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,328B, BPFP=1.1959 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,636B, BPFP=1.5452 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,804B, BPFP=0.1479 +⌛️ [2/4] FRONTEND: Frontend time: 0.193s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.255s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08608804 163.99396907 + layer.0.v_cache 0.00001767 0.01748247 + layer.1.k_cache 0.12276106 16.27356947 + layer.1.v_cache 0.00000594 0.00621698 + layer.2.k_cache 0.00886028 1.84141417 + layer.2.v_cache 0.00002088 0.01748909 + layer.3.k_cache 0.01513371 7.31189542 + layer.3.v_cache 0.00002085 0.01990670 + layer.4.k_cache 0.00068430 0.43733824 + layer.4.v_cache 0.00005331 0.03901517 + layer.4.output 0.08957551 366.17425796 + ------------------------------------------------------------------------------------- + TOTAL 0.05062792 161.95165309 + (elements=1,288,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1288192 +Total Bytes 121148 +BPFP 0.7524 bits/point +EBPFP 1.5047 equivalent bits/point +MSE 161.951653 +---------------------- -------------------------------------------------------- +Time: 0.454s Load: 0.006s, Pack+Encode: 0.193s, Decode+Unpack: 0.255s +---------------------- -------------------------------------------------------- +💾 Converting with 161.9517 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 170, 128) +Output shape: (1, 170, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.output: torch.Size([1, 170, 3584]) -> torch.Size([1, 1, 170, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,404B, BPFP=0.3129 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,888B, BPFP=1.5522 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,168B, BPFP=0.5669 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,064B, BPFP=1.4765 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,172B, BPFP=0.6592 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,316B, BPFP=1.4077 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,412B, BPFP=0.6813 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,976B, BPFP=1.4684 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,364B, BPFP=1.0445 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,600B, BPFP=1.3419 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,792B, BPFP=0.1548 +⌛️ [2/4] FRONTEND: Frontend time: 0.190s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.256s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13519069 156.86869256 + layer.0.v_cache 0.00001779 0.01874293 + layer.1.k_cache 0.14803200 15.73036822 + layer.1.v_cache 0.00000678 0.00683920 + layer.2.k_cache 0.02267574 1.75568058 + layer.2.v_cache 0.00002005 0.01799469 + layer.3.k_cache 0.04618364 6.74301327 + layer.3.v_cache 0.00002282 0.02098113 + layer.4.k_cache 0.00077641 0.42738401 + layer.4.v_cache 0.00005264 0.03815016 + layer.4.output 0.00945594 323.03765756 + ------------------------------------------------------------------------------------- + TOTAL 0.02465707 143.69949704 + (elements=1,479,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1479680 +Total Bytes 126156 +BPFP 0.6821 bits/point +EBPFP 1.3641 equivalent bits/point +MSE 143.699497 +---------------------- -------------------------------------------------------- +Time: 0.453s Load: 0.006s, Pack+Encode: 0.190s, Decode+Unpack: 0.256s +---------------------- -------------------------------------------------------- +💾 Converting with 143.6995 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,372B, BPFP=0.3174 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,012B, BPFP=1.6013 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,284B, BPFP=0.5915 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,620B, BPFP=1.5644 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,760B, BPFP=0.7304 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,776B, BPFP=1.4849 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,940B, BPFP=0.6532 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,468B, BPFP=1.5501 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,784B, BPFP=1.1092 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,964B, BPFP=1.4085 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,148B, BPFP=0.1499 +⌛️ [2/4] FRONTEND: Frontend time: 0.190s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.255s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12520085 153.80611116 + layer.0.v_cache 0.00001716 0.01858394 + layer.1.k_cache 0.15152623 15.47885536 + layer.1.v_cache 0.00000633 0.00644604 + layer.2.k_cache 0.01605446 1.82807003 + layer.2.v_cache 0.00002200 0.01860475 + layer.3.k_cache 0.04332644 7.27647087 + layer.3.v_cache 0.00002349 0.02128477 + layer.4.k_cache 0.00069336 0.43786299 + layer.4.v_cache 0.00005614 0.03791004 + layer.4.output 0.00966471 330.83675775 + ------------------------------------------------------------------------------------- + TOTAL 0.02379879 146.75220613 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 128128 +BPFP 0.7094 bits/point +EBPFP 1.4189 equivalent bits/point +MSE 146.752206 +---------------------- -------------------------------------------------------- +Time: 0.451s Load: 0.006s, Pack+Encode: 0.190s, Decode+Unpack: 0.255s +---------------------- -------------------------------------------------------- +💾 Converting with 146.7522 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 181, 128) +Output shape: (1, 181, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.output: torch.Size([1, 181, 3584]) -> torch.Size([1, 1, 181, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,432B, BPFP=0.2963 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,776B, BPFP=1.4482 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,072B, BPFP=0.5242 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,060B, BPFP=1.3864 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,804B, BPFP=0.5874 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,284B, BPFP=1.3194 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,508B, BPFP=0.5618 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,844B, BPFP=1.3677 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,312B, BPFP=0.9765 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,604B, BPFP=1.2607 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,312B, BPFP=0.1395 +⌛️ [2/4] FRONTEND: Frontend time: 0.194s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13566811 151.64352771 + layer.0.v_cache 0.00002016 0.01795564 + layer.1.k_cache 0.14782082 15.63599307 + layer.1.v_cache 0.00000634 0.00699677 + layer.2.k_cache 0.01546242 1.69175594 + layer.2.v_cache 0.00002121 0.01775345 + layer.3.k_cache 0.02895848 7.84279030 + layer.3.v_cache 0.00002197 0.02087223 + layer.4.k_cache 0.00067433 0.43171270 + layer.4.v_cache 0.00005229 0.03826411 + layer.4.output 0.00892717 303.15097178 + ------------------------------------------------------------------------------------- + TOTAL 0.02301155 135.25908379 + (elements=1,575,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1575424 +Total Bytes 124008 +BPFP 0.6297 bits/point +EBPFP 1.2594 equivalent bits/point +MSE 135.259084 +---------------------- -------------------------------------------------------- +Time: 0.473s Load: 0.006s, Pack+Encode: 0.194s, Decode+Unpack: 0.272s +---------------------- -------------------------------------------------------- +💾 Converting with 135.2591 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 186, 128) +Output shape: (1, 186, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.output: torch.Size([1, 186, 3584]) -> torch.Size([1, 1, 186, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,532B, BPFP=0.2967 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,992B, BPFP=1.4274 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,244B, BPFP=0.5245 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,180B, BPFP=1.3592 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,752B, BPFP=0.5672 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,580B, BPFP=1.3088 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,304B, BPFP=0.5296 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,296B, BPFP=1.3690 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,688B, BPFP=0.9819 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,936B, BPFP=1.2547 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,632B, BPFP=0.1396 +⌛️ [2/4] FRONTEND: Frontend time: 0.190s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.256s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09795722 158.21729041 + layer.0.v_cache 0.00001815 0.01761484 + layer.1.k_cache 0.13430934 15.37726027 + layer.1.v_cache 0.00000632 0.00620179 + layer.2.k_cache 0.01830538 1.74408714 + layer.2.v_cache 0.00002120 0.01724937 + layer.3.k_cache 0.04892764 6.76396392 + layer.3.v_cache 0.00002113 0.01939436 + layer.4.k_cache 0.00067235 0.41752645 + layer.4.v_cache 0.00008605 0.03680107 + layer.4.output 0.00869856 295.16553859 + ------------------------------------------------------------------------------------- + TOTAL 0.02124792 132.28095058 + (elements=1,618,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1618944 +Total Bytes 126136 +BPFP 0.6233 bits/point +EBPFP 1.2466 equivalent bits/point +MSE 132.280951 +---------------------- -------------------------------------------------------- +Time: 0.453s Load: 0.007s, Pack+Encode: 0.190s, Decode+Unpack: 0.256s +---------------------- -------------------------------------------------------- +💾 Converting with 132.2810 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 163, 128) +Output shape: (1, 163, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.output: torch.Size([1, 163, 3584]) -> torch.Size([1, 1, 163, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,392B, BPFP=0.3252 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,164B, BPFP=1.5495 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,028B, BPFP=0.5778 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,736B, BPFP=1.5084 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,164B, BPFP=0.6867 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,372B, BPFP=1.4735 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,072B, BPFP=0.6779 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,652B, BPFP=1.5004 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,612B, BPFP=1.1131 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,060B, BPFP=1.3478 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,952B, BPFP=0.1500 +⌛️ [2/4] FRONTEND: Frontend time: 0.191s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.257s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11530171 157.03367044 + layer.0.v_cache 0.00001813 0.01803011 + layer.1.k_cache 0.14319220 15.70541663 + layer.1.v_cache 0.00000612 0.00636507 + layer.2.k_cache 0.02368733 1.73670445 + layer.2.v_cache 0.00002106 0.01737980 + layer.3.k_cache 0.02525310 6.76706139 + layer.3.v_cache 0.00002161 0.02038988 + layer.4.k_cache 0.00069984 0.42515016 + layer.4.v_cache 0.00005709 0.03742231 + layer.4.output 0.00982084 336.81636174 + ------------------------------------------------------------------------------------- + TOTAL 0.02217671 149.38130132 + (elements=1,418,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1418752 +Total Bytes 123204 +BPFP 0.6947 bits/point +EBPFP 1.3894 equivalent bits/point +MSE 149.381301 +---------------------- -------------------------------------------------------- +Time: 0.454s Load: 0.006s, Pack+Encode: 0.191s, Decode+Unpack: 0.257s +---------------------- -------------------------------------------------------- +💾 Converting with 149.3813 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 191, 128) +Output shape: (1, 191, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.output: torch.Size([1, 191, 3584]) -> torch.Size([1, 1, 191, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,200B, BPFP=0.2618 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,872B, BPFP=1.3802 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,732B, BPFP=0.4689 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,132B, BPFP=1.3197 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,644B, BPFP=0.5435 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,636B, BPFP=1.2791 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,920B, BPFP=0.4843 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,212B, BPFP=1.3262 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,836B, BPFP=0.9683 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,172B, BPFP=1.2412 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,736B, BPFP=0.1255 +⌛️ [2/4] FRONTEND: Frontend time: 0.191s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.256s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15112510 145.47905759 + layer.0.v_cache 0.00002069 0.01821922 + layer.1.k_cache 0.18491925 14.84971419 + layer.1.v_cache 0.00000629 0.00671309 + layer.2.k_cache 0.02122488 1.68551843 + layer.2.v_cache 0.00002327 0.01769068 + layer.3.k_cache 0.01259250 6.79370692 + layer.3.v_cache 0.00002182 0.01995604 + layer.4.k_cache 0.00068006 0.43749149 + layer.4.v_cache 0.00005885 0.03873974 + layer.4.output 0.00845948 287.37927730 + ------------------------------------------------------------------------------------- + TOTAL 0.02528759 128.29422050 + (elements=1,662,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1662464 +Total Bytes 124092 +BPFP 0.5971 bits/point +EBPFP 1.1943 equivalent bits/point +MSE 128.294221 +---------------------- -------------------------------------------------------- +Time: 0.454s Load: 0.007s, Pack+Encode: 0.191s, Decode+Unpack: 0.256s +---------------------- -------------------------------------------------------- +💾 Converting with 128.2942 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 174, 128) +Output shape: (1, 174, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.output: torch.Size([1, 174, 3584]) -> torch.Size([1, 1, 174, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,404B, BPFP=0.3057 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,956B, BPFP=1.5226 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,008B, BPFP=0.5395 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,184B, BPFP=1.4533 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,336B, BPFP=0.6588 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,564B, BPFP=1.3976 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,844B, BPFP=0.6146 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,152B, BPFP=1.4504 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,728B, BPFP=1.0532 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,796B, BPFP=1.3287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,792B, BPFP=0.1384 +⌛️ [2/4] FRONTEND: Frontend time: 0.191s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.256s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14146894 159.15389278 + layer.0.v_cache 0.00001664 0.01724234 + layer.1.k_cache 0.16292955 16.32062680 + layer.1.v_cache 0.00000602 0.00608156 + layer.2.k_cache 0.01834387 1.81680140 + layer.2.v_cache 0.00002001 0.01749020 + layer.3.k_cache 0.03015787 7.25519220 + layer.3.v_cache 0.00002180 0.01967245 + layer.4.k_cache 0.00067479 0.43771020 + layer.4.v_cache 0.00005130 0.03829077 + layer.4.output 0.00923280 315.52047414 + ------------------------------------------------------------------------------------- + TOTAL 0.02460708 140.80743057 + (elements=1,514,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1514496 +Total Bytes 125764 +BPFP 0.6643 bits/point +EBPFP 1.3286 equivalent bits/point +MSE 140.807431 +---------------------- -------------------------------------------------------- +Time: 0.452s Load: 0.006s, Pack+Encode: 0.191s, Decode+Unpack: 0.256s +---------------------- -------------------------------------------------------- +💾 Converting with 140.8074 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.6523 bits/point +Avg EBPFP 1.3046 equivalent bits/point +Avg MSE 111.148866 +Avg Time 0.654s +------------------------ ---------------------------- diff --git a/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..0657d79be6db08a59cff4580a87b6f90aa2230f2 --- /dev/null +++ b/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 559 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa +Output output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,332B, BPFP=0.2893 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,104B, BPFP=1.3620 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,016B, BPFP=0.4891 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,048B, BPFP=1.3047 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,200B, BPFP=0.5534 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,096B, BPFP=1.2530 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,356B, BPFP=0.5076 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,492B, BPFP=1.2745 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,864B, BPFP=0.9149 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,040B, BPFP=1.1957 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,992B, BPFP=0.1239 +⌛️ [2/4] FRONTEND: Frontend time: 0.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.471s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14134069 155.81214735 + layer.0.v_cache 0.00001405 0.01305278 + layer.1.k_cache 0.62039534 15.95241970 + layer.1.v_cache 0.00000582 0.00481966 + layer.2.k_cache 0.00676800 1.71868568 + layer.2.v_cache 0.00001893 0.01380193 + layer.3.k_cache 0.03896382 7.09352790 + layer.3.v_cache 0.00001942 0.01572041 + layer.4.k_cache 0.00069280 0.37273762 + layer.4.v_cache 0.00005238 0.03405040 + layer.4.output 0.00804578 192.77198041 + ------------------------------------------------------------------------------------- + TOTAL 0.05085834 90.02557802 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 184540 +BPFP 0.5889 bits/point +EBPFP 1.1779 equivalent bits/point +MSE 90.025578 +---------------------- -------------------------------------------------------- +Time: 0.985s Load: 0.012s, Pack+Encode: 0.503s, Decode+Unpack: 0.471s +---------------------- -------------------------------------------------------- +💾 Converting with 90.0256 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,676B, BPFP=0.3027 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,984B, BPFP=1.3857 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,572B, BPFP=0.5105 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,972B, BPFP=1.3317 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,928B, BPFP=0.5828 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,056B, BPFP=1.2828 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,440B, BPFP=0.5567 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,324B, BPFP=1.2971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,744B, BPFP=0.9462 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,748B, BPFP=1.2131 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,536B, BPFP=0.1336 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15682776 153.87413343 + layer.0.v_cache 0.00001501 0.01337550 + layer.1.k_cache 0.54447479 16.02136105 + layer.1.v_cache 0.00000580 0.00522876 + layer.2.k_cache 0.00853899 1.70808822 + layer.2.v_cache 0.00001972 0.01424479 + layer.3.k_cache 0.02401569 7.38642878 + layer.3.v_cache 0.00001939 0.01624151 + layer.4.k_cache 0.00071059 0.37201415 + layer.4.v_cache 0.00005329 0.03504094 + layer.4.output 0.05121508 185.23648525 + ------------------------------------------------------------------------------------- + TOTAL 0.06430510 86.82950317 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 193980 +BPFP 0.6085 bits/point +EBPFP 1.2170 equivalent bits/point +MSE 86.829503 +---------------------- -------------------------------------------------------- +Time: 0.663s Load: 0.011s, Pack+Encode: 0.262s, Decode+Unpack: 0.390s +---------------------- -------------------------------------------------------- +💾 Converting with 86.8295 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,604B, BPFP=0.2978 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,208B, BPFP=1.3929 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,504B, BPFP=0.5051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,212B, BPFP=1.3399 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,448B, BPFP=0.5553 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,996B, BPFP=1.2753 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,076B, BPFP=0.5355 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,588B, BPFP=1.3068 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,404B, BPFP=0.9250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,796B, BPFP=1.2115 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,112B, BPFP=0.1375 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15363285 150.50279018 + layer.0.v_cache 0.00001412 0.01303057 + layer.1.k_cache 0.60852461 16.01613819 + layer.1.v_cache 0.00000580 0.00477700 + layer.2.k_cache 0.01182756 1.74698001 + layer.2.v_cache 0.00001863 0.01330544 + layer.3.k_cache 0.03467803 7.74376528 + layer.3.v_cache 0.00001919 0.01541205 + layer.4.k_cache 0.00072319 0.35445098 + layer.4.v_cache 0.00005301 0.03280888 + layer.4.output 0.05035121 184.63420190 + ------------------------------------------------------------------------------------- + TOTAL 0.06835032 86.40487481 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 193948 +BPFP 0.6063 bits/point +EBPFP 1.2127 equivalent bits/point +MSE 86.404875 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.011s, Pack+Encode: 0.254s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 86.4049 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,528B, BPFP=0.3052 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,660B, BPFP=1.4167 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,480B, BPFP=0.5234 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,084B, BPFP=1.3849 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,504B, BPFP=0.5799 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,264B, BPFP=1.3397 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,632B, BPFP=0.5318 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,556B, BPFP=1.3558 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,612B, BPFP=0.9724 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,920B, BPFP=1.2655 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,288B, BPFP=0.1285 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12197856 156.22529262 + layer.0.v_cache 0.00001365 0.01320320 + layer.1.k_cache 0.62524759 15.78932648 + layer.1.v_cache 0.00000551 0.00469830 + layer.2.k_cache 0.01486346 1.70401874 + layer.2.v_cache 0.00001981 0.01395896 + layer.3.k_cache 0.07387851 7.60002864 + layer.3.v_cache 0.00001943 0.01525438 + layer.4.k_cache 0.00069560 0.36429033 + layer.4.v_cache 0.00005405 0.03578256 + layer.4.output 0.01098318 196.20709238 + ------------------------------------------------------------------------------------- + TOTAL 0.05374461 91.48326476 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 191528 +BPFP 0.6220 bits/point +EBPFP 1.2441 equivalent bits/point +MSE 91.483265 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4833 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,504B, BPFP=0.3018 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,108B, BPFP=1.3765 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,396B, BPFP=0.5151 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,576B, BPFP=1.3474 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,468B, BPFP=0.5739 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,468B, BPFP=1.2866 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,828B, BPFP=0.5388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,892B, BPFP=1.3099 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,408B, BPFP=0.9544 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,076B, BPFP=1.2103 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,124B, BPFP=0.1341 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15362121 155.22309485 + layer.0.v_cache 0.00001420 0.01304150 + layer.1.k_cache 0.52952420 16.04210184 + layer.1.v_cache 0.00000567 0.00481344 + layer.2.k_cache 0.00939641 1.66422783 + layer.2.v_cache 0.00001891 0.01348925 + layer.3.k_cache 0.04339437 8.06343287 + layer.3.v_cache 0.00002732 0.01560270 + layer.4.k_cache 0.00068326 0.36340581 + layer.4.v_cache 0.00005156 0.03413027 + layer.4.output 0.00776138 194.65211466 + ------------------------------------------------------------------------------------- + TOTAL 0.04653334 90.82365547 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 188848 +BPFP 0.6090 bits/point +EBPFP 1.2181 equivalent bits/point +MSE 90.823655 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 90.8237 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,632B, BPFP=0.2993 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,988B, BPFP=1.3812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,572B, BPFP=0.5087 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,200B, BPFP=1.3393 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,800B, BPFP=0.5740 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,124B, BPFP=1.2821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,776B, BPFP=0.5196 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,660B, BPFP=1.3106 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,908B, BPFP=0.9517 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,968B, BPFP=1.2207 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,960B, BPFP=0.1288 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14921317 150.94459503 + layer.0.v_cache 0.00001395 0.01343503 + layer.1.k_cache 0.64083395 15.85821076 + layer.1.v_cache 0.00000574 0.00503423 + layer.2.k_cache 0.00578804 1.70969791 + layer.2.v_cache 0.00001912 0.01408391 + layer.3.k_cache 0.02848778 7.44325308 + layer.3.v_cache 0.00001956 0.01559736 + layer.4.k_cache 0.00070939 0.35876758 + layer.4.v_cache 0.00005206 0.03307978 + layer.4.output 0.04916091 184.58910350 + ------------------------------------------------------------------------------------- + TOTAL 0.06878054 86.38349877 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 193588 +BPFP 0.6052 bits/point +EBPFP 1.2104 equivalent bits/point +MSE 86.383499 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 86.3835 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,668B, BPFP=0.3023 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,952B, BPFP=1.3840 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,624B, BPFP=0.5132 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,352B, BPFP=1.3520 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,112B, BPFP=0.5926 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,348B, BPFP=1.2984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,036B, BPFP=0.5352 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,344B, BPFP=1.2982 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,928B, BPFP=0.9561 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,624B, BPFP=1.2065 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,892B, BPFP=0.1363 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12806672 153.85634866 + layer.0.v_cache 0.00001388 0.01293028 + layer.1.k_cache 0.56701790 16.06679121 + layer.1.v_cache 0.00000557 0.00473157 + layer.2.k_cache 0.01020011 1.64096548 + layer.2.v_cache 0.00001881 0.01368280 + layer.3.k_cache 0.04301871 7.41180691 + layer.3.v_cache 0.00002051 0.01564794 + layer.4.k_cache 0.00071755 0.36047952 + layer.4.v_cache 0.00005284 0.03252113 + layer.4.output 0.04940383 185.24171136 + ------------------------------------------------------------------------------------- + TOTAL 0.06440938 86.82987559 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 194880 +BPFP 0.6113 bits/point +EBPFP 1.2226 equivalent bits/point +MSE 86.829876 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.011s, Pack+Encode: 0.257s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 86.8299 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,524B, BPFP=0.3029 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,208B, BPFP=1.3820 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,420B, BPFP=0.5164 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,720B, BPFP=1.3553 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,520B, BPFP=0.5768 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,768B, BPFP=1.3031 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,920B, BPFP=0.5439 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,116B, BPFP=1.3221 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,624B, BPFP=0.9662 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,344B, BPFP=1.2250 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,828B, BPFP=0.1318 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12903612 156.23634868 + layer.0.v_cache 0.00001458 0.01304554 + layer.1.k_cache 0.59860808 16.14830387 + layer.1.v_cache 0.00000581 0.00487567 + layer.2.k_cache 0.00940504 1.66033132 + layer.2.v_cache 0.00001954 0.01409579 + layer.3.k_cache 0.02644065 7.64430082 + layer.3.v_cache 0.00002029 0.01543882 + layer.4.k_cache 0.00068693 0.36566424 + layer.4.v_cache 0.00005376 0.03401413 + layer.4.output 0.00862506 194.60748747 + ------------------------------------------------------------------------------------- + TOTAL 0.04850978 90.84640183 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 189992 +BPFP 0.6127 bits/point +EBPFP 1.2254 equivalent bits/point +MSE 90.846402 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 90.8464 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,668B, BPFP=0.2875 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,144B, BPFP=1.3263 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,544B, BPFP=0.4842 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,356B, BPFP=1.2863 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,112B, BPFP=0.5130 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,260B, BPFP=1.2307 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,676B, BPFP=0.4909 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,604B, BPFP=1.2482 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,460B, BPFP=0.8858 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,948B, BPFP=1.1642 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,928B, BPFP=0.1227 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16469599 153.10530388 + layer.0.v_cache 0.00001357 0.01272275 + layer.1.k_cache 0.70445866 16.21728357 + layer.1.v_cache 0.00000558 0.00491383 + layer.2.k_cache 0.01447187 1.62073358 + layer.2.v_cache 0.00001924 0.01364304 + layer.3.k_cache 0.02617162 7.57725941 + layer.3.v_cache 0.00001974 0.01562373 + layer.4.k_cache 0.00070375 0.36253040 + layer.4.v_cache 0.00005076 0.03380345 + layer.4.output 0.04814868 176.03658395 + ------------------------------------------------------------------------------------- + TOTAL 0.07339127 83.01293561 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 192700 +BPFP 0.5750 bits/point +EBPFP 1.1501 equivalent bits/point +MSE 83.012936 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 83.0129 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,492B, BPFP=0.3190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,580B, BPFP=1.4858 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,104B, BPFP=0.5288 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,580B, BPFP=1.4277 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,288B, BPFP=0.5976 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,104B, BPFP=1.3420 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,832B, BPFP=0.5711 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,592B, BPFP=1.3704 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,928B, BPFP=0.9833 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,652B, BPFP=1.3158 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,360B, BPFP=0.1275 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12132761 159.73893471 + layer.0.v_cache 0.00001366 0.01298459 + layer.1.k_cache 0.57241889 16.32403360 + layer.1.v_cache 0.00000556 0.00449516 + layer.2.k_cache 0.00926234 1.72913483 + layer.2.v_cache 0.00002145 0.01294582 + layer.3.k_cache 0.03109839 7.41757168 + layer.3.v_cache 0.00002257 0.01458164 + layer.4.k_cache 0.00070701 0.34894990 + layer.4.v_cache 0.00004920 0.03230823 + layer.4.output 0.01031505 206.30362122 + ------------------------------------------------------------------------------------- + TOTAL 0.04747836 95.86831110 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 186512 +BPFP 0.6373 bits/point +EBPFP 1.2745 equivalent bits/point +MSE 95.868311 +---------------------- -------------------------------------------------------- +Time: 0.649s Load: 0.010s, Pack+Encode: 0.258s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 95.8683 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,484B, BPFP=0.2875 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,128B, BPFP=1.3700 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,496B, BPFP=0.4979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,160B, BPFP=1.3192 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,752B, BPFP=0.5638 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,164B, BPFP=1.2670 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,808B, BPFP=0.5143 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,448B, BPFP=1.2819 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,452B, BPFP=0.9151 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,848B, BPFP=1.1980 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,896B, BPFP=0.1266 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14357947 155.68985948 + layer.0.v_cache 0.00001420 0.01404530 + layer.1.k_cache 0.63723908 16.11869003 + layer.1.v_cache 0.00000573 0.00503833 + layer.2.k_cache 0.00765126 1.72561871 + layer.2.v_cache 0.00002073 0.01391773 + layer.3.k_cache 0.02843528 7.54269020 + layer.3.v_cache 0.00002112 0.01576905 + layer.4.k_cache 0.00069171 0.36092797 + layer.4.v_cache 0.00005234 0.03373093 + layer.4.output 0.04871205 182.11377936 + ------------------------------------------------------------------------------------- + TOTAL 0.06815855 85.66569078 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 192636 +BPFP 0.5941 bits/point +EBPFP 1.1883 equivalent bits/point +MSE 85.665691 +---------------------- -------------------------------------------------------- +Time: 0.649s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 85.6657 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,628B, BPFP=0.2883 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,292B, BPFP=1.3469 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,476B, BPFP=0.4855 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,300B, BPFP=1.2961 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,256B, BPFP=0.5254 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,316B, BPFP=1.2457 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,532B, BPFP=0.4883 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,744B, BPFP=1.2676 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,748B, BPFP=0.9092 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,104B, BPFP=1.1836 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,060B, BPFP=0.1249 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16857396 155.02975154 + layer.0.v_cache 0.00001478 0.01347187 + layer.1.k_cache 0.58716421 16.05840644 + layer.1.v_cache 0.00000600 0.00518620 + layer.2.k_cache 0.00827454 1.76942979 + layer.2.v_cache 0.00002177 0.01436116 + layer.3.k_cache 0.02531170 7.67848040 + layer.3.v_cache 0.00001983 0.01550785 + layer.4.k_cache 0.00069463 0.36080355 + layer.4.v_cache 0.00005316 0.03505158 + layer.4.output 0.04836898 177.73498244 + ------------------------------------------------------------------------------------- + TOTAL 0.06639514 83.83090161 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 193456 +BPFP 0.5830 bits/point +EBPFP 1.1660 equivalent bits/point +MSE 83.830902 +---------------------- -------------------------------------------------------- +Time: 0.656s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.390s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8309 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,644B, BPFP=0.2959 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,220B, BPFP=1.3748 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,540B, BPFP=0.5002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,300B, BPFP=1.3266 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,628B, BPFP=0.5573 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,248B, BPFP=1.2714 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,956B, BPFP=0.5220 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,672B, BPFP=1.2936 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,744B, BPFP=0.9304 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,140B, BPFP=1.2133 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,568B, BPFP=0.1316 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14028978 156.27089451 + layer.0.v_cache 0.00001397 0.01320698 + layer.1.k_cache 0.59966084 15.84394007 + layer.1.v_cache 0.00000578 0.00493647 + layer.2.k_cache 0.01765286 1.68009017 + layer.2.v_cache 0.00001884 0.01385811 + layer.3.k_cache 0.02653811 7.23311251 + layer.3.v_cache 0.00001935 0.01559823 + layer.4.k_cache 0.00071527 0.36597192 + layer.4.v_cache 0.00005402 0.03401386 + layer.4.output 0.04907703 182.07628236 + ------------------------------------------------------------------------------------- + TOTAL 0.06638282 85.64762349 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 194660 +BPFP 0.6004 bits/point +EBPFP 1.2008 equivalent bits/point +MSE 85.647623 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.011s, Pack+Encode: 0.256s, Decode+Unpack: 0.386s +---------------------- -------------------------------------------------------- +💾 Converting with 85.6476 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,540B, BPFP=0.3182 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,100B, BPFP=1.4993 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,308B, BPFP=0.5347 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,940B, BPFP=1.4327 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,704B, BPFP=0.6149 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,856B, BPFP=1.3704 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,768B, BPFP=0.5611 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,464B, BPFP=1.4053 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,116B, BPFP=0.9832 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,992B, BPFP=1.3208 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,228B, BPFP=0.1332 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13255837 154.18248794 + layer.0.v_cache 0.00001394 0.01262548 + layer.1.k_cache 0.54921044 16.07827579 + layer.1.v_cache 0.00000578 0.00475376 + layer.2.k_cache 0.01292971 1.69298800 + layer.2.v_cache 0.00001901 0.01339815 + layer.3.k_cache 0.07190662 7.88052458 + layer.3.v_cache 0.00001944 0.01491340 + layer.4.k_cache 0.00068159 0.34246882 + layer.4.v_cache 0.00005125 0.03314469 + layer.4.output 0.00790709 204.14110097 + ------------------------------------------------------------------------------------- + TOTAL 0.04839681 94.66136985 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 191016 +BPFP 0.6455 bits/point +EBPFP 1.2909 equivalent bits/point +MSE 94.661370 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.010s, Pack+Encode: 0.256s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 94.6614 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,492B, BPFP=0.3043 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,676B, BPFP=1.4227 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,512B, BPFP=0.5270 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,284B, BPFP=1.4009 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,488B, BPFP=0.5811 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,304B, BPFP=1.3466 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,804B, BPFP=0.5432 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,688B, BPFP=1.3679 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,648B, BPFP=0.9778 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,832B, BPFP=1.2651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,336B, BPFP=0.1293 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13695534 160.63392066 + layer.0.v_cache 0.00001452 0.01381505 + layer.1.k_cache 0.52272531 16.01122874 + layer.1.v_cache 0.00000574 0.00510424 + layer.2.k_cache 0.00793556 1.63257236 + layer.2.v_cache 0.00001907 0.01413610 + layer.3.k_cache 0.04682703 7.27802336 + layer.3.v_cache 0.00002008 0.01580359 + layer.4.k_cache 0.00068945 0.36987521 + layer.4.v_cache 0.00005221 0.03470163 + layer.4.output 0.00941004 196.90303635 + ------------------------------------------------------------------------------------- + TOTAL 0.04594792 92.01943737 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 192064 +BPFP 0.6260 bits/point +EBPFP 1.2520 equivalent bits/point +MSE 92.019437 +---------------------- -------------------------------------------------------- +Time: 0.653s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 92.0194 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,540B, BPFP=0.3092 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,700B, BPFP=1.4342 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,352B, BPFP=0.5219 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,000B, BPFP=1.3951 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,360B, BPFP=0.5781 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,800B, BPFP=1.3281 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,544B, BPFP=0.5326 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,380B, BPFP=1.3605 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,180B, BPFP=0.9587 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,480B, BPFP=1.2545 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,032B, BPFP=0.1278 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14340984 151.25799386 + layer.0.v_cache 0.00001397 0.01326519 + layer.1.k_cache 0.51622396 15.94526367 + layer.1.v_cache 0.00000557 0.00477450 + layer.2.k_cache 0.00797994 1.66028006 + layer.2.v_cache 0.00001813 0.01338426 + layer.3.k_cache 0.02756977 6.89328221 + layer.3.v_cache 0.00001890 0.01537293 + layer.4.k_cache 0.00068956 0.35698588 + layer.4.v_cache 0.00005177 0.03357613 + layer.4.output 0.00947078 198.27064732 + ------------------------------------------------------------------------------------- + TOTAL 0.04483982 92.00521823 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 189368 +BPFP 0.6216 bits/point +EBPFP 1.2432 equivalent bits/point +MSE 92.005218 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.011s, Pack+Encode: 0.256s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 92.0052 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,644B, BPFP=0.3000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,864B, BPFP=1.3746 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,516B, BPFP=0.5057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,056B, BPFP=1.3316 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,608B, BPFP=0.5638 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,124B, BPFP=1.2821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,944B, BPFP=0.5285 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,420B, BPFP=1.2978 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,544B, BPFP=0.9324 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,768B, BPFP=1.2100 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,604B, BPFP=0.1337 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.393s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13622742 150.27032844 + layer.0.v_cache 0.00001380 0.01335198 + layer.1.k_cache 0.61520739 16.03210366 + layer.1.v_cache 0.00000584 0.00496741 + layer.2.k_cache 0.00900942 1.78537859 + layer.2.v_cache 0.00001916 0.01413526 + layer.3.k_cache 0.02667915 7.16623651 + layer.3.v_cache 0.00001881 0.01519035 + layer.4.k_cache 0.00071466 0.36109849 + layer.4.v_cache 0.00005760 0.03343823 + layer.4.output 0.04972813 184.64127794 + ------------------------------------------------------------------------------------- + TOTAL 0.06682648 86.36383380 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 193092 +BPFP 0.6037 bits/point +EBPFP 1.2073 equivalent bits/point +MSE 86.363834 +---------------------- -------------------------------------------------------- +Time: 0.661s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.393s +---------------------- -------------------------------------------------------- +💾 Converting with 86.3638 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,580B, BPFP=0.2916 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,092B, BPFP=1.3635 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,500B, BPFP=0.4964 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,132B, BPFP=1.3133 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,860B, BPFP=0.5675 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,168B, BPFP=1.2630 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,004B, BPFP=0.5228 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,604B, BPFP=1.2857 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,460B, BPFP=0.9124 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,012B, BPFP=1.2026 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,056B, BPFP=0.1273 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11478914 155.17737510 + layer.0.v_cache 0.00001377 0.01297512 + layer.1.k_cache 0.64218915 16.04402043 + layer.1.v_cache 0.00000556 0.00458857 + layer.2.k_cache 0.00898698 1.72707446 + layer.2.v_cache 0.00001980 0.01362715 + layer.3.k_cache 0.05442990 8.10418354 + layer.3.v_cache 0.00001964 0.01573312 + layer.4.k_cache 0.00070461 0.35831709 + layer.4.v_cache 0.00005015 0.03326044 + layer.4.output 0.05017195 181.40268156 + ------------------------------------------------------------------------------------- + TOTAL 0.06896543 85.37117211 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 193468 +BPFP 0.5947 bits/point +EBPFP 1.1894 equivalent bits/point +MSE 85.371172 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.011s, Pack+Encode: 0.254s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 85.3712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,516B, BPFP=0.3252 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,200B, BPFP=1.4858 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,248B, BPFP=0.5453 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,512B, BPFP=1.4453 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,292B, BPFP=0.6068 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,024B, BPFP=1.3575 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,792B, BPFP=0.5774 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,632B, BPFP=1.3934 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,804B, BPFP=0.9908 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,288B, BPFP=1.3142 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,436B, BPFP=0.1300 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14755792 150.99056604 + layer.0.v_cache 0.00001375 0.01353086 + layer.1.k_cache 0.51940866 16.26014151 + layer.1.v_cache 0.00000586 0.00514982 + layer.2.k_cache 0.00816260 1.68560814 + layer.2.v_cache 0.00001900 0.01419358 + layer.3.k_cache 0.02921419 7.59562574 + layer.3.v_cache 0.00001979 0.01589369 + layer.4.k_cache 0.00071017 0.36778049 + layer.4.v_cache 0.00005507 0.03384622 + layer.4.output 0.00937099 209.53013814 + ------------------------------------------------------------------------------------- + TOTAL 0.04533906 96.68784136 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 185744 +BPFP 0.6442 bits/point +EBPFP 1.2885 equivalent bits/point +MSE 96.687841 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.009s, Pack+Encode: 0.250s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 96.6878 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,520B, BPFP=0.3230 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,476B, BPFP=1.4909 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,048B, BPFP=0.5295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,720B, BPFP=1.4466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,296B, BPFP=0.6025 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,228B, BPFP=1.3593 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,752B, BPFP=0.5707 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,756B, BPFP=1.3902 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,044B, BPFP=0.9974 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,716B, BPFP=1.3294 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,556B, BPFP=0.1300 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12840328 160.69086493 + layer.0.v_cache 0.00001372 0.01286779 + layer.1.k_cache 0.60081110 16.34125922 + layer.1.v_cache 0.00000562 0.00469397 + layer.2.k_cache 0.00586274 1.74632463 + layer.2.v_cache 0.00001839 0.01339294 + layer.3.k_cache 0.02491910 7.57182946 + layer.3.v_cache 0.00001957 0.01544879 + layer.4.k_cache 0.00072047 0.37321846 + layer.4.v_cache 0.00005229 0.03477470 + layer.4.output 0.01050496 207.82838416 + ------------------------------------------------------------------------------------- + TOTAL 0.04908006 96.56490377 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 187112 +BPFP 0.6441 bits/point +EBPFP 1.2882 equivalent bits/point +MSE 96.564904 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 96.5649 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,556B, BPFP=0.3068 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,160B, BPFP=1.3891 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,448B, BPFP=0.5216 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,544B, BPFP=1.3551 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,420B, BPFP=0.5753 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,564B, BPFP=1.3010 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,668B, BPFP=0.5338 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,572B, BPFP=1.3015 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,084B, BPFP=0.9432 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,120B, BPFP=1.2213 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,840B, BPFP=0.1249 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14595435 156.78813770 + layer.0.v_cache 0.00001364 0.01277065 + layer.1.k_cache 0.59759904 15.91795840 + layer.1.v_cache 0.00000566 0.00460432 + layer.2.k_cache 0.01409306 1.70830544 + layer.2.v_cache 0.00002047 0.01349473 + layer.3.k_cache 0.06079736 8.19131567 + layer.3.v_cache 0.00002215 0.01529797 + layer.4.k_cache 0.00070776 0.35874901 + layer.4.v_cache 0.00005221 0.03375606 + layer.4.output 0.00769303 196.22231512 + ------------------------------------------------------------------------------------- + TOTAL 0.05135982 91.56474093 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 186976 +BPFP 0.6073 bits/point +EBPFP 1.2145 equivalent bits/point +MSE 91.564741 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.386s +---------------------- -------------------------------------------------------- +💾 Converting with 91.5647 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 315, 128) +Output shape: (1, 315, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.output: torch.Size([1, 315, 3584]) -> torch.Size([1, 1, 315, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,732B, BPFP=0.2843 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,700B, BPFP=1.3244 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,640B, BPFP=0.4782 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,736B, BPFP=1.2766 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,192B, BPFP=0.5056 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,504B, BPFP=1.2155 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,380B, BPFP=0.4653 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,912B, BPFP=1.2357 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,908B, BPFP=0.8883 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,348B, BPFP=1.1581 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,008B, BPFP=0.1205 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15500355 152.53599950 + layer.0.v_cache 0.00001366 0.01334682 + layer.1.k_cache 0.68167124 15.94832434 + layer.1.v_cache 0.00000571 0.00500994 + layer.2.k_cache 0.01007247 1.66481934 + layer.2.v_cache 0.00001936 0.01422315 + layer.3.k_cache 0.03272899 7.21710844 + layer.3.v_cache 0.00002006 0.01557826 + layer.4.k_cache 0.00071175 0.36667476 + layer.4.v_cache 0.00005427 0.03388410 + layer.4.output 0.04584261 172.14641440 + ------------------------------------------------------------------------------------- + TOTAL 0.07065879 81.34352173 + (elements=2,741,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2741760 +Total Bytes 195060 +BPFP 0.5692 bits/point +EBPFP 1.1383 equivalent bits/point +MSE 81.343522 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.012s, Pack+Encode: 0.256s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 81.3435 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,576B, BPFP=0.3215 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,632B, BPFP=1.4779 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,216B, BPFP=0.5314 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,568B, BPFP=1.4165 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,044B, BPFP=0.5791 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,204B, BPFP=1.3379 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,652B, BPFP=0.5565 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,600B, BPFP=1.3607 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,964B, BPFP=0.9781 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,312B, BPFP=1.2864 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,448B, BPFP=0.1272 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12889319 155.22215175 + layer.0.v_cache 0.00001350 0.01290762 + layer.1.k_cache 0.58775566 16.43226044 + layer.1.v_cache 0.00000553 0.00486224 + layer.2.k_cache 0.00723753 1.69114961 + layer.2.v_cache 0.00001924 0.01381640 + layer.3.k_cache 0.01761036 7.59783192 + layer.3.v_cache 0.00001847 0.01529530 + layer.4.k_cache 0.00071814 0.35857470 + layer.4.v_cache 0.00005270 0.03370241 + layer.4.output 0.01035140 204.92720414 + ------------------------------------------------------------------------------------- + TOTAL 0.04792848 95.05135184 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 186216 +BPFP 0.6316 bits/point +EBPFP 1.2631 equivalent bits/point +MSE 95.051352 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.010s, Pack+Encode: 0.261s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 95.0514 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,648B, BPFP=0.3107 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,956B, BPFP=1.3730 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,580B, BPFP=0.5271 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,200B, BPFP=1.3314 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,500B, BPFP=0.5777 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,264B, BPFP=1.2799 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,748B, BPFP=0.5363 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,476B, BPFP=1.2916 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,200B, BPFP=0.9463 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,720B, BPFP=1.1950 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,964B, BPFP=0.1255 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13033964 160.24045445 + layer.0.v_cache 0.00001368 0.01310755 + layer.1.k_cache 0.63029007 16.08694673 + layer.1.v_cache 0.00000551 0.00478614 + layer.2.k_cache 0.00637632 1.64039913 + layer.2.v_cache 0.00001960 0.01336662 + layer.3.k_cache 0.02368569 7.72053377 + layer.3.v_cache 0.00001974 0.01495837 + layer.4.k_cache 0.00068050 0.35391842 + layer.4.v_cache 0.00005209 0.03377375 + layer.4.output 0.00765416 195.43566084 + ------------------------------------------------------------------------------------- + TOTAL 0.04970952 91.42187475 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 186256 +BPFP 0.6028 bits/point +EBPFP 1.2056 equivalent bits/point +MSE 91.421875 +---------------------- -------------------------------------------------------- +Time: 0.649s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,680B, BPFP=0.3263 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,924B, BPFP=1.4892 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,240B, BPFP=0.5308 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,768B, BPFP=1.4228 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,540B, BPFP=0.6055 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,560B, BPFP=1.3534 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,816B, BPFP=0.5639 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,192B, BPFP=1.3897 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,136B, BPFP=0.9844 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,000B, BPFP=1.3212 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,740B, BPFP=0.1292 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15644295 155.26438994 + layer.0.v_cache 0.00001405 0.01312147 + layer.1.k_cache 0.56163917 16.01636819 + layer.1.v_cache 0.00000554 0.00502670 + layer.2.k_cache 0.01059573 1.75835867 + layer.2.v_cache 0.00001807 0.01400048 + layer.3.k_cache 0.04944291 7.60352460 + layer.3.v_cache 0.00001950 0.01599872 + layer.4.k_cache 0.00071267 0.36807509 + layer.4.v_cache 0.00005817 0.03635369 + layer.4.output 0.00956290 204.17082458 + ------------------------------------------------------------------------------------- + TOTAL 0.04975818 94.72299939 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 189596 +BPFP 0.6407 bits/point +EBPFP 1.2813 equivalent bits/point +MSE 94.722999 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 94.7230 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 312, 128) +Output shape: (1, 312, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.output: torch.Size([1, 312, 3584]) -> torch.Size([1, 1, 312, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,728B, BPFP=0.2869 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,664B, BPFP=1.3353 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,588B, BPFP=0.4802 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,804B, BPFP=1.2923 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,200B, BPFP=0.5108 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,624B, BPFP=1.2332 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,920B, BPFP=0.4968 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,032B, BPFP=1.2536 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,748B, BPFP=0.8888 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,484B, BPFP=1.1761 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,372B, BPFP=0.1243 +⌛️ [2/4] FRONTEND: Frontend time: 0.266s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14984696 152.26139323 + layer.0.v_cache 0.00001474 0.01334929 + layer.1.k_cache 0.67859346 15.77389292 + layer.1.v_cache 0.00000581 0.00494256 + layer.2.k_cache 0.01626646 1.67521863 + layer.2.v_cache 0.00001951 0.01399751 + layer.3.k_cache 0.02914289 6.80223436 + layer.3.v_cache 0.00001919 0.01566338 + layer.4.k_cache 0.00071665 0.35464282 + layer.4.v_cache 0.00005096 0.03283392 + layer.4.output 0.04642455 173.73340201 + ------------------------------------------------------------------------------------- + TOTAL 0.07056756 81.94599898 + (elements=2,715,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2715648 +Total Bytes 196164 +BPFP 0.5779 bits/point +EBPFP 1.1558 equivalent bits/point +MSE 81.945999 +---------------------- -------------------------------------------------------- +Time: 0.662s Load: 0.011s, Pack+Encode: 0.266s, Decode+Unpack: 0.386s +---------------------- -------------------------------------------------------- +💾 Converting with 81.9460 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,508B, BPFP=0.3199 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,496B, BPFP=1.4809 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,156B, BPFP=0.5318 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,604B, BPFP=1.4291 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,240B, BPFP=0.5948 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,192B, BPFP=1.3471 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,952B, BPFP=0.5781 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,460B, BPFP=1.3627 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,052B, BPFP=0.9905 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,664B, BPFP=1.3164 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,844B, BPFP=0.1315 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15823836 154.85995586 + layer.0.v_cache 0.00001502 0.01301230 + layer.1.k_cache 0.53668020 16.33168818 + layer.1.v_cache 0.00000576 0.00486035 + layer.2.k_cache 0.01018995 1.69954435 + layer.2.v_cache 0.00002218 0.01382313 + layer.3.k_cache 0.05410197 7.59873355 + layer.3.v_cache 0.00002001 0.01571257 + layer.4.k_cache 0.00069143 0.36409175 + layer.4.v_cache 0.00005304 0.03479191 + layer.4.output 0.01041225 206.18311869 + ------------------------------------------------------------------------------------- + TOTAL 0.04899433 95.54223793 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 187168 +BPFP 0.6395 bits/point +EBPFP 1.2790 equivalent bits/point +MSE 95.542238 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 95.5422 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 329, 128) +Output shape: (1, 329, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.output: torch.Size([1, 329, 3584]) -> torch.Size([1, 1, 329, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,528B, BPFP=0.3100 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,800B, BPFP=1.4628 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,136B, BPFP=0.5289 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,692B, BPFP=1.4101 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,108B, BPFP=0.5750 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,952B, BPFP=1.3275 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,692B, BPFP=0.5553 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,752B, BPFP=1.3655 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,376B, BPFP=0.9677 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,196B, BPFP=1.2916 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,820B, BPFP=0.1413 +⌛️ [2/4] FRONTEND: Frontend time: 0.347s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.442s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11391156 151.94711721 + layer.0.v_cache 0.00001522 0.01351728 + layer.1.k_cache 0.63933222 15.72476076 + layer.1.v_cache 0.00000652 0.00517649 + layer.2.k_cache 0.01643764 1.69026884 + layer.2.v_cache 0.00001873 0.01368770 + layer.3.k_cache 0.02826017 7.38320802 + layer.3.v_cache 0.00001876 0.01536696 + layer.4.k_cache 0.00074607 0.35954962 + layer.4.v_cache 0.00005483 0.03345907 + layer.4.output 3.44632065 163.34550043 + ------------------------------------------------------------------------------------- + TOTAL 1.46606155 77.68262441 + (elements=2,863,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2863616 +Total Bytes 227052 +BPFP 0.6343 bits/point +EBPFP 1.2686 equivalent bits/point +MSE 77.682624 +---------------------- -------------------------------------------------------- +Time: 0.801s Load: 0.012s, Pack+Encode: 0.347s, Decode+Unpack: 0.442s +---------------------- -------------------------------------------------------- +💾 Converting with 77.6826 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,524B, BPFP=0.3094 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,164B, BPFP=1.4653 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,212B, BPFP=0.5159 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,416B, BPFP=1.4234 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,584B, BPFP=0.5927 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,080B, BPFP=1.3486 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,660B, BPFP=0.5410 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,280B, BPFP=1.3598 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,376B, BPFP=0.9731 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,772B, BPFP=1.2753 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,460B, BPFP=0.1237 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13536828 156.61360327 + layer.0.v_cache 0.00001433 0.01303872 + layer.1.k_cache 0.57185238 16.22359431 + layer.1.v_cache 0.00000558 0.00478449 + layer.2.k_cache 0.00766427 1.79309738 + layer.2.v_cache 0.00001891 0.01375786 + layer.3.k_cache 0.04143807 7.40455239 + layer.3.v_cache 0.00001947 0.01512206 + layer.4.k_cache 0.00070404 0.35815394 + layer.4.v_cache 0.00005498 0.03383541 + layer.4.output 0.00948056 198.96028546 + ------------------------------------------------------------------------------------- + TOTAL 0.04844142 92.65856106 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 190528 +BPFP 0.6277 bits/point +EBPFP 1.2553 equivalent bits/point +MSE 92.658561 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.009s, Pack+Encode: 0.259s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 92.6586 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,496B, BPFP=0.3111 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,760B, BPFP=1.4583 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,384B, BPFP=0.5312 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,092B, BPFP=1.4205 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,644B, BPFP=0.6026 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,812B, BPFP=1.3481 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,712B, BPFP=0.5498 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,200B, BPFP=1.3700 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,152B, BPFP=0.9710 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,964B, BPFP=1.3000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,832B, BPFP=0.1280 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.392s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15132187 154.76265285 + layer.0.v_cache 0.00001456 0.01316960 + layer.1.k_cache 0.61798344 16.02286076 + layer.1.v_cache 0.00000569 0.00474850 + layer.2.k_cache 0.01686249 1.77258776 + layer.2.v_cache 0.00001898 0.01396901 + layer.3.k_cache 0.04401700 7.74940380 + layer.3.v_cache 0.00001935 0.01496782 + layer.4.k_cache 0.00071189 0.36142385 + layer.4.v_cache 0.00005111 0.03359277 + layer.4.output 0.01058244 201.06036491 + ------------------------------------------------------------------------------------- + TOTAL 0.05324020 93.42187830 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 190048 +BPFP 0.6329 bits/point +EBPFP 1.2658 equivalent bits/point +MSE 93.421878 +---------------------- -------------------------------------------------------- +Time: 0.660s Load: 0.010s, Pack+Encode: 0.258s, Decode+Unpack: 0.392s +---------------------- -------------------------------------------------------- +💾 Converting with 93.4219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,548B, BPFP=0.3141 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,092B, BPFP=1.4771 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,244B, BPFP=0.5233 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,476B, BPFP=1.4423 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,900B, BPFP=0.6171 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,936B, BPFP=1.3551 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,832B, BPFP=0.5566 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,608B, BPFP=1.3931 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,380B, BPFP=0.9839 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,916B, BPFP=1.2973 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,572B, BPFP=0.1259 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12581717 156.38291157 + layer.0.v_cache 0.00001375 0.01340673 + layer.1.k_cache 0.57814745 16.23612998 + layer.1.v_cache 0.00000592 0.00502772 + layer.2.k_cache 0.00733882 1.71036762 + layer.2.v_cache 0.00002058 0.01462403 + layer.3.k_cache 0.02353939 6.89630525 + layer.3.v_cache 0.00002077 0.01602844 + layer.4.k_cache 0.00069182 0.37205257 + layer.4.v_cache 0.00005027 0.03486611 + layer.4.output 0.01015823 201.13502847 + ------------------------------------------------------------------------------------- + TOTAL 0.04745609 93.50746584 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 191504 +BPFP 0.6377 bits/point +EBPFP 1.2755 equivalent bits/point +MSE 93.507466 +---------------------- -------------------------------------------------------- +Time: 0.649s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 93.5075 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,580B, BPFP=0.3182 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,956B, BPFP=1.4802 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,140B, BPFP=0.5212 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,352B, BPFP=1.4457 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,940B, BPFP=0.6239 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,808B, BPFP=1.3577 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,772B, BPFP=0.5573 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,568B, BPFP=1.4010 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,356B, BPFP=0.9897 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,320B, BPFP=1.3298 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,052B, BPFP=0.1308 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13625057 155.55260607 + layer.0.v_cache 0.00001362 0.01272628 + layer.1.k_cache 0.57512581 16.23083763 + layer.1.v_cache 0.00000584 0.00483042 + layer.2.k_cache 0.00872732 1.78423337 + layer.2.v_cache 0.00001851 0.01390183 + layer.3.k_cache 0.04362779 7.58895451 + layer.3.v_cache 0.00001925 0.01599285 + layer.4.k_cache 0.00073619 0.36450056 + layer.4.v_cache 0.00005148 0.03421578 + layer.4.output 0.00783937 202.52605253 + ------------------------------------------------------------------------------------- + TOTAL 0.04820306 94.07559806 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 191844 +BPFP 0.6435 bits/point +EBPFP 1.2871 equivalent bits/point +MSE 94.075598 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.010s, Pack+Encode: 0.256s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 94.0756 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,700B, BPFP=0.2930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,140B, BPFP=1.3435 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,428B, BPFP=0.4846 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,232B, BPFP=1.2969 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,664B, BPFP=0.5481 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,900B, BPFP=1.2284 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,776B, BPFP=0.5025 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,504B, BPFP=1.2595 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,404B, BPFP=0.8945 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,916B, BPFP=1.1778 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,872B, BPFP=0.1312 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16730670 151.40798468 + layer.0.v_cache 0.00001388 0.01276482 + layer.1.k_cache 0.64158324 15.75973511 + layer.1.v_cache 0.00000571 0.00471234 + layer.2.k_cache 0.02207795 1.72566022 + layer.2.v_cache 0.00001922 0.01337775 + layer.3.k_cache 0.04503629 7.58809220 + layer.3.v_cache 0.00001955 0.01566076 + layer.4.k_cache 0.00070478 0.34503324 + layer.4.v_cache 0.00005260 0.03314881 + layer.4.output 0.04763387 178.46968985 + ------------------------------------------------------------------------------------- + TOTAL 0.07119159 83.89376464 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 193536 +BPFP 0.5851 bits/point +EBPFP 1.1703 equivalent bits/point +MSE 83.893765 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.011s, Pack+Encode: 0.257s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8938 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,504B, BPFP=0.3105 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,836B, BPFP=1.4574 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,216B, BPFP=0.5199 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,964B, BPFP=1.4082 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,504B, BPFP=0.5925 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,492B, BPFP=1.3251 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,652B, BPFP=0.5444 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,932B, BPFP=1.3500 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,908B, BPFP=0.9537 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,600B, BPFP=1.2748 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,288B, BPFP=0.1313 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15225300 153.71141697 + layer.0.v_cache 0.00001427 0.01257646 + layer.1.k_cache 0.52731114 16.06880888 + layer.1.v_cache 0.00000556 0.00484439 + layer.2.k_cache 0.01483519 1.61732334 + layer.2.v_cache 0.00001864 0.01385095 + layer.3.k_cache 0.04699480 7.52284874 + layer.3.v_cache 0.00001968 0.01533213 + layer.4.k_cache 0.00070142 0.35801606 + layer.4.v_cache 0.00005177 0.03276787 + layer.4.output 0.01077793 200.31457904 + ------------------------------------------------------------------------------------- + TOTAL 0.04809712 93.03293171 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 188896 +BPFP 0.6268 bits/point +EBPFP 1.2536 equivalent bits/point +MSE 93.032932 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.009s, Pack+Encode: 0.254s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 93.0329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,516B, BPFP=0.3111 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,012B, BPFP=1.4673 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,112B, BPFP=0.5140 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,352B, BPFP=1.4301 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,568B, BPFP=0.5961 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,008B, BPFP=1.3542 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,664B, BPFP=0.5451 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,288B, BPFP=1.3700 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,080B, BPFP=0.9634 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,732B, BPFP=1.2823 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,092B, BPFP=0.1297 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15700974 151.65130020 + layer.0.v_cache 0.00001374 0.01237397 + layer.1.k_cache 0.52255304 16.03649594 + layer.1.v_cache 0.00000566 0.00471445 + layer.2.k_cache 0.01360668 1.66696795 + layer.2.v_cache 0.00001926 0.01373669 + layer.3.k_cache 0.08085050 8.02494818 + layer.3.v_cache 0.00001943 0.01538321 + layer.4.k_cache 0.00071029 0.35951712 + layer.4.v_cache 0.00005450 0.03383345 + layer.4.output 0.01014708 200.43642019 + ------------------------------------------------------------------------------------- + TOTAL 0.04975720 92.99260074 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 190424 +BPFP 0.6318 bits/point +EBPFP 1.2637 equivalent bits/point +MSE 92.992601 +---------------------- -------------------------------------------------------- +Time: 0.630s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 92.9926 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,496B, BPFP=0.3013 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,024B, BPFP=1.3719 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,360B, BPFP=0.5132 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,440B, BPFP=1.3399 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,620B, BPFP=0.5822 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,632B, BPFP=1.2956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,884B, BPFP=0.5419 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,008B, BPFP=1.3162 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,196B, BPFP=0.9428 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,964B, BPFP=1.2042 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,616B, BPFP=0.1301 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16436556 157.32724781 + layer.0.v_cache 0.00001441 0.01274770 + layer.1.k_cache 0.62741763 16.22239583 + layer.1.v_cache 0.00000560 0.00472022 + layer.2.k_cache 0.01487184 1.66406132 + layer.2.v_cache 0.00002167 0.01365374 + layer.3.k_cache 0.02142572 7.81066252 + layer.3.v_cache 0.00001943 0.01528089 + layer.4.k_cache 0.00070776 0.35737168 + layer.4.v_cache 0.00005251 0.03302940 + layer.4.output 0.00894435 194.72774123 + ------------------------------------------------------------------------------------- + TOTAL 0.05244192 90.97384469 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 188240 +BPFP 0.6071 bits/point +EBPFP 1.2141 equivalent bits/point +MSE 90.973845 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.010s, Pack+Encode: 0.262s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 90.9738 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,544B, BPFP=0.3083 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,592B, BPFP=1.4230 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,440B, BPFP=0.5249 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,908B, BPFP=1.3850 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,480B, BPFP=0.5827 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,788B, BPFP=1.3227 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,836B, BPFP=0.5469 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,324B, BPFP=1.3525 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,020B, BPFP=0.9464 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,568B, BPFP=1.2549 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,224B, BPFP=0.1209 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11705062 158.11779915 + layer.0.v_cache 0.00001445 0.01324600 + layer.1.k_cache 0.56620099 16.26635482 + layer.1.v_cache 0.00000566 0.00487164 + layer.2.k_cache 0.01299965 1.68995161 + layer.2.v_cache 0.00001972 0.01352204 + layer.3.k_cache 0.05069733 7.46954063 + layer.3.v_cache 0.00001944 0.01558103 + layer.4.k_cache 0.00068883 0.35321517 + layer.4.v_cache 0.00005144 0.03269709 + layer.4.output 0.00838925 197.47353203 + ------------------------------------------------------------------------------------- + TOTAL 0.04743958 92.13479432 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 188724 +BPFP 0.6173 bits/point +EBPFP 1.2346 equivalent bits/point +MSE 92.134794 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.386s +---------------------- -------------------------------------------------------- +💾 Converting with 92.1348 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,568B, BPFP=0.3074 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,272B, BPFP=1.3953 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,500B, BPFP=0.5245 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,996B, BPFP=1.3801 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,180B, BPFP=0.5621 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,760B, BPFP=1.3118 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,516B, BPFP=0.5254 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,332B, BPFP=1.3434 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,456B, BPFP=0.9638 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,424B, BPFP=1.2381 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,044B, BPFP=0.1265 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14370025 157.15068739 + layer.0.v_cache 0.00001382 0.01328091 + layer.1.k_cache 0.61448195 16.11728067 + layer.1.v_cache 0.00000570 0.00490806 + layer.2.k_cache 0.01208573 1.65744029 + layer.2.v_cache 0.00001823 0.01394544 + layer.3.k_cache 0.01319265 7.13375154 + layer.3.v_cache 0.00002242 0.01563916 + layer.4.k_cache 0.00071920 0.36642904 + layer.4.v_cache 0.00005145 0.03483989 + layer.4.output 0.00992508 196.20545179 + ------------------------------------------------------------------------------------- + TOTAL 0.05022159 91.52625676 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 189048 +BPFP 0.6140 bits/point +EBPFP 1.2280 equivalent bits/point +MSE 91.526257 +---------------------- -------------------------------------------------------- +Time: 0.646s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 91.5263 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 259, 128) +Output shape: (1, 259, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.output: torch.Size([1, 259, 3584]) -> torch.Size([1, 1, 259, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,376B, BPFP=0.3243 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,536B, BPFP=1.4802 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,040B, BPFP=0.5454 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,472B, BPFP=1.4160 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,360B, BPFP=0.6250 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,248B, BPFP=1.3422 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,564B, BPFP=0.5770 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,596B, BPFP=1.3632 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,628B, BPFP=1.0031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,476B, BPFP=1.2956 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,160B, BPFP=0.1393 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13349269 154.29986125 + layer.0.v_cache 0.00001436 0.01337741 + layer.1.k_cache 0.50562990 16.18965485 + layer.1.v_cache 0.00000590 0.00497573 + layer.2.k_cache 0.00820730 1.73911854 + layer.2.v_cache 0.00001794 0.01353436 + layer.3.k_cache 0.03685973 7.46662437 + layer.3.v_cache 0.00002061 0.01502120 + layer.4.k_cache 0.00069539 0.36304356 + layer.4.v_cache 0.00005422 0.03373967 + layer.4.output 0.01018517 214.35731522 + ------------------------------------------------------------------------------------- + TOTAL 0.04448790 98.86118574 + (elements=2,254,336) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2254336 +Total Bytes 181456 +BPFP 0.6439 bits/point +EBPFP 1.2879 equivalent bits/point +MSE 98.861186 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.010s, Pack+Encode: 0.256s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 98.8612 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,520B, BPFP=0.3194 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,816B, BPFP=1.4940 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,148B, BPFP=0.5294 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,804B, BPFP=1.4354 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,428B, BPFP=0.6035 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,344B, BPFP=1.3509 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,940B, BPFP=0.5752 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,652B, BPFP=1.3687 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,092B, BPFP=0.9891 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,696B, BPFP=1.3134 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,764B, BPFP=0.1303 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12649573 157.02586806 + layer.0.v_cache 0.00001373 0.01325845 + layer.1.k_cache 0.51574391 16.46226671 + layer.1.v_cache 0.00000589 0.00488503 + layer.2.k_cache 0.01194659 1.77199085 + layer.2.v_cache 0.00001984 0.01419641 + layer.3.k_cache 0.04962628 7.39760064 + layer.3.v_cache 0.00002059 0.01608710 + layer.4.k_cache 0.00068541 0.36410438 + layer.4.v_cache 0.00005704 0.03520258 + layer.4.output 0.00772497 205.59244378 + ------------------------------------------------------------------------------------- + TOTAL 0.04462881 95.42662157 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 188204 +BPFP 0.6407 bits/point +EBPFP 1.2813 equivalent bits/point +MSE 95.426622 +---------------------- -------------------------------------------------------- +Time: 0.630s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 95.4266 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,480B, BPFP=0.3183 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,496B, BPFP=1.4809 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,108B, BPFP=0.5290 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,528B, BPFP=1.4247 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,440B, BPFP=0.6064 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,344B, BPFP=1.3559 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,780B, BPFP=0.5681 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,744B, BPFP=1.3792 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,152B, BPFP=0.9963 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,692B, BPFP=1.3181 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,588B, BPFP=0.1293 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12210708 157.13856296 + layer.0.v_cache 0.00001342 0.01292188 + layer.1.k_cache 0.54942129 16.20426493 + layer.1.v_cache 0.00000547 0.00484680 + layer.2.k_cache 0.01053270 1.78637355 + layer.2.v_cache 0.00001820 0.01388557 + layer.3.k_cache 0.02850239 7.86878576 + layer.3.v_cache 0.00002491 0.01565449 + layer.4.k_cache 0.00067987 0.36956072 + layer.4.v_cache 0.00005179 0.03436080 + layer.4.output 0.01126663 206.29238582 + ------------------------------------------------------------------------------------- + TOTAL 0.04648373 95.73505401 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 187352 +BPFP 0.6401 bits/point +EBPFP 1.2803 equivalent bits/point +MSE 95.735054 +---------------------- -------------------------------------------------------- +Time: 0.624s Load: 0.009s, Pack+Encode: 0.245s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 95.7351 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,628B, BPFP=0.2902 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,232B, BPFP=1.3527 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,408B, BPFP=0.4851 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,328B, BPFP=1.3061 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,468B, BPFP=0.5398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,008B, BPFP=1.2380 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,996B, BPFP=0.5155 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,348B, BPFP=1.2556 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,296B, BPFP=0.8919 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,752B, BPFP=1.1733 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,756B, BPFP=0.1234 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14443027 155.32158880 + layer.0.v_cache 0.00001375 0.01345347 + layer.1.k_cache 0.62837083 15.92776499 + layer.1.v_cache 0.00000570 0.00486880 + layer.2.k_cache 0.01557094 1.65334156 + layer.2.v_cache 0.00001871 0.01367452 + layer.3.k_cache 0.04676386 7.28289855 + layer.3.v_cache 0.00001981 0.01538609 + layer.4.k_cache 0.00070582 0.36112369 + layer.4.v_cache 0.00005437 0.03319570 + layer.4.output 0.04838977 178.95299976 + ------------------------------------------------------------------------------------- + TOTAL 0.06909897 84.31166438 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 192220 +BPFP 0.5831 bits/point +EBPFP 1.1662 equivalent bits/point +MSE 84.311664 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 84.3117 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,544B, BPFP=0.3061 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,248B, BPFP=1.3940 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,472B, BPFP=0.5230 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,756B, BPFP=1.3668 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,568B, BPFP=0.5835 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,904B, BPFP=1.3198 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,656B, BPFP=0.5331 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,924B, BPFP=1.3209 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,160B, BPFP=0.9474 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,252B, BPFP=1.2286 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,496B, BPFP=0.1222 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13293209 156.31722063 + layer.0.v_cache 0.00001388 0.01313295 + layer.1.k_cache 0.56267701 16.07355448 + layer.1.v_cache 0.00000566 0.00487320 + layer.2.k_cache 0.01145046 1.66864035 + layer.2.v_cache 0.00001921 0.01393800 + layer.3.k_cache 0.02674564 7.02549156 + layer.3.v_cache 0.00001926 0.01556877 + layer.4.k_cache 0.00069135 0.36680700 + layer.4.v_cache 0.00005493 0.03525629 + layer.4.output 0.00973303 196.18199457 + ------------------------------------------------------------------------------------- + TOTAL 0.04722004 91.45932031 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 187980 +BPFP 0.6105 bits/point +EBPFP 1.2210 equivalent bits/point +MSE 91.459320 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,540B, BPFP=0.3125 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,840B, BPFP=1.4576 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,208B, BPFP=0.5194 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,444B, BPFP=1.4352 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,472B, BPFP=0.5907 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,884B, BPFP=1.3472 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,712B, BPFP=0.5478 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,416B, BPFP=1.3773 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,480B, BPFP=0.9860 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,908B, BPFP=1.2922 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,484B, BPFP=0.1248 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15182159 153.97393953 + layer.0.v_cache 0.00001420 0.01319467 + layer.1.k_cache 0.56224567 16.37176007 + layer.1.v_cache 0.00000588 0.00498505 + layer.2.k_cache 0.00652889 1.72523978 + layer.2.v_cache 0.00001844 0.01384107 + layer.3.k_cache 0.06011557 7.82855159 + layer.3.v_cache 0.00001921 0.01559951 + layer.4.k_cache 0.00068317 0.36812275 + layer.4.v_cache 0.00005096 0.03361269 + layer.4.output 0.01130176 200.12923865 + ------------------------------------------------------------------------------------- + TOTAL 0.05062446 93.01491278 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 190388 +BPFP 0.6317 bits/point +EBPFP 1.2635 equivalent bits/point +MSE 93.014913 +---------------------- -------------------------------------------------------- +Time: 0.630s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 93.0149 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,472B, BPFP=0.3000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,376B, BPFP=1.3912 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,336B, BPFP=0.5118 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,768B, BPFP=1.3579 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,608B, BPFP=0.5816 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,980B, BPFP=1.3147 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,768B, BPFP=0.5355 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,180B, BPFP=1.3257 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,732B, BPFP=0.9721 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,536B, BPFP=1.2355 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,640B, BPFP=0.1303 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13053865 156.71641996 + layer.0.v_cache 0.00001414 0.01400651 + layer.1.k_cache 0.56363723 16.05619518 + layer.1.v_cache 0.00000590 0.00509494 + layer.2.k_cache 0.01033439 1.64610360 + layer.2.v_cache 0.00002015 0.01438185 + layer.3.k_cache 0.03655195 7.52946478 + layer.3.v_cache 0.00002026 0.01568281 + layer.4.k_cache 0.00068516 0.37240638 + layer.4.v_cache 0.00005378 0.03482909 + layer.4.output 0.00863618 194.72363722 + ------------------------------------------------------------------------------------- + TOTAL 0.04719499 90.91000268 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 190396 +BPFP 0.6140 bits/point +EBPFP 1.2280 equivalent bits/point +MSE 90.910003 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.011s, Pack+Encode: 0.246s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 90.9100 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,516B, BPFP=0.3111 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,108B, BPFP=1.4727 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,220B, BPFP=0.5201 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,360B, BPFP=1.4305 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,660B, BPFP=0.6013 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,796B, BPFP=1.3423 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,912B, BPFP=0.5591 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,372B, BPFP=1.3748 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,216B, BPFP=0.9711 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,924B, BPFP=1.2931 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,788B, BPFP=0.1272 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15908699 153.64290952 + layer.0.v_cache 0.00001433 0.01301239 + layer.1.k_cache 0.53822481 16.08260414 + layer.1.v_cache 0.00000570 0.00484867 + layer.2.k_cache 0.01139215 1.69678957 + layer.2.v_cache 0.00001845 0.01366349 + layer.3.k_cache 0.02907990 7.81793103 + layer.3.v_cache 0.00001963 0.01593921 + layer.4.k_cache 0.00069750 0.35921343 + layer.4.v_cache 0.00005397 0.03459537 + layer.4.output 0.01069233 200.45759734 + ------------------------------------------------------------------------------------- + TOTAL 0.04784940 93.11086401 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 190872 +BPFP 0.6333 bits/point +EBPFP 1.2667 equivalent bits/point +MSE 93.110864 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.009s, Pack+Encode: 0.245s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 93.1109 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,460B, BPFP=0.2983 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,152B, BPFP=1.3741 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,276B, BPFP=0.5068 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,336B, BPFP=1.3295 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,532B, BPFP=0.5754 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,192B, BPFP=1.2670 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,996B, BPFP=0.5461 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,700B, BPFP=1.2948 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,172B, BPFP=0.9382 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,116B, BPFP=1.2083 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,564B, BPFP=0.1293 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16084659 156.20570640 + layer.0.v_cache 0.00001440 0.01294661 + layer.1.k_cache 0.59028764 16.27737277 + layer.1.v_cache 0.00000559 0.00475682 + layer.2.k_cache 0.00929020 1.79287357 + layer.2.v_cache 0.00002205 0.01405257 + layer.3.k_cache 0.04382113 7.66198901 + layer.3.v_cache 0.00002103 0.01558898 + layer.4.k_cache 0.00069271 0.36219617 + layer.4.v_cache 0.00005052 0.03360672 + layer.4.output 0.00984751 193.92498127 + ------------------------------------------------------------------------------------- + TOTAL 0.05141085 90.57976226 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 187496 +BPFP 0.6026 bits/point +EBPFP 1.2051 equivalent bits/point +MSE 90.579762 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 90.5798 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,376B, BPFP=0.2917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,016B, BPFP=1.3572 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,988B, BPFP=0.4876 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,132B, BPFP=1.3092 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,200B, BPFP=0.5534 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,044B, BPFP=1.2502 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,200B, BPFP=0.4991 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,524B, BPFP=1.2763 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,940B, BPFP=0.9191 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,088B, BPFP=1.1984 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,072B, BPFP=0.1323 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14211286 155.87917752 + layer.0.v_cache 0.00001516 0.01274504 + layer.1.k_cache 0.60291523 15.96898227 + layer.1.v_cache 0.00000570 0.00486907 + layer.2.k_cache 0.01041983 1.66468853 + layer.2.v_cache 0.00001835 0.01335156 + layer.3.k_cache 0.03896446 6.92378574 + layer.3.v_cache 0.00001926 0.01554266 + layer.4.k_cache 0.00071470 0.36730991 + layer.4.v_cache 0.00005236 0.03400208 + layer.4.output 0.00754456 192.76024616 + ------------------------------------------------------------------------------------- + TOTAL 0.04988528 90.01212809 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 185580 +BPFP 0.5923 bits/point +EBPFP 1.1845 equivalent bits/point +MSE 90.012128 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 90.0121 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 296, 128) +Output shape: (1, 296, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.output: torch.Size([1, 296, 3584]) -> torch.Size([1, 1, 296, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,696B, BPFP=0.3007 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,948B, BPFP=1.3697 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,588B, BPFP=0.5061 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,992B, BPFP=1.3193 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,884B, BPFP=0.5745 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,116B, BPFP=1.2730 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,996B, BPFP=0.5277 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,564B, BPFP=1.2967 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,704B, BPFP=0.9345 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,976B, BPFP=1.2128 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,984B, BPFP=0.1281 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14616069 152.18826541 + layer.0.v_cache 0.00001366 0.01303566 + layer.1.k_cache 0.61022733 15.97441802 + layer.1.v_cache 0.00000576 0.00477697 + layer.2.k_cache 0.01378242 1.69219847 + layer.2.v_cache 0.00001907 0.01370745 + layer.3.k_cache 0.03184220 7.62339494 + layer.3.v_cache 0.00001934 0.01523633 + layer.4.k_cache 0.00072724 0.35606119 + layer.4.v_cache 0.00005080 0.03254160 + layer.4.output 0.04784788 183.30796030 + ------------------------------------------------------------------------------------- + TOTAL 0.06692845 85.94525636 + (elements=2,576,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2576384 +Total Bytes 193448 +BPFP 0.6007 bits/point +EBPFP 1.2014 equivalent bits/point +MSE 85.945256 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 85.9453 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,576B, BPFP=0.2924 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,324B, BPFP=1.3802 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,536B, BPFP=0.5000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,320B, BPFP=1.3276 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,944B, BPFP=0.5738 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,236B, BPFP=1.2708 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,232B, BPFP=0.5365 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,672B, BPFP=1.2936 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,608B, BPFP=0.9232 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,188B, BPFP=1.2158 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,308B, BPFP=0.1296 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15387250 153.65552905 + layer.0.v_cache 0.00001382 0.01308227 + layer.1.k_cache 0.63844893 15.92773765 + layer.1.v_cache 0.00000583 0.00497380 + layer.2.k_cache 0.01105453 1.77496829 + layer.2.v_cache 0.00001885 0.01395570 + layer.3.k_cache 0.03893430 7.17539538 + layer.3.v_cache 0.00001923 0.01572233 + layer.4.k_cache 0.00070257 0.36010804 + layer.4.v_cache 0.00005146 0.03442326 + layer.4.output 0.04873301 182.11246105 + ------------------------------------------------------------------------------------- + TOTAL 0.06966195 85.51547783 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 194944 +BPFP 0.6013 bits/point +EBPFP 1.2025 equivalent bits/point +MSE 85.515478 +---------------------- -------------------------------------------------------- +Time: 0.649s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5155 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,564B, BPFP=0.3208 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,972B, BPFP=1.4975 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,304B, BPFP=0.5364 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,828B, BPFP=1.4315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,516B, BPFP=0.6063 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,276B, BPFP=1.3420 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,068B, BPFP=0.5805 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,604B, BPFP=1.3609 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,932B, BPFP=0.9762 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,792B, BPFP=1.3141 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,816B, BPFP=0.1385 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12629252 157.18871079 + layer.0.v_cache 0.00001451 0.01369193 + layer.1.k_cache 0.49097243 16.10845790 + layer.1.v_cache 0.00000577 0.00512666 + layer.2.k_cache 0.01060468 1.75053749 + layer.2.v_cache 0.00001909 0.01418098 + layer.3.k_cache 0.02748298 7.34855489 + layer.3.v_cache 0.00001876 0.01585756 + layer.4.k_cache 0.00069967 0.36996156 + layer.4.v_cache 0.00005414 0.03410301 + layer.4.output 0.01047103 204.80986426 + ------------------------------------------------------------------------------------- + TOTAL 0.04290952 95.08930780 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 189672 +BPFP 0.6433 bits/point +EBPFP 1.2866 equivalent bits/point +MSE 95.089308 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.253s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 95.0893 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,568B, BPFP=0.3118 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,760B, BPFP=1.4427 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,328B, BPFP=0.5224 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,928B, BPFP=1.3961 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,640B, BPFP=0.5959 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,572B, BPFP=1.3201 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,784B, BPFP=0.5479 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,928B, BPFP=1.3401 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,232B, BPFP=0.9651 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,776B, BPFP=1.2755 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,952B, BPFP=0.1276 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15022702 155.31304603 + layer.0.v_cache 0.00001395 0.01315500 + layer.1.k_cache 0.54271635 16.08004837 + layer.1.v_cache 0.00000573 0.00491088 + layer.2.k_cache 0.01332666 1.74039035 + layer.2.v_cache 0.00001832 0.01351757 + layer.3.k_cache 0.02579262 7.53094023 + layer.3.v_cache 0.00001838 0.01476759 + layer.4.k_cache 0.00069325 0.35380981 + layer.4.v_cache 0.00005155 0.03272034 + layer.4.output 0.01006869 199.06429211 + ------------------------------------------------------------------------------------- + TOTAL 0.04725557 92.62043241 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 189468 +BPFP 0.6242 bits/point +EBPFP 1.2483 equivalent bits/point +MSE 92.620432 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.012s, Pack+Encode: 0.245s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 92.6204 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,712B, BPFP=0.3025 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,888B, BPFP=1.3712 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,628B, BPFP=0.5100 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,108B, BPFP=1.3299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,700B, BPFP=0.5667 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,048B, BPFP=1.2737 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,108B, BPFP=0.5354 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,492B, BPFP=1.2972 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,768B, BPFP=0.9411 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,904B, BPFP=1.2131 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,744B, BPFP=0.1343 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12416320 151.34149894 + layer.0.v_cache 0.00001370 0.01261449 + layer.1.k_cache 0.62725913 15.75756422 + layer.1.v_cache 0.00000565 0.00477899 + layer.2.k_cache 0.01102555 1.69591354 + layer.2.v_cache 0.00002033 0.01375404 + layer.3.k_cache 0.04118381 7.22520111 + layer.3.v_cache 0.00001949 0.01553237 + layer.4.k_cache 0.00070768 0.36696940 + layer.4.v_cache 0.00005397 0.03312062 + layer.4.output 0.05105524 183.94921308 + ------------------------------------------------------------------------------------- + TOTAL 0.06834348 86.12420231 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 194100 +BPFP 0.6047 bits/point +EBPFP 1.2095 equivalent bits/point +MSE 86.124202 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 86.1242 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,364B, BPFP=0.2900 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,084B, BPFP=1.3562 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,064B, BPFP=0.4901 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,228B, BPFP=1.3099 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,052B, BPFP=0.5435 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,168B, BPFP=1.2526 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,380B, BPFP=0.5071 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,572B, BPFP=1.2744 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,132B, BPFP=0.9263 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,004B, BPFP=1.1897 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,768B, BPFP=0.1218 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15516999 154.93147167 + layer.0.v_cache 0.00001361 0.01349851 + layer.1.k_cache 0.56212566 16.22090080 + layer.1.v_cache 0.00000565 0.00527269 + layer.2.k_cache 0.00839357 1.69787819 + layer.2.v_cache 0.00001980 0.01437900 + layer.3.k_cache 0.02402391 7.15713786 + layer.3.v_cache 0.00001987 0.01629490 + layer.4.k_cache 0.00071212 0.36996323 + layer.4.v_cache 0.00005188 0.03509732 + layer.4.output 0.05073151 187.85416152 + ------------------------------------------------------------------------------------- + TOTAL 0.06503863 87.96711911 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 184816 +BPFP 0.5878 bits/point +EBPFP 1.1756 equivalent bits/point +MSE 87.967119 +---------------------- -------------------------------------------------------- +Time: 0.627s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.369s +---------------------- -------------------------------------------------------- +💾 Converting with 87.9671 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,652B, BPFP=0.2973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,236B, BPFP=1.3803 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,556B, BPFP=0.5027 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,336B, BPFP=1.3329 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,740B, BPFP=0.5650 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,264B, BPFP=1.2765 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,040B, BPFP=0.5282 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,644B, BPFP=1.2965 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,636B, BPFP=0.9278 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,036B, BPFP=1.2119 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,020B, BPFP=0.1279 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005571 152.26204756 + layer.0.v_cache 0.00001459 0.01323450 + layer.1.k_cache 0.64038677 15.88168633 + layer.1.v_cache 0.00000545 0.00473495 + layer.2.k_cache 0.01623841 1.76113409 + layer.2.v_cache 0.00002030 0.01376536 + layer.3.k_cache 0.01774169 7.46846640 + layer.3.v_cache 0.00001948 0.01558592 + layer.4.k_cache 0.00069419 0.36912311 + layer.4.v_cache 0.00005057 0.03403613 + layer.4.output 0.05052170 182.71264731 + ------------------------------------------------------------------------------------- + TOTAL 0.06875759 85.69484385 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 194160 +BPFP 0.6009 bits/point +EBPFP 1.2017 equivalent bits/point +MSE 85.694844 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.011s, Pack+Encode: 0.245s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 85.6948 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,464B, BPFP=0.3060 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,148B, BPFP=1.4644 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,204B, BPFP=0.5155 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,432B, BPFP=1.4243 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,668B, BPFP=0.5974 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,056B, BPFP=1.3472 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,780B, BPFP=0.5477 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,256B, BPFP=1.3584 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,448B, BPFP=0.9772 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,608B, BPFP=1.2661 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,788B, BPFP=0.1263 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14010120 155.72373432 + layer.0.v_cache 0.00001437 0.01319232 + layer.1.k_cache 0.58692467 16.18891059 + layer.1.v_cache 0.00000569 0.00485948 + layer.2.k_cache 0.00806839 1.76705156 + layer.2.v_cache 0.00001934 0.01374767 + layer.3.k_cache 0.03832107 7.44174840 + layer.3.v_cache 0.00001992 0.01555170 + layer.4.k_cache 0.00069804 0.35658237 + layer.4.v_cache 0.00005131 0.03403984 + layer.4.output 0.00759059 199.00771249 + ------------------------------------------------------------------------------------- + TOTAL 0.04866812 92.62431798 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 190852 +BPFP 0.6287 bits/point +EBPFP 1.2575 equivalent bits/point +MSE 92.624318 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 92.6243 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,504B, BPFP=0.3209 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,268B, BPFP=1.4732 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,232B, BPFP=0.5382 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,476B, BPFP=1.4270 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,472B, BPFP=0.6105 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,928B, BPFP=1.3368 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,968B, BPFP=0.5812 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,388B, BPFP=1.3636 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,016B, BPFP=0.9921 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,504B, BPFP=1.3120 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,804B, BPFP=0.1316 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13199043 156.45945079 + layer.0.v_cache 0.00001386 0.01345751 + layer.1.k_cache 0.56989129 16.31444219 + layer.1.v_cache 0.00000577 0.00497039 + layer.2.k_cache 0.00892691 1.79138776 + layer.2.v_cache 0.00001816 0.01404091 + layer.3.k_cache 0.04626510 7.38681167 + layer.3.v_cache 0.00001927 0.01578155 + layer.4.k_cache 0.00069392 0.37249431 + layer.4.v_cache 0.00005195 0.03485539 + layer.4.output 0.00911696 207.04192764 + ------------------------------------------------------------------------------------- + TOTAL 0.04833502 95.98242270 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 186560 +BPFP 0.6398 bits/point +EBPFP 1.2796 equivalent bits/point +MSE 95.982423 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 95.9824 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,604B, BPFP=0.3196 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,684B, BPFP=1.4646 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,204B, BPFP=0.5249 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,036B, BPFP=1.4277 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,564B, BPFP=0.6024 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,764B, BPFP=1.3552 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,812B, BPFP=0.5595 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,388B, BPFP=1.3907 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,040B, BPFP=0.9717 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,824B, BPFP=1.3016 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,584B, BPFP=0.1270 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16500157 155.69422901 + layer.0.v_cache 0.00001391 0.01309981 + layer.1.k_cache 0.57858722 16.48920257 + layer.1.v_cache 0.00000575 0.00492294 + layer.2.k_cache 0.01018238 1.75526038 + layer.2.v_cache 0.00001877 0.01396624 + layer.3.k_cache 0.03425542 8.12214961 + layer.3.v_cache 0.00002021 0.01582536 + layer.4.k_cache 0.00069807 0.36359536 + layer.4.v_cache 0.00005194 0.03429746 + layer.4.output 0.01023613 202.65077555 + ------------------------------------------------------------------------------------- + TOTAL 0.05061695 94.18011633 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 189504 +BPFP 0.6357 bits/point +EBPFP 1.2714 equivalent bits/point +MSE 94.180116 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.009s, Pack+Encode: 0.246s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 94.1801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,564B, BPFP=0.3220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,600B, BPFP=1.4815 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,320B, BPFP=0.5394 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,444B, BPFP=1.4146 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,440B, BPFP=0.6042 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,028B, BPFP=1.3326 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,048B, BPFP=0.5815 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,440B, BPFP=1.3565 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,700B, BPFP=0.9664 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,464B, BPFP=1.3000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,616B, BPFP=0.1374 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385483 158.70785590 + layer.0.v_cache 0.00001387 0.01293083 + layer.1.k_cache 0.55871571 16.26576787 + layer.1.v_cache 0.00000565 0.00471945 + layer.2.k_cache 0.00818202 1.68164741 + layer.2.v_cache 0.00001875 0.01367360 + layer.3.k_cache 0.04697755 7.38643392 + layer.3.v_cache 0.00001978 0.01566867 + layer.4.k_cache 0.00071636 0.36078946 + layer.4.v_cache 0.00005146 0.03357417 + layer.4.output 0.01106052 205.48900463 + ------------------------------------------------------------------------------------- + TOTAL 0.04741057 95.46506434 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 187664 +BPFP 0.6388 bits/point +EBPFP 1.2777 equivalent bits/point +MSE 95.465064 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 95.4651 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,492B, BPFP=0.3120 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,952B, BPFP=1.4745 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,260B, BPFP=0.5261 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,976B, BPFP=1.4191 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,636B, BPFP=0.6043 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,484B, BPFP=1.3343 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,948B, BPFP=0.5652 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,236B, BPFP=1.3770 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,056B, BPFP=0.9691 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,908B, BPFP=1.3016 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,452B, BPFP=0.1254 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203608 155.77301136 + layer.0.v_cache 0.00001420 0.01287216 + layer.1.k_cache 0.56673284 16.26996449 + layer.1.v_cache 0.00000552 0.00466709 + layer.2.k_cache 0.00530624 1.69303378 + layer.2.v_cache 0.00001825 0.01365769 + layer.3.k_cache 0.02188354 7.23614790 + layer.3.v_cache 0.00001854 0.01519499 + layer.4.k_cache 0.00069888 0.36414784 + layer.4.v_cache 0.00004918 0.03328869 + layer.4.output 0.00799176 201.86870130 + ------------------------------------------------------------------------------------- + TOTAL 0.04662974 93.79393501 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 189400 +BPFP 0.6330 bits/point +EBPFP 1.2660 equivalent bits/point +MSE 93.793935 +---------------------- -------------------------------------------------------- +Time: 0.646s Load: 0.009s, Pack+Encode: 0.253s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 93.7939 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,596B, BPFP=0.3090 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,060B, BPFP=1.3836 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,532B, BPFP=0.5263 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,656B, BPFP=1.3613 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,424B, BPFP=0.5755 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,748B, BPFP=1.3112 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,684B, BPFP=0.5347 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,868B, BPFP=1.3178 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,868B, BPFP=0.9313 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,128B, BPFP=1.2217 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,320B, BPFP=0.1287 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12732459 157.09880190 + layer.0.v_cache 0.00001480 0.01312252 + layer.1.k_cache 0.57322790 16.04288766 + layer.1.v_cache 0.00000557 0.00494612 + layer.2.k_cache 0.00950604 1.78333943 + layer.2.v_cache 0.00002039 0.01420216 + layer.3.k_cache 0.01510038 7.65377495 + layer.3.v_cache 0.00001916 0.01568389 + layer.4.k_cache 0.00068141 0.35727975 + layer.4.v_cache 0.00005165 0.03529569 + layer.4.output 0.00936729 196.18869889 + ------------------------------------------------------------------------------------- + TOTAL 0.04656017 91.54942508 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 187884 +BPFP 0.6102 bits/point +EBPFP 1.2204 equivalent bits/point +MSE 91.549425 +---------------------- -------------------------------------------------------- +Time: 0.646s Load: 0.009s, Pack+Encode: 0.253s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 91.5494 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,552B, BPFP=0.3143 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,960B, BPFP=1.4697 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,364B, BPFP=0.5301 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,120B, BPFP=1.4221 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,524B, BPFP=0.5958 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,876B, BPFP=1.3517 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,736B, BPFP=0.5512 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,588B, BPFP=1.3920 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,408B, BPFP=0.9855 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,000B, BPFP=1.3021 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,972B, BPFP=0.1292 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13630490 155.15003680 + layer.0.v_cache 0.00001544 0.01318002 + layer.1.k_cache 0.58776402 16.10805169 + layer.1.v_cache 0.00000559 0.00479117 + layer.2.k_cache 0.01000112 1.77259163 + layer.2.v_cache 0.00001940 0.01413149 + layer.3.k_cache 0.04424953 7.39001996 + layer.3.v_cache 0.00002139 0.01571931 + layer.4.k_cache 0.00071093 0.36910441 + layer.4.v_cache 0.00005158 0.03444153 + layer.4.output 0.01101471 201.00786102 + ------------------------------------------------------------------------------------- + TOTAL 0.05036746 93.40747619 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 191100 +BPFP 0.6364 bits/point +EBPFP 1.2728 equivalent bits/point +MSE 93.407476 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.009s, Pack+Encode: 0.254s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 93.4075 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,548B, BPFP=0.3118 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,128B, BPFP=1.4685 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,168B, BPFP=0.5153 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,268B, BPFP=1.4202 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,712B, BPFP=0.6021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,900B, BPFP=1.3433 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,876B, BPFP=0.5551 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,980B, BPFP=1.3478 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,432B, BPFP=0.9798 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,644B, BPFP=1.2727 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,064B, BPFP=0.1290 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12958164 153.26500674 + layer.0.v_cache 0.00001430 0.01326217 + layer.1.k_cache 0.54121377 16.15332734 + layer.1.v_cache 0.00000577 0.00481602 + layer.2.k_cache 0.01080061 1.66001190 + layer.2.v_cache 0.00001903 0.01395851 + layer.3.k_cache 0.01783211 7.50210857 + layer.3.v_cache 0.00002028 0.01537621 + layer.4.k_cache 0.00068935 0.35805437 + layer.4.v_cache 0.00005053 0.03404787 + layer.4.output 0.00833506 199.63836074 + ------------------------------------------------------------------------------------- + TOTAL 0.04462193 92.73461735 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 190720 +BPFP 0.6306 bits/point +EBPFP 1.2611 equivalent bits/point +MSE 92.734617 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.012s, Pack+Encode: 0.253s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 92.7346 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,484B, BPFP=0.3258 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,388B, BPFP=1.5083 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,300B, BPFP=0.5525 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,692B, BPFP=1.4670 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,296B, BPFP=0.6117 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,224B, BPFP=1.3798 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,664B, BPFP=0.5741 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,856B, BPFP=1.4173 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,120B, BPFP=1.0171 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,768B, BPFP=1.3527 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,692B, BPFP=0.1417 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12547956 153.38616920 + layer.0.v_cache 0.00001492 0.01338777 + layer.1.k_cache 0.50446328 15.99882107 + layer.1.v_cache 0.00000633 0.00478345 + layer.2.k_cache 0.00671430 1.73448291 + layer.2.v_cache 0.00001992 0.01393298 + layer.3.k_cache 0.02142675 6.51051244 + layer.3.v_cache 0.00001986 0.01564131 + layer.4.k_cache 0.00069859 0.36307940 + layer.4.v_cache 0.00005414 0.03435921 + layer.4.output 0.01038299 211.20119840 + ------------------------------------------------------------------------------------- + TOTAL 0.04303403 97.44020933 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 188484 +BPFP 0.6587 bits/point +EBPFP 1.3174 equivalent bits/point +MSE 97.440209 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 97.4402 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,620B, BPFP=0.2917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,120B, BPFP=1.3559 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,456B, BPFP=0.4909 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,200B, BPFP=1.3081 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,380B, BPFP=0.5388 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,072B, BPFP=1.2496 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,900B, BPFP=0.5139 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,472B, BPFP=1.2703 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,428B, BPFP=0.9047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,920B, BPFP=1.1898 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,716B, BPFP=0.1314 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13970643 155.58201827 + layer.0.v_cache 0.00001441 0.01317314 + layer.1.k_cache 0.60015088 16.03931556 + layer.1.v_cache 0.00000584 0.00481488 + layer.2.k_cache 0.00648033 1.62946631 + layer.2.v_cache 0.00001840 0.01399238 + layer.3.k_cache 0.03288257 7.12603739 + layer.3.v_cache 0.00002024 0.01550900 + layer.4.k_cache 0.00070287 0.35732338 + layer.4.v_cache 0.00005060 0.03376618 + layer.4.output 0.04874230 180.10988669 + ------------------------------------------------------------------------------------- + TOTAL 0.06595463 84.79909549 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 193284 +BPFP 0.5902 bits/point +EBPFP 1.1804 equivalent bits/point +MSE 84.799095 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 84.7991 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,548B, BPFP=0.3129 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,656B, BPFP=1.4648 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,084B, BPFP=0.5296 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,628B, BPFP=1.4157 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,984B, BPFP=0.5726 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,888B, BPFP=1.3326 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,456B, BPFP=0.5474 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,680B, BPFP=1.3704 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,340B, BPFP=0.9719 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,224B, BPFP=1.3008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,784B, BPFP=0.1282 +⌛️ [2/4] FRONTEND: Frontend time: 0.285s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.438s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13960611 155.49084958 + layer.0.v_cache 0.00001491 0.01349530 + layer.1.k_cache 0.68890666 16.02865778 + layer.1.v_cache 0.00000583 0.00488658 + layer.2.k_cache 0.00940153 1.71925153 + layer.2.v_cache 0.00001882 0.01391703 + layer.3.k_cache 0.04667800 7.16804864 + layer.3.v_cache 0.00001939 0.01595273 + layer.4.k_cache 0.00070827 0.36471567 + layer.4.v_cache 0.00005291 0.03333786 + layer.4.output 0.04427218 165.96537789 + ------------------------------------------------------------------------------------- + TOTAL 0.07031281 78.97710341 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 224272 +BPFP 0.6304 bits/point +EBPFP 1.2607 equivalent bits/point +MSE 78.977103 +---------------------- -------------------------------------------------------- +Time: 0.736s Load: 0.013s, Pack+Encode: 0.285s, Decode+Unpack: 0.438s +---------------------- -------------------------------------------------------- +💾 Converting with 78.9771 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,576B, BPFP=0.2914 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,084B, BPFP=1.3631 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,536B, BPFP=0.4983 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,112B, BPFP=1.3123 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,636B, BPFP=0.5558 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,100B, BPFP=1.2594 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,960B, BPFP=0.5205 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,420B, BPFP=1.2761 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,460B, BPFP=0.9124 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,980B, BPFP=1.2009 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,080B, BPFP=0.1350 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15817952 154.79648307 + layer.0.v_cache 0.00001474 0.01325843 + layer.1.k_cache 0.61609703 15.93105926 + layer.1.v_cache 0.00000566 0.00483639 + layer.2.k_cache 0.00844678 1.64775774 + layer.2.v_cache 0.00001876 0.01367517 + layer.3.k_cache 0.04682518 7.50210010 + layer.3.v_cache 0.00001896 0.01541997 + layer.4.k_cache 0.00069859 0.35717335 + layer.4.v_cache 0.00005216 0.03306436 + layer.4.output 0.04649196 181.42948220 + ------------------------------------------------------------------------------------- + TOTAL 0.06798830 85.31301196 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 193944 +BPFP 0.5962 bits/point +EBPFP 1.1924 equivalent bits/point +MSE 85.313012 +---------------------- -------------------------------------------------------- +Time: 0.653s Load: 0.011s, Pack+Encode: 0.257s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 85.3130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,500B, BPFP=0.2994 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,372B, BPFP=1.3813 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,164B, BPFP=0.4989 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,352B, BPFP=1.3258 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,372B, BPFP=0.5647 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,812B, BPFP=1.2964 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,676B, BPFP=0.5268 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,680B, BPFP=1.2892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,212B, BPFP=0.9371 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,092B, BPFP=1.2027 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,524B, BPFP=0.1285 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14002854 152.48174815 + layer.0.v_cache 0.00001601 0.01360862 + layer.1.k_cache 0.54217949 16.16116344 + layer.1.v_cache 0.00000579 0.00504269 + layer.2.k_cache 0.01082110 1.68545904 + layer.2.v_cache 0.00001971 0.01402009 + layer.3.k_cache 0.03102738 7.57867453 + layer.3.v_cache 0.00001907 0.01556497 + layer.4.k_cache 0.00066782 0.35989207 + layer.4.v_cache 0.00005141 0.03372012 + layer.4.output 0.00856273 193.29666501 + ------------------------------------------------------------------------------------- + TOTAL 0.04616326 90.08385581 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 187756 +BPFP 0.6013 bits/point +EBPFP 1.2026 equivalent bits/point +MSE 90.083856 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.386s +---------------------- -------------------------------------------------------- +💾 Converting with 90.0839 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,584B, BPFP=0.2938 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,136B, BPFP=1.3750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,548B, BPFP=0.5023 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,212B, BPFP=1.3264 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,548B, BPFP=0.5549 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,220B, BPFP=1.2742 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,912B, BPFP=0.5215 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,680B, BPFP=1.2984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,580B, BPFP=0.9249 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,912B, BPFP=1.2054 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,648B, BPFP=0.1326 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14705232 154.61109796 + layer.0.v_cache 0.00001374 0.01310585 + layer.1.k_cache 0.65666707 16.17733323 + layer.1.v_cache 0.00000579 0.00481980 + layer.2.k_cache 0.01062412 1.74997842 + layer.2.v_cache 0.00001918 0.01386772 + layer.3.k_cache 0.03205616 7.54422809 + layer.3.v_cache 0.00001999 0.01570770 + layer.4.k_cache 0.00071739 0.35387277 + layer.4.v_cache 0.00005111 0.03293932 + layer.4.output 0.04976816 182.72937710 + ------------------------------------------------------------------------------------- + TOTAL 0.07032965 85.86015239 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 193980 +BPFP 0.6003 bits/point +EBPFP 1.2006 equivalent bits/point +MSE 85.860152 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8602 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,528B, BPFP=0.3074 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,788B, BPFP=1.4339 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,476B, BPFP=0.5269 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,148B, BPFP=1.3984 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,568B, BPFP=0.5876 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,940B, BPFP=1.3312 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,892B, BPFP=0.5500 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,148B, BPFP=1.3427 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,132B, BPFP=0.9526 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,540B, BPFP=1.2533 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,360B, BPFP=0.1300 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.389s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131647 156.12216415 + layer.0.v_cache 0.00001392 0.01270161 + layer.1.k_cache 0.51967005 16.02189446 + layer.1.v_cache 0.00000557 0.00455278 + layer.2.k_cache 0.01013049 1.63316057 + layer.2.v_cache 0.00001908 0.01319005 + layer.3.k_cache 0.05186962 7.33321286 + layer.3.v_cache 0.00001891 0.01524533 + layer.4.k_cache 0.00069958 0.36038577 + layer.4.v_cache 0.00005017 0.03237304 + layer.4.output 0.01073789 197.56388218 + ------------------------------------------------------------------------------------- + TOTAL 0.04758583 92.02917976 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 190520 +BPFP 0.6232 bits/point +EBPFP 1.2463 equivalent bits/point +MSE 92.029180 +---------------------- -------------------------------------------------------- +Time: 0.658s Load: 0.012s, Pack+Encode: 0.258s, Decode+Unpack: 0.389s +---------------------- -------------------------------------------------------- +💾 Converting with 92.0292 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,344B, BPFP=0.2899 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,884B, BPFP=1.3500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,908B, BPFP=0.4833 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,300B, BPFP=1.3184 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,036B, BPFP=0.5445 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,140B, BPFP=1.2554 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,212B, BPFP=0.4998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,456B, BPFP=1.2726 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,012B, BPFP=0.9230 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,900B, BPFP=1.1882 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,580B, BPFP=0.1285 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15566496 155.42735460 + layer.0.v_cache 0.00001437 0.01276973 + layer.1.k_cache 0.62643353 16.31358676 + layer.1.v_cache 0.00000573 0.00468123 + layer.2.k_cache 0.01010078 1.70256212 + layer.2.v_cache 0.00001868 0.01375799 + layer.3.k_cache 0.06271097 8.00170305 + layer.3.v_cache 0.00001958 0.01544757 + layer.4.k_cache 0.00073784 0.36694007 + layer.4.v_cache 0.00005294 0.03286831 + layer.4.output 0.00677609 192.84852431 + ------------------------------------------------------------------------------------- + TOTAL 0.05312894 90.10772598 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 184772 +BPFP 0.5897 bits/point +EBPFP 1.1794 equivalent bits/point +MSE 90.107726 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.009s, Pack+Encode: 0.254s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 90.1077 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,488B, BPFP=0.2998 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,004B, BPFP=1.3660 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,148B, BPFP=0.4998 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,240B, BPFP=1.3243 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,304B, BPFP=0.5629 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,544B, BPFP=1.2863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,552B, BPFP=0.5219 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,500B, BPFP=1.2839 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,080B, BPFP=0.9331 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,028B, BPFP=1.2035 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,516B, BPFP=0.1289 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13918158 152.28112708 + layer.0.v_cache 0.00001566 0.01326837 + layer.1.k_cache 0.57997750 15.98907855 + layer.1.v_cache 0.00000571 0.00500652 + layer.2.k_cache 0.01637170 1.70605170 + layer.2.v_cache 0.00001887 0.01452990 + layer.3.k_cache 0.04393929 7.58277221 + layer.3.v_cache 0.00001903 0.01574363 + layer.4.k_cache 0.00071104 0.36788522 + layer.4.v_cache 0.00005175 0.03498502 + layer.4.output 0.01065216 193.90336226 + ------------------------------------------------------------------------------------- + TOTAL 0.05028572 90.31376377 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 186404 +BPFP 0.5990 bits/point +EBPFP 1.1981 equivalent bits/point +MSE 90.313764 +---------------------- -------------------------------------------------------- +Time: 0.646s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 90.3138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,640B, BPFP=0.2843 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,492B, BPFP=1.3353 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,532B, BPFP=0.4804 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,596B, BPFP=1.2901 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,212B, BPFP=0.5147 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,416B, BPFP=1.2306 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,916B, BPFP=0.4998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,920B, BPFP=1.2560 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,940B, BPFP=0.9042 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,316B, BPFP=1.1752 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,952B, BPFP=0.1221 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16496690 153.51020665 + layer.0.v_cache 0.00001433 0.01310856 + layer.1.k_cache 0.65265572 15.82238691 + layer.1.v_cache 0.00000578 0.00492805 + layer.2.k_cache 0.01241827 1.64368070 + layer.2.v_cache 0.00001920 0.01367248 + layer.3.k_cache 0.08444468 7.17747527 + layer.3.v_cache 0.00001975 0.01532389 + layer.4.k_cache 0.00068928 0.35784789 + layer.4.v_cache 0.00005259 0.03365938 + layer.4.output 0.04785494 174.82498560 + ------------------------------------------------------------------------------------- + TOTAL 0.07354536 82.49218759 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 194932 +BPFP 0.5780 bits/point +EBPFP 1.1559 equivalent bits/point +MSE 82.492188 +---------------------- -------------------------------------------------------- +Time: 0.646s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 82.4922 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,696B, BPFP=0.3027 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,876B, BPFP=1.3752 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,560B, BPFP=0.5081 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,896B, BPFP=1.3231 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,716B, BPFP=0.5695 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,004B, BPFP=1.2757 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,016B, BPFP=0.5323 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,396B, BPFP=1.2966 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,400B, BPFP=0.9247 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,724B, BPFP=1.2077 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,456B, BPFP=0.1325 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15001445 150.22645621 + layer.0.v_cache 0.00001441 0.01317620 + layer.1.k_cache 0.57475359 15.70008902 + layer.1.v_cache 0.00000584 0.00507043 + layer.2.k_cache 0.01434258 1.65322824 + layer.2.v_cache 0.00001954 0.01424388 + layer.3.k_cache 0.02649183 7.21783821 + layer.3.v_cache 0.00001945 0.01574684 + layer.4.k_cache 0.00069414 0.37181468 + layer.4.v_cache 0.00005306 0.03567013 + layer.4.output 0.05035525 184.57908163 + ------------------------------------------------------------------------------------- + TOTAL 0.06581739 86.31217090 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 192740 +BPFP 0.6026 bits/point +EBPFP 1.2051 equivalent bits/point +MSE 86.312171 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.013s, Pack+Encode: 0.255s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 86.3122 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,544B, BPFP=0.3072 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,476B, BPFP=1.4116 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,408B, BPFP=0.5213 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,940B, BPFP=1.3819 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,200B, BPFP=0.5652 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,824B, BPFP=1.3200 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,244B, BPFP=0.5122 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,892B, BPFP=1.3238 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,552B, BPFP=0.9725 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,216B, BPFP=1.2309 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,488B, BPFP=0.1226 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11854591 160.64936558 + layer.0.v_cache 0.00001566 0.01276293 + layer.1.k_cache 0.56985614 16.30182188 + layer.1.v_cache 0.00000555 0.00472401 + layer.2.k_cache 0.01267573 1.80342297 + layer.2.v_cache 0.00001902 0.01371491 + layer.3.k_cache 0.04477839 7.92480555 + layer.3.v_cache 0.00001896 0.01547685 + layer.4.k_cache 0.00071754 0.36357877 + layer.4.v_cache 0.00005321 0.03372700 + layer.4.output 0.00761333 196.87830864 + ------------------------------------------------------------------------------------- + TOTAL 0.04705761 92.07479770 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 187784 +BPFP 0.6120 bits/point +EBPFP 1.2241 equivalent bits/point +MSE 92.074798 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 92.0748 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,632B, BPFP=0.2993 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,264B, BPFP=1.3958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,628B, BPFP=0.5117 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,160B, BPFP=1.3372 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,640B, BPFP=0.5655 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,056B, BPFP=1.2785 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,888B, BPFP=0.5255 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,564B, BPFP=1.3055 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,948B, BPFP=0.9539 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,028B, BPFP=1.2239 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,328B, BPFP=0.1316 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582189 153.33237670 + layer.0.v_cache 0.00001509 0.01352437 + layer.1.k_cache 0.58311327 16.22673356 + layer.1.v_cache 0.00000606 0.00498951 + layer.2.k_cache 0.00657600 1.68579122 + layer.2.v_cache 0.00001909 0.01419088 + layer.3.k_cache 0.02234380 7.24604226 + layer.3.v_cache 0.00001905 0.01609774 + layer.4.k_cache 0.00071828 0.36664960 + layer.4.v_cache 0.00005370 0.03479577 + layer.4.output 0.05044473 184.61681548 + ------------------------------------------------------------------------------------- + TOTAL 0.06363526 86.54464117 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 194136 +BPFP 0.6069 bits/point +EBPFP 1.2138 equivalent bits/point +MSE 86.544641 +---------------------- -------------------------------------------------------- +Time: 0.649s Load: 0.011s, Pack+Encode: 0.259s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 86.5446 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 292, 128) +Output shape: (1, 292, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.output: torch.Size([1, 292, 3584]) -> torch.Size([1, 1, 292, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,700B, BPFP=0.3050 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,800B, BPFP=1.3806 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,520B, BPFP=0.5094 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,824B, BPFP=1.3283 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,972B, BPFP=0.5871 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,988B, BPFP=1.2836 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,112B, BPFP=0.5411 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,264B, BPFP=1.2984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,416B, BPFP=0.9319 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,496B, BPFP=1.2038 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,056B, BPFP=0.1304 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887579 153.43788795 + layer.0.v_cache 0.00001384 0.01323482 + layer.1.k_cache 0.57277983 16.32082085 + layer.1.v_cache 0.00000568 0.00486864 + layer.2.k_cache 0.01314942 1.77366931 + layer.2.v_cache 0.00002105 0.01399835 + layer.3.k_cache 0.03125376 7.88463687 + layer.3.v_cache 0.00001978 0.01545697 + layer.4.k_cache 0.00068035 0.36057399 + layer.4.v_cache 0.00005123 0.03452258 + layer.4.output 0.04859251 185.84177776 + ------------------------------------------------------------------------------------- + TOTAL 0.06511755 87.10306557 + (elements=2,541,568) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2541568 +Total Bytes 192148 +BPFP 0.6048 bits/point +EBPFP 1.2096 equivalent bits/point +MSE 87.103066 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.011s, Pack+Encode: 0.254s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 87.1031 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 338, 128) +Output shape: (1, 338, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.output: torch.Size([1, 338, 3584]) -> torch.Size([1, 1, 338, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,652B, BPFP=0.3075 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,660B, BPFP=1.4636 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,160B, BPFP=0.5159 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,068B, BPFP=1.3900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,732B, BPFP=0.5886 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,348B, BPFP=1.3105 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,720B, BPFP=0.5418 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,936B, BPFP=1.3376 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,184B, BPFP=0.9331 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,520B, BPFP=1.2722 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,644B, BPFP=0.1297 +⌛️ [2/4] FRONTEND: Frontend time: 0.279s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.427s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13976314 155.14848373 + layer.0.v_cache 0.00001472 0.01340832 + layer.1.k_cache 0.73750278 16.09745111 + layer.1.v_cache 0.00000589 0.00514589 + layer.2.k_cache 0.02011468 1.72205800 + layer.2.v_cache 0.00002041 0.01372775 + layer.3.k_cache 0.02722352 7.34026052 + layer.3.v_cache 0.00002029 0.01564679 + layer.4.k_cache 0.00072609 0.36996627 + layer.4.v_cache 0.00005477 0.03456385 + layer.4.output 0.04270584 160.52094780 + ------------------------------------------------------------------------------------- + TOTAL 0.07202277 76.72984393 + (elements=2,941,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2941952 +Total Bytes 228624 +BPFP 0.6217 bits/point +EBPFP 1.2434 equivalent bits/point +MSE 76.729844 +---------------------- -------------------------------------------------------- +Time: 0.718s Load: 0.012s, Pack+Encode: 0.279s, Decode+Unpack: 0.427s +---------------------- -------------------------------------------------------- +💾 Converting with 76.7298 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,596B, BPFP=0.3005 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,480B, BPFP=1.3681 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,536B, BPFP=0.5120 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,788B, BPFP=1.3310 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,612B, BPFP=0.5698 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,528B, BPFP=1.2633 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,840B, BPFP=0.5284 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,812B, BPFP=1.2786 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,352B, BPFP=0.9317 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,272B, BPFP=1.1959 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,028B, BPFP=0.1306 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150811 151.76123550 + layer.0.v_cache 0.00001374 0.01295881 + layer.1.k_cache 0.66021251 16.08122919 + layer.1.v_cache 0.00000574 0.00476048 + layer.2.k_cache 0.00963001 1.69914094 + layer.2.v_cache 0.00001988 0.01356697 + layer.3.k_cache 0.08055985 7.63056993 + layer.3.v_cache 0.00001942 0.01503590 + layer.4.k_cache 0.00069026 0.35711387 + layer.4.v_cache 0.00005126 0.03329136 + layer.4.output 0.05093874 186.45653841 + ------------------------------------------------------------------------------------- + TOTAL 0.07289894 87.22380423 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 189844 +BPFP 0.5996 bits/point +EBPFP 1.1992 equivalent bits/point +MSE 87.223804 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 87.2238 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,628B, BPFP=0.2902 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,040B, BPFP=1.3428 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,468B, BPFP=0.4882 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,976B, BPFP=1.2880 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,580B, BPFP=0.5456 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,880B, BPFP=1.2314 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,816B, BPFP=0.5062 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,420B, BPFP=1.2593 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,404B, BPFP=0.8975 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,928B, BPFP=1.1823 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,652B, BPFP=0.1300 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14893842 154.41557085 + layer.0.v_cache 0.00001415 0.01267624 + layer.1.k_cache 0.65110169 15.63065633 + layer.1.v_cache 0.00000575 0.00497838 + layer.2.k_cache 0.01411552 1.68883361 + layer.2.v_cache 0.00001904 0.01374035 + layer.3.k_cache 0.03939451 7.35758544 + layer.3.v_cache 0.00001876 0.01494355 + layer.4.k_cache 0.00070558 0.36136222 + layer.4.v_cache 0.00005345 0.03327341 + layer.4.output 0.04819407 179.02738979 + ------------------------------------------------------------------------------------- + TOTAL 0.07010149 84.27796170 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 192792 +BPFP 0.5848 bits/point +EBPFP 1.1696 equivalent bits/point +MSE 84.277962 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 84.2780 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,552B, BPFP=0.3298 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,112B, BPFP=1.4919 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,272B, BPFP=0.5509 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,484B, BPFP=1.4546 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,060B, BPFP=0.5977 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,104B, BPFP=1.3726 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,612B, BPFP=0.5711 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,600B, BPFP=1.4021 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,968B, BPFP=1.0081 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,500B, BPFP=1.3367 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,460B, BPFP=0.1312 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14492303 153.06276735 + layer.0.v_cache 0.00001375 0.01280449 + layer.1.k_cache 0.50395557 16.24783151 + layer.1.v_cache 0.00000564 0.00478318 + layer.2.k_cache 0.00856250 1.72467436 + layer.2.v_cache 0.00001902 0.01357610 + layer.3.k_cache 0.02692018 7.55459931 + layer.3.v_cache 0.00002118 0.01536261 + layer.4.k_cache 0.00070790 0.35399935 + layer.4.v_cache 0.00005130 0.03271734 + layer.4.output 0.01131245 211.18928232 + ------------------------------------------------------------------------------------- + TOTAL 0.04496278 97.49106423 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 185724 +BPFP 0.6491 bits/point +EBPFP 1.2981 equivalent bits/point +MSE 97.491064 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 97.4911 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,560B, BPFP=0.3148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,872B, BPFP=1.4647 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,304B, BPFP=0.5267 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,344B, BPFP=1.4348 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,800B, BPFP=0.6114 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,788B, BPFP=1.3467 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,900B, BPFP=0.5605 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,560B, BPFP=1.3904 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,088B, BPFP=0.9674 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,604B, BPFP=1.2797 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,036B, BPFP=0.1297 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14113761 155.31477865 + layer.0.v_cache 0.00001369 0.01319231 + layer.1.k_cache 0.59212339 16.09507154 + layer.1.v_cache 0.00000552 0.00470380 + layer.2.k_cache 0.01129283 1.78785728 + layer.2.v_cache 0.00001853 0.01396363 + layer.3.k_cache 0.01358872 7.68962915 + layer.3.v_cache 0.00001883 0.01539018 + layer.4.k_cache 0.00070763 0.36875987 + layer.4.v_cache 0.00005153 0.03394083 + layer.4.output 0.01019159 200.97533320 + ------------------------------------------------------------------------------------- + TOTAL 0.04884114 93.42144822 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 190856 +BPFP 0.6356 bits/point +EBPFP 1.2712 equivalent bits/point +MSE 93.421448 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.009s, Pack+Encode: 0.246s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 93.4214 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,476B, BPFP=0.2981 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,272B, BPFP=1.3759 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,140B, BPFP=0.4976 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,220B, BPFP=1.3186 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,492B, BPFP=0.5712 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,128B, BPFP=1.2591 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,808B, BPFP=0.5340 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,484B, BPFP=1.2785 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,292B, BPFP=0.9414 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,264B, BPFP=1.2121 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,812B, BPFP=0.1308 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11028925 152.61748693 + layer.0.v_cache 0.00001466 0.01288776 + layer.1.k_cache 0.56032453 15.98243889 + layer.1.v_cache 0.00000595 0.00468825 + layer.2.k_cache 0.00809042 1.66205674 + layer.2.v_cache 0.00001929 0.01369636 + layer.3.k_cache 0.03501172 7.79515240 + layer.3.v_cache 0.00001861 0.01531357 + layer.4.k_cache 0.00070074 0.35567185 + layer.4.v_cache 0.00005246 0.03317450 + layer.4.output 0.00889761 193.49454019 + ------------------------------------------------------------------------------------- + TOTAL 0.04569476 90.17378521 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 187388 +BPFP 0.6001 bits/point +EBPFP 1.2002 equivalent bits/point +MSE 90.173785 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 90.1738 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 311, 128) +Output shape: (1, 311, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.output: torch.Size([1, 311, 3584]) -> torch.Size([1, 1, 311, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,704B, BPFP=0.2866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,392B, BPFP=1.3260 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,528B, BPFP=0.4787 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,640B, BPFP=1.2882 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,300B, BPFP=0.5175 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,476B, BPFP=1.2297 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,880B, BPFP=0.4964 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,032B, BPFP=1.2576 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,936B, BPFP=0.9011 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,312B, BPFP=1.1712 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,500B, BPFP=0.1256 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14227538 153.68993670 + layer.0.v_cache 0.00001500 0.01315835 + layer.1.k_cache 0.64922478 15.89727881 + layer.1.v_cache 0.00000572 0.00490276 + layer.2.k_cache 0.00828172 1.72297026 + layer.2.v_cache 0.00001922 0.01358763 + layer.3.k_cache 0.02636586 7.38884301 + layer.3.v_cache 0.00001996 0.01568403 + layer.4.k_cache 0.00069494 0.35651984 + layer.4.v_cache 0.00005248 0.03370663 + layer.4.output 0.04686997 174.38911059 + ------------------------------------------------------------------------------------- + TOTAL 0.06794382 82.34472718 + (elements=2,706,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2706944 +Total Bytes 195700 +BPFP 0.5784 bits/point +EBPFP 1.1567 equivalent bits/point +MSE 82.344727 +---------------------- -------------------------------------------------------- +Time: 0.627s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 82.3447 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,364B, BPFP=0.2674 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,280B, BPFP=1.3039 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,448B, BPFP=0.4564 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,576B, BPFP=1.2608 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,184B, BPFP=0.5015 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,796B, BPFP=1.2130 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,376B, BPFP=0.4520 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,148B, BPFP=1.2346 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,492B, BPFP=0.8880 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,052B, BPFP=1.1674 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,208B, BPFP=0.1244 +⌛️ [2/4] FRONTEND: Frontend time: 0.282s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13149661 149.89509804 + layer.0.v_cache 0.00001370 0.01286549 + layer.1.k_cache 0.50214604 15.50008425 + layer.1.v_cache 0.00000625 0.00489403 + layer.2.k_cache 0.00807812 1.70003375 + layer.2.v_cache 0.00001969 0.01411979 + layer.3.k_cache 0.06061829 7.48211455 + layer.3.v_cache 0.00001868 0.01520756 + layer.4.k_cache 0.00068906 0.35154679 + layer.4.v_cache 0.00005044 0.03335067 + layer.4.output 1.20679682 211.99725140 + ------------------------------------------------------------------------------------- + TOTAL 0.53827733 97.58765145 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 156924 +BPFP 0.5656 bits/point +EBPFP 1.1312 equivalent bits/point +MSE 97.587651 +---------------------- -------------------------------------------------------- +Time: 0.611s Load: 0.011s, Pack+Encode: 0.282s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 97.5877 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 300, 128) +Output shape: (1, 300, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.output: torch.Size([1, 300, 3584]) -> torch.Size([1, 1, 300, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,616B, BPFP=0.2925 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,052B, BPFP=1.3569 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,464B, BPFP=0.4929 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,108B, BPFP=1.3077 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,600B, BPFP=0.5521 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,144B, BPFP=1.2575 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,076B, BPFP=0.5248 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,688B, BPFP=1.2858 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,576B, BPFP=0.9154 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,972B, BPFP=1.1965 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,036B, BPFP=0.1268 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13491969 156.69572917 + layer.0.v_cache 0.00001558 0.01281772 + layer.1.k_cache 0.64590566 16.09480469 + layer.1.v_cache 0.00000556 0.00463096 + layer.2.k_cache 0.01395306 1.72878947 + layer.2.v_cache 0.00001972 0.01379983 + layer.3.k_cache 0.02379198 7.80010661 + layer.3.v_cache 0.00001942 0.01563480 + layer.4.k_cache 0.00073123 0.36996493 + layer.4.v_cache 0.00005274 0.03346053 + layer.4.output 0.05011183 180.58733631 + ------------------------------------------------------------------------------------- + TOTAL 0.06883514 85.11065252 + (elements=2,611,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2611200 +Total Bytes 193332 +BPFP 0.5923 bits/point +EBPFP 1.1846 equivalent bits/point +MSE 85.110653 +---------------------- -------------------------------------------------------- +Time: 0.630s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1107 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,588B, BPFP=0.3000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,344B, BPFP=1.3608 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,468B, BPFP=0.5084 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,452B, BPFP=1.3129 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,652B, BPFP=0.5720 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,176B, BPFP=1.2444 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,676B, BPFP=0.5195 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,468B, BPFP=1.2601 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,956B, BPFP=0.9104 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,912B, BPFP=1.1765 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,204B, BPFP=0.1320 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14826349 153.83655498 + layer.0.v_cache 0.00001415 0.01311804 + layer.1.k_cache 0.61440683 16.02113704 + layer.1.v_cache 0.00000562 0.00504580 + layer.2.k_cache 0.00551110 1.74768821 + layer.2.v_cache 0.00001912 0.01379523 + layer.3.k_cache 0.01294636 6.96179703 + layer.3.v_cache 0.00002032 0.01534602 + layer.4.k_cache 0.00070633 0.37237182 + layer.4.v_cache 0.00005074 0.03434545 + layer.4.output 0.04972342 186.48982879 + ------------------------------------------------------------------------------------- + TOTAL 0.06647106 87.32058830 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 187896 +BPFP 0.5935 bits/point +EBPFP 1.1869 equivalent bits/point +MSE 87.320588 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 87.3206 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,508B, BPFP=0.3107 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,836B, BPFP=1.4574 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,168B, BPFP=0.5171 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,292B, BPFP=1.4267 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,668B, BPFP=0.6018 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,756B, BPFP=1.3400 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,772B, BPFP=0.5512 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,400B, BPFP=1.3764 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,160B, BPFP=0.9680 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,824B, BPFP=1.2875 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,992B, BPFP=0.1289 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13861307 153.09290388 + layer.0.v_cache 0.00001419 0.01299402 + layer.1.k_cache 0.53867822 16.03493767 + layer.1.v_cache 0.00000573 0.00471351 + layer.2.k_cache 0.00871976 1.72912532 + layer.2.v_cache 0.00001877 0.01339035 + layer.3.k_cache 0.02195022 7.72268181 + layer.3.v_cache 0.00001967 0.01579426 + layer.4.k_cache 0.00069227 0.35764552 + layer.4.v_cache 0.00004961 0.03277595 + layer.4.output 0.00954030 200.43601728 + ------------------------------------------------------------------------------------- + TOTAL 0.04562021 93.06288725 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 190376 +BPFP 0.6317 bits/point +EBPFP 1.2634 equivalent bits/point +MSE 93.062887 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 93.0629 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,540B, BPFP=0.3027 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,916B, BPFP=1.3612 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,188B, BPFP=0.5020 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,164B, BPFP=1.3201 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,284B, BPFP=0.5618 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,736B, BPFP=1.2968 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,548B, BPFP=0.5216 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,040B, BPFP=1.3134 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,344B, BPFP=0.9476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,116B, BPFP=1.2083 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,540B, BPFP=0.1291 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13928204 154.63628169 + layer.0.v_cache 0.00001373 0.01267323 + layer.1.k_cache 0.60860016 16.16113623 + layer.1.v_cache 0.00000555 0.00475548 + layer.2.k_cache 0.01026906 1.74126237 + layer.2.v_cache 0.00001872 0.01383718 + layer.3.k_cache 0.05166252 7.98082045 + layer.3.v_cache 0.00001937 0.01553928 + layer.4.k_cache 0.00073448 0.36103093 + layer.4.v_cache 0.00005026 0.03329067 + layer.4.output 0.00697322 194.06999251 + ------------------------------------------------------------------------------------- + TOTAL 0.05055697 90.55591618 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 187416 +BPFP 0.6023 bits/point +EBPFP 1.2046 equivalent bits/point +MSE 90.555916 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 90.5559 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,664B, BPFP=0.3000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,052B, BPFP=1.3799 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,692B, BPFP=0.5133 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,944B, BPFP=1.3212 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,512B, BPFP=0.5568 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,876B, BPFP=1.2646 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,832B, BPFP=0.5208 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,388B, BPFP=1.2917 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,652B, BPFP=0.9350 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,960B, BPFP=1.2161 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,144B, BPFP=0.1297 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13820392 153.22635064 + layer.0.v_cache 0.00001454 0.01286100 + layer.1.k_cache 0.61684219 15.61012646 + layer.1.v_cache 0.00000570 0.00477728 + layer.2.k_cache 0.01285210 1.70641055 + layer.2.v_cache 0.00002011 0.01373858 + layer.3.k_cache 0.05980752 7.94906896 + layer.3.v_cache 0.00002263 0.01549916 + layer.4.k_cache 0.00071023 0.36693498 + layer.4.v_cache 0.00005833 0.03366782 + layer.4.output 0.05086143 183.95009080 + ------------------------------------------------------------------------------------- + TOTAL 0.06968043 86.27000418 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 192716 +BPFP 0.6004 bits/point +EBPFP 1.2009 equivalent bits/point +MSE 86.270004 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2700 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,612B, BPFP=0.2913 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,360B, BPFP=1.3684 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,416B, BPFP=0.4888 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,368B, BPFP=1.3169 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,488B, BPFP=0.5444 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,244B, BPFP=1.2585 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,948B, BPFP=0.5164 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,848B, BPFP=1.2899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,700B, BPFP=0.9188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,276B, BPFP=1.2083 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,280B, BPFP=0.1281 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14825445 156.52557880 + layer.0.v_cache 0.00001396 0.01377848 + layer.1.k_cache 0.63210902 16.04133844 + layer.1.v_cache 0.00000581 0.00521381 + layer.2.k_cache 0.01184132 1.71783244 + layer.2.v_cache 0.00001890 0.01447581 + layer.3.k_cache 0.01767834 7.75498501 + layer.3.v_cache 0.00002009 0.01617003 + layer.4.k_cache 0.00068272 0.37199006 + layer.4.v_cache 0.00005247 0.03506118 + layer.4.output 0.04867053 180.28157629 + ------------------------------------------------------------------------------------- + TOTAL 0.06772769 84.96867401 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 194540 +BPFP 0.5940 bits/point +EBPFP 1.1881 equivalent bits/point +MSE 84.968674 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.012s, Pack+Encode: 0.253s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9687 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,608B, BPFP=0.3319 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,316B, BPFP=1.4983 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,228B, BPFP=0.5462 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,576B, BPFP=1.4545 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,020B, BPFP=0.5930 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,148B, BPFP=1.3700 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,560B, BPFP=0.5658 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,680B, BPFP=1.4015 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,160B, BPFP=1.0156 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,512B, BPFP=1.3324 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,460B, BPFP=0.1307 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14315909 150.44435073 + layer.0.v_cache 0.00001391 0.01329760 + layer.1.k_cache 0.49711875 16.32671194 + layer.1.v_cache 0.00000560 0.00497425 + layer.2.k_cache 0.00646762 1.73545537 + layer.2.v_cache 0.00001905 0.01392954 + layer.3.k_cache 0.04707547 7.82546812 + layer.3.v_cache 0.00001924 0.01566292 + layer.4.k_cache 0.00071833 0.36639410 + layer.4.v_cache 0.00005197 0.03254978 + layer.4.output 0.00997522 210.37987013 + ------------------------------------------------------------------------------------- + TOTAL 0.04496915 97.02575796 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 186268 +BPFP 0.6485 bits/point +EBPFP 1.2970 equivalent bits/point +MSE 97.025758 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.009s, Pack+Encode: 0.250s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 97.0258 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,484B, BPFP=0.3071 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,996B, BPFP=1.4559 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,372B, BPFP=0.5249 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,032B, BPFP=1.4019 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,860B, BPFP=0.6082 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,820B, BPFP=1.3340 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,860B, BPFP=0.5522 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,964B, BPFP=1.3421 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,136B, BPFP=0.9597 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,660B, BPFP=1.2690 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,856B, BPFP=0.1269 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15581964 153.40284778 + layer.0.v_cache 0.00001703 0.01304547 + layer.1.k_cache 0.56298232 16.28334663 + layer.1.v_cache 0.00000550 0.00480352 + layer.2.k_cache 0.00979712 1.74467802 + layer.2.v_cache 0.00002023 0.01394437 + layer.3.k_cache 0.04960944 7.20324007 + layer.3.v_cache 0.00001971 0.01546927 + layer.4.k_cache 0.00068561 0.35585435 + layer.4.v_cache 0.00005245 0.03349645 + layer.4.output 0.01122868 198.99972798 + ------------------------------------------------------------------------------------- + TOTAL 0.05044764 92.47463658 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 190040 +BPFP 0.6261 bits/point +EBPFP 1.2521 equivalent bits/point +MSE 92.474637 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 92.4746 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,568B, BPFP=0.3074 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,168B, BPFP=1.3896 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,580B, BPFP=0.5289 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,556B, BPFP=1.3558 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,572B, BPFP=0.5837 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,644B, BPFP=1.3054 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,772B, BPFP=0.5395 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,516B, BPFP=1.2984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,204B, BPFP=0.9499 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,040B, BPFP=1.2169 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,652B, BPFP=0.1313 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13599485 156.20751436 + layer.0.v_cache 0.00001418 0.01295148 + layer.1.k_cache 0.56373192 16.12943594 + layer.1.v_cache 0.00000588 0.00489258 + layer.2.k_cache 0.01549433 1.70687834 + layer.2.v_cache 0.00001867 0.01379682 + layer.3.k_cache 0.03475572 7.09189867 + layer.3.v_cache 0.00001907 0.01527275 + layer.4.k_cache 0.00072111 0.36375530 + layer.4.v_cache 0.00005295 0.03282190 + layer.4.output 0.00820773 196.16924849 + ------------------------------------------------------------------------------------- + TOTAL 0.04754487 91.45670338 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 188272 +BPFP 0.6115 bits/point +EBPFP 1.2229 equivalent bits/point +MSE 91.456703 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.009s, Pack+Encode: 0.246s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4567 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,620B, BPFP=0.2851 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,216B, BPFP=1.3300 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,524B, BPFP=0.4832 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,312B, BPFP=1.2841 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,424B, BPFP=0.5288 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,408B, BPFP=1.2382 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,776B, BPFP=0.4959 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,024B, BPFP=1.2695 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,792B, BPFP=0.9026 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,336B, BPFP=1.1838 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,848B, BPFP=0.1221 +⌛️ [2/4] FRONTEND: Frontend time: 0.264s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13331154 152.55339387 + layer.0.v_cache 0.00001460 0.01312039 + layer.1.k_cache 0.64535681 16.03940176 + layer.1.v_cache 0.00000563 0.00483881 + layer.2.k_cache 0.00762887 1.74689077 + layer.2.v_cache 0.00001910 0.01442533 + layer.3.k_cache 0.03538683 7.56513799 + layer.3.v_cache 0.00002103 0.01602239 + layer.4.k_cache 0.00072257 0.36505011 + layer.4.v_cache 0.00005300 0.03474462 + layer.4.output 0.04522907 175.96499594 + ------------------------------------------------------------------------------------- + TOTAL 0.06700726 82.94752927 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 194280 +BPFP 0.5798 bits/point +EBPFP 1.1595 equivalent bits/point +MSE 82.947529 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.012s, Pack+Encode: 0.264s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 82.9475 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,588B, BPFP=0.3074 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,120B, BPFP=1.3820 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,640B, BPFP=0.5304 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,448B, BPFP=1.3451 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,680B, BPFP=0.5876 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,428B, BPFP=1.2890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,040B, BPFP=0.5524 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,664B, BPFP=1.3019 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,592B, BPFP=0.9679 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,252B, BPFP=1.2243 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,024B, BPFP=0.1338 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14806804 157.03078235 + layer.0.v_cache 0.00001376 0.01275110 + layer.1.k_cache 0.60662186 15.80298367 + layer.1.v_cache 0.00000571 0.00472251 + layer.2.k_cache 0.01239752 1.64423155 + layer.2.v_cache 0.00002027 0.01364153 + layer.3.k_cache 0.02456893 7.37511777 + layer.3.v_cache 0.00001975 0.01507873 + layer.4.k_cache 0.00069718 0.36569904 + layer.4.v_cache 0.00005379 0.03312507 + layer.4.output 0.00818182 195.49388519 + ------------------------------------------------------------------------------------- + TOTAL 0.04998468 91.22090174 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 189476 +BPFP 0.6132 bits/point +EBPFP 1.2264 equivalent bits/point +MSE 91.220902 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 91.2209 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,564B, BPFP=0.3072 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,460B, BPFP=1.4057 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,596B, BPFP=0.5298 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,856B, BPFP=1.3723 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,496B, BPFP=0.5795 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,552B, BPFP=1.3004 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,664B, BPFP=0.5336 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,924B, BPFP=1.3209 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,392B, BPFP=0.9602 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,180B, BPFP=1.2246 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,564B, BPFP=0.1306 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14592592 155.01470020 + layer.0.v_cache 0.00001391 0.01304558 + layer.1.k_cache 0.53254910 16.06143199 + layer.1.v_cache 0.00000583 0.00503074 + layer.2.k_cache 0.01029040 1.75681372 + layer.2.v_cache 0.00002009 0.01430684 + layer.3.k_cache 0.03848921 7.47352584 + layer.3.v_cache 0.00001972 0.01595291 + layer.4.k_cache 0.00069417 0.36680978 + layer.4.v_cache 0.00005427 0.03470931 + layer.4.output 0.01004909 196.21471164 + ------------------------------------------------------------------------------------- + TOTAL 0.04696507 91.42701814 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 189248 +BPFP 0.6146 bits/point +EBPFP 1.2293 equivalent bits/point +MSE 91.427018 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.012s, Pack+Encode: 0.249s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4270 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,676B, BPFP=0.3261 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,996B, BPFP=1.4933 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,388B, BPFP=0.5393 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,852B, BPFP=1.4276 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,420B, BPFP=0.5986 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,484B, BPFP=1.3490 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,020B, BPFP=0.5756 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,836B, BPFP=1.3693 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,132B, BPFP=0.9841 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,900B, BPFP=1.3155 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,020B, BPFP=0.1315 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15384994 154.53077608 + layer.0.v_cache 0.00001404 0.01341570 + layer.1.k_cache 0.48768111 16.22193460 + layer.1.v_cache 0.00000584 0.00519424 + layer.2.k_cache 0.00538594 1.70833509 + layer.2.v_cache 0.00001948 0.01393737 + layer.3.k_cache 0.02347840 7.56764939 + layer.3.v_cache 0.00001899 0.01533743 + layer.4.k_cache 0.00068094 0.36354287 + layer.4.v_cache 0.00005116 0.03369251 + layer.4.output 0.00909345 204.13256631 + ------------------------------------------------------------------------------------- + TOTAL 0.04322588 94.67069291 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 189724 +BPFP 0.6411 bits/point +EBPFP 1.2822 equivalent bits/point +MSE 94.670693 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 94.6707 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,616B, BPFP=0.2985 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,848B, BPFP=1.3737 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,588B, BPFP=0.5096 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,228B, BPFP=1.3408 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,736B, BPFP=0.5706 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,144B, BPFP=1.2832 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,700B, BPFP=0.5155 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,400B, BPFP=1.2968 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,900B, BPFP=0.9513 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,832B, BPFP=1.2134 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,916B, BPFP=0.1284 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13453574 151.65974437 + layer.0.v_cache 0.00001386 0.01293213 + layer.1.k_cache 0.70462929 16.27884181 + layer.1.v_cache 0.00000580 0.00489770 + layer.2.k_cache 0.00896267 1.69096302 + layer.2.v_cache 0.00001881 0.01372627 + layer.3.k_cache 0.02468209 7.57675482 + layer.3.v_cache 0.00001940 0.01534781 + layer.4.k_cache 0.00068809 0.35756232 + layer.4.v_cache 0.00005209 0.03275209 + layer.4.output 0.04953035 184.59689322 + ------------------------------------------------------------------------------------- + TOTAL 0.07178355 86.46010440 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 192908 +BPFP 0.6031 bits/point +EBPFP 1.2062 equivalent bits/point +MSE 86.460104 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.012s, Pack+Encode: 0.250s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 86.4601 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,548B, BPFP=0.3235 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,524B, BPFP=1.4881 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,144B, BPFP=0.5331 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,540B, BPFP=1.4307 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,340B, BPFP=0.6028 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,148B, BPFP=1.3496 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,044B, BPFP=0.5856 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,612B, BPFP=1.3766 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,984B, BPFP=0.9902 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,760B, BPFP=1.3270 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,612B, BPFP=0.1300 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150386 160.27763526 + layer.0.v_cache 0.00001401 0.01334321 + layer.1.k_cache 0.57442873 16.44901287 + layer.1.v_cache 0.00000554 0.00480971 + layer.2.k_cache 0.00779028 1.78775355 + layer.2.v_cache 0.00001875 0.01392221 + layer.3.k_cache 0.05176373 8.03369687 + layer.3.v_cache 0.00002027 0.01540454 + layer.4.k_cache 0.00069872 0.36387355 + layer.4.v_cache 0.00005023 0.03349359 + layer.4.output 0.01044900 207.03884595 + ------------------------------------------------------------------------------------- + TOTAL 0.04937866 96.25087453 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 187256 +BPFP 0.6422 bits/point +EBPFP 1.2844 equivalent bits/point +MSE 96.250875 +---------------------- -------------------------------------------------------- +Time: 0.629s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 96.2509 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.6139 bits/point +Avg EBPFP 1.2278 equivalent bits/point +Avg MSE 89.968587 +Avg Time 0.648s +------------------------ ---------------------------- diff --git a/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..2db9015e6b6f2e693d958a36b3714af8e50e52f6 --- /dev/null +++ b/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 559 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.004_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa +Output output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,868B, BPFP=0.3475 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,984B, BPFP=1.6711 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,340B, BPFP=0.6213 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,532B, BPFP=1.5871 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,740B, BPFP=0.6957 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,568B, BPFP=1.4077 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,472B, BPFP=0.6458 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,796B, BPFP=1.4501 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,136B, BPFP=1.1414 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,792B, BPFP=1.4494 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,884B, BPFP=0.1829 +⌛️ [2/4] FRONTEND: Frontend time: 0.394s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.247s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14073507 164.20860073 + layer.0.v_cache 0.00001426 0.01360620 + layer.1.k_cache 0.05042761 16.81713577 + layer.1.v_cache 0.00000517 0.00450841 + layer.2.k_cache 0.00244098 1.99958838 + layer.2.v_cache 0.00001715 0.01286712 + layer.3.k_cache 0.02726505 8.36306036 + layer.3.v_cache 0.00001796 0.01557918 + layer.4.k_cache 0.00069123 0.37030670 + layer.4.v_cache 0.00005084 0.03592197 + layer.4.output 0.16186065 645.22401148 + ------------------------------------------------------------------------------------- + TOTAL 0.07968764 276.96525030 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 66112 +BPFP 0.7234 bits/point +EBPFP 1.4468 equivalent bits/point +MSE 276.965250 +---------------------- -------------------------------------------------------- +Time: 0.646s Load: 0.005s, Pack+Encode: 0.394s, Decode+Unpack: 0.247s +---------------------- -------------------------------------------------------- +💾 Converting with 276.9653 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,852B, BPFP=0.3486 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,272B, BPFP=1.5572 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,312B, BPFP=0.6235 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,616B, BPFP=1.4337 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,528B, BPFP=0.6642 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,768B, BPFP=1.2741 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,392B, BPFP=0.6386 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,200B, BPFP=1.3554 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,648B, BPFP=1.0633 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,248B, BPFP=1.3645 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,240B, BPFP=0.1947 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11392649 160.16392131 + layer.0.v_cache 0.00001670 0.01353160 + layer.1.k_cache 0.05249626 16.52404638 + layer.1.v_cache 0.00000506 0.00452395 + layer.2.k_cache 0.00416583 1.88698936 + layer.2.v_cache 0.00001711 0.01282466 + layer.3.k_cache 0.05931453 8.04831631 + layer.3.v_cache 0.00001789 0.01507690 + layer.4.k_cache 0.00069248 0.35952488 + layer.4.v_cache 0.00004729 0.03348895 + layer.4.output 0.16383441 652.91571644 + ------------------------------------------------------------------------------------- + TOTAL 0.08103179 279.85130937 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 62076 +BPFP 0.6874 bits/point +EBPFP 1.3748 equivalent bits/point +MSE 279.851309 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.005s, Pack+Encode: 0.166s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 279.8513 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,892B, BPFP=0.3145 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,572B, BPFP=1.4249 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,872B, BPFP=0.6436 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,204B, BPFP=1.3637 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,552B, BPFP=0.7566 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,800B, BPFP=1.2965 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,136B, BPFP=0.6875 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,400B, BPFP=1.3963 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,252B, BPFP=1.2055 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,832B, BPFP=1.3019 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,148B, BPFP=0.1697 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17268721 163.71072972 + layer.0.v_cache 0.00001453 0.01378890 + layer.1.k_cache 0.08276152 16.75098825 + layer.1.v_cache 0.00000545 0.00468879 + layer.2.k_cache 0.00739249 1.71739862 + layer.2.v_cache 0.00001625 0.01247598 + layer.3.k_cache 0.05929063 9.25258815 + layer.3.v_cache 0.00001766 0.01454117 + layer.4.k_cache 0.00068371 0.36046519 + layer.4.v_cache 0.00005055 0.03256829 + layer.4.output 0.14466690 576.72691869 + ------------------------------------------------------------------------------------- + TOTAL 0.07856402 248.76227435 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 69660 +BPFP 0.6811 bits/point +EBPFP 1.3622 equivalent bits/point +MSE 248.762274 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 248.7623 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 112, 128) +Output shape: (1, 112, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.output: torch.Size([1, 112, 3584]) -> torch.Size([1, 1, 112, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,156B, BPFP=0.3008 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,868B, BPFP=1.3767 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,132B, BPFP=0.5765 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,360B, BPFP=1.3058 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,332B, BPFP=0.6044 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,744B, BPFP=1.2199 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,376B, BPFP=0.6105 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,184B, BPFP=1.2812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,632B, BPFP=0.9252 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,388B, BPFP=1.1702 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,408B, BPFP=0.1676 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17321212 149.13138253 + layer.0.v_cache 0.00001450 0.01316277 + layer.1.k_cache 0.08643453 16.02306693 + layer.1.v_cache 0.00000659 0.00484644 + layer.2.k_cache 0.00690480 1.71152469 + layer.2.v_cache 0.00001899 0.01304702 + layer.3.k_cache 0.01205242 7.74182728 + layer.3.v_cache 0.00001923 0.01536943 + layer.4.k_cache 0.00071999 0.35228051 + layer.4.v_cache 0.00004947 0.03163518 + layer.4.output 10.17892331 479.45065370 + ------------------------------------------------------------------------------------- + TOTAL 4.20775858 207.71721875 + (elements=974,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 974848 +Total Bytes 75580 +BPFP 0.6202 bits/point +EBPFP 1.2405 equivalent bits/point +MSE 207.717219 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.005s, Pack+Encode: 0.166s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 207.7172 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,968B, BPFP=0.3494 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,048B, BPFP=1.6065 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,436B, BPFP=0.6101 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,384B, BPFP=1.4886 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,780B, BPFP=0.6712 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,544B, BPFP=1.3395 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,404B, BPFP=0.6044 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,968B, BPFP=1.4148 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,168B, BPFP=1.0952 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,840B, BPFP=1.3920 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,808B, BPFP=0.1727 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11687154 165.50639205 + layer.0.v_cache 0.00001423 0.01323008 + layer.1.k_cache 0.05324238 16.40700462 + layer.1.v_cache 0.00000539 0.00475035 + layer.2.k_cache 0.00397935 1.60112901 + layer.2.v_cache 0.00001766 0.01235349 + layer.3.k_cache 0.01210797 7.88500075 + layer.3.v_cache 0.00001803 0.01567762 + layer.4.k_cache 0.00075037 0.36546803 + layer.4.v_cache 0.00004481 0.03140146 + layer.4.output 0.15448949 616.16040990 + ------------------------------------------------------------------------------------- + TOTAL 0.07461636 264.99795746 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 66348 +BPFP 0.6930 bits/point +EBPFP 1.3859 equivalent bits/point +MSE 264.997957 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.004s, Pack+Encode: 0.168s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 264.9980 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3472 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,912B, BPFP=1.7191 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,292B, BPFP=0.6350 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,312B, BPFP=1.6034 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,700B, BPFP=0.7137 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,436B, BPFP=1.4344 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,352B, BPFP=0.6466 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,492B, BPFP=1.4452 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,560B, BPFP=1.0725 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,524B, BPFP=1.4514 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,076B, BPFP=0.1950 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203883 159.08159722 + layer.0.v_cache 0.00001438 0.01488424 + layer.1.k_cache 0.05633916 15.99725116 + layer.1.v_cache 0.00000602 0.00546598 + layer.2.k_cache 0.00244657 1.84940705 + layer.2.v_cache 0.00001847 0.01450908 + layer.3.k_cache 0.05492775 7.16872981 + layer.3.v_cache 0.00001883 0.01652312 + layer.4.k_cache 0.00062984 0.37550140 + layer.4.v_cache 0.00005178 0.03671734 + layer.4.output 0.18538215 669.52678571 + ------------------------------------------------------------------------------------- + TOTAL 0.09142157 286.54400508 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 64456 +BPFP 0.7314 bits/point +EBPFP 1.4628 equivalent bits/point +MSE 286.544005 +---------------------- -------------------------------------------------------- +Time: 0.382s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 286.5440 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,924B, BPFP=0.3340 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,916B, BPFP=1.5479 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,488B, BPFP=0.6056 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,252B, BPFP=1.4326 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,988B, BPFP=0.6924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,748B, BPFP=1.3451 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,680B, BPFP=0.6389 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,004B, BPFP=1.3896 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,616B, BPFP=1.1486 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,040B, BPFP=1.3958 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,232B, BPFP=0.1794 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13028670 180.72035590 + layer.0.v_cache 0.00001462 0.01271387 + layer.1.k_cache 0.04946813 16.92979601 + layer.1.v_cache 0.00000624 0.00427902 + layer.2.k_cache 0.00411291 1.79397922 + layer.2.v_cache 0.00001729 0.01211277 + layer.3.k_cache 0.01793825 8.55931193 + layer.3.v_cache 0.00001895 0.01539653 + layer.4.k_cache 0.00077120 0.36603139 + layer.4.v_cache 0.00004532 0.03047380 + layer.4.output 0.15109633 602.12341270 + ------------------------------------------------------------------------------------- + TOTAL 0.07413847 260.19460820 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 67888 +BPFP 0.6933 bits/point +EBPFP 1.3866 equivalent bits/point +MSE 260.194608 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.005s, Pack+Encode: 0.166s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 260.1946 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,384B, BPFP=0.3794 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,456B, BPFP=1.4956 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,360B, BPFP=0.6469 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,100B, BPFP=1.3980 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,432B, BPFP=0.6667 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,792B, BPFP=1.3136 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,276B, BPFP=0.6239 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,980B, BPFP=1.3651 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,844B, BPFP=1.0537 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,768B, BPFP=1.3070 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,172B, BPFP=0.2025 +⌛️ [2/4] FRONTEND: Frontend time: 0.193s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.153s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11683119 163.39103618 + layer.0.v_cache 0.00001514 0.01582417 + layer.1.k_cache 0.01218004 16.28929486 + layer.1.v_cache 0.00000521 0.00538811 + layer.2.k_cache 0.00491684 1.89425177 + layer.2.v_cache 0.00001706 0.01453516 + layer.3.k_cache 0.10823302 8.43665782 + layer.3.v_cache 0.00002043 0.01799029 + layer.4.k_cache 0.00062874 0.38378461 + layer.4.v_cache 0.00004794 0.03655866 + layer.4.output 0.23833110 950.51464599 + ------------------------------------------------------------------------------------- + TOTAL 0.11242431 402.59340256 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 42564 +BPFP 0.6863 bits/point +EBPFP 1.3727 equivalent bits/point +MSE 402.593403 +---------------------- -------------------------------------------------------- +Time: 0.349s Load: 0.003s, Pack+Encode: 0.193s, Decode+Unpack: 0.153s +---------------------- -------------------------------------------------------- +💾 Converting with 402.5934 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,268B, BPFP=0.3962 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,752B, BPFP=1.4850 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,164B, BPFP=0.6763 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,468B, BPFP=1.3962 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,152B, BPFP=0.6725 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,160B, BPFP=1.3000 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,068B, BPFP=0.6462 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,432B, BPFP=1.3850 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,356B, BPFP=1.0488 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,164B, BPFP=1.3013 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,796B, BPFP=0.2141 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14598725 167.14998047 + layer.0.v_cache 0.00001449 0.01649391 + layer.1.k_cache 0.01347294 17.20545410 + layer.1.v_cache 0.00000521 0.00552253 + layer.2.k_cache 0.00488209 2.10772308 + layer.2.v_cache 0.00001649 0.01565487 + layer.3.k_cache 0.04421538 9.66644775 + layer.3.v_cache 0.00001905 0.01921707 + layer.4.k_cache 0.00073779 0.43596294 + layer.4.v_cache 0.00005089 0.03801117 + layer.4.output 0.27158585 1083.71053571 + ------------------------------------------------------------------------------------- + TOTAL 0.12414721 457.80201282 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 37780 +BPFP 0.6945 bits/point +EBPFP 1.3890 equivalent bits/point +MSE 457.802013 +---------------------- -------------------------------------------------------- +Time: 0.275s Load: 0.003s, Pack+Encode: 0.129s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 457.8020 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 46, 128) +Output shape: (1, 46, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.output: torch.Size([1, 46, 3584]) -> torch.Size([1, 1, 46, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,248B, BPFP=0.4239 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,488B, BPFP=1.5245 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,060B, BPFP=0.6997 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,212B, BPFP=1.4307 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,036B, BPFP=0.6916 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,948B, BPFP=1.3410 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,952B, BPFP=0.6630 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,212B, BPFP=1.4307 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,204B, BPFP=1.0883 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,920B, BPFP=1.3315 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,852B, BPFP=0.2354 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.147s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14129612 175.61397121 + layer.0.v_cache 0.00001532 0.01663890 + layer.1.k_cache 0.01315293 16.99022509 + layer.1.v_cache 0.00000553 0.00552908 + layer.2.k_cache 0.00779452 2.25262766 + layer.2.v_cache 0.00001723 0.01468736 + layer.3.k_cache 0.01828384 8.70602815 + layer.3.v_cache 0.00001944 0.02052227 + layer.4.k_cache 0.00060559 0.41795838 + layer.4.v_cache 0.00005113 0.03947388 + layer.4.output 0.29517000 1178.32948370 + ------------------------------------------------------------------------------------- + TOTAL 0.13220186 497.19906164 + (elements=400,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 400384 +Total Bytes 36132 +BPFP 0.7219 bits/point +EBPFP 1.4439 equivalent bits/point +MSE 497.199062 +---------------------- -------------------------------------------------------- +Time: 0.278s Load: 0.002s, Pack+Encode: 0.129s, Decode+Unpack: 0.147s +---------------------- -------------------------------------------------------- +💾 Converting with 497.1991 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 110, 128) +Output shape: (1, 110, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.output: torch.Size([1, 110, 3584]) -> torch.Size([1, 1, 110, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,116B, BPFP=0.3006 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,072B, BPFP=1.4307 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,016B, BPFP=0.5705 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,388B, BPFP=1.3335 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,528B, BPFP=0.6432 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,852B, BPFP=1.2574 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,744B, BPFP=0.6739 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,056B, BPFP=1.2864 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,972B, BPFP=0.9903 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,500B, BPFP=1.2074 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,572B, BPFP=0.1739 +⌛️ [2/4] FRONTEND: Frontend time: 0.184s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14476244 161.06890980 + layer.0.v_cache 0.00001476 0.01442012 + layer.1.k_cache 0.05665370 16.68508856 + layer.1.v_cache 0.00000525 0.00442881 + layer.2.k_cache 0.00786287 1.66794281 + layer.2.v_cache 0.00001801 0.01255096 + layer.3.k_cache 0.01410385 7.40144043 + layer.3.v_cache 0.00002027 0.01584601 + layer.4.k_cache 0.00066873 0.35251160 + layer.4.v_cache 0.00004626 0.03011203 + layer.4.output 10.36401748 487.99618506 + ------------------------------------------------------------------------------------- + TOTAL 4.28072226 211.95450274 + (elements=957,440) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 957440 +Total Bytes 76816 +BPFP 0.6418 bits/point +EBPFP 1.2837 equivalent bits/point +MSE 211.954503 +---------------------- -------------------------------------------------------- +Time: 0.409s Load: 0.004s, Pack+Encode: 0.184s, Decode+Unpack: 0.222s +---------------------- -------------------------------------------------------- +💾 Converting with 211.9545 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3472 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,924B, BPFP=1.7215 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,448B, BPFP=0.6651 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,148B, BPFP=1.5718 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,868B, BPFP=0.7461 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,484B, BPFP=1.4437 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,440B, BPFP=0.6636 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,968B, BPFP=1.5370 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,848B, BPFP=1.1281 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,940B, BPFP=1.5316 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,800B, BPFP=0.2149 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10533771 168.23712384 + layer.0.v_cache 0.00001610 0.01581488 + layer.1.k_cache 0.05553475 16.75420012 + layer.1.v_cache 0.00000748 0.00594015 + layer.2.k_cache 0.02166171 1.92398862 + layer.2.v_cache 0.00002000 0.01531714 + layer.3.k_cache 0.06386927 8.66783462 + layer.3.v_cache 0.00001944 0.01778444 + layer.4.k_cache 0.00064315 0.40376428 + layer.4.v_cache 0.00005385 0.03582620 + layer.4.output 0.16801396 668.97508818 + ------------------------------------------------------------------------------------- + TOTAL 0.08372125 286.99430656 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 66668 +BPFP 0.7565 bits/point +EBPFP 1.5130 equivalent bits/point +MSE 286.994307 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.003s, Pack+Encode: 0.165s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 286.9943 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 49, 128) +Output shape: (1, 49, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.output: torch.Size([1, 49, 3584]) -> torch.Size([1, 1, 49, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,280B, BPFP=0.4082 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,776B, BPFP=1.5230 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,144B, BPFP=0.6837 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,488B, BPFP=1.4311 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,180B, BPFP=0.6952 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,192B, BPFP=1.3367 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,064B, BPFP=0.6582 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,380B, BPFP=1.3967 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,376B, BPFP=1.0765 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,252B, BPFP=1.3559 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,484B, BPFP=0.2498 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12410851 165.40270249 + layer.0.v_cache 0.00001740 0.01656302 + layer.1.k_cache 0.01401279 16.99067657 + layer.1.v_cache 0.00000579 0.00607775 + layer.2.k_cache 0.00545398 2.19599728 + layer.2.v_cache 0.00001955 0.01655375 + layer.3.k_cache 0.07546502 8.33772558 + layer.3.v_cache 0.00001978 0.01969550 + layer.4.k_cache 0.00063206 0.45076035 + layer.4.v_cache 0.00005205 0.03844803 + layer.4.output 0.27729562 1106.01466837 + ------------------------------------------------------------------------------------- + TOTAL 0.12710919 466.79869876 + (elements=426,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 426496 +Total Bytes 38616 +BPFP 0.7243 bits/point +EBPFP 1.4487 equivalent bits/point +MSE 466.798699 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 466.7987 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,396B, BPFP=0.3827 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,276B, BPFP=1.4463 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,384B, BPFP=0.6535 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,972B, BPFP=1.3629 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,428B, BPFP=0.6656 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,688B, BPFP=1.2851 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,328B, BPFP=0.6382 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,884B, BPFP=1.3388 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,776B, BPFP=1.0351 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,624B, BPFP=1.2675 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,392B, BPFP=0.2112 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19197876 167.72443805 + layer.0.v_cache 0.00001449 0.01454388 + layer.1.k_cache 0.01466701 15.83373595 + layer.1.v_cache 0.00000519 0.00508085 + layer.2.k_cache 0.00720649 1.59111853 + layer.2.v_cache 0.00001785 0.01405130 + layer.3.k_cache 0.10851213 8.30267762 + layer.3.v_cache 0.00001996 0.01832920 + layer.4.k_cache 0.00062085 0.36964212 + layer.4.v_cache 0.00004770 0.03434614 + layer.4.output 0.23832821 950.91643170 + ------------------------------------------------------------------------------------- + TOTAL 0.11714047 402.96017562 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 42148 +BPFP 0.6796 bits/point +EBPFP 1.3593 equivalent bits/point +MSE 402.960176 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.002s, Pack+Encode: 0.130s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 402.9602 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.4069 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,812B, BPFP=1.4743 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,208B, BPFP=0.6765 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,452B, BPFP=1.3640 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,164B, BPFP=0.6630 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,176B, BPFP=1.2794 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,096B, BPFP=0.6422 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,476B, BPFP=1.3713 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,420B, BPFP=1.0478 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,168B, BPFP=1.2770 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,056B, BPFP=0.2213 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12665685 163.58478860 + layer.0.v_cache 0.00001436 0.01609404 + layer.1.k_cache 0.01497617 16.88700597 + layer.1.v_cache 0.00000537 0.00555845 + layer.2.k_cache 0.00986420 2.16840916 + layer.2.v_cache 0.00001802 0.01528013 + layer.3.k_cache 0.09846177 9.51308307 + layer.3.v_cache 0.00001923 0.01932660 + layer.4.k_cache 0.00062397 0.42324133 + layer.4.v_cache 0.00005014 0.03907725 + layer.4.output 0.26631049 1062.61108193 + ------------------------------------------------------------------------------------- + TOTAL 0.12440374 448.87937871 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 38356 +BPFP 0.6912 bits/point +EBPFP 1.3825 equivalent bits/point +MSE 448.879379 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.002s, Pack+Encode: 0.131s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 448.8794 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,284B, BPFP=0.4180 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,936B, BPFP=1.6068 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,148B, BPFP=0.6992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,524B, BPFP=1.4727 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,224B, BPFP=0.7240 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,268B, BPFP=1.3893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,124B, BPFP=0.6914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,524B, BPFP=1.4727 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,468B, BPFP=1.1289 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,348B, BPFP=1.4154 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,216B, BPFP=0.2426 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18531116 163.65952555 + layer.0.v_cache 0.00001874 0.01815130 + layer.1.k_cache 0.01459585 16.26040904 + layer.1.v_cache 0.00000674 0.00699053 + layer.2.k_cache 0.01042949 2.02111785 + layer.2.v_cache 0.00001841 0.01803005 + layer.3.k_cache 0.15260293 8.77418137 + layer.3.v_cache 0.00002390 0.02398034 + layer.4.k_cache 0.00061422 0.44325852 + layer.4.v_cache 0.00004881 0.04180181 + layer.4.output 0.28297247 1129.30943080 + ------------------------------------------------------------------------------------- + TOTAL 0.13791044 476.26079188 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 39064 +BPFP 0.7480 bits/point +EBPFP 1.4960 equivalent bits/point +MSE 476.260792 +---------------------- -------------------------------------------------------- +Time: 0.276s Load: 0.002s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 476.2608 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,376B, BPFP=0.3644 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,544B, BPFP=1.4682 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,484B, BPFP=0.6578 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,220B, BPFP=1.3824 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,560B, BPFP=0.6780 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,908B, BPFP=1.2998 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,384B, BPFP=0.6314 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,196B, BPFP=1.3761 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,984B, BPFP=1.0551 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,948B, BPFP=1.3104 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,516B, BPFP=0.2087 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12052289 171.81413864 + layer.0.v_cache 0.00001546 0.01575042 + layer.1.k_cache 0.01434235 14.96589687 + layer.1.v_cache 0.00000574 0.00587632 + layer.2.k_cache 0.00254285 1.64480280 + layer.2.v_cache 0.00001710 0.01463853 + layer.3.k_cache 0.03737745 8.24373769 + layer.3.v_cache 0.00001946 0.01867108 + layer.4.k_cache 0.00063766 0.39842050 + layer.4.v_cache 0.00005039 0.03616039 + layer.4.output 0.23032015 918.72472760 + ------------------------------------------------------------------------------------- + TOTAL 0.10516308 389.89595214 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 44120 +BPFP 0.6873 bits/point +EBPFP 1.3746 equivalent bits/point +MSE 389.895952 +---------------------- -------------------------------------------------------- +Time: 0.276s Load: 0.002s, Pack+Encode: 0.129s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 389.8960 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,768B, BPFP=0.3733 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,820B, BPFP=1.6512 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,112B, BPFP=0.6571 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,276B, BPFP=1.5363 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,356B, BPFP=0.7086 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,820B, BPFP=1.4400 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,092B, BPFP=0.6529 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,268B, BPFP=1.5346 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,200B, BPFP=1.0980 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,980B, BPFP=1.4738 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,064B, BPFP=0.2131 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11335979 172.29922403 + layer.0.v_cache 0.00001513 0.01641965 + layer.1.k_cache 0.01655370 16.28417969 + layer.1.v_cache 0.00000667 0.00561337 + layer.2.k_cache 0.00428703 1.90257181 + layer.2.v_cache 0.00001789 0.01505709 + layer.3.k_cache 0.04938591 7.40522230 + layer.3.v_cache 0.00002111 0.01931002 + layer.4.k_cache 0.00066117 0.42404884 + layer.4.v_cache 0.00004716 0.03664852 + layer.4.output 0.18372514 732.67899373 + ------------------------------------------------------------------------------------- + TOTAL 0.08649597 313.36242655 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 59756 +BPFP 0.7422 bits/point +EBPFP 1.4844 equivalent bits/point +MSE 313.362427 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.003s, Pack+Encode: 0.164s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 313.3624 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,192B, BPFP=0.4233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,428B, BPFP=1.5724 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,992B, BPFP=0.7074 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,160B, BPFP=1.4773 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,032B, BPFP=0.7216 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,936B, BPFP=1.3977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,924B, BPFP=0.6832 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,200B, BPFP=1.4915 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,180B, BPFP=1.1293 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,968B, BPFP=1.4091 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,356B, BPFP=0.2717 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14777968 190.81527433 + layer.0.v_cache 0.00001531 0.01877771 + layer.1.k_cache 0.01728138 17.72126909 + layer.1.v_cache 0.00000547 0.00667929 + layer.2.k_cache 0.00823456 2.09741315 + layer.2.v_cache 0.00001932 0.01826545 + layer.3.k_cache 0.16770924 8.88590449 + layer.3.v_cache 0.00002046 0.02226676 + layer.4.k_cache 0.00061569 0.45699713 + layer.4.v_cache 0.00005592 0.04720083 + layer.4.output 0.30867824 1232.03601867 + ------------------------------------------------------------------------------------- + TOTAL 0.14720498 520.25542229 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 36368 +BPFP 0.7597 bits/point +EBPFP 1.5194 equivalent bits/point +MSE 520.255422 +---------------------- -------------------------------------------------------- +Time: 0.278s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 520.2554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,776B, BPFP=0.3558 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,260B, BPFP=1.8550 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,312B, BPFP=0.6635 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,632B, BPFP=1.7292 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,704B, BPFP=0.7420 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,804B, BPFP=1.5633 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,300B, BPFP=0.6611 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,984B, BPFP=1.5994 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,084B, BPFP=1.2188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,860B, BPFP=1.5745 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,016B, BPFP=0.2008 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14408422 170.34478916 + layer.0.v_cache 0.00002097 0.01776201 + layer.1.k_cache 0.05735342 16.62295767 + layer.1.v_cache 0.00000575 0.00610614 + layer.2.k_cache 0.01678689 2.05125936 + layer.2.v_cache 0.00001894 0.01613891 + layer.3.k_cache 0.03029843 7.05740278 + layer.3.v_cache 0.00002125 0.02094163 + layer.4.k_cache 0.00062582 0.43078555 + layer.4.v_cache 0.00005050 0.04136310 + layer.4.output 0.17433185 694.78725962 + ------------------------------------------------------------------------------------- + TOTAL 0.08644642 297.65413668 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 66732 +BPFP 0.7863 bits/point +EBPFP 1.5727 equivalent bits/point +MSE 297.654137 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.003s, Pack+Encode: 0.164s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 297.6541 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,212B, BPFP=0.4304 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,272B, BPFP=1.5170 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,972B, BPFP=0.7003 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,992B, BPFP=1.4176 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,968B, BPFP=0.6989 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,796B, BPFP=1.3480 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,840B, BPFP=0.6534 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,048B, BPFP=1.4375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,012B, BPFP=1.0696 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,776B, BPFP=1.3409 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,812B, BPFP=0.2441 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13243668 191.51251776 + layer.0.v_cache 0.00001576 0.01825734 + layer.1.k_cache 0.01981235 17.80127231 + layer.1.v_cache 0.00000577 0.00652404 + layer.2.k_cache 0.01756107 2.22153369 + layer.2.v_cache 0.00001728 0.01720197 + layer.3.k_cache 0.04920139 8.82317144 + layer.3.v_cache 0.00001837 0.01994913 + layer.4.k_cache 0.00061593 0.46276240 + layer.4.v_cache 0.00004907 0.04261926 + layer.4.output 0.30854983 1231.98153409 + ------------------------------------------------------------------------------------- + TOTAL 0.13997544 520.28214988 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 34700 +BPFP 0.7248 bits/point +EBPFP 1.4497 equivalent bits/point +MSE 520.282150 +---------------------- -------------------------------------------------------- +Time: 0.279s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 520.2821 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,272B, BPFP=0.3975 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,980B, BPFP=1.5562 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,164B, BPFP=0.6763 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,600B, BPFP=1.4375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,244B, BPFP=0.7013 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,352B, BPFP=1.3600 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,112B, BPFP=0.6600 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,576B, BPFP=1.4300 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,444B, BPFP=1.0762 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,336B, BPFP=1.3550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,876B, BPFP=0.2177 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20666794 163.58391602 + layer.0.v_cache 0.00002565 0.01803046 + layer.1.k_cache 0.01518770 17.29127686 + layer.1.v_cache 0.00000581 0.00624457 + layer.2.k_cache 0.01031687 2.21211212 + layer.2.v_cache 0.00001873 0.01695551 + layer.3.k_cache 0.09638391 9.98242188 + layer.3.v_cache 0.00001967 0.02102599 + layer.4.k_cache 0.00062172 0.45397484 + layer.4.v_cache 0.00004827 0.04100801 + layer.4.output 0.27162739 1083.69892857 + ------------------------------------------------------------------------------------- + TOTAL 0.13121694 457.61879213 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 38956 +BPFP 0.7161 bits/point +EBPFP 1.4322 equivalent bits/point +MSE 457.618792 +---------------------- -------------------------------------------------------- +Time: 0.276s Load: 0.002s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 457.6188 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,604B, BPFP=0.3686 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,628B, BPFP=1.7528 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,792B, BPFP=0.6415 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,176B, BPFP=1.6489 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,996B, BPFP=0.6884 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,428B, BPFP=1.4770 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,744B, BPFP=0.6305 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,476B, BPFP=1.4881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,736B, BPFP=1.0882 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,660B, BPFP=1.5303 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,196B, BPFP=0.2362 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13431002 161.21392463 + layer.0.v_cache 0.00001720 0.01729790 + layer.1.k_cache 0.01179634 15.30978573 + layer.1.v_cache 0.00000537 0.00556351 + layer.2.k_cache 0.00798416 1.99058937 + layer.2.v_cache 0.00001783 0.01579547 + layer.3.k_cache 0.05795463 8.40261302 + layer.3.v_cache 0.00001974 0.01956060 + layer.4.k_cache 0.00062059 0.41470875 + layer.4.v_cache 0.00005204 0.03864649 + layer.4.output 0.19989206 797.02370011 + ------------------------------------------------------------------------------------- + TOTAL 0.09482484 339.21143448 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 56436 +BPFP 0.7628 bits/point +EBPFP 1.5256 equivalent bits/point +MSE 339.211434 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.003s, Pack+Encode: 0.164s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 339.2114 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,268B, BPFP=0.3962 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,888B, BPFP=1.5275 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,176B, BPFP=0.6800 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,536B, BPFP=1.4175 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,208B, BPFP=0.6900 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,324B, BPFP=1.3513 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,116B, BPFP=0.6613 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,552B, BPFP=1.4225 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,420B, BPFP=1.0688 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,352B, BPFP=1.3600 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,908B, BPFP=0.2191 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15467889 166.14445313 + layer.0.v_cache 0.00001741 0.01738820 + layer.1.k_cache 0.01620592 16.90489014 + layer.1.v_cache 0.00000548 0.00614663 + layer.2.k_cache 0.00723796 2.21288300 + layer.2.v_cache 0.00001737 0.01661934 + layer.3.k_cache 0.06661982 9.37812500 + layer.3.v_cache 0.00002040 0.02029860 + layer.4.k_cache 0.00058776 0.42323578 + layer.4.v_cache 0.00004656 0.04118970 + layer.4.output 0.27162074 1083.33544643 + ------------------------------------------------------------------------------------- + TOTAL 0.12628134 457.55960909 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 38748 +BPFP 0.7123 bits/point +EBPFP 1.4246 equivalent bits/point +MSE 457.559609 +---------------------- -------------------------------------------------------- +Time: 0.276s Load: 0.003s, Pack+Encode: 0.129s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 457.5596 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,324B, BPFP=0.4056 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,756B, BPFP=1.4571 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,212B, BPFP=0.6777 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,444B, BPFP=1.3615 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,136B, BPFP=0.6544 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,164B, BPFP=1.2757 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,060B, BPFP=0.6311 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,444B, BPFP=1.3615 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,328B, BPFP=1.0196 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,164B, BPFP=1.2757 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 4,912B, BPFP=0.2150 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11261671 160.89005055 + layer.0.v_cache 0.00001829 0.01624327 + layer.1.k_cache 0.01190703 16.30815095 + layer.1.v_cache 0.00000528 0.00581159 + layer.2.k_cache 0.00973116 2.06460556 + layer.2.v_cache 0.00001668 0.01589032 + layer.3.k_cache 0.04200072 8.80331481 + layer.3.v_cache 0.00001874 0.01996098 + layer.4.k_cache 0.00062040 0.41703692 + layer.4.v_cache 0.00004776 0.03895752 + layer.4.output 0.26626884 1062.61677171 + ------------------------------------------------------------------------------------- + TOTAL 0.12005086 448.64102497 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 37944 +BPFP 0.6838 bits/point +EBPFP 1.3676 equivalent bits/point +MSE 448.641025 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.002s, Pack+Encode: 0.129s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 448.6410 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.3990 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,092B, BPFP=1.5300 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,256B, BPFP=0.6779 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,744B, BPFP=1.4255 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,300B, BPFP=0.6911 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,476B, BPFP=1.3450 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,200B, BPFP=0.6611 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,764B, BPFP=1.4315 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,632B, BPFP=1.0913 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,524B, BPFP=1.3594 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,348B, BPFP=0.2296 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12932784 168.56201172 + layer.0.v_cache 0.00001491 0.01756012 + layer.1.k_cache 0.01488123 16.43546589 + layer.1.v_cache 0.00000563 0.00634485 + layer.2.k_cache 0.00939683 2.16293100 + layer.2.v_cache 0.00001981 0.01625087 + layer.3.k_cache 0.07051506 9.05244915 + layer.3.v_cache 0.00001927 0.02026055 + layer.4.k_cache 0.00064413 0.42629913 + layer.4.v_cache 0.00005939 0.04180783 + layer.4.output 0.26127321 1041.79369849 + ------------------------------------------------------------------------------------- + TOTAL 0.12081156 440.54689827 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 40664 +BPFP 0.7188 bits/point +EBPFP 1.4375 equivalent bits/point +MSE 440.546898 +---------------------- -------------------------------------------------------- +Time: 0.275s Load: 0.002s, Pack+Encode: 0.129s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 440.5469 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,328B, BPFP=0.3990 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,116B, BPFP=1.5373 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,260B, BPFP=0.6791 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,820B, BPFP=1.4483 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,340B, BPFP=0.7031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,564B, BPFP=1.3714 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,248B, BPFP=0.6755 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,784B, BPFP=1.4375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,620B, BPFP=1.0877 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,512B, BPFP=1.3558 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,308B, BPFP=0.2279 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.148s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13190371 172.35199444 + layer.0.v_cache 0.00001552 0.01745356 + layer.1.k_cache 0.01668521 16.61436580 + layer.1.v_cache 0.00000581 0.00656241 + layer.2.k_cache 0.00490881 2.14354148 + layer.2.v_cache 0.00001873 0.01708750 + layer.3.k_cache 0.11824917 8.58918586 + layer.3.v_cache 0.00002115 0.02069829 + layer.4.k_cache 0.00061206 0.42215102 + layer.4.v_cache 0.00004844 0.03879626 + layer.4.output 0.26122982 1041.99304602 + ------------------------------------------------------------------------------------- + TOTAL 0.12359279 440.83371522 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 40900 +BPFP 0.7229 bits/point +EBPFP 1.4458 equivalent bits/point +MSE 440.833715 +---------------------- -------------------------------------------------------- +Time: 0.280s Load: 0.002s, Pack+Encode: 0.130s, Decode+Unpack: 0.148s +---------------------- -------------------------------------------------------- +💾 Converting with 440.8337 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,956B, BPFP=0.3322 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,120B, BPFP=1.5489 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,772B, BPFP=0.6406 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,772B, BPFP=1.4898 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,496B, BPFP=0.7636 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,012B, BPFP=1.3607 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,988B, BPFP=0.6773 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,420B, BPFP=1.4300 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,168B, BPFP=1.2174 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,396B, BPFP=1.4260 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,968B, BPFP=0.1933 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10326390 173.66550611 + layer.0.v_cache 0.00001541 0.01402939 + layer.1.k_cache 0.04976616 16.38886825 + layer.1.v_cache 0.00000534 0.00462504 + layer.2.k_cache 0.01168035 1.84078714 + layer.2.v_cache 0.00001798 0.01324262 + layer.3.k_cache 0.02907468 8.52081432 + layer.3.v_cache 0.00001946 0.01686086 + layer.4.k_cache 0.00073861 0.36187951 + layer.4.v_cache 0.00005077 0.03275989 + layer.4.output 0.14788401 589.25742430 + ------------------------------------------------------------------------------------- + TOTAL 0.07234240 254.45066725 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 72068 +BPFP 0.7200 bits/point +EBPFP 1.4400 equivalent bits/point +MSE 254.450667 +---------------------- -------------------------------------------------------- +Time: 0.382s Load: 0.003s, Pack+Encode: 0.165s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 254.4507 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 64, 128) +Output shape: (1, 64, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.output: torch.Size([1, 64, 3584]) -> torch.Size([1, 1, 64, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 984B, BPFP=0.2402 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,268B, BPFP=1.2861 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,768B, BPFP=0.4316 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,944B, BPFP=1.2070 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,888B, BPFP=0.4609 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,716B, BPFP=1.1514 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,772B, BPFP=0.4326 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,824B, BPFP=1.1777 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,500B, BPFP=0.8545 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,556B, BPFP=1.1123 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 3,716B, BPFP=0.1296 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09722494 113.91770935 + layer.0.v_cache 0.00001391 0.01266741 + layer.1.k_cache 0.01806639 12.73778439 + layer.1.v_cache 0.00000557 0.00469871 + layer.2.k_cache 0.00423524 1.52084196 + layer.2.v_cache 0.00001868 0.01343889 + layer.3.k_cache 0.03759564 5.39554977 + layer.3.v_cache 0.00001825 0.01467717 + layer.4.k_cache 0.00061753 0.31828406 + layer.4.v_cache 0.00004894 0.03259017 + layer.4.output 0.21408452 845.57498605 + ------------------------------------------------------------------------------------- + TOTAL 0.09743746 356.05842025 + (elements=557,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 557056 +Total Bytes 37936 +BPFP 0.5448 bits/point +EBPFP 1.0896 equivalent bits/point +MSE 356.058420 +---------------------- -------------------------------------------------------- +Time: 0.279s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 356.0584 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,992B, BPFP=0.3051 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,848B, BPFP=1.5086 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,368B, BPFP=0.6691 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,200B, BPFP=1.4093 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,916B, BPFP=0.7531 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,444B, BPFP=1.2935 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,348B, BPFP=0.6661 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,136B, BPFP=1.3995 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,680B, BPFP=1.1765 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,340B, BPFP=1.2776 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,960B, BPFP=0.1742 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11648470 158.88847082 + layer.0.v_cache 0.00001454 0.01344617 + layer.1.k_cache 0.02721847 16.50015319 + layer.1.v_cache 0.00000531 0.00444627 + layer.2.k_cache 0.01045542 1.77606904 + layer.2.v_cache 0.00001699 0.01205061 + layer.3.k_cache 0.01744144 8.32018086 + layer.3.v_cache 0.00001839 0.01464627 + layer.4.k_cache 0.00076342 0.35882942 + layer.4.v_cache 0.00004795 0.03068275 + layer.4.output 11.17676328 526.59515056 + ------------------------------------------------------------------------------------- + TOTAL 4.61234174 227.76970761 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 76232 +BPFP 0.6869 bits/point +EBPFP 1.3738 equivalent bits/point +MSE 227.769708 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 227.7697 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,784B, BPFP=0.3484 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,720B, BPFP=1.7031 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,256B, BPFP=0.6359 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,736B, BPFP=1.5109 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,532B, BPFP=0.6898 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,848B, BPFP=1.3375 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,296B, BPFP=0.6438 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,192B, BPFP=1.4047 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,492B, BPFP=1.0727 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,028B, BPFP=1.3727 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,940B, BPFP=0.1936 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11433966 161.58027344 + layer.0.v_cache 0.00001437 0.01326039 + layer.1.k_cache 0.05315235 16.46204834 + layer.1.v_cache 0.00000524 0.00422906 + layer.2.k_cache 0.00678293 1.81539688 + layer.2.v_cache 0.00001704 0.01256902 + layer.3.k_cache 0.01951189 7.36596069 + layer.3.v_cache 0.00001758 0.01487750 + layer.4.k_cache 0.00076624 0.36696486 + layer.4.v_cache 0.00004856 0.03196912 + layer.4.output 0.16992235 677.71925223 + ------------------------------------------------------------------------------------- + TOTAL 0.08141837 290.10013617 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 61824 +BPFP 0.7103 bits/point +EBPFP 1.4206 equivalent bits/point +MSE 290.100136 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1001 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,916B, BPFP=0.3326 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,488B, BPFP=1.6472 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,512B, BPFP=0.6097 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,720B, BPFP=1.5139 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,944B, BPFP=0.6847 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,144B, BPFP=1.4139 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,600B, BPFP=0.6250 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,520B, BPFP=1.4792 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,928B, BPFP=1.2028 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,352B, BPFP=1.4500 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,548B, BPFP=0.1872 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11570581 180.75211589 + layer.0.v_cache 0.00001517 0.01370706 + layer.1.k_cache 0.03236232 16.79860569 + layer.1.v_cache 0.00000662 0.00435019 + layer.2.k_cache 0.00641356 1.67759298 + layer.2.v_cache 0.00001676 0.01231912 + layer.3.k_cache 0.02648412 8.34485270 + layer.3.v_cache 0.00001896 0.01549241 + layer.4.k_cache 0.00074524 0.36987695 + layer.4.v_cache 0.00004843 0.03184930 + layer.4.output 0.15113260 601.98546627 + ------------------------------------------------------------------------------------- + TOTAL 0.07292619 260.11288389 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 70672 +BPFP 0.7217 bits/point +EBPFP 1.4435 equivalent bits/point +MSE 260.112884 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 260.1129 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 114, 128) +Output shape: (1, 114, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.output: torch.Size([1, 114, 3584]) -> torch.Size([1, 1, 114, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,112B, BPFP=0.2895 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,732B, BPFP=1.3339 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,212B, BPFP=0.5773 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,096B, BPFP=1.2467 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,664B, BPFP=0.6393 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,496B, BPFP=1.1645 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,600B, BPFP=0.6305 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,812B, BPFP=1.2078 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,044B, BPFP=0.9655 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,288B, BPFP=1.1360 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,404B, BPFP=0.1646 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10081583 164.30631510 + layer.0.v_cache 0.00001633 0.01319515 + layer.1.k_cache 0.08473034 16.63068162 + layer.1.v_cache 0.00000563 0.00452016 + layer.2.k_cache 0.00792928 1.67204097 + layer.2.v_cache 0.00001730 0.01198750 + layer.3.k_cache 0.01019665 8.49187858 + layer.3.v_cache 0.00001810 0.01386925 + layer.4.k_cache 0.00077074 0.33712093 + layer.4.v_cache 0.00005022 0.03226647 + layer.4.output 10.00037814 470.83161028 + ------------------------------------------------------------------------------------- + TOTAL 4.12983514 205.13736163 + (elements=992,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 992256 +Total Bytes 75460 +BPFP 0.6084 bits/point +EBPFP 1.2168 equivalent bits/point +MSE 205.137362 +---------------------- -------------------------------------------------------- +Time: 0.386s Load: 0.005s, Pack+Encode: 0.166s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 205.1374 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 98, 128) +Output shape: (1, 98, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.output: torch.Size([1, 98, 3584]) -> torch.Size([1, 1, 98, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,832B, BPFP=0.2921 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,940B, BPFP=1.4254 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,588B, BPFP=0.5721 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,772B, BPFP=1.3986 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,476B, BPFP=0.7136 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,248B, BPFP=1.3151 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,844B, BPFP=0.6129 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,524B, BPFP=1.3591 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,140B, BPFP=1.1384 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,036B, BPFP=1.2812 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,244B, BPFP=0.1650 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11008905 167.05436862 + layer.0.v_cache 0.00001718 0.01321119 + layer.1.k_cache 0.06240322 17.06813641 + layer.1.v_cache 0.00000548 0.00419989 + layer.2.k_cache 0.00473963 1.77331045 + layer.2.v_cache 0.00001718 0.01219331 + layer.3.k_cache 0.01141277 8.70917495 + layer.3.v_cache 0.00001769 0.01399920 + layer.4.k_cache 0.00076610 0.37823556 + layer.4.v_cache 0.00004814 0.03158011 + layer.4.output 0.02426045 568.22102770 + ------------------------------------------------------------------------------------- + TOTAL 0.02113762 245.44738845 + (elements=852,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 852992 +Total Bytes 70644 +BPFP 0.6626 bits/point +EBPFP 1.3251 equivalent bits/point +MSE 245.447388 +---------------------- -------------------------------------------------------- +Time: 0.382s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 245.4474 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,956B, BPFP=0.3322 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,896B, BPFP=1.5109 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,876B, BPFP=0.6583 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,620B, BPFP=1.4640 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,668B, BPFP=0.7928 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,236B, BPFP=1.3988 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,092B, BPFP=0.6950 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,572B, BPFP=1.4558 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,720B, BPFP=1.1413 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,340B, BPFP=1.4164 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,956B, BPFP=0.1930 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08919133 176.04256539 + layer.0.v_cache 0.00001414 0.01288561 + layer.1.k_cache 0.08591949 16.48264479 + layer.1.v_cache 0.00000528 0.00431043 + layer.2.k_cache 0.00389855 1.82540645 + layer.2.v_cache 0.00001738 0.01261572 + layer.3.k_cache 0.07679920 8.50219196 + layer.3.v_cache 0.00002028 0.01621992 + layer.4.k_cache 0.00072407 0.37525476 + layer.4.v_cache 0.00004779 0.03393554 + layer.4.output 0.14786680 589.25160132 + ------------------------------------------------------------------------------------- + TOTAL 0.07598265 254.59230822 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 71932 +BPFP 0.7186 bits/point +EBPFP 1.4373 equivalent bits/point +MSE 254.592308 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.005s, Pack+Encode: 0.166s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 254.5923 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,088B, BPFP=0.3049 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,448B, BPFP=1.5257 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,204B, BPFP=0.6139 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,988B, BPFP=1.4585 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,696B, BPFP=0.8318 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,540B, BPFP=1.3931 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,780B, BPFP=0.6980 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,756B, BPFP=1.4246 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,620B, BPFP=1.1127 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,980B, BPFP=1.3113 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,388B, BPFP=0.1750 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11353527 163.19960207 + layer.0.v_cache 0.00001674 0.01356592 + layer.1.k_cache 0.04276301 15.97557909 + layer.1.v_cache 0.00000542 0.00423575 + layer.2.k_cache 0.00247095 1.86072091 + layer.2.v_cache 0.00001885 0.01278605 + layer.3.k_cache 0.02722364 7.31949851 + layer.3.v_cache 0.00001942 0.01492758 + layer.4.k_cache 0.00068256 0.35513413 + layer.4.v_cache 0.00004843 0.03211468 + layer.4.output 10.65454330 501.96862483 + ------------------------------------------------------------------------------------- + TOTAL 4.39815220 217.79814932 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 81488 +BPFP 0.7000 bits/point +EBPFP 1.3999 equivalent bits/point +MSE 217.798149 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.006s, Pack+Encode: 0.166s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 217.7981 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,864B, BPFP=0.3467 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,752B, BPFP=1.6280 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,336B, BPFP=0.6205 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,820B, BPFP=1.4546 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,648B, BPFP=0.6786 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,364B, BPFP=1.3698 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,412B, BPFP=0.6347 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,660B, BPFP=1.4249 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,552B, BPFP=1.0327 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,292B, BPFP=1.3564 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,088B, BPFP=0.1884 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12550225 169.60456194 + layer.0.v_cache 0.00001468 0.01403844 + layer.1.k_cache 0.05234316 16.81608364 + layer.1.v_cache 0.00000508 0.00431522 + layer.2.k_cache 0.00733058 1.68043918 + layer.2.v_cache 0.00001689 0.01245660 + layer.3.k_cache 0.01592888 8.56604803 + layer.3.v_cache 0.00001847 0.01544885 + layer.4.k_cache 0.00069460 0.36200201 + layer.4.v_cache 0.00005205 0.03320231 + layer.4.output 0.16184913 645.14801233 + ------------------------------------------------------------------------------------- + TOTAL 0.07852062 277.24380485 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 63788 +BPFP 0.6980 bits/point +EBPFP 1.3959 equivalent bits/point +MSE 277.243805 +---------------------- -------------------------------------------------------- +Time: 0.385s Load: 0.004s, Pack+Encode: 0.167s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 277.2438 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,780B, BPFP=0.3566 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,552B, BPFP=1.7131 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,164B, BPFP=0.6338 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,868B, BPFP=1.5761 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,420B, BPFP=0.6851 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,920B, BPFP=1.3862 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,224B, BPFP=0.6458 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,436B, BPFP=1.4896 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,816B, BPFP=1.1651 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,828B, BPFP=1.5681 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,828B, BPFP=0.1954 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11879376 168.54534756 + layer.0.v_cache 0.00001438 0.01422641 + layer.1.k_cache 0.05518859 16.62360402 + layer.1.v_cache 0.00000514 0.00492579 + layer.2.k_cache 0.00243519 1.83481500 + layer.2.v_cache 0.00001742 0.01305210 + layer.3.k_cache 0.07959401 7.19458399 + layer.3.v_cache 0.00001812 0.01572344 + layer.4.k_cache 0.00069358 0.38497675 + layer.4.v_cache 0.00004620 0.03225011 + layer.4.output 0.17426339 694.96228251 + ------------------------------------------------------------------------------------- + TOTAL 0.08686177 297.61173428 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 62836 +BPFP 0.7404 bits/point +EBPFP 1.4809 equivalent bits/point +MSE 297.611734 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 297.6117 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,796B, BPFP=0.3422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,856B, BPFP=1.6875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,332B, BPFP=0.6349 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,120B, BPFP=1.5473 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,628B, BPFP=0.6913 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,352B, BPFP=1.4009 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,288B, BPFP=0.6265 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,808B, BPFP=1.4878 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,904B, BPFP=1.1250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,812B, BPFP=1.4886 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,740B, BPFP=0.1835 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11325797 161.94656298 + layer.0.v_cache 0.00001565 0.01366552 + layer.1.k_cache 0.07651102 16.56941633 + layer.1.v_cache 0.00000530 0.00500859 + layer.2.k_cache 0.00872486 1.93863213 + layer.2.v_cache 0.00001696 0.01321456 + layer.3.k_cache 0.12876377 8.01515012 + layer.3.v_cache 0.00001905 0.01601728 + layer.4.k_cache 0.00069190 0.38731659 + layer.4.v_cache 0.00004968 0.03572915 + layer.4.output 0.16578155 660.83340592 + ------------------------------------------------------------------------------------- + TOTAL 0.08756041 283.22203263 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 64636 +BPFP 0.7245 bits/point +EBPFP 1.4490 equivalent bits/point +MSE 283.222033 +---------------------- -------------------------------------------------------- +Time: 0.377s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 283.2220 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,776B, BPFP=0.3604 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,408B, BPFP=1.7062 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,172B, BPFP=0.6437 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,680B, BPFP=1.5584 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,396B, BPFP=0.6891 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,348B, BPFP=1.4911 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,040B, BPFP=0.6169 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,544B, BPFP=1.5308 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,628B, BPFP=1.1420 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,524B, BPFP=1.5268 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,544B, BPFP=0.1897 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12835005 174.17418324 + layer.0.v_cache 0.00001616 0.01412409 + layer.1.k_cache 0.05672838 16.62919478 + layer.1.v_cache 0.00000535 0.00522202 + layer.2.k_cache 0.00240846 1.87377454 + layer.2.v_cache 0.00001644 0.01376631 + layer.3.k_cache 0.04570918 6.89845652 + layer.3.v_cache 0.00001733 0.01613316 + layer.4.k_cache 0.00072090 0.38828622 + layer.4.v_cache 0.00004908 0.03534843 + layer.4.output 0.18444509 704.33418367 + ------------------------------------------------------------------------------------- + TOTAL 0.08971394 301.78751618 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 62060 +BPFP 0.7408 bits/point +EBPFP 1.4816 equivalent bits/point +MSE 301.787516 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 301.7875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,696B, BPFP=0.3732 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,356B, BPFP=1.6188 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,080B, BPFP=0.6778 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,616B, BPFP=1.4560 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,076B, BPFP=0.6769 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,140B, BPFP=1.3512 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,796B, BPFP=0.6153 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,356B, BPFP=1.3988 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,072B, BPFP=1.1162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,612B, BPFP=1.4551 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,676B, BPFP=0.2099 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10085608 154.28809969 + layer.0.v_cache 0.00001554 0.01320484 + layer.1.k_cache 0.06589249 15.84586474 + layer.1.v_cache 0.00000499 0.00472285 + layer.2.k_cache 0.00422055 1.80804228 + layer.2.v_cache 0.00001579 0.01200183 + layer.3.k_cache 0.05337993 7.38524359 + layer.3.v_cache 0.00001747 0.01561165 + layer.4.k_cache 0.00075412 0.37049280 + layer.4.v_cache 0.00004712 0.03310871 + layer.4.output 0.19135183 763.77621982 + ------------------------------------------------------------------------------------- + TOTAL 0.09203923 325.07117245 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 55476 +BPFP 0.7182 bits/point +EBPFP 1.4363 equivalent bits/point +MSE 325.071172 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 325.0712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,888B, BPFP=0.3512 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,480B, BPFP=1.5774 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,356B, BPFP=0.6243 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,788B, BPFP=1.4487 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,812B, BPFP=0.7091 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,184B, BPFP=1.3363 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,416B, BPFP=0.6354 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,512B, BPFP=1.3973 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,924B, BPFP=1.1019 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,540B, BPFP=1.4025 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,692B, BPFP=0.1778 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09057366 167.65397135 + layer.0.v_cache 0.00001789 0.01331047 + layer.1.k_cache 0.03614600 16.42707461 + layer.1.v_cache 0.00000481 0.00427173 + layer.2.k_cache 0.00554357 1.93348022 + layer.2.v_cache 0.00001714 0.01198989 + layer.3.k_cache 0.02742965 8.62528846 + layer.3.v_cache 0.00001729 0.01470648 + layer.4.k_cache 0.00093027 0.35991292 + layer.4.v_cache 0.00004906 0.03068907 + layer.4.output 0.16441821 645.05872662 + ------------------------------------------------------------------------------------- + TOTAL 0.07715628 277.08739891 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 63592 +BPFP 0.6958 bits/point +EBPFP 1.3916 equivalent bits/point +MSE 277.087399 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 277.0874 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,780B, BPFP=0.3612 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,332B, BPFP=1.6907 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,156B, BPFP=0.6404 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,828B, BPFP=1.5885 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,372B, BPFP=0.6843 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,888B, BPFP=1.3977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,124B, BPFP=0.6339 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,336B, BPFP=1.4886 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,664B, BPFP=1.1494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,448B, BPFP=1.5114 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,808B, BPFP=0.1974 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09325875 175.54682427 + layer.0.v_cache 0.00001406 0.01376835 + layer.1.k_cache 0.05775181 16.52322031 + layer.1.v_cache 0.00000524 0.00466242 + layer.2.k_cache 0.00413042 2.04726895 + layer.2.v_cache 0.00001743 0.01283322 + layer.3.k_cache 0.04387648 7.55827351 + layer.3.v_cache 0.00002497 0.01657445 + layer.4.k_cache 0.00071512 0.38176727 + layer.4.v_cache 0.00005147 0.03614605 + layer.4.output 0.17652170 703.87615955 + ------------------------------------------------------------------------------------- + TOTAL 0.08444104 301.72202681 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 61736 +BPFP 0.7369 bits/point +EBPFP 1.4738 equivalent bits/point +MSE 301.722027 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 301.7220 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,984B, BPFP=0.3523 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,260B, BPFP=1.4666 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,396B, BPFP=0.6030 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,300B, BPFP=1.2962 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,696B, BPFP=0.6562 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,088B, BPFP=1.2585 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,352B, BPFP=0.5952 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,572B, BPFP=1.3445 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,552B, BPFP=0.9858 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,896B, BPFP=1.2244 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,576B, BPFP=0.1668 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212358 165.49997781 + layer.0.v_cache 0.00001682 0.01115591 + layer.1.k_cache 0.05918076 16.18841553 + layer.1.v_cache 0.00000473 0.00393215 + layer.2.k_cache 0.00260097 1.69810451 + layer.2.v_cache 0.00001544 0.01137295 + layer.3.k_cache 0.02838851 7.70760554 + layer.3.v_cache 0.00001657 0.01396456 + layer.4.k_cache 0.00106615 0.32410041 + layer.4.v_cache 0.00004456 0.02901179 + layer.4.output 0.17173813 616.16512784 + ------------------------------------------------------------------------------------- + TOTAL 0.08091912 264.97903153 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 61672 +BPFP 0.6441 bits/point +EBPFP 1.2883 equivalent bits/point +MSE 264.979032 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 264.9790 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,968B, BPFP=0.3494 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,092B, BPFP=1.6143 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,496B, BPFP=0.6207 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,348B, BPFP=1.4822 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,956B, BPFP=0.7024 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,784B, BPFP=1.3821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,488B, BPFP=0.6193 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,068B, BPFP=1.4325 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,592B, BPFP=1.1705 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,116B, BPFP=1.4411 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,120B, BPFP=0.1806 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10731998 165.04896129 + layer.0.v_cache 0.00001618 0.01271783 + layer.1.k_cache 0.07375650 16.09716242 + layer.1.v_cache 0.00000540 0.00433796 + layer.2.k_cache 0.00246326 1.67326494 + layer.2.v_cache 0.00001746 0.01247102 + layer.3.k_cache 0.01642703 8.35669500 + layer.3.v_cache 0.00001778 0.01452301 + layer.4.k_cache 0.00066607 0.35249682 + layer.4.v_cache 0.00004728 0.03112665 + layer.4.output 0.15452130 615.95363231 + ------------------------------------------------------------------------------------- + TOTAL 0.07543447 264.89877547 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 68028 +BPFP 0.7105 bits/point +EBPFP 1.4210 equivalent bits/point +MSE 264.898775 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 264.8988 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 99, 128) +Output shape: (1, 99, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.output: torch.Size([1, 99, 3584]) -> torch.Size([1, 1, 99, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,944B, BPFP=0.3068 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,960B, BPFP=1.5720 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,956B, BPFP=0.6244 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,260B, BPFP=1.4615 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,748B, BPFP=0.7494 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,964B, BPFP=1.4148 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,240B, BPFP=0.6692 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,288B, BPFP=1.4659 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,964B, BPFP=1.0991 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,308B, BPFP=1.3112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,172B, BPFP=0.1843 +⌛️ [2/4] FRONTEND: Frontend time: 0.177s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09061131 163.33591777 + layer.0.v_cache 0.00001461 0.01287632 + layer.1.k_cache 0.05062923 16.34530979 + layer.1.v_cache 0.00000709 0.00456278 + layer.2.k_cache 0.00392854 1.87120133 + layer.2.v_cache 0.00001793 0.01243628 + layer.3.k_cache 0.05082532 7.43504071 + layer.3.v_cache 0.00001820 0.01452493 + layer.4.k_cache 0.00080398 0.35660168 + layer.4.v_cache 0.00005053 0.02975551 + layer.4.output 0.02403493 562.23624639 + ------------------------------------------------------------------------------------- + TOTAL 0.02147948 242.65129128 + (elements=861,696) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 861696 +Total Bytes 75804 +BPFP 0.7038 bits/point +EBPFP 1.4075 equivalent bits/point +MSE 242.651291 +---------------------- -------------------------------------------------------- +Time: 0.388s Load: 0.005s, Pack+Encode: 0.177s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 242.6513 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,676B, BPFP=0.3688 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,240B, BPFP=1.8134 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,072B, BPFP=0.6761 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,036B, BPFP=1.5484 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,088B, BPFP=0.6796 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,604B, BPFP=1.4533 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,872B, BPFP=0.6320 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,968B, BPFP=1.5335 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,864B, BPFP=1.0704 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,956B, BPFP=1.5308 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,220B, BPFP=0.2270 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11136492 160.20898438 + layer.0.v_cache 0.00001955 0.01559974 + layer.1.k_cache 0.06280328 16.21675217 + layer.1.v_cache 0.00000544 0.00504736 + layer.2.k_cache 0.00235174 1.90700477 + layer.2.v_cache 0.00001859 0.01370564 + layer.3.k_cache 0.03326307 7.33819580 + layer.3.v_cache 0.00001926 0.01651742 + layer.4.k_cache 0.00064473 0.36807023 + layer.4.v_cache 0.00005254 0.03259006 + layer.4.output 0.19494371 763.93907193 + ------------------------------------------------------------------------------------- + TOTAL 0.09265583 325.51152771 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 58596 +BPFP 0.7585 bits/point +EBPFP 1.5171 equivalent bits/point +MSE 325.511528 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 325.5115 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,752B, BPFP=0.3699 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,576B, BPFP=1.8108 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,096B, BPFP=0.6537 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,484B, BPFP=1.5802 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,192B, BPFP=0.6740 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,116B, BPFP=1.5025 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,888B, BPFP=0.6098 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,708B, BPFP=1.6275 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,360B, BPFP=1.1318 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,404B, BPFP=1.5633 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,824B, BPFP=0.2058 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08829598 176.07972181 + layer.0.v_cache 0.00001455 0.01514616 + layer.1.k_cache 0.05950350 16.05413983 + layer.1.v_cache 0.00000529 0.00505618 + layer.2.k_cache 0.00411868 1.91374227 + layer.2.v_cache 0.00001680 0.01383735 + layer.3.k_cache 0.04877868 8.62717252 + layer.3.v_cache 0.00001896 0.01770076 + layer.4.k_cache 0.00063741 0.39823171 + layer.4.v_cache 0.00004764 0.03694725 + layer.4.output 0.18365649 732.74710425 + ------------------------------------------------------------------------------------- + TOTAL 0.08747253 313.67008386 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 61400 +BPFP 0.7626 bits/point +EBPFP 1.5252 equivalent bits/point +MSE 313.670084 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 313.6701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,764B, BPFP=0.3580 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,376B, BPFP=1.6997 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,124B, BPFP=0.6339 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,660B, BPFP=1.5544 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,276B, BPFP=0.6648 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,252B, BPFP=1.4716 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,012B, BPFP=0.6112 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,376B, BPFP=1.4968 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,424B, BPFP=1.1006 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,300B, BPFP=1.4813 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,600B, BPFP=0.1913 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11195674 170.52745789 + layer.0.v_cache 0.00001346 0.01326579 + layer.1.k_cache 0.03755230 16.22564618 + layer.1.v_cache 0.00000535 0.00505979 + layer.2.k_cache 0.00239600 1.94924788 + layer.2.v_cache 0.00001753 0.01377494 + layer.3.k_cache 0.02867049 8.54165848 + layer.3.v_cache 0.00001751 0.01513865 + layer.4.k_cache 0.00071811 0.38575170 + layer.4.v_cache 0.00005486 0.03567255 + layer.4.output 0.19495771 704.00550788 + ------------------------------------------------------------------------------------- + TOTAL 0.09094743 301.51477818 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 61164 +BPFP 0.7301 bits/point +EBPFP 1.4602 equivalent bits/point +MSE 301.514778 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.003s, Pack+Encode: 0.164s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 301.5148 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,948B, BPFP=0.3382 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,936B, BPFP=1.5514 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,504B, BPFP=0.6083 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,228B, BPFP=1.4285 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,860B, BPFP=0.6701 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,700B, BPFP=1.3368 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,448B, BPFP=0.5986 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,332B, BPFP=1.4465 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,272B, BPFP=1.0889 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,980B, BPFP=1.3854 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,900B, BPFP=0.1711 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08657128 176.93187934 + layer.0.v_cache 0.00001574 0.01367522 + layer.1.k_cache 0.05618976 16.99788818 + layer.1.v_cache 0.00000506 0.00451581 + layer.2.k_cache 0.00392605 1.79089830 + layer.2.v_cache 0.00001570 0.01222904 + layer.3.k_cache 0.02849737 7.69334988 + layer.3.v_cache 0.00001769 0.01592743 + layer.4.k_cache 0.00099738 0.37314470 + layer.4.v_cache 0.00004751 0.03361771 + layer.4.output 0.15107292 602.25252976 + ------------------------------------------------------------------------------------- + TOTAL 0.07257612 259.97851964 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 67108 +BPFP 0.6853 bits/point +EBPFP 1.3707 equivalent bits/point +MSE 259.978520 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 259.9785 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,812B, BPFP=0.3539 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,644B, BPFP=1.6883 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,316B, BPFP=0.6477 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,968B, BPFP=1.5562 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,572B, BPFP=0.6977 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,372B, BPFP=1.4398 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,236B, BPFP=0.6320 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,664B, BPFP=1.4969 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,388B, BPFP=1.0523 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,532B, BPFP=1.4711 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,492B, BPFP=0.1811 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09254595 164.09948730 + layer.0.v_cache 0.00001776 0.01304047 + layer.1.k_cache 0.05775445 16.17884674 + layer.1.v_cache 0.00000504 0.00437987 + layer.2.k_cache 0.00891839 1.93481827 + layer.2.v_cache 0.00001732 0.01286714 + layer.3.k_cache 0.02836140 7.53549042 + layer.3.v_cache 0.00001672 0.01519988 + layer.4.k_cache 0.00083566 0.36160855 + layer.4.v_cache 0.00004876 0.03440848 + layer.4.output 0.16993865 677.82087054 + ------------------------------------------------------------------------------------- + TOTAL 0.08106423 290.29036711 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 62996 +BPFP 0.7238 bits/point +EBPFP 1.4475 equivalent bits/point +MSE 290.290367 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 290.2904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,792B, BPFP=0.2887 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,640B, BPFP=1.3918 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,520B, BPFP=0.5670 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,164B, BPFP=1.3151 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,952B, BPFP=0.6366 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,684B, BPFP=1.2378 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,692B, BPFP=0.5947 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,944B, BPFP=1.2796 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,440B, BPFP=1.0374 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,564B, BPFP=1.2184 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,092B, BPFP=0.1632 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13093194 168.21225636 + layer.0.v_cache 0.00001635 0.01294965 + layer.1.k_cache 0.05235285 16.39646298 + layer.1.v_cache 0.00000496 0.00441494 + layer.2.k_cache 0.00512381 1.78196512 + layer.2.v_cache 0.00001647 0.01244450 + layer.3.k_cache 0.01138228 8.73876890 + layer.3.v_cache 0.00001850 0.01477197 + layer.4.k_cache 0.00081880 0.35137900 + layer.4.v_cache 0.00004796 0.02981612 + layer.4.output 0.02447439 574.02144698 + ------------------------------------------------------------------------------------- + TOTAL 0.02188439 247.86502108 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 66484 +BPFP 0.6300 bits/point +EBPFP 1.2599 equivalent bits/point +MSE 247.865021 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 247.8650 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,788B, BPFP=0.3536 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,852B, BPFP=1.7508 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,240B, BPFP=0.6408 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,880B, BPFP=1.5585 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,408B, BPFP=0.6741 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,276B, BPFP=1.4391 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,132B, BPFP=0.6195 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,212B, BPFP=1.4264 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,404B, BPFP=1.0688 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,152B, BPFP=1.4146 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,880B, BPFP=0.1944 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510638 166.47362045 + layer.0.v_cache 0.00001399 0.01386156 + layer.1.k_cache 0.05467150 16.72546541 + layer.1.v_cache 0.00000511 0.00474932 + layer.2.k_cache 0.00545186 1.76035898 + layer.2.v_cache 0.00001652 0.01333868 + layer.3.k_cache 0.02986256 7.46407193 + layer.3.v_cache 0.00001747 0.01430014 + layer.4.k_cache 0.00068266 0.36119729 + layer.4.v_cache 0.00005277 0.03303712 + layer.4.output 0.17642128 686.80142405 + ------------------------------------------------------------------------------------- + TOTAL 0.08357822 294.14552760 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 62224 +BPFP 0.7239 bits/point +EBPFP 1.4479 equivalent bits/point +MSE 294.145528 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 294.1455 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,780B, BPFP=0.3612 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,220B, BPFP=1.6680 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,140B, BPFP=0.6372 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,728B, BPFP=1.5682 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,372B, BPFP=0.6843 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,000B, BPFP=1.4205 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,052B, BPFP=0.6193 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,636B, BPFP=1.5495 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,516B, BPFP=1.1193 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,732B, BPFP=1.5690 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,644B, BPFP=0.1926 +⌛️ [2/4] FRONTEND: Frontend time: 0.218s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08730992 173.54520089 + layer.0.v_cache 0.00001543 0.01408538 + layer.1.k_cache 0.05635426 16.38563121 + layer.1.v_cache 0.00000520 0.00511102 + layer.2.k_cache 0.00390536 1.95305604 + layer.2.v_cache 0.00001661 0.01360546 + layer.3.k_cache 0.06483487 6.98945281 + layer.3.v_cache 0.00002010 0.01593683 + layer.4.k_cache 0.00072652 0.38333843 + layer.4.v_cache 0.00004842 0.03436326 + layer.4.output 0.18845482 704.22912801 + ------------------------------------------------------------------------------------- + TOTAL 0.09014238 301.70256926 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 61820 +BPFP 0.7379 bits/point +EBPFP 1.4758 equivalent bits/point +MSE 301.702569 +---------------------- -------------------------------------------------------- +Time: 0.435s Load: 0.005s, Pack+Encode: 0.218s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 301.7026 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,616B, BPFP=0.3713 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,152B, BPFP=1.6434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,788B, BPFP=0.6406 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,624B, BPFP=1.5221 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,912B, BPFP=0.6691 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,948B, BPFP=1.3667 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,692B, BPFP=0.6186 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,508B, BPFP=1.4954 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,760B, BPFP=1.0938 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,412B, BPFP=1.4733 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,220B, BPFP=0.2370 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10784986 165.04499368 + layer.0.v_cache 0.00001387 0.01268230 + layer.1.k_cache 0.01409798 15.77117920 + layer.1.v_cache 0.00000489 0.00410538 + layer.2.k_cache 0.00254294 1.95411996 + layer.2.v_cache 0.00001653 0.01262920 + layer.3.k_cache 0.05709259 7.84678381 + layer.3.v_cache 0.00001736 0.01432940 + layer.4.k_cache 0.00062715 0.35298914 + layer.4.v_cache 0.00004758 0.03207444 + layer.4.output 0.22376562 797.67391019 + ------------------------------------------------------------------------------------- + TOTAL 0.10286295 339.69195634 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 54632 +BPFP 0.7384 bits/point +EBPFP 1.4769 equivalent bits/point +MSE 339.691956 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 339.6920 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,836B, BPFP=0.3631 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,320B, BPFP=1.6456 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,252B, BPFP=0.6432 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,596B, BPFP=1.5024 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,420B, BPFP=0.6764 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,100B, BPFP=1.4043 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,108B, BPFP=0.6147 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,584B, BPFP=1.5000 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,456B, BPFP=1.0791 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,788B, BPFP=1.3426 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,972B, BPFP=0.1970 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14563359 167.78406843 + layer.0.v_cache 0.00001368 0.01204082 + layer.1.k_cache 0.05947259 16.57423420 + layer.1.v_cache 0.00000496 0.00411768 + layer.2.k_cache 0.00553216 1.83250350 + layer.2.v_cache 0.00001557 0.01197928 + layer.3.k_cache 0.04377548 8.06885925 + layer.3.v_cache 0.00001730 0.01390593 + layer.4.k_cache 0.00079035 0.35156540 + layer.4.v_cache 0.00005153 0.03246482 + layer.4.output 0.17347779 686.95835217 + ------------------------------------------------------------------------------------- + TOTAL 0.08645010 294.31730615 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 61432 +BPFP 0.7147 bits/point +EBPFP 1.4294 equivalent bits/point +MSE 294.317306 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 294.3173 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,520B, BPFP=0.3598 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,288B, BPFP=1.7254 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,608B, BPFP=0.6174 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,924B, BPFP=1.6392 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,732B, BPFP=0.6468 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,948B, BPFP=1.4081 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,432B, BPFP=0.5758 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,152B, BPFP=1.4564 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,348B, BPFP=1.0294 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,032B, BPFP=1.4280 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,496B, BPFP=0.2197 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12373332 173.58068478 + layer.0.v_cache 0.00001436 0.01415777 + layer.1.k_cache 0.01270404 16.17343787 + layer.1.v_cache 0.00000490 0.00456604 + layer.2.k_cache 0.00252672 1.76996058 + layer.2.v_cache 0.00001629 0.01371045 + layer.3.k_cache 0.05185594 8.59041341 + layer.3.v_cache 0.00001849 0.01645048 + layer.4.k_cache 0.00068458 0.37128833 + layer.4.v_cache 0.00004975 0.03444630 + layer.4.output 0.22283003 822.28943452 + ------------------------------------------------------------------------------------- + TOTAL 0.10302462 350.38795045 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 52480 +BPFP 0.7308 bits/point +EBPFP 1.4617 equivalent bits/point +MSE 350.387950 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 350.3880 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.3452 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,864B, BPFP=1.6488 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,304B, BPFP=0.6146 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,144B, BPFP=1.5149 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,860B, BPFP=0.7180 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,536B, BPFP=1.4018 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,540B, BPFP=0.6585 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,996B, BPFP=1.4874 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,956B, BPFP=1.1079 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,744B, BPFP=1.4405 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,056B, BPFP=0.1875 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12002775 163.50573149 + layer.0.v_cache 0.00001466 0.01424415 + layer.1.k_cache 0.05271155 16.34442720 + layer.1.v_cache 0.00000524 0.00477064 + layer.2.k_cache 0.00808975 1.96113041 + layer.2.v_cache 0.00001826 0.01322346 + layer.3.k_cache 0.07501186 8.15039208 + layer.3.v_cache 0.00001895 0.01647130 + layer.4.k_cache 0.00071216 0.38834604 + layer.4.v_cache 0.00005170 0.03461026 + layer.4.output 0.17053346 645.58742560 + ------------------------------------------------------------------------------------- + TOTAL 0.08531742 277.03207801 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 65856 +BPFP 0.7206 bits/point +EBPFP 1.4412 equivalent bits/point +MSE 277.032078 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 277.0321 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 75, 128) +Output shape: (1, 75, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.output: torch.Size([1, 75, 3584]) -> torch.Size([1, 1, 75, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,776B, BPFP=0.3700 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,008B, BPFP=1.6683 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,152B, BPFP=0.6567 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,420B, BPFP=1.5458 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,424B, BPFP=0.7133 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,732B, BPFP=1.4025 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,980B, BPFP=0.6208 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,832B, BPFP=1.4233 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,232B, BPFP=1.0900 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,744B, BPFP=1.4050 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,992B, BPFP=0.2081 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08149677 167.78113281 + layer.0.v_cache 0.00001391 0.01484267 + layer.1.k_cache 0.05841089 15.58958333 + layer.1.v_cache 0.00000541 0.00505240 + layer.2.k_cache 0.00252637 1.86953084 + layer.2.v_cache 0.00001693 0.01389493 + layer.3.k_cache 0.02876982 8.11319987 + layer.3.v_cache 0.00001932 0.01588304 + layer.4.k_cache 0.00075246 0.37015704 + layer.4.v_cache 0.00005041 0.03605851 + layer.4.output 0.19266937 723.17309524 + ------------------------------------------------------------------------------------- + TOTAL 0.08945576 309.17770601 + (elements=652,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 652800 +Total Bytes 59292 +BPFP 0.7266 bits/point +EBPFP 1.4532 equivalent bits/point +MSE 309.177706 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 309.1777 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,616B, BPFP=0.3713 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,220B, BPFP=1.6590 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,804B, BPFP=0.6443 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,336B, BPFP=1.4559 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,920B, BPFP=0.6710 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,140B, BPFP=1.4108 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,620B, BPFP=0.6020 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,588B, BPFP=1.5138 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,496B, BPFP=1.0331 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,356B, BPFP=1.4605 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,860B, BPFP=0.2252 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10404026 161.96596392 + layer.0.v_cache 0.00001395 0.01396588 + layer.1.k_cache 0.01458383 16.00779814 + layer.1.v_cache 0.00000499 0.00458469 + layer.2.k_cache 0.00236740 1.91059943 + layer.2.v_cache 0.00001641 0.01331671 + layer.3.k_cache 0.03645969 7.52837192 + layer.3.v_cache 0.00002420 0.01544891 + layer.4.k_cache 0.00069996 0.38120287 + layer.4.v_cache 0.00005119 0.03661641 + layer.4.output 0.20980707 797.73654149 + ------------------------------------------------------------------------------------- + TOTAL 0.09570067 339.53139172 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 53956 +BPFP 0.7293 bits/point +EBPFP 1.4586 equivalent bits/point +MSE 339.531392 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 339.5314 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,656B, BPFP=0.3696 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,324B, BPFP=1.6348 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,048B, BPFP=0.6804 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,876B, BPFP=1.5348 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,080B, BPFP=0.6875 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,196B, BPFP=1.3830 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,736B, BPFP=0.6107 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,312B, BPFP=1.4089 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,944B, BPFP=1.1036 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,748B, BPFP=1.5063 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,036B, BPFP=0.2244 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11725240 164.26143973 + layer.0.v_cache 0.00001567 0.01410920 + layer.1.k_cache 0.03693308 15.87159947 + layer.1.v_cache 0.00000533 0.00530075 + layer.2.k_cache 0.00240566 1.72642375 + layer.2.v_cache 0.00001850 0.01527601 + layer.3.k_cache 0.06751672 7.66626936 + layer.3.v_cache 0.00001909 0.01609254 + layer.4.k_cache 0.00067984 0.36458639 + layer.4.v_cache 0.00004901 0.03500375 + layer.4.output 0.20094273 775.28724490 + ------------------------------------------------------------------------------------- + TOTAL 0.09597026 330.41098913 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 55956 +BPFP 0.7347 bits/point +EBPFP 1.4694 equivalent bits/point +MSE 330.410989 +---------------------- -------------------------------------------------------- +Time: 0.358s Load: 0.003s, Pack+Encode: 0.155s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 330.4110 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,648B, BPFP=0.3679 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,308B, BPFP=1.6313 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,116B, BPFP=0.6955 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,740B, BPFP=1.5045 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,072B, BPFP=0.6857 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,208B, BPFP=1.3857 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,804B, BPFP=0.6259 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,388B, BPFP=1.4259 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,904B, BPFP=1.0946 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,300B, BPFP=1.4062 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,688B, BPFP=0.2133 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07685234 161.98289621 + layer.0.v_cache 0.00001390 0.01343576 + layer.1.k_cache 0.01187498 15.92869001 + layer.1.v_cache 0.00000934 0.00504545 + layer.2.k_cache 0.00239603 1.87209429 + layer.2.v_cache 0.00001621 0.01262244 + layer.3.k_cache 0.07241479 7.80865130 + layer.3.v_cache 0.00001716 0.01516185 + layer.4.k_cache 0.00072067 0.34234540 + layer.4.v_cache 0.00004958 0.03258243 + layer.4.output 0.21417375 775.24279337 + ------------------------------------------------------------------------------------- + TOTAL 0.09785772 330.27723992 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 55176 +BPFP 0.7245 bits/point +EBPFP 1.4489 equivalent bits/point +MSE 330.277240 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 330.2772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,736B, BPFP=0.3569 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,204B, BPFP=1.6867 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,156B, BPFP=0.6488 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,260B, BPFP=1.4926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,324B, BPFP=0.6834 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,880B, BPFP=1.4145 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,040B, BPFP=0.6250 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,008B, BPFP=1.4408 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,088B, BPFP=1.0461 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,272B, BPFP=1.4951 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,836B, BPFP=0.2008 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09264422 175.42459910 + layer.0.v_cache 0.00001680 0.01487157 + layer.1.k_cache 0.05606840 15.87097168 + layer.1.v_cache 0.00000542 0.00552400 + layer.2.k_cache 0.00241068 1.93800876 + layer.2.v_cache 0.00001702 0.01413441 + layer.3.k_cache 0.15270803 7.71977154 + layer.3.v_cache 0.00001849 0.01805153 + layer.4.k_cache 0.00070970 0.39904788 + layer.4.v_cache 0.00005539 0.03725305 + layer.4.output 0.18838861 713.48607848 + ------------------------------------------------------------------------------------- + TOTAL 0.09549261 305.63792840 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 59804 +BPFP 0.7232 bits/point +EBPFP 1.4465 equivalent bits/point +MSE 305.637928 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 305.6379 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,784B, BPFP=0.3620 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,240B, BPFP=1.6721 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,112B, BPFP=0.6315 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,752B, BPFP=1.5731 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,320B, BPFP=0.6737 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,920B, BPFP=1.4042 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,056B, BPFP=0.6201 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,552B, BPFP=1.5325 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,248B, BPFP=1.0649 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,132B, BPFP=1.4472 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,428B, BPFP=0.1863 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10144231 174.73282772 + layer.0.v_cache 0.00001330 0.01274591 + layer.1.k_cache 0.09893422 16.17035626 + layer.1.v_cache 0.00000513 0.00481105 + layer.2.k_cache 0.00368564 1.77370780 + layer.2.v_cache 0.00001777 0.01299166 + layer.3.k_cache 0.06341467 7.78599647 + layer.3.v_cache 0.00001739 0.01558838 + layer.4.k_cache 0.00073693 0.38377301 + layer.4.v_cache 0.00004994 0.03350614 + layer.4.output 0.18261438 704.13532004 + ------------------------------------------------------------------------------------- + TOTAL 0.09097753 301.75726733 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 60544 +BPFP 0.7227 bits/point +EBPFP 1.4454 equivalent bits/point +MSE 301.757267 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 301.7573 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,940B, BPFP=0.3484 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,292B, BPFP=1.4892 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,424B, BPFP=0.6149 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,864B, BPFP=1.4124 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,748B, BPFP=0.6731 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,192B, BPFP=1.2917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,404B, BPFP=0.6114 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,896B, BPFP=1.4181 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,716B, BPFP=1.0266 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,296B, BPFP=1.3103 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,976B, BPFP=0.1790 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08947531 161.82834950 + layer.0.v_cache 0.00001432 0.01212090 + layer.1.k_cache 0.05320675 16.15479778 + layer.1.v_cache 0.00000497 0.00418553 + layer.2.k_cache 0.00562141 1.56676509 + layer.2.v_cache 0.00001574 0.01142807 + layer.3.k_cache 0.02742601 8.43894239 + layer.3.v_cache 0.00001759 0.01410120 + layer.4.k_cache 0.00114861 0.34386185 + layer.4.v_cache 0.00004734 0.03005601 + layer.4.output 0.15628038 623.11915025 + ------------------------------------------------------------------------------------- + TOTAL 0.07476122 267.66109765 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 63748 +BPFP 0.6735 bits/point +EBPFP 1.3469 equivalent bits/point +MSE 267.661098 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.155s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 267.6611 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,980B, BPFP=0.3363 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,116B, BPFP=1.5482 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,796B, BPFP=0.6447 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,240B, BPFP=1.3995 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,152B, BPFP=0.7052 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,172B, BPFP=1.3879 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,796B, BPFP=0.6447 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,468B, BPFP=1.4382 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,620B, BPFP=1.1243 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,772B, BPFP=1.3200 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,156B, BPFP=0.1736 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12311893 176.29498556 + layer.0.v_cache 0.00001357 0.01248594 + layer.1.k_cache 0.03743907 16.32870616 + layer.1.v_cache 0.00000515 0.00448186 + layer.2.k_cache 0.00511221 1.73159558 + layer.2.v_cache 0.00001545 0.01170095 + layer.3.k_cache 0.04225797 8.66998026 + layer.3.v_cache 0.00001799 0.01444009 + layer.4.k_cache 0.00119426 0.35308071 + layer.4.v_cache 0.00004621 0.02971397 + layer.4.output 0.14783020 589.18400621 + ------------------------------------------------------------------------------------- + TOTAL 0.07317837 254.57289497 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 69268 +BPFP 0.6920 bits/point +EBPFP 1.3840 equivalent bits/point +MSE 254.572895 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 254.5729 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,032B, BPFP=0.3113 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,540B, BPFP=1.4614 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,128B, BPFP=0.6324 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,148B, BPFP=1.4013 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,624B, BPFP=0.7083 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,148B, BPFP=1.4013 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,300B, BPFP=0.6587 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,196B, BPFP=1.4087 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,312B, BPFP=1.1201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,272B, BPFP=1.2672 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,344B, BPFP=0.1826 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09293266 156.11142387 + layer.0.v_cache 0.00001503 0.01315596 + layer.1.k_cache 0.07854857 15.76053634 + layer.1.v_cache 0.00000559 0.00455748 + layer.2.k_cache 0.00650115 1.67634149 + layer.2.v_cache 0.00001763 0.01228243 + layer.3.k_cache 0.03715762 8.49506752 + layer.3.v_cache 0.00001788 0.01466779 + layer.4.k_cache 0.00073365 0.31948052 + layer.4.v_cache 0.00004630 0.02839301 + layer.4.output 11.17680568 526.56867122 + ------------------------------------------------------------------------------------- + TOTAL 4.61491858 227.55391794 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 76044 +BPFP 0.6852 bits/point +EBPFP 1.3705 equivalent bits/point +MSE 227.553918 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 227.5539 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 96, 128) +Output shape: (1, 96, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.output: torch.Size([1, 96, 3584]) -> torch.Size([1, 1, 96, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,816B, BPFP=0.2956 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,012B, BPFP=1.4668 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,748B, BPFP=0.6100 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,716B, BPFP=1.4186 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,260B, BPFP=0.6934 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,384B, BPFP=1.3646 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,196B, BPFP=0.6829 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,724B, BPFP=1.4199 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,816B, BPFP=1.1094 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,952B, BPFP=1.2943 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,896B, BPFP=0.1603 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10158994 172.37721761 + layer.0.v_cache 0.00001467 0.01414357 + layer.1.k_cache 0.06416908 16.75825628 + layer.1.v_cache 0.00000542 0.00479732 + layer.2.k_cache 0.00400108 1.80474631 + layer.2.v_cache 0.00001872 0.01345979 + layer.3.k_cache 0.01491791 7.96084086 + layer.3.v_cache 0.00001877 0.01666048 + layer.4.k_cache 0.00070988 0.37601066 + layer.4.v_cache 0.00004832 0.03326837 + layer.4.output 0.14169503 564.81263951 + ------------------------------------------------------------------------------------- + TOTAL 0.06925641 244.29693399 + (elements=835,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 835584 +Total Bytes 70520 +BPFP 0.6752 bits/point +EBPFP 1.3503 equivalent bits/point +MSE 244.296934 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 244.2969 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,736B, BPFP=0.3767 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,660B, BPFP=1.6623 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,056B, BPFP=0.6632 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,820B, BPFP=1.4800 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,208B, BPFP=0.6962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,368B, BPFP=1.3819 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,868B, BPFP=0.6224 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,528B, BPFP=1.4167 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,208B, BPFP=1.1302 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,764B, BPFP=1.4679 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,684B, BPFP=0.2072 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10321087 161.96400282 + layer.0.v_cache 0.00001557 0.01391591 + layer.1.k_cache 0.06175087 16.01428053 + layer.1.v_cache 0.00000571 0.00478127 + layer.2.k_cache 0.01329466 1.79428503 + layer.2.v_cache 0.00001784 0.01395351 + layer.3.k_cache 0.05039416 7.78206550 + layer.3.v_cache 0.00001745 0.01572540 + layer.4.k_cache 0.00079786 0.37161652 + layer.4.v_cache 0.00005071 0.03483690 + layer.4.output 0.18870974 753.04439484 + ------------------------------------------------------------------------------------- + TOTAL 0.09120729 321.13648396 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 56900 +BPFP 0.7264 bits/point +EBPFP 1.4527 equivalent bits/point +MSE 321.136484 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 321.1365 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,744B, BPFP=0.3785 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,932B, BPFP=1.5043 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,080B, BPFP=0.6684 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,776B, BPFP=1.4705 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,176B, BPFP=0.6892 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,920B, BPFP=1.2847 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,904B, BPFP=0.6302 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,080B, BPFP=1.3194 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,784B, BPFP=1.0382 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,456B, BPFP=1.4010 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,208B, BPFP=0.2235 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09666174 155.48695204 + layer.0.v_cache 0.00001715 0.01332882 + layer.1.k_cache 0.08476152 15.65492757 + layer.1.v_cache 0.00000506 0.00453571 + layer.2.k_cache 0.00238475 1.82742924 + layer.2.v_cache 0.00001567 0.01248060 + layer.3.k_cache 0.07332428 7.74587758 + layer.3.v_cache 0.00001796 0.01568429 + layer.4.k_cache 0.00072290 0.33909257 + layer.4.v_cache 0.00005092 0.03000792 + layer.4.output 0.18875189 753.14794147 + ------------------------------------------------------------------------------------- + TOTAL 0.09289560 320.77446510 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 55060 +BPFP 0.7029 bits/point +EBPFP 1.4057 equivalent bits/point +MSE 320.774465 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 320.7745 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,608B, BPFP=0.3695 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,548B, BPFP=1.7344 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,772B, BPFP=0.6369 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,844B, BPFP=1.5726 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,016B, BPFP=0.6930 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,360B, BPFP=1.4614 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,772B, BPFP=0.6369 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,468B, BPFP=1.4862 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,684B, BPFP=1.0763 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,632B, BPFP=1.5239 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,260B, BPFP=0.2383 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10601801 163.18036248 + layer.0.v_cache 0.00001817 0.01461650 + layer.1.k_cache 0.01696859 16.14722398 + layer.1.v_cache 0.00000501 0.00497650 + layer.2.k_cache 0.00274148 1.80036814 + layer.2.v_cache 0.00001757 0.01419520 + layer.3.k_cache 0.01706722 7.81693582 + layer.3.v_cache 0.00001843 0.01564972 + layer.4.k_cache 0.00080906 0.37017110 + layer.4.v_cache 0.00005076 0.03384728 + layer.4.output 0.19983917 797.04227941 + ------------------------------------------------------------------------------------- + TOTAL 0.09074050 339.33495898 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 55964 +BPFP 0.7564 bits/point +EBPFP 1.5129 equivalent bits/point +MSE 339.334959 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 339.3350 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 65, 128) +Output shape: (1, 65, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.output: torch.Size([1, 65, 3584]) -> torch.Size([1, 1, 65, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,468B, BPFP=0.3529 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,120B, BPFP=1.7115 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,632B, BPFP=0.6327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,512B, BPFP=1.5654 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,732B, BPFP=0.6567 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,408B, BPFP=1.5404 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,340B, BPFP=0.5625 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,408B, BPFP=1.5404 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,788B, BPFP=1.1510 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,540B, BPFP=1.5721 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,904B, BPFP=0.2027 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09579474 172.62935697 + layer.0.v_cache 0.00001312 0.01384953 + layer.1.k_cache 0.01648773 16.64040715 + layer.1.v_cache 0.00000712 0.00459916 + layer.2.k_cache 0.00431219 1.84266228 + layer.2.v_cache 0.00001803 0.01374925 + layer.3.k_cache 0.03460605 7.90805946 + layer.3.v_cache 0.00001721 0.01595484 + layer.4.k_cache 0.00068162 0.38532421 + layer.4.v_cache 0.00004727 0.03560145 + layer.4.output 0.20898957 834.08489011 + ------------------------------------------------------------------------------------- + TOTAL 0.09499483 355.18139971 + (elements=565,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 565760 +Total Bytes 52852 +BPFP 0.7473 bits/point +EBPFP 1.4947 equivalent bits/point +MSE 355.181400 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.003s, Pack+Encode: 0.163s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 355.1814 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 41, 128) +Output shape: (1, 41, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.output: torch.Size([1, 41, 3584]) -> torch.Size([1, 1, 41, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,144B, BPFP=0.4360 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 3,924B, BPFP=1.4954 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 1,868B, BPFP=0.7119 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 3,676B, BPFP=1.4009 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 1,804B, BPFP=0.6875 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 3,584B, BPFP=1.3659 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,700B, BPFP=0.6479 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 3,628B, BPFP=1.3826 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 2,788B, BPFP=1.0625 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 3,480B, BPFP=1.3262 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,004B, BPFP=0.2724 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.143s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10098721 193.98668540 + layer.0.v_cache 0.00001362 0.01788012 + layer.1.k_cache 0.01708067 17.16714254 + layer.1.v_cache 0.00000533 0.00607759 + layer.2.k_cache 0.00250472 2.19485604 + layer.2.v_cache 0.00001755 0.01724164 + layer.3.k_cache 0.08199174 8.76487769 + layer.3.v_cache 0.00001815 0.01812156 + layer.4.k_cache 0.00059437 0.44859458 + layer.4.v_cache 0.00005232 0.04333493 + layer.4.output 0.35770382 1322.83983014 + ------------------------------------------------------------------------------------- + TOTAL 0.15924661 557.79668371 + (elements=356,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 356864 +Total Bytes 32600 +BPFP 0.7308 bits/point +EBPFP 1.4616 equivalent bits/point +MSE 557.796684 +---------------------- -------------------------------------------------------- +Time: 0.273s Load: 0.003s, Pack+Encode: 0.128s, Decode+Unpack: 0.143s +---------------------- -------------------------------------------------------- +💾 Converting with 557.7967 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,336B, BPFP=0.4014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,856B, BPFP=1.4591 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,264B, BPFP=0.6803 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,604B, BPFP=1.3834 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,240B, BPFP=0.6731 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,376B, BPFP=1.3149 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,108B, BPFP=0.6334 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,448B, BPFP=1.3365 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,404B, BPFP=1.0228 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,208B, BPFP=1.2644 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,356B, BPFP=0.2299 +⌛️ [2/4] FRONTEND: Frontend time: 0.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.143s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09210638 168.91218450 + layer.0.v_cache 0.00001369 0.01478946 + layer.1.k_cache 0.01682195 16.14776142 + layer.1.v_cache 0.00000520 0.00592843 + layer.2.k_cache 0.00487203 2.14614296 + layer.2.v_cache 0.00001755 0.01622684 + layer.3.k_cache 0.04965206 7.58662708 + layer.3.v_cache 0.00001944 0.01767030 + layer.4.k_cache 0.00060442 0.41213978 + layer.4.v_cache 0.00005207 0.04055854 + layer.4.output 0.26182200 1041.89405907 + ------------------------------------------------------------------------------------- + TOTAL 0.11746581 440.50343781 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 39200 +BPFP 0.6929 bits/point +EBPFP 1.3857 equivalent bits/point +MSE 440.503438 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.002s, Pack+Encode: 0.132s, Decode+Unpack: 0.143s +---------------------- -------------------------------------------------------- +💾 Converting with 440.5034 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 47, 128) +Output shape: (1, 47, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.output: torch.Size([1, 47, 3584]) -> torch.Size([1, 1, 47, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,276B, BPFP=0.4242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,764B, BPFP=1.5838 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,132B, BPFP=0.7088 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,600B, BPFP=1.5293 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,100B, BPFP=0.6981 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,332B, BPFP=1.4402 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,996B, BPFP=0.6636 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,448B, BPFP=1.4787 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,320B, BPFP=1.1037 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,164B, BPFP=1.3843 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,188B, BPFP=0.2464 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08461995 160.70236661 + layer.0.v_cache 0.00001366 0.01623306 + layer.1.k_cache 0.01838745 17.24405362 + layer.1.v_cache 0.00000609 0.00628999 + layer.2.k_cache 0.00780689 1.98315024 + layer.2.v_cache 0.00001932 0.01672737 + layer.3.k_cache 0.04292001 8.83582322 + layer.3.v_cache 0.00001942 0.01799199 + layer.4.k_cache 0.00059704 0.40731678 + layer.4.v_cache 0.00004850 0.03960231 + layer.4.output 0.29441516 1154.24373100 + ------------------------------------------------------------------------------------- + TOTAL 0.13031438 486.41033366 + (elements=409,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 409088 +Total Bytes 38320 +BPFP 0.7494 bits/point +EBPFP 1.4987 equivalent bits/point +MSE 486.410334 +---------------------- -------------------------------------------------------- +Time: 0.274s Load: 0.002s, Pack+Encode: 0.128s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 486.4103 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,384B, BPFP=0.3794 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,436B, BPFP=1.4901 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,364B, BPFP=0.6480 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,204B, BPFP=1.4265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,480B, BPFP=0.6798 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,916B, BPFP=1.3476 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,264B, BPFP=0.6206 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,984B, BPFP=1.3662 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,788B, BPFP=1.0384 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,824B, BPFP=1.3224 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,516B, BPFP=0.2160 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.143s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10653571 160.36006031 + layer.0.v_cache 0.00001396 0.01382906 + layer.1.k_cache 0.01513677 16.52280466 + layer.1.v_cache 0.00000574 0.00479763 + layer.2.k_cache 0.00241650 1.93221029 + layer.2.v_cache 0.00001997 0.01422272 + layer.3.k_cache 0.04250543 7.90271746 + layer.3.v_cache 0.00001889 0.01515678 + layer.4.k_cache 0.00060099 0.36234528 + layer.4.v_cache 0.00005317 0.03577241 + layer.4.output 0.25650722 951.12539160 + ------------------------------------------------------------------------------------- + TOTAL 0.11546222 402.64950928 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 43160 +BPFP 0.6959 bits/point +EBPFP 1.3919 equivalent bits/point +MSE 402.649509 +---------------------- -------------------------------------------------------- +Time: 0.274s Load: 0.002s, Pack+Encode: 0.129s, Decode+Unpack: 0.143s +---------------------- -------------------------------------------------------- +💾 Converting with 402.6495 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,276B, BPFP=0.4154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 4,768B, BPFP=1.5521 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,156B, BPFP=0.7018 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,464B, BPFP=1.4531 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,200B, BPFP=0.7161 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,272B, BPFP=1.3906 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 1,996B, BPFP=0.6497 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,336B, BPFP=1.4115 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,348B, BPFP=1.0898 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,416B, BPFP=1.4375 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,224B, BPFP=0.2429 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10088900 155.69284058 + layer.0.v_cache 0.00001321 0.01382701 + layer.1.k_cache 0.01601078 16.13207753 + layer.1.v_cache 0.00000529 0.00503891 + layer.2.k_cache 0.00240732 2.03764884 + layer.2.v_cache 0.00001858 0.01571620 + layer.3.k_cache 0.04236705 7.85604223 + layer.3.v_cache 0.00001968 0.01709569 + layer.4.k_cache 0.00061965 0.41979313 + layer.4.v_cache 0.00005235 0.03944288 + layer.4.output 0.30498679 1129.69335938 + ------------------------------------------------------------------------------------- + TOTAL 0.13513591 475.88723757 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 38456 +BPFP 0.7364 bits/point +EBPFP 1.4727 equivalent bits/point +MSE 475.887238 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.129s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 475.8872 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 53, 128) +Output shape: (1, 53, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.output: torch.Size([1, 53, 3584]) -> torch.Size([1, 1, 53, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,316B, BPFP=0.3880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,132B, BPFP=1.5130 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,252B, BPFP=0.6639 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,940B, BPFP=1.4564 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,312B, BPFP=0.6816 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,728B, BPFP=1.3939 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,160B, BPFP=0.6368 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,764B, BPFP=1.4045 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,644B, BPFP=1.0743 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,636B, BPFP=1.3667 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,392B, BPFP=0.2271 +⌛️ [2/4] FRONTEND: Frontend time: 0.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.167s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10150213 159.68993219 + layer.0.v_cache 0.00001370 0.01564462 + layer.1.k_cache 0.01753690 15.46654813 + layer.1.v_cache 0.00000536 0.00581426 + layer.2.k_cache 0.00234856 2.14438068 + layer.2.v_cache 0.00001946 0.01727389 + layer.3.k_cache 0.03772523 6.59774492 + layer.3.v_cache 0.00001792 0.01782130 + layer.4.k_cache 0.00061365 0.41200332 + layer.4.v_cache 0.00006478 0.04042741 + layer.4.output 0.25806696 1022.00876011 + ------------------------------------------------------------------------------------- + TOTAL 0.11566567 431.67464185 + (elements=461,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 461312 +Total Bytes 41276 +BPFP 0.7158 bits/point +EBPFP 1.4316 equivalent bits/point +MSE 431.674642 +---------------------- -------------------------------------------------------- +Time: 0.297s Load: 0.002s, Pack+Encode: 0.128s, Decode+Unpack: 0.167s +---------------------- -------------------------------------------------------- +💾 Converting with 431.6746 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,404B, BPFP=0.3656 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,816B, BPFP=1.5146 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,528B, BPFP=0.6583 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,400B, BPFP=1.4062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,556B, BPFP=0.6656 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,164B, BPFP=1.3448 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,388B, BPFP=0.6219 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,184B, BPFP=1.3500 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,984B, BPFP=1.0375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,988B, BPFP=1.2990 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,344B, BPFP=0.2360 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11850098 151.59023437 + layer.0.v_cache 0.00001385 0.01370371 + layer.1.k_cache 0.01468890 15.18467814 + layer.1.v_cache 0.00000523 0.00487930 + layer.2.k_cache 0.01477869 1.78366000 + layer.2.v_cache 0.00001915 0.01520792 + layer.3.k_cache 0.14661819 6.78069560 + layer.3.v_cache 0.00001925 0.01718236 + layer.4.k_cache 0.00060183 0.35404660 + layer.4.v_cache 0.00005752 0.03757556 + layer.4.output 0.22966646 903.65773810 + ------------------------------------------------------------------------------------- + TOTAL 0.11193934 382.43447237 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 45756 +BPFP 0.7009 bits/point +EBPFP 1.4018 equivalent bits/point +MSE 382.434472 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.002s, Pack+Encode: 0.129s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 382.4345 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,960B, BPFP=0.3329 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,424B, BPFP=1.6005 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,716B, BPFP=0.6311 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,844B, BPFP=1.5020 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,356B, BPFP=0.7398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,656B, BPFP=1.4701 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,804B, BPFP=0.6461 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,160B, BPFP=1.5557 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,060B, BPFP=1.1990 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,632B, BPFP=1.4660 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,568B, BPFP=0.1836 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10836371 171.46610691 + layer.0.v_cache 0.00001445 0.01391499 + layer.1.k_cache 0.06846340 16.05697234 + layer.1.v_cache 0.00000523 0.00449280 + layer.2.k_cache 0.00506382 1.76450829 + layer.2.v_cache 0.00001740 0.01266305 + layer.3.k_cache 0.03973729 8.96395343 + layer.3.v_cache 0.00001777 0.01529827 + layer.4.k_cache 0.00066797 0.37338547 + layer.4.v_cache 0.00005099 0.03569812 + layer.4.output 0.14786712 589.30784161 + ------------------------------------------------------------------------------------- + TOTAL 0.07396893 254.34481676 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 73180 +BPFP 0.7311 bits/point +EBPFP 1.4622 equivalent bits/point +MSE 254.344817 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 254.3448 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,500B, BPFP=0.3551 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,624B, BPFP=1.8049 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,600B, BPFP=0.6155 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,828B, BPFP=1.6165 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,780B, BPFP=0.6581 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,156B, BPFP=1.4574 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,480B, BPFP=0.5871 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,520B, BPFP=1.5436 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,600B, BPFP=1.0890 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,756B, BPFP=1.5994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,416B, BPFP=0.2170 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08728886 170.91959635 + layer.0.v_cache 0.00001734 0.01543552 + layer.1.k_cache 0.01499850 16.72379927 + layer.1.v_cache 0.00000528 0.00470524 + layer.2.k_cache 0.00240521 1.77507389 + layer.2.v_cache 0.00001739 0.01378279 + layer.3.k_cache 0.03126823 7.04631181 + layer.3.v_cache 0.00001843 0.01678376 + layer.4.k_cache 0.00061395 0.39933840 + layer.4.v_cache 0.00004830 0.03524902 + layer.4.output 0.21953388 822.31270292 + ------------------------------------------------------------------------------------- + TOTAL 0.09843639 350.18464685 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 54260 +BPFP 0.7556 bits/point +EBPFP 1.5113 equivalent bits/point +MSE 350.184647 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 350.1846 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,500B, BPFP=0.3551 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,568B, BPFP=1.7917 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,632B, BPFP=0.6231 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,092B, BPFP=1.6790 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,852B, BPFP=0.6752 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,508B, BPFP=1.5407 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,528B, BPFP=0.5985 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,024B, BPFP=1.6629 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,192B, BPFP=1.2292 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,236B, BPFP=1.7131 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,344B, BPFP=0.2146 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08223097 172.16215376 + layer.0.v_cache 0.00001442 0.01466537 + layer.1.k_cache 0.01480860 16.56618060 + layer.1.v_cache 0.00000502 0.00466686 + layer.2.k_cache 0.00576455 2.01089339 + layer.2.v_cache 0.00002304 0.01368860 + layer.3.k_cache 0.03215047 7.84144823 + layer.3.v_cache 0.00001933 0.01731137 + layer.4.k_cache 0.00061510 0.38712718 + layer.4.v_cache 0.00004916 0.03539327 + layer.4.output 0.22077460 822.31297348 + ------------------------------------------------------------------------------------- + TOTAL 0.09888840 350.30849077 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 56476 +BPFP 0.7865 bits/point +EBPFP 1.5730 equivalent bits/point +MSE 350.308491 +---------------------- -------------------------------------------------------- +Time: 0.372s Load: 0.003s, Pack+Encode: 0.165s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 350.3085 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,712B, BPFP=0.3664 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,964B, BPFP=1.7046 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,068B, BPFP=0.6567 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,188B, BPFP=1.5385 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,156B, BPFP=0.6755 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,456B, BPFP=1.3818 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,816B, BPFP=0.6027 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,712B, BPFP=1.4366 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,100B, BPFP=1.0916 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,004B, BPFP=1.4991 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,688B, BPFP=0.2045 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10114760 166.20186751 + layer.0.v_cache 0.00001420 0.01384756 + layer.1.k_cache 0.03683303 15.93798995 + layer.1.v_cache 0.00000544 0.00522590 + layer.2.k_cache 0.00248627 1.86676485 + layer.2.v_cache 0.00001713 0.01397375 + layer.3.k_cache 0.02909129 8.65602843 + layer.3.v_cache 0.00001766 0.01631187 + layer.4.k_cache 0.00072613 0.39807563 + layer.4.v_cache 0.00005154 0.03535456 + layer.4.output 0.19731793 743.33286448 + ------------------------------------------------------------------------------------- + TOTAL 0.09127152 317.43973479 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 57864 +BPFP 0.7285 bits/point +EBPFP 1.4571 equivalent bits/point +MSE 317.439735 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 317.4397 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,760B, BPFP=0.3618 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,140B, BPFP=1.6735 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,176B, BPFP=0.6530 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,420B, BPFP=1.5255 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,388B, BPFP=0.6965 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,540B, BPFP=1.3446 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,032B, BPFP=0.6234 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,700B, BPFP=1.3775 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,364B, BPFP=1.1028 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,084B, BPFP=1.4564 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,036B, BPFP=0.2066 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10230959 171.59174548 + layer.0.v_cache 0.00001409 0.01403579 + layer.1.k_cache 0.07973289 16.09728040 + layer.1.v_cache 0.00000506 0.00465126 + layer.2.k_cache 0.00553133 1.84072013 + layer.2.v_cache 0.00001928 0.01342238 + layer.3.k_cache 0.03946657 7.60762747 + layer.3.v_cache 0.00001980 0.01686397 + layer.4.k_cache 0.00066525 0.37640014 + layer.4.v_cache 0.00005119 0.03569054 + layer.4.output 0.17985767 713.52531720 + ------------------------------------------------------------------------------------- + TOTAL 0.08745993 305.42797988 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 59640 +BPFP 0.7213 bits/point +EBPFP 1.4425 equivalent bits/point +MSE 305.427980 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 305.4280 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,780B, BPFP=0.3612 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,880B, BPFP=1.5990 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,080B, BPFP=0.6250 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,000B, BPFP=1.4205 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,252B, BPFP=0.6599 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,620B, BPFP=1.3433 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,968B, BPFP=0.6023 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,796B, BPFP=1.3791 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,104B, BPFP=1.0357 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,724B, BPFP=1.3644 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,732B, BPFP=0.1952 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633976 172.70936485 + layer.0.v_cache 0.00001461 0.01388719 + layer.1.k_cache 0.07748085 16.59122933 + layer.1.v_cache 0.00000515 0.00446295 + layer.2.k_cache 0.00505460 1.85098683 + layer.2.v_cache 0.00001651 0.01262193 + layer.3.k_cache 0.02793909 7.68788127 + layer.3.v_cache 0.00001809 0.01479580 + layer.4.k_cache 0.00070819 0.37668273 + layer.4.v_cache 0.00004716 0.03175501 + layer.4.output 0.18765952 704.17862941 + ------------------------------------------------------------------------------------- + TOTAL 0.09007298 301.67906316 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 57936 +BPFP 0.6916 bits/point +EBPFP 1.3831 equivalent bits/point +MSE 301.679063 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 301.6791 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,632B, BPFP=0.3643 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,372B, BPFP=1.6455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,008B, BPFP=0.6714 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,040B, BPFP=1.5714 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,036B, BPFP=0.6777 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,652B, BPFP=1.4848 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,756B, BPFP=0.6152 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,644B, BPFP=1.4830 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,836B, BPFP=1.0795 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,628B, BPFP=1.4795 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,596B, BPFP=0.2103 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08399124 164.90058594 + layer.0.v_cache 0.00001671 0.01387642 + layer.1.k_cache 0.03849935 15.84732143 + layer.1.v_cache 0.00000503 0.00462347 + layer.2.k_cache 0.00738276 1.79037999 + layer.2.v_cache 0.00001692 0.01262715 + layer.3.k_cache 0.04914829 7.32424578 + layer.3.v_cache 0.00001713 0.01499480 + layer.4.k_cache 0.00063731 0.35550864 + layer.4.v_cache 0.00004711 0.03279569 + layer.4.output 0.21328995 775.14827806 + ------------------------------------------------------------------------------------- + TOTAL 0.09839950 330.37264151 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 56200 +BPFP 0.7379 bits/point +EBPFP 1.4758 equivalent bits/point +MSE 330.372642 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 330.3726 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 56, 128) +Output shape: (1, 56, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.output: torch.Size([1, 56, 3584]) -> torch.Size([1, 1, 56, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,372B, BPFP=0.3828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,324B, BPFP=1.4855 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,340B, BPFP=0.6529 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,044B, BPFP=1.4074 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,408B, BPFP=0.6719 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,720B, BPFP=1.3170 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,276B, BPFP=0.6350 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,964B, BPFP=1.3850 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,768B, BPFP=1.0513 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,648B, BPFP=1.2969 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,272B, BPFP=0.2101 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13768739 159.43727330 + layer.0.v_cache 0.00001432 0.01412993 + layer.1.k_cache 0.01638686 15.71203613 + layer.1.v_cache 0.00000533 0.00498423 + layer.2.k_cache 0.00251956 1.82134656 + layer.2.v_cache 0.00001745 0.01402675 + layer.3.k_cache 0.08599430 9.14250619 + layer.3.v_cache 0.00002012 0.01804183 + layer.4.k_cache 0.00061995 0.37389987 + layer.4.v_cache 0.00004798 0.03246558 + layer.4.output 0.24256272 967.69993622 + ------------------------------------------------------------------------------------- + TOTAL 0.11419131 409.43942729 + (elements=487,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 487424 +Total Bytes 42136 +BPFP 0.6916 bits/point +EBPFP 1.3831 equivalent bits/point +MSE 409.439427 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.129s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 409.4394 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,800B, BPFP=0.3516 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,520B, BPFP=1.6641 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,236B, BPFP=0.6320 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,360B, BPFP=1.4375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,452B, BPFP=0.6742 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,968B, BPFP=1.3609 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,144B, BPFP=0.6141 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,160B, BPFP=1.3984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,580B, BPFP=1.0898 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,184B, BPFP=1.4031 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,764B, BPFP=0.1887 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09278157 160.70502930 + layer.0.v_cache 0.00001504 0.01457602 + layer.1.k_cache 0.07523044 16.35033569 + layer.1.v_cache 0.00000517 0.00488376 + layer.2.k_cache 0.00381728 1.81614914 + layer.2.v_cache 0.00001789 0.01284992 + layer.3.k_cache 0.03327554 7.67462997 + layer.3.v_cache 0.00001939 0.01508378 + layer.4.k_cache 0.00067710 0.38400640 + layer.4.v_cache 0.00005192 0.03413948 + layer.4.output 0.16994183 677.81897321 + ------------------------------------------------------------------------------------- + TOTAL 0.08208730 290.10261741 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 61168 +BPFP 0.7028 bits/point +EBPFP 1.4055 equivalent bits/point +MSE 290.102617 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1026 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,380B, BPFP=0.3655 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,704B, BPFP=1.5106 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,508B, BPFP=0.6642 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,188B, BPFP=1.3739 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,516B, BPFP=0.6663 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,932B, BPFP=1.3061 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,372B, BPFP=0.6282 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,076B, BPFP=1.3443 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,864B, BPFP=1.0233 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,832B, BPFP=1.2797 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 5,700B, BPFP=0.2156 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.147s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12468526 162.35273769 + layer.0.v_cache 0.00001451 0.01495027 + layer.1.k_cache 0.01313243 14.49461235 + layer.1.v_cache 0.00000580 0.00524964 + layer.2.k_cache 0.00906488 1.71873151 + layer.2.v_cache 0.00002099 0.01469802 + layer.3.k_cache 0.01569995 7.24673307 + layer.3.v_cache 0.00002002 0.01713484 + layer.4.k_cache 0.00064964 0.37980616 + layer.4.v_cache 0.00005478 0.03570047 + layer.4.output 0.23031524 918.68530569 + ------------------------------------------------------------------------------------- + TOTAL 0.10444441 389.23985258 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 44072 +BPFP 0.6866 bits/point +EBPFP 1.3731 equivalent bits/point +MSE 389.239853 +---------------------- -------------------------------------------------------- +Time: 0.280s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.147s +---------------------- -------------------------------------------------------- +💾 Converting with 389.2399 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,412B, BPFP=0.3677 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,664B, BPFP=1.4750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,556B, BPFP=0.6656 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,380B, BPFP=1.4010 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,540B, BPFP=0.6615 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,084B, BPFP=1.3240 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,408B, BPFP=0.6271 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,152B, BPFP=1.3417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,928B, BPFP=1.0229 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,008B, BPFP=1.3042 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,244B, BPFP=0.2323 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15164433 154.72010091 + layer.0.v_cache 0.00001435 0.01534573 + layer.1.k_cache 0.01704682 15.41484172 + layer.1.v_cache 0.00000573 0.00529694 + layer.2.k_cache 0.00240009 1.95748405 + layer.2.v_cache 0.00001845 0.01446619 + layer.3.k_cache 0.02161660 7.53106384 + layer.3.v_cache 0.00001985 0.01758954 + layer.4.k_cache 0.00061459 0.37861112 + layer.4.v_cache 0.00005275 0.03916866 + layer.4.output 0.22653077 903.36927083 + ------------------------------------------------------------------------------------- + TOTAL 0.10465582 382.56934497 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 45376 +BPFP 0.6951 bits/point +EBPFP 1.3902 equivalent bits/point +MSE 382.569345 +---------------------- -------------------------------------------------------- +Time: 0.278s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 382.5693 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,720B, BPFP=0.3682 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,388B, BPFP=1.5813 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,080B, BPFP=0.6592 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,764B, BPFP=1.4478 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,068B, BPFP=0.6567 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,240B, BPFP=1.3356 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,924B, BPFP=0.6259 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,716B, BPFP=1.4375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,020B, BPFP=1.0745 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,560B, BPFP=1.4041 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,568B, BPFP=0.2008 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08869729 163.02144424 + layer.0.v_cache 0.00001594 0.01300693 + layer.1.k_cache 0.08105310 16.08208142 + layer.1.v_cache 0.00000494 0.00447026 + layer.2.k_cache 0.00256802 1.80690211 + layer.2.v_cache 0.00001661 0.01229738 + layer.3.k_cache 0.01835501 8.18196670 + layer.3.v_cache 0.00001861 0.01496024 + layer.4.k_cache 0.00078324 0.37963567 + layer.4.v_cache 0.00004769 0.03136719 + layer.4.output 0.18611241 742.81097114 + ------------------------------------------------------------------------------------- + TOTAL 0.08790278 317.01323118 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 56048 +BPFP 0.7057 bits/point +EBPFP 1.4114 equivalent bits/point +MSE 317.013231 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 317.0132 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,788B, BPFP=0.2880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,836B, BPFP=1.4233 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,564B, BPFP=0.5741 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,300B, BPFP=1.3370 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,416B, BPFP=0.7113 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,032B, BPFP=1.2938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,852B, BPFP=0.6205 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,236B, BPFP=1.3267 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,796B, BPFP=1.0947 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,756B, BPFP=1.2494 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,304B, BPFP=0.1681 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17228427 170.32357442 + layer.0.v_cache 0.00001543 0.01275154 + layer.1.k_cache 0.04621154 16.91422107 + layer.1.v_cache 0.00000545 0.00422733 + layer.2.k_cache 0.00739691 1.72548023 + layer.2.v_cache 0.00001797 0.01207748 + layer.3.k_cache 0.03096496 8.45490194 + layer.3.v_cache 0.00001890 0.01533280 + layer.4.k_cache 0.00071091 0.37132145 + layer.4.v_cache 0.00004900 0.02916695 + layer.4.output 0.02450962 574.03976436 + ------------------------------------------------------------------------------------- + TOTAL 0.02524957 248.00831798 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 68880 +BPFP 0.6527 bits/point +EBPFP 1.3053 equivalent bits/point +MSE 248.008318 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 248.0083 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,952B, BPFP=0.3280 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,480B, BPFP=1.5927 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,932B, BPFP=0.6606 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,952B, BPFP=1.5040 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,524B, BPFP=0.7601 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,284B, BPFP=1.3918 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,044B, BPFP=0.6794 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,968B, BPFP=1.5067 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,320B, BPFP=1.2298 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,576B, BPFP=1.4409 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,676B, BPFP=0.1842 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16174046 167.23071027 + layer.0.v_cache 0.00001612 0.01369141 + layer.1.k_cache 0.03352992 16.24863097 + layer.1.v_cache 0.00000563 0.00491919 + layer.2.k_cache 0.00496730 1.80086690 + layer.2.v_cache 0.00001740 0.01317273 + layer.3.k_cache 0.01521110 8.16752247 + layer.3.v_cache 0.00001939 0.01604896 + layer.4.k_cache 0.00066431 0.38451582 + layer.4.v_cache 0.00005067 0.03301934 + layer.4.output 0.14624557 582.96490975 + ------------------------------------------------------------------------------------- + TOTAL 0.07293772 251.45102743 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 73708 +BPFP 0.7285 bits/point +EBPFP 1.4569 equivalent bits/point +MSE 251.451027 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 251.4510 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,896B, BPFP=0.3445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,624B, BPFP=1.5669 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,456B, BPFP=0.6279 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,004B, BPFP=1.4542 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,760B, BPFP=0.6831 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,252B, BPFP=1.3176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,484B, BPFP=0.6330 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,568B, BPFP=1.3750 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,764B, BPFP=1.0472 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,464B, BPFP=1.3561 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,028B, BPFP=0.1824 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07121617 166.52456168 + layer.0.v_cache 0.00001681 0.01353297 + layer.1.k_cache 0.06877367 16.35883704 + layer.1.v_cache 0.00000530 0.00457771 + layer.2.k_cache 0.00486504 1.75792180 + layer.2.v_cache 0.00001749 0.01186846 + layer.3.k_cache 0.08833314 7.61357010 + layer.3.v_cache 0.00001723 0.01473419 + layer.4.k_cache 0.00077278 0.34984318 + layer.4.v_cache 0.00004595 0.03093901 + layer.4.output 0.15812202 630.33746885 + ------------------------------------------------------------------------------------- + TOTAL 0.07887751 270.88486283 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 64300 +BPFP 0.6872 bits/point +EBPFP 1.3744 equivalent bits/point +MSE 270.884863 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 270.8849 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,924B, BPFP=0.2976 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,816B, BPFP=1.5186 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,180B, BPFP=0.6467 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,268B, BPFP=1.4338 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,000B, BPFP=0.7735 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,016B, BPFP=1.3948 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,280B, BPFP=0.6621 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,188B, BPFP=1.4214 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,452B, BPFP=1.1528 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,380B, BPFP=1.2964 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,724B, BPFP=0.1707 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13964315 154.45018564 + layer.0.v_cache 0.00001432 0.01406903 + layer.1.k_cache 0.04810369 15.75427850 + layer.1.v_cache 0.00000572 0.00507246 + layer.2.k_cache 0.00492308 1.64621274 + layer.2.v_cache 0.00001777 0.01330915 + layer.3.k_cache 0.04255116 7.83498088 + layer.3.v_cache 0.00001880 0.01598338 + layer.4.k_cache 0.00073971 0.38426125 + layer.4.v_cache 0.00004573 0.03310012 + layer.4.output 11.28739116 531.77572489 + ------------------------------------------------------------------------------------- + TOTAL 4.66163537 229.56361926 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 76228 +BPFP 0.6937 bits/point +EBPFP 1.3874 equivalent bits/point +MSE 229.563619 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 229.5636 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,952B, BPFP=0.3466 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,000B, BPFP=1.5980 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,468B, BPFP=0.6158 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,316B, BPFP=1.4766 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,788B, BPFP=0.6726 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,760B, BPFP=1.3778 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,460B, BPFP=0.6143 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,912B, BPFP=1.4048 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,036B, BPFP=1.0717 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,020B, BPFP=1.4240 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,152B, BPFP=0.1814 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14145413 164.26433771 + layer.0.v_cache 0.00001516 0.01399292 + layer.1.k_cache 0.10647225 16.15937528 + layer.1.v_cache 0.00000560 0.00476364 + layer.2.k_cache 0.00244946 1.82465467 + layer.2.v_cache 0.00001685 0.01231337 + layer.3.k_cache 0.01581429 7.69142567 + layer.3.v_cache 0.00001841 0.01508885 + layer.4.k_cache 0.00072023 0.36010567 + layer.4.v_cache 0.00004652 0.02980716 + layer.4.output 0.15451367 615.79159903 + ------------------------------------------------------------------------------------- + TOTAL 0.07932992 264.75982695 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 66864 +BPFP 0.6984 bits/point +EBPFP 1.3967 equivalent bits/point +MSE 264.759827 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 264.7598 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,740B, BPFP=0.3674 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,840B, BPFP=1.6554 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,128B, BPFP=0.6605 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,104B, BPFP=1.5000 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,164B, BPFP=0.6681 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,572B, BPFP=1.3877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,928B, BPFP=0.6182 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,088B, BPFP=1.4966 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,120B, BPFP=1.0811 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,208B, BPFP=1.5220 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,472B, BPFP=0.1952 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.198s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10237078 169.76836993 + layer.0.v_cache 0.00001391 0.01345954 + layer.1.k_cache 0.05953315 15.99571104 + layer.1.v_cache 0.00000517 0.00462832 + layer.2.k_cache 0.00728044 1.89244286 + layer.2.v_cache 0.00001648 0.01277037 + layer.3.k_cache 0.03101032 8.41327564 + layer.3.v_cache 0.00001666 0.01549137 + layer.4.k_cache 0.00072555 0.36012222 + layer.4.v_cache 0.00004870 0.03257090 + layer.4.output 0.18364459 732.55773407 + ------------------------------------------------------------------------------------- + TOTAL 0.08744314 313.20076357 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 58364 +BPFP 0.7249 bits/point +EBPFP 1.4498 equivalent bits/point +MSE 313.200764 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.198s +---------------------- -------------------------------------------------------- +💾 Converting with 313.2008 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 111, 128) +Output shape: (1, 111, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.output: torch.Size([1, 111, 3584]) -> torch.Size([1, 1, 111, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,144B, BPFP=0.3018 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,188B, BPFP=1.4341 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,120B, BPFP=0.5800 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,664B, BPFP=1.3604 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,476B, BPFP=0.6301 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,000B, BPFP=1.2669 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,764B, BPFP=0.6706 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,360B, BPFP=1.3176 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,976B, BPFP=0.9820 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,728B, BPFP=1.2286 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,884B, BPFP=0.1585 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11107565 156.58593750 + layer.0.v_cache 0.00001597 0.01311023 + layer.1.k_cache 0.04333284 16.54688050 + layer.1.v_cache 0.00000558 0.00444205 + layer.2.k_cache 0.00477623 1.77992730 + layer.2.v_cache 0.00001758 0.01191004 + layer.3.k_cache 0.01020195 6.95750523 + layer.3.v_cache 0.00001870 0.01346775 + layer.4.k_cache 0.00079488 0.34585894 + layer.4.v_cache 0.00005041 0.03095970 + layer.4.output 10.27057757 483.65689350 + ------------------------------------------------------------------------------------- + TOTAL 4.23907840 209.87577963 + (elements=966,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 966144 +Total Bytes 77304 +BPFP 0.6401 bits/point +EBPFP 1.2802 equivalent bits/point +MSE 209.875780 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 209.8758 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,668B, BPFP=0.3671 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,756B, BPFP=1.7069 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,076B, BPFP=0.6769 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,876B, BPFP=1.5132 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,032B, BPFP=0.6673 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,004B, BPFP=1.3213 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,780B, BPFP=0.6118 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,316B, BPFP=1.3900 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,836B, BPFP=1.0643 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,488B, BPFP=1.4278 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,952B, BPFP=0.2186 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09375787 153.70265735 + layer.0.v_cache 0.00001622 0.01377486 + layer.1.k_cache 0.05984993 15.59642523 + layer.1.v_cache 0.00000512 0.00477780 + layer.2.k_cache 0.00243208 1.80384354 + layer.2.v_cache 0.00001865 0.01374537 + layer.3.k_cache 0.04944374 7.78671737 + layer.3.v_cache 0.00001908 0.01607617 + layer.4.k_cache 0.00084877 0.36574474 + layer.4.v_cache 0.00005139 0.03443983 + layer.4.output 0.19371969 763.75119467 + ------------------------------------------------------------------------------------- + TOTAL 0.09191063 325.03509205 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 55784 +BPFP 0.7221 bits/point +EBPFP 1.4443 equivalent bits/point +MSE 325.035092 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 325.0351 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 109, 128) +Output shape: (1, 109, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.output: torch.Size([1, 109, 3584]) -> torch.Size([1, 1, 109, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,120B, BPFP=0.3039 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,540B, BPFP=1.3675 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,120B, BPFP=0.5906 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,112B, BPFP=1.3062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,668B, BPFP=0.6692 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,588B, BPFP=1.2311 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,352B, BPFP=0.6239 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,984B, BPFP=1.2878 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,840B, BPFP=0.9805 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,384B, BPFP=1.2018 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,252B, BPFP=0.1690 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860953 161.66601562 + layer.0.v_cache 0.00001500 0.01300292 + layer.1.k_cache 0.05917679 16.65617273 + layer.1.v_cache 0.00000528 0.00471555 + layer.2.k_cache 0.01253135 1.70290466 + layer.2.v_cache 0.00001869 0.01298896 + layer.3.k_cache 0.07406004 7.45138914 + layer.3.v_cache 0.00001854 0.01518504 + layer.4.k_cache 0.00071234 0.35409903 + layer.4.v_cache 0.00004916 0.03217366 + layer.4.output 10.45909253 492.46580111 + ------------------------------------------------------------------------------------- + TOTAL 4.32287320 213.83348559 + (elements=948,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 948736 +Total Bytes 74960 +BPFP 0.6321 bits/point +EBPFP 1.2642 equivalent bits/point +MSE 213.833486 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 213.8335 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst + to output-fixed/kimiaudio/lambda0.004/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.7089 bits/point +Avg EBPFP 1.4178 equivalent bits/point +Avg MSE 325.955816 +Avg Time 0.350s +------------------------ ---------------------------- diff --git a/lambda0.007/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.007/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..7886bd99aca049e8670bd957b1e04c79a07eebdb --- /dev/null +++ b/lambda0.007/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 333 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- -------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench +Output output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench +---------------- -------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,600B, BPFP=0.5015 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,696B, BPFP=3.0278 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,228B, BPFP=0.8156 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,884B, BPFP=2.8711 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,140B, BPFP=0.9915 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,600B, BPFP=2.8164 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,636B, BPFP=0.8943 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,156B, BPFP=2.9236 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,776B, BPFP=2.0787 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,648B, BPFP=2.8256 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,360B, BPFP=0.2855 +⌛️ [2/4] FRONTEND: Frontend time: 3.051s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.422s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12089630 32.28008054 + layer.0.v_cache 0.00001394 0.00851398 + layer.1.k_cache 0.01174029 7.68651666 + layer.1.v_cache 0.00000579 0.00347127 + layer.2.k_cache 0.00835781 1.07636826 + layer.2.v_cache 0.00001852 0.01143221 + layer.3.k_cache 0.02896961 4.12392699 + layer.3.v_cache 0.00001872 0.01236843 + layer.4.k_cache 0.00063445 0.31101425 + layer.4.v_cache 0.00005063 0.02389580 + layer.4.output 0.17246709 653.00804674 + ------------------------------------------------------------------------------------- + TOTAL 0.08105739 271.56434797 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 112724 +BPFP 1.2791 bits/point +EBPFP 2.5582 equivalent bits/point +MSE 271.564348 +---------------------- -------------------------------------------------------- +Time: 4.478s Load: 0.006s, Pack+Encode: 3.051s, Decode+Unpack: 1.422s +---------------------- -------------------------------------------------------- +💾 Converting with 271.5643 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,576B, BPFP=0.5031 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,516B, BPFP=3.0305 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,456B, BPFP=0.8703 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,896B, BPFP=2.9094 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,244B, BPFP=1.0242 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,540B, BPFP=2.8398 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,120B, BPFP=1.0000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,832B, BPFP=2.8969 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,904B, BPFP=2.3250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,448B, BPFP=2.8219 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,140B, BPFP=0.2550 +⌛️ [2/4] FRONTEND: Frontend time: 2.039s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.254s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08029544 33.79498901 + layer.0.v_cache 0.00001376 0.00823542 + layer.1.k_cache 0.03632395 9.71218033 + layer.1.v_cache 0.00000563 0.00351647 + layer.2.k_cache 0.00522719 1.28301268 + layer.2.v_cache 0.00001997 0.01160795 + layer.3.k_cache 0.02707962 8.27893982 + layer.3.v_cache 0.00001843 0.01242046 + layer.4.k_cache 0.00063057 0.30961080 + layer.4.v_cache 0.00005085 0.02427261 + layer.4.output 0.18451144 665.19676339 + ------------------------------------------------------------------------------------- + TOTAL 0.08477915 277.04800761 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 112672 +BPFP 1.2945 bits/point +EBPFP 2.5890 equivalent bits/point +MSE 277.048008 +---------------------- -------------------------------------------------------- +Time: 3.297s Load: 0.004s, Pack+Encode: 2.039s, Decode+Unpack: 1.254s +---------------------- -------------------------------------------------------- +💾 Converting with 277.0480 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,668B, BPFP=0.4847 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,372B, BPFP=2.9746 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,392B, BPFP=0.7980 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,576B, BPFP=2.8299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,360B, BPFP=0.9738 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,580B, BPFP=2.8307 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,884B, BPFP=0.8874 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,640B, BPFP=2.8416 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,716B, BPFP=2.1286 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,788B, BPFP=2.6868 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,172B, BPFP=0.2640 +⌛️ [2/4] FRONTEND: Frontend time: 2.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.612s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10874708 32.42228379 + layer.0.v_cache 0.00001608 0.00826042 + layer.1.k_cache 0.03270756 8.36937341 + layer.1.v_cache 0.00000572 0.00354747 + layer.2.k_cache 0.00777350 1.05484541 + layer.2.v_cache 0.00002080 0.01176705 + layer.3.k_cache 0.05920834 5.02436438 + layer.3.v_cache 0.00001846 0.01220186 + layer.4.k_cache 0.00062493 0.32819089 + layer.4.v_cache 0.00005036 0.02338427 + layer.4.output 0.17137358 613.69440407 + ------------------------------------------------------------------------------------- + TOTAL 0.08286988 255.47759103 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 117148 +BPFP 1.2520 bits/point +EBPFP 2.5040 equivalent bits/point +MSE 255.477591 +---------------------- -------------------------------------------------------- +Time: 3.772s Load: 0.004s, Pack+Encode: 2.156s, Decode+Unpack: 1.612s +---------------------- -------------------------------------------------------- +💾 Converting with 255.4776 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,636B, BPFP=0.5280 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,640B, BPFP=2.9327 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,416B, BPFP=0.8846 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,636B, BPFP=2.9319 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,628B, BPFP=1.1274 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,656B, BPFP=2.7356 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,792B, BPFP=0.9599 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,244B, BPFP=2.8534 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,560B, BPFP=2.3157 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,068B, BPFP=2.8181 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,636B, BPFP=0.2758 +⌛️ [2/4] FRONTEND: Frontend time: 1.972s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12073612 35.82735815 + layer.0.v_cache 0.00001348 0.00842497 + layer.1.k_cache 0.03279715 10.97943272 + layer.1.v_cache 0.00000568 0.00362314 + layer.2.k_cache 0.00696802 1.18351922 + layer.2.v_cache 0.00001833 0.01128341 + layer.3.k_cache 0.03227850 5.15472647 + layer.3.v_cache 0.00001834 0.01206034 + layer.4.k_cache 0.00060115 0.32095968 + layer.4.v_cache 0.00005191 0.02483228 + layer.4.output 0.18320683 688.06330128 + ------------------------------------------------------------------------------------- + TOTAL 0.08681979 286.46878408 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 109912 +BPFP 1.2952 bits/point +EBPFP 2.5903 equivalent bits/point +MSE 286.468784 +---------------------- -------------------------------------------------------- +Time: 3.323s Load: 0.003s, Pack+Encode: 1.972s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 286.4688 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,648B, BPFP=0.5304 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,156B, BPFP=3.0361 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,552B, BPFP=0.9119 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,260B, BPFP=2.8566 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,404B, BPFP=1.2829 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,108B, BPFP=2.8261 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,260B, BPFP=1.0537 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,424B, BPFP=2.8894 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,052B, BPFP=2.0136 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,116B, BPFP=2.8277 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,672B, BPFP=0.2768 +⌛️ [2/4] FRONTEND: Frontend time: 1.873s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.285s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10595027 36.55903821 + layer.0.v_cache 0.00001376 0.00836699 + layer.1.k_cache 0.03307601 11.01818378 + layer.1.v_cache 0.00000550 0.00334345 + layer.2.k_cache 0.00695119 1.46702047 + layer.2.v_cache 0.00001929 0.01117191 + layer.3.k_cache 0.03450640 4.43847069 + layer.3.v_cache 0.00001844 0.01220485 + layer.4.k_cache 0.00061739 0.31100195 + layer.4.v_cache 0.00005297 0.02376979 + layer.4.output 0.18589628 688.78972070 + ------------------------------------------------------------------------------------- + TOTAL 0.08720501 286.78709512 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 110652 +BPFP 1.3039 bits/point +EBPFP 2.6077 equivalent bits/point +MSE 286.787095 +---------------------- -------------------------------------------------------- +Time: 3.161s Load: 0.003s, Pack+Encode: 1.873s, Decode+Unpack: 1.285s +---------------------- -------------------------------------------------------- +💾 Converting with 286.7871 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,580B, BPFP=0.5103 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,604B, BPFP=3.0862 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,500B, BPFP=0.8900 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,040B, BPFP=2.9747 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,244B, BPFP=1.2350 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,476B, BPFP=2.8631 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,512B, BPFP=0.8924 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,600B, BPFP=2.8877 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,960B, BPFP=2.3655 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,296B, BPFP=2.8275 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,412B, BPFP=0.2377 +⌛️ [2/4] FRONTEND: Frontend time: 1.899s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.335s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07772056 31.98093849 + layer.0.v_cache 0.00001351 0.00859599 + layer.1.k_cache 0.03518496 10.63023048 + layer.1.v_cache 0.00000556 0.00345576 + layer.2.k_cache 0.00230807 1.58467566 + layer.2.v_cache 0.00001833 0.01097428 + layer.3.k_cache 0.04514034 4.81983504 + layer.3.v_cache 0.00001901 0.01212976 + layer.4.k_cache 0.00062178 0.31691373 + layer.4.v_cache 0.00004990 0.02518165 + layer.4.output 0.18426369 665.60991184 + ------------------------------------------------------------------------------------- + TOTAL 0.08534870 276.98013610 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 112224 +BPFP 1.3057 bits/point +EBPFP 2.6113 equivalent bits/point +MSE 276.980136 +---------------------- -------------------------------------------------------- +Time: 3.238s Load: 0.004s, Pack+Encode: 1.899s, Decode+Unpack: 1.335s +---------------------- -------------------------------------------------------- +💾 Converting with 276.9801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,684B, BPFP=0.5446 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,508B, BPFP=3.1469 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,292B, BPFP=0.8709 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,952B, BPFP=3.0341 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,608B, BPFP=1.1380 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,444B, BPFP=2.9310 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,436B, BPFP=1.1031 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,900B, BPFP=3.0235 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,844B, BPFP=2.2005 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,216B, BPFP=2.8847 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,612B, BPFP=0.2786 +⌛️ [2/4] FRONTEND: Frontend time: 2.203s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.595s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14568528 31.67228084 + layer.0.v_cache 0.00001322 0.00817551 + layer.1.k_cache 0.01379078 10.65768789 + layer.1.v_cache 0.00000548 0.00341192 + layer.2.k_cache 0.00727733 1.35966541 + layer.2.v_cache 0.00001786 0.01130879 + layer.3.k_cache 0.03082054 5.38647263 + layer.3.v_cache 0.00001879 0.01263307 + layer.4.k_cache 0.00061820 0.32551055 + layer.4.v_cache 0.00005377 0.02597020 + layer.4.output 0.19662610 690.70251623 + ------------------------------------------------------------------------------------- + TOTAL 0.09262847 287.31651356 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 112496 +BPFP 1.3428 bits/point +EBPFP 2.6856 equivalent bits/point +MSE 287.316514 +---------------------- -------------------------------------------------------- +Time: 3.801s Load: 0.004s, Pack+Encode: 2.203s, Decode+Unpack: 1.595s +---------------------- -------------------------------------------------------- +💾 Converting with 287.3165 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,580B, BPFP=0.5039 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,456B, BPFP=3.0187 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,416B, BPFP=0.8625 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,912B, BPFP=2.9125 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,468B, BPFP=1.0680 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,328B, BPFP=2.7984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,076B, BPFP=0.9914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,852B, BPFP=2.9008 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,244B, BPFP=2.1961 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,720B, BPFP=2.8750 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,740B, BPFP=0.2718 +⌛️ [2/4] FRONTEND: Frontend time: 2.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10149734 34.80321350 + layer.0.v_cache 0.00001421 0.00886290 + layer.1.k_cache 0.01324784 8.52882233 + layer.1.v_cache 0.00000586 0.00375256 + layer.2.k_cache 0.00706157 1.32662163 + layer.2.v_cache 0.00001822 0.01073418 + layer.3.k_cache 0.02712160 7.98721695 + layer.3.v_cache 0.00002016 0.01224072 + layer.4.k_cache 0.00063687 0.32731490 + layer.4.v_cache 0.00005451 0.02388805 + layer.4.output 0.17858309 653.13409598 + ------------------------------------------------------------------------------------- + TOTAL 0.08233881 272.05713762 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 112792 +BPFP 1.2959 bits/point +EBPFP 2.5917 equivalent bits/point +MSE 272.057138 +---------------------- -------------------------------------------------------- +Time: 3.611s Load: 0.005s, Pack+Encode: 2.246s, Decode+Unpack: 1.359s +---------------------- -------------------------------------------------------- +💾 Converting with 272.0571 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,676B, BPFP=0.4806 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,564B, BPFP=2.9749 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,860B, BPFP=0.8728 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,824B, BPFP=2.8420 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,200B, BPFP=0.9339 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,644B, BPFP=2.8096 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,976B, BPFP=0.8937 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,824B, BPFP=2.8420 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,504B, BPFP=2.2457 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,516B, BPFP=2.7866 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,788B, BPFP=0.2511 +⌛️ [2/4] FRONTEND: Frontend time: 2.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.446s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10772817 30.92194796 + layer.0.v_cache 0.00001348 0.00846614 + layer.1.k_cache 0.03276526 10.71225361 + layer.1.v_cache 0.00000562 0.00345190 + layer.2.k_cache 0.01020141 1.04220195 + layer.2.v_cache 0.00001934 0.01063489 + layer.3.k_cache 0.04694761 4.24958082 + layer.3.v_cache 0.00001910 0.01206433 + layer.4.k_cache 0.00062564 0.32790120 + layer.4.v_cache 0.00005323 0.02253809 + layer.4.output 0.17166949 609.10745074 + ------------------------------------------------------------------------------------- + TOTAL 0.08235678 253.59195271 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 119376 +BPFP 1.2612 bits/point +EBPFP 2.5223 equivalent bits/point +MSE 253.591953 +---------------------- -------------------------------------------------------- +Time: 3.615s Load: 0.006s, Pack+Encode: 2.163s, Decode+Unpack: 1.446s +---------------------- -------------------------------------------------------- +💾 Converting with 253.5920 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,564B, BPFP=0.4827 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,256B, BPFP=3.0602 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,280B, BPFP=0.8057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,568B, BPFP=2.9307 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,332B, BPFP=1.0038 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,832B, BPFP=2.7922 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,752B, BPFP=0.8946 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,352B, BPFP=2.8901 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,200B, BPFP=2.2967 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,652B, BPFP=2.7583 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,316B, BPFP=0.2505 +⌛️ [2/4] FRONTEND: Frontend time: 1.949s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.476s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12335577 30.64030732 + layer.0.v_cache 0.00001400 0.00860607 + layer.1.k_cache 0.01519604 8.09932258 + layer.1.v_cache 0.00000547 0.00339677 + layer.2.k_cache 0.01098718 1.24333880 + layer.2.v_cache 0.00001779 0.01117089 + layer.3.k_cache 0.07416311 5.25136557 + layer.3.v_cache 0.00001944 0.01269659 + layer.4.k_cache 0.00061719 0.31200198 + layer.4.v_cache 0.00006444 0.02201425 + layer.4.output 0.17504946 634.58121773 + ------------------------------------------------------------------------------------- + TOTAL 0.08528157 263.98074970 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 115104 +BPFP 1.2746 bits/point +EBPFP 2.5493 equivalent bits/point +MSE 263.980750 +---------------------- -------------------------------------------------------- +Time: 3.429s Load: 0.004s, Pack+Encode: 1.949s, Decode+Unpack: 1.476s +---------------------- -------------------------------------------------------- +💾 Converting with 263.9807 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,680B, BPFP=0.5438 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,292B, BPFP=3.1031 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,428B, BPFP=0.8985 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,968B, BPFP=2.8344 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,288B, BPFP=1.0731 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,100B, BPFP=2.8612 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,324B, BPFP=0.8774 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,676B, BPFP=2.9781 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,496B, BPFP=2.3328 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,328B, BPFP=2.9075 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,084B, BPFP=0.2633 +⌛️ [2/4] FRONTEND: Frontend time: 2.074s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13955514 30.70091188 + layer.0.v_cache 0.00001392 0.00875477 + layer.1.k_cache 0.03454666 9.82794824 + layer.1.v_cache 0.00000526 0.00363960 + layer.2.k_cache 0.00236268 1.29928123 + layer.2.v_cache 0.00001660 0.01064056 + layer.3.k_cache 0.02874016 4.24035327 + layer.3.v_cache 0.00001899 0.01255710 + layer.4.k_cache 0.00062191 0.32775542 + layer.4.v_cache 0.00005265 0.02418883 + layer.4.output 0.18728454 698.36723098 + ------------------------------------------------------------------------------------- + TOTAL 0.08923092 290.29568516 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 109664 +BPFP 1.3090 bits/point +EBPFP 2.6180 equivalent bits/point +MSE 290.295685 +---------------------- -------------------------------------------------------- +Time: 3.281s Load: 0.005s, Pack+Encode: 2.074s, Decode+Unpack: 1.202s +---------------------- -------------------------------------------------------- +💾 Converting with 290.2957 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,660B, BPFP=0.5398 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,760B, BPFP=3.1981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,256B, BPFP=0.8636 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,272B, BPFP=2.8961 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,828B, BPFP=0.9797 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,636B, BPFP=2.9700 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,596B, BPFP=0.9326 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,632B, BPFP=2.9692 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,428B, BPFP=2.3190 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,144B, BPFP=2.8701 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,764B, BPFP=0.2830 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.414s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11454884 30.89802531 + layer.0.v_cache 0.00001351 0.00893926 + layer.1.k_cache 0.01665594 10.74319735 + layer.1.v_cache 0.00000570 0.00378612 + layer.2.k_cache 0.00836552 0.99217036 + layer.2.v_cache 0.00001974 0.01157044 + layer.3.k_cache 0.03124422 4.81914570 + layer.3.v_cache 0.00001876 0.01276937 + layer.4.k_cache 0.00062231 0.31718333 + layer.4.v_cache 0.00005487 0.02623531 + layer.4.output 0.19663499 690.43941327 + ------------------------------------------------------------------------------------- + TOTAL 0.09105849 287.11228914 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 110976 +BPFP 1.3247 bits/point +EBPFP 2.6494 equivalent bits/point +MSE 287.112289 +---------------------- -------------------------------------------------------- +Time: 3.274s Load: 0.004s, Pack+Encode: 1.856s, Decode+Unpack: 1.414s +---------------------- -------------------------------------------------------- +💾 Converting with 287.1123 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,584B, BPFP=0.4807 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,292B, BPFP=3.0305 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,384B, BPFP=0.8155 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,612B, BPFP=2.9040 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,952B, BPFP=0.9211 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,404B, BPFP=2.8653 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,704B, BPFP=0.8750 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,648B, BPFP=2.9107 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,068B, BPFP=2.2448 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,840B, BPFP=2.7604 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,700B, BPFP=0.2578 +⌛️ [2/4] FRONTEND: Frontend time: 2.367s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.493s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12290487 29.30471075 + layer.0.v_cache 0.00001422 0.00859082 + layer.1.k_cache 0.01283375 7.83175659 + layer.1.v_cache 0.00000579 0.00363068 + layer.2.k_cache 0.01100613 0.97590038 + layer.2.v_cache 0.00001929 0.01099665 + layer.3.k_cache 0.03011658 5.26462519 + layer.3.v_cache 0.00001858 0.01216878 + layer.4.k_cache 0.00063020 0.31823363 + layer.4.v_cache 0.00005223 0.02286312 + layer.4.output 0.17536174 620.92086522 + ------------------------------------------------------------------------------------- + TOTAL 0.08265493 258.24703136 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 116188 +BPFP 1.2713 bits/point +EBPFP 2.5426 equivalent bits/point +MSE 258.247031 +---------------------- -------------------------------------------------------- +Time: 3.865s Load: 0.005s, Pack+Encode: 2.367s, Decode+Unpack: 1.493s +---------------------- -------------------------------------------------------- +💾 Converting with 258.2470 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,596B, BPFP=0.4772 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,260B, BPFP=2.9890 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,336B, BPFP=0.7971 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,496B, BPFP=2.8485 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,432B, BPFP=0.9985 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,452B, BPFP=2.8404 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,708B, BPFP=0.8654 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,420B, BPFP=2.8346 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,452B, BPFP=2.1051 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,940B, BPFP=2.7463 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,764B, BPFP=0.2564 +⌛️ [2/4] FRONTEND: Frontend time: 1.944s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11218034 30.87873679 + layer.0.v_cache 0.00001438 0.00861185 + layer.1.k_cache 0.03157643 7.85438161 + layer.1.v_cache 0.00000545 0.00368280 + layer.2.k_cache 0.00802605 1.14329762 + layer.2.v_cache 0.00001826 0.01118281 + layer.3.k_cache 0.04519741 3.76594418 + layer.3.v_cache 0.00001794 0.01231757 + layer.4.k_cache 0.00061137 0.33197731 + layer.4.v_cache 0.00005913 0.02350262 + layer.4.output 0.17010492 620.89747899 + ------------------------------------------------------------------------------------- + TOTAL 0.08167301 258.25388165 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 115856 +BPFP 1.2528 bits/point +EBPFP 2.5055 equivalent bits/point +MSE 258.253882 +---------------------- -------------------------------------------------------- +Time: 3.261s Load: 0.004s, Pack+Encode: 1.944s, Decode+Unpack: 1.313s +---------------------- -------------------------------------------------------- +💾 Converting with 258.2539 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,664B, BPFP=0.4840 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,264B, BPFP=2.9549 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,392B, BPFP=0.7980 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,508B, BPFP=2.8176 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,292B, BPFP=0.9615 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,244B, BPFP=2.7696 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,100B, BPFP=0.9266 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,472B, BPFP=2.8110 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,588B, BPFP=2.1054 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,104B, BPFP=2.7442 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,072B, BPFP=0.2614 +⌛️ [2/4] FRONTEND: Frontend time: 1.889s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.442s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10146660 34.19216706 + layer.0.v_cache 0.00001344 0.00806585 + layer.1.k_cache 0.03090366 8.25895939 + layer.1.v_cache 0.00000566 0.00345112 + layer.2.k_cache 0.01188999 0.95465070 + layer.2.v_cache 0.00001845 0.01072868 + layer.3.k_cache 0.02542871 4.47540993 + layer.3.v_cache 0.00001884 0.01227134 + layer.4.k_cache 0.00060243 0.32919280 + layer.4.v_cache 0.00005224 0.02351222 + layer.4.output 0.16839122 617.75342608 + ------------------------------------------------------------------------------------- + TOTAL 0.07936109 257.20837598 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 116700 +BPFP 1.2472 bits/point +EBPFP 2.4944 equivalent bits/point +MSE 257.208376 +---------------------- -------------------------------------------------------- +Time: 3.335s Load: 0.004s, Pack+Encode: 1.889s, Decode+Unpack: 1.442s +---------------------- -------------------------------------------------------- +💾 Converting with 257.2084 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,712B, BPFP=0.4606 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,668B, BPFP=2.8308 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,924B, BPFP=1.0061 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,156B, BPFP=2.7439 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,992B, BPFP=1.0177 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,608B, BPFP=2.6508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,472B, BPFP=0.9293 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,144B, BPFP=2.7418 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,440B, BPFP=2.1128 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,748B, BPFP=2.6746 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,720B, BPFP=0.2358 +⌛️ [2/4] FRONTEND: Frontend time: 1.912s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.631s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12981709 30.18341330 + layer.0.v_cache 0.00001628 0.00844221 + layer.1.k_cache 0.06761242 13.55403734 + layer.1.v_cache 0.00000547 0.00323810 + layer.2.k_cache 0.00384689 1.51403543 + layer.2.v_cache 0.00001812 0.01085127 + layer.3.k_cache 0.05327560 8.51002436 + layer.3.v_cache 0.00001896 0.01189916 + layer.4.k_cache 0.00060800 0.29881552 + layer.4.v_cache 0.00005116 0.02149742 + layer.4.output 0.15518735 576.50460986 + ------------------------------------------------------------------------------------- + TOTAL 0.07891655 240.56756018 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 122584 +BPFP 1.2247 bits/point +EBPFP 2.4493 equivalent bits/point +MSE 240.567560 +---------------------- -------------------------------------------------------- +Time: 3.547s Load: 0.004s, Pack+Encode: 1.912s, Decode+Unpack: 1.631s +---------------------- -------------------------------------------------------- +💾 Converting with 240.5676 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,588B, BPFP=0.4992 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,848B, BPFP=3.0571 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,232B, BPFP=0.8164 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,180B, BPFP=2.9282 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,172B, BPFP=0.9977 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,836B, BPFP=2.8619 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,716B, BPFP=0.9097 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,412B, BPFP=2.9730 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,788B, BPFP=2.2739 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,864B, BPFP=2.8673 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,228B, BPFP=0.2543 +⌛️ [2/4] FRONTEND: Frontend time: 1.989s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.452s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887507 29.41532841 + layer.0.v_cache 0.00001345 0.00860976 + layer.1.k_cache 0.03259810 7.43721668 + layer.1.v_cache 0.00000523 0.00334950 + layer.2.k_cache 0.00244350 1.00049167 + layer.2.v_cache 0.00001834 0.01088076 + layer.3.k_cache 0.10864470 5.49999849 + layer.3.v_cache 0.00001904 0.01182296 + layer.4.k_cache 0.00060864 0.29984053 + layer.4.v_cache 0.00004940 0.02404423 + layer.4.output 0.18628448 647.77849427 + ------------------------------------------------------------------------------------- + TOTAL 0.09395687 269.30359076 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 113864 +BPFP 1.2920 bits/point +EBPFP 2.5841 equivalent bits/point +MSE 269.303591 +---------------------- -------------------------------------------------------- +Time: 3.447s Load: 0.005s, Pack+Encode: 1.989s, Decode+Unpack: 1.452s +---------------------- -------------------------------------------------------- +💾 Converting with 269.3036 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,716B, BPFP=0.4768 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,612B, BPFP=2.9164 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,084B, BPFP=0.8926 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,072B, BPFP=2.8216 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,516B, BPFP=0.9684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,704B, BPFP=2.7570 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,948B, BPFP=0.8687 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,840B, BPFP=2.7809 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,100B, BPFP=2.1243 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,100B, BPFP=2.6510 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,832B, BPFP=0.2717 +⌛️ [2/4] FRONTEND: Frontend time: 2.016s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422350 32.32381825 + layer.0.v_cache 0.00001493 0.00850011 + layer.1.k_cache 0.01422887 10.53944517 + layer.1.v_cache 0.00000595 0.00352474 + layer.2.k_cache 0.00668612 1.04058889 + layer.2.v_cache 0.00001902 0.01074495 + layer.3.k_cache 0.02497406 4.07466623 + layer.3.v_cache 0.00001869 0.01214295 + layer.4.k_cache 0.00061935 0.30950551 + layer.4.v_cache 0.00005212 0.02395775 + layer.4.output 0.16899524 591.11205859 + ------------------------------------------------------------------------------------- + TOTAL 0.07904761 246.24301792 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 120524 +BPFP 1.2447 bits/point +EBPFP 2.4893 equivalent bits/point +MSE 246.243018 +---------------------- -------------------------------------------------------- +Time: 3.393s Load: 0.004s, Pack+Encode: 2.016s, Decode+Unpack: 1.373s +---------------------- -------------------------------------------------------- +💾 Converting with 246.2430 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 113, 128) +Output shape: (1, 113, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.output: torch.Size([1, 113, 3584]) -> torch.Size([1, 1, 113, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,264B, BPFP=0.4513 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,208B, BPFP=2.2412 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,828B, BPFP=0.8059 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,944B, BPFP=2.2046 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,336B, BPFP=0.8761 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,748B, BPFP=2.1775 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,088B, BPFP=0.8418 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,036B, BPFP=2.2174 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,056B, BPFP=2.2201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,556B, BPFP=2.1510 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,204B, BPFP=0.2213 +⌛️ [2/4] FRONTEND: Frontend time: 2.012s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.270s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10214825 29.83241583 + layer.0.v_cache 0.00001637 0.00813537 + layer.1.k_cache 0.04284562 8.58960771 + layer.1.v_cache 0.00000578 0.00336812 + layer.2.k_cache 0.00913706 1.21785959 + layer.2.v_cache 0.00001928 0.01043682 + layer.3.k_cache 0.05585717 4.55684047 + layer.3.v_cache 0.00001870 0.01092755 + layer.4.k_cache 0.00065968 0.31534418 + layer.4.v_cache 0.00005753 0.02203632 + layer.4.output 10.09706179 469.22064633 + ------------------------------------------------------------------------------------- + TOTAL 4.17001165 195.83008802 + (elements=983,552) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 983552 +Total Bytes 128268 +BPFP 1.0433 bits/point +EBPFP 2.0866 equivalent bits/point +MSE 195.830088 +---------------------- -------------------------------------------------------- +Time: 3.288s Load: 0.006s, Pack+Encode: 2.012s, Decode+Unpack: 1.270s +---------------------- -------------------------------------------------------- +💾 Converting with 195.8301 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,676B, BPFP=0.4595 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,700B, BPFP=2.8674 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,288B, BPFP=0.9080 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,036B, BPFP=2.7534 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,584B, BPFP=1.1305 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,684B, BPFP=2.6930 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,388B, BPFP=0.9251 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,972B, BPFP=2.7424 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,112B, BPFP=2.2514 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,492B, BPFP=2.6600 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,828B, BPFP=0.2656 +⌛️ [2/4] FRONTEND: Frontend time: 2.078s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11481400 32.17666123 + layer.0.v_cache 0.00001396 0.00850517 + layer.1.k_cache 0.03135788 11.69288719 + layer.1.v_cache 0.00000545 0.00349904 + layer.2.k_cache 0.00370715 1.56397457 + layer.2.v_cache 0.00001965 0.01103796 + layer.3.k_cache 0.03930106 7.74834132 + layer.3.v_cache 0.00001886 0.01192380 + layer.4.k_cache 0.00060819 0.30841171 + layer.4.v_cache 0.00005437 0.02241806 + layer.4.output 0.16591480 579.86096939 + ------------------------------------------------------------------------------------- + TOTAL 0.07948848 241.91614387 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 123760 +BPFP 1.2500 bits/point +EBPFP 2.5000 equivalent bits/point +MSE 241.916144 +---------------------- -------------------------------------------------------- +Time: 3.425s Load: 0.007s, Pack+Encode: 2.078s, Decode+Unpack: 1.341s +---------------------- -------------------------------------------------------- +💾 Converting with 241.9161 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,720B, BPFP=0.4775 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,512B, BPFP=2.8989 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,744B, BPFP=0.8329 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,868B, BPFP=2.7858 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,052B, BPFP=0.8869 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,672B, BPFP=2.7514 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,584B, BPFP=0.8048 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,996B, BPFP=2.8083 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,840B, BPFP=2.2542 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,644B, BPFP=2.7465 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,208B, BPFP=0.2309 +⌛️ [2/4] FRONTEND: Frontend time: 2.030s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09788724 32.49262915 + layer.0.v_cache 0.00001308 0.00822821 + layer.1.k_cache 0.03299916 9.27393675 + layer.1.v_cache 0.00000544 0.00352087 + layer.2.k_cache 0.00367894 0.96425774 + layer.2.v_cache 0.00002012 0.01101957 + layer.3.k_cache 0.03939044 3.00395426 + layer.3.v_cache 0.00001875 0.01185502 + layer.4.k_cache 0.00061788 0.30473602 + layer.4.v_cache 0.00005010 0.02332929 + layer.4.output 0.17021053 602.01108547 + ------------------------------------------------------------------------------------- + TOTAL 0.08036205 250.59853325 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 118840 +BPFP 1.2273 bits/point +EBPFP 2.4546 equivalent bits/point +MSE 250.598533 +---------------------- -------------------------------------------------------- +Time: 3.350s Load: 0.005s, Pack+Encode: 2.030s, Decode+Unpack: 1.315s +---------------------- -------------------------------------------------------- +💾 Converting with 250.5985 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,168B, BPFP=0.4583 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,212B, BPFP=2.3455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,428B, BPFP=0.7853 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,096B, BPFP=2.3287 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,548B, BPFP=0.9473 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,984B, BPFP=2.3125 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,264B, BPFP=0.9062 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,344B, BPFP=2.3646 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,740B, BPFP=2.1325 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,768B, BPFP=2.2812 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,828B, BPFP=0.3065 +⌛️ [2/4] FRONTEND: Frontend time: 1.982s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.683s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12672405 30.78991247 + layer.0.v_cache 0.00001401 0.00796703 + layer.1.k_cache 0.12044146 10.52786594 + layer.1.v_cache 0.00000568 0.00330071 + layer.2.k_cache 0.00570256 1.35406000 + layer.2.v_cache 0.00001866 0.01071785 + layer.3.k_cache 0.03611955 8.05994104 + layer.3.v_cache 0.00001983 0.01174698 + layer.4.k_cache 0.00066233 0.32598435 + layer.4.v_cache 0.00005331 0.02306659 + layer.4.output 10.56037249 480.53869048 + ------------------------------------------------------------------------------------- + TOTAL 4.36543346 200.87561155 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 131380 +BPFP 1.1181 bits/point +EBPFP 2.2362 equivalent bits/point +MSE 200.875612 +---------------------- -------------------------------------------------------- +Time: 3.670s Load: 0.004s, Pack+Encode: 1.982s, Decode+Unpack: 1.683s +---------------------- -------------------------------------------------------- +💾 Converting with 200.8756 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,696B, BPFP=0.4787 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,316B, BPFP=2.8970 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,072B, BPFP=0.9006 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,028B, BPFP=2.8459 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,264B, BPFP=0.9347 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,384B, BPFP=2.7315 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,148B, BPFP=0.9141 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,972B, BPFP=2.8359 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,716B, BPFP=2.2578 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,968B, BPFP=2.6577 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,944B, BPFP=0.2269 +⌛️ [2/4] FRONTEND: Frontend time: 1.972s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.285s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13733016 32.01185747 + layer.0.v_cache 0.00001373 0.00863508 + layer.1.k_cache 0.03155638 11.64084140 + layer.1.v_cache 0.00000617 0.00372428 + layer.2.k_cache 0.00751131 1.02381255 + layer.2.v_cache 0.00001754 0.01115142 + layer.3.k_cache 0.02548880 7.34175942 + layer.3.v_cache 0.00001765 0.01266183 + layer.4.k_cache 0.00061011 0.30582814 + layer.4.v_cache 0.00005045 0.02373240 + layer.4.output 0.16239834 606.89280641 + ------------------------------------------------------------------------------------- + TOTAL 0.07878769 252.97844993 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 118508 +BPFP 1.2378 bits/point +EBPFP 2.4755 equivalent bits/point +MSE 252.978450 +---------------------- -------------------------------------------------------- +Time: 3.263s Load: 0.005s, Pack+Encode: 1.972s, Decode+Unpack: 1.285s +---------------------- -------------------------------------------------------- +💾 Converting with 252.9784 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,632B, BPFP=0.5206 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,288B, BPFP=3.0237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,436B, BPFP=0.8774 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,740B, BPFP=2.9153 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,836B, BPFP=1.1543 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,328B, BPFP=2.8339 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,000B, BPFP=0.9889 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,792B, BPFP=2.9256 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,232B, BPFP=2.2215 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,360B, BPFP=2.8402 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,432B, BPFP=0.2665 +⌛️ [2/4] FRONTEND: Frontend time: 1.934s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.302s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10979201 36.74627299 + layer.0.v_cache 0.00001339 0.00877312 + layer.1.k_cache 0.01160004 10.97278064 + layer.1.v_cache 0.00000567 0.00343537 + layer.2.k_cache 0.00714037 1.32551488 + layer.2.v_cache 0.00001829 0.01061129 + layer.3.k_cache 0.04622446 7.22081592 + layer.3.v_cache 0.00001896 0.01235155 + layer.4.k_cache 0.00064378 0.30752324 + layer.4.v_cache 0.00005251 0.02420348 + layer.4.output 0.18415536 659.40636302 + ------------------------------------------------------------------------------------- + TOTAL 0.08615276 274.85157786 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 112076 +BPFP 1.3039 bits/point +EBPFP 2.6079 equivalent bits/point +MSE 274.851578 +---------------------- -------------------------------------------------------- +Time: 3.239s Load: 0.003s, Pack+Encode: 1.934s, Decode+Unpack: 1.302s +---------------------- -------------------------------------------------------- +💾 Converting with 274.8516 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,564B, BPFP=0.5203 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,372B, BPFP=3.1193 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,392B, BPFP=0.8912 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,808B, BPFP=3.0049 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,048B, BPFP=1.0244 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,048B, BPFP=2.8506 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,092B, BPFP=1.0333 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,708B, BPFP=2.9846 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,724B, BPFP=2.1761 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,392B, BPFP=2.9205 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,356B, BPFP=0.3002 +⌛️ [2/4] FRONTEND: Frontend time: 1.833s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.259s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11290005 31.04714451 + layer.0.v_cache 0.00001570 0.00855228 + layer.1.k_cache 0.03434501 10.34199980 + layer.1.v_cache 0.00000534 0.00341145 + layer.2.k_cache 0.00238675 1.01779631 + layer.2.v_cache 0.00001838 0.01091569 + layer.3.k_cache 0.02838124 8.22408701 + layer.3.v_cache 0.00002047 0.01267068 + layer.4.k_cache 0.00062757 0.31399932 + layer.4.v_cache 0.00005455 0.02474622 + layer.4.output 0.18420308 686.36630334 + ------------------------------------------------------------------------------------- + TOTAL 0.08636333 285.62173216 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 111504 +BPFP 1.3310 bits/point +EBPFP 2.6620 equivalent bits/point +MSE 285.621732 +---------------------- -------------------------------------------------------- +Time: 3.095s Load: 0.003s, Pack+Encode: 1.833s, Decode+Unpack: 1.259s +---------------------- -------------------------------------------------------- +💾 Converting with 285.6217 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,624B, BPFP=0.5256 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,072B, BPFP=3.0192 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,548B, BPFP=0.9111 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,472B, BPFP=2.8990 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,680B, BPFP=1.1378 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,216B, BPFP=2.8478 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,056B, BPFP=1.0128 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,816B, BPFP=2.9679 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,172B, BPFP=2.2380 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,288B, BPFP=2.8622 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,116B, BPFP=0.2895 +⌛️ [2/4] FRONTEND: Frontend time: 1.926s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09752406 32.54828350 + layer.0.v_cache 0.00001393 0.00863287 + layer.1.k_cache 0.01493649 10.36157540 + layer.1.v_cache 0.00000588 0.00361074 + layer.2.k_cache 0.01003601 1.28853793 + layer.2.v_cache 0.00001805 0.01109152 + layer.3.k_cache 0.04451986 6.83934295 + layer.3.v_cache 0.00001914 0.01293120 + layer.4.k_cache 0.00061869 0.33666063 + layer.4.v_cache 0.00005251 0.02524294 + layer.4.output 0.18409089 678.76488095 + ------------------------------------------------------------------------------------- + TOTAL 0.08566946 282.51706331 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 112060 +BPFP 1.3205 bits/point +EBPFP 2.6409 equivalent bits/point +MSE 282.517063 +---------------------- -------------------------------------------------------- +Time: 3.287s Load: 0.004s, Pack+Encode: 1.926s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 282.5171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,556B, BPFP=0.4812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,112B, BPFP=3.0331 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,244B, BPFP=0.7989 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,372B, BPFP=2.8938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,276B, BPFP=0.9932 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,696B, BPFP=2.7666 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,612B, BPFP=0.8682 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,420B, BPFP=2.9029 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,728B, BPFP=2.2078 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,704B, BPFP=2.7681 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,812B, BPFP=0.2639 +⌛️ [2/4] FRONTEND: Frontend time: 2.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.308s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13658243 30.05781721 + layer.0.v_cache 0.00001379 0.00848861 + layer.1.k_cache 0.01234693 7.71126547 + layer.1.v_cache 0.00000566 0.00358893 + layer.2.k_cache 0.00375735 1.07923963 + layer.2.v_cache 0.00001852 0.01081986 + layer.3.k_cache 0.02700986 4.24162053 + layer.3.v_cache 0.00001930 0.01176213 + layer.4.k_cache 0.00062007 0.31400092 + layer.4.v_cache 0.00005146 0.02402087 + layer.4.output 0.17614663 606.91523236 + ------------------------------------------------------------------------------------- + TOTAL 0.08314423 252.46289710 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 114532 +BPFP 1.2683 bits/point +EBPFP 2.5366 equivalent bits/point +MSE 252.462897 +---------------------- -------------------------------------------------------- +Time: 3.534s Load: 0.005s, Pack+Encode: 2.222s, Decode+Unpack: 1.308s +---------------------- -------------------------------------------------------- +💾 Converting with 252.4629 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,424B, BPFP=0.5188 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,216B, BPFP=3.0428 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,184B, BPFP=0.8955 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,920B, BPFP=2.9795 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,676B, BPFP=1.2149 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,664B, BPFP=2.9247 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,992B, BPFP=1.0685 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,088B, BPFP=3.0154 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,304B, BPFP=2.2055 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,076B, BPFP=3.0128 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,456B, BPFP=0.2891 +⌛️ [2/4] FRONTEND: Frontend time: 1.823s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.283s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08738674 31.98844513 + layer.0.v_cache 0.00001368 0.00897095 + layer.1.k_cache 0.01286346 10.92717586 + layer.1.v_cache 0.00000597 0.00353612 + layer.2.k_cache 0.00251921 1.45068516 + layer.2.v_cache 0.00001769 0.01116854 + layer.3.k_cache 0.08324167 6.41689395 + layer.3.v_cache 0.00002021 0.01268369 + layer.4.k_cache 0.00062021 0.29875815 + layer.4.v_cache 0.00005375 0.02515124 + layer.4.output 0.19567458 734.13863748 + ------------------------------------------------------------------------------------- + TOTAL 0.09155674 305.30081948 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 107000 +BPFP 1.3472 bits/point +EBPFP 2.6944 equivalent bits/point +MSE 305.300819 +---------------------- -------------------------------------------------------- +Time: 3.110s Load: 0.004s, Pack+Encode: 1.823s, Decode+Unpack: 1.283s +---------------------- -------------------------------------------------------- +💾 Converting with 305.3008 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,568B, BPFP=0.4834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,940B, BPFP=3.0008 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,240B, BPFP=0.7982 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,336B, BPFP=2.8870 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,444B, BPFP=1.0248 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,164B, BPFP=2.8547 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,548B, BPFP=0.8562 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,100B, BPFP=2.8426 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,416B, BPFP=2.1491 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,956B, BPFP=2.8155 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,324B, BPFP=0.2776 +⌛️ [2/4] FRONTEND: Frontend time: 2.009s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12176956 31.11150461 + layer.0.v_cache 0.00001396 0.00854001 + layer.1.k_cache 0.01120206 7.69909668 + layer.1.v_cache 0.00000562 0.00352940 + layer.2.k_cache 0.00678912 1.07125395 + layer.2.v_cache 0.00001849 0.01062225 + layer.3.k_cache 0.02622183 3.99071530 + layer.3.v_cache 0.00001965 0.01176468 + layer.4.k_cache 0.00062413 0.32918015 + layer.4.v_cache 0.00005248 0.02289306 + layer.4.output 0.17940697 624.61752367 + ------------------------------------------------------------------------------------- + TOTAL 0.08368034 259.79892740 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 115036 +BPFP 1.2739 bits/point +EBPFP 2.5477 equivalent bits/point +MSE 259.798927 +---------------------- -------------------------------------------------------- +Time: 3.381s Load: 0.004s, Pack+Encode: 2.009s, Decode+Unpack: 1.368s +---------------------- -------------------------------------------------------- +💾 Converting with 259.7989 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.4583 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,452B, BPFP=2.4482 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,236B, BPFP=0.7792 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,856B, BPFP=2.6571 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,096B, BPFP=1.0560 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,056B, BPFP=2.3893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,480B, BPFP=0.9643 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,416B, BPFP=2.4429 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,908B, BPFP=2.3673 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,768B, BPFP=2.3464 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,272B, BPFP=0.2821 +⌛️ [2/4] FRONTEND: Frontend time: 2.144s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.310s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120816 31.89875837 + layer.0.v_cache 0.00001607 0.00822623 + layer.1.k_cache 0.10678518 8.20075567 + layer.1.v_cache 0.00000585 0.00324265 + layer.2.k_cache 0.00813519 1.55028309 + layer.2.v_cache 0.00001919 0.01012032 + layer.3.k_cache 0.06560528 4.54735747 + layer.3.v_cache 0.00001973 0.01152094 + layer.4.k_cache 0.00062300 0.31269578 + layer.4.v_cache 0.00005287 0.02350748 + layer.4.output 10.86395687 508.00488946 + ------------------------------------------------------------------------------------- + TOTAL 4.49059815 211.91768789 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 133620 +BPFP 1.1696 bits/point +EBPFP 2.3393 equivalent bits/point +MSE 211.917688 +---------------------- -------------------------------------------------------- +Time: 3.459s Load: 0.005s, Pack+Encode: 2.144s, Decode+Unpack: 1.310s +---------------------- -------------------------------------------------------- +💾 Converting with 211.9177 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,688B, BPFP=0.4828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,416B, BPFP=2.9483 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,596B, BPFP=0.8254 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,960B, BPFP=2.8664 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,144B, BPFP=0.9239 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,528B, BPFP=2.7888 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,924B, BPFP=0.8843 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,936B, BPFP=2.8621 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,388B, BPFP=2.2249 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,372B, BPFP=2.7608 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,692B, BPFP=0.2743 +⌛️ [2/4] FRONTEND: Frontend time: 2.119s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12848519 35.17629759 + layer.0.v_cache 0.00001371 0.00812046 + layer.1.k_cache 0.03007136 10.95734204 + layer.1.v_cache 0.00000570 0.00345167 + layer.2.k_cache 0.00936044 1.01458126 + layer.2.v_cache 0.00001842 0.01087855 + layer.3.k_cache 0.04065019 7.52567195 + layer.3.v_cache 0.00001877 0.01178939 + layer.4.k_cache 0.00062204 0.33407378 + layer.4.v_cache 0.00010837 0.02324742 + layer.4.output 0.16426703 606.54279557 + ------------------------------------------------------------------------------------- + TOTAL 0.07995432 252.99206018 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 119644 +BPFP 1.2640 bits/point +EBPFP 2.5280 equivalent bits/point +MSE 252.992060 +---------------------- -------------------------------------------------------- +Time: 3.440s Load: 0.004s, Pack+Encode: 2.119s, Decode+Unpack: 1.317s +---------------------- -------------------------------------------------------- +💾 Converting with 252.9921 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,584B, BPFP=0.5047 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,788B, BPFP=3.0836 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,300B, BPFP=0.8398 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,004B, BPFP=2.9305 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,556B, BPFP=1.0852 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,252B, BPFP=2.7836 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,920B, BPFP=0.9609 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,640B, BPFP=2.8594 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,592B, BPFP=2.0688 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,428B, BPFP=2.8180 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,676B, BPFP=0.2700 +⌛️ [2/4] FRONTEND: Frontend time: 1.831s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.413s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09257430 30.84225464 + layer.0.v_cache 0.00001329 0.00837710 + layer.1.k_cache 0.01202901 7.96757965 + layer.1.v_cache 0.00000544 0.00330379 + layer.2.k_cache 0.00549439 1.37939758 + layer.2.v_cache 0.00001849 0.01125534 + layer.3.k_cache 0.11014032 5.19351349 + layer.3.v_cache 0.00001916 0.01215511 + layer.4.k_cache 0.00063444 0.31719747 + layer.4.v_cache 0.00005494 0.02242913 + layer.4.output 0.19143520 666.87561384 + ------------------------------------------------------------------------------------- + TOTAL 0.09182531 277.28745648 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 111740 +BPFP 1.2838 bits/point +EBPFP 2.5676 equivalent bits/point +MSE 277.287456 +---------------------- -------------------------------------------------------- +Time: 3.249s Load: 0.005s, Pack+Encode: 1.831s, Decode+Unpack: 1.413s +---------------------- -------------------------------------------------------- +💾 Converting with 277.2875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,588B, BPFP=0.4814 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,352B, BPFP=3.0417 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,376B, BPFP=0.8140 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,424B, BPFP=2.8690 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,088B, BPFP=0.9464 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,064B, BPFP=2.8021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,852B, BPFP=0.9025 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,404B, BPFP=2.8653 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,844B, BPFP=2.2031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,004B, BPFP=2.7909 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,052B, BPFP=0.2405 +⌛️ [2/4] FRONTEND: Frontend time: 1.865s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.401s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12365316 30.21539597 + layer.0.v_cache 0.00001332 0.00838005 + layer.1.k_cache 0.03511774 7.44168381 + layer.1.v_cache 0.00000569 0.00366387 + layer.2.k_cache 0.01236608 0.93279602 + layer.2.v_cache 0.00001835 0.01086630 + layer.3.k_cache 0.02851222 4.45446414 + layer.3.v_cache 0.00001844 0.01184856 + layer.4.k_cache 0.00063207 0.30642691 + layer.4.v_cache 0.00005066 0.02315889 + layer.4.output 0.17964420 614.34566327 + ------------------------------------------------------------------------------------- + TOTAL 0.08575865 255.51931338 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 115048 +BPFP 1.2588 bits/point +EBPFP 2.5177 equivalent bits/point +MSE 255.519313 +---------------------- -------------------------------------------------------- +Time: 3.270s Load: 0.003s, Pack+Encode: 1.865s, Decode+Unpack: 1.401s +---------------------- -------------------------------------------------------- +💾 Converting with 255.5193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,620B, BPFP=0.5248 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,676B, BPFP=2.9399 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,496B, BPFP=0.9006 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,912B, BPFP=2.7869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,348B, BPFP=1.0713 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,272B, BPFP=2.8590 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,872B, BPFP=0.9760 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,564B, BPFP=2.9175 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,152B, BPFP=2.0337 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,200B, BPFP=2.8446 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,000B, BPFP=0.2862 +⌛️ [2/4] FRONTEND: Frontend time: 2.000s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.556s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12096892 32.39307893 + layer.0.v_cache 0.00001393 0.00876298 + layer.1.k_cache 0.03247837 11.16220484 + layer.1.v_cache 0.00000613 0.00353769 + layer.2.k_cache 0.00394396 1.15764853 + layer.2.v_cache 0.00001920 0.01127585 + layer.3.k_cache 0.02751121 6.48377756 + layer.3.v_cache 0.00001919 0.01242206 + layer.4.k_cache 0.00062481 0.32231839 + layer.4.v_cache 0.00005293 0.02458413 + layer.4.output 0.18564256 667.01711310 + ------------------------------------------------------------------------------------- + TOTAL 0.08736098 277.68820015 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 109112 +BPFP 1.2857 bits/point +EBPFP 2.5715 equivalent bits/point +MSE 277.688200 +---------------------- -------------------------------------------------------- +Time: 3.562s Load: 0.006s, Pack+Encode: 2.000s, Decode+Unpack: 1.556s +---------------------- -------------------------------------------------------- +💾 Converting with 277.6882 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,540B, BPFP=0.5154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,784B, BPFP=3.2029 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,412B, BPFP=0.8953 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,860B, BPFP=3.0154 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,848B, BPFP=0.9838 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,380B, BPFP=2.9180 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,612B, BPFP=0.9359 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,828B, BPFP=3.0089 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,164B, BPFP=2.2654 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,188B, BPFP=2.8791 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,480B, BPFP=0.3038 +⌛️ [2/4] FRONTEND: Frontend time: 2.190s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.479s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08564486 32.45460570 + layer.0.v_cache 0.00001408 0.00887063 + layer.1.k_cache 0.03507188 10.27716778 + layer.1.v_cache 0.00000566 0.00360618 + layer.2.k_cache 0.00239554 1.08147470 + layer.2.v_cache 0.00001938 0.01183265 + layer.3.k_cache 0.03313774 4.61753429 + layer.3.v_cache 0.00001993 0.01386572 + layer.4.k_cache 0.00061714 0.32289034 + layer.4.v_cache 0.00005134 0.02571474 + layer.4.output 0.18226704 695.68297774 + ------------------------------------------------------------------------------------- + TOTAL 0.08428511 289.32931805 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 112096 +BPFP 1.3380 bits/point +EBPFP 2.6761 equivalent bits/point +MSE 289.329318 +---------------------- -------------------------------------------------------- +Time: 3.674s Load: 0.005s, Pack+Encode: 2.190s, Decode+Unpack: 1.479s +---------------------- -------------------------------------------------------- +💾 Converting with 289.3293 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 117, 128) +Output shape: (1, 117, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.output: torch.Size([1, 117, 3584]) -> torch.Size([1, 1, 117, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,356B, BPFP=0.4482 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,188B, BPFP=2.1619 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,920B, BPFP=0.7906 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,820B, BPFP=2.1127 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,784B, BPFP=0.9060 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,792B, BPFP=2.1090 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,236B, BPFP=0.8328 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,080B, BPFP=2.1474 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,636B, BPFP=2.0881 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,564B, BPFP=2.0785 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,068B, BPFP=0.2493 +⌛️ [2/4] FRONTEND: Frontend time: 1.909s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11900647 30.93825538 + layer.0.v_cache 0.00001428 0.00843178 + layer.1.k_cache 0.12535267 11.99305034 + layer.1.v_cache 0.00000579 0.00344960 + layer.2.k_cache 0.00517510 1.17448673 + layer.2.v_cache 0.00001992 0.01108446 + layer.3.k_cache 0.09315677 5.39096356 + layer.3.v_cache 0.00001978 0.01184316 + layer.4.k_cache 0.00061896 0.33209630 + layer.4.v_cache 0.00005133 0.02219906 + layer.4.output 9.75369281 446.05235043 + ------------------------------------------------------------------------------------- + TOTAL 4.03642769 186.60307726 + (elements=1,018,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1018368 +Total Bytes 130444 +BPFP 1.0247 bits/point +EBPFP 2.0495 equivalent bits/point +MSE 186.603077 +---------------------- -------------------------------------------------------- +Time: 3.227s Load: 0.005s, Pack+Encode: 1.909s, Decode+Unpack: 1.313s +---------------------- -------------------------------------------------------- +💾 Converting with 186.6031 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,068B, BPFP=0.4565 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,400B, BPFP=2.4405 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,260B, BPFP=0.7827 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,084B, BPFP=2.3935 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,872B, BPFP=1.0226 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,828B, BPFP=2.3554 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,724B, BPFP=1.0006 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,152B, BPFP=2.4036 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,596B, BPFP=2.4696 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,612B, BPFP=2.3232 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,236B, BPFP=0.2176 +⌛️ [2/4] FRONTEND: Frontend time: 1.850s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.472s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14131550 34.55114397 + layer.0.v_cache 0.00001383 0.00830957 + layer.1.k_cache 0.02978072 8.74803757 + layer.1.v_cache 0.00000537 0.00331312 + layer.2.k_cache 0.00578868 1.57617449 + layer.2.v_cache 0.00001872 0.01010986 + layer.3.k_cache 0.04020810 7.34816313 + layer.3.v_cache 0.00001850 0.01160504 + layer.4.k_cache 0.00063358 0.31767549 + layer.4.v_cache 0.00005294 0.02370434 + layer.4.output 10.86925024 509.26275510 + ------------------------------------------------------------------------------------- + TOTAL 4.48838751 212.79044249 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 128832 +BPFP 1.1277 bits/point +EBPFP 2.2555 equivalent bits/point +MSE 212.790442 +---------------------- -------------------------------------------------------- +Time: 3.327s Load: 0.005s, Pack+Encode: 1.850s, Decode+Unpack: 1.472s +---------------------- -------------------------------------------------------- +💾 Converting with 212.7904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,580B, BPFP=0.4799 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,740B, BPFP=3.1138 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,364B, BPFP=0.8118 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,528B, BPFP=2.8884 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,416B, BPFP=1.0074 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,432B, BPFP=2.8705 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,840B, BPFP=0.9003 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,440B, BPFP=2.8720 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,820B, BPFP=2.1987 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,796B, BPFP=2.7522 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,296B, BPFP=0.2470 +⌛️ [2/4] FRONTEND: Frontend time: 1.832s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09514355 29.76288714 + layer.0.v_cache 0.00001423 0.00827580 + layer.1.k_cache 0.03283414 7.71769932 + layer.1.v_cache 0.00000553 0.00336114 + layer.2.k_cache 0.00489674 0.99134718 + layer.2.v_cache 0.00001883 0.01109523 + layer.3.k_cache 0.07199075 4.58339800 + layer.3.v_cache 0.00001932 0.01272474 + layer.4.k_cache 0.00061592 0.31837631 + layer.4.v_cache 0.00006816 0.02397855 + layer.4.output 0.17253765 632.88270621 + ------------------------------------------------------------------------------------- + TOTAL 0.08313946 263.15365217 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 116252 +BPFP 1.2720 bits/point +EBPFP 2.5440 equivalent bits/point +MSE 263.153652 +---------------------- -------------------------------------------------------- +Time: 3.057s Load: 0.003s, Pack+Encode: 1.832s, Decode+Unpack: 1.222s +---------------------- -------------------------------------------------------- +💾 Converting with 263.1537 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,708B, BPFP=0.4864 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,480B, BPFP=2.9598 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,496B, BPFP=0.8075 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,760B, BPFP=2.8305 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,188B, BPFP=0.9318 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,564B, BPFP=2.7953 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,932B, BPFP=0.8858 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,788B, BPFP=2.8355 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,624B, BPFP=2.2672 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,432B, BPFP=2.7716 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,324B, BPFP=0.2392 +⌛️ [2/4] FRONTEND: Frontend time: 2.062s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.428s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11660385 30.09997980 + layer.0.v_cache 0.00001381 0.00845202 + layer.1.k_cache 0.05175667 8.91824516 + layer.1.v_cache 0.00000569 0.00360915 + layer.2.k_cache 0.00227456 1.14165190 + layer.2.v_cache 0.00001743 0.01056096 + layer.3.k_cache 0.02781237 3.76624938 + layer.3.v_cache 0.00001883 0.01199879 + layer.4.k_cache 0.00061846 0.32577216 + layer.4.v_cache 0.00005344 0.02237525 + layer.4.output 0.17410791 616.21500411 + ------------------------------------------------------------------------------------- + TOTAL 0.08340767 256.34199549 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 118296 +BPFP 1.2497 bits/point +EBPFP 2.4995 equivalent bits/point +MSE 256.341995 +---------------------- -------------------------------------------------------- +Time: 3.494s Load: 0.005s, Pack+Encode: 2.062s, Decode+Unpack: 1.428s +---------------------- -------------------------------------------------------- +💾 Converting with 256.3420 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,572B, BPFP=0.4842 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,200B, BPFP=3.0497 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,228B, BPFP=0.7959 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,392B, BPFP=2.8976 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,992B, BPFP=0.9398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,904B, BPFP=2.8057 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,812B, BPFP=0.9059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,416B, BPFP=2.9021 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,232B, BPFP=2.3027 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,592B, BPFP=2.7470 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,612B, BPFP=0.2585 +⌛️ [2/4] FRONTEND: Frontend time: 1.962s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212080 35.40369094 + layer.0.v_cache 0.00001360 0.00830428 + layer.1.k_cache 0.03288655 7.95092185 + layer.1.v_cache 0.00000536 0.00339214 + layer.2.k_cache 0.00973156 1.05018873 + layer.2.v_cache 0.00001762 0.01045935 + layer.3.k_cache 0.02906220 6.71653509 + layer.3.v_cache 0.00001891 0.01237387 + layer.4.k_cache 0.00063934 0.31437826 + layer.4.v_cache 0.00004866 0.02285371 + layer.4.output 0.18425958 624.13161575 + ------------------------------------------------------------------------------------- + TOTAL 0.08496245 260.02437697 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 114952 +BPFP 1.2729 bits/point +EBPFP 2.5459 equivalent bits/point +MSE 260.024377 +---------------------- -------------------------------------------------------- +Time: 3.320s Load: 0.004s, Pack+Encode: 1.962s, Decode+Unpack: 1.353s +---------------------- -------------------------------------------------------- +💾 Converting with 260.0244 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,632B, BPFP=0.4838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,484B, BPFP=3.0301 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,348B, BPFP=0.7993 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,800B, BPFP=2.9044 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,112B, BPFP=0.9397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,084B, BPFP=2.7728 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,756B, BPFP=0.8743 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,584B, BPFP=2.8647 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,604B, BPFP=2.1331 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,780B, BPFP=2.7169 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,716B, BPFP=0.2551 +⌛️ [2/4] FRONTEND: Frontend time: 2.241s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.454s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10758395 28.84392808 + layer.0.v_cache 0.00001397 0.00854330 + layer.1.k_cache 0.03293324 8.03276511 + layer.1.v_cache 0.00000575 0.00347411 + layer.2.k_cache 0.00823421 0.96523527 + layer.2.v_cache 0.00001745 0.01101202 + layer.3.k_cache 0.04116619 4.87169764 + layer.3.v_cache 0.00001985 0.01233925 + layer.4.k_cache 0.00062779 0.31801924 + layer.4.v_cache 0.00005042 0.02393825 + layer.4.output 0.16954060 626.11444328 + ------------------------------------------------------------------------------------- + TOTAL 0.08102571 260.34659148 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 115900 +BPFP 1.2532 bits/point +EBPFP 2.5065 equivalent bits/point +MSE 260.346591 +---------------------- -------------------------------------------------------- +Time: 3.701s Load: 0.006s, Pack+Encode: 2.241s, Decode+Unpack: 1.454s +---------------------- -------------------------------------------------------- +💾 Converting with 260.3466 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,696B, BPFP=0.4681 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,652B, BPFP=2.8910 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,104B, BPFP=1.0597 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,908B, BPFP=2.7618 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,772B, BPFP=1.1757 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,516B, BPFP=2.6938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,512B, BPFP=1.1306 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,944B, BPFP=2.7681 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,952B, BPFP=2.4222 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,288B, BPFP=2.6542 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,872B, BPFP=0.2448 +⌛️ [2/4] FRONTEND: Frontend time: 2.053s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.284s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12307767 33.07890625 + layer.0.v_cache 0.00001410 0.00861561 + layer.1.k_cache 0.03088827 13.14660373 + layer.1.v_cache 0.00000584 0.00357716 + layer.2.k_cache 0.00371391 1.60498284 + layer.2.v_cache 0.00001950 0.01094329 + layer.3.k_cache 0.03079298 8.57948066 + layer.3.v_cache 0.00001950 0.01189327 + layer.4.k_cache 0.00063014 0.32322468 + layer.4.v_cache 0.00005282 0.02300581 + layer.4.output 0.15767360 597.50833333 + ------------------------------------------------------------------------------------- + TOTAL 0.07605470 249.37350392 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 125216 +BPFP 1.2788 bits/point +EBPFP 2.5575 equivalent bits/point +MSE 249.373504 +---------------------- -------------------------------------------------------- +Time: 3.341s Load: 0.004s, Pack+Encode: 2.053s, Decode+Unpack: 1.284s +---------------------- -------------------------------------------------------- +💾 Converting with 249.3735 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,548B, BPFP=0.4855 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,652B, BPFP=2.9825 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,228B, BPFP=0.8056 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,092B, BPFP=2.8758 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,260B, BPFP=1.0023 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,896B, BPFP=2.8384 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,704B, BPFP=0.8963 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,316B, BPFP=2.9184 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,096B, BPFP=2.1143 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,768B, BPFP=2.8140 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,096B, BPFP=0.2748 +⌛️ [2/4] FRONTEND: Frontend time: 1.896s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.333s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10295349 31.92191668 + layer.0.v_cache 0.00001348 0.00828803 + layer.1.k_cache 0.03334363 7.72992539 + layer.1.v_cache 0.00000550 0.00340205 + layer.2.k_cache 0.01126997 0.96080575 + layer.2.v_cache 0.00001933 0.01094289 + layer.3.k_cache 0.09131574 5.38230338 + layer.3.v_cache 0.00001769 0.01130056 + layer.4.k_cache 0.00062275 0.32851242 + layer.4.v_cache 0.00006541 0.02203152 + layer.4.output 0.17512874 642.81451437 + ------------------------------------------------------------------------------------- + TOTAL 0.08620754 267.41653113 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 113656 +BPFP 1.2739 bits/point +EBPFP 2.5479 equivalent bits/point +MSE 267.416531 +---------------------- -------------------------------------------------------- +Time: 3.233s Load: 0.004s, Pack+Encode: 1.896s, Decode+Unpack: 1.333s +---------------------- -------------------------------------------------------- +💾 Converting with 267.4165 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,704B, BPFP=0.4856 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,500B, BPFP=2.9634 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,608B, BPFP=0.8276 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,012B, BPFP=2.8757 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,104B, BPFP=0.9167 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,556B, BPFP=2.7938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,796B, BPFP=0.8614 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,012B, BPFP=2.8757 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,940B, BPFP=2.1444 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,044B, BPFP=2.7019 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,496B, BPFP=0.2436 +⌛️ [2/4] FRONTEND: Frontend time: 1.812s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.403s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12314748 31.12245196 + layer.0.v_cache 0.00001636 0.00838114 + layer.1.k_cache 0.01190477 10.56049005 + layer.1.v_cache 0.00000588 0.00344276 + layer.2.k_cache 0.01614088 0.96141999 + layer.2.v_cache 0.00001834 0.01056059 + layer.3.k_cache 0.05604688 4.19081835 + layer.3.v_cache 0.00001882 0.01207753 + layer.4.k_cache 0.00063774 0.32732462 + layer.4.v_cache 0.00005162 0.02296818 + layer.4.output 0.17348555 614.55618842 + ------------------------------------------------------------------------------------- + TOTAL 0.08366986 255.83019142 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 117772 +BPFP 1.2442 bits/point +EBPFP 2.4884 equivalent bits/point +MSE 255.830191 +---------------------- -------------------------------------------------------- +Time: 3.223s Load: 0.009s, Pack+Encode: 1.812s, Decode+Unpack: 1.403s +---------------------- -------------------------------------------------------- +💾 Converting with 255.8302 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,584B, BPFP=0.4985 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,008B, BPFP=3.0880 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,264B, BPFP=0.8225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,792B, BPFP=2.8534 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,632B, BPFP=1.0864 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,728B, BPFP=2.8410 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,752B, BPFP=0.9167 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,972B, BPFP=2.8881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,992B, BPFP=2.1204 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,272B, BPFP=2.7531 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,456B, BPFP=0.2881 +⌛️ [2/4] FRONTEND: Frontend time: 2.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.426s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11560643 31.98533951 + layer.0.v_cache 0.00001402 0.00889411 + layer.1.k_cache 0.03272506 8.15063778 + layer.1.v_cache 0.00000563 0.00368564 + layer.2.k_cache 0.00645547 0.99052938 + layer.2.v_cache 0.00001830 0.01136476 + layer.3.k_cache 0.02731901 4.50373219 + layer.3.v_cache 0.00001960 0.01268320 + layer.4.k_cache 0.00061302 0.30193621 + layer.4.v_cache 0.00004978 0.02310367 + layer.4.output 0.17251868 650.42520944 + ------------------------------------------------------------------------------------- + TOTAL 0.08179159 270.52755132 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 113452 +BPFP 1.2874 bits/point +EBPFP 2.5747 equivalent bits/point +MSE 270.527551 +---------------------- -------------------------------------------------------- +Time: 3.694s Load: 0.005s, Pack+Encode: 2.262s, Decode+Unpack: 1.426s +---------------------- -------------------------------------------------------- +💾 Converting with 270.5276 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,684B, BPFP=0.4766 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,336B, BPFP=2.9006 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,176B, BPFP=0.9190 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,824B, BPFP=2.8097 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,464B, BPFP=1.1477 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,464B, BPFP=2.7457 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,456B, BPFP=0.9688 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,844B, BPFP=2.8132 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,232B, BPFP=2.1719 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,048B, BPFP=2.6719 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,784B, BPFP=0.2735 +⌛️ [2/4] FRONTEND: Frontend time: 1.891s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.313s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11690623 32.33054421 + layer.0.v_cache 0.00001418 0.00836756 + layer.1.k_cache 0.03305150 11.15029006 + layer.1.v_cache 0.00000545 0.00350322 + layer.2.k_cache 0.01095151 1.57158800 + layer.2.v_cache 0.00001854 0.01062503 + layer.3.k_cache 0.04293454 6.25609935 + layer.3.v_cache 0.00002007 0.01182757 + layer.4.k_cache 0.00062489 0.32161483 + layer.4.v_cache 0.00005451 0.02226004 + layer.4.output 0.16672201 604.97813515 + ------------------------------------------------------------------------------------- + TOTAL 0.08068444 252.14903917 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 121312 +BPFP 1.2670 bits/point +EBPFP 2.5341 equivalent bits/point +MSE 252.149039 +---------------------- -------------------------------------------------------- +Time: 3.207s Load: 0.004s, Pack+Encode: 1.891s, Decode+Unpack: 1.313s +---------------------- -------------------------------------------------------- +💾 Converting with 252.1490 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,736B, BPFP=0.4698 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,556B, BPFP=2.8427 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,520B, BPFP=0.9478 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,080B, BPFP=2.7610 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,868B, BPFP=1.0076 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,608B, BPFP=2.6799 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,028B, BPFP=1.0350 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,152B, BPFP=2.7734 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,488B, BPFP=2.3159 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,516B, BPFP=2.6641 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,428B, BPFP=0.2313 +⌛️ [2/4] FRONTEND: Frontend time: 2.020s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11067300 32.35758285 + layer.0.v_cache 0.00001409 0.00838841 + layer.1.k_cache 0.05152405 13.18575346 + layer.1.v_cache 0.00000569 0.00350357 + layer.2.k_cache 0.01362713 1.37711812 + layer.2.v_cache 0.00001889 0.01122514 + layer.3.k_cache 0.08227781 8.07302823 + layer.3.v_cache 0.00001889 0.01197771 + layer.4.k_cache 0.00062414 0.30530879 + layer.4.v_cache 0.00005257 0.02178642 + layer.4.output 0.15746453 562.49690934 + ------------------------------------------------------------------------------------- + TOTAL 0.08006400 234.87259047 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 122980 +BPFP 1.2421 bits/point +EBPFP 2.4842 equivalent bits/point +MSE 234.872590 +---------------------- -------------------------------------------------------- +Time: 3.369s Load: 0.006s, Pack+Encode: 2.020s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 234.8726 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 116, 128) +Output shape: (1, 116, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.output: torch.Size([1, 116, 3584]) -> torch.Size([1, 1, 116, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,268B, BPFP=0.4402 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,084B, BPFP=2.1665 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,260B, BPFP=0.8432 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,756B, BPFP=2.1223 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,860B, BPFP=0.9240 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,596B, BPFP=2.1008 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,752B, BPFP=0.9095 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,808B, BPFP=2.1293 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,100B, BPFP=2.1686 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,444B, BPFP=2.0803 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,368B, BPFP=0.2188 +⌛️ [2/4] FRONTEND: Frontend time: 1.823s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.248s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16314862 31.17247062 + layer.0.v_cache 0.00001373 0.00812047 + layer.1.k_cache 0.05481785 8.58873196 + layer.1.v_cache 0.00000557 0.00346613 + layer.2.k_cache 0.00968372 1.35861956 + layer.2.v_cache 0.00001954 0.01103667 + layer.3.k_cache 0.02418142 8.74207701 + layer.3.v_cache 0.00001860 0.01177356 + layer.4.k_cache 0.00062333 0.30720329 + layer.4.v_cache 0.00005467 0.02384386 + layer.4.output 9.83417319 446.30961361 + ------------------------------------------------------------------------------------- + TOTAL 4.06422232 186.72909638 + (elements=1,009,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1009664 +Total Bytes 129296 +BPFP 1.0245 bits/point +EBPFP 2.0489 equivalent bits/point +MSE 186.729096 +---------------------- -------------------------------------------------------- +Time: 3.076s Load: 0.005s, Pack+Encode: 1.823s, Decode+Unpack: 1.248s +---------------------- -------------------------------------------------------- +💾 Converting with 186.7291 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,628B, BPFP=0.5264 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,320B, BPFP=3.0689 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,408B, BPFP=0.8830 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,344B, BPFP=2.8734 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,672B, BPFP=1.1362 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,028B, BPFP=2.8101 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,044B, BPFP=1.0104 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,904B, BPFP=2.9856 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,932B, BPFP=2.1899 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,852B, BPFP=2.7748 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,096B, BPFP=0.2889 +⌛️ [2/4] FRONTEND: Frontend time: 1.776s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11749315 36.90356132 + layer.0.v_cache 0.00001389 0.00853144 + layer.1.k_cache 0.03399578 10.04834689 + layer.1.v_cache 0.00000555 0.00359424 + layer.2.k_cache 0.00854440 1.34113918 + layer.2.v_cache 0.00001777 0.01051503 + layer.3.k_cache 0.02740619 5.54086069 + layer.3.v_cache 0.00001935 0.01225437 + layer.4.k_cache 0.00063531 0.32748296 + layer.4.v_cache 0.00005166 0.02424150 + layer.4.output 0.18490290 688.11372482 + ------------------------------------------------------------------------------------- + TOTAL 0.08720608 286.53038831 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 111228 +BPFP 1.3107 bits/point +EBPFP 2.6213 equivalent bits/point +MSE 286.530388 +---------------------- -------------------------------------------------------- +Time: 3.133s Load: 0.003s, Pack+Encode: 1.776s, Decode+Unpack: 1.353s +---------------------- -------------------------------------------------------- +💾 Converting with 286.5304 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,840B, BPFP=0.4721 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,556B, BPFP=2.7520 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,444B, BPFP=1.0711 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,020B, BPFP=2.6629 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,136B, BPFP=1.0199 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,732B, BPFP=2.6150 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,088B, BPFP=1.0120 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,236B, BPFP=2.6988 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,060B, BPFP=2.1709 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,364B, BPFP=2.5539 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,108B, BPFP=0.2400 +⌛️ [2/4] FRONTEND: Frontend time: 2.066s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11492289 38.50122849 + layer.0.v_cache 0.00001440 0.00829325 + layer.1.k_cache 0.01112092 12.88903549 + layer.1.v_cache 0.00000555 0.00336453 + layer.2.k_cache 0.01096075 1.47145470 + layer.2.v_cache 0.00001882 0.01094147 + layer.3.k_cache 0.03781782 9.39034580 + layer.3.v_cache 0.00001905 0.01160714 + layer.4.k_cache 0.00062716 0.33325512 + layer.4.v_cache 0.00006343 0.02253597 + layer.4.output 0.15224552 557.03229483 + ------------------------------------------------------------------------------------- + TOTAL 0.07301703 233.05106622 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 124584 +BPFP 1.2182 bits/point +EBPFP 2.4363 equivalent bits/point +MSE 233.051066 +---------------------- -------------------------------------------------------- +Time: 3.388s Load: 0.004s, Pack+Encode: 2.066s, Decode+Unpack: 1.318s +---------------------- -------------------------------------------------------- +💾 Converting with 233.0511 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,688B, BPFP=0.5455 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,100B, BPFP=3.0641 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,444B, BPFP=0.9018 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,452B, BPFP=2.9326 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,856B, BPFP=1.1883 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,044B, BPFP=2.8498 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,092B, BPFP=1.0333 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,812B, BPFP=3.0057 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,448B, BPFP=2.1201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,044B, BPFP=2.8498 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,640B, BPFP=0.2795 +⌛️ [2/4] FRONTEND: Frontend time: 1.889s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10405910 30.61926111 + layer.0.v_cache 0.00001453 0.00869955 + layer.1.k_cache 0.03397077 10.15794472 + layer.1.v_cache 0.00000547 0.00360820 + layer.2.k_cache 0.00568739 1.38963001 + layer.2.v_cache 0.00001916 0.01153056 + layer.3.k_cache 0.09581309 7.70955668 + layer.3.v_cache 0.00001862 0.01275254 + layer.4.k_cache 0.00061896 0.30877656 + layer.4.v_cache 0.00005653 0.02406985 + layer.4.output 0.19123089 692.27539425 + ------------------------------------------------------------------------------------- + TOTAL 0.09287528 288.01021115 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 110620 +BPFP 1.3204 bits/point +EBPFP 2.6409 equivalent bits/point +MSE 288.010211 +---------------------- -------------------------------------------------------- +Time: 3.224s Load: 0.005s, Pack+Encode: 1.889s, Decode+Unpack: 1.330s +---------------------- -------------------------------------------------------- +💾 Converting with 288.0102 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,644B, BPFP=0.4860 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,344B, BPFP=3.0044 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,372B, BPFP=0.8037 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,752B, BPFP=2.8956 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,704B, BPFP=1.0485 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,328B, BPFP=2.8176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,936B, BPFP=0.9074 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,640B, BPFP=2.8750 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,192B, BPFP=2.2412 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,176B, BPFP=2.7897 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,304B, BPFP=0.2706 +⌛️ [2/4] FRONTEND: Frontend time: 1.804s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.339s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09384689 30.17649357 + layer.0.v_cache 0.00001364 0.00828897 + layer.1.k_cache 0.01313462 7.57305406 + layer.1.v_cache 0.00000592 0.00362793 + layer.2.k_cache 0.00813576 1.38049747 + layer.2.v_cache 0.00001762 0.01081703 + layer.3.k_cache 0.02628646 5.28741670 + layer.3.v_cache 0.00002044 0.01241051 + layer.4.k_cache 0.00061596 0.33716040 + layer.4.v_cache 0.00005334 0.02363756 + layer.4.output 0.17343949 616.80992647 + ------------------------------------------------------------------------------------- + TOTAL 0.07977689 256.61664056 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 118392 +BPFP 1.2802 bits/point +EBPFP 2.5604 equivalent bits/point +MSE 256.616641 +---------------------- -------------------------------------------------------- +Time: 3.146s Load: 0.003s, Pack+Encode: 1.804s, Decode+Unpack: 1.339s +---------------------- -------------------------------------------------------- +💾 Converting with 256.6166 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 126, 128) +Output shape: (1, 126, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.output: torch.Size([1, 126, 3584]) -> torch.Size([1, 1, 126, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,448B, BPFP=0.4276 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,512B, BPFP=2.2956 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,004B, BPFP=0.7445 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,972B, BPFP=1.9807 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,304B, BPFP=0.9058 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,660B, BPFP=2.3140 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,264B, BPFP=0.9008 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,932B, BPFP=2.2237 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,080B, BPFP=1.9940 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,412B, BPFP=1.9112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,584B, BPFP=0.2229 +⌛️ [2/4] FRONTEND: Frontend time: 2.237s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13141363 34.60345750 + layer.0.v_cache 0.00001429 0.00801788 + layer.1.k_cache 0.06496696 9.60681346 + layer.1.v_cache 0.00000583 0.00382701 + layer.2.k_cache 0.01057190 1.73697311 + layer.2.v_cache 0.00002020 0.01216396 + layer.3.k_cache 0.05250616 10.51417178 + layer.3.v_cache 0.00001853 0.01328222 + layer.4.k_cache 0.00062861 0.36217378 + layer.4.v_cache 0.00004930 0.02761873 + layer.4.output 9.05470577 420.24248866 + ------------------------------------------------------------------------------------- + TOTAL 3.74371387 176.38740706 + (elements=1,096,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1096704 +Total Bytes 139172 +BPFP 1.0152 bits/point +EBPFP 2.0304 equivalent bits/point +MSE 176.387407 +---------------------- -------------------------------------------------------- +Time: 3.634s Load: 0.007s, Pack+Encode: 2.237s, Decode+Unpack: 1.390s +---------------------- -------------------------------------------------------- +💾 Converting with 176.3874 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,684B, BPFP=0.4820 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,508B, BPFP=2.9648 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,916B, BPFP=0.8829 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,868B, BPFP=2.8499 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,652B, BPFP=1.0151 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,572B, BPFP=2.7967 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,168B, BPFP=0.9282 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,188B, BPFP=2.9073 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,076B, BPFP=2.1688 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,128B, BPFP=2.7170 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,412B, BPFP=0.2415 +⌛️ [2/4] FRONTEND: Frontend time: 2.175s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13059476 30.78685120 + layer.0.v_cache 0.00001380 0.00810476 + layer.1.k_cache 0.03034586 10.06816715 + layer.1.v_cache 0.00000531 0.00341166 + layer.2.k_cache 0.00374181 1.26761803 + layer.2.v_cache 0.00001886 0.01088230 + layer.3.k_cache 0.04454902 6.48727242 + layer.3.v_cache 0.00001968 0.01166423 + layer.4.k_cache 0.00060938 0.31700937 + layer.4.v_cache 0.00007316 0.02342124 + layer.4.output 0.16485390 606.68996305 + ------------------------------------------------------------------------------------- + TOTAL 0.08023229 252.69494963 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 119172 +BPFP 1.2590 bits/point +EBPFP 2.5180 equivalent bits/point +MSE 252.694950 +---------------------- -------------------------------------------------------- +Time: 3.558s Load: 0.004s, Pack+Encode: 2.175s, Decode+Unpack: 1.378s +---------------------- -------------------------------------------------------- +💾 Converting with 252.6949 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,760B, BPFP=0.4688 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,348B, BPFP=2.7765 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,012B, BPFP=1.0211 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,160B, BPFP=2.7446 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,316B, BPFP=1.0727 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,416B, BPFP=2.6182 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,580B, BPFP=0.9477 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,196B, BPFP=2.7507 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,448B, BPFP=2.1141 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,680B, BPFP=2.6630 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,416B, BPFP=0.2770 +⌛️ [2/4] FRONTEND: Frontend time: 1.800s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.533s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15436328 31.57832934 + layer.0.v_cache 0.00001586 0.00827823 + layer.1.k_cache 0.01080767 13.04111912 + layer.1.v_cache 0.00000575 0.00335318 + layer.2.k_cache 0.00786338 1.48867632 + layer.2.v_cache 0.00001961 0.01055864 + layer.3.k_cache 0.05652578 7.09465956 + layer.3.v_cache 0.00002042 0.01212938 + layer.4.k_cache 0.00062828 0.32119861 + layer.4.v_cache 0.00005674 0.02343867 + layer.4.output 0.15192759 579.64552601 + ------------------------------------------------------------------------------------- + TOTAL 0.07610588 241.82943665 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 124332 +BPFP 1.2421 bits/point +EBPFP 2.4843 equivalent bits/point +MSE 241.829437 +---------------------- -------------------------------------------------------- +Time: 3.337s Load: 0.004s, Pack+Encode: 1.800s, Decode+Unpack: 1.533s +---------------------- -------------------------------------------------------- +💾 Converting with 241.8294 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,636B, BPFP=0.5280 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,420B, BPFP=3.0889 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,392B, BPFP=0.8798 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,216B, BPFP=2.8478 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,632B, BPFP=1.3285 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,284B, BPFP=2.8614 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,324B, BPFP=0.8662 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,616B, BPFP=2.9279 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,076B, BPFP=2.2188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,372B, BPFP=2.8790 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,076B, BPFP=0.2597 +⌛️ [2/4] FRONTEND: Frontend time: 2.075s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.395s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10254851 35.52737505 + layer.0.v_cache 0.00001399 0.00871939 + layer.1.k_cache 0.05476840 10.49426817 + layer.1.v_cache 0.00000542 0.00354352 + layer.2.k_cache 0.00244555 1.69183584 + layer.2.v_cache 0.00001844 0.01111407 + layer.3.k_cache 0.02811479 4.18446937 + layer.3.v_cache 0.00001905 0.01260954 + layer.4.k_cache 0.00060111 0.30913299 + layer.4.v_cache 0.00005313 0.02587030 + layer.4.output 0.19635826 676.66111493 + ------------------------------------------------------------------------------------- + TOTAL 0.09194683 281.69980840 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 111044 +BPFP 1.3085 bits/point +EBPFP 2.6170 equivalent bits/point +MSE 281.699808 +---------------------- -------------------------------------------------------- +Time: 3.474s Load: 0.004s, Pack+Encode: 2.075s, Decode+Unpack: 1.395s +---------------------- -------------------------------------------------------- +💾 Converting with 281.6998 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,580B, BPFP=0.4916 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,800B, BPFP=3.0107 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,244B, BPFP=0.8087 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,040B, BPFP=2.8659 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,148B, BPFP=0.9809 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,652B, BPFP=2.7919 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,548B, BPFP=0.8666 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,232B, BPFP=2.9024 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,696B, BPFP=2.0381 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,804B, BPFP=2.8209 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,820B, BPFP=0.2673 +⌛️ [2/4] FRONTEND: Frontend time: 1.797s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.399s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11987135 30.06102027 + layer.0.v_cache 0.00001390 0.00816254 + layer.1.k_cache 0.03139273 7.82234750 + layer.1.v_cache 0.00000562 0.00333262 + layer.2.k_cache 0.00816743 1.19114946 + layer.2.v_cache 0.00001857 0.01017293 + layer.3.k_cache 0.02640573 4.37563566 + layer.3.v_cache 0.00001968 0.01127791 + layer.4.k_cache 0.00062091 0.32176395 + layer.4.v_cache 0.00004972 0.02250710 + layer.4.output 0.16983549 639.95628267 + ------------------------------------------------------------------------------------- + TOTAL 0.08090671 266.08949109 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 112564 +BPFP 1.2617 bits/point +EBPFP 2.5234 equivalent bits/point +MSE 266.089491 +---------------------- -------------------------------------------------------- +Time: 3.200s Load: 0.004s, Pack+Encode: 1.797s, Decode+Unpack: 1.399s +---------------------- -------------------------------------------------------- +💾 Converting with 266.0895 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,696B, BPFP=0.4842 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,460B, BPFP=2.9562 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,644B, BPFP=0.8341 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,828B, BPFP=2.8427 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,404B, BPFP=0.9705 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,540B, BPFP=2.7909 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,992B, BPFP=0.8966 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,972B, BPFP=2.8685 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,356B, BPFP=2.5783 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,460B, BPFP=2.7766 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,844B, BPFP=0.2526 +⌛️ [2/4] FRONTEND: Frontend time: 1.998s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14089697 30.62173918 + layer.0.v_cache 0.00001393 0.00839464 + layer.1.k_cache 0.01618495 9.17863798 + layer.1.v_cache 0.00000582 0.00353784 + layer.2.k_cache 0.00661465 1.15168306 + layer.2.v_cache 0.00002083 0.01064563 + layer.3.k_cache 0.07087009 4.49906904 + layer.3.v_cache 0.00001822 0.01202343 + layer.4.k_cache 0.00061891 0.32906249 + layer.4.v_cache 0.00005311 0.02213254 + layer.4.output 0.16425455 613.43837233 + ------------------------------------------------------------------------------------- + TOTAL 0.08147525 255.28856071 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 121196 +BPFP 1.2804 bits/point +EBPFP 2.5608 equivalent bits/point +MSE 255.288561 +---------------------- -------------------------------------------------------- +Time: 3.375s Load: 0.004s, Pack+Encode: 1.998s, Decode+Unpack: 1.373s +---------------------- -------------------------------------------------------- +💾 Converting with 255.2886 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,708B, BPFP=0.4754 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,640B, BPFP=2.9213 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,808B, BPFP=0.8441 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,308B, BPFP=2.8631 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,008B, BPFP=1.0548 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,480B, BPFP=2.7177 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,444B, BPFP=1.1313 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,192B, BPFP=2.8427 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,272B, BPFP=2.1545 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,248B, BPFP=2.6770 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,792B, BPFP=0.2707 +⌛️ [2/4] FRONTEND: Frontend time: 1.989s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12786581 32.49736657 + layer.0.v_cache 0.00001426 0.00847103 + layer.1.k_cache 0.03183105 10.18206856 + layer.1.v_cache 0.00000593 0.00354570 + layer.2.k_cache 0.00241907 1.39197609 + layer.2.v_cache 0.00001785 0.01003825 + layer.3.k_cache 0.02946699 7.63913135 + layer.3.v_cache 0.00001957 0.01227447 + layer.4.k_cache 0.00063747 0.32896505 + layer.4.v_cache 0.00005003 0.02293048 + layer.4.output 0.16568538 602.56520867 + ------------------------------------------------------------------------------------- + TOTAL 0.07953681 251.17960166 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 122900 +BPFP 1.2692 bits/point +EBPFP 2.5384 equivalent bits/point +MSE 251.179602 +---------------------- -------------------------------------------------------- +Time: 3.323s Load: 0.004s, Pack+Encode: 1.989s, Decode+Unpack: 1.330s +---------------------- -------------------------------------------------------- +💾 Converting with 251.1796 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,640B, BPFP=0.5288 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,168B, BPFP=3.0385 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,400B, BPFP=0.8814 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,896B, BPFP=2.9840 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,280B, BPFP=1.0577 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,680B, BPFP=2.9407 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,056B, BPFP=1.0128 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,208B, BPFP=3.0465 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,220B, BPFP=2.2476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,620B, BPFP=2.9287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,608B, BPFP=0.3036 +⌛️ [2/4] FRONTEND: Frontend time: 1.873s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.415s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12225728 37.20752579 + layer.0.v_cache 0.00001337 0.00856484 + layer.1.k_cache 0.03480693 11.84882374 + layer.1.v_cache 0.00000545 0.00356484 + layer.2.k_cache 0.00853881 1.26200358 + layer.2.v_cache 0.00001924 0.01163672 + layer.3.k_cache 0.06130744 6.75260730 + layer.3.v_cache 0.00001928 0.01271938 + layer.4.k_cache 0.00058936 0.33490098 + layer.4.v_cache 0.00007197 0.02622658 + layer.4.output 0.18657821 676.41763965 + ------------------------------------------------------------------------------------- + TOTAL 0.09021627 281.90541478 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 113776 +BPFP 1.3407 bits/point +EBPFP 2.6814 equivalent bits/point +MSE 281.905415 +---------------------- -------------------------------------------------------- +Time: 3.292s Load: 0.003s, Pack+Encode: 1.873s, Decode+Unpack: 1.415s +---------------------- -------------------------------------------------------- +💾 Converting with 281.9054 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,688B, BPFP=0.5455 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,084B, BPFP=3.0609 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,544B, BPFP=0.9221 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,612B, BPFP=2.9651 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,988B, BPFP=1.0122 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,360B, BPFP=2.9140 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,944B, BPFP=1.0032 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,176B, BPFP=3.0795 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,540B, BPFP=2.3417 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,492B, BPFP=2.9407 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,476B, BPFP=0.2747 +⌛️ [2/4] FRONTEND: Frontend time: 1.737s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.338s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582910 32.28506113 + layer.0.v_cache 0.00001386 0.00889167 + layer.1.k_cache 0.03408930 10.91244229 + layer.1.v_cache 0.00000538 0.00359788 + layer.2.k_cache 0.00228146 1.03078926 + layer.2.v_cache 0.00001859 0.01154781 + layer.3.k_cache 0.06276476 7.81649424 + layer.3.v_cache 0.00001913 0.01354644 + layer.4.k_cache 0.00060520 0.31538872 + layer.4.v_cache 0.00005125 0.02502950 + layer.4.output 0.19356344 697.27533627 + ------------------------------------------------------------------------------------- + TOTAL 0.09238954 290.19706723 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 111904 +BPFP 1.3358 bits/point +EBPFP 2.6715 equivalent bits/point +MSE 290.197067 +---------------------- -------------------------------------------------------- +Time: 3.079s Load: 0.004s, Pack+Encode: 1.737s, Decode+Unpack: 1.338s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1971 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,568B, BPFP=0.4834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,352B, BPFP=3.0783 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,248B, BPFP=0.7997 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,528B, BPFP=2.9232 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,324B, BPFP=1.0023 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,264B, BPFP=2.8735 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,768B, BPFP=0.8976 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,704B, BPFP=2.9563 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,816B, BPFP=2.0361 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,948B, BPFP=2.8140 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,184B, BPFP=0.2739 +⌛️ [2/4] FRONTEND: Frontend time: 1.973s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.602s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09700135 30.97013248 + layer.0.v_cache 0.00001355 0.00841289 + layer.1.k_cache 0.01103399 7.71585635 + layer.1.v_cache 0.00000598 0.00355858 + layer.2.k_cache 0.00794396 1.01240659 + layer.2.v_cache 0.00001935 0.01141413 + layer.3.k_cache 0.04207947 4.29271652 + layer.3.v_cache 0.00002010 0.01258709 + layer.4.k_cache 0.00062246 0.31351765 + layer.4.v_cache 0.00005226 0.02340734 + layer.4.output 0.17535226 628.24429862 + ------------------------------------------------------------------------------------- + TOTAL 0.08154460 261.29847647 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 115704 +BPFP 1.2813 bits/point +EBPFP 2.5625 equivalent bits/point +MSE 261.298476 +---------------------- -------------------------------------------------------- +Time: 3.581s Load: 0.006s, Pack+Encode: 1.973s, Decode+Unpack: 1.602s +---------------------- -------------------------------------------------------- +💾 Converting with 261.2985 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,632B, BPFP=0.4838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,284B, BPFP=2.9934 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,408B, BPFP=0.8103 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,080B, BPFP=2.7721 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,356B, BPFP=0.9846 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,776B, BPFP=2.7162 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,928B, BPFP=0.9059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,448B, BPFP=2.8397 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,936B, BPFP=2.1941 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,900B, BPFP=2.7390 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,864B, BPFP=0.2590 +⌛️ [2/4] FRONTEND: Frontend time: 1.910s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11944316 33.48917739 + layer.0.v_cache 0.00001525 0.00795334 + layer.1.k_cache 0.01376067 7.79644273 + layer.1.v_cache 0.00000561 0.00346456 + layer.2.k_cache 0.00522648 0.99513675 + layer.2.v_cache 0.00001966 0.01139349 + layer.3.k_cache 0.04600189 3.93846112 + layer.3.v_cache 0.00001912 0.01298255 + layer.4.k_cache 0.00062079 0.34201158 + layer.4.v_cache 0.00005114 0.02396478 + layer.4.output 0.17175110 628.12599790 + ------------------------------------------------------------------------------------- + TOTAL 0.08161303 261.38252786 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 115612 +BPFP 1.2501 bits/point +EBPFP 2.5003 equivalent bits/point +MSE 261.382528 +---------------------- -------------------------------------------------------- +Time: 3.268s Load: 0.004s, Pack+Encode: 1.910s, Decode+Unpack: 1.354s +---------------------- -------------------------------------------------------- +💾 Converting with 261.3825 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,644B, BPFP=0.5296 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,260B, BPFP=3.0569 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,232B, BPFP=0.8478 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,384B, BPFP=2.8814 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,788B, BPFP=1.3598 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,196B, BPFP=2.8438 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,032B, BPFP=1.0080 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,776B, BPFP=2.9599 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,652B, BPFP=2.1338 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,280B, BPFP=2.8606 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,060B, BPFP=0.2879 +⌛️ [2/4] FRONTEND: Frontend time: 2.064s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.234s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120220 37.10851738 + layer.0.v_cache 0.00001368 0.00846133 + layer.1.k_cache 0.03273370 9.96434921 + layer.1.v_cache 0.00000573 0.00345574 + layer.2.k_cache 0.00892877 1.54479139 + layer.2.v_cache 0.00001953 0.01111696 + layer.3.k_cache 0.02974504 5.28772207 + layer.3.v_cache 0.00001879 0.01239118 + layer.4.k_cache 0.00063874 0.32255865 + layer.4.v_cache 0.00005200 0.02480479 + layer.4.output 0.17853031 675.16866987 + ------------------------------------------------------------------------------------- + TOTAL 0.08429825 281.20405046 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 112304 +BPFP 1.3233 bits/point +EBPFP 2.6467 equivalent bits/point +MSE 281.204050 +---------------------- -------------------------------------------------------- +Time: 3.303s Load: 0.005s, Pack+Encode: 2.064s, Decode+Unpack: 1.234s +---------------------- -------------------------------------------------------- +💾 Converting with 281.2041 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,572B, BPFP=0.4842 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,428B, BPFP=3.0926 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,340B, BPFP=0.8170 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,592B, BPFP=2.9352 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,012B, BPFP=0.9435 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,836B, BPFP=2.7929 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,940B, BPFP=0.9300 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,720B, BPFP=2.9593 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,352B, BPFP=2.1370 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,004B, BPFP=2.8245 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,224B, BPFP=0.2750 +⌛️ [2/4] FRONTEND: Frontend time: 2.071s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.238s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09802409 31.25415921 + layer.0.v_cache 0.00001402 0.00847930 + layer.1.k_cache 0.03453328 7.87797381 + layer.1.v_cache 0.00000594 0.00345767 + layer.2.k_cache 0.01027090 0.98868864 + layer.2.v_cache 0.00001933 0.01068309 + layer.3.k_cache 0.02613793 4.64680867 + layer.3.v_cache 0.00001931 0.01170058 + layer.4.k_cache 0.00063818 0.31791370 + layer.4.v_cache 0.00005393 0.02212806 + layer.4.output 0.18084440 633.33864028 + ------------------------------------------------------------------------------------- + TOTAL 0.08444869 263.44191027 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 116020 +BPFP 1.2848 bits/point +EBPFP 2.5695 equivalent bits/point +MSE 263.441910 +---------------------- -------------------------------------------------------- +Time: 3.314s Load: 0.005s, Pack+Encode: 2.071s, Decode+Unpack: 1.238s +---------------------- -------------------------------------------------------- +💾 Converting with 263.4419 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,012B, BPFP=0.4569 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,280B, BPFP=2.4697 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,272B, BPFP=0.7998 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,136B, BPFP=2.4478 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,260B, BPFP=0.9496 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,840B, BPFP=2.4029 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,884B, BPFP=0.8926 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,020B, BPFP=2.4302 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,600B, BPFP=2.5182 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,524B, BPFP=2.3550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,000B, BPFP=0.2384 +⌛️ [2/4] FRONTEND: Frontend time: 1.820s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.575s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13030120 33.40620970 + layer.0.v_cache 0.00001385 0.00835843 + layer.1.k_cache 0.02890817 8.14630838 + layer.1.v_cache 0.00000611 0.00355244 + layer.2.k_cache 0.00721959 1.13459541 + layer.2.v_cache 0.00001830 0.01046902 + layer.3.k_cache 0.02195997 8.01497317 + layer.3.v_cache 0.00001903 0.01157337 + layer.4.k_cache 0.00063766 0.30618342 + layer.4.v_cache 0.00005061 0.02145769 + layer.4.output 11.07892277 516.77297157 + ------------------------------------------------------------------------------------- + TOTAL 4.57303494 215.79261659 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 127828 +BPFP 1.1407 bits/point +EBPFP 2.2813 equivalent bits/point +MSE 215.792617 +---------------------- -------------------------------------------------------- +Time: 3.399s Load: 0.004s, Pack+Encode: 1.820s, Decode+Unpack: 1.575s +---------------------- -------------------------------------------------------- +💾 Converting with 215.7926 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,568B, BPFP=0.4834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,020B, BPFP=3.0158 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,280B, BPFP=0.8057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,244B, BPFP=2.8697 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,884B, BPFP=0.9194 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,028B, BPFP=2.8291 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,612B, BPFP=0.8682 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,868B, BPFP=2.7989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,276B, BPFP=2.3110 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,204B, BPFP=2.8622 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,208B, BPFP=0.2476 +⌛️ [2/4] FRONTEND: Frontend time: 1.865s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.249s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12633393 32.52814971 + layer.0.v_cache 0.00001372 0.00850895 + layer.1.k_cache 0.01411889 7.48560306 + layer.1.v_cache 0.00000549 0.00347655 + layer.2.k_cache 0.00833277 0.96852360 + layer.2.v_cache 0.00001851 0.01095819 + layer.3.k_cache 0.05797521 4.29518293 + layer.3.v_cache 0.00001879 0.01247777 + layer.4.k_cache 0.00064672 0.31910078 + layer.4.v_cache 0.00005349 0.02429070 + layer.4.output 0.17447430 631.34267427 + ------------------------------------------------------------------------------------- + TOTAL 0.08404927 262.65029365 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 114192 +BPFP 1.2645 bits/point +EBPFP 2.5291 equivalent bits/point +MSE 262.650294 +---------------------- -------------------------------------------------------- +Time: 3.118s Load: 0.004s, Pack+Encode: 1.865s, Decode+Unpack: 1.249s +---------------------- -------------------------------------------------------- +💾 Converting with 262.6503 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,348B, BPFP=0.5241 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,924B, BPFP=3.1080 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,804B, BPFP=0.8491 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,544B, BPFP=3.0232 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,404B, BPFP=0.9830 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,888B, BPFP=3.1000 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,848B, BPFP=0.8589 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,420B, BPFP=3.2188 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,748B, BPFP=2.3991 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,744B, BPFP=3.0679 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,740B, BPFP=0.2787 +⌛️ [2/4] FRONTEND: Frontend time: 1.734s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.441s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08086312 31.41448800 + layer.0.v_cache 0.00001331 0.00902826 + layer.1.k_cache 0.01551799 7.78095529 + layer.1.v_cache 0.00000513 0.00359884 + layer.2.k_cache 0.00232868 1.01564157 + layer.2.v_cache 0.00002262 0.01162135 + layer.3.k_cache 0.10623371 4.00341099 + layer.3.v_cache 0.00001836 0.01310233 + layer.4.k_cache 0.00061897 0.30876912 + layer.4.v_cache 0.00005922 0.02660183 + layer.4.output 0.20964142 761.00433673 + ------------------------------------------------------------------------------------- + TOTAL 0.09842183 315.97750440 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 103412 +BPFP 1.3578 bits/point +EBPFP 2.7157 equivalent bits/point +MSE 315.977504 +---------------------- -------------------------------------------------------- +Time: 3.178s Load: 0.003s, Pack+Encode: 1.734s, Decode+Unpack: 1.441s +---------------------- -------------------------------------------------------- +💾 Converting with 315.9775 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,780B, BPFP=0.4721 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,708B, BPFP=2.8376 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,412B, BPFP=0.9192 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,376B, BPFP=2.7812 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,848B, BPFP=0.9932 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,856B, BPFP=2.6929 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,260B, BPFP=0.8933 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,804B, BPFP=2.6841 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,680B, BPFP=2.3234 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,520B, BPFP=2.6359 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,460B, BPFP=0.2538 +⌛️ [2/4] FRONTEND: Frontend time: 1.776s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.277s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13670269 30.53746497 + layer.0.v_cache 0.00001568 0.00805364 + layer.1.k_cache 0.03139175 12.01816194 + layer.1.v_cache 0.00000602 0.00354037 + layer.2.k_cache 0.00629326 1.44988947 + layer.2.v_cache 0.00001871 0.01050043 + layer.3.k_cache 0.03837063 7.70790896 + layer.3.v_cache 0.00001860 0.01114034 + layer.4.k_cache 0.00062597 0.29264506 + layer.4.v_cache 0.00005354 0.02241835 + layer.4.output 0.15974679 564.54585598 + ------------------------------------------------------------------------------------- + TOTAL 0.07833673 235.52251267 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 123704 +BPFP 1.2359 bits/point +EBPFP 2.4717 equivalent bits/point +MSE 235.522513 +---------------------- -------------------------------------------------------- +Time: 3.057s Load: 0.004s, Pack+Encode: 1.776s, Decode+Unpack: 1.277s +---------------------- -------------------------------------------------------- +💾 Converting with 235.5225 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,128B, BPFP=0.4568 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,296B, BPFP=2.3797 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,604B, BPFP=0.8183 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,140B, BPFP=2.3569 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,544B, BPFP=0.9556 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,948B, BPFP=2.3289 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,756B, BPFP=0.9866 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,200B, BPFP=2.3657 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,960B, BPFP=2.3306 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,540B, BPFP=2.2693 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,596B, BPFP=0.2210 +⌛️ [2/4] FRONTEND: Frontend time: 2.215s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.280s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12281527 32.66507557 + layer.0.v_cache 0.00001417 0.00819094 + layer.1.k_cache 0.04442539 10.67567515 + layer.1.v_cache 0.00000572 0.00347736 + layer.2.k_cache 0.01033308 1.25880760 + layer.2.v_cache 0.00001923 0.01055403 + layer.3.k_cache 0.04555132 9.50575841 + layer.3.v_cache 0.00001841 0.01177128 + layer.4.k_cache 0.00060580 0.29682701 + layer.4.v_cache 0.00007068 0.02359670 + layer.4.output 10.66341841 482.23560581 + ------------------------------------------------------------------------------------- + TOTAL 4.40398753 201.77111616 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 128712 +BPFP 1.1056 bits/point +EBPFP 2.2112 equivalent bits/point +MSE 201.771116 +---------------------- -------------------------------------------------------- +Time: 3.501s Load: 0.006s, Pack+Encode: 2.215s, Decode+Unpack: 1.280s +---------------------- -------------------------------------------------------- +💾 Converting with 201.7711 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,660B, BPFP=0.4833 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,252B, BPFP=2.9528 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,500B, BPFP=0.8176 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,660B, BPFP=2.8452 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,172B, BPFP=0.9397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,040B, BPFP=2.7326 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,876B, BPFP=0.8859 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,364B, BPFP=2.7914 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,484B, BPFP=2.2682 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,784B, BPFP=2.6860 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,632B, BPFP=0.2500 +⌛️ [2/4] FRONTEND: Frontend time: 2.149s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.405s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11362342 33.70807027 + layer.0.v_cache 0.00001418 0.00856734 + layer.1.k_cache 0.05429294 10.03468926 + layer.1.v_cache 0.00000580 0.00346397 + layer.2.k_cache 0.00513919 1.04882972 + layer.2.v_cache 0.00001816 0.01087260 + layer.3.k_cache 0.05885953 4.54114621 + layer.3.v_cache 0.00001841 0.01194333 + layer.4.k_cache 0.00061763 0.32914694 + layer.4.v_cache 0.00005080 0.02318237 + layer.4.output 0.16727475 621.94684385 + ------------------------------------------------------------------------------------- + TOTAL 0.08256255 259.02045994 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 116424 +BPFP 1.2443 bits/point +EBPFP 2.4885 equivalent bits/point +MSE 259.020460 +---------------------- -------------------------------------------------------- +Time: 3.557s Load: 0.003s, Pack+Encode: 2.149s, Decode+Unpack: 1.405s +---------------------- -------------------------------------------------------- +💾 Converting with 259.0205 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,944B, BPFP=0.4946 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,640B, BPFP=2.7957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,824B, BPFP=0.9785 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,924B, BPFP=2.6754 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,380B, BPFP=1.0719 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,964B, BPFP=2.6821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,680B, BPFP=0.9543 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,212B, BPFP=2.7238 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,712B, BPFP=2.1358 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,572B, BPFP=2.6163 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,264B, BPFP=0.2464 +⌛️ [2/4] FRONTEND: Frontend time: 1.898s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.239s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702597 41.36973391 + layer.0.v_cache 0.00001402 0.00849910 + layer.1.k_cache 0.01332537 14.05437248 + layer.1.v_cache 0.00000581 0.00350150 + layer.2.k_cache 0.00507878 1.58427413 + layer.2.v_cache 0.00001944 0.01077757 + layer.3.k_cache 0.04051446 8.63136406 + layer.3.v_cache 0.00001931 0.01223971 + layer.4.k_cache 0.00062870 0.30532847 + layer.4.v_cache 0.00006514 0.02267352 + layer.4.output 0.16549673 558.36976767 + ------------------------------------------------------------------------------------- + TOTAL 0.07795142 233.79947871 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 124116 +BPFP 1.2266 bits/point +EBPFP 2.4533 equivalent bits/point +MSE 233.799479 +---------------------- -------------------------------------------------------- +Time: 3.141s Load: 0.005s, Pack+Encode: 1.898s, Decode+Unpack: 1.239s +---------------------- -------------------------------------------------------- +💾 Converting with 233.7995 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,612B, BPFP=0.4801 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,336B, BPFP=3.0029 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,296B, BPFP=0.7897 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,620B, BPFP=2.8713 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,820B, BPFP=1.0699 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,148B, BPFP=2.7846 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,588B, BPFP=0.8434 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,400B, BPFP=2.8309 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,168B, BPFP=2.0529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,112B, BPFP=2.7779 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,228B, BPFP=0.2686 +⌛️ [2/4] FRONTEND: Frontend time: 2.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.457s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13925676 31.34698415 + layer.0.v_cache 0.00001387 0.00807300 + layer.1.k_cache 0.03290265 7.68085435 + layer.1.v_cache 0.00000589 0.00337954 + layer.2.k_cache 0.00800494 1.09348863 + layer.2.v_cache 0.00001921 0.01060213 + layer.3.k_cache 0.08301921 3.43704762 + layer.3.v_cache 0.00001931 0.01171556 + layer.4.k_cache 0.00065893 0.32224767 + layer.4.v_cache 0.00004891 0.02397112 + layer.4.output 0.17367665 619.53471639 + ------------------------------------------------------------------------------------- + TOTAL 0.08704037 257.68713991 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 116328 +BPFP 1.2579 bits/point +EBPFP 2.5157 equivalent bits/point +MSE 257.687140 +---------------------- -------------------------------------------------------- +Time: 3.708s Load: 0.006s, Pack+Encode: 2.245s, Decode+Unpack: 1.457s +---------------------- -------------------------------------------------------- +💾 Converting with 257.6871 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,008B, BPFP=0.4563 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,500B, BPFP=2.5030 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,224B, BPFP=0.7925 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,032B, BPFP=2.4320 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,168B, BPFP=0.9357 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,928B, BPFP=2.4163 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,708B, BPFP=0.8659 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,080B, BPFP=2.4393 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,256B, BPFP=2.3143 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,584B, BPFP=2.3641 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,404B, BPFP=0.2038 +⌛️ [2/4] FRONTEND: Frontend time: 1.919s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.248s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10583916 30.65692316 + layer.0.v_cache 0.00001406 0.00829774 + layer.1.k_cache 0.01351924 8.07523620 + layer.1.v_cache 0.00000536 0.00325722 + layer.2.k_cache 0.00567923 1.42212862 + layer.2.v_cache 0.00001856 0.01041675 + layer.3.k_cache 0.05293757 8.26316226 + layer.3.v_cache 0.00001841 0.01114087 + layer.4.k_cache 0.00064880 0.30581239 + layer.4.v_cache 0.00004836 0.02102870 + layer.4.output 11.07886250 514.10939667 + ------------------------------------------------------------------------------------- + TOTAL 4.57239802 214.56136357 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 124892 +BPFP 1.1145 bits/point +EBPFP 2.2289 equivalent bits/point +MSE 214.561364 +---------------------- -------------------------------------------------------- +Time: 3.172s Load: 0.005s, Pack+Encode: 1.919s, Decode+Unpack: 1.248s +---------------------- -------------------------------------------------------- +💾 Converting with 214.5614 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,556B, BPFP=0.4812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,348B, BPFP=3.0776 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,244B, BPFP=0.7989 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,116B, BPFP=2.8456 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,792B, BPFP=0.9021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,320B, BPFP=2.8840 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,820B, BPFP=0.9074 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,200B, BPFP=2.8614 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,260B, BPFP=2.1197 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,492B, BPFP=2.7282 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,500B, BPFP=0.2555 +⌛️ [2/4] FRONTEND: Frontend time: 1.758s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860243 31.88251835 + layer.0.v_cache 0.00001403 0.00845439 + layer.1.k_cache 0.01400426 7.93315253 + layer.1.v_cache 0.00000584 0.00365736 + layer.2.k_cache 0.01277778 0.99421168 + layer.2.v_cache 0.00001980 0.01157057 + layer.3.k_cache 0.07376247 4.86787065 + layer.3.v_cache 0.00001877 0.01275641 + layer.4.k_cache 0.00062152 0.31374644 + layer.4.v_cache 0.00005140 0.02335771 + layer.4.output 0.18179212 623.91679217 + ------------------------------------------------------------------------------------- + TOTAL 0.08837783 259.61581419 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 113648 +BPFP 1.2585 bits/point +EBPFP 2.5170 equivalent bits/point +MSE 259.615814 +---------------------- -------------------------------------------------------- +Time: 3.109s Load: 0.004s, Pack+Encode: 1.758s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 259.6158 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,708B, BPFP=0.4701 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,708B, BPFP=2.9007 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,196B, BPFP=0.9021 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,288B, BPFP=2.8278 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,404B, BPFP=0.9382 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,704B, BPFP=2.7264 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,220B, BPFP=0.9062 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,288B, BPFP=2.8278 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,352B, BPFP=2.1444 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,592B, BPFP=2.7069 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,212B, BPFP=0.2533 +⌛️ [2/4] FRONTEND: Frontend time: 1.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.238s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13781073 31.29459093 + layer.0.v_cache 0.00001414 0.00873573 + layer.1.k_cache 0.01252146 11.12650825 + layer.1.v_cache 0.00000561 0.00348562 + layer.2.k_cache 0.00238173 1.07489183 + layer.2.v_cache 0.00001849 0.01074732 + layer.3.k_cache 0.04360520 8.05830621 + layer.3.v_cache 0.00001973 0.01286737 + layer.4.k_cache 0.00061668 0.32114351 + layer.4.v_cache 0.00005448 0.02297547 + layer.4.output 0.16772948 587.88804563 + ------------------------------------------------------------------------------------- + TOTAL 0.08065615 245.12650422 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 121672 +BPFP 1.2426 bits/point +EBPFP 2.4851 equivalent bits/point +MSE 245.126504 +---------------------- -------------------------------------------------------- +Time: 3.090s Load: 0.005s, Pack+Encode: 1.847s, Decode+Unpack: 1.238s +---------------------- -------------------------------------------------------- +💾 Converting with 245.1265 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,656B, BPFP=0.5253 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,800B, BPFP=3.1250 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,316B, BPFP=0.8536 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,940B, BPFP=2.9549 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,064B, BPFP=1.0016 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,448B, BPFP=2.8576 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,688B, BPFP=0.9272 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,076B, BPFP=2.9818 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,024B, BPFP=2.3782 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,532B, BPFP=2.8742 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,896B, BPFP=0.2796 +⌛️ [2/4] FRONTEND: Frontend time: 1.954s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.302s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07550411 40.38523845 + layer.0.v_cache 0.00001381 0.00865204 + layer.1.k_cache 0.01271453 10.83467353 + layer.1.v_cache 0.00000531 0.00349231 + layer.2.k_cache 0.00681068 1.12378567 + layer.2.v_cache 0.00001904 0.01140629 + layer.3.k_cache 0.02844424 7.29073440 + layer.3.v_cache 0.00001798 0.01243582 + layer.4.k_cache 0.00061127 0.31020708 + layer.4.v_cache 0.00009323 0.02358825 + layer.4.output 0.18255964 657.94998870 + ------------------------------------------------------------------------------------- + TOTAL 0.08247951 274.45024322 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 113440 +BPFP 1.3198 bits/point +EBPFP 2.6396 equivalent bits/point +MSE 274.450243 +---------------------- -------------------------------------------------------- +Time: 3.261s Load: 0.005s, Pack+Encode: 1.954s, Decode+Unpack: 1.302s +---------------------- -------------------------------------------------------- +💾 Converting with 274.4502 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,584B, BPFP=0.5047 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,648B, BPFP=3.0562 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,268B, BPFP=0.8336 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,740B, BPFP=2.8789 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,848B, BPFP=1.1422 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,548B, BPFP=2.8414 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,120B, BPFP=1.0000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,976B, BPFP=2.9250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,332B, BPFP=2.2133 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,396B, BPFP=2.8117 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,960B, BPFP=0.2779 +⌛️ [2/4] FRONTEND: Frontend time: 1.851s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08590809 32.69654541 + layer.0.v_cache 0.00001388 0.00866443 + layer.1.k_cache 0.01594085 8.18312454 + layer.1.v_cache 0.00000552 0.00344469 + layer.2.k_cache 0.00791870 1.33352757 + layer.2.v_cache 0.00001801 0.01128122 + layer.3.k_cache 0.02757127 7.06531830 + layer.3.v_cache 0.00001862 0.01302577 + layer.4.k_cache 0.00063773 0.31339145 + layer.4.v_cache 0.00005492 0.02401945 + layer.4.output 0.18408036 660.15770089 + ------------------------------------------------------------------------------------- + TOTAL 0.08392059 274.75036759 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 113420 +BPFP 1.3031 bits/point +EBPFP 2.6062 equivalent bits/point +MSE 274.750368 +---------------------- -------------------------------------------------------- +Time: 3.171s Load: 0.005s, Pack+Encode: 1.851s, Decode+Unpack: 1.316s +---------------------- -------------------------------------------------------- +💾 Converting with 274.7504 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,512B, BPFP=0.5164 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,404B, BPFP=3.1669 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,612B, BPFP=0.9482 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,704B, BPFP=2.8174 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,880B, BPFP=1.2089 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,736B, BPFP=2.8240 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,940B, BPFP=1.0156 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,528B, BPFP=2.9868 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,656B, BPFP=2.1908 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,116B, BPFP=2.9021 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,784B, BPFP=0.2874 +⌛️ [2/4] FRONTEND: Frontend time: 1.925s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.430s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08991440 33.09064042 + layer.0.v_cache 0.00001347 0.00876452 + layer.1.k_cache 0.03345199 11.32307755 + layer.1.v_cache 0.00000518 0.00358187 + layer.2.k_cache 0.00398020 1.38244589 + layer.2.v_cache 0.00001815 0.01107215 + layer.3.k_cache 0.04607732 6.48202675 + layer.3.v_cache 0.00001896 0.01240644 + layer.4.k_cache 0.00062119 0.30937265 + layer.4.v_cache 0.00005040 0.02413196 + layer.4.output 0.18612516 698.70541588 + ------------------------------------------------------------------------------------- + TOTAL 0.08688396 290.79914302 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 109872 +BPFP 1.3288 bits/point +EBPFP 2.6575 equivalent bits/point +MSE 290.799143 +---------------------- -------------------------------------------------------- +Time: 3.363s Load: 0.008s, Pack+Encode: 1.925s, Decode+Unpack: 1.430s +---------------------- -------------------------------------------------------- +💾 Converting with 290.7991 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,692B, BPFP=0.4622 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,700B, BPFP=2.8674 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,728B, BPFP=0.9835 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,744B, BPFP=2.7033 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,352B, BPFP=1.0907 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,664B, BPFP=2.6896 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,412B, BPFP=0.9293 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,936B, BPFP=2.7363 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,196B, BPFP=2.2658 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,616B, BPFP=2.6813 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,232B, BPFP=0.2510 +⌛️ [2/4] FRONTEND: Frontend time: 2.101s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.393s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11186004 30.87437221 + layer.0.v_cache 0.00001364 0.00820352 + layer.1.k_cache 0.01386243 13.41820752 + layer.1.v_cache 0.00000608 0.00343464 + layer.2.k_cache 0.00506424 1.58404910 + layer.2.v_cache 0.00001872 0.01042005 + layer.3.k_cache 0.02496567 8.43416386 + layer.3.v_cache 0.00001876 0.01199386 + layer.4.k_cache 0.00064518 0.31457633 + layer.4.v_cache 0.00005436 0.02347585 + layer.4.output 0.16682192 584.04817504 + ------------------------------------------------------------------------------------- + TOTAL 0.07789780 243.70706601 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 123272 +BPFP 1.2451 bits/point +EBPFP 2.4901 equivalent bits/point +MSE 243.707066 +---------------------- -------------------------------------------------------- +Time: 3.500s Load: 0.006s, Pack+Encode: 2.101s, Decode+Unpack: 1.393s +---------------------- -------------------------------------------------------- +💾 Converting with 243.7071 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,568B, BPFP=0.4834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,128B, BPFP=3.0361 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,232B, BPFP=0.7967 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,304B, BPFP=2.8810 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,228B, BPFP=0.9842 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,804B, BPFP=2.7869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,616B, BPFP=0.8690 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,252B, BPFP=2.8712 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,596B, BPFP=2.1830 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,560B, BPFP=2.7410 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,156B, BPFP=0.2731 +⌛️ [2/4] FRONTEND: Frontend time: 2.048s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.275s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08542032 30.61424899 + layer.0.v_cache 0.00001402 0.00848952 + layer.1.k_cache 0.01083013 7.81005124 + layer.1.v_cache 0.00000556 0.00366661 + layer.2.k_cache 0.00382075 1.07907178 + layer.2.v_cache 0.00001860 0.01133664 + layer.3.k_cache 0.02610962 4.16320286 + layer.3.v_cache 0.00002051 0.01255432 + layer.4.k_cache 0.00063312 0.32311118 + layer.4.v_cache 0.00005614 0.02427679 + layer.4.output 0.16811641 618.47875430 + ------------------------------------------------------------------------------------- + TOTAL 0.07669080 257.25889941 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 114444 +BPFP 1.2673 bits/point +EBPFP 2.5346 equivalent bits/point +MSE 257.258899 +---------------------- -------------------------------------------------------- +Time: 3.327s Load: 0.004s, Pack+Encode: 2.048s, Decode+Unpack: 1.275s +---------------------- -------------------------------------------------------- +💾 Converting with 257.2589 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,592B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,820B, BPFP=3.0517 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,256B, BPFP=0.8210 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,260B, BPFP=2.9437 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,112B, BPFP=0.9861 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,728B, BPFP=2.8410 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,836B, BPFP=0.9329 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,068B, BPFP=2.9066 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,356B, BPFP=2.1906 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,632B, BPFP=2.8225 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,416B, BPFP=0.2870 +⌛️ [2/4] FRONTEND: Frontend time: 1.905s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10540552 32.35525174 + layer.0.v_cache 0.00001541 0.00831137 + layer.1.k_cache 0.03402772 7.77335763 + layer.1.v_cache 0.00000555 0.00356466 + layer.2.k_cache 0.00394259 0.97983796 + layer.2.v_cache 0.00001868 0.01104914 + layer.3.k_cache 0.02873192 6.74427098 + layer.3.v_cache 0.00001983 0.01222487 + layer.4.k_cache 0.00063741 0.30941130 + layer.4.v_cache 0.00005806 0.02513683 + layer.4.output 0.17961645 662.18253968 + ------------------------------------------------------------------------------------- + TOTAL 0.08412811 275.50001143 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 114076 +BPFP 1.2944 bits/point +EBPFP 2.5889 equivalent bits/point +MSE 275.500011 +---------------------- -------------------------------------------------------- +Time: 3.261s Load: 0.005s, Pack+Encode: 1.905s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 275.5000 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,624B, BPFP=0.5125 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,752B, BPFP=3.0766 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,332B, BPFP=0.8461 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,972B, BPFP=2.9242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,448B, BPFP=1.0641 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,568B, BPFP=2.8453 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,744B, BPFP=0.9266 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,180B, BPFP=2.9648 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,288B, BPFP=2.2047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,536B, BPFP=2.8391 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,384B, BPFP=0.2897 +⌛️ [2/4] FRONTEND: Frontend time: 2.057s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.278s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10022384 30.16390381 + layer.0.v_cache 0.00001350 0.00871059 + layer.1.k_cache 0.03445883 10.50532150 + layer.1.v_cache 0.00000556 0.00375680 + layer.2.k_cache 0.00550650 1.27882862 + layer.2.v_cache 0.00001848 0.01137021 + layer.3.k_cache 0.03016714 4.80983810 + layer.3.v_cache 0.00001937 0.01397737 + layer.4.k_cache 0.00063435 0.33360534 + layer.4.v_cache 0.00006510 0.02438599 + layer.4.output 0.18593751 641.51651786 + ------------------------------------------------------------------------------------- + TOTAL 0.08662796 266.92760725 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 113828 +BPFP 1.3078 bits/point +EBPFP 2.6155 equivalent bits/point +MSE 266.927607 +---------------------- -------------------------------------------------------- +Time: 3.341s Load: 0.005s, Pack+Encode: 2.057s, Decode+Unpack: 1.278s +---------------------- -------------------------------------------------------- +💾 Converting with 266.9276 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,720B, BPFP=0.4381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,304B, BPFP=2.6263 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,696B, BPFP=0.9175 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,840B, BPFP=2.5515 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,028B, BPFP=1.1321 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,752B, BPFP=2.5374 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,012B, BPFP=1.1295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,920B, BPFP=2.5644 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,176B, BPFP=2.2835 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,448B, BPFP=2.4884 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,000B, BPFP=0.2301 +⌛️ [2/4] FRONTEND: Frontend time: 2.065s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11898830 31.91586461 + layer.0.v_cache 0.00001403 0.00815248 + layer.1.k_cache 0.01243949 12.01042128 + layer.1.v_cache 0.00000551 0.00335293 + layer.2.k_cache 0.00346032 1.58904085 + layer.2.v_cache 0.00001901 0.01052771 + layer.3.k_cache 0.02403709 9.70207041 + layer.3.v_cache 0.00001994 0.01198300 + layer.4.k_cache 0.00063949 0.29455964 + layer.4.v_cache 0.00006591 0.02251060 + layer.4.output 0.03773703 565.66034610 + ------------------------------------------------------------------------------------- + TOTAL 0.02493225 236.18770036 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 125896 +BPFP 1.1929 bits/point +EBPFP 2.3858 equivalent bits/point +MSE 236.187700 +---------------------- -------------------------------------------------------- +Time: 3.457s Load: 0.004s, Pack+Encode: 2.065s, Decode+Unpack: 1.388s +---------------------- -------------------------------------------------------- +💾 Converting with 236.1877 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,560B, BPFP=0.4878 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,800B, BPFP=3.0107 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,268B, BPFP=0.8133 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,012B, BPFP=2.8605 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,448B, BPFP=1.0381 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,732B, BPFP=2.8072 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,684B, BPFP=0.8925 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,952B, BPFP=2.8491 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,712B, BPFP=2.2317 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,416B, BPFP=2.7470 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,100B, BPFP=0.2749 +⌛️ [2/4] FRONTEND: Frontend time: 1.921s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.479s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10693085 31.72776534 + layer.0.v_cache 0.00001341 0.00831146 + layer.1.k_cache 0.01172146 8.16796428 + layer.1.v_cache 0.00000553 0.00346543 + layer.2.k_cache 0.00236719 1.17541001 + layer.2.v_cache 0.00001835 0.01043317 + layer.3.k_cache 0.02622769 4.85730725 + layer.3.v_cache 0.00001864 0.01118463 + layer.4.k_cache 0.00061713 0.31207336 + layer.4.v_cache 0.00004943 0.02314387 + layer.4.output 0.18273007 645.16071429 + ------------------------------------------------------------------------------------- + TOTAL 0.08394589 268.37776816 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 113684 +BPFP 1.2743 bits/point +EBPFP 2.5485 equivalent bits/point +MSE 268.377768 +---------------------- -------------------------------------------------------- +Time: 3.404s Load: 0.004s, Pack+Encode: 1.921s, Decode+Unpack: 1.479s +---------------------- -------------------------------------------------------- +💾 Converting with 268.3778 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,956B, BPFP=0.4573 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,164B, BPFP=2.5006 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,316B, BPFP=0.8224 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,952B, BPFP=2.4678 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,804B, BPFP=0.8979 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,828B, BPFP=2.4486 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,156B, BPFP=0.9524 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,916B, BPFP=2.4623 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,896B, BPFP=2.3045 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,404B, BPFP=2.3830 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,744B, BPFP=0.2153 +⌛️ [2/4] FRONTEND: Frontend time: 2.200s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.492s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13156438 30.10321105 + layer.0.v_cache 0.00001398 0.00865770 + layer.1.k_cache 0.04507210 8.75639540 + layer.1.v_cache 0.00000603 0.00361286 + layer.2.k_cache 0.00848899 1.03730411 + layer.2.v_cache 0.00001983 0.01084043 + layer.3.k_cache 0.06093804 8.27078277 + layer.3.v_cache 0.00001891 0.01177973 + layer.4.k_cache 0.00062755 0.30712502 + layer.4.v_cache 0.00004781 0.02212828 + layer.4.output 11.29576791 509.52090700 + ------------------------------------------------------------------------------------- + TOTAL 4.66571606 212.65754037 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 124136 +BPFP 1.1297 bits/point +EBPFP 2.2593 equivalent bits/point +MSE 212.657540 +---------------------- -------------------------------------------------------- +Time: 3.698s Load: 0.005s, Pack+Encode: 2.200s, Decode+Unpack: 1.492s +---------------------- -------------------------------------------------------- +💾 Converting with 212.6575 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,580B, BPFP=0.4857 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,236B, BPFP=3.0565 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,248B, BPFP=0.7997 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,508B, BPFP=2.9194 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,568B, BPFP=1.0482 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,188B, BPFP=2.8592 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,604B, BPFP=0.8667 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,796B, BPFP=2.9736 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,228B, BPFP=2.1137 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,984B, BPFP=2.8208 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,852B, BPFP=0.2650 +⌛️ [2/4] FRONTEND: Frontend time: 1.962s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.441s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12340720 33.07296275 + layer.0.v_cache 0.00001379 0.00868243 + layer.1.k_cache 0.01365945 7.83199742 + layer.1.v_cache 0.00000586 0.00354327 + layer.2.k_cache 0.00515996 1.09625676 + layer.2.v_cache 0.00002118 0.01109741 + layer.3.k_cache 0.04682968 4.07010586 + layer.3.v_cache 0.00001899 0.01186755 + layer.4.k_cache 0.00062596 0.31690687 + layer.4.v_cache 0.00004975 0.02288677 + layer.4.output 0.18155291 631.51462995 + ------------------------------------------------------------------------------------- + TOTAL 0.08592130 262.76757157 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 115792 +BPFP 1.2822 bits/point +EBPFP 2.5645 equivalent bits/point +MSE 262.767572 +---------------------- -------------------------------------------------------- +Time: 3.408s Load: 0.004s, Pack+Encode: 1.962s, Decode+Unpack: 1.441s +---------------------- -------------------------------------------------------- +💾 Converting with 262.7676 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,612B, BPFP=0.4859 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,240B, BPFP=3.0208 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,316B, BPFP=0.8028 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,612B, BPFP=2.9040 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,160B, BPFP=0.9598 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,856B, BPFP=2.7634 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,760B, BPFP=0.8854 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,124B, BPFP=2.8132 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,268B, BPFP=2.2820 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,040B, BPFP=2.7976 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,900B, BPFP=0.2896 +⌛️ [2/4] FRONTEND: Frontend time: 1.803s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.289s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10400932 30.96861630 + layer.0.v_cache 0.00001381 0.00832571 + layer.1.k_cache 0.01300498 7.72157651 + layer.1.v_cache 0.00000559 0.00340618 + layer.2.k_cache 0.00535372 0.92236301 + layer.2.v_cache 0.00001862 0.01064077 + layer.3.k_cache 0.04415199 3.80790928 + layer.3.v_cache 0.00001841 0.01193014 + layer.4.k_cache 0.00060249 0.31031886 + layer.4.v_cache 0.00005274 0.02367971 + layer.4.output 0.16694923 618.52003614 + ------------------------------------------------------------------------------------- + TOTAL 0.07858096 257.26053056 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 116888 +BPFP 1.2790 bits/point +EBPFP 2.5579 equivalent bits/point +MSE 257.260531 +---------------------- -------------------------------------------------------- +Time: 3.095s Load: 0.003s, Pack+Encode: 1.803s, Decode+Unpack: 1.289s +---------------------- -------------------------------------------------------- +💾 Converting with 257.2605 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,576B, BPFP=0.5031 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,880B, BPFP=3.1016 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,280B, BPFP=0.8359 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,416B, BPFP=2.8156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,820B, BPFP=1.1367 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,572B, BPFP=2.8461 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,824B, BPFP=0.9422 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,832B, BPFP=2.8969 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,552B, BPFP=2.2563 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,460B, BPFP=2.8242 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,184B, BPFP=0.2562 +⌛️ [2/4] FRONTEND: Frontend time: 1.944s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.433s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09321334 30.03128662 + layer.0.v_cache 0.00001862 0.00908398 + layer.1.k_cache 0.03257937 10.11335220 + layer.1.v_cache 0.00000532 0.00349612 + layer.2.k_cache 0.01442847 1.53798962 + layer.2.v_cache 0.00001865 0.01165157 + layer.3.k_cache 0.04385926 4.85398178 + layer.3.v_cache 0.00001828 0.01248649 + layer.4.k_cache 0.00061064 0.31221542 + layer.4.v_cache 0.00005353 0.02356744 + layer.4.output 0.18500509 649.41238839 + ------------------------------------------------------------------------------------- + TOTAL 0.08704948 270.16446059 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 112396 +BPFP 1.2913 bits/point +EBPFP 2.5826 equivalent bits/point +MSE 270.164461 +---------------------- -------------------------------------------------------- +Time: 3.381s Load: 0.004s, Pack+Encode: 1.944s, Decode+Unpack: 1.433s +---------------------- -------------------------------------------------------- +💾 Converting with 270.1645 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,036B, BPFP=0.4606 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,320B, BPFP=2.4757 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,448B, BPFP=0.8265 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,128B, BPFP=2.4466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,052B, BPFP=0.9181 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,840B, BPFP=2.4029 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,700B, BPFP=0.8647 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,044B, BPFP=2.4339 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,900B, BPFP=2.4120 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,552B, BPFP=2.3592 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,636B, BPFP=0.2305 +⌛️ [2/4] FRONTEND: Frontend time: 2.045s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.227s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11527186 32.10658990 + layer.0.v_cache 0.00001510 0.00839193 + layer.1.k_cache 0.02920971 8.25565434 + layer.1.v_cache 0.00000551 0.00345753 + layer.2.k_cache 0.00857845 1.15575090 + layer.2.v_cache 0.00001837 0.01030656 + layer.3.k_cache 0.03550095 4.53372637 + layer.3.v_cache 0.00001950 0.01173975 + layer.4.k_cache 0.00063242 0.30787129 + layer.4.v_cache 0.00005150 0.02299114 + layer.4.output 11.08123415 515.58581831 + ------------------------------------------------------------------------------------- + TOTAL 4.57399661 215.03042399 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 126656 +BPFP 1.1302 bits/point +EBPFP 2.2604 equivalent bits/point +MSE 215.030424 +---------------------- -------------------------------------------------------- +Time: 3.277s Load: 0.005s, Pack+Encode: 2.045s, Decode+Unpack: 1.227s +---------------------- -------------------------------------------------------- +💾 Converting with 215.0304 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,724B, BPFP=0.4626 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,624B, BPFP=2.8234 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,220B, BPFP=1.0564 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,392B, BPFP=2.7840 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,096B, BPFP=1.0353 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,912B, BPFP=2.7024 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,784B, BPFP=0.9823 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,164B, BPFP=2.7452 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,236B, BPFP=2.4178 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,640B, BPFP=2.6562 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,516B, BPFP=0.2309 +⌛️ [2/4] FRONTEND: Frontend time: 2.086s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.487s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09673090 33.20548680 + layer.0.v_cache 0.00001344 0.00812083 + layer.1.k_cache 0.04918879 13.36920962 + layer.1.v_cache 0.00000572 0.00343520 + layer.2.k_cache 0.00880511 1.50028162 + layer.2.v_cache 0.00001944 0.01031181 + layer.3.k_cache 0.06586690 9.26292419 + layer.3.v_cache 0.00001823 0.01176282 + layer.4.k_cache 0.00062484 0.33562706 + layer.4.v_cache 0.00006325 0.02241924 + layer.4.output 0.15294319 584.58792702 + ------------------------------------------------------------------------------------- + TOTAL 0.07599641 244.10853343 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 125308 +BPFP 1.2519 bits/point +EBPFP 2.5038 equivalent bits/point +MSE 244.108533 +---------------------- -------------------------------------------------------- +Time: 3.578s Load: 0.005s, Pack+Encode: 2.086s, Decode+Unpack: 1.487s +---------------------- -------------------------------------------------------- +💾 Converting with 244.1085 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,720B, BPFP=0.4620 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,720B, BPFP=2.8397 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,320B, BPFP=0.9035 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,180B, BPFP=2.7480 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,952B, BPFP=1.0109 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,568B, BPFP=2.6440 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,584B, BPFP=0.9484 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,184B, BPFP=2.7486 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,632B, BPFP=2.3152 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,388B, BPFP=2.6135 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,468B, BPFP=0.2540 +⌛️ [2/4] FRONTEND: Frontend time: 1.918s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.405s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12676482 30.70474641 + layer.0.v_cache 0.00001339 0.00806088 + layer.1.k_cache 0.03133728 11.96151601 + layer.1.v_cache 0.00000590 0.00346175 + layer.2.k_cache 0.00372887 1.48791802 + layer.2.v_cache 0.00001820 0.01071688 + layer.3.k_cache 0.02418065 8.76133264 + layer.3.v_cache 0.00002034 0.01300607 + layer.4.k_cache 0.00062996 0.31445259 + layer.4.v_cache 0.00006053 0.02294956 + layer.4.output 0.15389095 566.66624612 + ------------------------------------------------------------------------------------- + TOTAL 0.07435274 236.46775786 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 123716 +BPFP 1.2360 bits/point +EBPFP 2.4719 equivalent bits/point +MSE 236.467758 +---------------------- -------------------------------------------------------- +Time: 3.327s Load: 0.004s, Pack+Encode: 1.918s, Decode+Unpack: 1.405s +---------------------- -------------------------------------------------------- +💾 Converting with 236.4678 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,692B, BPFP=0.4674 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,716B, BPFP=2.9021 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,560B, BPFP=0.9653 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,072B, BPFP=2.7903 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,084B, BPFP=1.0562 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,484B, BPFP=2.6882 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,120B, BPFP=0.8889 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,896B, BPFP=2.7597 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,788B, BPFP=2.0465 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,196B, BPFP=2.6382 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,360B, BPFP=0.2569 +⌛️ [2/4] FRONTEND: Frontend time: 1.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.483s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12863920 30.70604926 + layer.0.v_cache 0.00001425 0.00829866 + layer.1.k_cache 0.01389826 11.04383409 + layer.1.v_cache 0.00000571 0.00349823 + layer.2.k_cache 0.00659207 1.45914273 + layer.2.v_cache 0.00001883 0.01075421 + layer.3.k_cache 0.05540012 6.59711100 + layer.3.v_cache 0.00001922 0.01176355 + layer.4.k_cache 0.00063279 0.32782709 + layer.4.v_cache 0.00005416 0.02326858 + layer.4.output 0.16196846 597.00610119 + ------------------------------------------------------------------------------------- + TOTAL 0.07876787 248.77848563 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 120968 +BPFP 1.2354 bits/point +EBPFP 2.4708 equivalent bits/point +MSE 248.778486 +---------------------- -------------------------------------------------------- +Time: 3.333s Load: 0.004s, Pack+Encode: 1.847s, Decode+Unpack: 1.483s +---------------------- -------------------------------------------------------- +💾 Converting with 248.7785 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,164B, BPFP=0.4578 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,136B, BPFP=2.3345 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,844B, BPFP=0.8455 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,028B, BPFP=2.3189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,360B, BPFP=1.0648 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,864B, BPFP=2.2951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,644B, BPFP=0.9612 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,080B, BPFP=2.3264 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,008B, BPFP=2.3160 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,640B, BPFP=2.2627 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,800B, BPFP=0.2232 +⌛️ [2/4] FRONTEND: Frontend time: 1.948s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13160179 30.95779984 + layer.0.v_cache 0.00001577 0.00835684 + layer.1.k_cache 0.05775581 11.33265630 + layer.1.v_cache 0.00000549 0.00329928 + layer.2.k_cache 0.01027601 1.60808705 + layer.2.v_cache 0.00002197 0.01076445 + layer.3.k_cache 0.07329086 9.14031756 + layer.3.v_cache 0.00001983 0.01261563 + layer.4.k_cache 0.00064477 0.30782830 + layer.4.v_cache 0.00005260 0.02328513 + layer.4.output 10.56448284 490.73540840 + ------------------------------------------------------------------------------------- + TOTAL 4.36618028 205.20899230 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 129568 +BPFP 1.1027 bits/point +EBPFP 2.2053 equivalent bits/point +MSE 205.208992 +---------------------- -------------------------------------------------------- +Time: 3.279s Load: 0.006s, Pack+Encode: 1.948s, Decode+Unpack: 1.324s +---------------------- -------------------------------------------------------- +💾 Converting with 205.2090 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,548B, BPFP=0.4797 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,412B, BPFP=3.0896 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,252B, BPFP=0.8005 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,396B, BPFP=2.8983 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,240B, BPFP=0.9864 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,904B, BPFP=2.8057 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,492B, BPFP=0.8456 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,164B, BPFP=2.8547 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,472B, BPFP=2.1596 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,720B, BPFP=2.7711 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,660B, BPFP=0.2598 +⌛️ [2/4] FRONTEND: Frontend time: 1.918s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.298s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10139442 31.69169451 + layer.0.v_cache 0.00001346 0.00829546 + layer.1.k_cache 0.03266963 7.72775673 + layer.1.v_cache 0.00000526 0.00346921 + layer.2.k_cache 0.01283502 1.02840488 + layer.2.v_cache 0.00001884 0.01163285 + layer.3.k_cache 0.02666297 4.40139587 + layer.3.v_cache 0.00001780 0.01213731 + layer.4.k_cache 0.00061512 0.32148711 + layer.4.v_cache 0.00004947 0.02292429 + layer.4.output 0.18393325 620.85192556 + ------------------------------------------------------------------------------------- + TOTAL 0.08598910 258.30545160 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 114260 +BPFP 1.2653 bits/point +EBPFP 2.5306 equivalent bits/point +MSE 258.305452 +---------------------- -------------------------------------------------------- +Time: 3.220s Load: 0.004s, Pack+Encode: 1.918s, Decode+Unpack: 1.298s +---------------------- -------------------------------------------------------- +💾 Converting with 258.3055 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,968B, BPFP=0.4592 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,340B, BPFP=2.5278 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,100B, BPFP=0.9437 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,060B, BPFP=2.4845 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,716B, BPFP=1.0390 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,968B, BPFP=2.4703 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,216B, BPFP=0.9616 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,124B, BPFP=2.4944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,032B, BPFP=2.3255 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,404B, BPFP=2.3830 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,368B, BPFP=0.2291 +⌛️ [2/4] FRONTEND: Frontend time: 2.074s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.424s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11149573 30.91707196 + layer.0.v_cache 0.00001357 0.00835000 + layer.1.k_cache 0.08092193 9.36224184 + layer.1.v_cache 0.00000571 0.00336393 + layer.2.k_cache 0.01109533 1.52203414 + layer.2.v_cache 0.00001920 0.01080739 + layer.3.k_cache 0.03617269 7.59697399 + layer.3.v_cache 0.00001858 0.01173379 + layer.4.k_cache 0.00062559 0.30450677 + layer.4.v_cache 0.00005684 0.02327437 + layer.4.output 11.30219499 523.39471358 + ------------------------------------------------------------------------------------- + TOTAL 4.66798765 218.44255019 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 127296 +BPFP 1.1584 bits/point +EBPFP 2.3168 equivalent bits/point +MSE 218.442550 +---------------------- -------------------------------------------------------- +Time: 3.501s Load: 0.004s, Pack+Encode: 2.074s, Decode+Unpack: 1.424s +---------------------- -------------------------------------------------------- +💾 Converting with 218.4426 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,656B, BPFP=0.5321 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,028B, BPFP=3.0104 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,360B, BPFP=0.8734 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,844B, BPFP=2.9736 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,548B, BPFP=1.1114 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,228B, BPFP=2.8502 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,472B, BPFP=0.8958 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,672B, BPFP=2.9391 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,068B, BPFP=2.2171 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,372B, BPFP=2.8790 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,564B, BPFP=0.2737 +⌛️ [2/4] FRONTEND: Frontend time: 1.925s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10167113 36.44545961 + layer.0.v_cache 0.00001386 0.00885445 + layer.1.k_cache 0.03545215 10.50738682 + layer.1.v_cache 0.00000569 0.00361291 + layer.2.k_cache 0.01047806 1.54451341 + layer.2.v_cache 0.00001755 0.01120771 + layer.3.k_cache 0.06687863 4.55459399 + layer.3.v_cache 0.00001960 0.01251553 + layer.4.k_cache 0.00059181 0.31684186 + layer.4.v_cache 0.00004882 0.02505338 + layer.4.output 0.19633365 681.66964286 + ------------------------------------------------------------------------------------- + TOTAL 0.09350076 283.83044351 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 110812 +BPFP 1.3058 bits/point +EBPFP 2.6115 equivalent bits/point +MSE 283.830444 +---------------------- -------------------------------------------------------- +Time: 3.293s Load: 0.004s, Pack+Encode: 1.925s, Decode+Unpack: 1.363s +---------------------- -------------------------------------------------------- +💾 Converting with 283.8304 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,692B, BPFP=0.4835 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,620B, BPFP=2.9849 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,616B, BPFP=0.8290 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,680B, BPFP=2.8161 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,252B, BPFP=0.9432 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,380B, BPFP=2.7622 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,964B, BPFP=0.8915 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,168B, BPFP=2.9037 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,232B, BPFP=2.1968 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,440B, BPFP=2.7730 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,276B, BPFP=0.2636 +⌛️ [2/4] FRONTEND: Frontend time: 2.088s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11445767 31.76266725 + layer.0.v_cache 0.00001583 0.00846831 + layer.1.k_cache 0.03092440 10.94774408 + layer.1.v_cache 0.00000551 0.00343117 + layer.2.k_cache 0.00639992 1.02873782 + layer.2.v_cache 0.00001897 0.01120817 + layer.3.k_cache 0.05613507 4.43330120 + layer.3.v_cache 0.00002026 0.01208956 + layer.4.k_cache 0.00063356 0.31992928 + layer.4.v_cache 0.00005123 0.02239711 + layer.4.output 0.16055210 610.19370895 + ------------------------------------------------------------------------------------- + TOTAL 0.07838395 254.11211392 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 119320 +BPFP 1.2606 bits/point +EBPFP 2.5211 equivalent bits/point +MSE 254.112114 +---------------------- -------------------------------------------------------- +Time: 3.457s Load: 0.004s, Pack+Encode: 2.088s, Decode+Unpack: 1.365s +---------------------- -------------------------------------------------------- +💾 Converting with 254.1121 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,580B, BPFP=0.4799 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,512B, BPFP=3.0714 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,340B, BPFP=0.8073 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,840B, BPFP=2.9464 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,192B, BPFP=0.9658 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,840B, BPFP=2.7604 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,716B, BPFP=0.8772 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,596B, BPFP=2.9010 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,968B, BPFP=2.2262 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,792B, BPFP=2.7515 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,228B, BPFP=0.2452 +⌛️ [2/4] FRONTEND: Frontend time: 2.139s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.703s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13956541 28.88876488 + layer.0.v_cache 0.00001349 0.00830317 + layer.1.k_cache 0.03336124 7.85951160 + layer.1.v_cache 0.00000558 0.00356679 + layer.2.k_cache 0.00698305 1.02923938 + layer.2.v_cache 0.00001831 0.01082696 + layer.3.k_cache 0.07289400 4.94938514 + layer.3.v_cache 0.00001855 0.01210094 + layer.4.k_cache 0.00061006 0.32750152 + layer.4.v_cache 0.00005429 0.02357994 + layer.4.output 0.17891866 634.09725765 + ------------------------------------------------------------------------------------- + TOTAL 0.08858556 263.63491670 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 115604 +BPFP 1.2649 bits/point +EBPFP 2.5298 equivalent bits/point +MSE 263.634917 +---------------------- -------------------------------------------------------- +Time: 3.846s Load: 0.004s, Pack+Encode: 2.139s, Decode+Unpack: 1.703s +---------------------- -------------------------------------------------------- +💾 Converting with 263.6349 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,592B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,520B, BPFP=2.9938 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,216B, BPFP=0.8133 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,096B, BPFP=2.9120 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,336B, BPFP=1.0293 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,344B, BPFP=2.7670 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,840B, BPFP=0.9336 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,972B, BPFP=2.8881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,220B, BPFP=2.1644 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,332B, BPFP=2.7647 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,256B, BPFP=0.2826 +⌛️ [2/4] FRONTEND: Frontend time: 2.085s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.257s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10837562 30.67544367 + layer.0.v_cache 0.00001644 0.00844731 + layer.1.k_cache 0.01241452 7.45520020 + layer.1.v_cache 0.00000532 0.00343749 + layer.2.k_cache 0.01724591 1.12524791 + layer.2.v_cache 0.00001958 0.01067227 + layer.3.k_cache 0.04641535 7.15551532 + layer.3.v_cache 0.00001847 0.01154057 + layer.4.k_cache 0.00063678 0.32685320 + layer.4.v_cache 0.00004974 0.02312632 + layer.4.output 0.18014439 646.88343254 + ------------------------------------------------------------------------------------- + TOTAL 0.08507109 269.11644188 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 112724 +BPFP 1.2791 bits/point +EBPFP 2.5582 equivalent bits/point +MSE 269.116442 +---------------------- -------------------------------------------------------- +Time: 3.346s Load: 0.004s, Pack+Encode: 2.085s, Decode+Unpack: 1.257s +---------------------- -------------------------------------------------------- +💾 Converting with 269.1164 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.2525 bits/point +Avg EBPFP 2.5049 equivalent bits/point +Avg MSE 255.803251 +Avg Time 3.375s +------------------------ ---------------------------- diff --git a/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..af8de3b7e1ad6e3695c6b1a55f1ec74f56aad990 --- /dev/null +++ b/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 333 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean +Output output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean +---------------- ------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,904B, BPFP=0.4212 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,820B, BPFP=2.4130 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,072B, BPFP=0.7900 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,988B, BPFP=2.2822 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,940B, BPFP=0.9946 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,964B, BPFP=2.2805 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,732B, BPFP=0.9084 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,936B, BPFP=2.3499 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,708B, BPFP=1.9769 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,488B, BPFP=2.2466 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,556B, BPFP=0.2197 +⌛️ [2/4] FRONTEND: Frontend time: 2.806s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.554s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12261864 30.71233769 + layer.0.v_cache 0.00001806 0.00864819 + layer.1.k_cache 0.30198812 12.22137646 + layer.1.v_cache 0.00000607 0.00333515 + layer.2.k_cache 0.01966049 1.63404979 + layer.2.v_cache 0.00002021 0.01015360 + layer.3.k_cache 0.01413035 9.10915761 + layer.3.v_cache 0.00002054 0.01108215 + layer.4.k_cache 0.00067868 0.33430174 + layer.4.v_cache 0.00004987 0.01993184 + layer.4.output 1.39792533 245.35594015 + ------------------------------------------------------------------------------------- + TOTAL 0.60262755 104.20917384 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 255108 +BPFP 1.0707 bits/point +EBPFP 2.1413 equivalent bits/point +MSE 104.209174 +---------------------- -------------------------------------------------------- +Time: 4.369s Load: 0.009s, Pack+Encode: 2.806s, Decode+Unpack: 1.554s +---------------------- -------------------------------------------------------- +💾 Converting with 104.2092 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,868B, BPFP=0.4245 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,900B, BPFP=2.3799 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,460B, BPFP=0.8290 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,868B, BPFP=2.3053 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,128B, BPFP=1.0220 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,820B, BPFP=2.3018 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,052B, BPFP=0.9442 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,528B, BPFP=2.3530 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,952B, BPFP=2.0943 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,388B, BPFP=2.2705 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,016B, BPFP=0.1862 +⌛️ [2/4] FRONTEND: Frontend time: 2.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.548s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11787021 28.31277127 + layer.0.v_cache 0.00001738 0.00833810 + layer.1.k_cache 0.25265321 11.43208935 + layer.1.v_cache 0.00000586 0.00332246 + layer.2.k_cache 0.01010228 1.62521984 + layer.2.v_cache 0.00002017 0.01036608 + layer.3.k_cache 0.01159736 9.13798919 + layer.3.v_cache 0.00002197 0.01206388 + layer.4.k_cache 0.00067244 0.33888623 + layer.4.v_cache 0.00004845 0.02030570 + layer.4.output 1.41728832 249.19051753 + ------------------------------------------------------------------------------------- + TOTAL 0.60670750 105.60205734 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 251980 +BPFP 1.0722 bits/point +EBPFP 2.1444 equivalent bits/point +MSE 105.602057 +---------------------- -------------------------------------------------------- +Time: 3.715s Load: 0.011s, Pack+Encode: 2.157s, Decode+Unpack: 1.548s +---------------------- -------------------------------------------------------- +💾 Converting with 105.6021 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,908B, BPFP=0.4140 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,776B, BPFP=2.4367 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,152B, BPFP=0.7814 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,848B, BPFP=2.2315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,772B, BPFP=0.9650 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,644B, BPFP=2.2172 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,948B, BPFP=0.9072 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,260B, BPFP=2.2604 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,192B, BPFP=1.9753 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,176B, BPFP=2.1844 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,464B, BPFP=0.1648 +⌛️ [2/4] FRONTEND: Frontend time: 2.120s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.538s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13311632 33.49755202 + layer.0.v_cache 0.00001526 0.00809620 + layer.1.k_cache 0.36474110 12.36231564 + layer.1.v_cache 0.00000573 0.00328995 + layer.2.k_cache 0.01296863 1.65759510 + layer.2.v_cache 0.00001984 0.01024518 + layer.3.k_cache 0.01490937 10.06515599 + layer.3.v_cache 0.00002017 0.01168920 + layer.4.k_cache 0.00067370 0.33320392 + layer.4.v_cache 0.00005228 0.01984337 + layer.4.output 1.37278975 241.57463165 + ------------------------------------------------------------------------------------- + TOTAL 0.59623828 102.88184753 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 250140 +BPFP 1.0310 bits/point +EBPFP 2.0620 equivalent bits/point +MSE 102.881848 +---------------------- -------------------------------------------------------- +Time: 3.667s Load: 0.010s, Pack+Encode: 2.120s, Decode+Unpack: 1.538s +---------------------- -------------------------------------------------------- +💾 Converting with 102.8818 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,476B, BPFP=0.4147 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,664B, BPFP=2.0917 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,572B, BPFP=0.7410 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,692B, BPFP=2.0295 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,276B, BPFP=0.9142 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,508B, BPFP=2.0177 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,668B, BPFP=0.8753 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,260B, BPFP=2.0658 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,912B, BPFP=1.9155 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,260B, BPFP=2.0018 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,884B, BPFP=0.1910 +⌛️ [2/4] FRONTEND: Frontend time: 2.061s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.867s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11263794 32.13230821 + layer.0.v_cache 0.00001639 0.00832592 + layer.1.k_cache 0.42101050 12.43320653 + layer.1.v_cache 0.00000577 0.00327752 + layer.2.k_cache 0.00499437 1.62647835 + layer.2.v_cache 0.00001973 0.01014717 + layer.3.k_cache 0.03029772 9.86316118 + layer.3.v_cache 0.00002073 0.01180404 + layer.4.k_cache 0.00068836 0.34060685 + layer.4.v_cache 0.00005112 0.02095462 + layer.4.output 1.25471843 220.51489315 + ------------------------------------------------------------------------------------- + TOTAL 0.55016304 94.12085426 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 256172 +BPFP 0.9650 bits/point +EBPFP 1.9299 equivalent bits/point +MSE 94.120854 +---------------------- -------------------------------------------------------- +Time: 3.937s Load: 0.009s, Pack+Encode: 2.061s, Decode+Unpack: 1.867s +---------------------- -------------------------------------------------------- +💾 Converting with 94.1209 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,880B, BPFP=0.4214 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,312B, BPFP=2.3876 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,884B, BPFP=0.7801 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,748B, BPFP=2.3472 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,304B, BPFP=1.0252 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,220B, BPFP=2.5244 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,372B, BPFP=0.9584 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,720B, BPFP=2.5602 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,376B, BPFP=2.0338 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,968B, BPFP=2.3630 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,808B, BPFP=0.2028 +⌛️ [2/4] FRONTEND: Frontend time: 2.282s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.637s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10233516 31.75038749 + layer.0.v_cache 0.00001822 0.00854442 + layer.1.k_cache 0.31275415 11.44815147 + layer.1.v_cache 0.00000587 0.00322987 + layer.2.k_cache 0.01225604 1.65125632 + layer.2.v_cache 0.00002018 0.01028265 + layer.3.k_cache 0.02847941 10.01922887 + layer.3.v_cache 0.00002101 0.01136179 + layer.4.k_cache 0.00067306 0.36078469 + layer.4.v_cache 0.00005947 0.02076922 + layer.4.output 1.40430263 247.13399001 + ------------------------------------------------------------------------------------- + TOTAL 0.60510241 105.01305452 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 262592 +BPFP 1.1071 bits/point +EBPFP 2.2142 equivalent bits/point +MSE 105.013055 +---------------------- -------------------------------------------------------- +Time: 3.928s Load: 0.009s, Pack+Encode: 2.282s, Decode+Unpack: 1.637s +---------------------- -------------------------------------------------------- +💾 Converting with 105.0131 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,864B, BPFP=0.4179 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,004B, BPFP=2.2324 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,360B, BPFP=0.7632 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,992B, BPFP=2.1254 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,768B, BPFP=0.9443 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,568B, BPFP=2.1029 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,320B, BPFP=0.8673 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,560B, BPFP=2.1556 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,024B, BPFP=1.9145 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,024B, BPFP=2.0740 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,464B, BPFP=0.2085 +⌛️ [2/4] FRONTEND: Frontend time: 3.367s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.978s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16298727 35.31097205 + layer.0.v_cache 0.00001770 0.00833557 + layer.1.k_cache 0.61736183 12.87822116 + layer.1.v_cache 0.00000612 0.00328513 + layer.2.k_cache 0.02593859 1.61177208 + layer.2.v_cache 0.00002024 0.00994532 + layer.3.k_cache 0.01326758 9.75288983 + layer.3.v_cache 0.00002270 0.01104307 + layer.4.k_cache 0.00075427 0.36071030 + layer.4.v_cache 0.00005043 0.01956677 + layer.4.output 0.04538776 183.70210763 + ------------------------------------------------------------------------------------- + TOTAL 0.06694947 79.16949969 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 320948 +BPFP 1.0034 bits/point +EBPFP 2.0067 equivalent bits/point +MSE 79.169500 +---------------------- -------------------------------------------------------- +Time: 5.357s Load: 0.012s, Pack+Encode: 3.367s, Decode+Unpack: 1.978s +---------------------- -------------------------------------------------------- +💾 Converting with 79.1695 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 239, 128) +Output shape: (1, 239, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.output: torch.Size([1, 239, 3584]) -> torch.Size([1, 1, 239, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,468B, BPFP=0.4229 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,228B, BPFP=2.1723 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,972B, BPFP=0.7827 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,872B, BPFP=2.0837 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,700B, BPFP=0.9610 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,580B, BPFP=2.0646 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,640B, BPFP=0.8917 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,272B, BPFP=2.1098 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,636B, BPFP=1.9375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,180B, BPFP=2.0384 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,540B, BPFP=0.2292 +⌛️ [2/4] FRONTEND: Frontend time: 2.098s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.520s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11863919 35.04640919 + layer.0.v_cache 0.00001702 0.00853792 + layer.1.k_cache 0.43747612 10.66714337 + layer.1.v_cache 0.00000797 0.00347810 + layer.2.k_cache 0.02286501 1.62439744 + layer.2.v_cache 0.00002174 0.00984634 + layer.3.k_cache 0.01903599 9.64094473 + layer.3.v_cache 0.00002123 0.01093712 + layer.4.k_cache 0.00068597 0.34721451 + layer.4.v_cache 0.00006921 0.02022634 + layer.4.output 1.28101462 225.46824567 + ------------------------------------------------------------------------------------- + TOTAL 0.56270246 96.21510910 + (elements=2,080,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2080256 +Total Bytes 261088 +BPFP 1.0041 bits/point +EBPFP 2.0081 equivalent bits/point +MSE 96.215109 +---------------------- -------------------------------------------------------- +Time: 3.629s Load: 0.010s, Pack+Encode: 2.098s, Decode+Unpack: 1.520s +---------------------- -------------------------------------------------------- +💾 Converting with 96.2151 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,172B, BPFP=0.4261 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,568B, BPFP=2.4102 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,596B, BPFP=0.8077 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,976B, BPFP=2.3156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,856B, BPFP=1.0014 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,804B, BPFP=2.3054 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,616B, BPFP=0.9278 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,684B, BPFP=2.3577 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,108B, BPFP=1.9670 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,388B, BPFP=2.2807 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,556B, BPFP=0.2254 +⌛️ [2/4] FRONTEND: Frontend time: 2.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.595s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15151563 31.13073498 + layer.0.v_cache 0.00001631 0.00835303 + layer.1.k_cache 0.46755480 12.46812062 + layer.1.v_cache 0.00000667 0.00340203 + layer.2.k_cache 0.02171374 1.63472269 + layer.2.v_cache 0.00002115 0.01001385 + layer.3.k_cache 0.01385514 9.47051561 + layer.3.v_cache 0.00002107 0.01095655 + layer.4.k_cache 0.00069081 0.33288969 + layer.4.v_cache 0.00005638 0.02084943 + layer.4.output 0.00504555 210.21818984 + ------------------------------------------------------------------------------------- + TOTAL 0.04063356 89.80105220 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 309324 +BPFP 1.0810 bits/point +EBPFP 2.1620 equivalent bits/point +MSE 89.801052 +---------------------- -------------------------------------------------------- +Time: 3.834s Load: 0.011s, Pack+Encode: 2.229s, Decode+Unpack: 1.595s +---------------------- -------------------------------------------------------- +💾 Converting with 89.8011 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,944B, BPFP=0.4128 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,880B, BPFP=2.2833 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,688B, BPFP=0.8117 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,900B, BPFP=2.2153 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,936B, BPFP=0.9678 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,836B, BPFP=2.2108 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,100B, BPFP=0.9097 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,328B, BPFP=2.2450 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,968B, BPFP=2.0811 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,324B, BPFP=2.1753 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,024B, BPFP=0.2185 +⌛️ [2/4] FRONTEND: Frontend time: 2.032s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.477s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14529029 31.77209201 + layer.0.v_cache 0.00001764 0.00857722 + layer.1.k_cache 0.31089128 11.14914062 + layer.1.v_cache 0.00000630 0.00342788 + layer.2.k_cache 0.01355181 1.61543538 + layer.2.v_cache 0.00002032 0.00993340 + layer.3.k_cache 0.02426839 9.99728733 + layer.3.v_cache 0.00002043 0.01122027 + layer.4.k_cache 0.00068925 0.35660129 + layer.4.v_cache 0.00005165 0.02035473 + layer.4.output 1.36066019 240.05567460 + ------------------------------------------------------------------------------------- + TOTAL 0.58937816 102.07845837 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 256928 +BPFP 1.0495 bits/point +EBPFP 2.0991 equivalent bits/point +MSE 102.078458 +---------------------- -------------------------------------------------------- +Time: 3.518s Load: 0.009s, Pack+Encode: 2.032s, Decode+Unpack: 1.477s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0785 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 262, 128) +Output shape: (1, 262, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.output: torch.Size([1, 262, 3584]) -> torch.Size([1, 1, 262, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,172B, BPFP=0.4277 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,512B, BPFP=2.4160 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,012B, BPFP=0.8356 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,896B, BPFP=2.3197 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,352B, BPFP=0.9752 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,676B, BPFP=2.3065 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,440B, BPFP=0.9208 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,544B, BPFP=2.3583 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,636B, BPFP=2.0060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,992B, BPFP=2.2657 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,116B, BPFP=0.2225 +⌛️ [2/4] FRONTEND: Frontend time: 2.716s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.689s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12917162 30.55024079 + layer.0.v_cache 0.00001830 0.00845988 + layer.1.k_cache 0.40634074 12.15245930 + layer.1.v_cache 0.00000633 0.00345629 + layer.2.k_cache 0.02363680 1.54385050 + layer.2.v_cache 0.00002248 0.01021433 + layer.3.k_cache 0.02343226 8.53573865 + layer.3.v_cache 0.00002042 0.01133619 + layer.4.k_cache 0.00071170 0.35191462 + layer.4.v_cache 0.00005093 0.02039895 + layer.4.output 0.00505940 210.51758451 + ------------------------------------------------------------------------------------- + TOTAL 0.03640161 89.81242124 + (elements=2,280,448) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2280448 +Total Bytes 308348 +BPFP 1.0817 bits/point +EBPFP 2.1634 equivalent bits/point +MSE 89.812421 +---------------------- -------------------------------------------------------- +Time: 4.416s Load: 0.011s, Pack+Encode: 2.716s, Decode+Unpack: 1.689s +---------------------- -------------------------------------------------------- +💾 Converting with 89.8124 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,984B, BPFP=0.4156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,860B, BPFP=2.2819 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,744B, BPFP=0.8156 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,084B, BPFP=2.2281 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,052B, BPFP=0.9758 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,020B, BPFP=2.2236 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,264B, BPFP=0.9211 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,476B, BPFP=2.2553 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,476B, BPFP=2.0469 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,256B, BPFP=2.1706 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,332B, BPFP=0.2215 +⌛️ [2/4] FRONTEND: Frontend time: 2.882s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.908s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18074451 31.55622830 + layer.0.v_cache 0.00001764 0.00863176 + layer.1.k_cache 0.33421082 12.91388021 + layer.1.v_cache 0.00000655 0.00345390 + layer.2.k_cache 0.01487562 1.60668254 + layer.2.v_cache 0.00002082 0.01006515 + layer.3.k_cache 0.01319740 9.76944661 + layer.3.v_cache 0.00002096 0.01124754 + layer.4.k_cache 0.00067848 0.32878760 + layer.4.v_cache 0.00005239 0.01977548 + layer.4.output 1.36066546 239.76988095 + ------------------------------------------------------------------------------------- + TOTAL 0.59226373 102.03631563 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 257548 +BPFP 1.0521 bits/point +EBPFP 2.1042 equivalent bits/point +MSE 102.036316 +---------------------- -------------------------------------------------------- +Time: 4.798s Load: 0.008s, Pack+Encode: 2.882s, Decode+Unpack: 1.908s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0363 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 232, 128) +Output shape: (1, 232, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.output: torch.Size([1, 232, 3584]) -> torch.Size([1, 1, 232, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,264B, BPFP=0.4219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,828B, BPFP=2.2109 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,424B, BPFP=0.7694 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,068B, BPFP=2.1598 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,508B, BPFP=0.9771 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,084B, BPFP=2.1608 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,408B, BPFP=0.9030 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,088B, BPFP=2.2284 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,108B, BPFP=1.9604 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,452B, BPFP=2.1183 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,864B, BPFP=0.2296 +⌛️ [2/4] FRONTEND: Frontend time: 2.361s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.738s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11978169 32.67474576 + layer.0.v_cache 0.00001746 0.00840798 + layer.1.k_cache 0.32299256 11.00236038 + layer.1.v_cache 0.00000596 0.00316994 + layer.2.k_cache 0.02091339 1.62805965 + layer.2.v_cache 0.00002092 0.00996225 + layer.3.k_cache 0.01769928 9.84074454 + layer.3.v_cache 0.00002180 0.01100800 + layer.4.k_cache 0.00071208 0.35290468 + layer.4.v_cache 0.00005763 0.02053265 + layer.4.output 1.31963615 232.95687731 + ------------------------------------------------------------------------------------- + TOTAL 0.57174564 99.19117865 + (elements=2,019,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2019328 +Total Bytes 260096 +BPFP 1.0304 bits/point +EBPFP 2.0609 equivalent bits/point +MSE 99.191179 +---------------------- -------------------------------------------------------- +Time: 4.112s Load: 0.012s, Pack+Encode: 2.361s, Decode+Unpack: 1.738s +---------------------- -------------------------------------------------------- +💾 Converting with 99.1912 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,948B, BPFP=0.4224 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,020B, BPFP=2.3452 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,996B, BPFP=0.7810 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,124B, BPFP=2.2815 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,884B, BPFP=0.9861 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,060B, BPFP=2.2770 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,272B, BPFP=0.9426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,736B, BPFP=2.3250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,744B, BPFP=1.9705 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,612B, BPFP=2.2452 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,784B, BPFP=0.2515 +⌛️ [2/4] FRONTEND: Frontend time: 2.281s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.437s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15639258 31.97949441 + layer.0.v_cache 0.00001700 0.00854265 + layer.1.k_cache 0.33150226 12.03480447 + layer.1.v_cache 0.00000613 0.00331037 + layer.2.k_cache 0.01694729 1.60662939 + layer.2.v_cache 0.00002208 0.00978639 + layer.3.k_cache 0.03858858 9.91890869 + layer.3.v_cache 0.00002090 0.01119355 + layer.4.k_cache 0.00067984 0.34405240 + layer.4.v_cache 0.00005217 0.02106681 + layer.4.output 1.39161012 243.83587662 + ------------------------------------------------------------------------------------- + TOTAL 0.60502939 103.69346620 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 258180 +BPFP 1.0786 bits/point +EBPFP 2.1573 equivalent bits/point +MSE 103.693466 +---------------------- -------------------------------------------------------- +Time: 3.728s Load: 0.011s, Pack+Encode: 2.281s, Decode+Unpack: 1.437s +---------------------- -------------------------------------------------------- +💾 Converting with 103.6935 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,468B, BPFP=0.4274 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,084B, BPFP=2.3514 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,232B, BPFP=0.7573 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,920B, BPFP=2.2848 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,808B, BPFP=0.9620 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,512B, BPFP=2.2614 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,584B, BPFP=0.8919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,320B, BPFP=2.3077 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,704B, BPFP=1.9863 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,960B, BPFP=2.2299 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,916B, BPFP=0.2446 +⌛️ [2/4] FRONTEND: Frontend time: 2.283s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.635s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13605902 30.50373455 + layer.0.v_cache 0.00001757 0.00820085 + layer.1.k_cache 0.58683352 11.64819962 + layer.1.v_cache 0.00000623 0.00333759 + layer.2.k_cache 0.01500385 1.57087098 + layer.2.v_cache 0.00002229 0.00989349 + layer.3.k_cache 0.00978726 9.10797276 + layer.3.v_cache 0.00001998 0.01056267 + layer.4.k_cache 0.00067570 0.33841574 + layer.4.v_cache 0.00005472 0.01978755 + layer.4.output 0.00490606 199.76496272 + ------------------------------------------------------------------------------------- + TOTAL 0.04604839 85.38680675 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 317508 +BPFP 1.0690 bits/point +EBPFP 2.1379 equivalent bits/point +MSE 85.386807 +---------------------- -------------------------------------------------------- +Time: 3.928s Load: 0.009s, Pack+Encode: 2.283s, Decode+Unpack: 1.635s +---------------------- -------------------------------------------------------- +💾 Converting with 85.3868 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,532B, BPFP=0.4002 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,816B, BPFP=2.0108 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,316B, BPFP=0.7547 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,036B, BPFP=1.9630 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,572B, BPFP=0.8316 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,732B, BPFP=1.9444 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,776B, BPFP=0.7828 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,700B, BPFP=2.2488 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,748B, BPFP=1.8228 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,308B, BPFP=1.9184 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,568B, BPFP=0.2238 +⌛️ [2/4] FRONTEND: Frontend time: 2.290s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.538s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12967140 35.69498315 + layer.0.v_cache 0.00001601 0.00772216 + layer.1.k_cache 0.49353925 13.48804573 + layer.1.v_cache 0.00000620 0.00311693 + layer.2.k_cache 0.03189284 1.68784024 + layer.2.v_cache 0.00002040 0.00962347 + layer.3.k_cache 0.04285583 10.08059513 + layer.3.v_cache 0.00002221 0.01044393 + layer.4.k_cache 0.00071132 0.35439737 + layer.4.v_cache 0.00005107 0.01905941 + layer.4.output 1.20063613 208.37354692 + ------------------------------------------------------------------------------------- + TOTAL 0.53548467 89.41003859 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 265104 +BPFP 0.9555 bits/point +EBPFP 1.9111 equivalent bits/point +MSE 89.410039 +---------------------- -------------------------------------------------------- +Time: 3.836s Load: 0.009s, Pack+Encode: 2.290s, Decode+Unpack: 1.538s +---------------------- -------------------------------------------------------- +💾 Converting with 89.4100 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,608B, BPFP=0.4147 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,728B, BPFP=2.0537 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,628B, BPFP=0.7297 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,332B, BPFP=2.0289 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,548B, BPFP=0.9129 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,632B, BPFP=1.9849 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,560B, BPFP=0.8509 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,212B, BPFP=2.0213 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,456B, BPFP=1.9739 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,888B, BPFP=1.9383 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,528B, BPFP=0.1840 +⌛️ [2/4] FRONTEND: Frontend time: 2.078s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.580s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16529650 35.89866497 + layer.0.v_cache 0.00001613 0.00867308 + layer.1.k_cache 0.54577502 11.75974798 + layer.1.v_cache 0.00000662 0.00354909 + layer.2.k_cache 0.01453520 1.59393752 + layer.2.v_cache 0.00002096 0.01066169 + layer.3.k_cache 0.02409829 9.81861234 + layer.3.v_cache 0.00002071 0.01198720 + layer.4.k_cache 0.00068699 0.35371295 + layer.4.v_cache 0.00004971 0.02139555 + layer.4.output 1.22953407 213.58824584 + ------------------------------------------------------------------------------------- + TOTAL 0.55042615 91.44698019 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 258120 +BPFP 0.9528 bits/point +EBPFP 1.9056 equivalent bits/point +MSE 91.446980 +---------------------- -------------------------------------------------------- +Time: 3.666s Load: 0.008s, Pack+Encode: 2.078s, Decode+Unpack: 1.580s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4470 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,860B, BPFP=0.4234 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,620B, BPFP=2.3232 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,396B, BPFP=0.7834 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,160B, BPFP=2.2534 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,232B, BPFP=0.9667 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,564B, BPFP=2.2250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,748B, BPFP=0.8958 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,712B, BPFP=2.2798 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,572B, BPFP=1.9386 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,928B, BPFP=2.1946 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,944B, BPFP=0.2249 +⌛️ [2/4] FRONTEND: Frontend time: 2.801s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.741s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19043710 34.43156895 + layer.0.v_cache 0.00001603 0.00822709 + layer.1.k_cache 0.64205214 12.65212425 + layer.1.v_cache 0.00000630 0.00334886 + layer.2.k_cache 0.01513053 1.57530992 + layer.2.v_cache 0.00002191 0.01008448 + layer.3.k_cache 0.02775778 9.59733073 + layer.3.v_cache 0.00002125 0.01072418 + layer.4.k_cache 0.00071623 0.34436726 + layer.4.v_cache 0.00005146 0.01911803 + layer.4.output 0.04089398 165.34780472 + ------------------------------------------------------------------------------------- + TOTAL 0.06838051 71.53451981 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 373736 +BPFP 1.0505 bits/point +EBPFP 2.1010 equivalent bits/point +MSE 71.534520 +---------------------- -------------------------------------------------------- +Time: 4.553s Load: 0.011s, Pack+Encode: 2.801s, Decode+Unpack: 1.741s +---------------------- -------------------------------------------------------- +💾 Converting with 71.5345 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,560B, BPFP=0.4174 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,500B, BPFP=2.2913 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,036B, BPFP=0.7750 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,300B, BPFP=2.2250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,228B, BPFP=0.9512 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,916B, BPFP=2.2038 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,152B, BPFP=0.8918 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,684B, BPFP=2.2462 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,052B, BPFP=1.9353 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,300B, BPFP=2.1698 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,960B, BPFP=0.2126 +⌛️ [2/4] FRONTEND: Frontend time: 2.182s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.577s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14392723 31.44363198 + layer.0.v_cache 0.00001682 0.00826258 + layer.1.k_cache 0.53349789 13.09804187 + layer.1.v_cache 0.00000638 0.00332360 + layer.2.k_cache 0.01523476 1.61818498 + layer.2.v_cache 0.00002109 0.00971917 + layer.3.k_cache 0.03867486 9.79546364 + layer.3.v_cache 0.00001994 0.01054346 + layer.4.k_cache 0.00070525 0.33846987 + layer.4.v_cache 0.00005078 0.01953532 + layer.4.output 0.00472849 193.91356954 + ------------------------------------------------------------------------------------- + TOTAL 0.04501497 83.16118607 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 318688 +BPFP 1.0350 bits/point +EBPFP 2.0700 equivalent bits/point +MSE 83.161186 +---------------------- -------------------------------------------------------- +Time: 3.770s Load: 0.011s, Pack+Encode: 2.182s, Decode+Unpack: 1.577s +---------------------- -------------------------------------------------------- +💾 Converting with 83.1612 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,900B, BPFP=0.4097 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,740B, BPFP=2.3431 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,736B, BPFP=0.8150 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,008B, BPFP=2.2228 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,032B, BPFP=0.9744 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,616B, BPFP=2.1956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,020B, BPFP=0.9042 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,292B, BPFP=2.2425 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,572B, BPFP=2.0536 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,016B, BPFP=2.1539 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,804B, BPFP=0.2064 +⌛️ [2/4] FRONTEND: Frontend time: 2.048s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.541s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14673459 30.82989149 + layer.0.v_cache 0.00001655 0.00848723 + layer.1.k_cache 0.37581726 11.45552517 + layer.1.v_cache 0.00000603 0.00331039 + layer.2.k_cache 0.01062915 1.61688409 + layer.2.v_cache 0.00002006 0.01009386 + layer.3.k_cache 0.04027582 9.62880534 + layer.3.v_cache 0.00002066 0.01137915 + layer.4.k_cache 0.00066373 0.32335782 + layer.4.v_cache 0.00004861 0.02018248 + layer.4.output 1.36064577 239.80898810 + ------------------------------------------------------------------------------------- + TOTAL 0.59404429 101.91593139 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 255736 +BPFP 1.0447 bits/point +EBPFP 2.0893 equivalent bits/point +MSE 101.915931 +---------------------- -------------------------------------------------------- +Time: 3.597s Load: 0.008s, Pack+Encode: 2.048s, Decode+Unpack: 1.541s +---------------------- -------------------------------------------------------- +💾 Converting with 101.9159 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,304B, BPFP=0.4209 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,100B, BPFP=2.2102 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,932B, BPFP=0.7967 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,148B, BPFP=2.1466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,688B, BPFP=0.9808 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,780B, BPFP=2.1221 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,488B, BPFP=0.9006 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,464B, BPFP=2.1677 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,156B, BPFP=1.9468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,280B, BPFP=2.0887 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,040B, BPFP=0.2293 +⌛️ [2/4] FRONTEND: Frontend time: 2.097s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.530s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12385076 34.85006427 + layer.0.v_cache 0.00002004 0.00913225 + layer.1.k_cache 0.38583452 12.88205921 + layer.1.v_cache 0.00000646 0.00353778 + layer.2.k_cache 0.02500911 1.61823957 + layer.2.v_cache 0.00002014 0.00988038 + layer.3.k_cache 0.03425702 9.77397795 + layer.3.v_cache 0.00002057 0.01161822 + layer.4.k_cache 0.00068928 0.33073634 + layer.4.v_cache 0.00004882 0.01995376 + layer.4.output 1.30836928 230.14278083 + ------------------------------------------------------------------------------------- + TOTAL 0.57225539 98.26521562 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 260380 +BPFP 1.0227 bits/point +EBPFP 2.0455 equivalent bits/point +MSE 98.265216 +---------------------- -------------------------------------------------------- +Time: 3.636s Load: 0.009s, Pack+Encode: 2.097s, Decode+Unpack: 1.530s +---------------------- -------------------------------------------------------- +💾 Converting with 98.2652 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 247, 128) +Output shape: (1, 247, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.output: torch.Size([1, 247, 3584]) -> torch.Size([1, 1, 247, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,584B, BPFP=0.4165 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,416B, BPFP=2.1139 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,724B, BPFP=0.7416 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,904B, BPFP=2.0182 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,392B, BPFP=0.9104 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,376B, BPFP=2.0481 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,448B, BPFP=0.8507 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,172B, BPFP=2.0984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,144B, BPFP=1.8436 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,020B, BPFP=1.9623 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,112B, BPFP=0.2089 +⌛️ [2/4] FRONTEND: Frontend time: 2.023s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.674s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13907220 33.19609533 + layer.0.v_cache 0.00001735 0.00833824 + layer.1.k_cache 0.47208408 11.25263514 + layer.1.v_cache 0.00000611 0.00324187 + layer.2.k_cache 0.03408500 1.61538795 + layer.2.v_cache 0.00001989 0.00969588 + layer.3.k_cache 0.02387486 9.78649952 + layer.3.v_cache 0.00002031 0.01058792 + layer.4.k_cache 0.00069623 0.35322373 + layer.4.v_cache 0.00004922 0.01994057 + layer.4.output 1.23951819 218.14855046 + ------------------------------------------------------------------------------------- + TOTAL 0.54979721 93.13502938 + (elements=2,149,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2149888 +Total Bytes 260292 +BPFP 0.9686 bits/point +EBPFP 1.9372 equivalent bits/point +MSE 93.135029 +---------------------- -------------------------------------------------------- +Time: 3.706s Load: 0.008s, Pack+Encode: 2.023s, Decode+Unpack: 1.674s +---------------------- -------------------------------------------------------- +💾 Converting with 93.1350 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 427, 128) +Output shape: (1, 427, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.output: torch.Size([1, 427, 3584]) -> torch.Size([1, 1, 427, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,340B, BPFP=0.4150 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,584B, BPFP=2.1071 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,244B, BPFP=0.7408 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,908B, BPFP=2.0458 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,464B, BPFP=0.8586 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,156B, BPFP=2.0183 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,024B, BPFP=0.8059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,584B, BPFP=2.0706 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,768B, BPFP=1.7480 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,324B, BPFP=1.9879 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,448B, BPFP=0.2219 +⌛️ [2/4] FRONTEND: Frontend time: 3.091s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.077s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15738957 34.92653140 + layer.0.v_cache 0.00001693 0.00752462 + layer.1.k_cache 1.00627419 12.06668527 + layer.1.v_cache 0.00000663 0.00306233 + layer.2.k_cache 0.03590691 1.58172107 + layer.2.v_cache 0.00002108 0.00968426 + layer.3.k_cache 0.02176077 9.72988213 + layer.3.v_cache 0.00002004 0.00944232 + layer.4.k_cache 0.00078997 0.37417556 + layer.4.v_cache 0.00005212 0.02029750 + layer.4.output 0.00621442 129.05149088 + ------------------------------------------------------------------------------------- + TOTAL 0.07445524 56.59349663 + (elements=3,716,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3716608 +Total Bytes 446844 +BPFP 0.9618 bits/point +EBPFP 1.9237 equivalent bits/point +MSE 56.593497 +---------------------- -------------------------------------------------------- +Time: 5.183s Load: 0.015s, Pack+Encode: 3.091s, Decode+Unpack: 2.077s +---------------------- -------------------------------------------------------- +💾 Converting with 56.5935 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,416B, BPFP=0.4276 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,208B, BPFP=2.3759 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,704B, BPFP=0.7901 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,672B, BPFP=2.2874 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,700B, BPFP=0.9629 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,972B, BPFP=2.2470 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,556B, BPFP=0.8969 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,924B, BPFP=2.3019 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,176B, BPFP=1.8552 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,536B, BPFP=2.2219 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,056B, BPFP=0.2229 +⌛️ [2/4] FRONTEND: Frontend time: 2.362s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.710s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15504150 30.47541657 + layer.0.v_cache 0.00001776 0.00776732 + layer.1.k_cache 0.59434689 12.67807851 + layer.1.v_cache 0.00000612 0.00314458 + layer.2.k_cache 0.01687181 1.56744430 + layer.2.v_cache 0.00002031 0.00993884 + layer.3.k_cache 0.01346795 8.28805170 + layer.3.v_cache 0.00002023 0.01023936 + layer.4.k_cache 0.00071492 0.35558632 + layer.4.v_cache 0.00006004 0.01989275 + layer.4.output 0.00488892 203.85592383 + ------------------------------------------------------------------------------------- + TOTAL 0.04792882 87.08276630 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 310920 +BPFP 1.0545 bits/point +EBPFP 2.1090 equivalent bits/point +MSE 87.082766 +---------------------- -------------------------------------------------------- +Time: 4.083s Load: 0.010s, Pack+Encode: 2.362s, Decode+Unpack: 1.710s +---------------------- -------------------------------------------------------- +💾 Converting with 87.0828 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 361, 128) +Output shape: (1, 361, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.output: torch.Size([1, 361, 3584]) -> torch.Size([1, 1, 361, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,484B, BPFP=0.4105 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,780B, BPFP=2.1113 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,184B, BPFP=0.7438 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,712B, BPFP=2.0651 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,464B, BPFP=0.8857 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,748B, BPFP=2.0667 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,232B, BPFP=0.8324 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,452B, BPFP=2.0971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,856B, BPFP=1.8982 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,488B, BPFP=2.0121 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,772B, BPFP=0.1779 +⌛️ [2/4] FRONTEND: Frontend time: 2.414s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.687s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21057716 35.22003766 + layer.0.v_cache 0.00001554 0.00732829 + layer.1.k_cache 0.82271198 10.31338391 + layer.1.v_cache 0.00000592 0.00294108 + layer.2.k_cache 0.01184860 1.61377477 + layer.2.v_cache 0.00001987 0.00992001 + layer.3.k_cache 0.01345186 9.79631368 + layer.3.v_cache 0.00002111 0.00990758 + layer.4.k_cache 0.00076976 0.34407756 + layer.4.v_cache 0.00005077 0.02088622 + layer.4.output 0.03703259 149.21123120 + ------------------------------------------------------------------------------------- + TOTAL 0.07757063 64.81277583 + (elements=3,142,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3142144 +Total Bytes 378172 +BPFP 0.9628 bits/point +EBPFP 1.9257 equivalent bits/point +MSE 64.812776 +---------------------- -------------------------------------------------------- +Time: 4.115s Load: 0.014s, Pack+Encode: 2.414s, Decode+Unpack: 1.687s +---------------------- -------------------------------------------------------- +💾 Converting with 64.8128 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,500B, BPFP=0.4156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,728B, BPFP=2.2566 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,308B, BPFP=0.7928 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,972B, BPFP=2.2148 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,676B, BPFP=0.9240 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,520B, BPFP=2.1897 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,748B, BPFP=0.8726 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,296B, BPFP=2.2327 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,136B, BPFP=1.7806 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,696B, BPFP=2.1441 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,404B, BPFP=0.2011 +⌛️ [2/4] FRONTEND: Frontend time: 2.187s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.878s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21757852 30.24698027 + layer.0.v_cache 0.00001544 0.00718831 + layer.1.k_cache 0.59705023 12.43276263 + layer.1.v_cache 0.00000603 0.00298752 + layer.2.k_cache 0.00787334 1.57080100 + layer.2.v_cache 0.00001970 0.01018812 + layer.3.k_cache 0.01914395 9.50803240 + layer.3.v_cache 0.00001916 0.00991483 + layer.4.k_cache 0.00071861 0.33342921 + layer.4.v_cache 0.00005003 0.01994136 + layer.4.output 0.00471288 195.77449025 + ------------------------------------------------------------------------------------- + TOTAL 0.05149795 83.79786220 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 310984 +BPFP 1.0136 bits/point +EBPFP 2.0272 equivalent bits/point +MSE 83.797862 +---------------------- -------------------------------------------------------- +Time: 4.074s Load: 0.009s, Pack+Encode: 2.187s, Decode+Unpack: 1.878s +---------------------- -------------------------------------------------------- +💾 Converting with 83.7979 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 372, 128) +Output shape: (1, 372, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.output: torch.Size([1, 372, 3584]) -> torch.Size([1, 1, 372, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,732B, BPFP=0.4088 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,048B, BPFP=2.0601 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,992B, BPFP=0.7557 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,924B, BPFP=2.0129 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,716B, BPFP=0.8701 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,304B, BPFP=1.9869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,232B, BPFP=0.8078 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,544B, BPFP=2.0390 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,760B, BPFP=1.7960 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,524B, BPFP=1.9541 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,272B, BPFP=0.2416 +⌛️ [2/4] FRONTEND: Frontend time: 2.402s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.844s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18192377 36.33337534 + layer.0.v_cache 0.00001786 0.00750773 + layer.1.k_cache 0.82695811 12.77852114 + layer.1.v_cache 0.00000687 0.00309812 + layer.2.k_cache 0.01216373 1.55365745 + layer.2.v_cache 0.00002073 0.01035523 + layer.3.k_cache 0.01743993 9.93588552 + layer.3.v_cache 0.00002103 0.01013832 + layer.4.k_cache 0.00069676 0.34449149 + layer.4.v_cache 0.00005544 0.01970002 + layer.4.output 0.03606073 144.67144297 + ------------------------------------------------------------------------------------- + TOTAL 0.07598408 63.15863713 + (elements=3,237,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3237888 +Total Bytes 390048 +BPFP 0.9637 bits/point +EBPFP 1.9274 equivalent bits/point +MSE 63.158637 +---------------------- -------------------------------------------------------- +Time: 4.258s Load: 0.013s, Pack+Encode: 2.402s, Decode+Unpack: 1.844s +---------------------- -------------------------------------------------------- +💾 Converting with 63.1586 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,568B, BPFP=0.4145 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,536B, BPFP=2.2995 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,260B, BPFP=0.7866 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,440B, BPFP=2.2465 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,992B, BPFP=0.9187 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,356B, BPFP=2.1941 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,988B, BPFP=0.8702 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 49,064B, BPFP=2.3735 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,024B, BPFP=1.7910 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,928B, BPFP=2.1734 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,244B, BPFP=0.2297 +⌛️ [2/4] FRONTEND: Frontend time: 2.326s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.811s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16470559 34.75416022 + layer.0.v_cache 0.00001558 0.00764853 + layer.1.k_cache 0.70577214 12.25487374 + layer.1.v_cache 0.00000663 0.00307836 + layer.2.k_cache 0.01487649 1.54036965 + layer.2.v_cache 0.00002061 0.00988639 + layer.3.k_cache 0.03012228 9.50008088 + layer.3.v_cache 0.00002030 0.01024495 + layer.4.k_cache 0.00074575 0.34728283 + layer.4.v_cache 0.00005065 0.02070612 + layer.4.output 0.04139163 166.43756911 + ------------------------------------------------------------------------------------- + TOTAL 0.07094573 71.97125385 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 365400 +BPFP 1.0398 bits/point +EBPFP 2.0795 equivalent bits/point +MSE 71.971254 +---------------------- -------------------------------------------------------- +Time: 4.148s Load: 0.011s, Pack+Encode: 2.326s, Decode+Unpack: 1.811s +---------------------- -------------------------------------------------------- +💾 Converting with 71.9713 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,244B, BPFP=0.4271 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,076B, BPFP=2.4809 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,004B, BPFP=0.8257 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,372B, BPFP=2.3215 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,704B, BPFP=0.9849 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,956B, BPFP=2.2969 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,324B, BPFP=0.9035 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,724B, BPFP=2.3422 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,824B, BPFP=1.9354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,356B, BPFP=2.2616 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,452B, BPFP=0.2481 +⌛️ [2/4] FRONTEND: Frontend time: 2.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.054s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16585089 29.65170622 + layer.0.v_cache 0.00001549 0.00783661 + layer.1.k_cache 0.52552974 12.30174952 + layer.1.v_cache 0.00000676 0.00316808 + layer.2.k_cache 0.01473994 1.58157061 + layer.2.v_cache 0.00002188 0.01013908 + layer.3.k_cache 0.05129043 8.04068120 + layer.3.v_cache 0.00002012 0.01035800 + layer.4.k_cache 0.00075129 0.34116741 + layer.4.v_cache 0.00005717 0.02040650 + layer.4.output 0.00501330 205.45392520 + ------------------------------------------------------------------------------------- + TOTAL 0.04666922 87.65566233 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 314036 +BPFP 1.0892 bits/point +EBPFP 2.1784 equivalent bits/point +MSE 87.655662 +---------------------- -------------------------------------------------------- +Time: 4.330s Load: 0.009s, Pack+Encode: 2.267s, Decode+Unpack: 2.054s +---------------------- -------------------------------------------------------- +💾 Converting with 87.6557 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.017s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 368, 128) +Output shape: (1, 368, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.output: torch.Size([1, 368, 3584]) -> torch.Size([1, 1, 368, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,732B, BPFP=0.4132 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,068B, BPFP=2.0834 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,428B, BPFP=0.7400 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,840B, BPFP=2.0312 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,352B, BPFP=0.8641 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,368B, BPFP=2.0112 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,280B, BPFP=0.8186 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,436B, BPFP=2.0566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,180B, BPFP=1.7060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,836B, BPFP=1.9886 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,956B, BPFP=0.2181 +⌛️ [2/4] FRONTEND: Frontend time: 2.862s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.241s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14147143 35.27085279 + layer.0.v_cache 0.00001611 0.00737732 + layer.1.k_cache 0.89062865 11.79511958 + layer.1.v_cache 0.00000637 0.00297453 + layer.2.k_cache 0.02201829 1.51547075 + layer.2.v_cache 0.00002088 0.01028411 + layer.3.k_cache 0.00965744 9.88286425 + layer.3.v_cache 0.00002101 0.01024524 + layer.4.k_cache 0.00077426 0.35062197 + layer.4.v_cache 0.00005366 0.02185478 + layer.4.output 0.03638816 146.90457589 + ------------------------------------------------------------------------------------- + TOTAL 0.07761090 63.95292333 + (elements=3,203,072) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3203072 +Total Bytes 382476 +BPFP 0.9553 bits/point +EBPFP 1.9105 equivalent bits/point +MSE 63.952923 +---------------------- -------------------------------------------------------- +Time: 5.119s Load: 0.017s, Pack+Encode: 2.862s, Decode+Unpack: 2.241s +---------------------- -------------------------------------------------------- +💾 Converting with 63.9529 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.018s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 453, 128) +Output shape: (1, 453, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.output: torch.Size([1, 453, 3584]) -> torch.Size([1, 1, 453, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,252B, BPFP=0.4226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 64,496B, BPFP=2.2246 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,384B, BPFP=0.7721 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 62,720B, BPFP=2.1634 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,708B, BPFP=0.8867 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 62,060B, BPFP=2.1406 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,092B, BPFP=0.8310 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 63,344B, BPFP=2.1849 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 49,352B, BPFP=1.7023 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 61,064B, BPFP=2.1062 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,844B, BPFP=0.2111 +⌛️ [2/4] FRONTEND: Frontend time: 3.898s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.664s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13665763 32.62338102 + layer.0.v_cache 0.00001897 0.00774798 + layer.1.k_cache 1.07282356 12.14934077 + layer.1.v_cache 0.00000630 0.00301377 + layer.2.k_cache 0.02241544 1.57196678 + layer.2.v_cache 0.00001979 0.00969262 + layer.3.k_cache 0.01806638 9.26517983 + layer.3.v_cache 0.00002079 0.01033944 + layer.4.k_cache 0.00076567 0.35175201 + layer.4.v_cache 0.00005104 0.01974009 + layer.4.output 0.00584071 121.35386116 + ------------------------------------------------------------------------------------- + TOTAL 0.07598415 53.26406956 + (elements=3,942,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3942912 +Total Bytes 490316 +BPFP 0.9948 bits/point +EBPFP 1.9897 equivalent bits/point +MSE 53.264070 +---------------------- -------------------------------------------------------- +Time: 6.580s Load: 0.018s, Pack+Encode: 3.898s, Decode+Unpack: 2.664s +---------------------- -------------------------------------------------------- +💾 Converting with 53.2641 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.018s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 402, 128) +Output shape: (1, 402, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.output: torch.Size([1, 402, 3584]) -> torch.Size([1, 1, 402, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,820B, BPFP=0.4206 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,792B, BPFP=2.2463 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,996B, BPFP=0.7383 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,684B, BPFP=2.1643 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,984B, BPFP=0.8933 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,260B, BPFP=2.1479 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,868B, BPFP=0.8111 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,808B, BPFP=2.2469 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,280B, BPFP=1.7211 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,816B, BPFP=2.0917 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,524B, BPFP=0.1861 +⌛️ [2/4] FRONTEND: Frontend time: 2.527s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.941s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18062694 33.30830224 + layer.0.v_cache 0.00001752 0.00753557 + layer.1.k_cache 0.89169403 11.73849017 + layer.1.v_cache 0.00000593 0.00287866 + layer.2.k_cache 0.03854401 1.53773643 + layer.2.v_cache 0.00001967 0.00993299 + layer.3.k_cache 0.02336847 9.50147760 + layer.3.v_cache 0.00001924 0.01045820 + layer.4.k_cache 0.00085529 0.36186287 + layer.4.v_cache 0.00004959 0.02099649 + layer.4.output 0.00648942 136.76671331 + ------------------------------------------------------------------------------------- + TOTAL 0.06944863 59.63921555 + (elements=3,499,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3499008 +Total Bytes 431832 +BPFP 0.9873 bits/point +EBPFP 1.9746 equivalent bits/point +MSE 59.639216 +---------------------- -------------------------------------------------------- +Time: 4.486s Load: 0.018s, Pack+Encode: 2.527s, Decode+Unpack: 1.941s +---------------------- -------------------------------------------------------- +💾 Converting with 59.6392 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,468B, BPFP=0.4182 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,392B, BPFP=2.3181 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,952B, BPFP=0.7814 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,036B, BPFP=2.2422 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,340B, BPFP=0.9711 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,680B, BPFP=2.2222 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,784B, BPFP=0.8840 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,732B, BPFP=2.2811 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,988B, BPFP=2.0155 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,272B, BPFP=2.1994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,532B, BPFP=0.2123 +⌛️ [2/4] FRONTEND: Frontend time: 2.323s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.640s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13958113 30.31591622 + layer.0.v_cache 0.00001688 0.00837468 + layer.1.k_cache 0.56864935 12.08902820 + layer.1.v_cache 0.00000621 0.00346555 + layer.2.k_cache 0.01018704 1.64416088 + layer.2.v_cache 0.00002046 0.00997404 + layer.3.k_cache 0.04374025 8.19609480 + layer.3.v_cache 0.00002178 0.01149762 + layer.4.k_cache 0.00068741 0.33724401 + layer.4.v_cache 0.00005272 0.01963168 + layer.4.output 0.00479460 198.29969918 + ------------------------------------------------------------------------------------- + TOTAL 0.04685444 84.74901658 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 318176 +BPFP 1.0482 bits/point +EBPFP 2.0964 equivalent bits/point +MSE 84.749017 +---------------------- -------------------------------------------------------- +Time: 3.973s Load: 0.011s, Pack+Encode: 2.323s, Decode+Unpack: 1.640s +---------------------- -------------------------------------------------------- +💾 Converting with 84.7490 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.017s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 381, 128) +Output shape: (1, 381, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.output: torch.Size([1, 381, 3584]) -> torch.Size([1, 1, 381, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,956B, BPFP=0.4083 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,904B, BPFP=2.0056 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,860B, BPFP=0.7324 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,768B, BPFP=1.9590 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,656B, BPFP=0.8471 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,624B, BPFP=1.9531 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,612B, BPFP=0.8043 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,504B, BPFP=1.9892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,316B, BPFP=1.7354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,788B, BPFP=1.9188 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,380B, BPFP=0.2190 +⌛️ [2/4] FRONTEND: Frontend time: 2.366s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.858s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15713709 38.41368315 + layer.0.v_cache 0.00001719 0.00809464 + layer.1.k_cache 0.86767426 13.38395439 + layer.1.v_cache 0.00000622 0.00332792 + layer.2.k_cache 0.01635739 1.64382334 + layer.2.v_cache 0.00002098 0.01003261 + layer.3.k_cache 0.01709203 9.96933760 + layer.3.v_cache 0.00002073 0.01072818 + layer.4.k_cache 0.00073405 0.35031108 + layer.4.v_cache 0.00005083 0.01930364 + layer.4.output 0.03519400 141.93951537 + ------------------------------------------------------------------------------------- + TOTAL 0.07679228 62.19936495 + (elements=3,316,224) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3316224 +Total Bytes 387368 +BPFP 0.9345 bits/point +EBPFP 1.8690 equivalent bits/point +MSE 62.199365 +---------------------- -------------------------------------------------------- +Time: 4.242s Load: 0.017s, Pack+Encode: 2.366s, Decode+Unpack: 1.858s +---------------------- -------------------------------------------------------- +💾 Converting with 62.1994 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,332B, BPFP=0.4228 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,876B, BPFP=2.1952 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,688B, BPFP=0.7804 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,052B, BPFP=2.1402 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,504B, BPFP=0.9685 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,700B, BPFP=2.1167 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,416B, BPFP=0.8958 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,368B, BPFP=2.1613 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,004B, BPFP=2.0035 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,272B, BPFP=2.0881 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,704B, BPFP=0.2261 +⌛️ [2/4] FRONTEND: Frontend time: 2.341s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.611s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422791 32.98152669 + layer.0.v_cache 0.00001930 0.00858835 + layer.1.k_cache 0.38928963 12.26997780 + layer.1.v_cache 0.00000631 0.00341963 + layer.2.k_cache 0.01623082 1.62871141 + layer.2.v_cache 0.00002095 0.01035555 + layer.3.k_cache 0.01313608 9.77809391 + layer.3.v_cache 0.00002100 0.01128290 + layer.4.k_cache 0.00066994 0.33704357 + layer.4.v_cache 0.00005661 0.02063158 + layer.4.output 1.30836332 230.57392781 + ------------------------------------------------------------------------------------- + TOTAL 0.57013069 98.29806624 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 259916 +BPFP 1.0209 bits/point +EBPFP 2.0418 equivalent bits/point +MSE 98.298066 +---------------------- -------------------------------------------------------- +Time: 3.962s Load: 0.010s, Pack+Encode: 2.341s, Decode+Unpack: 1.611s +---------------------- -------------------------------------------------------- +💾 Converting with 98.2981 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,404B, BPFP=0.4240 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,524B, BPFP=2.2195 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,836B, BPFP=0.7836 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,804B, BPFP=2.1057 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,592B, BPFP=0.9661 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,436B, BPFP=2.0813 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,608B, BPFP=0.9010 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,420B, BPFP=2.1465 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,224B, BPFP=1.8686 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,096B, BPFP=2.0588 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,796B, BPFP=0.2062 +⌛️ [2/4] FRONTEND: Frontend time: 1.969s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.857s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13591524 35.13397527 + layer.0.v_cache 0.00001656 0.00846721 + layer.1.k_cache 0.34520291 12.30935100 + layer.1.v_cache 0.00000598 0.00340005 + layer.2.k_cache 0.01250870 1.61096437 + layer.2.v_cache 0.00001982 0.00989676 + layer.3.k_cache 0.02633334 9.76338429 + layer.3.v_cache 0.00002177 0.01149248 + layer.4.k_cache 0.00067446 0.33226935 + layer.4.v_cache 0.00005295 0.01992938 + layer.4.output 1.29725540 228.50030266 + ------------------------------------------------------------------------------------- + TOTAL 0.56479644 97.57089699 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 256740 +BPFP 0.9999 bits/point +EBPFP 1.9998 equivalent bits/point +MSE 97.570897 +---------------------- -------------------------------------------------------- +Time: 3.833s Load: 0.007s, Pack+Encode: 1.969s, Decode+Unpack: 1.857s +---------------------- -------------------------------------------------------- +💾 Converting with 97.5709 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,864B, BPFP=0.4222 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,308B, BPFP=2.3983 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,788B, BPFP=0.7768 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,480B, BPFP=2.3387 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,988B, BPFP=1.0072 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,912B, BPFP=2.2978 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,880B, BPFP=0.9274 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,572B, BPFP=2.3453 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,556B, BPFP=2.2722 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,380B, BPFP=2.2595 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,240B, BPFP=0.1979 +⌛️ [2/4] FRONTEND: Frontend time: 2.055s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.486s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150146 28.99582598 + layer.0.v_cache 0.00001564 0.00839884 + layer.1.k_cache 0.36060091 11.68563238 + layer.1.v_cache 0.00000598 0.00333100 + layer.2.k_cache 0.01419523 1.64887684 + layer.2.v_cache 0.00002081 0.01015144 + layer.3.k_cache 0.03680107 9.91309944 + layer.3.v_cache 0.00002092 0.01096579 + layer.4.k_cache 0.00067164 0.32983722 + layer.4.v_cache 0.00004838 0.01978812 + layer.4.output 1.41076765 248.11995968 + ------------------------------------------------------------------------------------- + TOTAL 0.61289739 105.26268381 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 255968 +BPFP 1.0842 bits/point +EBPFP 2.1683 equivalent bits/point +MSE 105.262684 +---------------------- -------------------------------------------------------- +Time: 3.550s Load: 0.009s, Pack+Encode: 2.055s, Decode+Unpack: 1.486s +---------------------- -------------------------------------------------------- +💾 Converting with 105.2627 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 343, 128) +Output shape: (1, 343, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.output: torch.Size([1, 343, 3584]) -> torch.Size([1, 1, 343, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,124B, BPFP=0.4156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,464B, BPFP=2.2533 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,168B, BPFP=0.7821 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,264B, BPFP=2.1986 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,288B, BPFP=0.9242 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,760B, BPFP=2.1757 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,412B, BPFP=0.8387 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,836B, BPFP=2.2247 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,592B, BPFP=1.8491 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,980B, BPFP=2.1401 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,748B, BPFP=0.2326 +⌛️ [2/4] FRONTEND: Frontend time: 2.401s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.154s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16677478 32.05730970 + layer.0.v_cache 0.00001698 0.00797588 + layer.1.k_cache 0.72144124 13.06954235 + layer.1.v_cache 0.00000659 0.00330283 + layer.2.k_cache 0.01351720 1.57825490 + layer.2.v_cache 0.00002117 0.00966336 + layer.3.k_cache 0.02347133 9.66627685 + layer.3.v_cache 0.00002006 0.01050929 + layer.4.k_cache 0.00069292 0.33975175 + layer.4.v_cache 0.00005144 0.01935215 + layer.4.output 0.03905215 157.06878644 + ------------------------------------------------------------------------------------- + TOTAL 0.07055169 68.01432024 + (elements=2,985,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2985472 +Total Bytes 382636 +BPFP 1.0253 bits/point +EBPFP 2.0507 equivalent bits/point +MSE 68.014320 +---------------------- -------------------------------------------------------- +Time: 4.567s Load: 0.011s, Pack+Encode: 2.401s, Decode+Unpack: 2.154s +---------------------- -------------------------------------------------------- +💾 Converting with 68.0143 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,532B, BPFP=0.4280 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,256B, BPFP=2.3441 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,844B, BPFP=0.7866 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,852B, BPFP=2.2643 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,768B, BPFP=0.9527 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,280B, BPFP=2.2318 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,252B, BPFP=0.8666 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,448B, BPFP=2.2982 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,236B, BPFP=1.7748 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,960B, BPFP=2.2136 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,500B, BPFP=0.2151 +⌛️ [2/4] FRONTEND: Frontend time: 2.392s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.706s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13317820 30.62671165 + layer.0.v_cache 0.00001784 0.00789229 + layer.1.k_cache 0.51782770 12.10229759 + layer.1.v_cache 0.00000612 0.00322462 + layer.2.k_cache 0.01694398 1.58603815 + layer.2.v_cache 0.00002016 0.00986305 + layer.3.k_cache 0.01920085 9.37121715 + layer.3.v_cache 0.00002016 0.01066221 + layer.4.k_cache 0.00071488 0.33778354 + layer.4.v_cache 0.00005349 0.02164815 + layer.4.output 0.00484358 200.31107143 + ------------------------------------------------------------------------------------- + TOTAL 0.04246403 85.66204932 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 310928 +BPFP 1.0392 bits/point +EBPFP 2.0784 equivalent bits/point +MSE 85.662049 +---------------------- -------------------------------------------------------- +Time: 4.109s Load: 0.011s, Pack+Encode: 2.392s, Decode+Unpack: 1.706s +---------------------- -------------------------------------------------------- +💾 Converting with 85.6620 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 374, 128) +Output shape: (1, 374, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.output: torch.Size([1, 374, 3584]) -> torch.Size([1, 1, 374, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,884B, BPFP=0.4129 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,920B, BPFP=2.0438 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,560B, BPFP=0.7336 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,632B, BPFP=1.9900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,880B, BPFP=0.8723 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,032B, BPFP=1.9649 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,436B, BPFP=0.8120 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,108B, BPFP=2.0099 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,680B, BPFP=1.7831 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,512B, BPFP=1.9432 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,148B, BPFP=0.1799 +⌛️ [2/4] FRONTEND: Frontend time: 2.420s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.798s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18781370 35.54924329 + layer.0.v_cache 0.00001670 0.00783692 + layer.1.k_cache 0.85217587 12.37774430 + layer.1.v_cache 0.00000605 0.00318586 + layer.2.k_cache 0.02919085 1.59991553 + layer.2.v_cache 0.00002159 0.00965831 + layer.3.k_cache 0.02976600 9.96146952 + layer.3.v_cache 0.00002022 0.01054742 + layer.4.k_cache 0.00072409 0.35067329 + layer.4.v_cache 0.00004949 0.02001265 + layer.4.output 0.03578218 144.21411860 + ------------------------------------------------------------------------------------- + TOTAL 0.07942705 62.90524219 + (elements=3,255,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3255296 +Total Bytes 378792 +BPFP 0.9309 bits/point +EBPFP 1.8618 equivalent bits/point +MSE 62.905242 +---------------------- -------------------------------------------------------- +Time: 4.231s Load: 0.013s, Pack+Encode: 2.420s, Decode+Unpack: 1.798s +---------------------- -------------------------------------------------------- +💾 Converting with 62.9052 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,496B, BPFP=0.4212 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,636B, BPFP=2.1159 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,080B, BPFP=0.7832 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,864B, BPFP=2.0659 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,308B, BPFP=0.9276 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,420B, BPFP=2.0371 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,540B, BPFP=0.8779 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,148B, BPFP=2.0843 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,432B, BPFP=1.8434 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,084B, BPFP=2.0153 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,028B, BPFP=0.2040 +⌛️ [2/4] FRONTEND: Frontend time: 2.058s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.496s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13979621 35.55455783 + layer.0.v_cache 0.00001743 0.00807585 + layer.1.k_cache 0.46959303 11.25414330 + layer.1.v_cache 0.00000602 0.00334956 + layer.2.k_cache 0.01642122 1.61818508 + layer.2.v_cache 0.00002053 0.01062993 + layer.3.k_cache 0.03876302 9.72368144 + layer.3.v_cache 0.00002036 0.01150792 + layer.4.k_cache 0.00069112 0.34873772 + layer.4.v_cache 0.00005399 0.02174943 + layer.4.output 1.27033806 222.14204209 + ------------------------------------------------------------------------------------- + TOTAL 0.56222055 94.91464192 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 256036 +BPFP 0.9765 bits/point +EBPFP 1.9529 equivalent bits/point +MSE 94.914642 +---------------------- -------------------------------------------------------- +Time: 3.561s Load: 0.007s, Pack+Encode: 2.058s, Decode+Unpack: 1.496s +---------------------- -------------------------------------------------------- +💾 Converting with 94.9146 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,084B, BPFP=0.4188 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,804B, BPFP=2.3268 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,204B, BPFP=0.7712 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,408B, BPFP=2.2307 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,028B, BPFP=0.9656 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,884B, BPFP=2.1947 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,204B, BPFP=0.9089 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,524B, BPFP=2.2387 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,836B, BPFP=2.0537 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,264B, BPFP=2.1520 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,144B, BPFP=0.2177 +⌛️ [2/4] FRONTEND: Frontend time: 2.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.567s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13893990 31.98073117 + layer.0.v_cache 0.00001711 0.00841577 + layer.1.k_cache 0.36820191 12.16508854 + layer.1.v_cache 0.00000637 0.00342109 + layer.2.k_cache 0.01566778 1.60774937 + layer.2.v_cache 0.00002161 0.01020599 + layer.3.k_cache 0.01884081 9.43753549 + layer.3.v_cache 0.00002076 0.01092159 + layer.4.k_cache 0.00070054 0.34405202 + layer.4.v_cache 0.00005210 0.02039352 + layer.4.output 1.34866800 237.24392306 + ------------------------------------------------------------------------------------- + TOTAL 0.58724382 100.95858682 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 258384 +BPFP 1.0462 bits/point +EBPFP 2.0924 equivalent bits/point +MSE 100.958587 +---------------------- -------------------------------------------------------- +Time: 3.729s Load: 0.008s, Pack+Encode: 2.155s, Decode+Unpack: 1.567s +---------------------- -------------------------------------------------------- +💾 Converting with 100.9586 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,320B, BPFP=0.4284 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,648B, BPFP=2.4373 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,116B, BPFP=0.8261 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,180B, BPFP=2.2928 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,752B, BPFP=0.9803 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,348B, BPFP=2.3027 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,556B, BPFP=0.9103 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,808B, BPFP=2.4466 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,772B, BPFP=2.0349 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,568B, BPFP=2.2570 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,808B, BPFP=0.1990 +⌛️ [2/4] FRONTEND: Frontend time: 2.149s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.684s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11943251 30.86974412 + layer.0.v_cache 0.00001561 0.00794006 + layer.1.k_cache 0.49619702 12.35233150 + layer.1.v_cache 0.00000608 0.00320663 + layer.2.k_cache 0.01349212 1.61700005 + layer.2.v_cache 0.00002006 0.00983048 + layer.3.k_cache 0.01080775 9.73963180 + layer.3.v_cache 0.00002073 0.01051987 + layer.4.k_cache 0.00068886 0.35263499 + layer.4.v_cache 0.00007406 0.02017319 + layer.4.output 0.00494179 205.80666466 + ------------------------------------------------------------------------------------- + TOTAL 0.03972631 87.97821561 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 312876 +BPFP 1.0770 bits/point +EBPFP 2.1541 equivalent bits/point +MSE 87.978216 +---------------------- -------------------------------------------------------- +Time: 3.846s Load: 0.013s, Pack+Encode: 2.149s, Decode+Unpack: 1.684s +---------------------- -------------------------------------------------------- +💾 Converting with 87.9782 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,992B, BPFP=0.4236 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,752B, BPFP=2.3156 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,940B, BPFP=0.7735 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,568B, BPFP=2.3026 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,952B, BPFP=0.9864 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,980B, BPFP=2.2610 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,400B, BPFP=0.8767 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,480B, BPFP=2.2964 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,436B, BPFP=2.0105 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,432B, BPFP=2.2223 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,528B, BPFP=0.2073 +⌛️ [2/4] FRONTEND: Frontend time: 2.328s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.808s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11815372 30.80048519 + layer.0.v_cache 0.00001627 0.00846405 + layer.1.k_cache 0.37623358 12.61811104 + layer.1.v_cache 0.00000616 0.00342414 + layer.2.k_cache 0.01429399 1.65263829 + layer.2.v_cache 0.00002051 0.01029508 + layer.3.k_cache 0.01320818 8.46022607 + layer.3.v_cache 0.00002004 0.01169827 + layer.4.k_cache 0.00067885 0.34878347 + layer.4.v_cache 0.00005318 0.02145489 + layer.4.output 1.38524649 242.87075792 + ------------------------------------------------------------------------------------- + TOTAL 0.60114176 103.17828741 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 253460 +BPFP 1.0541 bits/point +EBPFP 2.1082 equivalent bits/point +MSE 103.178287 +---------------------- -------------------------------------------------------- +Time: 4.144s Load: 0.008s, Pack+Encode: 2.328s, Decode+Unpack: 1.808s +---------------------- -------------------------------------------------------- +💾 Converting with 103.1783 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,580B, BPFP=0.4260 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,704B, BPFP=2.3440 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,416B, BPFP=0.7540 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,948B, BPFP=2.2453 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,776B, BPFP=0.9429 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,396B, BPFP=2.2143 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,468B, BPFP=0.8694 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,412B, BPFP=2.2714 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,752B, BPFP=1.8408 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,104B, BPFP=2.1978 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,380B, BPFP=0.1797 +⌛️ [2/4] FRONTEND: Frontend time: 2.817s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.730s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09820673 30.63033245 + layer.0.v_cache 0.00001666 0.00790363 + layer.1.k_cache 0.50544519 10.62270262 + layer.1.v_cache 0.00000613 0.00333623 + layer.2.k_cache 0.01930715 1.61758840 + layer.2.v_cache 0.00002033 0.01053867 + layer.3.k_cache 0.01337509 9.47609266 + layer.3.v_cache 0.00001987 0.01128617 + layer.4.k_cache 0.00074230 0.33359917 + layer.4.v_cache 0.00004869 0.01981567 + layer.4.output 0.00473525 198.77130974 + ------------------------------------------------------------------------------------- + TOTAL 0.03943146 84.94896258 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 308936 +BPFP 1.0214 bits/point +EBPFP 2.0428 equivalent bits/point +MSE 84.948963 +---------------------- -------------------------------------------------------- +Time: 4.559s Load: 0.012s, Pack+Encode: 2.817s, Decode+Unpack: 1.730s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9490 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,956B, BPFP=0.4269 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,004B, BPFP=2.3655 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,928B, BPFP=0.7833 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,108B, BPFP=2.3013 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,808B, BPFP=0.9897 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,964B, BPFP=2.2910 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,640B, BPFP=0.9060 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,812B, BPFP=2.3518 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,672B, BPFP=1.9834 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,416B, BPFP=2.2517 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,088B, BPFP=0.1750 +⌛️ [2/4] FRONTEND: Frontend time: 2.090s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.600s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13853801 30.74058603 + layer.0.v_cache 0.00001803 0.00855432 + layer.1.k_cache 0.23740530 11.41990172 + layer.1.v_cache 0.00000581 0.00337191 + layer.2.k_cache 0.01896443 1.64571374 + layer.2.v_cache 0.00001977 0.01070869 + layer.3.k_cache 0.00876656 9.41995435 + layer.3.v_cache 0.00002034 0.01212339 + layer.4.k_cache 0.00073522 0.33492489 + layer.4.v_cache 0.00004914 0.02103892 + layer.4.output 1.40427464 246.95498853 + ------------------------------------------------------------------------------------- + TOTAL 0.60202618 104.84128222 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 249396 +BPFP 1.0515 bits/point +EBPFP 2.1030 equivalent bits/point +MSE 104.841282 +---------------------- -------------------------------------------------------- +Time: 3.698s Load: 0.008s, Pack+Encode: 2.090s, Decode+Unpack: 1.600s +---------------------- -------------------------------------------------------- +💾 Converting with 104.8413 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,508B, BPFP=0.4250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,120B, BPFP=2.3279 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,616B, BPFP=0.7708 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,280B, BPFP=2.2803 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,564B, BPFP=0.9377 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,356B, BPFP=2.2280 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,412B, BPFP=0.8725 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,024B, BPFP=2.3225 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,624B, BPFP=2.1300 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,276B, BPFP=2.2235 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,296B, BPFP=0.2046 +⌛️ [2/4] FRONTEND: Frontend time: 2.241s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.587s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10827675 29.27332958 + layer.0.v_cache 0.00001706 0.00791177 + layer.1.k_cache 0.56919524 10.90098329 + layer.1.v_cache 0.00000592 0.00316217 + layer.2.k_cache 0.01911232 1.56667461 + layer.2.v_cache 0.00002018 0.01033514 + layer.3.k_cache 0.01342056 9.07160397 + layer.3.v_cache 0.00002333 0.01129642 + layer.4.k_cache 0.00074072 0.35824646 + layer.4.v_cache 0.00005062 0.02090742 + layer.4.output 0.00480139 199.36713898 + ------------------------------------------------------------------------------------- + TOTAL 0.04379250 85.10555433 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 317076 +BPFP 1.0559 bits/point +EBPFP 2.1118 equivalent bits/point +MSE 85.105554 +---------------------- -------------------------------------------------------- +Time: 3.837s Load: 0.009s, Pack+Encode: 2.241s, Decode+Unpack: 1.587s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1056 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,264B, BPFP=0.4251 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,552B, BPFP=2.4316 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,500B, BPFP=0.8485 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,704B, BPFP=2.3235 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,684B, BPFP=0.9764 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,264B, BPFP=2.3563 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,624B, BPFP=0.9143 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,044B, BPFP=2.3434 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,816B, BPFP=1.8619 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,708B, BPFP=2.2652 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,724B, BPFP=0.2067 +⌛️ [2/4] FRONTEND: Frontend time: 2.346s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.657s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15963619 29.52936637 + layer.0.v_cache 0.00001721 0.00837566 + layer.1.k_cache 0.49929055 13.12175211 + layer.1.v_cache 0.00000624 0.00321979 + layer.2.k_cache 0.03270629 1.59290499 + layer.2.v_cache 0.00002062 0.00989473 + layer.3.k_cache 0.00946706 9.59206113 + layer.3.v_cache 0.00002056 0.01058353 + layer.4.k_cache 0.00068536 0.36702600 + layer.4.v_cache 0.00004941 0.01886391 + layer.4.output 0.00495045 206.09970238 + ------------------------------------------------------------------------------------- + TOTAL 0.04332662 88.05599793 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 310884 +BPFP 1.0702 bits/point +EBPFP 2.1404 equivalent bits/point +MSE 88.055998 +---------------------- -------------------------------------------------------- +Time: 4.015s Load: 0.011s, Pack+Encode: 2.346s, Decode+Unpack: 1.657s +---------------------- -------------------------------------------------------- +💾 Converting with 88.0560 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,396B, BPFP=0.4235 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,664B, BPFP=2.1626 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,656B, BPFP=0.7717 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,984B, BPFP=2.1176 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,276B, BPFP=0.9452 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,408B, BPFP=2.0794 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,424B, BPFP=0.8888 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,288B, BPFP=2.1377 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,544B, BPFP=1.9560 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,192B, BPFP=2.0651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,568B, BPFP=0.1945 +⌛️ [2/4] FRONTEND: Frontend time: 2.081s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.540s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14874478 34.39546229 + layer.0.v_cache 0.00001714 0.00812988 + layer.1.k_cache 0.45078779 11.27864928 + layer.1.v_cache 0.00000615 0.00322397 + layer.2.k_cache 0.01724745 1.59577683 + layer.2.v_cache 0.00001987 0.00952283 + layer.3.k_cache 0.02849353 9.81904111 + layer.3.v_cache 0.00002056 0.01049777 + layer.4.k_cache 0.00066912 0.34151439 + layer.4.v_cache 0.00005068 0.01946648 + layer.4.output 1.29722614 228.35173275 + ------------------------------------------------------------------------------------- + TOTAL 0.57215530 97.40843612 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 255400 +BPFP 0.9947 bits/point +EBPFP 1.9893 equivalent bits/point +MSE 97.408436 +---------------------- -------------------------------------------------------- +Time: 3.631s Load: 0.010s, Pack+Encode: 2.081s, Decode+Unpack: 1.540s +---------------------- -------------------------------------------------------- +💾 Converting with 97.4084 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 341, 128) +Output shape: (1, 341, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.output: torch.Size([1, 341, 3584]) -> torch.Size([1, 1, 341, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,120B, BPFP=0.4179 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,792B, BPFP=2.2815 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,604B, BPFP=0.7608 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,148B, BPFP=2.2062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,096B, BPFP=0.9208 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,540B, BPFP=2.1783 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,156B, BPFP=0.8777 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,620B, BPFP=2.2278 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,032B, BPFP=1.8343 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,876B, BPFP=2.1479 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,108B, BPFP=0.1905 +⌛️ [2/4] FRONTEND: Frontend time: 2.361s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.754s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17009405 30.89410454 + layer.0.v_cache 0.00001631 0.00793559 + layer.1.k_cache 0.79816426 11.11625819 + layer.1.v_cache 0.00000611 0.00313702 + layer.2.k_cache 0.02878924 1.58731446 + layer.2.v_cache 0.00002016 0.00946780 + layer.3.k_cache 0.01035325 9.96496833 + layer.3.v_cache 0.00002082 0.01038825 + layer.4.k_cache 0.00072160 0.35343750 + layer.4.v_cache 0.00005837 0.01905947 + layer.4.output 0.03919861 157.67692187 + ------------------------------------------------------------------------------------- + TOTAL 0.07544909 68.10026613 + (elements=2,968,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2968064 +Total Bytes 375092 +BPFP 1.0110 bits/point +EBPFP 2.0220 equivalent bits/point +MSE 68.100266 +---------------------- -------------------------------------------------------- +Time: 4.128s Load: 0.013s, Pack+Encode: 2.361s, Decode+Unpack: 1.754s +---------------------- -------------------------------------------------------- +💾 Converting with 68.1003 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 335, 128) +Output shape: (1, 335, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.output: torch.Size([1, 335, 3584]) -> torch.Size([1, 1, 335, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,056B, BPFP=0.4224 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 50,992B, BPFP=2.3784 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,880B, BPFP=0.7873 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,532B, BPFP=2.2170 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,572B, BPFP=0.9595 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,760B, BPFP=2.1810 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,016B, BPFP=0.8869 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,208B, BPFP=2.2485 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,104B, BPFP=1.9172 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,424B, BPFP=2.1653 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,588B, BPFP=0.2038 +⌛️ [2/4] FRONTEND: Frontend time: 2.370s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.776s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15195324 32.96620511 + layer.0.v_cache 0.00001666 0.00800278 + layer.1.k_cache 0.76788093 12.48105031 + layer.1.v_cache 0.00000624 0.00322261 + layer.2.k_cache 0.02686723 1.57917098 + layer.2.v_cache 0.00001960 0.00967989 + layer.3.k_cache 0.01080669 9.77009897 + layer.3.v_cache 0.00002024 0.01041284 + layer.4.k_cache 0.00071790 0.33601081 + layer.4.v_cache 0.00005106 0.01908711 + layer.4.output 3.38143948 159.02813166 + ------------------------------------------------------------------------------------- + TOTAL 1.44873036 68.84587430 + (elements=2,915,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2915840 +Total Bytes 377132 +BPFP 1.0347 bits/point +EBPFP 2.0694 equivalent bits/point +MSE 68.845874 +---------------------- -------------------------------------------------------- +Time: 4.159s Load: 0.012s, Pack+Encode: 2.370s, Decode+Unpack: 1.776s +---------------------- -------------------------------------------------------- +💾 Converting with 68.8459 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,180B, BPFP=0.4233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,684B, BPFP=2.3988 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,436B, BPFP=0.8512 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,264B, BPFP=2.4330 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,516B, BPFP=0.9738 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,692B, BPFP=2.2814 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,320B, BPFP=0.9033 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,040B, BPFP=2.4198 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,528B, BPFP=1.8000 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,260B, BPFP=2.2559 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,608B, BPFP=0.2410 +⌛️ [2/4] FRONTEND: Frontend time: 2.189s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.638s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17469128 28.73369878 + layer.0.v_cache 0.00001775 0.00824809 + layer.1.k_cache 0.51265420 12.44468050 + layer.1.v_cache 0.00000620 0.00325599 + layer.2.k_cache 0.03428403 1.59797525 + layer.2.v_cache 0.00002024 0.00957213 + layer.3.k_cache 0.01443097 7.88404702 + layer.3.v_cache 0.00002182 0.01072441 + layer.4.k_cache 0.00070119 0.36648208 + layer.4.v_cache 0.00005097 0.02092256 + layer.4.output 0.00503371 206.80399259 + ------------------------------------------------------------------------------------- + TOTAL 0.04541851 88.15926794 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 312528 +BPFP 1.0840 bits/point +EBPFP 2.1679 equivalent bits/point +MSE 88.159268 +---------------------- -------------------------------------------------------- +Time: 3.836s Load: 0.009s, Pack+Encode: 2.189s, Decode+Unpack: 1.638s +---------------------- -------------------------------------------------------- +💾 Converting with 88.1593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,224B, BPFP=0.4259 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,472B, BPFP=2.3863 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,096B, BPFP=0.8311 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,124B, BPFP=2.3068 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,552B, BPFP=0.9759 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,584B, BPFP=2.2750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,152B, BPFP=0.8934 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,552B, BPFP=2.3321 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,588B, BPFP=1.9215 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,148B, BPFP=2.2493 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,012B, BPFP=0.2023 +⌛️ [2/4] FRONTEND: Frontend time: 2.172s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.622s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18043976 29.84665758 + layer.0.v_cache 0.00001814 0.00805291 + layer.1.k_cache 0.58196952 12.50500719 + layer.1.v_cache 0.00000608 0.00332332 + layer.2.k_cache 0.02409397 1.61393179 + layer.2.v_cache 0.00002020 0.00961117 + layer.3.k_cache 0.01875559 8.89251456 + layer.3.v_cache 0.00002085 0.01091237 + layer.4.k_cache 0.00069073 0.33683319 + layer.4.v_cache 0.00005389 0.02019642 + layer.4.output 0.00498411 208.22587601 + ------------------------------------------------------------------------------------- + TOTAL 0.04946809 88.87224545 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 305504 +BPFP 1.0596 bits/point +EBPFP 2.1192 equivalent bits/point +MSE 88.872245 +---------------------- -------------------------------------------------------- +Time: 3.806s Load: 0.012s, Pack+Encode: 2.172s, Decode+Unpack: 1.622s +---------------------- -------------------------------------------------------- +💾 Converting with 88.8722 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 386, 128) +Output shape: (1, 386, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.output: torch.Size([1, 386, 3584]) -> torch.Size([1, 1, 386, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,188B, BPFP=0.4124 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,144B, BPFP=2.4346 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,080B, BPFP=0.7723 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,336B, BPFP=2.2804 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,624B, BPFP=0.9158 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,244B, BPFP=2.1958 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,108B, BPFP=0.8544 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,552B, BPFP=2.2892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,956B, BPFP=1.9007 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,772B, BPFP=2.1767 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,904B, BPFP=0.2076 +⌛️ [2/4] FRONTEND: Frontend time: 2.463s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.946s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16808204 32.20470875 + layer.0.v_cache 0.00001827 0.00828985 + layer.1.k_cache 0.89383947 11.89160030 + layer.1.v_cache 0.00000605 0.00333437 + layer.2.k_cache 0.02643975 1.58493738 + layer.2.v_cache 0.00002094 0.00970313 + layer.3.k_cache 0.04368651 9.82978718 + layer.3.v_cache 0.00002149 0.01109465 + layer.4.k_cache 0.00075087 0.34900179 + layer.4.v_cache 0.00005208 0.01942552 + layer.4.output 0.03469837 139.87416728 + ------------------------------------------------------------------------------------- + TOTAL 0.08092977 60.88417964 + (elements=3,359,744) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3359744 +Total Bytes 436908 +BPFP 1.0403 bits/point +EBPFP 2.0807 equivalent bits/point +MSE 60.884180 +---------------------- -------------------------------------------------------- +Time: 4.422s Load: 0.012s, Pack+Encode: 2.463s, Decode+Unpack: 1.946s +---------------------- -------------------------------------------------------- +💾 Converting with 60.8842 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,504B, BPFP=0.4071 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,232B, BPFP=2.2370 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,076B, BPFP=0.7637 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,876B, BPFP=2.1634 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,084B, BPFP=0.9269 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,588B, BPFP=2.1478 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,028B, BPFP=0.8696 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,176B, BPFP=2.1797 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,996B, BPFP=1.9529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,868B, BPFP=2.1087 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,296B, BPFP=0.2193 +⌛️ [2/4] FRONTEND: Frontend time: 2.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.692s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14364293 34.03257243 + layer.0.v_cache 0.00001600 0.00843090 + layer.1.k_cache 0.62366660 12.50445472 + layer.1.v_cache 0.00000628 0.00349111 + layer.2.k_cache 0.01981349 1.57935026 + layer.2.v_cache 0.00002090 0.01020077 + layer.3.k_cache 0.02158341 9.87627411 + layer.3.v_cache 0.00002097 0.01129264 + layer.4.k_cache 0.00071696 0.33849981 + layer.4.v_cache 0.00005883 0.02025553 + layer.4.output 0.00463133 191.36247520 + ------------------------------------------------------------------------------------- + TOTAL 0.04952739 82.23071463 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 318724 +BPFP 1.0172 bits/point +EBPFP 2.0343 equivalent bits/point +MSE 82.230715 +---------------------- -------------------------------------------------------- +Time: 4.038s Load: 0.012s, Pack+Encode: 2.334s, Decode+Unpack: 1.692s +---------------------- -------------------------------------------------------- +💾 Converting with 82.2307 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,488B, BPFP=0.4224 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,152B, BPFP=2.3213 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,704B, BPFP=0.7730 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,036B, BPFP=2.2583 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,952B, BPFP=0.9562 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,852B, BPFP=2.2480 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,552B, BPFP=0.8773 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,912B, BPFP=2.3078 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,704B, BPFP=1.8448 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,312B, BPFP=2.2175 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,440B, BPFP=0.2453 +⌛️ [2/4] FRONTEND: Frontend time: 2.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.664s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12300906 29.57363528 + layer.0.v_cache 0.00001720 0.00762684 + layer.1.k_cache 0.58060381 11.40672947 + layer.1.v_cache 0.00000644 0.00305445 + layer.2.k_cache 0.02305070 1.53910084 + layer.2.v_cache 0.00002053 0.00993359 + layer.3.k_cache 0.05279791 9.51501509 + layer.3.v_cache 0.00002021 0.01040183 + layer.4.k_cache 0.00074809 0.37165103 + layer.4.v_cache 0.00005348 0.02022794 + layer.4.output 0.00483522 199.56125903 + ------------------------------------------------------------------------------------- + TOTAL 0.04789259 85.25801115 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 318104 +BPFP 1.0555 bits/point +EBPFP 2.1110 equivalent bits/point +MSE 85.258011 +---------------------- -------------------------------------------------------- +Time: 3.805s Load: 0.010s, Pack+Encode: 2.131s, Decode+Unpack: 1.664s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2580 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,880B, BPFP=0.4120 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,076B, BPFP=2.3876 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,588B, BPFP=0.8119 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,840B, BPFP=2.2309 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,772B, BPFP=0.9650 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,796B, BPFP=2.2279 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,688B, BPFP=0.8890 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,140B, BPFP=2.4622 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,232B, BPFP=2.0482 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,244B, BPFP=2.1892 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,408B, BPFP=0.2043 +⌛️ [2/4] FRONTEND: Frontend time: 2.205s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.460s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16177035 31.07584781 + layer.0.v_cache 0.00001782 0.00885899 + layer.1.k_cache 0.31952496 13.37871247 + layer.1.v_cache 0.00000626 0.00349052 + layer.2.k_cache 0.01271096 1.59926232 + layer.2.v_cache 0.00002040 0.01024264 + layer.3.k_cache 0.01363577 9.66508723 + layer.3.v_cache 0.00002033 0.01152674 + layer.4.k_cache 0.00066992 0.33035579 + layer.4.v_cache 0.00005030 0.01937873 + layer.4.output 1.37284199 240.61883408 + ------------------------------------------------------------------------------------- + TOTAL 0.59519535 102.37850599 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 257664 +BPFP 1.0620 bits/point +EBPFP 2.1240 equivalent bits/point +MSE 102.378506 +---------------------- -------------------------------------------------------- +Time: 3.673s Load: 0.008s, Pack+Encode: 2.205s, Decode+Unpack: 1.460s +---------------------- -------------------------------------------------------- +💾 Converting with 102.3785 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,900B, BPFP=0.4184 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,884B, BPFP=2.1247 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,856B, BPFP=0.7561 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,772B, BPFP=2.0297 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,504B, BPFP=0.9822 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,776B, BPFP=2.0301 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,924B, BPFP=0.9327 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,376B, BPFP=2.0813 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,604B, BPFP=1.8446 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,584B, BPFP=2.0137 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,972B, BPFP=0.2314 +⌛️ [2/4] FRONTEND: Frontend time: 2.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.405s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09962382 32.43759606 + layer.0.v_cache 0.00001686 0.00893629 + layer.1.k_cache 0.10988800 10.54005240 + layer.1.v_cache 0.00000590 0.00353979 + layer.2.k_cache 0.01214521 1.63786975 + layer.2.v_cache 0.00002216 0.01070556 + layer.3.k_cache 0.02260447 9.88084403 + layer.3.v_cache 0.00002006 0.01247315 + layer.4.k_cache 0.00069723 0.36203086 + layer.4.v_cache 0.00005242 0.02182270 + layer.4.output 0.00882482 299.06328064 + ------------------------------------------------------------------------------------- + TOTAL 0.01804999 126.37404912 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 197152 +BPFP 0.9902 bits/point +EBPFP 1.9804 equivalent bits/point +MSE 126.374049 +---------------------- -------------------------------------------------------- +Time: 3.573s Load: 0.006s, Pack+Encode: 2.162s, Decode+Unpack: 1.405s +---------------------- -------------------------------------------------------- +💾 Converting with 126.3740 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,544B, BPFP=0.4195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,388B, BPFP=2.3014 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,840B, BPFP=0.7696 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,312B, BPFP=2.2415 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,736B, BPFP=0.9306 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,600B, BPFP=2.2020 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,096B, BPFP=0.8950 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,796B, BPFP=2.2685 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,408B, BPFP=1.8020 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,112B, BPFP=2.1748 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,720B, BPFP=0.2043 +⌛️ [2/4] FRONTEND: Frontend time: 2.268s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.046s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11148910 32.56997192 + layer.0.v_cache 0.00001684 0.00813055 + layer.1.k_cache 0.60671574 11.98839333 + layer.1.v_cache 0.00000630 0.00314462 + layer.2.k_cache 0.04149003 1.61584060 + layer.2.v_cache 0.00001976 0.00936437 + layer.3.k_cache 0.02200458 9.68812990 + layer.3.v_cache 0.00001954 0.01038338 + layer.4.k_cache 0.00078895 0.35647950 + layer.4.v_cache 0.00004864 0.01960660 + layer.4.output 0.00472744 196.60968480 + ------------------------------------------------------------------------------------- + TOTAL 0.04798185 84.26689637 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 313552 +BPFP 1.0256 bits/point +EBPFP 2.0512 equivalent bits/point +MSE 84.266896 +---------------------- -------------------------------------------------------- +Time: 4.326s Load: 0.012s, Pack+Encode: 2.268s, Decode+Unpack: 2.046s +---------------------- -------------------------------------------------------- +💾 Converting with 84.2669 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,524B, BPFP=0.4316 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,656B, BPFP=2.5513 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,832B, BPFP=0.8462 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,948B, BPFP=2.4178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,644B, BPFP=1.0659 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,668B, BPFP=2.3959 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,580B, BPFP=0.9828 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,336B, BPFP=2.4481 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,620B, BPFP=2.0797 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,280B, BPFP=2.3656 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,472B, BPFP=0.1950 +⌛️ [2/4] FRONTEND: Frontend time: 2.006s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.429s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14565937 29.93160645 + layer.0.v_cache 0.00001550 0.00833001 + layer.1.k_cache 0.22559690 12.85841064 + layer.1.v_cache 0.00000590 0.00346428 + layer.2.k_cache 0.00679684 1.58715866 + layer.2.v_cache 0.00002314 0.01061699 + layer.3.k_cache 0.06866968 9.39696411 + layer.3.v_cache 0.00002119 0.01211696 + layer.4.k_cache 0.00065507 0.34241150 + layer.4.v_cache 0.00005191 0.02091885 + layer.4.output 1.53062134 270.29024554 + ------------------------------------------------------------------------------------- + TOTAL 0.65657911 114.48257160 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 242560 +BPFP 1.1147 bits/point +EBPFP 2.2294 equivalent bits/point +MSE 114.482572 +---------------------- -------------------------------------------------------- +Time: 3.445s Load: 0.011s, Pack+Encode: 2.006s, Decode+Unpack: 1.429s +---------------------- -------------------------------------------------------- +💾 Converting with 114.4826 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,516B, BPFP=0.4288 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,000B, BPFP=2.4876 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,204B, BPFP=0.8710 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,980B, BPFP=2.4083 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,492B, BPFP=1.0488 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,336B, BPFP=2.3582 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,400B, BPFP=0.9639 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,148B, BPFP=2.4213 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,180B, BPFP=2.1906 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,164B, BPFP=2.3448 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,696B, BPFP=0.1965 +⌛️ [2/4] FRONTEND: Frontend time: 1.973s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.458s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10663509 30.32279277 + layer.0.v_cache 0.00001613 0.00882666 + layer.1.k_cache 0.26620468 13.13641145 + layer.1.v_cache 0.00000602 0.00357881 + layer.2.k_cache 0.00649704 1.56738205 + layer.2.v_cache 0.00002002 0.01052114 + layer.3.k_cache 0.02565914 8.50628450 + layer.3.v_cache 0.00002186 0.01221386 + layer.4.k_cache 0.00070101 0.33647729 + layer.4.v_cache 0.00005421 0.02224619 + layer.4.output 1.52303164 267.82054016 + ------------------------------------------------------------------------------------- + TOTAL 0.65100216 113.45120681 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 243116 +BPFP 1.1117 bits/point +EBPFP 2.2234 equivalent bits/point +MSE 113.451207 +---------------------- -------------------------------------------------------- +Time: 3.438s Load: 0.007s, Pack+Encode: 1.973s, Decode+Unpack: 1.458s +---------------------- -------------------------------------------------------- +💾 Converting with 113.4512 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,332B, BPFP=0.4291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,200B, BPFP=2.4110 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,328B, BPFP=0.8385 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,524B, BPFP=2.3130 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,792B, BPFP=0.9827 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,040B, BPFP=2.2846 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,876B, BPFP=0.9291 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,100B, BPFP=2.3467 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,428B, BPFP=1.9562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,592B, BPFP=2.2584 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,720B, BPFP=0.2150 +⌛️ [2/4] FRONTEND: Frontend time: 2.344s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.052s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258070 29.21417625 + layer.0.v_cache 0.00001713 0.00846564 + layer.1.k_cache 0.49703808 12.68032300 + layer.1.v_cache 0.00000670 0.00351602 + layer.2.k_cache 0.01458295 1.54254779 + layer.2.v_cache 0.00002181 0.01002498 + layer.3.k_cache 0.02940617 9.28886042 + layer.3.v_cache 0.00002120 0.01134903 + layer.4.k_cache 0.00070157 0.34454857 + layer.4.v_cache 0.00005095 0.01955177 + layer.4.output 0.00497170 206.93174826 + ------------------------------------------------------------------------------------- + TOTAL 0.04171936 88.33209419 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 311932 +BPFP 1.0738 bits/point +EBPFP 2.1476 equivalent bits/point +MSE 88.332094 +---------------------- -------------------------------------------------------- +Time: 4.405s Load: 0.010s, Pack+Encode: 2.344s, Decode+Unpack: 2.052s +---------------------- -------------------------------------------------------- +💾 Converting with 88.3321 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,004B, BPFP=0.4151 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,388B, BPFP=2.2392 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,376B, BPFP=0.7865 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,720B, BPFP=2.1930 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,388B, BPFP=0.9947 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,572B, BPFP=2.1828 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,212B, BPFP=0.9134 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,348B, BPFP=2.2364 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,228B, BPFP=2.0899 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,076B, BPFP=2.1485 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,692B, BPFP=0.1945 +⌛️ [2/4] FRONTEND: Frontend time: 1.997s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.710s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15433978 33.32590613 + layer.0.v_cache 0.00001575 0.00839658 + layer.1.k_cache 0.29362690 11.71983351 + layer.1.v_cache 0.00000607 0.00344798 + layer.2.k_cache 0.00940348 1.64920030 + layer.2.v_cache 0.00002159 0.01024921 + layer.3.k_cache 0.02320663 9.72257307 + layer.3.v_cache 0.00002067 0.01194801 + layer.4.k_cache 0.00067785 0.33303114 + layer.4.v_cache 0.00005016 0.02039008 + layer.4.output 1.35461534 239.10872314 + ------------------------------------------------------------------------------------- + TOTAL 0.58609860 101.79800223 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 254004 +BPFP 1.0330 bits/point +EBPFP 2.0660 equivalent bits/point +MSE 101.798002 +---------------------- -------------------------------------------------------- +Time: 3.717s Load: 0.009s, Pack+Encode: 1.997s, Decode+Unpack: 1.710s +---------------------- -------------------------------------------------------- +💾 Converting with 101.7980 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,648B, BPFP=0.4193 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,628B, BPFP=2.3371 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,416B, BPFP=0.7355 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,428B, BPFP=2.2164 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,096B, BPFP=0.9373 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,108B, BPFP=2.1989 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,896B, BPFP=0.8715 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,888B, BPFP=2.2417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,508B, BPFP=2.1660 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,148B, BPFP=2.1463 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,656B, BPFP=0.2479 +⌛️ [2/4] FRONTEND: Frontend time: 2.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.754s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14751465 30.77996162 + layer.0.v_cache 0.00001719 0.00824594 + layer.1.k_cache 0.52245269 12.43323482 + layer.1.v_cache 0.00000618 0.00324720 + layer.2.k_cache 0.02707210 1.64673055 + layer.2.v_cache 0.00002227 0.00997446 + layer.3.k_cache 0.03364062 10.04047766 + layer.3.v_cache 0.00002083 0.01087791 + layer.4.k_cache 0.00069748 0.35907304 + layer.4.v_cache 0.00005305 0.01894196 + layer.4.output 0.00471531 189.41799812 + ------------------------------------------------------------------------------------- + TOTAL 0.04497084 81.24922071 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 328420 +BPFP 1.0591 bits/point +EBPFP 2.1183 equivalent bits/point +MSE 81.249221 +---------------------- -------------------------------------------------------- +Time: 3.920s Load: 0.010s, Pack+Encode: 2.156s, Decode+Unpack: 1.754s +---------------------- -------------------------------------------------------- +💾 Converting with 81.2492 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,524B, BPFP=0.3998 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,516B, BPFP=1.9924 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,000B, BPFP=0.7353 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,348B, BPFP=1.9821 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,396B, BPFP=0.8208 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,904B, BPFP=1.9549 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,708B, BPFP=0.7787 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,940B, BPFP=2.1409 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,196B, BPFP=1.9115 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,212B, BPFP=1.9125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,452B, BPFP=0.2666 +⌛️ [2/4] FRONTEND: Frontend time: 2.175s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.568s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17432048 33.91587776 + layer.0.v_cache 0.00001704 0.00739579 + layer.1.k_cache 0.39057378 12.42836052 + layer.1.v_cache 0.00000668 0.00319546 + layer.2.k_cache 0.01299327 1.67432706 + layer.2.v_cache 0.00002114 0.00929797 + layer.3.k_cache 0.02499930 10.18789637 + layer.3.v_cache 0.00002096 0.01010538 + layer.4.k_cache 0.00069121 0.35075133 + layer.4.v_cache 0.00005416 0.01876744 + layer.4.output 1.20070616 205.16285014 + ------------------------------------------------------------------------------------- + TOTAL 0.52992007 87.92623094 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 269196 +BPFP 0.9703 bits/point +EBPFP 1.9406 equivalent bits/point +MSE 87.926231 +---------------------- -------------------------------------------------------- +Time: 3.751s Load: 0.009s, Pack+Encode: 2.175s, Decode+Unpack: 1.568s +---------------------- -------------------------------------------------------- +💾 Converting with 87.9262 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,160B, BPFP=0.4203 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,552B, BPFP=2.2211 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,476B, BPFP=0.7830 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,664B, BPFP=2.1605 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,032B, BPFP=0.9574 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,640B, BPFP=2.1588 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,372B, BPFP=0.9124 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,268B, BPFP=2.2017 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,020B, BPFP=1.9801 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,100B, BPFP=2.1220 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,360B, BPFP=0.2277 +⌛️ [2/4] FRONTEND: Frontend time: 1.992s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.509s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16239141 32.90001748 + layer.0.v_cache 0.00001645 0.00837453 + layer.1.k_cache 0.33751179 12.58142037 + layer.1.v_cache 0.00000668 0.00343429 + layer.2.k_cache 0.00660169 1.54527609 + layer.2.v_cache 0.00002045 0.00975514 + layer.3.k_cache 0.02213682 9.83368480 + layer.3.v_cache 0.00002049 0.01120804 + layer.4.k_cache 0.00067824 0.34742110 + layer.4.v_cache 0.00005183 0.01975934 + layer.4.output 1.33690934 235.72333125 + ------------------------------------------------------------------------------------- + TOTAL 0.58163537 100.43080412 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 256644 +BPFP 1.0301 bits/point +EBPFP 2.0601 equivalent bits/point +MSE 100.430804 +---------------------- -------------------------------------------------------- +Time: 3.509s Load: 0.008s, Pack+Encode: 1.992s, Decode+Unpack: 1.509s +---------------------- -------------------------------------------------------- +💾 Converting with 100.4308 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,368B, BPFP=0.4296 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,560B, BPFP=2.3647 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,404B, BPFP=0.8398 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,728B, BPFP=2.3162 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,404B, BPFP=0.9564 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,036B, BPFP=2.2759 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,264B, BPFP=0.8899 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,788B, BPFP=2.3197 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,428B, BPFP=1.9489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,688B, BPFP=2.2556 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,480B, BPFP=0.1872 +⌛️ [2/4] FRONTEND: Frontend time: 2.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.656s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15796844 30.21738901 + layer.0.v_cache 0.00001527 0.00757484 + layer.1.k_cache 0.50235566 12.97765567 + layer.1.v_cache 0.00000626 0.00323572 + layer.2.k_cache 0.02709456 1.41549318 + layer.2.v_cache 0.00002063 0.00902200 + layer.3.k_cache 0.00856834 9.38481664 + layer.3.v_cache 0.00001945 0.00989422 + layer.4.k_cache 0.00071483 0.34187664 + layer.4.v_cache 0.00005464 0.02045342 + layer.4.output 0.00491706 204.97532982 + ------------------------------------------------------------------------------------- + TOTAL 0.04301397 87.60086589 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 307148 +BPFP 1.0534 bits/point +EBPFP 2.1068 equivalent bits/point +MSE 87.600866 +---------------------- -------------------------------------------------------- +Time: 3.896s Load: 0.009s, Pack+Encode: 2.231s, Decode+Unpack: 1.656s +---------------------- -------------------------------------------------------- +💾 Converting with 87.6009 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,616B, BPFP=0.4168 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,836B, BPFP=2.0688 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,784B, BPFP=0.7424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,076B, BPFP=2.0209 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,372B, BPFP=0.9055 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,328B, BPFP=2.0998 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,600B, BPFP=0.8569 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,612B, BPFP=2.0547 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,480B, BPFP=1.7944 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,368B, BPFP=1.9763 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,056B, BPFP=0.2435 +⌛️ [2/4] FRONTEND: Frontend time: 2.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.573s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13184587 34.38630529 + layer.0.v_cache 0.00001706 0.00848110 + layer.1.k_cache 0.45546043 12.37466923 + layer.1.v_cache 0.00000678 0.00342975 + layer.2.k_cache 0.02272066 1.61125749 + layer.2.v_cache 0.00001996 0.00958150 + layer.3.k_cache 0.01643563 9.54401029 + layer.3.v_cache 0.00002069 0.01102250 + layer.4.k_cache 0.00071781 0.37918208 + layer.4.v_cache 0.00005262 0.01979337 + layer.4.output 1.23455937 217.68890409 + ------------------------------------------------------------------------------------- + TOTAL 0.54524783 93.06882713 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 264128 +BPFP 0.9789 bits/point +EBPFP 1.9578 equivalent bits/point +MSE 93.068827 +---------------------- -------------------------------------------------------- +Time: 3.836s Load: 0.010s, Pack+Encode: 2.253s, Decode+Unpack: 1.573s +---------------------- -------------------------------------------------------- +💾 Converting with 93.0688 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,188B, BPFP=0.4254 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,964B, BPFP=2.4837 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,200B, BPFP=0.8404 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,672B, BPFP=2.2888 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,692B, BPFP=0.9879 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,492B, BPFP=2.2782 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,680B, BPFP=0.9280 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,784B, BPFP=2.5322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,412B, BPFP=1.9775 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,080B, BPFP=2.2538 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,884B, BPFP=0.2019 +⌛️ [2/4] FRONTEND: Frontend time: 2.424s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.852s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17705467 31.42129794 + layer.0.v_cache 0.00001633 0.00779672 + layer.1.k_cache 0.50406676 12.89870753 + layer.1.v_cache 0.00000599 0.00316685 + layer.2.k_cache 0.00944760 1.61236087 + layer.2.v_cache 0.00001883 0.00929145 + layer.3.k_cache 0.04120483 9.32930409 + layer.3.v_cache 0.00002392 0.01079981 + layer.4.k_cache 0.00070100 0.35283707 + layer.4.v_cache 0.00005002 0.01966415 + layer.4.output 0.00499618 209.61106602 + ------------------------------------------------------------------------------------- + TOTAL 0.04515078 89.58486404 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 311048 +BPFP 1.0829 bits/point +EBPFP 2.1658 equivalent bits/point +MSE 89.584864 +---------------------- -------------------------------------------------------- +Time: 4.284s Load: 0.009s, Pack+Encode: 2.424s, Decode+Unpack: 1.852s +---------------------- -------------------------------------------------------- +💾 Converting with 89.5849 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,332B, BPFP=0.4243 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,016B, BPFP=2.3736 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,724B, BPFP=0.7942 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,012B, BPFP=2.2576 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,816B, BPFP=0.9731 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,628B, BPFP=2.2354 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,712B, BPFP=0.9093 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,660B, BPFP=2.2951 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,920B, BPFP=1.9630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,148B, BPFP=2.2076 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,456B, BPFP=0.1939 +⌛️ [2/4] FRONTEND: Frontend time: 2.494s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.679s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20501094 31.20174334 + layer.0.v_cache 0.00001516 0.00776470 + layer.1.k_cache 0.57085758 12.38476834 + layer.1.v_cache 0.00000574 0.00316673 + layer.2.k_cache 0.01259345 1.63287150 + layer.2.v_cache 0.00002099 0.00949805 + layer.3.k_cache 0.01648921 9.69318215 + layer.3.v_cache 0.00001970 0.01044028 + layer.4.k_cache 0.00070444 0.32834820 + layer.4.v_cache 0.00005121 0.01933527 + layer.4.output 0.00487421 204.34750331 + ------------------------------------------------------------------------------------- + TOTAL 0.04940517 87.39550834 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 307424 +BPFP 1.0465 bits/point +EBPFP 2.0930 equivalent bits/point +MSE 87.395508 +---------------------- -------------------------------------------------------- +Time: 4.185s Load: 0.012s, Pack+Encode: 2.494s, Decode+Unpack: 1.679s +---------------------- -------------------------------------------------------- +💾 Converting with 87.3955 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,424B, BPFP=0.4280 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,556B, BPFP=2.4536 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,428B, BPFP=0.7742 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,832B, BPFP=2.2966 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,956B, BPFP=0.9776 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,360B, BPFP=2.2694 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,996B, BPFP=0.9223 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,332B, BPFP=2.3254 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,692B, BPFP=1.9426 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,784B, BPFP=2.2362 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,152B, BPFP=0.2236 +⌛️ [2/4] FRONTEND: Frontend time: 2.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.035s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14466163 29.86592265 + layer.0.v_cache 0.00001631 0.00801712 + layer.1.k_cache 0.53789613 12.61137317 + layer.1.v_cache 0.00000653 0.00327091 + layer.2.k_cache 0.02095415 1.53184492 + layer.2.v_cache 0.00002169 0.00955747 + layer.3.k_cache 0.03083320 8.80454966 + layer.3.v_cache 0.00002083 0.01091664 + layer.4.k_cache 0.00069737 0.33516949 + layer.4.v_cache 0.00004943 0.01958951 + layer.4.output 0.00490546 203.16120849 + ------------------------------------------------------------------------------------- + TOTAL 0.04526444 86.78403947 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 315512 +BPFP 1.0701 bits/point +EBPFP 2.1402 equivalent bits/point +MSE 86.784039 +---------------------- -------------------------------------------------------- +Time: 4.267s Load: 0.011s, Pack+Encode: 2.221s, Decode+Unpack: 2.035s +---------------------- -------------------------------------------------------- +💾 Converting with 86.7840 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,040B, BPFP=0.4202 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,648B, BPFP=2.1242 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,612B, BPFP=0.7636 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,768B, BPFP=2.0782 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,844B, BPFP=0.8802 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,348B, BPFP=2.0562 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,948B, BPFP=0.8334 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,164B, BPFP=2.0989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,804B, BPFP=1.7665 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,644B, BPFP=2.0194 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,976B, BPFP=0.2014 +⌛️ [2/4] FRONTEND: Frontend time: 3.079s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.186s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14240686 41.60783014 + layer.0.v_cache 0.00001673 0.00763202 + layer.1.k_cache 0.61438330 12.84426114 + layer.1.v_cache 0.00000638 0.00313034 + layer.2.k_cache 0.02027981 1.56363823 + layer.2.v_cache 0.00002012 0.00999365 + layer.3.k_cache 0.02199794 9.83047006 + layer.3.v_cache 0.00002003 0.01056289 + layer.4.k_cache 0.00078937 0.34794198 + layer.4.v_cache 0.00005038 0.02089897 + layer.4.output 0.04464151 180.29379181 + ------------------------------------------------------------------------------------- + TOTAL 0.06543891 78.13546483 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 314796 +BPFP 0.9677 bits/point +EBPFP 1.9353 equivalent bits/point +MSE 78.135465 +---------------------- -------------------------------------------------------- +Time: 5.277s Load: 0.011s, Pack+Encode: 3.079s, Decode+Unpack: 2.186s +---------------------- -------------------------------------------------------- +💾 Converting with 78.1355 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 399, 128) +Output shape: (1, 399, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.output: torch.Size([1, 399, 3584]) -> torch.Size([1, 1, 399, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,700B, BPFP=0.4190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,888B, BPFP=2.2669 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,900B, BPFP=0.7793 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,424B, BPFP=2.1704 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,032B, BPFP=0.9019 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,964B, BPFP=2.1524 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,808B, BPFP=0.8540 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,740B, BPFP=2.2611 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,640B, BPFP=1.8264 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,536B, BPFP=2.1357 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,868B, BPFP=0.2007 +⌛️ [2/4] FRONTEND: Frontend time: 3.340s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.097s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17889012 32.73452675 + layer.0.v_cache 0.00001635 0.00744857 + layer.1.k_cache 0.94984295 12.87889646 + layer.1.v_cache 0.00000590 0.00301234 + layer.2.k_cache 0.02305104 1.51011134 + layer.2.v_cache 0.00002114 0.00973939 + layer.3.k_cache 0.01620252 9.59001324 + layer.3.v_cache 0.00002049 0.01045844 + layer.4.k_cache 0.00073175 0.34057571 + layer.4.v_cache 0.00005255 0.01982869 + layer.4.output 0.00656733 138.21538221 + ------------------------------------------------------------------------------------- + TOTAL 0.07145918 60.27131096 + (elements=3,472,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3472896 +Total Bytes 438500 +BPFP 1.0101 bits/point +EBPFP 2.0202 equivalent bits/point +MSE 60.271311 +---------------------- -------------------------------------------------------- +Time: 5.452s Load: 0.014s, Pack+Encode: 3.340s, Decode+Unpack: 2.097s +---------------------- -------------------------------------------------------- +💾 Converting with 60.2713 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,060B, BPFP=0.4190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,752B, BPFP=2.2644 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,920B, BPFP=0.8241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,944B, BPFP=2.2085 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,340B, BPFP=0.9914 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,940B, BPFP=2.2082 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,416B, BPFP=0.9275 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,344B, BPFP=2.2362 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,740B, BPFP=1.9870 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,252B, BPFP=2.1607 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,144B, BPFP=0.2088 +⌛️ [2/4] FRONTEND: Frontend time: 2.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.510s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10670873 33.12817599 + layer.0.v_cache 0.00001672 0.00889256 + layer.1.k_cache 0.29393762 12.44367805 + layer.1.v_cache 0.00000627 0.00347389 + layer.2.k_cache 0.01243114 1.60975647 + layer.2.v_cache 0.00002105 0.01051186 + layer.3.k_cache 0.00919804 9.91070287 + layer.3.v_cache 0.00002002 0.01226316 + layer.4.k_cache 0.00070102 0.33696565 + layer.4.v_cache 0.00005024 0.02142699 + layer.4.output 1.35463108 238.58148309 + ------------------------------------------------------------------------------------- + TOTAL 0.58267697 101.62095465 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 255852 +BPFP 1.0405 bits/point +EBPFP 2.0810 equivalent bits/point +MSE 101.620955 +---------------------- -------------------------------------------------------- +Time: 3.684s Load: 0.012s, Pack+Encode: 2.162s, Decode+Unpack: 1.510s +---------------------- -------------------------------------------------------- +💾 Converting with 101.6210 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,736B, BPFP=0.4144 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,864B, BPFP=2.0217 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,668B, BPFP=0.7178 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,004B, BPFP=1.9688 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,220B, BPFP=0.8748 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,820B, BPFP=2.0189 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,384B, BPFP=0.8233 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,304B, BPFP=2.0487 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,156B, BPFP=1.9781 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,380B, BPFP=1.9304 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,264B, BPFP=0.2132 +⌛️ [2/4] FRONTEND: Frontend time: 2.067s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.533s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633087 35.79264579 + layer.0.v_cache 0.00001891 0.00819388 + layer.1.k_cache 0.36012592 11.13310662 + layer.1.v_cache 0.00000581 0.00318476 + layer.2.k_cache 0.01407751 1.70038947 + layer.2.v_cache 0.00002023 0.01042849 + layer.3.k_cache 0.01230825 10.20880079 + layer.3.v_cache 0.00002000 0.01103253 + layer.4.k_cache 0.00071033 0.36419801 + layer.4.v_cache 0.00004856 0.02058604 + layer.4.output 1.20534739 212.22156215 + ------------------------------------------------------------------------------------- + TOTAL 0.52535871 90.87079420 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 264800 +BPFP 0.9582 bits/point +EBPFP 1.9164 equivalent bits/point +MSE 90.870794 +---------------------- -------------------------------------------------------- +Time: 3.608s Load: 0.008s, Pack+Encode: 2.067s, Decode+Unpack: 1.533s +---------------------- -------------------------------------------------------- +💾 Converting with 90.8708 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,516B, BPFP=0.4225 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,476B, BPFP=2.1055 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,048B, BPFP=0.7811 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,720B, BPFP=2.0565 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,512B, BPFP=0.9409 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,584B, BPFP=2.0477 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,424B, BPFP=0.8703 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,100B, BPFP=2.0812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,932B, BPFP=2.0054 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,372B, BPFP=2.0340 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,332B, BPFP=0.2068 +⌛️ [2/4] FRONTEND: Frontend time: 1.996s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.544s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11142051 33.65993946 + layer.0.v_cache 0.00001650 0.00808903 + layer.1.k_cache 0.42032433 10.37805024 + layer.1.v_cache 0.00000619 0.00327907 + layer.2.k_cache 0.01443043 1.62795577 + layer.2.v_cache 0.00002030 0.00997959 + layer.3.k_cache 0.02895402 9.98018814 + layer.3.v_cache 0.00001914 0.01060753 + layer.4.k_cache 0.00067703 0.34677371 + layer.4.v_cache 0.00005693 0.02116288 + layer.4.output 1.27033357 222.46556387 + ------------------------------------------------------------------------------------- + TOTAL 0.55695649 94.90029250 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 259016 +BPFP 0.9878 bits/point +EBPFP 1.9757 equivalent bits/point +MSE 94.900293 +---------------------- -------------------------------------------------------- +Time: 3.555s Load: 0.015s, Pack+Encode: 1.996s, Decode+Unpack: 1.544s +---------------------- -------------------------------------------------------- +💾 Converting with 94.9003 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 302, 128) +Output shape: (1, 302, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.output: torch.Size([1, 302, 3584]) -> torch.Size([1, 1, 302, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,144B, BPFP=0.4214 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,264B, BPFP=2.0832 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,544B, BPFP=0.7525 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,356B, BPFP=2.0362 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,692B, BPFP=0.9154 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,316B, BPFP=2.0341 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,420B, BPFP=0.8495 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,488B, BPFP=2.0948 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,060B, BPFP=1.8657 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,664B, BPFP=2.0004 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,976B, BPFP=0.1550 +⌛️ [2/4] FRONTEND: Frontend time: 2.844s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.787s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15012774 35.31959463 + layer.0.v_cache 0.00001504 0.00763696 + layer.1.k_cache 0.60100308 9.47601359 + layer.1.v_cache 0.00000575 0.00310996 + layer.2.k_cache 0.00814273 1.62496130 + layer.2.v_cache 0.00002032 0.01029691 + layer.3.k_cache 0.01797006 9.79052007 + layer.3.v_cache 0.00001899 0.01064196 + layer.4.k_cache 0.00070977 0.33139528 + layer.4.v_cache 0.00004868 0.02017560 + layer.4.output 0.04414053 178.47315516 + ------------------------------------------------------------------------------------- + TOTAL 0.06394387 76.81802543 + (elements=2,628,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2628608 +Total Bytes 311924 +BPFP 0.9493 bits/point +EBPFP 1.8986 equivalent bits/point +MSE 76.818025 +---------------------- -------------------------------------------------------- +Time: 4.646s Load: 0.015s, Pack+Encode: 2.844s, Decode+Unpack: 1.787s +---------------------- -------------------------------------------------------- +💾 Converting with 76.8180 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,324B, BPFP=0.4270 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,024B, BPFP=2.3335 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,244B, BPFP=0.8305 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,092B, BPFP=2.2792 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,980B, BPFP=0.9900 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,408B, BPFP=2.2393 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,244B, BPFP=0.8888 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,588B, BPFP=2.3081 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,676B, BPFP=1.8468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,364B, BPFP=2.2367 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,908B, BPFP=0.2324 +⌛️ [2/4] FRONTEND: Frontend time: 2.332s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.778s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14350806 28.92262018 + layer.0.v_cache 0.00001655 0.00777100 + layer.1.k_cache 0.46812200 13.06023259 + layer.1.v_cache 0.00000598 0.00314894 + layer.2.k_cache 0.01096032 1.57082333 + layer.2.v_cache 0.00002019 0.00949313 + layer.3.k_cache 0.00658789 9.63876115 + layer.3.v_cache 0.00002014 0.01038506 + layer.4.k_cache 0.00067528 0.34935393 + layer.4.v_cache 0.00005785 0.01955421 + layer.4.output 0.00496594 205.22962753 + ------------------------------------------------------------------------------------- + TOTAL 0.03910211 87.65879625 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 308852 +BPFP 1.0592 bits/point +EBPFP 2.1184 equivalent bits/point +MSE 87.658796 +---------------------- -------------------------------------------------------- +Time: 4.121s Load: 0.010s, Pack+Encode: 2.332s, Decode+Unpack: 1.778s +---------------------- -------------------------------------------------------- +💾 Converting with 87.6588 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 260, 128) +Output shape: (1, 260, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.output: torch.Size([1, 260, 3584]) -> torch.Size([1, 1, 260, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,032B, BPFP=0.4226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,336B, BPFP=2.4240 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,712B, BPFP=0.8240 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,648B, BPFP=2.3226 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,064B, BPFP=0.9654 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,012B, BPFP=2.2844 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,016B, BPFP=0.9024 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,680B, BPFP=2.3846 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,636B, BPFP=2.0214 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,532B, BPFP=2.2555 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,816B, BPFP=0.1873 +⌛️ [2/4] FRONTEND: Frontend time: 2.389s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.614s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14548651 31.70550631 + layer.0.v_cache 0.00001584 0.00803877 + layer.1.k_cache 0.47062912 12.16021071 + layer.1.v_cache 0.00000604 0.00330542 + layer.2.k_cache 0.00619981 1.57419445 + layer.2.v_cache 0.00002003 0.01033812 + layer.3.k_cache 0.01299995 7.99045598 + layer.3.v_cache 0.00001981 0.01078027 + layer.4.k_cache 0.00067777 0.33927882 + layer.4.v_cache 0.00006192 0.02085966 + layer.4.output 0.00504520 212.09043613 + ------------------------------------------------------------------------------------- + TOTAL 0.03949607 90.49741302 + (elements=2,263,040) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2263040 +Total Bytes 301484 +BPFP 1.0658 bits/point +EBPFP 2.1315 equivalent bits/point +MSE 90.497413 +---------------------- -------------------------------------------------------- +Time: 4.013s Load: 0.010s, Pack+Encode: 2.389s, Decode+Unpack: 1.614s +---------------------- -------------------------------------------------------- +💾 Converting with 90.4974 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,464B, BPFP=0.4288 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,000B, BPFP=2.3552 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,236B, BPFP=0.7603 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,620B, BPFP=2.2760 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,252B, BPFP=0.9910 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,348B, BPFP=2.2603 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,600B, BPFP=0.8961 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,372B, BPFP=2.3766 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,992B, BPFP=1.7803 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,840B, BPFP=2.2312 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,984B, BPFP=0.2050 +⌛️ [2/4] FRONTEND: Frontend time: 2.276s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.626s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14555165 30.28383502 + layer.0.v_cache 0.00001661 0.00804179 + layer.1.k_cache 0.55468144 9.75846234 + layer.1.v_cache 0.00000621 0.00311502 + layer.2.k_cache 0.02243149 1.67934608 + layer.2.v_cache 0.00002103 0.00945850 + layer.3.k_cache 0.02028967 9.08873704 + layer.3.v_cache 0.00002036 0.01021633 + layer.4.k_cache 0.00072975 0.34885794 + layer.4.v_cache 0.00005124 0.01894228 + layer.4.output 0.00486695 202.04813879 + ------------------------------------------------------------------------------------- + TOTAL 0.04575695 86.20858728 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 309708 +BPFP 1.0465 bits/point +EBPFP 2.0931 equivalent bits/point +MSE 86.208587 +---------------------- -------------------------------------------------------- +Time: 3.913s Load: 0.011s, Pack+Encode: 2.276s, Decode+Unpack: 1.626s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2086 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 356, 128) +Output shape: (1, 356, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.output: torch.Size([1, 356, 3584]) -> torch.Size([1, 1, 356, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,492B, BPFP=0.4166 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,116B, BPFP=2.1557 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,824B, BPFP=0.7384 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,196B, BPFP=2.1153 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,168B, BPFP=0.8852 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,788B, BPFP=2.0974 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,220B, BPFP=0.8436 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,460B, BPFP=2.1269 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,568B, BPFP=1.6928 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,660B, BPFP=2.0479 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,504B, BPFP=0.2101 +⌛️ [2/4] FRONTEND: Frontend time: 2.371s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.991s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15223843 35.19062994 + layer.0.v_cache 0.00001740 0.00764671 + layer.1.k_cache 0.82972649 11.92095330 + layer.1.v_cache 0.00000631 0.00306673 + layer.2.k_cache 0.03086275 1.54922691 + layer.2.v_cache 0.00001960 0.00967120 + layer.3.k_cache 0.01100341 9.97549473 + layer.3.v_cache 0.00002062 0.01001496 + layer.4.k_cache 0.00073857 0.36082090 + layer.4.v_cache 0.00005279 0.01867742 + layer.4.output 0.03758925 151.64694522 + ------------------------------------------------------------------------------------- + TOTAL 0.07575360 65.91616585 + (elements=3,098,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3098624 +Total Bytes 377996 +BPFP 0.9759 bits/point +EBPFP 1.9518 equivalent bits/point +MSE 65.916166 +---------------------- -------------------------------------------------------- +Time: 4.377s Load: 0.015s, Pack+Encode: 2.371s, Decode+Unpack: 1.991s +---------------------- -------------------------------------------------------- +💾 Converting with 65.9162 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,260B, BPFP=0.4265 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,640B, BPFP=2.3872 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,312B, BPFP=0.8407 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,172B, BPFP=2.3597 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,008B, BPFP=0.9991 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,808B, BPFP=2.2796 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,484B, BPFP=0.9095 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,624B, BPFP=2.3275 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,736B, BPFP=1.8642 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,264B, BPFP=2.2477 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,296B, BPFP=0.2374 +⌛️ [2/4] FRONTEND: Frontend time: 2.329s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.687s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12347319 31.10826627 + layer.0.v_cache 0.00001696 0.00802382 + layer.1.k_cache 0.44284580 12.67872287 + layer.1.v_cache 0.00000660 0.00326939 + layer.2.k_cache 0.00882101 1.58413398 + layer.2.v_cache 0.00002210 0.01015255 + layer.3.k_cache 0.01145097 9.51322673 + layer.3.v_cache 0.00001984 0.01071885 + layer.4.k_cache 0.00068754 0.33554791 + layer.4.v_cache 0.00006483 0.02030929 + layer.4.output 0.00500964 207.46635003 + ------------------------------------------------------------------------------------- + TOTAL 0.03661626 88.67863658 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 311604 +BPFP 1.0767 bits/point +EBPFP 2.1534 equivalent bits/point +MSE 88.678637 +---------------------- -------------------------------------------------------- +Time: 4.028s Load: 0.012s, Pack+Encode: 2.329s, Decode+Unpack: 1.687s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6786 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,528B, BPFP=0.4127 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,984B, BPFP=2.2469 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,876B, BPFP=0.7607 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,896B, BPFP=2.1873 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,968B, BPFP=0.9303 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,640B, BPFP=2.1732 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,260B, BPFP=0.8914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,492B, BPFP=2.2200 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,536B, BPFP=2.0031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,224B, BPFP=2.1504 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,392B, BPFP=0.1989 +⌛️ [2/4] FRONTEND: Frontend time: 2.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.760s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13002123 32.05006853 + layer.0.v_cache 0.00001833 0.00849301 + layer.1.k_cache 0.50397591 12.49987407 + layer.1.v_cache 0.00000608 0.00325744 + layer.2.k_cache 0.01926375 1.62784734 + layer.2.v_cache 0.00002060 0.00971621 + layer.3.k_cache 0.04992905 9.97331243 + layer.3.v_cache 0.00002096 0.01151944 + layer.4.k_cache 0.00069540 0.33413723 + layer.4.v_cache 0.00005467 0.01999996 + layer.4.output 0.00466491 190.11578947 + ------------------------------------------------------------------------------------- + TOTAL 0.04333296 81.60875012 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 316796 +BPFP 1.0217 bits/point +EBPFP 2.0433 equivalent bits/point +MSE 81.608750 +---------------------- -------------------------------------------------------- +Time: 4.022s Load: 0.011s, Pack+Encode: 2.252s, Decode+Unpack: 1.760s +---------------------- -------------------------------------------------------- +💾 Converting with 81.6088 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,508B, BPFP=0.4190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,216B, BPFP=2.3000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,200B, BPFP=0.7924 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,272B, BPFP=2.2473 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,260B, BPFP=0.9632 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,200B, BPFP=2.2433 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,784B, BPFP=0.8808 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,924B, BPFP=2.2837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,036B, BPFP=1.7877 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,396B, BPFP=2.1984 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,108B, BPFP=0.2560 +⌛️ [2/4] FRONTEND: Frontend time: 2.320s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.705s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15974760 29.58416225 + layer.0.v_cache 0.00001739 0.00766170 + layer.1.k_cache 0.45644542 12.39776001 + layer.1.v_cache 0.00000643 0.00303520 + layer.2.k_cache 0.01631587 1.64883183 + layer.2.v_cache 0.00002005 0.00981810 + layer.3.k_cache 0.01511352 9.48787231 + layer.3.v_cache 0.00002063 0.01057295 + layer.4.k_cache 0.00073517 0.36059410 + layer.4.v_cache 0.00005091 0.02042114 + layer.4.output 0.00481773 197.48883929 + ------------------------------------------------------------------------------------- + TOTAL 0.04012924 84.46780027 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 320904 +BPFP 1.0534 bits/point +EBPFP 2.1068 equivalent bits/point +MSE 84.467800 +---------------------- -------------------------------------------------------- +Time: 4.035s Load: 0.011s, Pack+Encode: 2.320s, Decode+Unpack: 1.705s +---------------------- -------------------------------------------------------- +💾 Converting with 84.4678 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,284B, BPFP=0.4263 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,532B, BPFP=2.3720 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,204B, BPFP=0.8312 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,276B, BPFP=2.2985 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,260B, BPFP=0.9515 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,588B, BPFP=2.2582 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,836B, BPFP=0.8682 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,420B, BPFP=2.3069 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,268B, BPFP=2.0639 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,228B, BPFP=2.2371 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,020B, BPFP=0.1841 +⌛️ [2/4] FRONTEND: Frontend time: 2.424s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.638s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14148096 29.52456768 + layer.0.v_cache 0.00001588 0.00772054 + layer.1.k_cache 0.55266568 12.53426655 + layer.1.v_cache 0.00000632 0.00319987 + layer.2.k_cache 0.00687984 1.53232383 + layer.2.v_cache 0.00001885 0.00975986 + layer.3.k_cache 0.01675770 7.79355853 + layer.3.v_cache 0.00001964 0.01052167 + layer.4.k_cache 0.00068991 0.32819209 + layer.4.v_cache 0.00004859 0.01914155 + layer.4.output 0.00492290 206.65248796 + ------------------------------------------------------------------------------------- + TOTAL 0.04429669 88.13709811 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 305916 +BPFP 1.0531 bits/point +EBPFP 2.1062 equivalent bits/point +MSE 88.137098 +---------------------- -------------------------------------------------------- +Time: 4.073s Load: 0.012s, Pack+Encode: 2.424s, Decode+Unpack: 1.638s +---------------------- -------------------------------------------------------- +💾 Converting with 88.1371 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,620B, BPFP=0.4170 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,048B, BPFP=2.3727 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,112B, BPFP=0.7794 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,924B, BPFP=2.2699 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,988B, BPFP=0.9669 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,016B, BPFP=2.2260 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,172B, BPFP=0.8791 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,228B, BPFP=2.2846 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,848B, BPFP=1.7825 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,320B, BPFP=2.1923 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,132B, BPFP=0.1944 +⌛️ [2/4] FRONTEND: Frontend time: 2.415s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.173s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13144862 33.17240712 + layer.0.v_cache 0.00001769 0.00803256 + layer.1.k_cache 0.64513603 12.69783251 + layer.1.v_cache 0.00000607 0.00311412 + layer.2.k_cache 0.02726801 1.61794524 + layer.2.v_cache 0.00001956 0.00956340 + layer.3.k_cache 0.02271994 9.09102061 + layer.3.v_cache 0.00002018 0.01033373 + layer.4.k_cache 0.00073057 0.34454556 + layer.4.v_cache 0.00004940 0.01889062 + layer.4.output 0.04135688 166.57180175 + ------------------------------------------------------------------------------------- + TOTAL 0.06570084 71.93978222 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 362408 +BPFP 1.0313 bits/point +EBPFP 2.0625 equivalent bits/point +MSE 71.939782 +---------------------- -------------------------------------------------------- +Time: 4.599s Load: 0.011s, Pack+Encode: 2.415s, Decode+Unpack: 2.173s +---------------------- -------------------------------------------------------- +💾 Converting with 71.9398 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,852B, BPFP=0.4253 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,112B, BPFP=2.4064 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,000B, BPFP=0.7994 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,868B, BPFP=2.3160 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,200B, BPFP=1.0320 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,716B, BPFP=2.3049 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,656B, BPFP=0.9198 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,596B, BPFP=2.3689 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,716B, BPFP=2.0869 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,472B, BPFP=2.2872 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,128B, BPFP=0.1778 +⌛️ [2/4] FRONTEND: Frontend time: 2.170s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.595s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12396691 29.52554960 + layer.0.v_cache 0.00001598 0.00832225 + layer.1.k_cache 0.24854917 11.91295308 + layer.1.v_cache 0.00000584 0.00323536 + layer.2.k_cache 0.02290044 1.63894994 + layer.2.v_cache 0.00002299 0.00986452 + layer.3.k_cache 0.01294550 9.34347861 + layer.3.v_cache 0.00002077 0.01123901 + layer.4.k_cache 0.00065144 0.33439409 + layer.4.v_cache 0.00005054 0.01935117 + layer.4.output 1.42385976 250.18658638 + ------------------------------------------------------------------------------------- + TOTAL 0.61036164 106.12432013 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 250316 +BPFP 1.0701 bits/point +EBPFP 2.1402 equivalent bits/point +MSE 106.124320 +---------------------- -------------------------------------------------------- +Time: 3.773s Load: 0.007s, Pack+Encode: 2.170s, Decode+Unpack: 1.595s +---------------------- -------------------------------------------------------- +💾 Converting with 106.1243 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,540B, BPFP=0.4077 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,244B, BPFP=2.2299 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,636B, BPFP=0.7913 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,844B, BPFP=2.1542 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,380B, BPFP=0.9397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,824B, BPFP=2.1531 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,256B, BPFP=0.8789 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,988B, BPFP=2.2701 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,712B, BPFP=1.8227 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,992B, BPFP=2.1081 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,696B, BPFP=0.1907 +⌛️ [2/4] FRONTEND: Frontend time: 2.510s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.040s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14783513 33.93106618 + layer.0.v_cache 0.00001600 0.00828277 + layer.1.k_cache 0.54442752 11.20118032 + layer.1.v_cache 0.00000628 0.00327502 + layer.2.k_cache 0.01848349 1.59581073 + layer.2.v_cache 0.00002061 0.00985948 + layer.3.k_cache 0.01777397 9.83508427 + layer.3.v_cache 0.00002092 0.01142622 + layer.4.k_cache 0.00070447 0.32398322 + layer.4.v_cache 0.00005049 0.01990953 + layer.4.output 0.04615090 186.68172269 + ------------------------------------------------------------------------------------- + TOTAL 0.06190560 80.21834921 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 316112 +BPFP 1.0053 bits/point +EBPFP 2.0107 equivalent bits/point +MSE 80.218349 +---------------------- -------------------------------------------------------- +Time: 4.564s Load: 0.013s, Pack+Encode: 2.510s, Decode+Unpack: 2.040s +---------------------- -------------------------------------------------------- +💾 Converting with 80.2183 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,988B, BPFP=0.4188 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,380B, BPFP=2.1697 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,180B, BPFP=0.7435 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,812B, BPFP=2.0875 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,660B, BPFP=0.9260 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,668B, BPFP=2.0799 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,480B, BPFP=0.8641 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,608B, BPFP=2.1292 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,312B, BPFP=1.7991 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,948B, BPFP=2.0422 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,388B, BPFP=0.1902 +⌛️ [2/4] FRONTEND: Frontend time: 2.435s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.749s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16274841 39.10137898 + layer.0.v_cache 0.00001794 0.00823811 + layer.1.k_cache 0.56737488 12.13657620 + layer.1.v_cache 0.00000586 0.00309978 + layer.2.k_cache 0.02027899 1.58490318 + layer.2.v_cache 0.00002117 0.00963554 + layer.3.k_cache 0.01891868 9.98003946 + layer.3.v_cache 0.00002140 0.01040854 + layer.4.k_cache 0.00069658 0.34014181 + layer.4.v_cache 0.00005005 0.01904823 + layer.4.output 0.04476916 181.39549377 + ------------------------------------------------------------------------------------- + TOTAL 0.06373636 78.40952507 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 316424 +BPFP 0.9759 bits/point +EBPFP 1.9519 equivalent bits/point +MSE 78.409525 +---------------------- -------------------------------------------------------- +Time: 4.194s Load: 0.010s, Pack+Encode: 2.435s, Decode+Unpack: 1.749s +---------------------- -------------------------------------------------------- +💾 Converting with 78.4095 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,452B, BPFP=0.4265 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,676B, BPFP=2.3853 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,236B, BPFP=0.7576 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,464B, BPFP=2.2587 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,192B, BPFP=0.9840 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,160B, BPFP=2.2413 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,444B, BPFP=0.8839 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,084B, BPFP=2.2942 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,696B, BPFP=1.8713 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,564B, BPFP=2.2072 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,464B, BPFP=0.2000 +⌛️ [2/4] FRONTEND: Frontend time: 2.275s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.624s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13366390 30.42430746 + layer.0.v_cache 0.00001719 0.00830535 + layer.1.k_cache 0.57618574 11.37271509 + layer.1.v_cache 0.00000658 0.00322662 + layer.2.k_cache 0.02858028 1.63480229 + layer.2.v_cache 0.00002021 0.00970849 + layer.3.k_cache 0.00873641 7.82808118 + layer.3.v_cache 0.00002023 0.01090694 + layer.4.k_cache 0.00067337 0.33754800 + layer.4.v_cache 0.00005011 0.01975880 + layer.4.output 0.00484872 200.45893838 + ------------------------------------------------------------------------------------- + TOTAL 0.04599383 85.58011347 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 309432 +BPFP 1.0418 bits/point +EBPFP 2.0835 equivalent bits/point +MSE 85.580113 +---------------------- -------------------------------------------------------- +Time: 3.908s Load: 0.009s, Pack+Encode: 2.275s, Decode+Unpack: 1.624s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,000B, BPFP=0.4242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,716B, BPFP=2.3838 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,536B, BPFP=0.8156 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,132B, BPFP=2.2718 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,040B, BPFP=0.9926 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,916B, BPFP=2.2565 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,104B, BPFP=0.9265 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,808B, BPFP=2.3196 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,720B, BPFP=2.1012 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,544B, BPFP=2.2302 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,728B, BPFP=0.2397 +⌛️ [2/4] FRONTEND: Frontend time: 2.663s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.537s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13057878 31.29778970 + layer.0.v_cache 0.00001797 0.00908196 + layer.1.k_cache 0.31658528 12.58644346 + layer.1.v_cache 0.00000647 0.00341936 + layer.2.k_cache 0.02849985 1.70060530 + layer.2.v_cache 0.00001991 0.00983164 + layer.3.k_cache 0.03599559 10.02382857 + layer.3.v_cache 0.00002239 0.01195785 + layer.4.k_cache 0.00078811 0.34378856 + layer.4.v_cache 0.00005540 0.02058157 + layer.4.output 1.38530015 243.18117728 + ------------------------------------------------------------------------------------- + TOTAL 0.60056887 103.42797464 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 260244 +BPFP 1.0823 bits/point +EBPFP 2.1647 equivalent bits/point +MSE 103.427975 +---------------------- -------------------------------------------------------- +Time: 4.214s Load: 0.013s, Pack+Encode: 2.663s, Decode+Unpack: 1.537s +---------------------- -------------------------------------------------------- +💾 Converting with 103.4280 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,904B, BPFP=0.4251 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,256B, BPFP=2.3946 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,392B, BPFP=0.8203 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,192B, BPFP=2.3180 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,368B, BPFP=1.0346 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,972B, BPFP=2.3021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,372B, BPFP=0.9628 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,848B, BPFP=2.3652 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,700B, BPFP=2.1385 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,660B, BPFP=2.2797 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,100B, BPFP=0.2273 +⌛️ [2/4] FRONTEND: Frontend time: 2.181s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.545s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12725546 29.70253996 + layer.0.v_cache 0.00001736 0.00901547 + layer.1.k_cache 0.32262681 12.69659958 + layer.1.v_cache 0.00000608 0.00343978 + layer.2.k_cache 0.02157613 1.65761598 + layer.2.v_cache 0.00002208 0.00996763 + layer.3.k_cache 0.01760322 9.77013541 + layer.3.v_cache 0.00002192 0.01205737 + layer.4.k_cache 0.00069785 0.33506209 + layer.4.v_cache 0.00005172 0.02200342 + layer.4.output 1.41081570 248.71502633 + ------------------------------------------------------------------------------------- + TOTAL 0.60974050 105.60138947 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 258764 +BPFP 1.0960 bits/point +EBPFP 2.1920 equivalent bits/point +MSE 105.601389 +---------------------- -------------------------------------------------------- +Time: 3.734s Load: 0.008s, Pack+Encode: 2.181s, Decode+Unpack: 1.545s +---------------------- -------------------------------------------------------- +💾 Converting with 105.6014 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,824B, BPFP=0.4272 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,416B, BPFP=2.5246 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,960B, BPFP=0.8040 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,312B, BPFP=2.3703 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,416B, BPFP=1.0575 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,032B, BPFP=2.3498 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,272B, BPFP=0.9736 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,828B, BPFP=2.4082 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,480B, BPFP=2.1626 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,660B, BPFP=2.3225 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,948B, BPFP=0.2300 +⌛️ [2/4] FRONTEND: Frontend time: 2.071s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.401s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13788442 29.90962671 + layer.0.v_cache 0.00001717 0.00927709 + layer.1.k_cache 0.30723858 12.20753558 + layer.1.v_cache 0.00000664 0.00362943 + layer.2.k_cache 0.01337219 1.69454755 + layer.2.v_cache 0.00002159 0.01024630 + layer.3.k_cache 0.03522867 9.64707352 + layer.3.v_cache 0.00002126 0.01190574 + layer.4.k_cache 0.00073009 0.35370973 + layer.4.v_cache 0.00005119 0.02123945 + layer.4.output 1.43730973 253.40580986 + ------------------------------------------------------------------------------------- + TOTAL 0.62092588 107.51232118 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 259148 +BPFP 1.1183 bits/point +EBPFP 2.2365 equivalent bits/point +MSE 107.512321 +---------------------- -------------------------------------------------------- +Time: 3.481s Load: 0.009s, Pack+Encode: 2.071s, Decode+Unpack: 1.401s +---------------------- -------------------------------------------------------- +💾 Converting with 107.5123 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,716B, BPFP=0.4378 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,612B, BPFP=2.4979 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,048B, BPFP=0.8462 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,416B, BPFP=2.4062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,836B, BPFP=1.0597 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,240B, BPFP=2.3928 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,164B, BPFP=1.0083 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,220B, BPFP=2.4678 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,276B, BPFP=1.9360 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,832B, BPFP=2.3615 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,444B, BPFP=0.2565 +⌛️ [2/4] FRONTEND: Frontend time: 2.114s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.561s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11546621 30.03176700 + layer.0.v_cache 0.00001823 0.00885882 + layer.1.k_cache 0.26924565 13.08351883 + layer.1.v_cache 0.00000737 0.00357537 + layer.2.k_cache 0.01761903 1.64042305 + layer.2.v_cache 0.00002145 0.01000407 + layer.3.k_cache 0.02374693 9.70679549 + layer.3.v_cache 0.00002104 0.01146817 + layer.4.k_cache 0.00067272 0.33527352 + layer.4.v_cache 0.00005153 0.01990890 + layer.4.output 1.50071327 263.80514706 + ------------------------------------------------------------------------------------- + TOTAL 0.64305077 111.85221310 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 250804 +BPFP 1.1300 bits/point +EBPFP 2.2600 equivalent bits/point +MSE 111.852213 +---------------------- -------------------------------------------------------- +Time: 3.686s Load: 0.010s, Pack+Encode: 2.114s, Decode+Unpack: 1.561s +---------------------- -------------------------------------------------------- +💾 Converting with 111.8522 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,124B, BPFP=0.4121 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,044B, BPFP=2.0822 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,600B, BPFP=0.7407 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,892B, BPFP=2.0237 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,436B, BPFP=0.8845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,704B, BPFP=2.0142 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,500B, BPFP=0.8371 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,648B, BPFP=2.2143 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,344B, BPFP=1.9959 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,284B, BPFP=1.9929 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,324B, BPFP=0.2053 +⌛️ [2/4] FRONTEND: Frontend time: 2.539s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.730s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18206066 32.05363484 + layer.0.v_cache 0.00001652 0.00820174 + layer.1.k_cache 0.62405802 12.65299849 + layer.1.v_cache 0.00000616 0.00322015 + layer.2.k_cache 0.02036299 1.58774834 + layer.2.v_cache 0.00002087 0.00977183 + layer.3.k_cache 0.02546444 10.04545217 + layer.3.v_cache 0.00002064 0.01058392 + layer.4.k_cache 0.00069880 0.35768038 + layer.4.v_cache 0.00007146 0.01959111 + layer.4.output 0.04336799 174.82411004 + ------------------------------------------------------------------------------------- + TOTAL 0.06802097 75.32456784 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 327900 +BPFP 0.9785 bits/point +EBPFP 1.9570 equivalent bits/point +MSE 75.324568 +---------------------- -------------------------------------------------------- +Time: 4.282s Load: 0.013s, Pack+Encode: 2.539s, Decode+Unpack: 1.730s +---------------------- -------------------------------------------------------- +💾 Converting with 75.3246 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,112B, BPFP=0.4170 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,756B, BPFP=2.2350 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,148B, BPFP=0.7606 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,068B, BPFP=2.1880 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,328B, BPFP=0.9776 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,740B, BPFP=2.1657 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,072B, BPFP=0.8919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,412B, BPFP=2.2115 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,244B, BPFP=1.9954 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,176B, BPFP=2.1272 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,820B, BPFP=0.2419 +⌛️ [2/4] FRONTEND: Frontend time: 2.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.608s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14018573 32.92068521 + layer.0.v_cache 0.00001653 0.00864598 + layer.1.k_cache 0.40671090 11.07030610 + layer.1.v_cache 0.00000662 0.00337426 + layer.2.k_cache 0.03036383 1.62049646 + layer.2.v_cache 0.00002088 0.00948870 + layer.3.k_cache 0.05940984 9.82892245 + layer.3.v_cache 0.00002025 0.01071884 + layer.4.k_cache 0.00071161 0.36473044 + layer.4.v_cache 0.00005094 0.01976338 + layer.4.output 1.33692960 235.43385449 + ------------------------------------------------------------------------------------- + TOTAL 0.58800025 100.22906549 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 258876 +BPFP 1.0390 bits/point +EBPFP 2.0781 equivalent bits/point +MSE 100.229065 +---------------------- -------------------------------------------------------- +Time: 3.839s Load: 0.011s, Pack+Encode: 2.221s, Decode+Unpack: 1.608s +---------------------- -------------------------------------------------------- +💾 Converting with 100.2291 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,944B, BPFP=0.4241 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,908B, BPFP=2.5619 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,932B, BPFP=0.7800 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,384B, BPFP=2.3105 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,392B, BPFP=1.0268 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,940B, BPFP=2.2788 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,848B, BPFP=0.9167 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,560B, BPFP=2.3231 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,504B, BPFP=2.1050 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,604B, BPFP=2.2549 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,200B, BPFP=0.2263 +⌛️ [2/4] FRONTEND: Frontend time: 2.426s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.577s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351684 31.33623627 + layer.0.v_cache 0.00001872 0.00912925 + layer.1.k_cache 0.28219827 12.15808496 + layer.1.v_cache 0.00000627 0.00352059 + layer.2.k_cache 0.01639774 1.64443161 + layer.2.v_cache 0.00002172 0.01092326 + layer.3.k_cache 0.02752649 9.92296081 + layer.3.v_cache 0.00002090 0.01201326 + layer.4.k_cache 0.00067264 0.35456820 + layer.4.v_cache 0.00005058 0.02103919 + layer.4.output 1.39794763 244.56941047 + ------------------------------------------------------------------------------------- + TOTAL 0.60212139 103.96816357 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 260216 +BPFP 1.0921 bits/point +EBPFP 2.1842 equivalent bits/point +MSE 103.968164 +---------------------- -------------------------------------------------------- +Time: 4.015s Load: 0.012s, Pack+Encode: 2.426s, Decode+Unpack: 1.577s +---------------------- -------------------------------------------------------- +💾 Converting with 103.9682 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,360B, BPFP=0.4291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,680B, BPFP=2.3717 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,672B, BPFP=0.8554 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,904B, BPFP=2.3265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,592B, BPFP=0.9674 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,016B, BPFP=2.2747 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,712B, BPFP=0.9160 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,116B, BPFP=2.3389 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,852B, BPFP=1.7987 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,236B, BPFP=2.2292 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,240B, BPFP=0.2352 +⌛️ [2/4] FRONTEND: Frontend time: 2.434s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.746s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15646637 29.35770063 + layer.0.v_cache 0.00001727 0.00806909 + layer.1.k_cache 0.53079696 13.03561993 + layer.1.v_cache 0.00000635 0.00328676 + layer.2.k_cache 0.02340859 1.64570082 + layer.2.v_cache 0.00002061 0.00958193 + layer.3.k_cache 0.01050839 9.47196573 + layer.3.v_cache 0.00002156 0.01018441 + layer.4.k_cache 0.00068596 0.34469488 + layer.4.v_cache 0.00006072 0.01981005 + layer.4.output 0.00497423 205.63696029 + ------------------------------------------------------------------------------------- + TOTAL 0.04451838 87.84501978 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 311380 +BPFP 1.0679 bits/point +EBPFP 2.1358 equivalent bits/point +MSE 87.845020 +---------------------- -------------------------------------------------------- +Time: 4.191s Load: 0.011s, Pack+Encode: 2.434s, Decode+Unpack: 1.746s +---------------------- -------------------------------------------------------- +💾 Converting with 87.8450 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,132B, BPFP=0.4125 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,816B, BPFP=2.0706 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,952B, BPFP=0.7585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,960B, BPFP=2.0272 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,776B, BPFP=0.9018 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,376B, BPFP=1.9976 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,548B, BPFP=0.8395 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,192B, BPFP=2.0390 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,780B, BPFP=1.8659 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,868B, BPFP=1.9718 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,768B, BPFP=0.2375 +⌛️ [2/4] FRONTEND: Frontend time: 2.320s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.759s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15292851 34.28132610 + layer.0.v_cache 0.00001614 0.00850128 + layer.1.k_cache 0.61736927 12.98275321 + layer.1.v_cache 0.00000669 0.00347882 + layer.2.k_cache 0.02003183 1.60201174 + layer.2.v_cache 0.00002344 0.01005076 + layer.3.k_cache 0.01028479 9.78130073 + layer.3.v_cache 0.00002212 0.01060160 + layer.4.k_cache 0.00070434 0.34607736 + layer.4.v_cache 0.00006490 0.01966528 + layer.4.output 0.04342050 174.40033627 + ------------------------------------------------------------------------------------- + TOTAL 0.06502327 75.28518358 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 326168 +BPFP 0.9733 bits/point +EBPFP 1.9467 equivalent bits/point +MSE 75.285184 +---------------------- -------------------------------------------------------- +Time: 4.091s Load: 0.012s, Pack+Encode: 2.320s, Decode+Unpack: 1.759s +---------------------- -------------------------------------------------------- +💾 Converting with 75.2852 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,324B, BPFP=0.4254 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,580B, BPFP=2.3571 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,572B, BPFP=0.8464 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,408B, BPFP=2.2890 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,088B, BPFP=0.9926 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,056B, BPFP=2.2686 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,620B, BPFP=0.9073 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,896B, BPFP=2.3174 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,348B, BPFP=1.9951 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,624B, BPFP=2.2435 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,692B, BPFP=0.1883 +⌛️ [2/4] FRONTEND: Frontend time: 2.410s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.897s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12950333 30.60412552 + layer.0.v_cache 0.00001690 0.00833432 + layer.1.k_cache 0.52417032 12.70507268 + layer.1.v_cache 0.00000602 0.00331010 + layer.2.k_cache 0.01578268 1.64091917 + layer.2.v_cache 0.00002268 0.00989496 + layer.3.k_cache 0.03345619 9.32507914 + layer.3.v_cache 0.00002007 0.01102228 + layer.4.k_cache 0.00067301 0.34947922 + layer.4.v_cache 0.00005077 0.02063562 + layer.4.output 0.00489311 205.29943906 + ------------------------------------------------------------------------------------- + TOTAL 0.04340905 87.75140861 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 309208 +BPFP 1.0565 bits/point +EBPFP 2.1130 equivalent bits/point +MSE 87.751409 +---------------------- -------------------------------------------------------- +Time: 4.316s Load: 0.009s, Pack+Encode: 2.410s, Decode+Unpack: 1.897s +---------------------- -------------------------------------------------------- +💾 Converting with 87.7514 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,008B, BPFP=0.4248 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,136B, BPFP=2.3428 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,632B, BPFP=0.8224 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,080B, BPFP=2.2681 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,700B, BPFP=0.9686 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,648B, BPFP=2.2376 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,612B, BPFP=0.8917 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,244B, BPFP=2.2797 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,772B, BPFP=1.8221 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,380B, BPFP=2.2186 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,572B, BPFP=0.1977 +⌛️ [2/4] FRONTEND: Frontend time: 2.211s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.528s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13888381 30.40093414 + layer.0.v_cache 0.00001731 0.00835537 + layer.1.k_cache 0.27925331 12.82604815 + layer.1.v_cache 0.00000589 0.00333495 + layer.2.k_cache 0.03172064 1.64041345 + layer.2.v_cache 0.00001980 0.01023397 + layer.3.k_cache 0.01738505 9.71988785 + layer.3.v_cache 0.00001961 0.01134570 + layer.4.k_cache 0.00068340 0.32827400 + layer.4.v_cache 0.00005063 0.02074977 + layer.4.output 1.38525280 242.92723820 + ------------------------------------------------------------------------------------- + TOTAL 0.59792994 103.26236734 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 249784 +BPFP 1.0388 bits/point +EBPFP 2.0777 equivalent bits/point +MSE 103.262367 +---------------------- -------------------------------------------------------- +Time: 3.747s Load: 0.008s, Pack+Encode: 2.211s, Decode+Unpack: 1.528s +---------------------- -------------------------------------------------------- +💾 Converting with 103.2624 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.0325 bits/point +Avg EBPFP 2.0649 equivalent bits/point +Avg MSE 88.396566 +Avg Time 4.074s +------------------------ ---------------------------- diff --git a/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..eb3a47a265f3c535f63f1803b411edb6d59b3df7 --- /dev/null +++ b/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 333 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other +Output output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other +---------------- ------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,544B, BPFP=0.4310 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,432B, BPFP=2.5211 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,104B, BPFP=0.8632 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,292B, BPFP=2.4325 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,840B, BPFP=1.0759 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,960B, BPFP=2.4067 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,696B, BPFP=0.9869 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,760B, BPFP=2.4689 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,008B, BPFP=2.0218 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,844B, BPFP=2.3977 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,828B, BPFP=0.2535 +⌛️ [2/4] FRONTEND: Frontend time: 2.645s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.804s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09475397 29.48650060 + layer.0.v_cache 0.00001927 0.00914222 + layer.1.k_cache 0.17401673 12.69733486 + layer.1.v_cache 0.00000670 0.00370863 + layer.2.k_cache 0.00814445 1.60965512 + layer.2.v_cache 0.00002157 0.01055311 + layer.3.k_cache 0.01691350 9.26038509 + layer.3.v_cache 0.00002161 0.01304835 + layer.4.k_cache 0.00069969 0.35498920 + layer.4.v_cache 0.00005477 0.02187075 + layer.4.output 1.52309445 267.46768390 + ------------------------------------------------------------------------------------- + TOTAL 0.64448903 113.27888090 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 249308 +BPFP 1.1400 bits/point +EBPFP 2.2800 equivalent bits/point +MSE 113.278881 +---------------------- -------------------------------------------------------- +Time: 4.458s Load: 0.008s, Pack+Encode: 2.645s, Decode+Unpack: 1.804s +---------------------- -------------------------------------------------------- +💾 Converting with 113.2789 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 212, 128) +Output shape: (1, 212, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.output: torch.Size([1, 212, 3584]) -> torch.Size([1, 1, 212, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,852B, BPFP=0.4313 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,080B, BPFP=2.5855 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,504B, BPFP=0.7742 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,992B, BPFP=2.3579 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,640B, BPFP=1.0053 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,920B, BPFP=2.3526 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,412B, BPFP=0.9885 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,812B, BPFP=2.4183 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,872B, BPFP=2.1279 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,584B, BPFP=2.3278 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,872B, BPFP=0.1987 +⌛️ [2/4] FRONTEND: Frontend time: 2.297s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.550s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09403566 29.42552329 + layer.0.v_cache 0.00001799 0.00915113 + layer.1.k_cache 0.24957257 11.93011244 + layer.1.v_cache 0.00000621 0.00350614 + layer.2.k_cache 0.01169950 1.61428344 + layer.2.v_cache 0.00002031 0.01030798 + layer.3.k_cache 0.03645871 9.85143582 + layer.3.v_cache 0.00002058 0.01208817 + layer.4.k_cache 0.00069161 0.34551898 + layer.4.v_cache 0.00005130 0.02040544 + layer.4.output 1.44404146 253.46245367 + ------------------------------------------------------------------------------------- + TOTAL 0.61769792 107.49761815 + (elements=1,845,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1845248 +Total Bytes 254540 +BPFP 1.1035 bits/point +EBPFP 2.2071 equivalent bits/point +MSE 107.497618 +---------------------- -------------------------------------------------------- +Time: 3.854s Load: 0.007s, Pack+Encode: 2.297s, Decode+Unpack: 1.550s +---------------------- -------------------------------------------------------- +💾 Converting with 107.4976 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 250, 128) +Output shape: (1, 250, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.output: torch.Size([1, 250, 3584]) -> torch.Size([1, 1, 250, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,612B, BPFP=0.4133 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,828B, BPFP=2.0518 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,508B, BPFP=0.7192 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,980B, BPFP=1.9988 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,264B, BPFP=0.8915 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,736B, BPFP=1.9835 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,140B, BPFP=0.8213 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,444B, BPFP=2.0278 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,220B, BPFP=1.7637 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,280B, BPFP=1.9550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,348B, BPFP=0.2085 +⌛️ [2/4] FRONTEND: Frontend time: 2.035s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.782s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09121272 33.58999219 + layer.0.v_cache 0.00001793 0.00856388 + layer.1.k_cache 0.34710925 11.70233594 + layer.1.v_cache 0.00000615 0.00342703 + layer.2.k_cache 0.02211576 1.62921411 + layer.2.v_cache 0.00002179 0.01094959 + layer.3.k_cache 0.02174057 9.63645605 + layer.3.v_cache 0.00001938 0.01138599 + layer.4.k_cache 0.00073173 0.35500000 + layer.4.v_cache 0.00005145 0.02102119 + layer.4.output 1.22463198 214.97628571 + ------------------------------------------------------------------------------------- + TOTAL 0.53267356 91.87072623 + (elements=2,176,000) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2176000 +Total Bytes 257360 +BPFP 0.9462 bits/point +EBPFP 1.8924 equivalent bits/point +MSE 91.870726 +---------------------- -------------------------------------------------------- +Time: 3.826s Load: 0.009s, Pack+Encode: 2.035s, Decode+Unpack: 1.782s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8707 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,092B, BPFP=0.4193 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,208B, BPFP=2.2858 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,712B, BPFP=0.8062 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,100B, BPFP=2.2095 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,064B, BPFP=0.9681 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,920B, BPFP=2.1971 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,076B, BPFP=0.9001 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,360B, BPFP=2.2274 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,972B, BPFP=1.9254 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,252B, BPFP=2.1512 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,676B, BPFP=0.2033 +⌛️ [2/4] FRONTEND: Frontend time: 2.636s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.886s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12780563 30.40364297 + layer.0.v_cache 0.00001798 0.00848912 + layer.1.k_cache 0.36413978 12.72125056 + layer.1.v_cache 0.00000619 0.00335804 + layer.2.k_cache 0.01529831 1.59316869 + layer.2.v_cache 0.00002052 0.01008829 + layer.3.k_cache 0.01269315 9.88124763 + layer.3.v_cache 0.00002023 0.01121308 + layer.4.k_cache 0.00069261 0.33043600 + layer.4.v_cache 0.00005545 0.01964968 + layer.4.output 1.34865724 236.89065450 + ------------------------------------------------------------------------------------- + TOTAL 0.58596180 100.77747797 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 254432 +BPFP 1.0302 bits/point +EBPFP 2.0604 equivalent bits/point +MSE 100.777478 +---------------------- -------------------------------------------------------- +Time: 4.532s Load: 0.010s, Pack+Encode: 2.636s, Decode+Unpack: 1.886s +---------------------- -------------------------------------------------------- +💾 Converting with 100.7775 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 214, 128) +Output shape: (1, 214, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.output: torch.Size([1, 214, 3584]) -> torch.Size([1, 1, 214, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,844B, BPFP=0.4267 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,588B, BPFP=2.4524 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,900B, BPFP=0.7959 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,484B, BPFP=2.3718 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,216B, BPFP=1.0380 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,652B, BPFP=2.3110 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,540B, BPFP=0.9156 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,820B, BPFP=2.3963 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,808B, BPFP=2.1034 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,476B, BPFP=2.2982 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,984B, BPFP=0.1876 +⌛️ [2/4] FRONTEND: Frontend time: 1.985s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.466s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12820032 30.23208190 + layer.0.v_cache 0.00001693 0.00845728 + layer.1.k_cache 0.24875343 11.14924750 + layer.1.v_cache 0.00000587 0.00320477 + layer.2.k_cache 0.01453762 1.68506116 + layer.2.v_cache 0.00001988 0.01004866 + layer.3.k_cache 0.02276520 9.12612487 + layer.3.v_cache 0.00001932 0.01094160 + layer.4.k_cache 0.00071399 0.35321266 + layer.4.v_cache 0.00004895 0.02056928 + layer.4.output 1.43051836 251.41764019 + ------------------------------------------------------------------------------------- + TOTAL 0.61345353 106.61896653 + (elements=1,862,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1862656 +Total Bytes 252312 +BPFP 1.0837 bits/point +EBPFP 2.1673 equivalent bits/point +MSE 106.618967 +---------------------- -------------------------------------------------------- +Time: 3.459s Load: 0.008s, Pack+Encode: 1.985s, Decode+Unpack: 1.466s +---------------------- -------------------------------------------------------- +💾 Converting with 106.6190 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,524B, BPFP=0.4244 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,236B, BPFP=2.3824 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,928B, BPFP=0.7856 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,048B, BPFP=2.3154 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,420B, BPFP=0.9826 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,704B, BPFP=2.2396 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,448B, BPFP=0.8714 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,568B, BPFP=2.3448 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,832B, BPFP=2.0776 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,396B, BPFP=2.2222 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,832B, BPFP=0.2162 +⌛️ [2/4] FRONTEND: Frontend time: 2.685s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.587s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11875103 30.62772168 + layer.0.v_cache 0.00001689 0.00842964 + layer.1.k_cache 0.49171178 11.91817764 + layer.1.v_cache 0.00000616 0.00334481 + layer.2.k_cache 0.02532451 1.61984848 + layer.2.v_cache 0.00002086 0.01063949 + layer.3.k_cache 0.02370731 9.70964188 + layer.3.v_cache 0.00002129 0.01177898 + layer.4.k_cache 0.00073268 0.35447112 + layer.4.v_cache 0.00004879 0.02038681 + layer.4.output 0.00480151 198.23646209 + ------------------------------------------------------------------------------------- + TOTAL 0.04082070 84.81998089 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 321936 +BPFP 1.0682 bits/point +EBPFP 2.1364 equivalent bits/point +MSE 84.819981 +---------------------- -------------------------------------------------------- +Time: 4.283s Load: 0.011s, Pack+Encode: 2.685s, Decode+Unpack: 1.587s +---------------------- -------------------------------------------------------- +💾 Converting with 84.8200 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,624B, BPFP=0.4224 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,460B, BPFP=2.1339 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,228B, BPFP=0.7798 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,884B, BPFP=2.0334 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,576B, BPFP=0.9296 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,828B, BPFP=2.0298 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,672B, BPFP=0.8719 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,304B, BPFP=2.0602 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,896B, BPFP=1.9704 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,312B, BPFP=1.9969 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,572B, BPFP=0.1965 +⌛️ [2/4] FRONTEND: Frontend time: 2.059s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.459s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10657067 33.82767459 + layer.0.v_cache 0.00001604 0.00840828 + layer.1.k_cache 0.39270948 12.19732143 + layer.1.v_cache 0.00000611 0.00339622 + layer.2.k_cache 0.01541229 1.62059949 + layer.2.v_cache 0.00002285 0.01065751 + layer.3.k_cache 0.01321528 9.79617546 + layer.3.v_cache 0.00001985 0.01146490 + layer.4.k_cache 0.00076706 0.35044923 + layer.4.v_cache 0.00005205 0.02054923 + layer.4.output 1.24960368 220.16587099 + ------------------------------------------------------------------------------------- + TOTAL 0.54564808 94.05928196 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 260356 +BPFP 0.9767 bits/point +EBPFP 1.9535 equivalent bits/point +MSE 94.059282 +---------------------- -------------------------------------------------------- +Time: 3.526s Load: 0.008s, Pack+Encode: 2.059s, Decode+Unpack: 1.459s +---------------------- -------------------------------------------------------- +💾 Converting with 94.0593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 197, 128) +Output shape: (1, 197, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.output: torch.Size([1, 197, 3584]) -> torch.Size([1, 1, 197, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,488B, BPFP=0.4353 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,236B, BPFP=2.5568 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,656B, BPFP=0.8452 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,916B, BPFP=2.4521 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,348B, BPFP=1.0587 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,688B, BPFP=2.4340 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,532B, BPFP=0.9940 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,440B, BPFP=2.4937 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,428B, BPFP=2.0961 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,760B, BPFP=2.3604 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,304B, BPFP=0.2527 +⌛️ [2/4] FRONTEND: Frontend time: 2.053s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.749s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09219642 28.51955108 + layer.0.v_cache 0.00001621 0.00865432 + layer.1.k_cache 0.17840836 12.29666680 + layer.1.v_cache 0.00000639 0.00333292 + layer.2.k_cache 0.00922235 1.59484848 + layer.2.v_cache 0.00002142 0.01067068 + layer.3.k_cache 0.01957878 9.29777535 + layer.3.v_cache 0.00002124 0.01136163 + layer.4.k_cache 0.00068177 0.34671776 + layer.4.v_cache 0.00005086 0.02158709 + layer.4.output 1.55401793 273.41900834 + ------------------------------------------------------------------------------------- + TOTAL 0.65754878 115.64966026 + (elements=1,714,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1714688 +Total Bytes 245796 +BPFP 1.1468 bits/point +EBPFP 2.2936 equivalent bits/point +MSE 115.649660 +---------------------- -------------------------------------------------------- +Time: 3.809s Load: 0.007s, Pack+Encode: 2.053s, Decode+Unpack: 1.749s +---------------------- -------------------------------------------------------- +💾 Converting with 115.6497 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,288B, BPFP=0.4177 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,856B, BPFP=2.0593 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,384B, BPFP=0.7250 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,472B, BPFP=1.9895 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,056B, BPFP=0.9101 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,280B, BPFP=1.9798 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,300B, BPFP=0.8216 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,868B, BPFP=2.0095 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,560B, BPFP=1.7419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,048B, BPFP=1.9681 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,096B, BPFP=0.2023 +⌛️ [2/4] FRONTEND: Frontend time: 2.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.706s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11169702 34.64494708 + layer.0.v_cache 0.00001595 0.00807445 + layer.1.k_cache 0.54955656 11.37665543 + layer.1.v_cache 0.00000603 0.00332122 + layer.2.k_cache 0.01751511 1.62883104 + layer.2.v_cache 0.00002011 0.01094848 + layer.3.k_cache 0.01650532 9.68830094 + layer.3.v_cache 0.00001967 0.01141390 + layer.4.k_cache 0.00077751 0.34153895 + layer.4.v_cache 0.00005008 0.02134322 + layer.4.output 0.04306356 174.20914459 + ------------------------------------------------------------------------------------- + TOTAL 0.05868284 75.12937569 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 318208 +BPFP 0.9435 bits/point +EBPFP 1.8869 equivalent bits/point +MSE 75.129376 +---------------------- -------------------------------------------------------- +Time: 3.970s Load: 0.014s, Pack+Encode: 2.250s, Decode+Unpack: 1.706s +---------------------- -------------------------------------------------------- +💾 Converting with 75.1294 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,912B, BPFP=0.4237 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,420B, BPFP=2.3954 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,528B, BPFP=0.8263 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,496B, BPFP=2.3291 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,352B, BPFP=1.0287 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,052B, BPFP=2.2973 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,200B, BPFP=0.9461 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,824B, BPFP=2.3526 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,536B, BPFP=2.0453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,668B, BPFP=2.2698 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,904B, BPFP=0.2345 +⌛️ [2/4] FRONTEND: Frontend time: 2.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.534s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09832749 32.25235181 + layer.0.v_cache 0.00001658 0.00862035 + layer.1.k_cache 0.33493462 12.58300333 + layer.1.v_cache 0.00000656 0.00344801 + layer.2.k_cache 0.02762392 1.66241343 + layer.2.v_cache 0.00002012 0.01045586 + layer.3.k_cache 0.01647770 10.03539143 + layer.3.v_cache 0.00001951 0.01128046 + layer.4.k_cache 0.00067889 0.35210552 + layer.4.v_cache 0.00004874 0.02090048 + layer.4.output 1.40435175 247.53086091 + ------------------------------------------------------------------------------------- + TOTAL 0.60638920 105.27388218 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 258892 +BPFP 1.0915 bits/point +EBPFP 2.1830 equivalent bits/point +MSE 105.273882 +---------------------- -------------------------------------------------------- +Time: 3.766s Load: 0.010s, Pack+Encode: 2.222s, Decode+Unpack: 1.534s +---------------------- -------------------------------------------------------- +💾 Converting with 105.2739 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,064B, BPFP=0.4192 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,624B, BPFP=2.2555 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,536B, BPFP=0.8667 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,748B, BPFP=2.1950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,092B, BPFP=0.9743 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,836B, BPFP=2.2011 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,896B, BPFP=0.8916 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,364B, BPFP=2.2376 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,104B, BPFP=1.8739 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,304B, BPFP=2.1643 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,060B, BPFP=0.2278 +⌛️ [2/4] FRONTEND: Frontend time: 2.097s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.517s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102250 32.71009921 + layer.0.v_cache 0.00001721 0.00833153 + layer.1.k_cache 0.37371607 12.47478870 + layer.1.v_cache 0.00000658 0.00339060 + layer.2.k_cache 0.01127678 1.64864140 + layer.2.v_cache 0.00001998 0.01046959 + layer.3.k_cache 0.02136800 9.58709015 + layer.3.v_cache 0.00001963 0.01128980 + layer.4.k_cache 0.00069450 0.33254708 + layer.4.v_cache 0.00005147 0.02006589 + layer.4.output 1.35464589 238.25126422 + ------------------------------------------------------------------------------------- + TOTAL 0.58945376 101.44503315 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 255628 +BPFP 1.0396 bits/point +EBPFP 2.0792 equivalent bits/point +MSE 101.445033 +---------------------- -------------------------------------------------------- +Time: 3.622s Load: 0.007s, Pack+Encode: 2.097s, Decode+Unpack: 1.517s +---------------------- -------------------------------------------------------- +💾 Converting with 101.4450 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,612B, BPFP=0.4116 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,884B, BPFP=2.0471 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,756B, BPFP=0.7318 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,968B, BPFP=1.9900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,200B, BPFP=0.8840 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,948B, BPFP=2.0510 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,688B, BPFP=0.8521 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,548B, BPFP=2.0261 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,212B, BPFP=1.6940 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,264B, BPFP=1.9462 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,168B, BPFP=0.2149 +⌛️ [2/4] FRONTEND: Frontend time: 2.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.547s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11871604 32.65442916 + layer.0.v_cache 0.00001755 0.00812839 + layer.1.k_cache 0.46315337 11.74109519 + layer.1.v_cache 0.00000588 0.00317682 + layer.2.k_cache 0.02202094 1.55967025 + layer.2.v_cache 0.00001989 0.01034888 + layer.3.k_cache 0.01278766 10.02906149 + layer.3.v_cache 0.00002092 0.01119216 + layer.4.k_cache 0.00074254 0.35251791 + layer.4.v_cache 0.00004925 0.02111586 + layer.4.output 1.21975157 213.05595475 + ------------------------------------------------------------------------------------- + TOTAL 0.53857618 91.04602467 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 259248 +BPFP 0.9493 bits/point +EBPFP 1.8986 equivalent bits/point +MSE 91.046025 +---------------------- -------------------------------------------------------- +Time: 3.812s Load: 0.011s, Pack+Encode: 2.254s, Decode+Unpack: 1.547s +---------------------- -------------------------------------------------------- +💾 Converting with 91.0460 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 195, 128) +Output shape: (1, 195, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.output: torch.Size([1, 195, 3584]) -> torch.Size([1, 1, 195, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,332B, BPFP=0.4272 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,852B, BPFP=2.6324 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,480B, BPFP=0.8397 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,756B, BPFP=2.4644 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,800B, BPFP=1.0256 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,624B, BPFP=2.4538 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,148B, BPFP=0.9734 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,376B, BPFP=2.5141 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,696B, BPFP=1.8987 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,168B, BPFP=2.4173 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,556B, BPFP=0.2124 +⌛️ [2/4] FRONTEND: Frontend time: 2.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.478s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605874 29.82940956 + layer.0.v_cache 0.00001676 0.00848973 + layer.1.k_cache 0.13872500 12.56360427 + layer.1.v_cache 0.00000593 0.00333175 + layer.2.k_cache 0.01222654 1.59980844 + layer.2.v_cache 0.00001968 0.00950411 + layer.3.k_cache 0.01415822 9.60058656 + layer.3.v_cache 0.00002127 0.01152872 + layer.4.k_cache 0.00071127 0.34840014 + layer.4.v_cache 0.00005419 0.02115348 + layer.4.output 1.56987251 275.36785714 + ------------------------------------------------------------------------------------- + TOTAL 0.66300619 116.56298922 + (elements=1,697,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1697280 +Total Bytes 238788 +BPFP 1.1255 bits/point +EBPFP 2.2510 equivalent bits/point +MSE 116.562989 +---------------------- -------------------------------------------------------- +Time: 3.651s Load: 0.008s, Pack+Encode: 2.166s, Decode+Unpack: 1.478s +---------------------- -------------------------------------------------------- +💾 Converting with 116.5630 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,916B, BPFP=0.4220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,212B, BPFP=2.1645 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,872B, BPFP=0.7617 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,048B, BPFP=2.0646 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,332B, BPFP=0.9729 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,940B, BPFP=2.0553 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,872B, BPFP=0.9334 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,388B, BPFP=2.0938 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,012B, BPFP=2.1473 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,636B, BPFP=2.0292 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,032B, BPFP=0.2334 +⌛️ [2/4] FRONTEND: Frontend time: 2.605s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.323s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11176265 32.84636311 + layer.0.v_cache 0.00001802 0.00904559 + layer.1.k_cache 0.18491925 11.32030713 + layer.1.v_cache 0.00000608 0.00351125 + layer.2.k_cache 0.01999094 1.66523558 + layer.2.v_cache 0.00002176 0.01047593 + layer.3.k_cache 0.03910406 9.60689872 + layer.3.v_cache 0.00002004 0.01204777 + layer.4.k_cache 0.00067358 0.36193437 + layer.4.v_cache 0.00005212 0.02131487 + layer.4.output 0.00886795 299.48964874 + ------------------------------------------------------------------------------------- + TOTAL 0.02462613 126.60498091 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 201260 +BPFP 1.0164 bits/point +EBPFP 2.0328 equivalent bits/point +MSE 126.604981 +---------------------- -------------------------------------------------------- +Time: 3.935s Load: 0.007s, Pack+Encode: 2.605s, Decode+Unpack: 1.323s +---------------------- -------------------------------------------------------- +💾 Converting with 126.6050 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,868B, BPFP=0.4305 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,044B, BPFP=2.4240 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,864B, BPFP=0.7969 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,044B, BPFP=2.3506 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,024B, BPFP=1.0288 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,848B, BPFP=2.3363 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,116B, BPFP=0.9621 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,820B, BPFP=2.4076 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,876B, BPFP=1.9715 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,516B, BPFP=2.3119 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,328B, BPFP=0.2025 +⌛️ [2/4] FRONTEND: Frontend time: 2.038s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.489s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15089922 28.75061666 + layer.0.v_cache 0.00001686 0.00869462 + layer.1.k_cache 0.31217133 11.57219227 + layer.1.v_cache 0.00000593 0.00344292 + layer.2.k_cache 0.01305961 1.63217579 + layer.2.v_cache 0.00001999 0.01023740 + layer.3.k_cache 0.01178671 9.28579124 + layer.3.v_cache 0.00002039 0.01162220 + layer.4.k_cache 0.00067606 0.33346393 + layer.4.v_cache 0.00005285 0.02071869 + layer.4.output 1.43724915 251.70127012 + ------------------------------------------------------------------------------------- + TOTAL 0.62055606 106.67869686 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 251348 +BPFP 1.0846 bits/point +EBPFP 2.1692 equivalent bits/point +MSE 106.678697 +---------------------- -------------------------------------------------------- +Time: 3.534s Load: 0.007s, Pack+Encode: 2.038s, Decode+Unpack: 1.489s +---------------------- -------------------------------------------------------- +💾 Converting with 106.6787 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 187, 128) +Output shape: (1, 187, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.output: torch.Size([1, 187, 3584]) -> torch.Size([1, 1, 187, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,952B, BPFP=0.4138 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,652B, BPFP=2.1434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,660B, BPFP=0.7236 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,260B, BPFP=2.0271 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,264B, BPFP=0.9412 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,164B, BPFP=2.0191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,916B, BPFP=0.9121 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,832B, BPFP=2.0749 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,696B, BPFP=1.8964 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,760B, BPFP=1.9853 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,204B, BPFP=0.2531 +⌛️ [2/4] FRONTEND: Frontend time: 2.185s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.500s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08258281 31.50633460 + layer.0.v_cache 0.00001733 0.00950050 + layer.1.k_cache 0.10049684 10.79641022 + layer.1.v_cache 0.00000647 0.00360178 + layer.2.k_cache 0.01472783 1.50958529 + layer.2.v_cache 0.00002140 0.01128574 + layer.3.k_cache 0.02125903 10.05445380 + layer.3.v_cache 0.00002180 0.01279552 + layer.4.k_cache 0.00070831 0.36921215 + layer.4.v_cache 0.00005280 0.02426587 + layer.4.output 0.00866398 291.52938789 + ------------------------------------------------------------------------------------- + TOTAL 0.01650250 123.23548004 + (elements=1,627,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1627648 +Total Bytes 202360 +BPFP 0.9946 bits/point +EBPFP 1.9892 equivalent bits/point +MSE 123.235480 +---------------------- -------------------------------------------------------- +Time: 3.697s Load: 0.011s, Pack+Encode: 2.185s, Decode+Unpack: 1.500s +---------------------- -------------------------------------------------------- +💾 Converting with 123.2355 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,828B, BPFP=0.4316 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,176B, BPFP=2.4568 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,176B, BPFP=0.8276 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,396B, BPFP=2.3990 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,556B, BPFP=1.0039 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,792B, BPFP=2.4283 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,640B, BPFP=0.9360 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,832B, BPFP=2.4313 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,208B, BPFP=2.0889 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,644B, BPFP=2.3433 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,004B, BPFP=0.2434 +⌛️ [2/4] FRONTEND: Frontend time: 2.050s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.546s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11071491 28.82636858 + layer.0.v_cache 0.00001804 0.00862662 + layer.1.k_cache 0.26784508 12.29390482 + layer.1.v_cache 0.00000610 0.00335676 + layer.2.k_cache 0.01369454 1.61163128 + layer.2.v_cache 0.00002102 0.00997861 + layer.3.k_cache 0.01444715 9.55921567 + layer.3.v_cache 0.00002041 0.01147522 + layer.4.k_cache 0.00067564 0.35149351 + layer.4.v_cache 0.00005248 0.02107247 + layer.4.output 1.45091435 255.17042569 + ------------------------------------------------------------------------------------- + TOTAL 0.62140564 108.17000608 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 257252 +BPFP 1.1206 bits/point +EBPFP 2.2412 equivalent bits/point +MSE 108.170006 +---------------------- -------------------------------------------------------- +Time: 3.604s Load: 0.008s, Pack+Encode: 2.050s, Decode+Unpack: 1.546s +---------------------- -------------------------------------------------------- +💾 Converting with 108.1700 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,920B, BPFP=0.4148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,920B, BPFP=2.3066 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,956B, BPFP=0.7677 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,924B, BPFP=2.2368 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,536B, BPFP=0.9484 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,864B, BPFP=2.2326 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,972B, BPFP=0.9089 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,492B, BPFP=2.2766 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,468B, BPFP=1.9246 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,520B, BPFP=2.2085 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,724B, BPFP=0.2174 +⌛️ [2/4] FRONTEND: Frontend time: 2.330s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.434s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13389712 29.07196346 + layer.0.v_cache 0.00001748 0.00831979 + layer.1.k_cache 0.36644789 12.33298592 + layer.1.v_cache 0.00000587 0.00318598 + layer.2.k_cache 0.02260907 1.63917685 + layer.2.v_cache 0.00001979 0.00934100 + layer.3.k_cache 0.02550721 9.83003543 + layer.3.v_cache 0.00001968 0.01078890 + layer.4.k_cache 0.00067529 0.34584757 + layer.4.v_cache 0.00006705 0.02058159 + layer.4.output 1.37284074 241.39958360 + ------------------------------------------------------------------------------------- + TOTAL 0.59759716 102.53348892 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 253296 +BPFP 1.0440 bits/point +EBPFP 2.0880 equivalent bits/point +MSE 102.533489 +---------------------- -------------------------------------------------------- +Time: 3.772s Load: 0.008s, Pack+Encode: 2.330s, Decode+Unpack: 1.434s +---------------------- -------------------------------------------------------- +💾 Converting with 102.5335 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.017s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 406, 128) +Output shape: (1, 406, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.output: torch.Size([1, 406, 3584]) -> torch.Size([1, 1, 406, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,820B, BPFP=0.4164 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,748B, BPFP=2.2224 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,920B, BPFP=0.7666 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,352B, BPFP=2.1687 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,924B, BPFP=0.9207 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,012B, BPFP=2.1556 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,184B, BPFP=0.8538 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,108B, BPFP=2.2363 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 50,520B, BPFP=1.9443 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,020B, BPFP=2.1175 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,252B, BPFP=0.2378 +⌛️ [2/4] FRONTEND: Frontend time: 3.080s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.976s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10474718 32.39786984 + layer.0.v_cache 0.00001775 0.00809694 + layer.1.k_cache 0.89873824 11.67790034 + layer.1.v_cache 0.00000687 0.00326283 + layer.2.k_cache 0.02602503 1.57829352 + layer.2.v_cache 0.00002164 0.01026156 + layer.3.k_cache 0.01558245 9.55950236 + layer.3.v_cache 0.00002201 0.01086162 + layer.4.k_cache 0.00072958 0.36728954 + layer.4.v_cache 0.00005327 0.01981466 + layer.4.output 0.00651461 135.33372185 + ------------------------------------------------------------------------------------- + TOTAL 0.06420861 58.99818860 + (elements=3,533,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3533824 +Total Bytes 453860 +BPFP 1.0275 bits/point +EBPFP 2.0549 equivalent bits/point +MSE 58.998189 +---------------------- -------------------------------------------------------- +Time: 5.072s Load: 0.017s, Pack+Encode: 3.080s, Decode+Unpack: 1.976s +---------------------- -------------------------------------------------------- +💾 Converting with 58.9982 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,604B, BPFP=0.4212 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,440B, BPFP=2.1327 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,768B, BPFP=0.7505 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,276B, BPFP=2.0584 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,364B, BPFP=0.9161 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,856B, BPFP=2.0316 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,764B, BPFP=0.8778 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,616B, BPFP=2.0801 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,504B, BPFP=1.7541 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,428B, BPFP=2.0043 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,936B, BPFP=0.2272 +⌛️ [2/4] FRONTEND: Frontend time: 2.336s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.614s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14414167 34.06937580 + layer.0.v_cache 0.00001717 0.00858427 + layer.1.k_cache 0.44179790 12.31233060 + layer.1.v_cache 0.00000667 0.00346328 + layer.2.k_cache 0.03016641 1.59886612 + layer.2.v_cache 0.00002043 0.01067261 + layer.3.k_cache 0.03726396 9.94188457 + layer.3.v_cache 0.00002099 0.01156158 + layer.4.k_cache 0.00070653 0.34387624 + layer.4.v_cache 0.00004910 0.02122369 + layer.4.output 1.24963187 220.42629373 + ------------------------------------------------------------------------------------- + TOTAL 0.55303611 94.19446440 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 260556 +BPFP 0.9775 bits/point +EBPFP 1.9550 equivalent bits/point +MSE 94.194464 +---------------------- -------------------------------------------------------- +Time: 3.958s Load: 0.009s, Pack+Encode: 2.336s, Decode+Unpack: 1.614s +---------------------- -------------------------------------------------------- +💾 Converting with 94.1945 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,512B, BPFP=0.4328 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,404B, BPFP=2.6228 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,260B, BPFP=0.8056 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,420B, BPFP=2.4670 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,280B, BPFP=1.0427 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,608B, BPFP=2.4033 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,684B, BPFP=0.9959 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,464B, BPFP=2.4705 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,856B, BPFP=2.1087 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,264B, BPFP=2.3763 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,472B, BPFP=0.1960 +⌛️ [2/4] FRONTEND: Frontend time: 2.067s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.498s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15985105 31.28919009 + layer.0.v_cache 0.00001905 0.00907590 + layer.1.k_cache 0.22745673 11.88058702 + layer.1.v_cache 0.00000590 0.00345284 + layer.2.k_cache 0.01549998 1.60794942 + layer.2.v_cache 0.00002003 0.01062071 + layer.3.k_cache 0.01333464 9.45745819 + layer.3.v_cache 0.00002034 0.01226302 + layer.4.k_cache 0.00067076 0.34509691 + layer.4.v_cache 0.00004949 0.02077993 + layer.4.output 1.53832061 271.63437724 + ------------------------------------------------------------------------------------- + TOTAL 0.65795131 115.06335969 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 243224 +BPFP 1.1234 bits/point +EBPFP 2.2467 equivalent bits/point +MSE 115.063360 +---------------------- -------------------------------------------------------- +Time: 3.571s Load: 0.006s, Pack+Encode: 2.067s, Decode+Unpack: 1.498s +---------------------- -------------------------------------------------------- +💾 Converting with 115.0634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 237, 128) +Output shape: (1, 237, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.output: torch.Size([1, 237, 3584]) -> torch.Size([1, 1, 237, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,452B, BPFP=0.4254 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,648B, BPFP=2.1524 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,476B, BPFP=0.8225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,048B, BPFP=2.1129 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,724B, BPFP=0.9707 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,092B, BPFP=2.1158 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,676B, BPFP=0.9016 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,436B, BPFP=2.1384 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,036B, BPFP=1.9143 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,440B, BPFP=2.0728 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,856B, BPFP=0.2247 +⌛️ [2/4] FRONTEND: Frontend time: 2.331s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.659s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351868 35.00358073 + layer.0.v_cache 0.00001655 0.00844078 + layer.1.k_cache 0.34375901 11.23534538 + layer.1.v_cache 0.00000619 0.00337038 + layer.2.k_cache 0.01282014 1.65141457 + layer.2.v_cache 0.00002104 0.00998803 + layer.3.k_cache 0.03241006 9.90920441 + layer.3.v_cache 0.00001985 0.01098381 + layer.4.k_cache 0.00067667 0.34620103 + layer.4.v_cache 0.00005279 0.02001301 + layer.4.output 1.29178675 227.59358047 + ------------------------------------------------------------------------------------- + TOTAL 0.56210637 97.13844738 + (elements=2,062,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2062848 +Total Bytes 260884 +BPFP 1.0117 bits/point +EBPFP 2.0235 equivalent bits/point +MSE 97.138447 +---------------------- -------------------------------------------------------- +Time: 3.998s Load: 0.009s, Pack+Encode: 2.331s, Decode+Unpack: 1.659s +---------------------- -------------------------------------------------------- +💾 Converting with 97.1384 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,832B, BPFP=0.4238 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,428B, BPFP=2.5747 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,800B, BPFP=0.7849 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,160B, BPFP=2.3372 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,068B, BPFP=1.0224 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,924B, BPFP=2.3201 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,992B, BPFP=0.9442 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,640B, BPFP=2.3721 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,516B, BPFP=2.0724 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,432B, BPFP=2.2843 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,148B, BPFP=0.2092 +⌛️ [2/4] FRONTEND: Frontend time: 2.032s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.589s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11023432 29.87946493 + layer.0.v_cache 0.00001677 0.00834693 + layer.1.k_cache 0.30291741 11.33223565 + layer.1.v_cache 0.00000597 0.00329392 + layer.2.k_cache 0.03464115 1.61528845 + layer.2.v_cache 0.00001998 0.00982996 + layer.3.k_cache 0.01330193 9.95540743 + layer.3.v_cache 0.00002112 0.01066129 + layer.4.k_cache 0.00069379 0.34241649 + layer.4.v_cache 0.00005021 0.01969037 + layer.4.output 1.42388847 250.61306063 + ------------------------------------------------------------------------------------- + TOTAL 0.61347776 106.32165058 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 255940 +BPFP 1.0941 bits/point +EBPFP 2.1883 equivalent bits/point +MSE 106.321651 +---------------------- -------------------------------------------------------- +Time: 3.629s Load: 0.008s, Pack+Encode: 2.032s, Decode+Unpack: 1.589s +---------------------- -------------------------------------------------------- +💾 Converting with 106.3217 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 231, 128) +Output shape: (1, 231, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.output: torch.Size([1, 231, 3584]) -> torch.Size([1, 1, 231, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,284B, BPFP=0.4251 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,940B, BPFP=2.3634 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,740B, BPFP=0.7941 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,332B, BPFP=2.1870 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,564B, BPFP=0.9851 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,220B, BPFP=2.1794 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,800B, BPFP=0.9334 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,640B, BPFP=2.2078 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,456B, BPFP=1.9924 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,500B, BPFP=2.1307 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,612B, BPFP=0.2475 +⌛️ [2/4] FRONTEND: Frontend time: 2.537s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.694s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13257201 34.52636085 + layer.0.v_cache 0.00001701 0.00859462 + layer.1.k_cache 0.40462213 11.43121470 + layer.1.v_cache 0.00000658 0.00336597 + layer.2.k_cache 0.02649644 1.56099499 + layer.2.v_cache 0.00002056 0.00960409 + layer.3.k_cache 0.06147716 9.92796583 + layer.3.v_cache 0.00002105 0.01138856 + layer.4.k_cache 0.00069437 0.34749871 + layer.4.v_cache 0.00005136 0.02091420 + layer.4.output 1.32536713 233.87961889 + ------------------------------------------------------------------------------------- + TOTAL 0.58256168 99.70619028 + (elements=2,010,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2010624 +Total Bytes 265088 +BPFP 1.0547 bits/point +EBPFP 2.1095 equivalent bits/point +MSE 99.706190 +---------------------- -------------------------------------------------------- +Time: 4.239s Load: 0.009s, Pack+Encode: 2.537s, Decode+Unpack: 1.694s +---------------------- -------------------------------------------------------- +💾 Converting with 99.7062 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,580B, BPFP=0.4032 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,796B, BPFP=2.0096 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,652B, BPFP=0.7140 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,980B, BPFP=1.9596 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,952B, BPFP=0.8549 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,324B, BPFP=2.1645 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,980B, BPFP=0.7953 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,444B, BPFP=1.9880 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,464B, BPFP=1.8667 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,504B, BPFP=1.9304 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,956B, BPFP=0.2360 +⌛️ [2/4] FRONTEND: Frontend time: 2.191s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.588s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12841698 35.63269378 + layer.0.v_cache 0.00001718 0.00816612 + layer.1.k_cache 0.45450200 12.65668945 + layer.1.v_cache 0.00000627 0.00330324 + layer.2.k_cache 0.01434461 1.67734806 + layer.2.v_cache 0.00002215 0.01043527 + layer.3.k_cache 0.03530993 10.18805722 + layer.3.v_cache 0.00002139 0.01126596 + layer.4.k_cache 0.00068344 0.34435802 + layer.4.v_cache 0.00005073 0.01943708 + layer.4.output 1.20064817 207.59094888 + ------------------------------------------------------------------------------------- + TOTAL 0.53164187 89.04049390 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 266632 +BPFP 0.9610 bits/point +EBPFP 1.9221 equivalent bits/point +MSE 89.040494 +---------------------- -------------------------------------------------------- +Time: 3.790s Load: 0.011s, Pack+Encode: 2.191s, Decode+Unpack: 1.588s +---------------------- -------------------------------------------------------- +💾 Converting with 89.0405 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,528B, BPFP=0.4171 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,352B, BPFP=2.2912 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,172B, BPFP=0.7852 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,244B, BPFP=2.2298 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,052B, BPFP=0.9448 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,980B, BPFP=2.2152 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,824B, BPFP=0.8768 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,676B, BPFP=2.3092 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,144B, BPFP=2.0581 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,400B, BPFP=2.1831 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,188B, BPFP=0.2073 +⌛️ [2/4] FRONTEND: Frontend time: 2.427s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.809s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13731835 33.01862741 + layer.0.v_cache 0.00001709 0.00824683 + layer.1.k_cache 0.55879569 11.69543197 + layer.1.v_cache 0.00000637 0.00326795 + layer.2.k_cache 0.01348169 1.59262301 + layer.2.v_cache 0.00002375 0.01019419 + layer.3.k_cache 0.01306405 9.66148257 + layer.3.v_cache 0.00002005 0.01096242 + layer.4.k_cache 0.00072533 0.35677730 + layer.4.v_cache 0.00005384 0.02044908 + layer.4.output 0.00472113 195.54592515 + ------------------------------------------------------------------------------------- + TOTAL 0.04450318 83.83526699 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 320560 +BPFP 1.0448 bits/point +EBPFP 2.0896 equivalent bits/point +MSE 83.835267 +---------------------- -------------------------------------------------------- +Time: 4.247s Load: 0.011s, Pack+Encode: 2.427s, Decode+Unpack: 1.809s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8353 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,076B, BPFP=0.4192 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,020B, BPFP=2.1294 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,988B, BPFP=0.7780 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,072B, BPFP=2.0801 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,876B, BPFP=0.9279 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,936B, BPFP=2.0731 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,108B, BPFP=0.8881 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,964B, BPFP=2.1265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,552B, BPFP=1.9493 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,160B, BPFP=2.0328 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,964B, BPFP=0.2074 +⌛️ [2/4] FRONTEND: Frontend time: 2.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.786s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13972237 35.51563149 + layer.0.v_cache 0.00001689 0.00837409 + layer.1.k_cache 0.51719427 12.80659845 + layer.1.v_cache 0.00000613 0.00338496 + layer.2.k_cache 0.02019965 1.59900054 + layer.2.v_cache 0.00002053 0.00987039 + layer.3.k_cache 0.01759115 9.89066637 + layer.3.v_cache 0.00002125 0.01106603 + layer.4.k_cache 0.00069604 0.35718323 + layer.4.v_cache 0.00005872 0.02009615 + layer.4.output 0.04434862 179.37359101 + ------------------------------------------------------------------------------------- + TOTAL 0.05917455 77.40217699 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 324716 +BPFP 0.9915 bits/point +EBPFP 1.9831 equivalent bits/point +MSE 77.402177 +---------------------- -------------------------------------------------------- +Time: 4.054s Load: 0.011s, Pack+Encode: 2.257s, Decode+Unpack: 1.786s +---------------------- -------------------------------------------------------- +💾 Converting with 77.4022 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,528B, BPFP=0.4340 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,652B, BPFP=2.5638 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,940B, BPFP=0.8590 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,936B, BPFP=2.4290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,024B, BPFP=1.0226 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,616B, BPFP=2.4039 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,924B, BPFP=0.9362 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,632B, BPFP=2.4837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,712B, BPFP=2.0974 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,296B, BPFP=2.3788 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,356B, BPFP=0.2059 +⌛️ [2/4] FRONTEND: Frontend time: 2.080s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.616s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13044639 29.72352485 + layer.0.v_cache 0.00001737 0.00873189 + layer.1.k_cache 0.22565776 12.22904807 + layer.1.v_cache 0.00000620 0.00350937 + layer.2.k_cache 0.01998205 1.56579866 + layer.2.v_cache 0.00001989 0.01021768 + layer.3.k_cache 0.00629339 7.90612486 + layer.3.v_cache 0.00002050 0.01159650 + layer.4.k_cache 0.00068118 0.35803644 + layer.4.v_cache 0.00005692 0.02193824 + layer.4.output 1.53833376 271.34947057 + ------------------------------------------------------------------------------------- + TOTAL 0.65597165 114.78146003 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 242616 +BPFP 1.1206 bits/point +EBPFP 2.2411 equivalent bits/point +MSE 114.781460 +---------------------- -------------------------------------------------------- +Time: 3.703s Load: 0.007s, Pack+Encode: 2.080s, Decode+Unpack: 1.616s +---------------------- -------------------------------------------------------- +💾 Converting with 114.7815 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 246, 128) +Output shape: (1, 246, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.output: torch.Size([1, 246, 3584]) -> torch.Size([1, 1, 246, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,548B, BPFP=0.4159 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,604B, BPFP=2.0709 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,524B, BPFP=0.7320 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,672B, BPFP=2.0117 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,604B, BPFP=0.9276 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,448B, BPFP=1.9975 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,592B, BPFP=0.8633 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,464B, BPFP=2.1255 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,944B, BPFP=1.9654 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,024B, BPFP=1.9705 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,928B, BPFP=0.1990 +⌛️ [2/4] FRONTEND: Frontend time: 2.368s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.566s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131485 32.98197131 + layer.0.v_cache 0.00001613 0.00873542 + layer.1.k_cache 0.49701033 12.03151101 + layer.1.v_cache 0.00000620 0.00348206 + layer.2.k_cache 0.01298508 1.60530996 + layer.2.v_cache 0.00002120 0.01024241 + layer.3.k_cache 0.02762915 9.86039523 + layer.3.v_cache 0.00002136 0.01177600 + layer.4.k_cache 0.00068392 0.34003455 + layer.4.v_cache 0.00006372 0.01991857 + layer.4.output 1.24453647 218.55208333 + ------------------------------------------------------------------------------------- + TOTAL 0.55302984 93.33752705 + (elements=2,141,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2141184 +Total Bytes 259352 +BPFP 0.9690 bits/point +EBPFP 1.9380 equivalent bits/point +MSE 93.337527 +---------------------- -------------------------------------------------------- +Time: 3.943s Load: 0.009s, Pack+Encode: 2.368s, Decode+Unpack: 1.566s +---------------------- -------------------------------------------------------- +💾 Converting with 93.3375 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,488B, BPFP=0.4288 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,696B, BPFP=2.6325 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,552B, BPFP=0.9025 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,724B, BPFP=2.4003 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,308B, BPFP=1.0397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,676B, BPFP=2.3966 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,716B, BPFP=0.9934 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,280B, BPFP=2.4438 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,180B, BPFP=2.1234 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,244B, BPFP=2.3628 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,776B, BPFP=0.1872 +⌛️ [2/4] FRONTEND: Frontend time: 2.141s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.550s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11599884 29.36478516 + layer.0.v_cache 0.00001671 0.00835534 + layer.1.k_cache 0.22559120 12.75709961 + layer.1.v_cache 0.00000595 0.00334084 + layer.2.k_cache 0.01146694 1.60523514 + layer.2.v_cache 0.00002201 0.01034130 + layer.3.k_cache 0.03668072 9.51519897 + layer.3.v_cache 0.00002131 0.01143933 + layer.4.k_cache 0.00066762 0.33045803 + layer.4.v_cache 0.00005146 0.02074425 + layer.4.output 1.53061454 270.12370536 + ------------------------------------------------------------------------------------- + TOTAL 0.65322497 114.38193738 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 243640 +BPFP 1.1197 bits/point +EBPFP 2.2393 equivalent bits/point +MSE 114.381937 +---------------------- -------------------------------------------------------- +Time: 3.699s Load: 0.008s, Pack+Encode: 2.141s, Decode+Unpack: 1.550s +---------------------- -------------------------------------------------------- +💾 Converting with 114.3819 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,180B, BPFP=0.4217 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,756B, BPFP=2.2350 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,344B, BPFP=0.7740 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,036B, BPFP=2.1859 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,496B, BPFP=0.9891 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,968B, BPFP=2.1812 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,452B, BPFP=0.9178 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,308B, BPFP=2.2044 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,572B, BPFP=1.8813 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,148B, BPFP=2.1253 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,640B, BPFP=0.2597 +⌛️ [2/4] FRONTEND: Frontend time: 1.966s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.474s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605523 32.36276312 + layer.0.v_cache 0.00001866 0.00876180 + layer.1.k_cache 0.31881074 11.37495629 + layer.1.v_cache 0.00000653 0.00356144 + layer.2.k_cache 0.02091764 1.60756100 + layer.2.v_cache 0.00002113 0.01046115 + layer.3.k_cache 0.03221333 9.86724827 + layer.3.v_cache 0.00002185 0.01194519 + layer.4.k_cache 0.00068277 0.35043325 + layer.4.v_cache 0.00005596 0.02118948 + layer.4.output 1.33695214 235.57639972 + ------------------------------------------------------------------------------------- + TOTAL 0.57926287 100.27374582 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 259900 +BPFP 1.0431 bits/point +EBPFP 2.0863 equivalent bits/point +MSE 100.273746 +---------------------- -------------------------------------------------------- +Time: 3.446s Load: 0.007s, Pack+Encode: 1.966s, Decode+Unpack: 1.474s +---------------------- -------------------------------------------------------- +💾 Converting with 100.2737 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 192, 128) +Output shape: (1, 192, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.output: torch.Size([1, 192, 3584]) -> torch.Size([1, 1, 192, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,696B, BPFP=0.3822 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,044B, BPFP=1.9567 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,872B, BPFP=0.7220 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,520B, BPFP=1.9141 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,916B, BPFP=0.8070 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,672B, BPFP=1.9264 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,232B, BPFP=0.7513 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,908B, BPFP=1.9456 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,680B, BPFP=1.9271 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,256B, BPFP=1.8926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,708B, BPFP=0.1942 +⌛️ [2/4] FRONTEND: Frontend time: 1.899s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.446s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12444629 34.34461466 + layer.0.v_cache 0.00001540 0.00744538 + layer.1.k_cache 0.18411521 13.91219203 + layer.1.v_cache 0.00000585 0.00303132 + layer.2.k_cache 0.01124177 1.60590776 + layer.2.v_cache 0.00002179 0.00965008 + layer.3.k_cache 0.01616090 10.00791105 + layer.3.v_cache 0.00002117 0.01052136 + layer.4.k_cache 0.00068507 0.33654769 + layer.4.v_cache 0.00005275 0.01784333 + layer.4.output 0.00839530 281.43836031 + ------------------------------------------------------------------------------------- + TOTAL 0.02326666 119.43083452 + (elements=1,671,168) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1671168 +Total Bytes 191504 +BPFP 0.9167 bits/point +EBPFP 1.8335 equivalent bits/point +MSE 119.430835 +---------------------- -------------------------------------------------------- +Time: 3.353s Load: 0.007s, Pack+Encode: 1.899s, Decode+Unpack: 1.446s +---------------------- -------------------------------------------------------- +💾 Converting with 119.4308 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,580B, BPFP=0.4311 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,464B, BPFP=2.3968 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,264B, BPFP=0.7779 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,536B, BPFP=2.3095 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,172B, BPFP=1.0516 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,700B, BPFP=2.2308 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,432B, BPFP=0.9819 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,088B, BPFP=2.2673 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,768B, BPFP=2.2372 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,320B, BPFP=2.1950 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,672B, BPFP=0.1838 +⌛️ [2/4] FRONTEND: Frontend time: 2.058s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.575s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09999813 31.23005106 + layer.0.v_cache 0.00001899 0.00895177 + layer.1.k_cache 0.12989934 11.67006306 + layer.1.v_cache 0.00000610 0.00353004 + layer.2.k_cache 0.00966061 1.67319020 + layer.2.v_cache 0.00001994 0.01094528 + layer.3.k_cache 0.03384685 9.62242623 + layer.3.v_cache 0.00002040 0.01216613 + layer.4.k_cache 0.00066502 0.33291336 + layer.4.v_cache 0.00005200 0.02196788 + layer.4.output 0.00959846 326.72595740 + ------------------------------------------------------------------------------------- + TOTAL 0.02008098 137.74517099 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 192996 +BPFP 1.0686 bits/point +EBPFP 2.1372 equivalent bits/point +MSE 137.745171 +---------------------- -------------------------------------------------------- +Time: 3.644s Load: 0.011s, Pack+Encode: 2.058s, Decode+Unpack: 1.575s +---------------------- -------------------------------------------------------- +💾 Converting with 137.7452 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 428, 128) +Output shape: (1, 428, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.output: torch.Size([1, 428, 3584]) -> torch.Size([1, 1, 428, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,312B, BPFP=0.4130 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,104B, BPFP=2.0847 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,348B, BPFP=0.7428 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,560B, BPFP=2.0283 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,272B, BPFP=0.8861 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,404B, BPFP=2.0226 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,972B, BPFP=0.8386 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,052B, BPFP=2.0463 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 49,184B, BPFP=1.7956 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,456B, BPFP=1.9880 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,532B, BPFP=0.2166 +⌛️ [2/4] FRONTEND: Frontend time: 2.644s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.935s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13760449 35.75435802 + layer.0.v_cache 0.00001624 0.00778396 + layer.1.k_cache 0.89392254 11.59850390 + layer.1.v_cache 0.00000615 0.00304362 + layer.2.k_cache 0.03917003 1.57279954 + layer.2.v_cache 0.00002113 0.01003412 + layer.3.k_cache 0.03749864 10.00879591 + layer.3.v_cache 0.00002059 0.01017226 + layer.4.k_cache 0.00086919 0.33503819 + layer.4.v_cache 0.00005177 0.01885968 + layer.4.output 0.00617707 128.05719084 + ------------------------------------------------------------------------------------- + TOTAL 0.06778943 56.21880736 + (elements=3,725,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3725312 +Total Bytes 448196 +BPFP 0.9625 bits/point +EBPFP 1.9250 equivalent bits/point +MSE 56.218807 +---------------------- -------------------------------------------------------- +Time: 4.594s Load: 0.016s, Pack+Encode: 2.644s, Decode+Unpack: 1.935s +---------------------- -------------------------------------------------------- +💾 Converting with 56.2188 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,600B, BPFP=0.4211 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,508B, BPFP=2.2999 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,084B, BPFP=0.7804 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,372B, BPFP=2.2369 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,316B, BPFP=0.9594 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,056B, BPFP=2.2194 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,160B, BPFP=0.8954 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,144B, BPFP=2.3351 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,892B, BPFP=1.8225 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,424B, BPFP=2.1844 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,396B, BPFP=0.2327 +⌛️ [2/4] FRONTEND: Frontend time: 2.172s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.702s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385377 31.62427970 + layer.0.v_cache 0.00001705 0.00819773 + layer.1.k_cache 0.49372572 12.67583493 + layer.1.v_cache 0.00000626 0.00320710 + layer.2.k_cache 0.02422354 1.57367593 + layer.2.v_cache 0.00002046 0.01004805 + layer.3.k_cache 0.04153839 9.76150837 + layer.3.v_cache 0.00002075 0.01094628 + layer.4.k_cache 0.00074884 0.35660415 + layer.4.v_cache 0.00004954 0.01961970 + layer.4.output 0.00474379 195.90333713 + ------------------------------------------------------------------------------------- + TOTAL 0.04161240 83.96278129 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 320952 +BPFP 1.0461 bits/point +EBPFP 2.0921 equivalent bits/point +MSE 83.962781 +---------------------- -------------------------------------------------------- +Time: 3.883s Load: 0.009s, Pack+Encode: 2.172s, Decode+Unpack: 1.702s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9628 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,664B, BPFP=0.4191 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,220B, BPFP=2.3810 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,536B, BPFP=0.7999 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,144B, BPFP=2.3289 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,176B, BPFP=0.9276 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,536B, BPFP=2.2512 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,872B, BPFP=0.8646 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 51,096B, BPFP=2.4717 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,864B, BPFP=1.7833 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,620B, BPFP=2.2068 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,116B, BPFP=0.2358 +⌛️ [2/4] FRONTEND: Frontend time: 3.034s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.903s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12985682 33.41133840 + layer.0.v_cache 0.00001645 0.00780341 + layer.1.k_cache 0.61935935 12.83560039 + layer.1.v_cache 0.00000656 0.00316550 + layer.2.k_cache 0.01492825 1.43964436 + layer.2.v_cache 0.00002178 0.00997230 + layer.3.k_cache 0.02579401 8.28740794 + layer.3.v_cache 0.00002068 0.01022831 + layer.4.k_cache 0.00071613 0.35086423 + layer.4.v_cache 0.00005057 0.01947856 + layer.4.output 0.04141478 166.71685648 + ------------------------------------------------------------------------------------- + TOTAL 0.06356906 71.96432346 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 373844 +BPFP 1.0638 bits/point +EBPFP 2.1276 equivalent bits/point +MSE 71.964323 +---------------------- -------------------------------------------------------- +Time: 4.948s Load: 0.011s, Pack+Encode: 3.034s, Decode+Unpack: 1.903s +---------------------- -------------------------------------------------------- +💾 Converting with 71.9643 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,544B, BPFP=0.4122 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,172B, BPFP=2.2493 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,820B, BPFP=0.7550 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,116B, BPFP=2.1917 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,272B, BPFP=0.9436 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,008B, BPFP=2.1858 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,456B, BPFP=0.8990 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,700B, BPFP=2.2236 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,464B, BPFP=1.8282 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,272B, BPFP=2.1455 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,916B, BPFP=0.2257 +⌛️ [2/4] FRONTEND: Frontend time: 2.322s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.724s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12359551 33.27349896 + layer.0.v_cache 0.00001766 0.00854435 + layer.1.k_cache 0.52692803 12.86349039 + layer.1.v_cache 0.00000637 0.00332029 + layer.2.k_cache 0.02102186 1.61042103 + layer.2.v_cache 0.00002187 0.00998049 + layer.3.k_cache 0.03868174 10.05502725 + layer.3.v_cache 0.00002097 0.01138808 + layer.4.k_cache 0.00070799 0.33911432 + layer.4.v_cache 0.00005204 0.01970507 + layer.4.output 0.00469152 192.62336101 + ------------------------------------------------------------------------------------- + TOTAL 0.04375851 82.73870690 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 318740 +BPFP 1.0243 bits/point +EBPFP 2.0487 equivalent bits/point +MSE 82.738707 +---------------------- -------------------------------------------------------- +Time: 4.058s Load: 0.012s, Pack+Encode: 2.322s, Decode+Unpack: 1.724s +---------------------- -------------------------------------------------------- +💾 Converting with 82.7387 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,448B, BPFP=0.4278 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,868B, BPFP=2.5807 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,424B, BPFP=0.8185 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,288B, BPFP=2.4567 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,312B, BPFP=1.0452 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,996B, BPFP=2.4337 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,228B, BPFP=0.9601 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,916B, BPFP=2.5060 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,204B, BPFP=1.9790 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,180B, BPFP=2.3697 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,196B, BPFP=0.2265 +⌛️ [2/4] FRONTEND: Frontend time: 2.380s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.756s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10671347 31.77554030 + layer.0.v_cache 0.00001788 0.00897793 + layer.1.k_cache 0.15257932 11.28138741 + layer.1.v_cache 0.00000642 0.00346867 + layer.2.k_cache 0.02197224 1.56204423 + layer.2.v_cache 0.00002094 0.01004849 + layer.3.k_cache 0.03325213 7.91825767 + layer.3.v_cache 0.00002126 0.01171374 + layer.4.k_cache 0.00067229 0.35735785 + layer.4.v_cache 0.00005023 0.02007986 + layer.4.output 1.53834953 271.39931802 + ------------------------------------------------------------------------------------- + TOTAL 0.65198546 114.86730013 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 244060 +BPFP 1.1272 bits/point +EBPFP 2.2545 equivalent bits/point +MSE 114.867300 +---------------------- -------------------------------------------------------- +Time: 4.147s Load: 0.010s, Pack+Encode: 2.380s, Decode+Unpack: 1.756s +---------------------- -------------------------------------------------------- +💾 Converting with 114.8673 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,552B, BPFP=0.4079 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,860B, BPFP=2.0456 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,036B, BPFP=0.7493 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,060B, BPFP=1.9958 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,912B, BPFP=0.9283 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,932B, BPFP=1.9878 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,968B, BPFP=0.8695 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,848B, BPFP=2.1693 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,112B, BPFP=2.1235 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,464B, BPFP=1.9587 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,080B, BPFP=0.2319 +⌛️ [2/4] FRONTEND: Frontend time: 2.444s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.685s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15214830 33.74894173 + layer.0.v_cache 0.00001722 0.00887040 + layer.1.k_cache 0.49754258 12.28073740 + layer.1.v_cache 0.00000639 0.00360124 + layer.2.k_cache 0.01862443 1.58394510 + layer.2.v_cache 0.00002219 0.01045650 + layer.3.k_cache 0.01200796 9.92687162 + layer.3.v_cache 0.00002210 0.01187319 + layer.4.k_cache 0.00075226 0.37497015 + layer.4.v_cache 0.00005409 0.02174927 + layer.4.output 1.21979868 214.71770063 + ------------------------------------------------------------------------------------- + TOTAL 0.54234049 91.82328947 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 270824 +BPFP 0.9917 bits/point +EBPFP 1.9834 equivalent bits/point +MSE 91.823289 +---------------------- -------------------------------------------------------- +Time: 4.139s Load: 0.009s, Pack+Encode: 2.444s, Decode+Unpack: 1.685s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8233 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,128B, BPFP=0.4164 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,936B, BPFP=2.1996 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,168B, BPFP=0.7258 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,740B, BPFP=2.0359 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,956B, BPFP=0.9199 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,536B, BPFP=2.0254 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,932B, BPFP=0.8674 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,460B, BPFP=2.0727 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,052B, BPFP=1.9494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,056B, BPFP=2.0008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,888B, BPFP=0.1895 +⌛️ [2/4] FRONTEND: Frontend time: 2.270s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.752s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18432382 34.67709721 + layer.0.v_cache 0.00001629 0.00843177 + layer.1.k_cache 0.53789903 9.65709929 + layer.1.v_cache 0.00000607 0.00335472 + layer.2.k_cache 0.01514517 1.59971864 + layer.2.v_cache 0.00002151 0.01014118 + layer.3.k_cache 0.04182221 9.99797643 + layer.3.v_cache 0.00002017 0.01109962 + layer.4.k_cache 0.00068627 0.34435818 + layer.4.v_cache 0.00005063 0.02014607 + layer.4.output 0.04375917 176.21191452 + ------------------------------------------------------------------------------------- + TOTAL 0.06390032 75.87134263 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 322852 +BPFP 0.9729 bits/point +EBPFP 1.9458 equivalent bits/point +MSE 75.871343 +---------------------- -------------------------------------------------------- +Time: 4.033s Load: 0.010s, Pack+Encode: 2.270s, Decode+Unpack: 1.752s +---------------------- -------------------------------------------------------- +💾 Converting with 75.8713 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 371, 128) +Output shape: (1, 371, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.output: torch.Size([1, 371, 3584]) -> torch.Size([1, 1, 371, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,676B, BPFP=0.4075 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,252B, BPFP=2.0743 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,284B, BPFP=0.7700 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,196B, BPFP=2.0298 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,128B, BPFP=0.8898 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,740B, BPFP=2.0106 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,668B, BPFP=0.8283 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,904B, BPFP=2.0596 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,924B, BPFP=2.0184 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,960B, BPFP=1.9778 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,556B, BPFP=0.2260 +⌛️ [2/4] FRONTEND: Frontend time: 2.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16952868 32.62434984 + layer.0.v_cache 0.00001698 0.00816091 + layer.1.k_cache 0.82091874 12.38294674 + layer.1.v_cache 0.00000666 0.00330017 + layer.2.k_cache 0.02418445 1.61029135 + layer.2.v_cache 0.00002238 0.01012473 + layer.3.k_cache 0.03055777 9.88647527 + layer.3.v_cache 0.00002117 0.01090088 + layer.4.k_cache 0.00071103 0.35199032 + layer.4.v_cache 0.00005915 0.01949159 + layer.4.output 0.03610923 144.96891846 + ------------------------------------------------------------------------------------- + TOTAL 0.07639951 63.04061536 + (elements=3,229,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3229184 +Total Bytes 395288 +BPFP 0.9793 bits/point +EBPFP 1.9586 equivalent bits/point +MSE 63.040615 +---------------------- -------------------------------------------------------- +Time: 4.513s Load: 0.012s, Pack+Encode: 2.485s, Decode+Unpack: 2.016s +---------------------- -------------------------------------------------------- +💾 Converting with 63.0406 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,860B, BPFP=0.4200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,716B, BPFP=2.4166 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,160B, BPFP=0.7999 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,420B, BPFP=2.3237 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,976B, BPFP=1.0017 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,140B, BPFP=2.3036 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,024B, BPFP=0.9335 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,980B, BPFP=2.3638 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,280B, BPFP=2.0986 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,684B, BPFP=2.2709 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,356B, BPFP=0.2289 +⌛️ [2/4] FRONTEND: Frontend time: 2.285s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.526s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13674593 29.89581243 + layer.0.v_cache 0.00001768 0.00908337 + layer.1.k_cache 0.26206151 12.78606338 + layer.1.v_cache 0.00000624 0.00353963 + layer.2.k_cache 0.02938290 1.62368102 + layer.2.v_cache 0.00002158 0.01032395 + layer.3.k_cache 0.05653211 9.84794743 + layer.3.v_cache 0.00002129 0.01184217 + layer.4.k_cache 0.00070776 0.34530391 + layer.4.v_cache 0.00005193 0.02160230 + layer.4.output 1.40434551 247.22927588 + ------------------------------------------------------------------------------------- + TOTAL 0.60682162 105.00941946 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 258596 +BPFP 1.0903 bits/point +EBPFP 2.1806 equivalent bits/point +MSE 105.009419 +---------------------- -------------------------------------------------------- +Time: 3.821s Load: 0.010s, Pack+Encode: 2.285s, Decode+Unpack: 1.526s +---------------------- -------------------------------------------------------- +💾 Converting with 105.0094 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 193, 128) +Output shape: (1, 193, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.output: torch.Size([1, 193, 3584]) -> torch.Size([1, 1, 193, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,136B, BPFP=0.4158 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,248B, BPFP=2.6108 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,616B, BPFP=0.7785 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,920B, BPFP=2.5032 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,624B, BPFP=1.0220 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,780B, BPFP=2.4919 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,904B, BPFP=0.9637 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,216B, BPFP=2.5272 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,244B, BPFP=2.0437 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,212B, BPFP=2.4459 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,644B, BPFP=0.2272 +⌛️ [2/4] FRONTEND: Frontend time: 2.219s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.710s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14305212 29.36023518 + layer.0.v_cache 0.00001770 0.00914414 + layer.1.k_cache 0.14179654 11.86036295 + layer.1.v_cache 0.00000608 0.00352834 + layer.2.k_cache 0.01555968 1.54695153 + layer.2.v_cache 0.00002047 0.01075628 + layer.3.k_cache 0.02655400 7.66937287 + layer.3.v_cache 0.00002031 0.01271506 + layer.4.k_cache 0.00069483 0.35224847 + layer.4.v_cache 0.00005004 0.02175294 + layer.4.output 1.58615237 279.37319578 + ------------------------------------------------------------------------------------- + TOTAL 0.67240225 118.02702578 + (elements=1,679,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1679872 +Total Bytes 239544 +BPFP 1.1408 bits/point +EBPFP 2.2815 equivalent bits/point +MSE 118.027026 +---------------------- -------------------------------------------------------- +Time: 3.935s Load: 0.006s, Pack+Encode: 2.219s, Decode+Unpack: 1.710s +---------------------- -------------------------------------------------------- +💾 Converting with 118.0270 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,656B, BPFP=0.4177 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,908B, BPFP=2.0650 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,684B, BPFP=0.7332 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,888B, BPFP=2.0010 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,788B, BPFP=0.9280 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,992B, BPFP=2.0075 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,816B, BPFP=0.8670 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,572B, BPFP=2.0439 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,284B, BPFP=1.7121 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,484B, BPFP=1.9757 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,828B, BPFP=0.2405 +⌛️ [2/4] FRONTEND: Frontend time: 2.204s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.517s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10821181 35.51729182 + layer.0.v_cache 0.00001942 0.00841933 + layer.1.k_cache 0.30751975 11.85460102 + layer.1.v_cache 0.00000599 0.00325901 + layer.2.k_cache 0.01421910 1.60334992 + layer.2.v_cache 0.00002020 0.01077149 + layer.3.k_cache 0.01730646 9.66470766 + layer.3.v_cache 0.00002068 0.01227682 + layer.4.k_cache 0.00075739 0.36055078 + layer.4.v_cache 0.00005209 0.02183289 + layer.4.output 1.22958295 214.87166523 + ------------------------------------------------------------------------------------- + TOTAL 0.53265962 91.95051279 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 261900 +BPFP 0.9667 bits/point +EBPFP 1.9335 equivalent bits/point +MSE 91.950513 +---------------------- -------------------------------------------------------- +Time: 3.730s Load: 0.009s, Pack+Encode: 2.204s, Decode+Unpack: 1.517s +---------------------- -------------------------------------------------------- +💾 Converting with 91.9505 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 202, 128) +Output shape: (1, 202, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.output: torch.Size([1, 202, 3584]) -> torch.Size([1, 1, 202, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,656B, BPFP=0.4375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,532B, BPFP=2.5164 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,808B, BPFP=0.8360 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,588B, BPFP=2.4434 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,272B, BPFP=1.1040 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,284B, BPFP=2.4972 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,724B, BPFP=0.9842 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,120B, BPFP=2.4845 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,656B, BPFP=2.0619 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,524B, BPFP=2.3611 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,696B, BPFP=0.2397 +⌛️ [2/4] FRONTEND: Frontend time: 2.149s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.946s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15282421 28.57180635 + layer.0.v_cache 0.00001801 0.00926351 + layer.1.k_cache 0.08597872 12.15635878 + layer.1.v_cache 0.00000649 0.00362914 + layer.2.k_cache 0.01425038 1.69960233 + layer.2.v_cache 0.00002047 0.01046732 + layer.3.k_cache 0.05704349 8.75731818 + layer.3.v_cache 0.00002256 0.01229564 + layer.4.k_cache 0.00069767 0.35229001 + layer.4.v_cache 0.00005056 0.02094911 + layer.4.output 1.51553867 266.11845827 + ------------------------------------------------------------------------------------- + TOTAL 0.64233431 112.61312872 + (elements=1,758,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1758208 +Total Bytes 250860 +BPFP 1.1414 bits/point +EBPFP 2.2829 equivalent bits/point +MSE 112.613129 +---------------------- -------------------------------------------------------- +Time: 4.103s Load: 0.008s, Pack+Encode: 2.149s, Decode+Unpack: 1.946s +---------------------- -------------------------------------------------------- +💾 Converting with 112.6131 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,640B, BPFP=0.4218 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,828B, BPFP=2.2542 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,716B, BPFP=0.7573 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,812B, BPFP=2.1981 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,184B, BPFP=0.9488 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,392B, BPFP=2.2853 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,012B, BPFP=0.8841 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,432B, BPFP=2.2323 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,896B, BPFP=1.7610 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,120B, BPFP=2.1599 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,628B, BPFP=0.2100 +⌛️ [2/4] FRONTEND: Frontend time: 2.512s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.727s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14837397 31.49074164 + layer.0.v_cache 0.00001603 0.00745191 + layer.1.k_cache 0.60911565 12.77043880 + layer.1.v_cache 0.00000597 0.00291929 + layer.2.k_cache 0.02477714 1.60952198 + layer.2.v_cache 0.00001938 0.00968415 + layer.3.k_cache 0.01052417 9.73128140 + layer.3.v_cache 0.00002039 0.01043062 + layer.4.k_cache 0.00073933 0.36311297 + layer.4.v_cache 0.00005963 0.02382426 + layer.4.output 0.00470031 193.83504228 + ------------------------------------------------------------------------------------- + TOTAL 0.04862081 83.10968841 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 314660 +BPFP 1.0219 bits/point +EBPFP 2.0439 equivalent bits/point +MSE 83.109688 +---------------------- -------------------------------------------------------- +Time: 4.251s Load: 0.012s, Pack+Encode: 2.512s, Decode+Unpack: 1.727s +---------------------- -------------------------------------------------------- +💾 Converting with 83.1097 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 210, 128) +Output shape: (1, 210, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.output: torch.Size([1, 210, 3584]) -> torch.Size([1, 1, 210, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,892B, BPFP=0.4384 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,108B, BPFP=2.4634 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,652B, BPFP=0.7926 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,088B, BPFP=2.3875 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,356B, BPFP=0.9938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,888B, BPFP=2.3726 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,564B, BPFP=0.9348 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,600B, BPFP=2.4256 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,324B, BPFP=2.0330 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,144B, BPFP=2.3173 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,548B, BPFP=0.2397 +⌛️ [2/4] FRONTEND: Frontend time: 2.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.566s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11415643 29.80416667 + layer.0.v_cache 0.00001882 0.00856504 + layer.1.k_cache 0.27145578 12.40533622 + layer.1.v_cache 0.00000623 0.00344226 + layer.2.k_cache 0.01201470 1.54838475 + layer.2.v_cache 0.00002208 0.01043230 + layer.3.k_cache 0.00887682 9.00014416 + layer.3.v_cache 0.00002165 0.01189865 + layer.4.k_cache 0.00072611 0.36960805 + layer.4.v_cache 0.00005233 0.02078670 + layer.4.output 1.45781997 252.65846088 + ------------------------------------------------------------------------------------- + TOTAL 0.62424063 107.16423476 + (elements=1,827,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1827840 +Total Bytes 253164 +BPFP 1.1080 bits/point +EBPFP 2.2161 equivalent bits/point +MSE 107.164235 +---------------------- -------------------------------------------------------- +Time: 3.829s Load: 0.007s, Pack+Encode: 2.255s, Decode+Unpack: 1.566s +---------------------- -------------------------------------------------------- +💾 Converting with 107.1642 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 165, 128) +Output shape: (1, 165, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.output: torch.Size([1, 165, 3584]) -> torch.Size([1, 1, 165, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,524B, BPFP=0.4284 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,048B, BPFP=2.3720 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,436B, BPFP=0.7989 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,912B, BPFP=2.2644 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,956B, BPFP=1.0375 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,848B, BPFP=2.2583 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,596B, BPFP=1.0034 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,284B, BPFP=2.2996 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,560B, BPFP=2.1364 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,444B, BPFP=2.2201 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,392B, BPFP=0.2353 +⌛️ [2/4] FRONTEND: Frontend time: 2.077s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.394s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10245823 30.40272550 + layer.0.v_cache 0.00002070 0.00982097 + layer.1.k_cache 0.13127365 11.35032256 + layer.1.v_cache 0.00000615 0.00371441 + layer.2.k_cache 0.01458293 1.63085198 + layer.2.v_cache 0.00002074 0.01082371 + layer.3.k_cache 0.02375244 9.82116995 + layer.3.v_cache 0.00002214 0.01322162 + layer.4.k_cache 0.00067489 0.33746934 + layer.4.v_cache 0.00005196 0.02247941 + layer.4.output 0.00973386 331.32064394 + ------------------------------------------------------------------------------------- + TOTAL 0.02005887 139.57924159 + (elements=1,436,160) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1436160 +Total Bytes 195000 +BPFP 1.0862 bits/point +EBPFP 2.1725 equivalent bits/point +MSE 139.579242 +---------------------- -------------------------------------------------------- +Time: 3.476s Load: 0.006s, Pack+Encode: 2.077s, Decode+Unpack: 1.394s +---------------------- -------------------------------------------------------- +💾 Converting with 139.5792 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,936B, BPFP=0.4214 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,780B, BPFP=2.1158 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,220B, BPFP=0.7872 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,008B, BPFP=2.0499 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,520B, BPFP=0.9836 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,780B, BPFP=2.0304 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,772B, BPFP=0.9197 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,372B, BPFP=2.0809 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,160B, BPFP=2.0628 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,520B, BPFP=2.0082 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,964B, BPFP=0.2313 +⌛️ [2/4] FRONTEND: Frontend time: 2.051s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.393s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12273688 31.98309159 + layer.0.v_cache 0.00001959 0.00929580 + layer.1.k_cache 0.11126613 12.95509680 + layer.1.v_cache 0.00000674 0.00383090 + layer.2.k_cache 0.01416154 1.67548875 + layer.2.v_cache 0.00002164 0.01075830 + layer.3.k_cache 0.04491856 9.71030047 + layer.3.v_cache 0.00002114 0.01281954 + layer.4.k_cache 0.00068813 0.36314309 + layer.4.v_cache 0.00006151 0.02174276 + layer.4.output 0.00884247 298.77254098 + ------------------------------------------------------------------------------------- + TOTAL 0.02092936 126.36196205 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 200032 +BPFP 1.0047 bits/point +EBPFP 2.0093 equivalent bits/point +MSE 126.361962 +---------------------- -------------------------------------------------------- +Time: 3.451s Load: 0.006s, Pack+Encode: 2.051s, Decode+Unpack: 1.393s +---------------------- -------------------------------------------------------- +💾 Converting with 126.3620 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 257, 128) +Output shape: (1, 257, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.output: torch.Size([1, 257, 3584]) -> torch.Size([1, 1, 257, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,788B, BPFP=0.4127 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,872B, BPFP=2.4241 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,692B, BPFP=0.8324 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,608B, BPFP=2.3473 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,412B, BPFP=0.9370 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,188B, BPFP=2.3217 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,884B, BPFP=0.9049 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,400B, BPFP=2.3954 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,824B, BPFP=1.9348 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,868B, BPFP=2.3023 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,804B, BPFP=0.2328 +⌛️ [2/4] FRONTEND: Frontend time: 2.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.938s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12694686 30.83529342 + layer.0.v_cache 0.00002055 0.00864881 + layer.1.k_cache 0.54844232 12.85308244 + layer.1.v_cache 0.00000676 0.00331234 + layer.2.k_cache 0.01835482 1.59265339 + layer.2.v_cache 0.00002099 0.00991986 + layer.3.k_cache 0.02141366 9.02152427 + layer.3.v_cache 0.00002047 0.01103692 + layer.4.k_cache 0.00070767 0.34110349 + layer.4.v_cache 0.00005046 0.01960500 + layer.4.output 0.00513101 215.17607004 + ------------------------------------------------------------------------------------- + TOTAL 0.04422951 91.81933354 + (elements=2,236,928) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2236928 +Total Bytes 303340 +BPFP 1.0848 bits/point +EBPFP 2.1697 equivalent bits/point +MSE 91.819334 +---------------------- -------------------------------------------------------- +Time: 4.113s Load: 0.008s, Pack+Encode: 2.167s, Decode+Unpack: 1.938s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,960B, BPFP=0.4258 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,752B, BPFP=2.1250 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,132B, BPFP=0.7840 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,852B, BPFP=2.0477 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,320B, BPFP=0.9718 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,692B, BPFP=2.0340 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,672B, BPFP=0.9162 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,880B, BPFP=2.1360 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,960B, BPFP=1.9712 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,588B, BPFP=2.0251 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,168B, BPFP=0.2106 +⌛️ [2/4] FRONTEND: Frontend time: 1.920s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.699s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12731336 33.36323564 + layer.0.v_cache 0.00001750 0.00928312 + layer.1.k_cache 0.10052182 13.17493212 + layer.1.v_cache 0.00000572 0.00338474 + layer.2.k_cache 0.01141533 1.74340099 + layer.2.v_cache 0.00001926 0.01044682 + layer.3.k_cache 0.03745206 9.78722340 + layer.3.v_cache 0.00002083 0.01247758 + layer.4.k_cache 0.00066429 0.34637170 + layer.4.v_cache 0.00005101 0.02178033 + layer.4.output 0.00884527 300.30781495 + ------------------------------------------------------------------------------------- + TOTAL 0.01996459 127.09572007 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 196976 +BPFP 0.9947 bits/point +EBPFP 1.9895 equivalent bits/point +MSE 127.095720 +---------------------- -------------------------------------------------------- +Time: 3.627s Load: 0.008s, Pack+Encode: 1.920s, Decode+Unpack: 1.699s +---------------------- -------------------------------------------------------- +💾 Converting with 127.0957 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,932B, BPFP=0.4175 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,560B, BPFP=2.3620 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,040B, BPFP=0.7770 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,888B, BPFP=2.2444 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,808B, BPFP=0.9718 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,880B, BPFP=2.2438 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,964B, BPFP=0.9124 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,796B, BPFP=2.3787 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,148B, BPFP=1.8404 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,404B, BPFP=2.2103 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,672B, BPFP=0.2380 +⌛️ [2/4] FRONTEND: Frontend time: 2.180s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.624s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12752426 32.39010813 + layer.0.v_cache 0.00002004 0.00899117 + layer.1.k_cache 0.30660921 12.44361891 + layer.1.v_cache 0.00000613 0.00331246 + layer.2.k_cache 0.01808885 1.66774715 + layer.2.v_cache 0.00002107 0.01014249 + layer.3.k_cache 0.01678786 10.01465174 + layer.3.v_cache 0.00002050 0.01116597 + layer.4.k_cache 0.00068290 0.35079351 + layer.4.v_cache 0.00004851 0.01933812 + layer.4.output 1.37905047 239.63982062 + ------------------------------------------------------------------------------------- + TOTAL 0.59548016 102.02344788 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 256092 +BPFP 1.0603 bits/point +EBPFP 2.1205 equivalent bits/point +MSE 102.023448 +---------------------- -------------------------------------------------------- +Time: 3.813s Load: 0.009s, Pack+Encode: 2.180s, Decode+Unpack: 1.624s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0234 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,876B, BPFP=0.4231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,216B, BPFP=2.3917 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,196B, BPFP=0.8062 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,296B, BPFP=2.3255 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,164B, BPFP=1.0199 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,036B, BPFP=2.3067 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,080B, BPFP=0.9418 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,440B, BPFP=2.4078 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,192B, BPFP=1.9579 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,804B, BPFP=2.2900 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,136B, BPFP=0.2586 +⌛️ [2/4] FRONTEND: Frontend time: 2.197s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.613s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12766143 30.47447437 + layer.0.v_cache 0.00001653 0.00870798 + layer.1.k_cache 0.34631467 12.19605842 + layer.1.v_cache 0.00000659 0.00341118 + layer.2.k_cache 0.01786991 1.61590857 + layer.2.v_cache 0.00002041 0.00977889 + layer.3.k_cache 0.06874437 8.80218478 + layer.3.v_cache 0.00002276 0.01189947 + layer.4.k_cache 0.00067882 0.35054108 + layer.4.v_cache 0.00005621 0.02194899 + layer.4.output 1.41085444 248.28966425 + ------------------------------------------------------------------------------------- + TOTAL 0.61396310 105.38368021 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 259436 +BPFP 1.0989 bits/point +EBPFP 2.1977 equivalent bits/point +MSE 105.383680 +---------------------- -------------------------------------------------------- +Time: 3.818s Load: 0.008s, Pack+Encode: 2.197s, Decode+Unpack: 1.613s +---------------------- -------------------------------------------------------- +💾 Converting with 105.3837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,044B, BPFP=0.4179 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,040B, BPFP=2.2843 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,288B, BPFP=0.7804 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,232B, BPFP=2.2284 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,564B, BPFP=1.0069 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,160B, BPFP=2.2235 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,332B, BPFP=0.9217 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,196B, BPFP=2.2951 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,280B, BPFP=1.8861 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,460B, BPFP=2.1751 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,852B, BPFP=0.2751 +⌛️ [2/4] FRONTEND: Frontend time: 2.172s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.632s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12070838 33.98764605 + layer.0.v_cache 0.00001763 0.00877665 + layer.1.k_cache 0.31015487 11.19430353 + layer.1.v_cache 0.00000677 0.00344189 + layer.2.k_cache 0.03544329 1.61952844 + layer.2.v_cache 0.00002212 0.01008482 + layer.3.k_cache 0.05682086 9.75022686 + layer.3.v_cache 0.00002136 0.01178344 + layer.4.k_cache 0.00073636 0.35520432 + layer.4.v_cache 0.00005031 0.01998582 + layer.4.output 1.35470685 239.05351217 + ------------------------------------------------------------------------------------- + TOTAL 0.58864294 101.78444512 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 262448 +BPFP 1.0673 bits/point +EBPFP 2.1347 equivalent bits/point +MSE 101.784445 +---------------------- -------------------------------------------------------- +Time: 3.811s Load: 0.008s, Pack+Encode: 2.172s, Decode+Unpack: 1.632s +---------------------- -------------------------------------------------------- +💾 Converting with 101.7844 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,852B, BPFP=0.4293 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,272B, BPFP=2.4407 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,796B, BPFP=0.7920 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,596B, BPFP=2.3911 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,208B, BPFP=1.0423 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,184B, BPFP=2.3609 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,376B, BPFP=0.9812 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,148B, BPFP=2.4316 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,548B, BPFP=2.0942 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,744B, BPFP=2.3286 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,496B, BPFP=0.2672 +⌛️ [2/4] FRONTEND: Frontend time: 2.058s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.798s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13227103 30.00507537 + layer.0.v_cache 0.00002050 0.00874113 + layer.1.k_cache 0.22864332 8.90138610 + layer.1.v_cache 0.00000662 0.00349909 + layer.2.k_cache 0.02950178 1.60691611 + layer.2.v_cache 0.00002095 0.00998852 + layer.3.k_cache 0.01835468 9.88533758 + layer.3.v_cache 0.00002123 0.01195400 + layer.4.k_cache 0.00070999 0.36755830 + layer.4.v_cache 0.00005457 0.02024042 + layer.4.output 1.43734575 252.05879024 + ------------------------------------------------------------------------------------- + TOTAL 0.61594264 106.77836637 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 261220 +BPFP 1.1272 bits/point +EBPFP 2.2544 equivalent bits/point +MSE 106.778366 +---------------------- -------------------------------------------------------- +Time: 3.863s Load: 0.007s, Pack+Encode: 2.058s, Decode+Unpack: 1.798s +---------------------- -------------------------------------------------------- +💾 Converting with 106.7784 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,964B, BPFP=0.4198 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,484B, BPFP=2.3567 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,384B, BPFP=0.8012 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,012B, BPFP=2.2531 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,032B, BPFP=0.9876 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,060B, BPFP=2.2565 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,956B, BPFP=0.9119 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,396B, BPFP=2.2801 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,748B, BPFP=2.1641 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,324B, BPFP=2.2047 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,820B, BPFP=0.2294 +⌛️ [2/4] FRONTEND: Frontend time: 2.146s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.506s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12518616 31.51811260 + layer.0.v_cache 0.00001676 0.00860175 + layer.1.k_cache 0.33007166 12.30764138 + layer.1.v_cache 0.00000630 0.00345947 + layer.2.k_cache 0.01131002 1.59534131 + layer.2.v_cache 0.00002125 0.01059307 + layer.3.k_cache 0.02232737 10.06527353 + layer.3.v_cache 0.00001983 0.01135268 + layer.4.k_cache 0.00071526 0.34611556 + layer.4.v_cache 0.00005006 0.01973870 + layer.4.output 1.37904434 240.68388031 + ------------------------------------------------------------------------------------- + TOTAL 0.59664912 102.39255248 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 259180 +BPFP 1.0730 bits/point +EBPFP 2.1461 equivalent bits/point +MSE 102.392552 +---------------------- -------------------------------------------------------- +Time: 3.661s Load: 0.009s, Pack+Encode: 2.146s, Decode+Unpack: 1.506s +---------------------- -------------------------------------------------------- +💾 Converting with 102.3926 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,508B, BPFP=0.4168 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,776B, BPFP=2.0989 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,608B, BPFP=0.7433 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,088B, BPFP=2.0548 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,748B, BPFP=0.9444 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,732B, BPFP=2.0320 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,508B, BPFP=0.8650 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,388B, BPFP=2.0740 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,780B, BPFP=1.8430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,344B, BPFP=2.0072 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,128B, BPFP=0.2573 +⌛️ [2/4] FRONTEND: Frontend time: 2.390s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.623s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14655781 32.65691638 + layer.0.v_cache 0.00001709 0.00818723 + layer.1.k_cache 0.36683986 11.31887067 + layer.1.v_cache 0.00000651 0.00334900 + layer.2.k_cache 0.01755252 1.60211582 + layer.2.v_cache 0.00002060 0.01054442 + layer.3.k_cache 0.02713263 9.44823118 + layer.3.v_cache 0.00001997 0.01077240 + layer.4.k_cache 0.00071102 0.35402573 + layer.4.v_cache 0.00005162 0.01991652 + layer.4.output 1.25479176 220.12509148 + ------------------------------------------------------------------------------------- + TOTAL 0.54955600 93.90050410 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 263608 +BPFP 0.9930 bits/point +EBPFP 1.9860 equivalent bits/point +MSE 93.900504 +---------------------- -------------------------------------------------------- +Time: 4.021s Load: 0.008s, Pack+Encode: 2.390s, Decode+Unpack: 1.623s +---------------------- -------------------------------------------------------- +💾 Converting with 93.9005 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 155, 128) +Output shape: (1, 155, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.output: torch.Size([1, 155, 3584]) -> torch.Size([1, 1, 155, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,292B, BPFP=0.4327 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,424B, BPFP=2.5629 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,692B, BPFP=0.7754 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,208B, BPFP=2.4403 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,500B, BPFP=1.0585 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,956B, BPFP=2.4149 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,240B, BPFP=1.0323 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,564B, BPFP=2.4762 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,332B, BPFP=2.3520 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,732B, BPFP=2.3923 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,536B, BPFP=0.2093 +⌛️ [2/4] FRONTEND: Frontend time: 2.010s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.559s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102877 28.68066406 + layer.0.v_cache 0.00001822 0.00912485 + layer.1.k_cache 0.12773433 12.11570218 + layer.1.v_cache 0.00000585 0.00357835 + layer.2.k_cache 0.00931232 1.65995562 + layer.2.v_cache 0.00002110 0.01117429 + layer.3.k_cache 0.01537714 9.82759498 + layer.3.v_cache 0.00002091 0.01341065 + layer.4.k_cache 0.00066471 0.36540227 + layer.4.v_cache 0.00005180 0.02258919 + layer.4.output 0.01742802 352.07759217 + ------------------------------------------------------------------------------------- + TOTAL 0.02389596 148.07366715 + (elements=1,349,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1349120 +Total Bytes 192476 +BPFP 1.1413 bits/point +EBPFP 2.2827 equivalent bits/point +MSE 148.073667 +---------------------- -------------------------------------------------------- +Time: 3.574s Load: 0.006s, Pack+Encode: 2.010s, Decode+Unpack: 1.559s +---------------------- -------------------------------------------------------- +💾 Converting with 148.0737 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,840B, BPFP=0.4244 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,732B, BPFP=2.5241 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,984B, BPFP=0.7983 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,244B, BPFP=2.3433 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,072B, BPFP=1.0227 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,896B, BPFP=2.3180 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,228B, BPFP=0.9613 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,740B, BPFP=2.3794 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,128B, BPFP=2.1895 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,620B, BPFP=2.2980 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,404B, BPFP=0.2222 +⌛️ [2/4] FRONTEND: Frontend time: 2.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.640s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11659910 31.19280523 + layer.0.v_cache 0.00001673 0.00884796 + layer.1.k_cache 0.29571861 10.96760651 + layer.1.v_cache 0.00000615 0.00341413 + layer.2.k_cache 0.04707494 1.65372868 + layer.2.v_cache 0.00002002 0.00980872 + layer.3.k_cache 0.03274950 9.48147768 + layer.3.v_cache 0.00002078 0.01132418 + layer.4.k_cache 0.00068608 0.33734024 + layer.4.v_cache 0.00005136 0.01968898 + layer.4.output 1.42391751 251.01465947 + ------------------------------------------------------------------------------------- + TOTAL 0.61531564 106.51697992 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 258888 +BPFP 1.1067 bits/point +EBPFP 2.2135 equivalent bits/point +MSE 106.516980 +---------------------- -------------------------------------------------------- +Time: 3.777s Load: 0.008s, Pack+Encode: 2.129s, Decode+Unpack: 1.640s +---------------------- -------------------------------------------------------- +💾 Converting with 106.5170 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,596B, BPFP=0.4156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,556B, BPFP=2.1142 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,836B, BPFP=0.7457 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,052B, BPFP=2.0194 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,944B, BPFP=0.9415 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,964B, BPFP=2.0139 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,356B, BPFP=0.8415 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,736B, BPFP=2.0625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,324B, BPFP=1.9105 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,568B, BPFP=1.9889 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,184B, BPFP=0.2447 +⌛️ [2/4] FRONTEND: Frontend time: 2.356s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.637s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10631144 33.76471538 + layer.0.v_cache 0.00001780 0.00879258 + layer.1.k_cache 0.46192560 11.57050939 + layer.1.v_cache 0.00000619 0.00346627 + layer.2.k_cache 0.02890189 1.62374558 + layer.2.v_cache 0.00002115 0.01002462 + layer.3.k_cache 0.04994748 9.75861186 + layer.3.v_cache 0.00002140 0.01151536 + layer.4.k_cache 0.00068365 0.35728233 + layer.4.v_cache 0.00005040 0.01952909 + layer.4.output 1.23456514 217.69963278 + ------------------------------------------------------------------------------------- + TOTAL 0.54646135 93.00150717 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 266116 +BPFP 0.9863 bits/point +EBPFP 1.9725 equivalent bits/point +MSE 93.001507 +---------------------- -------------------------------------------------------- +Time: 4.001s Load: 0.008s, Pack+Encode: 2.356s, Decode+Unpack: 1.637s +---------------------- -------------------------------------------------------- +💾 Converting with 93.0015 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 194, 128) +Output shape: (1, 194, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.output: torch.Size([1, 194, 3584]) -> torch.Size([1, 1, 194, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,276B, BPFP=0.4249 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,952B, BPFP=2.5735 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,976B, BPFP=0.8035 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,692B, BPFP=2.4720 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,640B, BPFP=1.0180 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,444B, BPFP=2.4520 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,776B, BPFP=0.9485 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,240B, BPFP=2.5161 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,264B, BPFP=2.1153 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,020B, BPFP=2.4178 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,300B, BPFP=0.1875 +⌛️ [2/4] FRONTEND: Frontend time: 2.139s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.683s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15027005 27.29418945 + layer.0.v_cache 0.00001782 0.00933853 + layer.1.k_cache 0.21402168 11.05854341 + layer.1.v_cache 0.00000588 0.00358847 + layer.2.k_cache 0.02177384 1.60651107 + layer.2.v_cache 0.00002019 0.01098175 + layer.3.k_cache 0.04024272 8.37642772 + layer.3.v_cache 0.00002076 0.01286844 + layer.4.k_cache 0.00069591 0.35001857 + layer.4.v_cache 0.00005073 0.02158809 + layer.4.output 1.57794378 278.22072901 + ------------------------------------------------------------------------------------- + TOTAL 0.67486624 117.42877403 + (elements=1,688,576) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1688576 +Total Bytes 236580 +BPFP 1.1208 bits/point +EBPFP 2.2417 equivalent bits/point +MSE 117.428774 +---------------------- -------------------------------------------------------- +Time: 3.831s Load: 0.009s, Pack+Encode: 2.139s, Decode+Unpack: 1.683s +---------------------- -------------------------------------------------------- +💾 Converting with 117.4288 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 258, 128) +Output shape: (1, 258, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.output: torch.Size([1, 258, 3584]) -> torch.Size([1, 1, 258, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,916B, BPFP=0.4188 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,820B, BPFP=2.5327 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,164B, BPFP=0.7972 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,312B, BPFP=2.3203 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,740B, BPFP=0.9532 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,992B, BPFP=2.3009 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,740B, BPFP=0.8927 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,056B, BPFP=2.3653 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,268B, BPFP=1.8331 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,840B, BPFP=2.2917 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,228B, BPFP=0.1837 +⌛️ [2/4] FRONTEND: Frontend time: 2.278s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.673s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09035230 31.32088595 + layer.0.v_cache 0.00001726 0.00854238 + layer.1.k_cache 0.42455262 12.53239311 + layer.1.v_cache 0.00000600 0.00328070 + layer.2.k_cache 0.01891134 1.62929328 + layer.2.v_cache 0.00001996 0.01012313 + layer.3.k_cache 0.02093011 8.12159812 + layer.3.v_cache 0.00001928 0.01139116 + layer.4.k_cache 0.00072082 0.33194641 + layer.4.v_cache 0.00004951 0.02085344 + layer.4.output 0.00505962 214.26351398 + ------------------------------------------------------------------------------------- + TOTAL 0.03476450 91.40205327 + (elements=2,245,632) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2245632 +Total Bytes 297076 +BPFP 1.0583 bits/point +EBPFP 2.1166 equivalent bits/point +MSE 91.402053 +---------------------- -------------------------------------------------------- +Time: 3.961s Load: 0.010s, Pack+Encode: 2.278s, Decode+Unpack: 1.673s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4021 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,240B, BPFP=0.4239 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,876B, BPFP=2.2334 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,104B, BPFP=0.7543 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,148B, BPFP=2.1840 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,424B, BPFP=0.9799 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,948B, BPFP=2.1704 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,200B, BPFP=0.8967 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,528B, BPFP=2.2098 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,560B, BPFP=2.0082 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,268B, BPFP=2.1242 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,668B, BPFP=0.2782 +⌛️ [2/4] FRONTEND: Frontend time: 2.373s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.587s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15008528 35.18906462 + layer.0.v_cache 0.00001847 0.00864468 + layer.1.k_cache 0.25612781 11.78679093 + layer.1.v_cache 0.00000678 0.00333728 + layer.2.k_cache 0.03931687 1.59722608 + layer.2.v_cache 0.00002107 0.01014500 + layer.3.k_cache 0.03786138 9.54032460 + layer.3.v_cache 0.00002177 0.01163159 + layer.4.k_cache 0.00069674 0.37661846 + layer.4.v_cache 0.00005115 0.02130936 + layer.4.output 1.33117570 234.78272516 + ------------------------------------------------------------------------------------- + TOTAL 0.57661395 100.11906875 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 263964 +BPFP 1.0548 bits/point +EBPFP 2.1097 equivalent bits/point +MSE 100.119069 +---------------------- -------------------------------------------------------- +Time: 3.968s Load: 0.008s, Pack+Encode: 2.373s, Decode+Unpack: 1.587s +---------------------- -------------------------------------------------------- +💾 Converting with 100.1191 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,884B, BPFP=0.4240 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,832B, BPFP=2.1556 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,872B, BPFP=0.7701 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,076B, BPFP=2.0899 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,532B, BPFP=1.0010 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,932B, BPFP=2.0774 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,628B, BPFP=0.9226 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,544B, BPFP=2.1306 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,296B, BPFP=2.0222 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,680B, BPFP=2.0556 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,960B, BPFP=0.2599 +⌛️ [2/4] FRONTEND: Frontend time: 1.968s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.482s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15161699 31.13350152 + layer.0.v_cache 0.00001786 0.00916335 + layer.1.k_cache 0.08662784 11.32394002 + layer.1.v_cache 0.00000652 0.00353450 + layer.2.k_cache 0.02063684 1.62674289 + layer.2.v_cache 0.00002327 0.01061274 + layer.3.k_cache 0.08087930 9.69145508 + layer.3.v_cache 0.00002116 0.01199816 + layer.4.k_cache 0.00074161 0.37399373 + layer.4.v_cache 0.00006096 0.02178463 + layer.4.output 0.00898438 302.09290675 + ------------------------------------------------------------------------------------- + TOTAL 0.02373665 127.57982787 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 201236 +BPFP 1.0276 bits/point +EBPFP 2.0551 equivalent bits/point +MSE 127.579828 +---------------------- -------------------------------------------------------- +Time: 3.456s Load: 0.006s, Pack+Encode: 1.968s, Decode+Unpack: 1.482s +---------------------- -------------------------------------------------------- +💾 Converting with 127.5798 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,960B, BPFP=0.4233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,584B, BPFP=2.3852 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,088B, BPFP=0.7875 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,360B, BPFP=2.2983 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,100B, BPFP=1.0014 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,004B, BPFP=2.2730 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,936B, BPFP=0.9187 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,844B, BPFP=2.3327 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,580B, BPFP=2.1719 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,768B, BPFP=2.2563 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,144B, BPFP=0.2247 +⌛️ [2/4] FRONTEND: Frontend time: 2.269s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.550s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14327769 30.31520774 + layer.0.v_cache 0.00001763 0.00870653 + layer.1.k_cache 0.34925704 12.48164950 + layer.1.v_cache 0.00000642 0.00353534 + layer.2.k_cache 0.02352169 1.62739452 + layer.2.v_cache 0.00002271 0.01031965 + layer.3.k_cache 0.00845094 9.29314575 + layer.3.v_cache 0.00002026 0.01150790 + layer.4.k_cache 0.00069169 0.34159425 + layer.4.v_cache 0.00005086 0.02028183 + layer.4.output 1.39158238 240.82938312 + ------------------------------------------------------------------------------------- + TOTAL 0.60390550 102.34817793 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 259368 +BPFP 1.0836 bits/point +EBPFP 2.1672 equivalent bits/point +MSE 102.348178 +---------------------- -------------------------------------------------------- +Time: 3.828s Load: 0.009s, Pack+Encode: 2.269s, Decode+Unpack: 1.550s +---------------------- -------------------------------------------------------- +💾 Converting with 102.3482 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,456B, BPFP=0.4236 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,512B, BPFP=2.3586 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,312B, BPFP=0.7564 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,288B, BPFP=2.2891 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,112B, BPFP=0.9723 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,012B, BPFP=2.2734 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,928B, BPFP=0.9050 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,920B, BPFP=2.3250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,332B, BPFP=1.9507 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,456B, BPFP=2.2418 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,932B, BPFP=0.2024 +⌛️ [2/4] FRONTEND: Frontend time: 2.396s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.676s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12716211 30.65626776 + layer.0.v_cache 0.00001677 0.00817768 + layer.1.k_cache 0.51313377 10.30870739 + layer.1.v_cache 0.00000586 0.00319981 + layer.2.k_cache 0.02995576 1.59713312 + layer.2.v_cache 0.00002133 0.01024229 + layer.3.k_cache 0.03339291 9.66575018 + layer.3.v_cache 0.00002012 0.01086812 + layer.4.k_cache 0.00068476 0.34686174 + layer.4.v_cache 0.00005875 0.01972610 + layer.4.output 0.00481428 199.65894481 + ------------------------------------------------------------------------------------- + TOTAL 0.04342071 85.30820870 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 315260 +BPFP 1.0537 bits/point +EBPFP 2.1074 equivalent bits/point +MSE 85.308209 +---------------------- -------------------------------------------------------- +Time: 4.085s Load: 0.013s, Pack+Encode: 2.396s, Decode+Unpack: 1.676s +---------------------- -------------------------------------------------------- +💾 Converting with 85.3082 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,696B, BPFP=0.4119 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,956B, BPFP=2.0273 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,864B, BPFP=0.7298 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,688B, BPFP=2.0108 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,120B, BPFP=0.8686 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,056B, BPFP=1.9719 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,292B, BPFP=0.8177 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,152B, BPFP=2.1009 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,400B, BPFP=2.0546 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,872B, BPFP=1.9606 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,388B, BPFP=0.2319 +⌛️ [2/4] FRONTEND: Frontend time: 2.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.785s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13396681 34.88684563 + layer.0.v_cache 0.00001641 0.00801183 + layer.1.k_cache 0.47111181 11.54049659 + layer.1.v_cache 0.00000608 0.00331531 + layer.2.k_cache 0.02933135 1.67265092 + layer.2.v_cache 0.00002255 0.01043730 + layer.3.k_cache 0.03744583 10.25158980 + layer.3.v_cache 0.00002260 0.01118794 + layer.4.k_cache 0.00071777 0.38709773 + layer.4.v_cache 0.00004917 0.01930042 + layer.4.output 1.20538323 211.72122821 + ------------------------------------------------------------------------------------- + TOTAL 0.53590429 90.63761946 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 269484 +BPFP 0.9751 bits/point +EBPFP 1.9503 equivalent bits/point +MSE 90.637619 +---------------------- -------------------------------------------------------- +Time: 4.014s Load: 0.009s, Pack+Encode: 2.220s, Decode+Unpack: 1.785s +---------------------- -------------------------------------------------------- +💾 Converting with 90.6376 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,508B, BPFP=0.4073 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,136B, BPFP=2.2318 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,764B, BPFP=0.7467 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,964B, BPFP=2.1682 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,152B, BPFP=0.9306 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,896B, BPFP=2.1645 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,164B, BPFP=0.8770 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,548B, BPFP=2.1999 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,164B, BPFP=2.0163 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,052B, BPFP=2.1187 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,952B, BPFP=0.2244 +⌛️ [2/4] FRONTEND: Frontend time: 2.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.696s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14060297 32.92227851 + layer.0.v_cache 0.00001598 0.00808201 + layer.1.k_cache 0.58610762 12.46117147 + layer.1.v_cache 0.00000669 0.00320684 + layer.2.k_cache 0.02834670 1.56377909 + layer.2.v_cache 0.00002256 0.00954903 + layer.3.k_cache 0.01962393 9.90216997 + layer.3.v_cache 0.00002003 0.01056520 + layer.4.k_cache 0.00067797 0.34161544 + layer.4.v_cache 0.00005035 0.01906399 + layer.4.output 0.00463834 191.74530320 + ------------------------------------------------------------------------------------- + TOTAL 0.04752607 82.32109435 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 321300 +BPFP 1.0254 bits/point +EBPFP 2.0508 equivalent bits/point +MSE 82.321094 +---------------------- -------------------------------------------------------- +Time: 4.189s Load: 0.009s, Pack+Encode: 2.484s, Decode+Unpack: 1.696s +---------------------- -------------------------------------------------------- +💾 Converting with 82.3211 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,076B, BPFP=0.4151 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,944B, BPFP=2.1044 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,588B, BPFP=0.7498 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,780B, BPFP=2.0446 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,892B, BPFP=0.9196 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,872B, BPFP=2.0493 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,424B, BPFP=0.8442 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,444B, BPFP=2.1815 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,064B, BPFP=1.8022 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,060B, BPFP=2.0076 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,252B, BPFP=0.2001 +⌛️ [2/4] FRONTEND: Frontend time: 2.345s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.064s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12404394 35.10810290 + layer.0.v_cache 0.00001645 0.00800633 + layer.1.k_cache 0.57408182 12.09647088 + layer.1.v_cache 0.00000616 0.00313622 + layer.2.k_cache 0.03522489 1.59193461 + layer.2.v_cache 0.00002326 0.00988988 + layer.3.k_cache 0.02908128 9.68974063 + layer.3.v_cache 0.00002078 0.01072506 + layer.4.k_cache 0.00072493 0.35165719 + layer.4.v_cache 0.00004967 0.02042937 + layer.4.output 0.04390477 177.61582178 + ------------------------------------------------------------------------------------- + TOTAL 0.06297686 76.60004974 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 321396 +BPFP 0.9717 bits/point +EBPFP 1.9434 equivalent bits/point +MSE 76.600050 +---------------------- -------------------------------------------------------- +Time: 4.420s Load: 0.012s, Pack+Encode: 2.345s, Decode+Unpack: 2.064s +---------------------- -------------------------------------------------------- +💾 Converting with 76.6000 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 242, 128) +Output shape: (1, 242, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.output: torch.Size([1, 242, 3584]) -> torch.Size([1, 1, 242, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,508B, BPFP=0.4202 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,060B, BPFP=2.1346 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,208B, BPFP=0.7882 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,968B, BPFP=2.0640 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,232B, BPFP=0.9189 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,652B, BPFP=2.0436 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,284B, BPFP=0.8577 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,404B, BPFP=2.0922 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,600B, BPFP=1.8466 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,396B, BPFP=2.0271 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,560B, BPFP=0.2358 +⌛️ [2/4] FRONTEND: Frontend time: 2.232s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.450s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11466674 32.52789862 + layer.0.v_cache 0.00001794 0.00835793 + layer.1.k_cache 0.39062487 10.74927767 + layer.1.v_cache 0.00000616 0.00324345 + layer.2.k_cache 0.02414889 1.62164105 + layer.2.v_cache 0.00002046 0.01039198 + layer.3.k_cache 0.03495786 9.60954852 + layer.3.v_cache 0.00002053 0.01158130 + layer.4.k_cache 0.00069125 0.34549943 + layer.4.v_cache 0.00005034 0.02169621 + layer.4.output 1.26514320 222.46011659 + ------------------------------------------------------------------------------------- + TOTAL 0.55418868 94.83117366 + (elements=2,106,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2106368 +Total Bytes 260872 +BPFP 0.9908 bits/point +EBPFP 1.9816 equivalent bits/point +MSE 94.831174 +---------------------- -------------------------------------------------------- +Time: 3.691s Load: 0.009s, Pack+Encode: 2.232s, Decode+Unpack: 1.450s +---------------------- -------------------------------------------------------- +💾 Converting with 94.8312 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 189, 128) +Output shape: (1, 189, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.output: torch.Size([1, 189, 3584]) -> torch.Size([1, 1, 189, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,016B, BPFP=0.4147 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,600B, BPFP=2.0337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,012B, BPFP=0.7450 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,072B, BPFP=1.9901 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,240B, BPFP=0.9292 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,012B, BPFP=1.9851 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,736B, BPFP=0.8876 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,424B, BPFP=2.0192 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,464B, BPFP=1.8571 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,668B, BPFP=1.9567 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,236B, BPFP=0.2508 +⌛️ [2/4] FRONTEND: Frontend time: 1.946s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.623s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14361043 33.00030485 + layer.0.v_cache 0.00001648 0.00910412 + layer.1.k_cache 0.18932698 10.75540985 + layer.1.v_cache 0.00000663 0.00385517 + layer.2.k_cache 0.00768016 1.68757847 + layer.2.v_cache 0.00002453 0.01139330 + layer.3.k_cache 0.01421528 10.00114837 + layer.3.v_cache 0.00002024 0.01284955 + layer.4.k_cache 0.00072804 0.36949880 + layer.4.v_cache 0.00005086 0.02290489 + layer.4.output 0.00858909 289.32820767 + ------------------------------------------------------------------------------------- + TOTAL 0.02445902 122.42185300 + (elements=1,645,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1645056 +Total Bytes 200480 +BPFP 0.9749 bits/point +EBPFP 1.9499 equivalent bits/point +MSE 122.421853 +---------------------- -------------------------------------------------------- +Time: 3.576s Load: 0.007s, Pack+Encode: 1.946s, Decode+Unpack: 1.623s +---------------------- -------------------------------------------------------- +💾 Converting with 122.4219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,832B, BPFP=0.4278 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,412B, BPFP=2.4510 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,804B, BPFP=0.7925 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,464B, BPFP=2.3815 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,388B, BPFP=0.9821 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,088B, BPFP=2.3539 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,328B, BPFP=0.9777 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,792B, BPFP=2.4055 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,832B, BPFP=1.9683 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,488B, BPFP=2.3099 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,976B, BPFP=0.2093 +⌛️ [2/4] FRONTEND: Frontend time: 2.435s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.504s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16894778 30.02586974 + layer.0.v_cache 0.00001779 0.00876319 + layer.1.k_cache 0.37628758 9.76641015 + layer.1.v_cache 0.00000652 0.00338946 + layer.2.k_cache 0.02181128 1.56177574 + layer.2.v_cache 0.00002030 0.01022095 + layer.3.k_cache 0.01815432 9.72603296 + layer.3.v_cache 0.00002063 0.01169217 + layer.4.k_cache 0.00071780 0.35685805 + layer.4.v_cache 0.00005059 0.02025440 + layer.4.output 1.43727796 252.72206154 + ------------------------------------------------------------------------------------- + TOTAL 0.62629296 107.09092339 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 252404 +BPFP 1.0892 bits/point +EBPFP 2.1783 equivalent bits/point +MSE 107.090923 +---------------------- -------------------------------------------------------- +Time: 3.948s Load: 0.010s, Pack+Encode: 2.435s, Decode+Unpack: 1.504s +---------------------- -------------------------------------------------------- +💾 Converting with 107.0909 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,904B, BPFP=0.4271 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,336B, BPFP=2.4115 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,700B, BPFP=0.7740 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,316B, BPFP=2.3377 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,856B, BPFP=1.0023 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,916B, BPFP=2.3087 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,812B, BPFP=0.9268 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,960B, BPFP=2.3843 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,652B, BPFP=1.9280 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,488B, BPFP=2.2778 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,724B, BPFP=0.2038 +⌛️ [2/4] FRONTEND: Frontend time: 2.074s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.544s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09877178 30.72132252 + layer.0.v_cache 0.00001652 0.00829076 + layer.1.k_cache 0.28824160 11.14185475 + layer.1.v_cache 0.00000619 0.00339125 + layer.2.k_cache 0.02137197 1.59604970 + layer.2.v_cache 0.00001995 0.01049825 + layer.3.k_cache 0.03811082 9.44089084 + layer.3.v_cache 0.00002058 0.01157266 + layer.4.k_cache 0.00070031 0.35534608 + layer.4.v_cache 0.00004695 0.02040797 + layer.4.output 1.41729720 248.96412037 + ------------------------------------------------------------------------------------- + TOTAL 0.60990512 105.65049808 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 251664 +BPFP 1.0709 bits/point +EBPFP 2.1417 equivalent bits/point +MSE 105.650498 +---------------------- -------------------------------------------------------- +Time: 3.625s Load: 0.008s, Pack+Encode: 2.074s, Decode+Unpack: 1.544s +---------------------- -------------------------------------------------------- +💾 Converting with 105.6505 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,940B, BPFP=0.4107 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,804B, BPFP=2.2680 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,268B, BPFP=0.8482 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,064B, BPFP=2.2168 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,952B, BPFP=0.9646 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,848B, BPFP=2.2019 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,944B, BPFP=0.8949 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,524B, BPFP=2.2486 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,172B, BPFP=1.9477 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,184B, BPFP=2.1560 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,520B, BPFP=0.2422 +⌛️ [2/4] FRONTEND: Frontend time: 2.468s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.648s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12161526 31.97164080 + layer.0.v_cache 0.00001916 0.00866122 + layer.1.k_cache 0.26028682 12.01835484 + layer.1.v_cache 0.00000628 0.00337911 + layer.2.k_cache 0.01296707 1.60484030 + layer.2.v_cache 0.00002040 0.01002989 + layer.3.k_cache 0.04595671 9.82981886 + layer.3.v_cache 0.00002073 0.01116713 + layer.4.k_cache 0.00067352 0.35969007 + layer.4.v_cache 0.00005047 0.02109107 + layer.4.output 1.35465500 238.85370575 + ------------------------------------------------------------------------------------- + TOTAL 0.58377655 101.63615374 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 258220 +BPFP 1.0502 bits/point +EBPFP 2.1003 equivalent bits/point +MSE 101.636154 +---------------------- -------------------------------------------------------- +Time: 4.125s Load: 0.009s, Pack+Encode: 2.468s, Decode+Unpack: 1.648s +---------------------- -------------------------------------------------------- +💾 Converting with 101.6362 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,028B, BPFP=0.4168 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,040B, BPFP=2.2843 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,540B, BPFP=0.7978 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,120B, BPFP=2.2207 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,024B, BPFP=0.9696 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,956B, BPFP=2.2093 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,960B, BPFP=0.8960 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,288B, BPFP=2.3014 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,364B, BPFP=1.8227 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,352B, BPFP=2.1676 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,376B, BPFP=0.2210 +⌛️ [2/4] FRONTEND: Frontend time: 2.281s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.573s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10229862 31.57045510 + layer.0.v_cache 0.00001727 0.00830005 + layer.1.k_cache 0.34669285 10.89702235 + layer.1.v_cache 0.00000643 0.00328214 + layer.2.k_cache 0.01639456 1.63103627 + layer.2.v_cache 0.00002164 0.01018044 + layer.3.k_cache 0.03343437 9.45590561 + layer.3.v_cache 0.00002043 0.01110932 + layer.4.k_cache 0.00071218 0.35472370 + layer.4.v_cache 0.00005088 0.02036188 + layer.4.output 1.35462105 238.86069848 + ------------------------------------------------------------------------------------- + TOTAL 0.58717627 101.52866272 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 255048 +BPFP 1.0373 bits/point +EBPFP 2.0745 equivalent bits/point +MSE 101.528663 +---------------------- -------------------------------------------------------- +Time: 3.865s Load: 0.010s, Pack+Encode: 2.281s, Decode+Unpack: 1.573s +---------------------- -------------------------------------------------------- +💾 Converting with 101.5287 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,048B, BPFP=0.4163 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,760B, BPFP=2.3926 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,264B, BPFP=0.7753 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,000B, BPFP=2.2026 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,112B, BPFP=0.9714 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,976B, BPFP=2.2010 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,048B, BPFP=0.8981 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,408B, BPFP=2.2996 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,176B, BPFP=1.9394 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,360B, BPFP=2.1586 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,800B, BPFP=0.2045 +⌛️ [2/4] FRONTEND: Frontend time: 2.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.865s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11800324 31.32746249 + layer.0.v_cache 0.00001819 0.00837016 + layer.1.k_cache 0.35011090 12.65437539 + layer.1.v_cache 0.00000612 0.00310861 + layer.2.k_cache 0.02017111 1.60149783 + layer.2.v_cache 0.00002006 0.00978375 + layer.3.k_cache 0.03026835 9.92748701 + layer.3.v_cache 0.00001975 0.01088720 + layer.4.k_cache 0.00074701 0.34773375 + layer.4.v_cache 0.00005118 0.01888520 + layer.4.output 1.34863711 237.35686753 + ------------------------------------------------------------------------------------- + TOTAL 0.58587504 101.02398024 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 256952 +BPFP 1.0404 bits/point +EBPFP 2.0808 equivalent bits/point +MSE 101.023980 +---------------------- -------------------------------------------------------- +Time: 4.119s Load: 0.008s, Pack+Encode: 2.246s, Decode+Unpack: 1.865s +---------------------- -------------------------------------------------------- +💾 Converting with 101.0240 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,184B, BPFP=0.4219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,676B, BPFP=2.2295 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,516B, BPFP=0.7858 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,020B, BPFP=2.1848 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,664B, BPFP=1.0005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,100B, BPFP=2.1902 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,576B, BPFP=0.9263 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,704B, BPFP=2.2314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,928B, BPFP=1.9738 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,504B, BPFP=2.1496 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,028B, BPFP=0.2147 +⌛️ [2/4] FRONTEND: Frontend time: 2.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.575s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10957497 33.02275092 + layer.0.v_cache 0.00001697 0.00873896 + layer.1.k_cache 0.31262030 12.18187624 + layer.1.v_cache 0.00000629 0.00355009 + layer.2.k_cache 0.01311818 1.59874382 + layer.2.v_cache 0.00002149 0.01045813 + layer.3.k_cache 0.06227440 9.82735100 + layer.3.v_cache 0.00002139 0.01258977 + layer.4.k_cache 0.00068925 0.34524639 + layer.4.v_cache 0.00005302 0.02182382 + layer.4.output 1.33688878 235.59940346 + ------------------------------------------------------------------------------------- + TOTAL 0.57980104 100.36640902 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 257900 +BPFP 1.0351 bits/point +EBPFP 2.0702 equivalent bits/point +MSE 100.366409 +---------------------- -------------------------------------------------------- +Time: 3.812s Load: 0.007s, Pack+Encode: 2.230s, Decode+Unpack: 1.575s +---------------------- -------------------------------------------------------- +💾 Converting with 100.3664 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,832B, BPFP=0.4278 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,692B, BPFP=2.4715 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,468B, BPFP=0.7679 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,052B, BPFP=2.3512 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,020B, BPFP=1.0285 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,856B, BPFP=2.3369 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,780B, BPFP=0.9375 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,752B, BPFP=2.4026 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,852B, BPFP=1.9698 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,152B, BPFP=2.2852 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,468B, BPFP=0.2145 +⌛️ [2/4] FRONTEND: Frontend time: 2.569s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.690s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13987277 29.29888314 + layer.0.v_cache 0.00001887 0.00873442 + layer.1.k_cache 0.23498858 9.38720245 + layer.1.v_cache 0.00000603 0.00329050 + layer.2.k_cache 0.02028183 1.60727430 + layer.2.v_cache 0.00001992 0.01044875 + layer.3.k_cache 0.05552547 9.70802042 + layer.3.v_cache 0.00002044 0.01176086 + layer.4.k_cache 0.00067773 0.35980228 + layer.4.v_cache 0.00004777 0.02091929 + layer.4.output 1.43727485 253.04449614 + ------------------------------------------------------------------------------------- + TOTAL 0.61837549 107.16045938 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 251924 +BPFP 1.0871 bits/point +EBPFP 2.1742 equivalent bits/point +MSE 107.160459 +---------------------- -------------------------------------------------------- +Time: 4.266s Load: 0.008s, Pack+Encode: 2.569s, Decode+Unpack: 1.690s +---------------------- -------------------------------------------------------- +💾 Converting with 107.1605 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,888B, BPFP=0.4360 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,012B, BPFP=2.5187 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,532B, BPFP=0.7799 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,188B, BPFP=2.3836 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,808B, BPFP=1.0225 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,600B, BPFP=2.3400 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,116B, BPFP=0.9713 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,596B, BPFP=2.4138 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,772B, BPFP=1.9085 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,296B, BPFP=2.3175 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,860B, BPFP=0.2101 +⌛️ [2/4] FRONTEND: Frontend time: 2.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.548s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17004660 29.46782898 + layer.0.v_cache 0.00001740 0.00832494 + layer.1.k_cache 0.16740718 12.06882336 + layer.1.v_cache 0.00000654 0.00342916 + layer.2.k_cache 0.01191964 1.59218478 + layer.2.v_cache 0.00001991 0.01030122 + layer.3.k_cache 0.00660476 9.57192545 + layer.3.v_cache 0.00002076 0.01142345 + layer.4.k_cache 0.00070551 0.35487431 + layer.4.v_cache 0.00005179 0.02149135 + layer.4.output 1.45088344 254.63134733 + ------------------------------------------------------------------------------------- + TOTAL 0.61841083 107.97235519 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 250668 +BPFP 1.0919 bits/point +EBPFP 2.1838 equivalent bits/point +MSE 107.972355 +---------------------- -------------------------------------------------------- +Time: 3.814s Load: 0.007s, Pack+Encode: 2.259s, Decode+Unpack: 1.548s +---------------------- -------------------------------------------------------- +💾 Converting with 107.9724 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,216B, BPFP=0.4223 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,528B, BPFP=2.2098 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,228B, BPFP=0.7628 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,948B, BPFP=2.1704 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,604B, BPFP=0.9921 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,964B, BPFP=2.1715 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,468B, BPFP=0.9149 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,328B, BPFP=2.1962 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,684B, BPFP=1.9486 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,120B, BPFP=2.1141 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,248B, BPFP=0.1965 +⌛️ [2/4] FRONTEND: Frontend time: 2.376s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.476s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14712874 34.95512483 + layer.0.v_cache 0.00001826 0.00814784 + layer.1.k_cache 0.36788522 12.26669922 + layer.1.v_cache 0.00000598 0.00326807 + layer.2.k_cache 0.01156470 1.60125427 + layer.2.v_cache 0.00002030 0.01009341 + layer.3.k_cache 0.01315050 10.01778724 + layer.3.v_cache 0.00002017 0.01085422 + layer.4.k_cache 0.00068370 0.33832129 + layer.4.v_cache 0.00004817 0.01986746 + layer.4.output 1.33105469 234.32723214 + ------------------------------------------------------------------------------------- + TOTAL 0.57987697 99.97188487 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 254336 +BPFP 1.0164 bits/point +EBPFP 2.0327 equivalent bits/point +MSE 99.971885 +---------------------- -------------------------------------------------------- +Time: 3.863s Load: 0.010s, Pack+Encode: 2.376s, Decode+Unpack: 1.476s +---------------------- -------------------------------------------------------- +💾 Converting with 99.9719 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 188, 128) +Output shape: (1, 188, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.output: torch.Size([1, 188, 3584]) -> torch.Size([1, 1, 188, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,000B, BPFP=0.4156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,012B, BPFP=2.0788 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,768B, BPFP=0.7287 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,944B, BPFP=1.9900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,116B, BPFP=0.9239 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,840B, BPFP=1.9814 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,500B, BPFP=0.8727 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,408B, BPFP=2.0286 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,448B, BPFP=1.7826 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,640B, BPFP=1.9648 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,852B, BPFP=0.2238 +⌛️ [2/4] FRONTEND: Frontend time: 2.322s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.577s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005002 32.95979991 + layer.0.v_cache 0.00001663 0.00882353 + layer.1.k_cache 0.14149688 10.86339423 + layer.1.v_cache 0.00000584 0.00346449 + layer.2.k_cache 0.00881058 1.60110230 + layer.2.v_cache 0.00002285 0.01119886 + layer.3.k_cache 0.03192479 9.93645331 + layer.3.v_cache 0.00001958 0.01212396 + layer.4.k_cache 0.00066612 0.34254886 + layer.4.v_cache 0.00005162 0.02300855 + layer.4.output 0.00857985 285.01904445 + ------------------------------------------------------------------------------------- + TOTAL 0.02253670 120.64089583 + (elements=1,636,352) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1636352 +Total Bytes 196528 +BPFP 0.9608 bits/point +EBPFP 1.9216 equivalent bits/point +MSE 120.640896 +---------------------- -------------------------------------------------------- +Time: 3.907s Load: 0.007s, Pack+Encode: 2.322s, Decode+Unpack: 1.577s +---------------------- -------------------------------------------------------- +💾 Converting with 120.6409 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,516B, BPFP=0.4331 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,032B, BPFP=2.5936 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,584B, BPFP=0.8310 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,324B, BPFP=2.4595 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,532B, BPFP=1.0625 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,140B, BPFP=2.4450 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,824B, BPFP=1.0069 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,564B, BPFP=2.4783 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,472B, BPFP=2.0000 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,480B, BPFP=2.3932 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,136B, BPFP=0.2146 +⌛️ [2/4] FRONTEND: Frontend time: 2.381s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.397s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11435634 29.96756733 + layer.0.v_cache 0.00001764 0.00886364 + layer.1.k_cache 0.25025516 12.80159832 + layer.1.v_cache 0.00000690 0.00343529 + layer.2.k_cache 0.01528106 1.57600595 + layer.2.v_cache 0.00002105 0.01043529 + layer.3.k_cache 0.02533426 8.73672746 + layer.3.v_cache 0.00002127 0.01193894 + layer.4.k_cache 0.00070809 0.33925234 + layer.4.v_cache 0.00005072 0.02090156 + layer.4.output 1.53833902 271.22381999 + ------------------------------------------------------------------------------------- + TOTAL 0.65731915 114.82608624 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 244604 +BPFP 1.1297 bits/point +EBPFP 2.2595 equivalent bits/point +MSE 114.826086 +---------------------- -------------------------------------------------------- +Time: 3.784s Load: 0.007s, Pack+Encode: 2.381s, Decode+Unpack: 1.397s +---------------------- -------------------------------------------------------- +💾 Converting with 114.8261 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,944B, BPFP=0.4292 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,744B, BPFP=2.1479 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,852B, BPFP=0.7684 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,932B, BPFP=2.0774 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,220B, BPFP=0.9740 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,928B, BPFP=2.0771 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,764B, BPFP=0.9344 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,568B, BPFP=2.1326 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,440B, BPFP=1.9479 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,772B, BPFP=2.0635 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,540B, BPFP=0.2423 +⌛️ [2/4] FRONTEND: Frontend time: 2.172s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.490s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258550 31.76770562 + layer.0.v_cache 0.00001767 0.00897311 + layer.1.k_cache 0.14790291 10.59525418 + layer.1.v_cache 0.00000629 0.00338235 + layer.2.k_cache 0.00999225 1.67132738 + layer.2.v_cache 0.00002110 0.01037048 + layer.3.k_cache 0.02117847 9.89272597 + layer.3.v_cache 0.00002045 0.01220899 + layer.4.k_cache 0.00067024 0.36784168 + layer.4.v_cache 0.00006105 0.02108647 + layer.4.output 0.00897616 302.71944444 + ------------------------------------------------------------------------------------- + TOTAL 0.02207583 127.84629337 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 198704 +BPFP 1.0146 bits/point +EBPFP 2.0292 equivalent bits/point +MSE 127.846293 +---------------------- -------------------------------------------------------- +Time: 3.669s Load: 0.007s, Pack+Encode: 2.172s, Decode+Unpack: 1.490s +---------------------- -------------------------------------------------------- +💾 Converting with 127.8463 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,736B, BPFP=0.4393 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,772B, BPFP=2.5867 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,460B, BPFP=0.8778 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,500B, BPFP=2.4127 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,912B, BPFP=1.0656 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,156B, BPFP=2.3863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,568B, BPFP=0.9626 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,968B, BPFP=2.4485 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,096B, BPFP=1.9222 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,764B, BPFP=2.3563 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,888B, BPFP=0.2286 +⌛️ [2/4] FRONTEND: Frontend time: 2.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.412s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09364704 29.50250843 + layer.0.v_cache 0.00001853 0.00849035 + layer.1.k_cache 0.25163280 12.44589413 + layer.1.v_cache 0.00000592 0.00334625 + layer.2.k_cache 0.01595125 1.67766182 + layer.2.v_cache 0.00002050 0.01031189 + layer.3.k_cache 0.04325477 9.76938584 + layer.3.v_cache 0.00002016 0.01144592 + layer.4.k_cache 0.00069210 0.36188638 + layer.4.v_cache 0.00005026 0.02040849 + layer.4.output 1.50065788 264.07913165 + ------------------------------------------------------------------------------------- + TOTAL 0.64175873 111.90383889 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 248820 +BPFP 1.1211 bits/point +EBPFP 2.2421 equivalent bits/point +MSE 111.903839 +---------------------- -------------------------------------------------------- +Time: 3.917s Load: 0.010s, Pack+Encode: 2.495s, Decode+Unpack: 1.412s +---------------------- -------------------------------------------------------- +💾 Converting with 111.9038 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,048B, BPFP=0.4181 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,788B, BPFP=2.2669 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,344B, BPFP=0.7843 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,920B, BPFP=2.2069 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,080B, BPFP=0.9735 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,764B, BPFP=2.1961 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,008B, BPFP=0.8993 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,512B, BPFP=2.3861 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,756B, BPFP=1.8498 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,144B, BPFP=2.1532 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,640B, BPFP=0.2335 +⌛️ [2/4] FRONTEND: Frontend time: 2.149s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.775s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510114 34.31126201 + layer.0.v_cache 0.00001717 0.00832582 + layer.1.k_cache 0.31797088 10.59756551 + layer.1.v_cache 0.00000673 0.00335279 + layer.2.k_cache 0.01711024 1.59649334 + layer.2.v_cache 0.00002022 0.01059302 + layer.3.k_cache 0.01307247 9.64676308 + layer.3.v_cache 0.00002089 0.01139132 + layer.4.k_cache 0.00069413 0.35148661 + layer.4.v_cache 0.00005461 0.02211627 + layer.4.output 1.35466395 238.82020386 + ------------------------------------------------------------------------------------- + TOTAL 0.58392448 101.66475157 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 257004 +BPFP 1.0452 bits/point +EBPFP 2.0904 equivalent bits/point +MSE 101.664752 +---------------------- -------------------------------------------------------- +Time: 3.933s Load: 0.008s, Pack+Encode: 2.149s, Decode+Unpack: 1.775s +---------------------- -------------------------------------------------------- +💾 Converting with 101.6648 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,624B, BPFP=0.4173 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,044B, BPFP=2.0819 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,812B, BPFP=0.7442 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,288B, BPFP=2.0343 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,584B, BPFP=0.9189 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,196B, BPFP=2.0285 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,748B, BPFP=0.8662 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,476B, BPFP=2.1721 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,144B, BPFP=1.8362 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,488B, BPFP=1.9839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,052B, BPFP=0.2435 +⌛️ [2/4] FRONTEND: Frontend time: 2.308s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.592s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14860459 33.33544922 + layer.0.v_cache 0.00001629 0.00815480 + layer.1.k_cache 0.38266554 12.18940390 + layer.1.v_cache 0.00000669 0.00330431 + layer.2.k_cache 0.01628031 1.57550799 + layer.2.v_cache 0.00002106 0.01026005 + layer.3.k_cache 0.01842327 9.84810023 + layer.3.v_cache 0.00002289 0.01137382 + layer.4.k_cache 0.00072690 0.36560443 + layer.4.v_cache 0.00004923 0.02081484 + layer.4.output 1.23454798 217.57463278 + ------------------------------------------------------------------------------------- + TOTAL 0.54168545 92.96414135 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 266456 +BPFP 0.9875 bits/point +EBPFP 1.9750 equivalent bits/point +MSE 92.964141 +---------------------- -------------------------------------------------------- +Time: 3.910s Load: 0.010s, Pack+Encode: 2.308s, Decode+Unpack: 1.592s +---------------------- -------------------------------------------------------- +💾 Converting with 92.9641 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 435, 128) +Output shape: (1, 435, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.output: torch.Size([1, 435, 3584]) -> torch.Size([1, 1, 435, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,448B, BPFP=0.4112 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,908B, BPFP=2.0800 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,716B, BPFP=0.7441 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,860B, BPFP=2.0065 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,036B, BPFP=0.8634 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,568B, BPFP=1.9960 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,028B, BPFP=0.7912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,608B, BPFP=2.0333 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 48,660B, BPFP=1.7478 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,584B, BPFP=1.9606 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,192B, BPFP=0.2216 +⌛️ [2/4] FRONTEND: Frontend time: 2.723s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.976s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15864419 34.79133217 + layer.0.v_cache 0.00001607 0.00779324 + layer.1.k_cache 0.94329666 12.10976001 + layer.1.v_cache 0.00000639 0.00313881 + layer.2.k_cache 0.02505282 1.53953436 + layer.2.v_cache 0.00002131 0.00968493 + layer.3.k_cache 0.01724711 9.69658315 + layer.3.v_cache 0.00002032 0.00997469 + layer.4.k_cache 0.00074513 0.34252951 + layer.4.v_cache 0.00005173 0.01821054 + layer.4.output 0.00608059 126.00418719 + ------------------------------------------------------------------------------------- + TOTAL 0.06986270 55.32693246 + (elements=3,786,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3786240 +Total Bytes 450608 +BPFP 0.9521 bits/point +EBPFP 1.9042 equivalent bits/point +MSE 55.326932 +---------------------- -------------------------------------------------------- +Time: 4.714s Load: 0.015s, Pack+Encode: 2.723s, Decode+Unpack: 1.976s +---------------------- -------------------------------------------------------- +💾 Converting with 55.3269 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.019s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 437, 128) +Output shape: (1, 437, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.output: torch.Size([1, 437, 3584]) -> torch.Size([1, 1, 437, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,436B, BPFP=0.4089 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,424B, BPFP=2.0532 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,508B, BPFP=0.7333 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,036B, BPFP=2.0036 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,524B, BPFP=0.8769 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 59,708B, BPFP=2.1349 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,984B, BPFP=0.8218 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 61,620B, BPFP=2.2032 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 52,096B, BPFP=1.8627 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,028B, BPFP=1.9675 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 49,768B, BPFP=0.2542 +⌛️ [2/4] FRONTEND: Frontend time: 2.625s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.076s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14454122 36.72189645 + layer.0.v_cache 0.00001702 0.00786361 + layer.1.k_cache 0.94267751 12.13341140 + layer.1.v_cache 0.00000668 0.00321715 + layer.2.k_cache 0.01808156 1.57773714 + layer.2.v_cache 0.00002154 0.00966260 + layer.3.k_cache 0.03034366 9.83497435 + layer.3.v_cache 0.00002210 0.01088080 + layer.4.k_cache 0.00074160 0.36755500 + layer.4.v_cache 0.00005552 0.01976260 + layer.4.output 0.00611241 125.73209178 + ------------------------------------------------------------------------------------- + TOTAL 0.06937031 55.34185903 + (elements=3,803,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3803648 +Total Bytes 471132 +BPFP 0.9909 bits/point +EBPFP 1.9818 equivalent bits/point +MSE 55.341859 +---------------------- -------------------------------------------------------- +Time: 4.719s Load: 0.019s, Pack+Encode: 2.625s, Decode+Unpack: 2.076s +---------------------- -------------------------------------------------------- +💾 Converting with 55.3419 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,116B, BPFP=0.4117 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,020B, BPFP=2.0810 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,176B, BPFP=0.7192 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,884B, BPFP=2.0233 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,640B, BPFP=0.8949 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,792B, BPFP=2.0187 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,648B, BPFP=0.8446 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,652B, BPFP=2.0623 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,132B, BPFP=1.8837 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,136B, BPFP=1.9854 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,056B, BPFP=0.2468 +⌛️ [2/4] FRONTEND: Frontend time: 2.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.753s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12827574 32.85224736 + layer.0.v_cache 0.00001866 0.00818061 + layer.1.k_cache 0.53653341 11.66614879 + layer.1.v_cache 0.00000626 0.00314202 + layer.2.k_cache 0.01950284 1.56673788 + layer.2.v_cache 0.00002083 0.00977831 + layer.3.k_cache 0.04340698 9.91184780 + layer.3.v_cache 0.00002181 0.01081210 + layer.4.k_cache 0.00072091 0.35403393 + layer.4.v_cache 0.00005393 0.01996719 + layer.4.output 0.04342448 174.64888683 + ------------------------------------------------------------------------------------- + TOTAL 0.06073722 75.23206493 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 328252 +BPFP 0.9796 bits/point +EBPFP 1.9591 equivalent bits/point +MSE 75.232065 +---------------------- -------------------------------------------------------- +Time: 3.988s Load: 0.014s, Pack+Encode: 2.221s, Decode+Unpack: 1.753s +---------------------- -------------------------------------------------------- +💾 Converting with 75.2321 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,240B, BPFP=0.4253 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,792B, BPFP=2.3961 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,944B, BPFP=0.8191 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,672B, BPFP=2.3304 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,868B, BPFP=0.9908 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,632B, BPFP=2.3280 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,524B, BPFP=0.9119 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,332B, BPFP=2.3691 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,404B, BPFP=1.9622 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,688B, BPFP=2.2726 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,824B, BPFP=0.2503 +⌛️ [2/4] FRONTEND: Frontend time: 2.241s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.780s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702504 30.23624552 + layer.0.v_cache 0.00001641 0.00838856 + layer.1.k_cache 0.51130361 12.60249868 + layer.1.v_cache 0.00000656 0.00342658 + layer.2.k_cache 0.01810413 1.57368309 + layer.2.v_cache 0.00002122 0.00994388 + layer.3.k_cache 0.02476560 9.61090501 + layer.3.v_cache 0.00002197 0.01097941 + layer.4.k_cache 0.00071356 0.35079546 + layer.4.v_cache 0.00005290 0.02007329 + layer.4.output 0.00501064 206.05023161 + ------------------------------------------------------------------------------------- + TOTAL 0.04100620 88.04579769 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 315920 +BPFP 1.0916 bits/point +EBPFP 2.1832 equivalent bits/point +MSE 88.045798 +---------------------- -------------------------------------------------------- +Time: 4.030s Load: 0.009s, Pack+Encode: 2.241s, Decode+Unpack: 1.780s +---------------------- -------------------------------------------------------- +💾 Converting with 88.0458 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,328B, BPFP=0.4225 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,884B, BPFP=2.1958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,568B, BPFP=0.7724 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,020B, BPFP=2.1381 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,480B, BPFP=0.9669 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,068B, BPFP=2.1413 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,864B, BPFP=0.9257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,852B, BPFP=2.1936 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,076B, BPFP=1.8080 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,504B, BPFP=2.1036 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,012B, BPFP=0.2481 +⌛️ [2/4] FRONTEND: Frontend time: 2.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.571s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11892304 33.07284155 + layer.0.v_cache 0.00001803 0.00874259 + layer.1.k_cache 0.31174104 11.23991929 + layer.1.v_cache 0.00000657 0.00346127 + layer.2.k_cache 0.01452858 1.65217603 + layer.2.v_cache 0.00002161 0.01021173 + layer.3.k_cache 0.02263702 9.99912360 + layer.3.v_cache 0.00002141 0.01166156 + layer.4.k_cache 0.00068312 0.35702619 + layer.4.v_cache 0.00005251 0.02145662 + layer.4.output 1.30838121 230.34193758 + ------------------------------------------------------------------------------------- + TOTAL 0.56631185 98.16295197 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 260656 +BPFP 1.0238 bits/point +EBPFP 2.0476 equivalent bits/point +MSE 98.162952 +---------------------- -------------------------------------------------------- +Time: 3.903s Load: 0.008s, Pack+Encode: 2.324s, Decode+Unpack: 1.571s +---------------------- -------------------------------------------------------- +💾 Converting with 98.1630 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 233, 128) +Output shape: (1, 233, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.output: torch.Size([1, 233, 3584]) -> torch.Size([1, 1, 233, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,272B, BPFP=0.4206 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,900B, BPFP=2.2063 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,056B, BPFP=0.8085 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,496B, BPFP=2.1792 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,752B, BPFP=0.9893 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,072B, BPFP=2.1508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,652B, BPFP=0.9155 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,552B, BPFP=2.1829 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,340B, BPFP=1.9675 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,428B, BPFP=2.1076 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,548B, BPFP=0.2256 +⌛️ [2/4] FRONTEND: Frontend time: 2.175s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.773s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13365312 32.47706126 + layer.0.v_cache 0.00001735 0.00875713 + layer.1.k_cache 0.40872042 12.20352736 + layer.1.v_cache 0.00000792 0.00362157 + layer.2.k_cache 0.01308113 1.58118173 + layer.2.v_cache 0.00002137 0.01023067 + layer.3.k_cache 0.00601081 9.81121119 + layer.3.v_cache 0.00002094 0.01108311 + layer.4.k_cache 0.00068546 0.35835142 + layer.4.v_cache 0.00005338 0.02170049 + layer.4.output 1.31399239 231.64552039 + ------------------------------------------------------------------------------------- + TOTAL 0.57413051 98.70619815 + (elements=2,028,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2028032 +Total Bytes 261068 +BPFP 1.0298 bits/point +EBPFP 2.0597 equivalent bits/point +MSE 98.706198 +---------------------- -------------------------------------------------------- +Time: 3.957s Load: 0.009s, Pack+Encode: 2.175s, Decode+Unpack: 1.773s +---------------------- -------------------------------------------------------- +💾 Converting with 98.7062 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 148, 128) +Output shape: (1, 148, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.output: torch.Size([1, 148, 3584]) -> torch.Size([1, 1, 148, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,252B, BPFP=0.4489 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,176B, BPFP=2.6579 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,444B, BPFP=0.7859 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,272B, BPFP=2.5625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,228B, BPFP=1.0798 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,080B, BPFP=2.5422 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,624B, BPFP=0.9105 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,712B, BPFP=2.6090 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,724B, BPFP=2.1879 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,820B, BPFP=2.5148 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,128B, BPFP=0.2131 +⌛️ [2/4] FRONTEND: Frontend time: 2.060s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.510s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08608804 29.61669922 + layer.0.v_cache 0.00001767 0.00914085 + layer.1.k_cache 0.12276106 9.73254889 + layer.1.v_cache 0.00000594 0.00354628 + layer.2.k_cache 0.00886028 1.63201574 + layer.2.v_cache 0.00002088 0.01045700 + layer.3.k_cache 0.01513371 7.15797053 + layer.3.v_cache 0.00002085 0.01226450 + layer.4.k_cache 0.00068430 0.34803122 + layer.4.v_cache 0.00005331 0.02130826 + layer.4.output 0.08957551 363.78731298 + ------------------------------------------------------------------------------------- + TOTAL 0.05062792 152.65030432 + (elements=1,288,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1288192 +Total Bytes 187460 +BPFP 1.1642 bits/point +EBPFP 2.3283 equivalent bits/point +MSE 152.650304 +---------------------- -------------------------------------------------------- +Time: 3.577s Load: 0.007s, Pack+Encode: 2.060s, Decode+Unpack: 1.510s +---------------------- -------------------------------------------------------- +💾 Converting with 152.6503 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 170, 128) +Output shape: (1, 170, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.output: torch.Size([1, 170, 3584]) -> torch.Size([1, 1, 170, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,732B, BPFP=0.4349 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,716B, BPFP=2.2717 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,172B, BPFP=0.8430 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,040B, BPFP=2.2096 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,060B, BPFP=1.0165 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,844B, BPFP=2.1915 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,256B, BPFP=0.9426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,640B, BPFP=2.6324 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,096B, BPFP=2.3066 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,512B, BPFP=2.1610 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,780B, BPFP=0.2335 +⌛️ [2/4] FRONTEND: Frontend time: 1.993s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.424s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13519069 31.52120290 + layer.0.v_cache 0.00001779 0.00957975 + layer.1.k_cache 0.14803200 12.45130256 + layer.1.v_cache 0.00000678 0.00373164 + layer.2.k_cache 0.02267574 1.63996169 + layer.2.v_cache 0.00002005 0.01051247 + layer.3.k_cache 0.04618364 9.75949348 + layer.3.v_cache 0.00002282 0.01239777 + layer.4.k_cache 0.00077641 0.35536656 + layer.4.v_cache 0.00005264 0.02263792 + layer.4.output 0.00945594 321.75856092 + ------------------------------------------------------------------------------------- + TOTAL 0.02465707 135.77035960 + (elements=1,479,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1479680 +Total Bytes 202848 +BPFP 1.0967 bits/point +EBPFP 2.1934 equivalent bits/point +MSE 135.770360 +---------------------- -------------------------------------------------------- +Time: 3.423s Load: 0.006s, Pack+Encode: 1.993s, Decode+Unpack: 1.424s +---------------------- -------------------------------------------------------- +💾 Converting with 135.7704 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,572B, BPFP=0.4303 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,728B, BPFP=2.3276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,496B, BPFP=0.7997 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,432B, BPFP=2.2997 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,232B, BPFP=1.0572 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,036B, BPFP=2.2624 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,204B, BPFP=0.9605 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,460B, BPFP=2.3023 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,976B, BPFP=2.1627 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,496B, BPFP=2.2116 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,176B, BPFP=0.2310 +⌛️ [2/4] FRONTEND: Frontend time: 2.047s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.565s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12520085 32.89496070 + layer.0.v_cache 0.00001716 0.00974452 + layer.1.k_cache 0.15152623 11.98940562 + layer.1.v_cache 0.00000633 0.00372419 + layer.2.k_cache 0.01605446 1.65570068 + layer.2.v_cache 0.00002200 0.01129270 + layer.3.k_cache 0.04332644 9.61319301 + layer.3.v_cache 0.00002349 0.01315982 + layer.4.k_cache 0.00069336 0.35994454 + layer.4.v_cache 0.00005614 0.02212548 + layer.4.output 0.00966471 329.60222139 + ------------------------------------------------------------------------------------- + TOTAL 0.02379879 139.04640006 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 195808 +BPFP 1.0842 bits/point +EBPFP 2.1683 equivalent bits/point +MSE 139.046400 +---------------------- -------------------------------------------------------- +Time: 3.618s Load: 0.006s, Pack+Encode: 2.047s, Decode+Unpack: 1.565s +---------------------- -------------------------------------------------------- +💾 Converting with 139.0464 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 181, 128) +Output shape: (1, 181, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.output: torch.Size([1, 181, 3584]) -> torch.Size([1, 1, 181, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,904B, BPFP=0.4233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,684B, BPFP=2.1309 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,132B, BPFP=0.7883 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,144B, BPFP=2.0843 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,116B, BPFP=0.9596 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,852B, BPFP=2.0590 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,656B, BPFP=0.9199 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,456B, BPFP=2.1112 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,724B, BPFP=2.0480 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,476B, BPFP=2.0266 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,016B, BPFP=0.2222 +⌛️ [2/4] FRONTEND: Frontend time: 1.993s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.416s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13566811 31.85346437 + layer.0.v_cache 0.00002016 0.00957476 + layer.1.k_cache 0.14782082 12.95453184 + layer.1.v_cache 0.00000634 0.00381552 + layer.2.k_cache 0.01546242 1.55150726 + layer.2.v_cache 0.00002121 0.01071417 + layer.3.k_cache 0.02895848 9.79661063 + layer.3.v_cache 0.00002197 0.01265624 + layer.4.k_cache 0.00067433 0.35409630 + layer.4.v_cache 0.00005229 0.02212342 + layer.4.output 0.00892717 301.79595008 + ------------------------------------------------------------------------------------- + TOTAL 0.02301155 127.59651442 + (elements=1,575,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1575424 +Total Bytes 198160 +BPFP 1.0063 bits/point +EBPFP 2.0125 equivalent bits/point +MSE 127.596514 +---------------------- -------------------------------------------------------- +Time: 3.414s Load: 0.006s, Pack+Encode: 1.993s, Decode+Unpack: 1.416s +---------------------- -------------------------------------------------------- +💾 Converting with 127.5965 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 186, 128) +Output shape: (1, 186, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.output: torch.Size([1, 186, 3584]) -> torch.Size([1, 1, 186, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,000B, BPFP=0.4200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,860B, BPFP=2.0884 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,764B, BPFP=0.7362 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,080B, BPFP=2.0228 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,020B, BPFP=0.9257 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,928B, BPFP=2.0101 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,384B, BPFP=0.8723 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,940B, BPFP=2.0951 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,212B, BPFP=1.9499 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,628B, BPFP=1.9849 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,076B, BPFP=0.2409 +⌛️ [2/4] FRONTEND: Frontend time: 2.049s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09795722 35.47864163 + layer.0.v_cache 0.00001815 0.00946081 + layer.1.k_cache 0.13430934 9.87363229 + layer.1.v_cache 0.00000632 0.00362681 + layer.2.k_cache 0.01830538 1.49264789 + layer.2.v_cache 0.00002120 0.01106937 + layer.3.k_cache 0.04892764 9.62205128 + layer.3.v_cache 0.00002113 0.01283018 + layer.4.k_cache 0.00067235 0.36159630 + layer.4.v_cache 0.00008605 0.02206666 + layer.4.output 0.00869856 291.50506432 + ------------------------------------------------------------------------------------- + TOTAL 0.02124792 123.37782785 + (elements=1,618,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1618944 +Total Bytes 199892 +BPFP 0.9878 bits/point +EBPFP 1.9755 equivalent bits/point +MSE 123.377828 +---------------------- -------------------------------------------------------- +Time: 3.365s Load: 0.009s, Pack+Encode: 2.049s, Decode+Unpack: 1.307s +---------------------- -------------------------------------------------------- +💾 Converting with 123.3778 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 163, 128) +Output shape: (1, 163, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.output: torch.Size([1, 163, 3584]) -> torch.Size([1, 1, 163, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,452B, BPFP=0.4268 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,576B, BPFP=2.4517 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,404B, BPFP=0.8056 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,084B, BPFP=2.3087 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,960B, BPFP=1.0506 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,916B, BPFP=2.2926 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,012B, BPFP=0.9597 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,512B, BPFP=2.3497 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,564B, BPFP=2.0671 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,552B, BPFP=2.2577 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,068B, BPFP=0.2200 +⌛️ [2/4] FRONTEND: Frontend time: 2.086s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11530171 31.12867259 + layer.0.v_cache 0.00001813 0.00929857 + layer.1.k_cache 0.14319220 11.66360268 + layer.1.v_cache 0.00000612 0.00362425 + layer.2.k_cache 0.02368733 1.66062618 + layer.2.v_cache 0.00002106 0.01032089 + layer.3.k_cache 0.02525310 9.60466144 + layer.3.v_cache 0.00002161 0.01233247 + layer.4.k_cache 0.00069984 0.35355564 + layer.4.v_cache 0.00005709 0.02175982 + layer.4.output 0.00982084 334.89222721 + ------------------------------------------------------------------------------------- + TOTAL 0.02217671 141.10082618 + (elements=1,418,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1418752 +Total Bytes 193100 +BPFP 1.0888 bits/point +EBPFP 2.1777 equivalent bits/point +MSE 141.100826 +---------------------- -------------------------------------------------------- +Time: 3.434s Load: 0.007s, Pack+Encode: 2.086s, Decode+Unpack: 1.341s +---------------------- -------------------------------------------------------- +💾 Converting with 141.1008 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 191, 128) +Output shape: (1, 191, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.output: torch.Size([1, 191, 3584]) -> torch.Size([1, 1, 191, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,876B, BPFP=0.3989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,872B, BPFP=2.0347 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,880B, BPFP=0.7264 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,992B, BPFP=1.9627 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,572B, BPFP=0.8649 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,904B, BPFP=1.9555 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,912B, BPFP=0.8109 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,784B, BPFP=2.1911 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,696B, BPFP=1.8567 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,780B, BPFP=1.9454 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,704B, BPFP=0.2303 +⌛️ [2/4] FRONTEND: Frontend time: 2.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.590s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15112510 36.60767752 + layer.0.v_cache 0.00002069 0.00862578 + layer.1.k_cache 0.18491925 12.81436621 + layer.1.v_cache 0.00000629 0.00346249 + layer.2.k_cache 0.02122488 1.70553030 + layer.2.v_cache 0.00002327 0.01020756 + layer.3.k_cache 0.01259250 10.21639807 + layer.3.v_cache 0.00002182 0.01153786 + layer.4.k_cache 0.00068006 0.36975952 + layer.4.v_cache 0.00005885 0.01947884 + layer.4.output 0.00845948 283.12906694 + ------------------------------------------------------------------------------------- + TOTAL 0.02528759 120.21591251 + (elements=1,662,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1662464 +Total Bytes 199972 +BPFP 0.9623 bits/point +EBPFP 1.9246 equivalent bits/point +MSE 120.215913 +---------------------- -------------------------------------------------------- +Time: 3.844s Load: 0.006s, Pack+Encode: 2.248s, Decode+Unpack: 1.590s +---------------------- -------------------------------------------------------- +💾 Converting with 120.2159 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 174, 128) +Output shape: (1, 174, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.output: torch.Size([1, 174, 3584]) -> torch.Size([1, 1, 174, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,804B, BPFP=0.4314 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,560B, BPFP=2.2055 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,388B, BPFP=0.7532 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,988B, BPFP=2.1541 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,520B, BPFP=1.0345 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,920B, BPFP=2.1480 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,644B, BPFP=0.9558 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,388B, BPFP=2.1900 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,508B, BPFP=2.0212 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,612B, BPFP=2.1203 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,796B, BPFP=0.2155 +⌛️ [2/4] FRONTEND: Frontend time: 2.306s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14146894 30.28641063 + layer.0.v_cache 0.00001664 0.00923101 + layer.1.k_cache 0.16292955 9.52768751 + layer.1.v_cache 0.00000602 0.00356446 + layer.2.k_cache 0.01834387 1.67165822 + layer.2.v_cache 0.00002001 0.01057046 + layer.3.k_cache 0.03015787 10.00975510 + layer.3.v_cache 0.00002180 0.01275688 + layer.4.k_cache 0.00067479 0.34587468 + layer.4.v_cache 0.00005130 0.02212906 + layer.4.output 0.00923280 313.49686987 + ------------------------------------------------------------------------------------- + TOTAL 0.02460708 132.13986630 + (elements=1,514,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1514496 +Total Bytes 195128 +BPFP 1.0307 bits/point +EBPFP 2.0614 equivalent bits/point +MSE 132.139866 +---------------------- -------------------------------------------------------- +Time: 3.782s Load: 0.006s, Pack+Encode: 2.306s, Decode+Unpack: 1.470s +---------------------- -------------------------------------------------------- +💾 Converting with 132.1399 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.0449 bits/point +Avg EBPFP 2.0899 equivalent bits/point +Avg MSE 103.119351 +Avg Time 3.898s +------------------------ ---------------------------- diff --git a/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..5d7a6a6e323c8927b32ee487ff20875c246499ea --- /dev/null +++ b/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 333 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- -------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa +Output output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa +---------------- -------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,524B, BPFP=0.4082 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,904B, BPFP=2.1649 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,828B, BPFP=0.8045 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,832B, BPFP=2.1068 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,804B, BPFP=0.9117 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,960B, BPFP=2.1137 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,552B, BPFP=0.8438 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,488B, BPFP=2.1424 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,580B, BPFP=1.7133 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,340B, BPFP=2.0801 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,932B, BPFP=0.1855 +⌛️ [2/4] FRONTEND: Frontend time: 3.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.879s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14134069 31.66402181 + layer.0.v_cache 0.00001405 0.00675414 + layer.1.k_cache 0.62039534 13.04715051 + layer.1.v_cache 0.00000582 0.00281361 + layer.2.k_cache 0.00676800 1.60490905 + layer.2.v_cache 0.00001893 0.01024223 + layer.3.k_cache 0.03896382 9.84652710 + layer.3.v_cache 0.00001942 0.01051294 + layer.4.k_cache 0.00069280 0.29278474 + layer.4.v_cache 0.00005238 0.02084926 + layer.4.output 0.00804578 189.67069692 + ------------------------------------------------------------------------------------- + TOTAL 0.05085834 81.42361435 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 305744 +BPFP 0.9757 bits/point +EBPFP 1.9515 equivalent bits/point +MSE 81.423614 +---------------------- -------------------------------------------------------- +Time: 5.046s Load: 0.012s, Pack+Encode: 3.155s, Decode+Unpack: 1.879s +---------------------- -------------------------------------------------------- +💾 Converting with 81.4236 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,864B, BPFP=0.4194 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,092B, BPFP=2.1380 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,220B, BPFP=0.7583 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,184B, BPFP=2.0896 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,144B, BPFP=0.9142 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,348B, BPFP=2.0983 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,508B, BPFP=0.8270 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,128B, BPFP=2.1399 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,928B, BPFP=1.7560 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,780B, BPFP=2.0680 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,372B, BPFP=0.1933 +⌛️ [2/4] FRONTEND: Frontend time: 2.211s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.988s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15682776 34.33948712 + layer.0.v_cache 0.00001501 0.00694440 + layer.1.k_cache 0.54447479 12.94600826 + layer.1.v_cache 0.00000580 0.00293335 + layer.2.k_cache 0.00853899 1.59813037 + layer.2.v_cache 0.00001972 0.01009478 + layer.3.k_cache 0.02401569 9.73328512 + layer.3.v_cache 0.00001939 0.01044894 + layer.4.k_cache 0.00071059 0.32450164 + layer.4.v_cache 0.00005329 0.02055072 + layer.4.output 0.05121508 180.30244393 + ------------------------------------------------------------------------------------- + TOTAL 0.06430510 77.71232307 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 310568 +BPFP 0.9742 bits/point +EBPFP 1.9485 equivalent bits/point +MSE 77.712323 +---------------------- -------------------------------------------------------- +Time: 4.209s Load: 0.010s, Pack+Encode: 2.211s, Decode+Unpack: 1.988s +---------------------- -------------------------------------------------------- +💾 Converting with 77.7123 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,892B, BPFP=0.4194 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,036B, BPFP=2.1278 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,432B, BPFP=0.7670 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,004B, BPFP=2.0729 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,472B, BPFP=0.9286 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,628B, BPFP=2.1061 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,016B, BPFP=0.8512 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,400B, BPFP=2.1471 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,272B, BPFP=1.6620 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,684B, BPFP=2.0559 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,812B, BPFP=0.1960 +⌛️ [2/4] FRONTEND: Frontend time: 2.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.682s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15363285 33.20332762 + layer.0.v_cache 0.00001412 0.00673382 + layer.1.k_cache 0.60852461 13.00051236 + layer.1.v_cache 0.00000580 0.00284427 + layer.2.k_cache 0.01182756 1.61142860 + layer.2.v_cache 0.00001863 0.00965454 + layer.3.k_cache 0.03467803 9.81785781 + layer.3.v_cache 0.00001919 0.01009442 + layer.4.k_cache 0.00072319 0.32395956 + layer.4.v_cache 0.00005301 0.02223539 + layer.4.output 0.05035121 181.56365403 + ------------------------------------------------------------------------------------- + TOTAL 0.06835032 78.17377804 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 310648 +BPFP 0.9712 bits/point +EBPFP 1.9423 equivalent bits/point +MSE 78.173778 +---------------------- -------------------------------------------------------- +Time: 3.823s Load: 0.010s, Pack+Encode: 2.131s, Decode+Unpack: 1.682s +---------------------- -------------------------------------------------------- +💾 Converting with 78.1738 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,596B, BPFP=0.4194 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,392B, BPFP=2.2301 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,980B, BPFP=0.7719 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,216B, BPFP=2.1652 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,176B, BPFP=0.9483 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,484B, BPFP=2.1800 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,040B, BPFP=0.8856 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,128B, BPFP=2.2155 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,444B, BPFP=1.6809 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,104B, BPFP=2.1590 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,372B, BPFP=0.1843 +⌛️ [2/4] FRONTEND: Frontend time: 2.375s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.698s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12197856 29.76678790 + layer.0.v_cache 0.00001365 0.00660405 + layer.1.k_cache 0.62524759 12.66381491 + layer.1.v_cache 0.00000551 0.00270781 + layer.2.k_cache 0.01486346 1.64648082 + layer.2.v_cache 0.00001981 0.00992292 + layer.3.k_cache 0.07387851 9.91771512 + layer.3.v_cache 0.00001943 0.01001863 + layer.4.k_cache 0.00069560 0.32096456 + layer.4.v_cache 0.00005405 0.02062323 + layer.4.output 0.01098318 191.73725391 + ------------------------------------------------------------------------------------- + TOTAL 0.05374461 82.14861278 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 306932 +BPFP 0.9968 bits/point +EBPFP 1.9937 equivalent bits/point +MSE 82.148613 +---------------------- -------------------------------------------------------- +Time: 4.085s Load: 0.012s, Pack+Encode: 2.375s, Decode+Unpack: 1.698s +---------------------- -------------------------------------------------------- +💾 Converting with 82.1486 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,672B, BPFP=0.4206 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,192B, BPFP=2.2035 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,108B, BPFP=0.7735 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,164B, BPFP=2.1471 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,864B, BPFP=0.9246 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,240B, BPFP=2.1513 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,452B, BPFP=0.8471 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,004B, BPFP=2.1932 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,536B, BPFP=1.7289 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,612B, BPFP=2.1169 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,636B, BPFP=0.2164 +⌛️ [2/4] FRONTEND: Frontend time: 2.218s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.658s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15362121 31.85852522 + layer.0.v_cache 0.00001420 0.00671136 + layer.1.k_cache 0.52952420 13.14873904 + layer.1.v_cache 0.00000567 0.00281888 + layer.2.k_cache 0.00939641 1.61587417 + layer.2.v_cache 0.00001891 0.00972931 + layer.3.k_cache 0.04339437 9.77337754 + layer.3.v_cache 0.00002732 0.01052691 + layer.4.k_cache 0.00068326 0.31253539 + layer.4.v_cache 0.00005156 0.02150262 + layer.4.output 0.00776138 189.94409461 + ------------------------------------------------------------------------------------- + TOTAL 0.04653334 81.55111781 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 310480 +BPFP 1.0013 bits/point +EBPFP 2.0026 equivalent bits/point +MSE 81.551118 +---------------------- -------------------------------------------------------- +Time: 3.887s Load: 0.010s, Pack+Encode: 2.218s, Decode+Unpack: 1.658s +---------------------- -------------------------------------------------------- +💾 Converting with 81.5511 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,900B, BPFP=0.4199 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,064B, BPFP=2.1293 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,160B, BPFP=0.8057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,196B, BPFP=2.0831 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,172B, BPFP=0.9126 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,312B, BPFP=2.0893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,972B, BPFP=0.8489 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,944B, BPFP=2.1229 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,140B, BPFP=1.7081 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,684B, BPFP=2.0559 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,048B, BPFP=0.1826 +⌛️ [2/4] FRONTEND: Frontend time: 2.378s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.855s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14921317 34.73445140 + layer.0.v_cache 0.00001395 0.00682783 + layer.1.k_cache 0.64083395 13.03958151 + layer.1.v_cache 0.00000574 0.00288296 + layer.2.k_cache 0.00578804 1.59301540 + layer.2.v_cache 0.00001912 0.00980360 + layer.3.k_cache 0.02848778 9.83702949 + layer.3.v_cache 0.00001956 0.01062280 + layer.4.k_cache 0.00070939 0.31285619 + layer.4.v_cache 0.00005206 0.02152484 + layer.4.output 0.04916091 179.54521987 + ------------------------------------------------------------------------------------- + TOTAL 0.06878054 77.43441971 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 309592 +BPFP 0.9679 bits/point +EBPFP 1.9357 equivalent bits/point +MSE 77.434420 +---------------------- -------------------------------------------------------- +Time: 4.244s Load: 0.011s, Pack+Encode: 2.378s, Decode+Unpack: 1.855s +---------------------- -------------------------------------------------------- +💾 Converting with 77.4344 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,880B, BPFP=0.4202 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,048B, BPFP=2.1357 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,336B, BPFP=0.7645 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,112B, BPFP=2.0858 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,796B, BPFP=0.8957 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,652B, BPFP=2.1145 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,108B, BPFP=0.8590 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,100B, BPFP=2.1384 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,288B, BPFP=1.6685 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,684B, BPFP=2.0629 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,432B, BPFP=0.1937 +⌛️ [2/4] FRONTEND: Frontend time: 2.218s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.697s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12806672 35.39863748 + layer.0.v_cache 0.00001388 0.00682999 + layer.1.k_cache 0.56701790 12.84959938 + layer.1.v_cache 0.00000557 0.00280684 + layer.2.k_cache 0.01020011 1.58803051 + layer.2.v_cache 0.00001881 0.00966817 + layer.3.k_cache 0.04301871 9.91088701 + layer.3.v_cache 0.00002051 0.01038247 + layer.4.k_cache 0.00071755 0.31756537 + layer.4.v_cache 0.00005284 0.02123319 + layer.4.output 0.04940383 180.59690395 + ------------------------------------------------------------------------------------- + TOTAL 0.06440938 77.89964518 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 309436 +BPFP 0.9707 bits/point +EBPFP 1.9414 equivalent bits/point +MSE 77.899645 +---------------------- -------------------------------------------------------- +Time: 3.926s Load: 0.010s, Pack+Encode: 2.218s, Decode+Unpack: 1.697s +---------------------- -------------------------------------------------------- +💾 Converting with 77.8996 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,648B, BPFP=0.4193 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,292B, BPFP=2.2090 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,852B, BPFP=0.8143 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,340B, BPFP=2.1568 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,012B, BPFP=0.9327 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,344B, BPFP=2.1570 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,712B, BPFP=0.8614 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,044B, BPFP=2.1954 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,752B, BPFP=1.7408 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,824B, BPFP=2.1285 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,972B, BPFP=0.1878 +⌛️ [2/4] FRONTEND: Frontend time: 2.179s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.737s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12903612 31.75963199 + layer.0.v_cache 0.00001458 0.00676655 + layer.1.k_cache 0.59860808 13.08534985 + layer.1.v_cache 0.00000581 0.00280797 + layer.2.k_cache 0.00940504 1.61825390 + layer.2.v_cache 0.00001954 0.01026733 + layer.3.k_cache 0.02644065 9.85095772 + layer.3.v_cache 0.00002029 0.01021852 + layer.4.k_cache 0.00068693 0.31704731 + layer.4.v_cache 0.00005376 0.02135985 + layer.4.output 0.00862506 187.49741541 + ------------------------------------------------------------------------------------- + TOTAL 0.04850978 80.53909229 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 308792 +BPFP 0.9958 bits/point +EBPFP 1.9917 equivalent bits/point +MSE 80.539092 +---------------------- -------------------------------------------------------- +Time: 3.925s Load: 0.009s, Pack+Encode: 2.179s, Decode+Unpack: 1.737s +---------------------- -------------------------------------------------------- +💾 Converting with 80.5391 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,188B, BPFP=0.4154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,012B, BPFP=2.0298 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,348B, BPFP=0.7279 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,952B, BPFP=1.9761 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,144B, BPFP=0.8697 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,020B, BPFP=1.9795 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,420B, BPFP=0.8330 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,920B, BPFP=2.1774 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,192B, BPFP=1.6331 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,724B, BPFP=1.9645 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,252B, BPFP=0.1758 +⌛️ [2/4] FRONTEND: Frontend time: 2.399s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.708s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16469599 33.92909459 + layer.0.v_cache 0.00001357 0.00661739 + layer.1.k_cache 0.70445866 11.31809225 + layer.1.v_cache 0.00000558 0.00279258 + layer.2.k_cache 0.01447187 1.59102502 + layer.2.v_cache 0.00001924 0.00987348 + layer.3.k_cache 0.02617162 9.88711667 + layer.3.v_cache 0.00001974 0.01012342 + layer.4.k_cache 0.00070375 0.31237129 + layer.4.v_cache 0.00005076 0.02175385 + layer.4.output 0.04814868 169.05765886 + ------------------------------------------------------------------------------------- + TOTAL 0.07339127 72.97014544 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 312172 +BPFP 0.9316 bits/point +EBPFP 1.8631 equivalent bits/point +MSE 72.970145 +---------------------- -------------------------------------------------------- +Time: 4.119s Load: 0.012s, Pack+Encode: 2.399s, Decode+Unpack: 1.708s +---------------------- -------------------------------------------------------- +💾 Converting with 72.9701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,480B, BPFP=0.4345 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,684B, BPFP=2.3051 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,328B, BPFP=0.7742 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,196B, BPFP=2.2186 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,996B, BPFP=0.9872 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,156B, BPFP=2.2163 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,596B, BPFP=0.9059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,712B, BPFP=2.2486 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,136B, BPFP=1.7505 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,632B, BPFP=2.1859 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,308B, BPFP=0.1768 +⌛️ [2/4] FRONTEND: Frontend time: 2.114s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.822s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12132761 29.86886472 + layer.0.v_cache 0.00001366 0.00665635 + layer.1.k_cache 0.57241889 12.26390149 + layer.1.v_cache 0.00000556 0.00275019 + layer.2.k_cache 0.00926234 1.60115533 + layer.2.v_cache 0.00002145 0.00997942 + layer.3.k_cache 0.03109839 9.78720649 + layer.3.v_cache 0.00002257 0.01052613 + layer.4.k_cache 0.00070701 0.32208797 + layer.4.v_cache 0.00004920 0.02148462 + layer.4.output 0.01031505 203.60237985 + ------------------------------------------------------------------------------------- + TOTAL 0.04747836 87.00654539 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 297224 +BPFP 1.0156 bits/point +EBPFP 2.0311 equivalent bits/point +MSE 87.006545 +---------------------- -------------------------------------------------------- +Time: 3.945s Load: 0.010s, Pack+Encode: 2.114s, Decode+Unpack: 1.822s +---------------------- -------------------------------------------------------- +💾 Converting with 87.0065 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,988B, BPFP=0.4188 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,144B, BPFP=2.1049 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,920B, BPFP=0.7823 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,180B, BPFP=2.0543 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,544B, BPFP=0.9199 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,836B, BPFP=2.1411 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,988B, BPFP=0.8383 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,924B, BPFP=2.0933 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,952B, BPFP=1.7278 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,880B, BPFP=2.0386 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,896B, BPFP=0.1790 +⌛️ [2/4] FRONTEND: Frontend time: 2.406s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.830s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14357947 37.87667130 + layer.0.v_cache 0.00001420 0.00694222 + layer.1.k_cache 0.63723908 12.62223171 + layer.1.v_cache 0.00000573 0.00280231 + layer.2.k_cache 0.00765126 1.64089208 + layer.2.v_cache 0.00002073 0.00955195 + layer.3.k_cache 0.02843528 9.93572138 + layer.3.v_cache 0.00002112 0.01009155 + layer.4.k_cache 0.00069171 0.31651214 + layer.4.v_cache 0.00005234 0.02013393 + layer.4.output 0.04871205 181.25037452 + ------------------------------------------------------------------------------------- + TOTAL 0.06815855 78.30553954 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 312252 +BPFP 0.9631 bits/point +EBPFP 1.9261 equivalent bits/point +MSE 78.305540 +---------------------- -------------------------------------------------------- +Time: 4.249s Load: 0.012s, Pack+Encode: 2.406s, Decode+Unpack: 1.830s +---------------------- -------------------------------------------------------- +💾 Converting with 78.3055 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,132B, BPFP=0.4166 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,880B, BPFP=2.0430 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,172B, BPFP=0.7260 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,848B, BPFP=1.9902 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,068B, BPFP=0.8744 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,052B, BPFP=2.0006 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,196B, BPFP=0.8297 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,404B, BPFP=2.0186 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,900B, BPFP=1.7367 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,540B, BPFP=1.9744 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,288B, BPFP=0.1778 +⌛️ [2/4] FRONTEND: Frontend time: 2.288s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.767s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16857396 33.81541368 + layer.0.v_cache 0.00001478 0.00699190 + layer.1.k_cache 0.58716421 10.47617988 + layer.1.v_cache 0.00000600 0.00300393 + layer.2.k_cache 0.00827454 1.59633209 + layer.2.v_cache 0.00002177 0.01015836 + layer.3.k_cache 0.02531170 9.87067191 + layer.3.v_cache 0.00001983 0.01098831 + layer.4.k_cache 0.00069463 0.30753669 + layer.4.v_cache 0.00005316 0.02105518 + layer.4.output 0.04836898 172.16151932 + ------------------------------------------------------------------------------------- + TOTAL 0.06639514 74.19111572 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 309480 +BPFP 0.9326 bits/point +EBPFP 1.8652 equivalent bits/point +MSE 74.191116 +---------------------- -------------------------------------------------------- +Time: 4.065s Load: 0.010s, Pack+Encode: 2.288s, Decode+Unpack: 1.767s +---------------------- -------------------------------------------------------- +💾 Converting with 74.1911 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,988B, BPFP=0.4188 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,908B, BPFP=2.0925 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,164B, BPFP=0.7951 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,024B, BPFP=2.0461 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,412B, BPFP=0.9130 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,172B, BPFP=2.0539 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,480B, BPFP=0.8117 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,740B, BPFP=2.0837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,400B, BPFP=1.8037 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,608B, BPFP=2.0243 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,676B, BPFP=0.1923 +⌛️ [2/4] FRONTEND: Frontend time: 2.299s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.905s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14028978 39.49692612 + layer.0.v_cache 0.00001397 0.00697867 + layer.1.k_cache 0.59966084 12.99129041 + layer.1.v_cache 0.00000578 0.00298806 + layer.2.k_cache 0.01765286 1.63098247 + layer.2.v_cache 0.00001884 0.01006576 + layer.3.k_cache 0.02653811 9.71826008 + layer.3.v_cache 0.00001935 0.01058633 + layer.4.k_cache 0.00071527 0.31128964 + layer.4.v_cache 0.00005402 0.02237149 + layer.4.output 0.04907703 180.07875419 + ------------------------------------------------------------------------------------- + TOTAL 0.06638282 77.92664814 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 312572 +BPFP 0.9641 bits/point +EBPFP 1.9281 equivalent bits/point +MSE 77.926648 +---------------------- -------------------------------------------------------- +Time: 4.217s Load: 0.013s, Pack+Encode: 2.299s, Decode+Unpack: 1.905s +---------------------- -------------------------------------------------------- +💾 Converting with 77.9266 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,492B, BPFP=0.4304 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,160B, BPFP=2.3070 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,668B, BPFP=0.7852 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,804B, BPFP=2.2291 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,712B, BPFP=0.9600 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,752B, BPFP=2.2261 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,588B, BPFP=0.8955 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,060B, BPFP=2.2438 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,796B, BPFP=1.6542 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,020B, BPFP=2.1841 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,788B, BPFP=0.1952 +⌛️ [2/4] FRONTEND: Frontend time: 2.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.694s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13255837 30.67529656 + layer.0.v_cache 0.00001394 0.00668641 + layer.1.k_cache 0.54921044 12.59268907 + layer.1.v_cache 0.00000578 0.00278066 + layer.2.k_cache 0.01292971 1.56434070 + layer.2.v_cache 0.00001901 0.01027352 + layer.3.k_cache 0.07190662 9.45413567 + layer.3.v_cache 0.00001944 0.01045228 + layer.4.k_cache 0.00068159 0.30149895 + layer.4.v_cache 0.00005125 0.02058068 + layer.4.output 0.00790709 198.89412093 + ------------------------------------------------------------------------------------- + TOTAL 0.04839681 85.11162241 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 300840 +BPFP 1.0166 bits/point +EBPFP 2.0331 equivalent bits/point +MSE 85.111622 +---------------------- -------------------------------------------------------- +Time: 3.930s Load: 0.009s, Pack+Encode: 2.227s, Decode+Unpack: 1.694s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1116 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,632B, BPFP=0.4229 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,508B, BPFP=2.2445 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,948B, BPFP=0.8282 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,392B, BPFP=2.1826 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,816B, BPFP=0.9317 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,848B, BPFP=2.2079 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,596B, BPFP=0.8641 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,224B, BPFP=2.2287 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,900B, BPFP=1.7121 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,016B, BPFP=2.1618 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,992B, BPFP=0.2137 +⌛️ [2/4] FRONTEND: Frontend time: 2.404s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.831s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13695534 30.99289049 + layer.0.v_cache 0.00001452 0.00675028 + layer.1.k_cache 0.52272531 13.06554137 + layer.1.v_cache 0.00000574 0.00288764 + layer.2.k_cache 0.00793556 1.64418030 + layer.2.v_cache 0.00001907 0.00965498 + layer.3.k_cache 0.04682703 10.00464646 + layer.3.v_cache 0.00002008 0.01023446 + layer.4.k_cache 0.00068945 0.33149184 + layer.4.v_cache 0.00005221 0.02022986 + layer.4.output 0.00941004 194.55194086 + ------------------------------------------------------------------------------------- + TOTAL 0.04594792 83.40894669 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 311872 +BPFP 1.0165 bits/point +EBPFP 2.0330 equivalent bits/point +MSE 83.408947 +---------------------- -------------------------------------------------------- +Time: 4.245s Load: 0.010s, Pack+Encode: 2.404s, Decode+Unpack: 1.831s +---------------------- -------------------------------------------------------- +💾 Converting with 83.4089 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,628B, BPFP=0.4257 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,376B, BPFP=2.2531 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,928B, BPFP=0.7772 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,224B, BPFP=2.1888 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,704B, BPFP=0.9321 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,228B, BPFP=2.1891 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,596B, BPFP=0.8703 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,140B, BPFP=2.2400 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,096B, BPFP=1.7353 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,712B, BPFP=2.1603 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,240B, BPFP=0.1853 +⌛️ [2/4] FRONTEND: Frontend time: 2.370s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.669s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14340984 31.69652623 + layer.0.v_cache 0.00001397 0.00666171 + layer.1.k_cache 0.51622396 11.46303362 + layer.1.v_cache 0.00000557 0.00276137 + layer.2.k_cache 0.00797994 1.58643690 + layer.2.v_cache 0.00001813 0.00973467 + layer.3.k_cache 0.02756977 9.18518938 + layer.3.v_cache 0.00001890 0.00968840 + layer.4.k_cache 0.00068956 0.31790150 + layer.4.v_cache 0.00005177 0.02038106 + layer.4.output 0.00947078 196.02992666 + ------------------------------------------------------------------------------------- + TOTAL 0.04483982 83.91222361 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 305872 +BPFP 1.0040 bits/point +EBPFP 2.0081 equivalent bits/point +MSE 83.912224 +---------------------- -------------------------------------------------------- +Time: 4.049s Load: 0.011s, Pack+Encode: 2.370s, Decode+Unpack: 1.669s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9122 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,860B, BPFP=0.4177 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,996B, BPFP=2.1256 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,096B, BPFP=0.7491 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,140B, BPFP=2.0801 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,404B, BPFP=0.9250 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,336B, BPFP=2.0906 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,744B, BPFP=0.8367 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,892B, BPFP=2.1201 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,564B, BPFP=1.8369 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,780B, BPFP=2.0610 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,672B, BPFP=0.1873 +⌛️ [2/4] FRONTEND: Frontend time: 2.198s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.608s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13622742 35.66189015 + layer.0.v_cache 0.00001380 0.00678778 + layer.1.k_cache 0.61520739 10.61517874 + layer.1.v_cache 0.00000584 0.00290630 + layer.2.k_cache 0.00900942 1.60612550 + layer.2.v_cache 0.00001916 0.01006903 + layer.3.k_cache 0.02667915 9.71607940 + layer.3.v_cache 0.00001881 0.00979225 + layer.4.k_cache 0.00071466 0.31176218 + layer.4.v_cache 0.00005760 0.02055237 + layer.4.output 0.04972813 179.48047255 + ------------------------------------------------------------------------------------- + TOTAL 0.06682648 77.31320303 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 311484 +BPFP 0.9738 bits/point +EBPFP 1.9476 equivalent bits/point +MSE 77.313203 +---------------------- -------------------------------------------------------- +Time: 3.817s Load: 0.011s, Pack+Encode: 2.198s, Decode+Unpack: 1.608s +---------------------- -------------------------------------------------------- +💾 Converting with 77.3132 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,072B, BPFP=0.4218 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,948B, BPFP=2.0876 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,812B, BPFP=0.7740 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,988B, BPFP=2.0374 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,504B, BPFP=0.9147 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,448B, BPFP=2.0615 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,900B, BPFP=0.8309 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,140B, BPFP=2.0976 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,372B, BPFP=1.6917 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,804B, BPFP=2.0278 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,208B, BPFP=0.1807 +⌛️ [2/4] FRONTEND: Frontend time: 2.266s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.841s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11478914 39.09817883 + layer.0.v_cache 0.00001377 0.00664748 + layer.1.k_cache 0.64218915 12.42591026 + layer.1.v_cache 0.00000556 0.00271943 + layer.2.k_cache 0.00898698 1.61874400 + layer.2.v_cache 0.00001980 0.00965920 + layer.3.k_cache 0.05442990 9.92048364 + layer.3.v_cache 0.00001964 0.01007968 + layer.4.k_cache 0.00070461 0.32323594 + layer.4.v_cache 0.00005015 0.02071131 + layer.4.output 0.05017195 176.66589525 + ------------------------------------------------------------------------------------- + TOTAL 0.06896543 76.47633156 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 310196 +BPFP 0.9535 bits/point +EBPFP 1.9071 equivalent bits/point +MSE 76.476332 +---------------------- -------------------------------------------------------- +Time: 4.117s Load: 0.010s, Pack+Encode: 2.266s, Decode+Unpack: 1.841s +---------------------- -------------------------------------------------------- +💾 Converting with 76.4763 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,248B, BPFP=0.4274 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,320B, BPFP=2.3184 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,592B, BPFP=0.8604 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,976B, BPFP=2.2392 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,152B, BPFP=0.9524 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,348B, BPFP=2.2021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,080B, BPFP=0.8892 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,768B, BPFP=2.2858 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,564B, BPFP=1.7432 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,228B, BPFP=2.1950 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,892B, BPFP=0.1676 +⌛️ [2/4] FRONTEND: Frontend time: 2.172s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.689s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14755792 30.15864350 + layer.0.v_cache 0.00001375 0.00688526 + layer.1.k_cache 0.51940866 12.70733527 + layer.1.v_cache 0.00000586 0.00299025 + layer.2.k_cache 0.00816260 1.58548457 + layer.2.v_cache 0.00001900 0.01039783 + layer.3.k_cache 0.02921419 9.19003906 + layer.3.v_cache 0.00001979 0.01084006 + layer.4.k_cache 0.00071017 0.31867451 + layer.4.v_cache 0.00005507 0.02357987 + layer.4.output 0.00937099 208.12840296 + ------------------------------------------------------------------------------------- + TOTAL 0.04533906 88.87727594 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 293168 +BPFP 1.0168 bits/point +EBPFP 2.0336 equivalent bits/point +MSE 88.877276 +---------------------- -------------------------------------------------------- +Time: 3.872s Load: 0.011s, Pack+Encode: 2.172s, Decode+Unpack: 1.689s +---------------------- -------------------------------------------------------- +💾 Converting with 88.8773 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,324B, BPFP=0.4286 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,392B, BPFP=2.3052 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,840B, BPFP=0.8684 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,396B, BPFP=2.2470 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,684B, BPFP=0.9764 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,872B, BPFP=2.2163 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,336B, BPFP=0.8975 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,816B, BPFP=2.2715 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,904B, BPFP=1.7500 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,584B, BPFP=2.1994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,256B, BPFP=0.1693 +⌛️ [2/4] FRONTEND: Frontend time: 2.294s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.581s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12840328 28.93040438 + layer.0.v_cache 0.00001372 0.00670899 + layer.1.k_cache 0.60081110 13.04221895 + layer.1.v_cache 0.00000562 0.00276332 + layer.2.k_cache 0.00586274 1.62325055 + layer.2.v_cache 0.00001839 0.00975343 + layer.3.k_cache 0.02491910 9.68555876 + layer.3.v_cache 0.00001957 0.01039715 + layer.4.k_cache 0.00072047 0.30819399 + layer.4.v_cache 0.00005229 0.02254163 + layer.4.output 0.01050496 203.04455926 + ------------------------------------------------------------------------------------- + TOTAL 0.04908006 86.76198270 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 296404 +BPFP 1.0203 bits/point +EBPFP 2.0407 equivalent bits/point +MSE 86.761983 +---------------------- -------------------------------------------------------- +Time: 3.886s Load: 0.011s, Pack+Encode: 2.294s, Decode+Unpack: 1.581s +---------------------- -------------------------------------------------------- +💾 Converting with 86.7620 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,644B, BPFP=0.4220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,364B, BPFP=2.2286 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,600B, BPFP=0.8061 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,264B, BPFP=2.1678 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,968B, BPFP=0.9368 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,260B, BPFP=2.1676 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,692B, BPFP=0.8664 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,428B, BPFP=2.3425 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,848B, BPFP=1.7584 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,968B, BPFP=2.1515 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,872B, BPFP=0.1804 +⌛️ [2/4] FRONTEND: Frontend time: 2.271s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.699s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14595435 32.36361942 + layer.0.v_cache 0.00001364 0.00634690 + layer.1.k_cache 0.59759904 12.67096659 + layer.1.v_cache 0.00000566 0.00269069 + layer.2.k_cache 0.01409306 1.62416276 + layer.2.v_cache 0.00002047 0.00998248 + layer.3.k_cache 0.06079736 9.81299173 + layer.3.v_cache 0.00002215 0.01020852 + layer.4.k_cache 0.00070776 0.33036529 + layer.4.v_cache 0.00005221 0.02086900 + layer.4.output 0.00769303 192.24062973 + ------------------------------------------------------------------------------------- + TOTAL 0.05135982 82.50215362 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 309908 +BPFP 1.0065 bits/point +EBPFP 2.0130 equivalent bits/point +MSE 82.502154 +---------------------- -------------------------------------------------------- +Time: 3.979s Load: 0.009s, Pack+Encode: 2.271s, Decode+Unpack: 1.699s +---------------------- -------------------------------------------------------- +💾 Converting with 82.5022 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 315, 128) +Output shape: (1, 315, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.output: torch.Size([1, 315, 3584]) -> torch.Size([1, 1, 315, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,304B, BPFP=0.4119 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,256B, BPFP=1.9968 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,716B, BPFP=0.7300 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,244B, BPFP=1.9466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,660B, BPFP=0.8264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,248B, BPFP=1.9468 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,880B, BPFP=0.7877 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,992B, BPFP=1.9837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,860B, BPFP=1.7292 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,724B, BPFP=1.9208 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,740B, BPFP=0.1824 +⌛️ [2/4] FRONTEND: Frontend time: 2.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.666s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15500355 36.93503224 + layer.0.v_cache 0.00001366 0.00674910 + layer.1.k_cache 0.68167124 11.71620009 + layer.1.v_cache 0.00000571 0.00288127 + layer.2.k_cache 0.01007247 1.53035559 + layer.2.v_cache 0.00001936 0.01026502 + layer.3.k_cache 0.03272899 9.81045542 + layer.3.v_cache 0.00002006 0.01041833 + layer.4.k_cache 0.00071175 0.33521041 + layer.4.v_cache 0.00005427 0.02199376 + layer.4.output 0.04584261 165.69630102 + ------------------------------------------------------------------------------------- + TOTAL 0.07065879 71.77962755 + (elements=2,741,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2741760 +Total Bytes 313624 +BPFP 0.9151 bits/point +EBPFP 1.8302 equivalent bits/point +MSE 71.779628 +---------------------- -------------------------------------------------------- +Time: 3.928s Load: 0.011s, Pack+Encode: 2.251s, Decode+Unpack: 1.666s +---------------------- -------------------------------------------------------- +💾 Converting with 71.7796 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,484B, BPFP=0.4315 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,884B, BPFP=2.2996 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,620B, BPFP=0.7853 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,680B, BPFP=2.2302 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,676B, BPFP=0.9615 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,572B, BPFP=2.2239 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,216B, BPFP=0.8773 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,056B, BPFP=2.3095 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,668B, BPFP=1.7106 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,680B, BPFP=2.1725 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,984B, BPFP=0.1728 +⌛️ [2/4] FRONTEND: Frontend time: 2.427s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.668s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12889319 31.17023538 + layer.0.v_cache 0.00001350 0.00677412 + layer.1.k_cache 0.58775566 12.67911634 + layer.1.v_cache 0.00000553 0.00281372 + layer.2.k_cache 0.00723753 1.58920468 + layer.2.v_cache 0.00001924 0.01017231 + layer.3.k_cache 0.01761036 8.67017593 + layer.3.v_cache 0.00001847 0.01113271 + layer.4.k_cache 0.00071814 0.30492212 + layer.4.v_cache 0.00005270 0.02156043 + layer.4.output 0.01035140 199.27347457 + ------------------------------------------------------------------------------------- + TOTAL 0.04792848 85.25767234 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 298520 +BPFP 1.0125 bits/point +EBPFP 2.0249 equivalent bits/point +MSE 85.257672 +---------------------- -------------------------------------------------------- +Time: 4.104s Load: 0.009s, Pack+Encode: 2.427s, Decode+Unpack: 1.668s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2577 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,648B, BPFP=0.4208 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,096B, BPFP=2.2060 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,644B, BPFP=0.7507 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,952B, BPFP=2.1430 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,760B, BPFP=0.9221 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,068B, BPFP=2.1494 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,492B, BPFP=0.8523 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,396B, BPFP=2.1675 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,492B, BPFP=1.6776 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,548B, BPFP=2.1208 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,724B, BPFP=0.1707 +⌛️ [2/4] FRONTEND: Frontend time: 2.318s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.613s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13033964 31.34054522 + layer.0.v_cache 0.00001368 0.00670349 + layer.1.k_cache 0.63029007 12.37443435 + layer.1.v_cache 0.00000551 0.00274867 + layer.2.k_cache 0.00637632 1.61754297 + layer.2.v_cache 0.00001960 0.00971264 + layer.3.k_cache 0.02368569 9.64454844 + layer.3.v_cache 0.00001974 0.01044672 + layer.4.k_cache 0.00068050 0.30846996 + layer.4.v_cache 0.00005209 0.02092158 + layer.4.output 0.00765416 189.13185362 + ------------------------------------------------------------------------------------- + TOTAL 0.04970952 81.13288526 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 301820 +BPFP 0.9768 bits/point +EBPFP 1.9536 equivalent bits/point +MSE 81.132885 +---------------------- -------------------------------------------------------- +Time: 3.941s Load: 0.010s, Pack+Encode: 2.318s, Decode+Unpack: 1.613s +---------------------- -------------------------------------------------------- +💾 Converting with 81.1329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,512B, BPFP=0.4315 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,724B, BPFP=2.2819 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,872B, BPFP=0.7969 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,548B, BPFP=2.2144 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,916B, BPFP=0.9717 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,392B, BPFP=2.2054 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,524B, BPFP=0.8918 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,176B, BPFP=2.2505 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,632B, BPFP=1.7597 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,928B, BPFP=2.1788 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,384B, BPFP=0.1755 +⌛️ [2/4] FRONTEND: Frontend time: 2.326s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.874s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15644295 30.35442756 + layer.0.v_cache 0.00001405 0.00667337 + layer.1.k_cache 0.56163917 12.93277247 + layer.1.v_cache 0.00000554 0.00284575 + layer.2.k_cache 0.01059573 1.61458273 + layer.2.v_cache 0.00001807 0.01006039 + layer.3.k_cache 0.04944291 8.56522325 + layer.3.v_cache 0.00001950 0.01019172 + layer.4.k_cache 0.00071267 0.30755329 + layer.4.v_cache 0.00005817 0.02136496 + layer.4.output 0.00956290 199.91386555 + ------------------------------------------------------------------------------------- + TOTAL 0.04975818 85.48369143 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 299608 +BPFP 1.0124 bits/point +EBPFP 2.0248 equivalent bits/point +MSE 85.483691 +---------------------- -------------------------------------------------------- +Time: 4.210s Load: 0.010s, Pack+Encode: 2.326s, Decode+Unpack: 1.874s +---------------------- -------------------------------------------------------- +💾 Converting with 85.4837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 312, 128) +Output shape: (1, 312, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.output: torch.Size([1, 312, 3584]) -> torch.Size([1, 1, 312, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,308B, BPFP=0.4161 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,460B, BPFP=2.0262 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,528B, BPFP=0.7276 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,132B, BPFP=1.9597 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,132B, BPFP=0.8580 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,468B, BPFP=1.9766 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,988B, BPFP=0.8007 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,976B, BPFP=2.0020 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,512B, BPFP=1.7284 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,724B, BPFP=1.9393 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,480B, BPFP=0.1823 +⌛️ [2/4] FRONTEND: Frontend time: 2.621s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.673s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14984696 35.94080842 + layer.0.v_cache 0.00001474 0.00691546 + layer.1.k_cache 0.67859346 12.21377955 + layer.1.v_cache 0.00000581 0.00293105 + layer.2.k_cache 0.01626646 1.51984601 + layer.2.v_cache 0.00001951 0.01027935 + layer.3.k_cache 0.02914289 9.88103309 + layer.3.v_cache 0.00001919 0.01029707 + layer.4.k_cache 0.00071665 0.31831387 + layer.4.v_cache 0.00005096 0.02050780 + layer.4.output 0.04642455 171.54245364 + ------------------------------------------------------------------------------------- + TOTAL 0.07056756 74.16011101 + (elements=2,715,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2715648 +Total Bytes 313708 +BPFP 0.9241 bits/point +EBPFP 1.8483 equivalent bits/point +MSE 74.160111 +---------------------- -------------------------------------------------------- +Time: 4.304s Load: 0.010s, Pack+Encode: 2.621s, Decode+Unpack: 1.673s +---------------------- -------------------------------------------------------- +💾 Converting with 74.1601 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,448B, BPFP=0.4326 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,100B, BPFP=2.3292 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,728B, BPFP=0.7974 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,676B, BPFP=2.2465 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,756B, BPFP=0.9733 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,508B, BPFP=2.2368 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,452B, BPFP=0.8975 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,492B, BPFP=2.2939 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,112B, BPFP=1.6910 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,124B, BPFP=2.2145 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,900B, BPFP=0.1983 +⌛️ [2/4] FRONTEND: Frontend time: 2.330s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.612s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15823836 31.11675912 + layer.0.v_cache 0.00001502 0.00676613 + layer.1.k_cache 0.53668020 12.57210226 + layer.1.v_cache 0.00000576 0.00286075 + layer.2.k_cache 0.01018995 1.58648455 + layer.2.v_cache 0.00002218 0.01016855 + layer.3.k_cache 0.05410197 9.62863307 + layer.3.v_cache 0.00002001 0.01052328 + layer.4.k_cache 0.00069143 0.32474878 + layer.4.v_cache 0.00005304 0.02168167 + layer.4.output 0.01041225 201.67541822 + ------------------------------------------------------------------------------------- + TOTAL 0.04899433 86.29462680 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 301296 +BPFP 1.0295 bits/point +EBPFP 2.0589 equivalent bits/point +MSE 86.294627 +---------------------- -------------------------------------------------------- +Time: 3.952s Load: 0.009s, Pack+Encode: 2.330s, Decode+Unpack: 1.612s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2946 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 329, 128) +Output shape: (1, 329, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.output: torch.Size([1, 329, 3584]) -> torch.Size([1, 1, 329, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,976B, BPFP=0.4263 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,316B, BPFP=2.2472 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,284B, BPFP=0.8209 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,984B, BPFP=2.2314 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,408B, BPFP=0.9217 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,176B, BPFP=2.1455 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,924B, BPFP=0.8513 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,632B, BPFP=2.2147 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,884B, BPFP=1.6567 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,720B, BPFP=2.1714 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,588B, BPFP=0.2279 +⌛️ [2/4] FRONTEND: Frontend time: 3.147s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.922s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11391156 32.50267442 + layer.0.v_cache 0.00001522 0.00668789 + layer.1.k_cache 0.63933222 12.85334940 + layer.1.v_cache 0.00000652 0.00294005 + layer.2.k_cache 0.01643764 1.53536152 + layer.2.v_cache 0.00001873 0.00957080 + layer.3.k_cache 0.02826017 9.18886837 + layer.3.v_cache 0.00001876 0.00961093 + layer.4.k_cache 0.00074607 0.33131571 + layer.4.v_cache 0.00005483 0.02230479 + layer.4.output 3.44632065 161.96372937 + ------------------------------------------------------------------------------------- + TOTAL 1.46606155 70.01228174 + (elements=2,863,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2863616 +Total Bytes 363892 +BPFP 1.0166 bits/point +EBPFP 2.0332 equivalent bits/point +MSE 70.012282 +---------------------- -------------------------------------------------------- +Time: 5.080s Load: 0.011s, Pack+Encode: 3.147s, Decode+Unpack: 1.922s +---------------------- -------------------------------------------------------- +💾 Converting with 70.0123 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,508B, BPFP=0.4205 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,372B, BPFP=2.2610 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,248B, BPFP=0.7979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,060B, BPFP=2.1875 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,064B, BPFP=0.9556 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,944B, BPFP=2.1810 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,608B, BPFP=0.8741 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,012B, BPFP=2.2408 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,684B, BPFP=1.7184 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,592B, BPFP=2.1613 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,996B, BPFP=0.1680 +⌛️ [2/4] FRONTEND: Frontend time: 2.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.943s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13536828 30.02515611 + layer.0.v_cache 0.00001433 0.00680138 + layer.1.k_cache 0.57185238 12.32465278 + layer.1.v_cache 0.00000558 0.00279885 + layer.2.k_cache 0.00766427 1.60396497 + layer.2.v_cache 0.00001891 0.00998514 + layer.3.k_cache 0.04143807 9.69526000 + layer.3.v_cache 0.00001947 0.01023727 + layer.4.k_cache 0.00070404 0.30084950 + layer.4.v_cache 0.00005498 0.02098326 + layer.4.output 0.00948056 196.05032322 + ------------------------------------------------------------------------------------- + TOTAL 0.04844142 83.90311481 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 303088 +BPFP 0.9985 bits/point +EBPFP 1.9969 equivalent bits/point +MSE 83.903115 +---------------------- -------------------------------------------------------- +Time: 4.220s Load: 0.010s, Pack+Encode: 2.267s, Decode+Unpack: 1.943s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9031 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,504B, BPFP=0.4248 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,204B, BPFP=2.2760 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,524B, BPFP=0.7656 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,024B, BPFP=2.2092 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,528B, BPFP=0.9357 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,120B, BPFP=2.2147 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,412B, BPFP=0.8725 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,360B, BPFP=2.2283 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,272B, BPFP=1.7138 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,472B, BPFP=2.1780 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,288B, BPFP=0.1803 +⌛️ [2/4] FRONTEND: Frontend time: 2.390s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.854s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15132187 28.88604294 + layer.0.v_cache 0.00001456 0.00669783 + layer.1.k_cache 0.61798344 12.10295746 + layer.1.v_cache 0.00000569 0.00283646 + layer.2.k_cache 0.01686249 1.57215152 + layer.2.v_cache 0.00001898 0.01072981 + layer.3.k_cache 0.04401700 9.61036129 + layer.3.v_cache 0.00001935 0.01058049 + layer.4.k_cache 0.00071189 0.29857929 + layer.4.v_cache 0.00005111 0.02184155 + layer.4.output 0.01058244 194.15117107 + ------------------------------------------------------------------------------------- + TOTAL 0.05324020 83.03417506 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 301708 +BPFP 1.0047 bits/point +EBPFP 2.0095 equivalent bits/point +MSE 83.034175 +---------------------- -------------------------------------------------------- +Time: 4.257s Load: 0.013s, Pack+Encode: 2.390s, Decode+Unpack: 1.854s +---------------------- -------------------------------------------------------- +💾 Converting with 83.0342 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,476B, BPFP=0.4232 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,300B, BPFP=2.2815 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,612B, BPFP=0.7706 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,048B, BPFP=2.2106 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,196B, BPFP=0.9169 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,956B, BPFP=2.2054 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,308B, BPFP=0.8666 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,308B, BPFP=2.2253 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,876B, BPFP=1.7480 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,360B, BPFP=2.1716 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,300B, BPFP=0.1723 +⌛️ [2/4] FRONTEND: Frontend time: 2.386s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.758s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12581717 29.87117867 + layer.0.v_cache 0.00001375 0.00678868 + layer.1.k_cache 0.57814745 10.91733628 + layer.1.v_cache 0.00000592 0.00282435 + layer.2.k_cache 0.00733882 1.55700628 + layer.2.v_cache 0.00002058 0.00968204 + layer.3.k_cache 0.02353939 7.73874034 + layer.3.v_cache 0.00002077 0.01037592 + layer.4.k_cache 0.00069182 0.32679812 + layer.4.v_cache 0.00005027 0.02089270 + layer.4.output 0.01015823 188.98152821 + ------------------------------------------------------------------------------------- + TOTAL 0.04745609 80.78425417 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 300740 +BPFP 1.0015 bits/point +EBPFP 2.0030 equivalent bits/point +MSE 80.784254 +---------------------- -------------------------------------------------------- +Time: 4.153s Load: 0.010s, Pack+Encode: 2.386s, Decode+Unpack: 1.758s +---------------------- -------------------------------------------------------- +💾 Converting with 80.7843 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,524B, BPFP=0.4291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,068B, BPFP=2.2849 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,392B, BPFP=0.7637 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,656B, BPFP=2.2044 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,612B, BPFP=0.9473 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,604B, BPFP=2.2014 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,188B, BPFP=0.8661 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,256B, BPFP=2.2386 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,068B, BPFP=1.7146 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,420B, BPFP=2.1909 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,008B, BPFP=0.2037 +⌛️ [2/4] FRONTEND: Frontend time: 2.637s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.928s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13625057 28.81320391 + layer.0.v_cache 0.00001362 0.00652444 + layer.1.k_cache 0.57512581 11.31359596 + layer.1.v_cache 0.00000584 0.00279396 + layer.2.k_cache 0.00872732 1.59343547 + layer.2.v_cache 0.00001851 0.00956207 + layer.3.k_cache 0.04362779 7.49846566 + layer.3.v_cache 0.00001925 0.01009184 + layer.4.k_cache 0.00073619 0.31961388 + layer.4.v_cache 0.00005148 0.02095361 + layer.4.output 0.00783937 196.85421011 + ------------------------------------------------------------------------------------- + TOTAL 0.04820306 83.97457127 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 302796 +BPFP 1.0157 bits/point +EBPFP 2.0314 equivalent bits/point +MSE 83.974571 +---------------------- -------------------------------------------------------- +Time: 4.578s Load: 0.013s, Pack+Encode: 2.637s, Decode+Unpack: 1.928s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9746 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,160B, BPFP=0.4194 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,936B, BPFP=2.0526 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,464B, BPFP=0.7434 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,768B, BPFP=1.9926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,316B, BPFP=0.8900 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,460B, BPFP=2.0282 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,148B, BPFP=0.8300 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,996B, BPFP=2.1071 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,332B, BPFP=1.7132 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,628B, BPFP=1.9854 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,728B, BPFP=0.2036 +⌛️ [2/4] FRONTEND: Frontend time: 2.692s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.750s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16730670 34.35702354 + layer.0.v_cache 0.00001388 0.00676147 + layer.1.k_cache 0.64158324 10.27014642 + layer.1.v_cache 0.00000571 0.00284754 + layer.2.k_cache 0.02207795 1.56084673 + layer.2.v_cache 0.00001922 0.00986918 + layer.3.k_cache 0.04503629 9.85492746 + layer.3.v_cache 0.00001955 0.01037494 + layer.4.k_cache 0.00070478 0.32564193 + layer.4.v_cache 0.00005260 0.02177877 + layer.4.output 0.04763387 176.05061971 + ------------------------------------------------------------------------------------- + TOTAL 0.07119159 75.81026800 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 314936 +BPFP 0.9522 bits/point +EBPFP 1.9044 equivalent bits/point +MSE 75.810268 +---------------------- -------------------------------------------------------- +Time: 4.453s Load: 0.010s, Pack+Encode: 2.692s, Decode+Unpack: 1.750s +---------------------- -------------------------------------------------------- +💾 Converting with 75.8103 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,544B, BPFP=0.4255 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,136B, BPFP=2.2640 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,140B, BPFP=0.7976 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,092B, BPFP=2.2051 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,884B, BPFP=0.9524 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,060B, BPFP=2.2033 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,284B, BPFP=0.8621 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,584B, BPFP=2.2329 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,156B, BPFP=1.7010 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,484B, BPFP=2.1708 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,520B, BPFP=0.1895 +⌛️ [2/4] FRONTEND: Frontend time: 2.665s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.741s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15225300 30.75566547 + layer.0.v_cache 0.00001427 0.00643728 + layer.1.k_cache 0.52731114 11.87977176 + layer.1.v_cache 0.00000556 0.00278129 + layer.2.k_cache 0.01483519 1.59962888 + layer.2.v_cache 0.00001864 0.01031175 + layer.3.k_cache 0.04699480 9.36511275 + layer.3.v_cache 0.00001968 0.01027689 + layer.4.k_cache 0.00070142 0.32527681 + layer.4.v_cache 0.00005177 0.02165580 + layer.4.output 0.01077793 197.48625258 + ------------------------------------------------------------------------------------- + TOTAL 0.04809712 84.49298157 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 303884 +BPFP 1.0083 bits/point +EBPFP 2.0166 equivalent bits/point +MSE 84.492982 +---------------------- -------------------------------------------------------- +Time: 4.416s Load: 0.010s, Pack+Encode: 2.665s, Decode+Unpack: 1.741s +---------------------- -------------------------------------------------------- +💾 Converting with 84.4930 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,496B, BPFP=0.4228 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,140B, BPFP=2.2642 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,536B, BPFP=0.7635 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,928B, BPFP=2.1958 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,012B, BPFP=0.9032 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,728B, BPFP=2.1846 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,064B, BPFP=0.8497 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,636B, BPFP=2.2358 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,660B, BPFP=1.6731 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,524B, BPFP=2.1731 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,968B, BPFP=0.1931 +⌛️ [2/4] FRONTEND: Frontend time: 2.445s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.731s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15700974 31.95049851 + layer.0.v_cache 0.00001374 0.00658997 + layer.1.k_cache 0.52255304 11.01206601 + layer.1.v_cache 0.00000566 0.00278588 + layer.2.k_cache 0.01360668 1.56178961 + layer.2.v_cache 0.00001926 0.01054422 + layer.3.k_cache 0.08085050 9.33165843 + layer.3.v_cache 0.00001943 0.01066505 + layer.4.k_cache 0.00071029 0.32090297 + layer.4.v_cache 0.00005450 0.02327449 + layer.4.output 0.01014708 197.66545255 + ------------------------------------------------------------------------------------- + TOTAL 0.04975720 84.58170253 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 301692 +BPFP 1.0010 bits/point +EBPFP 2.0021 equivalent bits/point +MSE 84.581703 +---------------------- -------------------------------------------------------- +Time: 4.186s Load: 0.010s, Pack+Encode: 2.445s, Decode+Unpack: 1.731s +---------------------- -------------------------------------------------------- +💾 Converting with 84.5817 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,660B, BPFP=0.4200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,248B, BPFP=2.2066 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,408B, BPFP=0.7899 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,384B, BPFP=2.1592 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,804B, BPFP=0.9213 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,812B, BPFP=2.1827 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,584B, BPFP=0.8544 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,192B, BPFP=2.2035 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,448B, BPFP=1.6693 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,800B, BPFP=2.1272 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,068B, BPFP=0.1963 +⌛️ [2/4] FRONTEND: Frontend time: 2.214s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.838s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16436556 33.06691338 + layer.0.v_cache 0.00001441 0.00666948 + layer.1.k_cache 0.62741763 12.80159591 + layer.1.v_cache 0.00000560 0.00280355 + layer.2.k_cache 0.01487184 1.60408743 + layer.2.v_cache 0.00002167 0.00986263 + layer.3.k_cache 0.02142572 9.76617153 + layer.3.v_cache 0.00001943 0.01050989 + layer.4.k_cache 0.00070776 0.32730338 + layer.4.v_cache 0.00005251 0.02112333 + layer.4.output 0.00894435 190.07708333 + ------------------------------------------------------------------------------------- + TOTAL 0.05244192 81.65627199 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 308408 +BPFP 0.9946 bits/point +EBPFP 1.9892 equivalent bits/point +MSE 81.656272 +---------------------- -------------------------------------------------------- +Time: 4.062s Load: 0.010s, Pack+Encode: 2.214s, Decode+Unpack: 1.838s +---------------------- -------------------------------------------------------- +💾 Converting with 81.6563 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,588B, BPFP=0.4219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,316B, BPFP=2.2418 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,832B, BPFP=0.7691 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,052B, BPFP=2.1715 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,924B, BPFP=0.9411 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,352B, BPFP=2.1882 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,132B, BPFP=0.8414 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,240B, BPFP=2.2375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,928B, BPFP=1.7754 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,776B, BPFP=2.1561 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,276B, BPFP=0.1690 +⌛️ [2/4] FRONTEND: Frontend time: 2.272s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.807s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11705062 30.24081475 + layer.0.v_cache 0.00001445 0.00664259 + layer.1.k_cache 0.56620099 12.25842242 + layer.1.v_cache 0.00000566 0.00277709 + layer.2.k_cache 0.01299965 1.61740471 + layer.2.v_cache 0.00001972 0.01008980 + layer.3.k_cache 0.05069733 9.73054173 + layer.3.v_cache 0.00001944 0.01023099 + layer.4.k_cache 0.00068883 0.33816718 + layer.4.v_cache 0.00005144 0.02051566 + layer.4.output 0.00838925 194.61221085 + ------------------------------------------------------------------------------------- + TOTAL 0.04743958 83.32476958 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 304416 +BPFP 0.9957 bits/point +EBPFP 1.9914 equivalent bits/point +MSE 83.324770 +---------------------- -------------------------------------------------------- +Time: 4.091s Load: 0.012s, Pack+Encode: 2.272s, Decode+Unpack: 1.807s +---------------------- -------------------------------------------------------- +💾 Converting with 83.3248 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,688B, BPFP=0.4245 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,156B, BPFP=2.2171 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,136B, BPFP=0.7805 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,924B, BPFP=2.1491 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,872B, BPFP=0.9315 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,268B, BPFP=2.1681 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,072B, BPFP=0.8874 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,368B, BPFP=2.1736 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,940B, BPFP=1.7083 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,700B, BPFP=2.1367 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,412B, BPFP=0.1768 +⌛️ [2/4] FRONTEND: Frontend time: 2.303s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.800s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14370025 31.15243692 + layer.0.v_cache 0.00001382 0.00665425 + layer.1.k_cache 0.61448195 12.90473598 + layer.1.v_cache 0.00000570 0.00285640 + layer.2.k_cache 0.01208573 1.60464704 + layer.2.v_cache 0.00001823 0.00986753 + layer.3.k_cache 0.01319265 10.05096328 + layer.3.v_cache 0.00002242 0.01045759 + layer.4.k_cache 0.00071920 0.32173262 + layer.4.v_cache 0.00005145 0.02137576 + layer.4.output 0.00992508 191.54128281 + ------------------------------------------------------------------------------------- + TOTAL 0.05022159 82.16910042 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 304536 +BPFP 0.9891 bits/point +EBPFP 1.9781 equivalent bits/point +MSE 82.169100 +---------------------- -------------------------------------------------------- +Time: 4.114s Load: 0.011s, Pack+Encode: 2.303s, Decode+Unpack: 1.800s +---------------------- -------------------------------------------------------- +💾 Converting with 82.1691 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 259, 128) +Output shape: (1, 259, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.output: torch.Size([1, 259, 3584]) -> torch.Size([1, 1, 259, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,020B, BPFP=0.4235 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,560B, BPFP=2.3263 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,936B, BPFP=0.8407 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,432B, BPFP=2.2582 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,936B, BPFP=0.9614 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,116B, BPFP=2.2391 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,880B, BPFP=0.8977 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,624B, BPFP=2.2698 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,408B, BPFP=1.7741 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,576B, BPFP=2.2066 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,560B, BPFP=0.2030 +⌛️ [2/4] FRONTEND: Frontend time: 2.232s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.030s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13349269 31.06279976 + layer.0.v_cache 0.00001436 0.00691630 + layer.1.k_cache 0.50562990 12.56187975 + layer.1.v_cache 0.00000590 0.00291329 + layer.2.k_cache 0.00820730 1.57227788 + layer.2.v_cache 0.00001794 0.00987291 + layer.3.k_cache 0.03685973 8.98567677 + layer.3.v_cache 0.00002061 0.01087637 + layer.4.k_cache 0.00069539 0.32351861 + layer.4.v_cache 0.00005422 0.02098049 + layer.4.output 0.01018517 200.56015582 + ------------------------------------------------------------------------------------- + TOTAL 0.04448790 85.79287076 + (elements=2,254,336) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2254336 +Total Bytes 292048 +BPFP 1.0364 bits/point +EBPFP 2.0728 equivalent bits/point +MSE 85.792871 +---------------------- -------------------------------------------------------- +Time: 4.274s Load: 0.012s, Pack+Encode: 2.232s, Decode+Unpack: 2.030s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7929 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,432B, BPFP=0.4301 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,732B, BPFP=2.2993 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,008B, BPFP=0.8106 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,500B, BPFP=2.2280 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,988B, BPFP=0.9831 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,340B, BPFP=2.2188 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,464B, BPFP=0.8949 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,076B, BPFP=2.2613 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,912B, BPFP=1.7310 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,892B, BPFP=2.1928 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,192B, BPFP=0.1917 +⌛️ [2/4] FRONTEND: Frontend time: 2.363s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.604s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12649573 29.90908203 + layer.0.v_cache 0.00001373 0.00674764 + layer.1.k_cache 0.51574391 10.81951226 + layer.1.v_cache 0.00000589 0.00285280 + layer.2.k_cache 0.01194659 1.59649353 + layer.2.v_cache 0.00001984 0.01001826 + layer.3.k_cache 0.04962628 9.61491880 + layer.3.v_cache 0.00002059 0.01017869 + layer.4.k_cache 0.00068541 0.30558726 + layer.4.v_cache 0.00005704 0.02143573 + layer.4.output 0.00772497 200.26035053 + ------------------------------------------------------------------------------------- + TOTAL 0.04462881 85.53642828 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 300536 +BPFP 1.0231 bits/point +EBPFP 2.0461 equivalent bits/point +MSE 85.536428 +---------------------- -------------------------------------------------------- +Time: 3.982s Load: 0.014s, Pack+Encode: 2.363s, Decode+Unpack: 1.604s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5364 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,428B, BPFP=0.4315 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,616B, BPFP=2.3011 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,360B, BPFP=0.7760 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,168B, BPFP=2.2170 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,244B, BPFP=0.9435 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,148B, BPFP=2.2158 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,188B, BPFP=0.8822 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,228B, BPFP=2.2786 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,668B, BPFP=1.7233 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,812B, BPFP=2.1963 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,416B, BPFP=0.1860 +⌛️ [2/4] FRONTEND: Frontend time: 2.444s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.887s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12210708 30.22065956 + layer.0.v_cache 0.00001342 0.00671541 + layer.1.k_cache 0.54942129 11.83751307 + layer.1.v_cache 0.00000547 0.00283700 + layer.2.k_cache 0.01053270 1.51224844 + layer.2.v_cache 0.00001820 0.01004976 + layer.3.k_cache 0.02850239 9.33766010 + layer.3.v_cache 0.00002491 0.01056124 + layer.4.k_cache 0.00067987 0.31499217 + layer.4.v_cache 0.00005179 0.02064969 + layer.4.output 0.01126663 201.89294012 + ------------------------------------------------------------------------------------- + TOTAL 0.04648373 86.26614513 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 297276 +BPFP 1.0157 bits/point +EBPFP 2.0315 equivalent bits/point +MSE 86.266145 +---------------------- -------------------------------------------------------- +Time: 4.341s Load: 0.009s, Pack+Encode: 2.444s, Decode+Unpack: 1.887s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2661 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,116B, BPFP=0.4185 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,944B, BPFP=2.0598 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,080B, BPFP=0.7261 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,916B, BPFP=2.0068 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,008B, BPFP=0.8771 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,504B, BPFP=2.0887 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,140B, BPFP=0.8323 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,688B, BPFP=2.0466 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,908B, BPFP=1.6970 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,580B, BPFP=1.9895 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,888B, BPFP=0.1760 +⌛️ [2/4] FRONTEND: Frontend time: 2.313s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.682s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14443027 34.71058942 + layer.0.v_cache 0.00001375 0.00671047 + layer.1.k_cache 0.62837083 10.54254009 + layer.1.v_cache 0.00000570 0.00275942 + layer.2.k_cache 0.01557094 1.59992120 + layer.2.v_cache 0.00001871 0.00967568 + layer.3.k_cache 0.04676386 9.67807440 + layer.3.v_cache 0.00001981 0.00979635 + layer.4.k_cache 0.00070582 0.32757949 + layer.4.v_cache 0.00005437 0.02045291 + layer.4.output 0.04838977 173.52562176 + ------------------------------------------------------------------------------------- + TOTAL 0.06909897 74.79926187 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 309772 +BPFP 0.9397 bits/point +EBPFP 1.8793 equivalent bits/point +MSE 74.799262 +---------------------- -------------------------------------------------------- +Time: 4.005s Load: 0.010s, Pack+Encode: 2.313s, Decode+Unpack: 1.682s +---------------------- -------------------------------------------------------- +💾 Converting with 74.7993 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,576B, BPFP=0.4183 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,440B, BPFP=2.2328 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,336B, BPFP=0.7915 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,416B, BPFP=2.1762 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,508B, BPFP=0.9114 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,176B, BPFP=2.1630 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,768B, BPFP=0.8706 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,056B, BPFP=2.2116 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,640B, BPFP=1.7469 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,848B, BPFP=2.1449 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,252B, BPFP=0.1676 +⌛️ [2/4] FRONTEND: Frontend time: 2.850s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.720s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13293209 31.53699895 + layer.0.v_cache 0.00001388 0.00655938 + layer.1.k_cache 0.56267701 12.72382605 + layer.1.v_cache 0.00000566 0.00269230 + layer.2.k_cache 0.01145046 1.62805467 + layer.2.v_cache 0.00001921 0.00964177 + layer.3.k_cache 0.02674564 10.10672482 + layer.3.v_cache 0.00001926 0.00983693 + layer.4.k_cache 0.00069135 0.32886618 + layer.4.v_cache 0.00005493 0.02260077 + layer.4.output 0.00973303 190.20292781 + ------------------------------------------------------------------------------------- + TOTAL 0.04722004 81.63507627 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 305016 +BPFP 0.9906 bits/point +EBPFP 1.9812 equivalent bits/point +MSE 81.635076 +---------------------- -------------------------------------------------------- +Time: 4.584s Load: 0.014s, Pack+Encode: 2.850s, Decode+Unpack: 1.720s +---------------------- -------------------------------------------------------- +💾 Converting with 81.6351 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,508B, BPFP=0.4235 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,196B, BPFP=2.2674 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,148B, BPFP=0.7981 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,148B, BPFP=2.2083 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,028B, BPFP=0.9605 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,576B, BPFP=2.1760 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,368B, BPFP=0.8669 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,580B, BPFP=2.2326 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,228B, BPFP=1.8179 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,484B, BPFP=2.1708 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,560B, BPFP=0.1657 +⌛️ [2/4] FRONTEND: Frontend time: 2.341s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.814s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15182159 29.22815250 + layer.0.v_cache 0.00001420 0.00676943 + layer.1.k_cache 0.56224567 12.71989579 + layer.1.v_cache 0.00000588 0.00287184 + layer.2.k_cache 0.00652889 1.60465326 + layer.2.v_cache 0.00001844 0.00982133 + layer.3.k_cache 0.06011557 8.96192728 + layer.3.v_cache 0.00001921 0.01087543 + layer.4.k_cache 0.00068317 0.32225254 + layer.4.v_cache 0.00005096 0.02181337 + layer.4.output 0.01130176 192.36054345 + ------------------------------------------------------------------------------------- + TOTAL 0.05062446 82.31840217 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 302824 +BPFP 1.0048 bits/point +EBPFP 2.0096 equivalent bits/point +MSE 82.318402 +---------------------- -------------------------------------------------------- +Time: 4.167s Load: 0.013s, Pack+Encode: 2.341s, Decode+Unpack: 1.814s +---------------------- -------------------------------------------------------- +💾 Converting with 82.3184 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,620B, BPFP=0.4178 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,400B, BPFP=2.2149 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,468B, BPFP=0.7932 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,268B, BPFP=2.1529 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,652B, BPFP=0.9129 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,960B, BPFP=2.1360 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,772B, BPFP=0.8647 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,848B, BPFP=2.1846 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,256B, BPFP=1.7684 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,860B, BPFP=2.1305 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,216B, BPFP=0.2210 +⌛️ [2/4] FRONTEND: Frontend time: 2.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.789s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13053865 34.07526727 + layer.0.v_cache 0.00001414 0.00715754 + layer.1.k_cache 0.56363723 12.70026984 + layer.1.v_cache 0.00000590 0.00296462 + layer.2.k_cache 0.01033439 1.63136083 + layer.2.v_cache 0.00002015 0.00980134 + layer.3.k_cache 0.03655195 10.05156421 + layer.3.v_cache 0.00002026 0.01037102 + layer.4.k_cache 0.00068516 0.30779258 + layer.4.v_cache 0.00005378 0.02038508 + layer.4.output 0.00863618 187.38117168 + ------------------------------------------------------------------------------------- + TOTAL 0.04719499 80.61677271 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 312320 +BPFP 1.0072 bits/point +EBPFP 2.0144 equivalent bits/point +MSE 80.616773 +---------------------- -------------------------------------------------------- +Time: 4.056s Load: 0.011s, Pack+Encode: 2.256s, Decode+Unpack: 1.789s +---------------------- -------------------------------------------------------- +💾 Converting with 80.6168 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,524B, BPFP=0.4244 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,316B, BPFP=2.2741 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,696B, BPFP=0.7726 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,104B, BPFP=2.2058 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,444B, BPFP=0.9276 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,328B, BPFP=2.2184 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,448B, BPFP=0.8714 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,344B, BPFP=2.2193 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,812B, BPFP=1.7380 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,512B, BPFP=2.1724 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,724B, BPFP=0.1831 +⌛️ [2/4] FRONTEND: Frontend time: 2.468s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.823s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15908699 28.75149128 + layer.0.v_cache 0.00001433 0.00665139 + layer.1.k_cache 0.53822481 11.39001950 + layer.1.v_cache 0.00000570 0.00279856 + layer.2.k_cache 0.01139215 1.56798777 + layer.2.v_cache 0.00001845 0.00986743 + layer.3.k_cache 0.02907990 8.68445485 + layer.3.v_cache 0.00001963 0.01039540 + layer.4.k_cache 0.00069750 0.33376348 + layer.4.v_cache 0.00005397 0.02154528 + layer.4.output 0.01069233 196.33324201 + ------------------------------------------------------------------------------------- + TOTAL 0.04784940 83.83009818 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 303252 +BPFP 1.0062 bits/point +EBPFP 2.0124 equivalent bits/point +MSE 83.830098 +---------------------- -------------------------------------------------------- +Time: 4.302s Load: 0.011s, Pack+Encode: 2.468s, Decode+Unpack: 1.823s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8301 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,648B, BPFP=0.4178 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,280B, BPFP=2.2006 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,108B, BPFP=0.8254 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,400B, BPFP=2.1525 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,864B, BPFP=0.9213 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,512B, BPFP=2.1587 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,488B, BPFP=0.8462 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,048B, BPFP=2.1879 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,376B, BPFP=1.7142 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,752B, BPFP=2.1171 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,812B, BPFP=0.2015 +⌛️ [2/4] FRONTEND: Frontend time: 2.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.732s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16084659 33.41091770 + layer.0.v_cache 0.00001440 0.00662110 + layer.1.k_cache 0.59028764 13.23210432 + layer.1.v_cache 0.00000559 0.00274862 + layer.2.k_cache 0.00929020 1.62664966 + layer.2.v_cache 0.00002205 0.00956332 + layer.3.k_cache 0.04382113 9.86655325 + layer.3.v_cache 0.00002103 0.00985781 + layer.4.k_cache 0.00069271 0.31820380 + layer.4.v_cache 0.00005052 0.02105140 + layer.4.output 0.00984751 186.10455170 + ------------------------------------------------------------------------------------- + TOTAL 0.05141085 80.07271370 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 310288 +BPFP 0.9972 bits/point +EBPFP 1.9943 equivalent bits/point +MSE 80.072714 +---------------------- -------------------------------------------------------- +Time: 4.068s Load: 0.012s, Pack+Encode: 2.324s, Decode+Unpack: 1.732s +---------------------- -------------------------------------------------------- +💾 Converting with 80.0727 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,548B, BPFP=0.4095 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,244B, BPFP=2.1834 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,240B, BPFP=0.7726 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,268B, BPFP=2.1304 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,900B, BPFP=0.9169 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,404B, BPFP=2.1378 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,480B, BPFP=0.8398 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,856B, BPFP=2.1623 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,940B, BPFP=1.6243 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,616B, BPFP=2.0951 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,436B, BPFP=0.2281 +⌛️ [2/4] FRONTEND: Frontend time: 2.541s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.667s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14211286 32.09095595 + layer.0.v_cache 0.00001516 0.00651938 + layer.1.k_cache 0.60291523 12.89053175 + layer.1.v_cache 0.00000570 0.00275720 + layer.2.k_cache 0.01041983 1.61570782 + layer.2.v_cache 0.00001835 0.00921640 + layer.3.k_cache 0.03896446 9.81487359 + layer.3.v_cache 0.00001926 0.00987539 + layer.4.k_cache 0.00071470 0.32080666 + layer.4.v_cache 0.00005236 0.02125705 + layer.4.output 0.00754456 186.10436818 + ------------------------------------------------------------------------------------- + TOTAL 0.04988528 79.97135756 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 310932 +BPFP 0.9923 bits/point +EBPFP 1.9846 equivalent bits/point +MSE 79.971358 +---------------------- -------------------------------------------------------- +Time: 4.220s Load: 0.012s, Pack+Encode: 2.541s, Decode+Unpack: 1.667s +---------------------- -------------------------------------------------------- +💾 Converting with 79.9714 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 296, 128) +Output shape: (1, 296, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.output: torch.Size([1, 296, 3584]) -> torch.Size([1, 1, 296, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,960B, BPFP=0.4202 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,984B, BPFP=2.1106 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,224B, BPFP=0.7508 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,040B, BPFP=2.0608 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,056B, BPFP=0.9003 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,876B, BPFP=2.1577 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,892B, BPFP=0.8389 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,660B, BPFP=2.1463 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,660B, BPFP=1.9880 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,688B, BPFP=2.0422 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,036B, BPFP=0.1737 +⌛️ [2/4] FRONTEND: Frontend time: 2.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.754s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14616069 35.96702452 + layer.0.v_cache 0.00001366 0.00675976 + layer.1.k_cache 0.61022733 13.12988941 + layer.1.v_cache 0.00000576 0.00283582 + layer.2.k_cache 0.01378242 1.58737183 + layer.2.v_cache 0.00001907 0.01060101 + layer.3.k_cache 0.03184220 9.87032009 + layer.3.v_cache 0.00001934 0.01079162 + layer.4.k_cache 0.00072724 0.34335386 + layer.4.v_cache 0.00005080 0.02143369 + layer.4.output 0.04784788 182.17128680 + ------------------------------------------------------------------------------------- + TOTAL 0.06692845 78.59702290 + (elements=2,576,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2576384 +Total Bytes 315076 +BPFP 0.9784 bits/point +EBPFP 1.9567 equivalent bits/point +MSE 78.597023 +---------------------- -------------------------------------------------------- +Time: 4.025s Load: 0.012s, Pack+Encode: 2.260s, Decode+Unpack: 1.754s +---------------------- -------------------------------------------------------- +💾 Converting with 78.5970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,048B, BPFP=0.4220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,188B, BPFP=2.1072 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,376B, BPFP=0.7538 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,316B, BPFP=2.0615 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,504B, BPFP=0.9178 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,732B, BPFP=2.0833 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,136B, BPFP=0.8461 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,368B, BPFP=2.1166 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,724B, BPFP=1.7158 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,932B, BPFP=2.0413 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,012B, BPFP=0.1874 +⌛️ [2/4] FRONTEND: Frontend time: 2.325s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.716s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15387250 38.14238150 + layer.0.v_cache 0.00001382 0.00675336 + layer.1.k_cache 0.63844893 12.91006122 + layer.1.v_cache 0.00000583 0.00290221 + layer.2.k_cache 0.01105453 1.67444227 + layer.2.v_cache 0.00001885 0.00991389 + layer.3.k_cache 0.03893430 9.91982356 + layer.3.v_cache 0.00001923 0.01004605 + layer.4.k_cache 0.00070257 0.31948469 + layer.4.v_cache 0.00005146 0.02161809 + layer.4.output 0.04873301 180.96064537 + ------------------------------------------------------------------------------------- + TOTAL 0.06966195 78.22011438 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 312336 +BPFP 0.9633 bits/point +EBPFP 1.9267 equivalent bits/point +MSE 78.220114 +---------------------- -------------------------------------------------------- +Time: 4.055s Load: 0.013s, Pack+Encode: 2.325s, Decode+Unpack: 1.716s +---------------------- -------------------------------------------------------- +💾 Converting with 78.2201 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,464B, BPFP=0.4304 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,124B, BPFP=2.3134 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,676B, BPFP=0.7885 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,820B, BPFP=2.2382 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,460B, BPFP=0.9490 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,592B, BPFP=2.2251 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,188B, BPFP=0.8757 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,408B, BPFP=2.2721 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,584B, BPFP=1.7057 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,188B, BPFP=2.2018 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,288B, BPFP=0.2083 +⌛️ [2/4] FRONTEND: Frontend time: 2.185s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.821s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12629252 31.98958574 + layer.0.v_cache 0.00001451 0.00676165 + layer.1.k_cache 0.49097243 12.93173432 + layer.1.v_cache 0.00000577 0.00284613 + layer.2.k_cache 0.01060468 1.57403869 + layer.2.v_cache 0.00001909 0.00985875 + layer.3.k_cache 0.02748298 9.49881533 + layer.3.v_cache 0.00001876 0.00969244 + layer.4.k_cache 0.00069967 0.31401822 + layer.4.v_cache 0.00005414 0.02089204 + layer.4.output 0.01047103 202.24449789 + ------------------------------------------------------------------------------------- + TOTAL 0.04290952 86.59233697 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 302792 +BPFP 1.0269 bits/point +EBPFP 2.0539 equivalent bits/point +MSE 86.592337 +---------------------- -------------------------------------------------------- +Time: 4.017s Load: 0.011s, Pack+Encode: 2.185s, Decode+Unpack: 1.821s +---------------------- -------------------------------------------------------- +💾 Converting with 86.5923 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,560B, BPFP=0.4234 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,432B, BPFP=2.2643 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,364B, BPFP=0.8044 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,408B, BPFP=2.2070 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,672B, BPFP=0.9337 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,368B, BPFP=2.2047 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,464B, BPFP=0.8660 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,932B, BPFP=2.2363 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,188B, BPFP=1.8026 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,804B, BPFP=2.1732 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,980B, BPFP=0.1839 +⌛️ [2/4] FRONTEND: Frontend time: 2.474s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.760s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15022702 29.18255768 + layer.0.v_cache 0.00001395 0.00657822 + layer.1.k_cache 0.54271635 12.59811128 + layer.1.v_cache 0.00000573 0.00276135 + layer.2.k_cache 0.01332666 1.60496056 + layer.2.v_cache 0.00001832 0.00953386 + layer.3.k_cache 0.02579262 9.10948876 + layer.3.v_cache 0.00001838 0.00990735 + layer.4.k_cache 0.00069325 0.32522999 + layer.4.v_cache 0.00005155 0.02053190 + layer.4.output 0.01006869 196.72251024 + ------------------------------------------------------------------------------------- + TOTAL 0.04725557 84.11336663 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 307172 +BPFP 1.0119 bits/point +EBPFP 2.0239 equivalent bits/point +MSE 84.113367 +---------------------- -------------------------------------------------------- +Time: 4.246s Load: 0.013s, Pack+Encode: 2.474s, Decode+Unpack: 1.760s +---------------------- -------------------------------------------------------- +💾 Converting with 84.1134 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,916B, BPFP=0.4193 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,036B, BPFP=2.1206 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,400B, BPFP=0.7627 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,240B, BPFP=2.0784 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,816B, BPFP=0.8907 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,556B, BPFP=2.2540 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,648B, BPFP=0.8288 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,060B, BPFP=2.1218 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,244B, BPFP=1.7608 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,756B, BPFP=2.0528 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,736B, BPFP=0.1872 +⌛️ [2/4] FRONTEND: Frontend time: 2.432s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.047s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12416320 34.17300053 + layer.0.v_cache 0.00001370 0.00665093 + layer.1.k_cache 0.62725913 12.53031730 + layer.1.v_cache 0.00000565 0.00279736 + layer.2.k_cache 0.01102555 1.56822272 + layer.2.v_cache 0.00002033 0.00987724 + layer.3.k_cache 0.04118381 9.83405720 + layer.3.v_cache 0.00001949 0.00999603 + layer.4.k_cache 0.00070768 0.31953792 + layer.4.v_cache 0.00005397 0.02078147 + layer.4.output 0.05105524 181.96312046 + ------------------------------------------------------------------------------------- + TOTAL 0.06834348 78.36571070 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 313408 +BPFP 0.9765 bits/point +EBPFP 1.9529 equivalent bits/point +MSE 78.365711 +---------------------- -------------------------------------------------------- +Time: 4.490s Load: 0.010s, Pack+Encode: 2.432s, Decode+Unpack: 2.047s +---------------------- -------------------------------------------------------- +💾 Converting with 78.3657 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,572B, BPFP=0.4094 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,180B, BPFP=2.1724 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,804B, BPFP=0.8545 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,204B, BPFP=2.1196 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,908B, BPFP=0.9141 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,104B, BPFP=2.1142 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,612B, BPFP=0.8441 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,844B, BPFP=2.1542 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,068B, BPFP=1.6797 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,640B, BPFP=2.0891 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,428B, BPFP=0.1887 +⌛️ [2/4] FRONTEND: Frontend time: 2.278s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.897s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15516999 32.76806472 + layer.0.v_cache 0.00001361 0.00674615 + layer.1.k_cache 0.56212566 12.56348754 + layer.1.v_cache 0.00000565 0.00289973 + layer.2.k_cache 0.00839357 1.62354698 + layer.2.v_cache 0.00001980 0.00957058 + layer.3.k_cache 0.02402391 9.91745681 + layer.3.v_cache 0.00001987 0.00977237 + layer.4.k_cache 0.00071212 0.31099387 + layer.4.v_cache 0.00005188 0.02056790 + layer.4.output 0.05073151 182.69890015 + ------------------------------------------------------------------------------------- + TOTAL 0.06503863 78.59561222 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 308364 +BPFP 0.9807 bits/point +EBPFP 1.9614 equivalent bits/point +MSE 78.595612 +---------------------- -------------------------------------------------------- +Time: 4.184s Load: 0.009s, Pack+Encode: 2.278s, Decode+Unpack: 1.897s +---------------------- -------------------------------------------------------- +💾 Converting with 78.5956 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,960B, BPFP=0.4188 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,112B, BPFP=2.1103 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,420B, BPFP=0.7586 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,116B, BPFP=2.0579 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,412B, BPFP=0.9160 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,436B, BPFP=2.1799 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,308B, BPFP=0.8580 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,352B, BPFP=2.1229 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,768B, BPFP=1.9870 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,800B, BPFP=2.0412 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,972B, BPFP=0.1802 +⌛️ [2/4] FRONTEND: Frontend time: 2.277s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.728s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005571 35.39597143 + layer.0.v_cache 0.00001459 0.00664829 + layer.1.k_cache 0.64038677 12.58887048 + layer.1.v_cache 0.00000545 0.00274550 + layer.2.k_cache 0.01623841 1.61547739 + layer.2.v_cache 0.00002030 0.01012905 + layer.3.k_cache 0.01774169 9.89660603 + layer.3.v_cache 0.00001948 0.01016378 + layer.4.k_cache 0.00069419 0.33058552 + layer.4.v_cache 0.00005057 0.02144953 + layer.4.output 0.05052170 181.65883538 + ------------------------------------------------------------------------------------- + TOTAL 0.06875759 78.32297027 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 317656 +BPFP 0.9830 bits/point +EBPFP 1.9661 equivalent bits/point +MSE 78.322970 +---------------------- -------------------------------------------------------- +Time: 4.019s Load: 0.014s, Pack+Encode: 2.277s, Decode+Unpack: 1.728s +---------------------- -------------------------------------------------------- +💾 Converting with 78.3230 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,536B, BPFP=0.4220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,212B, BPFP=2.2520 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,132B, BPFP=0.7914 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,116B, BPFP=2.1906 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,292B, BPFP=0.9684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,836B, BPFP=2.1750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,628B, BPFP=0.8752 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,844B, BPFP=2.2314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,800B, BPFP=1.8369 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,856B, BPFP=2.1761 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,204B, BPFP=0.1776 +⌛️ [2/4] FRONTEND: Frontend time: 2.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.599s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14010120 29.95393005 + layer.0.v_cache 0.00001437 0.00679059 + layer.1.k_cache 0.58692467 12.09878595 + layer.1.v_cache 0.00000569 0.00292302 + layer.2.k_cache 0.00806839 1.62693858 + layer.2.v_cache 0.00001934 0.01021795 + layer.3.k_cache 0.03832107 9.74880117 + layer.3.v_cache 0.00001992 0.01085710 + layer.4.k_cache 0.00069804 0.31448465 + layer.4.v_cache 0.00005131 0.02119317 + layer.4.output 0.00759059 197.20364503 + ------------------------------------------------------------------------------------- + TOTAL 0.04866812 84.36590809 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 306456 +BPFP 1.0096 bits/point +EBPFP 2.0191 equivalent bits/point +MSE 84.365908 +---------------------- -------------------------------------------------------- +Time: 3.832s Load: 0.013s, Pack+Encode: 2.221s, Decode+Unpack: 1.599s +---------------------- -------------------------------------------------------- +💾 Converting with 84.3659 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,384B, BPFP=0.4305 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,992B, BPFP=2.3316 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,944B, BPFP=0.8130 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,588B, BPFP=2.2498 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,656B, BPFP=0.9711 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,040B, BPFP=2.2178 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,304B, BPFP=0.8923 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,984B, BPFP=2.2729 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,924B, BPFP=1.7446 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,776B, BPFP=2.2024 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,128B, BPFP=0.1843 +⌛️ [2/4] FRONTEND: Frontend time: 2.278s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.747s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13199043 29.32853494 + layer.0.v_cache 0.00001386 0.00687978 + layer.1.k_cache 0.56989129 12.93344617 + layer.1.v_cache 0.00000577 0.00290777 + layer.2.k_cache 0.00892691 1.64239593 + layer.2.v_cache 0.00001816 0.00962568 + layer.3.k_cache 0.04626510 9.45518471 + layer.3.v_cache 0.00001927 0.01019583 + layer.4.k_cache 0.00069392 0.31538448 + layer.4.v_cache 0.00005195 0.02093281 + layer.4.output 0.00911696 201.87543310 + ------------------------------------------------------------------------------------- + TOTAL 0.04833502 86.28550117 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 298720 +BPFP 1.0245 bits/point +EBPFP 2.0489 equivalent bits/point +MSE 86.285501 +---------------------- -------------------------------------------------------- +Time: 4.037s Load: 0.012s, Pack+Encode: 2.278s, Decode+Unpack: 1.747s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2855 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,552B, BPFP=0.4307 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,184B, BPFP=2.3485 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,252B, BPFP=0.7557 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,792B, BPFP=2.2121 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,460B, BPFP=0.9386 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,596B, BPFP=2.2010 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,044B, BPFP=0.8579 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,108B, BPFP=2.2302 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,588B, BPFP=1.6873 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,996B, BPFP=2.1667 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,984B, BPFP=0.1872 +⌛️ [2/4] FRONTEND: Frontend time: 2.424s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.709s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16500157 30.81297402 + layer.0.v_cache 0.00001391 0.00680672 + layer.1.k_cache 0.57858722 11.03267831 + layer.1.v_cache 0.00000575 0.00281748 + layer.2.k_cache 0.01018238 1.60428870 + layer.2.v_cache 0.00001877 0.01004181 + layer.3.k_cache 0.03425542 8.47805229 + layer.3.v_cache 0.00002021 0.01008245 + layer.4.k_cache 0.00069807 0.32261624 + layer.4.v_cache 0.00005194 0.02106606 + layer.4.output 0.01023613 197.35492701 + ------------------------------------------------------------------------------------- + TOTAL 0.05061695 84.34034783 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 300556 +BPFP 1.0082 bits/point +EBPFP 2.0164 equivalent bits/point +MSE 84.340348 +---------------------- -------------------------------------------------------- +Time: 4.143s Load: 0.010s, Pack+Encode: 2.424s, Decode+Unpack: 1.709s +---------------------- -------------------------------------------------------- +💾 Converting with 84.3403 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,428B, BPFP=0.4299 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,480B, BPFP=2.2847 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,648B, BPFP=0.7898 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,200B, BPFP=2.2106 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,008B, BPFP=0.9264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,464B, BPFP=2.2259 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,432B, BPFP=0.8931 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,764B, BPFP=2.3590 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,832B, BPFP=1.6685 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,976B, BPFP=2.1977 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,032B, BPFP=0.2069 +⌛️ [2/4] FRONTEND: Frontend time: 2.352s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.734s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385483 31.97591146 + layer.0.v_cache 0.00001387 0.00655376 + layer.1.k_cache 0.55871571 12.00832248 + layer.1.v_cache 0.00000565 0.00270925 + layer.2.k_cache 0.00818202 1.57028130 + layer.2.v_cache 0.00001875 0.01007135 + layer.3.k_cache 0.04697755 9.70318920 + layer.3.v_cache 0.00001978 0.00986380 + layer.4.k_cache 0.00071636 0.31745408 + layer.4.v_cache 0.00005146 0.02112127 + layer.4.output 0.01106052 198.89432870 + ------------------------------------------------------------------------------------- + TOTAL 0.04741057 85.16975170 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 301264 +BPFP 1.0255 bits/point +EBPFP 2.0511 equivalent bits/point +MSE 85.169752 +---------------------- -------------------------------------------------------- +Time: 4.095s Load: 0.010s, Pack+Encode: 2.352s, Decode+Unpack: 1.734s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1698 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,556B, BPFP=0.4293 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,244B, BPFP=2.2866 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,696B, BPFP=0.7782 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,204B, BPFP=2.2275 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,412B, BPFP=0.9325 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,948B, BPFP=2.2130 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,072B, BPFP=0.8564 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,612B, BPFP=2.2507 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,676B, BPFP=1.6861 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,228B, BPFP=2.1720 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,504B, BPFP=0.1745 +⌛️ [2/4] FRONTEND: Frontend time: 2.432s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.635s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203608 31.90325284 + layer.0.v_cache 0.00001420 0.00658825 + layer.1.k_cache 0.56673284 12.25869851 + layer.1.v_cache 0.00000552 0.00276464 + layer.2.k_cache 0.00530624 1.61596147 + layer.2.v_cache 0.00001825 0.01008170 + layer.3.k_cache 0.02188354 7.78304954 + layer.3.v_cache 0.00001854 0.00990904 + layer.4.k_cache 0.00069888 0.31685397 + layer.4.v_cache 0.00004918 0.02141086 + layer.4.output 0.00799176 191.46727273 + ------------------------------------------------------------------------------------- + TOTAL 0.04662974 82.01173411 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 300152 +BPFP 1.0032 bits/point +EBPFP 2.0064 equivalent bits/point +MSE 82.011734 +---------------------- -------------------------------------------------------- +Time: 4.077s Load: 0.010s, Pack+Encode: 2.432s, Decode+Unpack: 1.635s +---------------------- -------------------------------------------------------- +💾 Converting with 82.0117 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,580B, BPFP=0.4185 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,304B, BPFP=2.2253 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,692B, BPFP=0.7560 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,164B, BPFP=2.1623 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,816B, BPFP=0.9284 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,412B, BPFP=2.1760 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,596B, BPFP=0.8611 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,008B, BPFP=2.2089 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,576B, BPFP=1.6330 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,872B, BPFP=2.1462 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,016B, BPFP=0.1973 +⌛️ [2/4] FRONTEND: Frontend time: 2.411s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.608s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12732459 31.44459819 + layer.0.v_cache 0.00001480 0.00654049 + layer.1.k_cache 0.57322790 12.98761957 + layer.1.v_cache 0.00000557 0.00284292 + layer.2.k_cache 0.00950604 1.62984690 + layer.2.v_cache 0.00002039 0.01028836 + layer.3.k_cache 0.01510038 9.88923003 + layer.3.v_cache 0.00001916 0.01020840 + layer.4.k_cache 0.00068141 0.31528858 + layer.4.v_cache 0.00005165 0.01981701 + layer.4.output 0.00936729 187.78940560 + ------------------------------------------------------------------------------------- + TOTAL 0.04656017 80.63777175 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 306036 +BPFP 0.9939 bits/point +EBPFP 1.9879 equivalent bits/point +MSE 80.637772 +---------------------- -------------------------------------------------------- +Time: 4.030s Load: 0.011s, Pack+Encode: 2.411s, Decode+Unpack: 1.608s +---------------------- -------------------------------------------------------- +💾 Converting with 80.6378 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,524B, BPFP=0.4260 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,400B, BPFP=2.2871 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,988B, BPFP=0.7919 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,164B, BPFP=2.2172 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,524B, BPFP=0.9355 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,988B, BPFP=2.2072 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,204B, BPFP=0.8607 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,668B, BPFP=2.2457 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,512B, BPFP=1.7274 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,768B, BPFP=2.1947 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,340B, BPFP=0.1888 +⌛️ [2/4] FRONTEND: Frontend time: 2.301s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.660s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13630490 31.03991522 + layer.0.v_cache 0.00001544 0.00670795 + layer.1.k_cache 0.58776402 11.54801255 + layer.1.v_cache 0.00000559 0.00271034 + layer.2.k_cache 0.01000112 1.52427707 + layer.2.v_cache 0.00001940 0.00999642 + layer.3.k_cache 0.04424953 8.96365931 + layer.3.v_cache 0.00002139 0.01039063 + layer.4.k_cache 0.00071093 0.31075016 + layer.4.v_cache 0.00005158 0.02168776 + layer.4.output 0.01101471 195.09582039 + ------------------------------------------------------------------------------------- + TOTAL 0.05036746 83.47699119 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 304080 +BPFP 1.0126 bits/point +EBPFP 2.0253 equivalent bits/point +MSE 83.476991 +---------------------- -------------------------------------------------------- +Time: 3.973s Load: 0.013s, Pack+Encode: 2.301s, Decode+Unpack: 1.660s +---------------------- -------------------------------------------------------- +💾 Converting with 83.4770 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,532B, BPFP=0.4233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,304B, BPFP=2.2653 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,024B, BPFP=0.7882 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,248B, BPFP=2.2059 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,596B, BPFP=0.9328 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,772B, BPFP=2.1792 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,136B, BPFP=0.8507 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,652B, BPFP=2.2286 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,672B, BPFP=1.7239 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,732B, BPFP=2.1769 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,408B, BPFP=0.1960 +⌛️ [2/4] FRONTEND: Frontend time: 2.326s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.868s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12958164 29.67704586 + layer.0.v_cache 0.00001430 0.00676940 + layer.1.k_cache 0.54121377 11.64668109 + layer.1.v_cache 0.00000577 0.00283212 + layer.2.k_cache 0.01080061 1.60442226 + layer.2.v_cache 0.00001903 0.01009155 + layer.3.k_cache 0.01783211 8.00614303 + layer.3.v_cache 0.00002028 0.01021802 + layer.4.k_cache 0.00068935 0.32034779 + layer.4.v_cache 0.00005053 0.02125353 + layer.4.output 0.00833506 193.63286871 + ------------------------------------------------------------------------------------- + TOTAL 0.04462193 82.74916974 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 305076 +BPFP 1.0086 bits/point +EBPFP 2.0173 equivalent bits/point +MSE 82.749170 +---------------------- -------------------------------------------------------- +Time: 4.204s Load: 0.011s, Pack+Encode: 2.326s, Decode+Unpack: 1.868s +---------------------- -------------------------------------------------------- +💾 Converting with 82.7492 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,232B, BPFP=0.4297 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,172B, BPFP=2.3272 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,444B, BPFP=0.8581 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,696B, BPFP=2.2395 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,048B, BPFP=0.9534 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,308B, BPFP=2.2165 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,912B, BPFP=0.8859 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,396B, BPFP=2.2811 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,580B, BPFP=1.7574 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,088B, BPFP=2.2034 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,120B, BPFP=0.2047 +⌛️ [2/4] FRONTEND: Frontend time: 2.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.960s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12547956 31.69590289 + layer.0.v_cache 0.00001492 0.00683396 + layer.1.k_cache 0.50446328 12.78205947 + layer.1.v_cache 0.00000633 0.00285532 + layer.2.k_cache 0.00671430 1.58026065 + layer.2.v_cache 0.00001992 0.01010332 + layer.3.k_cache 0.02142675 8.29293997 + layer.3.v_cache 0.00001986 0.01050024 + layer.4.k_cache 0.00069859 0.30745007 + layer.4.v_cache 0.00005414 0.02081699 + layer.4.output 0.01038299 209.51045627 + ------------------------------------------------------------------------------------- + TOTAL 0.04303403 89.48723040 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 295996 +BPFP 1.0344 bits/point +EBPFP 2.0689 equivalent bits/point +MSE 89.487230 +---------------------- -------------------------------------------------------- +Time: 4.230s Load: 0.012s, Pack+Encode: 2.258s, Decode+Unpack: 1.960s +---------------------- -------------------------------------------------------- +💾 Converting with 89.4872 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.018s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,112B, BPFP=0.4211 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,068B, BPFP=2.0799 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,316B, BPFP=0.7431 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,268B, BPFP=2.0384 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,716B, BPFP=0.9196 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,456B, BPFP=2.1520 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,296B, BPFP=0.8459 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,260B, BPFP=2.0899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,376B, BPFP=1.7326 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,876B, BPFP=2.0181 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,472B, BPFP=0.1963 +⌛️ [2/4] FRONTEND: Frontend time: 2.357s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.846s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13970643 35.38939537 + layer.0.v_cache 0.00001441 0.00666286 + layer.1.k_cache 0.60015088 11.18744485 + layer.1.v_cache 0.00000584 0.00280318 + layer.2.k_cache 0.00648033 1.61252267 + layer.2.v_cache 0.00001840 0.01006371 + layer.3.k_cache 0.03288257 9.95770041 + layer.3.v_cache 0.00002024 0.00993013 + layer.4.k_cache 0.00070287 0.31899989 + layer.4.v_cache 0.00005060 0.02156237 + layer.4.output 0.04874230 178.78924419 + ------------------------------------------------------------------------------------- + TOTAL 0.06595463 77.06128204 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 316216 +BPFP 0.9656 bits/point +EBPFP 1.9312 equivalent bits/point +MSE 77.061282 +---------------------- -------------------------------------------------------- +Time: 4.221s Load: 0.018s, Pack+Encode: 2.357s, Decode+Unpack: 1.846s +---------------------- -------------------------------------------------------- +💾 Converting with 77.0613 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,868B, BPFP=0.4237 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,584B, BPFP=2.2737 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,976B, BPFP=0.8112 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,788B, BPFP=2.1879 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,476B, BPFP=0.9306 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,488B, BPFP=2.1735 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,944B, BPFP=0.8574 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,656B, BPFP=2.2294 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,460B, BPFP=1.6944 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,876B, BPFP=2.1443 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,476B, BPFP=0.1739 +⌛️ [2/4] FRONTEND: Frontend time: 2.443s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.130s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13960611 33.55176080 + layer.0.v_cache 0.00001491 0.00680881 + layer.1.k_cache 0.68890666 12.99104519 + layer.1.v_cache 0.00000583 0.00287089 + layer.2.k_cache 0.00940153 1.56876590 + layer.2.v_cache 0.00001882 0.00978022 + layer.3.k_cache 0.04667800 9.16199216 + layer.3.v_cache 0.00001939 0.00994362 + layer.4.k_cache 0.00070827 0.31091215 + layer.4.v_cache 0.00005291 0.02108949 + layer.4.output 0.04427218 162.50184305 + ------------------------------------------------------------------------------------- + TOTAL 0.07031281 70.30281592 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 354592 +BPFP 0.9967 bits/point +EBPFP 1.9933 equivalent bits/point +MSE 70.302816 +---------------------- -------------------------------------------------------- +Time: 4.586s Load: 0.013s, Pack+Encode: 2.443s, Decode+Unpack: 2.130s +---------------------- -------------------------------------------------------- +💾 Converting with 70.3028 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,000B, BPFP=0.4181 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,916B, BPFP=2.0859 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,676B, BPFP=0.7669 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,124B, BPFP=2.0445 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,664B, BPFP=0.9231 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,620B, BPFP=2.2272 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,804B, BPFP=0.8259 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,828B, BPFP=2.1336 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,244B, BPFP=1.7372 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,860B, BPFP=2.0307 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,988B, BPFP=0.1940 +⌛️ [2/4] FRONTEND: Frontend time: 2.647s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.932s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15817952 35.11382669 + layer.0.v_cache 0.00001474 0.00663754 + layer.1.k_cache 0.61609703 12.54773807 + layer.1.v_cache 0.00000566 0.00277700 + layer.2.k_cache 0.00844678 1.64074054 + layer.2.v_cache 0.00001876 0.00988124 + layer.3.k_cache 0.04682518 9.80192847 + layer.3.v_cache 0.00001896 0.00985082 + layer.4.k_cache 0.00069859 0.32471826 + layer.4.v_cache 0.00005216 0.02208984 + layer.4.output 0.04649196 174.56691949 + ------------------------------------------------------------------------------------- + TOTAL 0.06798830 75.37933088 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 316724 +BPFP 0.9736 bits/point +EBPFP 1.9472 equivalent bits/point +MSE 75.379331 +---------------------- -------------------------------------------------------- +Time: 4.589s Load: 0.010s, Pack+Encode: 2.647s, Decode+Unpack: 1.932s +---------------------- -------------------------------------------------------- +💾 Converting with 75.3793 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,552B, BPFP=0.4111 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,228B, BPFP=2.1901 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,632B, BPFP=0.7966 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,116B, BPFP=2.1296 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,788B, BPFP=0.9140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,008B, BPFP=2.1237 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,336B, BPFP=0.8349 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,796B, BPFP=2.1666 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,144B, BPFP=1.6956 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,592B, BPFP=2.1010 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,204B, BPFP=0.1805 +⌛️ [2/4] FRONTEND: Frontend time: 2.645s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.666s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14002854 31.98800904 + layer.0.v_cache 0.00001601 0.00675031 + layer.1.k_cache 0.54217949 13.38716875 + layer.1.v_cache 0.00000579 0.00284678 + layer.2.k_cache 0.01082110 1.63297413 + layer.2.v_cache 0.00001971 0.00963584 + layer.3.k_cache 0.03102738 9.91675485 + layer.3.v_cache 0.00001907 0.00972929 + layer.4.k_cache 0.00066782 0.30830811 + layer.4.v_cache 0.00005141 0.02009070 + layer.4.output 0.00856273 190.35984943 + ------------------------------------------------------------------------------------- + TOTAL 0.04616326 81.75301258 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 305396 +BPFP 0.9780 bits/point +EBPFP 1.9561 equivalent bits/point +MSE 81.753013 +---------------------- -------------------------------------------------------- +Time: 4.320s Load: 0.009s, Pack+Encode: 2.645s, Decode+Unpack: 1.666s +---------------------- -------------------------------------------------------- +💾 Converting with 81.7530 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,924B, BPFP=0.4169 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,024B, BPFP=2.1056 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,668B, BPFP=0.7717 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,128B, BPFP=2.0585 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,700B, BPFP=0.9312 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,960B, BPFP=2.1549 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,176B, BPFP=0.8510 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,500B, BPFP=2.1833 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,296B, BPFP=1.7517 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,672B, BPFP=2.0345 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,232B, BPFP=0.1896 +⌛️ [2/4] FRONTEND: Frontend time: 2.644s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.838s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14705232 36.47710175 + layer.0.v_cache 0.00001374 0.00665202 + layer.1.k_cache 0.65666707 12.55201100 + layer.1.v_cache 0.00000579 0.00277909 + layer.2.k_cache 0.01062412 1.62428659 + layer.2.v_cache 0.00001918 0.00993513 + layer.3.k_cache 0.03205616 10.02874776 + layer.3.v_cache 0.00001999 0.01062666 + layer.4.k_cache 0.00071739 0.31692225 + layer.4.v_cache 0.00005111 0.02122691 + layer.4.output 0.04976816 181.25225469 + ------------------------------------------------------------------------------------- + TOTAL 0.07032965 78.22447482 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 315280 +BPFP 0.9757 bits/point +EBPFP 1.9514 equivalent bits/point +MSE 78.224475 +---------------------- -------------------------------------------------------- +Time: 4.494s Load: 0.011s, Pack+Encode: 2.644s, Decode+Unpack: 1.838s +---------------------- -------------------------------------------------------- +💾 Converting with 78.2245 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,588B, BPFP=0.4219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,372B, BPFP=2.2449 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,176B, BPFP=0.7883 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,276B, BPFP=2.1839 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,920B, BPFP=0.9408 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,196B, BPFP=2.1795 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,316B, BPFP=0.8516 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,956B, BPFP=2.2218 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,924B, BPFP=1.6639 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,588B, BPFP=2.1457 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,076B, BPFP=0.1992 +⌛️ [2/4] FRONTEND: Frontend time: 2.318s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.693s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131647 30.11568269 + layer.0.v_cache 0.00001392 0.00655897 + layer.1.k_cache 0.51967005 12.43456684 + layer.1.v_cache 0.00000557 0.00273785 + layer.2.k_cache 0.01013049 1.58439446 + layer.2.v_cache 0.00001908 0.01026375 + layer.3.k_cache 0.05186962 9.30533912 + layer.3.v_cache 0.00001891 0.01020329 + layer.4.k_cache 0.00069958 0.32176179 + layer.4.v_cache 0.00005017 0.02065347 + layer.4.output 0.01073789 193.70715239 + ------------------------------------------------------------------------------------- + TOTAL 0.04758583 82.92718994 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 306388 +BPFP 1.0022 bits/point +EBPFP 2.0043 equivalent bits/point +MSE 82.927190 +---------------------- -------------------------------------------------------- +Time: 4.022s Load: 0.011s, Pack+Encode: 2.318s, Decode+Unpack: 1.693s +---------------------- -------------------------------------------------------- +💾 Converting with 82.9272 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,524B, BPFP=0.4082 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,912B, BPFP=2.1654 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,016B, BPFP=0.8147 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,928B, BPFP=2.1120 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,132B, BPFP=0.9295 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,292B, BPFP=2.1317 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,616B, BPFP=0.8472 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,584B, BPFP=2.1476 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,644B, BPFP=1.6625 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,544B, BPFP=2.0911 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,584B, BPFP=0.2060 +⌛️ [2/4] FRONTEND: Frontend time: 2.415s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.749s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15566496 33.59281413 + layer.0.v_cache 0.00001437 0.00652203 + layer.1.k_cache 0.62643353 12.75777859 + layer.1.v_cache 0.00000573 0.00279269 + layer.2.k_cache 0.01010078 1.62052748 + layer.2.v_cache 0.00001868 0.01002888 + layer.3.k_cache 0.06271097 10.03432719 + layer.3.v_cache 0.00001958 0.01064368 + layer.4.k_cache 0.00073784 0.30203247 + layer.4.v_cache 0.00005294 0.02133392 + layer.4.output 0.00677609 188.54999070 + ------------------------------------------------------------------------------------- + TOTAL 0.05312894 81.07110212 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 308776 +BPFP 0.9854 bits/point +EBPFP 1.9708 equivalent bits/point +MSE 81.071102 +---------------------- -------------------------------------------------------- +Time: 4.174s Load: 0.010s, Pack+Encode: 2.415s, Decode+Unpack: 1.749s +---------------------- -------------------------------------------------------- +💾 Converting with 81.0711 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,636B, BPFP=0.4172 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,172B, BPFP=2.1947 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,364B, BPFP=0.7847 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,172B, BPFP=2.1401 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,928B, BPFP=0.9248 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,040B, BPFP=2.1329 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,312B, BPFP=0.8365 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,656B, BPFP=2.1665 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,984B, BPFP=1.8020 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,596B, BPFP=2.1086 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,068B, BPFP=0.1956 +⌛️ [2/4] FRONTEND: Frontend time: 2.429s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.677s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13918158 32.63604267 + layer.0.v_cache 0.00001566 0.00666399 + layer.1.k_cache 0.57997750 12.62383308 + layer.1.v_cache 0.00000571 0.00281311 + layer.2.k_cache 0.01637170 1.65507091 + layer.2.v_cache 0.00001887 0.00974838 + layer.3.k_cache 0.04393929 9.81923606 + layer.3.v_cache 0.00001903 0.00980514 + layer.4.k_cache 0.00071104 0.30888583 + layer.4.v_cache 0.00005175 0.02150050 + layer.4.output 0.01065216 187.98078484 + ------------------------------------------------------------------------------------- + TOTAL 0.05028572 80.76229962 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 308928 +BPFP 0.9928 bits/point +EBPFP 1.9856 equivalent bits/point +MSE 80.762300 +---------------------- -------------------------------------------------------- +Time: 4.116s Load: 0.011s, Pack+Encode: 2.429s, Decode+Unpack: 1.677s +---------------------- -------------------------------------------------------- +💾 Converting with 80.7623 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,276B, BPFP=0.4171 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,108B, BPFP=2.0216 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,472B, BPFP=0.7294 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,200B, BPFP=1.9758 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,164B, BPFP=0.8651 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,368B, BPFP=1.9843 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,188B, BPFP=0.8159 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,084B, BPFP=2.0204 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,676B, BPFP=1.6974 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,928B, BPFP=1.9621 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,668B, BPFP=0.1848 +⌛️ [2/4] FRONTEND: Frontend time: 2.313s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.743s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16496690 35.65546875 + layer.0.v_cache 0.00001433 0.00686299 + layer.1.k_cache 0.65265572 11.15627993 + layer.1.v_cache 0.00000578 0.00282528 + layer.2.k_cache 0.01241827 1.59274331 + layer.2.v_cache 0.00001920 0.00962511 + layer.3.k_cache 0.08444468 9.84535975 + layer.3.v_cache 0.00001975 0.01006697 + layer.4.k_cache 0.00068928 0.31930434 + layer.4.v_cache 0.00005259 0.02058596 + layer.4.output 0.04785494 169.74732143 + ------------------------------------------------------------------------------------- + TOTAL 0.07354536 73.34413955 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 313132 +BPFP 0.9284 bits/point +EBPFP 1.8568 equivalent bits/point +MSE 73.344140 +---------------------- -------------------------------------------------------- +Time: 4.070s Load: 0.014s, Pack+Encode: 2.313s, Decode+Unpack: 1.743s +---------------------- -------------------------------------------------------- +💾 Converting with 73.3441 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,928B, BPFP=0.4213 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,060B, BPFP=2.1290 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,964B, BPFP=0.7421 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,188B, BPFP=2.0827 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,948B, BPFP=0.9007 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,000B, BPFP=2.1790 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,904B, BPFP=0.8452 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,892B, BPFP=2.1201 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,192B, BPFP=1.8172 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,640B, BPFP=2.0536 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,048B, BPFP=0.1902 +⌛️ [2/4] FRONTEND: Frontend time: 2.449s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.872s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15001445 35.81467235 + layer.0.v_cache 0.00001441 0.00672256 + layer.1.k_cache 0.57475359 11.95180548 + layer.1.v_cache 0.00000584 0.00288237 + layer.2.k_cache 0.01434258 1.61140566 + layer.2.v_cache 0.00001954 0.00987533 + layer.3.k_cache 0.02649183 9.94242516 + layer.3.v_cache 0.00001945 0.01013293 + layer.4.k_cache 0.00069414 0.31541448 + layer.4.v_cache 0.00005306 0.02126465 + layer.4.output 0.05035525 178.78814383 + ------------------------------------------------------------------------------------- + TOTAL 0.06581739 77.12962399 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 312764 +BPFP 0.9778 bits/point +EBPFP 1.9556 equivalent bits/point +MSE 77.129624 +---------------------- -------------------------------------------------------- +Time: 4.334s Load: 0.014s, Pack+Encode: 2.449s, Decode+Unpack: 1.872s +---------------------- -------------------------------------------------------- +💾 Converting with 77.1296 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,628B, BPFP=0.4227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,316B, BPFP=2.2338 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,448B, BPFP=0.8005 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,288B, BPFP=2.1769 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,412B, BPFP=0.9094 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,176B, BPFP=2.1707 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,788B, BPFP=0.8748 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,352B, BPFP=2.2358 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,724B, BPFP=1.7023 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,936B, BPFP=2.1574 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,932B, BPFP=0.1736 +⌛️ [2/4] FRONTEND: Frontend time: 2.305s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.787s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11854591 32.53251053 + layer.0.v_cache 0.00001566 0.00662751 + layer.1.k_cache 0.56985614 12.82543253 + layer.1.v_cache 0.00000555 0.00277868 + layer.2.k_cache 0.01267573 1.58717509 + layer.2.v_cache 0.00001902 0.00984972 + layer.3.k_cache 0.04477839 9.81329216 + layer.3.v_cache 0.00001896 0.00991779 + layer.4.k_cache 0.00071754 0.31801078 + layer.4.v_cache 0.00005321 0.02159369 + layer.4.output 0.00761333 194.85676292 + ------------------------------------------------------------------------------------- + TOTAL 0.04705761 83.59556052 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 305000 +BPFP 0.9941 bits/point +EBPFP 1.9882 equivalent bits/point +MSE 83.595561 +---------------------- -------------------------------------------------------- +Time: 4.104s Load: 0.012s, Pack+Encode: 2.305s, Decode+Unpack: 1.787s +---------------------- -------------------------------------------------------- +💾 Converting with 83.5956 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,836B, BPFP=0.4165 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,188B, BPFP=2.1358 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,500B, BPFP=0.7706 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,180B, BPFP=2.0823 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,160B, BPFP=0.9120 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,148B, BPFP=2.0806 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,228B, BPFP=0.8625 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,904B, BPFP=2.1207 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,320B, BPFP=1.7708 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,724B, BPFP=2.0580 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,748B, BPFP=0.1803 +⌛️ [2/4] FRONTEND: Frontend time: 2.286s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.636s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582189 33.50249455 + layer.0.v_cache 0.00001509 0.00698213 + layer.1.k_cache 0.58311327 12.50083539 + layer.1.v_cache 0.00000606 0.00290643 + layer.2.k_cache 0.00657600 1.60179200 + layer.2.v_cache 0.00001909 0.00983283 + layer.3.k_cache 0.02234380 9.89287126 + layer.3.v_cache 0.00001905 0.01022562 + layer.4.k_cache 0.00071828 0.30840431 + layer.4.v_cache 0.00005370 0.02024473 + layer.4.output 0.05044473 182.63787658 + ------------------------------------------------------------------------------------- + TOTAL 0.06363526 78.60716031 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 309936 +BPFP 0.9689 bits/point +EBPFP 1.9379 equivalent bits/point +MSE 78.607160 +---------------------- -------------------------------------------------------- +Time: 3.933s Load: 0.010s, Pack+Encode: 2.286s, Decode+Unpack: 1.636s +---------------------- -------------------------------------------------------- +💾 Converting with 78.6072 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 292, 128) +Output shape: (1, 292, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.output: torch.Size([1, 292, 3584]) -> torch.Size([1, 1, 292, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,880B, BPFP=0.4217 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,964B, BPFP=2.1385 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,200B, BPFP=0.7598 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,044B, BPFP=2.0893 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,124B, BPFP=0.9163 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,528B, BPFP=2.1152 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,232B, BPFP=0.8686 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,848B, BPFP=2.1323 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,456B, BPFP=1.7902 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,460B, BPFP=2.0580 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,252B, BPFP=0.1777 +⌛️ [2/4] FRONTEND: Frontend time: 2.352s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.825s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887579 36.37393648 + layer.0.v_cache 0.00001384 0.00674164 + layer.1.k_cache 0.57277983 11.09943379 + layer.1.v_cache 0.00000568 0.00282074 + layer.2.k_cache 0.01314942 1.60446794 + layer.2.v_cache 0.00002105 0.01015909 + layer.3.k_cache 0.03125376 9.93999324 + layer.3.v_cache 0.00001978 0.01053321 + layer.4.k_cache 0.00068035 0.32481755 + layer.4.v_cache 0.00005123 0.02156059 + layer.4.output 0.04859251 181.77661754 + ------------------------------------------------------------------------------------- + TOTAL 0.06511755 78.34298747 + (elements=2,541,568) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2541568 +Total Bytes 308988 +BPFP 0.9726 bits/point +EBPFP 1.9452 equivalent bits/point +MSE 78.342987 +---------------------- -------------------------------------------------------- +Time: 4.187s Load: 0.010s, Pack+Encode: 2.352s, Decode+Unpack: 1.825s +---------------------- -------------------------------------------------------- +💾 Converting with 78.3430 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 338, 128) +Output shape: (1, 338, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.output: torch.Size([1, 338, 3584]) -> torch.Size([1, 1, 338, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,172B, BPFP=0.4240 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,276B, BPFP=2.2317 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,132B, BPFP=0.7457 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,836B, BPFP=2.1651 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,904B, BPFP=0.9201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,648B, BPFP=2.1564 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,308B, BPFP=0.8463 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,552B, BPFP=2.1982 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,096B, BPFP=1.6686 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,148B, BPFP=2.1333 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,952B, BPFP=0.1912 +⌛️ [2/4] FRONTEND: Frontend time: 2.508s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.876s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13976314 32.19363385 + layer.0.v_cache 0.00001472 0.00676381 + layer.1.k_cache 0.73750278 11.92315349 + layer.1.v_cache 0.00000589 0.00285670 + layer.2.k_cache 0.02011468 1.62121113 + layer.2.v_cache 0.00002041 0.00964642 + layer.3.k_cache 0.02722352 9.44930037 + layer.3.v_cache 0.00002029 0.00977235 + layer.4.k_cache 0.00072609 0.32638394 + layer.4.v_cache 0.00005477 0.02178644 + layer.4.output 0.04270584 156.14452134 + ------------------------------------------------------------------------------------- + TOTAL 0.07202277 67.56330341 + (elements=2,941,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2941952 +Total Bytes 364024 +BPFP 0.9899 bits/point +EBPFP 1.9798 equivalent bits/point +MSE 67.563303 +---------------------- -------------------------------------------------------- +Time: 4.394s Load: 0.011s, Pack+Encode: 2.508s, Decode+Unpack: 1.876s +---------------------- -------------------------------------------------------- +💾 Converting with 67.5633 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,736B, BPFP=0.4154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,164B, BPFP=2.1566 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,744B, BPFP=0.7917 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,200B, BPFP=2.1048 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,200B, BPFP=0.9235 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,688B, BPFP=2.1310 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,092B, BPFP=0.8640 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,916B, BPFP=2.1433 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,204B, BPFP=1.7292 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,824B, BPFP=2.0846 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,312B, BPFP=0.1788 +⌛️ [2/4] FRONTEND: Frontend time: 2.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.663s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150811 32.40866959 + layer.0.v_cache 0.00001374 0.00658345 + layer.1.k_cache 0.66021251 12.03710434 + layer.1.v_cache 0.00000574 0.00286247 + layer.2.k_cache 0.00963001 1.60834695 + layer.2.v_cache 0.00001988 0.00987693 + layer.3.k_cache 0.08055985 9.96350140 + layer.3.v_cache 0.00001942 0.01015459 + layer.4.k_cache 0.00069026 0.31932954 + layer.4.v_cache 0.00005126 0.02074889 + layer.4.output 0.05093874 182.57724288 + ------------------------------------------------------------------------------------- + TOTAL 0.07289894 78.49575755 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 309080 +BPFP 0.9762 bits/point +EBPFP 1.9524 equivalent bits/point +MSE 78.495758 +---------------------- -------------------------------------------------------- +Time: 4.160s Load: 0.009s, Pack+Encode: 2.488s, Decode+Unpack: 1.663s +---------------------- -------------------------------------------------------- +💾 Converting with 78.4958 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,188B, BPFP=0.4222 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,920B, BPFP=2.0586 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,192B, BPFP=0.7318 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,080B, BPFP=2.0153 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,548B, BPFP=0.9049 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,248B, BPFP=2.0239 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,096B, BPFP=0.8300 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,880B, BPFP=2.1081 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,724B, BPFP=1.7391 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,640B, BPFP=1.9926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,672B, BPFP=0.1891 +⌛️ [2/4] FRONTEND: Frontend time: 2.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.784s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14893842 37.86714882 + layer.0.v_cache 0.00001415 0.00653604 + layer.1.k_cache 0.65110169 9.40909750 + layer.1.v_cache 0.00000575 0.00280388 + layer.2.k_cache 0.01411552 1.62977605 + layer.2.v_cache 0.00001904 0.01015837 + layer.3.k_cache 0.03939451 9.76094685 + layer.3.v_cache 0.00001876 0.00988045 + layer.4.k_cache 0.00070558 0.33661468 + layer.4.v_cache 0.00005345 0.02153289 + layer.4.output 0.04819407 176.88375177 + ------------------------------------------------------------------------------------- + TOTAL 0.07010149 76.30827988 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 313188 +BPFP 0.9500 bits/point +EBPFP 1.9000 equivalent bits/point +MSE 76.308280 +---------------------- -------------------------------------------------------- +Time: 4.285s Load: 0.012s, Pack+Encode: 2.489s, Decode+Unpack: 1.784s +---------------------- -------------------------------------------------------- +💾 Converting with 76.3083 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,188B, BPFP=0.4270 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,840B, BPFP=2.3075 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,048B, BPFP=0.8346 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,812B, BPFP=2.2464 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,332B, BPFP=0.9703 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,520B, BPFP=2.2291 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,924B, BPFP=0.8866 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,508B, BPFP=2.2878 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,756B, BPFP=1.7678 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,200B, BPFP=2.2101 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,608B, BPFP=0.1749 +⌛️ [2/4] FRONTEND: Frontend time: 2.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.770s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14492303 30.41810985 + layer.0.v_cache 0.00001375 0.00669921 + layer.1.k_cache 0.50395557 13.05523705 + layer.1.v_cache 0.00000564 0.00276320 + layer.2.k_cache 0.00856250 1.55417613 + layer.2.v_cache 0.00001902 0.00992050 + layer.3.k_cache 0.02692018 8.53696178 + layer.3.v_cache 0.00002118 0.01001187 + layer.4.k_cache 0.00070790 0.30615034 + layer.4.v_cache 0.00005130 0.01979158 + layer.4.output 0.01131245 209.11568102 + ------------------------------------------------------------------------------------- + TOTAL 0.04496278 89.27821110 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 292736 +BPFP 1.0230 bits/point +EBPFP 2.0461 equivalent bits/point +MSE 89.278211 +---------------------- -------------------------------------------------------- +Time: 3.941s Load: 0.010s, Pack+Encode: 2.162s, Decode+Unpack: 1.770s +---------------------- -------------------------------------------------------- +💾 Converting with 89.2782 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,544B, BPFP=0.4271 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,384B, BPFP=2.2862 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,212B, BPFP=0.8046 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,208B, BPFP=2.2197 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,032B, BPFP=0.9642 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,052B, BPFP=2.2108 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,716B, BPFP=0.8897 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,628B, BPFP=2.2434 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,596B, BPFP=1.7321 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,572B, BPFP=2.1837 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,472B, BPFP=0.1898 +⌛️ [2/4] FRONTEND: Frontend time: 2.364s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.701s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14113761 29.30787371 + layer.0.v_cache 0.00001369 0.00661509 + layer.1.k_cache 0.59212339 12.32413295 + layer.1.v_cache 0.00000552 0.00274227 + layer.2.k_cache 0.01129283 1.59365889 + layer.2.v_cache 0.00001853 0.00990650 + layer.3.k_cache 0.01358872 8.89711596 + layer.3.v_cache 0.00001883 0.00992614 + layer.4.k_cache 0.00070763 0.32480151 + layer.4.v_cache 0.00005153 0.02026993 + layer.4.output 0.01019159 195.57511646 + ------------------------------------------------------------------------------------- + TOTAL 0.04884114 83.61899166 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 305416 +BPFP 1.0171 bits/point +EBPFP 2.0342 equivalent bits/point +MSE 83.618992 +---------------------- -------------------------------------------------------- +Time: 4.075s Load: 0.009s, Pack+Encode: 2.364s, Decode+Unpack: 1.701s +---------------------- -------------------------------------------------------- +💾 Converting with 83.6190 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,564B, BPFP=0.4118 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,304B, BPFP=2.1943 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,000B, BPFP=0.8166 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,292B, BPFP=2.1392 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,696B, BPFP=0.9090 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,300B, BPFP=2.1396 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,332B, BPFP=0.8347 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,040B, BPFP=2.1799 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,920B, BPFP=1.6834 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,916B, BPFP=2.1187 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,204B, BPFP=0.2116 +⌛️ [2/4] FRONTEND: Frontend time: 2.717s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.955s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11028925 33.14153011 + layer.0.v_cache 0.00001466 0.00658935 + layer.1.k_cache 0.56032453 13.12611607 + layer.1.v_cache 0.00000595 0.00274843 + layer.2.k_cache 0.00809042 1.65137591 + layer.2.v_cache 0.00001929 0.00975171 + layer.3.k_cache 0.03501172 9.65444234 + layer.3.v_cache 0.00001861 0.00980849 + layer.4.k_cache 0.00070074 0.32031792 + layer.4.v_cache 0.00005246 0.02050112 + layer.4.output 0.00889761 185.90777439 + ------------------------------------------------------------------------------------- + TOTAL 0.04569476 79.95868248 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 310568 +BPFP 0.9946 bits/point +EBPFP 1.9892 equivalent bits/point +MSE 79.958682 +---------------------- -------------------------------------------------------- +Time: 4.683s Load: 0.011s, Pack+Encode: 2.717s, Decode+Unpack: 1.955s +---------------------- -------------------------------------------------------- +💾 Converting with 79.9587 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 311, 128) +Output shape: (1, 311, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.output: torch.Size([1, 311, 3584]) -> torch.Size([1, 1, 311, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,304B, BPFP=0.4172 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,904B, BPFP=2.0048 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,392B, BPFP=0.7231 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,796B, BPFP=1.9492 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,536B, BPFP=0.8810 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,536B, BPFP=1.9863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,496B, BPFP=0.8288 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,848B, BPFP=2.0020 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,772B, BPFP=1.7972 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,656B, BPFP=1.9421 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,656B, BPFP=0.1841 +⌛️ [2/4] FRONTEND: Frontend time: 2.242s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.719s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14227538 35.37740844 + layer.0.v_cache 0.00001500 0.00677982 + layer.1.k_cache 0.64922478 10.78415300 + layer.1.v_cache 0.00000572 0.00289182 + layer.2.k_cache 0.00828172 1.60580101 + layer.2.v_cache 0.00001922 0.01008068 + layer.3.k_cache 0.02636586 9.72825186 + layer.3.v_cache 0.00001996 0.01050282 + layer.4.k_cache 0.00069494 0.31788267 + layer.4.v_cache 0.00005248 0.02186899 + layer.4.output 0.04686997 172.30253790 + ------------------------------------------------------------------------------------- + TOTAL 0.06794382 74.35196390 + (elements=2,706,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2706944 +Total Bytes 314896 +BPFP 0.9306 bits/point +EBPFP 1.8613 equivalent bits/point +MSE 74.351964 +---------------------- -------------------------------------------------------- +Time: 3.974s Load: 0.012s, Pack+Encode: 2.242s, Decode+Unpack: 1.719s +---------------------- -------------------------------------------------------- +💾 Converting with 74.3520 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,628B, BPFP=0.4061 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,956B, BPFP=1.9581 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,992B, BPFP=0.7348 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,188B, BPFP=1.9110 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,440B, BPFP=0.8235 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,288B, BPFP=1.9784 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,540B, BPFP=0.7684 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,976B, BPFP=1.9593 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,084B, BPFP=1.5983 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,096B, BPFP=1.9054 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,640B, BPFP=0.1982 +⌛️ [2/4] FRONTEND: Frontend time: 2.814s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.925s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13149661 35.36179534 + layer.0.v_cache 0.00001370 0.00646708 + layer.1.k_cache 0.50214604 13.27892923 + layer.1.v_cache 0.00000625 0.00286591 + layer.2.k_cache 0.00807812 1.65763119 + layer.2.v_cache 0.00001969 0.01002985 + layer.3.k_cache 0.06061829 10.27213350 + layer.3.v_cache 0.00001868 0.00991052 + layer.4.k_cache 0.00068906 0.32177339 + layer.4.v_cache 0.00005044 0.02058102 + layer.4.output 1.20679682 202.35558473 + ------------------------------------------------------------------------------------- + TOTAL 0.53827733 86.90771825 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 251828 +BPFP 0.9077 bits/point +EBPFP 1.8154 equivalent bits/point +MSE 86.907718 +---------------------- -------------------------------------------------------- +Time: 4.748s Load: 0.009s, Pack+Encode: 2.814s, Decode+Unpack: 1.925s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9077 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 300, 128) +Output shape: (1, 300, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.output: torch.Size([1, 300, 3584]) -> torch.Size([1, 1, 300, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,064B, BPFP=0.4200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,924B, BPFP=2.0794 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,468B, BPFP=0.7535 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,988B, BPFP=2.0306 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,508B, BPFP=0.9119 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,120B, BPFP=2.0375 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,168B, BPFP=0.8421 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,568B, BPFP=2.1650 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,472B, BPFP=1.7954 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,672B, BPFP=2.0142 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,624B, BPFP=0.1907 +⌛️ [2/4] FRONTEND: Frontend time: 2.439s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.879s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13491969 37.98871419 + layer.0.v_cache 0.00001558 0.00658251 + layer.1.k_cache 0.64590566 12.59720378 + layer.1.v_cache 0.00000556 0.00274226 + layer.2.k_cache 0.01395306 1.60225006 + layer.2.v_cache 0.00001972 0.01006103 + layer.3.k_cache 0.02379198 9.74874674 + layer.3.v_cache 0.00001942 0.01058380 + layer.4.k_cache 0.00073123 0.32474190 + layer.4.v_cache 0.00005274 0.02234126 + layer.4.output 0.05011183 172.31559524 + ------------------------------------------------------------------------------------- + TOTAL 0.06883514 74.61900789 + (elements=2,611,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2611200 +Total Bytes 314576 +BPFP 0.9638 bits/point +EBPFP 1.9275 equivalent bits/point +MSE 74.619008 +---------------------- -------------------------------------------------------- +Time: 4.331s Load: 0.012s, Pack+Encode: 2.439s, Decode+Unpack: 1.879s +---------------------- -------------------------------------------------------- +💾 Converting with 74.6190 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,756B, BPFP=0.4165 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,224B, BPFP=2.1598 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,156B, BPFP=0.7601 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,380B, BPFP=2.1145 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,124B, BPFP=0.9195 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,512B, BPFP=2.1216 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,652B, BPFP=0.8404 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,752B, BPFP=2.2418 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,416B, BPFP=1.6869 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,732B, BPFP=2.0797 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,732B, BPFP=0.1820 +⌛️ [2/4] FRONTEND: Frontend time: 2.653s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.722s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14826349 32.92154277 + layer.0.v_cache 0.00001415 0.00648068 + layer.1.k_cache 0.61440683 12.57100213 + layer.1.v_cache 0.00000562 0.00282073 + layer.2.k_cache 0.00551110 1.66109548 + layer.2.v_cache 0.00001912 0.00972955 + layer.3.k_cache 0.01294636 9.78151511 + layer.3.v_cache 0.00002032 0.00961712 + layer.4.k_cache 0.00070633 0.32468747 + layer.4.v_cache 0.00005074 0.02064892 + layer.4.output 0.04972342 183.89037187 + ------------------------------------------------------------------------------------- + TOTAL 0.06647106 79.09069077 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 309436 +BPFP 0.9773 bits/point +EBPFP 1.9547 equivalent bits/point +MSE 79.090691 +---------------------- -------------------------------------------------------- +Time: 4.385s Load: 0.010s, Pack+Encode: 2.653s, Decode+Unpack: 1.722s +---------------------- -------------------------------------------------------- +💾 Converting with 79.0907 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,560B, BPFP=0.4264 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,116B, BPFP=2.2629 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,132B, BPFP=0.7972 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,028B, BPFP=2.2015 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,988B, BPFP=0.9583 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,932B, BPFP=2.1961 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,220B, BPFP=0.8585 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,472B, BPFP=2.2265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,788B, BPFP=1.7367 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,048B, BPFP=2.1462 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,000B, BPFP=0.1773 +⌛️ [2/4] FRONTEND: Frontend time: 2.367s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.688s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13861307 29.32430512 + layer.0.v_cache 0.00001419 0.00665239 + layer.1.k_cache 0.53867822 11.99233028 + layer.1.v_cache 0.00000573 0.00285631 + layer.2.k_cache 0.00871976 1.60252419 + layer.2.v_cache 0.00001877 0.00984114 + layer.3.k_cache 0.02195022 7.60395102 + layer.3.v_cache 0.00001967 0.01062120 + layer.4.k_cache 0.00069227 0.32146820 + layer.4.v_cache 0.00004961 0.02194016 + layer.4.output 0.00954030 196.30616619 + ------------------------------------------------------------------------------------- + TOTAL 0.04562021 83.82586196 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 302284 +BPFP 1.0030 bits/point +EBPFP 2.0060 equivalent bits/point +MSE 83.825862 +---------------------- -------------------------------------------------------- +Time: 4.065s Load: 0.010s, Pack+Encode: 2.367s, Decode+Unpack: 1.688s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8259 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,612B, BPFP=0.4159 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,256B, BPFP=2.1993 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,992B, BPFP=0.7644 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,784B, BPFP=2.2281 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,752B, BPFP=0.9152 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,100B, BPFP=2.1361 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,732B, BPFP=0.8595 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,712B, BPFP=2.1696 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,128B, BPFP=1.7006 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,672B, BPFP=2.1128 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,052B, BPFP=0.1955 +⌛️ [2/4] FRONTEND: Frontend time: 2.535s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.737s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13928204 32.34315245 + layer.0.v_cache 0.00001373 0.00649263 + layer.1.k_cache 0.60860016 12.56711989 + layer.1.v_cache 0.00000555 0.00281489 + layer.2.k_cache 0.01026906 1.62485061 + layer.2.v_cache 0.00001872 0.01025519 + layer.3.k_cache 0.05166252 9.67931275 + layer.3.v_cache 0.00001937 0.01073650 + layer.4.k_cache 0.00073448 0.31211560 + layer.4.v_cache 0.00005026 0.02106393 + layer.4.output 0.00697322 185.32053884 + ------------------------------------------------------------------------------------- + TOTAL 0.05055697 79.63656978 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 308792 +BPFP 0.9924 bits/point +EBPFP 1.9847 equivalent bits/point +MSE 79.636570 +---------------------- -------------------------------------------------------- +Time: 4.287s Load: 0.015s, Pack+Encode: 2.535s, Decode+Unpack: 1.737s +---------------------- -------------------------------------------------------- +💾 Converting with 79.6366 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,880B, BPFP=0.4174 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,076B, BPFP=2.1227 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,264B, BPFP=0.7555 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,988B, BPFP=2.0650 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,232B, BPFP=0.9127 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,292B, BPFP=2.0811 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,036B, BPFP=0.8494 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,796B, BPFP=2.1078 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,632B, BPFP=1.7814 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,708B, BPFP=2.0502 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,604B, BPFP=0.1786 +⌛️ [2/4] FRONTEND: Frontend time: 2.637s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.850s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13820392 35.28598385 + layer.0.v_cache 0.00001454 0.00674696 + layer.1.k_cache 0.61684219 12.74062334 + layer.1.v_cache 0.00000570 0.00283609 + layer.2.k_cache 0.01285210 1.59359214 + layer.2.v_cache 0.00002011 0.01016523 + layer.3.k_cache 0.05980752 9.87093154 + layer.3.v_cache 0.00002263 0.01041850 + layer.4.k_cache 0.00071023 0.32824606 + layer.4.v_cache 0.00005833 0.02050757 + layer.4.output 0.05086143 179.76843220 + ------------------------------------------------------------------------------------- + TOTAL 0.06968043 77.54406333 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 309508 +BPFP 0.9643 bits/point +EBPFP 1.9286 equivalent bits/point +MSE 77.544063 +---------------------- -------------------------------------------------------- +Time: 4.498s Load: 0.011s, Pack+Encode: 2.637s, Decode+Unpack: 1.850s +---------------------- -------------------------------------------------------- +💾 Converting with 77.5441 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,072B, BPFP=0.4190 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,000B, BPFP=2.0764 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,804B, BPFP=0.7685 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,124B, BPFP=2.0309 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,348B, BPFP=0.9005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,300B, BPFP=2.0401 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,468B, BPFP=0.8549 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,824B, BPFP=2.0673 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,892B, BPFP=1.7593 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,728B, BPFP=2.0104 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,840B, BPFP=0.1916 +⌛️ [2/4] FRONTEND: Frontend time: 2.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.635s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14825445 37.96312098 + layer.0.v_cache 0.00001396 0.00705158 + layer.1.k_cache 0.63210902 11.52050619 + layer.1.v_cache 0.00000581 0.00295986 + layer.2.k_cache 0.01184132 1.61613186 + layer.2.v_cache 0.00001890 0.01017260 + layer.3.k_cache 0.01767834 9.91584205 + layer.3.v_cache 0.00002009 0.01035120 + layer.4.k_cache 0.00068272 0.30437399 + layer.4.v_cache 0.00005247 0.02223881 + layer.4.output 0.04867053 177.21370729 + ------------------------------------------------------------------------------------- + TOTAL 0.06772769 76.58051177 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 313400 +BPFP 0.9570 bits/point +EBPFP 1.9140 equivalent bits/point +MSE 76.580512 +---------------------- -------------------------------------------------------- +Time: 3.956s Load: 0.012s, Pack+Encode: 2.309s, Decode+Unpack: 1.635s +---------------------- -------------------------------------------------------- +💾 Converting with 76.5805 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,232B, BPFP=0.4280 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,104B, BPFP=2.3144 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,020B, BPFP=0.8890 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,888B, BPFP=2.2424 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,644B, BPFP=0.9851 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,164B, BPFP=2.1996 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,000B, BPFP=0.8878 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,820B, BPFP=2.2976 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,764B, BPFP=1.7616 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,028B, BPFP=2.1915 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,184B, BPFP=0.1622 +⌛️ [2/4] FRONTEND: Frontend time: 2.207s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.680s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14315909 30.41496323 + layer.0.v_cache 0.00001391 0.00706976 + layer.1.k_cache 0.49711875 12.85116855 + layer.1.v_cache 0.00000560 0.00303192 + layer.2.k_cache 0.00646762 1.60936298 + layer.2.v_cache 0.00001905 0.01031560 + layer.3.k_cache 0.04707547 7.78535323 + layer.3.v_cache 0.00001924 0.01100205 + layer.4.k_cache 0.00071833 0.28623592 + layer.4.v_cache 0.00005197 0.02030348 + layer.4.output 0.00997522 208.08272457 + ------------------------------------------------------------------------------------- + TOTAL 0.04496915 88.79869875 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 292848 +BPFP 1.0196 bits/point +EBPFP 2.0391 equivalent bits/point +MSE 88.798699 +---------------------- -------------------------------------------------------- +Time: 3.896s Load: 0.009s, Pack+Encode: 2.207s, Decode+Unpack: 1.680s +---------------------- -------------------------------------------------------- +💾 Converting with 88.7987 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,556B, BPFP=0.4232 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,340B, BPFP=2.2592 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,808B, BPFP=0.7733 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,384B, BPFP=2.2056 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,072B, BPFP=0.9561 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,776B, BPFP=2.2836 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,604B, BPFP=0.8739 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,188B, BPFP=2.3627 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,564B, BPFP=1.7117 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,428B, BPFP=2.2081 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,340B, BPFP=0.1867 +⌛️ [2/4] FRONTEND: Frontend time: 2.197s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.748s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15581964 29.40025412 + layer.0.v_cache 0.00001703 0.00658873 + layer.1.k_cache 0.56298232 12.28986230 + layer.1.v_cache 0.00000550 0.00269085 + layer.2.k_cache 0.00979712 1.59314085 + layer.2.v_cache 0.00002023 0.00980950 + layer.3.k_cache 0.04960944 9.51649743 + layer.3.v_cache 0.00001971 0.01018869 + layer.4.k_cache 0.00068561 0.31646001 + layer.4.v_cache 0.00005245 0.02068166 + layer.4.output 0.01122868 192.58005312 + ------------------------------------------------------------------------------------- + TOTAL 0.05044764 82.42509094 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 310060 +BPFP 1.0214 bits/point +EBPFP 2.0429 equivalent bits/point +MSE 82.425091 +---------------------- -------------------------------------------------------- +Time: 3.955s Load: 0.009s, Pack+Encode: 2.197s, Decode+Unpack: 1.748s +---------------------- -------------------------------------------------------- +💾 Converting with 82.4251 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,564B, BPFP=0.4176 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,244B, BPFP=2.2220 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,196B, BPFP=0.7838 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,308B, BPFP=2.1703 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,308B, BPFP=0.9004 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,224B, BPFP=2.1656 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,704B, BPFP=0.8670 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,984B, BPFP=2.2076 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,832B, BPFP=1.6471 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,672B, BPFP=2.1352 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,144B, BPFP=0.1983 +⌛️ [2/4] FRONTEND: Frontend time: 2.471s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.629s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13599485 30.81212042 + layer.0.v_cache 0.00001418 0.00658889 + layer.1.k_cache 0.56373192 12.60177386 + layer.1.v_cache 0.00000588 0.00282809 + layer.2.k_cache 0.01549433 1.59474177 + layer.2.v_cache 0.00001867 0.00990408 + layer.3.k_cache 0.03475572 9.88563349 + layer.3.v_cache 0.00001907 0.00975533 + layer.4.k_cache 0.00072111 0.31779305 + layer.4.v_cache 0.00005295 0.02187863 + layer.4.output 0.00820773 182.24859604 + ------------------------------------------------------------------------------------- + TOTAL 0.04754487 78.29430529 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 306180 +BPFP 0.9944 bits/point +EBPFP 1.9888 equivalent bits/point +MSE 78.294305 +---------------------- -------------------------------------------------------- +Time: 4.110s Load: 0.010s, Pack+Encode: 2.471s, Decode+Unpack: 1.629s +---------------------- -------------------------------------------------------- +💾 Converting with 78.2943 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,168B, BPFP=0.4144 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,796B, BPFP=2.0189 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,488B, BPFP=0.7350 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,728B, BPFP=1.9647 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,728B, BPFP=0.8994 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,844B, BPFP=1.9706 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,364B, BPFP=0.8302 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,500B, BPFP=2.0039 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,192B, BPFP=1.6838 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,496B, BPFP=1.9529 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,992B, BPFP=0.1739 +⌛️ [2/4] FRONTEND: Frontend time: 2.424s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.944s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13331154 34.50169630 + layer.0.v_cache 0.00001460 0.00677974 + layer.1.k_cache 0.64535681 10.44506123 + layer.1.v_cache 0.00000563 0.00290255 + layer.2.k_cache 0.00762887 1.58208634 + layer.2.v_cache 0.00001910 0.01051227 + layer.3.k_cache 0.03538683 9.69783000 + layer.3.v_cache 0.00002103 0.01092299 + layer.4.k_cache 0.00072257 0.32189652 + layer.4.v_cache 0.00005300 0.02154001 + layer.4.output 0.04522907 168.54544005 + ------------------------------------------------------------------------------------- + TOTAL 0.06700726 72.73054755 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 309296 +BPFP 0.9230 bits/point +EBPFP 1.8460 equivalent bits/point +MSE 72.730548 +---------------------- -------------------------------------------------------- +Time: 4.381s Load: 0.013s, Pack+Encode: 2.424s, Decode+Unpack: 1.944s +---------------------- -------------------------------------------------------- +💾 Converting with 72.7305 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,640B, BPFP=0.4203 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,356B, BPFP=2.2203 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,044B, BPFP=0.7727 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,316B, BPFP=2.1631 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,020B, BPFP=0.9364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,952B, BPFP=2.1430 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,900B, BPFP=0.8748 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,000B, BPFP=2.2007 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,572B, BPFP=1.6820 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,608B, BPFP=2.1241 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,612B, BPFP=0.2249 +⌛️ [2/4] FRONTEND: Frontend time: 2.271s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.682s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14806804 31.23416868 + layer.0.v_cache 0.00001376 0.00662995 + layer.1.k_cache 0.60662186 12.87208837 + layer.1.v_cache 0.00000571 0.00277393 + layer.2.k_cache 0.01239752 1.61675735 + layer.2.v_cache 0.00002027 0.00977495 + layer.3.k_cache 0.02456893 9.94261406 + layer.3.v_cache 0.00001975 0.01002294 + layer.4.k_cache 0.00069718 0.32439256 + layer.4.v_cache 0.00005379 0.02078729 + layer.4.output 0.00818182 184.53337211 + ------------------------------------------------------------------------------------- + TOTAL 0.04998468 79.28080087 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 311020 +BPFP 1.0066 bits/point +EBPFP 2.0131 equivalent bits/point +MSE 79.280801 +---------------------- -------------------------------------------------------- +Time: 3.967s Load: 0.013s, Pack+Encode: 2.271s, Decode+Unpack: 1.682s +---------------------- -------------------------------------------------------- +💾 Converting with 79.2808 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,568B, BPFP=0.4178 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,456B, BPFP=2.2337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,396B, BPFP=0.7396 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,440B, BPFP=2.1776 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,776B, BPFP=0.9262 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,216B, BPFP=2.1652 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,984B, BPFP=0.8825 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,224B, BPFP=2.2208 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,260B, BPFP=1.7259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,184B, BPFP=2.1634 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,540B, BPFP=0.2172 +⌛️ [2/4] FRONTEND: Frontend time: 2.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.707s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14592592 30.67638168 + layer.0.v_cache 0.00001391 0.00664618 + layer.1.k_cache 0.53254910 12.39133240 + layer.1.v_cache 0.00000583 0.00287551 + layer.2.k_cache 0.01029040 1.63218166 + layer.2.v_cache 0.00002009 0.01002804 + layer.3.k_cache 0.03848921 9.76295153 + layer.3.v_cache 0.00001972 0.00979804 + layer.4.k_cache 0.00069417 0.33467054 + layer.4.v_cache 0.00005427 0.02100123 + layer.4.output 0.01004909 190.93369826 + ------------------------------------------------------------------------------------- + TOTAL 0.04696507 81.84610321 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 311044 +BPFP 1.0102 bits/point +EBPFP 2.0204 equivalent bits/point +MSE 81.846103 +---------------------- -------------------------------------------------------- +Time: 3.966s Load: 0.010s, Pack+Encode: 2.249s, Decode+Unpack: 1.707s +---------------------- -------------------------------------------------------- +💾 Converting with 81.8461 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,520B, BPFP=0.4320 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,976B, BPFP=2.2964 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,228B, BPFP=0.7599 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,584B, BPFP=2.2165 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,276B, BPFP=0.9350 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,004B, BPFP=2.2406 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,216B, BPFP=0.8741 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,188B, BPFP=2.2511 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,884B, BPFP=1.7167 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,880B, BPFP=2.1760 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,480B, BPFP=0.1845 +⌛️ [2/4] FRONTEND: Frontend time: 2.468s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.669s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15384994 31.18882482 + layer.0.v_cache 0.00001404 0.00687384 + layer.1.k_cache 0.48768111 12.29203348 + layer.1.v_cache 0.00000584 0.00290602 + layer.2.k_cache 0.00538594 1.57188135 + layer.2.v_cache 0.00001948 0.00958256 + layer.3.k_cache 0.02347840 9.18795238 + layer.3.v_cache 0.00001899 0.00964707 + layer.4.k_cache 0.00068094 0.31237411 + layer.4.v_cache 0.00005116 0.02069471 + layer.4.output 0.00909345 193.15636489 + ------------------------------------------------------------------------------------- + TOTAL 0.04322588 82.74690145 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 299236 +BPFP 1.0112 bits/point +EBPFP 2.0223 equivalent bits/point +MSE 82.746901 +---------------------- -------------------------------------------------------- +Time: 4.147s Load: 0.011s, Pack+Encode: 2.468s, Decode+Unpack: 1.669s +---------------------- -------------------------------------------------------- +💾 Converting with 82.7469 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,856B, BPFP=0.4175 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,872B, BPFP=2.1190 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,912B, BPFP=0.7925 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,996B, BPFP=2.0725 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,312B, BPFP=0.9201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,936B, BPFP=2.0693 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,056B, BPFP=0.8533 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,684B, BPFP=2.1091 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,096B, BPFP=1.7589 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,284B, BPFP=2.0347 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,516B, BPFP=0.1709 +⌛️ [2/4] FRONTEND: Frontend time: 2.669s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.870s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13453574 35.73243184 + layer.0.v_cache 0.00001386 0.00677964 + layer.1.k_cache 0.70462929 12.45105894 + layer.1.v_cache 0.00000580 0.00290330 + layer.2.k_cache 0.00896267 1.57189890 + layer.2.v_cache 0.00001881 0.00981611 + layer.3.k_cache 0.02468209 9.82631304 + layer.3.v_cache 0.00001940 0.01048625 + layer.4.k_cache 0.00068809 0.31032933 + layer.4.v_cache 0.00005209 0.02187357 + layer.4.output 0.04953035 175.55179483 + ------------------------------------------------------------------------------------- + TOTAL 0.07178355 75.81214439 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 307520 +BPFP 0.9614 bits/point +EBPFP 1.9228 equivalent bits/point +MSE 75.812144 +---------------------- -------------------------------------------------------- +Time: 4.553s Load: 0.013s, Pack+Encode: 2.669s, Decode+Unpack: 1.870s +---------------------- -------------------------------------------------------- +💾 Converting with 75.8121 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,360B, BPFP=0.4291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,016B, BPFP=2.3330 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,396B, BPFP=0.8393 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,576B, BPFP=2.2491 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,428B, BPFP=0.9578 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,196B, BPFP=2.2269 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,432B, BPFP=0.8997 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,388B, BPFP=2.2964 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,512B, BPFP=1.7206 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,060B, BPFP=2.2190 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,764B, BPFP=0.1896 +⌛️ [2/4] FRONTEND: Frontend time: 2.291s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.790s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150386 30.06379358 + layer.0.v_cache 0.00001401 0.00665281 + layer.1.k_cache 0.57442873 13.17776990 + layer.1.v_cache 0.00000554 0.00274890 + layer.2.k_cache 0.00779028 1.58867554 + layer.2.v_cache 0.00001875 0.00970726 + layer.3.k_cache 0.05176373 9.62578161 + layer.3.v_cache 0.00002027 0.01027068 + layer.4.k_cache 0.00069872 0.30473999 + layer.4.v_cache 0.00005023 0.02182542 + layer.4.output 0.01044900 201.51875666 + ------------------------------------------------------------------------------------- + TOTAL 0.04937866 86.20254484 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 300128 +BPFP 1.0293 bits/point +EBPFP 2.0586 equivalent bits/point +MSE 86.202545 +---------------------- -------------------------------------------------------- +Time: 4.091s Load: 0.010s, Pack+Encode: 2.291s, Decode+Unpack: 1.790s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2025 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.9891 bits/point +Avg EBPFP 1.9781 equivalent bits/point +Avg MSE 80.660868 +Avg Time 4.173s +------------------------ ---------------------------- diff --git a/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..fe958fb7904f9f8b3165500ce1146b8406700677 --- /dev/null +++ b/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 333 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- -------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.007_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa +Output output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa +---------------- -------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,620B, BPFP=0.4874 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,572B, BPFP=2.8966 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,264B, BPFP=0.7932 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,604B, BPFP=2.7165 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,892B, BPFP=0.9100 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,172B, BPFP=2.6362 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,448B, BPFP=0.8274 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,960B, BPFP=2.7827 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,292B, BPFP=1.9144 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,264B, BPFP=2.6533 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,012B, BPFP=0.2661 +⌛️ [2/4] FRONTEND: Frontend time: 2.461s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.269s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14073507 30.27146112 + layer.0.v_cache 0.00001426 0.00832492 + layer.1.k_cache 0.05042761 6.99375843 + layer.1.v_cache 0.00000517 0.00312943 + layer.2.k_cache 0.00244098 0.89380573 + layer.2.v_cache 0.00001715 0.01016634 + layer.3.k_cache 0.02726505 3.74358768 + layer.3.v_cache 0.00001796 0.01107943 + layer.4.k_cache 0.00069123 0.33844505 + layer.4.v_cache 0.00005084 0.02486840 + layer.4.output 0.16186065 637.84959609 + ------------------------------------------------------------------------------------- + TOTAL 0.07968764 265.13210583 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 110100 +BPFP 1.2047 bits/point +EBPFP 2.4094 equivalent bits/point +MSE 265.132106 +---------------------- -------------------------------------------------------- +Time: 3.735s Load: 0.004s, Pack+Encode: 2.461s, Decode+Unpack: 1.269s +---------------------- -------------------------------------------------------- +💾 Converting with 265.1321 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,592B, BPFP=0.4880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,896B, BPFP=2.8042 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,140B, BPFP=0.7794 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,952B, BPFP=2.6265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,616B, BPFP=0.8690 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,352B, BPFP=2.5136 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,336B, BPFP=0.8163 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,600B, BPFP=2.7485 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,336B, BPFP=1.7575 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,792B, BPFP=2.5964 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,672B, BPFP=0.2870 +⌛️ [2/4] FRONTEND: Frontend time: 1.892s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.461s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11392649 31.04247753 + layer.0.v_cache 0.00001670 0.00861108 + layer.1.k_cache 0.05249626 7.80382124 + layer.1.v_cache 0.00000506 0.00303126 + layer.2.k_cache 0.00416583 1.01233646 + layer.2.v_cache 0.00001711 0.00938487 + layer.3.k_cache 0.05931453 3.58798181 + layer.3.v_cache 0.00001789 0.01071679 + layer.4.k_cache 0.00069248 0.33022226 + layer.4.v_cache 0.00004729 0.02217478 + layer.4.output 0.16383441 646.51204819 + ------------------------------------------------------------------------------------- + TOTAL 0.08103179 268.78912326 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 106284 +BPFP 1.1770 bits/point +EBPFP 2.3539 equivalent bits/point +MSE 268.789123 +---------------------- -------------------------------------------------------- +Time: 3.360s Load: 0.007s, Pack+Encode: 1.892s, Decode+Unpack: 1.461s +---------------------- -------------------------------------------------------- +💾 Converting with 268.7891 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,872B, BPFP=0.4774 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,652B, BPFP=2.6017 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,848B, BPFP=0.9721 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,880B, BPFP=2.4734 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,244B, BPFP=1.0379 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,252B, BPFP=2.3690 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,104B, BPFP=1.0146 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,200B, BPFP=2.5266 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,584B, BPFP=1.7593 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,064B, BPFP=2.3378 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,908B, BPFP=0.2353 +⌛️ [2/4] FRONTEND: Frontend time: 2.032s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.329s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17268721 44.87648043 + layer.0.v_cache 0.00001453 0.00769775 + layer.1.k_cache 0.08276152 14.13484614 + layer.1.v_cache 0.00000545 0.00299413 + layer.2.k_cache 0.00739249 1.48052459 + layer.2.v_cache 0.00001625 0.00928268 + layer.3.k_cache 0.05929063 9.30178314 + layer.3.v_cache 0.00001766 0.01047411 + layer.4.k_cache 0.00068371 0.32687991 + layer.4.v_cache 0.00005055 0.02179862 + layer.4.output 0.14466690 573.45592705 + ------------------------------------------------------------------------------------- + TOTAL 0.07856402 240.25672064 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 115608 +BPFP 1.1304 bits/point +EBPFP 2.2608 equivalent bits/point +MSE 240.256721 +---------------------- -------------------------------------------------------- +Time: 3.368s Load: 0.006s, Pack+Encode: 2.032s, Decode+Unpack: 1.329s +---------------------- -------------------------------------------------------- +💾 Converting with 240.2567 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 112, 128) +Output shape: (1, 112, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.output: torch.Size([1, 112, 3584]) -> torch.Size([1, 1, 112, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,276B, BPFP=0.4570 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,204B, BPFP=2.2606 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,412B, BPFP=0.7550 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,756B, BPFP=2.1981 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,152B, BPFP=0.8583 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,740B, BPFP=2.1959 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,176B, BPFP=0.8616 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,148B, BPFP=2.2528 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,096B, BPFP=1.6875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,320B, BPFP=2.1373 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,780B, BPFP=0.2746 +⌛️ [2/4] FRONTEND: Frontend time: 1.968s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.279s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17321212 30.25192697 + layer.0.v_cache 0.00001450 0.00774435 + layer.1.k_cache 0.08643453 8.20371628 + layer.1.v_cache 0.00000659 0.00299473 + layer.2.k_cache 0.00690480 0.99733121 + layer.2.v_cache 0.00001899 0.00940096 + layer.3.k_cache 0.01205242 8.56242589 + layer.3.v_cache 0.00001923 0.01053361 + layer.4.k_cache 0.00071999 0.33978483 + layer.4.v_cache 0.00004947 0.02111722 + layer.4.output 10.17892331 477.23401626 + ------------------------------------------------------------------------------------- + TOTAL 4.20775858 199.35559352 + (elements=974,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 974848 +Total Bytes 126060 +BPFP 1.0345 bits/point +EBPFP 2.0690 equivalent bits/point +MSE 199.355594 +---------------------- -------------------------------------------------------- +Time: 3.251s Load: 0.004s, Pack+Encode: 1.968s, Decode+Unpack: 1.279s +---------------------- -------------------------------------------------------- +💾 Converting with 199.3556 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,712B, BPFP=0.4815 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,660B, BPFP=2.7805 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,400B, BPFP=0.7812 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,580B, BPFP=2.5888 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,852B, BPFP=0.8615 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,276B, BPFP=2.5348 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,812B, BPFP=0.8544 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,788B, BPFP=2.6257 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,240B, BPFP=1.8182 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,308B, BPFP=2.5405 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,284B, BPFP=0.2355 +⌛️ [2/4] FRONTEND: Frontend time: 1.940s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.436s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11687154 30.74972257 + layer.0.v_cache 0.00001423 0.00828937 + layer.1.k_cache 0.05324238 8.34708335 + layer.1.v_cache 0.00000539 0.00315041 + layer.2.k_cache 0.00397935 0.98277560 + layer.2.v_cache 0.00001766 0.01004817 + layer.3.k_cache 0.01210797 5.78882079 + layer.3.v_cache 0.00001803 0.01193321 + layer.4.k_cache 0.00075037 0.34068905 + layer.4.v_cache 0.00004481 0.02143816 + layer.4.output 0.15448949 613.45200893 + ------------------------------------------------------------------------------------- + TOTAL 0.07461636 255.31929489 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 109912 +BPFP 1.1480 bits/point +EBPFP 2.2960 equivalent bits/point +MSE 255.319295 +---------------------- -------------------------------------------------------- +Time: 3.380s Load: 0.005s, Pack+Encode: 1.940s, Decode+Unpack: 1.436s +---------------------- -------------------------------------------------------- +💾 Converting with 255.3193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,580B, BPFP=0.4977 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,388B, BPFP=2.9684 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,248B, BPFP=0.8194 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,496B, BPFP=2.7963 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,188B, BPFP=1.0008 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,980B, BPFP=2.6968 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,640B, BPFP=0.8951 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,428B, BPFP=2.7832 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,496B, BPFP=2.0247 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,208B, BPFP=2.7407 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,212B, BPFP=0.3090 +⌛️ [2/4] FRONTEND: Frontend time: 2.046s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203883 30.61408601 + layer.0.v_cache 0.00001438 0.00849540 + layer.1.k_cache 0.05633916 7.61314185 + layer.1.v_cache 0.00000602 0.00348860 + layer.2.k_cache 0.00244657 1.04746331 + layer.2.v_cache 0.00001847 0.01100230 + layer.3.k_cache 0.05492775 3.84554865 + layer.3.v_cache 0.00001883 0.01165624 + layer.4.k_cache 0.00062984 0.34377703 + layer.4.v_cache 0.00005178 0.02495769 + layer.4.output 0.18538215 662.36772487 + ------------------------------------------------------------------------------------- + TOTAL 0.09142157 275.29986419 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 110864 +BPFP 1.2580 bits/point +EBPFP 2.5160 equivalent bits/point +MSE 275.299864 +---------------------- -------------------------------------------------------- +Time: 3.397s Load: 0.005s, Pack+Encode: 2.046s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 275.2999 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,716B, BPFP=0.4715 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,720B, BPFP=2.7292 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,600B, BPFP=0.7986 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,860B, BPFP=2.5799 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,396B, BPFP=0.9368 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,068B, BPFP=2.4424 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,644B, BPFP=0.8063 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,888B, BPFP=2.5847 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,004B, BPFP=1.7368 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,228B, BPFP=2.4701 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,504B, BPFP=0.2605 +⌛️ [2/4] FRONTEND: Frontend time: 1.868s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.491s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13028670 30.80128581 + layer.0.v_cache 0.00001462 0.00818352 + layer.1.k_cache 0.04946813 11.17579617 + layer.1.v_cache 0.00000624 0.00279528 + layer.2.k_cache 0.00411291 1.02363264 + layer.2.v_cache 0.00001729 0.00947737 + layer.3.k_cache 0.01793825 4.59271410 + layer.3.v_cache 0.00001895 0.01038636 + layer.4.k_cache 0.00077120 0.35773167 + layer.4.v_cache 0.00004532 0.02169241 + layer.4.output 0.15109633 597.92777778 + ------------------------------------------------------------------------------------- + TOTAL 0.07413847 249.02930234 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 111628 +BPFP 1.1400 bits/point +EBPFP 2.2800 equivalent bits/point +MSE 249.029302 +---------------------- -------------------------------------------------------- +Time: 3.362s Load: 0.004s, Pack+Encode: 1.868s, Decode+Unpack: 1.491s +---------------------- -------------------------------------------------------- +💾 Converting with 249.0293 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,000B, BPFP=0.5482 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,020B, BPFP=2.1985 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,060B, BPFP=0.8388 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,736B, BPFP=2.1206 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,468B, BPFP=0.9507 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,672B, BPFP=2.1031 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,584B, BPFP=0.9825 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,072B, BPFP=2.2127 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,864B, BPFP=2.1557 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,608B, BPFP=2.0855 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,580B, BPFP=0.2968 +⌛️ [2/4] FRONTEND: Frontend time: 2.648s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11683119 35.02968664 + layer.0.v_cache 0.00001514 0.01180571 + layer.1.k_cache 0.01218004 8.82025842 + layer.1.v_cache 0.00000521 0.00390474 + layer.2.k_cache 0.00491684 1.11921658 + layer.2.v_cache 0.00001706 0.01088918 + layer.3.k_cache 0.10823302 5.92072925 + layer.3.v_cache 0.00002043 0.01353735 + layer.4.k_cache 0.00062874 0.32843259 + layer.4.v_cache 0.00004794 0.02760428 + layer.4.output 0.23833110 940.10581140 + ------------------------------------------------------------------------------------- + TOTAL 0.11242431 390.11922027 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 66664 +BPFP 1.0749 bits/point +EBPFP 2.1499 equivalent bits/point +MSE 390.119220 +---------------------- -------------------------------------------------------- +Time: 3.977s Load: 0.004s, Pack+Encode: 2.648s, Decode+Unpack: 1.325s +---------------------- -------------------------------------------------------- +💾 Converting with 390.1192 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,856B, BPFP=0.5800 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,008B, BPFP=2.5025 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,956B, BPFP=0.9237 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,752B, BPFP=2.4225 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,272B, BPFP=1.0225 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,736B, BPFP=2.4175 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,056B, BPFP=0.9550 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,996B, BPFP=2.4987 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,668B, BPFP=2.3963 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,624B, BPFP=2.3825 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,936B, BPFP=0.3096 +⌛️ [2/4] FRONTEND: Frontend time: 1.686s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.180s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14598725 36.54078125 + layer.0.v_cache 0.00001449 0.01159852 + layer.1.k_cache 0.01347294 10.24533813 + layer.1.v_cache 0.00000521 0.00383228 + layer.2.k_cache 0.00488209 1.06228806 + layer.2.v_cache 0.00001649 0.01094802 + layer.3.k_cache 0.04421538 5.18045532 + layer.3.v_cache 0.00001905 0.01385267 + layer.4.k_cache 0.00073779 0.33864697 + layer.4.v_cache 0.00005089 0.02709081 + layer.4.output 0.27158585 1077.24017857 + ------------------------------------------------------------------------------------- + TOTAL 0.12414721 446.71271071 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 64860 +BPFP 1.1923 bits/point +EBPFP 2.3846 equivalent bits/point +MSE 446.712711 +---------------------- -------------------------------------------------------- +Time: 2.869s Load: 0.003s, Pack+Encode: 1.686s, Decode+Unpack: 1.180s +---------------------- -------------------------------------------------------- +💾 Converting with 446.7127 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 46, 128) +Output shape: (1, 46, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.output: torch.Size([1, 46, 3584]) -> torch.Size([1, 1, 46, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,832B, BPFP=0.6223 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,008B, BPFP=2.7201 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,104B, BPFP=1.0543 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,804B, BPFP=2.6508 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,084B, BPFP=1.0476 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,776B, BPFP=2.6413 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,108B, BPFP=1.0557 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,976B, BPFP=2.7092 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,372B, BPFP=2.5041 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,624B, BPFP=2.5897 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,892B, BPFP=0.3344 +⌛️ [2/4] FRONTEND: Frontend time: 2.074s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.500s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14129612 35.22630244 + layer.0.v_cache 0.00001532 0.01156581 + layer.1.k_cache 0.01315293 11.36441703 + layer.1.v_cache 0.00000553 0.00393192 + layer.2.k_cache 0.00779452 1.06538341 + layer.2.v_cache 0.00001723 0.01055698 + layer.3.k_cache 0.01828384 5.12502455 + layer.3.v_cache 0.00001944 0.01602972 + layer.4.k_cache 0.00060559 0.34920979 + layer.4.v_cache 0.00005113 0.03035551 + layer.4.output 0.29517000 1170.57259317 + ------------------------------------------------------------------------------------- + TOTAL 0.13220186 485.13005467 + (elements=400,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 400384 +Total Bytes 64580 +BPFP 1.2904 bits/point +EBPFP 2.5807 equivalent bits/point +MSE 485.130055 +---------------------- -------------------------------------------------------- +Time: 3.578s Load: 0.004s, Pack+Encode: 2.074s, Decode+Unpack: 1.500s +---------------------- -------------------------------------------------------- +💾 Converting with 485.1301 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 110, 128) +Output shape: (1, 110, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.output: torch.Size([1, 110, 3584]) -> torch.Size([1, 1, 110, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,224B, BPFP=0.4580 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,196B, BPFP=2.3006 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,444B, BPFP=0.7733 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,836B, BPFP=2.2494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,176B, BPFP=0.8773 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,036B, BPFP=2.2778 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,540B, BPFP=0.9290 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,524B, BPFP=2.3472 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,352B, BPFP=1.7545 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,640B, BPFP=2.2216 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,304B, BPFP=0.2903 +⌛️ [2/4] FRONTEND: Frontend time: 2.031s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.420s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14476244 29.40703569 + layer.0.v_cache 0.00001476 0.00860113 + layer.1.k_cache 0.05665370 8.19770008 + layer.1.v_cache 0.00000525 0.00285922 + layer.2.k_cache 0.00786287 1.36728155 + layer.2.v_cache 0.00001801 0.00918765 + layer.3.k_cache 0.01410385 8.94114324 + layer.3.v_cache 0.00002027 0.01126082 + layer.4.k_cache 0.00066873 0.35911747 + layer.4.v_cache 0.00004626 0.02158298 + layer.4.output 10.36401748 486.83530844 + ------------------------------------------------------------------------------------- + TOTAL 4.28072226 203.30428994 + (elements=957,440) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 957440 +Total Bytes 128272 +BPFP 1.0718 bits/point +EBPFP 2.1436 equivalent bits/point +MSE 203.304290 +---------------------- -------------------------------------------------------- +Time: 3.457s Load: 0.006s, Pack+Encode: 2.031s, Decode+Unpack: 1.420s +---------------------- -------------------------------------------------------- +💾 Converting with 203.3043 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,620B, BPFP=0.5054 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,416B, BPFP=3.1667 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,328B, BPFP=0.8349 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,608B, BPFP=3.0108 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,840B, BPFP=0.9336 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,180B, BPFP=2.9282 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,844B, BPFP=0.9344 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,040B, BPFP=3.0941 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,948B, BPFP=2.1119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,100B, BPFP=2.9128 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,592B, BPFP=0.3746 +⌛️ [2/4] FRONTEND: Frontend time: 2.011s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.511s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10533771 33.30526620 + layer.0.v_cache 0.00001610 0.00942920 + layer.1.k_cache 0.05553475 7.62705259 + layer.1.v_cache 0.00000748 0.00361800 + layer.2.k_cache 0.02166171 0.87361578 + layer.2.v_cache 0.00002000 0.01035022 + layer.3.k_cache 0.06386927 5.61266035 + layer.3.v_cache 0.00001944 0.01212037 + layer.4.k_cache 0.00064315 0.35303135 + layer.4.v_cache 0.00005385 0.02622972 + layer.4.output 0.16801396 661.20827822 + ------------------------------------------------------------------------------------- + TOTAL 0.08372125 275.07596008 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 119516 +BPFP 1.3562 bits/point +EBPFP 2.7123 equivalent bits/point +MSE 275.075960 +---------------------- -------------------------------------------------------- +Time: 3.525s Load: 0.004s, Pack+Encode: 2.011s, Decode+Unpack: 1.511s +---------------------- -------------------------------------------------------- +💾 Converting with 275.0760 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 49, 128) +Output shape: (1, 49, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.output: torch.Size([1, 49, 3584]) -> torch.Size([1, 1, 49, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,872B, BPFP=0.5969 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,936B, BPFP=2.5306 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,872B, BPFP=0.9158 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,848B, BPFP=2.5026 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,372B, BPFP=1.0753 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,808B, BPFP=2.4898 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,896B, BPFP=0.9235 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,024B, BPFP=2.5587 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,332B, BPFP=2.3380 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,676B, BPFP=2.4477 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,172B, BPFP=0.4178 +⌛️ [2/4] FRONTEND: Frontend time: 1.918s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.269s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12410851 36.50080218 + layer.0.v_cache 0.00001740 0.01167383 + layer.1.k_cache 0.01401279 11.47211441 + layer.1.v_cache 0.00000579 0.00402447 + layer.2.k_cache 0.00545398 0.96616971 + layer.2.v_cache 0.00001955 0.01155520 + layer.3.k_cache 0.07546502 3.72012018 + layer.3.v_cache 0.00001978 0.01388071 + layer.4.k_cache 0.00063206 0.33181159 + layer.4.v_cache 0.00005205 0.02717984 + layer.4.output 0.27729562 1098.67037172 + ------------------------------------------------------------------------------------- + TOTAL 0.12710919 455.51481966 + (elements=426,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 426496 +Total Bytes 66808 +BPFP 1.2532 bits/point +EBPFP 2.5063 equivalent bits/point +MSE 455.514820 +---------------------- -------------------------------------------------------- +Time: 3.191s Load: 0.003s, Pack+Encode: 1.918s, Decode+Unpack: 1.269s +---------------------- -------------------------------------------------------- +💾 Converting with 455.5148 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,008B, BPFP=0.5504 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,020B, BPFP=2.1985 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,036B, BPFP=0.8322 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,908B, BPFP=2.1678 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,416B, BPFP=0.9364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,724B, BPFP=2.1173 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,296B, BPFP=0.9035 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,948B, BPFP=2.1787 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,992B, BPFP=2.1908 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,772B, BPFP=2.1305 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,076B, BPFP=0.3163 +⌛️ [2/4] FRONTEND: Frontend time: 1.831s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19197876 34.99285782 + layer.0.v_cache 0.00001449 0.01101121 + layer.1.k_cache 0.01466701 10.28522906 + layer.1.v_cache 0.00000519 0.00356476 + layer.2.k_cache 0.00720649 1.07310064 + layer.2.v_cache 0.00001785 0.01119540 + layer.3.k_cache 0.10851213 5.52017212 + layer.3.v_cache 0.00001996 0.01355624 + layer.4.k_cache 0.00062085 0.32093322 + layer.4.v_cache 0.00004770 0.02658669 + layer.4.output 0.23832821 945.23386591 + ------------------------------------------------------------------------------------- + TOTAL 0.11714047 392.28795697 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 67196 +BPFP 1.0835 bits/point +EBPFP 2.1671 equivalent bits/point +MSE 392.287957 +---------------------- -------------------------------------------------------- +Time: 3.035s Load: 0.003s, Pack+Encode: 1.831s, Decode+Unpack: 1.201s +---------------------- -------------------------------------------------------- +💾 Converting with 392.2880 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.5662 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,972B, BPFP=2.4424 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,792B, BPFP=0.8554 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,632B, BPFP=2.3382 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,204B, BPFP=0.9816 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,660B, BPFP=2.3468 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,996B, BPFP=0.9179 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,916B, BPFP=2.4252 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,172B, BPFP=2.1973 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,636B, BPFP=2.3395 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,448B, BPFP=0.3260 +⌛️ [2/4] FRONTEND: Frontend time: 1.791s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.125s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12665685 35.53604665 + layer.0.v_cache 0.00001436 0.01151659 + layer.1.k_cache 0.01497617 7.39807009 + layer.1.v_cache 0.00000537 0.00380725 + layer.2.k_cache 0.00986420 1.18111083 + layer.2.v_cache 0.00001802 0.01130468 + layer.3.k_cache 0.09846177 3.87462242 + layer.3.v_cache 0.00001923 0.01446989 + layer.4.k_cache 0.00062397 0.31027783 + layer.4.v_cache 0.00005014 0.02611485 + layer.4.output 0.26631049 1036.40353641 + ------------------------------------------------------------------------------------- + TOTAL 0.12440374 429.59953506 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 64276 +BPFP 1.1584 bits/point +EBPFP 2.3168 equivalent bits/point +MSE 429.599535 +---------------------- -------------------------------------------------------- +Time: 2.919s Load: 0.003s, Pack+Encode: 1.791s, Decode+Unpack: 1.125s +---------------------- -------------------------------------------------------- +💾 Converting with 429.5995 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,852B, BPFP=0.6029 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,032B, BPFP=2.6146 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,128B, BPFP=1.0182 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,812B, BPFP=2.5430 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,516B, BPFP=1.1445 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,828B, BPFP=2.5482 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,180B, BPFP=1.0352 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,032B, BPFP=2.6146 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,876B, BPFP=2.5638 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,668B, BPFP=2.4961 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,444B, BPFP=0.3927 +⌛️ [2/4] FRONTEND: Frontend time: 1.802s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.137s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18531116 41.00394948 + layer.0.v_cache 0.00001874 0.01175553 + layer.1.k_cache 0.01459585 8.96615918 + layer.1.v_cache 0.00000674 0.00439773 + layer.2.k_cache 0.01042949 0.97438121 + layer.2.v_cache 0.00001841 0.01157352 + layer.3.k_cache 0.15260293 5.94107691 + layer.3.v_cache 0.00002390 0.01560708 + layer.4.k_cache 0.00061422 0.34490983 + layer.4.v_cache 0.00004881 0.02770909 + layer.4.output 0.28297247 1124.22051711 + ------------------------------------------------------------------------------------- + TOTAL 0.13791044 466.28500820 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 67368 +BPFP 1.2900 bits/point +EBPFP 2.5800 equivalent bits/point +MSE 466.285008 +---------------------- -------------------------------------------------------- +Time: 2.942s Load: 0.002s, Pack+Encode: 1.802s, Decode+Unpack: 1.137s +---------------------- -------------------------------------------------------- +💾 Converting with 466.2850 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,996B, BPFP=0.5286 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,016B, BPFP=2.1229 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,000B, BPFP=0.7945 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,728B, BPFP=2.0466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,476B, BPFP=0.9206 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,740B, BPFP=2.0498 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,544B, BPFP=0.9386 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,992B, BPFP=2.1165 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,484B, BPFP=1.9820 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,728B, BPFP=2.0466 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,244B, BPFP=0.3497 +⌛️ [2/4] FRONTEND: Frontend time: 1.771s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.242s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12052289 35.54544326 + layer.0.v_cache 0.00001546 0.01291629 + layer.1.k_cache 0.01434235 8.72064106 + layer.1.v_cache 0.00000574 0.00406853 + layer.2.k_cache 0.00254285 1.19550155 + layer.2.v_cache 0.00001710 0.01173787 + layer.3.k_cache 0.03737745 8.17054102 + layer.3.v_cache 0.00001946 0.01600385 + layer.4.k_cache 0.00063766 0.35262567 + layer.4.v_cache 0.00005039 0.03013105 + layer.4.output 0.23032015 906.54706416 + ------------------------------------------------------------------------------------- + TOTAL 0.10516308 376.46406231 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 67948 +BPFP 1.0585 bits/point +EBPFP 2.1170 equivalent bits/point +MSE 376.464062 +---------------------- -------------------------------------------------------- +Time: 3.016s Load: 0.003s, Pack+Encode: 1.771s, Decode+Unpack: 1.242s +---------------------- -------------------------------------------------------- +💾 Converting with 376.4641 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,504B, BPFP=0.5287 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,208B, BPFP=3.2111 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,464B, BPFP=0.9426 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,644B, BPFP=2.8809 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,664B, BPFP=1.1959 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,064B, BPFP=2.9696 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,968B, BPFP=1.2601 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,712B, BPFP=3.1064 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,268B, BPFP=1.9569 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,644B, BPFP=2.8809 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,028B, BPFP=0.3025 +⌛️ [2/4] FRONTEND: Frontend time: 1.861s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11335979 34.48954814 + layer.0.v_cache 0.00001513 0.00979720 + layer.1.k_cache 0.01655370 10.14540306 + layer.1.v_cache 0.00000667 0.00371944 + layer.2.k_cache 0.00428703 1.12886130 + layer.2.v_cache 0.00001789 0.01113990 + layer.3.k_cache 0.04938591 8.18074489 + layer.3.v_cache 0.00002111 0.01331043 + layer.4.k_cache 0.00066117 0.31240365 + layer.4.v_cache 0.00004716 0.02699717 + layer.4.output 0.18372514 727.44099903 + ------------------------------------------------------------------------------------- + TOTAL 0.08649597 302.72993638 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 109168 +BPFP 1.3559 bits/point +EBPFP 2.7118 equivalent bits/point +MSE 302.729936 +---------------------- -------------------------------------------------------- +Time: 3.185s Load: 0.004s, Pack+Encode: 1.861s, Decode+Unpack: 1.321s +---------------------- -------------------------------------------------------- +💾 Converting with 302.7299 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,780B, BPFP=0.6321 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,108B, BPFP=2.8793 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,824B, BPFP=1.0028 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,980B, BPFP=2.8338 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,496B, BPFP=1.2415 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,028B, BPFP=2.8509 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,072B, BPFP=1.0909 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,152B, BPFP=2.8949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,128B, BPFP=2.8864 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,972B, BPFP=2.8310 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,716B, BPFP=0.5944 +⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14777968 32.95348566 + layer.0.v_cache 0.00001531 0.01222777 + layer.1.k_cache 0.01728138 10.90094202 + layer.1.v_cache 0.00000547 0.00430903 + layer.2.k_cache 0.00823456 0.98283785 + layer.2.v_cache 0.00001932 0.01312070 + layer.3.k_cache 0.16770924 5.99733110 + layer.3.v_cache 0.00002046 0.01670642 + layer.4.k_cache 0.00061569 0.33110970 + layer.4.v_cache 0.00005592 0.03163488 + layer.4.output 0.30867824 1211.93060065 + ------------------------------------------------------------------------------------- + TOTAL 0.14720498 502.04458292 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 71256 +BPFP 1.4885 bits/point +EBPFP 2.9769 equivalent bits/point +MSE 502.044583 +---------------------- -------------------------------------------------------- +Time: 3.091s Load: 0.003s, Pack+Encode: 1.735s, Decode+Unpack: 1.354s +---------------------- -------------------------------------------------------- +💾 Converting with 502.0446 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,676B, BPFP=0.5361 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,512B, BPFP=3.1074 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,440B, BPFP=0.8894 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,512B, BPFP=2.9071 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,768B, BPFP=1.1554 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,532B, BPFP=2.9111 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,708B, BPFP=1.1434 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,008B, BPFP=3.0064 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,708B, BPFP=2.1450 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,440B, BPFP=2.8926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,504B, BPFP=0.3006 +⌛️ [2/4] FRONTEND: Frontend time: 2.065s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.397s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14408422 37.98055639 + layer.0.v_cache 0.00002097 0.00951841 + layer.1.k_cache 0.05735342 11.21242582 + layer.1.v_cache 0.00000575 0.00365638 + layer.2.k_cache 0.01678689 1.13900639 + layer.2.v_cache 0.00001894 0.01069373 + layer.3.k_cache 0.03029843 7.25925543 + layer.3.v_cache 0.00002125 0.01295364 + layer.4.k_cache 0.00062582 0.33688690 + layer.4.v_cache 0.00005050 0.02543937 + layer.4.output 0.17433185 690.77003205 + ------------------------------------------------------------------------------------- + TOTAL 0.08644642 287.84591864 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 113808 +BPFP 1.3411 bits/point +EBPFP 2.6821 equivalent bits/point +MSE 287.845919 +---------------------- -------------------------------------------------------- +Time: 3.467s Load: 0.004s, Pack+Encode: 2.065s, Decode+Unpack: 1.397s +---------------------- -------------------------------------------------------- +💾 Converting with 287.8459 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,772B, BPFP=0.6293 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,144B, BPFP=2.8920 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,896B, BPFP=1.0284 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,960B, BPFP=2.8267 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,160B, BPFP=1.1222 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,036B, BPFP=2.8537 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,204B, BPFP=1.1378 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,068B, BPFP=2.8651 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,252B, BPFP=2.9304 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,708B, BPFP=2.7372 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,596B, BPFP=0.3346 +⌛️ [2/4] FRONTEND: Frontend time: 1.852s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13243668 33.93382124 + layer.0.v_cache 0.00001576 0.01201362 + layer.1.k_cache 0.01981235 9.96846078 + layer.1.v_cache 0.00000577 0.00425720 + layer.2.k_cache 0.01756107 1.05676590 + layer.2.v_cache 0.00001728 0.01208384 + layer.3.k_cache 0.04920139 4.19513598 + layer.3.v_cache 0.00001837 0.01558568 + layer.4.k_cache 0.00061593 0.31333074 + layer.4.v_cache 0.00004907 0.02959601 + layer.4.output 0.30854983 1216.73193994 + ------------------------------------------------------------------------------------- + TOTAL 0.13997544 503.92144886 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 65796 +BPFP 1.3744 bits/point +EBPFP 2.7488 equivalent bits/point +MSE 503.921449 +---------------------- -------------------------------------------------------- +Time: 3.222s Load: 0.003s, Pack+Encode: 1.852s, Decode+Unpack: 1.368s +---------------------- -------------------------------------------------------- +💾 Converting with 503.9214 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,844B, BPFP=0.5763 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,096B, BPFP=2.5300 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,876B, BPFP=0.8988 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,852B, BPFP=2.4537 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,660B, BPFP=1.1438 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,840B, BPFP=2.4500 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,076B, BPFP=0.9613 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,996B, BPFP=2.4987 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,896B, BPFP=2.4675 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,728B, BPFP=2.4150 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,248B, BPFP=0.3236 +⌛️ [2/4] FRONTEND: Frontend time: 1.887s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20666794 34.70549805 + layer.0.v_cache 0.00002565 0.01161922 + layer.1.k_cache 0.01518770 10.94135620 + layer.1.v_cache 0.00000581 0.00395008 + layer.2.k_cache 0.01031687 0.97151413 + layer.2.v_cache 0.00001873 0.01152898 + layer.3.k_cache 0.09638391 6.35959717 + layer.3.v_cache 0.00001967 0.01473441 + layer.4.k_cache 0.00062172 0.31810780 + layer.4.v_cache 0.00004827 0.02795703 + layer.4.output 0.27162739 1066.83714286 + ------------------------------------------------------------------------------------- + TOTAL 0.13121694 442.42505077 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 66112 +BPFP 1.2153 bits/point +EBPFP 2.4306 equivalent bits/point +MSE 442.425051 +---------------------- -------------------------------------------------------- +Time: 3.266s Load: 0.003s, Pack+Encode: 1.887s, Decode+Unpack: 1.376s +---------------------- -------------------------------------------------------- +💾 Converting with 442.4251 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,212B, BPFP=0.5083 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,488B, BPFP=3.3290 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,124B, BPFP=0.9476 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,120B, BPFP=3.2445 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,352B, BPFP=1.0000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,840B, BPFP=3.1801 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,204B, BPFP=0.9660 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,124B, BPFP=3.2454 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,348B, BPFP=1.9182 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,828B, BPFP=3.1774 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,804B, BPFP=0.3218 +⌛️ [2/4] FRONTEND: Frontend time: 2.199s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.494s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13431002 31.44847556 + layer.0.v_cache 0.00001720 0.01065596 + layer.1.k_cache 0.01179634 10.75933299 + layer.1.v_cache 0.00000537 0.00352252 + layer.2.k_cache 0.00798416 0.99103479 + layer.2.v_cache 0.00001783 0.01013838 + layer.3.k_cache 0.05795463 4.78330186 + layer.3.v_cache 0.00001974 0.01271937 + layer.4.k_cache 0.00062059 0.31480781 + layer.4.v_cache 0.00005204 0.02485213 + layer.4.output 0.19989206 793.68940389 + ------------------------------------------------------------------------------------- + TOTAL 0.09482484 329.65792168 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 103444 +BPFP 1.3982 bits/point +EBPFP 2.7964 equivalent bits/point +MSE 329.657922 +---------------------- -------------------------------------------------------- +Time: 3.697s Load: 0.004s, Pack+Encode: 2.199s, Decode+Unpack: 1.494s +---------------------- -------------------------------------------------------- +💾 Converting with 329.6579 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,868B, BPFP=0.5837 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,164B, BPFP=2.5513 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,832B, BPFP=0.8850 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,828B, BPFP=2.4463 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,884B, BPFP=1.2138 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,832B, BPFP=2.4475 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,092B, BPFP=0.9663 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,268B, BPFP=2.8963 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,004B, BPFP=2.5013 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,788B, BPFP=2.4337 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,292B, BPFP=0.3255 +⌛️ [2/4] FRONTEND: Frontend time: 1.806s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15467889 35.97903320 + layer.0.v_cache 0.00001741 0.01155765 + layer.1.k_cache 0.01620592 9.96381897 + layer.1.v_cache 0.00000548 0.00387326 + layer.2.k_cache 0.00723796 1.16286636 + layer.2.v_cache 0.00001737 0.01182663 + layer.3.k_cache 0.06661982 4.29639709 + layer.3.v_cache 0.00002040 0.01436098 + layer.4.k_cache 0.00058776 0.32877922 + layer.4.v_cache 0.00004656 0.02752357 + layer.4.output 0.27162074 1071.67696429 + ------------------------------------------------------------------------------------- + TOTAL 0.12628134 444.32581100 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 67852 +BPFP 1.2473 bits/point +EBPFP 2.4946 equivalent bits/point +MSE 444.325811 +---------------------- -------------------------------------------------------- +Time: 3.027s Load: 0.003s, Pack+Encode: 1.806s, Decode+Unpack: 1.219s +---------------------- -------------------------------------------------------- +💾 Converting with 444.3258 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,868B, BPFP=0.5723 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,060B, BPFP=2.4694 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,808B, BPFP=0.8603 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,788B, BPFP=2.3860 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,244B, BPFP=0.9939 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,664B, BPFP=2.3480 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,880B, BPFP=0.8824 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,960B, BPFP=2.4387 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,400B, BPFP=2.2672 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,664B, BPFP=2.3480 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,780B, BPFP=0.2967 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.462s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11261671 31.88520604 + layer.0.v_cache 0.00001829 0.01167796 + layer.1.k_cache 0.01190703 7.57567581 + layer.1.v_cache 0.00000528 0.00380233 + layer.2.k_cache 0.00973116 1.02740703 + layer.2.v_cache 0.00001668 0.01055852 + layer.3.k_cache 0.04200072 3.74322480 + layer.3.v_cache 0.00001874 0.01445803 + layer.4.k_cache 0.00062040 0.29957657 + layer.4.v_cache 0.00004776 0.02679863 + layer.4.output 0.26626884 1054.14084384 + ------------------------------------------------------------------------------------- + TOTAL 0.12005086 436.68142898 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 64116 +BPFP 1.1555 bits/point +EBPFP 2.3110 equivalent bits/point +MSE 436.681429 +---------------------- -------------------------------------------------------- +Time: 3.294s Load: 0.002s, Pack+Encode: 1.830s, Decode+Unpack: 1.462s +---------------------- -------------------------------------------------------- +💾 Converting with 436.6814 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,888B, BPFP=0.5673 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,996B, BPFP=2.4026 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,852B, BPFP=0.8570 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,836B, BPFP=2.3546 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,548B, BPFP=1.0661 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,848B, BPFP=2.3582 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,052B, BPFP=0.9171 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,920B, BPFP=2.3798 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,896B, BPFP=2.3726 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,748B, BPFP=2.3281 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,452B, BPFP=0.3628 +⌛️ [2/4] FRONTEND: Frontend time: 1.724s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.155s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12932784 31.51527053 + layer.0.v_cache 0.00001491 0.01175707 + layer.1.k_cache 0.01488123 7.57189002 + layer.1.v_cache 0.00000563 0.00404234 + layer.2.k_cache 0.00939683 1.06002441 + layer.2.v_cache 0.00001981 0.01164338 + layer.3.k_cache 0.07051506 3.60202555 + layer.3.v_cache 0.00001927 0.01521748 + layer.4.k_cache 0.00064413 0.34247651 + layer.4.v_cache 0.00005939 0.02910172 + layer.4.output 0.26127321 1019.16938530 + ------------------------------------------------------------------------------------- + TOTAL 0.12081156 422.25583212 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 67036 +BPFP 1.1849 bits/point +EBPFP 2.3698 equivalent bits/point +MSE 422.255832 +---------------------- -------------------------------------------------------- +Time: 2.881s Load: 0.002s, Pack+Encode: 1.724s, Decode+Unpack: 1.155s +---------------------- -------------------------------------------------------- +💾 Converting with 422.2558 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,896B, BPFP=0.5697 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,092B, BPFP=2.4315 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,944B, BPFP=0.8846 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,844B, BPFP=2.3570 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,284B, BPFP=0.9868 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,888B, BPFP=2.3702 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,244B, BPFP=0.9748 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,220B, BPFP=2.4700 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,040B, BPFP=2.4159 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,844B, BPFP=2.3570 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,048B, BPFP=0.3455 +⌛️ [2/4] FRONTEND: Frontend time: 1.812s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13190371 34.23462853 + layer.0.v_cache 0.00001552 0.01143925 + layer.1.k_cache 0.01668521 8.52119563 + layer.1.v_cache 0.00000581 0.00423331 + layer.2.k_cache 0.00490881 1.01720957 + layer.2.v_cache 0.00001873 0.01188839 + layer.3.k_cache 0.11824917 5.34491906 + layer.3.v_cache 0.00002115 0.01584157 + layer.4.k_cache 0.00061206 0.31031583 + layer.4.v_cache 0.00004844 0.02772967 + layer.4.output 0.26122982 1023.76210508 + ------------------------------------------------------------------------------------- + TOTAL 0.12359279 424.46083155 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 67344 +BPFP 1.1903 bits/point +EBPFP 2.3807 equivalent bits/point +MSE 424.460832 +---------------------- -------------------------------------------------------- +Time: 3.018s Load: 0.003s, Pack+Encode: 1.812s, Decode+Unpack: 1.203s +---------------------- -------------------------------------------------------- +💾 Converting with 424.4608 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,772B, BPFP=0.4708 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,984B, BPFP=2.7147 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,608B, BPFP=0.9524 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,860B, BPFP=2.5238 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,852B, BPFP=0.9939 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,460B, BPFP=2.4558 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,352B, BPFP=1.0788 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,112B, BPFP=2.5666 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,548B, BPFP=1.7914 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,384B, BPFP=2.4429 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,880B, BPFP=0.2882 +⌛️ [2/4] FRONTEND: Frontend time: 2.067s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10326390 32.22992209 + layer.0.v_cache 0.00001541 0.00854584 + layer.1.k_cache 0.04976616 12.31934921 + layer.1.v_cache 0.00000534 0.00311611 + layer.2.k_cache 0.01168035 1.43986445 + layer.2.v_cache 0.00001798 0.00948562 + layer.3.k_cache 0.02907468 8.77705649 + layer.3.v_cache 0.00001946 0.01156763 + layer.4.k_cache 0.00073861 0.36236630 + layer.4.v_cache 0.00005077 0.02243597 + layer.4.output 0.14788401 584.47428183 + ------------------------------------------------------------------------------------- + TOTAL 0.07234240 243.91198133 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 117812 +BPFP 1.1770 bits/point +EBPFP 2.3540 equivalent bits/point +MSE 243.911981 +---------------------- -------------------------------------------------------- +Time: 3.276s Load: 0.006s, Pack+Encode: 2.067s, Decode+Unpack: 1.203s +---------------------- -------------------------------------------------------- +💾 Converting with 243.9120 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 64, 128) +Output shape: (1, 64, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.output: torch.Size([1, 64, 3584]) -> torch.Size([1, 1, 64, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,740B, BPFP=0.4248 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,828B, BPFP=2.1553 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,088B, BPFP=0.7539 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,680B, BPFP=1.8750 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,440B, BPFP=0.8398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,768B, BPFP=1.8965 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,020B, BPFP=0.7373 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,024B, BPFP=1.9590 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,980B, BPFP=1.7041 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,644B, BPFP=1.8662 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,240B, BPFP=0.2525 +⌛️ [2/4] FRONTEND: Frontend time: 2.050s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.191s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09722494 36.96508408 + layer.0.v_cache 0.00001391 0.00675809 + layer.1.k_cache 0.01806639 14.52134228 + layer.1.v_cache 0.00000557 0.00282023 + layer.2.k_cache 0.00423524 1.47742856 + layer.2.v_cache 0.00001868 0.01052097 + layer.3.k_cache 0.03759564 10.15134430 + layer.3.v_cache 0.00001825 0.01023351 + layer.4.k_cache 0.00061753 0.30082059 + layer.4.v_cache 0.00004894 0.01845010 + layer.4.output 0.21408452 795.86090960 + ------------------------------------------------------------------------------------- + TOTAL 0.09743746 331.44065705 + (elements=557,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 557056 +Total Bytes 65452 +BPFP 0.9400 bits/point +EBPFP 1.8799 equivalent bits/point +MSE 331.440657 +---------------------- -------------------------------------------------------- +Time: 3.244s Load: 0.003s, Pack+Encode: 2.050s, Decode+Unpack: 1.191s +---------------------- -------------------------------------------------------- +💾 Converting with 331.4407 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,072B, BPFP=0.4706 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,096B, BPFP=2.4657 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,212B, BPFP=0.7984 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,496B, BPFP=2.3738 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,852B, BPFP=0.8964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,216B, BPFP=2.3309 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,448B, BPFP=0.8346 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,856B, BPFP=2.4289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,252B, BPFP=2.0300 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,996B, BPFP=2.2972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,368B, BPFP=0.2488 +⌛️ [2/4] FRONTEND: Frontend time: 1.896s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.417s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11648470 37.16138174 + layer.0.v_cache 0.00001454 0.00768511 + layer.1.k_cache 0.02721847 8.59696931 + layer.1.v_cache 0.00000531 0.00294661 + layer.2.k_cache 0.01045542 1.28129982 + layer.2.v_cache 0.00001699 0.00930761 + layer.3.k_cache 0.01744144 8.69043687 + layer.3.v_cache 0.00001839 0.01072547 + layer.4.k_cache 0.00076342 0.35380180 + layer.4.v_cache 0.00004795 0.02431597 + layer.4.output 11.17676328 522.06775210 + ------------------------------------------------------------------------------------- + TOTAL 4.61234174 218.27136088 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 121864 +BPFP 1.0981 bits/point +EBPFP 2.1962 equivalent bits/point +MSE 218.271361 +---------------------- -------------------------------------------------------- +Time: 3.317s Load: 0.004s, Pack+Encode: 1.896s, Decode+Unpack: 1.417s +---------------------- -------------------------------------------------------- +💾 Converting with 218.2714 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,576B, BPFP=0.5031 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,616B, BPFP=2.8547 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,140B, BPFP=0.8086 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,728B, BPFP=2.6812 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,024B, BPFP=0.9812 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,068B, BPFP=2.5523 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,308B, BPFP=0.8414 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,316B, BPFP=2.7961 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,820B, BPFP=1.7227 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,664B, BPFP=2.6688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,248B, BPFP=0.2859 +⌛️ [2/4] FRONTEND: Frontend time: 2.037s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.488s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11433966 31.93243713 + layer.0.v_cache 0.00001437 0.00837241 + layer.1.k_cache 0.05315235 7.87515564 + layer.1.v_cache 0.00000524 0.00287378 + layer.2.k_cache 0.00678293 0.97724800 + layer.2.v_cache 0.00001704 0.00937884 + layer.3.k_cache 0.01951189 4.29553566 + layer.3.v_cache 0.00001758 0.01063680 + layer.4.k_cache 0.00076624 0.31605544 + layer.4.v_cache 0.00004856 0.02409325 + layer.4.output 0.16992235 672.47600446 + ------------------------------------------------------------------------------------- + TOTAL 0.08141837 279.57551872 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 104508 +BPFP 1.2007 bits/point +EBPFP 2.4014 equivalent bits/point +MSE 279.575519 +---------------------- -------------------------------------------------------- +Time: 3.532s Load: 0.006s, Pack+Encode: 2.037s, Decode+Unpack: 1.488s +---------------------- -------------------------------------------------------- +💾 Converting with 279.5755 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,700B, BPFP=0.4688 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,800B, BPFP=2.7431 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,744B, BPFP=0.8236 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,884B, BPFP=2.5840 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,088B, BPFP=0.8833 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,228B, BPFP=2.4701 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,112B, BPFP=0.8875 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,240B, BPFP=2.6458 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,260B, BPFP=1.9549 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,528B, BPFP=2.5222 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,064B, BPFP=0.2744 +⌛️ [2/4] FRONTEND: Frontend time: 1.958s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11570581 32.01932780 + layer.0.v_cache 0.00001517 0.00812021 + layer.1.k_cache 0.03236232 10.54931776 + layer.1.v_cache 0.00000662 0.00290193 + layer.2.k_cache 0.00641356 1.10695513 + layer.2.v_cache 0.00001676 0.00921159 + layer.3.k_cache 0.02648412 8.15587294 + layer.3.v_cache 0.00001896 0.01035486 + layer.4.k_cache 0.00074524 0.36771605 + layer.4.v_cache 0.00004843 0.02228947 + layer.4.output 0.15113260 599.03670635 + ------------------------------------------------------------------------------------- + TOTAL 0.07292619 249.73582425 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 114648 +BPFP 1.1708 bits/point +EBPFP 2.3417 equivalent bits/point +MSE 249.735824 +---------------------- -------------------------------------------------------- +Time: 3.307s Load: 0.008s, Pack+Encode: 1.958s, Decode+Unpack: 1.341s +---------------------- -------------------------------------------------------- +💾 Converting with 249.7358 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 114, 128) +Output shape: (1, 114, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.output: torch.Size([1, 114, 3584]) -> torch.Size([1, 1, 114, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,304B, BPFP=0.4529 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,176B, BPFP=2.2171 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,748B, BPFP=0.7878 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,732B, BPFP=2.1562 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,024B, BPFP=0.8257 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,868B, BPFP=2.1749 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,148B, BPFP=0.8427 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,200B, BPFP=2.2204 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,056B, BPFP=1.6524 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,916B, BPFP=2.1815 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,268B, BPFP=0.2598 +⌛️ [2/4] FRONTEND: Frontend time: 2.080s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10081583 30.74013372 + layer.0.v_cache 0.00001633 0.00833555 + layer.1.k_cache 0.08473034 8.72943758 + layer.1.v_cache 0.00000563 0.00298055 + layer.2.k_cache 0.00792928 1.02223848 + layer.2.v_cache 0.00001730 0.00970666 + layer.3.k_cache 0.01019665 8.48153901 + layer.3.v_cache 0.00001810 0.01042828 + layer.4.k_cache 0.00077074 0.34082650 + layer.4.v_cache 0.00005022 0.02155510 + layer.4.output 10.00037814 464.13619987 + ------------------------------------------------------------------------------------- + TOTAL 4.12983514 194.01885768 + (elements=992,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 992256 +Total Bytes 126440 +BPFP 1.0194 bits/point +EBPFP 2.0388 equivalent bits/point +MSE 194.018858 +---------------------- -------------------------------------------------------- +Time: 3.394s Load: 0.005s, Pack+Encode: 2.080s, Decode+Unpack: 1.309s +---------------------- -------------------------------------------------------- +💾 Converting with 194.0189 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 98, 128) +Output shape: (1, 98, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.output: torch.Size([1, 98, 3584]) -> torch.Size([1, 1, 98, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,820B, BPFP=0.4496 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,852B, BPFP=2.5274 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,904B, BPFP=0.7819 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,264B, BPFP=2.4337 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,592B, BPFP=1.0510 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,872B, BPFP=2.3712 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,252B, BPFP=0.9968 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,408B, BPFP=2.4566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,264B, BPFP=1.7959 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,636B, BPFP=2.3335 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,432B, BPFP=0.2376 +⌛️ [2/4] FRONTEND: Frontend time: 1.897s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11008905 32.17716637 + layer.0.v_cache 0.00001718 0.00822192 + layer.1.k_cache 0.06240322 9.52876998 + layer.1.v_cache 0.00000548 0.00303878 + layer.2.k_cache 0.00473963 1.32429131 + layer.2.v_cache 0.00001718 0.00911916 + layer.3.k_cache 0.01141277 9.30881672 + layer.3.v_cache 0.00001769 0.01022486 + layer.4.k_cache 0.00076610 0.35014997 + layer.4.v_cache 0.00004814 0.02158911 + layer.4.output 0.02426045 564.29805940 + ------------------------------------------------------------------------------------- + TOTAL 0.02113762 235.46045906 + (elements=852,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 852992 +Total Bytes 118296 +BPFP 1.1095 bits/point +EBPFP 2.2189 equivalent bits/point +MSE 235.460459 +---------------------- -------------------------------------------------------- +Time: 3.290s Load: 0.005s, Pack+Encode: 1.897s, Decode+Unpack: 1.387s +---------------------- -------------------------------------------------------- +💾 Converting with 235.4605 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,764B, BPFP=0.4694 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,768B, BPFP=2.6780 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,076B, BPFP=0.8621 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,860B, BPFP=2.5238 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,644B, BPFP=0.9586 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,576B, BPFP=2.4755 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,376B, BPFP=0.9130 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,432B, BPFP=3.1304 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,076B, BPFP=1.7113 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,644B, BPFP=2.4871 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,920B, BPFP=0.2892 +⌛️ [2/4] FRONTEND: Frontend time: 1.882s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.540s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08919133 30.03169051 + layer.0.v_cache 0.00001414 0.00756375 + layer.1.k_cache 0.08591949 11.29995329 + layer.1.v_cache 0.00000528 0.00279237 + layer.2.k_cache 0.00389855 1.15667600 + layer.2.v_cache 0.00001738 0.00938876 + layer.3.k_cache 0.07679920 6.13746842 + layer.3.v_cache 0.00002028 0.01065894 + layer.4.k_cache 0.00072407 0.35346110 + layer.4.v_cache 0.00004779 0.02145814 + layer.4.output 0.14786680 583.29828222 + ------------------------------------------------------------------------------------- + TOTAL 0.07598265 243.06582864 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 119136 +BPFP 1.1902 bits/point +EBPFP 2.3804 equivalent bits/point +MSE 243.065829 +---------------------- -------------------------------------------------------- +Time: 3.427s Load: 0.005s, Pack+Encode: 1.882s, Decode+Unpack: 1.540s +---------------------- -------------------------------------------------------- +💾 Converting with 243.0658 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,156B, BPFP=0.4609 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,272B, BPFP=2.3762 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,272B, BPFP=0.7699 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,900B, BPFP=2.3218 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,572B, BPFP=0.9597 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,944B, BPFP=2.4743 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,868B, BPFP=0.8569 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,352B, BPFP=2.3879 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,740B, BPFP=1.8604 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,680B, BPFP=2.2897 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,736B, BPFP=0.2657 +⌛️ [2/4] FRONTEND: Frontend time: 2.065s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.427s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11353527 31.64277207 + layer.0.v_cache 0.00001674 0.00851307 + layer.1.k_cache 0.04276301 8.28749441 + layer.1.v_cache 0.00000542 0.00299592 + layer.2.k_cache 0.00247095 1.46417479 + layer.2.v_cache 0.00001885 0.01068276 + layer.3.k_cache 0.02722364 5.32399971 + layer.3.v_cache 0.00001942 0.01089021 + layer.4.k_cache 0.00068256 0.35766402 + layer.4.v_cache 0.00004843 0.02454944 + layer.4.output 10.65454330 491.96762350 + ------------------------------------------------------------------------------------- + TOTAL 4.39815220 205.34747652 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 127492 +BPFP 1.0951 bits/point +EBPFP 2.1903 equivalent bits/point +MSE 205.347477 +---------------------- -------------------------------------------------------- +Time: 3.498s Load: 0.006s, Pack+Encode: 2.065s, Decode+Unpack: 1.427s +---------------------- -------------------------------------------------------- +💾 Converting with 205.3475 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,604B, BPFP=0.4844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,320B, BPFP=2.8497 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,212B, BPFP=0.7835 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,100B, BPFP=2.6228 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,624B, BPFP=0.8601 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,552B, BPFP=2.5208 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,316B, BPFP=0.8028 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,696B, BPFP=2.7336 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,720B, BPFP=1.6220 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,864B, BPFP=2.5789 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,448B, BPFP=0.2776 +⌛️ [2/4] FRONTEND: Frontend time: 2.007s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.282s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12550225 28.40822928 + layer.0.v_cache 0.00001468 0.00839171 + layer.1.k_cache 0.05234316 7.44689069 + layer.1.v_cache 0.00000508 0.00281714 + layer.2.k_cache 0.00733058 0.88967805 + layer.2.v_cache 0.00001689 0.00911544 + layer.3.k_cache 0.01592888 3.25389026 + layer.3.v_cache 0.00001847 0.01122119 + layer.4.k_cache 0.00069460 0.31753451 + layer.4.v_cache 0.00005205 0.02139165 + layer.4.output 0.16184913 622.36782526 + ------------------------------------------------------------------------------------- + TOTAL 0.07852062 258.64376098 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 106456 +BPFP 1.1648 bits/point +EBPFP 2.3297 equivalent bits/point +MSE 258.643761 +---------------------- -------------------------------------------------------- +Time: 3.293s Load: 0.004s, Pack+Encode: 2.007s, Decode+Unpack: 1.282s +---------------------- -------------------------------------------------------- +💾 Converting with 258.6438 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,672B, BPFP=0.5353 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,284B, BPFP=2.8614 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,380B, BPFP=0.8774 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,272B, BPFP=2.6587 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,996B, BPFP=1.0008 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,832B, BPFP=2.5705 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,604B, BPFP=0.9223 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,068B, BPFP=2.8181 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,136B, BPFP=2.0304 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,616B, BPFP=2.7276 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,608B, BPFP=0.2750 +⌛️ [2/4] FRONTEND: Frontend time: 2.017s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.250s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11879376 37.38120368 + layer.0.v_cache 0.00001438 0.00926644 + layer.1.k_cache 0.05518859 10.89391777 + layer.1.v_cache 0.00000514 0.00317536 + layer.2.k_cache 0.00243519 1.04023821 + layer.2.v_cache 0.00001742 0.00962218 + layer.3.k_cache 0.07959401 5.97304320 + layer.3.v_cache 0.00001812 0.01144336 + layer.4.k_cache 0.00069358 0.33191671 + layer.4.v_cache 0.00004620 0.02310474 + layer.4.output 0.17426339 690.56982601 + ------------------------------------------------------------------------------------- + TOTAL 0.08686177 287.62739492 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 104468 +BPFP 1.2310 bits/point +EBPFP 2.4620 equivalent bits/point +MSE 287.627395 +---------------------- -------------------------------------------------------- +Time: 3.270s Load: 0.003s, Pack+Encode: 2.017s, Decode+Unpack: 1.250s +---------------------- -------------------------------------------------------- +💾 Converting with 287.6274 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,592B, BPFP=0.4939 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,724B, BPFP=2.8056 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,180B, BPFP=0.7965 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,088B, BPFP=2.6845 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,716B, BPFP=0.8986 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,712B, BPFP=2.6128 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,384B, BPFP=0.8354 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,404B, BPFP=2.7447 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,476B, BPFP=1.9962 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,136B, BPFP=2.6936 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,868B, BPFP=0.2686 +⌛️ [2/4] FRONTEND: Frontend time: 1.814s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11325797 31.77290754 + layer.0.v_cache 0.00001565 0.00845100 + layer.1.k_cache 0.07651102 7.47332168 + layer.1.v_cache 0.00000530 0.00339758 + layer.2.k_cache 0.00872486 0.95646612 + layer.2.v_cache 0.00001696 0.00992999 + layer.3.k_cache 0.12876377 4.84452410 + layer.3.v_cache 0.00001905 0.01156969 + layer.4.k_cache 0.00069190 0.35755362 + layer.4.v_cache 0.00004968 0.02242740 + layer.4.output 0.16578155 649.60469294 + ------------------------------------------------------------------------------------- + TOTAL 0.08756041 270.15843525 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 107280 +BPFP 1.2025 bits/point +EBPFP 2.4049 equivalent bits/point +MSE 270.158435 +---------------------- -------------------------------------------------------- +Time: 3.146s Load: 0.004s, Pack+Encode: 1.814s, Decode+Unpack: 1.328s +---------------------- -------------------------------------------------------- +💾 Converting with 270.1584 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,684B, BPFP=0.5446 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,236B, BPFP=2.6859 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,264B, BPFP=0.8653 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,616B, BPFP=2.5601 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,592B, BPFP=1.1347 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,380B, BPFP=2.5122 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,876B, BPFP=0.9894 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,508B, BPFP=2.5381 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,748B, BPFP=2.1810 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,312B, BPFP=2.4984 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,260B, BPFP=0.2684 +⌛️ [2/4] FRONTEND: Frontend time: 2.072s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.607s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12835005 33.04247096 + layer.0.v_cache 0.00001616 0.00851119 + layer.1.k_cache 0.05672838 10.68425483 + layer.1.v_cache 0.00000535 0.00338153 + layer.2.k_cache 0.00240846 1.12376830 + layer.2.v_cache 0.00001644 0.00997990 + layer.3.k_cache 0.04570918 5.66014357 + layer.3.v_cache 0.00001733 0.01151174 + layer.4.k_cache 0.00072090 0.34880943 + layer.4.v_cache 0.00004908 0.02264420 + layer.4.output 0.18444509 693.61850649 + ------------------------------------------------------------------------------------- + TOTAL 0.08971394 288.60264830 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 100476 +BPFP 1.1993 bits/point +EBPFP 2.3987 equivalent bits/point +MSE 288.602648 +---------------------- -------------------------------------------------------- +Time: 3.684s Load: 0.005s, Pack+Encode: 2.072s, Decode+Unpack: 1.607s +---------------------- -------------------------------------------------------- +💾 Converting with 288.6026 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,428B, BPFP=0.5343 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,880B, BPFP=3.0546 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,796B, BPFP=0.8354 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,300B, BPFP=2.4868 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,044B, BPFP=0.8900 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,192B, BPFP=2.6831 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,212B, BPFP=0.9269 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,324B, BPFP=2.9322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,580B, BPFP=2.1083 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,976B, BPFP=2.8556 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,736B, BPFP=0.2746 +⌛️ [2/4] FRONTEND: Frontend time: 1.831s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10085608 31.13920142 + layer.0.v_cache 0.00001554 0.00891082 + layer.1.k_cache 0.06589249 7.65882379 + layer.1.v_cache 0.00000499 0.00322387 + layer.2.k_cache 0.00422055 1.02702546 + layer.2.v_cache 0.00001579 0.00936148 + layer.3.k_cache 0.05337993 5.53909775 + layer.3.v_cache 0.00001747 0.01153525 + layer.4.k_cache 0.00075412 0.32965641 + layer.4.v_cache 0.00004712 0.02199619 + layer.4.output 0.19135183 759.82224598 + ------------------------------------------------------------------------------------- + TOTAL 0.09203923 315.55909143 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 96468 +BPFP 1.2488 bits/point +EBPFP 2.4976 equivalent bits/point +MSE 315.559091 +---------------------- -------------------------------------------------------- +Time: 3.146s Load: 0.003s, Pack+Encode: 1.831s, Decode+Unpack: 1.311s +---------------------- -------------------------------------------------------- +💾 Converting with 315.5591 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,608B, BPFP=0.4851 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,764B, BPFP=2.7463 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,248B, BPFP=0.7902 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,816B, BPFP=2.5699 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,936B, BPFP=0.9182 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,872B, BPFP=2.5804 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,396B, BPFP=0.8177 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,604B, BPFP=2.7165 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,308B, BPFP=1.9174 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,008B, BPFP=2.6057 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,520B, BPFP=0.2530 +⌛️ [2/4] FRONTEND: Frontend time: 1.884s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09057366 30.68012928 + layer.0.v_cache 0.00001789 0.00798600 + layer.1.k_cache 0.03614600 7.51018415 + layer.1.v_cache 0.00000481 0.00298978 + layer.2.k_cache 0.00554357 0.95750009 + layer.2.v_cache 0.00001714 0.00946888 + layer.3.k_cache 0.02742965 3.62143526 + layer.3.v_cache 0.00001729 0.01121927 + layer.4.k_cache 0.00093027 0.35101055 + layer.4.v_cache 0.00004906 0.02311864 + layer.4.output 0.16441821 623.35283801 + ------------------------------------------------------------------------------------- + TOTAL 0.07715628 259.21440635 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 107080 +BPFP 1.1717 bits/point +EBPFP 2.3433 equivalent bits/point +MSE 259.214406 +---------------------- -------------------------------------------------------- +Time: 3.205s Load: 0.003s, Pack+Encode: 1.884s, Decode+Unpack: 1.318s +---------------------- -------------------------------------------------------- +💾 Converting with 259.2144 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,584B, BPFP=0.5244 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,000B, BPFP=2.8409 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,324B, BPFP=0.8774 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,888B, BPFP=2.6153 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,872B, BPFP=0.9886 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,148B, BPFP=2.6680 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,512B, BPFP=0.9156 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,976B, BPFP=2.8360 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,528B, BPFP=2.1364 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,040B, BPFP=2.6461 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,896B, BPFP=0.2869 +⌛️ [2/4] FRONTEND: Frontend time: 2.024s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.260s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09325875 31.19321035 + layer.0.v_cache 0.00001406 0.00848169 + layer.1.k_cache 0.05775181 11.47270222 + layer.1.v_cache 0.00000524 0.00321884 + layer.2.k_cache 0.00413042 0.97234136 + layer.2.v_cache 0.00001743 0.01005165 + layer.3.k_cache 0.04387648 4.81195544 + layer.3.v_cache 0.00002497 0.01206842 + layer.4.k_cache 0.00071512 0.35809690 + layer.4.v_cache 0.00005147 0.02472256 + layer.4.output 0.17652170 699.80049861 + ------------------------------------------------------------------------------------- + TOTAL 0.08444104 291.02766704 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 103768 +BPFP 1.2386 bits/point +EBPFP 2.4773 equivalent bits/point +MSE 291.027667 +---------------------- -------------------------------------------------------- +Time: 3.287s Load: 0.003s, Pack+Encode: 2.024s, Decode+Unpack: 1.260s +---------------------- -------------------------------------------------------- +💾 Converting with 291.0277 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,744B, BPFP=0.4872 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,404B, BPFP=2.5575 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,492B, BPFP=0.7976 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,988B, BPFP=2.3061 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,028B, BPFP=0.8928 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,956B, BPFP=2.3004 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,652B, BPFP=0.8260 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,916B, BPFP=2.4709 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,968B, BPFP=1.7699 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,880B, BPFP=2.2869 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,524B, BPFP=0.2162 +⌛️ [2/4] FRONTEND: Frontend time: 2.061s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.338s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212358 33.69768732 + layer.0.v_cache 0.00001682 0.00827068 + layer.1.k_cache 0.05918076 11.09712081 + layer.1.v_cache 0.00000473 0.00301198 + layer.2.k_cache 0.00260097 1.11671933 + layer.2.v_cache 0.00001544 0.00907258 + layer.3.k_cache 0.02838851 6.82286627 + layer.3.v_cache 0.00001657 0.01078109 + layer.4.k_cache 0.00106615 0.34153806 + layer.4.v_cache 0.00004456 0.02331587 + layer.4.output 0.17173813 611.20403815 + ------------------------------------------------------------------------------------- + TOTAL 0.08091912 254.79756771 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 102552 +BPFP 1.0711 bits/point +EBPFP 2.1422 equivalent bits/point +MSE 254.797568 +---------------------- -------------------------------------------------------- +Time: 3.403s Load: 0.004s, Pack+Encode: 2.061s, Decode+Unpack: 1.338s +---------------------- -------------------------------------------------------- +💾 Converting with 254.7976 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,736B, BPFP=0.4858 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,764B, BPFP=2.7990 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,972B, BPFP=0.8828 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,644B, BPFP=2.4226 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,888B, BPFP=0.8679 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,496B, BPFP=2.5739 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,808B, BPFP=0.8537 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,120B, BPFP=2.6847 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,676B, BPFP=1.8956 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,380B, BPFP=2.5533 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,280B, BPFP=0.2608 +⌛️ [2/4] FRONTEND: Frontend time: 1.911s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.546s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10731998 33.99677069 + layer.0.v_cache 0.00001618 0.00878765 + layer.1.k_cache 0.07375650 10.96841223 + layer.1.v_cache 0.00000540 0.00315030 + layer.2.k_cache 0.00246326 0.97602029 + layer.2.v_cache 0.00001746 0.00976292 + layer.3.k_cache 0.01642703 3.74872936 + layer.3.v_cache 0.00001778 0.01066363 + layer.4.k_cache 0.00066607 0.34485093 + layer.4.v_cache 0.00004728 0.02282075 + layer.4.output 0.15452130 612.47701907 + ------------------------------------------------------------------------------------- + TOTAL 0.07543447 255.14288837 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 111764 +BPFP 1.1673 bits/point +EBPFP 2.3346 equivalent bits/point +MSE 255.142888 +---------------------- -------------------------------------------------------- +Time: 3.461s Load: 0.004s, Pack+Encode: 1.911s, Decode+Unpack: 1.546s +---------------------- -------------------------------------------------------- +💾 Converting with 255.1429 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 99, 128) +Output shape: (1, 99, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.output: torch.Size([1, 99, 3584]) -> torch.Size([1, 1, 99, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,876B, BPFP=0.4539 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,052B, BPFP=2.5335 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,596B, BPFP=0.8832 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,528B, BPFP=2.4508 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,412B, BPFP=1.0120 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,608B, BPFP=2.3056 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,728B, BPFP=0.9040 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,060B, BPFP=2.5347 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,636B, BPFP=1.8365 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,884B, BPFP=2.3491 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,720B, BPFP=0.2642 +⌛️ [2/4] FRONTEND: Frontend time: 2.241s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.356s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09061131 31.41358408 + layer.0.v_cache 0.00001461 0.00821088 + layer.1.k_cache 0.05062923 8.66953779 + layer.1.v_cache 0.00000709 0.00334480 + layer.2.k_cache 0.00392854 1.11361679 + layer.2.v_cache 0.00001793 0.01057447 + layer.3.k_cache 0.05082532 8.23627079 + layer.3.v_cache 0.00001820 0.01154140 + layer.4.k_cache 0.00080398 0.35168538 + layer.4.v_cache 0.00005053 0.02338777 + layer.4.output 0.02403493 559.23250361 + ------------------------------------------------------------------------------------- + TOTAL 0.02147948 233.20407526 + (elements=861,696) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 861696 +Total Bytes 121100 +BPFP 1.1243 bits/point +EBPFP 2.2486 equivalent bits/point +MSE 233.204075 +---------------------- -------------------------------------------------------- +Time: 3.601s Load: 0.005s, Pack+Encode: 2.241s, Decode+Unpack: 1.356s +---------------------- -------------------------------------------------------- +💾 Converting with 233.2041 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,372B, BPFP=0.5220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,956B, BPFP=3.2914 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,960B, BPFP=0.8715 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,484B, BPFP=2.9674 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,156B, BPFP=0.9146 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,624B, BPFP=2.9982 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,792B, BPFP=1.2746 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,452B, BPFP=3.1805 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,052B, BPFP=1.9921 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,460B, BPFP=2.9621 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,208B, BPFP=0.3209 +⌛️ [2/4] FRONTEND: Frontend time: 1.971s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.361s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11136492 30.86405150 + layer.0.v_cache 0.00001955 0.00904586 + layer.1.k_cache 0.06280328 9.02173367 + layer.1.v_cache 0.00000544 0.00340811 + layer.2.k_cache 0.00235174 0.94415111 + layer.2.v_cache 0.00001859 0.01068814 + layer.3.k_cache 0.03326307 8.84257572 + layer.3.v_cache 0.00001926 0.01231062 + layer.4.k_cache 0.00064473 0.35357983 + layer.4.v_cache 0.00005254 0.02317160 + layer.4.output 0.19494371 757.32903672 + ------------------------------------------------------------------------------------- + TOTAL 0.09265583 314.78752784 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 105516 +BPFP 1.3659 bits/point +EBPFP 2.7319 equivalent bits/point +MSE 314.787528 +---------------------- -------------------------------------------------------- +Time: 3.335s Load: 0.004s, Pack+Encode: 1.971s, Decode+Unpack: 1.361s +---------------------- -------------------------------------------------------- +💾 Converting with 314.7875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,496B, BPFP=0.5270 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,808B, BPFP=3.1267 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,064B, BPFP=0.8581 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,304B, BPFP=3.0203 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,684B, BPFP=1.2002 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,076B, BPFP=2.9721 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,028B, BPFP=1.0617 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,184B, BPFP=2.9949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,852B, BPFP=1.8691 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,924B, BPFP=2.9400 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,980B, BPFP=0.2709 +⌛️ [2/4] FRONTEND: Frontend time: 1.892s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.247s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08829598 33.06401763 + layer.0.v_cache 0.00001455 0.00914468 + layer.1.k_cache 0.05950350 9.65379375 + layer.1.v_cache 0.00000529 0.00342206 + layer.2.k_cache 0.00411868 1.18773981 + layer.2.v_cache 0.00001680 0.01089185 + layer.3.k_cache 0.04877868 4.14858761 + layer.3.v_cache 0.00001896 0.01260106 + layer.4.k_cache 0.00063741 0.31272125 + layer.4.v_cache 0.00004764 0.02658882 + layer.4.output 0.18365649 713.82227317 + ------------------------------------------------------------------------------------- + TOTAL 0.08747253 296.77561298 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 106400 +BPFP 1.3215 bits/point +EBPFP 2.6431 equivalent bits/point +MSE 296.775613 +---------------------- -------------------------------------------------------- +Time: 3.144s Load: 0.004s, Pack+Encode: 1.892s, Decode+Unpack: 1.247s +---------------------- -------------------------------------------------------- +💾 Converting with 296.7756 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,700B, BPFP=0.5479 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,064B, BPFP=2.8539 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,428B, BPFP=0.8985 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,540B, BPFP=2.5446 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,028B, BPFP=1.0203 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,120B, BPFP=2.6623 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,440B, BPFP=0.9010 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,936B, BPFP=2.8279 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,788B, BPFP=1.9862 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,288B, BPFP=2.6964 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,060B, BPFP=0.2626 +⌛️ [2/4] FRONTEND: Frontend time: 2.120s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11195674 33.55206220 + layer.0.v_cache 0.00001346 0.00822411 + layer.1.k_cache 0.03755230 10.41934164 + layer.1.v_cache 0.00000535 0.00332037 + layer.2.k_cache 0.00239600 1.05177614 + layer.2.v_cache 0.00001753 0.01138842 + layer.3.k_cache 0.02867049 5.30721689 + layer.3.v_cache 0.00001751 0.01229739 + layer.4.k_cache 0.00071811 0.34982813 + layer.4.v_cache 0.00005486 0.02602543 + layer.4.output 0.19495771 685.80096243 + ------------------------------------------------------------------------------------- + TOTAL 0.09094743 285.37342457 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 102392 +BPFP 1.2222 bits/point +EBPFP 2.4444 equivalent bits/point +MSE 285.373425 +---------------------- -------------------------------------------------------- +Time: 3.489s Load: 0.006s, Pack+Encode: 2.120s, Decode+Unpack: 1.363s +---------------------- -------------------------------------------------------- +💾 Converting with 285.3734 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,748B, BPFP=0.4771 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,664B, BPFP=2.7194 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,080B, BPFP=0.8819 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,824B, BPFP=2.5736 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,528B, BPFP=0.9597 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,892B, BPFP=2.4118 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,468B, BPFP=0.9493 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,972B, BPFP=2.5993 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,948B, BPFP=1.9007 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,428B, BPFP=2.5049 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,496B, BPFP=0.2355 +⌛️ [2/4] FRONTEND: Frontend time: 2.010s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.514s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08657128 30.63572591 + layer.0.v_cache 0.00001574 0.00837675 + layer.1.k_cache 0.05618976 10.80531684 + layer.1.v_cache 0.00000506 0.00317060 + layer.2.k_cache 0.00392605 1.19404373 + layer.2.v_cache 0.00001570 0.00970748 + layer.3.k_cache 0.02849737 6.68964708 + layer.3.v_cache 0.00001769 0.01185885 + layer.4.k_cache 0.00099738 0.36987330 + layer.4.v_cache 0.00004751 0.02482578 + layer.4.output 0.15107292 592.88566468 + ------------------------------------------------------------------------------------- + TOTAL 0.07257612 247.05601171 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 113048 +BPFP 1.1545 bits/point +EBPFP 2.3090 equivalent bits/point +MSE 247.056012 +---------------------- -------------------------------------------------------- +Time: 3.528s Load: 0.004s, Pack+Encode: 2.010s, Decode+Unpack: 1.514s +---------------------- -------------------------------------------------------- +💾 Converting with 247.0560 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,612B, BPFP=0.5102 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,336B, BPFP=2.8000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,296B, BPFP=0.8391 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,392B, BPFP=2.6156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,184B, BPFP=1.0125 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,364B, BPFP=2.6102 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,220B, BPFP=0.8242 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,996B, BPFP=2.7336 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,520B, BPFP=1.8594 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,740B, BPFP=2.6836 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,176B, BPFP=0.2560 +⌛️ [2/4] FRONTEND: Frontend time: 1.911s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.334s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09254595 33.02170410 + layer.0.v_cache 0.00001776 0.00825607 + layer.1.k_cache 0.05775445 9.98044586 + layer.1.v_cache 0.00000504 0.00313427 + layer.2.k_cache 0.00891839 0.98642845 + layer.2.v_cache 0.00001732 0.00975438 + layer.3.k_cache 0.02836140 3.73454781 + layer.3.v_cache 0.00001672 0.01140916 + layer.4.k_cache 0.00083566 0.31693168 + layer.4.v_cache 0.00004876 0.02236634 + layer.4.output 0.16993865 667.99787946 + ------------------------------------------------------------------------------------- + TOTAL 0.08106423 277.88706673 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 103836 +BPFP 1.1930 bits/point +EBPFP 2.3859 equivalent bits/point +MSE 277.887067 +---------------------- -------------------------------------------------------- +Time: 3.249s Load: 0.004s, Pack+Encode: 1.911s, Decode+Unpack: 1.334s +---------------------- -------------------------------------------------------- +💾 Converting with 277.8871 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,752B, BPFP=0.4433 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,436B, BPFP=2.4865 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,984B, BPFP=0.8028 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,896B, BPFP=2.3995 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,468B, BPFP=1.0419 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,940B, BPFP=2.2455 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,148B, BPFP=0.8293 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,100B, BPFP=2.4323 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,488B, BPFP=1.6894 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,056B, BPFP=2.2642 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,252B, BPFP=0.2359 +⌛️ [2/4] FRONTEND: Frontend time: 1.875s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.335s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13093194 30.93016571 + layer.0.v_cache 0.00001635 0.00857937 + layer.1.k_cache 0.05235285 10.16039913 + layer.1.v_cache 0.00000496 0.00320263 + layer.2.k_cache 0.00512381 1.41223616 + layer.2.v_cache 0.00001647 0.00918605 + layer.3.k_cache 0.01138228 7.17281758 + layer.3.v_cache 0.00001850 0.01086376 + layer.4.k_cache 0.00081880 0.35007426 + layer.4.v_cache 0.00004796 0.02488228 + layer.4.output 0.02447439 572.14814985 + ------------------------------------------------------------------------------------- + TOTAL 0.02188439 238.53643858 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 113520 +BPFP 1.0757 bits/point +EBPFP 2.1513 equivalent bits/point +MSE 238.536439 +---------------------- -------------------------------------------------------- +Time: 3.215s Load: 0.004s, Pack+Encode: 1.875s, Decode+Unpack: 1.335s +---------------------- -------------------------------------------------------- +💾 Converting with 238.5364 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,668B, BPFP=0.5277 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,760B, BPFP=2.9193 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,320B, BPFP=0.8544 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,176B, BPFP=2.6060 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,672B, BPFP=0.9241 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,328B, BPFP=2.6361 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,252B, BPFP=0.8410 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,012B, BPFP=2.7714 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,324B, BPFP=2.0419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,560B, BPFP=2.6820 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,148B, BPFP=0.2867 +⌛️ [2/4] FRONTEND: Frontend time: 1.909s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.473s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510638 38.00080968 + layer.0.v_cache 0.00001399 0.00845632 + layer.1.k_cache 0.05467150 11.44925583 + layer.1.v_cache 0.00000511 0.00306141 + layer.2.k_cache 0.00545186 1.00601216 + layer.2.v_cache 0.00001652 0.01003771 + layer.3.k_cache 0.02986256 4.07449225 + layer.3.v_cache 0.00001747 0.01120444 + layer.4.k_cache 0.00068266 0.34530236 + layer.4.v_cache 0.00005277 0.02453679 + layer.4.output 0.17642128 677.41563065 + ------------------------------------------------------------------------------------- + TOTAL 0.08357822 282.16721079 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 105220 +BPFP 1.2242 bits/point +EBPFP 2.4483 equivalent bits/point +MSE 282.167211 +---------------------- -------------------------------------------------------- +Time: 3.385s Load: 0.003s, Pack+Encode: 1.909s, Decode+Unpack: 1.473s +---------------------- -------------------------------------------------------- +💾 Converting with 282.1672 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,684B, BPFP=0.5446 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,552B, BPFP=2.9529 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,396B, BPFP=0.8920 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,688B, BPFP=2.7776 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,916B, BPFP=0.9976 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,000B, BPFP=2.6380 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,552B, BPFP=0.9237 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,788B, BPFP=2.7979 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,404B, BPFP=2.1112 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,584B, BPFP=2.7565 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,476B, BPFP=0.2747 +⌛️ [2/4] FRONTEND: Frontend time: 1.846s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.420s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08730992 34.57814719 + layer.0.v_cache 0.00001543 0.00911621 + layer.1.k_cache 0.05635426 11.15914640 + layer.1.v_cache 0.00000520 0.00345631 + layer.2.k_cache 0.00390536 1.04148845 + layer.2.v_cache 0.00001661 0.01022998 + layer.3.k_cache 0.06483487 3.40495875 + layer.3.v_cache 0.00002010 0.01188762 + layer.4.k_cache 0.00072652 0.35685998 + layer.4.v_cache 0.00004842 0.02433657 + layer.4.output 0.18845482 693.76692950 + ------------------------------------------------------------------------------------- + TOTAL 0.09014238 288.64518435 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 105040 +BPFP 1.2538 bits/point +EBPFP 2.5076 equivalent bits/point +MSE 288.645184 +---------------------- -------------------------------------------------------- +Time: 3.269s Load: 0.003s, Pack+Encode: 1.846s, Decode+Unpack: 1.420s +---------------------- -------------------------------------------------------- +💾 Converting with 288.6452 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,228B, BPFP=0.5119 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,308B, BPFP=3.0579 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,904B, BPFP=0.8971 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,792B, BPFP=2.7096 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,892B, BPFP=0.8943 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,180B, BPFP=2.5689 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,496B, BPFP=0.8033 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,084B, BPFP=2.7767 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,688B, BPFP=1.5368 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,264B, BPFP=2.8180 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,820B, BPFP=0.3223 +⌛️ [2/4] FRONTEND: Frontend time: 2.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.562s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10784986 31.39754710 + layer.0.v_cache 0.00001387 0.00892754 + layer.1.k_cache 0.01409798 10.67402559 + layer.1.v_cache 0.00000489 0.00306854 + layer.2.k_cache 0.00254294 0.97016032 + layer.2.v_cache 0.00001653 0.00987391 + layer.3.k_cache 0.05709259 3.51735553 + layer.3.v_cache 0.00001736 0.01125849 + layer.4.k_cache 0.00062715 0.26294237 + layer.4.v_cache 0.00004758 0.02295882 + layer.4.output 0.22376562 782.66819853 + ------------------------------------------------------------------------------------- + TOTAL 0.10286295 325.03267694 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 90656 +BPFP 1.2253 bits/point +EBPFP 2.4507 equivalent bits/point +MSE 325.032677 +---------------------- -------------------------------------------------------- +Time: 3.795s Load: 0.003s, Pack+Encode: 2.230s, Decode+Unpack: 1.562s +---------------------- -------------------------------------------------------- +💾 Converting with 325.0327 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,620B, BPFP=0.5182 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,820B, BPFP=2.7334 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,168B, BPFP=0.8244 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,108B, BPFP=2.3948 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,544B, BPFP=1.0965 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,576B, BPFP=2.4873 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,524B, BPFP=0.8948 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,456B, BPFP=2.6614 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,312B, BPFP=2.0396 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,080B, BPFP=2.5870 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,532B, BPFP=0.2976 +⌛️ [2/4] FRONTEND: Frontend time: 2.147s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.403s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14563359 36.92404445 + layer.0.v_cache 0.00001368 0.00793751 + layer.1.k_cache 0.05947259 8.92250853 + layer.1.v_cache 0.00000496 0.00318309 + layer.2.k_cache 0.00553216 1.15583029 + layer.2.v_cache 0.00001557 0.00989793 + layer.3.k_cache 0.04377548 3.26056565 + layer.3.v_cache 0.00001730 0.01114666 + layer.4.k_cache 0.00079035 0.34121907 + layer.4.v_cache 0.00005153 0.02361278 + layer.4.output 0.17347779 681.06803797 + ------------------------------------------------------------------------------------- + TOTAL 0.08645010 283.41977716 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 102740 +BPFP 1.1953 bits/point +EBPFP 2.3906 equivalent bits/point +MSE 283.419777 +---------------------- -------------------------------------------------------- +Time: 3.555s Load: 0.005s, Pack+Encode: 2.147s, Decode+Unpack: 1.403s +---------------------- -------------------------------------------------------- +💾 Converting with 283.4198 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,096B, BPFP=0.4962 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,892B, BPFP=3.0521 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,400B, BPFP=0.8049 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,448B, BPFP=2.4735 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,928B, BPFP=0.9299 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,192B, BPFP=2.6496 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,336B, BPFP=0.7898 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,892B, BPFP=2.8153 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,504B, BPFP=1.5398 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,000B, BPFP=2.8409 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,176B, BPFP=0.3103 +⌛️ [2/4] FRONTEND: Frontend time: 2.011s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.302s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12373332 33.32285008 + layer.0.v_cache 0.00001436 0.00863157 + layer.1.k_cache 0.01270404 7.91928193 + layer.1.v_cache 0.00000490 0.00310143 + layer.2.k_cache 0.00252672 1.10932206 + layer.2.v_cache 0.00001629 0.01070626 + layer.3.k_cache 0.05185594 3.38183339 + layer.3.v_cache 0.00001849 0.01208608 + layer.4.k_cache 0.00068458 0.23832945 + layer.4.v_cache 0.00004975 0.02338276 + layer.4.output 0.22283003 808.01778950 + ------------------------------------------------------------------------------------- + TOTAL 0.10302462 335.42082656 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 86864 +BPFP 1.2097 bits/point +EBPFP 2.4193 equivalent bits/point +MSE 335.420827 +---------------------- -------------------------------------------------------- +Time: 3.319s Load: 0.005s, Pack+Encode: 2.011s, Decode+Unpack: 1.302s +---------------------- -------------------------------------------------------- +💾 Converting with 335.4208 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,604B, BPFP=0.4844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,560B, BPFP=2.8943 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,328B, BPFP=0.8051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,640B, BPFP=2.7232 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,988B, BPFP=0.9278 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,480B, BPFP=2.6935 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,808B, BPFP=0.8943 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,140B, BPFP=2.8162 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,948B, BPFP=2.0365 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,712B, BPFP=2.7366 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,992B, BPFP=0.2921 +⌛️ [2/4] FRONTEND: Frontend time: 2.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.402s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12002775 31.60391090 + layer.0.v_cache 0.00001466 0.00880169 + layer.1.k_cache 0.05271155 7.79686483 + layer.1.v_cache 0.00000524 0.00317174 + layer.2.k_cache 0.00808975 0.99812916 + layer.2.v_cache 0.00001826 0.01045665 + layer.3.k_cache 0.07501186 5.93154253 + layer.3.v_cache 0.00001895 0.01170407 + layer.4.k_cache 0.00071216 0.34142372 + layer.4.v_cache 0.00005170 0.02200279 + layer.4.output 0.17053346 609.28252551 + ------------------------------------------------------------------------------------- + TOTAL 0.08531742 253.62974627 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 113200 +BPFP 1.2386 bits/point +EBPFP 2.4772 equivalent bits/point +MSE 253.629746 +---------------------- -------------------------------------------------------- +Time: 3.564s Load: 0.003s, Pack+Encode: 2.158s, Decode+Unpack: 1.402s +---------------------- -------------------------------------------------------- +💾 Converting with 253.6297 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 75, 128) +Output shape: (1, 75, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.output: torch.Size([1, 75, 3584]) -> torch.Size([1, 1, 75, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,540B, BPFP=0.5292 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,332B, BPFP=2.9858 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,476B, BPFP=0.9325 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,920B, BPFP=2.6917 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,528B, BPFP=1.1517 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,848B, BPFP=2.6767 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,616B, BPFP=0.9617 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,260B, BPFP=2.7625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,460B, BPFP=1.9708 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,476B, BPFP=2.8075 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,020B, BPFP=0.2982 +⌛️ [2/4] FRONTEND: Frontend time: 2.040s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08149677 32.01474284 + layer.0.v_cache 0.00001391 0.00914252 + layer.1.k_cache 0.05841089 9.55062419 + layer.1.v_cache 0.00000541 0.00322156 + layer.2.k_cache 0.00252637 1.16688741 + layer.2.v_cache 0.00001693 0.00978483 + layer.3.k_cache 0.02876982 6.45200033 + layer.3.v_cache 0.00001932 0.01188649 + layer.4.k_cache 0.00075246 0.34491241 + layer.4.v_cache 0.00005041 0.02267667 + layer.4.output 0.19266937 699.77017857 + ------------------------------------------------------------------------------------- + TOTAL 0.08945576 291.05747819 + (elements=652,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 652800 +Total Bytes 103476 +BPFP 1.2681 bits/point +EBPFP 2.5362 equivalent bits/point +MSE 291.057478 +---------------------- -------------------------------------------------------- +Time: 3.404s Load: 0.003s, Pack+Encode: 2.040s, Decode+Unpack: 1.362s +---------------------- -------------------------------------------------------- +💾 Converting with 291.0575 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,256B, BPFP=0.5184 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,440B, BPFP=3.0882 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,744B, BPFP=0.8603 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,532B, BPFP=2.4200 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,972B, BPFP=1.1425 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,816B, BPFP=2.7151 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,500B, BPFP=0.8042 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,556B, BPFP=2.8851 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,292B, BPFP=1.6756 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,024B, BPFP=2.7629 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,496B, BPFP=0.2789 +⌛️ [2/4] FRONTEND: Frontend time: 2.088s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.395s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10404026 30.99991742 + layer.0.v_cache 0.00001395 0.00915234 + layer.1.k_cache 0.01458383 7.52338365 + layer.1.v_cache 0.00000499 0.00297145 + layer.2.k_cache 0.00236740 1.32073548 + layer.2.v_cache 0.00001641 0.00998366 + layer.3.k_cache 0.03645969 2.84720926 + layer.3.v_cache 0.00002420 0.01163448 + layer.4.k_cache 0.00069996 0.25287084 + layer.4.v_cache 0.00005119 0.02391675 + layer.4.output 0.20980707 757.01404937 + ------------------------------------------------------------------------------------- + TOTAL 0.09570067 314.24118358 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 90628 +BPFP 1.2250 bits/point +EBPFP 2.4499 equivalent bits/point +MSE 314.241184 +---------------------- -------------------------------------------------------- +Time: 3.487s Load: 0.004s, Pack+Encode: 2.088s, Decode+Unpack: 1.395s +---------------------- -------------------------------------------------------- +💾 Converting with 314.2412 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,388B, BPFP=0.5330 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,544B, BPFP=3.0232 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,056B, BPFP=0.9054 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,904B, BPFP=2.8804 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,388B, BPFP=0.9795 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,604B, BPFP=2.8134 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,952B, BPFP=0.8821 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,080B, BPFP=2.9196 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,152B, BPFP=2.0429 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,196B, BPFP=2.9455 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,092B, BPFP=0.3218 +⌛️ [2/4] FRONTEND: Frontend time: 1.956s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.394s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11725240 32.86812919 + layer.0.v_cache 0.00001567 0.00849320 + layer.1.k_cache 0.03693308 7.63443952 + layer.1.v_cache 0.00000533 0.00335746 + layer.2.k_cache 0.00240566 0.93848201 + layer.2.v_cache 0.00001850 0.01049158 + layer.3.k_cache 0.06751672 5.57475673 + layer.3.v_cache 0.00001909 0.01125386 + layer.4.k_cache 0.00067984 0.31913376 + layer.4.v_cache 0.00004901 0.02426555 + layer.4.output 0.20094273 762.88692602 + ------------------------------------------------------------------------------------- + TOTAL 0.09597026 316.91772265 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 99356 +BPFP 1.3046 bits/point +EBPFP 2.6091 equivalent bits/point +MSE 316.917723 +---------------------- -------------------------------------------------------- +Time: 3.354s Load: 0.004s, Pack+Encode: 1.956s, Decode+Unpack: 1.394s +---------------------- -------------------------------------------------------- +💾 Converting with 316.9177 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,404B, BPFP=0.5366 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,332B, BPFP=2.9759 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,920B, BPFP=0.8750 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,176B, BPFP=2.4946 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,928B, BPFP=0.8768 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,716B, BPFP=2.6152 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,184B, BPFP=0.9339 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,576B, BPFP=2.8071 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,536B, BPFP=1.6821 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,440B, BPFP=2.7768 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,668B, BPFP=0.2764 +⌛️ [2/4] FRONTEND: Frontend time: 1.807s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.432s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07685234 32.89944894 + layer.0.v_cache 0.00001390 0.00901491 + layer.1.k_cache 0.01187498 8.05845250 + layer.1.v_cache 0.00000934 0.00326410 + layer.2.k_cache 0.00239603 1.02582049 + layer.2.v_cache 0.00001621 0.00952513 + layer.3.k_cache 0.07241479 4.68666469 + layer.3.v_cache 0.00001716 0.01164034 + layer.4.k_cache 0.00072067 0.26562740 + layer.4.v_cache 0.00004958 0.02190826 + layer.4.output 0.21417375 760.39834184 + ------------------------------------------------------------------------------------- + TOTAL 0.09785772 315.86939762 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 91880 +BPFP 1.2064 bits/point +EBPFP 2.4128 equivalent bits/point +MSE 315.869398 +---------------------- -------------------------------------------------------- +Time: 3.243s Load: 0.004s, Pack+Encode: 1.807s, Decode+Unpack: 1.432s +---------------------- -------------------------------------------------------- +💾 Converting with 315.8694 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,548B, BPFP=0.5238 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,324B, BPFP=2.9449 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,860B, BPFP=0.9992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,560B, BPFP=2.7878 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,692B, BPFP=0.9646 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,464B, BPFP=2.7681 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,700B, BPFP=0.9663 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,732B, BPFP=2.8232 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,144B, BPFP=2.0855 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,436B, BPFP=2.7623 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,148B, BPFP=0.2980 +⌛️ [2/4] FRONTEND: Frontend time: 2.016s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09264422 31.07852976 + layer.0.v_cache 0.00001680 0.00912795 + layer.1.k_cache 0.05606840 11.20397467 + layer.1.v_cache 0.00000542 0.00336554 + layer.2.k_cache 0.00241068 0.99715082 + layer.2.v_cache 0.00001702 0.01011791 + layer.3.k_cache 0.15270803 4.33924946 + layer.3.v_cache 0.00001849 0.01266499 + layer.4.k_cache 0.00070970 0.33861743 + layer.4.v_cache 0.00005539 0.02453698 + layer.4.output 0.18838861 697.85573308 + ------------------------------------------------------------------------------------- + TOTAL 0.09549261 290.17690983 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 105608 +BPFP 1.2772 bits/point +EBPFP 2.5544 equivalent bits/point +MSE 290.176910 +---------------------- -------------------------------------------------------- +Time: 3.338s Load: 0.004s, Pack+Encode: 2.016s, Decode+Unpack: 1.317s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1769 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,684B, BPFP=0.5446 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,236B, BPFP=2.8888 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,280B, BPFP=0.8685 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,768B, BPFP=2.5909 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,708B, BPFP=1.1583 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,616B, BPFP=2.5601 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,240B, BPFP=0.8604 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,640B, BPFP=2.7679 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,200B, BPFP=1.8669 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,280B, BPFP=2.6948 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,800B, BPFP=0.2551 +⌛️ [2/4] FRONTEND: Frontend time: 2.302s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.493s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10144231 30.70122261 + layer.0.v_cache 0.00001330 0.00877676 + layer.1.k_cache 0.09893422 11.07872742 + layer.1.v_cache 0.00000513 0.00305061 + layer.2.k_cache 0.00368564 1.21209836 + layer.2.v_cache 0.00001777 0.01042177 + layer.3.k_cache 0.06341467 4.78120125 + layer.3.v_cache 0.00001739 0.01183676 + layer.4.k_cache 0.00073693 0.36223226 + layer.4.v_cache 0.00004994 0.02318440 + layer.4.output 0.18261438 695.03490260 + ------------------------------------------------------------------------------------- + TOTAL 0.09097753 289.02571002 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 101452 +BPFP 1.2110 bits/point +EBPFP 2.4220 equivalent bits/point +MSE 289.025710 +---------------------- -------------------------------------------------------- +Time: 3.801s Load: 0.005s, Pack+Encode: 2.302s, Decode+Unpack: 1.493s +---------------------- -------------------------------------------------------- +💾 Converting with 289.0257 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,740B, BPFP=0.4921 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,912B, BPFP=2.6782 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,480B, BPFP=0.8046 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,000B, BPFP=2.5144 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,600B, BPFP=0.8261 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,736B, BPFP=2.2874 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,028B, BPFP=0.9030 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,100B, BPFP=2.5323 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,020B, BPFP=1.7996 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,676B, BPFP=2.4562 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,992B, BPFP=0.2564 +⌛️ [2/4] FRONTEND: Frontend time: 2.106s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.506s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08947531 30.99834153 + layer.0.v_cache 0.00001432 0.00829354 + layer.1.k_cache 0.05320675 10.08682847 + layer.1.v_cache 0.00000497 0.00330592 + layer.2.k_cache 0.00562141 0.96057436 + layer.2.v_cache 0.00001574 0.00902076 + layer.3.k_cache 0.02742601 7.46998754 + layer.3.v_cache 0.00001759 0.01092720 + layer.4.k_cache 0.00114861 0.35330325 + layer.4.v_cache 0.00004734 0.02208163 + layer.4.output 0.15628038 620.06460386 + ------------------------------------------------------------------------------------- + TOTAL 0.07476122 258.25734654 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 106284 +BPFP 1.1228 bits/point +EBPFP 2.2457 equivalent bits/point +MSE 258.257347 +---------------------- -------------------------------------------------------- +Time: 3.616s Load: 0.004s, Pack+Encode: 2.106s, Decode+Unpack: 1.506s +---------------------- -------------------------------------------------------- +💾 Converting with 258.2573 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,760B, BPFP=0.4688 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,080B, BPFP=2.5611 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,964B, BPFP=0.8431 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,676B, BPFP=2.3227 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,332B, BPFP=0.9056 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,428B, BPFP=2.2806 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,932B, BPFP=0.8376 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,644B, BPFP=2.4871 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,996B, BPFP=1.6977 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,652B, BPFP=2.3186 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,328B, BPFP=0.2263 +⌛️ [2/4] FRONTEND: Frontend time: 2.140s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12311893 30.39363961 + layer.0.v_cache 0.00001357 0.00807771 + layer.1.k_cache 0.03743907 10.61816274 + layer.1.v_cache 0.00000515 0.00297372 + layer.2.k_cache 0.00511221 1.36589681 + layer.2.v_cache 0.00001545 0.00882816 + layer.3.k_cache 0.04225797 7.91463968 + layer.3.v_cache 0.00001799 0.01093566 + layer.4.k_cache 0.00119426 0.36478901 + layer.4.v_cache 0.00004621 0.02080983 + layer.4.output 0.14783020 571.85782220 + ------------------------------------------------------------------------------------- + TOTAL 0.07317837 238.45373579 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 107792 +BPFP 1.0769 bits/point +EBPFP 2.1538 equivalent bits/point +MSE 238.453736 +---------------------- -------------------------------------------------------- +Time: 3.349s Load: 0.004s, Pack+Encode: 2.140s, Decode+Unpack: 1.205s +---------------------- -------------------------------------------------------- +💾 Converting with 238.4537 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,056B, BPFP=0.4681 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,248B, BPFP=2.4890 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,236B, BPFP=0.8021 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,424B, BPFP=2.3627 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,456B, BPFP=0.8358 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,312B, BPFP=2.3456 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,988B, BPFP=0.7641 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,852B, BPFP=2.4283 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,752B, BPFP=1.9534 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,304B, BPFP=2.3444 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,648B, BPFP=0.2768 +⌛️ [2/4] FRONTEND: Frontend time: 1.976s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.404s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09293266 33.14517032 + layer.0.v_cache 0.00001503 0.00765002 + layer.1.k_cache 0.07854857 8.19009220 + layer.1.v_cache 0.00000559 0.00288363 + layer.2.k_cache 0.00650115 0.99758829 + layer.2.v_cache 0.00001763 0.00983754 + layer.3.k_cache 0.03715762 3.09349240 + layer.3.v_cache 0.00001788 0.01099580 + layer.4.k_cache 0.00073365 0.34006960 + layer.4.v_cache 0.00004630 0.02123838 + layer.4.output 11.17680568 524.59646359 + ------------------------------------------------------------------------------------- + TOTAL 4.61491858 218.70554490 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 122276 +BPFP 1.1018 bits/point +EBPFP 2.2036 equivalent bits/point +MSE 218.705545 +---------------------- -------------------------------------------------------- +Time: 3.384s Load: 0.004s, Pack+Encode: 1.976s, Decode+Unpack: 1.404s +---------------------- -------------------------------------------------------- +💾 Converting with 218.7055 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 96, 128) +Output shape: (1, 96, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.output: torch.Size([1, 96, 3584]) -> torch.Size([1, 1, 96, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,676B, BPFP=0.4355 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,136B, BPFP=2.6263 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,552B, BPFP=0.7409 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,444B, BPFP=2.5137 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,288B, BPFP=1.0234 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,000B, BPFP=2.4414 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,336B, BPFP=1.0312 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,796B, BPFP=2.5710 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,784B, BPFP=1.9180 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,068B, BPFP=2.4525 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,456B, BPFP=0.2431 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10158994 31.32756042 + layer.0.v_cache 0.00001467 0.00851914 + layer.1.k_cache 0.06416908 9.48211861 + layer.1.v_cache 0.00000542 0.00322310 + layer.2.k_cache 0.00400108 1.46692181 + layer.2.v_cache 0.00001872 0.00991150 + layer.3.k_cache 0.01491791 9.56751823 + layer.3.v_cache 0.00001877 0.01126903 + layer.4.k_cache 0.00070988 0.33089566 + layer.4.v_cache 0.00004832 0.02272363 + layer.4.output 0.14169503 563.12193080 + ------------------------------------------------------------------------------------- + TOTAL 0.06925641 234.94612805 + (elements=835,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 835584 +Total Bytes 119536 +BPFP 1.1445 bits/point +EBPFP 2.2889 equivalent bits/point +MSE 234.946128 +---------------------- -------------------------------------------------------- +Time: 3.230s Load: 0.004s, Pack+Encode: 1.838s, Decode+Unpack: 1.388s +---------------------- -------------------------------------------------------- +💾 Converting with 234.9461 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,424B, BPFP=0.5260 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,128B, BPFP=3.0660 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,928B, BPFP=0.8524 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,916B, BPFP=2.8030 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,632B, BPFP=1.0052 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,724B, BPFP=2.7613 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,564B, BPFP=0.9905 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,488B, BPFP=2.9271 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,688B, BPFP=2.1024 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,352B, BPFP=2.8976 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,224B, BPFP=0.2550 +⌛️ [2/4] FRONTEND: Frontend time: 1.900s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.471s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10321087 32.12683105 + layer.0.v_cache 0.00001557 0.00971346 + layer.1.k_cache 0.06175087 7.88626268 + layer.1.v_cache 0.00000571 0.00336136 + layer.2.k_cache 0.01329466 1.02955702 + layer.2.v_cache 0.00001784 0.01021079 + layer.3.k_cache 0.05039416 5.59070757 + layer.3.v_cache 0.00001745 0.01176364 + layer.4.k_cache 0.00079786 0.33620734 + layer.4.v_cache 0.00005071 0.02341517 + layer.4.output 0.18870974 750.41722470 + ------------------------------------------------------------------------------------- + TOTAL 0.09120729 311.76168253 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 100068 +BPFP 1.2774 bits/point +EBPFP 2.5548 equivalent bits/point +MSE 311.761683 +---------------------- -------------------------------------------------------- +Time: 3.374s Load: 0.003s, Pack+Encode: 1.900s, Decode+Unpack: 1.471s +---------------------- -------------------------------------------------------- +💾 Converting with 311.7617 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,416B, BPFP=0.5243 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,172B, BPFP=2.8585 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,920B, BPFP=0.8507 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,100B, BPFP=2.6259 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,136B, BPFP=0.8976 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,924B, BPFP=2.5877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,796B, BPFP=0.8238 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,448B, BPFP=2.7014 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,772B, BPFP=1.6866 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,680B, BPFP=2.5347 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,548B, BPFP=0.2960 +⌛️ [2/4] FRONTEND: Frontend time: 1.872s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09666174 31.79808214 + layer.0.v_cache 0.00001715 0.00843720 + layer.1.k_cache 0.08476152 9.48204295 + layer.1.v_cache 0.00000506 0.00297529 + layer.2.k_cache 0.00238475 0.90508461 + layer.2.v_cache 0.00001567 0.00941531 + layer.3.k_cache 0.07332428 4.36195204 + layer.3.v_cache 0.00001796 0.01106873 + layer.4.k_cache 0.00072290 0.27184486 + layer.4.v_cache 0.00005092 0.02150694 + layer.4.output 0.18875189 746.84288194 + ------------------------------------------------------------------------------------- + TOTAL 0.09289560 310.28074022 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 92912 +BPFP 1.1861 bits/point +EBPFP 2.3721 equivalent bits/point +MSE 310.280740 +---------------------- -------------------------------------------------------- +Time: 3.184s Load: 0.003s, Pack+Encode: 1.872s, Decode+Unpack: 1.309s +---------------------- -------------------------------------------------------- +💾 Converting with 310.2807 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,236B, BPFP=0.5138 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,608B, BPFP=3.1268 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,936B, BPFP=0.9044 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,480B, BPFP=2.6379 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,840B, BPFP=1.1121 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,784B, BPFP=2.4779 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,616B, BPFP=0.8309 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,076B, BPFP=3.0046 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,488B, BPFP=1.7206 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,368B, BPFP=2.6121 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,884B, BPFP=0.3244 +⌛️ [2/4] FRONTEND: Frontend time: 1.926s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.337s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10601801 33.00840131 + layer.0.v_cache 0.00001817 0.00965364 + layer.1.k_cache 0.01696859 10.54297414 + layer.1.v_cache 0.00000501 0.00322627 + layer.2.k_cache 0.00274148 1.32166212 + layer.2.v_cache 0.00001757 0.00946259 + layer.3.k_cache 0.01706722 3.98141120 + layer.3.v_cache 0.00001843 0.01210530 + layer.4.k_cache 0.00080906 0.29123539 + layer.4.v_cache 0.00005076 0.02227449 + layer.4.output 0.19983917 783.15158876 + ------------------------------------------------------------------------------------- + TOTAL 0.09074050 325.36844281 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 92316 +BPFP 1.2478 bits/point +EBPFP 2.4956 equivalent bits/point +MSE 325.368443 +---------------------- -------------------------------------------------------- +Time: 3.266s Load: 0.003s, Pack+Encode: 1.926s, Decode+Unpack: 1.337s +---------------------- -------------------------------------------------------- +💾 Converting with 325.3684 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 65, 128) +Output shape: (1, 65, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.output: torch.Size([1, 65, 3584]) -> torch.Size([1, 1, 65, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,008B, BPFP=0.4827 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,784B, BPFP=2.8327 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,332B, BPFP=0.8010 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,388B, BPFP=2.7375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,056B, BPFP=0.9750 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,996B, BPFP=2.6433 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,232B, BPFP=0.7769 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,424B, BPFP=2.7462 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,012B, BPFP=1.6856 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,412B, BPFP=2.9837 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,932B, BPFP=0.2724 +⌛️ [2/4] FRONTEND: Frontend time: 2.297s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.469s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09579474 35.02058293 + layer.0.v_cache 0.00001312 0.00825999 + layer.1.k_cache 0.01648773 7.50854539 + layer.1.v_cache 0.00000712 0.00309339 + layer.2.k_cache 0.00431219 1.21165419 + layer.2.v_cache 0.00001803 0.01016982 + layer.3.k_cache 0.03460605 3.13758874 + layer.3.v_cache 0.00001721 0.01231984 + layer.4.k_cache 0.00068162 0.26454315 + layer.4.v_cache 0.00004727 0.02354821 + layer.4.output 0.20898957 830.74223901 + ------------------------------------------------------------------------------------- + TOTAL 0.09499483 344.84682228 + (elements=565,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 565760 +Total Bytes 85576 +BPFP 1.2101 bits/point +EBPFP 2.4201 equivalent bits/point +MSE 344.846822 +---------------------- -------------------------------------------------------- +Time: 3.769s Load: 0.003s, Pack+Encode: 2.297s, Decode+Unpack: 1.469s +---------------------- -------------------------------------------------------- +💾 Converting with 344.8468 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 41, 128) +Output shape: (1, 41, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.output: torch.Size([1, 41, 3584]) -> torch.Size([1, 1, 41, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,652B, BPFP=0.6296 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,124B, BPFP=3.0960 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,660B, BPFP=1.0137 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,996B, BPFP=3.0473 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,924B, BPFP=1.1143 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,092B, BPFP=3.0838 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,576B, BPFP=0.9817 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,096B, BPFP=3.0854 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,972B, BPFP=3.0381 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,780B, BPFP=2.9649 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 6,644B, BPFP=0.3617 +⌛️ [2/4] FRONTEND: Frontend time: 1.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.152s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10098721 39.77108244 + layer.0.v_cache 0.00001362 0.01149820 + layer.1.k_cache 0.01708067 9.04381580 + layer.1.v_cache 0.00000533 0.00407550 + layer.2.k_cache 0.00250472 0.98659348 + layer.2.v_cache 0.00001755 0.01316094 + layer.3.k_cache 0.08199174 3.69014907 + layer.3.v_cache 0.00001815 0.01450254 + layer.4.k_cache 0.00059437 0.33232489 + layer.4.v_cache 0.00005232 0.03129839 + layer.4.output 0.35770382 1317.73442944 + ------------------------------------------------------------------------------------- + TOTAL 0.15924661 545.76702984 + (elements=356,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 356864 +Total Bytes 64516 +BPFP 1.4463 bits/point +EBPFP 2.8926 equivalent bits/point +MSE 545.767030 +---------------------- -------------------------------------------------------- +Time: 3.001s Load: 0.002s, Pack+Encode: 1.847s, Decode+Unpack: 1.152s +---------------------- -------------------------------------------------------- +💾 Converting with 545.7670 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,872B, BPFP=0.5625 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,016B, BPFP=2.4087 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,904B, BPFP=0.8726 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,852B, BPFP=2.3594 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,136B, BPFP=0.9423 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,848B, BPFP=2.3582 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,812B, BPFP=0.8450 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,876B, BPFP=2.3666 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,824B, BPFP=2.3510 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,672B, BPFP=2.3053 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,180B, BPFP=0.3511 +⌛️ [2/4] FRONTEND: Frontend time: 2.013s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09210638 33.68676993 + layer.0.v_cache 0.00001369 0.00948135 + layer.1.k_cache 0.01682195 7.62812042 + layer.1.v_cache 0.00000520 0.00365472 + layer.2.k_cache 0.00487203 0.98350415 + layer.2.v_cache 0.00001755 0.01242678 + layer.3.k_cache 0.04965206 3.83354422 + layer.3.v_cache 0.00001944 0.01361653 + layer.4.k_cache 0.00060442 0.35953496 + layer.4.v_cache 0.00005207 0.02806372 + layer.4.output 0.26182200 1010.69239354 + ------------------------------------------------------------------------------------- + TOTAL 0.11746581 418.90620421 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 65992 +BPFP 1.1664 bits/point +EBPFP 2.3329 equivalent bits/point +MSE 418.906204 +---------------------- -------------------------------------------------------- +Time: 3.236s Load: 0.005s, Pack+Encode: 2.013s, Decode+Unpack: 1.218s +---------------------- -------------------------------------------------------- +💾 Converting with 418.9062 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 47, 128) +Output shape: (1, 47, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.output: torch.Size([1, 47, 3584]) -> torch.Size([1, 1, 47, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,820B, BPFP=0.6051 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,912B, BPFP=2.6303 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,956B, BPFP=0.9827 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,780B, BPFP=2.5864 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,300B, BPFP=1.0971 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,712B, BPFP=2.5638 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,012B, BPFP=1.0013 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,824B, BPFP=2.6011 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,228B, BPFP=2.7354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,652B, BPFP=2.5439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,156B, BPFP=0.3873 +⌛️ [2/4] FRONTEND: Frontend time: 1.760s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08461995 34.87937375 + layer.0.v_cache 0.00001366 0.01045831 + layer.1.k_cache 0.01838745 8.90241517 + layer.1.v_cache 0.00000609 0.00406937 + layer.2.k_cache 0.00780689 1.09333119 + layer.2.v_cache 0.00001932 0.01261572 + layer.3.k_cache 0.04292001 7.96486453 + layer.3.v_cache 0.00001942 0.01359001 + layer.4.k_cache 0.00059704 0.31637372 + layer.4.v_cache 0.00004850 0.02772776 + layer.4.output 0.29441516 1133.20241261 + ------------------------------------------------------------------------------------- + TOTAL 0.13031438 469.74362987 + (elements=409,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 409088 +Total Bytes 66352 +BPFP 1.2976 bits/point +EBPFP 2.5951 equivalent bits/point +MSE 469.743630 +---------------------- -------------------------------------------------------- +Time: 2.968s Load: 0.002s, Pack+Encode: 1.760s, Decode+Unpack: 1.206s +---------------------- -------------------------------------------------------- +💾 Converting with 469.7436 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,984B, BPFP=0.5439 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,048B, BPFP=2.2061 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,060B, BPFP=0.8388 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,864B, BPFP=2.1557 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,340B, BPFP=0.9156 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,008B, BPFP=2.1952 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,400B, BPFP=0.9320 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,976B, BPFP=2.1864 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,868B, BPFP=2.1568 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,736B, BPFP=2.1206 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,848B, BPFP=0.3465 +⌛️ [2/4] FRONTEND: Frontend time: 1.779s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.133s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10653571 33.54624966 + layer.0.v_cache 0.00001396 0.01006196 + layer.1.k_cache 0.01513677 9.00729959 + layer.1.v_cache 0.00000574 0.00375176 + layer.2.k_cache 0.00241650 1.00156329 + layer.2.v_cache 0.00001997 0.01235784 + layer.3.k_cache 0.04250543 5.23815704 + layer.3.v_cache 0.00001889 0.01299024 + layer.4.k_cache 0.00060099 0.33446352 + layer.4.v_cache 0.00005317 0.03077650 + layer.4.output 0.25650722 938.13377193 + ------------------------------------------------------------------------------------- + TOTAL 0.11546222 389.18435735 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 68132 +BPFP 1.0986 bits/point +EBPFP 2.1972 equivalent bits/point +MSE 389.184357 +---------------------- -------------------------------------------------------- +Time: 2.916s Load: 0.003s, Pack+Encode: 1.779s, Decode+Unpack: 1.133s +---------------------- -------------------------------------------------------- +💾 Converting with 389.1844 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,844B, BPFP=0.6003 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,912B, BPFP=2.5755 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,832B, BPFP=0.9219 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,676B, BPFP=2.4987 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,516B, BPFP=1.1445 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,796B, BPFP=2.5378 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,976B, BPFP=0.9688 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,024B, BPFP=2.9375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,960B, BPFP=2.5911 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,660B, BPFP=2.4935 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,984B, BPFP=0.3713 +⌛️ [2/4] FRONTEND: Frontend time: 1.928s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.189s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10088900 37.26086934 + layer.0.v_cache 0.00001321 0.00987166 + layer.1.k_cache 0.01601078 8.91693052 + layer.1.v_cache 0.00000529 0.00377240 + layer.2.k_cache 0.00240732 1.04283977 + layer.2.v_cache 0.00001858 0.01290416 + layer.3.k_cache 0.04236705 4.69611486 + layer.3.v_cache 0.00001968 0.01395623 + layer.4.k_cache 0.00061965 0.32149339 + layer.4.v_cache 0.00005235 0.03104200 + layer.4.output 0.30498679 1108.33138021 + ------------------------------------------------------------------------------------- + TOTAL 0.13513591 459.44879152 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 67180 +BPFP 1.2864 bits/point +EBPFP 2.5728 equivalent bits/point +MSE 459.448792 +---------------------- -------------------------------------------------------- +Time: 3.119s Load: 0.003s, Pack+Encode: 1.928s, Decode+Unpack: 1.189s +---------------------- -------------------------------------------------------- +💾 Converting with 459.4488 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 53, 128) +Output shape: (1, 53, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.output: torch.Size([1, 53, 3584]) -> torch.Size([1, 1, 53, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,912B, BPFP=0.5637 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,020B, BPFP=2.3644 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,932B, BPFP=0.8644 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,900B, BPFP=2.3290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,284B, BPFP=0.9682 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,916B, BPFP=2.3337 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,240B, BPFP=0.9552 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,968B, BPFP=2.3491 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,840B, BPFP=2.3113 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,844B, BPFP=2.3125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,220B, BPFP=0.3462 +⌛️ [2/4] FRONTEND: Frontend time: 1.942s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.175s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10150213 32.94001050 + layer.0.v_cache 0.00001370 0.01085328 + layer.1.k_cache 0.01753690 8.86099819 + layer.1.v_cache 0.00000536 0.00433830 + layer.2.k_cache 0.00234856 1.12513755 + layer.2.v_cache 0.00001946 0.01330877 + layer.3.k_cache 0.03772523 6.57457575 + layer.3.v_cache 0.00001792 0.01451018 + layer.4.k_cache 0.00061365 0.37735849 + layer.4.v_cache 0.00006478 0.03211952 + layer.4.output 0.25806696 1011.17511792 + ------------------------------------------------------------------------------------- + TOTAL 0.11566567 419.30464918 + (elements=461,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 461312 +Total Bytes 67076 +BPFP 1.1632 bits/point +EBPFP 2.3264 equivalent bits/point +MSE 419.304649 +---------------------- -------------------------------------------------------- +Time: 3.120s Load: 0.003s, Pack+Encode: 1.942s, Decode+Unpack: 1.175s +---------------------- -------------------------------------------------------- +💾 Converting with 419.3046 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,976B, BPFP=0.5146 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,748B, BPFP=2.2781 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,108B, BPFP=0.8094 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,840B, BPFP=2.0417 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,560B, BPFP=0.9271 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,840B, BPFP=2.0417 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,668B, BPFP=0.9552 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,488B, BPFP=2.7313 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,736B, BPFP=2.0146 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,880B, BPFP=2.0521 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,568B, BPFP=0.3560 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11850098 34.40835775 + layer.0.v_cache 0.00001385 0.01071651 + layer.1.k_cache 0.01468890 11.43140361 + layer.1.v_cache 0.00000523 0.00399016 + layer.2.k_cache 0.01477869 1.00012576 + layer.2.v_cache 0.00001915 0.01329995 + layer.3.k_cache 0.14661819 6.81411641 + layer.3.v_cache 0.00001925 0.01520657 + layer.4.k_cache 0.00060183 0.36411393 + layer.4.v_cache 0.00005752 0.03354601 + layer.4.output 0.22966646 886.59441964 + ------------------------------------------------------------------------------------- + TOTAL 0.11193934 368.25034201 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 72412 +BPFP 1.1093 bits/point +EBPFP 2.2185 equivalent bits/point +MSE 368.250342 +---------------------- -------------------------------------------------------- +Time: 3.048s Load: 0.003s, Pack+Encode: 1.841s, Decode+Unpack: 1.204s +---------------------- -------------------------------------------------------- +💾 Converting with 368.2503 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,852B, BPFP=0.4844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,024B, BPFP=2.7215 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,648B, BPFP=0.7894 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,068B, BPFP=2.5591 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,236B, BPFP=0.8893 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,468B, BPFP=2.4572 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,316B, BPFP=0.9029 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,556B, BPFP=2.6420 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,604B, BPFP=1.8010 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,416B, BPFP=2.4484 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,476B, BPFP=0.2542 +⌛️ [2/4] FRONTEND: Frontend time: 2.198s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.491s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10836371 31.06291928 + layer.0.v_cache 0.00001445 0.00843630 + layer.1.k_cache 0.06846340 11.80231509 + layer.1.v_cache 0.00000523 0.00318532 + layer.2.k_cache 0.00506382 1.23518421 + layer.2.v_cache 0.00001740 0.00979297 + layer.3.k_cache 0.03973729 8.02682760 + layer.3.v_cache 0.00001777 0.01162586 + layer.4.k_cache 0.00066797 0.34013537 + layer.4.v_cache 0.00005099 0.02302668 + layer.4.output 0.14786712 564.79187694 + ------------------------------------------------------------------------------------- + TOTAL 0.07396893 235.65097572 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 114664 +BPFP 1.1455 bits/point +EBPFP 2.2911 equivalent bits/point +MSE 235.650976 +---------------------- -------------------------------------------------------- +Time: 3.694s Load: 0.005s, Pack+Encode: 2.198s, Decode+Unpack: 1.491s +---------------------- -------------------------------------------------------- +💾 Converting with 235.6510 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,108B, BPFP=0.4991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,016B, BPFP=3.3182 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,392B, BPFP=0.8030 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,216B, BPFP=3.1288 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,080B, BPFP=0.9659 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,908B, BPFP=3.0559 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,716B, BPFP=0.8797 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,464B, BPFP=3.1875 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,216B, BPFP=1.9451 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,064B, BPFP=3.0928 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,664B, BPFP=0.2930 +⌛️ [2/4] FRONTEND: Frontend time: 1.953s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.535s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08728886 35.33183890 + layer.0.v_cache 0.00001734 0.00970064 + layer.1.k_cache 0.01499850 7.96580274 + layer.1.v_cache 0.00000528 0.00324722 + layer.2.k_cache 0.00240521 1.02598630 + layer.2.v_cache 0.00001739 0.01033415 + layer.3.k_cache 0.03126823 5.19821861 + layer.3.v_cache 0.00001843 0.01160623 + layer.4.k_cache 0.00061395 0.32539755 + layer.4.v_cache 0.00004830 0.02278248 + layer.4.output 0.21953388 816.83982684 + ------------------------------------------------------------------------------------- + TOTAL 0.09843639 339.28139428 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 96844 +BPFP 1.3487 bits/point +EBPFP 2.6973 equivalent bits/point +MSE 339.281394 +---------------------- -------------------------------------------------------- +Time: 3.491s Load: 0.003s, Pack+Encode: 1.953s, Decode+Unpack: 1.535s +---------------------- -------------------------------------------------------- +💾 Converting with 339.2814 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,088B, BPFP=0.4943 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,184B, BPFP=3.3580 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,448B, BPFP=0.8163 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,532B, BPFP=2.9669 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,380B, BPFP=1.0369 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,264B, BPFP=2.9034 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,728B, BPFP=0.8826 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,556B, BPFP=2.9725 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,120B, BPFP=1.9223 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,060B, BPFP=3.0919 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,528B, BPFP=0.2884 +⌛️ [2/4] FRONTEND: Frontend time: 1.936s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.458s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08223097 36.08445046 + layer.0.v_cache 0.00001442 0.00961772 + layer.1.k_cache 0.01480860 9.59286222 + layer.1.v_cache 0.00000502 0.00316669 + layer.2.k_cache 0.00576455 1.20907662 + layer.2.v_cache 0.00002304 0.01020024 + layer.3.k_cache 0.03215047 4.68243316 + layer.3.v_cache 0.00001933 0.01279302 + layer.4.k_cache 0.00061510 0.27607481 + layer.4.v_cache 0.00004916 0.02410998 + layer.4.output 0.22077460 815.77110390 + ------------------------------------------------------------------------------------- + TOTAL 0.09888840 338.95897131 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 94888 +BPFP 1.3214 bits/point +EBPFP 2.6428 equivalent bits/point +MSE 338.958971 +---------------------- -------------------------------------------------------- +Time: 3.398s Load: 0.004s, Pack+Encode: 1.936s, Decode+Unpack: 1.458s +---------------------- -------------------------------------------------------- +💾 Converting with 338.9590 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,456B, BPFP=0.5257 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,388B, BPFP=3.0796 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,156B, BPFP=0.8896 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,560B, BPFP=2.9024 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,436B, BPFP=0.9495 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,524B, BPFP=2.6807 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,460B, BPFP=0.9546 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,836B, BPFP=2.9615 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,496B, BPFP=2.0325 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,320B, BPFP=2.8510 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,100B, BPFP=0.2783 +⌛️ [2/4] FRONTEND: Frontend time: 2.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.408s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10114760 31.32766682 + layer.0.v_cache 0.00001420 0.00906845 + layer.1.k_cache 0.03683303 8.49712717 + layer.1.v_cache 0.00000544 0.00332335 + layer.2.k_cache 0.00248627 1.11474923 + layer.2.v_cache 0.00001713 0.00983742 + layer.3.k_cache 0.02909129 4.56083616 + layer.3.v_cache 0.00001766 0.01158265 + layer.4.k_cache 0.00072613 0.33686928 + layer.4.v_cache 0.00005154 0.02400412 + layer.4.output 0.19731793 737.06311155 + ------------------------------------------------------------------------------------- + TOTAL 0.09127152 306.19628503 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 101732 +BPFP 1.2809 bits/point +EBPFP 2.5617 equivalent bits/point +MSE 306.196285 +---------------------- -------------------------------------------------------- +Time: 3.577s Load: 0.004s, Pack+Encode: 2.165s, Decode+Unpack: 1.408s +---------------------- -------------------------------------------------------- +💾 Converting with 306.1963 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,564B, BPFP=0.5271 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,836B, BPFP=3.0502 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,664B, BPFP=0.9589 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,448B, BPFP=2.7648 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,736B, BPFP=1.1793 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,288B, BPFP=2.7319 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,668B, BPFP=0.9597 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,124B, BPFP=2.9038 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,996B, BPFP=2.0551 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,888B, BPFP=2.8553 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,300B, BPFP=0.3025 +⌛️ [2/4] FRONTEND: Frontend time: 1.956s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.288s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10230959 34.31320030 + layer.0.v_cache 0.00001409 0.00947363 + layer.1.k_cache 0.07973289 9.82547318 + layer.1.v_cache 0.00000506 0.00301948 + layer.2.k_cache 0.00553133 1.14656348 + layer.2.v_cache 0.00001928 0.01006822 + layer.3.k_cache 0.03946657 6.64774523 + layer.3.v_cache 0.00001980 0.01159593 + layer.4.k_cache 0.00066525 0.34470799 + layer.4.v_cache 0.00005119 0.02188592 + layer.4.output 0.17985767 693.53912124 + ------------------------------------------------------------------------------------- + TOTAL 0.08745993 288.65338718 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 107512 +BPFP 1.3002 bits/point +EBPFP 2.6004 equivalent bits/point +MSE 288.653387 +---------------------- -------------------------------------------------------- +Time: 3.248s Load: 0.004s, Pack+Encode: 1.956s, Decode+Unpack: 1.288s +---------------------- -------------------------------------------------------- +💾 Converting with 288.6534 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,688B, BPFP=0.5455 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,160B, BPFP=2.8734 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,172B, BPFP=0.8466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,312B, BPFP=2.4984 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,276B, BPFP=1.0706 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,312B, BPFP=2.4984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,716B, BPFP=0.9570 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,980B, BPFP=2.6339 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,852B, BPFP=1.7963 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,740B, BPFP=2.5852 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,544B, BPFP=0.2767 +⌛️ [2/4] FRONTEND: Frontend time: 2.038s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.438s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633976 32.87470830 + layer.0.v_cache 0.00001461 0.00924614 + layer.1.k_cache 0.07748085 9.63189222 + layer.1.v_cache 0.00000515 0.00311648 + layer.2.k_cache 0.00505460 1.24668310 + layer.2.v_cache 0.00001651 0.00958500 + layer.3.k_cache 0.02793909 7.39226661 + layer.3.v_cache 0.00001809 0.01114958 + layer.4.k_cache 0.00070819 0.31783324 + layer.4.v_cache 0.00004716 0.02164954 + layer.4.output 0.18765952 690.13410250 + ------------------------------------------------------------------------------------- + TOTAL 0.09007298 287.20334399 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 99752 +BPFP 1.1907 bits/point +EBPFP 2.3814 equivalent bits/point +MSE 287.203344 +---------------------- -------------------------------------------------------- +Time: 3.480s Load: 0.004s, Pack+Encode: 2.038s, Decode+Unpack: 1.438s +---------------------- -------------------------------------------------------- +💾 Converting with 287.2033 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,368B, BPFP=0.5286 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,180B, BPFP=3.1652 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,816B, BPFP=0.8518 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,336B, BPFP=2.9768 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,120B, BPFP=0.9196 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,020B, BPFP=2.9062 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,360B, BPFP=0.9732 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,684B, BPFP=3.0545 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,536B, BPFP=1.9054 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,944B, BPFP=2.8893 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,316B, BPFP=0.2652 +⌛️ [2/4] FRONTEND: Frontend time: 2.103s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.405s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08399124 33.31875698 + layer.0.v_cache 0.00001671 0.00902342 + layer.1.k_cache 0.03849935 7.52685024 + layer.1.v_cache 0.00000503 0.00313260 + layer.2.k_cache 0.00738276 0.93764779 + layer.2.v_cache 0.00001692 0.00998809 + layer.3.k_cache 0.04914829 4.77521798 + layer.3.v_cache 0.00001713 0.01126334 + layer.4.k_cache 0.00063731 0.30332366 + layer.4.v_cache 0.00004711 0.02249691 + layer.4.output 0.21328995 761.18737245 + ------------------------------------------------------------------------------------- + TOTAL 0.09839950 316.18995930 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 98680 +BPFP 1.2957 bits/point +EBPFP 2.5914 equivalent bits/point +MSE 316.189959 +---------------------- -------------------------------------------------------- +Time: 3.512s Load: 0.004s, Pack+Encode: 2.103s, Decode+Unpack: 1.405s +---------------------- -------------------------------------------------------- +💾 Converting with 316.1900 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 56, 128) +Output shape: (1, 56, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.output: torch.Size([1, 56, 3584]) -> torch.Size([1, 1, 56, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,016B, BPFP=0.5625 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,048B, BPFP=2.2455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,984B, BPFP=0.8326 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,808B, BPFP=2.1786 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,288B, BPFP=0.9174 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,740B, BPFP=2.1596 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,276B, BPFP=0.9141 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,004B, BPFP=2.2333 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,912B, BPFP=2.2076 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,756B, BPFP=2.1641 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,820B, BPFP=0.3117 +⌛️ [2/4] FRONTEND: Frontend time: 2.061s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.273s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13768739 34.59713745 + layer.0.v_cache 0.00001432 0.01034053 + layer.1.k_cache 0.01638686 9.25617872 + layer.1.v_cache 0.00000533 0.00357276 + layer.2.k_cache 0.00251956 0.96229635 + layer.2.v_cache 0.00001745 0.01121861 + layer.3.k_cache 0.08599430 6.13358198 + layer.3.v_cache 0.00002012 0.01387628 + layer.4.k_cache 0.00061995 0.31656350 + layer.4.v_cache 0.00004798 0.02580741 + layer.4.output 0.24256272 961.93463010 + ------------------------------------------------------------------------------------- + TOTAL 0.11419131 399.11017555 + (elements=487,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 487424 +Total Bytes 66652 +BPFP 1.0939 bits/point +EBPFP 2.1879 equivalent bits/point +MSE 399.110176 +---------------------- -------------------------------------------------------- +Time: 3.339s Load: 0.005s, Pack+Encode: 2.061s, Decode+Unpack: 1.273s +---------------------- -------------------------------------------------------- +💾 Converting with 399.1102 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,612B, BPFP=0.5102 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,856B, BPFP=2.9016 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,228B, BPFP=0.8258 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,044B, BPFP=2.7430 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,448B, BPFP=0.8688 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,712B, BPFP=2.6781 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,260B, BPFP=0.8320 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,076B, BPFP=2.7492 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,204B, BPFP=1.9930 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,700B, BPFP=2.6758 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,100B, BPFP=0.2818 +⌛️ [2/4] FRONTEND: Frontend time: 2.007s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.529s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09278157 30.40851440 + layer.0.v_cache 0.00001504 0.00875351 + layer.1.k_cache 0.07523044 8.28502808 + layer.1.v_cache 0.00000517 0.00319797 + layer.2.k_cache 0.00381728 0.95439034 + layer.2.v_cache 0.00001789 0.00958122 + layer.3.k_cache 0.03327554 4.43764267 + layer.3.v_cache 0.00001939 0.01142979 + layer.4.k_cache 0.00067710 0.34701083 + layer.4.v_cache 0.00005192 0.02347557 + layer.4.output 0.16994183 674.69994420 + ------------------------------------------------------------------------------------- + TOTAL 0.08208730 280.43462551 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 106240 +BPFP 1.2206 bits/point +EBPFP 2.4412 equivalent bits/point +MSE 280.434626 +---------------------- -------------------------------------------------------- +Time: 3.539s Load: 0.004s, Pack+Encode: 2.007s, Decode+Unpack: 1.529s +---------------------- -------------------------------------------------------- +💾 Converting with 280.4346 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,996B, BPFP=0.5286 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,188B, BPFP=2.1684 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,056B, BPFP=0.8093 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,268B, BPFP=2.1896 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,616B, BPFP=0.9576 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,864B, BPFP=2.0826 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,432B, BPFP=0.9089 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,816B, BPFP=2.3347 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,560B, BPFP=2.0021 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,784B, BPFP=2.0614 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,212B, BPFP=0.3485 +⌛️ [2/4] FRONTEND: Frontend time: 1.693s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.179s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12468526 38.76652708 + layer.0.v_cache 0.00001451 0.01102987 + layer.1.k_cache 0.01313243 8.46221458 + layer.1.v_cache 0.00000580 0.00413330 + layer.2.k_cache 0.00906488 1.07743803 + layer.2.v_cache 0.00002099 0.01304351 + layer.3.k_cache 0.01569995 5.10312614 + layer.3.v_cache 0.00002002 0.01482312 + layer.4.k_cache 0.00064964 0.36823845 + layer.4.v_cache 0.00005478 0.03087243 + layer.4.output 0.23031524 908.21950666 + ------------------------------------------------------------------------------------- + TOTAL 0.10444441 377.14047018 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 69792 +BPFP 1.0872 bits/point +EBPFP 2.1745 equivalent bits/point +MSE 377.140470 +---------------------- -------------------------------------------------------- +Time: 2.876s Load: 0.004s, Pack+Encode: 1.693s, Decode+Unpack: 1.179s +---------------------- -------------------------------------------------------- +💾 Converting with 377.1405 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,996B, BPFP=0.5198 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,148B, BPFP=2.1219 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,048B, BPFP=0.7937 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,932B, BPFP=2.0656 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,532B, BPFP=0.9198 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,860B, BPFP=2.0469 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,280B, BPFP=0.8542 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,084B, BPFP=2.8865 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,544B, BPFP=1.9646 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,876B, BPFP=2.0510 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,008B, BPFP=0.3723 +⌛️ [2/4] FRONTEND: Frontend time: 2.076s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.245s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15164433 36.87751872 + layer.0.v_cache 0.00001435 0.01194533 + layer.1.k_cache 0.01704682 9.71457520 + layer.1.v_cache 0.00000573 0.00432566 + layer.2.k_cache 0.00240009 1.00761681 + layer.2.v_cache 0.00001845 0.01256500 + layer.3.k_cache 0.02161660 5.06041667 + layer.3.v_cache 0.00001985 0.01547817 + layer.4.k_cache 0.00061459 0.35194492 + layer.4.v_cache 0.00005275 0.03344124 + layer.4.output 0.22653077 897.81659226 + ------------------------------------------------------------------------------------- + TOTAL 0.10465582 372.81211609 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 72308 +BPFP 1.1077 bits/point +EBPFP 2.2153 equivalent bits/point +MSE 372.812116 +---------------------- -------------------------------------------------------- +Time: 3.324s Load: 0.003s, Pack+Encode: 2.076s, Decode+Unpack: 1.245s +---------------------- -------------------------------------------------------- +💾 Converting with 372.8121 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,460B, BPFP=0.5265 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,692B, BPFP=2.9307 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,948B, BPFP=0.8450 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,324B, BPFP=2.6378 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,236B, BPFP=0.9067 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,780B, BPFP=2.5214 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,856B, BPFP=0.8253 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,860B, BPFP=2.7526 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,828B, BPFP=1.8896 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,444B, BPFP=2.6635 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,808B, BPFP=0.2693 +⌛️ [2/4] FRONTEND: Frontend time: 1.972s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.329s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08869729 33.48279310 + layer.0.v_cache 0.00001594 0.00898424 + layer.1.k_cache 0.08105310 7.89371856 + layer.1.v_cache 0.00000494 0.00310373 + layer.2.k_cache 0.00256802 0.94352983 + layer.2.v_cache 0.00001661 0.00939195 + layer.3.k_cache 0.01835501 4.33232138 + layer.3.v_cache 0.00001861 0.01132310 + layer.4.k_cache 0.00078324 0.31888450 + layer.4.v_cache 0.00004769 0.02261122 + layer.4.output 0.18611241 723.13931018 + ------------------------------------------------------------------------------------- + TOTAL 0.08790278 300.52951958 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 95236 +BPFP 1.1991 bits/point +EBPFP 2.3982 equivalent bits/point +MSE 300.529520 +---------------------- -------------------------------------------------------- +Time: 3.306s Load: 0.005s, Pack+Encode: 1.972s, Decode+Unpack: 1.329s +---------------------- -------------------------------------------------------- +💾 Converting with 300.5295 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,736B, BPFP=0.4407 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,040B, BPFP=2.5838 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,028B, BPFP=0.9710 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,128B, BPFP=2.5979 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,308B, BPFP=1.0161 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,452B, BPFP=2.4890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,468B, BPFP=1.0419 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,016B, BPFP=2.5799 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,472B, BPFP=1.6869 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,792B, BPFP=2.3827 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,252B, BPFP=0.2589 +⌛️ [2/4] FRONTEND: Frontend time: 1.917s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.285s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17228427 30.57926264 + layer.0.v_cache 0.00001543 0.00800496 + layer.1.k_cache 0.04621154 12.07296910 + layer.1.v_cache 0.00000545 0.00287225 + layer.2.k_cache 0.00739691 1.21716238 + layer.2.v_cache 0.00001797 0.00923967 + layer.3.k_cache 0.03096496 9.11803130 + layer.3.v_cache 0.00001890 0.01051965 + layer.4.k_cache 0.00071091 0.36247179 + layer.4.v_cache 0.00004900 0.02388992 + layer.4.output 0.02450962 570.68239138 + ------------------------------------------------------------------------------------- + TOTAL 0.02524957 238.12830373 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 121692 +BPFP 1.1531 bits/point +EBPFP 2.3062 equivalent bits/point +MSE 238.128304 +---------------------- -------------------------------------------------------- +Time: 3.209s Load: 0.006s, Pack+Encode: 1.917s, Decode+Unpack: 1.285s +---------------------- -------------------------------------------------------- +💾 Converting with 238.1283 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,940B, BPFP=0.4940 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,236B, BPFP=2.7278 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,864B, BPFP=0.8172 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,588B, BPFP=2.6190 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,956B, BPFP=1.0007 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,140B, BPFP=2.5437 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,764B, BPFP=0.9684 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,944B, BPFP=2.6788 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,108B, BPFP=1.8663 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,040B, BPFP=2.5269 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,068B, BPFP=0.2416 +⌛️ [2/4] FRONTEND: Frontend time: 2.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16174046 40.80664588 + layer.0.v_cache 0.00001612 0.00834122 + layer.1.k_cache 0.03352992 10.74074169 + layer.1.v_cache 0.00000563 0.00324123 + layer.2.k_cache 0.00496730 1.40984976 + layer.2.v_cache 0.00001740 0.00982358 + layer.3.k_cache 0.01521110 8.91892037 + layer.3.v_cache 0.00001939 0.01133879 + layer.4.k_cache 0.00066431 0.31491692 + layer.4.v_cache 0.00005067 0.02220994 + layer.4.output 0.14624557 578.61160714 + ------------------------------------------------------------------------------------- + TOTAL 0.07293772 241.91336938 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 118648 +BPFP 1.1726 bits/point +EBPFP 2.3452 equivalent bits/point +MSE 241.913369 +---------------------- -------------------------------------------------------- +Time: 3.582s Load: 0.005s, Pack+Encode: 2.230s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 241.9134 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,720B, BPFP=0.4942 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,140B, BPFP=2.7507 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,268B, BPFP=0.7754 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,208B, BPFP=2.5814 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,680B, BPFP=0.8503 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,636B, BPFP=2.4775 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,552B, BPFP=0.8270 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,716B, BPFP=2.6737 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,156B, BPFP=1.6635 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,056B, BPFP=2.5538 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,056B, BPFP=0.2870 +⌛️ [2/4] FRONTEND: Frontend time: 2.004s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07121617 31.06029138 + layer.0.v_cache 0.00001681 0.00853717 + layer.1.k_cache 0.06877367 8.26468037 + layer.1.v_cache 0.00000530 0.00311558 + layer.2.k_cache 0.00486504 0.95582465 + layer.2.v_cache 0.00001749 0.00933635 + layer.3.k_cache 0.08833314 5.14844176 + layer.3.v_cache 0.00001723 0.01047991 + layer.4.k_cache 0.00077278 0.33867015 + layer.4.v_cache 0.00004595 0.02289719 + layer.4.output 0.15812202 627.48473837 + ------------------------------------------------------------------------------------- + TOTAL 0.07887751 261.07149665 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 108188 +BPFP 1.1562 bits/point +EBPFP 2.3125 equivalent bits/point +MSE 261.071497 +---------------------- -------------------------------------------------------- +Time: 3.390s Load: 0.007s, Pack+Encode: 2.004s, Decode+Unpack: 1.380s +---------------------- -------------------------------------------------------- +💾 Converting with 261.0715 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,056B, BPFP=0.4728 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,368B, BPFP=2.5322 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,544B, BPFP=0.8577 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,656B, BPFP=2.4220 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,968B, BPFP=0.9233 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,312B, BPFP=2.3688 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,604B, BPFP=0.8670 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,944B, BPFP=2.4666 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,976B, BPFP=1.8527 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,196B, BPFP=2.3509 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,988B, BPFP=0.2207 +⌛️ [2/4] FRONTEND: Frontend time: 2.025s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.265s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13964315 37.21067644 + layer.0.v_cache 0.00001432 0.00840159 + layer.1.k_cache 0.04810369 8.36581814 + layer.1.v_cache 0.00000572 0.00319481 + layer.2.k_cache 0.00492308 1.20195581 + layer.2.v_cache 0.00001777 0.01020015 + layer.3.k_cache 0.04255116 7.42077758 + layer.3.v_cache 0.00001880 0.01148922 + layer.4.k_cache 0.00073971 0.31768599 + layer.4.v_cache 0.00004573 0.02252837 + layer.4.output 11.28739116 528.81559406 + ------------------------------------------------------------------------------------- + TOTAL 4.66163537 220.95775803 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 120612 +BPFP 1.0976 bits/point +EBPFP 2.1952 equivalent bits/point +MSE 220.957758 +---------------------- -------------------------------------------------------- +Time: 3.294s Load: 0.004s, Pack+Encode: 2.025s, Decode+Unpack: 1.265s +---------------------- -------------------------------------------------------- +💾 Converting with 220.9578 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,708B, BPFP=0.4808 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,604B, BPFP=2.7706 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,408B, BPFP=0.7827 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,524B, BPFP=2.5788 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,920B, BPFP=0.8736 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,092B, BPFP=2.5021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,188B, BPFP=0.9212 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,920B, BPFP=2.6491 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,528B, BPFP=1.8693 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,084B, BPFP=2.5007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,260B, BPFP=0.2602 +⌛️ [2/4] FRONTEND: Frontend time: 1.851s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.535s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14145413 32.79384544 + layer.0.v_cache 0.00001516 0.00803999 + layer.1.k_cache 0.10647225 8.32658594 + layer.1.v_cache 0.00000560 0.00309759 + layer.2.k_cache 0.00244946 0.98759998 + layer.2.v_cache 0.00001685 0.00997700 + layer.3.k_cache 0.01581429 7.83381722 + layer.3.v_cache 0.00001841 0.01155884 + layer.4.k_cache 0.00072023 0.34718050 + layer.4.v_cache 0.00004652 0.02101397 + layer.4.output 0.15451367 612.73625203 + ------------------------------------------------------------------------------------- + TOTAL 0.07932992 255.26449886 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 111236 +BPFP 1.1618 bits/point +EBPFP 2.3236 equivalent bits/point +MSE 255.264499 +---------------------- -------------------------------------------------------- +Time: 3.390s Load: 0.004s, Pack+Encode: 1.851s, Decode+Unpack: 1.535s +---------------------- -------------------------------------------------------- +💾 Converting with 255.2645 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,488B, BPFP=0.5253 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,272B, BPFP=3.0135 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,100B, BPFP=0.8657 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,292B, BPFP=2.8066 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,956B, BPFP=1.0465 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,392B, BPFP=2.6166 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,692B, BPFP=0.9907 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,244B, BPFP=2.7965 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,552B, BPFP=1.5946 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,056B, BPFP=2.7568 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,024B, BPFP=0.2420 +⌛️ [2/4] FRONTEND: Frontend time: 2.306s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.233s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10237078 34.33172772 + layer.0.v_cache 0.00001391 0.00848286 + layer.1.k_cache 0.05953315 10.86727988 + layer.1.v_cache 0.00000517 0.00313431 + layer.2.k_cache 0.00728044 1.04787115 + layer.2.v_cache 0.00001648 0.00929183 + layer.3.k_cache 0.03101032 7.12550643 + layer.3.v_cache 0.00001666 0.01076450 + layer.4.k_cache 0.00072555 0.25906547 + layer.4.v_cache 0.00004870 0.02430365 + layer.4.output 0.18364459 727.62572394 + ------------------------------------------------------------------------------------- + TOTAL 0.08744314 302.76867620 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 98068 +BPFP 1.2181 bits/point +EBPFP 2.4361 equivalent bits/point +MSE 302.768676 +---------------------- -------------------------------------------------------- +Time: 3.544s Load: 0.005s, Pack+Encode: 2.306s, Decode+Unpack: 1.233s +---------------------- -------------------------------------------------------- +💾 Converting with 302.7687 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 111, 128) +Output shape: (1, 111, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.output: torch.Size([1, 111, 3584]) -> torch.Size([1, 1, 111, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,244B, BPFP=0.4566 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,084B, BPFP=2.2641 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,488B, BPFP=0.7725 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,816B, BPFP=2.2264 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,360B, BPFP=0.8953 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,236B, BPFP=2.2855 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,840B, BPFP=0.8221 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,304B, BPFP=2.2950 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,536B, BPFP=1.7646 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,504B, BPFP=2.1824 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,812B, BPFP=0.2375 +⌛️ [2/4] FRONTEND: Frontend time: 1.850s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11107565 31.52220580 + layer.0.v_cache 0.00001597 0.00778091 + layer.1.k_cache 0.04333284 8.72297297 + layer.1.v_cache 0.00000558 0.00292750 + layer.2.k_cache 0.00477623 1.18513558 + layer.2.v_cache 0.00001758 0.00966597 + layer.3.k_cache 0.01020195 5.63934601 + layer.3.v_cache 0.00001870 0.01018023 + layer.4.k_cache 0.00079488 0.34793788 + layer.4.v_cache 0.00005041 0.02258330 + layer.4.output 10.27057757 478.93967181 + ------------------------------------------------------------------------------------- + TOTAL 4.23907840 200.00284934 + (elements=966,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 966144 +Total Bytes 125224 +BPFP 1.0369 bits/point +EBPFP 2.0738 equivalent bits/point +MSE 200.002849 +---------------------- -------------------------------------------------------- +Time: 3.235s Load: 0.007s, Pack+Encode: 1.850s, Decode+Unpack: 1.378s +---------------------- -------------------------------------------------------- +💾 Converting with 200.0028 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,420B, BPFP=0.5326 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,084B, BPFP=3.0995 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,988B, BPFP=0.8776 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,228B, BPFP=2.9111 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,928B, BPFP=1.0845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,760B, BPFP=2.8081 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,504B, BPFP=0.9912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,528B, BPFP=2.9771 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,040B, BPFP=1.9894 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,088B, BPFP=2.8803 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,940B, BPFP=0.3125 +⌛️ [2/4] FRONTEND: Frontend time: 1.971s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.516s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09375787 34.51876788 + layer.0.v_cache 0.00001622 0.00832757 + layer.1.k_cache 0.05984993 10.67483993 + layer.1.v_cache 0.00000512 0.00291799 + layer.2.k_cache 0.00243208 1.10386507 + layer.2.v_cache 0.00001865 0.00948055 + layer.3.k_cache 0.04944374 6.45214414 + layer.3.v_cache 0.00001908 0.01094426 + layer.4.k_cache 0.00084877 0.32462023 + layer.4.v_cache 0.00005139 0.02328514 + layer.4.output 0.19371969 756.07344064 + ------------------------------------------------------------------------------------- + TOTAL 0.09191063 314.44960455 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 101508 +BPFP 1.3141 bits/point +EBPFP 2.6281 equivalent bits/point +MSE 314.449605 +---------------------- -------------------------------------------------------- +Time: 3.490s Load: 0.004s, Pack+Encode: 1.971s, Decode+Unpack: 1.516s +---------------------- -------------------------------------------------------- +💾 Converting with 314.4496 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 109, 128) +Output shape: (1, 109, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.output: torch.Size([1, 109, 3584]) -> torch.Size([1, 1, 109, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,196B, BPFP=0.4581 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,144B, BPFP=2.3142 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,384B, BPFP=0.7718 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,716B, BPFP=2.2529 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,192B, BPFP=0.8876 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,680B, BPFP=2.2477 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,980B, BPFP=0.8572 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,128B, BPFP=2.3119 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,256B, BPFP=1.9002 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,452B, BPFP=2.2150 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,868B, BPFP=0.2635 +⌛️ [2/4] FRONTEND: Frontend time: 2.376s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.507s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860953 32.76237726 + layer.0.v_cache 0.00001500 0.00795155 + layer.1.k_cache 0.05917679 8.12195384 + layer.1.v_cache 0.00000528 0.00303554 + layer.2.k_cache 0.01253135 1.07905320 + layer.2.v_cache 0.00001869 0.00956237 + layer.3.k_cache 0.07406004 7.55671790 + layer.3.v_cache 0.00001854 0.01080126 + layer.4.k_cache 0.00071234 0.34469339 + layer.4.v_cache 0.00004916 0.02165960 + layer.4.output 10.45909253 490.45490662 + ------------------------------------------------------------------------------------- + TOTAL 4.32287320 204.88836190 + (elements=948,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 948736 +Total Bytes 125996 +BPFP 1.0624 bits/point +EBPFP 2.1249 equivalent bits/point +MSE 204.888362 +---------------------- -------------------------------------------------------- +Time: 3.890s Load: 0.007s, Pack+Encode: 2.376s, Decode+Unpack: 1.507s +---------------------- -------------------------------------------------------- +💾 Converting with 204.8884 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst + to output-fixed/kimiaudio/lambda0.007/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.1976 bits/point +Avg EBPFP 2.3953 equivalent bits/point +Avg MSE 313.256818 +Avg Time 3.342s +------------------------ ---------------------------- diff --git a/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..07fc4052c9594f0a421c84e03da1cd0f4c0d00af --- /dev/null +++ b/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 599 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench +Output output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,932B, BPFP=0.5656 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,880B, BPFP=2.0988 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,988B, BPFP=0.9622 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,548B, BPFP=2.0347 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,316B, BPFP=1.0255 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,928B, BPFP=1.9151 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,092B, BPFP=0.9823 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,960B, BPFP=1.9213 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,312B, BPFP=1.6034 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,564B, BPFP=1.8449 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,612B, BPFP=0.2924 +⌛️ [2/4] FRONTEND: Frontend time: 0.400s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.240s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12089630 68.47574870 + layer.0.v_cache 0.00001394 0.00957370 + layer.1.k_cache 0.01174029 10.99494312 + layer.1.v_cache 0.00000579 0.00397846 + layer.2.k_cache 0.00835781 1.16004982 + layer.2.v_cache 0.00001852 0.01096530 + layer.3.k_cache 0.02896961 4.92476550 + layer.3.v_cache 0.00001872 0.01180413 + layer.4.k_cache 0.00063445 0.24437968 + layer.4.v_cache 0.00005063 0.02359109 + layer.4.output 0.17246709 668.56266534 + ------------------------------------------------------------------------------------- + TOTAL 0.08105739 280.34108570 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 88132 +BPFP 1.0000 bits/point +EBPFP 2.0001 equivalent bits/point +MSE 280.341086 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.005s, Pack+Encode: 0.400s, Decode+Unpack: 0.240s +---------------------- -------------------------------------------------------- +💾 Converting with 280.3411 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,788B, BPFP=0.5445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,944B, BPFP=2.1375 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,880B, BPFP=0.9531 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,836B, BPFP=2.1164 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,648B, BPFP=1.1031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,224B, BPFP=1.9969 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,176B, BPFP=1.0109 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,348B, BPFP=2.0211 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,324B, BPFP=1.6258 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,992B, BPFP=1.9516 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,264B, BPFP=0.2864 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08029544 74.60729980 + layer.0.v_cache 0.00001376 0.01017275 + layer.1.k_cache 0.03632395 11.10391541 + layer.1.v_cache 0.00000563 0.00427991 + layer.2.k_cache 0.00522719 1.17931824 + layer.2.v_cache 0.00001997 0.01258738 + layer.3.k_cache 0.02707962 4.86192703 + layer.3.v_cache 0.00001843 0.01347745 + layer.4.k_cache 0.00063057 0.25283709 + layer.4.v_cache 0.00005085 0.02683750 + layer.4.output 0.18451144 676.48504464 + ------------------------------------------------------------------------------------- + TOTAL 0.08477915 283.96870383 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 89424 +BPFP 1.0274 bits/point +EBPFP 2.0548 equivalent bits/point +MSE 283.968704 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 283.9687 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,052B, BPFP=0.5545 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,292B, BPFP=2.0516 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,148B, BPFP=0.9353 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,852B, BPFP=1.9717 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,476B, BPFP=0.9949 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,312B, BPFP=1.8735 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,228B, BPFP=0.9499 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,572B, BPFP=1.9208 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,552B, BPFP=1.5538 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,024B, BPFP=1.8212 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,428B, BPFP=0.2707 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10874708 70.55427871 + layer.0.v_cache 0.00001608 0.01005624 + layer.1.k_cache 0.03270756 10.74274959 + layer.1.v_cache 0.00000572 0.00402497 + layer.2.k_cache 0.00777350 1.19725835 + layer.2.v_cache 0.00002080 0.01128176 + layer.3.k_cache 0.05920834 5.00591012 + layer.3.v_cache 0.00001846 0.01229779 + layer.4.k_cache 0.00062493 0.23979446 + layer.4.v_cache 0.00005036 0.02384293 + layer.4.output 0.17137358 629.63600498 + ------------------------------------------------------------------------------------- + TOTAL 0.08286988 264.42667822 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 90936 +BPFP 0.9719 bits/point +EBPFP 1.9437 equivalent bits/point +MSE 264.426678 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 264.4267 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,576B, BPFP=0.5160 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,440B, BPFP=2.0913 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,712B, BPFP=0.9439 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,276B, BPFP=2.0585 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,300B, BPFP=1.0617 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,876B, BPFP=1.9784 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,872B, BPFP=0.9760 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,900B, BPFP=1.9832 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,048B, BPFP=1.6122 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,616B, BPFP=1.9263 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,560B, BPFP=0.3022 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12073612 76.92611929 + layer.0.v_cache 0.00001348 0.01014913 + layer.1.k_cache 0.03279715 11.45763378 + layer.1.v_cache 0.00000568 0.00463546 + layer.2.k_cache 0.00696802 1.26394419 + layer.2.v_cache 0.00001833 0.01231676 + layer.3.k_cache 0.03227850 4.66605866 + layer.3.v_cache 0.00001834 0.01374064 + layer.4.k_cache 0.00060115 0.27654653 + layer.4.v_cache 0.00005191 0.02704346 + layer.4.output 0.18320683 693.48489011 + ------------------------------------------------------------------------------------- + TOTAL 0.08681979 291.12073051 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 86176 +BPFP 1.0155 bits/point +EBPFP 2.0309 equivalent bits/point +MSE 291.120731 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 291.1207 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,520B, BPFP=0.5048 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,628B, BPFP=2.1290 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,792B, BPFP=0.9599 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,416B, BPFP=2.0865 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,284B, BPFP=1.0585 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,752B, BPFP=1.9535 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,032B, BPFP=1.0080 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,876B, BPFP=1.9784 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,016B, BPFP=1.6058 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,416B, BPFP=1.8862 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,612B, BPFP=0.3037 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10595027 78.94379758 + layer.0.v_cache 0.00001376 0.00989113 + layer.1.k_cache 0.03307601 11.03725023 + layer.1.v_cache 0.00000550 0.00414363 + layer.2.k_cache 0.00695119 1.23276833 + layer.2.v_cache 0.00001929 0.01149347 + layer.3.k_cache 0.03450640 4.40346625 + layer.3.v_cache 0.00001844 0.01268807 + layer.4.k_cache 0.00061739 0.25010422 + layer.4.v_cache 0.00005297 0.02444983 + layer.4.output 0.18589628 692.43543956 + ------------------------------------------------------------------------------------- + TOTAL 0.08720501 290.76341939 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 86344 +BPFP 1.0174 bits/point +EBPFP 2.0349 equivalent bits/point +MSE 290.763419 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 290.7634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,704B, BPFP=0.5348 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,792B, BPFP=2.1345 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,836B, BPFP=0.9565 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,372B, BPFP=2.0514 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,500B, BPFP=1.0878 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,812B, BPFP=1.9407 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,120B, BPFP=1.0127 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,988B, BPFP=1.9755 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,304B, BPFP=1.6424 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,068B, BPFP=1.9913 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,740B, BPFP=0.2752 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07772056 74.02196648 + layer.0.v_cache 0.00001351 0.01031818 + layer.1.k_cache 0.03518496 10.96655892 + layer.1.v_cache 0.00000556 0.00412662 + layer.2.k_cache 0.00230807 1.16270968 + layer.2.v_cache 0.00001833 0.01165205 + layer.3.k_cache 0.04514034 5.00165181 + layer.3.v_cache 0.00001901 0.01310007 + layer.4.k_cache 0.00062178 0.26182295 + layer.4.v_cache 0.00004990 0.02667461 + layer.4.output 0.18426369 683.86036392 + ------------------------------------------------------------------------------------- + TOTAL 0.08534870 286.97077228 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 87236 +BPFP 1.0149 bits/point +EBPFP 2.0299 equivalent bits/point +MSE 286.970772 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 286.9708 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,580B, BPFP=0.5235 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,712B, BPFP=2.1737 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,744B, BPFP=0.9627 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,484B, BPFP=2.1274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,336B, BPFP=1.0828 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,908B, BPFP=2.0106 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,024B, BPFP=1.0195 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,084B, BPFP=2.0463 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,300B, BPFP=1.6843 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,952B, BPFP=2.0195 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,464B, BPFP=0.3033 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14568528 74.44735440 + layer.0.v_cache 0.00001322 0.01006663 + layer.1.k_cache 0.01379078 11.09792734 + layer.1.v_cache 0.00000548 0.00397765 + layer.2.k_cache 0.00727733 1.21760103 + layer.2.v_cache 0.00001786 0.01231654 + layer.3.k_cache 0.03082054 4.98400800 + layer.3.v_cache 0.00001879 0.01396218 + layer.4.k_cache 0.00061820 0.25939055 + layer.4.v_cache 0.00005377 0.02637729 + layer.4.output 0.19662610 702.01229128 + ------------------------------------------------------------------------------------- + TOTAL 0.09262847 294.47994239 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 87588 +BPFP 1.0455 bits/point +EBPFP 2.0910 equivalent bits/point +MSE 294.479942 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 294.4799 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,000B, BPFP=0.5859 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,144B, BPFP=2.1766 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,844B, BPFP=0.9461 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,888B, BPFP=2.1266 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,532B, BPFP=1.0805 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,128B, BPFP=1.9781 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,124B, BPFP=1.0008 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,188B, BPFP=1.9898 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,316B, BPFP=1.6242 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,764B, BPFP=1.9070 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,692B, BPFP=0.2983 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10149734 75.35426636 + layer.0.v_cache 0.00001421 0.01036704 + layer.1.k_cache 0.01324784 11.23240967 + layer.1.v_cache 0.00000586 0.00430702 + layer.2.k_cache 0.00706157 1.21807890 + layer.2.v_cache 0.00001822 0.01173034 + layer.3.k_cache 0.02712160 4.91351395 + layer.3.v_cache 0.00002016 0.01258254 + layer.4.k_cache 0.00063687 0.24762263 + layer.4.v_cache 0.00005451 0.02573890 + layer.4.output 0.17858309 675.03303571 + ------------------------------------------------------------------------------------- + TOTAL 0.08233881 283.42716867 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 89620 +BPFP 1.0296 bits/point +EBPFP 2.0593 equivalent bits/point +MSE 283.427169 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 283.4272 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,192B, BPFP=0.5733 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,296B, BPFP=2.0287 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,292B, BPFP=0.9504 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,136B, BPFP=2.0000 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,712B, BPFP=1.0259 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,632B, BPFP=1.9095 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,356B, BPFP=0.9619 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,744B, BPFP=1.9296 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,796B, BPFP=1.5797 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,356B, BPFP=1.8599 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,424B, BPFP=0.2674 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10772817 69.13146552 + layer.0.v_cache 0.00001348 0.01013572 + layer.1.k_cache 0.03276526 10.92762002 + layer.1.v_cache 0.00000562 0.00411399 + layer.2.k_cache 0.01020141 1.20489967 + layer.2.v_cache 0.00001934 0.01142393 + layer.3.k_cache 0.04694761 5.05674866 + layer.3.v_cache 0.00001910 0.01334320 + layer.4.k_cache 0.00062564 0.24545507 + layer.4.v_cache 0.00005323 0.02551536 + layer.4.output 0.17166949 621.27191092 + ------------------------------------------------------------------------------------- + TOTAL 0.08235678 260.91377045 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 92936 +BPFP 0.9818 bits/point +EBPFP 1.9637 equivalent bits/point +MSE 260.913770 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 260.9138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,996B, BPFP=0.5640 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,088B, BPFP=2.0873 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,192B, BPFP=0.9774 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,828B, BPFP=2.0384 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,552B, BPFP=1.0452 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,332B, BPFP=1.9450 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,240B, BPFP=0.9864 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,528B, BPFP=1.9819 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,440B, BPFP=1.5889 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,976B, BPFP=1.8780 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,104B, BPFP=0.2717 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12335577 73.24427005 + layer.0.v_cache 0.00001400 0.01027178 + layer.1.k_cache 0.01519604 10.69812453 + layer.1.v_cache 0.00000547 0.00410855 + layer.2.k_cache 0.01098718 1.22171149 + layer.2.v_cache 0.00001779 0.01207639 + layer.3.k_cache 0.07416311 5.36753606 + layer.3.v_cache 0.00001944 0.01362900 + layer.4.k_cache 0.00061719 0.26448565 + layer.4.v_cache 0.00006444 0.02627857 + layer.4.output 0.17504946 652.49112522 + ------------------------------------------------------------------------------------- + TOTAL 0.08528157 274.01766874 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 90276 +BPFP 0.9997 bits/point +EBPFP 1.9994 equivalent bits/point +MSE 274.017669 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 274.0177 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,596B, BPFP=0.5268 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,612B, BPFP=2.1534 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,780B, BPFP=0.9700 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,372B, BPFP=2.1047 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,292B, BPFP=1.0739 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,752B, BPFP=1.9789 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,956B, BPFP=1.0057 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,952B, BPFP=2.0195 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,232B, BPFP=1.6705 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,744B, BPFP=1.9773 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,256B, BPFP=0.2973 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13955514 79.27835836 + layer.0.v_cache 0.00001392 0.01043010 + layer.1.k_cache 0.03454666 10.92203987 + layer.1.v_cache 0.00000526 0.00439090 + layer.2.k_cache 0.00236268 1.17347262 + layer.2.v_cache 0.00001660 0.01178843 + layer.3.k_cache 0.02874016 4.79308289 + layer.3.v_cache 0.00001899 0.01349417 + layer.4.k_cache 0.00062191 0.26718377 + layer.4.v_cache 0.00005265 0.02612470 + layer.4.output 0.18728454 701.86902829 + ------------------------------------------------------------------------------------- + TOTAL 0.08923092 294.68138611 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 86544 +BPFP 1.0330 bits/point +EBPFP 2.0661 equivalent bits/point +MSE 294.681386 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 294.6814 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,600B, BPFP=0.5276 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,684B, BPFP=2.1680 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,852B, BPFP=0.9846 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,560B, BPFP=2.1429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,404B, BPFP=1.0966 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,812B, BPFP=1.9911 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,028B, BPFP=1.0203 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,944B, BPFP=2.0179 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,360B, BPFP=1.6964 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,872B, BPFP=2.0032 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,616B, BPFP=0.3077 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11454884 76.26253678 + layer.0.v_cache 0.00001351 0.01084814 + layer.1.k_cache 0.01665594 11.53178584 + layer.1.v_cache 0.00000570 0.00471278 + layer.2.k_cache 0.00836552 1.12102717 + layer.2.v_cache 0.00001974 0.01282753 + layer.3.k_cache 0.03124422 5.05506679 + layer.3.v_cache 0.00001876 0.01377117 + layer.4.k_cache 0.00062231 0.26908151 + layer.4.v_cache 0.00005487 0.02779875 + layer.4.output 0.19663499 702.23915816 + ------------------------------------------------------------------------------------- + TOTAL 0.09105849 294.70491551 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 87732 +BPFP 1.0472 bits/point +EBPFP 2.0944 equivalent bits/point +MSE 294.704916 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 294.7049 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,040B, BPFP=0.5655 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,068B, BPFP=2.0588 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,024B, BPFP=0.9345 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,756B, BPFP=2.0007 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,448B, BPFP=1.0134 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,144B, BPFP=1.8869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,172B, BPFP=0.9621 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,324B, BPFP=1.9204 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,316B, BPFP=1.5469 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,796B, BPFP=1.8222 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,424B, BPFP=0.2770 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12290487 74.30310640 + layer.0.v_cache 0.00001422 0.01047849 + layer.1.k_cache 0.01283375 10.50112043 + layer.1.v_cache 0.00000579 0.00455139 + layer.2.k_cache 0.01100613 1.19695191 + layer.2.v_cache 0.00001929 0.01207180 + layer.3.k_cache 0.03011658 5.13405064 + layer.3.v_cache 0.00001858 0.01346555 + layer.4.k_cache 0.00063020 0.26985922 + layer.4.v_cache 0.00005223 0.02703229 + layer.4.output 0.17536174 644.01562500 + ------------------------------------------------------------------------------------- + TOTAL 0.08265493 270.56365077 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 89512 +BPFP 0.9794 bits/point +EBPFP 1.9589 equivalent bits/point +MSE 270.563651 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 270.5637 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,036B, BPFP=0.5581 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,072B, BPFP=2.0353 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,192B, BPFP=0.9544 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,936B, BPFP=2.0103 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,528B, BPFP=1.0162 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,400B, BPFP=1.9118 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,248B, BPFP=0.9647 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,528B, BPFP=1.9353 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,512B, BPFP=1.5647 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,016B, BPFP=1.8412 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,540B, BPFP=0.2768 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11218034 72.96616498 + layer.0.v_cache 0.00001438 0.01021136 + layer.1.k_cache 0.03157643 10.41336598 + layer.1.v_cache 0.00000545 0.00408499 + layer.2.k_cache 0.00802605 1.21281173 + layer.2.v_cache 0.00001826 0.01166747 + layer.3.k_cache 0.04519741 5.54530855 + layer.3.v_cache 0.00001794 0.01225135 + layer.4.k_cache 0.00061137 0.24677766 + layer.4.v_cache 0.00005913 0.02542556 + layer.4.output 0.17010492 636.43419118 + ------------------------------------------------------------------------------------- + TOTAL 0.08167301 267.38161223 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 91008 +BPFP 0.9841 bits/point +EBPFP 1.9682 equivalent bits/point +MSE 267.381612 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 267.3816 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.5596 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,272B, BPFP=2.0480 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,352B, BPFP=0.9724 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,036B, BPFP=2.0051 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,572B, BPFP=1.0124 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,452B, BPFP=1.8990 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,328B, BPFP=0.9680 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,572B, BPFP=1.9208 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,536B, BPFP=1.5509 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,244B, BPFP=1.8612 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,464B, BPFP=0.2716 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10146660 67.36658566 + layer.0.v_cache 0.00001344 0.01002489 + layer.1.k_cache 0.03090366 10.85856877 + layer.1.v_cache 0.00000566 0.00425233 + layer.2.k_cache 0.01188999 1.19969851 + layer.2.v_cache 0.00001845 0.01185328 + layer.3.k_cache 0.02542871 5.12924975 + layer.3.v_cache 0.00001884 0.01358300 + layer.4.k_cache 0.00060243 0.25732280 + layer.4.v_cache 0.00005224 0.02499539 + layer.4.output 0.16839122 629.25726744 + ------------------------------------------------------------------------------------- + TOTAL 0.07936109 264.09864744 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 91908 +BPFP 0.9823 bits/point +EBPFP 1.9645 equivalent bits/point +MSE 264.098647 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 264.0986 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,048B, BPFP=0.5177 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,808B, BPFP=2.0054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,580B, BPFP=0.9477 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,532B, BPFP=1.9586 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,080B, BPFP=1.0326 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,920B, BPFP=1.8546 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,756B, BPFP=0.9776 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,184B, BPFP=1.8995 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,268B, BPFP=1.5740 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,752B, BPFP=1.8261 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,028B, BPFP=0.2676 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12981709 71.86169964 + layer.0.v_cache 0.00001628 0.01038550 + layer.1.k_cache 0.06761242 10.63999209 + layer.1.v_cache 0.00000547 0.00390125 + layer.2.k_cache 0.00384689 1.14154020 + layer.2.v_cache 0.00001812 0.01143658 + layer.3.k_cache 0.05327560 5.04840585 + layer.3.v_cache 0.00001896 0.01311267 + layer.4.k_cache 0.00060800 0.24252203 + layer.4.v_cache 0.00005116 0.02405871 + layer.4.output 0.15518735 584.57181677 + ------------------------------------------------------------------------------------- + TOTAL 0.07891655 245.94116305 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 96956 +BPFP 0.9686 bits/point +EBPFP 1.9373 equivalent bits/point +MSE 245.941163 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 245.9412 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,912B, BPFP=0.5617 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,960B, BPFP=2.1142 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,952B, BPFP=0.9552 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,628B, BPFP=2.0502 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,452B, BPFP=1.0517 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,088B, BPFP=1.9460 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,108B, BPFP=0.9853 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,360B, BPFP=1.9985 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,516B, BPFP=1.6427 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,988B, BPFP=1.9267 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,948B, BPFP=0.2741 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887507 70.53656684 + layer.0.v_cache 0.00001345 0.00993242 + layer.1.k_cache 0.03259810 10.94092098 + layer.1.v_cache 0.00000523 0.00399397 + layer.2.k_cache 0.00244350 1.18353074 + layer.2.v_cache 0.00001834 0.01154766 + layer.3.k_cache 0.10864470 4.73607532 + layer.3.v_cache 0.00001904 0.01277657 + layer.4.k_cache 0.00060864 0.24665585 + layer.4.v_cache 0.00004940 0.02422456 + layer.4.output 0.18628448 668.09826940 + ------------------------------------------------------------------------------------- + TOTAL 0.09395687 280.25847710 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 88912 +BPFP 1.0089 bits/point +EBPFP 2.0178 equivalent bits/point +MSE 280.258477 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 280.2585 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,268B, BPFP=0.5737 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,544B, BPFP=2.0267 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,420B, BPFP=0.9515 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,224B, BPFP=1.9705 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,696B, BPFP=1.0000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,512B, BPFP=1.8455 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,536B, BPFP=0.9719 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,680B, BPFP=1.8750 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,676B, BPFP=1.5232 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,324B, BPFP=1.8125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,352B, BPFP=0.2847 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422350 72.88441450 + layer.0.v_cache 0.00001493 0.01014337 + layer.1.k_cache 0.01422887 10.60814195 + layer.1.v_cache 0.00000595 0.00409031 + layer.2.k_cache 0.00668612 1.09472245 + layer.2.v_cache 0.00001902 0.01134461 + layer.3.k_cache 0.02497406 5.10461769 + layer.3.v_cache 0.00001869 0.01237406 + layer.4.k_cache 0.00061935 0.23062991 + layer.4.v_cache 0.00005212 0.02424052 + layer.4.output 0.16899524 608.48129013 + ------------------------------------------------------------------------------------- + TOTAL 0.07904761 255.84433825 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 94232 +BPFP 0.9731 bits/point +EBPFP 1.9463 equivalent bits/point +MSE 255.844338 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 255.8443 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 113, 128) +Output shape: (1, 113, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.output: torch.Size([1, 113, 3584]) -> torch.Size([1, 1, 113, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,016B, BPFP=0.5553 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,452B, BPFP=1.8601 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,380B, BPFP=0.8822 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,104B, BPFP=1.8119 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,884B, BPFP=0.9519 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,512B, BPFP=1.7301 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,508B, BPFP=0.8999 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,636B, BPFP=1.7472 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,488B, BPFP=1.4502 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,184B, BPFP=1.6847 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,404B, BPFP=0.2450 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10214825 61.50871128 + layer.0.v_cache 0.00001637 0.00961079 + layer.1.k_cache 0.04284562 10.67884705 + layer.1.v_cache 0.00000578 0.00416664 + layer.2.k_cache 0.00913706 1.07394558 + layer.2.v_cache 0.00001928 0.01099926 + layer.3.k_cache 0.05585717 4.48752965 + layer.3.v_cache 0.00001870 0.01184410 + layer.4.k_cache 0.00065968 0.24626170 + layer.4.v_cache 0.00005753 0.02477300 + layer.4.output 10.09706179 473.08284608 + ------------------------------------------------------------------------------------- + TOTAL 4.17001165 199.39038892 + (elements=983,552) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 983552 +Total Bytes 110568 +BPFP 0.8993 bits/point +EBPFP 1.7987 equivalent bits/point +MSE 199.390389 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 199.3904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,212B, BPFP=0.5515 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,540B, BPFP=1.9815 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,392B, BPFP=0.9258 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,240B, BPFP=1.9299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,800B, BPFP=0.9959 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,712B, BPFP=1.8393 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,484B, BPFP=0.9416 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,900B, BPFP=1.8716 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,816B, BPFP=1.5137 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,484B, BPFP=1.8001 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,848B, BPFP=0.2906 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11481400 80.58937157 + layer.0.v_cache 0.00001396 0.01053425 + layer.1.k_cache 0.03135788 10.74486366 + layer.1.v_cache 0.00000545 0.00421230 + layer.2.k_cache 0.00370715 1.05151350 + layer.2.v_cache 0.00001965 0.01178748 + layer.3.k_cache 0.03930106 4.72910653 + layer.3.v_cache 0.00001886 0.01291309 + layer.4.k_cache 0.00060819 0.24093133 + layer.4.v_cache 0.00005437 0.02519986 + layer.4.output 0.16591480 593.13245683 + ------------------------------------------------------------------------------------- + TOTAL 0.07948848 249.96162537 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 95428 +BPFP 0.9638 bits/point +EBPFP 1.9277 equivalent bits/point +MSE 249.961625 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 249.9616 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,240B, BPFP=0.5688 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,260B, BPFP=1.9768 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,100B, BPFP=0.8954 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,860B, BPFP=1.9066 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,552B, BPFP=0.9747 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,424B, BPFP=1.8301 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,200B, BPFP=0.9129 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,392B, BPFP=1.8244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,720B, BPFP=1.5309 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,032B, BPFP=1.7612 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,388B, BPFP=0.2605 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09788724 71.86099895 + layer.0.v_cache 0.00001308 0.01008527 + layer.1.k_cache 0.03299916 10.93650286 + layer.1.v_cache 0.00000544 0.00423157 + layer.2.k_cache 0.00367894 1.20048403 + layer.2.v_cache 0.00002012 0.01240942 + layer.3.k_cache 0.03939044 4.81420693 + layer.3.v_cache 0.00001875 0.01266754 + layer.4.k_cache 0.00061788 0.25231115 + layer.4.v_cache 0.00005010 0.02504806 + layer.4.output 0.17021053 607.61110554 + ------------------------------------------------------------------------------------- + TOTAL 0.08036205 255.43568733 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 91168 +BPFP 0.9415 bits/point +EBPFP 1.8830 equivalent bits/point +MSE 255.435687 +---------------------- -------------------------------------------------------- +Time: 0.367s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 255.4357 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,616B, BPFP=0.5231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,992B, BPFP=1.8796 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,384B, BPFP=0.9236 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,712B, BPFP=1.8391 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,828B, BPFP=0.9878 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,212B, BPFP=1.7668 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,472B, BPFP=0.9363 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,408B, BPFP=1.7951 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,308B, BPFP=1.4913 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,980B, BPFP=1.7332 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,052B, BPFP=0.3111 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12672405 72.45889395 + layer.0.v_cache 0.00001401 0.00983773 + layer.1.k_cache 0.12044146 10.84533691 + layer.1.v_cache 0.00000568 0.00408063 + layer.2.k_cache 0.00570256 1.14076402 + layer.2.v_cache 0.00001866 0.01110840 + layer.3.k_cache 0.03611955 4.39423144 + layer.3.v_cache 0.00001983 0.01301958 + layer.4.k_cache 0.00066233 0.24231188 + layer.4.v_cache 0.00005331 0.02432805 + layer.4.output 10.56037249 494.19080688 + ------------------------------------------------------------------------------------- + TOTAL 4.36543346 208.73409181 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 110964 +BPFP 0.9443 bits/point +EBPFP 1.8887 equivalent bits/point +MSE 208.734092 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.006s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 208.7341 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,120B, BPFP=0.5540 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,396B, BPFP=2.0234 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,360B, BPFP=0.9517 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,888B, BPFP=1.9332 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,516B, BPFP=0.9794 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,328B, BPFP=1.8338 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,156B, BPFP=0.9155 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,568B, BPFP=1.8764 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,576B, BPFP=1.5227 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,120B, BPFP=1.7969 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,424B, BPFP=0.2644 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13733016 65.94775391 + layer.0.v_cache 0.00001373 0.01074757 + layer.1.k_cache 0.03155638 10.29112660 + layer.1.v_cache 0.00000617 0.00489951 + layer.2.k_cache 0.00751131 1.11551684 + layer.2.v_cache 0.00001754 0.01288261 + layer.3.k_cache 0.02548880 4.78225500 + layer.3.v_cache 0.00001765 0.01408641 + layer.4.k_cache 0.00061011 0.26156161 + layer.4.v_cache 0.00005045 0.02771143 + layer.4.output 0.16239834 614.09801136 + ------------------------------------------------------------------------------------- + TOTAL 0.07878769 257.71497771 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 91452 +BPFP 0.9552 bits/point +EBPFP 1.9103 equivalent bits/point +MSE 257.714978 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 257.7150 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,808B, BPFP=0.5554 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,580B, BPFP=2.0926 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,668B, BPFP=0.9233 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,304B, BPFP=2.0380 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,356B, BPFP=1.0593 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,832B, BPFP=1.9446 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,088B, BPFP=1.0063 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,808B, BPFP=1.9399 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,120B, BPFP=1.6060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,568B, BPFP=1.8924 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,416B, BPFP=0.2943 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10979201 79.20342168 + layer.0.v_cache 0.00001339 0.00994797 + layer.1.k_cache 0.01160004 11.14826135 + layer.1.v_cache 0.00000567 0.00433640 + layer.2.k_cache 0.00714037 1.24956464 + layer.2.v_cache 0.00001829 0.01130042 + layer.3.k_cache 0.04622446 5.00426860 + layer.3.v_cache 0.00001896 0.01266669 + layer.4.k_cache 0.00064378 0.25492521 + layer.4.v_cache 0.00005251 0.02522533 + layer.4.output 0.18415536 684.13991863 + ------------------------------------------------------------------------------------- + TOTAL 0.08615276 287.40607933 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 86548 +BPFP 1.0069 bits/point +EBPFP 2.0139 equivalent bits/point +MSE 287.406079 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 287.4061 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,612B, BPFP=0.5300 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,648B, BPFP=2.1607 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,828B, BPFP=0.9797 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,208B, BPFP=2.0714 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,412B, BPFP=1.0982 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,740B, BPFP=1.9765 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,996B, BPFP=1.0138 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,824B, BPFP=1.9935 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,192B, BPFP=1.6623 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,716B, BPFP=1.9716 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,052B, BPFP=0.3204 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11290005 77.38562804 + layer.0.v_cache 0.00001570 0.01045011 + layer.1.k_cache 0.03434501 11.31516969 + layer.1.v_cache 0.00000534 0.00391260 + layer.2.k_cache 0.00238675 1.19504141 + layer.2.v_cache 0.00001838 0.01142208 + layer.3.k_cache 0.02838124 4.98406982 + layer.3.v_cache 0.00002047 0.01276805 + layer.4.k_cache 0.00062757 0.25446592 + layer.4.v_cache 0.00005455 0.02472865 + layer.4.output 0.18420308 702.41964286 + ------------------------------------------------------------------------------------- + TOTAL 0.08636333 294.83147979 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 87228 +BPFP 1.0412 bits/point +EBPFP 2.0824 equivalent bits/point +MSE 294.831480 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 294.8315 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,588B, BPFP=0.5184 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,648B, BPFP=2.1330 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,764B, BPFP=0.9543 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,544B, BPFP=2.1122 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,296B, BPFP=1.0609 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,932B, BPFP=1.9896 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,072B, BPFP=1.0160 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,060B, BPFP=2.0152 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,096B, BPFP=1.6218 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,600B, BPFP=1.9231 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,828B, BPFP=0.3099 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09752406 71.19022937 + layer.0.v_cache 0.00001393 0.01057982 + layer.1.k_cache 0.01493649 11.21379676 + layer.1.v_cache 0.00000588 0.00428412 + layer.2.k_cache 0.01003601 1.22568326 + layer.2.v_cache 0.00001805 0.01169575 + layer.3.k_cache 0.04451986 5.31915948 + layer.3.v_cache 0.00001914 0.01368344 + layer.4.k_cache 0.00061869 0.26778935 + layer.4.v_cache 0.00005251 0.02683204 + layer.4.output 0.18409089 692.96468636 + ------------------------------------------------------------------------------------- + TOTAL 0.08566946 290.59038458 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 87428 +BPFP 1.0302 bits/point +EBPFP 2.0604 equivalent bits/point +MSE 290.590385 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 290.5904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,936B, BPFP=0.5527 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,876B, BPFP=2.0474 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,148B, BPFP=0.9691 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,580B, BPFP=1.9917 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,360B, BPFP=1.0090 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,988B, BPFP=1.8803 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,156B, BPFP=0.9706 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,020B, BPFP=1.8863 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,396B, BPFP=1.5806 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,728B, BPFP=1.8313 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,556B, BPFP=0.2839 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13658243 71.32656015 + layer.0.v_cache 0.00001379 0.01031279 + layer.1.k_cache 0.01234693 10.67431788 + layer.1.v_cache 0.00000566 0.00423358 + layer.2.k_cache 0.00375735 1.25691269 + layer.2.v_cache 0.00001852 0.01123444 + layer.3.k_cache 0.02700986 5.21172728 + layer.3.v_cache 0.00001930 0.01245541 + layer.4.k_cache 0.00062007 0.26044944 + layer.4.v_cache 0.00005146 0.02558632 + layer.4.output 0.17614663 652.08708046 + ------------------------------------------------------------------------------------- + TOTAL 0.08314423 273.72960901 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 88744 +BPFP 0.9827 bits/point +EBPFP 1.9655 equivalent bits/point +MSE 273.729609 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 273.7296 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,396B, BPFP=0.5128 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,232B, BPFP=2.1901 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,664B, BPFP=0.9983 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,772B, BPFP=2.0916 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,120B, BPFP=1.0959 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,476B, BPFP=2.0283 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,888B, BPFP=1.0462 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,468B, BPFP=2.0265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,152B, BPFP=1.7449 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,644B, BPFP=2.0642 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,664B, BPFP=0.3261 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08738674 72.79415935 + layer.0.v_cache 0.00001368 0.01065412 + layer.1.k_cache 0.01286346 10.93382953 + layer.1.v_cache 0.00000597 0.00449130 + layer.2.k_cache 0.00251921 1.22021568 + layer.2.v_cache 0.00001769 0.01301624 + layer.3.k_cache 0.08324167 5.31348242 + layer.3.v_cache 0.00002021 0.01422206 + layer.4.k_cache 0.00062021 0.26977649 + layer.4.v_cache 0.00005375 0.02720775 + layer.4.output 0.19567458 742.22908513 + ------------------------------------------------------------------------------------- + TOTAL 0.09155674 310.95321475 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 84476 +BPFP 1.0636 bits/point +EBPFP 2.1272 equivalent bits/point +MSE 310.953215 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 310.9532 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,004B, BPFP=0.5655 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,916B, BPFP=2.0550 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,228B, BPFP=0.9842 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,644B, BPFP=2.0038 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,440B, BPFP=1.0241 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,252B, BPFP=1.9300 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,096B, BPFP=0.9593 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,164B, BPFP=1.9134 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,332B, BPFP=1.5685 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,896B, BPFP=1.8630 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,840B, BPFP=0.2915 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12176956 72.33878483 + layer.0.v_cache 0.00001396 0.00973331 + layer.1.k_cache 0.01120206 10.41416839 + layer.1.v_cache 0.00000562 0.00411845 + layer.2.k_cache 0.00678912 1.22869340 + layer.2.v_cache 0.00001849 0.01114606 + layer.3.k_cache 0.02622183 5.04452882 + layer.3.v_cache 0.00001965 0.01313522 + layer.4.k_cache 0.00062413 0.25063177 + layer.4.v_cache 0.00005248 0.02488254 + layer.4.output 0.17940697 651.95896084 + ------------------------------------------------------------------------------------- + TOTAL 0.08368034 273.70897345 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 89812 +BPFP 0.9946 bits/point +EBPFP 1.9891 equivalent bits/point +MSE 273.708973 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 273.7090 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,496B, BPFP=0.5202 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,840B, BPFP=1.9107 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,972B, BPFP=0.8887 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,460B, BPFP=1.8542 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,484B, BPFP=0.9649 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,012B, BPFP=1.7875 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,224B, BPFP=0.9262 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,220B, BPFP=1.8185 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,188B, BPFP=1.5161 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,764B, BPFP=1.7506 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,832B, BPFP=0.2940 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120816 70.39684245 + layer.0.v_cache 0.00001607 0.01016520 + layer.1.k_cache 0.10678518 10.54884208 + layer.1.v_cache 0.00000585 0.00395480 + layer.2.k_cache 0.00813519 1.10047331 + layer.2.v_cache 0.00001919 0.01125300 + layer.3.k_cache 0.06560528 4.58614211 + layer.3.v_cache 0.00001973 0.01338767 + layer.4.k_cache 0.00062300 0.24061287 + layer.4.v_cache 0.00005287 0.02415123 + layer.4.output 10.86395687 510.47465986 + ------------------------------------------------------------------------------------- + TOTAL 4.49059815 215.30932022 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 107492 +BPFP 0.9409 bits/point +EBPFP 1.8819 equivalent bits/point +MSE 215.309320 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 215.3093 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,020B, BPFP=0.5424 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,156B, BPFP=2.0036 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,208B, BPFP=0.9353 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,728B, BPFP=1.9267 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,528B, BPFP=0.9928 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,352B, BPFP=1.8592 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,124B, BPFP=0.9203 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,548B, BPFP=1.8944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,436B, BPFP=1.5151 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,032B, BPFP=1.8017 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,864B, BPFP=0.2787 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12848519 67.25300826 + layer.0.v_cache 0.00001371 0.00965521 + layer.1.k_cache 0.03007136 10.57498064 + layer.1.v_cache 0.00000570 0.00433004 + layer.2.k_cache 0.00936044 1.17798334 + layer.2.v_cache 0.00001842 0.01167516 + layer.3.k_cache 0.04065019 5.10397444 + layer.3.v_cache 0.00001877 0.01280954 + layer.4.k_cache 0.00062204 0.25429697 + layer.4.v_cache 0.00010837 0.02610898 + layer.4.output 0.16426703 620.67220854 + ------------------------------------------------------------------------------------- + TOTAL 0.07995432 260.53731073 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 90996 +BPFP 0.9613 bits/point +EBPFP 1.9227 equivalent bits/point +MSE 260.537311 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 260.5373 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,788B, BPFP=0.5445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,968B, BPFP=2.1422 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,800B, BPFP=0.9375 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,740B, BPFP=2.0977 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,700B, BPFP=1.1133 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,104B, BPFP=1.9734 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,368B, BPFP=1.0484 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,388B, BPFP=2.0289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,420B, BPFP=1.6445 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,860B, BPFP=1.9258 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,716B, BPFP=0.2990 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09257430 80.86270142 + layer.0.v_cache 0.00001329 0.00953926 + layer.1.k_cache 0.01202901 10.83438644 + layer.1.v_cache 0.00000544 0.00371250 + layer.2.k_cache 0.00549439 1.11923733 + layer.2.v_cache 0.00001849 0.01097208 + layer.3.k_cache 0.11014032 5.19002838 + layer.3.v_cache 0.00001916 0.01175387 + layer.4.k_cache 0.00063444 0.23479319 + layer.4.v_cache 0.00005494 0.02264440 + layer.4.output 0.19143520 675.65435268 + ------------------------------------------------------------------------------------- + TOTAL 0.09182531 283.99295515 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 89852 +BPFP 1.0323 bits/point +EBPFP 2.0646 equivalent bits/point +MSE 283.992955 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 283.9930 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,908B, BPFP=0.5409 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,112B, BPFP=2.0670 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,160B, BPFP=0.9598 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,712B, BPFP=1.9926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,460B, BPFP=1.0156 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,156B, BPFP=1.8891 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,180B, BPFP=0.9635 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,308B, BPFP=1.9174 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,412B, BPFP=1.5647 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,812B, BPFP=1.8251 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,968B, BPFP=0.2649 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12365316 74.85430618 + layer.0.v_cache 0.00001332 0.01072918 + layer.1.k_cache 0.03511774 10.46504139 + layer.1.v_cache 0.00000569 0.00471833 + layer.2.k_cache 0.01236608 1.15079934 + layer.2.v_cache 0.00001835 0.01171899 + layer.3.k_cache 0.02851222 5.23856281 + layer.3.v_cache 0.00001844 0.01412247 + layer.4.k_cache 0.00063207 0.26254668 + layer.4.v_cache 0.00005066 0.02707735 + layer.4.output 0.17964420 644.44855442 + ------------------------------------------------------------------------------------- + TOTAL 0.08575865 270.77526492 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 89188 +BPFP 0.9759 bits/point +EBPFP 1.9518 equivalent bits/point +MSE 270.775265 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 270.7753 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,552B, BPFP=0.5112 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,656B, BPFP=2.1346 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,816B, BPFP=0.9647 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,444B, BPFP=2.0921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,300B, BPFP=1.0617 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,832B, BPFP=1.9696 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,928B, BPFP=0.9872 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,068B, BPFP=2.0168 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,072B, BPFP=1.6170 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,668B, BPFP=1.9367 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,772B, BPFP=0.3083 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12096892 76.69052985 + layer.0.v_cache 0.00001393 0.01035575 + layer.1.k_cache 0.03247837 11.31262364 + layer.1.v_cache 0.00000613 0.00427462 + layer.2.k_cache 0.00394396 1.13335732 + layer.2.v_cache 0.00001920 0.01179562 + layer.3.k_cache 0.02751121 4.71729963 + layer.3.v_cache 0.00001919 0.01284008 + layer.4.k_cache 0.00062481 0.25853059 + layer.4.v_cache 0.00005293 0.02483192 + layer.4.output 0.18564256 693.43240614 + ------------------------------------------------------------------------------------- + TOTAL 0.08736098 291.07078129 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 87108 +BPFP 1.0264 bits/point +EBPFP 2.0529 equivalent bits/point +MSE 291.070781 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 291.0708 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,576B, BPFP=0.5227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,656B, BPFP=2.1623 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,772B, BPFP=0.9683 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,296B, BPFP=2.0893 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,392B, BPFP=1.0942 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,780B, BPFP=1.9846 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,996B, BPFP=1.0138 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,888B, BPFP=2.0065 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,292B, BPFP=1.6826 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,700B, BPFP=1.9683 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,084B, BPFP=0.3213 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08564486 75.37629363 + layer.0.v_cache 0.00001408 0.01080216 + layer.1.k_cache 0.03507188 11.40870884 + layer.1.v_cache 0.00000566 0.00433998 + layer.2.k_cache 0.00239554 1.23212700 + layer.2.v_cache 0.00001938 0.01248949 + layer.3.k_cache 0.03313774 4.29151144 + layer.3.v_cache 0.00001993 0.01425563 + layer.4.k_cache 0.00061714 0.26167228 + layer.4.v_cache 0.00005134 0.02676343 + layer.4.output 0.18226704 702.26785714 + ------------------------------------------------------------------------------------- + TOTAL 0.08428511 294.61846846 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 87432 +BPFP 1.0436 bits/point +EBPFP 2.0873 equivalent bits/point +MSE 294.618468 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 294.6185 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 117, 128) +Output shape: (1, 117, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.output: torch.Size([1, 117, 3584]) -> torch.Size([1, 1, 117, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,132B, BPFP=0.5518 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,748B, BPFP=1.8360 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,484B, BPFP=0.8659 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,472B, BPFP=1.7991 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,080B, BPFP=0.9455 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,948B, BPFP=1.7292 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,772B, BPFP=0.9044 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,112B, BPFP=1.7511 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,904B, BPFP=1.4562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,544B, BPFP=1.6752 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,224B, BPFP=0.2904 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11900647 62.89291617 + layer.0.v_cache 0.00001428 0.01006786 + layer.1.k_cache 0.12535267 10.14626945 + layer.1.v_cache 0.00000579 0.00406188 + layer.2.k_cache 0.00517510 0.96112334 + layer.2.v_cache 0.00001992 0.01170076 + layer.3.k_cache 0.09315677 4.45201802 + layer.3.v_cache 0.00001978 0.01262068 + layer.4.k_cache 0.00061896 0.23212808 + layer.4.v_cache 0.00005133 0.02331934 + layer.4.output 9.75369281 457.50633394 + ------------------------------------------------------------------------------------- + TOTAL 4.03642769 193.01709195 + (elements=1,018,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1018368 +Total Bytes 116420 +BPFP 0.9146 bits/point +EBPFP 1.8291 equivalent bits/point +MSE 193.017092 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.006s, Pack+Encode: 0.159s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 193.0171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,468B, BPFP=0.5161 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,940B, BPFP=1.9256 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,084B, BPFP=0.9054 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,732B, BPFP=1.8946 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,588B, BPFP=0.9804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,176B, BPFP=1.8119 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,392B, BPFP=0.9512 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,220B, BPFP=1.8185 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,196B, BPFP=1.5173 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,844B, BPFP=1.7625 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,644B, BPFP=0.2475 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14131550 70.36934059 + layer.0.v_cache 0.00001383 0.00991846 + layer.1.k_cache 0.02978072 10.42615444 + layer.1.v_cache 0.00000537 0.00402094 + layer.2.k_cache 0.00578868 1.05911298 + layer.2.v_cache 0.00001872 0.01121115 + layer.3.k_cache 0.04020810 4.64574149 + layer.3.v_cache 0.00001850 0.01266626 + layer.4.k_cache 0.00063358 0.24363414 + layer.4.v_cache 0.00005294 0.02657299 + layer.4.output 10.86925024 510.39417517 + ------------------------------------------------------------------------------------- + TOTAL 4.48838751 215.26868233 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 106284 +BPFP 0.9304 bits/point +EBPFP 1.8607 equivalent bits/point +MSE 215.268682 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 215.2687 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,052B, BPFP=0.5677 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,184B, BPFP=2.0804 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,064B, BPFP=0.9420 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,832B, BPFP=2.0149 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,584B, BPFP=1.0387 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,420B, BPFP=1.9382 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,260B, BPFP=0.9784 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,560B, BPFP=1.9643 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,464B, BPFP=1.5744 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,920B, BPFP=1.8452 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,100B, BPFP=0.2684 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09514355 72.38376581 + layer.0.v_cache 0.00001423 0.01015043 + layer.1.k_cache 0.03283414 10.56422715 + layer.1.v_cache 0.00000553 0.00384832 + layer.2.k_cache 0.00489674 1.26239450 + layer.2.v_cache 0.00001883 0.01134047 + layer.3.k_cache 0.07199075 5.17096965 + layer.3.v_cache 0.00001932 0.01260508 + layer.4.k_cache 0.00061592 0.25076494 + layer.4.v_cache 0.00006816 0.02405121 + layer.4.output 0.17253765 644.43728741 + ------------------------------------------------------------------------------------- + TOTAL 0.08313946 270.63265467 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 90440 +BPFP 0.9896 bits/point +EBPFP 1.9792 equivalent bits/point +MSE 270.632655 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.003s, Pack+Encode: 0.160s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 270.6327 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,084B, BPFP=0.5539 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,232B, BPFP=2.0172 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,248B, BPFP=0.9425 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,924B, BPFP=1.9619 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,540B, BPFP=0.9950 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,376B, BPFP=1.8635 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,204B, BPFP=0.9346 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,472B, BPFP=1.8807 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,612B, BPFP=1.5467 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,204B, BPFP=1.8326 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,048B, BPFP=0.2578 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11660385 67.57907911 + layer.0.v_cache 0.00001381 0.01041977 + layer.1.k_cache 0.05175667 10.35925714 + layer.1.v_cache 0.00000569 0.00458149 + layer.2.k_cache 0.00227456 1.12819926 + layer.2.v_cache 0.00001743 0.01184674 + layer.3.k_cache 0.02781237 4.84459432 + layer.3.v_cache 0.00001883 0.01302027 + layer.4.k_cache 0.00061846 0.25734864 + layer.4.v_cache 0.00005344 0.02613610 + layer.4.output 0.17410791 621.12936166 + ------------------------------------------------------------------------------------- + TOTAL 0.08340767 260.71411850 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 90944 +BPFP 0.9608 bits/point +EBPFP 1.9216 equivalent bits/point +MSE 260.714118 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 260.7141 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,976B, BPFP=0.5602 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,952B, BPFP=2.0617 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,076B, BPFP=0.9556 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,584B, BPFP=1.9925 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,444B, BPFP=1.0248 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,056B, BPFP=1.8931 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,124B, BPFP=0.9646 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,256B, BPFP=1.9307 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,424B, BPFP=1.5858 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,864B, BPFP=1.8569 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,396B, BPFP=0.2796 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212080 74.42036898 + layer.0.v_cache 0.00001360 0.00985906 + layer.1.k_cache 0.03288655 10.38728351 + layer.1.v_cache 0.00000536 0.00421215 + layer.2.k_cache 0.00973156 1.24875898 + layer.2.v_cache 0.00001762 0.01097798 + layer.3.k_cache 0.02906220 5.65260922 + layer.3.v_cache 0.00001891 0.01262755 + layer.4.k_cache 0.00063934 0.25428547 + layer.4.v_cache 0.00004866 0.02550364 + layer.4.output 0.18425958 652.04894578 + ------------------------------------------------------------------------------------- + TOTAL 0.08496245 273.90406512 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 89152 +BPFP 0.9872 bits/point +EBPFP 1.9745 equivalent bits/point +MSE 273.904065 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 273.9041 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,160B, BPFP=0.5809 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,312B, BPFP=2.0794 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,332B, BPFP=0.9801 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,884B, BPFP=2.0007 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,512B, BPFP=1.0132 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,500B, BPFP=1.9301 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,380B, BPFP=0.9890 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,604B, BPFP=1.9493 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,580B, BPFP=1.5772 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,088B, BPFP=1.8544 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,268B, BPFP=0.2696 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10758395 69.99701287 + layer.0.v_cache 0.00001397 0.00993985 + layer.1.k_cache 0.03293324 10.49651453 + layer.1.v_cache 0.00000575 0.00394966 + layer.2.k_cache 0.00823421 1.17384850 + layer.2.v_cache 0.00001745 0.01092023 + layer.3.k_cache 0.04116619 5.46593377 + layer.3.v_cache 0.00001985 0.01339789 + layer.4.k_cache 0.00062779 0.24707518 + layer.4.v_cache 0.00005042 0.02442561 + layer.4.output 0.16954060 636.55000000 + ------------------------------------------------------------------------------------- + TOTAL 0.08102571 267.25253048 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 91620 +BPFP 0.9907 bits/point +EBPFP 1.9814 equivalent bits/point +MSE 267.252530 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 267.2525 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,084B, BPFP=0.5354 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,548B, BPFP=2.0049 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,248B, BPFP=0.9111 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,188B, BPFP=1.9424 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,736B, BPFP=0.9958 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,676B, BPFP=1.8535 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,496B, BPFP=0.9542 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,688B, BPFP=1.8556 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,880B, BPFP=1.5417 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,436B, BPFP=1.8118 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,772B, BPFP=0.2672 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12307767 77.26386719 + layer.0.v_cache 0.00001410 0.01051193 + layer.1.k_cache 0.03088827 10.47891100 + layer.1.v_cache 0.00000584 0.00438827 + layer.2.k_cache 0.00371391 1.05564185 + layer.2.v_cache 0.00001950 0.01231731 + layer.3.k_cache 0.03079298 5.09982503 + layer.3.v_cache 0.00001950 0.01367217 + layer.4.k_cache 0.00063014 0.25329789 + layer.4.v_cache 0.00005282 0.02649574 + layer.4.output 0.15767360 600.82351190 + ------------------------------------------------------------------------------------- + TOTAL 0.07605470 252.94020657 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 93752 +BPFP 0.9574 bits/point +EBPFP 1.9149 equivalent bits/point +MSE 252.940207 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 252.9402 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,984B, BPFP=0.5686 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,832B, BPFP=2.0640 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,968B, BPFP=0.9466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,520B, BPFP=2.0046 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,504B, BPFP=1.0488 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,000B, BPFP=1.9055 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,228B, BPFP=0.9962 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,084B, BPFP=1.9215 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,388B, BPFP=1.5983 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,872B, BPFP=1.8811 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,664B, BPFP=0.2903 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10295349 70.83223371 + layer.0.v_cache 0.00001348 0.00961117 + layer.1.k_cache 0.03334363 10.69707508 + layer.1.v_cache 0.00000550 0.00419049 + layer.2.k_cache 0.01126997 1.20857890 + layer.2.v_cache 0.00001933 0.01128694 + layer.3.k_cache 0.09131574 5.29051916 + layer.3.v_cache 0.00001769 0.01206468 + layer.4.k_cache 0.00062275 0.26177695 + layer.4.v_cache 0.00006541 0.02403571 + layer.4.output 0.17512874 660.59826873 + ------------------------------------------------------------------------------------- + TOTAL 0.08620754 277.20819140 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 89044 +BPFP 0.9981 bits/point +EBPFP 1.9961 equivalent bits/point +MSE 277.208191 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 277.2082 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,212B, BPFP=0.5769 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,328B, BPFP=2.0345 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,196B, BPFP=0.9332 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,148B, BPFP=2.0022 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,548B, BPFP=0.9964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,560B, BPFP=1.8966 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,248B, BPFP=0.9425 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,752B, BPFP=1.9310 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,504B, BPFP=1.5273 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,124B, BPFP=1.8182 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,164B, BPFP=0.2608 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12314748 66.59341325 + layer.0.v_cache 0.00001636 0.01005284 + layer.1.k_cache 0.01190477 10.48508848 + layer.1.v_cache 0.00000588 0.00413182 + layer.2.k_cache 0.01614088 1.12212696 + layer.2.v_cache 0.00001834 0.01107147 + layer.3.k_cache 0.05604688 5.17537435 + layer.3.v_cache 0.00001882 0.01258880 + layer.4.k_cache 0.00063774 0.24701564 + layer.4.v_cache 0.00005162 0.02433528 + layer.4.output 0.17348555 618.54987685 + ------------------------------------------------------------------------------------- + TOTAL 0.08366986 259.61966687 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 91784 +BPFP 0.9697 bits/point +EBPFP 1.9393 equivalent bits/point +MSE 259.619667 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 259.6197 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,876B, BPFP=0.5548 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,820B, BPFP=2.0872 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,956B, BPFP=0.9560 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,508B, BPFP=2.0270 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,572B, BPFP=1.0748 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,968B, BPFP=1.9228 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,112B, BPFP=0.9861 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,092B, BPFP=1.9468 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,308B, BPFP=1.6026 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,736B, BPFP=1.8781 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,636B, BPFP=0.2931 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11560643 70.32927035 + layer.0.v_cache 0.00001402 0.01034254 + layer.1.k_cache 0.03272506 11.15553265 + layer.1.v_cache 0.00000563 0.00446451 + layer.2.k_cache 0.00645547 1.21739253 + layer.2.v_cache 0.00001830 0.01207868 + layer.3.k_cache 0.02731901 4.60137563 + layer.3.v_cache 0.00001960 0.01305238 + layer.4.k_cache 0.00061302 0.26154017 + layer.4.v_cache 0.00004978 0.02474644 + layer.4.output 0.17251868 668.39737654 + ------------------------------------------------------------------------------------- + TOTAL 0.08179159 280.37714304 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 88584 +BPFP 1.0052 bits/point +EBPFP 2.0103 equivalent bits/point +MSE 280.377143 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 280.3771 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,160B, BPFP=0.5611 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,348B, BPFP=2.0149 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,312B, BPFP=0.9432 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,980B, BPFP=1.9496 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,568B, BPFP=0.9886 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,528B, BPFP=1.8693 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,312B, BPFP=0.9432 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,704B, BPFP=1.9006 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,592B, BPFP=1.5256 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,208B, BPFP=1.8125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,484B, BPFP=0.2913 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11690623 65.55060924 + layer.0.v_cache 0.00001418 0.00994352 + layer.1.k_cache 0.03305150 10.82234885 + layer.1.v_cache 0.00000545 0.00410642 + layer.2.k_cache 0.01095151 1.24788874 + layer.2.v_cache 0.00001854 0.01109079 + layer.3.k_cache 0.04293454 4.67660800 + layer.3.v_cache 0.00002007 0.01277300 + layer.4.k_cache 0.00062489 0.24241460 + layer.4.v_cache 0.00005451 0.02422743 + layer.4.output 0.16672201 614.06102881 + ------------------------------------------------------------------------------------- + TOTAL 0.08068444 257.70760072 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 93196 +BPFP 0.9734 bits/point +EBPFP 1.9468 equivalent bits/point +MSE 257.707601 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 257.7076 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,212B, BPFP=0.5515 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,508B, BPFP=1.9760 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,420B, BPFP=0.9306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,160B, BPFP=1.9162 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,744B, BPFP=0.9863 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,744B, BPFP=1.8448 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,444B, BPFP=0.9348 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,628B, BPFP=1.8249 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,848B, BPFP=1.5192 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,244B, BPFP=1.7589 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,760B, BPFP=0.2639 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11067300 79.00069754 + layer.0.v_cache 0.00001409 0.01060376 + layer.1.k_cache 0.05152405 10.54020406 + layer.1.v_cache 0.00000569 0.00431574 + layer.2.k_cache 0.01362713 1.15403789 + layer.2.v_cache 0.00001889 0.01210499 + layer.3.k_cache 0.08227781 4.86509487 + layer.3.v_cache 0.00001889 0.01297901 + layer.4.k_cache 0.00062414 0.24971012 + layer.4.v_cache 0.00005257 0.02437305 + layer.4.output 0.15746453 592.71065542 + ------------------------------------------------------------------------------------- + TOTAL 0.08006400 249.69698288 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 93712 +BPFP 0.9465 bits/point +EBPFP 1.8930 equivalent bits/point +MSE 249.696983 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 249.6970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 116, 128) +Output shape: (1, 116, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.output: torch.Size([1, 116, 3584]) -> torch.Size([1, 1, 116, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,964B, BPFP=0.5339 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,560B, BPFP=1.8265 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,412B, BPFP=0.8637 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,292B, BPFP=1.7904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,032B, BPFP=0.9472 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,832B, BPFP=1.7284 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,544B, BPFP=0.8815 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,884B, BPFP=1.7355 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,720B, BPFP=1.4440 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,396B, BPFP=1.6697 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,284B, BPFP=0.2749 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16314862 61.85399970 + layer.0.v_cache 0.00001373 0.00997331 + layer.1.k_cache 0.05481785 10.04653615 + layer.1.v_cache 0.00000557 0.00427742 + layer.2.k_cache 0.00968372 1.06316876 + layer.2.v_cache 0.00001954 0.01207282 + layer.3.k_cache 0.02418142 4.21493004 + layer.3.v_cache 0.00001860 0.01265905 + layer.4.k_cache 0.00062333 0.24825159 + layer.4.v_cache 0.00005467 0.02577897 + layer.4.output 9.83417319 461.63462131 + ------------------------------------------------------------------------------------- + TOTAL 4.06422232 194.64317629 + (elements=1,009,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1009664 +Total Bytes 113920 +BPFP 0.9026 bits/point +EBPFP 1.8053 equivalent bits/point +MSE 194.643176 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 194.6432 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,568B, BPFP=0.5144 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,604B, BPFP=2.1242 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,784B, BPFP=0.9583 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,288B, BPFP=2.0609 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,212B, BPFP=1.0441 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,664B, BPFP=1.9359 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,036B, BPFP=1.0088 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,832B, BPFP=1.9696 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,084B, BPFP=1.6194 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,488B, BPFP=1.9006 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,968B, BPFP=0.3139 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11749315 75.89580829 + layer.0.v_cache 0.00001389 0.01006500 + layer.1.k_cache 0.03399578 11.29842748 + layer.1.v_cache 0.00000555 0.00416232 + layer.2.k_cache 0.00854440 1.23401485 + layer.2.v_cache 0.00001777 0.01143770 + layer.3.k_cache 0.02740619 5.09333214 + layer.3.v_cache 0.00001935 0.01338806 + layer.4.k_cache 0.00063531 0.25924881 + layer.4.v_cache 0.00005166 0.02478155 + layer.4.output 0.18490290 692.81490385 + ------------------------------------------------------------------------------------- + TOTAL 0.08720608 290.79699960 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 86528 +BPFP 1.0196 bits/point +EBPFP 2.0392 equivalent bits/point +MSE 290.797000 +---------------------- -------------------------------------------------------- +Time: 0.367s Load: 0.003s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 290.7970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,180B, BPFP=0.5286 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,724B, BPFP=1.9488 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,356B, BPFP=0.8903 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,464B, BPFP=1.9056 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,884B, BPFP=0.9781 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,736B, BPFP=1.7846 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,652B, BPFP=0.9395 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,868B, BPFP=1.8065 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,100B, BPFP=1.5126 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,720B, BPFP=1.7819 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,284B, BPFP=0.2680 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11492289 73.87249626 + layer.0.v_cache 0.00001440 0.01010330 + layer.1.k_cache 0.01112092 10.65631818 + layer.1.v_cache 0.00000555 0.00391774 + layer.2.k_cache 0.01096075 1.08618010 + layer.2.v_cache 0.00001882 0.01155592 + layer.3.k_cache 0.03781782 4.60052847 + layer.3.v_cache 0.00001905 0.01243742 + layer.4.k_cache 0.00062716 0.23811551 + layer.4.v_cache 0.00006343 0.02332119 + layer.4.output 0.15224552 575.98988412 + ------------------------------------------------------------------------------------- + TOTAL 0.07301703 242.49671547 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 95968 +BPFP 0.9384 bits/point +EBPFP 1.8767 equivalent bits/point +MSE 242.496715 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 242.4967 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,564B, BPFP=0.5203 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,608B, BPFP=2.1526 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,812B, BPFP=0.9765 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,324B, BPFP=2.0950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,408B, BPFP=1.0974 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,768B, BPFP=1.9821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,088B, BPFP=1.0325 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,968B, BPFP=2.0227 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,388B, BPFP=1.7021 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,572B, BPFP=1.9424 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,608B, BPFP=0.3075 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10405910 79.52044440 + layer.0.v_cache 0.00001453 0.00980935 + layer.1.k_cache 0.03397077 11.25257378 + layer.1.v_cache 0.00000547 0.00405427 + layer.2.k_cache 0.00568739 1.16230031 + layer.2.v_cache 0.00001916 0.01212905 + layer.3.k_cache 0.09581309 4.59884842 + layer.3.v_cache 0.00001862 0.01282258 + layer.4.k_cache 0.00061896 0.24743412 + layer.4.v_cache 0.00005653 0.02406884 + layer.4.output 0.19123089 701.84519944 + ------------------------------------------------------------------------------------- + TOTAL 0.09287528 294.69181654 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 87108 +BPFP 1.0398 bits/point +EBPFP 2.0795 equivalent bits/point +MSE 294.691817 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 294.6918 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,096B, BPFP=0.5691 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,112B, BPFP=2.0426 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,144B, BPFP=0.9456 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,792B, BPFP=1.9838 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,520B, BPFP=1.0147 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,280B, BPFP=1.8897 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,208B, BPFP=0.9574 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,316B, BPFP=1.8963 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,496B, BPFP=1.5618 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,924B, BPFP=1.8243 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,736B, BPFP=0.2819 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09384689 70.32742992 + layer.0.v_cache 0.00001364 0.01020457 + layer.1.k_cache 0.01313462 10.82325655 + layer.1.v_cache 0.00000592 0.00462998 + layer.2.k_cache 0.00813576 1.27615392 + layer.2.v_cache 0.00001762 0.01212040 + layer.3.k_cache 0.02628646 5.57007877 + layer.3.v_cache 0.00002044 0.01396859 + layer.4.k_cache 0.00061596 0.26278682 + layer.4.v_cache 0.00005334 0.02593586 + layer.4.output 0.17343949 636.76039916 + ------------------------------------------------------------------------------------- + TOTAL 0.07977689 267.39113879 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 90624 +BPFP 0.9799 bits/point +EBPFP 1.9599 equivalent bits/point +MSE 267.391139 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 267.3911 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 126, 128) +Output shape: (1, 126, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.output: torch.Size([1, 126, 3584]) -> torch.Size([1, 1, 126, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,444B, BPFP=0.5511 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,944B, BPFP=1.7292 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,872B, BPFP=0.8522 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,624B, BPFP=1.6895 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,552B, BPFP=0.9365 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,188B, BPFP=1.6354 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,148B, BPFP=0.8864 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,260B, BPFP=1.6443 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,208B, BPFP=1.3899 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,768B, BPFP=1.5833 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,872B, BPFP=0.3520 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13141363 68.73350694 + layer.0.v_cache 0.00001429 0.00985347 + layer.1.k_cache 0.06496696 10.03060671 + layer.1.v_cache 0.00000583 0.00449922 + layer.2.k_cache 0.01057190 1.00807759 + layer.2.v_cache 0.00002020 0.01191897 + layer.3.k_cache 0.05250616 4.09635804 + layer.3.v_cache 0.00001853 0.01270952 + layer.4.k_cache 0.00062861 0.24199577 + layer.4.v_cache 0.00004930 0.02449403 + layer.4.output 9.05470577 422.60455641 + ------------------------------------------------------------------------------------- + TOTAL 3.74371387 178.96505383 + (elements=1,096,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1096704 +Total Bytes 123880 +BPFP 0.9037 bits/point +EBPFP 1.8073 equivalent bits/point +MSE 178.965054 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.006s, Pack+Encode: 0.160s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 178.9651 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,128B, BPFP=0.5618 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,292B, BPFP=2.0280 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,272B, BPFP=0.9468 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,056B, BPFP=1.9856 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,564B, BPFP=0.9993 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,520B, BPFP=1.8894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,408B, BPFP=0.9713 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,784B, BPFP=1.9368 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,640B, BPFP=1.5517 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,272B, BPFP=1.8448 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,064B, BPFP=0.2582 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13059476 65.09164534 + layer.0.v_cache 0.00001380 0.00992604 + layer.1.k_cache 0.03034586 10.26939094 + layer.1.v_cache 0.00000531 0.00391304 + layer.2.k_cache 0.00374181 1.13417334 + layer.2.v_cache 0.00001886 0.01179013 + layer.3.k_cache 0.04454902 5.05545605 + layer.3.v_cache 0.00001968 0.01300508 + layer.4.k_cache 0.00060938 0.25244106 + layer.4.v_cache 0.00007316 0.02462966 + layer.4.output 0.16485390 620.73440066 + ------------------------------------------------------------------------------------- + TOTAL 0.08023229 260.41218678 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 92000 +BPFP 0.9719 bits/point +EBPFP 1.9439 equivalent bits/point +MSE 260.412187 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 260.4122 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,028B, BPFP=0.5143 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,636B, BPFP=1.9762 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,728B, BPFP=0.9728 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,336B, BPFP=1.9253 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,804B, BPFP=0.9857 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,964B, BPFP=1.8621 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,580B, BPFP=0.9477 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,132B, BPFP=1.8906 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,092B, BPFP=1.5442 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,736B, BPFP=1.8234 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,112B, BPFP=0.2939 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15436328 72.81190557 + layer.0.v_cache 0.00001586 0.01011959 + layer.1.k_cache 0.01080767 10.81672071 + layer.1.v_cache 0.00000575 0.00419616 + layer.2.k_cache 0.00786338 1.07735211 + layer.2.v_cache 0.00001961 0.01123657 + layer.3.k_cache 0.05652578 4.74767569 + layer.3.v_cache 0.00002042 0.01294382 + layer.4.k_cache 0.00062828 0.25314451 + layer.4.v_cache 0.00005674 0.02712406 + layer.4.output 0.15192759 587.83127911 + ------------------------------------------------------------------------------------- + TOTAL 0.07610588 247.32890427 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 97148 +BPFP 0.9705 bits/point +EBPFP 1.9411 equivalent bits/point +MSE 247.328904 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 247.3289 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,540B, BPFP=0.5088 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,724B, BPFP=2.1482 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,740B, BPFP=0.9495 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,548B, BPFP=2.1130 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,476B, BPFP=1.0970 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,976B, BPFP=1.9984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,108B, BPFP=1.0232 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,216B, BPFP=2.0465 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,308B, BPFP=1.6643 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,872B, BPFP=1.9776 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,044B, BPFP=0.2874 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10254851 76.17173102 + layer.0.v_cache 0.00001399 0.01021299 + layer.1.k_cache 0.05476840 10.59798960 + layer.1.v_cache 0.00000542 0.00416279 + layer.2.k_cache 0.00244555 1.16094736 + layer.2.v_cache 0.00001844 0.01203432 + layer.3.k_cache 0.02811479 5.10571054 + layer.3.v_cache 0.00001905 0.01350473 + layer.4.k_cache 0.00060111 0.24720226 + layer.4.v_cache 0.00005313 0.02507637 + layer.4.output 0.19635826 693.35313645 + ------------------------------------------------------------------------------------- + TOTAL 0.09194683 290.98944277 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 87552 +BPFP 1.0317 bits/point +EBPFP 2.0633 equivalent bits/point +MSE 290.989443 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 290.9894 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,020B, BPFP=0.5755 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,840B, BPFP=2.0655 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,132B, BPFP=0.9779 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,824B, BPFP=2.0625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,516B, BPFP=1.0511 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,196B, BPFP=1.9428 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,172B, BPFP=0.9855 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,328B, BPFP=1.9680 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,412B, BPFP=1.6029 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,864B, BPFP=1.8796 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,476B, BPFP=0.2852 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11987135 70.32735090 + layer.0.v_cache 0.00001390 0.00983731 + layer.1.k_cache 0.03139273 11.07842794 + layer.1.v_cache 0.00000562 0.00390344 + layer.2.k_cache 0.00816743 1.20471005 + layer.2.v_cache 0.00001857 0.01065727 + layer.3.k_cache 0.02640573 5.62175397 + layer.3.v_cache 0.00001968 0.01209107 + layer.4.k_cache 0.00062091 0.25141911 + layer.4.v_cache 0.00004972 0.02371294 + layer.4.output 0.16983549 660.39285714 + ------------------------------------------------------------------------------------- + TOTAL 0.08090671 277.13493318 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 89780 +BPFP 1.0063 bits/point +EBPFP 2.0126 equivalent bits/point +MSE 277.134933 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 277.1349 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,212B, BPFP=0.5769 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,372B, BPFP=2.0424 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,300B, BPFP=0.9519 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,148B, BPFP=2.0022 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,588B, BPFP=1.0036 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,508B, BPFP=1.8872 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,408B, BPFP=0.9713 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,716B, BPFP=1.9246 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,592B, BPFP=1.5431 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,140B, BPFP=1.8211 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,348B, BPFP=0.2655 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14089697 66.17941810 + layer.0.v_cache 0.00001393 0.01023950 + layer.1.k_cache 0.01618495 10.67152703 + layer.1.v_cache 0.00000582 0.00425276 + layer.2.k_cache 0.00661465 1.03818459 + layer.2.v_cache 0.00002083 0.01158582 + layer.3.k_cache 0.07087009 4.97531865 + layer.3.v_cache 0.00001822 0.01304892 + layer.4.k_cache 0.00061891 0.25745302 + layer.4.v_cache 0.00005311 0.02586638 + layer.4.output 0.16425455 617.67574918 + ------------------------------------------------------------------------------------- + TOTAL 0.08147525 259.23041994 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 92332 +BPFP 0.9754 bits/point +EBPFP 1.9509 equivalent bits/point +MSE 259.230420 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 259.2304 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,192B, BPFP=0.5604 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,532B, BPFP=2.0246 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,156B, BPFP=0.9052 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,120B, BPFP=1.9522 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,708B, BPFP=1.0021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,564B, BPFP=1.8546 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,476B, BPFP=0.9614 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,724B, BPFP=1.8827 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,804B, BPFP=1.5456 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,320B, BPFP=1.8118 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,356B, BPFP=0.2848 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12786581 72.07051549 + layer.0.v_cache 0.00001426 0.01053706 + layer.1.k_cache 0.03183105 11.21592935 + layer.1.v_cache 0.00000593 0.00431170 + layer.2.k_cache 0.00241907 1.22386581 + layer.2.v_cache 0.00001785 0.01120848 + layer.3.k_cache 0.02946699 4.45879064 + layer.3.v_cache 0.00001957 0.01399859 + layer.4.k_cache 0.00063747 0.24509865 + layer.4.v_cache 0.00005003 0.02531904 + layer.4.output 0.16568538 608.59349920 + ------------------------------------------------------------------------------------- + TOTAL 0.07953681 255.84906289 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 93952 +BPFP 0.9703 bits/point +EBPFP 1.9405 equivalent bits/point +MSE 255.849063 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 255.8491 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,572B, BPFP=0.5152 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,812B, BPFP=2.1659 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,884B, BPFP=0.9784 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,628B, BPFP=2.1290 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,488B, BPFP=1.0994 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,268B, BPFP=2.0569 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,132B, BPFP=1.0280 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,312B, BPFP=2.0657 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,244B, BPFP=1.6514 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,772B, BPFP=1.9575 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,064B, BPFP=0.3166 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12225728 82.35817308 + layer.0.v_cache 0.00001337 0.01006161 + layer.1.k_cache 0.03480693 10.40809357 + layer.1.v_cache 0.00000545 0.00432015 + layer.2.k_cache 0.00853881 1.19710619 + layer.2.v_cache 0.00001924 0.01255910 + layer.3.k_cache 0.06130744 5.22268207 + layer.3.v_cache 0.00001928 0.01425725 + layer.4.k_cache 0.00058936 0.25537735 + layer.4.v_cache 0.00007197 0.02688287 + layer.4.output 0.18657821 692.56227106 + ------------------------------------------------------------------------------------- + TOTAL 0.09021627 291.02620063 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 89176 +BPFP 1.0508 bits/point +EBPFP 2.1016 equivalent bits/point +MSE 291.026201 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 291.0262 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,652B, BPFP=0.5381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,480B, BPFP=2.1266 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,748B, BPFP=0.9635 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,356B, BPFP=2.1015 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,220B, BPFP=1.0593 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,800B, BPFP=1.9886 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,912B, BPFP=0.9968 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,820B, BPFP=1.9927 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,180B, BPFP=1.6599 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,632B, BPFP=1.9545 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,392B, BPFP=0.3013 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582910 77.12847504 + layer.0.v_cache 0.00001386 0.01079707 + layer.1.k_cache 0.03408930 11.31550895 + layer.1.v_cache 0.00000538 0.00462530 + layer.2.k_cache 0.00228146 1.13915302 + layer.2.v_cache 0.00001859 0.01269693 + layer.3.k_cache 0.06276476 5.10201194 + layer.3.v_cache 0.00001913 0.01455728 + layer.4.k_cache 0.00060520 0.27524309 + layer.4.v_cache 0.00005125 0.02879651 + layer.4.output 0.19356344 701.96956169 + ------------------------------------------------------------------------------------- + TOTAL 0.09238954 294.63639982 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 86192 +BPFP 1.0288 bits/point +EBPFP 2.0577 equivalent bits/point +MSE 294.636400 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 294.6364 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,968B, BPFP=0.5587 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,932B, BPFP=2.0580 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,076B, BPFP=0.9556 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,628B, BPFP=2.0008 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,568B, BPFP=1.0482 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,236B, BPFP=1.9270 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,308B, BPFP=0.9992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,500B, BPFP=1.9767 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,412B, BPFP=1.5836 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,860B, BPFP=1.8562 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,804B, BPFP=0.2906 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09700135 74.77340809 + layer.0.v_cache 0.00001355 0.01039014 + layer.1.k_cache 0.01103399 10.76993570 + layer.1.v_cache 0.00000598 0.00424450 + layer.2.k_cache 0.00794396 1.21554731 + layer.2.v_cache 0.00001935 0.01213499 + layer.3.k_cache 0.04207947 5.38459465 + layer.3.v_cache 0.00002010 0.01425586 + layer.4.k_cache 0.00062246 0.25699262 + layer.4.v_cache 0.00005226 0.02655258 + layer.4.output 0.17535226 651.56255379 + ------------------------------------------------------------------------------------- + TOTAL 0.08154460 273.72976076 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 90292 +BPFP 0.9999 bits/point +EBPFP 1.9997 equivalent bits/point +MSE 273.729761 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 273.7298 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,072B, BPFP=0.5647 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,012B, BPFP=2.0243 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,364B, BPFP=0.9860 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,728B, BPFP=1.9721 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,656B, BPFP=1.0397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,336B, BPFP=1.9000 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,352B, BPFP=0.9838 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,648B, BPFP=1.9574 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,468B, BPFP=1.5566 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,256B, BPFP=1.8853 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,412B, BPFP=0.2734 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11944316 71.09601907 + layer.0.v_cache 0.00001525 0.00959322 + layer.1.k_cache 0.01376067 10.83739732 + layer.1.v_cache 0.00000561 0.00400771 + layer.2.k_cache 0.00522648 1.18038931 + layer.2.v_cache 0.00001966 0.01160436 + layer.3.k_cache 0.04600189 5.34267363 + layer.3.v_cache 0.00001912 0.01318842 + layer.4.k_cache 0.00062079 0.24140991 + layer.4.v_cache 0.00005114 0.02420920 + layer.4.output 0.17175110 636.20057773 + ------------------------------------------------------------------------------------- + TOTAL 0.08161303 267.18614919 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 91304 +BPFP 0.9873 bits/point +EBPFP 1.9746 equivalent bits/point +MSE 267.186149 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 267.1861 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,576B, BPFP=0.5160 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,528B, BPFP=2.1090 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,780B, BPFP=0.9575 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,384B, BPFP=2.0801 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,244B, BPFP=1.0505 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,720B, BPFP=1.9471 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,924B, BPFP=0.9864 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,900B, BPFP=1.9832 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,188B, BPFP=1.6402 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,604B, BPFP=1.9239 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,972B, BPFP=0.3140 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120220 81.41804387 + layer.0.v_cache 0.00001368 0.01000771 + layer.1.k_cache 0.03273370 10.76376891 + layer.1.v_cache 0.00000573 0.00424118 + layer.2.k_cache 0.00892877 1.10393329 + layer.2.v_cache 0.00001953 0.01157618 + layer.3.k_cache 0.02974504 5.14127839 + layer.3.v_cache 0.00001879 0.01309489 + layer.4.k_cache 0.00063874 0.26846582 + layer.4.v_cache 0.00005200 0.02622612 + layer.4.output 0.17853031 692.77621337 + ------------------------------------------------------------------------------------- + TOTAL 0.08429825 291.07024294 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 86820 +BPFP 1.0230 bits/point +EBPFP 2.0461 equivalent bits/point +MSE 291.070243 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 291.0702 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,908B, BPFP=0.5474 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,140B, BPFP=2.0971 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,196B, BPFP=0.9782 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,796B, BPFP=2.0324 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,560B, BPFP=1.0467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,328B, BPFP=1.9443 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,284B, BPFP=0.9947 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,432B, BPFP=1.9639 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,576B, BPFP=1.6145 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,160B, BPFP=1.9127 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,948B, BPFP=0.2944 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09802409 71.10603351 + layer.0.v_cache 0.00001402 0.01044099 + layer.1.k_cache 0.03453328 10.56035715 + layer.1.v_cache 0.00000594 0.00426753 + layer.2.k_cache 0.01027090 1.09988863 + layer.2.v_cache 0.00001933 0.01149202 + layer.3.k_cache 0.02613793 5.43950534 + layer.3.v_cache 0.00001931 0.01310228 + layer.4.k_cache 0.00063818 0.25210332 + layer.4.v_cache 0.00005393 0.02513157 + layer.4.output 0.18084440 651.45342083 + ------------------------------------------------------------------------------------- + TOTAL 0.08444869 273.45272165 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 91328 +BPFP 1.0113 bits/point +EBPFP 2.0227 equivalent bits/point +MSE 273.452722 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 273.4527 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,384B, BPFP=0.5133 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,880B, BPFP=1.9539 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,856B, BPFP=0.8883 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,660B, BPFP=1.9205 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,420B, BPFP=0.9739 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,036B, BPFP=1.8258 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,124B, BPFP=0.9290 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,232B, BPFP=1.8556 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,940B, BPFP=1.5079 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,672B, BPFP=1.7706 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,160B, BPFP=0.2635 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13030120 68.34114267 + layer.0.v_cache 0.00001385 0.01014771 + layer.1.k_cache 0.02890817 10.46207050 + layer.1.v_cache 0.00000611 0.00462761 + layer.2.k_cache 0.00721959 1.09643362 + layer.2.v_cache 0.00001830 0.01201694 + layer.3.k_cache 0.02195997 4.15353838 + layer.3.v_cache 0.00001903 0.01350370 + layer.4.k_cache 0.00063766 0.26047705 + layer.4.v_cache 0.00005061 0.02611669 + layer.4.output 11.07892277 519.84093273 + ------------------------------------------------------------------------------------- + TOTAL 4.57303494 219.01568259 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 105364 +BPFP 0.9402 bits/point +EBPFP 1.8804 equivalent bits/point +MSE 219.015683 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 219.0157 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,008B, BPFP=0.5663 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,032B, BPFP=2.0768 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,016B, BPFP=0.9443 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,772B, BPFP=2.0279 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,516B, BPFP=1.0384 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,256B, BPFP=1.9307 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,120B, BPFP=0.9639 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,392B, BPFP=1.9563 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,568B, BPFP=1.6130 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,104B, BPFP=1.9021 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,152B, BPFP=0.2730 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12633393 72.57143614 + layer.0.v_cache 0.00001372 0.01045902 + layer.1.k_cache 0.01411889 11.06333243 + layer.1.v_cache 0.00000549 0.00436509 + layer.2.k_cache 0.00833277 1.19015246 + layer.2.v_cache 0.00001851 0.01192076 + layer.3.k_cache 0.05797521 5.89399848 + layer.3.v_cache 0.00001879 0.01342808 + layer.4.k_cache 0.00064672 0.27933930 + layer.4.v_cache 0.00005349 0.02775153 + layer.4.output 0.17447430 652.20793890 + ------------------------------------------------------------------------------------- + TOTAL 0.08404927 273.91304445 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 89936 +BPFP 0.9959 bits/point +EBPFP 1.9918 equivalent bits/point +MSE 273.913044 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 273.9130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,248B, BPFP=0.5018 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,016B, BPFP=2.2357 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,508B, BPFP=1.0063 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,740B, BPFP=2.1741 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,068B, BPFP=1.1313 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,012B, BPFP=2.0116 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,664B, BPFP=1.0411 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,224B, BPFP=2.0589 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,660B, BPFP=1.7098 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,156B, BPFP=2.0438 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,228B, BPFP=0.3261 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08086312 69.78753488 + layer.0.v_cache 0.00001331 0.01034748 + layer.1.k_cache 0.01551799 11.09392003 + layer.1.v_cache 0.00000513 0.00401626 + layer.2.k_cache 0.00232868 1.25242026 + layer.2.v_cache 0.00002262 0.01277414 + layer.3.k_cache 0.10623371 5.20413775 + layer.3.v_cache 0.00001836 0.01368329 + layer.4.k_cache 0.00061897 0.26020333 + layer.4.v_cache 0.00005922 0.02663836 + layer.4.output 0.20964142 774.08845663 + ------------------------------------------------------------------------------------- + TOTAL 0.09842183 323.89911013 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 81524 +BPFP 1.0704 bits/point +EBPFP 2.1409 equivalent bits/point +MSE 323.899110 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 323.8991 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,108B, BPFP=0.5279 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,572B, BPFP=1.9654 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,348B, BPFP=0.9083 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,292B, BPFP=1.9178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,624B, BPFP=0.9552 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,492B, BPFP=1.7819 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,340B, BPFP=0.9069 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,500B, BPFP=1.7833 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,800B, BPFP=1.4946 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,292B, BPFP=1.7480 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,620B, BPFP=0.2819 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13670269 72.89536982 + layer.0.v_cache 0.00001568 0.01033823 + layer.1.k_cache 0.03139175 10.76787202 + layer.1.v_cache 0.00000602 0.00457595 + layer.2.k_cache 0.00629326 1.09054068 + layer.2.v_cache 0.00001871 0.01139874 + layer.3.k_cache 0.03837063 5.02473284 + layer.3.v_cache 0.00001860 0.01209757 + layer.4.k_cache 0.00062597 0.25415931 + layer.4.v_cache 0.00005354 0.02533211 + layer.4.output 0.15974679 587.32545613 + ------------------------------------------------------------------------------------- + TOTAL 0.07833673 247.13968295 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 93988 +BPFP 0.9390 bits/point +EBPFP 1.8780 equivalent bits/point +MSE 247.139683 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 247.1397 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,652B, BPFP=0.5333 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,200B, BPFP=1.9276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,368B, BPFP=0.9299 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,920B, BPFP=1.8867 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,824B, BPFP=0.9965 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,524B, BPFP=1.8289 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,480B, BPFP=0.9463 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,608B, BPFP=1.8411 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,428B, BPFP=1.5228 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,972B, BPFP=1.7482 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,136B, BPFP=0.2532 +⌛️ [2/4] FRONTEND: Frontend time: 0.171s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12281527 68.65956301 + layer.0.v_cache 0.00001417 0.01015708 + layer.1.k_cache 0.04442539 10.05084286 + layer.1.v_cache 0.00000572 0.00420682 + layer.2.k_cache 0.01033308 0.99064572 + layer.2.v_cache 0.00001923 0.01187338 + layer.3.k_cache 0.04555132 4.07862469 + layer.3.v_cache 0.00001841 0.01296873 + layer.4.k_cache 0.00060580 0.23774309 + layer.4.v_cache 0.00007068 0.02538416 + layer.4.output 10.66341841 499.39999166 + ------------------------------------------------------------------------------------- + TOTAL 4.40398753 210.58129124 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 109112 +BPFP 0.9373 bits/point +EBPFP 1.8745 equivalent bits/point +MSE 210.581291 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.006s, Pack+Encode: 0.171s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 210.5813 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.5596 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,248B, BPFP=2.0436 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,388B, BPFP=0.9789 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,044B, BPFP=2.0065 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,640B, BPFP=1.0247 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,360B, BPFP=1.8823 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,296B, BPFP=0.9622 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,504B, BPFP=1.9084 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,556B, BPFP=1.5545 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,052B, BPFP=1.8263 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,048B, BPFP=0.2608 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11362342 65.70419241 + layer.0.v_cache 0.00001418 0.01001964 + layer.1.k_cache 0.05429294 10.53585603 + layer.1.v_cache 0.00000580 0.00423279 + layer.2.k_cache 0.00513919 1.15507232 + layer.2.v_cache 0.00001816 0.01193473 + layer.3.k_cache 0.05885953 5.29255428 + layer.3.v_cache 0.00001841 0.01294345 + layer.4.k_cache 0.00061763 0.25283922 + layer.4.v_cache 0.00005080 0.02487487 + layer.4.output 0.16727475 629.38751038 + ------------------------------------------------------------------------------------- + TOTAL 0.08256255 264.04218191 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 91216 +BPFP 0.9749 bits/point +EBPFP 1.9497 equivalent bits/point +MSE 264.042182 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 264.0422 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,156B, BPFP=0.5302 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,584B, BPFP=1.9462 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,592B, BPFP=0.9395 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,428B, BPFP=1.9200 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,020B, BPFP=1.0114 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,812B, BPFP=1.8165 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,780B, BPFP=0.9711 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,948B, BPFP=1.8394 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,184B, BPFP=1.5430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,460B, BPFP=1.7574 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,596B, BPFP=0.2783 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702597 71.55362168 + layer.0.v_cache 0.00001402 0.01014812 + layer.1.k_cache 0.01332537 10.62373664 + layer.1.v_cache 0.00000581 0.00426219 + layer.2.k_cache 0.00507878 1.16447153 + layer.2.v_cache 0.00001944 0.01146855 + layer.3.k_cache 0.04051446 5.11216490 + layer.3.v_cache 0.00001931 0.01245832 + layer.4.k_cache 0.00062870 0.23537166 + layer.4.v_cache 0.00006514 0.02395911 + layer.4.output 0.16549673 581.83328533 + ------------------------------------------------------------------------------------- + TOTAL 0.07795142 244.79909765 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 96560 +BPFP 0.9543 bits/point +EBPFP 1.9086 equivalent bits/point +MSE 244.799098 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 244.7991 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,076B, BPFP=0.5654 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,984B, BPFP=2.0191 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,232B, BPFP=0.9618 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,648B, BPFP=1.9574 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,460B, BPFP=1.0037 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,012B, BPFP=1.8404 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,148B, BPFP=0.9463 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,132B, BPFP=1.8625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,388B, BPFP=1.5419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,684B, BPFP=1.7801 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,700B, BPFP=0.2810 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13925676 68.40895565 + layer.0.v_cache 0.00001387 0.01005266 + layer.1.k_cache 0.03290265 10.60947553 + layer.1.v_cache 0.00000589 0.00425195 + layer.2.k_cache 0.00800494 1.09017953 + layer.2.v_cache 0.00001921 0.01121186 + layer.3.k_cache 0.08301921 5.70162856 + layer.3.v_cache 0.00001931 0.01252007 + layer.4.k_cache 0.00065893 0.24255355 + layer.4.v_cache 0.00004891 0.02380018 + layer.4.output 0.17367665 635.77752101 + ------------------------------------------------------------------------------------- + TOTAL 0.08704037 266.85631039 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 89464 +BPFP 0.9674 bits/point +EBPFP 1.9348 equivalent bits/point +MSE 266.856310 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 266.8563 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,176B, BPFP=0.4818 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,840B, BPFP=1.9478 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,856B, BPFP=0.8883 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,440B, BPFP=1.8871 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,340B, BPFP=0.9618 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,012B, BPFP=1.8222 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,064B, BPFP=0.9199 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,140B, BPFP=1.8416 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,952B, BPFP=1.5097 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,596B, BPFP=1.7591 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,168B, BPFP=0.2420 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10583916 65.25916831 + layer.0.v_cache 0.00001406 0.01010752 + layer.1.k_cache 0.01351924 10.27406918 + layer.1.v_cache 0.00000536 0.00428787 + layer.2.k_cache 0.00567923 1.00486518 + layer.2.v_cache 0.00001856 0.01195001 + layer.3.k_cache 0.05293757 4.20261568 + layer.3.v_cache 0.00001841 0.01288600 + layer.4.k_cache 0.00064880 0.25430853 + layer.4.v_cache 0.00004836 0.02624214 + layer.4.output 11.07886250 517.13761269 + ------------------------------------------------------------------------------------- + TOTAL 4.57239802 217.70728172 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 103584 +BPFP 0.9243 bits/point +EBPFP 1.8487 equivalent bits/point +MSE 217.707282 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 217.7073 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,980B, BPFP=0.5610 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,020B, BPFP=2.0745 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,940B, BPFP=0.9300 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,764B, BPFP=2.0264 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,564B, BPFP=1.0474 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,320B, BPFP=1.9428 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,220B, BPFP=0.9827 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,512B, BPFP=1.9789 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,388B, BPFP=1.5791 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,008B, BPFP=1.8840 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,408B, BPFP=0.2799 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860243 76.40350268 + layer.0.v_cache 0.00001403 0.00997384 + layer.1.k_cache 0.01400426 11.05059594 + layer.1.v_cache 0.00000584 0.00456448 + layer.2.k_cache 0.01277778 1.18053547 + layer.2.v_cache 0.00001980 0.01213613 + layer.3.k_cache 0.07376247 5.31392808 + layer.3.v_cache 0.00001877 0.01303515 + layer.4.k_cache 0.00062152 0.26103564 + layer.4.v_cache 0.00005140 0.02479152 + layer.4.output 0.18179212 652.30755164 + ------------------------------------------------------------------------------------- + TOTAL 0.08837783 274.14276237 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 90124 +BPFP 0.9980 bits/point +EBPFP 1.9960 equivalent bits/point +MSE 274.142762 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 274.1428 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,188B, BPFP=0.5535 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,592B, BPFP=2.0125 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,348B, BPFP=0.9285 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,324B, BPFP=1.9660 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,796B, BPFP=1.0063 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,680B, BPFP=1.8542 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,440B, BPFP=0.9444 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,984B, BPFP=1.9069 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,944B, BPFP=1.5528 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,484B, BPFP=1.8201 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,980B, BPFP=0.2723 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13781073 76.75595161 + layer.0.v_cache 0.00001414 0.01049015 + layer.1.k_cache 0.01252146 10.23820394 + layer.1.v_cache 0.00000561 0.00431857 + layer.2.k_cache 0.00238173 1.08762521 + layer.2.v_cache 0.00001849 0.01192177 + layer.3.k_cache 0.04360520 5.07010329 + layer.3.v_cache 0.00001973 0.01375615 + layer.4.k_cache 0.00061668 0.24458021 + layer.4.v_cache 0.00005448 0.02649152 + layer.4.output 0.16772948 601.27509921 + ------------------------------------------------------------------------------------- + TOTAL 0.08065615 253.08171393 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 94760 +BPFP 0.9677 bits/point +EBPFP 1.9355 equivalent bits/point +MSE 253.081714 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 253.0817 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,684B, BPFP=0.5309 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,664B, BPFP=2.1092 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,812B, BPFP=0.9517 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,160B, BPFP=2.0095 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,368B, BPFP=1.0617 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,856B, BPFP=1.9494 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,992B, BPFP=0.9873 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,044B, BPFP=1.9866 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,268B, BPFP=1.6353 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,636B, BPFP=1.9059 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,784B, BPFP=0.3047 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07550411 78.38794254 + layer.0.v_cache 0.00001381 0.01022491 + layer.1.k_cache 0.01271453 11.06547295 + layer.1.v_cache 0.00000531 0.00416959 + layer.2.k_cache 0.00681068 1.20778724 + layer.2.v_cache 0.00001904 0.01215878 + layer.3.k_cache 0.02844424 4.94415476 + layer.3.v_cache 0.00001798 0.01340087 + layer.4.k_cache 0.00061127 0.26397044 + layer.4.v_cache 0.00009323 0.02753822 + layer.4.output 0.18255964 684.35296112 + ------------------------------------------------------------------------------------- + TOTAL 0.08247951 287.43573813 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 87268 +BPFP 1.0153 bits/point +EBPFP 2.0306 equivalent bits/point +MSE 287.435738 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 287.4357 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,944B, BPFP=0.5750 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,828B, BPFP=2.1148 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,896B, BPFP=0.9563 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,688B, BPFP=2.0875 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,588B, BPFP=1.0914 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,048B, BPFP=1.9625 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,248B, BPFP=1.0250 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,364B, BPFP=2.0242 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,320B, BPFP=1.6250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,884B, BPFP=1.9305 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,912B, BPFP=0.3045 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08590809 72.43169556 + layer.0.v_cache 0.00001388 0.00976404 + layer.1.k_cache 0.01594085 10.90036392 + layer.1.v_cache 0.00000552 0.00415887 + layer.2.k_cache 0.00791870 1.24864616 + layer.2.v_cache 0.00001801 0.01110443 + layer.3.k_cache 0.02757127 4.76848221 + layer.3.v_cache 0.00001862 0.01264535 + layer.4.k_cache 0.00063773 0.25134463 + layer.4.v_cache 0.00005492 0.02512332 + layer.4.output 0.18408036 675.53387277 + ------------------------------------------------------------------------------------- + TOTAL 0.08392059 283.43531987 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 89720 +BPFP 1.0308 bits/point +EBPFP 2.0616 equivalent bits/point +MSE 283.435320 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 283.4353 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,488B, BPFP=0.5115 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,496B, BPFP=2.1579 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,840B, BPFP=0.9951 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,960B, BPFP=2.0477 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,256B, BPFP=1.0806 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,592B, BPFP=1.9720 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,780B, BPFP=0.9827 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,588B, BPFP=1.9712 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,104B, BPFP=1.6661 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,244B, BPFP=1.9005 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,796B, BPFP=0.3171 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08991440 72.71381579 + layer.0.v_cache 0.00001347 0.01005293 + layer.1.k_cache 0.03345199 11.51418586 + layer.1.v_cache 0.00000518 0.00425359 + layer.2.k_cache 0.00398020 1.24545870 + layer.2.v_cache 0.00001815 0.01171403 + layer.3.k_cache 0.04607732 4.98244516 + layer.3.v_cache 0.00001896 0.01306070 + layer.4.k_cache 0.00062119 0.25365639 + layer.4.v_cache 0.00005040 0.02485834 + layer.4.output 0.18612516 711.71957237 + ------------------------------------------------------------------------------------- + TOTAL 0.08688396 298.40061812 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 85144 +BPFP 1.0297 bits/point +EBPFP 2.0594 equivalent bits/point +MSE 298.400618 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 298.4006 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,136B, BPFP=0.5385 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,548B, BPFP=1.9828 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,396B, BPFP=0.9265 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,296B, BPFP=1.9396 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,772B, BPFP=0.9911 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,764B, BPFP=1.8482 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,500B, BPFP=0.9444 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,872B, BPFP=1.8668 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,824B, BPFP=1.5151 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,456B, BPFP=1.7953 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,168B, BPFP=0.2739 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11186004 80.08996180 + layer.0.v_cache 0.00001364 0.01042292 + layer.1.k_cache 0.01386243 10.54976847 + layer.1.v_cache 0.00000608 0.00426674 + layer.2.k_cache 0.00506424 1.08238438 + layer.2.v_cache 0.00001872 0.01149766 + layer.3.k_cache 0.02496567 4.63733749 + layer.3.v_cache 0.00001876 0.01317637 + layer.4.k_cache 0.00064518 0.25448036 + layer.4.v_cache 0.00005436 0.02649393 + layer.4.output 0.16682192 591.59051217 + ------------------------------------------------------------------------------------- + TOTAL 0.07789780 249.28313972 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 94732 +BPFP 0.9568 bits/point +EBPFP 1.9136 equivalent bits/point +MSE 249.283140 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 249.2831 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,988B, BPFP=0.5625 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,000B, BPFP=2.0708 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,200B, BPFP=0.9789 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,636B, BPFP=2.0023 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,452B, BPFP=1.0264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,272B, BPFP=1.9337 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,164B, BPFP=0.9721 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,372B, BPFP=1.9526 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,472B, BPFP=1.5949 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,924B, BPFP=1.8682 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,628B, BPFP=0.2858 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08542032 72.18914721 + layer.0.v_cache 0.00001402 0.01014128 + layer.1.k_cache 0.01083013 11.13170798 + layer.1.v_cache 0.00000556 0.00416753 + layer.2.k_cache 0.00382075 1.18441120 + layer.2.v_cache 0.00001860 0.01157408 + layer.3.k_cache 0.02610962 5.67092013 + layer.3.v_cache 0.00002051 0.01362682 + layer.4.k_cache 0.00063312 0.25611108 + layer.4.v_cache 0.00005614 0.02776466 + layer.4.output 0.16811641 652.17028830 + ------------------------------------------------------------------------------------- + TOTAL 0.07669080 273.86421118 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 90108 +BPFP 0.9978 bits/point +EBPFP 1.9957 equivalent bits/point +MSE 273.864211 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 273.8642 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,888B, BPFP=0.5571 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,976B, BPFP=2.1173 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,020B, BPFP=0.9684 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,540B, BPFP=2.0332 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,516B, BPFP=1.0640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,156B, BPFP=1.9591 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,152B, BPFP=0.9938 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,232B, BPFP=1.9738 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,284B, BPFP=1.5980 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,908B, BPFP=1.9113 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,732B, BPFP=0.2957 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10540552 72.05482615 + layer.0.v_cache 0.00001541 0.01000601 + layer.1.k_cache 0.03402772 10.46361853 + layer.1.v_cache 0.00000555 0.00435293 + layer.2.k_cache 0.00394259 1.21869179 + layer.2.v_cache 0.00001868 0.01200465 + layer.3.k_cache 0.02873192 4.91424561 + layer.3.v_cache 0.00001983 0.01297322 + layer.4.k_cache 0.00063741 0.26278531 + layer.4.v_cache 0.00005806 0.02651884 + layer.4.output 0.17961645 668.89572310 + ------------------------------------------------------------------------------------- + TOTAL 0.08412811 280.66176969 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 89404 +BPFP 1.0145 bits/point +EBPFP 2.0290 equivalent bits/point +MSE 280.661770 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 280.6618 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,812B, BPFP=0.5492 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,064B, BPFP=2.1609 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,808B, BPFP=0.9391 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,676B, BPFP=2.0852 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,664B, BPFP=1.1062 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,112B, BPFP=1.9750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,188B, BPFP=1.0133 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,336B, BPFP=2.0187 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,312B, BPFP=1.6234 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,864B, BPFP=1.9266 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,248B, BPFP=0.3138 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10022384 75.84978027 + layer.0.v_cache 0.00001350 0.01010755 + layer.1.k_cache 0.03445883 11.34320526 + layer.1.v_cache 0.00000556 0.00428351 + layer.2.k_cache 0.00550650 1.26080208 + layer.2.v_cache 0.00001848 0.01145646 + layer.3.k_cache 0.03016714 4.93417053 + layer.3.v_cache 0.00001937 0.01384518 + layer.4.k_cache 0.00063435 0.25309162 + layer.4.v_cache 0.00006510 0.02614925 + layer.4.output 0.18593751 675.62460937 + ------------------------------------------------------------------------------------- + TOTAL 0.08662796 283.71053867 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 90084 +BPFP 1.0350 bits/point +EBPFP 2.0699 equivalent bits/point +MSE 283.710539 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 283.7105 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,176B, BPFP=0.5116 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,800B, BPFP=1.9008 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,544B, BPFP=0.8930 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,504B, BPFP=1.8531 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,888B, BPFP=0.9485 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,932B, BPFP=1.7610 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,524B, BPFP=0.8898 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,124B, BPFP=1.7919 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,108B, BPFP=1.4671 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,664B, BPFP=1.7178 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,868B, BPFP=0.2501 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11898830 70.64452118 + layer.0.v_cache 0.00001403 0.01008333 + layer.1.k_cache 0.01243949 10.28596040 + layer.1.v_cache 0.00000551 0.00414274 + layer.2.k_cache 0.00346032 1.14535900 + layer.2.v_cache 0.00001901 0.01158983 + layer.3.k_cache 0.02403709 5.17777213 + layer.3.v_cache 0.00001994 0.01281119 + layer.4.k_cache 0.00063949 0.23962481 + layer.4.v_cache 0.00006591 0.02481835 + layer.4.output 0.03773703 573.54008652 + ------------------------------------------------------------------------------------- + TOTAL 0.02493225 241.31395815 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 96132 +BPFP 0.9109 bits/point +EBPFP 1.8218 equivalent bits/point +MSE 241.313958 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 241.3140 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,020B, BPFP=0.5755 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,988B, BPFP=2.0938 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,032B, BPFP=0.9588 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,712B, BPFP=2.0412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,552B, BPFP=1.0579 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,224B, BPFP=1.9482 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,112B, BPFP=0.9741 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,384B, BPFP=1.9787 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,360B, BPFP=1.5930 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,964B, BPFP=1.8986 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,664B, BPFP=0.2903 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10693085 73.71527248 + layer.0.v_cache 0.00001341 0.01027173 + layer.1.k_cache 0.01172146 10.60356028 + layer.1.v_cache 0.00000553 0.00442568 + layer.2.k_cache 0.00236719 1.21635288 + layer.2.v_cache 0.00001835 0.01180382 + layer.3.k_cache 0.02622769 4.98694741 + layer.3.v_cache 0.00001864 0.01285746 + layer.4.k_cache 0.00061713 0.26212194 + layer.4.v_cache 0.00004943 0.02638267 + layer.4.output 0.18273007 660.08150044 + ------------------------------------------------------------------------------------- + TOTAL 0.08394589 277.14238232 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 90012 +BPFP 1.0089 bits/point +EBPFP 2.0178 equivalent bits/point +MSE 277.142382 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 277.1424 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,056B, BPFP=0.4728 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,552B, BPFP=1.9418 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,836B, BPFP=0.9028 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,296B, BPFP=1.9022 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,156B, BPFP=0.9524 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,636B, BPFP=1.8001 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,860B, BPFP=0.9066 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,772B, BPFP=1.8212 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,664B, BPFP=1.4950 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,416B, BPFP=1.7661 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,396B, BPFP=0.2519 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13156438 71.81702506 + layer.0.v_cache 0.00001398 0.01049320 + layer.1.k_cache 0.04507210 10.49487184 + layer.1.v_cache 0.00000603 0.00453439 + layer.2.k_cache 0.00848899 1.19390575 + layer.2.v_cache 0.00001983 0.01263826 + layer.3.k_cache 0.06093804 4.36585953 + layer.3.v_cache 0.00001891 0.01349054 + layer.4.k_cache 0.00062755 0.24974188 + layer.4.v_cache 0.00004781 0.02488699 + layer.4.output 11.29576791 531.77864215 + ------------------------------------------------------------------------------------- + TOTAL 4.66571606 224.15517309 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 101640 +BPFP 0.9249 bits/point +EBPFP 1.8499 equivalent bits/point +MSE 224.155173 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 224.1552 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,988B, BPFP=0.5625 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,148B, BPFP=2.0986 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,112B, BPFP=0.9623 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,816B, BPFP=2.0361 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,608B, BPFP=1.0557 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,380B, BPFP=1.9541 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,208B, BPFP=0.9804 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,568B, BPFP=1.9895 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,468B, BPFP=1.5941 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,280B, BPFP=1.9352 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,568B, BPFP=0.2842 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12340720 74.53823889 + layer.0.v_cache 0.00001379 0.01049478 + layer.1.k_cache 0.01365945 10.76283356 + layer.1.v_cache 0.00000586 0.00422238 + layer.2.k_cache 0.00515996 1.20720976 + layer.2.v_cache 0.00002118 0.01204328 + layer.3.k_cache 0.04682968 4.82371760 + layer.3.v_cache 0.00001899 0.01286381 + layer.4.k_cache 0.00062596 0.26657920 + layer.4.v_cache 0.00004975 0.02600747 + layer.4.output 0.18155291 652.32610800 + ------------------------------------------------------------------------------------- + TOTAL 0.08592130 273.99688040 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 91144 +BPFP 1.0093 bits/point +EBPFP 2.0186 equivalent bits/point +MSE 273.996880 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 273.9969 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,040B, BPFP=0.5655 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,152B, BPFP=2.0744 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,192B, BPFP=0.9658 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,784B, BPFP=2.0060 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,460B, BPFP=1.0156 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,264B, BPFP=1.9092 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,188B, BPFP=0.9650 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,408B, BPFP=1.9360 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,420B, BPFP=1.5662 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,072B, BPFP=1.8735 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,084B, BPFP=0.2945 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10400932 73.89790853 + layer.0.v_cache 0.00001381 0.00986541 + layer.1.k_cache 0.01300498 10.79235259 + layer.1.v_cache 0.00000559 0.00397466 + layer.2.k_cache 0.00535372 1.18562544 + layer.2.v_cache 0.00001862 0.01115672 + layer.3.k_cache 0.04415199 4.95055644 + layer.3.v_cache 0.00001841 0.01258734 + layer.4.k_cache 0.00060249 0.25576192 + layer.4.v_cache 0.00005274 0.02582025 + layer.4.output 0.16694923 644.62877338 + ------------------------------------------------------------------------------------- + TOTAL 0.07858096 270.79688370 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 91064 +BPFP 0.9964 bits/point +EBPFP 1.9928 equivalent bits/point +MSE 270.796884 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 270.7969 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,812B, BPFP=0.5492 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,988B, BPFP=2.1461 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,852B, BPFP=0.9477 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,588B, BPFP=2.0680 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,524B, BPFP=1.0789 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,396B, BPFP=2.0305 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,160B, BPFP=1.0078 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,556B, BPFP=2.0617 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,400B, BPFP=1.6406 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,056B, BPFP=1.9641 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,288B, BPFP=0.2871 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09321334 70.53258057 + layer.0.v_cache 0.00001862 0.01075342 + layer.1.k_cache 0.03257937 10.87195358 + layer.1.v_cache 0.00000532 0.00411998 + layer.2.k_cache 0.01442847 1.16426907 + layer.2.v_cache 0.00001865 0.01170414 + layer.3.k_cache 0.04385926 4.56638641 + layer.3.v_cache 0.00001828 0.01311800 + layer.4.k_cache 0.00061064 0.25694802 + layer.4.v_cache 0.00005353 0.02766927 + layer.4.output 0.18500509 675.97488839 + ------------------------------------------------------------------------------------- + TOTAL 0.08704948 283.48727772 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 89620 +BPFP 1.0296 bits/point +EBPFP 2.0593 equivalent bits/point +MSE 283.487278 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 283.4873 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,448B, BPFP=0.5231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,952B, BPFP=1.9648 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,796B, BPFP=0.8792 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,556B, BPFP=1.9047 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,548B, BPFP=0.9933 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,076B, BPFP=1.8319 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,204B, BPFP=0.9411 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,136B, BPFP=1.8410 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,132B, BPFP=1.5370 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,784B, BPFP=1.7876 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,652B, BPFP=0.2525 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11527186 67.50973718 + layer.0.v_cache 0.00001510 0.01027871 + layer.1.k_cache 0.02920971 10.48303697 + layer.1.v_cache 0.00000551 0.00435804 + layer.2.k_cache 0.00857845 1.14028323 + layer.2.v_cache 0.00001837 0.01174607 + layer.3.k_cache 0.03550095 4.26009599 + layer.3.v_cache 0.00001950 0.01301973 + layer.4.k_cache 0.00063242 0.25723137 + layer.4.v_cache 0.00005150 0.02671726 + layer.4.output 11.08123415 519.78536755 + ------------------------------------------------------------------------------------- + TOTAL 4.57399661 218.95376926 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 105284 +BPFP 0.9395 bits/point +EBPFP 1.8790 equivalent bits/point +MSE 218.953769 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.006s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 218.9538 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,060B, BPFP=0.5197 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,632B, BPFP=1.9755 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,492B, BPFP=0.9327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,344B, BPFP=1.9266 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,892B, BPFP=1.0007 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,736B, BPFP=1.8234 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,648B, BPFP=0.9592 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,908B, BPFP=1.8526 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,012B, BPFP=1.5306 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,452B, BPFP=1.7751 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,900B, BPFP=0.2645 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09673090 70.86312203 + layer.0.v_cache 0.00001344 0.00970562 + layer.1.k_cache 0.04918879 10.99426800 + layer.1.v_cache 0.00000572 0.00421791 + layer.2.k_cache 0.00880511 1.08852237 + layer.2.v_cache 0.00001944 0.01153141 + layer.3.k_cache 0.06586690 4.61220716 + layer.3.v_cache 0.00001823 0.01278883 + layer.4.k_cache 0.00062484 0.25357143 + layer.4.v_cache 0.00006325 0.02488436 + layer.4.output 0.15294319 587.31128688 + ------------------------------------------------------------------------------------- + TOTAL 0.07599641 247.00316631 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 95076 +BPFP 0.9498 bits/point +EBPFP 1.8997 equivalent bits/point +MSE 247.003166 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 247.0032 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,096B, BPFP=0.5258 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,720B, BPFP=1.9905 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,544B, BPFP=0.9416 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,364B, BPFP=1.9300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,836B, BPFP=0.9912 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,732B, BPFP=1.8227 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,560B, BPFP=0.9443 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,896B, BPFP=1.8505 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,900B, BPFP=1.5115 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,404B, BPFP=1.7670 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,448B, BPFP=0.2778 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12676482 73.90635615 + layer.0.v_cache 0.00001339 0.00990674 + layer.1.k_cache 0.03133728 10.48062665 + layer.1.v_cache 0.00000590 0.00420188 + layer.2.k_cache 0.00372887 1.08971355 + layer.2.v_cache 0.00001820 0.01158170 + layer.3.k_cache 0.02418065 4.59064052 + layer.3.v_cache 0.00002034 0.01354747 + layer.4.k_cache 0.00062996 0.25554242 + layer.4.v_cache 0.00006053 0.02536840 + layer.4.output 0.15389095 588.14164402 + ------------------------------------------------------------------------------------- + TOTAL 0.07435274 247.49288198 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 95500 +BPFP 0.9541 bits/point +EBPFP 1.9082 equivalent bits/point +MSE 247.492882 +---------------------- -------------------------------------------------------- +Time: 0.372s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 247.4929 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,164B, BPFP=0.5493 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,500B, BPFP=1.9965 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,248B, BPFP=0.9111 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,004B, BPFP=1.9104 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,628B, BPFP=0.9771 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,528B, BPFP=1.8278 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,384B, BPFP=0.9347 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,640B, BPFP=1.8472 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,820B, BPFP=1.5312 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,348B, BPFP=1.7965 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,180B, BPFP=0.2773 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12863920 74.20754666 + layer.0.v_cache 0.00001425 0.01009664 + layer.1.k_cache 0.01389826 10.31337484 + layer.1.v_cache 0.00000571 0.00424801 + layer.2.k_cache 0.00659207 1.06098921 + layer.2.v_cache 0.00001883 0.01150508 + layer.3.k_cache 0.05540012 5.26467251 + layer.3.v_cache 0.00001922 0.01234221 + layer.4.k_cache 0.00063279 0.23835672 + layer.4.v_cache 0.00005416 0.02381265 + layer.4.output 0.16196846 601.64295635 + ------------------------------------------------------------------------------------- + TOTAL 0.07876787 253.09691994 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 93444 +BPFP 0.9543 bits/point +EBPFP 1.9086 equivalent bits/point +MSE 253.096920 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 253.0969 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,608B, BPFP=0.5220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,204B, BPFP=1.9103 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,432B, BPFP=0.9306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,924B, BPFP=1.8698 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,868B, BPFP=0.9936 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,376B, BPFP=1.7905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,488B, BPFP=0.9387 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,536B, BPFP=1.8137 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,512B, BPFP=1.5208 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,012B, BPFP=1.7378 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,132B, BPFP=0.2507 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13160179 70.13298430 + layer.0.v_cache 0.00001577 0.01022524 + layer.1.k_cache 0.05775581 10.73384942 + layer.1.v_cache 0.00000549 0.00409047 + layer.2.k_cache 0.01027601 1.04763137 + layer.2.v_cache 0.00002197 0.01225694 + layer.3.k_cache 0.07329086 4.53435884 + layer.3.v_cache 0.00001983 0.01368057 + layer.4.k_cache 0.00064477 0.25059530 + layer.4.v_cache 0.00005260 0.02533944 + layer.4.output 10.56448284 492.90186839 + ------------------------------------------------------------------------------------- + TOTAL 4.36618028 208.06341709 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 109092 +BPFP 0.9284 bits/point +EBPFP 1.8568 equivalent bits/point +MSE 208.063417 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 208.0634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,928B, BPFP=0.5512 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,032B, BPFP=2.0768 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,072B, BPFP=0.9548 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,772B, BPFP=2.0279 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,560B, BPFP=1.0467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,196B, BPFP=1.9194 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,156B, BPFP=0.9706 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,180B, BPFP=1.9164 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,408B, BPFP=1.5828 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,752B, BPFP=1.8358 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,340B, BPFP=0.2781 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10139442 72.92454584 + layer.0.v_cache 0.00001346 0.00995530 + layer.1.k_cache 0.03266963 10.72961573 + layer.1.v_cache 0.00000526 0.00398649 + layer.2.k_cache 0.01283502 1.23469157 + layer.2.v_cache 0.00001884 0.01114127 + layer.3.k_cache 0.02666297 5.49509292 + layer.3.v_cache 0.00001780 0.01222980 + layer.4.k_cache 0.00061512 0.25090937 + layer.4.v_cache 0.00004947 0.02414201 + layer.4.output 0.18393325 652.12284854 + ------------------------------------------------------------------------------------- + TOTAL 0.08598910 273.85625000 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 89396 +BPFP 0.9899 bits/point +EBPFP 1.9799 equivalent bits/point +MSE 273.856250 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 273.8563 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,148B, BPFP=0.4870 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,756B, BPFP=1.9734 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,788B, BPFP=0.8954 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,420B, BPFP=1.9214 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,452B, BPFP=0.9981 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,936B, BPFP=1.8465 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,196B, BPFP=0.9585 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,020B, BPFP=1.8595 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,916B, BPFP=1.5340 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,508B, BPFP=1.7803 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,708B, BPFP=0.2588 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11149573 71.66318263 + layer.0.v_cache 0.00001357 0.01008326 + layer.1.k_cache 0.08092193 10.39661007 + layer.1.v_cache 0.00000571 0.00414174 + layer.2.k_cache 0.01109533 1.02749437 + layer.2.v_cache 0.00001920 0.01166465 + layer.3.k_cache 0.03617269 4.72479611 + layer.3.v_cache 0.00001858 0.01240349 + layer.4.k_cache 0.00062559 0.23300541 + layer.4.v_cache 0.00005684 0.02457845 + layer.4.output 11.30219499 531.39245934 + ------------------------------------------------------------------------------------- + TOTAL 4.66798765 223.99148091 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 103848 +BPFP 0.9450 bits/point +EBPFP 1.8901 equivalent bits/point +MSE 223.991481 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 223.9915 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,540B, BPFP=0.5088 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,708B, BPFP=2.1450 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,836B, BPFP=0.9688 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,452B, BPFP=2.0938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,432B, BPFP=1.0881 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,788B, BPFP=1.9607 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,064B, BPFP=1.0144 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,912B, BPFP=1.9856 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,184B, BPFP=1.6394 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,580B, BPFP=1.9191 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,628B, BPFP=0.3041 +⌛️ [2/4] FRONTEND: Frontend time: 0.171s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10167113 79.80186423 + layer.0.v_cache 0.00001386 0.01042250 + layer.1.k_cache 0.03545215 11.09788944 + layer.1.v_cache 0.00000569 0.00413542 + layer.2.k_cache 0.01047806 1.17344450 + layer.2.v_cache 0.00001755 0.01163314 + layer.3.k_cache 0.06687863 5.61266581 + layer.3.v_cache 0.00001960 0.01345853 + layer.4.k_cache 0.00059181 0.25529015 + layer.4.v_cache 0.00004882 0.02493336 + layer.4.output 0.19633365 692.88043727 + ------------------------------------------------------------------------------------- + TOTAL 0.09350076 291.06875282 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 87124 +BPFP 1.0266 bits/point +EBPFP 2.0533 equivalent bits/point +MSE 291.068753 +---------------------- -------------------------------------------------------- +Time: 0.392s Load: 0.003s, Pack+Encode: 0.171s, Decode+Unpack: 0.218s +---------------------- -------------------------------------------------------- +💾 Converting with 291.0688 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,124B, BPFP=0.5611 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,188B, BPFP=2.0093 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,248B, BPFP=0.9425 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,980B, BPFP=1.9720 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,596B, BPFP=1.0050 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,472B, BPFP=1.8807 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,280B, BPFP=0.9483 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,800B, BPFP=1.9397 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,528B, BPFP=1.5316 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,296B, BPFP=1.8491 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,740B, BPFP=0.2756 +⌛️ [2/4] FRONTEND: Frontend time: 0.172s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11445767 70.97341954 + layer.0.v_cache 0.00001583 0.01018109 + layer.1.k_cache 0.03092440 10.40304811 + layer.1.v_cache 0.00000551 0.00408895 + layer.2.k_cache 0.00639992 1.09004247 + layer.2.v_cache 0.00001897 0.01178295 + layer.3.k_cache 0.05613507 4.97641412 + layer.3.v_cache 0.00002026 0.01349637 + layer.4.k_cache 0.00063356 0.25479994 + layer.4.v_cache 0.00005123 0.02527420 + layer.4.output 0.16055210 618.96946839 + ------------------------------------------------------------------------------------- + TOTAL 0.07838395 260.03228391 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 92252 +BPFP 0.9746 bits/point +EBPFP 1.9492 equivalent bits/point +MSE 260.032284 +---------------------- -------------------------------------------------------- +Time: 0.387s Load: 0.004s, Pack+Encode: 0.172s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 260.0323 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,984B, BPFP=0.5551 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,132B, BPFP=2.0707 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,172B, BPFP=0.9621 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,996B, BPFP=2.0454 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,492B, BPFP=1.0216 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,212B, BPFP=1.8996 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,200B, BPFP=0.9673 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,604B, BPFP=1.9725 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,636B, BPFP=1.6064 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,952B, BPFP=1.8512 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,072B, BPFP=0.2676 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13956541 72.12895275 + layer.0.v_cache 0.00001349 0.01047779 + layer.1.k_cache 0.03336124 10.29785738 + layer.1.v_cache 0.00000558 0.00443795 + layer.2.k_cache 0.00698305 1.12989644 + layer.2.v_cache 0.00001831 0.01180399 + layer.3.k_cache 0.07289400 5.14147949 + layer.3.v_cache 0.00001855 0.01343489 + layer.4.k_cache 0.00061006 0.26816177 + layer.4.v_cache 0.00005429 0.02769195 + layer.4.output 0.17891866 643.99994685 + ------------------------------------------------------------------------------------- + TOTAL 0.08858556 270.41375426 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 90452 +BPFP 0.9897 bits/point +EBPFP 1.9794 equivalent bits/point +MSE 270.413754 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 270.4138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,884B, BPFP=0.5563 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,828B, BPFP=2.0887 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,872B, BPFP=0.9398 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,464B, BPFP=2.0185 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,524B, BPFP=1.0656 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,904B, BPFP=1.9105 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,144B, BPFP=0.9923 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,936B, BPFP=1.9167 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,140B, BPFP=1.5702 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,636B, BPFP=1.8588 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,524B, BPFP=0.2900 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10837562 73.52281057 + layer.0.v_cache 0.00001644 0.00971977 + layer.1.k_cache 0.01241452 11.04627294 + layer.1.v_cache 0.00000532 0.00400551 + layer.2.k_cache 0.01724591 1.13033116 + layer.2.v_cache 0.00001958 0.01090285 + layer.3.k_cache 0.04641535 4.47386188 + layer.3.v_cache 0.00001847 0.01185499 + layer.4.k_cache 0.00063678 0.23803080 + layer.4.v_cache 0.00004974 0.02348976 + layer.4.output 0.18014439 668.84926146 + ------------------------------------------------------------------------------------- + TOTAL 0.08507109 280.73035944 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 87856 +BPFP 0.9969 bits/point +EBPFP 1.9938 equivalent bits/point +MSE 280.730359 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 280.7304 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.9860 bits/point +Avg EBPFP 1.9720 equivalent bits/point +Avg MSE 264.096001 +Avg Time 0.369s +------------------------ ---------------------------- diff --git a/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..59770e239d52817fa21d7ed681dd176309b70c37 --- /dev/null +++ b/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 599 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean +Output output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,744B, BPFP=0.5525 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,348B, BPFP=1.8799 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,624B, BPFP=0.8293 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,384B, BPFP=1.8111 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,140B, BPFP=0.9375 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,684B, BPFP=1.7611 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,872B, BPFP=0.8470 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,920B, BPFP=1.7780 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,172B, BPFP=1.4392 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,724B, BPFP=1.6926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,596B, BPFP=0.2813 +⌛️ [2/4] FRONTEND: Frontend time: 0.464s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.452s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12261864 63.21074130 + layer.0.v_cache 0.00001806 0.01170276 + layer.1.k_cache 0.30198812 10.13442227 + layer.1.v_cache 0.00000607 0.00442059 + layer.2.k_cache 0.01966049 0.97409113 + layer.2.v_cache 0.00002021 0.01205195 + layer.3.k_cache 0.01413035 3.75260918 + layer.3.v_cache 0.00002054 0.01353598 + layer.4.k_cache 0.00067868 0.26434149 + layer.4.v_cache 0.00004987 0.02472758 + layer.4.output 1.39792533 246.93211840 + ------------------------------------------------------------------------------------- + TOTAL 0.60262755 106.28985135 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 217208 +BPFP 0.9116 bits/point +EBPFP 1.8232 equivalent bits/point +MSE 106.289851 +---------------------- -------------------------------------------------------- +Time: 0.925s Load: 0.009s, Pack+Encode: 0.464s, Decode+Unpack: 0.452s +---------------------- -------------------------------------------------------- +💾 Converting with 106.2899 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,988B, BPFP=0.5778 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,968B, BPFP=1.8785 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,284B, BPFP=0.8163 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,224B, BPFP=1.8247 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,996B, BPFP=0.9401 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,524B, BPFP=1.7740 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,020B, BPFP=0.8695 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,868B, BPFP=1.7989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,204B, BPFP=1.4615 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,784B, BPFP=1.7205 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,596B, BPFP=0.2748 +⌛️ [2/4] FRONTEND: Frontend time: 0.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11787021 61.20408348 + layer.0.v_cache 0.00001738 0.01106207 + layer.1.k_cache 0.25265321 9.93490035 + layer.1.v_cache 0.00000586 0.00452503 + layer.2.k_cache 0.01010228 0.97881035 + layer.2.v_cache 0.00002017 0.01259529 + layer.3.k_cache 0.01159736 3.96087618 + layer.3.v_cache 0.00002197 0.01428614 + layer.4.k_cache 0.00067244 0.27433279 + layer.4.v_cache 0.00004845 0.02490192 + layer.4.output 1.41728832 250.43057622 + ------------------------------------------------------------------------------------- + TOTAL 0.60670750 107.61378866 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 215456 +BPFP 0.9168 bits/point +EBPFP 1.8336 equivalent bits/point +MSE 107.613789 +---------------------- -------------------------------------------------------- +Time: 0.571s Load: 0.010s, Pack+Encode: 0.231s, Decode+Unpack: 0.330s +---------------------- -------------------------------------------------------- +💾 Converting with 107.6138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,856B, BPFP=0.5504 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,228B, BPFP=1.8377 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,368B, BPFP=0.7965 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,468B, BPFP=1.7845 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,152B, BPFP=0.9215 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,632B, BPFP=1.7259 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,224B, BPFP=0.8565 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,192B, BPFP=1.7651 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,312B, BPFP=1.4232 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,924B, BPFP=1.6763 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,360B, BPFP=0.2238 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13311632 67.52582855 + layer.0.v_cache 0.00001526 0.01096874 + layer.1.k_cache 0.36474110 9.72226212 + layer.1.v_cache 0.00000573 0.00442028 + layer.2.k_cache 0.01296863 1.00875362 + layer.2.v_cache 0.00001984 0.01216878 + layer.3.k_cache 0.01490937 4.04124909 + layer.3.v_cache 0.00002017 0.01401917 + layer.4.k_cache 0.00067370 0.26118644 + layer.4.v_cache 0.00005228 0.02462326 + layer.4.output 1.37278975 242.41427771 + ------------------------------------------------------------------------------------- + TOTAL 0.59623828 104.67796612 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 212716 +BPFP 0.8767 bits/point +EBPFP 1.7535 equivalent bits/point +MSE 104.677966 +---------------------- -------------------------------------------------------- +Time: 0.551s Load: 0.008s, Pack+Encode: 0.224s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 104.6780 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,440B, BPFP=0.5405 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,120B, BPFP=1.8007 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,480B, BPFP=0.7992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,404B, BPFP=1.7549 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,500B, BPFP=0.9285 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,716B, BPFP=1.7108 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,368B, BPFP=0.8560 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,076B, BPFP=1.7339 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,016B, BPFP=1.4098 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,752B, BPFP=1.6491 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,652B, BPFP=0.2713 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11263794 63.14686860 + layer.0.v_cache 0.00001639 0.01111428 + layer.1.k_cache 0.42101050 9.66182721 + layer.1.v_cache 0.00000577 0.00431621 + layer.2.k_cache 0.00499437 0.95717771 + layer.2.v_cache 0.00001973 0.01193520 + layer.3.k_cache 0.03029772 3.61534169 + layer.3.v_cache 0.00002073 0.01381278 + layer.4.k_cache 0.00068836 0.25897858 + layer.4.v_cache 0.00005112 0.02435341 + layer.4.output 1.25471843 221.34852532 + ------------------------------------------------------------------------------------- + TOTAL 0.55016304 95.71443547 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 235524 +BPFP 0.8872 bits/point +EBPFP 1.7744 equivalent bits/point +MSE 95.714435 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.009s, Pack+Encode: 0.227s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 95.7144 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,804B, BPFP=0.5593 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,212B, BPFP=1.8787 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,292B, BPFP=0.8093 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,320B, BPFP=1.8148 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,044B, BPFP=0.9349 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,708B, BPFP=1.7709 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,084B, BPFP=0.8661 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,176B, BPFP=1.8045 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,188B, BPFP=1.4470 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,012B, BPFP=1.7210 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,140B, BPFP=0.2677 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10233516 65.86385464 + layer.0.v_cache 0.00001822 0.01144272 + layer.1.k_cache 0.31275415 10.08483999 + layer.1.v_cache 0.00000587 0.00419339 + layer.2.k_cache 0.01225604 0.96627234 + layer.2.v_cache 0.00002018 0.01199178 + layer.3.k_cache 0.02847941 4.22989340 + layer.3.v_cache 0.00002101 0.01362194 + layer.4.k_cache 0.00067306 0.26332400 + layer.4.v_cache 0.00005947 0.02482716 + layer.4.output 1.40430263 248.22108453 + ------------------------------------------------------------------------------------- + TOTAL 0.60510241 107.00128548 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 215980 +BPFP 0.9106 bits/point +EBPFP 1.8212 equivalent bits/point +MSE 107.001285 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.008s, Pack+Encode: 0.225s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 107.0013 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,076B, BPFP=0.5355 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,312B, BPFP=1.8236 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,160B, BPFP=0.8057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,308B, BPFP=1.7702 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,380B, BPFP=0.9237 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,228B, BPFP=1.7128 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,140B, BPFP=0.8578 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,668B, BPFP=1.7362 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,392B, BPFP=1.4026 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,120B, BPFP=1.6539 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,108B, BPFP=0.2741 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.394s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16298727 63.21457802 + layer.0.v_cache 0.00001770 0.01105220 + layer.1.k_cache 0.61736183 10.01647285 + layer.1.v_cache 0.00000612 0.00434054 + layer.2.k_cache 0.02593859 0.95709322 + layer.2.v_cache 0.00002024 0.01182712 + layer.3.k_cache 0.01326758 3.74293611 + layer.3.v_cache 0.00002270 0.01374178 + layer.4.k_cache 0.00075427 0.26423809 + layer.4.v_cache 0.00005043 0.02319278 + layer.4.output 0.04538776 184.40023688 + ------------------------------------------------------------------------------------- + TOTAL 0.06694947 80.53300770 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 284892 +BPFP 0.8906 bits/point +EBPFP 1.7813 equivalent bits/point +MSE 80.533008 +---------------------- -------------------------------------------------------- +Time: 0.746s Load: 0.010s, Pack+Encode: 0.342s, Decode+Unpack: 0.394s +---------------------- -------------------------------------------------------- +💾 Converting with 80.5330 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 239, 128) +Output shape: (1, 239, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.output: torch.Size([1, 239, 3584]) -> torch.Size([1, 1, 239, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,392B, BPFP=0.5486 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,960B, BPFP=1.8279 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,512B, BPFP=0.8180 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,092B, BPFP=1.7712 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,092B, BPFP=0.9213 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,208B, BPFP=1.7134 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,096B, BPFP=0.8562 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,560B, BPFP=1.7364 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,516B, BPFP=1.4066 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,308B, BPFP=1.6546 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,084B, BPFP=0.2810 +⌛️ [2/4] FRONTEND: Frontend time: 0.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11863919 59.05405825 + layer.0.v_cache 0.00001702 0.01146524 + layer.1.k_cache 0.43747612 9.98855808 + layer.1.v_cache 0.00000797 0.00453265 + layer.2.k_cache 0.02286501 0.96656914 + layer.2.v_cache 0.00002174 0.01173958 + layer.3.k_cache 0.01903599 3.78168210 + layer.3.v_cache 0.00002123 0.01347660 + layer.4.k_cache 0.00068597 0.26399288 + layer.4.v_cache 0.00006921 0.02373538 + layer.4.output 1.28101462 226.05065750 + ------------------------------------------------------------------------------------- + TOTAL 0.56270246 97.43967132 + (elements=2,080,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2080256 +Total Bytes 232820 +BPFP 0.8954 bits/point +EBPFP 1.7907 equivalent bits/point +MSE 97.439671 +---------------------- -------------------------------------------------------- +Time: 0.561s Load: 0.010s, Pack+Encode: 0.231s, Decode+Unpack: 0.320s +---------------------- -------------------------------------------------------- +💾 Converting with 97.4397 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,492B, BPFP=0.5045 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,084B, BPFP=1.9061 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,724B, BPFP=0.8154 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,752B, BPFP=1.8270 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,760B, BPFP=0.9363 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,856B, BPFP=1.7738 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,296B, BPFP=0.8493 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,308B, BPFP=1.8006 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,516B, BPFP=1.4565 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,976B, BPFP=1.7215 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,796B, BPFP=0.2868 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15151563 63.69549073 + layer.0.v_cache 0.00001631 0.01125239 + layer.1.k_cache 0.46755480 10.05296645 + layer.1.v_cache 0.00000667 0.00453377 + layer.2.k_cache 0.02171374 1.00801684 + layer.2.v_cache 0.00002115 0.01231161 + layer.3.k_cache 0.01385514 3.86985727 + layer.3.v_cache 0.00002107 0.01373509 + layer.4.k_cache 0.00069081 0.26669990 + layer.4.v_cache 0.00005638 0.02555282 + layer.4.output 0.00504555 210.96425177 + ------------------------------------------------------------------------------------- + TOTAL 0.04063356 91.51236348 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 262560 +BPFP 0.9176 bits/point +EBPFP 1.8352 equivalent bits/point +MSE 91.512363 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.011s, Pack+Encode: 0.258s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 91.5124 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,872B, BPFP=0.5467 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,436B, BPFP=1.8358 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,648B, BPFP=0.8089 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,660B, BPFP=1.7819 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,008B, BPFP=0.9033 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,700B, BPFP=1.7153 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,032B, BPFP=0.8356 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,088B, BPFP=1.7422 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,128B, BPFP=1.3978 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,936B, BPFP=1.6622 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,496B, BPFP=0.2926 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14529029 64.52283420 + layer.0.v_cache 0.00001764 0.01127020 + layer.1.k_cache 0.31089128 10.11457574 + layer.1.v_cache 0.00000630 0.00442940 + layer.2.k_cache 0.01355181 0.98588270 + layer.2.v_cache 0.00002032 0.01225606 + layer.3.k_cache 0.02426839 3.88717801 + layer.3.v_cache 0.00002043 0.01335726 + layer.4.k_cache 0.00068925 0.25375888 + layer.4.v_cache 0.00005165 0.02368824 + layer.4.output 1.36066019 240.56343254 + ------------------------------------------------------------------------------------- + TOTAL 0.58937816 103.75136815 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 220004 +BPFP 0.8987 bits/point +EBPFP 1.7974 equivalent bits/point +MSE 103.751368 +---------------------- -------------------------------------------------------- +Time: 0.559s Load: 0.009s, Pack+Encode: 0.230s, Decode+Unpack: 0.320s +---------------------- -------------------------------------------------------- +💾 Converting with 103.7514 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 262, 128) +Output shape: (1, 262, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.output: torch.Size([1, 262, 3584]) -> torch.Size([1, 1, 262, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,572B, BPFP=0.5112 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,832B, BPFP=1.8984 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,596B, BPFP=0.8108 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,596B, BPFP=1.8247 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,784B, BPFP=0.9413 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,880B, BPFP=1.7820 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,576B, BPFP=0.8693 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,520B, BPFP=1.8201 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,556B, BPFP=1.4645 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,996B, BPFP=1.7292 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,544B, BPFP=0.2858 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12917162 62.63107556 + layer.0.v_cache 0.00001830 0.01108042 + layer.1.k_cache 0.40634074 10.12528607 + layer.1.v_cache 0.00000633 0.00450594 + layer.2.k_cache 0.02363680 1.01414350 + layer.2.v_cache 0.00002248 0.01235622 + layer.3.k_cache 0.02343226 3.97190146 + layer.3.v_cache 0.00002042 0.01379701 + layer.4.k_cache 0.00071170 0.26870457 + layer.4.v_cache 0.00005093 0.02455826 + layer.4.output 0.00505940 211.74251977 + ------------------------------------------------------------------------------------- + TOTAL 0.03640161 91.78088514 + (elements=2,280,448) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2280448 +Total Bytes 262452 +BPFP 0.9207 bits/point +EBPFP 1.8414 equivalent bits/point +MSE 91.780885 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.009s, Pack+Encode: 0.256s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 91.7809 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,836B, BPFP=0.5442 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,512B, BPFP=1.8411 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,540B, BPFP=0.8014 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,608B, BPFP=1.7783 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,076B, BPFP=0.9081 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,760B, BPFP=1.7194 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,192B, BPFP=0.8467 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,128B, BPFP=1.7450 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,416B, BPFP=1.4178 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,036B, BPFP=1.6692 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,524B, BPFP=0.2731 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18074451 63.99275608 + layer.0.v_cache 0.00001764 0.01146308 + layer.1.k_cache 0.33421082 10.05507378 + layer.1.v_cache 0.00000655 0.00439686 + layer.2.k_cache 0.01487562 0.95936462 + layer.2.v_cache 0.00002082 0.01236850 + layer.3.k_cache 0.01319740 3.96641710 + layer.3.v_cache 0.00002096 0.01379633 + layer.4.k_cache 0.00067848 0.26452925 + layer.4.v_cache 0.00005239 0.02437273 + layer.4.output 1.36066546 240.55361111 + ------------------------------------------------------------------------------------- + TOTAL 0.59226373 103.71645977 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 218628 +BPFP 0.8931 bits/point +EBPFP 1.7862 equivalent bits/point +MSE 103.716460 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.009s, Pack+Encode: 0.227s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 103.7165 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 232, 128) +Output shape: (1, 232, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.output: torch.Size([1, 232, 3584]) -> torch.Size([1, 1, 232, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,064B, BPFP=0.5431 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,360B, BPFP=1.8427 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,140B, BPFP=0.8176 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,432B, BPFP=1.7802 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,760B, BPFP=0.9267 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,800B, BPFP=1.7376 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,772B, BPFP=0.8602 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,112B, BPFP=1.7586 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,416B, BPFP=1.4423 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,984B, BPFP=1.6827 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,992B, BPFP=0.2789 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11978169 63.85493838 + layer.0.v_cache 0.00001746 0.01122608 + layer.1.k_cache 0.32299256 10.11132497 + layer.1.v_cache 0.00000596 0.00419547 + layer.2.k_cache 0.02091339 0.95420910 + layer.2.v_cache 0.00002092 0.01200610 + layer.3.k_cache 0.01769928 3.68754788 + layer.3.v_cache 0.00002180 0.01388431 + layer.4.k_cache 0.00071208 0.26802778 + layer.4.v_cache 0.00005763 0.02473054 + layer.4.output 1.31963615 233.37332589 + ------------------------------------------------------------------------------------- + TOTAL 0.57174564 100.73855129 + (elements=2,019,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2019328 +Total Bytes 227832 +BPFP 0.9026 bits/point +EBPFP 1.8052 equivalent bits/point +MSE 100.738551 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.010s, Pack+Encode: 0.222s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 100.7386 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,856B, BPFP=0.5580 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,348B, BPFP=1.8713 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,776B, BPFP=0.8364 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,516B, BPFP=1.8122 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,020B, BPFP=0.9247 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,500B, BPFP=1.7401 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,228B, BPFP=0.8685 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,984B, BPFP=1.7744 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,272B, BPFP=1.4398 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,876B, BPFP=1.6957 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,920B, BPFP=0.2833 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15639258 65.07821378 + layer.0.v_cache 0.00001700 0.01188418 + layer.1.k_cache 0.33150226 10.45162908 + layer.1.v_cache 0.00000613 0.00445237 + layer.2.k_cache 0.01694729 1.02055955 + layer.2.v_cache 0.00002208 0.01192992 + layer.3.k_cache 0.03858858 4.29070435 + layer.3.v_cache 0.00002090 0.01414309 + layer.4.k_cache 0.00067984 0.26368693 + layer.4.v_cache 0.00005217 0.02535364 + layer.4.output 1.39161012 245.83084416 + ------------------------------------------------------------------------------------- + TOTAL 0.60502939 105.99932153 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 218296 +BPFP 0.9120 bits/point +EBPFP 1.8240 equivalent bits/point +MSE 105.999322 +---------------------- -------------------------------------------------------- +Time: 0.555s Load: 0.010s, Pack+Encode: 0.226s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 105.9993 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,652B, BPFP=0.5524 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,660B, BPFP=1.8693 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,248B, BPFP=0.8155 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,524B, BPFP=1.8043 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,988B, BPFP=0.9151 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,488B, BPFP=1.7450 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,848B, BPFP=0.8498 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,756B, BPFP=1.7603 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,896B, BPFP=1.4249 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,404B, BPFP=1.6829 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,204B, BPFP=0.3124 +⌛️ [2/4] FRONTEND: Frontend time: 0.314s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.394s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13605902 63.04278274 + layer.0.v_cache 0.00001757 0.01099565 + layer.1.k_cache 0.58683352 10.07653853 + layer.1.v_cache 0.00000623 0.00438011 + layer.2.k_cache 0.01500385 0.98832518 + layer.2.v_cache 0.00002229 0.01213374 + layer.3.k_cache 0.00978726 4.24371271 + layer.3.v_cache 0.00001998 0.01279725 + layer.4.k_cache 0.00067570 0.25744567 + layer.4.v_cache 0.00005472 0.02418759 + layer.4.output 0.00490606 203.17121272 + ------------------------------------------------------------------------------------- + TOTAL 0.04604839 88.28657578 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 272668 +BPFP 0.9180 bits/point +EBPFP 1.8360 equivalent bits/point +MSE 88.286576 +---------------------- -------------------------------------------------------- +Time: 0.717s Load: 0.010s, Pack+Encode: 0.314s, Decode+Unpack: 0.394s +---------------------- -------------------------------------------------------- +💾 Converting with 88.2866 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,352B, BPFP=0.5118 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,348B, BPFP=1.7370 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,432B, BPFP=0.7618 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,460B, BPFP=1.6826 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,404B, BPFP=0.8826 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,644B, BPFP=1.6326 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,480B, BPFP=0.8260 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,064B, BPFP=1.6583 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,924B, BPFP=1.3434 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,672B, BPFP=1.5730 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,888B, BPFP=0.3317 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12967140 58.38170190 + layer.0.v_cache 0.00001601 0.01094245 + layer.1.k_cache 0.49353925 9.45260608 + layer.1.v_cache 0.00000620 0.00429483 + layer.2.k_cache 0.03189284 0.92479308 + layer.2.v_cache 0.00002040 0.01145182 + layer.3.k_cache 0.04285583 3.42855679 + layer.3.v_cache 0.00002221 0.01323984 + layer.4.k_cache 0.00071132 0.24939934 + layer.4.v_cache 0.00005107 0.02281055 + layer.4.output 1.20063613 210.57909664 + ------------------------------------------------------------------------------------- + TOTAL 0.53548467 90.97373372 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 243668 +BPFP 0.8783 bits/point +EBPFP 1.7565 equivalent bits/point +MSE 90.973734 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.008s, Pack+Encode: 0.227s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 90.9737 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,596B, BPFP=0.5394 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,276B, BPFP=1.7743 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,584B, BPFP=0.7897 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,508B, BPFP=1.7262 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,500B, BPFP=0.9099 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,716B, BPFP=1.6765 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,412B, BPFP=0.8416 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,096B, BPFP=1.7003 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,124B, BPFP=1.3883 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,736B, BPFP=1.6150 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,420B, BPFP=0.2727 +⌛️ [2/4] FRONTEND: Frontend time: 0.285s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.331s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16529650 60.34238909 + layer.0.v_cache 0.00001613 0.01144043 + layer.1.k_cache 0.54577502 9.81757793 + layer.1.v_cache 0.00000662 0.00461351 + layer.2.k_cache 0.01453520 0.96034259 + layer.2.v_cache 0.00002096 0.01294316 + layer.3.k_cache 0.02409829 3.99344326 + layer.3.v_cache 0.00002071 0.01410901 + layer.4.k_cache 0.00068699 0.27135648 + layer.4.v_cache 0.00004971 0.02458317 + layer.4.output 1.22953407 217.23684022 + ------------------------------------------------------------------------------------- + TOTAL 0.55042615 93.88886354 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 236968 +BPFP 0.8747 bits/point +EBPFP 1.7494 equivalent bits/point +MSE 93.888864 +---------------------- -------------------------------------------------------- +Time: 0.624s Load: 0.009s, Pack+Encode: 0.285s, Decode+Unpack: 0.331s +---------------------- -------------------------------------------------------- +💾 Converting with 93.8889 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,780B, BPFP=0.5151 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,096B, BPFP=1.8681 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,752B, BPFP=0.8005 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,844B, BPFP=1.8083 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,464B, BPFP=0.9300 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,652B, BPFP=1.7513 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,716B, BPFP=0.8465 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,092B, BPFP=1.7724 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,896B, BPFP=1.4285 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,332B, BPFP=1.6883 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,960B, BPFP=0.3069 +⌛️ [2/4] FRONTEND: Frontend time: 0.346s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.459s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19043710 61.65361000 + layer.0.v_cache 0.00001603 0.01118022 + layer.1.k_cache 0.64205214 9.90991584 + layer.1.v_cache 0.00000630 0.00455139 + layer.2.k_cache 0.01513053 0.99812611 + layer.2.v_cache 0.00002191 0.01264209 + layer.3.k_cache 0.02775778 3.64635596 + layer.3.v_cache 0.00002125 0.01352977 + layer.4.k_cache 0.00071623 0.25950938 + layer.4.v_cache 0.00005146 0.02401425 + layer.4.output 0.04089398 165.83637779 + ------------------------------------------------------------------------------------- + TOTAL 0.06838051 72.78753409 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 325584 +BPFP 0.9151 bits/point +EBPFP 1.8303 equivalent bits/point +MSE 72.787534 +---------------------- -------------------------------------------------------- +Time: 0.817s Load: 0.012s, Pack+Encode: 0.346s, Decode+Unpack: 0.459s +---------------------- -------------------------------------------------------- +💾 Converting with 72.7875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,820B, BPFP=0.5422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,300B, BPFP=1.8386 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,508B, BPFP=0.8010 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,264B, BPFP=1.7814 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,720B, BPFP=0.9231 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,260B, BPFP=1.7259 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,340B, BPFP=0.8470 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,652B, BPFP=1.7476 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,480B, BPFP=1.4068 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,180B, BPFP=1.6663 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,156B, BPFP=0.3010 +⌛️ [2/4] FRONTEND: Frontend time: 0.273s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14392723 62.82295578 + layer.0.v_cache 0.00001682 0.01129879 + layer.1.k_cache 0.53349789 9.99329950 + layer.1.v_cache 0.00000638 0.00439659 + layer.2.k_cache 0.01523476 0.94790380 + layer.2.v_cache 0.00002109 0.01183071 + layer.3.k_cache 0.03867486 3.90736416 + layer.3.v_cache 0.00001994 0.01309890 + layer.4.k_cache 0.00070525 0.25180709 + layer.4.v_cache 0.00005078 0.02314813 + layer.4.output 0.00472849 195.77900997 + ------------------------------------------------------------------------------------- + TOTAL 0.04501497 85.20236313 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 278680 +BPFP 0.9051 bits/point +EBPFP 1.8102 equivalent bits/point +MSE 85.202363 +---------------------- -------------------------------------------------------- +Time: 0.662s Load: 0.010s, Pack+Encode: 0.273s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2024 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,800B, BPFP=0.5417 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,488B, BPFP=1.8394 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,588B, BPFP=0.8047 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,776B, BPFP=1.7900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,084B, BPFP=0.9086 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,984B, BPFP=1.7350 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,204B, BPFP=0.8475 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,300B, BPFP=1.7569 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,496B, BPFP=1.4233 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,136B, BPFP=1.6761 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,252B, BPFP=0.2803 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14673459 62.72482639 + layer.0.v_cache 0.00001655 0.01130918 + layer.1.k_cache 0.37581726 9.74045573 + layer.1.v_cache 0.00000603 0.00442958 + layer.2.k_cache 0.01062915 0.96955804 + layer.2.v_cache 0.00002006 0.01241640 + layer.3.k_cache 0.04027582 4.21674859 + layer.3.v_cache 0.00002066 0.01378730 + layer.4.k_cache 0.00066373 0.25551446 + layer.4.v_cache 0.00004861 0.02413788 + layer.4.output 1.36064577 240.56240079 + ------------------------------------------------------------------------------------- + TOTAL 0.59404429 103.64176406 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 220108 +BPFP 0.8991 bits/point +EBPFP 1.7983 equivalent bits/point +MSE 103.641764 +---------------------- -------------------------------------------------------- +Time: 0.551s Load: 0.007s, Pack+Encode: 0.224s, Decode+Unpack: 0.320s +---------------------- -------------------------------------------------------- +💾 Converting with 103.6418 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,312B, BPFP=0.5550 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,040B, BPFP=1.8723 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,364B, BPFP=0.8256 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,140B, BPFP=1.8122 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,124B, BPFP=0.9431 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,336B, BPFP=1.7585 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,220B, BPFP=0.8827 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,788B, BPFP=1.7887 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,716B, BPFP=1.4501 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,416B, BPFP=1.6971 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,544B, BPFP=0.2914 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12385076 60.82576289 + layer.0.v_cache 0.00002004 0.01207456 + layer.1.k_cache 0.38583452 9.92003456 + layer.1.v_cache 0.00000646 0.00467454 + layer.2.k_cache 0.02500911 0.97434619 + layer.2.v_cache 0.00002014 0.01232552 + layer.3.k_cache 0.03425702 3.45760013 + layer.3.v_cache 0.00002057 0.01411702 + layer.4.k_cache 0.00068928 0.27280943 + layer.4.v_cache 0.00004882 0.02417176 + layer.4.output 1.30836928 231.25076313 + ------------------------------------------------------------------------------------- + TOTAL 0.57225539 99.66313285 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 234000 +BPFP 0.9191 bits/point +EBPFP 1.8382 equivalent bits/point +MSE 99.663133 +---------------------- -------------------------------------------------------- +Time: 0.591s Load: 0.009s, Pack+Encode: 0.223s, Decode+Unpack: 0.359s +---------------------- -------------------------------------------------------- +💾 Converting with 99.6631 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 247, 128) +Output shape: (1, 247, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.output: torch.Size([1, 247, 3584]) -> torch.Size([1, 1, 247, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,460B, BPFP=0.5352 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,384B, BPFP=1.7955 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,564B, BPFP=0.7948 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,388B, BPFP=1.7325 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,276B, BPFP=0.9031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,456B, BPFP=1.6736 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,172B, BPFP=0.8332 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,884B, BPFP=1.7007 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,872B, BPFP=1.3836 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,520B, BPFP=1.6144 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,852B, BPFP=0.2969 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.329s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13907220 64.24066137 + layer.0.v_cache 0.00001735 0.01107669 + layer.1.k_cache 0.47208408 10.29817082 + layer.1.v_cache 0.00000611 0.00414811 + layer.2.k_cache 0.03408500 0.99924244 + layer.2.v_cache 0.00001989 0.01106341 + layer.3.k_cache 0.02387486 3.82592304 + layer.3.v_cache 0.00002031 0.01258707 + layer.4.k_cache 0.00069623 0.25100164 + layer.4.v_cache 0.00004922 0.02174790 + layer.4.output 1.23951819 218.88542872 + ------------------------------------------------------------------------------------- + TOTAL 0.54979721 94.81609550 + (elements=2,149,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2149888 +Total Bytes 237828 +BPFP 0.8850 bits/point +EBPFP 1.7700 equivalent bits/point +MSE 94.816096 +---------------------- -------------------------------------------------------- +Time: 0.565s Load: 0.008s, Pack+Encode: 0.228s, Decode+Unpack: 0.329s +---------------------- -------------------------------------------------------- +💾 Converting with 94.8161 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 427, 128) +Output shape: (1, 427, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.output: torch.Size([1, 427, 3584]) -> torch.Size([1, 1, 427, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,708B, BPFP=0.5382 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,436B, BPFP=1.7724 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,596B, BPFP=0.7903 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,828B, BPFP=1.7136 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,880B, BPFP=0.8738 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,024B, BPFP=1.6475 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,024B, BPFP=0.8059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,568B, BPFP=1.6674 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,944B, BPFP=1.3519 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,712B, BPFP=1.5995 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 57,440B, BPFP=0.3003 +⌛️ [2/4] FRONTEND: Frontend time: 0.384s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.520s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15738957 61.33734631 + layer.0.v_cache 0.00001693 0.01038630 + layer.1.k_cache 1.00627419 10.02598640 + layer.1.v_cache 0.00000663 0.00406785 + layer.2.k_cache 0.03590691 0.89681885 + layer.2.v_cache 0.00002108 0.01064512 + layer.3.k_cache 0.02176077 3.42318804 + layer.3.v_cache 0.00002004 0.01177075 + layer.4.k_cache 0.00078997 0.24367529 + layer.4.v_cache 0.00005212 0.02190608 + layer.4.output 0.00621442 129.65139888 + ------------------------------------------------------------------------------------- + TOTAL 0.07445524 57.85562254 + (elements=3,716,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3716608 +Total Bytes 406160 +BPFP 0.8743 bits/point +EBPFP 1.7485 equivalent bits/point +MSE 57.855623 +---------------------- -------------------------------------------------------- +Time: 0.918s Load: 0.014s, Pack+Encode: 0.384s, Decode+Unpack: 0.520s +---------------------- -------------------------------------------------------- +💾 Converting with 57.8556 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,660B, BPFP=0.5570 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,112B, BPFP=1.8515 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,996B, BPFP=0.8070 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,084B, BPFP=1.7922 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,648B, BPFP=0.9022 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,940B, BPFP=1.7262 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,548B, BPFP=0.8388 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,216B, BPFP=1.7422 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,476B, BPFP=1.4112 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,840B, BPFP=1.6628 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,464B, BPFP=0.2921 +⌛️ [2/4] FRONTEND: Frontend time: 0.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.396s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15504150 64.68803333 + layer.0.v_cache 0.00001776 0.01041129 + layer.1.k_cache 0.59434689 10.12188023 + layer.1.v_cache 0.00000612 0.00400183 + layer.2.k_cache 0.01687181 0.96933659 + layer.2.v_cache 0.00002031 0.01082847 + layer.3.k_cache 0.01346795 3.97185806 + layer.3.v_cache 0.00002023 0.01202438 + layer.4.k_cache 0.00071492 0.24603516 + layer.4.v_cache 0.00006004 0.02196134 + layer.4.output 0.00488892 204.50497496 + ------------------------------------------------------------------------------------- + TOTAL 0.04792882 88.91712914 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 265984 +BPFP 0.9021 bits/point +EBPFP 1.8042 equivalent bits/point +MSE 88.917129 +---------------------- -------------------------------------------------------- +Time: 0.679s Load: 0.016s, Pack+Encode: 0.267s, Decode+Unpack: 0.396s +---------------------- -------------------------------------------------------- +💾 Converting with 88.9171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 361, 128) +Output shape: (1, 361, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.output: torch.Size([1, 361, 3584]) -> torch.Size([1, 1, 361, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,512B, BPFP=0.5416 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,636B, BPFP=1.7588 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,116B, BPFP=0.7841 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,412B, BPFP=1.7059 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,944B, BPFP=0.8632 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,276B, BPFP=1.6567 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,732B, BPFP=0.8108 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,516B, BPFP=1.6671 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,420B, BPFP=1.3599 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,788B, BPFP=1.5923 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,244B, BPFP=0.2736 +⌛️ [2/4] FRONTEND: Frontend time: 0.346s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.591s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21057716 65.03339249 + layer.0.v_cache 0.00001554 0.00975688 + layer.1.k_cache 0.82271198 10.07718022 + layer.1.v_cache 0.00000592 0.00396234 + layer.2.k_cache 0.01184860 0.95909968 + layer.2.v_cache 0.00001987 0.01124883 + layer.3.k_cache 0.01345186 3.72206889 + layer.3.v_cache 0.00002111 0.01198148 + layer.4.k_cache 0.00076976 0.24916709 + layer.4.v_cache 0.00005077 0.02190069 + layer.4.output 0.03703259 150.13355758 + ------------------------------------------------------------------------------------- + TOTAL 0.07757063 66.53145068 + (elements=3,142,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3142144 +Total Bytes 338596 +BPFP 0.8621 bits/point +EBPFP 1.7242 equivalent bits/point +MSE 66.531451 +---------------------- -------------------------------------------------------- +Time: 0.951s Load: 0.013s, Pack+Encode: 0.346s, Decode+Unpack: 0.591s +---------------------- -------------------------------------------------------- +💾 Converting with 66.5315 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,708B, BPFP=0.5379 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,460B, BPFP=1.7985 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,488B, BPFP=0.8027 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,732B, BPFP=1.7582 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,892B, BPFP=0.8805 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,628B, BPFP=1.6970 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,760B, BPFP=0.8178 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,708B, BPFP=1.7015 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,208B, BPFP=1.3967 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,428B, BPFP=1.6305 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,632B, BPFP=0.3058 +⌛️ [2/4] FRONTEND: Frontend time: 0.352s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.484s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21757852 66.45813941 + layer.0.v_cache 0.00001544 0.00952360 + layer.1.k_cache 0.59705023 10.03869975 + layer.1.v_cache 0.00000603 0.00376038 + layer.2.k_cache 0.00787334 0.92186288 + layer.2.v_cache 0.00001970 0.01052676 + layer.3.k_cache 0.01914395 3.69345547 + layer.3.v_cache 0.00001916 0.01135512 + layer.4.k_cache 0.00071861 0.23996402 + layer.4.v_cache 0.00005003 0.02154173 + layer.4.output 0.00471288 196.66004939 + ------------------------------------------------------------------------------------- + TOTAL 0.05149795 85.76642205 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 273644 +BPFP 0.8919 bits/point +EBPFP 1.7838 equivalent bits/point +MSE 85.766422 +---------------------- -------------------------------------------------------- +Time: 0.845s Load: 0.009s, Pack+Encode: 0.352s, Decode+Unpack: 0.484s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7664 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 372, 128) +Output shape: (1, 372, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.output: torch.Size([1, 372, 3584]) -> torch.Size([1, 1, 372, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,620B, BPFP=0.5301 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,760B, BPFP=1.7540 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,640B, BPFP=0.7829 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,572B, BPFP=1.7041 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,700B, BPFP=0.8695 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,148B, BPFP=1.6443 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,264B, BPFP=0.8091 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,380B, BPFP=1.6541 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,992B, BPFP=1.3438 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,796B, BPFP=1.5875 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 56,600B, BPFP=0.3396 +⌛️ [2/4] FRONTEND: Frontend time: 0.412s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.566s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18192377 60.59494708 + layer.0.v_cache 0.00001786 0.01006781 + layer.1.k_cache 0.82695811 10.16550503 + layer.1.v_cache 0.00000687 0.00412101 + layer.2.k_cache 0.01216373 0.85487710 + layer.2.v_cache 0.00002073 0.01086145 + layer.3.k_cache 0.01743993 3.52882763 + layer.3.v_cache 0.00002103 0.01200042 + layer.4.k_cache 0.00069676 0.24435445 + layer.4.v_cache 0.00005544 0.02177101 + layer.4.output 0.03606073 145.44086022 + ------------------------------------------------------------------------------------- + TOTAL 0.07598408 64.32549144 + (elements=3,237,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3237888 +Total Bytes 358472 +BPFP 0.8857 bits/point +EBPFP 1.7714 equivalent bits/point +MSE 64.325491 +---------------------- -------------------------------------------------------- +Time: 0.994s Load: 0.016s, Pack+Encode: 0.412s, Decode+Unpack: 0.566s +---------------------- -------------------------------------------------------- +💾 Converting with 64.3255 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,332B, BPFP=0.4998 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,816B, BPFP=1.8293 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,428B, BPFP=0.7947 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,552B, BPFP=1.7682 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,116B, BPFP=0.8764 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,436B, BPFP=1.7142 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,040B, BPFP=0.8243 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,620B, BPFP=1.7231 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,772B, BPFP=1.3918 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,020B, BPFP=1.6457 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,832B, BPFP=0.3098 +⌛️ [2/4] FRONTEND: Frontend time: 0.375s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.548s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16470559 66.67465049 + layer.0.v_cache 0.00001558 0.01049779 + layer.1.k_cache 0.70577214 9.96233583 + layer.1.v_cache 0.00000663 0.00407170 + layer.2.k_cache 0.01487649 0.93824631 + layer.2.v_cache 0.00002061 0.01096814 + layer.3.k_cache 0.03012228 4.13127547 + layer.3.v_cache 0.00002030 0.01218133 + layer.4.k_cache 0.00074575 0.24622987 + layer.4.v_cache 0.00005065 0.02149160 + layer.4.output 0.04139163 167.83776537 + ------------------------------------------------------------------------------------- + TOTAL 0.07094573 73.93390036 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 314964 +BPFP 0.8963 bits/point +EBPFP 1.7925 equivalent bits/point +MSE 73.933900 +---------------------- -------------------------------------------------------- +Time: 0.933s Load: 0.010s, Pack+Encode: 0.375s, Decode+Unpack: 0.548s +---------------------- -------------------------------------------------------- +💾 Converting with 73.9339 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,704B, BPFP=0.5132 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,628B, BPFP=1.8649 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,624B, BPFP=0.8033 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,696B, BPFP=1.8099 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,484B, BPFP=0.9130 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,708B, BPFP=1.7517 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,280B, BPFP=0.8420 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,948B, BPFP=1.7658 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,188B, BPFP=1.4262 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,744B, BPFP=1.6948 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,788B, BPFP=0.2930 +⌛️ [2/4] FRONTEND: Frontend time: 0.381s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.487s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16585089 62.89379422 + layer.0.v_cache 0.00001549 0.01014183 + layer.1.k_cache 0.52552974 10.04920585 + layer.1.v_cache 0.00000676 0.00409622 + layer.2.k_cache 0.01473994 0.92013227 + layer.2.v_cache 0.00002188 0.01169965 + layer.3.k_cache 0.05129043 3.78189421 + layer.3.v_cache 0.00002012 0.01280872 + layer.4.k_cache 0.00075129 0.26225574 + layer.4.v_cache 0.00005717 0.02271594 + layer.4.output 0.00501330 209.36809299 + ------------------------------------------------------------------------------------- + TOTAL 0.04666922 90.79678798 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 261792 +BPFP 0.9080 bits/point +EBPFP 1.8160 equivalent bits/point +MSE 90.796788 +---------------------- -------------------------------------------------------- +Time: 0.877s Load: 0.009s, Pack+Encode: 0.381s, Decode+Unpack: 0.487s +---------------------- -------------------------------------------------------- +💾 Converting with 90.7968 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 368, 128) +Output shape: (1, 368, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.output: torch.Size([1, 368, 3584]) -> torch.Size([1, 1, 368, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,740B, BPFP=0.5409 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,444B, BPFP=1.7597 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,676B, BPFP=0.7930 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,388B, BPFP=1.7148 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,708B, BPFP=0.8792 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,992B, BPFP=1.6556 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,228B, BPFP=0.8164 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,428B, BPFP=1.6741 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,876B, BPFP=1.3534 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,724B, BPFP=1.6017 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 48,292B, BPFP=0.2929 +⌛️ [2/4] FRONTEND: Frontend time: 0.340s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.531s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14147143 60.87871518 + layer.0.v_cache 0.00001611 0.00988325 + layer.1.k_cache 0.89062865 9.94720923 + layer.1.v_cache 0.00000637 0.00384602 + layer.2.k_cache 0.02201829 0.89864640 + layer.2.v_cache 0.00002088 0.01074948 + layer.3.k_cache 0.00965744 3.70518560 + layer.3.v_cache 0.00002101 0.01193804 + layer.4.k_cache 0.00077426 0.24237969 + layer.4.v_cache 0.00005366 0.02218452 + layer.4.output 0.03638816 147.09364082 + ------------------------------------------------------------------------------------- + TOTAL 0.07761090 65.02271901 + (elements=3,203,072) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3203072 +Total Bytes 349496 +BPFP 0.8729 bits/point +EBPFP 1.7458 equivalent bits/point +MSE 65.022719 +---------------------- -------------------------------------------------------- +Time: 0.883s Load: 0.012s, Pack+Encode: 0.340s, Decode+Unpack: 0.531s +---------------------- -------------------------------------------------------- +💾 Converting with 65.0227 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 453, 128) +Output shape: (1, 453, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.output: torch.Size([1, 453, 3584]) -> torch.Size([1, 1, 453, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,012B, BPFP=0.5178 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 52,360B, BPFP=1.8060 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,600B, BPFP=0.7795 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,384B, BPFP=1.7379 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,288B, BPFP=0.8722 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 48,968B, BPFP=1.6890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,520B, BPFP=0.8113 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 49,564B, BPFP=1.7096 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,904B, BPFP=1.3764 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 47,312B, BPFP=1.6319 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 60,952B, BPFP=0.3003 +⌛️ [2/4] FRONTEND: Frontend time: 0.457s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.589s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13665763 60.16127294 + layer.0.v_cache 0.00001897 0.01046184 + layer.1.k_cache 1.07282356 9.92555597 + layer.1.v_cache 0.00000630 0.00405708 + layer.2.k_cache 0.02241544 0.96808319 + layer.2.v_cache 0.00001979 0.01100333 + layer.3.k_cache 0.01806638 3.54629833 + layer.3.v_cache 0.00002079 0.01268960 + layer.4.k_cache 0.00076567 0.25048911 + layer.4.v_cache 0.00005104 0.02246478 + layer.4.output 0.00584071 122.31980251 + ------------------------------------------------------------------------------------- + TOTAL 0.07598415 54.77358787 + (elements=3,942,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3942912 +Total Bytes 435864 +BPFP 0.8843 bits/point +EBPFP 1.7687 equivalent bits/point +MSE 54.773588 +---------------------- -------------------------------------------------------- +Time: 1.061s Load: 0.016s, Pack+Encode: 0.457s, Decode+Unpack: 0.589s +---------------------- -------------------------------------------------------- +💾 Converting with 54.7736 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 402, 128) +Output shape: (1, 402, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.output: torch.Size([1, 402, 3584]) -> torch.Size([1, 1, 402, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,008B, BPFP=0.5445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,280B, BPFP=1.7988 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,336B, BPFP=0.7904 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,420B, BPFP=1.7265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,352B, BPFP=0.8688 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,956B, BPFP=1.6696 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,548B, BPFP=0.7987 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,608B, BPFP=1.6950 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,284B, BPFP=1.3326 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,240B, BPFP=1.6029 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 54,016B, BPFP=0.2999 +⌛️ [2/4] FRONTEND: Frontend time: 0.335s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.506s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18062694 61.62589397 + layer.0.v_cache 0.00001752 0.01007023 + layer.1.k_cache 0.89169403 9.87323636 + layer.1.v_cache 0.00000593 0.00367223 + layer.2.k_cache 0.03854401 0.89599078 + layer.2.v_cache 0.00001967 0.01020309 + layer.3.k_cache 0.02336847 3.79730012 + layer.3.v_cache 0.00001924 0.01171002 + layer.4.k_cache 0.00085529 0.23673840 + layer.4.v_cache 0.00004959 0.02069969 + layer.4.output 0.00648942 137.76679104 + ------------------------------------------------------------------------------------- + TOTAL 0.06944863 61.22665013 + (elements=3,499,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3499008 +Total Bytes 384048 +BPFP 0.8781 bits/point +EBPFP 1.7561 equivalent bits/point +MSE 61.226650 +---------------------- -------------------------------------------------------- +Time: 0.857s Load: 0.016s, Pack+Encode: 0.335s, Decode+Unpack: 0.506s +---------------------- -------------------------------------------------------- +💾 Converting with 61.2267 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,896B, BPFP=0.5542 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,192B, BPFP=1.8589 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,364B, BPFP=0.8044 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,080B, BPFP=1.7966 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,568B, BPFP=0.9279 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,308B, BPFP=1.7534 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,120B, BPFP=0.8468 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,752B, BPFP=1.7782 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,576B, BPFP=1.4323 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,200B, BPFP=1.6913 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,796B, BPFP=0.3184 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13958113 59.11537438 + layer.0.v_cache 0.00001688 0.01149637 + layer.1.k_cache 0.56864935 10.10736675 + layer.1.v_cache 0.00000621 0.00469665 + layer.2.k_cache 0.01018704 0.92629589 + layer.2.v_cache 0.00002046 0.01244148 + layer.3.k_cache 0.04374025 3.99570566 + layer.3.v_cache 0.00002178 0.01446928 + layer.4.k_cache 0.00068741 0.27075854 + layer.4.v_cache 0.00005272 0.02493684 + layer.4.output 0.00479460 198.73692716 + ------------------------------------------------------------------------------------- + TOTAL 0.04685444 86.21423718 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 279852 +BPFP 0.9219 bits/point +EBPFP 1.8438 equivalent bits/point +MSE 86.214237 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2142 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 381, 128) +Output shape: (1, 381, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.output: torch.Size([1, 381, 3584]) -> torch.Size([1, 1, 381, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,176B, BPFP=0.5404 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,288B, BPFP=1.7343 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,044B, BPFP=0.7810 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,072B, BPFP=1.6844 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,676B, BPFP=0.8889 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,988B, BPFP=1.6399 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,096B, BPFP=0.8241 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,436B, BPFP=1.6583 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,732B, BPFP=1.3424 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,476B, BPFP=1.5779 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 55,000B, BPFP=0.3222 +⌛️ [2/4] FRONTEND: Frontend time: 0.284s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.437s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15713709 61.07482365 + layer.0.v_cache 0.00001719 0.01103172 + layer.1.k_cache 0.86767426 9.91406442 + layer.1.v_cache 0.00000622 0.00439813 + layer.2.k_cache 0.01635739 0.92763489 + layer.2.v_cache 0.00002098 0.01188988 + layer.3.k_cache 0.01709203 3.72606909 + layer.3.v_cache 0.00002073 0.01301945 + layer.4.k_cache 0.00073405 0.25604434 + layer.4.v_cache 0.00005083 0.02324600 + layer.4.output 0.03519400 141.80288011 + ------------------------------------------------------------------------------------- + TOTAL 0.07679228 62.85778720 + (elements=3,316,224) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3316224 +Total Bytes 363984 +BPFP 0.8781 bits/point +EBPFP 1.7561 equivalent bits/point +MSE 62.857787 +---------------------- -------------------------------------------------------- +Time: 0.735s Load: 0.014s, Pack+Encode: 0.284s, Decode+Unpack: 0.437s +---------------------- -------------------------------------------------------- +💾 Converting with 62.8578 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,372B, BPFP=0.5590 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,752B, BPFP=1.8531 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,280B, BPFP=0.8200 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,972B, BPFP=1.8010 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,084B, BPFP=0.9404 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,120B, BPFP=1.7441 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,120B, BPFP=0.8761 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,456B, BPFP=1.7666 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,668B, BPFP=1.4468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,312B, BPFP=1.6902 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,984B, BPFP=0.2765 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.322s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422791 63.59407552 + layer.0.v_cache 0.00001930 0.01133027 + layer.1.k_cache 0.38928963 10.36440805 + layer.1.v_cache 0.00000631 0.00438422 + layer.2.k_cache 0.01623082 0.95440948 + layer.2.v_cache 0.00002095 0.01235377 + layer.3.k_cache 0.01313608 3.69588373 + layer.3.v_cache 0.00002100 0.01380376 + layer.4.k_cache 0.00066994 0.26362016 + layer.4.v_cache 0.00005661 0.02464827 + layer.4.output 1.30836332 231.28558074 + ------------------------------------------------------------------------------------- + TOTAL 0.57013069 99.87870485 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 231120 +BPFP 0.9078 bits/point +EBPFP 1.8156 equivalent bits/point +MSE 99.878705 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.008s, Pack+Encode: 0.223s, Decode+Unpack: 0.322s +---------------------- -------------------------------------------------------- +💾 Converting with 99.8787 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,292B, BPFP=0.5490 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,732B, BPFP=1.8361 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,368B, BPFP=0.8189 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,940B, BPFP=1.7836 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,992B, BPFP=0.9264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,108B, BPFP=1.7285 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,156B, BPFP=0.8710 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,520B, BPFP=1.7558 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,556B, BPFP=1.4272 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,172B, BPFP=1.6666 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,104B, BPFP=0.2564 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13591524 63.26733398 + layer.0.v_cache 0.00001656 0.01160492 + layer.1.k_cache 0.34520291 10.26714778 + layer.1.v_cache 0.00000598 0.00447004 + layer.2.k_cache 0.01250870 1.01083458 + layer.2.v_cache 0.00001982 0.01206310 + layer.3.k_cache 0.02633334 3.69061409 + layer.3.v_cache 0.00002177 0.01444680 + layer.4.k_cache 0.00067446 0.26835771 + layer.4.v_cache 0.00005295 0.02449175 + layer.4.output 1.29725540 229.18984564 + ------------------------------------------------------------------------------------- + TOTAL 0.56479644 98.99413437 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 228940 +BPFP 0.8916 bits/point +EBPFP 1.7832 equivalent bits/point +MSE 98.994134 +---------------------- -------------------------------------------------------- +Time: 0.552s Load: 0.009s, Pack+Encode: 0.223s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 98.9941 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,924B, BPFP=0.5706 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,920B, BPFP=1.8664 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,584B, BPFP=0.8341 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,364B, BPFP=1.8263 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,836B, BPFP=0.9243 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,576B, BPFP=1.7696 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,864B, BPFP=0.8543 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,780B, BPFP=1.7843 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,900B, BPFP=1.4329 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,464B, BPFP=1.6895 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,912B, BPFP=0.2871 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150146 64.80203233 + layer.0.v_cache 0.00001564 0.01118677 + layer.1.k_cache 0.36060091 10.48584884 + layer.1.v_cache 0.00000598 0.00439757 + layer.2.k_cache 0.01419523 0.98884421 + layer.2.v_cache 0.00002081 0.01198622 + layer.3.k_cache 0.03680107 4.04694194 + layer.3.v_cache 0.00002092 0.01337811 + layer.4.k_cache 0.00067164 0.26150187 + layer.4.v_cache 0.00004838 0.02344769 + layer.4.output 1.41076765 249.36442561 + ------------------------------------------------------------------------------------- + TOTAL 0.61289739 107.42356146 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 216124 +BPFP 0.9154 bits/point +EBPFP 1.8308 equivalent bits/point +MSE 107.423561 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.009s, Pack+Encode: 0.222s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 107.4236 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 343, 128) +Output shape: (1, 343, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.output: torch.Size([1, 343, 3584]) -> torch.Size([1, 1, 343, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,188B, BPFP=0.5552 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,120B, BPFP=1.8276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,404B, BPFP=0.7928 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,948B, BPFP=1.7742 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,648B, BPFP=0.8950 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,596B, BPFP=1.7126 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,048B, BPFP=0.8222 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,008B, BPFP=1.7314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,528B, BPFP=1.3907 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,292B, BPFP=1.6532 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 50,480B, BPFP=0.3285 +⌛️ [2/4] FRONTEND: Frontend time: 0.280s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.438s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16677478 60.44711188 + layer.0.v_cache 0.00001698 0.01087676 + layer.1.k_cache 0.72144124 10.02104805 + layer.1.v_cache 0.00000659 0.00442840 + layer.2.k_cache 0.01351720 0.92759851 + layer.2.v_cache 0.00002117 0.01177243 + layer.3.k_cache 0.02347133 3.86115765 + layer.3.v_cache 0.00002006 0.01317582 + layer.4.k_cache 0.00069292 0.25380840 + layer.4.v_cache 0.00005144 0.02369433 + layer.4.output 0.03905215 157.91702676 + ------------------------------------------------------------------------------------- + TOTAL 0.07055169 69.47022703 + (elements=2,985,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2985472 +Total Bytes 339260 +BPFP 0.9091 bits/point +EBPFP 1.8182 equivalent bits/point +MSE 69.470227 +---------------------- -------------------------------------------------------- +Time: 0.729s Load: 0.011s, Pack+Encode: 0.280s, Decode+Unpack: 0.438s +---------------------- -------------------------------------------------------- +💾 Converting with 69.4702 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,744B, BPFP=0.5536 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,508B, BPFP=1.8470 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,116B, BPFP=0.8020 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,476B, BPFP=1.7884 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,760B, BPFP=0.8955 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,408B, BPFP=1.7277 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,488B, BPFP=0.8232 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,648B, BPFP=1.7414 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,688B, BPFP=1.4027 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,280B, BPFP=1.6636 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,984B, BPFP=0.2921 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13317820 61.05575284 + layer.0.v_cache 0.00001784 0.01106543 + layer.1.k_cache 0.51782770 10.13923384 + layer.1.v_cache 0.00000612 0.00432493 + layer.2.k_cache 0.01694398 0.93584473 + layer.2.v_cache 0.00002016 0.01160237 + layer.3.k_cache 0.01920085 3.98102894 + layer.3.v_cache 0.00002016 0.01266083 + layer.4.k_cache 0.00071488 0.25234289 + layer.4.v_cache 0.00005349 0.02372803 + layer.4.output 0.00484358 201.61538961 + ------------------------------------------------------------------------------------- + TOTAL 0.04246403 87.51384189 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 269100 +BPFP 0.8994 bits/point +EBPFP 1.7988 equivalent bits/point +MSE 87.513842 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 87.5138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 374, 128) +Output shape: (1, 374, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.output: torch.Size([1, 374, 3584]) -> torch.Size([1, 1, 374, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,744B, BPFP=0.5324 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,876B, BPFP=1.7495 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,724B, BPFP=0.7823 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,528B, BPFP=1.6932 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,840B, BPFP=0.8707 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,052B, BPFP=1.6315 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,260B, BPFP=0.8046 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,584B, BPFP=1.6537 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,844B, BPFP=1.3304 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,560B, BPFP=1.5692 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 47,588B, BPFP=0.2840 +⌛️ [2/4] FRONTEND: Frontend time: 0.281s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.438s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18781370 58.88187772 + layer.0.v_cache 0.00001670 0.01047995 + layer.1.k_cache 0.85217587 10.04158420 + layer.1.v_cache 0.00000605 0.00427446 + layer.2.k_cache 0.02919085 0.89054960 + layer.2.v_cache 0.00002159 0.01086368 + layer.3.k_cache 0.02976600 3.64245573 + layer.3.v_cache 0.00002022 0.01224700 + layer.4.k_cache 0.00072409 0.24777116 + layer.4.v_cache 0.00004949 0.02211424 + layer.4.output 0.03578218 144.85337089 + ------------------------------------------------------------------------------------- + TOTAL 0.07942705 63.98457729 + (elements=3,255,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3255296 +Total Bytes 349600 +BPFP 0.8592 bits/point +EBPFP 1.7183 equivalent bits/point +MSE 63.984577 +---------------------- -------------------------------------------------------- +Time: 0.732s Load: 0.013s, Pack+Encode: 0.281s, Decode+Unpack: 0.438s +---------------------- -------------------------------------------------------- +💾 Converting with 63.9846 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,472B, BPFP=0.5493 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,668B, BPFP=1.7938 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,504B, BPFP=0.8107 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,000B, BPFP=1.7505 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,188B, BPFP=0.9199 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,228B, BPFP=1.7005 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,204B, BPFP=0.8561 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,512B, BPFP=1.7189 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,556B, BPFP=1.3976 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,376B, BPFP=1.6452 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,260B, BPFP=0.2617 +⌛️ [2/4] FRONTEND: Frontend time: 0.275s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.334s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13979621 61.02479901 + layer.0.v_cache 0.00001743 0.01041962 + layer.1.k_cache 0.46959303 10.11537823 + layer.1.v_cache 0.00000602 0.00423684 + layer.2.k_cache 0.01642122 0.90614876 + layer.2.v_cache 0.00002053 0.01154263 + layer.3.k_cache 0.03876302 3.91488052 + layer.3.v_cache 0.00002036 0.01313612 + layer.4.k_cache 0.00069112 0.25254556 + layer.4.v_cache 0.00005399 0.02342110 + layer.4.output 1.27033806 224.36940575 + ------------------------------------------------------------------------------------- + TOTAL 0.56222055 96.87425580 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 230968 +BPFP 0.8809 bits/point +EBPFP 1.7617 equivalent bits/point +MSE 96.874256 +---------------------- -------------------------------------------------------- +Time: 0.619s Load: 0.010s, Pack+Encode: 0.275s, Decode+Unpack: 0.334s +---------------------- -------------------------------------------------------- +💾 Converting with 96.8743 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,728B, BPFP=0.5319 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,892B, BPFP=1.8510 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,900B, BPFP=0.8191 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,092B, BPFP=1.7960 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,584B, BPFP=0.9350 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,392B, BPFP=1.7478 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,668B, BPFP=0.8720 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,756B, BPFP=1.7729 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,852B, BPFP=1.4353 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,548B, BPFP=1.6897 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,060B, BPFP=0.2661 +⌛️ [2/4] FRONTEND: Frontend time: 0.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13893990 61.11742841 + layer.0.v_cache 0.00001711 0.01141258 + layer.1.k_cache 0.36820191 9.64014102 + layer.1.v_cache 0.00000637 0.00460086 + layer.2.k_cache 0.01566778 0.96082672 + layer.2.v_cache 0.00002161 0.01262193 + layer.3.k_cache 0.01884081 4.17449682 + layer.3.v_cache 0.00002076 0.01402874 + layer.4.k_cache 0.00070054 0.26838264 + layer.4.v_cache 0.00005210 0.02504674 + layer.4.output 1.34866800 238.42811910 + ------------------------------------------------------------------------------------- + TOTAL 0.58724382 102.66034236 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 222472 +BPFP 0.9008 bits/point +EBPFP 1.8016 equivalent bits/point +MSE 102.660342 +---------------------- -------------------------------------------------------- +Time: 0.564s Load: 0.008s, Pack+Encode: 0.231s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 102.6603 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,000B, BPFP=0.5267 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,192B, BPFP=1.8839 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,036B, BPFP=0.8214 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,284B, BPFP=1.8308 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,624B, BPFP=0.9143 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,220B, BPFP=1.7685 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,500B, BPFP=0.8485 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,680B, BPFP=1.7954 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,776B, BPFP=1.4499 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,204B, BPFP=1.7090 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,200B, BPFP=0.2776 +⌛️ [2/4] FRONTEND: Frontend time: 0.304s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.389s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11943251 63.14141503 + layer.0.v_cache 0.00001561 0.01084878 + layer.1.k_cache 0.49619702 10.07685583 + layer.1.v_cache 0.00000608 0.00430279 + layer.2.k_cache 0.01349212 0.97856454 + layer.2.v_cache 0.00002006 0.01196801 + layer.3.k_cache 0.01080775 4.02952593 + layer.3.v_cache 0.00002073 0.01280217 + layer.4.k_cache 0.00068886 0.26001862 + layer.4.v_cache 0.00007406 0.02432777 + layer.4.output 0.00494179 207.69458935 + ------------------------------------------------------------------------------------- + TOTAL 0.03972631 90.14192676 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 264716 +BPFP 0.9113 bits/point +EBPFP 1.8225 equivalent bits/point +MSE 90.141927 +---------------------- -------------------------------------------------------- +Time: 0.702s Load: 0.009s, Pack+Encode: 0.304s, Decode+Unpack: 0.389s +---------------------- -------------------------------------------------------- +💾 Converting with 90.1419 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,864B, BPFP=0.5560 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,084B, BPFP=1.8442 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,532B, BPFP=0.8153 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,416B, BPFP=1.7969 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,204B, BPFP=0.9335 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,688B, BPFP=1.7455 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,256B, BPFP=0.8665 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,088B, BPFP=1.7738 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,252B, BPFP=1.4318 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,900B, BPFP=1.6898 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,436B, BPFP=0.2670 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11815372 62.47377422 + layer.0.v_cache 0.00001627 0.01166485 + layer.1.k_cache 0.37623358 10.00721264 + layer.1.v_cache 0.00000616 0.00466934 + layer.2.k_cache 0.01429399 1.00178438 + layer.2.v_cache 0.00002051 0.01265668 + layer.3.k_cache 0.01320818 3.91989619 + layer.3.v_cache 0.00002004 0.01427142 + layer.4.k_cache 0.00067885 0.26899802 + layer.4.v_cache 0.00005318 0.02557234 + layer.4.output 1.38524649 244.84055834 + ------------------------------------------------------------------------------------- + TOTAL 0.60114176 105.38967108 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 216720 +BPFP 0.9013 bits/point +EBPFP 1.8026 equivalent bits/point +MSE 105.389671 +---------------------- -------------------------------------------------------- +Time: 0.546s Load: 0.009s, Pack+Encode: 0.221s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 105.3897 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,940B, BPFP=0.5587 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,448B, BPFP=1.8237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,336B, BPFP=0.8058 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,792B, BPFP=1.7869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,120B, BPFP=0.9060 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,668B, BPFP=1.7237 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,796B, BPFP=0.8316 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,072B, BPFP=1.7464 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,100B, BPFP=1.4107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,836B, BPFP=1.6769 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,244B, BPFP=0.2990 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.418s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09820673 60.09012478 + layer.0.v_cache 0.00001666 0.01038174 + layer.1.k_cache 0.50544519 9.80614181 + layer.1.v_cache 0.00000613 0.00420517 + layer.2.k_cache 0.01930715 0.95105244 + layer.2.v_cache 0.00002033 0.01149810 + layer.3.k_cache 0.01337509 3.95296912 + layer.3.v_cache 0.00001987 0.01290429 + layer.4.k_cache 0.00074230 0.25564981 + layer.4.v_cache 0.00004869 0.02332262 + layer.4.output 0.00473525 199.46809160 + ------------------------------------------------------------------------------------- + TOTAL 0.03943146 86.55264065 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 273352 +BPFP 0.9038 bits/point +EBPFP 1.8075 equivalent bits/point +MSE 86.552641 +---------------------- -------------------------------------------------------- +Time: 0.675s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.418s +---------------------- -------------------------------------------------------- +💾 Converting with 86.5526 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,868B, BPFP=0.5639 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,980B, BPFP=1.8621 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,404B, BPFP=0.8174 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,248B, BPFP=1.8096 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,980B, BPFP=0.9303 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,548B, BPFP=1.7595 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,824B, BPFP=0.8475 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,196B, BPFP=1.8059 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,156B, BPFP=1.4447 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,704B, BPFP=1.6990 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,756B, BPFP=0.2740 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13853801 66.96385823 + layer.0.v_cache 0.00001803 0.01136090 + layer.1.k_cache 0.23740530 9.81736377 + layer.1.v_cache 0.00000581 0.00449355 + layer.2.k_cache 0.01896443 0.99502325 + layer.2.v_cache 0.00001977 0.01254258 + layer.3.k_cache 0.00876656 4.07487194 + layer.3.v_cache 0.00002034 0.01435305 + layer.4.k_cache 0.00073522 0.27076056 + layer.4.v_cache 0.00004914 0.02494042 + layer.4.output 1.40427464 248.26136550 + ------------------------------------------------------------------------------------- + TOTAL 0.60202618 107.05994863 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 215664 +BPFP 0.9093 bits/point +EBPFP 1.8185 equivalent bits/point +MSE 107.059949 +---------------------- -------------------------------------------------------- +Time: 0.552s Load: 0.008s, Pack+Encode: 0.226s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 107.0599 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,716B, BPFP=0.5500 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,328B, BPFP=1.8302 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,208B, BPFP=0.8043 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,364B, BPFP=1.7756 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,980B, BPFP=0.9047 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,556B, BPFP=1.7298 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,744B, BPFP=0.8347 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,088B, BPFP=1.7600 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,012B, BPFP=1.4160 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,612B, BPFP=1.6764 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,068B, BPFP=0.3079 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10827675 61.64380944 + layer.0.v_cache 0.00001706 0.01054952 + layer.1.k_cache 0.56919524 10.11736707 + layer.1.v_cache 0.00000592 0.00410285 + layer.2.k_cache 0.01911232 1.03174536 + layer.2.v_cache 0.00002018 0.01150565 + layer.3.k_cache 0.01342056 4.00270766 + layer.3.v_cache 0.00002333 0.01353753 + layer.4.k_cache 0.00074072 0.25899036 + layer.4.v_cache 0.00005062 0.02337226 + layer.4.output 0.00480139 200.88032156 + ------------------------------------------------------------------------------------- + TOTAL 0.04379250 87.25176109 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 272676 +BPFP 0.9080 bits/point +EBPFP 1.8161 equivalent bits/point +MSE 87.251761 +---------------------- -------------------------------------------------------- +Time: 0.653s Load: 0.009s, Pack+Encode: 0.256s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 87.2518 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,160B, BPFP=0.5360 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,780B, BPFP=1.9183 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,224B, BPFP=0.8324 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,548B, BPFP=1.8462 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,868B, BPFP=0.9286 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,436B, BPFP=1.7811 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,920B, BPFP=0.8731 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,904B, BPFP=1.8085 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,580B, BPFP=1.4384 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,132B, BPFP=1.7048 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,040B, BPFP=0.2929 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15963619 60.11856273 + layer.0.v_cache 0.00001721 0.01151407 + layer.1.k_cache 0.49929055 10.29339668 + layer.1.v_cache 0.00000624 0.00432383 + layer.2.k_cache 0.03270629 1.02213039 + layer.2.v_cache 0.00002062 0.01146094 + layer.3.k_cache 0.00946706 3.64201464 + layer.3.v_cache 0.00002056 0.01308011 + layer.4.k_cache 0.00068536 0.25623767 + layer.4.v_cache 0.00004941 0.02284408 + layer.4.output 0.00495045 207.67740102 + ------------------------------------------------------------------------------------- + TOTAL 0.04332662 89.94925719 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 268592 +BPFP 0.9246 bits/point +EBPFP 1.8492 equivalent bits/point +MSE 89.949257 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.390s +---------------------- -------------------------------------------------------- +💾 Converting with 89.9493 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,320B, BPFP=0.5508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,852B, BPFP=1.8440 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,340B, BPFP=0.8170 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,068B, BPFP=1.7921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,032B, BPFP=0.9290 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,968B, BPFP=1.7193 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,196B, BPFP=0.8737 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,412B, BPFP=1.7487 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,632B, BPFP=1.4322 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,196B, BPFP=1.6682 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,664B, BPFP=0.2522 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14874478 64.46723550 + layer.0.v_cache 0.00001714 0.01082158 + layer.1.k_cache 0.45078779 10.07252632 + layer.1.v_cache 0.00000615 0.00422719 + layer.2.k_cache 0.01724745 1.00393114 + layer.2.v_cache 0.00001987 0.01136224 + layer.3.k_cache 0.02849353 3.95051963 + layer.3.v_cache 0.00002056 0.01317873 + layer.4.k_cache 0.00066912 0.25156273 + layer.4.v_cache 0.00005068 0.02323024 + layer.4.output 1.29722614 229.19398835 + ------------------------------------------------------------------------------------- + TOTAL 0.57215530 99.06861846 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 228680 +BPFP 0.8906 bits/point +EBPFP 1.7812 equivalent bits/point +MSE 99.068618 +---------------------- -------------------------------------------------------- +Time: 0.569s Load: 0.010s, Pack+Encode: 0.229s, Decode+Unpack: 0.330s +---------------------- -------------------------------------------------------- +💾 Converting with 99.0686 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 341, 128) +Output shape: (1, 341, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.output: torch.Size([1, 341, 3584]) -> torch.Size([1, 1, 341, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,072B, BPFP=0.5532 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,904B, BPFP=1.8284 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,316B, BPFP=0.7934 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,764B, BPFP=1.7762 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,548B, BPFP=0.8957 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,336B, BPFP=1.7108 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,164B, BPFP=0.8323 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,772B, BPFP=1.7308 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,352B, BPFP=1.3908 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,996B, BPFP=1.6494 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 47,360B, BPFP=0.3100 +⌛️ [2/4] FRONTEND: Frontend time: 0.276s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17009405 62.18046646 + layer.0.v_cache 0.00001631 0.01091443 + layer.1.k_cache 0.79816426 10.03000639 + layer.1.v_cache 0.00000611 0.00421402 + layer.2.k_cache 0.02878924 0.97768619 + layer.2.v_cache 0.00002016 0.01122330 + layer.3.k_cache 0.01035325 3.91390356 + layer.3.v_cache 0.00002082 0.01273857 + layer.4.k_cache 0.00072160 0.25202935 + layer.4.v_cache 0.00005837 0.02276338 + layer.4.output 0.03919861 158.93662285 + ------------------------------------------------------------------------------------- + TOTAL 0.07544909 69.99837092 + (elements=2,968,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2968064 +Total Bytes 334584 +BPFP 0.9018 bits/point +EBPFP 1.8036 equivalent bits/point +MSE 69.998371 +---------------------- -------------------------------------------------------- +Time: 0.729s Load: 0.013s, Pack+Encode: 0.276s, Decode+Unpack: 0.439s +---------------------- -------------------------------------------------------- +💾 Converting with 69.9984 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 335, 128) +Output shape: (1, 335, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.output: torch.Size([1, 335, 3584]) -> torch.Size([1, 1, 335, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,980B, BPFP=0.5588 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,632B, BPFP=1.8485 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,080B, BPFP=0.7966 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,140B, BPFP=1.7789 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,148B, BPFP=0.8931 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,464B, BPFP=1.7007 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,716B, BPFP=0.8263 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,056B, BPFP=1.7284 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,748B, BPFP=1.3875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,452B, BPFP=1.6535 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,668B, BPFP=0.2976 +⌛️ [2/4] FRONTEND: Frontend time: 0.296s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.446s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15195324 62.21429571 + layer.0.v_cache 0.00001666 0.01094969 + layer.1.k_cache 0.76788093 10.21473225 + layer.1.v_cache 0.00000624 0.00430799 + layer.2.k_cache 0.02686723 0.94609011 + layer.2.v_cache 0.00001960 0.01098344 + layer.3.k_cache 0.01080669 4.02631581 + layer.3.v_cache 0.00002024 0.01267151 + layer.4.k_cache 0.00071790 0.25790931 + layer.4.v_cache 0.00005106 0.02304074 + layer.4.output 3.38143948 160.15394456 + ------------------------------------------------------------------------------------- + TOTAL 1.44873036 70.51758285 + (elements=2,915,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2915840 +Total Bytes 327084 +BPFP 0.8974 bits/point +EBPFP 1.7948 equivalent bits/point +MSE 70.517583 +---------------------- -------------------------------------------------------- +Time: 0.753s Load: 0.012s, Pack+Encode: 0.296s, Decode+Unpack: 0.446s +---------------------- -------------------------------------------------------- +💾 Converting with 70.5176 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,780B, BPFP=0.5177 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,140B, BPFP=1.8950 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,756B, BPFP=0.8111 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,956B, BPFP=1.8252 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,612B, BPFP=0.9205 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,664B, BPFP=1.7491 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,512B, BPFP=0.8557 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,108B, BPFP=1.7752 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,092B, BPFP=1.4205 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,676B, BPFP=1.6908 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,356B, BPFP=0.2978 +⌛️ [2/4] FRONTEND: Frontend time: 0.384s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.523s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17469128 60.62934110 + layer.0.v_cache 0.00001775 0.01124627 + layer.1.k_cache 0.51265420 10.35983472 + layer.1.v_cache 0.00000620 0.00428696 + layer.2.k_cache 0.03428403 0.96619643 + layer.2.v_cache 0.00002024 0.01134621 + layer.3.k_cache 0.01443097 3.55348476 + layer.3.v_cache 0.00002182 0.01349372 + layer.4.k_cache 0.00070119 0.25250981 + layer.4.v_cache 0.00005097 0.02228352 + layer.4.output 0.00503371 209.36681267 + ------------------------------------------------------------------------------------- + TOTAL 0.04541851 90.67010072 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 263652 +BPFP 0.9144 bits/point +EBPFP 1.8289 equivalent bits/point +MSE 90.670101 +---------------------- -------------------------------------------------------- +Time: 0.917s Load: 0.010s, Pack+Encode: 0.384s, Decode+Unpack: 0.523s +---------------------- -------------------------------------------------------- +💾 Converting with 90.6701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,692B, BPFP=0.5125 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,948B, BPFP=1.8837 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,780B, BPFP=0.8125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,772B, BPFP=1.8144 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,372B, BPFP=0.9064 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,468B, BPFP=1.7375 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,240B, BPFP=0.8396 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,928B, BPFP=1.7646 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,440B, BPFP=1.4410 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,852B, BPFP=1.7012 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,824B, BPFP=0.2681 +⌛️ [2/4] FRONTEND: Frontend time: 0.335s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.471s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18043976 61.96959758 + layer.0.v_cache 0.00001814 0.01094386 + layer.1.k_cache 0.58196952 10.12731427 + layer.1.v_cache 0.00000608 0.00440631 + layer.2.k_cache 0.02409397 1.00953323 + layer.2.v_cache 0.00002020 0.01140755 + layer.3.k_cache 0.01875559 3.50328208 + layer.3.v_cache 0.00002085 0.01321813 + layer.4.k_cache 0.00069073 0.26221639 + layer.4.v_cache 0.00005389 0.02464971 + layer.4.output 0.00498411 209.33652291 + ------------------------------------------------------------------------------------- + TOTAL 0.04946809 90.72307232 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 259316 +BPFP 0.8994 bits/point +EBPFP 1.7988 equivalent bits/point +MSE 90.723072 +---------------------- -------------------------------------------------------- +Time: 0.815s Load: 0.009s, Pack+Encode: 0.335s, Decode+Unpack: 0.471s +---------------------- -------------------------------------------------------- +💾 Converting with 90.7231 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 386, 128) +Output shape: (1, 386, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.output: torch.Size([1, 386, 3584]) -> torch.Size([1, 1, 386, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,784B, BPFP=0.5175 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,540B, BPFP=1.8434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,520B, BPFP=0.7902 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,940B, BPFP=1.7787 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,824B, BPFP=0.9239 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,916B, BPFP=1.7372 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,080B, BPFP=0.8533 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,496B, BPFP=1.7607 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,144B, BPFP=1.4226 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,472B, BPFP=1.6788 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 50,552B, BPFP=0.2923 +⌛️ [2/4] FRONTEND: Frontend time: 0.412s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.638s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16808204 60.80335674 + layer.0.v_cache 0.00001827 0.01145014 + layer.1.k_cache 0.89383947 10.08815058 + layer.1.v_cache 0.00000605 0.00435215 + layer.2.k_cache 0.02643975 0.97858836 + layer.2.v_cache 0.00002094 0.01211943 + layer.3.k_cache 0.04368651 3.93569709 + layer.3.v_cache 0.00002149 0.01383746 + layer.4.k_cache 0.00075087 0.26445231 + layer.4.v_cache 0.00005208 0.02427281 + layer.4.output 0.03469837 140.43608901 + ------------------------------------------------------------------------------------- + TOTAL 0.08092977 62.30522942 + (elements=3,359,744) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3359744 +Total Bytes 379268 +BPFP 0.9031 bits/point +EBPFP 1.8062 equivalent bits/point +MSE 62.305229 +---------------------- -------------------------------------------------------- +Time: 1.062s Load: 0.013s, Pack+Encode: 0.412s, Decode+Unpack: 0.638s +---------------------- -------------------------------------------------------- +💾 Converting with 62.3052 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,000B, BPFP=0.5425 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,672B, BPFP=1.8268 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,612B, BPFP=0.7928 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,848B, BPFP=1.7821 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,804B, BPFP=0.9117 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,764B, BPFP=1.7233 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,440B, BPFP=0.8377 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,112B, BPFP=1.7422 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,920B, BPFP=1.4062 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,592B, BPFP=1.6597 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,756B, BPFP=0.2849 +⌛️ [2/4] FRONTEND: Frontend time: 0.361s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.458s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14364293 62.28771294 + layer.0.v_cache 0.00001600 0.01151214 + layer.1.k_cache 0.62366660 9.97448561 + layer.1.v_cache 0.00000628 0.00459055 + layer.2.k_cache 0.01981349 0.98573346 + layer.2.v_cache 0.00002090 0.01267383 + layer.3.k_cache 0.02158341 4.11311807 + layer.3.v_cache 0.00002097 0.01388010 + layer.4.k_cache 0.00071696 0.26602348 + layer.4.v_cache 0.00005883 0.02480909 + layer.4.output 0.00463133 192.34864831 + ------------------------------------------------------------------------------------- + TOTAL 0.04952739 83.77265162 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 280520 +BPFP 0.8952 bits/point +EBPFP 1.7905 equivalent bits/point +MSE 83.772652 +---------------------- -------------------------------------------------------- +Time: 0.828s Load: 0.009s, Pack+Encode: 0.361s, Decode+Unpack: 0.458s +---------------------- -------------------------------------------------------- +💾 Converting with 83.7727 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,996B, BPFP=0.5639 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,512B, BPFP=1.8339 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,128B, BPFP=0.7969 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,456B, BPFP=1.7744 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,008B, BPFP=0.9030 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,536B, BPFP=1.7225 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,936B, BPFP=0.8425 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,020B, BPFP=1.7498 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,764B, BPFP=1.3969 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,532B, BPFP=1.6658 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,216B, BPFP=0.3321 +⌛️ [2/4] FRONTEND: Frontend time: 0.347s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.498s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12300906 61.47141528 + layer.0.v_cache 0.00001720 0.00986912 + layer.1.k_cache 0.58060381 10.16819791 + layer.1.v_cache 0.00000644 0.00379206 + layer.2.k_cache 0.02305070 0.90066424 + layer.2.v_cache 0.00002053 0.01093094 + layer.3.k_cache 0.05279791 3.50324963 + layer.3.v_cache 0.00002021 0.01240421 + layer.4.k_cache 0.00074809 0.25535947 + layer.4.v_cache 0.00005348 0.02267367 + layer.4.output 0.00483522 200.22426186 + ------------------------------------------------------------------------------------- + TOTAL 0.04789259 86.93696409 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 276104 +BPFP 0.9161 bits/point +EBPFP 1.8323 equivalent bits/point +MSE 86.936964 +---------------------- -------------------------------------------------------- +Time: 0.855s Load: 0.009s, Pack+Encode: 0.347s, Decode+Unpack: 0.498s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9370 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,708B, BPFP=0.5401 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,688B, BPFP=1.8700 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,468B, BPFP=0.8035 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,748B, BPFP=1.8041 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,512B, BPFP=0.9467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,928B, BPFP=1.7466 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,524B, BPFP=0.8775 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,236B, BPFP=1.7682 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,600B, BPFP=1.4434 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,136B, BPFP=1.6911 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,724B, BPFP=0.2575 +⌛️ [2/4] FRONTEND: Frontend time: 0.339s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.404s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16177035 63.69201058 + layer.0.v_cache 0.00001782 0.01177664 + layer.1.k_cache 0.31952496 9.75139697 + layer.1.v_cache 0.00000626 0.00442723 + layer.2.k_cache 0.01271096 0.92981162 + layer.2.v_cache 0.00002040 0.01242824 + layer.3.k_cache 0.01363577 4.05702079 + layer.3.v_cache 0.00002033 0.01391826 + layer.4.k_cache 0.00066992 0.26898816 + layer.4.v_cache 0.00005030 0.02339395 + layer.4.output 1.37284199 242.39790199 + ------------------------------------------------------------------------------------- + TOTAL 0.59519535 104.44414626 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 218272 +BPFP 0.8996 bits/point +EBPFP 1.7993 equivalent bits/point +MSE 104.444146 +---------------------- -------------------------------------------------------- +Time: 0.750s Load: 0.008s, Pack+Encode: 0.339s, Decode+Unpack: 0.404s +---------------------- -------------------------------------------------------- +💾 Converting with 104.4441 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,440B, BPFP=0.5499 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,412B, BPFP=1.8282 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,516B, BPFP=0.8125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,768B, BPFP=1.7732 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,312B, BPFP=0.9658 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,336B, BPFP=1.7363 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,376B, BPFP=0.8859 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,712B, BPFP=1.7684 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,008B, BPFP=1.4522 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,712B, BPFP=1.6831 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,492B, BPFP=0.2743 +⌛️ [2/4] FRONTEND: Frontend time: 0.270s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.274s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09962382 63.31359930 + layer.0.v_cache 0.00001686 0.01177453 + layer.1.k_cache 0.10988800 10.01975738 + layer.1.v_cache 0.00000590 0.00454021 + layer.2.k_cache 0.01214521 0.97606171 + layer.2.v_cache 0.00002216 0.01293468 + layer.3.k_cache 0.02260447 4.07974760 + layer.3.v_cache 0.00002006 0.01502343 + layer.4.k_cache 0.00069723 0.26479244 + layer.4.v_cache 0.00005242 0.02593902 + layer.4.output 0.00882482 299.68645101 + ------------------------------------------------------------------------------------- + TOTAL 0.01804999 128.03113691 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 180084 +BPFP 0.9045 bits/point +EBPFP 1.8089 equivalent bits/point +MSE 128.031137 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.007s, Pack+Encode: 0.270s, Decode+Unpack: 0.274s +---------------------- -------------------------------------------------------- +💾 Converting with 128.0311 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,980B, BPFP=0.5549 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,068B, BPFP=1.8387 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,452B, BPFP=0.8036 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,876B, BPFP=1.7725 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,168B, BPFP=0.8990 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,612B, BPFP=1.7022 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,096B, BPFP=0.8394 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,368B, BPFP=1.7442 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,108B, BPFP=1.3961 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,764B, BPFP=1.6550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,684B, BPFP=0.3073 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.391s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11148910 61.98798237 + layer.0.v_cache 0.00001684 0.01089177 + layer.1.k_cache 0.60671574 10.08559344 + layer.1.v_cache 0.00000630 0.00421997 + layer.2.k_cache 0.04149003 0.98802842 + layer.2.v_cache 0.00001976 0.01088882 + layer.3.k_cache 0.02200458 3.86519267 + layer.3.v_cache 0.00001954 0.01258002 + layer.4.k_cache 0.00078895 0.25568550 + layer.4.v_cache 0.00004864 0.02237870 + layer.4.output 0.00472744 197.42369408 + ------------------------------------------------------------------------------------- + TOTAL 0.04798185 85.83584119 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 276176 +BPFP 0.9033 bits/point +EBPFP 1.8067 equivalent bits/point +MSE 85.835841 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.009s, Pack+Encode: 0.256s, Decode+Unpack: 0.391s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8358 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,520B, BPFP=0.5094 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,904B, BPFP=1.9456 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,968B, BPFP=0.8569 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,256B, BPFP=1.8950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,632B, BPFP=0.9869 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,492B, BPFP=1.8353 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,412B, BPFP=0.8916 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,684B, BPFP=1.8503 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,320B, BPFP=1.5094 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,516B, BPFP=1.7591 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,568B, BPFP=0.2407 +⌛️ [2/4] FRONTEND: Frontend time: 0.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14565937 64.91497559 + layer.0.v_cache 0.00001550 0.01107993 + layer.1.k_cache 0.22559690 10.07079285 + layer.1.v_cache 0.00000590 0.00439426 + layer.2.k_cache 0.00679684 0.99166840 + layer.2.v_cache 0.00002314 0.01281221 + layer.3.k_cache 0.06866968 3.92960938 + layer.3.v_cache 0.00002119 0.01422853 + layer.4.k_cache 0.00065507 0.26677101 + layer.4.v_cache 0.00005191 0.02530718 + layer.4.output 1.53062134 270.69747768 + ------------------------------------------------------------------------------------- + TOTAL 0.65657911 116.18376371 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 201272 +BPFP 0.9250 bits/point +EBPFP 1.8499 equivalent bits/point +MSE 116.183764 +---------------------- -------------------------------------------------------- +Time: 0.565s Load: 0.007s, Pack+Encode: 0.231s, Decode+Unpack: 0.327s +---------------------- -------------------------------------------------------- +💾 Converting with 116.1838 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,800B, BPFP=0.5286 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,956B, BPFP=1.9400 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,880B, BPFP=0.8458 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,160B, BPFP=1.8781 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,400B, BPFP=0.9639 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,304B, BPFP=1.8116 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,344B, BPFP=0.8818 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,728B, BPFP=1.8445 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,228B, BPFP=1.4947 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,648B, BPFP=1.7606 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,924B, BPFP=0.2435 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10663509 66.20947994 + layer.0.v_cache 0.00001613 0.01171081 + layer.1.k_cache 0.26620468 10.04098344 + layer.1.v_cache 0.00000602 0.00482960 + layer.2.k_cache 0.00649704 0.98773596 + layer.2.v_cache 0.00002002 0.01328317 + layer.3.k_cache 0.02565914 4.31748849 + layer.3.v_cache 0.00002186 0.01491853 + layer.4.k_cache 0.00070101 0.28187918 + layer.4.v_cache 0.00005421 0.02702911 + layer.4.output 1.52303164 269.29928483 + ------------------------------------------------------------------------------------- + TOTAL 0.65100216 115.70613718 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 201372 +BPFP 0.9208 bits/point +EBPFP 1.8416 equivalent bits/point +MSE 115.706137 +---------------------- -------------------------------------------------------- +Time: 0.562s Load: 0.007s, Pack+Encode: 0.228s, Decode+Unpack: 0.327s +---------------------- -------------------------------------------------------- +💾 Converting with 115.7061 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,104B, BPFP=0.5328 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,708B, BPFP=1.9141 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,152B, BPFP=0.8282 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,556B, BPFP=1.8467 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,208B, BPFP=0.9485 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,496B, BPFP=1.7846 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,980B, BPFP=0.8766 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,152B, BPFP=1.8230 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,164B, BPFP=1.4726 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,528B, BPFP=1.7280 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,112B, BPFP=0.2852 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258070 63.10909703 + layer.0.v_cache 0.00001713 0.01185784 + layer.1.k_cache 0.49703808 9.80798294 + layer.1.v_cache 0.00000670 0.00480208 + layer.2.k_cache 0.01458295 0.98689116 + layer.2.v_cache 0.00002181 0.01263198 + layer.3.k_cache 0.02940617 3.74012876 + layer.3.v_cache 0.00002120 0.01451647 + layer.4.k_cache 0.00070157 0.27657504 + layer.4.v_cache 0.00005095 0.02446195 + layer.4.output 0.00497170 207.72035514 + ------------------------------------------------------------------------------------- + TOTAL 0.04171936 90.11949595 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 269160 +BPFP 0.9266 bits/point +EBPFP 1.8531 equivalent bits/point +MSE 90.119496 +---------------------- -------------------------------------------------------- +Time: 0.657s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.390s +---------------------- -------------------------------------------------------- +💾 Converting with 90.1195 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,876B, BPFP=0.5445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,704B, BPFP=1.8462 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,976B, BPFP=0.8280 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,968B, BPFP=1.7954 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,600B, BPFP=0.9403 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,320B, BPFP=1.7506 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,584B, BPFP=0.8700 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,768B, BPFP=1.7815 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,992B, BPFP=1.4513 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,472B, BPFP=1.6919 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,392B, BPFP=0.2607 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15433978 65.08236397 + layer.0.v_cache 0.00001575 0.01142114 + layer.1.k_cache 0.29362690 9.59727492 + layer.1.v_cache 0.00000607 0.00466173 + layer.2.k_cache 0.00940348 1.00457291 + layer.2.v_cache 0.00002159 0.01282534 + layer.3.k_cache 0.02320663 3.93283432 + layer.3.v_cache 0.00002067 0.01503594 + layer.4.k_cache 0.00067785 0.27593260 + layer.4.v_cache 0.00005016 0.02535185 + layer.4.output 1.35461534 239.52972898 + ------------------------------------------------------------------------------------- + TOTAL 0.58609860 103.33355162 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 221652 +BPFP 0.9014 bits/point +EBPFP 1.8029 equivalent bits/point +MSE 103.333552 +---------------------- -------------------------------------------------------- +Time: 0.562s Load: 0.008s, Pack+Encode: 0.229s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 103.3336 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,840B, BPFP=0.5395 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,372B, BPFP=1.8296 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,608B, BPFP=0.8009 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,316B, BPFP=1.7717 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,788B, BPFP=0.9204 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,340B, BPFP=1.7182 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,540B, BPFP=0.8520 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,844B, BPFP=1.7458 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,420B, BPFP=1.3936 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,124B, BPFP=1.6515 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,976B, BPFP=0.2974 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14751465 60.87088816 + layer.0.v_cache 0.00001719 0.01100913 + layer.1.k_cache 0.52245269 10.02510965 + layer.1.v_cache 0.00000618 0.00415788 + layer.2.k_cache 0.02707210 0.92325354 + layer.2.v_cache 0.00002227 0.01161496 + layer.3.k_cache 0.03364062 4.09165810 + layer.3.v_cache 0.00002083 0.01324065 + layer.4.k_cache 0.00069748 0.25218808 + layer.4.v_cache 0.00005305 0.02257974 + layer.4.output 0.00471531 194.58115602 + ------------------------------------------------------------------------------------- + TOTAL 0.04497084 84.60551718 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 279168 +BPFP 0.9003 bits/point +EBPFP 1.8006 equivalent bits/point +MSE 84.605517 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.009s, Pack+Encode: 0.257s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 84.6055 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,356B, BPFP=0.5120 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,928B, BPFP=1.7113 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,552B, BPFP=0.7691 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,152B, BPFP=1.6637 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,008B, BPFP=0.8583 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,132B, BPFP=1.6012 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,228B, BPFP=0.8105 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,308B, BPFP=1.6120 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,572B, BPFP=1.3218 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,336B, BPFP=1.5525 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,888B, BPFP=0.3404 +⌛️ [2/4] FRONTEND: Frontend time: 0.236s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17432048 59.47900582 + layer.0.v_cache 0.00001704 0.01016958 + layer.1.k_cache 0.39057378 9.79523686 + layer.1.v_cache 0.00000668 0.00421943 + layer.2.k_cache 0.01299327 0.88448636 + layer.2.v_cache 0.00002114 0.01049210 + layer.3.k_cache 0.02499930 3.39022504 + layer.3.v_cache 0.00002096 0.01243479 + layer.4.k_cache 0.00069121 0.24483275 + layer.4.v_cache 0.00005416 0.02243317 + layer.4.output 1.20070616 208.64266457 + ------------------------------------------------------------------------------------- + TOTAL 0.52992007 90.25601105 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 241460 +BPFP 0.8703 bits/point +EBPFP 1.7406 equivalent bits/point +MSE 90.256011 +---------------------- -------------------------------------------------------- +Time: 0.572s Load: 0.008s, Pack+Encode: 0.236s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 90.2560 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,624B, BPFP=0.5202 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,176B, BPFP=1.8543 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,032B, BPFP=0.8210 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,300B, BPFP=1.7945 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,412B, BPFP=0.9151 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,468B, BPFP=1.7377 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,780B, BPFP=0.8720 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,848B, BPFP=1.7636 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,056B, BPFP=1.4367 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,680B, BPFP=1.6840 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,840B, BPFP=0.2714 +⌛️ [2/4] FRONTEND: Frontend time: 0.233s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.335s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16239141 63.42510320 + layer.0.v_cache 0.00001645 0.01111872 + layer.1.k_cache 0.33751179 10.25667389 + layer.1.v_cache 0.00000668 0.00453685 + layer.2.k_cache 0.00660169 1.02567928 + layer.2.v_cache 0.00002045 0.01204705 + layer.3.k_cache 0.02213682 3.82766524 + layer.3.v_cache 0.00002049 0.01399096 + layer.4.k_cache 0.00067824 0.26770096 + layer.4.v_cache 0.00005183 0.02421183 + layer.4.output 1.33690934 236.40921319 + ------------------------------------------------------------------------------------- + TOTAL 0.58163537 101.98430708 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 224216 +BPFP 0.8999 bits/point +EBPFP 1.7998 equivalent bits/point +MSE 101.984307 +---------------------- -------------------------------------------------------- +Time: 0.576s Load: 0.008s, Pack+Encode: 0.233s, Decode+Unpack: 0.335s +---------------------- -------------------------------------------------------- +💾 Converting with 101.9843 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,228B, BPFP=0.5380 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,068B, BPFP=1.8696 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,448B, BPFP=0.8424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,276B, BPFP=1.8235 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,396B, BPFP=0.8976 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,756B, BPFP=1.7348 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,356B, BPFP=0.8370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,988B, BPFP=1.7484 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,444B, BPFP=1.4251 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,768B, BPFP=1.6772 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,660B, BPFP=0.2804 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.394s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15796844 63.24943155 + layer.0.v_cache 0.00001527 0.01048532 + layer.1.k_cache 0.50235566 10.06564468 + layer.1.v_cache 0.00000626 0.00443000 + layer.2.k_cache 0.02709456 1.01171852 + layer.2.v_cache 0.00002063 0.01095855 + layer.3.k_cache 0.00856834 3.92308842 + layer.3.v_cache 0.00001945 0.01268471 + layer.4.k_cache 0.00071483 0.26010052 + layer.4.v_cache 0.00005464 0.02353771 + layer.4.output 0.00491706 206.90083622 + ------------------------------------------------------------------------------------- + TOTAL 0.04301397 89.81634903 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 263388 +BPFP 0.9033 bits/point +EBPFP 1.8066 equivalent bits/point +MSE 89.816349 +---------------------- -------------------------------------------------------- +Time: 0.661s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.394s +---------------------- -------------------------------------------------------- +💾 Converting with 89.8163 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,676B, BPFP=0.5466 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,472B, BPFP=1.7939 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,940B, BPFP=0.8153 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,728B, BPFP=1.7470 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,612B, BPFP=0.9206 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,680B, BPFP=1.6809 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,608B, BPFP=0.8574 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,124B, BPFP=1.7089 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,008B, BPFP=1.3866 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,852B, BPFP=1.6288 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,084B, BPFP=0.3068 +⌛️ [2/4] FRONTEND: Frontend time: 0.232s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.334s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13184587 57.23269752 + layer.0.v_cache 0.00001706 0.01110031 + layer.1.k_cache 0.45546043 10.06142598 + layer.1.v_cache 0.00000678 0.00446233 + layer.2.k_cache 0.02272066 0.98393329 + layer.2.v_cache 0.00001996 0.01157023 + layer.3.k_cache 0.01643563 3.77411996 + layer.3.v_cache 0.00002069 0.01358474 + layer.4.k_cache 0.00071781 0.25620485 + layer.4.v_cache 0.00005262 0.02292582 + layer.4.output 1.23455937 218.10552275 + ------------------------------------------------------------------------------------- + TOTAL 0.54524783 94.06533437 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 241784 +BPFP 0.8961 bits/point +EBPFP 1.7922 equivalent bits/point +MSE 94.065334 +---------------------- -------------------------------------------------------- +Time: 0.575s Load: 0.008s, Pack+Encode: 0.232s, Decode+Unpack: 0.334s +---------------------- -------------------------------------------------------- +💾 Converting with 94.0653 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,564B, BPFP=0.5069 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,704B, BPFP=1.8764 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,808B, BPFP=0.8172 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,584B, BPFP=1.8101 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,280B, BPFP=0.9044 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,760B, BPFP=1.7614 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,456B, BPFP=0.8556 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,052B, BPFP=1.7786 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,392B, BPFP=1.4437 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,648B, BPFP=1.6955 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,540B, BPFP=0.2667 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17705467 62.83242335 + layer.0.v_cache 0.00001633 0.01040632 + layer.1.k_cache 0.50406676 10.42962646 + layer.1.v_cache 0.00000599 0.00413397 + layer.2.k_cache 0.00944760 0.97730394 + layer.2.v_cache 0.00001883 0.01080077 + layer.3.k_cache 0.04120483 3.83922508 + layer.3.v_cache 0.00002392 0.01317825 + layer.4.k_cache 0.00070100 0.25523963 + layer.4.v_cache 0.00005002 0.02317208 + layer.4.output 0.00499618 210.17069129 + ------------------------------------------------------------------------------------- + TOTAL 0.04515078 91.15237346 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 258788 +BPFP 0.9010 bits/point +EBPFP 1.8019 equivalent bits/point +MSE 91.152373 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.010s, Pack+Encode: 0.258s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 91.1524 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,480B, BPFP=0.5486 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,160B, BPFP=1.8611 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,876B, BPFP=0.8030 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,168B, BPFP=1.8037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,468B, BPFP=0.8951 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,904B, BPFP=1.7306 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,516B, BPFP=0.8400 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,304B, BPFP=1.7537 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,648B, BPFP=1.4264 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,888B, BPFP=1.6718 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,872B, BPFP=0.2800 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20501094 65.34363426 + layer.0.v_cache 0.00001516 0.01032394 + layer.1.k_cache 0.57085758 10.27799841 + layer.1.v_cache 0.00000574 0.00423380 + layer.2.k_cache 0.01259345 1.00255681 + layer.2.v_cache 0.00002099 0.01132058 + layer.3.k_cache 0.01648921 4.29366591 + layer.3.v_cache 0.00001970 0.01306430 + layer.4.k_cache 0.00070444 0.25231402 + layer.4.v_cache 0.00005121 0.02370441 + layer.4.output 0.00487421 205.39707341 + ------------------------------------------------------------------------------------- + TOTAL 0.04940517 89.35366649 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 264284 +BPFP 0.8997 bits/point +EBPFP 1.7993 equivalent bits/point +MSE 89.353666 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.009s, Pack+Encode: 0.259s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 89.3537 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,636B, BPFP=0.5556 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,636B, BPFP=1.8817 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,120B, BPFP=0.8141 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,252B, BPFP=1.8019 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,916B, BPFP=0.9177 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,336B, BPFP=1.7491 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,876B, BPFP=0.8577 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,708B, BPFP=1.7705 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,896B, BPFP=1.4354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,276B, BPFP=1.6880 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,748B, BPFP=0.2944 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.392s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14466163 62.34230858 + layer.0.v_cache 0.00001631 0.01113253 + layer.1.k_cache 0.53789613 10.17113627 + layer.1.v_cache 0.00000653 0.00453887 + layer.2.k_cache 0.02095415 0.92025982 + layer.2.v_cache 0.00002169 0.01171365 + layer.3.k_cache 0.03083320 4.02841536 + layer.3.v_cache 0.00002083 0.01334507 + layer.4.k_cache 0.00069737 0.26954003 + layer.4.v_cache 0.00004943 0.02377775 + layer.4.output 0.00490546 204.51278334 + ------------------------------------------------------------------------------------- + TOTAL 0.04526444 88.78739125 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 269400 +BPFP 0.9137 bits/point +EBPFP 1.8274 equivalent bits/point +MSE 88.787391 +---------------------- -------------------------------------------------------- +Time: 0.659s Load: 0.011s, Pack+Encode: 0.256s, Decode+Unpack: 0.392s +---------------------- -------------------------------------------------------- +💾 Converting with 88.7874 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,528B, BPFP=0.5502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,096B, BPFP=1.7818 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,440B, BPFP=0.8069 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,080B, BPFP=1.7287 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,096B, BPFP=0.8934 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,928B, BPFP=1.6685 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,764B, BPFP=0.8238 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,360B, BPFP=1.6911 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,228B, BPFP=1.3706 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,936B, BPFP=1.6166 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,088B, BPFP=0.2918 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14240686 59.56847042 + layer.0.v_cache 0.00001673 0.01005744 + layer.1.k_cache 0.61438330 10.19242364 + layer.1.v_cache 0.00000638 0.00405721 + layer.2.k_cache 0.02027981 0.93277314 + layer.2.v_cache 0.00002012 0.01076531 + layer.3.k_cache 0.02199794 3.86634321 + layer.3.v_cache 0.00002003 0.01227206 + layer.4.k_cache 0.00078937 0.24297238 + layer.4.v_cache 0.00005038 0.02179597 + layer.4.output 0.04464151 181.12801601 + ------------------------------------------------------------------------------------- + TOTAL 0.06543891 78.98576723 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 286544 +BPFP 0.8808 bits/point +EBPFP 1.7617 equivalent bits/point +MSE 78.985767 +---------------------- -------------------------------------------------------- +Time: 0.656s Load: 0.011s, Pack+Encode: 0.258s, Decode+Unpack: 0.388s +---------------------- -------------------------------------------------------- +💾 Converting with 78.9858 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 399, 128) +Output shape: (1, 399, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.output: torch.Size([1, 399, 3584]) -> torch.Size([1, 1, 399, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,960B, BPFP=0.5467 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,148B, BPFP=1.8072 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,228B, BPFP=0.7921 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,404B, BPFP=1.7389 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,936B, BPFP=0.8982 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,508B, BPFP=1.7038 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,212B, BPFP=0.8307 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,132B, BPFP=1.7282 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,364B, BPFP=1.3849 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,976B, BPFP=1.6438 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 55,288B, BPFP=0.3093 +⌛️ [2/4] FRONTEND: Frontend time: 0.330s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.520s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17889012 62.09851778 + layer.0.v_cache 0.00001635 0.01025013 + layer.1.k_cache 0.94984295 9.90010365 + layer.1.v_cache 0.00000590 0.00395651 + layer.2.k_cache 0.02305104 0.95186223 + layer.2.v_cache 0.00002114 0.01159473 + layer.3.k_cache 0.01620252 3.75723060 + layer.3.v_cache 0.00002049 0.01290063 + layer.4.k_cache 0.00073175 0.25105691 + layer.4.v_cache 0.00005255 0.02281277 + layer.4.output 0.00656733 138.73552184 + ------------------------------------------------------------------------------------- + TOTAL 0.07145918 61.65699640 + (elements=3,472,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3472896 +Total Bytes 389156 +BPFP 0.8964 bits/point +EBPFP 1.7929 equivalent bits/point +MSE 61.656996 +---------------------- -------------------------------------------------------- +Time: 0.865s Load: 0.015s, Pack+Encode: 0.330s, Decode+Unpack: 0.520s +---------------------- -------------------------------------------------------- +💾 Converting with 61.6570 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,996B, BPFP=0.5528 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,692B, BPFP=1.8454 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,796B, BPFP=0.8155 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,872B, BPFP=1.7887 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,612B, BPFP=0.9411 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,356B, BPFP=1.7530 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,456B, BPFP=0.8612 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,844B, BPFP=1.7868 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,624B, BPFP=1.4259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,548B, BPFP=1.6972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,248B, BPFP=0.2592 +⌛️ [2/4] FRONTEND: Frontend time: 0.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.331s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10670873 64.06421114 + layer.0.v_cache 0.00001672 0.01188236 + layer.1.k_cache 0.29393762 9.64121591 + layer.1.v_cache 0.00000627 0.00459447 + layer.2.k_cache 0.01243114 0.98311156 + layer.2.v_cache 0.00002105 0.01281599 + layer.3.k_cache 0.00919804 3.87313168 + layer.3.v_cache 0.00002002 0.01442393 + layer.4.k_cache 0.00070102 0.26744512 + layer.4.v_cache 0.00005024 0.02611147 + layer.4.output 1.35463108 239.47998973 + ------------------------------------------------------------------------------------- + TOTAL 0.58267697 103.25052187 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 221044 +BPFP 0.8990 bits/point +EBPFP 1.7979 equivalent bits/point +MSE 103.250522 +---------------------- -------------------------------------------------------- +Time: 0.570s Load: 0.008s, Pack+Encode: 0.231s, Decode+Unpack: 0.331s +---------------------- -------------------------------------------------------- +💾 Converting with 103.2505 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,020B, BPFP=0.5549 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,136B, BPFP=1.7308 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,932B, BPFP=0.7955 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,228B, BPFP=1.6750 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,900B, BPFP=0.9166 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,600B, BPFP=1.6363 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,748B, BPFP=0.8457 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,036B, BPFP=1.6631 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,960B, BPFP=1.3509 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,780B, BPFP=1.5859 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,064B, BPFP=0.3257 +⌛️ [2/4] FRONTEND: Frontend time: 0.235s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.323s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633087 59.79504875 + layer.0.v_cache 0.00001891 0.01123690 + layer.1.k_cache 0.36012592 9.40283972 + layer.1.v_cache 0.00000581 0.00403835 + layer.2.k_cache 0.01407751 0.88596849 + layer.2.v_cache 0.00002023 0.01128962 + layer.3.k_cache 0.01230825 3.65396863 + layer.3.v_cache 0.00002000 0.01305041 + layer.4.k_cache 0.00071033 0.24851602 + layer.4.v_cache 0.00004856 0.02261025 + layer.4.output 1.20534739 212.43586544 + ------------------------------------------------------------------------------------- + TOTAL 0.52535871 91.82938972 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 244404 +BPFP 0.8844 bits/point +EBPFP 1.7688 equivalent bits/point +MSE 91.829390 +---------------------- -------------------------------------------------------- +Time: 0.566s Load: 0.008s, Pack+Encode: 0.235s, Decode+Unpack: 0.323s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8294 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,416B, BPFP=0.5456 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,676B, BPFP=1.7943 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,472B, BPFP=0.8086 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,992B, BPFP=1.7500 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,072B, BPFP=0.9123 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,204B, BPFP=1.6989 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,160B, BPFP=0.8532 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,508B, BPFP=1.7186 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,688B, BPFP=1.4061 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,456B, BPFP=1.6504 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,540B, BPFP=0.2736 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.322s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11142051 60.66302516 + layer.0.v_cache 0.00001650 0.01073435 + layer.1.k_cache 0.42032433 9.73129842 + layer.1.v_cache 0.00000619 0.00427189 + layer.2.k_cache 0.01443043 0.96960380 + layer.2.v_cache 0.00002030 0.01189123 + layer.3.k_cache 0.02895402 3.74359815 + layer.3.v_cache 0.00001914 0.01258502 + layer.4.k_cache 0.00067703 0.25025443 + layer.4.v_cache 0.00005693 0.02411024 + layer.4.output 1.27033357 224.35543865 + ------------------------------------------------------------------------------------- + TOTAL 0.55695649 96.81820254 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 232184 +BPFP 0.8855 bits/point +EBPFP 1.7710 equivalent bits/point +MSE 96.818203 +---------------------- -------------------------------------------------------- +Time: 0.555s Load: 0.008s, Pack+Encode: 0.224s, Decode+Unpack: 0.322s +---------------------- -------------------------------------------------------- +💾 Converting with 96.8182 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 302, 128) +Output shape: (1, 302, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.output: torch.Size([1, 302, 3584]) -> torch.Size([1, 1, 302, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,612B, BPFP=0.5490 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,352B, BPFP=1.7773 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,300B, BPFP=0.7916 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,388B, BPFP=1.7274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,412B, BPFP=0.9009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,536B, BPFP=1.6834 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,008B, BPFP=0.8282 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,900B, BPFP=1.7022 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,984B, BPFP=1.3961 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,488B, BPFP=1.6291 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,508B, BPFP=0.2551 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15012774 63.11050031 + layer.0.v_cache 0.00001504 0.01041874 + layer.1.k_cache 0.60100308 9.87763866 + layer.1.v_cache 0.00000575 0.00416575 + layer.2.k_cache 0.00814273 0.89261359 + layer.2.v_cache 0.00002032 0.01163522 + layer.3.k_cache 0.01797006 3.96553858 + layer.3.v_cache 0.00001899 0.01283051 + layer.4.k_cache 0.00070977 0.25328026 + layer.4.v_cache 0.00004868 0.02336124 + layer.4.output 0.04414053 179.48419761 + ------------------------------------------------------------------------------------- + TOTAL 0.06394387 78.50302154 + (elements=2,628,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2628608 +Total Bytes 285488 +BPFP 0.8689 bits/point +EBPFP 1.7377 equivalent bits/point +MSE 78.503022 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.010s, Pack+Encode: 0.258s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 78.5030 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,244B, BPFP=0.5389 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,000B, BPFP=1.8657 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,104B, BPFP=0.8223 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,940B, BPFP=1.8039 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,596B, BPFP=0.9093 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,996B, BPFP=1.7488 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,636B, BPFP=0.8533 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,260B, BPFP=1.7642 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,528B, BPFP=1.4300 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,168B, BPFP=1.7006 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,072B, BPFP=0.2921 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14350806 62.93204874 + layer.0.v_cache 0.00001655 0.01036928 + layer.1.k_cache 0.46812200 10.14359568 + layer.1.v_cache 0.00000598 0.00398105 + layer.2.k_cache 0.01096032 0.95374418 + layer.2.v_cache 0.00002019 0.01101832 + layer.3.k_cache 0.00658789 3.85760452 + layer.3.v_cache 0.00002014 0.01269079 + layer.4.k_cache 0.00067528 0.24855207 + layer.4.v_cache 0.00005785 0.02393417 + layer.4.output 0.00496594 206.93375200 + ------------------------------------------------------------------------------------- + TOTAL 0.03910211 89.80787075 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 265544 +BPFP 0.9107 bits/point +EBPFP 1.8214 equivalent bits/point +MSE 89.807871 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.011s, Pack+Encode: 0.250s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 89.8079 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 260, 128) +Output shape: (1, 260, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.output: torch.Size([1, 260, 3584]) -> torch.Size([1, 1, 260, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,096B, BPFP=0.4865 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,204B, BPFP=1.8752 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,304B, BPFP=0.7995 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,252B, BPFP=1.8180 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,452B, BPFP=0.9286 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,328B, BPFP=1.7625 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,176B, BPFP=0.8519 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,840B, BPFP=1.7933 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,760B, BPFP=1.4880 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,892B, BPFP=1.7363 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,324B, BPFP=0.2689 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14548651 66.38131761 + layer.0.v_cache 0.00001584 0.01105821 + layer.1.k_cache 0.47062912 9.75604811 + layer.1.v_cache 0.00000604 0.00443490 + layer.2.k_cache 0.00619981 0.94415753 + layer.2.v_cache 0.00002003 0.01261075 + layer.3.k_cache 0.01299995 3.71431885 + layer.3.v_cache 0.00001981 0.01381331 + layer.4.k_cache 0.00067777 0.26518822 + layer.4.v_cache 0.00006192 0.02591993 + layer.4.output 0.00504520 213.32280220 + ------------------------------------------------------------------------------------- + TOTAL 0.03949607 92.61108722 + (elements=2,263,040) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2263040 +Total Bytes 256628 +BPFP 0.9072 bits/point +EBPFP 1.8144 equivalent bits/point +MSE 92.611087 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 92.6111 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,540B, BPFP=0.5480 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,608B, BPFP=1.8732 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,056B, BPFP=0.8074 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,400B, BPFP=1.8038 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,992B, BPFP=0.9187 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,392B, BPFP=1.7459 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,852B, BPFP=0.8532 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,880B, BPFP=1.7739 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,776B, BPFP=1.4233 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,464B, BPFP=1.6926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,600B, BPFP=0.2921 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14555165 61.54892865 + layer.0.v_cache 0.00001661 0.01096352 + layer.1.k_cache 0.55468144 10.23326021 + layer.1.v_cache 0.00000621 0.00424893 + layer.2.k_cache 0.02243149 0.98334941 + layer.2.v_cache 0.00002103 0.01126662 + layer.3.k_cache 0.02028967 4.22193954 + layer.3.v_cache 0.00002036 0.01278207 + layer.4.k_cache 0.00072975 0.25526661 + layer.4.v_cache 0.00005124 0.02301491 + layer.4.output 0.00486695 203.79646468 + ------------------------------------------------------------------------------------- + TOTAL 0.04575695 88.46354548 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 269560 +BPFP 0.9109 bits/point +EBPFP 1.8217 equivalent bits/point +MSE 88.463545 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 88.4635 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 356, 128) +Output shape: (1, 356, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.output: torch.Size([1, 356, 3584]) -> torch.Size([1, 1, 356, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,008B, BPFP=0.5270 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,620B, BPFP=1.7828 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,228B, BPFP=0.8000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,420B, BPFP=1.7302 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,984B, BPFP=0.8771 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,896B, BPFP=1.6633 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,440B, BPFP=0.8093 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,936B, BPFP=1.6650 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,680B, BPFP=1.3466 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,356B, BPFP=1.5957 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 46,792B, BPFP=0.2934 +⌛️ [2/4] FRONTEND: Frontend time: 0.288s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.436s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15223843 60.35010863 + layer.0.v_cache 0.00001740 0.01034555 + layer.1.k_cache 0.82972649 10.04259088 + layer.1.v_cache 0.00000631 0.00409351 + layer.2.k_cache 0.03086275 0.92806338 + layer.2.v_cache 0.00001960 0.01057001 + layer.3.k_cache 0.01100341 3.50229739 + layer.3.v_cache 0.00002062 0.01190094 + layer.4.k_cache 0.00073857 0.24331984 + layer.4.v_cache 0.00005279 0.02128358 + layer.4.output 0.03758925 152.24752959 + ------------------------------------------------------------------------------------- + TOTAL 0.07575360 67.10925182 + (elements=3,098,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3098624 +Total Bytes 338360 +BPFP 0.8736 bits/point +EBPFP 1.7471 equivalent bits/point +MSE 67.109252 +---------------------- -------------------------------------------------------- +Time: 0.736s Load: 0.012s, Pack+Encode: 0.288s, Decode+Unpack: 0.436s +---------------------- -------------------------------------------------------- +💾 Converting with 67.1093 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,004B, BPFP=0.5289 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,952B, BPFP=1.8769 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,068B, BPFP=0.8264 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,944B, BPFP=1.8177 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,964B, BPFP=0.9377 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,116B, BPFP=1.7690 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,516B, BPFP=0.8527 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,552B, BPFP=1.7946 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,688B, BPFP=1.4502 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,196B, BPFP=1.7150 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,024B, BPFP=0.2939 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12347319 64.41156602 + layer.0.v_cache 0.00001696 0.01068897 + layer.1.k_cache 0.44284580 9.80553557 + layer.1.v_cache 0.00000660 0.00430530 + layer.2.k_cache 0.00882101 0.98877160 + layer.2.v_cache 0.00002210 0.01203293 + layer.3.k_cache 0.01145097 3.66897514 + layer.3.v_cache 0.00001984 0.01274111 + layer.4.k_cache 0.00068754 0.26846041 + layer.4.v_cache 0.00006483 0.02367824 + layer.4.output 0.00500964 208.58319347 + ------------------------------------------------------------------------------------- + TOTAL 0.03661626 90.54641821 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 266024 +BPFP 0.9192 bits/point +EBPFP 1.8384 equivalent bits/point +MSE 90.546418 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.008s, Pack+Encode: 0.252s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 90.5464 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,004B, BPFP=0.5485 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,520B, BPFP=1.8377 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,544B, BPFP=0.7974 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,288B, BPFP=1.7702 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,920B, BPFP=0.9276 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,448B, BPFP=1.7241 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,640B, BPFP=0.8575 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,896B, BPFP=1.7487 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,936B, BPFP=1.4219 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,528B, BPFP=1.6737 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,888B, BPFP=0.2889 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13002123 61.85891584 + layer.0.v_cache 0.00001833 0.01133777 + layer.1.k_cache 0.50397591 9.63706140 + layer.1.v_cache 0.00000608 0.00422832 + layer.2.k_cache 0.01926375 0.91463998 + layer.2.v_cache 0.00002060 0.01189369 + layer.3.k_cache 0.04992905 4.09109187 + layer.3.v_cache 0.00002096 0.01414018 + layer.4.k_cache 0.00069540 0.25992057 + layer.4.v_cache 0.00005467 0.02465259 + layer.4.output 0.00466491 194.62603383 + ------------------------------------------------------------------------------------- + TOTAL 0.04333296 84.65941877 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 279612 +BPFP 0.9017 bits/point +EBPFP 1.8035 equivalent bits/point +MSE 84.659419 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 84.6594 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,096B, BPFP=0.5634 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,704B, BPFP=1.8250 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,360B, BPFP=0.8013 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,592B, BPFP=1.7629 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,896B, BPFP=0.8871 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,364B, BPFP=1.6944 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,936B, BPFP=0.8335 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,608B, BPFP=1.7080 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,796B, BPFP=1.3837 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,484B, BPFP=1.6453 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,500B, BPFP=0.3468 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15974760 58.21864537 + layer.0.v_cache 0.00001739 0.01030018 + layer.1.k_cache 0.45644542 10.12105713 + layer.1.v_cache 0.00000643 0.00401789 + layer.2.k_cache 0.01631587 0.97115239 + layer.2.v_cache 0.00002005 0.01029637 + layer.3.k_cache 0.01511352 3.89179077 + layer.3.v_cache 0.00002063 0.01248821 + layer.4.k_cache 0.00073517 0.24295973 + layer.4.v_cache 0.00005091 0.02225147 + layer.4.output 0.00481773 198.01122449 + ------------------------------------------------------------------------------------- + TOTAL 0.04012924 85.85785476 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 278336 +BPFP 0.9137 bits/point +EBPFP 1.8273 equivalent bits/point +MSE 85.857855 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8579 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,140B, BPFP=0.5349 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,208B, BPFP=1.8848 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,232B, BPFP=0.8329 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,332B, BPFP=1.8336 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,756B, BPFP=0.9221 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,020B, BPFP=1.7568 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,616B, BPFP=0.8553 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,524B, BPFP=1.7863 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,684B, BPFP=1.4445 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,928B, BPFP=1.6929 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,968B, BPFP=0.2756 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14148096 65.29725538 + layer.0.v_cache 0.00001588 0.01045973 + layer.1.k_cache 0.55266568 9.70616441 + layer.1.v_cache 0.00000632 0.00424311 + layer.2.k_cache 0.00687984 0.94324245 + layer.2.v_cache 0.00001885 0.01097763 + layer.3.k_cache 0.01675770 3.53993757 + layer.3.v_cache 0.00001964 0.01287672 + layer.4.k_cache 0.00068991 0.25332233 + layer.4.v_cache 0.00004859 0.02308840 + layer.4.output 0.00492290 207.66716827 + ------------------------------------------------------------------------------------- + TOTAL 0.04429669 90.20422033 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 264408 +BPFP 0.9102 bits/point +EBPFP 1.8204 equivalent bits/point +MSE 90.204220 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 90.2042 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,316B, BPFP=0.4990 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,072B, BPFP=1.8417 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,416B, BPFP=0.7941 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,640B, BPFP=1.7724 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,376B, BPFP=0.8889 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,236B, BPFP=1.7045 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,864B, BPFP=0.8158 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,852B, BPFP=1.7343 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,832B, BPFP=1.3947 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,084B, BPFP=1.6488 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,060B, BPFP=0.2976 +⌛️ [2/4] FRONTEND: Frontend time: 0.279s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.441s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13144862 62.13157290 + layer.0.v_cache 0.00001769 0.01125595 + layer.1.k_cache 0.64513603 9.99603253 + layer.1.v_cache 0.00000607 0.00425179 + layer.2.k_cache 0.02726801 0.98113088 + layer.2.v_cache 0.00001956 0.01116110 + layer.3.k_cache 0.02271994 3.86095247 + layer.3.v_cache 0.00002018 0.01258242 + layer.4.k_cache 0.00073057 0.25385473 + layer.4.v_cache 0.00004940 0.02206841 + layer.4.output 0.04135688 167.80106977 + ------------------------------------------------------------------------------------- + TOTAL 0.06570084 73.64072656 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 313748 +BPFP 0.8928 bits/point +EBPFP 1.7856 equivalent bits/point +MSE 73.640727 +---------------------- -------------------------------------------------------- +Time: 0.732s Load: 0.012s, Pack+Encode: 0.279s, Decode+Unpack: 0.441s +---------------------- -------------------------------------------------------- +💾 Converting with 73.6407 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,868B, BPFP=0.5718 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,180B, BPFP=1.9026 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,196B, BPFP=0.8137 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,068B, BPFP=1.8218 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,708B, BPFP=0.9235 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,404B, BPFP=1.7735 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,804B, BPFP=0.8578 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,668B, BPFP=1.7927 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,980B, BPFP=1.4520 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,560B, BPFP=1.7122 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,344B, BPFP=0.2839 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12396691 62.33389807 + layer.0.v_cache 0.00001598 0.01128318 + layer.1.k_cache 0.24854917 10.01303597 + layer.1.v_cache 0.00000584 0.00426842 + layer.2.k_cache 0.02290044 0.95478828 + layer.2.v_cache 0.00002299 0.01169056 + layer.3.k_cache 0.01294550 4.01139285 + layer.3.v_cache 0.00002077 0.01365214 + layer.4.k_cache 0.00065144 0.25644260 + layer.4.v_cache 0.00005054 0.02399035 + layer.4.output 1.42385976 251.52466777 + ------------------------------------------------------------------------------------- + TOTAL 0.61036164 108.13571276 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 214780 +BPFP 0.9182 bits/point +EBPFP 1.8364 equivalent bits/point +MSE 108.135713 +---------------------- -------------------------------------------------------- +Time: 0.557s Load: 0.008s, Pack+Encode: 0.225s, Decode+Unpack: 0.324s +---------------------- -------------------------------------------------------- +💾 Converting with 108.1357 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,808B, BPFP=0.5303 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,468B, BPFP=1.8095 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,632B, BPFP=0.7911 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,280B, BPFP=1.7452 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,596B, BPFP=0.8973 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,424B, BPFP=1.6990 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,312B, BPFP=0.8279 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,856B, BPFP=1.7223 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,656B, BPFP=1.3871 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,224B, BPFP=1.6341 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,156B, BPFP=0.2638 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14783513 62.20620675 + layer.0.v_cache 0.00001600 0.01115137 + layer.1.k_cache 0.54442752 9.89440960 + layer.1.v_cache 0.00000628 0.00435249 + layer.2.k_cache 0.01848349 0.95187484 + layer.2.v_cache 0.00002061 0.01178255 + layer.3.k_cache 0.01777397 3.83603803 + layer.3.v_cache 0.00002092 0.01375641 + layer.4.k_cache 0.00070447 0.25965000 + layer.4.v_cache 0.00005049 0.02374279 + layer.4.output 0.04615090 187.56282439 + ------------------------------------------------------------------------------------- + TOTAL 0.06190560 81.77369033 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 275412 +BPFP 0.8759 bits/point +EBPFP 1.7518 equivalent bits/point +MSE 81.773690 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 81.7737 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,532B, BPFP=0.5522 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,552B, BPFP=1.8117 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,276B, BPFP=0.8010 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,248B, BPFP=1.7433 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,204B, BPFP=0.9021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,244B, BPFP=1.6906 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,876B, BPFP=0.8324 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,732B, BPFP=1.7162 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,480B, BPFP=1.3884 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,128B, BPFP=1.6321 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,232B, BPFP=0.2864 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16274841 61.38451657 + layer.0.v_cache 0.00001794 0.01099675 + layer.1.k_cache 0.56737488 9.94892300 + layer.1.v_cache 0.00000586 0.00404093 + layer.2.k_cache 0.02027899 0.99170495 + layer.2.v_cache 0.00002117 0.01128509 + layer.3.k_cache 0.01891868 3.92658372 + layer.3.v_cache 0.00002140 0.01273835 + layer.4.k_cache 0.00069658 0.25179939 + layer.4.v_cache 0.00005005 0.02234235 + layer.4.output 0.04476916 181.86662572 + ------------------------------------------------------------------------------------- + TOTAL 0.06373636 79.39007712 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 287504 +BPFP 0.8867 bits/point +EBPFP 1.7735 equivalent bits/point +MSE 79.390077 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.012s, Pack+Encode: 0.251s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 79.3901 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,824B, BPFP=0.5623 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,500B, BPFP=1.8601 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,088B, BPFP=0.8063 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,224B, BPFP=1.7871 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,788B, BPFP=0.9036 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,156B, BPFP=1.7260 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,528B, BPFP=0.8315 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,588B, BPFP=1.7507 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,532B, BPFP=1.4041 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,084B, BPFP=1.6646 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,556B, BPFP=0.2825 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13366390 60.89731427 + layer.0.v_cache 0.00001719 0.01116181 + layer.1.k_cache 0.57618574 9.93912492 + layer.1.v_cache 0.00000658 0.00422299 + layer.2.k_cache 0.02858028 0.97632719 + layer.2.v_cache 0.00002021 0.01131039 + layer.3.k_cache 0.00873641 3.98834564 + layer.3.v_cache 0.00002023 0.01306607 + layer.4.k_cache 0.00067337 0.25283425 + layer.4.v_cache 0.00005011 0.02277481 + layer.4.output 0.00484872 203.15172030 + ------------------------------------------------------------------------------------- + TOTAL 0.04599383 88.12814850 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 266868 +BPFP 0.8985 bits/point +EBPFP 1.7969 equivalent bits/point +MSE 88.128148 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 88.1281 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,756B, BPFP=0.5484 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,512B, BPFP=1.8744 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,748B, BPFP=0.8306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,580B, BPFP=1.8085 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,596B, BPFP=0.9613 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,836B, BPFP=1.7559 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,664B, BPFP=0.8954 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,280B, BPFP=1.7873 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,628B, BPFP=1.4584 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,124B, BPFP=1.7056 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,224B, BPFP=0.2851 +⌛️ [2/4] FRONTEND: Frontend time: 0.233s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13057878 63.81968503 + layer.0.v_cache 0.00001797 0.01227328 + layer.1.k_cache 0.31658528 10.22491759 + layer.1.v_cache 0.00000647 0.00454539 + layer.2.k_cache 0.02849985 1.00115242 + layer.2.v_cache 0.00001991 0.01237259 + layer.3.k_cache 0.03599559 4.07739424 + layer.3.v_cache 0.00002239 0.01505300 + layer.4.k_cache 0.00078811 0.27522799 + layer.4.v_cache 0.00005540 0.02542806 + layer.4.output 1.38530015 244.76609971 + ------------------------------------------------------------------------------------- + TOTAL 0.60056887 105.46063221 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 220948 +BPFP 0.9189 bits/point +EBPFP 1.8378 equivalent bits/point +MSE 105.460632 +---------------------- -------------------------------------------------------- +Time: 0.567s Load: 0.014s, Pack+Encode: 0.233s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 105.4606 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,760B, BPFP=0.5588 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,256B, BPFP=1.8906 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,356B, BPFP=0.8177 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,296B, BPFP=1.8214 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,112B, BPFP=0.9441 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,656B, BPFP=1.7753 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,084B, BPFP=0.8701 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,056B, BPFP=1.8041 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,252B, BPFP=1.4582 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,792B, BPFP=1.7131 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,028B, BPFP=0.2883 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12725546 61.53592580 + layer.0.v_cache 0.00001736 0.01195781 + layer.1.k_cache 0.32262681 10.23403298 + layer.1.v_cache 0.00000608 0.00444750 + layer.2.k_cache 0.02157613 1.03011459 + layer.2.v_cache 0.00002208 0.01232317 + layer.3.k_cache 0.01760322 4.02124220 + layer.3.v_cache 0.00002192 0.01463333 + layer.4.k_cache 0.00069785 0.27679820 + layer.4.v_cache 0.00005172 0.02543849 + layer.4.output 1.41081570 249.40236175 + ------------------------------------------------------------------------------------- + TOTAL 0.60974050 107.23432037 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 217648 +BPFP 0.9219 bits/point +EBPFP 1.8437 equivalent bits/point +MSE 107.234320 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.007s, Pack+Encode: 0.223s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 107.2343 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,808B, BPFP=0.5728 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,476B, BPFP=1.9422 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,224B, BPFP=0.8234 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,528B, BPFP=1.8727 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,340B, BPFP=0.9786 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,652B, BPFP=1.8084 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,084B, BPFP=0.8864 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,188B, BPFP=1.8477 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,220B, BPFP=1.4833 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,936B, BPFP=1.7559 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,272B, BPFP=0.2963 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13788442 59.39857963 + layer.0.v_cache 0.00001717 0.01243596 + layer.1.k_cache 0.30723858 10.19630969 + layer.1.v_cache 0.00000664 0.00482657 + layer.2.k_cache 0.01337219 1.01969881 + layer.2.v_cache 0.00002159 0.01279095 + layer.3.k_cache 0.03522867 4.06009527 + layer.3.v_cache 0.00002126 0.01488170 + layer.4.k_cache 0.00073009 0.27588168 + layer.4.v_cache 0.00005119 0.02561173 + layer.4.output 1.43730973 254.04279846 + ------------------------------------------------------------------------------------- + TOTAL 0.62092588 109.01886478 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 218728 +BPFP 0.9438 bits/point +EBPFP 1.8877 equivalent bits/point +MSE 109.018865 +---------------------- -------------------------------------------------------- +Time: 0.546s Load: 0.007s, Pack+Encode: 0.222s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 109.0189 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,324B, BPFP=0.5610 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,788B, BPFP=1.9752 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,236B, BPFP=0.8606 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,676B, BPFP=1.8900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,784B, BPFP=0.9792 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,860B, BPFP=1.8275 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,816B, BPFP=0.9050 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,536B, BPFP=1.8793 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,596B, BPFP=1.5009 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,092B, BPFP=1.7687 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,044B, BPFP=0.2959 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11546621 63.21460440 + layer.0.v_cache 0.00001823 0.01192864 + layer.1.k_cache 0.26924565 10.26970598 + layer.1.v_cache 0.00000737 0.00472032 + layer.2.k_cache 0.01761903 0.97615650 + layer.2.v_cache 0.00002145 0.01246999 + layer.3.k_cache 0.02374693 3.99174948 + layer.3.v_cache 0.00002104 0.01422080 + layer.4.k_cache 0.00067272 0.26570204 + layer.4.v_cache 0.00005153 0.02436552 + layer.4.output 1.50071327 265.14375438 + ------------------------------------------------------------------------------------- + TOTAL 0.64305077 113.81128849 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 211752 +BPFP 0.9540 bits/point +EBPFP 1.9081 equivalent bits/point +MSE 113.811288 +---------------------- -------------------------------------------------------- +Time: 0.546s Load: 0.009s, Pack+Encode: 0.221s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 113.8113 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,544B, BPFP=0.5349 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,356B, BPFP=1.7936 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,592B, BPFP=0.7910 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,340B, BPFP=1.7421 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,100B, BPFP=0.9182 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,324B, BPFP=1.6905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,740B, BPFP=0.8492 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,680B, BPFP=1.7086 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,536B, BPFP=1.3969 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,172B, BPFP=1.6321 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,424B, BPFP=0.2857 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18206066 62.54276583 + layer.0.v_cache 0.00001652 0.01106402 + layer.1.k_cache 0.62405802 9.94934003 + layer.1.v_cache 0.00000616 0.00425062 + layer.2.k_cache 0.02036299 0.95061176 + layer.2.v_cache 0.00002087 0.01184680 + layer.3.k_cache 0.02546444 3.71907658 + layer.3.v_cache 0.00002064 0.01306938 + layer.4.k_cache 0.00069880 0.25475012 + layer.4.v_cache 0.00007146 0.02426817 + layer.4.output 0.04336799 175.68748551 + ------------------------------------------------------------------------------------- + TOTAL 0.06802097 76.89961423 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 296808 +BPFP 0.8857 bits/point +EBPFP 1.7714 equivalent bits/point +MSE 76.899614 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.254s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 76.8996 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,700B, BPFP=0.5254 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,348B, BPFP=1.8660 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,136B, BPFP=0.8281 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,496B, BPFP=1.8079 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,700B, BPFP=0.9348 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,488B, BPFP=1.7391 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,600B, BPFP=0.8597 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,800B, BPFP=1.7604 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,836B, BPFP=1.4217 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,632B, BPFP=1.6807 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,592B, BPFP=0.2884 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14018573 62.99574406 + layer.0.v_cache 0.00001653 0.01143023 + layer.1.k_cache 0.40671090 10.31780393 + layer.1.v_cache 0.00000662 0.00452153 + layer.2.k_cache 0.03036383 0.87900073 + layer.2.v_cache 0.00002088 0.01181825 + layer.3.k_cache 0.05940984 4.15956589 + layer.3.v_cache 0.00002025 0.01343401 + layer.4.k_cache 0.00071161 0.26579966 + layer.4.v_cache 0.00005094 0.02261196 + layer.4.output 1.33692960 236.40256550 + ------------------------------------------------------------------------------------- + TOTAL 0.58800025 101.97056993 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 226328 +BPFP 0.9084 bits/point +EBPFP 1.8168 equivalent bits/point +MSE 101.970570 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.008s, Pack+Encode: 0.225s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 101.9706 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,648B, BPFP=0.5457 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,660B, BPFP=1.9021 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,536B, BPFP=0.8231 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,628B, BPFP=1.8285 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,460B, BPFP=0.9603 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,924B, BPFP=1.7783 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,132B, BPFP=0.8656 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,380B, BPFP=1.8108 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,512B, BPFP=1.4635 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,236B, BPFP=1.7292 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,812B, BPFP=0.3039 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351684 67.98731807 + layer.0.v_cache 0.00001872 0.01232586 + layer.1.k_cache 0.28219827 10.14476982 + layer.1.v_cache 0.00000627 0.00475505 + layer.2.k_cache 0.01639774 0.95747619 + layer.2.v_cache 0.00002172 0.01353362 + layer.3.k_cache 0.02752649 4.12981230 + layer.3.v_cache 0.00002090 0.01468956 + layer.4.k_cache 0.00067264 0.27114448 + layer.4.v_cache 0.00005058 0.02626971 + layer.4.output 1.39794763 246.84748043 + ------------------------------------------------------------------------------------- + TOTAL 0.60212139 106.55849751 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 221928 +BPFP 0.9314 bits/point +EBPFP 1.8628 equivalent bits/point +MSE 106.558498 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.010s, Pack+Encode: 0.226s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 106.5585 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,248B, BPFP=0.5392 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,260B, BPFP=1.8808 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,100B, BPFP=0.8221 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,320B, BPFP=1.8260 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,628B, BPFP=0.9111 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,904B, BPFP=1.7435 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,544B, BPFP=0.8479 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,176B, BPFP=1.7593 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,392B, BPFP=1.4221 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,944B, BPFP=1.6875 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,904B, BPFP=0.2907 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15646637 62.05328096 + layer.0.v_cache 0.00001727 0.01118765 + layer.1.k_cache 0.53079696 10.33771025 + layer.1.v_cache 0.00000635 0.00448645 + layer.2.k_cache 0.02340859 1.02377684 + layer.2.v_cache 0.00002061 0.01194508 + layer.3.k_cache 0.01050839 3.85653504 + layer.3.v_cache 0.00002156 0.01284985 + layer.4.k_cache 0.00068596 0.26184156 + layer.4.v_cache 0.00006072 0.02371252 + layer.4.output 0.00497423 206.92215818 + ------------------------------------------------------------------------------------- + TOTAL 0.04451838 89.76779021 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 265420 +BPFP 0.9103 bits/point +EBPFP 1.8205 equivalent bits/point +MSE 89.767790 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 89.7678 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,584B, BPFP=0.5369 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,312B, BPFP=1.7914 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,620B, BPFP=0.7924 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,308B, BPFP=1.7405 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,744B, BPFP=0.9002 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,020B, BPFP=1.6751 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,264B, BPFP=0.8251 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,204B, BPFP=1.6845 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,100B, BPFP=1.3748 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,664B, BPFP=1.6063 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,372B, BPFP=0.2998 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15292851 61.29455408 + layer.0.v_cache 0.00001614 0.01156681 + layer.1.k_cache 0.61736927 10.09304136 + layer.1.v_cache 0.00000669 0.00465635 + layer.2.k_cache 0.02003183 0.94749916 + layer.2.v_cache 0.00002344 0.01232482 + layer.3.k_cache 0.01028479 3.79413991 + layer.3.v_cache 0.00002212 0.01307511 + layer.4.k_cache 0.00070434 0.26300547 + layer.4.v_cache 0.00006490 0.02346075 + layer.4.output 0.04342050 175.57289251 + ------------------------------------------------------------------------------------- + TOTAL 0.06502327 76.79221008 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 296192 +BPFP 0.8839 bits/point +EBPFP 1.7678 equivalent bits/point +MSE 76.792210 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 76.7922 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,384B, BPFP=0.5451 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,524B, BPFP=1.8892 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,172B, BPFP=0.8232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,316B, BPFP=1.8190 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,072B, BPFP=0.9336 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,296B, BPFP=1.7598 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,900B, BPFP=0.8655 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,996B, BPFP=1.8004 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,852B, BPFP=1.4435 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,224B, BPFP=1.6975 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,148B, BPFP=0.2834 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12950333 63.81791647 + layer.0.v_cache 0.00001690 0.01136463 + layer.1.k_cache 0.52417032 10.07976138 + layer.1.v_cache 0.00000602 0.00446471 + layer.2.k_cache 0.01578268 1.06084422 + layer.2.v_cache 0.00002268 0.01244808 + layer.3.k_cache 0.03345619 4.17312928 + layer.3.v_cache 0.00002007 0.01406469 + layer.4.k_cache 0.00067301 0.25613911 + layer.4.v_cache 0.00005077 0.02421111 + layer.4.output 0.00489311 206.15731213 + ------------------------------------------------------------------------------------- + TOTAL 0.04340905 89.56208992 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 267884 +BPFP 0.9153 bits/point +EBPFP 1.8306 equivalent bits/point +MSE 89.562090 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.009s, Pack+Encode: 0.252s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 89.5621 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,868B, BPFP=0.5563 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,172B, BPFP=1.8504 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,772B, BPFP=0.8323 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,200B, BPFP=1.7817 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,176B, BPFP=0.9316 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,360B, BPFP=1.7223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,312B, BPFP=0.8705 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,980B, BPFP=1.7661 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,124B, BPFP=1.4228 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,680B, BPFP=1.6742 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,628B, BPFP=0.2588 +⌛️ [2/4] FRONTEND: Frontend time: 0.236s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13888381 66.07745334 + layer.0.v_cache 0.00001731 0.01117478 + layer.1.k_cache 0.27925331 9.98805699 + layer.1.v_cache 0.00000589 0.00428606 + layer.2.k_cache 0.03172064 0.91327277 + layer.2.v_cache 0.00001980 0.01194903 + layer.3.k_cache 0.01738505 4.08698504 + layer.3.v_cache 0.00001961 0.01327675 + layer.4.k_cache 0.00068340 0.25730389 + layer.4.v_cache 0.00005063 0.02425779 + layer.4.output 1.38525280 244.89621041 + ------------------------------------------------------------------------------------- + TOTAL 0.59792994 105.62714643 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 215272 +BPFP 0.8953 bits/point +EBPFP 1.7906 equivalent bits/point +MSE 105.627146 +---------------------- -------------------------------------------------------- +Time: 0.571s Load: 0.009s, Pack+Encode: 0.236s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 105.6271 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.9009 bits/point +Avg EBPFP 1.8019 equivalent bits/point +Avg MSE 90.191382 +Avg Time 0.667s +------------------------ ---------------------------- diff --git a/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..afcd875c149a92c313c2bea792526796c58f85cd --- /dev/null +++ b/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 599 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other +Output output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,592B, BPFP=0.5124 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,264B, BPFP=1.9639 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,020B, BPFP=0.8567 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,216B, BPFP=1.8825 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,596B, BPFP=0.9792 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,596B, BPFP=1.8343 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,636B, BPFP=0.9045 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,248B, BPFP=1.8850 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,584B, BPFP=1.5224 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,168B, BPFP=1.8010 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,172B, BPFP=0.2684 +⌛️ [2/4] FRONTEND: Frontend time: 0.472s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.395s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09475397 63.30023224 + layer.0.v_cache 0.00001927 0.01193357 + layer.1.k_cache 0.17401673 9.81945376 + layer.1.v_cache 0.00000670 0.00460013 + layer.2.k_cache 0.00814445 1.02673780 + layer.2.v_cache 0.00002157 0.01294463 + layer.3.k_cache 0.01691350 3.87213955 + layer.3.v_cache 0.00002161 0.01581721 + layer.4.k_cache 0.00069969 0.29261472 + layer.4.v_cache 0.00005477 0.02657354 + layer.4.output 1.52309445 269.30679193 + ------------------------------------------------------------------------------------- + TOTAL 0.64448903 115.50179945 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 206092 +BPFP 0.9424 bits/point +EBPFP 1.8848 equivalent bits/point +MSE 115.501799 +---------------------- -------------------------------------------------------- +Time: 0.877s Load: 0.009s, Pack+Encode: 0.472s, Decode+Unpack: 0.395s +---------------------- -------------------------------------------------------- +💾 Converting with 115.5018 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 212, 128) +Output shape: (1, 212, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.output: torch.Size([1, 212, 3584]) -> torch.Size([1, 1, 212, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,584B, BPFP=0.5590 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,300B, BPFP=1.9384 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,952B, BPFP=0.8072 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,280B, BPFP=1.8632 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,080B, BPFP=0.9640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,728B, BPFP=1.8225 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,220B, BPFP=0.9006 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,288B, BPFP=1.8638 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,208B, BPFP=1.4894 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,868B, BPFP=1.7591 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,392B, BPFP=0.2884 +⌛️ [2/4] FRONTEND: Frontend time: 0.233s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09403566 62.58337633 + layer.0.v_cache 0.00001799 0.01248997 + layer.1.k_cache 0.24957257 10.36869610 + layer.1.v_cache 0.00000621 0.00463893 + layer.2.k_cache 0.01169950 0.99101804 + layer.2.v_cache 0.00002031 0.01302014 + layer.3.k_cache 0.03645871 4.03244537 + layer.3.v_cache 0.00002058 0.01542770 + layer.4.k_cache 0.00069161 0.28222793 + layer.4.v_cache 0.00005130 0.02616017 + layer.4.output 1.44404146 255.15277544 + ------------------------------------------------------------------------------------- + TOTAL 0.61769792 109.67052522 + (elements=1,845,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1845248 +Total Bytes 216900 +BPFP 0.9404 bits/point +EBPFP 1.8807 equivalent bits/point +MSE 109.670525 +---------------------- -------------------------------------------------------- +Time: 0.573s Load: 0.010s, Pack+Encode: 0.233s, Decode+Unpack: 0.330s +---------------------- -------------------------------------------------------- +💾 Converting with 109.6705 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 250, 128) +Output shape: (1, 250, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.output: torch.Size([1, 250, 3584]) -> torch.Size([1, 1, 250, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,560B, BPFP=0.5350 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,228B, BPFP=1.7643 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,500B, BPFP=0.7812 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,392B, BPFP=1.7120 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,480B, BPFP=0.9050 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,616B, BPFP=1.6635 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,324B, BPFP=0.8327 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,904B, BPFP=1.6815 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,016B, BPFP=1.3760 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,768B, BPFP=1.6105 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,984B, BPFP=0.3124 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.331s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09121272 67.12390625 + layer.0.v_cache 0.00001793 0.01093153 + layer.1.k_cache 0.34710925 9.81217578 + layer.1.v_cache 0.00000615 0.00432605 + layer.2.k_cache 0.02211576 0.91791669 + layer.2.v_cache 0.00002179 0.01184161 + layer.3.k_cache 0.02174057 3.72916968 + layer.3.v_cache 0.00001938 0.01284504 + layer.4.k_cache 0.00073173 0.25375223 + layer.4.v_cache 0.00005145 0.02324653 + layer.4.output 1.22463198 216.44303571 + ------------------------------------------------------------------------------------- + TOTAL 0.53267356 93.94125655 + (elements=2,176,000) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2176000 +Total Bytes 240772 +BPFP 0.8852 bits/point +EBPFP 1.7704 equivalent bits/point +MSE 93.941257 +---------------------- -------------------------------------------------------- +Time: 0.568s Load: 0.009s, Pack+Encode: 0.229s, Decode+Unpack: 0.331s +---------------------- -------------------------------------------------------- +💾 Converting with 93.9413 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,796B, BPFP=0.5366 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,904B, BPFP=1.8519 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,820B, BPFP=0.8136 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,092B, BPFP=1.7960 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,524B, BPFP=0.9309 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,324B, BPFP=1.7431 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,464B, BPFP=0.8579 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,608B, BPFP=1.7627 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,764B, BPFP=1.4292 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,464B, BPFP=1.6839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,824B, BPFP=0.2736 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12780563 61.71742497 + layer.0.v_cache 0.00001798 0.01166546 + layer.1.k_cache 0.36413978 9.85821775 + layer.1.v_cache 0.00000619 0.00446429 + layer.2.k_cache 0.01529831 0.96100438 + layer.2.v_cache 0.00002052 0.01213990 + layer.3.k_cache 0.01269315 4.06133092 + layer.3.v_cache 0.00002023 0.01382875 + layer.4.k_cache 0.00069261 0.26367046 + layer.4.v_cache 0.00005545 0.02448631 + layer.4.output 1.34865724 238.40111705 + ------------------------------------------------------------------------------------- + TOTAL 0.58596180 102.69035603 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 222584 +BPFP 0.9012 bits/point +EBPFP 1.8025 equivalent bits/point +MSE 102.690356 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.009s, Pack+Encode: 0.226s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 102.6904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 214, 128) +Output shape: (1, 214, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.output: torch.Size([1, 214, 3584]) -> torch.Size([1, 1, 214, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,716B, BPFP=0.5634 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,968B, BPFP=1.8960 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,280B, BPFP=0.8236 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,168B, BPFP=1.8376 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,788B, BPFP=0.9337 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,352B, BPFP=1.7780 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,756B, BPFP=0.8584 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,668B, BPFP=1.8011 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,920B, BPFP=1.4544 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,524B, BPFP=1.7176 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,788B, BPFP=0.3003 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12820032 61.17161032 + layer.0.v_cache 0.00001693 0.01106897 + layer.1.k_cache 0.24875343 10.23264092 + layer.1.v_cache 0.00000587 0.00404869 + layer.2.k_cache 0.01453762 0.97729727 + layer.2.v_cache 0.00001988 0.01134094 + layer.3.k_cache 0.02276520 3.72132004 + layer.3.v_cache 0.00001932 0.01279988 + layer.4.k_cache 0.00071399 0.26305707 + layer.4.v_cache 0.00004895 0.02331025 + layer.4.output 1.43051836 252.84251919 + ------------------------------------------------------------------------------------- + TOTAL 0.61345353 108.60741933 + (elements=1,862,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1862656 +Total Bytes 215928 +BPFP 0.9274 bits/point +EBPFP 1.8548 equivalent bits/point +MSE 108.607419 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.008s, Pack+Encode: 0.226s, Decode+Unpack: 0.314s +---------------------- -------------------------------------------------------- +💾 Converting with 108.6074 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,940B, BPFP=0.5607 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,788B, BPFP=1.8495 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,200B, BPFP=0.8010 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,612B, BPFP=1.7832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,228B, BPFP=0.9154 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,884B, BPFP=1.7421 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,868B, BPFP=0.8387 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,292B, BPFP=1.7651 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,996B, BPFP=1.4100 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,964B, BPFP=1.6902 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,952B, BPFP=0.3058 +⌛️ [2/4] FRONTEND: Frontend time: 0.310s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11875103 62.12145335 + layer.0.v_cache 0.00001689 0.01089972 + layer.1.k_cache 0.49171178 10.12028024 + layer.1.v_cache 0.00000616 0.00419604 + layer.2.k_cache 0.02532451 0.96300950 + layer.2.v_cache 0.00002086 0.01182025 + layer.3.k_cache 0.02370731 3.82871552 + layer.3.v_cache 0.00002129 0.01317950 + layer.4.k_cache 0.00073268 0.26079127 + layer.4.v_cache 0.00004879 0.02366267 + layer.4.output 0.00480151 200.21103017 + ------------------------------------------------------------------------------------- + TOTAL 0.04082070 86.99030701 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 274724 +BPFP 0.9116 bits/point +EBPFP 1.8231 equivalent bits/point +MSE 86.990307 +---------------------- -------------------------------------------------------- +Time: 0.697s Load: 0.010s, Pack+Encode: 0.310s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9903 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,528B, BPFP=0.5439 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,156B, BPFP=1.7957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,584B, BPFP=0.8026 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,344B, BPFP=1.7439 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,268B, BPFP=0.9099 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,680B, BPFP=1.7015 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,452B, BPFP=0.8579 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,024B, BPFP=1.7235 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,040B, BPFP=1.4056 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,844B, BPFP=1.6482 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,792B, BPFP=0.2623 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10657067 60.88516422 + layer.0.v_cache 0.00001604 0.01094793 + layer.1.k_cache 0.39270948 10.26575056 + layer.1.v_cache 0.00000611 0.00433092 + layer.2.k_cache 0.01541229 0.87886041 + layer.2.v_cache 0.00002285 0.01263087 + layer.3.k_cache 0.01321528 3.65835833 + layer.3.v_cache 0.00001985 0.01350205 + layer.4.k_cache 0.00076706 0.26362199 + layer.4.v_cache 0.00005205 0.02344799 + layer.4.output 1.24960368 220.75010933 + ------------------------------------------------------------------------------------- + TOTAL 0.54564808 95.36866945 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 234712 +BPFP 0.8805 bits/point +EBPFP 1.7610 equivalent bits/point +MSE 95.368669 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.009s, Pack+Encode: 0.224s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 95.3687 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 197, 128) +Output shape: (1, 197, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.output: torch.Size([1, 197, 3584]) -> torch.Size([1, 1, 197, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,368B, BPFP=0.5051 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,568B, BPFP=1.9486 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,568B, BPFP=0.8382 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,392B, BPFP=1.8553 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,192B, BPFP=0.9670 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,144B, BPFP=1.8357 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,184B, BPFP=0.8871 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,408B, BPFP=1.8566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,068B, BPFP=1.5124 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,504B, BPFP=1.7849 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,228B, BPFP=0.2745 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09219642 63.09247105 + layer.0.v_cache 0.00001621 0.01122015 + layer.1.k_cache 0.17840836 10.08081241 + layer.1.v_cache 0.00000639 0.00447452 + layer.2.k_cache 0.00922235 0.98915882 + layer.2.v_cache 0.00002142 0.01242947 + layer.3.k_cache 0.01957878 4.11210330 + layer.3.v_cache 0.00002124 0.01392832 + layer.4.k_cache 0.00068177 0.27252410 + layer.4.v_cache 0.00005086 0.02558281 + layer.4.output 1.55401793 274.76661077 + ------------------------------------------------------------------------------------- + TOTAL 0.65754878 117.76358708 + (elements=1,714,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1714688 +Total Bytes 200624 +BPFP 0.9360 bits/point +EBPFP 1.8721 equivalent bits/point +MSE 117.763587 +---------------------- -------------------------------------------------------- +Time: 0.545s Load: 0.007s, Pack+Encode: 0.223s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 117.7636 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,596B, BPFP=0.5341 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,696B, BPFP=1.7488 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,692B, BPFP=0.7909 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,560B, BPFP=1.6915 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,548B, BPFP=0.8845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,812B, BPFP=1.6538 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,032B, BPFP=0.8081 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,284B, BPFP=1.6776 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,768B, BPFP=1.3492 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,920B, BPFP=1.6089 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,716B, BPFP=0.2788 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11169702 58.26185736 + layer.0.v_cache 0.00001595 0.01060397 + layer.1.k_cache 0.54955656 9.72878339 + layer.1.v_cache 0.00000603 0.00422443 + layer.2.k_cache 0.01751511 0.93047938 + layer.2.v_cache 0.00002011 0.01187959 + layer.3.k_cache 0.01650532 3.57881962 + layer.3.v_cache 0.00001967 0.01268861 + layer.4.k_cache 0.00077751 0.26149166 + layer.4.v_cache 0.00005008 0.02384657 + layer.4.output 0.04306356 174.72199021 + ------------------------------------------------------------------------------------- + TOTAL 0.05868284 76.22815330 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 291624 +BPFP 0.8646 bits/point +EBPFP 1.7293 equivalent bits/point +MSE 76.228153 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 76.2282 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,652B, BPFP=0.5485 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,184B, BPFP=1.8767 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,412B, BPFP=0.8179 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,416B, BPFP=1.8217 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,976B, BPFP=0.9300 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,564B, BPFP=1.7606 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,952B, BPFP=0.8567 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,208B, BPFP=1.8068 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,132B, BPFP=1.4429 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,924B, BPFP=1.7147 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,328B, BPFP=0.3105 +⌛️ [2/4] FRONTEND: Frontend time: 0.244s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09832749 65.23037020 + layer.0.v_cache 0.00001658 0.01150660 + layer.1.k_cache 0.33493462 10.39371819 + layer.1.v_cache 0.00000656 0.00438505 + layer.2.k_cache 0.02762392 0.91949169 + layer.2.v_cache 0.00002012 0.01176184 + layer.3.k_cache 0.01647770 4.07547529 + layer.3.v_cache 0.00001951 0.01351934 + layer.4.k_cache 0.00067889 0.26544758 + layer.4.v_cache 0.00004874 0.02348983 + layer.4.output 1.40435175 248.25329702 + ------------------------------------------------------------------------------------- + TOTAL 0.60638920 106.98366146 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 219748 +BPFP 0.9265 bits/point +EBPFP 1.8530 equivalent bits/point +MSE 106.983661 +---------------------- -------------------------------------------------------- +Time: 0.567s Load: 0.007s, Pack+Encode: 0.244s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 106.9837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,936B, BPFP=0.5487 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,332B, BPFP=1.8205 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,832B, BPFP=0.8180 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,476B, BPFP=1.7613 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,188B, BPFP=0.9118 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,784B, BPFP=1.7135 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,184B, BPFP=0.8424 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,152B, BPFP=1.7389 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,424B, BPFP=1.4121 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,196B, BPFP=1.6728 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,116B, BPFP=0.2777 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102250 65.14571522 + layer.0.v_cache 0.00001721 0.01092750 + layer.1.k_cache 0.37371607 10.13323597 + layer.1.v_cache 0.00000658 0.00417332 + layer.2.k_cache 0.01127678 0.95869203 + layer.2.v_cache 0.00001998 0.01177466 + layer.3.k_cache 0.02136800 4.08801162 + layer.3.v_cache 0.00001963 0.01301695 + layer.4.k_cache 0.00069450 0.25923401 + layer.4.v_cache 0.00005147 0.02441440 + layer.4.output 1.35464589 239.51080515 + ------------------------------------------------------------------------------------- + TOTAL 0.58945376 103.36616657 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 219620 +BPFP 0.8932 bits/point +EBPFP 1.7863 equivalent bits/point +MSE 103.366167 +---------------------- -------------------------------------------------------- +Time: 0.547s Load: 0.009s, Pack+Encode: 0.221s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 103.3662 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,712B, BPFP=0.5423 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,028B, BPFP=1.7448 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,700B, BPFP=0.7906 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,284B, BPFP=1.6985 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,328B, BPFP=0.8919 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,424B, BPFP=1.6449 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,388B, BPFP=0.8334 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,856B, BPFP=1.6718 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,692B, BPFP=1.3503 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,500B, BPFP=1.5874 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,056B, BPFP=0.3118 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.339s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11871604 63.00575433 + layer.0.v_cache 0.00001755 0.01051958 + layer.1.k_cache 0.46315337 9.68904655 + layer.1.v_cache 0.00000588 0.00390922 + layer.2.k_cache 0.02202094 1.00364351 + layer.2.v_cache 0.00001989 0.01113761 + layer.3.k_cache 0.01278766 3.98173614 + layer.3.v_cache 0.00002092 0.01283658 + layer.4.k_cache 0.00074254 0.24944713 + layer.4.v_cache 0.00004925 0.02228236 + layer.4.output 1.21975157 215.01778600 + ------------------------------------------------------------------------------------- + TOTAL 0.53857618 93.12440088 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 239968 +BPFP 0.8787 bits/point +EBPFP 1.7574 equivalent bits/point +MSE 93.124401 +---------------------- -------------------------------------------------------- +Time: 0.576s Load: 0.009s, Pack+Encode: 0.228s, Decode+Unpack: 0.339s +---------------------- -------------------------------------------------------- +💾 Converting with 93.1244 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 195, 128) +Output shape: (1, 195, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.output: torch.Size([1, 195, 3584]) -> torch.Size([1, 1, 195, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,088B, BPFP=0.4878 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,272B, BPFP=1.9449 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,384B, BPFP=0.8321 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,484B, BPFP=1.8817 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,868B, BPFP=0.9510 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,640B, BPFP=1.8141 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,196B, BPFP=0.8971 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,092B, BPFP=1.8503 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,988B, BPFP=1.5215 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,172B, BPFP=1.7766 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,216B, BPFP=0.2658 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605874 63.93751002 + layer.0.v_cache 0.00001676 0.01133199 + layer.1.k_cache 0.13872500 10.04858524 + layer.1.v_cache 0.00000593 0.00413464 + layer.2.k_cache 0.01222654 1.01466299 + layer.2.v_cache 0.00001968 0.01152991 + layer.3.k_cache 0.01415822 4.14513597 + layer.3.v_cache 0.00002127 0.01428735 + layer.4.k_cache 0.00071127 0.26558042 + layer.4.v_cache 0.00005419 0.02503995 + layer.4.output 1.56987251 277.51634615 + ------------------------------------------------------------------------------------- + TOTAL 0.66300619 118.94660127 + (elements=1,697,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1697280 +Total Bytes 197400 +BPFP 0.9304 bits/point +EBPFP 1.8609 equivalent bits/point +MSE 118.946601 +---------------------- -------------------------------------------------------- +Time: 0.552s Load: 0.007s, Pack+Encode: 0.228s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 118.9466 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,312B, BPFP=0.5419 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,312B, BPFP=1.8297 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,512B, BPFP=0.8166 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,636B, BPFP=1.7716 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,052B, BPFP=0.9488 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,084B, BPFP=1.7242 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,324B, BPFP=0.8863 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,376B, BPFP=1.7493 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,688B, BPFP=1.4327 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,344B, BPFP=1.6607 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,052B, BPFP=0.2705 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.268s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11176265 57.03633671 + layer.0.v_cache 0.00001802 0.01166598 + layer.1.k_cache 0.18491925 10.13414052 + layer.1.v_cache 0.00000608 0.00452319 + layer.2.k_cache 0.01999094 0.96227591 + layer.2.v_cache 0.00002176 0.01250598 + layer.3.k_cache 0.03910406 4.12360894 + layer.3.v_cache 0.00002004 0.01374246 + layer.4.k_cache 0.00067358 0.26612519 + layer.4.v_cache 0.00005212 0.02444716 + layer.4.output 0.00886795 301.26111166 + ------------------------------------------------------------------------------------- + TOTAL 0.02462613 128.31865610 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 177692 +BPFP 0.8974 bits/point +EBPFP 1.7947 equivalent bits/point +MSE 128.318656 +---------------------- -------------------------------------------------------- +Time: 0.527s Load: 0.007s, Pack+Encode: 0.252s, Decode+Unpack: 0.268s +---------------------- -------------------------------------------------------- +💾 Converting with 128.3187 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,796B, BPFP=0.5719 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,996B, BPFP=1.9070 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,204B, BPFP=0.8219 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,120B, BPFP=1.8427 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,800B, BPFP=0.9390 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,356B, BPFP=1.7867 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,800B, BPFP=0.8656 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,808B, BPFP=1.8198 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,756B, BPFP=1.4492 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,512B, BPFP=1.7248 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,132B, BPFP=0.3053 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15089922 60.99617169 + layer.0.v_cache 0.00001686 0.01156573 + layer.1.k_cache 0.31217133 10.66466540 + layer.1.v_cache 0.00000593 0.00453045 + layer.2.k_cache 0.01305961 0.93878160 + layer.2.v_cache 0.00001999 0.01261318 + layer.3.k_cache 0.01178671 3.84781500 + layer.3.v_cache 0.00002039 0.01433543 + layer.4.k_cache 0.00067606 0.26612028 + layer.4.v_cache 0.00005285 0.02537903 + layer.4.output 1.43724915 254.04808015 + ------------------------------------------------------------------------------------- + TOTAL 0.62055606 109.12461993 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 216280 +BPFP 0.9333 bits/point +EBPFP 1.8665 equivalent bits/point +MSE 109.124620 +---------------------- -------------------------------------------------------- +Time: 0.571s Load: 0.007s, Pack+Encode: 0.245s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 109.1246 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 187, 128) +Output shape: (1, 187, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.output: torch.Size([1, 187, 3584]) -> torch.Size([1, 1, 187, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,392B, BPFP=0.5341 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,588B, BPFP=1.8038 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,656B, BPFP=0.8068 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,736B, BPFP=1.7326 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,428B, BPFP=0.9549 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,472B, BPFP=1.7106 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,564B, BPFP=0.8827 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,852B, BPFP=1.7423 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,928B, BPFP=1.4144 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,856B, BPFP=1.6591 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,856B, BPFP=0.2967 +⌛️ [2/4] FRONTEND: Frontend time: 0.195s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08258281 66.28681693 + layer.0.v_cache 0.00001733 0.01216901 + layer.1.k_cache 0.10049684 10.05787830 + layer.1.v_cache 0.00000647 0.00458384 + layer.2.k_cache 0.01472783 1.03951871 + layer.2.v_cache 0.00002140 0.01313160 + layer.3.k_cache 0.02125903 3.92033509 + layer.3.v_cache 0.00002180 0.01545269 + layer.4.k_cache 0.00070831 0.28421094 + layer.4.v_cache 0.00005280 0.02625480 + layer.4.output 0.00866398 292.78549943 + ------------------------------------------------------------------------------------- + TOTAL 0.01650250 125.36228517 + (elements=1,627,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1627648 +Total Bytes 183328 +BPFP 0.9011 bits/point +EBPFP 1.8021 equivalent bits/point +MSE 125.362285 +---------------------- -------------------------------------------------------- +Time: 0.474s Load: 0.006s, Pack+Encode: 0.195s, Decode+Unpack: 0.272s +---------------------- -------------------------------------------------------- +💾 Converting with 125.3623 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,576B, BPFP=0.5610 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,864B, BPFP=1.9153 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,176B, BPFP=0.8276 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,176B, BPFP=1.8643 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,700B, BPFP=0.9405 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,360B, BPFP=1.8039 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,904B, BPFP=0.8815 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,888B, BPFP=1.8430 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,820B, BPFP=1.4677 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,632B, BPFP=1.7500 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,736B, BPFP=0.2934 +⌛️ [2/4] FRONTEND: Frontend time: 0.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11071491 61.41452995 + layer.0.v_cache 0.00001804 0.01149975 + layer.1.k_cache 0.26784508 10.29641565 + layer.1.v_cache 0.00000610 0.00444419 + layer.2.k_cache 0.01369454 0.98824068 + layer.2.v_cache 0.00002102 0.01198486 + layer.3.k_cache 0.01444715 4.41598843 + layer.3.v_cache 0.00002041 0.01374891 + layer.4.k_cache 0.00067564 0.26402829 + layer.4.v_cache 0.00005248 0.02397485 + layer.4.output 1.45091435 256.36260156 + ------------------------------------------------------------------------------------- + TOTAL 0.62140564 110.11665097 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 214832 +BPFP 0.9358 bits/point +EBPFP 1.8716 equivalent bits/point +MSE 110.116651 +---------------------- -------------------------------------------------------- +Time: 0.605s Load: 0.007s, Pack+Encode: 0.231s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 110.1167 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,772B, BPFP=0.5446 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,420B, BPFP=1.8512 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,500B, BPFP=0.8058 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,476B, BPFP=1.7850 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,060B, BPFP=0.9151 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,532B, BPFP=1.7189 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,100B, BPFP=0.8478 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,116B, BPFP=1.7598 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,152B, BPFP=1.4120 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,848B, BPFP=1.6710 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,328B, BPFP=0.2535 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13389712 64.76089984 + layer.0.v_cache 0.00001748 0.01112734 + layer.1.k_cache 0.36644789 9.95894058 + layer.1.v_cache 0.00000587 0.00426449 + layer.2.k_cache 0.02260907 0.97561468 + layer.2.v_cache 0.00001979 0.01168282 + layer.3.k_cache 0.02550721 4.02609841 + layer.3.v_cache 0.00001968 0.01314114 + layer.4.k_cache 0.00067529 0.25715596 + layer.4.v_cache 0.00006705 0.02431476 + layer.4.output 1.37284074 242.31564302 + ------------------------------------------------------------------------------------- + TOTAL 0.59759716 104.48545536 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 215304 +BPFP 0.8874 bits/point +EBPFP 1.7748 equivalent bits/point +MSE 104.485455 +---------------------- -------------------------------------------------------- +Time: 0.562s Load: 0.008s, Pack+Encode: 0.228s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 104.4855 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 406, 128) +Output shape: (1, 406, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.output: torch.Size([1, 406, 3584]) -> torch.Size([1, 1, 406, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,220B, BPFP=0.5473 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,100B, BPFP=1.8127 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,716B, BPFP=0.7973 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,248B, BPFP=1.7414 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,188B, BPFP=0.8924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,272B, BPFP=1.7038 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,268B, BPFP=0.8185 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,608B, BPFP=1.7167 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,184B, BPFP=1.3541 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,408B, BPFP=1.6321 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 59,508B, BPFP=0.3272 +⌛️ [2/4] FRONTEND: Frontend time: 0.385s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.504s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10474718 59.63108548 + layer.0.v_cache 0.00001775 0.01120867 + layer.1.k_cache 0.89873824 10.07303894 + layer.1.v_cache 0.00000687 0.00442131 + layer.2.k_cache 0.02602503 0.91743770 + layer.2.v_cache 0.00002164 0.01221665 + layer.3.k_cache 0.01558245 3.72986231 + layer.3.v_cache 0.00002201 0.01329723 + layer.4.k_cache 0.00072958 0.25769780 + layer.4.v_cache 0.00005327 0.02376584 + layer.4.output 0.00651461 136.41471675 + ------------------------------------------------------------------------------------- + TOTAL 0.06420861 60.56335583 + (elements=3,533,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3533824 +Total Bytes 397720 +BPFP 0.9004 bits/point +EBPFP 1.8007 equivalent bits/point +MSE 60.563356 +---------------------- -------------------------------------------------------- +Time: 0.904s Load: 0.015s, Pack+Encode: 0.385s, Decode+Unpack: 0.504s +---------------------- -------------------------------------------------------- +💾 Converting with 60.5634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,484B, BPFP=0.5411 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,148B, BPFP=1.7952 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,624B, BPFP=0.8051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,404B, BPFP=1.7477 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,380B, BPFP=0.9171 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,632B, BPFP=1.6985 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,312B, BPFP=0.8490 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,912B, BPFP=1.7163 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,888B, BPFP=1.3959 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,764B, BPFP=1.6431 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,132B, BPFP=0.2927 +⌛️ [2/4] FRONTEND: Frontend time: 0.236s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.331s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14414167 58.83523597 + layer.0.v_cache 0.00001717 0.01117575 + layer.1.k_cache 0.44179790 9.98715521 + layer.1.v_cache 0.00000667 0.00445442 + layer.2.k_cache 0.03016641 0.99912633 + layer.2.v_cache 0.00002043 0.01203637 + layer.3.k_cache 0.03726396 3.65065046 + layer.3.v_cache 0.00002099 0.01381929 + layer.4.k_cache 0.00070653 0.25336825 + layer.4.v_cache 0.00004910 0.02335703 + layer.4.output 1.24963187 220.72416181 + ------------------------------------------------------------------------------------- + TOTAL 0.55303611 95.22703010 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 237680 +BPFP 0.8917 bits/point +EBPFP 1.7833 equivalent bits/point +MSE 95.227030 +---------------------- -------------------------------------------------------- +Time: 0.576s Load: 0.010s, Pack+Encode: 0.236s, Decode+Unpack: 0.331s +---------------------- -------------------------------------------------------- +💾 Converting with 95.2270 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,484B, BPFP=0.5091 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,192B, BPFP=1.9780 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,760B, BPFP=0.8448 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,848B, BPFP=1.8725 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,644B, BPFP=0.9928 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,244B, BPFP=1.8251 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,408B, BPFP=0.8957 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,800B, BPFP=1.8687 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,244B, BPFP=1.5110 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,396B, BPFP=1.7585 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,232B, BPFP=0.2494 +⌛️ [2/4] FRONTEND: Frontend time: 0.243s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.332s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15985105 62.36049878 + layer.0.v_cache 0.00001905 0.01243777 + layer.1.k_cache 0.22745673 9.93795393 + layer.1.v_cache 0.00000590 0.00461622 + layer.2.k_cache 0.01549998 0.93960556 + layer.2.v_cache 0.00002003 0.01248540 + layer.3.k_cache 0.01333464 4.26714996 + layer.3.v_cache 0.00002034 0.01466269 + layer.4.k_cache 0.00067076 0.27840726 + layer.4.v_cache 0.00004949 0.02540980 + layer.4.output 1.53832061 272.06427225 + ------------------------------------------------------------------------------------- + TOTAL 0.65795131 116.60606666 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 201252 +BPFP 0.9295 bits/point +EBPFP 1.8590 equivalent bits/point +MSE 116.606067 +---------------------- -------------------------------------------------------- +Time: 0.584s Load: 0.008s, Pack+Encode: 0.243s, Decode+Unpack: 0.332s +---------------------- -------------------------------------------------------- +💾 Converting with 116.6061 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 237, 128) +Output shape: (1, 237, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.output: torch.Size([1, 237, 3584]) -> torch.Size([1, 1, 237, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,464B, BPFP=0.5580 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,700B, BPFP=1.8262 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,280B, BPFP=0.8096 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,904B, BPFP=1.7737 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,192B, BPFP=0.9357 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,124B, BPFP=1.7223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,000B, BPFP=0.8571 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,360B, BPFP=1.7379 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,540B, BPFP=1.4201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,328B, BPFP=1.6698 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,632B, BPFP=0.2697 +⌛️ [2/4] FRONTEND: Frontend time: 0.314s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.419s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351868 63.38859771 + layer.0.v_cache 0.00001655 0.01125066 + layer.1.k_cache 0.34375901 10.23880971 + layer.1.v_cache 0.00000619 0.00431647 + layer.2.k_cache 0.01282014 0.97300984 + layer.2.v_cache 0.00002104 0.01216344 + layer.3.k_cache 0.03241006 4.16290154 + layer.3.v_cache 0.00001985 0.01320543 + layer.4.k_cache 0.00067667 0.25774583 + layer.4.v_cache 0.00005279 0.02426165 + layer.4.output 1.29178675 228.34529461 + ------------------------------------------------------------------------------------- + TOTAL 0.56210637 98.67666615 + (elements=2,062,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2062848 +Total Bytes 230524 +BPFP 0.8940 bits/point +EBPFP 1.7880 equivalent bits/point +MSE 98.676666 +---------------------- -------------------------------------------------------- +Time: 0.741s Load: 0.008s, Pack+Encode: 0.314s, Decode+Unpack: 0.419s +---------------------- -------------------------------------------------------- +💾 Converting with 98.6767 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,732B, BPFP=0.5619 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,028B, BPFP=1.8916 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,304B, BPFP=0.8215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,252B, BPFP=1.8352 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,800B, BPFP=0.9302 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,336B, BPFP=1.7686 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,728B, BPFP=0.8523 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,664B, BPFP=1.7924 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,864B, BPFP=1.4436 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,336B, BPFP=1.6959 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,068B, BPFP=0.2914 +⌛️ [2/4] FRONTEND: Frontend time: 0.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.428s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11023432 58.79214662 + layer.0.v_cache 0.00001677 0.01072171 + layer.1.k_cache 0.30291741 9.80432186 + layer.1.v_cache 0.00000597 0.00421726 + layer.2.k_cache 0.03464115 0.97273865 + layer.2.v_cache 0.00001998 0.01162953 + layer.3.k_cache 0.01330193 3.98993062 + layer.3.v_cache 0.00002112 0.01327066 + layer.4.k_cache 0.00069379 0.26130098 + layer.4.v_cache 0.00005021 0.02394227 + layer.4.output 1.42388847 251.55066445 + ------------------------------------------------------------------------------------- + TOTAL 0.61347776 107.92581596 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 215112 +BPFP 0.9196 bits/point +EBPFP 1.8392 equivalent bits/point +MSE 107.925816 +---------------------- -------------------------------------------------------- +Time: 0.760s Load: 0.008s, Pack+Encode: 0.324s, Decode+Unpack: 0.428s +---------------------- -------------------------------------------------------- +💾 Converting with 107.9258 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 231, 128) +Output shape: (1, 231, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.output: torch.Size([1, 231, 3584]) -> torch.Size([1, 1, 231, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,064B, BPFP=0.5455 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,428B, BPFP=1.8552 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,100B, BPFP=0.8185 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,540B, BPFP=1.7952 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,024B, BPFP=0.9486 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,924B, BPFP=1.7535 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,032B, BPFP=0.8815 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,284B, BPFP=1.7779 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,212B, BPFP=1.4348 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,112B, BPFP=1.6986 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,652B, BPFP=0.2865 +⌛️ [2/4] FRONTEND: Frontend time: 0.303s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.417s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13257201 62.98871246 + layer.0.v_cache 0.00001701 0.01154689 + layer.1.k_cache 0.40462213 10.09075901 + layer.1.v_cache 0.00000658 0.00451116 + layer.2.k_cache 0.02649644 0.92838654 + layer.2.v_cache 0.00002056 0.01185973 + layer.3.k_cache 0.06147716 3.75841123 + layer.3.v_cache 0.00002105 0.01423554 + layer.4.k_cache 0.00069437 0.26281887 + layer.4.v_cache 0.00005136 0.02423036 + layer.4.output 1.32536713 234.38429576 + ------------------------------------------------------------------------------------- + TOTAL 0.58256168 101.10503189 + (elements=2,010,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2010624 +Total Bytes 229372 +BPFP 0.9126 bits/point +EBPFP 1.8253 equivalent bits/point +MSE 101.105032 +---------------------- -------------------------------------------------------- +Time: 0.729s Load: 0.008s, Pack+Encode: 0.303s, Decode+Unpack: 0.417s +---------------------- -------------------------------------------------------- +💾 Converting with 101.1050 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,404B, BPFP=0.5150 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,540B, BPFP=1.7488 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,408B, BPFP=0.7603 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,636B, BPFP=1.6934 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,944B, BPFP=0.9157 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,172B, BPFP=1.6650 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,780B, BPFP=0.8444 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,500B, BPFP=1.6850 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,524B, BPFP=1.3801 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,288B, BPFP=1.6108 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,428B, BPFP=0.3451 +⌛️ [2/4] FRONTEND: Frontend time: 0.328s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.430s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12841698 57.37342984 + layer.0.v_cache 0.00001718 0.01186258 + layer.1.k_cache 0.45450200 9.41077953 + layer.1.v_cache 0.00000627 0.00460142 + layer.2.k_cache 0.01434461 0.93271269 + layer.2.v_cache 0.00002215 0.01291206 + layer.3.k_cache 0.03530993 3.83298819 + layer.3.v_cache 0.00002139 0.01418891 + layer.4.k_cache 0.00068344 0.26248869 + layer.4.v_cache 0.00005073 0.02542948 + layer.4.output 1.20064817 210.39662115 + ------------------------------------------------------------------------------------- + TOTAL 0.53164187 90.86222008 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 248624 +BPFP 0.8961 bits/point +EBPFP 1.7923 equivalent bits/point +MSE 90.862220 +---------------------- -------------------------------------------------------- +Time: 0.767s Load: 0.009s, Pack+Encode: 0.328s, Decode+Unpack: 0.430s +---------------------- -------------------------------------------------------- +💾 Converting with 90.8622 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,796B, BPFP=0.5428 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,004B, BPFP=1.8287 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,412B, BPFP=0.7985 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,120B, BPFP=1.7797 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,572B, BPFP=0.9182 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,116B, BPFP=1.7241 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,256B, BPFP=0.8453 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,516B, BPFP=1.7462 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,620B, BPFP=1.4195 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,196B, BPFP=1.6731 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,684B, BPFP=0.2904 +⌛️ [2/4] FRONTEND: Frontend time: 0.326s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.478s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13731835 62.20828485 + layer.0.v_cache 0.00001709 0.01091066 + layer.1.k_cache 0.55879569 9.83502067 + layer.1.v_cache 0.00000637 0.00422581 + layer.2.k_cache 0.01348169 0.96315251 + layer.2.v_cache 0.00002375 0.01206955 + layer.3.k_cache 0.01306405 4.09981369 + layer.3.v_cache 0.00002005 0.01310267 + layer.4.k_cache 0.00072533 0.25604908 + layer.4.v_cache 0.00005384 0.02406645 + layer.4.output 0.00472113 196.69582700 + ------------------------------------------------------------------------------------- + TOTAL 0.04450318 85.54691088 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 276292 +BPFP 0.9005 bits/point +EBPFP 1.8010 equivalent bits/point +MSE 85.546911 +---------------------- -------------------------------------------------------- +Time: 0.813s Load: 0.010s, Pack+Encode: 0.326s, Decode+Unpack: 0.478s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5469 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,368B, BPFP=0.5382 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,972B, BPFP=1.8154 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,424B, BPFP=0.8007 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,996B, BPFP=1.7647 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,008B, BPFP=0.9348 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,008B, BPFP=1.7135 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,660B, BPFP=0.8648 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,484B, BPFP=1.7382 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,180B, BPFP=1.4109 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,836B, BPFP=1.6526 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,868B, BPFP=0.2734 +⌛️ [2/4] FRONTEND: Frontend time: 0.348s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.492s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13972237 60.80115111 + layer.0.v_cache 0.00001689 0.01152339 + layer.1.k_cache 0.51719427 10.04816708 + layer.1.v_cache 0.00000613 0.00465302 + layer.2.k_cache 0.02019965 0.96999260 + layer.2.v_cache 0.00002053 0.01273976 + layer.3.k_cache 0.01759115 3.92174198 + layer.3.v_cache 0.00002125 0.01435499 + layer.4.k_cache 0.00069604 0.27128994 + layer.4.v_cache 0.00005872 0.02502288 + layer.4.output 0.04434862 180.07016789 + ------------------------------------------------------------------------------------- + TOTAL 0.05917455 78.62187129 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 291804 +BPFP 0.8910 bits/point +EBPFP 1.7821 equivalent bits/point +MSE 78.621871 +---------------------- -------------------------------------------------------- +Time: 0.851s Load: 0.011s, Pack+Encode: 0.348s, Decode+Unpack: 0.492s +---------------------- -------------------------------------------------------- +💾 Converting with 78.6219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,436B, BPFP=0.5053 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,052B, BPFP=1.9670 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,604B, BPFP=0.8326 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,052B, BPFP=1.8885 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,476B, BPFP=0.9796 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,288B, BPFP=1.8285 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,412B, BPFP=0.8960 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,908B, BPFP=1.8772 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,164B, BPFP=1.5047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,604B, BPFP=1.7748 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,424B, BPFP=0.2515 +⌛️ [2/4] FRONTEND: Frontend time: 0.317s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.433s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13044639 63.50109434 + layer.0.v_cache 0.00001737 0.01180808 + layer.1.k_cache 0.22565776 9.96716554 + layer.1.v_cache 0.00000620 0.00455964 + layer.2.k_cache 0.01998205 1.02525314 + layer.2.v_cache 0.00001989 0.01220132 + layer.3.k_cache 0.00629339 3.86170998 + layer.3.v_cache 0.00002050 0.01445100 + layer.4.k_cache 0.00068118 0.27632001 + layer.4.v_cache 0.00005692 0.02494005 + layer.4.output 1.53833376 272.06335248 + ------------------------------------------------------------------------------------- + TOTAL 0.65597165 116.65546885 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 201420 +BPFP 0.9303 bits/point +EBPFP 1.8606 equivalent bits/point +MSE 116.655469 +---------------------- -------------------------------------------------------- +Time: 0.757s Load: 0.007s, Pack+Encode: 0.317s, Decode+Unpack: 0.433s +---------------------- -------------------------------------------------------- +💾 Converting with 116.6555 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 246, 128) +Output shape: (1, 246, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.output: torch.Size([1, 246, 3584]) -> torch.Size([1, 1, 246, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,508B, BPFP=0.5404 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,224B, BPFP=1.7927 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,584B, BPFP=0.7993 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,324B, BPFP=1.7355 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,412B, BPFP=0.9154 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,560B, BPFP=1.6870 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,484B, BPFP=0.8565 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,960B, BPFP=1.7124 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,156B, BPFP=1.4073 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,616B, BPFP=1.6270 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,380B, BPFP=0.2666 +⌛️ [2/4] FRONTEND: Frontend time: 0.244s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.331s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131485 59.42596386 + layer.0.v_cache 0.00001613 0.01150093 + layer.1.k_cache 0.49701033 10.00722597 + layer.1.v_cache 0.00000620 0.00445920 + layer.2.k_cache 0.01298508 0.91347777 + layer.2.v_cache 0.00002120 0.01225870 + layer.3.k_cache 0.02762915 3.95190281 + layer.3.v_cache 0.00002136 0.01410907 + layer.4.k_cache 0.00068392 0.26071626 + layer.4.v_cache 0.00006372 0.02341864 + layer.4.output 1.24453647 219.91096835 + ------------------------------------------------------------------------------------- + TOTAL 0.55302984 94.94128304 + (elements=2,141,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2141184 +Total Bytes 235208 +BPFP 0.8788 bits/point +EBPFP 1.7576 equivalent bits/point +MSE 94.941283 +---------------------- -------------------------------------------------------- +Time: 0.584s Load: 0.009s, Pack+Encode: 0.244s, Decode+Unpack: 0.331s +---------------------- -------------------------------------------------------- +💾 Converting with 94.9413 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,584B, BPFP=0.5144 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,944B, BPFP=1.9487 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,684B, BPFP=0.8347 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,968B, BPFP=1.8725 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,572B, BPFP=0.9822 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,316B, BPFP=1.8216 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,552B, BPFP=0.9025 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,668B, BPFP=1.8491 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,524B, BPFP=1.5253 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,608B, BPFP=1.7663 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,996B, BPFP=0.2343 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.333s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11599884 62.66605469 + layer.0.v_cache 0.00001671 0.01101535 + layer.1.k_cache 0.22559120 9.79611328 + layer.1.v_cache 0.00000595 0.00449789 + layer.2.k_cache 0.01146694 1.01086929 + layer.2.v_cache 0.00002201 0.01279245 + layer.3.k_cache 0.03668072 4.15956299 + layer.3.v_cache 0.00002131 0.01398399 + layer.4.k_cache 0.00066762 0.26582705 + layer.4.v_cache 0.00005146 0.02546408 + layer.4.output 1.53061454 270.69919643 + ------------------------------------------------------------------------------------- + TOTAL 0.65322497 116.05062094 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 200416 +BPFP 0.9210 bits/point +EBPFP 1.8421 equivalent bits/point +MSE 116.050621 +---------------------- -------------------------------------------------------- +Time: 0.571s Load: 0.008s, Pack+Encode: 0.230s, Decode+Unpack: 0.333s +---------------------- -------------------------------------------------------- +💾 Converting with 116.0506 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,652B, BPFP=0.5221 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,236B, BPFP=1.8584 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,096B, BPFP=0.8253 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,272B, BPFP=1.7926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,908B, BPFP=0.9490 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,536B, BPFP=1.7424 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,672B, BPFP=0.8646 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,020B, BPFP=1.7754 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,136B, BPFP=1.4421 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,836B, BPFP=1.6946 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,208B, BPFP=0.2847 +⌛️ [2/4] FRONTEND: Frontend time: 0.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605523 65.24310436 + layer.0.v_cache 0.00001866 0.01165734 + layer.1.k_cache 0.31881074 10.05614595 + layer.1.v_cache 0.00000653 0.00454614 + layer.2.k_cache 0.02091764 0.95821857 + layer.2.v_cache 0.00002113 0.01269677 + layer.3.k_cache 0.03221333 4.02661879 + layer.3.v_cache 0.00002185 0.01408995 + layer.4.k_cache 0.00068277 0.27495863 + layer.4.v_cache 0.00005596 0.02472260 + layer.4.output 1.33695214 236.40584061 + ------------------------------------------------------------------------------------- + TOTAL 0.57926287 102.08633196 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 226572 +BPFP 0.9094 bits/point +EBPFP 1.8187 equivalent bits/point +MSE 102.086332 +---------------------- -------------------------------------------------------- +Time: 0.567s Load: 0.008s, Pack+Encode: 0.231s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0863 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 192, 128) +Output shape: (1, 192, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.output: torch.Size([1, 192, 3584]) -> torch.Size([1, 1, 192, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,976B, BPFP=0.4863 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,904B, BPFP=1.7012 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,904B, BPFP=0.7246 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,384B, BPFP=1.6589 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,604B, BPFP=0.8630 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,972B, BPFP=1.6253 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,740B, BPFP=0.7926 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,180B, BPFP=1.6423 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,640B, BPFP=1.3542 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,332B, BPFP=1.5732 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,244B, BPFP=0.3051 +⌛️ [2/4] FRONTEND: Frontend time: 0.201s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.270s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12444629 57.12383525 + layer.0.v_cache 0.00001540 0.01068884 + layer.1.k_cache 0.18411521 9.41827520 + layer.1.v_cache 0.00000585 0.00418103 + layer.2.k_cache 0.01124177 0.99425753 + layer.2.v_cache 0.00002179 0.01238192 + layer.3.k_cache 0.01616090 3.57931074 + layer.3.v_cache 0.00002117 0.01387477 + layer.4.k_cache 0.00068507 0.26575635 + layer.4.v_cache 0.00005275 0.02417144 + layer.4.output 0.00839530 281.91706194 + ------------------------------------------------------------------------------------- + TOTAL 0.02326666 120.28624510 + (elements=1,671,168) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1671168 +Total Bytes 178880 +BPFP 0.8563 bits/point +EBPFP 1.7126 equivalent bits/point +MSE 120.286245 +---------------------- -------------------------------------------------------- +Time: 0.478s Load: 0.008s, Pack+Encode: 0.201s, Decode+Unpack: 0.270s +---------------------- -------------------------------------------------------- +💾 Converting with 120.2862 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,740B, BPFP=0.5403 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,272B, BPFP=1.9081 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,016B, BPFP=0.8486 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,664B, BPFP=1.8509 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,532B, BPFP=0.9913 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,100B, BPFP=1.7978 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,592B, BPFP=0.9029 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,376B, BPFP=1.8238 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,180B, BPFP=1.5230 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,516B, BPFP=1.7428 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,360B, BPFP=0.2334 +⌛️ [2/4] FRONTEND: Frontend time: 0.201s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.270s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09999813 61.89360763 + layer.0.v_cache 0.00001899 0.01169932 + layer.1.k_cache 0.12989934 9.77046957 + layer.1.v_cache 0.00000610 0.00455583 + layer.2.k_cache 0.00966061 1.01022091 + layer.2.v_cache 0.00001994 0.01334767 + layer.3.k_cache 0.03384685 4.01919225 + layer.3.v_cache 0.00002040 0.01453216 + layer.4.k_cache 0.00066502 0.27850149 + layer.4.v_cache 0.00005200 0.02542796 + layer.4.output 0.00959846 330.65493223 + ------------------------------------------------------------------------------------- + TOTAL 0.02008098 140.68388708 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 165348 +BPFP 0.9155 bits/point +EBPFP 1.8310 equivalent bits/point +MSE 140.683887 +---------------------- -------------------------------------------------------- +Time: 0.476s Load: 0.006s, Pack+Encode: 0.201s, Decode+Unpack: 0.270s +---------------------- -------------------------------------------------------- +💾 Converting with 140.6839 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 428, 128) +Output shape: (1, 428, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.output: torch.Size([1, 428, 3584]) -> torch.Size([1, 1, 428, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,784B, BPFP=0.5397 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,768B, BPFP=1.7804 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,556B, BPFP=0.7869 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,048B, BPFP=1.7176 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,672B, BPFP=0.9007 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,004B, BPFP=1.6795 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,780B, BPFP=0.8316 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,564B, BPFP=1.6999 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,728B, BPFP=1.3773 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,488B, BPFP=1.6241 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 56,888B, BPFP=0.2967 +⌛️ [2/4] FRONTEND: Frontend time: 0.327s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.517s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13760449 60.98150007 + layer.0.v_cache 0.00001624 0.01064197 + layer.1.k_cache 0.89392254 10.01502035 + layer.1.v_cache 0.00000615 0.00409728 + layer.2.k_cache 0.03917003 0.93035889 + layer.2.v_cache 0.00002113 0.01184402 + layer.3.k_cache 0.03749864 3.92633855 + layer.3.v_cache 0.00002059 0.01265878 + layer.4.k_cache 0.00086919 0.26734885 + layer.4.v_cache 0.00005177 0.02374444 + layer.4.output 0.00617707 129.34601343 + ------------------------------------------------------------------------------------- + TOTAL 0.06778943 57.74150866 + (elements=3,725,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3725312 +Total Bytes 411280 +BPFP 0.8832 bits/point +EBPFP 1.7664 equivalent bits/point +MSE 57.741509 +---------------------- -------------------------------------------------------- +Time: 0.857s Load: 0.013s, Pack+Encode: 0.327s, Decode+Unpack: 0.517s +---------------------- -------------------------------------------------------- +💾 Converting with 57.7415 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,952B, BPFP=0.5514 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,064B, BPFP=1.8320 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,604B, BPFP=0.8092 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,028B, BPFP=1.7746 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,404B, BPFP=0.9089 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,252B, BPFP=1.7316 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,152B, BPFP=0.8395 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,592B, BPFP=1.7504 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,220B, BPFP=1.3974 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,148B, BPFP=1.6704 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,724B, BPFP=0.2986 +⌛️ [2/4] FRONTEND: Frontend time: 0.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385377 62.43798482 + layer.0.v_cache 0.00001705 0.01090877 + layer.1.k_cache 0.49372572 9.80854527 + layer.1.v_cache 0.00000626 0.00414599 + layer.2.k_cache 0.02422354 1.00585743 + layer.2.v_cache 0.00002046 0.01145766 + layer.3.k_cache 0.04153839 4.05944132 + layer.3.v_cache 0.00002075 0.01293259 + layer.4.k_cache 0.00074884 0.26301799 + layer.4.v_cache 0.00004954 0.02281749 + layer.4.output 0.00474379 196.72465489 + ------------------------------------------------------------------------------------- + TOTAL 0.04161240 85.57115844 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 277140 +BPFP 0.9033 bits/point +EBPFP 1.8066 equivalent bits/point +MSE 85.571158 +---------------------- -------------------------------------------------------- +Time: 0.666s Load: 0.013s, Pack+Encode: 0.267s, Decode+Unpack: 0.386s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,324B, BPFP=0.4994 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,108B, BPFP=1.8435 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,600B, BPFP=0.8030 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,648B, BPFP=1.7728 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,592B, BPFP=0.8994 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,508B, BPFP=1.7177 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,212B, BPFP=0.8326 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,012B, BPFP=1.7421 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,296B, BPFP=1.4172 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,648B, BPFP=1.6761 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,392B, BPFP=0.3068 +⌛️ [2/4] FRONTEND: Frontend time: 0.348s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.450s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12985682 61.70256869 + layer.0.v_cache 0.00001645 0.01065224 + layer.1.k_cache 0.61935935 10.07892847 + layer.1.v_cache 0.00000656 0.00417027 + layer.2.k_cache 0.01492825 0.91653509 + layer.2.v_cache 0.00002178 0.01116745 + layer.3.k_cache 0.02579401 3.75943910 + layer.3.v_cache 0.00002068 0.01219861 + layer.4.k_cache 0.00071613 0.26261356 + layer.4.v_cache 0.00005057 0.02288393 + layer.4.output 0.04141478 167.80426249 + ------------------------------------------------------------------------------------- + TOTAL 0.06356906 73.61241146 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 317340 +BPFP 0.9030 bits/point +EBPFP 1.8060 equivalent bits/point +MSE 73.612411 +---------------------- -------------------------------------------------------- +Time: 0.809s Load: 0.012s, Pack+Encode: 0.348s, Decode+Unpack: 0.450s +---------------------- -------------------------------------------------------- +💾 Converting with 73.6124 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,084B, BPFP=0.5509 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,528B, BPFP=1.8317 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,640B, BPFP=0.7998 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,400B, BPFP=1.7701 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,856B, BPFP=0.9209 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,552B, BPFP=1.7238 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,688B, BPFP=0.8571 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,052B, BPFP=1.7511 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,828B, BPFP=1.4111 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,560B, BPFP=1.6696 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,892B, BPFP=0.2801 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12359551 62.56350388 + layer.0.v_cache 0.00001766 0.01136432 + layer.1.k_cache 0.52692803 10.39111499 + layer.1.v_cache 0.00000637 0.00439857 + layer.2.k_cache 0.02102186 0.92554789 + layer.2.v_cache 0.00002187 0.01197240 + layer.3.k_cache 0.03868174 3.64829494 + layer.3.v_cache 0.00002097 0.01386933 + layer.4.k_cache 0.00070799 0.26554849 + layer.4.v_cache 0.00005204 0.02418317 + layer.4.output 0.00469152 193.88313249 + ------------------------------------------------------------------------------------- + TOTAL 0.04375851 84.41421914 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 279080 +BPFP 0.8969 bits/point +EBPFP 1.7938 equivalent bits/point +MSE 84.414219 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 84.4142 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,424B, BPFP=0.5044 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,988B, BPFP=1.9620 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,720B, BPFP=0.8417 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,008B, BPFP=1.8851 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,472B, BPFP=0.9793 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,232B, BPFP=1.8241 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,540B, BPFP=0.9061 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,776B, BPFP=1.8668 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,108B, BPFP=1.5003 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,444B, BPFP=1.7622 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,424B, BPFP=0.2627 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10671347 63.03553412 + layer.0.v_cache 0.00001788 0.01167043 + layer.1.k_cache 0.15257932 10.29604566 + layer.1.v_cache 0.00000642 0.00439069 + layer.2.k_cache 0.02197224 0.93312203 + layer.2.v_cache 0.00002094 0.01215831 + layer.3.k_cache 0.03325213 3.99941112 + layer.3.v_cache 0.00002126 0.01485855 + layer.4.k_cache 0.00067229 0.26898159 + layer.4.v_cache 0.00005023 0.02395209 + layer.4.output 1.53834953 272.06990309 + ------------------------------------------------------------------------------------- + TOTAL 0.65198546 116.65232036 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 202136 +BPFP 0.9336 bits/point +EBPFP 1.8672 equivalent bits/point +MSE 116.652320 +---------------------- -------------------------------------------------------- +Time: 0.559s Load: 0.007s, Pack+Encode: 0.227s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 116.6523 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,644B, BPFP=0.5381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,596B, BPFP=1.7801 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,696B, BPFP=0.7903 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,776B, BPFP=1.7291 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,988B, BPFP=0.9330 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,160B, BPFP=1.6907 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,816B, BPFP=0.8601 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,432B, BPFP=1.7077 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,372B, BPFP=1.3927 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,156B, BPFP=1.6282 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,468B, BPFP=0.2976 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15214830 63.63572273 + layer.0.v_cache 0.00001722 0.01189678 + layer.1.k_cache 0.49754258 10.09477325 + layer.1.v_cache 0.00000639 0.00472148 + layer.2.k_cache 0.01862443 0.98729715 + layer.2.v_cache 0.00002219 0.01301122 + layer.3.k_cache 0.01200796 3.70853111 + layer.3.v_cache 0.00002210 0.01469553 + layer.4.k_cache 0.00075226 0.26926595 + layer.4.v_cache 0.00005409 0.02579701 + layer.4.output 1.21979868 215.12487550 + ------------------------------------------------------------------------------------- + TOTAL 0.54234049 93.21410828 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 243104 +BPFP 0.8902 bits/point +EBPFP 1.7804 equivalent bits/point +MSE 93.214108 +---------------------- -------------------------------------------------------- +Time: 0.562s Load: 0.008s, Pack+Encode: 0.228s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 93.2141 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,640B, BPFP=0.5451 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,084B, BPFP=1.7973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,768B, BPFP=0.8078 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,164B, BPFP=1.7502 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,032B, BPFP=0.9238 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,308B, BPFP=1.7064 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,648B, BPFP=0.8529 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,768B, BPFP=1.7299 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,500B, BPFP=1.4088 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,232B, BPFP=1.6512 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,088B, BPFP=0.2714 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18432382 60.33116035 + layer.0.v_cache 0.00001629 0.01126098 + layer.1.k_cache 0.53789903 9.83532034 + layer.1.v_cache 0.00000607 0.00443243 + layer.2.k_cache 0.01514517 0.91259545 + layer.2.v_cache 0.00002151 0.01253576 + layer.3.k_cache 0.04182221 3.84003346 + layer.3.v_cache 0.00002017 0.01366962 + layer.4.k_cache 0.00068627 0.26518912 + layer.4.v_cache 0.00005063 0.02440311 + layer.4.output 0.04375917 177.58583138 + ------------------------------------------------------------------------------------- + TOTAL 0.06390032 77.55008355 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 294232 +BPFP 0.8867 bits/point +EBPFP 1.7733 equivalent bits/point +MSE 77.550084 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 77.5501 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 371, 128) +Output shape: (1, 371, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.output: torch.Size([1, 371, 3584]) -> torch.Size([1, 1, 371, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,648B, BPFP=0.5327 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,300B, BPFP=1.7815 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,648B, BPFP=0.7854 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,104B, BPFP=1.7311 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,588B, BPFP=0.9092 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,072B, BPFP=1.6877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,932B, BPFP=0.8395 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,456B, BPFP=1.7038 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,888B, BPFP=1.3851 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,584B, BPFP=1.6250 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 50,944B, BPFP=0.3065 +⌛️ [2/4] FRONTEND: Frontend time: 0.287s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.465s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16952868 59.09536093 + layer.0.v_cache 0.00001698 0.01141586 + layer.1.k_cache 0.82091874 10.12843771 + layer.1.v_cache 0.00000666 0.00452285 + layer.2.k_cache 0.02418445 0.97142979 + layer.2.v_cache 0.00002238 0.01273941 + layer.3.k_cache 0.03055777 4.06372366 + layer.3.v_cache 0.00002117 0.01377682 + layer.4.k_cache 0.00071103 0.25810649 + layer.4.v_cache 0.00005915 0.02513486 + layer.4.output 0.03610923 145.80365566 + ------------------------------------------------------------------------------------- + TOTAL 0.07639951 64.42413165 + (elements=3,229,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3229184 +Total Bytes 359164 +BPFP 0.8898 bits/point +EBPFP 1.7796 equivalent bits/point +MSE 64.424132 +---------------------- -------------------------------------------------------- +Time: 0.765s Load: 0.013s, Pack+Encode: 0.287s, Decode+Unpack: 0.465s +---------------------- -------------------------------------------------------- +💾 Converting with 64.4241 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,688B, BPFP=0.5510 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,408B, BPFP=1.8928 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,436B, BPFP=0.8197 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,572B, BPFP=1.8329 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,136B, BPFP=0.9415 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,796B, BPFP=1.7772 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,304B, BPFP=0.8819 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,308B, BPFP=1.8139 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,116B, BPFP=1.4418 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,996B, BPFP=1.7199 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,180B, BPFP=0.2885 +⌛️ [2/4] FRONTEND: Frontend time: 0.234s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13674593 63.99398832 + layer.0.v_cache 0.00001768 0.01204870 + layer.1.k_cache 0.26206151 10.13803577 + layer.1.v_cache 0.00000624 0.00464817 + layer.2.k_cache 0.02938290 1.05906131 + layer.2.v_cache 0.00002158 0.01267890 + layer.3.k_cache 0.05653211 3.85960066 + layer.3.v_cache 0.00002129 0.01447929 + layer.4.k_cache 0.00070776 0.27346053 + layer.4.v_cache 0.00005193 0.02564647 + layer.4.output 1.40434551 248.28649246 + ------------------------------------------------------------------------------------- + TOTAL 0.60682162 106.90582914 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 218940 +BPFP 0.9231 bits/point +EBPFP 1.8462 equivalent bits/point +MSE 106.905829 +---------------------- -------------------------------------------------------- +Time: 0.572s Load: 0.009s, Pack+Encode: 0.234s, Decode+Unpack: 0.330s +---------------------- -------------------------------------------------------- +💾 Converting with 106.9058 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 193, 128) +Output shape: (1, 193, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.output: torch.Size([1, 193, 3584]) -> torch.Size([1, 1, 193, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,052B, BPFP=0.4900 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,228B, BPFP=1.9615 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,232B, BPFP=0.8284 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,224B, BPFP=1.8802 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,248B, BPFP=0.9916 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,788B, BPFP=1.8449 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,132B, BPFP=0.9012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,124B, BPFP=1.8721 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,996B, BPFP=1.5379 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,164B, BPFP=1.7944 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,632B, BPFP=0.2618 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.323s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14305212 62.27161087 + layer.0.v_cache 0.00001770 0.01207732 + layer.1.k_cache 0.14179654 9.97682498 + layer.1.v_cache 0.00000608 0.00454849 + layer.2.k_cache 0.01555968 0.96195035 + layer.2.v_cache 0.00002047 0.01286832 + layer.3.k_cache 0.02655400 4.24975238 + layer.3.v_cache 0.00002031 0.01495075 + layer.4.k_cache 0.00069483 0.28070343 + layer.4.v_cache 0.00005004 0.02551790 + layer.4.output 1.58615237 280.41268042 + ------------------------------------------------------------------------------------- + TOTAL 0.67240225 120.04115104 + (elements=1,679,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1679872 +Total Bytes 196820 +BPFP 0.9373 bits/point +EBPFP 1.8746 equivalent bits/point +MSE 120.041151 +---------------------- -------------------------------------------------------- +Time: 0.555s Load: 0.007s, Pack+Encode: 0.225s, Decode+Unpack: 0.323s +---------------------- -------------------------------------------------------- +💾 Converting with 120.0412 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,488B, BPFP=0.5326 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,032B, BPFP=1.7590 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,388B, BPFP=0.7774 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,212B, BPFP=1.7076 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,748B, BPFP=0.9255 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,704B, BPFP=1.6757 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,548B, BPFP=0.8502 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,028B, BPFP=1.6960 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,748B, BPFP=1.3647 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,724B, BPFP=1.6142 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,008B, BPFP=0.3049 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10821181 59.76553087 + layer.0.v_cache 0.00001942 0.01065811 + layer.1.k_cache 0.30751975 10.10312735 + layer.1.v_cache 0.00000599 0.00385672 + layer.2.k_cache 0.01421910 0.98872474 + layer.2.v_cache 0.00002020 0.01125175 + layer.3.k_cache 0.01730646 3.66766235 + layer.3.v_cache 0.00002068 0.01319240 + layer.4.k_cache 0.00075739 0.25559601 + layer.4.v_cache 0.00005209 0.02341894 + layer.4.output 1.22958295 217.24584050 + ------------------------------------------------------------------------------------- + TOTAL 0.53265962 93.85670016 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 239628 +BPFP 0.8845 bits/point +EBPFP 1.7690 equivalent bits/point +MSE 93.856700 +---------------------- -------------------------------------------------------- +Time: 0.560s Load: 0.008s, Pack+Encode: 0.227s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 93.8567 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 202, 128) +Output shape: (1, 202, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.output: torch.Size([1, 202, 3584]) -> torch.Size([1, 1, 202, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,924B, BPFP=0.5356 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,664B, BPFP=1.9851 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,972B, BPFP=0.8487 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,664B, BPFP=1.9078 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,704B, BPFP=0.9827 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,832B, BPFP=1.8434 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,664B, BPFP=0.9022 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,436B, BPFP=1.8902 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,428B, BPFP=1.5028 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,136B, BPFP=1.7896 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,276B, BPFP=0.2793 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15282421 65.42066638 + layer.0.v_cache 0.00001801 0.01240217 + layer.1.k_cache 0.08597872 10.29619455 + layer.1.v_cache 0.00000649 0.00475468 + layer.2.k_cache 0.01425038 1.04141175 + layer.2.v_cache 0.00002047 0.01264969 + layer.3.k_cache 0.05704349 3.95395774 + layer.3.v_cache 0.00002256 0.01505656 + layer.4.k_cache 0.00069767 0.28646250 + layer.4.v_cache 0.00005056 0.02521768 + layer.4.output 1.51553867 267.98870668 + ------------------------------------------------------------------------------------- + TOTAL 0.64233431 115.11704238 + (elements=1,758,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1758208 +Total Bytes 208700 +BPFP 0.9496 bits/point +EBPFP 1.8992 equivalent bits/point +MSE 115.117042 +---------------------- -------------------------------------------------------- +Time: 0.559s Load: 0.007s, Pack+Encode: 0.226s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 115.1170 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,868B, BPFP=0.5448 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,636B, BPFP=1.8019 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,436B, BPFP=0.7970 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,432B, BPFP=1.7354 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,220B, BPFP=0.8955 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,580B, BPFP=1.6884 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,004B, BPFP=0.8284 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,908B, BPFP=1.7065 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,780B, BPFP=1.3682 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,532B, BPFP=1.6305 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,836B, BPFP=0.3142 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14837397 67.86203208 + layer.0.v_cache 0.00001603 0.01012168 + layer.1.k_cache 0.60911565 10.10150643 + layer.1.v_cache 0.00000597 0.00379205 + layer.2.k_cache 0.02477714 0.96367869 + layer.2.v_cache 0.00001938 0.01061308 + layer.3.k_cache 0.01052417 3.83597934 + layer.3.v_cache 0.00002039 0.01237687 + layer.4.k_cache 0.00073933 0.23942137 + layer.4.v_cache 0.00005963 0.02291912 + layer.4.output 0.00470031 195.76738390 + ------------------------------------------------------------------------------------- + TOTAL 0.04862081 85.49612518 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 275232 +BPFP 0.8939 bits/point +EBPFP 1.7878 equivalent bits/point +MSE 85.496125 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.009s, Pack+Encode: 0.254s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 85.4961 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 210, 128) +Output shape: (1, 210, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.output: torch.Size([1, 210, 3584]) -> torch.Size([1, 1, 210, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,704B, BPFP=0.5732 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,796B, BPFP=1.9193 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,140B, BPFP=0.8289 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,976B, BPFP=1.8583 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,924B, BPFP=0.9616 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,272B, BPFP=1.8060 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,888B, BPFP=0.8845 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,680B, BPFP=1.8363 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,916B, BPFP=1.4818 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,584B, BPFP=1.7548 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,072B, BPFP=0.2878 +⌛️ [2/4] FRONTEND: Frontend time: 0.225s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11415643 60.34877232 + layer.0.v_cache 0.00001882 0.01139433 + layer.1.k_cache 0.27145578 10.41122001 + layer.1.v_cache 0.00000623 0.00442261 + layer.2.k_cache 0.01201470 0.98683028 + layer.2.v_cache 0.00002208 0.01242009 + layer.3.k_cache 0.00887682 3.93262358 + layer.3.v_cache 0.00002165 0.01435721 + layer.4.k_cache 0.00072611 0.27314588 + layer.4.v_cache 0.00005233 0.02483763 + layer.4.output 1.45781997 257.64606718 + ------------------------------------------------------------------------------------- + TOTAL 0.62424063 110.56132319 + (elements=1,827,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1827840 +Total Bytes 213952 +BPFP 0.9364 bits/point +EBPFP 1.8728 equivalent bits/point +MSE 110.561323 +---------------------- -------------------------------------------------------- +Time: 0.558s Load: 0.008s, Pack+Encode: 0.225s, Decode+Unpack: 0.325s +---------------------- -------------------------------------------------------- +💾 Converting with 110.5613 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 165, 128) +Output shape: (1, 165, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.output: torch.Size([1, 165, 3584]) -> torch.Size([1, 1, 165, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,460B, BPFP=0.5170 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,468B, BPFP=1.9383 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,920B, BPFP=0.8447 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,852B, BPFP=1.8799 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,564B, BPFP=1.0004 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,168B, BPFP=1.8152 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,844B, BPFP=0.9322 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,716B, BPFP=1.8670 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,160B, BPFP=1.5303 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,744B, BPFP=1.7750 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,844B, BPFP=0.2549 +⌛️ [2/4] FRONTEND: Frontend time: 0.196s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.268s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10245823 63.42044271 + layer.0.v_cache 0.00002070 0.01247729 + layer.1.k_cache 0.13127365 10.31867898 + layer.1.v_cache 0.00000615 0.00461558 + layer.2.k_cache 0.01458293 1.07035485 + layer.2.v_cache 0.00002074 0.01306136 + layer.3.k_cache 0.02375244 4.23196615 + layer.3.v_cache 0.00002214 0.01560372 + layer.4.k_cache 0.00067489 0.28717822 + layer.4.v_cache 0.00005196 0.02652869 + layer.4.output 0.00973386 332.66217532 + ------------------------------------------------------------------------------------- + TOTAL 0.02005887 141.64918440 + (elements=1,436,160) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1436160 +Total Bytes 167740 +BPFP 0.9344 bits/point +EBPFP 1.8688 equivalent bits/point +MSE 141.649184 +---------------------- -------------------------------------------------------- +Time: 0.470s Load: 0.005s, Pack+Encode: 0.196s, Decode+Unpack: 0.268s +---------------------- -------------------------------------------------------- +💾 Converting with 141.6492 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,348B, BPFP=0.5420 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,544B, BPFP=1.8395 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,676B, BPFP=0.8262 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,000B, BPFP=1.7930 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,528B, BPFP=0.9843 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,404B, BPFP=1.7421 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,604B, BPFP=0.9054 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,756B, BPFP=1.7722 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,060B, BPFP=1.4566 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,784B, BPFP=1.6892 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,816B, BPFP=0.2661 +⌛️ [2/4] FRONTEND: Frontend time: 0.195s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.271s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12273688 60.23276341 + layer.0.v_cache 0.00001959 0.01245964 + layer.1.k_cache 0.11126613 10.24792547 + layer.1.v_cache 0.00000674 0.00507703 + layer.2.k_cache 0.01416154 0.94019414 + layer.2.v_cache 0.00002164 0.01344357 + layer.3.k_cache 0.04491856 3.91303324 + layer.3.v_cache 0.00002114 0.01553740 + layer.4.k_cache 0.00068813 0.29359686 + layer.4.v_cache 0.00006151 0.02690108 + layer.4.output 0.00884247 299.67471702 + ------------------------------------------------------------------------------------- + TOTAL 0.02092936 127.84846770 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 180520 +BPFP 0.9067 bits/point +EBPFP 1.8133 equivalent bits/point +MSE 127.848468 +---------------------- -------------------------------------------------------- +Time: 0.474s Load: 0.008s, Pack+Encode: 0.195s, Decode+Unpack: 0.271s +---------------------- -------------------------------------------------------- +💾 Converting with 127.8485 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 257, 128) +Output shape: (1, 257, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.output: torch.Size([1, 257, 3584]) -> torch.Size([1, 1, 257, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,152B, BPFP=0.4956 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,068B, BPFP=1.8889 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,136B, BPFP=0.7986 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,816B, BPFP=1.8127 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,336B, BPFP=0.9324 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,076B, BPFP=1.7678 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,008B, BPFP=0.8517 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,516B, BPFP=1.7945 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,744B, BPFP=1.4436 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,104B, BPFP=1.7087 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,384B, BPFP=0.2900 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12694686 61.55157162 + layer.0.v_cache 0.00002055 0.01127865 + layer.1.k_cache 0.54844232 10.08222124 + layer.1.v_cache 0.00000676 0.00427488 + layer.2.k_cache 0.01835482 0.93449705 + layer.2.v_cache 0.00002099 0.01179712 + layer.3.k_cache 0.02141366 3.83324720 + layer.3.v_cache 0.00002047 0.01339344 + layer.4.k_cache 0.00070767 0.26185750 + layer.4.v_cache 0.00005046 0.02315710 + layer.4.output 0.00513101 215.83301487 + ------------------------------------------------------------------------------------- + TOTAL 0.04422951 93.38578823 + (elements=2,236,928) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2236928 +Total Bytes 255340 +BPFP 0.9132 bits/point +EBPFP 1.8264 equivalent bits/point +MSE 93.385788 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.008s, Pack+Encode: 0.254s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 93.3858 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,340B, BPFP=0.5443 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,352B, BPFP=1.8331 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,544B, BPFP=0.8194 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,612B, BPFP=1.7696 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,928B, BPFP=0.9382 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,008B, BPFP=1.7177 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,228B, BPFP=0.8781 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,492B, BPFP=1.7593 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,724B, BPFP=1.4358 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,408B, BPFP=1.6662 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,100B, BPFP=0.2588 +⌛️ [2/4] FRONTEND: Frontend time: 0.200s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.268s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12731336 61.35176640 + layer.0.v_cache 0.00001750 0.01206628 + layer.1.k_cache 0.10052182 10.11505194 + layer.1.v_cache 0.00000572 0.00423344 + layer.2.k_cache 0.01141533 0.95095138 + layer.2.v_cache 0.00001926 0.01186604 + layer.3.k_cache 0.03745206 3.95259983 + layer.3.v_cache 0.00002083 0.01442639 + layer.4.k_cache 0.00066429 0.25999002 + layer.4.v_cache 0.00005101 0.02440483 + layer.4.output 0.00884527 301.30519035 + ------------------------------------------------------------------------------------- + TOTAL 0.01996459 128.57845229 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 176736 +BPFP 0.8925 bits/point +EBPFP 1.7851 equivalent bits/point +MSE 128.578452 +---------------------- -------------------------------------------------------- +Time: 0.475s Load: 0.006s, Pack+Encode: 0.200s, Decode+Unpack: 0.268s +---------------------- -------------------------------------------------------- +💾 Converting with 128.5785 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,868B, BPFP=0.5538 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,684B, BPFP=1.8781 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,488B, BPFP=0.8086 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,608B, BPFP=1.8024 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,292B, BPFP=0.9355 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,876B, BPFP=1.7508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,364B, BPFP=0.8702 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,240B, BPFP=1.7765 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,256B, BPFP=1.4257 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,008B, BPFP=1.6898 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,832B, BPFP=0.2899 +⌛️ [2/4] FRONTEND: Frontend time: 0.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12752426 62.86138971 + layer.0.v_cache 0.00002004 0.01182779 + layer.1.k_cache 0.30660921 10.15758178 + layer.1.v_cache 0.00000613 0.00418671 + layer.2.k_cache 0.01808885 0.93171458 + layer.2.v_cache 0.00002107 0.01179234 + layer.3.k_cache 0.01678786 4.04109096 + layer.3.v_cache 0.00002050 0.01348418 + layer.4.k_cache 0.00068290 0.26467246 + layer.4.v_cache 0.00004851 0.02273133 + layer.4.output 1.37905047 243.72562741 + ------------------------------------------------------------------------------------- + TOTAL 0.59548016 104.96469787 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 220516 +BPFP 0.9130 bits/point +EBPFP 1.8259 equivalent bits/point +MSE 104.964698 +---------------------- -------------------------------------------------------- +Time: 0.543s Load: 0.007s, Pack+Encode: 0.220s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 104.9647 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,000B, BPFP=0.5760 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,248B, BPFP=1.8900 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,536B, BPFP=0.8306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,404B, BPFP=1.8292 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,020B, BPFP=0.9375 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,532B, BPFP=1.7664 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,152B, BPFP=0.8750 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,080B, BPFP=1.8059 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,116B, BPFP=1.4484 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,728B, BPFP=1.7085 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,200B, BPFP=0.3106 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12766143 65.29399482 + layer.0.v_cache 0.00001653 0.01125150 + layer.1.k_cache 0.34631467 10.03980280 + layer.1.v_cache 0.00000659 0.00439762 + layer.2.k_cache 0.01786991 1.00375922 + layer.2.v_cache 0.00002041 0.01159278 + layer.3.k_cache 0.06874437 4.29707948 + layer.3.v_cache 0.00002276 0.01466635 + layer.4.k_cache 0.00067882 0.26658484 + layer.4.v_cache 0.00005621 0.02497305 + layer.4.output 1.41085444 249.43729427 + ------------------------------------------------------------------------------------- + TOTAL 0.61396310 107.47230367 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 220016 +BPFP 0.9319 bits/point +EBPFP 1.8638 equivalent bits/point +MSE 107.472304 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.007s, Pack+Encode: 0.224s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 107.4723 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,848B, BPFP=0.5426 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,908B, BPFP=1.8603 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,900B, BPFP=0.8227 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,036B, BPFP=1.8001 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,600B, BPFP=0.9403 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,328B, BPFP=1.7511 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,532B, BPFP=0.8664 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,856B, BPFP=1.7876 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,580B, BPFP=1.4228 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,496B, BPFP=1.6936 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,164B, BPFP=0.2880 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12070838 62.78646986 + layer.0.v_cache 0.00001763 0.01176089 + layer.1.k_cache 0.31015487 10.10601321 + layer.1.v_cache 0.00000677 0.00451795 + layer.2.k_cache 0.03544329 0.98560995 + layer.2.v_cache 0.00002212 0.01255630 + layer.3.k_cache 0.05682086 3.88431488 + layer.3.v_cache 0.00002136 0.01471534 + layer.4.k_cache 0.00073636 0.27823436 + layer.4.v_cache 0.00005031 0.02466733 + layer.4.output 1.35470685 239.47515013 + ------------------------------------------------------------------------------------- + TOTAL 0.58864294 103.20205359 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 224248 +BPFP 0.9120 bits/point +EBPFP 1.8240 equivalent bits/point +MSE 103.202054 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.007s, Pack+Encode: 0.222s, Decode+Unpack: 0.320s +---------------------- -------------------------------------------------------- +💾 Converting with 103.2021 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,732B, BPFP=0.5672 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,156B, BPFP=1.9187 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,292B, BPFP=0.8283 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,452B, BPFP=1.8671 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,220B, BPFP=0.9698 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,588B, BPFP=1.8037 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,300B, BPFP=0.9023 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,904B, BPFP=1.8269 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,060B, BPFP=1.4715 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,808B, BPFP=1.7465 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,020B, BPFP=0.3251 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.322s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13227103 60.49870250 + layer.0.v_cache 0.00002050 0.01162528 + layer.1.k_cache 0.22864332 10.55134747 + layer.1.v_cache 0.00000662 0.00450886 + layer.2.k_cache 0.02950178 0.99597612 + layer.2.v_cache 0.00002095 0.01202350 + layer.3.k_cache 0.01835468 3.69809519 + layer.3.v_cache 0.00002123 0.01474123 + layer.4.k_cache 0.00070999 0.27578685 + layer.4.v_cache 0.00005457 0.02453806 + layer.4.output 1.43734575 254.05149648 + ------------------------------------------------------------------------------------- + TOTAL 0.61594264 109.08516591 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 220532 +BPFP 0.9516 bits/point +EBPFP 1.9032 equivalent bits/point +MSE 109.085166 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.007s, Pack+Encode: 0.224s, Decode+Unpack: 0.322s +---------------------- -------------------------------------------------------- +💾 Converting with 109.0852 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,952B, BPFP=0.5597 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,356B, BPFP=1.8550 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,608B, BPFP=0.8170 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,532B, BPFP=1.7970 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,508B, BPFP=0.9507 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,900B, BPFP=1.7525 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,472B, BPFP=0.8778 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,200B, BPFP=1.7736 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,528B, BPFP=1.4448 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,052B, BPFP=1.6928 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,328B, BPFP=0.2848 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12518616 61.62659241 + layer.0.v_cache 0.00001676 0.01164922 + layer.1.k_cache 0.33007166 10.04419055 + layer.1.v_cache 0.00000630 0.00455598 + layer.2.k_cache 0.01131002 0.98587174 + layer.2.v_cache 0.00002125 0.01255167 + layer.3.k_cache 0.02232737 4.40473952 + layer.3.v_cache 0.00001983 0.01351005 + layer.4.k_cache 0.00071526 0.27499641 + layer.4.v_cache 0.00005006 0.02381005 + layer.4.output 1.37904434 243.72751770 + ------------------------------------------------------------------------------------- + TOTAL 0.59664912 104.91147597 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 220436 +BPFP 0.9126 bits/point +EBPFP 1.8253 equivalent bits/point +MSE 104.911476 +---------------------- -------------------------------------------------------- +Time: 0.547s Load: 0.008s, Pack+Encode: 0.224s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 104.9115 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,496B, BPFP=0.5441 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,908B, BPFP=1.7871 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,688B, BPFP=0.8125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,956B, BPFP=1.7262 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,220B, BPFP=0.9106 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,264B, BPFP=1.6819 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,168B, BPFP=0.8432 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,648B, BPFP=1.7065 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,632B, BPFP=1.3852 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,528B, BPFP=1.6347 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,912B, BPFP=0.3102 +⌛️ [2/4] FRONTEND: Frontend time: 0.224s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14655781 60.51138656 + layer.0.v_cache 0.00001709 0.01084554 + layer.1.k_cache 0.36683986 9.99815294 + layer.1.v_cache 0.00000651 0.00439066 + layer.2.k_cache 0.01755252 0.99706206 + layer.2.v_cache 0.00002060 0.01166701 + layer.3.k_cache 0.02713263 4.05942173 + layer.3.v_cache 0.00001997 0.01290699 + layer.4.k_cache 0.00071102 0.26084112 + layer.4.v_cache 0.00005162 0.02288749 + layer.4.output 1.25479176 221.39646150 + ------------------------------------------------------------------------------------- + TOTAL 0.54955600 95.62734074 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 237420 +BPFP 0.8943 bits/point +EBPFP 1.7887 equivalent bits/point +MSE 95.627341 +---------------------- -------------------------------------------------------- +Time: 0.552s Load: 0.008s, Pack+Encode: 0.224s, Decode+Unpack: 0.320s +---------------------- -------------------------------------------------------- +💾 Converting with 95.6273 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 155, 128) +Output shape: (1, 155, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.output: torch.Size([1, 155, 3584]) -> torch.Size([1, 1, 155, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,704B, BPFP=0.5750 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,452B, BPFP=1.9609 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,660B, BPFP=0.8730 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,780B, BPFP=1.8931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,032B, BPFP=1.0113 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,272B, BPFP=1.8419 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,316B, BPFP=0.9391 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,784B, BPFP=1.8935 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,384B, BPFP=1.5508 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,768B, BPFP=1.7911 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,580B, BPFP=0.2676 +⌛️ [2/4] FRONTEND: Frontend time: 0.194s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.258s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102877 70.17321699 + layer.0.v_cache 0.00001822 0.01206489 + layer.1.k_cache 0.12773433 10.26688193 + layer.1.v_cache 0.00000585 0.00458202 + layer.2.k_cache 0.00931232 1.07052041 + layer.2.v_cache 0.00002110 0.01339795 + layer.3.k_cache 0.01537714 4.17554538 + layer.3.v_cache 0.00002091 0.01595516 + layer.4.k_cache 0.00066471 0.29744824 + layer.4.v_cache 0.00005180 0.02753415 + layer.4.output 0.01742802 357.00051843 + ------------------------------------------------------------------------------------- + TOTAL 0.02389596 152.06239860 + (elements=1,349,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1349120 +Total Bytes 160732 +BPFP 0.9531 bits/point +EBPFP 1.9062 equivalent bits/point +MSE 152.062399 +---------------------- -------------------------------------------------------- +Time: 0.458s Load: 0.005s, Pack+Encode: 0.194s, Decode+Unpack: 0.258s +---------------------- -------------------------------------------------------- +💾 Converting with 152.0624 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,720B, BPFP=0.5610 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,340B, BPFP=1.9142 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,288B, BPFP=0.8203 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,316B, BPFP=1.8398 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,048B, BPFP=0.9483 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,568B, BPFP=1.7855 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,112B, BPFP=0.8802 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,016B, BPFP=1.8180 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,992B, BPFP=1.4529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,796B, BPFP=1.7294 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,536B, BPFP=0.2963 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11659910 61.99036610 + layer.0.v_cache 0.00001673 0.01177502 + layer.1.k_cache 0.29571861 10.44495253 + layer.1.v_cache 0.00000615 0.00441677 + layer.2.k_cache 0.04707494 1.00191296 + layer.2.v_cache 0.00002002 0.01193076 + layer.3.k_cache 0.03274950 4.14908930 + layer.3.v_cache 0.00002078 0.01369909 + layer.4.k_cache 0.00068608 0.26680617 + layer.4.v_cache 0.00005136 0.02430797 + layer.4.output 1.42391751 251.48963870 + ------------------------------------------------------------------------------------- + TOTAL 0.61531564 108.13804280 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 217732 +BPFP 0.9308 bits/point +EBPFP 1.8616 equivalent bits/point +MSE 108.138043 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.009s, Pack+Encode: 0.223s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 108.1380 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,640B, BPFP=0.5444 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,700B, BPFP=1.8082 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,812B, BPFP=0.8072 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,852B, BPFP=1.7548 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,032B, BPFP=0.9471 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,052B, BPFP=1.7044 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,984B, BPFP=0.8810 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,608B, BPFP=1.7394 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,368B, BPFP=1.4093 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,144B, BPFP=1.6472 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,236B, BPFP=0.3171 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10631144 57.90278478 + layer.0.v_cache 0.00001780 0.01170595 + layer.1.k_cache 0.46192560 9.90842955 + layer.1.v_cache 0.00000619 0.00457168 + layer.2.k_cache 0.02890189 0.98829368 + layer.2.v_cache 0.00002115 0.01230265 + layer.3.k_cache 0.04994748 3.84875341 + layer.3.v_cache 0.00002140 0.01462351 + layer.4.k_cache 0.00068365 0.26384514 + layer.4.v_cache 0.00005040 0.02430305 + layer.4.output 1.23456514 218.05164531 + ------------------------------------------------------------------------------------- + TOTAL 0.54646135 94.07889003 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 245428 +BPFP 0.9096 bits/point +EBPFP 1.8192 equivalent bits/point +MSE 94.078890 +---------------------- -------------------------------------------------------- +Time: 0.548s Load: 0.008s, Pack+Encode: 0.222s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 94.0789 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 194, 128) +Output shape: (1, 194, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.output: torch.Size([1, 194, 3584]) -> torch.Size([1, 1, 194, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,216B, BPFP=0.5006 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,400B, BPFP=1.9652 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,208B, BPFP=0.8222 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,372B, BPFP=1.8824 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,396B, BPFP=0.9984 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,896B, BPFP=1.8441 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,324B, BPFP=0.9120 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,352B, BPFP=1.8808 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,124B, BPFP=1.5403 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,140B, BPFP=1.7832 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,848B, BPFP=0.2399 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15027005 64.90528351 + layer.0.v_cache 0.00001782 0.01232289 + layer.1.k_cache 0.21402168 9.97188419 + layer.1.v_cache 0.00000588 0.00457094 + layer.2.k_cache 0.02177384 0.93679196 + layer.2.v_cache 0.00002019 0.01331088 + layer.3.k_cache 0.04024272 4.52854778 + layer.3.v_cache 0.00002076 0.01512372 + layer.4.k_cache 0.00069591 0.28624898 + layer.4.v_cache 0.00005073 0.02636897 + layer.4.output 1.57794378 278.92974503 + ------------------------------------------------------------------------------------- + TOTAL 0.67486624 119.60050994 + (elements=1,688,576) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1688576 +Total Bytes 196276 +BPFP 0.9299 bits/point +EBPFP 1.8598 equivalent bits/point +MSE 119.600510 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.007s, Pack+Encode: 0.228s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 119.6005 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 258, 128) +Output shape: (1, 258, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.output: torch.Size([1, 258, 3584]) -> torch.Size([1, 1, 258, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,220B, BPFP=0.4978 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,096B, BPFP=1.8832 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,328B, BPFP=0.8072 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,860B, BPFP=1.8084 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,408B, BPFP=0.9331 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,184B, BPFP=1.7674 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,156B, BPFP=0.8573 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,636B, BPFP=1.7948 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,012B, BPFP=1.4542 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,484B, BPFP=1.7250 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,984B, BPFP=0.2681 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09035230 63.17962694 + layer.0.v_cache 0.00001726 0.01133942 + layer.1.k_cache 0.42455262 9.95654581 + layer.1.v_cache 0.00000600 0.00417401 + layer.2.k_cache 0.01891134 0.90681806 + layer.2.v_cache 0.00001996 0.01168720 + layer.3.k_cache 0.02093011 3.97861735 + layer.3.v_cache 0.00001928 0.01323914 + layer.4.k_cache 0.00072082 0.25990828 + layer.4.v_cache 0.00004951 0.02431512 + layer.4.output 0.00505962 214.95753738 + ------------------------------------------------------------------------------------- + TOTAL 0.03476450 93.12053135 + (elements=2,245,632) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2245632 +Total Bytes 254368 +BPFP 0.9062 bits/point +EBPFP 1.8124 equivalent bits/point +MSE 93.120531 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.008s, Pack+Encode: 0.248s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 93.1205 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,844B, BPFP=0.5329 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,324B, BPFP=1.8562 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,360B, BPFP=0.8397 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,428B, BPFP=1.7954 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,900B, BPFP=0.9443 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,536B, BPFP=1.7348 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,912B, BPFP=0.8772 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,952B, BPFP=1.7630 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,836B, BPFP=1.4155 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,708B, BPFP=1.6785 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,408B, BPFP=0.2854 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15008528 62.30619056 + layer.0.v_cache 0.00001847 0.01141763 + layer.1.k_cache 0.25612781 10.22240680 + layer.1.v_cache 0.00000678 0.00430538 + layer.2.k_cache 0.03931687 0.92864931 + layer.2.v_cache 0.00002107 0.01186654 + layer.3.k_cache 0.03786138 3.98352130 + layer.3.v_cache 0.00002177 0.01419560 + layer.4.k_cache 0.00069674 0.26517115 + layer.4.v_cache 0.00005115 0.02339686 + layer.4.output 1.33117570 235.38255047 + ------------------------------------------------------------------------------------- + TOTAL 0.57661395 101.49699849 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 227208 +BPFP 0.9080 bits/point +EBPFP 1.8159 equivalent bits/point +MSE 101.496998 +---------------------- -------------------------------------------------------- +Time: 0.547s Load: 0.008s, Pack+Encode: 0.223s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 101.4970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,228B, BPFP=0.5406 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,356B, BPFP=1.8538 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,552B, BPFP=0.8292 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,728B, BPFP=1.7993 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,152B, BPFP=0.9681 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,188B, BPFP=1.7524 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,412B, BPFP=0.9038 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,440B, BPFP=1.7743 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,752B, BPFP=1.4542 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,516B, BPFP=1.6941 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,668B, BPFP=0.3059 +⌛️ [2/4] FRONTEND: Frontend time: 0.191s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.261s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15161699 60.18126085 + layer.0.v_cache 0.00001786 0.01198531 + layer.1.k_cache 0.08662784 10.10196737 + layer.1.v_cache 0.00000652 0.00453631 + layer.2.k_cache 0.02063684 0.96329134 + layer.2.v_cache 0.00002327 0.01274065 + layer.3.k_cache 0.08087930 4.36071710 + layer.3.v_cache 0.00002116 0.01474398 + layer.4.k_cache 0.00074161 0.27716859 + layer.4.v_cache 0.00006096 0.02582561 + layer.4.output 0.00898438 304.16250000 + ------------------------------------------------------------------------------------- + TOTAL 0.02373665 129.71127865 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 180992 +BPFP 0.9242 bits/point +EBPFP 1.8484 equivalent bits/point +MSE 129.711279 +---------------------- -------------------------------------------------------- +Time: 0.459s Load: 0.006s, Pack+Encode: 0.191s, Decode+Unpack: 0.261s +---------------------- -------------------------------------------------------- +💾 Converting with 129.7113 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,664B, BPFP=0.5443 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,576B, BPFP=1.8875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,672B, BPFP=0.8290 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,688B, BPFP=1.8244 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,592B, BPFP=0.9653 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,944B, BPFP=1.7716 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,432B, BPFP=0.8830 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,388B, BPFP=1.8031 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,540B, BPFP=1.4588 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,192B, BPFP=1.7182 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,056B, BPFP=0.2847 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14327769 62.60241477 + layer.0.v_cache 0.00001763 0.01200189 + layer.1.k_cache 0.34925704 10.30521684 + layer.1.v_cache 0.00000642 0.00468857 + layer.2.k_cache 0.02352169 0.99870619 + layer.2.v_cache 0.00002271 0.01291183 + layer.3.k_cache 0.00845094 3.88922729 + layer.3.v_cache 0.00002026 0.01456376 + layer.4.k_cache 0.00069169 0.27281752 + layer.4.v_cache 0.00005086 0.02588109 + layer.4.output 1.39158238 245.73985390 + ------------------------------------------------------------------------------------- + TOTAL 0.60390550 105.78337688 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 220744 +BPFP 0.9222 bits/point +EBPFP 1.8445 equivalent bits/point +MSE 105.783377 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.007s, Pack+Encode: 0.223s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 105.7834 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,572B, BPFP=0.5439 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,800B, BPFP=1.8636 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,164B, BPFP=0.8048 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,752B, BPFP=1.8041 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,596B, BPFP=0.9430 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,028B, BPFP=1.7630 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,272B, BPFP=0.8677 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,376B, BPFP=1.7827 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,208B, BPFP=1.4323 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,780B, BPFP=1.6920 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,688B, BPFP=0.3059 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12716211 61.10450994 + layer.0.v_cache 0.00001677 0.01106705 + layer.1.k_cache 0.51313377 10.09213690 + layer.1.v_cache 0.00000586 0.00422076 + layer.2.k_cache 0.02995576 0.97626121 + layer.2.v_cache 0.00002133 0.01234027 + layer.3.k_cache 0.03339291 4.13066806 + layer.3.v_cache 0.00002012 0.01350002 + layer.4.k_cache 0.00068476 0.25556993 + layer.4.v_cache 0.00005875 0.02353155 + layer.4.output 0.00481428 201.61394481 + ------------------------------------------------------------------------------------- + TOTAL 0.04342071 87.52478937 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 275236 +BPFP 0.9199 bits/point +EBPFP 1.8398 equivalent bits/point +MSE 87.524789 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.009s, Pack+Encode: 0.247s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 87.5248 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,072B, BPFP=0.5581 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,312B, BPFP=1.7416 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,016B, BPFP=0.8007 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,564B, BPFP=1.6956 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,936B, BPFP=0.9188 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,824B, BPFP=1.6501 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,036B, BPFP=0.8634 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,140B, BPFP=1.6695 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,112B, BPFP=1.3602 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,892B, BPFP=1.5928 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,848B, BPFP=0.3326 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13396681 61.38590751 + layer.0.v_cache 0.00001641 0.01108906 + layer.1.k_cache 0.47111181 9.78956135 + layer.1.v_cache 0.00000608 0.00430618 + layer.2.k_cache 0.02933135 0.93236211 + layer.2.v_cache 0.00002255 0.01186161 + layer.3.k_cache 0.03744583 4.14956497 + layer.3.v_cache 0.00002260 0.01320549 + layer.4.k_cache 0.00071777 0.25140393 + layer.4.v_cache 0.00004917 0.02338830 + layer.4.output 1.20538323 212.56028543 + ------------------------------------------------------------------------------------- + TOTAL 0.53590429 92.02909697 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 246752 +BPFP 0.8929 bits/point +EBPFP 1.7858 equivalent bits/point +MSE 92.029097 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.009s, Pack+Encode: 0.223s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 92.0291 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,044B, BPFP=0.5449 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,460B, BPFP=1.8153 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,640B, BPFP=0.7943 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,416B, BPFP=1.7587 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,960B, BPFP=0.9201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,400B, BPFP=1.7036 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,756B, BPFP=0.8548 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,840B, BPFP=1.7274 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,792B, BPFP=1.3993 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,488B, BPFP=1.6541 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,348B, BPFP=0.2740 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14060297 59.07516140 + layer.0.v_cache 0.00001598 0.01085684 + layer.1.k_cache 0.58610762 9.69959344 + layer.1.v_cache 0.00000669 0.00417019 + layer.2.k_cache 0.02834670 0.91476303 + layer.2.v_cache 0.00002256 0.01191007 + layer.3.k_cache 0.01962393 3.70558082 + layer.3.v_cache 0.00002003 0.01300195 + layer.4.k_cache 0.00067797 0.25874681 + layer.4.v_cache 0.00005035 0.02329860 + layer.4.output 0.00463834 192.39967758 + ------------------------------------------------------------------------------------- + TOTAL 0.04752607 83.55969566 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 278144 +BPFP 0.8877 bits/point +EBPFP 1.7753 equivalent bits/point +MSE 83.559696 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.009s, Pack+Encode: 0.256s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 83.5597 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,516B, BPFP=0.5405 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,940B, BPFP=1.7958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,436B, BPFP=0.7934 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,928B, BPFP=1.7438 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,796B, BPFP=0.9147 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,044B, BPFP=1.6984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,472B, BPFP=0.8466 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,368B, BPFP=1.7150 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,956B, BPFP=1.3855 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,824B, BPFP=1.6357 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,900B, BPFP=0.2783 +⌛️ [2/4] FRONTEND: Frontend time: 0.263s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12404394 59.85010408 + layer.0.v_cache 0.00001645 0.01066311 + layer.1.k_cache 0.57408182 9.96207629 + layer.1.v_cache 0.00000616 0.00395917 + layer.2.k_cache 0.03522489 0.88773577 + layer.2.v_cache 0.00002326 0.01167208 + layer.3.k_cache 0.02908128 3.92269576 + layer.3.v_cache 0.00002078 0.01285377 + layer.4.k_cache 0.00072493 0.25006367 + layer.4.v_cache 0.00004967 0.02219210 + layer.4.output 0.04390477 178.02198367 + ------------------------------------------------------------------------------------- + TOTAL 0.06297686 77.71105303 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 292180 +BPFP 0.8834 bits/point +EBPFP 1.7668 equivalent bits/point +MSE 77.711053 +---------------------- -------------------------------------------------------- +Time: 0.660s Load: 0.011s, Pack+Encode: 0.263s, Decode+Unpack: 0.386s +---------------------- -------------------------------------------------------- +💾 Converting with 77.7111 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 242, 128) +Output shape: (1, 242, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.output: torch.Size([1, 242, 3584]) -> torch.Size([1, 1, 242, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,356B, BPFP=0.5395 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,104B, BPFP=1.8146 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,476B, BPFP=0.8055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,248B, BPFP=1.7593 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,360B, BPFP=0.9272 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,580B, BPFP=1.7162 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,396B, BPFP=0.8649 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,964B, BPFP=1.7410 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,900B, BPFP=1.4140 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,680B, BPFP=1.6581 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,260B, BPFP=0.2883 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11466674 59.66578157 + layer.0.v_cache 0.00001794 0.01109353 + layer.1.k_cache 0.39062487 10.04900568 + layer.1.v_cache 0.00000616 0.00417847 + layer.2.k_cache 0.02414889 0.95679133 + layer.2.v_cache 0.00002046 0.01178703 + layer.3.k_cache 0.03495786 3.73712133 + layer.3.v_cache 0.00002053 0.01335318 + layer.4.k_cache 0.00069125 0.25515337 + layer.4.v_cache 0.00005034 0.02350223 + layer.4.output 1.26514320 223.39896325 + ------------------------------------------------------------------------------------- + TOTAL 0.55418868 96.38355944 + (elements=2,106,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2106368 +Total Bytes 236324 +BPFP 0.8976 bits/point +EBPFP 1.7951 equivalent bits/point +MSE 96.383559 +---------------------- -------------------------------------------------------- +Time: 0.566s Load: 0.009s, Pack+Encode: 0.229s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 96.3836 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 189, 128) +Output shape: (1, 189, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.output: torch.Size([1, 189, 3584]) -> torch.Size([1, 1, 189, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,652B, BPFP=0.5499 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,568B, BPFP=1.7831 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,016B, BPFP=0.8280 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,952B, BPFP=1.7321 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,664B, BPFP=0.9643 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,540B, BPFP=1.6981 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,948B, BPFP=0.9051 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,788B, BPFP=1.7186 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,072B, BPFP=1.4114 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,856B, BPFP=1.6415 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,036B, BPFP=0.3429 +⌛️ [2/4] FRONTEND: Frontend time: 0.202s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14361043 62.13789166 + layer.0.v_cache 0.00001648 0.01156750 + layer.1.k_cache 0.18932698 9.52134745 + layer.1.v_cache 0.00000663 0.00488954 + layer.2.k_cache 0.00768016 0.95099232 + layer.2.v_cache 0.00002453 0.01299253 + layer.3.k_cache 0.01421528 3.57250137 + layer.3.v_cache 0.00002024 0.01438400 + layer.4.k_cache 0.00072804 0.27404147 + layer.4.v_cache 0.00005086 0.02507072 + layer.4.output 0.00858909 289.57397959 + ------------------------------------------------------------------------------------- + TOTAL 0.02445902 123.73785504 + (elements=1,645,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1645056 +Total Bytes 189092 +BPFP 0.9196 bits/point +EBPFP 1.8391 equivalent bits/point +MSE 123.737855 +---------------------- -------------------------------------------------------- +Time: 0.480s Load: 0.006s, Pack+Encode: 0.202s, Decode+Unpack: 0.272s +---------------------- -------------------------------------------------------- +💾 Converting with 123.7379 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,656B, BPFP=0.5616 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,208B, BPFP=1.9225 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,328B, BPFP=0.8310 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,400B, BPFP=1.8633 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,916B, BPFP=0.9475 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,564B, BPFP=1.8019 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,028B, BPFP=0.8823 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,092B, BPFP=1.8407 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,080B, BPFP=1.4730 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,748B, BPFP=1.7421 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,652B, BPFP=0.2898 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16894778 61.49859705 + layer.0.v_cache 0.00001779 0.01147032 + layer.1.k_cache 0.37628758 10.22747946 + layer.1.v_cache 0.00000652 0.00444978 + layer.2.k_cache 0.02181128 0.94314439 + layer.2.v_cache 0.00002030 0.01213694 + layer.3.k_cache 0.01815432 3.61210725 + layer.3.v_cache 0.00002063 0.01427263 + layer.4.k_cache 0.00071780 0.26359460 + layer.4.v_cache 0.00005059 0.02427444 + layer.4.output 1.43727796 254.04768192 + ------------------------------------------------------------------------------------- + TOTAL 0.62629296 109.11442943 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 216672 +BPFP 0.9350 bits/point +EBPFP 1.8699 equivalent bits/point +MSE 109.114429 +---------------------- -------------------------------------------------------- +Time: 0.565s Load: 0.010s, Pack+Encode: 0.227s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 109.1144 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,868B, BPFP=0.5692 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,116B, BPFP=1.8892 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,540B, BPFP=0.8348 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,236B, BPFP=1.8255 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,016B, BPFP=0.9416 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,428B, BPFP=1.7671 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,036B, BPFP=0.8707 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,012B, BPFP=1.8093 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,812B, BPFP=1.4332 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,440B, BPFP=1.6956 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,108B, BPFP=0.3111 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09877178 59.26760073 + layer.0.v_cache 0.00001652 0.01090100 + layer.1.k_cache 0.28824160 10.12660613 + layer.1.v_cache 0.00000619 0.00435317 + layer.2.k_cache 0.02137197 1.04187669 + layer.2.v_cache 0.00001995 0.01155659 + layer.3.k_cache 0.03811082 4.06086420 + layer.3.v_cache 0.00002058 0.01325361 + layer.4.k_cache 0.00070031 0.25571941 + layer.4.v_cache 0.00004695 0.02210614 + layer.4.output 1.41729720 250.45184358 + ------------------------------------------------------------------------------------- + TOTAL 0.60990512 107.52810252 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 218612 +BPFP 0.9302 bits/point +EBPFP 1.8605 equivalent bits/point +MSE 107.528103 +---------------------- -------------------------------------------------------- +Time: 0.562s Load: 0.007s, Pack+Encode: 0.227s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 107.5281 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,888B, BPFP=0.5454 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,748B, BPFP=1.8493 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,912B, BPFP=0.8236 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,924B, BPFP=1.7923 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,116B, BPFP=0.9068 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,016B, BPFP=1.7295 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,320B, BPFP=0.8518 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,248B, BPFP=1.7456 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,360B, BPFP=1.4076 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,284B, BPFP=1.6789 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,864B, BPFP=0.2752 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12161526 63.36243432 + layer.0.v_cache 0.00001916 0.01126398 + layer.1.k_cache 0.26028682 10.00924278 + layer.1.v_cache 0.00000628 0.00428653 + layer.2.k_cache 0.01296707 0.96939073 + layer.2.v_cache 0.00002040 0.01155143 + layer.3.k_cache 0.04595671 4.00686402 + layer.3.v_cache 0.00002073 0.01319541 + layer.4.k_cache 0.00067352 0.25575915 + layer.4.v_cache 0.00005047 0.02347675 + layer.4.output 1.35465500 239.50882980 + ------------------------------------------------------------------------------------- + TOTAL 0.58377655 103.24878081 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 220680 +BPFP 0.8975 bits/point +EBPFP 1.7950 equivalent bits/point +MSE 103.248781 +---------------------- -------------------------------------------------------- +Time: 0.560s Load: 0.007s, Pack+Encode: 0.227s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 103.2488 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,940B, BPFP=0.5489 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,452B, BPFP=1.8288 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,868B, BPFP=0.8205 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,712B, BPFP=1.7777 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,192B, BPFP=0.9121 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,956B, BPFP=1.7254 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,160B, BPFP=0.8407 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,284B, BPFP=1.7481 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,244B, BPFP=1.3996 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,040B, BPFP=1.6621 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,048B, BPFP=0.2671 +⌛️ [2/4] FRONTEND: Frontend time: 0.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10229862 64.65205856 + layer.0.v_cache 0.00001727 0.01093745 + layer.1.k_cache 0.34669285 9.88231342 + layer.1.v_cache 0.00000643 0.00416649 + layer.2.k_cache 0.01639456 0.90864455 + layer.2.v_cache 0.00002164 0.01173715 + layer.3.k_cache 0.03343437 4.37308928 + layer.3.v_cache 0.00002043 0.01317492 + layer.4.k_cache 0.00071218 0.25027385 + layer.4.v_cache 0.00005088 0.02299508 + layer.4.output 1.35462105 239.50268647 + ------------------------------------------------------------------------------------- + TOTAL 0.58717627 103.33224683 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 218896 +BPFP 0.8902 bits/point +EBPFP 1.7805 equivalent bits/point +MSE 103.332247 +---------------------- -------------------------------------------------------- +Time: 0.562s Load: 0.007s, Pack+Encode: 0.227s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 103.3322 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,596B, BPFP=0.5229 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,008B, BPFP=1.8590 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,072B, BPFP=0.8309 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,240B, BPFP=1.8062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,796B, BPFP=0.9496 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,432B, BPFP=1.7506 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,652B, BPFP=0.8709 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,836B, BPFP=1.7784 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,828B, BPFP=1.4336 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,612B, BPFP=1.6941 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,144B, BPFP=0.2669 +⌛️ [2/4] FRONTEND: Frontend time: 0.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11800324 63.20004044 + layer.0.v_cache 0.00001819 0.01115208 + layer.1.k_cache 0.35011090 10.28568755 + layer.1.v_cache 0.00000612 0.00414826 + layer.2.k_cache 0.02017111 0.93949971 + layer.2.v_cache 0.00002006 0.01165561 + layer.3.k_cache 0.03026835 4.15233434 + layer.3.v_cache 0.00001975 0.01361664 + layer.4.k_cache 0.00074701 0.25301238 + layer.4.v_cache 0.00005118 0.02317991 + layer.4.output 1.34863711 238.41862020 + ------------------------------------------------------------------------------------- + TOTAL 0.58587504 102.81321578 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 223216 +BPFP 0.9038 bits/point +EBPFP 1.8076 equivalent bits/point +MSE 102.813216 +---------------------- -------------------------------------------------------- +Time: 0.567s Load: 0.007s, Pack+Encode: 0.231s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 102.8132 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,548B, BPFP=0.5150 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,244B, BPFP=1.8589 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,904B, BPFP=0.8122 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,456B, BPFP=1.8051 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,012B, BPFP=0.9561 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,860B, BPFP=1.7645 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,216B, BPFP=0.9017 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,392B, BPFP=1.8008 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,472B, BPFP=1.4651 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,168B, BPFP=1.7172 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,952B, BPFP=0.2530 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10957497 60.71674570 + layer.0.v_cache 0.00001697 0.01165398 + layer.1.k_cache 0.31262030 10.13662813 + layer.1.v_cache 0.00000629 0.00464612 + layer.2.k_cache 0.01311818 0.97333397 + layer.2.v_cache 0.00002149 0.01280945 + layer.3.k_cache 0.06227440 4.05420828 + layer.3.v_cache 0.00002139 0.01530444 + layer.4.k_cache 0.00068925 0.27484489 + layer.4.v_cache 0.00005302 0.02643480 + layer.4.output 1.33688878 236.40989551 + ------------------------------------------------------------------------------------- + TOTAL 0.57980104 101.82916931 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 225224 +BPFP 0.9040 bits/point +EBPFP 1.8079 equivalent bits/point +MSE 101.829169 +---------------------- -------------------------------------------------------- +Time: 0.563s Load: 0.007s, Pack+Encode: 0.229s, Decode+Unpack: 0.327s +---------------------- -------------------------------------------------------- +💾 Converting with 101.8292 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,804B, BPFP=0.5725 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,940B, BPFP=1.9029 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,316B, BPFP=0.8301 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,888B, BPFP=1.8257 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,652B, BPFP=0.9281 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,048B, BPFP=1.7641 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,808B, BPFP=0.8662 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,524B, BPFP=1.7990 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,664B, BPFP=1.4425 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,368B, BPFP=1.7142 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,580B, BPFP=0.2890 +⌛️ [2/4] FRONTEND: Frontend time: 0.232s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13987277 60.55809401 + layer.0.v_cache 0.00001887 0.01164229 + layer.1.k_cache 0.23498858 10.43671141 + layer.1.v_cache 0.00000603 0.00421158 + layer.2.k_cache 0.02028183 0.95487496 + layer.2.v_cache 0.00001992 0.01175731 + layer.3.k_cache 0.05552547 3.73598828 + layer.3.v_cache 0.00002044 0.01387515 + layer.4.k_cache 0.00067773 0.26676764 + layer.4.v_cache 0.00004777 0.02381918 + layer.4.output 1.43727485 254.01563548 + ------------------------------------------------------------------------------------- + TOTAL 0.61837549 109.06630530 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 213592 +BPFP 0.9217 bits/point +EBPFP 1.8433 equivalent bits/point +MSE 109.066305 +---------------------- -------------------------------------------------------- +Time: 0.567s Load: 0.007s, Pack+Encode: 0.232s, Decode+Unpack: 0.328s +---------------------- -------------------------------------------------------- +💾 Converting with 109.0663 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,784B, BPFP=0.5764 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,708B, BPFP=1.9037 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,140B, BPFP=0.8249 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,032B, BPFP=1.8537 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,584B, BPFP=0.9319 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,024B, BPFP=1.7790 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,904B, BPFP=0.8815 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,368B, BPFP=1.8045 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,736B, BPFP=1.4615 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 23,152B, BPFP=1.7145 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,100B, BPFP=0.2973 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17004660 62.83951052 + layer.0.v_cache 0.00001740 0.01149669 + layer.1.k_cache 0.16740718 10.45101567 + layer.1.v_cache 0.00000654 0.00462026 + layer.2.k_cache 0.01191964 0.99537507 + layer.2.v_cache 0.00001991 0.01237107 + layer.3.k_cache 0.00660476 4.07196450 + layer.3.v_cache 0.00002076 0.01395584 + layer.4.k_cache 0.00070551 0.26876274 + layer.4.v_cache 0.00005179 0.02469811 + layer.4.output 1.45088344 256.30903013 + ------------------------------------------------------------------------------------- + TOTAL 0.61841083 110.16805773 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 213532 +BPFP 0.9301 bits/point +EBPFP 1.8603 equivalent bits/point +MSE 110.168058 +---------------------- -------------------------------------------------------- +Time: 0.563s Load: 0.007s, Pack+Encode: 0.226s, Decode+Unpack: 0.330s +---------------------- -------------------------------------------------------- +💾 Converting with 110.1681 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,964B, BPFP=0.5410 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,024B, BPFP=1.8359 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,124B, BPFP=0.8236 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,276B, BPFP=1.7851 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,888B, BPFP=0.9435 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,672B, BPFP=1.7440 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,692B, BPFP=0.8622 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,704B, BPFP=1.7462 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,184B, BPFP=1.4391 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,628B, BPFP=1.6731 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,784B, BPFP=0.2502 +⌛️ [2/4] FRONTEND: Frontend time: 0.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14712874 63.45986328 + layer.0.v_cache 0.00001826 0.01087308 + layer.1.k_cache 0.36788522 9.99528384 + layer.1.v_cache 0.00000598 0.00433844 + layer.2.k_cache 0.01156470 0.98118114 + layer.2.v_cache 0.00002030 0.01220683 + layer.3.k_cache 0.01315050 4.04673966 + layer.3.v_cache 0.00002017 0.01333607 + layer.4.k_cache 0.00068370 0.26590472 + layer.4.v_cache 0.00004817 0.02345808 + layer.4.output 1.33105469 235.36125776 + ------------------------------------------------------------------------------------- + TOTAL 0.57987697 101.54952879 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 222940 +BPFP 0.8909 bits/point +EBPFP 1.7818 equivalent bits/point +MSE 101.549529 +---------------------- -------------------------------------------------------- +Time: 0.564s Load: 0.008s, Pack+Encode: 0.230s, Decode+Unpack: 0.327s +---------------------- -------------------------------------------------------- +💾 Converting with 101.5495 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 188, 128) +Output shape: (1, 188, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.output: torch.Size([1, 188, 3584]) -> torch.Size([1, 1, 188, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,504B, BPFP=0.5406 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,304B, BPFP=1.7706 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,760B, BPFP=0.8112 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,788B, BPFP=1.7277 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,360B, BPFP=0.9441 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,304B, BPFP=1.6875 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,552B, BPFP=0.8770 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,568B, BPFP=1.7094 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,040B, BPFP=1.4162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,656B, BPFP=1.6336 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,976B, BPFP=0.2965 +⌛️ [2/4] FRONTEND: Frontend time: 0.197s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.269s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005002 65.25363094 + layer.0.v_cache 0.00001663 0.01135626 + layer.1.k_cache 0.14149688 9.89013802 + layer.1.v_cache 0.00000584 0.00418818 + layer.2.k_cache 0.00881058 0.99994262 + layer.2.v_cache 0.00002285 0.01235307 + layer.3.k_cache 0.03192479 4.00829624 + layer.3.v_cache 0.00001958 0.01363782 + layer.4.k_cache 0.00066612 0.26518712 + layer.4.v_cache 0.00005162 0.02492483 + layer.4.output 0.00857985 291.03894377 + ------------------------------------------------------------------------------------- + TOTAL 0.02253670 124.57389773 + (elements=1,636,352) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1636352 +Total Bytes 182812 +BPFP 0.8938 bits/point +EBPFP 1.7875 equivalent bits/point +MSE 124.573898 +---------------------- -------------------------------------------------------- +Time: 0.473s Load: 0.006s, Pack+Encode: 0.197s, Decode+Unpack: 0.269s +---------------------- -------------------------------------------------------- +💾 Converting with 124.5739 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,440B, BPFP=0.5057 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,052B, BPFP=1.9670 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,716B, BPFP=0.8414 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,892B, BPFP=1.8759 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,572B, BPFP=0.9871 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,272B, BPFP=1.8273 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,292B, BPFP=0.8866 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,868B, BPFP=1.8741 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,060B, BPFP=1.4965 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,472B, BPFP=1.7644 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,720B, BPFP=0.2548 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11435634 66.55137013 + layer.0.v_cache 0.00001764 0.01205922 + layer.1.k_cache 0.25025516 10.01436504 + layer.1.v_cache 0.00000690 0.00461354 + layer.2.k_cache 0.01528106 1.05264275 + layer.2.v_cache 0.00002105 0.01279030 + layer.3.k_cache 0.02533426 4.01626173 + layer.3.v_cache 0.00002127 0.01464616 + layer.4.k_cache 0.00070809 0.27447374 + layer.4.v_cache 0.00005072 0.02505021 + layer.4.output 1.53833902 272.06732322 + ------------------------------------------------------------------------------------- + TOTAL 0.65731915 116.84997267 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 201356 +BPFP 0.9300 bits/point +EBPFP 1.8600 equivalent bits/point +MSE 116.849973 +---------------------- -------------------------------------------------------- +Time: 0.554s Load: 0.009s, Pack+Encode: 0.228s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 116.8500 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,228B, BPFP=0.5406 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,284B, BPFP=1.8476 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,560B, BPFP=0.8299 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,556B, BPFP=1.7844 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,064B, BPFP=0.9604 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,112B, BPFP=1.7458 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,336B, BPFP=0.8972 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,540B, BPFP=1.7830 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,820B, BPFP=1.4601 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,552B, BPFP=1.6972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,764B, BPFP=0.2947 +⌛️ [2/4] FRONTEND: Frontend time: 0.192s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.260s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258550 60.42642687 + layer.0.v_cache 0.00001767 0.01205672 + layer.1.k_cache 0.14790291 10.11526964 + layer.1.v_cache 0.00000629 0.00444097 + layer.2.k_cache 0.00999225 0.99611528 + layer.2.v_cache 0.00002110 0.01278007 + layer.3.k_cache 0.02117847 3.70526598 + layer.3.v_cache 0.00002045 0.01510751 + layer.4.k_cache 0.00067024 0.26941066 + layer.4.v_cache 0.00006105 0.02519554 + layer.4.output 0.00897616 304.31483135 + ------------------------------------------------------------------------------------- + TOTAL 0.02207583 129.75211110 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 179816 +BPFP 0.9182 bits/point +EBPFP 1.8364 equivalent bits/point +MSE 129.752111 +---------------------- -------------------------------------------------------- +Time: 0.458s Load: 0.006s, Pack+Encode: 0.192s, Decode+Unpack: 0.260s +---------------------- -------------------------------------------------------- +💾 Converting with 129.7521 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,180B, BPFP=0.5499 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 25,368B, BPFP=1.9430 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,208B, BPFP=0.8585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 24,564B, BPFP=1.8814 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,660B, BPFP=0.9697 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,640B, BPFP=1.8107 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,624B, BPFP=0.8903 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 24,144B, BPFP=1.8493 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,472B, BPFP=1.4914 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,964B, BPFP=1.7589 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,740B, BPFP=0.2707 +⌛️ [2/4] FRONTEND: Frontend time: 0.222s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09364704 61.65197515 + layer.0.v_cache 0.00001853 0.01152093 + layer.1.k_cache 0.25163280 10.54417629 + layer.1.v_cache 0.00000592 0.00432890 + layer.2.k_cache 0.01595125 1.01958877 + layer.2.v_cache 0.00002050 0.01195035 + layer.3.k_cache 0.04325477 4.10622062 + layer.3.v_cache 0.00002016 0.01401059 + layer.4.k_cache 0.00069210 0.26397262 + layer.4.v_cache 0.00005026 0.02460581 + layer.4.output 1.50065788 265.13572304 + ------------------------------------------------------------------------------------- + TOTAL 0.64175873 113.74131831 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 207564 +BPFP 0.9352 bits/point +EBPFP 1.8704 equivalent bits/point +MSE 113.741318 +---------------------- -------------------------------------------------------- +Time: 0.546s Load: 0.007s, Pack+Encode: 0.222s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 113.7413 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,984B, BPFP=0.5520 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,448B, BPFP=1.8285 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,780B, BPFP=0.8144 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,592B, BPFP=1.7694 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,976B, BPFP=0.8971 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 24,788B, BPFP=1.7138 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,280B, BPFP=0.8490 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,172B, BPFP=1.7403 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,212B, BPFP=1.3974 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,040B, BPFP=1.6621 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,476B, BPFP=0.2911 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510114 63.02826846 + layer.0.v_cache 0.00001717 0.01089249 + layer.1.k_cache 0.31797088 10.06675842 + layer.1.v_cache 0.00000673 0.00432761 + layer.2.k_cache 0.01711024 1.02248909 + layer.2.v_cache 0.00002022 0.01189952 + layer.3.k_cache 0.01307247 3.77389553 + layer.3.v_cache 0.00002089 0.01330673 + layer.4.k_cache 0.00069413 0.25922417 + layer.4.v_cache 0.00005461 0.02432660 + layer.4.output 1.35466395 239.51461757 + ------------------------------------------------------------------------------------- + TOTAL 0.58392448 103.22457127 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 220748 +BPFP 0.8978 bits/point +EBPFP 1.7955 equivalent bits/point +MSE 103.224571 +---------------------- -------------------------------------------------------- +Time: 0.545s Load: 0.008s, Pack+Encode: 0.221s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 103.2246 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,584B, BPFP=0.5408 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,180B, BPFP=1.7755 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,704B, BPFP=0.8004 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,496B, BPFP=1.7324 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,632B, BPFP=0.9219 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,812B, BPFP=1.6893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,556B, BPFP=0.8541 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,080B, BPFP=1.7061 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,956B, BPFP=1.3833 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,908B, BPFP=1.6323 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,732B, BPFP=0.3036 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14860459 57.69366652 + layer.0.v_cache 0.00001629 0.01063934 + layer.1.k_cache 0.38266554 9.99492818 + layer.1.v_cache 0.00000669 0.00421165 + layer.2.k_cache 0.01628031 0.91132182 + layer.2.v_cache 0.00002106 0.01153814 + layer.3.k_cache 0.01842327 3.77041035 + layer.3.v_cache 0.00002289 0.01377083 + layer.4.k_cache 0.00072690 0.26241952 + layer.4.v_cache 0.00004923 0.02320238 + layer.4.output 1.23454798 218.11256120 + ------------------------------------------------------------------------------------- + TOTAL 0.54168545 94.08729630 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 240640 +BPFP 0.8918 bits/point +EBPFP 1.7837 equivalent bits/point +MSE 94.087296 +---------------------- -------------------------------------------------------- +Time: 0.555s Load: 0.008s, Pack+Encode: 0.229s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 94.0873 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 435, 128) +Output shape: (1, 435, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.output: torch.Size([1, 435, 3584]) -> torch.Size([1, 1, 435, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,076B, BPFP=0.5415 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,124B, BPFP=1.7645 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,724B, BPFP=0.7803 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,632B, BPFP=1.7109 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,508B, BPFP=0.8803 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,176B, BPFP=1.6586 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,776B, BPFP=0.8181 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,464B, BPFP=1.6690 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,348B, BPFP=1.3415 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,404B, BPFP=1.5950 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 60,968B, BPFP=0.3128 +⌛️ [2/4] FRONTEND: Frontend time: 0.323s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.507s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15864419 61.10314745 + layer.0.v_cache 0.00001607 0.01066575 + layer.1.k_cache 0.94329666 10.06435996 + layer.1.v_cache 0.00000639 0.00419186 + layer.2.k_cache 0.02505282 0.93555340 + layer.2.v_cache 0.00002131 0.01139401 + layer.3.k_cache 0.01724711 3.69509474 + layer.3.v_cache 0.00002032 0.01249103 + layer.4.k_cache 0.00074513 0.25119224 + layer.4.v_cache 0.00005173 0.02251967 + layer.4.output 0.00608059 127.07740148 + ------------------------------------------------------------------------------------- + TOTAL 0.06986270 56.80308356 + (elements=3,786,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3786240 +Total Bytes 416200 +BPFP 0.8794 bits/point +EBPFP 1.7588 equivalent bits/point +MSE 56.803084 +---------------------- -------------------------------------------------------- +Time: 0.844s Load: 0.014s, Pack+Encode: 0.323s, Decode+Unpack: 0.507s +---------------------- -------------------------------------------------------- +💾 Converting with 56.8031 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 437, 128) +Output shape: (1, 437, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.output: torch.Size([1, 437, 3584]) -> torch.Size([1, 1, 437, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,972B, BPFP=0.5353 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,104B, BPFP=1.7557 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,820B, BPFP=0.7802 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,544B, BPFP=1.6999 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,580B, BPFP=0.8789 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,288B, BPFP=1.6550 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,812B, BPFP=0.8156 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,744B, BPFP=1.6713 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,496B, BPFP=1.3407 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,596B, BPFP=1.5945 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 60,504B, BPFP=0.3090 +⌛️ [2/4] FRONTEND: Frontend time: 0.322s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.502s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14454122 58.29351402 + layer.0.v_cache 0.00001702 0.01081199 + layer.1.k_cache 0.94267751 10.03543112 + layer.1.v_cache 0.00000668 0.00434074 + layer.2.k_cache 0.01808156 0.91271239 + layer.2.v_cache 0.00002154 0.01159829 + layer.3.k_cache 0.03034366 3.54446907 + layer.3.v_cache 0.00002210 0.01340339 + layer.4.k_cache 0.00074160 0.26125591 + layer.4.v_cache 0.00005552 0.02390142 + layer.4.output 0.00611241 126.66794704 + ------------------------------------------------------------------------------------- + TOTAL 0.06937031 56.45806280 + (elements=3,803,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3803648 +Total Bytes 416460 +BPFP 0.8759 bits/point +EBPFP 1.7518 equivalent bits/point +MSE 56.458063 +---------------------- -------------------------------------------------------- +Time: 0.838s Load: 0.015s, Pack+Encode: 0.322s, Decode+Unpack: 0.502s +---------------------- -------------------------------------------------------- +💾 Converting with 56.4581 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,540B, BPFP=0.5347 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,152B, BPFP=1.7833 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,464B, BPFP=0.7845 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,980B, BPFP=1.7238 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,696B, BPFP=0.8977 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,080B, BPFP=1.6782 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,464B, BPFP=0.8352 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,532B, BPFP=1.7011 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,892B, BPFP=1.3642 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,860B, BPFP=1.6163 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,980B, BPFP=0.3187 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12827574 59.09907036 + layer.0.v_cache 0.00001866 0.01111065 + layer.1.k_cache 0.53653341 10.20294744 + layer.1.v_cache 0.00000626 0.00407115 + layer.2.k_cache 0.01950284 0.95340679 + layer.2.v_cache 0.00002083 0.01124931 + layer.3.k_cache 0.04340698 3.57455088 + layer.3.v_cache 0.00002181 0.01324801 + layer.4.k_cache 0.00072091 0.25289258 + layer.4.v_cache 0.00005393 0.02300110 + layer.4.output 0.04342448 175.67310703 + ------------------------------------------------------------------------------------- + TOTAL 0.06073722 76.69748809 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 298640 +BPFP 0.8912 bits/point +EBPFP 1.7824 equivalent bits/point +MSE 76.697488 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 76.6975 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,792B, BPFP=0.5164 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,432B, BPFP=1.9051 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,952B, BPFP=0.8195 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,412B, BPFP=1.8452 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,064B, BPFP=0.9436 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,304B, BPFP=1.7801 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,720B, BPFP=0.8647 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,696B, BPFP=1.8031 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,932B, BPFP=1.4645 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,352B, BPFP=1.7242 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,772B, BPFP=0.2918 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702504 62.12096158 + layer.0.v_cache 0.00001641 0.01151999 + layer.1.k_cache 0.51130361 9.98887325 + layer.1.v_cache 0.00000656 0.00462745 + layer.2.k_cache 0.01810413 0.97242955 + layer.2.v_cache 0.00002122 0.01259817 + layer.3.k_cache 0.02476560 4.09277894 + layer.3.v_cache 0.00002197 0.01402980 + layer.4.k_cache 0.00071356 0.27206802 + layer.4.v_cache 0.00005290 0.02532031 + layer.4.output 0.00501064 208.55917696 + ------------------------------------------------------------------------------------- + TOTAL 0.04100620 90.43702622 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 267428 +BPFP 0.9241 bits/point +EBPFP 1.8481 equivalent bits/point +MSE 90.437026 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 90.4370 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,308B, BPFP=0.5548 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,796B, BPFP=1.8560 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,280B, BPFP=0.8200 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,848B, BPFP=1.7927 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,104B, BPFP=0.9418 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,236B, BPFP=1.7519 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,200B, BPFP=0.8814 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,520B, BPFP=1.7708 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,332B, BPFP=1.4244 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,316B, BPFP=1.6904 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,200B, BPFP=0.2881 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11892304 61.75415665 + layer.0.v_cache 0.00001803 0.01170209 + layer.1.k_cache 0.31174104 10.13743031 + layer.1.v_cache 0.00000657 0.00443206 + layer.2.k_cache 0.01452858 0.96801667 + layer.2.v_cache 0.00002161 0.01245525 + layer.3.k_cache 0.02263702 3.79582593 + layer.3.v_cache 0.00002141 0.01418562 + layer.4.k_cache 0.00068312 0.26338202 + layer.4.v_cache 0.00005251 0.02433159 + layer.4.output 1.30838121 231.27184447 + ------------------------------------------------------------------------------------- + TOTAL 0.56631185 99.75816644 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 232140 +BPFP 0.9118 bits/point +EBPFP 1.8236 equivalent bits/point +MSE 99.758166 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.008s, Pack+Encode: 0.223s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 99.7582 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 233, 128) +Output shape: (1, 233, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.output: torch.Size([1, 233, 3584]) -> torch.Size([1, 1, 233, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,072B, BPFP=0.5413 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,500B, BPFP=1.8442 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,308B, BPFP=0.8254 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,712B, BPFP=1.7913 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,904B, BPFP=0.9324 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,924B, BPFP=1.7385 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,644B, BPFP=0.8479 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,244B, BPFP=1.7599 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,204B, BPFP=1.4219 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,932B, BPFP=1.6719 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,164B, BPFP=0.2794 +⌛️ [2/4] FRONTEND: Frontend time: 0.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13365312 62.99503336 + layer.0.v_cache 0.00001735 0.01159997 + layer.1.k_cache 0.40872042 10.34432630 + layer.1.v_cache 0.00000792 0.00488796 + layer.2.k_cache 0.01308113 0.97565215 + layer.2.v_cache 0.00002137 0.01244658 + layer.3.k_cache 0.00601081 3.85077402 + layer.3.v_cache 0.00002094 0.01370962 + layer.4.k_cache 0.00068546 0.26737238 + layer.4.v_cache 0.00005338 0.02504657 + layer.4.output 1.31399239 232.26389102 + ------------------------------------------------------------------------------------- + TOTAL 0.57413051 100.25576977 + (elements=2,028,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2028032 +Total Bytes 228608 +BPFP 0.9018 bits/point +EBPFP 1.8036 equivalent bits/point +MSE 100.255770 +---------------------- -------------------------------------------------------- +Time: 0.548s Load: 0.008s, Pack+Encode: 0.223s, Decode+Unpack: 0.317s +---------------------- -------------------------------------------------------- +💾 Converting with 100.2558 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 148, 128) +Output shape: (1, 148, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.output: torch.Size([1, 148, 3584]) -> torch.Size([1, 1, 148, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,712B, BPFP=0.6030 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,080B, BPFP=2.0144 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,272B, BPFP=0.8733 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,384B, BPFP=1.9409 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,324B, BPFP=0.9844 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,852B, BPFP=1.8847 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,708B, BPFP=0.9193 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,096B, BPFP=1.9105 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,612B, BPFP=1.5427 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,204B, BPFP=1.8163 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,560B, BPFP=0.2799 +⌛️ [2/4] FRONTEND: Frontend time: 0.192s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.259s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08608804 65.96802418 + layer.0.v_cache 0.00001767 0.01216275 + layer.1.k_cache 0.12276106 10.57214933 + layer.1.v_cache 0.00000594 0.00455454 + layer.2.k_cache 0.00886028 1.09155768 + layer.2.v_cache 0.00002088 0.01291408 + layer.3.k_cache 0.01513371 4.60352965 + layer.3.v_cache 0.00002085 0.01522635 + layer.4.k_cache 0.00068430 0.28329009 + layer.4.v_cache 0.00005331 0.02584423 + layer.4.output 0.08957551 365.91370053 + ------------------------------------------------------------------------------------- + TOTAL 0.05062792 155.52853862 + (elements=1,288,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1288192 +Total Bytes 155804 +BPFP 0.9676 bits/point +EBPFP 1.9352 equivalent bits/point +MSE 155.528539 +---------------------- -------------------------------------------------------- +Time: 0.457s Load: 0.005s, Pack+Encode: 0.192s, Decode+Unpack: 0.259s +---------------------- -------------------------------------------------------- +💾 Converting with 155.5285 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 170, 128) +Output shape: (1, 170, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.output: torch.Size([1, 170, 3584]) -> torch.Size([1, 1, 170, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,248B, BPFP=0.5743 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,848B, BPFP=1.9162 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,404B, BPFP=0.8643 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,088B, BPFP=1.8463 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,544B, BPFP=0.9691 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,372B, BPFP=1.7805 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,800B, BPFP=0.9007 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,736B, BPFP=1.8140 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,060B, BPFP=1.4761 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,696B, BPFP=1.7184 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,380B, BPFP=0.2676 +⌛️ [2/4] FRONTEND: Frontend time: 0.192s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.257s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13519069 63.29761029 + layer.0.v_cache 0.00001779 0.01257726 + layer.1.k_cache 0.14803200 10.31665111 + layer.1.v_cache 0.00000678 0.00499705 + layer.2.k_cache 0.02267574 1.01729126 + layer.2.v_cache 0.00002005 0.01304727 + layer.3.k_cache 0.04618364 4.12082017 + layer.3.v_cache 0.00002282 0.01547780 + layer.4.k_cache 0.00077641 0.28371198 + layer.4.v_cache 0.00005264 0.02588294 + layer.4.output 0.00945594 322.78321954 + ------------------------------------------------------------------------------------- + TOTAL 0.02465707 137.56415317 + (elements=1,479,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1479680 +Total Bytes 171176 +BPFP 0.9255 bits/point +EBPFP 1.8510 equivalent bits/point +MSE 137.564153 +---------------------- -------------------------------------------------------- +Time: 0.457s Load: 0.008s, Pack+Encode: 0.192s, Decode+Unpack: 0.257s +---------------------- -------------------------------------------------------- +💾 Converting with 137.5642 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,696B, BPFP=0.5361 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,620B, BPFP=1.9409 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,152B, BPFP=0.8614 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,828B, BPFP=1.8663 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,752B, BPFP=1.0120 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,432B, BPFP=1.8291 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,992B, BPFP=0.9405 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,876B, BPFP=1.8709 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,112B, BPFP=1.5166 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,720B, BPFP=1.7620 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,732B, BPFP=0.2519 +⌛️ [2/4] FRONTEND: Frontend time: 0.196s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.266s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12520085 62.07084784 + layer.0.v_cache 0.00001716 0.01263771 + layer.1.k_cache 0.15152623 10.04769419 + layer.1.v_cache 0.00000633 0.00482823 + layer.2.k_cache 0.01605446 1.01833150 + layer.2.v_cache 0.00002200 0.01416448 + layer.3.k_cache 0.04332644 4.15179774 + layer.3.v_cache 0.00002349 0.01592321 + layer.4.k_cache 0.00069336 0.28834619 + layer.4.v_cache 0.00005614 0.02629135 + layer.4.output 0.00966471 330.67211704 + ------------------------------------------------------------------------------------- + TOTAL 0.02379879 140.72680481 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 168912 +BPFP 0.9352 bits/point +EBPFP 1.8705 equivalent bits/point +MSE 140.726805 +---------------------- -------------------------------------------------------- +Time: 0.468s Load: 0.006s, Pack+Encode: 0.196s, Decode+Unpack: 0.266s +---------------------- -------------------------------------------------------- +💾 Converting with 140.7268 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 181, 128) +Output shape: (1, 181, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.output: torch.Size([1, 181, 3584]) -> torch.Size([1, 1, 181, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,348B, BPFP=0.5480 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,400B, BPFP=1.8474 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,404B, BPFP=0.8118 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,708B, BPFP=1.7876 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,996B, BPFP=0.9492 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,084B, BPFP=1.7338 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,224B, BPFP=0.8826 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,356B, BPFP=1.7573 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,560B, BPFP=1.4296 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,400B, BPFP=1.6747 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,664B, BPFP=0.3042 +⌛️ [2/4] FRONTEND: Frontend time: 0.191s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.261s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13566811 61.65084384 + layer.0.v_cache 0.00002016 0.01227295 + layer.1.k_cache 0.14782082 10.16321678 + layer.1.v_cache 0.00000634 0.00507043 + layer.2.k_cache 0.01546242 1.04695172 + layer.2.v_cache 0.00002121 0.01313991 + layer.3.k_cache 0.02895848 4.11433470 + layer.3.v_cache 0.00002197 0.01504899 + layer.4.k_cache 0.00067433 0.28420140 + layer.4.v_cache 0.00005229 0.02601004 + layer.4.output 0.00892717 302.99452447 + ------------------------------------------------------------------------------------- + TOTAL 0.02301155 129.31133894 + (elements=1,575,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1575424 +Total Bytes 180144 +BPFP 0.9148 bits/point +EBPFP 1.8295 equivalent bits/point +MSE 129.311339 +---------------------- -------------------------------------------------------- +Time: 0.461s Load: 0.009s, Pack+Encode: 0.191s, Decode+Unpack: 0.261s +---------------------- -------------------------------------------------------- +💾 Converting with 129.3113 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 186, 128) +Output shape: (1, 186, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.output: torch.Size([1, 186, 3584]) -> torch.Size([1, 1, 186, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,464B, BPFP=0.5430 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,480B, BPFP=1.8044 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,680B, BPFP=0.8132 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,776B, BPFP=1.7453 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,144B, BPFP=0.9362 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,136B, BPFP=1.6915 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,300B, BPFP=0.8653 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,640B, BPFP=1.7339 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,752B, BPFP=1.4073 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,604B, BPFP=1.6468 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,636B, BPFP=0.2837 +⌛️ [2/4] FRONTEND: Frontend time: 0.192s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.264s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09795722 64.56391759 + layer.0.v_cache 0.00001815 0.01218084 + layer.1.k_cache 0.13430934 10.34472394 + layer.1.v_cache 0.00000632 0.00457581 + layer.2.k_cache 0.01830538 1.06868826 + layer.2.v_cache 0.00002120 0.01238370 + layer.3.k_cache 0.04892764 3.81123401 + layer.3.v_cache 0.00002113 0.01459936 + layer.4.k_cache 0.00067235 0.26930261 + layer.4.v_cache 0.00008605 0.02500420 + layer.4.output 0.00869856 294.87696813 + ------------------------------------------------------------------------------------- + TOTAL 0.02124792 126.13325807 + (elements=1,618,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1618944 +Total Bytes 180612 +BPFP 0.8925 bits/point +EBPFP 1.7850 equivalent bits/point +MSE 126.133258 +---------------------- -------------------------------------------------------- +Time: 0.462s Load: 0.007s, Pack+Encode: 0.192s, Decode+Unpack: 0.264s +---------------------- -------------------------------------------------------- +💾 Converting with 126.1333 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 163, 128) +Output shape: (1, 163, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.output: torch.Size([1, 163, 3584]) -> torch.Size([1, 1, 163, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,632B, BPFP=0.5399 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,040B, BPFP=1.9210 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,960B, BPFP=0.8589 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,440B, BPFP=1.8635 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,172B, BPFP=0.9751 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,884B, BPFP=1.8102 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,608B, BPFP=0.9210 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,312B, BPFP=1.8512 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,720B, BPFP=1.5069 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,280B, BPFP=1.7523 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,320B, BPFP=0.2509 +⌛️ [2/4] FRONTEND: Frontend time: 0.192s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.259s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11530171 60.77724190 + layer.0.v_cache 0.00001813 0.01241742 + layer.1.k_cache 0.14319220 10.06504850 + layer.1.v_cache 0.00000612 0.00464013 + layer.2.k_cache 0.02368733 1.05094198 + layer.2.v_cache 0.00002106 0.01280034 + layer.3.k_cache 0.02525310 3.96919784 + layer.3.v_cache 0.00002161 0.01530615 + layer.4.k_cache 0.00069984 0.27343214 + layer.4.v_cache 0.00005709 0.02586483 + layer.4.output 0.00982084 336.61051161 + ------------------------------------------------------------------------------------- + TOTAL 0.02217671 143.08708662 + (elements=1,418,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1418752 +Total Bytes 164368 +BPFP 0.9268 bits/point +EBPFP 1.8537 equivalent bits/point +MSE 143.087087 +---------------------- -------------------------------------------------------- +Time: 0.456s Load: 0.006s, Pack+Encode: 0.192s, Decode+Unpack: 0.259s +---------------------- -------------------------------------------------------- +💾 Converting with 143.0871 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 191, 128) +Output shape: (1, 191, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.output: torch.Size([1, 191, 3584]) -> torch.Size([1, 1, 191, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,188B, BPFP=0.5062 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,540B, BPFP=1.7621 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,436B, BPFP=0.7719 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,868B, BPFP=1.7071 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,408B, BPFP=0.9332 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,376B, BPFP=1.6669 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,580B, BPFP=0.8655 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,768B, BPFP=1.6990 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,152B, BPFP=1.4031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,816B, BPFP=1.6211 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,388B, BPFP=0.3318 +⌛️ [2/4] FRONTEND: Frontend time: 0.191s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.263s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15112510 59.19931078 + layer.0.v_cache 0.00002069 0.01232823 + layer.1.k_cache 0.18491925 9.38755854 + layer.1.v_cache 0.00000629 0.00475966 + layer.2.k_cache 0.02122488 0.94909189 + layer.2.v_cache 0.00002327 0.01333725 + layer.3.k_cache 0.01259250 3.57547141 + layer.3.v_cache 0.00002182 0.01530134 + layer.4.k_cache 0.00068006 0.27935473 + layer.4.v_cache 0.00005885 0.02622587 + layer.4.output 0.00845948 284.30471672 + ------------------------------------------------------------------------------------- + TOTAL 0.02528759 121.38798569 + (elements=1,662,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1662464 +Total Bytes 186520 +BPFP 0.8976 bits/point +EBPFP 1.7951 equivalent bits/point +MSE 121.387986 +---------------------- -------------------------------------------------------- +Time: 0.462s Load: 0.009s, Pack+Encode: 0.191s, Decode+Unpack: 0.263s +---------------------- -------------------------------------------------------- +💾 Converting with 121.3880 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 174, 128) +Output shape: (1, 174, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.output: torch.Size([1, 174, 3584]) -> torch.Size([1, 1, 174, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,248B, BPFP=0.5611 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,004B, BPFP=1.8861 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,412B, BPFP=0.8452 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,364B, BPFP=1.8287 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,176B, BPFP=1.0036 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,912B, BPFP=1.7881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,412B, BPFP=0.9350 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,244B, BPFP=1.8179 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,608B, BPFP=1.4914 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,296B, BPFP=1.7328 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,280B, BPFP=0.2473 +⌛️ [2/4] FRONTEND: Frontend time: 0.190s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.261s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14146894 64.53126684 + layer.0.v_cache 0.00001664 0.01185594 + layer.1.k_cache 0.16292955 10.58953717 + layer.1.v_cache 0.00000602 0.00458830 + layer.2.k_cache 0.01834387 1.04398399 + layer.2.v_cache 0.00002001 0.01290818 + layer.3.k_cache 0.03015787 4.16131346 + layer.3.v_cache 0.00002180 0.01524670 + layer.4.k_cache 0.00067479 0.27932318 + layer.4.v_cache 0.00005130 0.02616764 + layer.4.output 0.00923280 315.38090107 + ------------------------------------------------------------------------------------- + TOTAL 0.02460708 134.60838229 + (elements=1,514,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1514496 +Total Bytes 173956 +BPFP 0.9189 bits/point +EBPFP 1.8378 equivalent bits/point +MSE 134.608382 +---------------------- -------------------------------------------------------- +Time: 0.457s Load: 0.006s, Pack+Encode: 0.190s, Decode+Unpack: 0.261s +---------------------- -------------------------------------------------------- +💾 Converting with 134.6084 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.9101 bits/point +Avg EBPFP 1.8202 equivalent bits/point +Avg MSE 104.988866 +Avg Time 0.588s +------------------------ ---------------------------- diff --git a/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..1c35fda795ed7e6729ed91fc85b70e6878a50216 --- /dev/null +++ b/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 599 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa +Output output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,100B, BPFP=0.5480 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,136B, BPFP=1.7435 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,524B, BPFP=0.7880 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,068B, BPFP=1.6855 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,940B, BPFP=0.8648 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,236B, BPFP=1.6404 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,652B, BPFP=0.7949 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,648B, BPFP=1.6628 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,756B, BPFP=1.3431 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,416B, BPFP=1.5959 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,076B, BPFP=0.2641 +⌛️ [2/4] FRONTEND: Frontend time: 0.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.483s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14134069 61.38089328 + layer.0.v_cache 0.00001405 0.00886013 + layer.1.k_cache 0.62039534 9.96760644 + layer.1.v_cache 0.00000582 0.00361321 + layer.2.k_cache 0.00676800 0.88824791 + layer.2.v_cache 0.00001893 0.01031157 + layer.3.k_cache 0.03896382 3.53809844 + layer.3.v_cache 0.00001942 0.01173831 + layer.4.k_cache 0.00069280 0.22581943 + layer.4.v_cache 0.00005238 0.02285550 + layer.4.output 0.00804578 191.72986421 + ------------------------------------------------------------------------------------- + TOTAL 0.05085834 83.42159375 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 267552 +BPFP 0.8539 bits/point +EBPFP 1.7077 equivalent bits/point +MSE 83.421594 +---------------------- -------------------------------------------------------- +Time: 0.997s Load: 0.011s, Pack+Encode: 0.503s, Decode+Unpack: 0.483s +---------------------- -------------------------------------------------------- +💾 Converting with 83.4216 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,080B, BPFP=0.5375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,140B, BPFP=1.7673 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,868B, BPFP=0.7929 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,156B, BPFP=1.7148 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,528B, BPFP=0.8814 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,324B, BPFP=1.6704 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,280B, BPFP=0.8148 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,700B, BPFP=1.6905 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,548B, BPFP=1.3624 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,380B, BPFP=1.6201 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,128B, BPFP=0.2676 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.391s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15682776 62.36226802 + layer.0.v_cache 0.00001501 0.00923145 + layer.1.k_cache 0.54447479 10.14097246 + layer.1.v_cache 0.00000580 0.00377834 + layer.2.k_cache 0.00853899 0.94479203 + layer.2.v_cache 0.00001972 0.01051916 + layer.3.k_cache 0.02401569 3.66529154 + layer.3.v_cache 0.00001939 0.01203249 + layer.4.k_cache 0.00071059 0.23598082 + layer.4.v_cache 0.00005329 0.02303223 + layer.4.output 0.05121508 184.93347757 + ------------------------------------------------------------------------------------- + TOTAL 0.06430510 80.70248480 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 276132 +BPFP 0.8662 bits/point +EBPFP 1.7324 equivalent bits/point +MSE 80.702485 +---------------------- -------------------------------------------------------- +Time: 0.658s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.391s +---------------------- -------------------------------------------------------- +💾 Converting with 80.7025 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,184B, BPFP=0.5412 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,036B, BPFP=1.7557 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,040B, BPFP=0.7993 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,196B, BPFP=1.7111 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,448B, BPFP=0.8741 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,172B, BPFP=1.6567 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,316B, BPFP=0.8140 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,712B, BPFP=1.6854 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,300B, BPFP=1.3446 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,176B, BPFP=1.6037 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,912B, BPFP=0.2727 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15363285 63.12748459 + layer.0.v_cache 0.00001412 0.00879614 + layer.1.k_cache 0.60852461 10.16911306 + layer.1.v_cache 0.00000580 0.00354741 + layer.2.k_cache 0.01182756 0.95175140 + layer.2.v_cache 0.00001863 0.00976945 + layer.3.k_cache 0.03467803 3.73563141 + layer.3.v_cache 0.00001919 0.01151890 + layer.4.k_cache 0.00072319 0.21897237 + layer.4.v_cache 0.00005301 0.02136227 + layer.4.output 0.05035121 184.41059281 + ------------------------------------------------------------------------------------- + TOTAL 0.06835032 80.53718216 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 276492 +BPFP 0.8644 bits/point +EBPFP 1.7288 equivalent bits/point +MSE 80.537182 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.011s, Pack+Encode: 0.258s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 80.5372 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,084B, BPFP=0.5568 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,148B, BPFP=1.7750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,516B, BPFP=0.8015 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,956B, BPFP=1.7091 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,112B, BPFP=0.8896 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,384B, BPFP=1.6776 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,632B, BPFP=0.8079 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,808B, BPFP=1.7010 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,728B, BPFP=1.3653 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,420B, BPFP=1.6243 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,648B, BPFP=0.2891 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12197856 66.15515956 + layer.0.v_cache 0.00001365 0.00876767 + layer.1.k_cache 0.62524759 10.00062459 + layer.1.v_cache 0.00000551 0.00344269 + layer.2.k_cache 0.01486346 0.97755944 + layer.2.v_cache 0.00001981 0.01033914 + layer.3.k_cache 0.07387851 3.80215034 + layer.3.v_cache 0.00001943 0.01164938 + layer.4.k_cache 0.00069560 0.22903887 + layer.4.v_cache 0.00005405 0.02313160 + layer.4.output 0.01098318 195.30989715 + ------------------------------------------------------------------------------------- + TOTAL 0.05374461 85.19947902 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 270436 +BPFP 0.8783 bits/point +EBPFP 1.7566 equivalent bits/point +MSE 85.199479 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1995 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,332B, BPFP=0.5664 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,972B, BPFP=1.7529 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,664B, BPFP=0.8039 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,036B, BPFP=1.7015 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,876B, BPFP=0.8704 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,136B, BPFP=1.6522 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,932B, BPFP=0.8186 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,500B, BPFP=1.6721 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,604B, BPFP=1.3489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,360B, BPFP=1.6096 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,984B, BPFP=0.2740 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15362121 62.67493147 + layer.0.v_cache 0.00001420 0.00912878 + layer.1.k_cache 0.52952420 9.87030393 + layer.1.v_cache 0.00000567 0.00359561 + layer.2.k_cache 0.00939641 0.94675711 + layer.2.v_cache 0.00001891 0.01007352 + layer.3.k_cache 0.04339437 3.97121582 + layer.3.v_cache 0.00002732 0.01167881 + layer.4.k_cache 0.00068326 0.22984769 + layer.4.v_cache 0.00005156 0.02238404 + layer.4.output 0.00776138 194.44910714 + ------------------------------------------------------------------------------------- + TOTAL 0.04653334 84.64080393 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 268396 +BPFP 0.8656 bits/point +EBPFP 1.7311 equivalent bits/point +MSE 84.640804 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 84.6408 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,280B, BPFP=0.5463 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,000B, BPFP=1.7538 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,992B, BPFP=0.7968 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,252B, BPFP=1.7141 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,532B, BPFP=0.8786 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,344B, BPFP=1.6658 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,280B, BPFP=0.8121 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,740B, BPFP=1.6869 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,560B, BPFP=1.3584 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,288B, BPFP=1.6097 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,836B, BPFP=0.2645 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14921317 65.10599357 + layer.0.v_cache 0.00001395 0.00920259 + layer.1.k_cache 0.64083395 9.98823973 + layer.1.v_cache 0.00000574 0.00375640 + layer.2.k_cache 0.00578804 0.91053419 + layer.2.v_cache 0.00001912 0.01065256 + layer.3.k_cache 0.02848778 3.73399426 + layer.3.v_cache 0.00001956 0.01202717 + layer.4.k_cache 0.00070939 0.22879386 + layer.4.v_cache 0.00005206 0.02299376 + layer.4.output 0.04916091 184.36585884 + ------------------------------------------------------------------------------------- + TOTAL 0.06878054 80.62277647 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 276104 +BPFP 0.8632 bits/point +EBPFP 1.7263 equivalent bits/point +MSE 80.622776 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 80.6228 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,136B, BPFP=0.5405 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,016B, BPFP=1.7607 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,852B, BPFP=0.7920 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,032B, BPFP=1.7082 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,380B, BPFP=0.8735 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,136B, BPFP=1.6604 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,024B, BPFP=0.8012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,656B, BPFP=1.6881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,344B, BPFP=1.3515 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,072B, BPFP=1.6037 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,972B, BPFP=0.2664 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12806672 62.81112681 + layer.0.v_cache 0.00001388 0.00877491 + layer.1.k_cache 0.56701790 10.11450695 + layer.1.v_cache 0.00000557 0.00344386 + layer.2.k_cache 0.01020011 0.98400421 + layer.2.v_cache 0.00001881 0.01004020 + layer.3.k_cache 0.04301871 4.00272471 + layer.3.v_cache 0.00002051 0.01172568 + layer.4.k_cache 0.00071755 0.22099348 + layer.4.v_cache 0.00005284 0.02124800 + layer.4.output 0.04940383 184.95086238 + ------------------------------------------------------------------------------------- + TOTAL 0.06440938 80.75556621 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 274620 +BPFP 0.8615 bits/point +EBPFP 1.7229 equivalent bits/point +MSE 80.755566 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 80.7556 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,228B, BPFP=0.5607 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,168B, BPFP=1.7636 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,704B, BPFP=0.8061 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,208B, BPFP=1.7110 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,092B, BPFP=0.8822 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,368B, BPFP=1.6649 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,000B, BPFP=0.8224 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,848B, BPFP=1.6912 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,812B, BPFP=1.3603 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,596B, BPFP=1.6226 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,448B, BPFP=0.2698 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12903612 62.92763158 + layer.0.v_cache 0.00001458 0.00900207 + layer.1.k_cache 0.59860808 10.09327628 + layer.1.v_cache 0.00000581 0.00358651 + layer.2.k_cache 0.00940504 0.96453547 + layer.2.v_cache 0.00001954 0.01050132 + layer.3.k_cache 0.02644065 3.85066946 + layer.3.v_cache 0.00002029 0.01190210 + layer.4.k_cache 0.00068693 0.22597753 + layer.4.v_cache 0.00005376 0.02238664 + layer.4.output 0.00862506 194.33353697 + ------------------------------------------------------------------------------------- + TOTAL 0.04850978 84.61495457 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 269472 +BPFP 0.8690 bits/point +EBPFP 1.7381 equivalent bits/point +MSE 84.614955 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 84.6150 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,692B, BPFP=0.5424 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,876B, BPFP=1.7185 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,588B, BPFP=0.7908 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,976B, BPFP=1.6729 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,800B, BPFP=0.8523 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,164B, BPFP=1.6317 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,680B, BPFP=0.7955 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,444B, BPFP=1.6459 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,984B, BPFP=1.3182 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,880B, BPFP=1.5666 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,524B, BPFP=0.3009 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16469599 61.68181818 + layer.0.v_cache 0.00001357 0.00877371 + layer.1.k_cache 0.70445866 9.83689028 + layer.1.v_cache 0.00000558 0.00357867 + layer.2.k_cache 0.01447187 0.98950701 + layer.2.v_cache 0.00001924 0.01011506 + layer.3.k_cache 0.02617162 3.90419977 + layer.3.v_cache 0.00001974 0.01159359 + layer.4.k_cache 0.00070375 0.22144823 + layer.4.v_cache 0.00005076 0.02195703 + layer.4.output 0.04814868 175.62521742 + ------------------------------------------------------------------------------------- + TOTAL 0.07339127 76.82743550 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 288608 +BPFP 0.8612 bits/point +EBPFP 1.7225 equivalent bits/point +MSE 76.827435 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.390s +---------------------- -------------------------------------------------------- +💾 Converting with 76.8274 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,752B, BPFP=0.5664 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,424B, BPFP=1.8253 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,056B, BPFP=0.8164 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,468B, BPFP=1.7697 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,192B, BPFP=0.8824 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,540B, BPFP=1.7158 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,968B, BPFP=0.8113 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,820B, BPFP=1.7321 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,976B, BPFP=1.3927 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,372B, BPFP=1.6480 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,508B, BPFP=0.2697 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12132761 67.61028259 + layer.0.v_cache 0.00001366 0.00867403 + layer.1.k_cache 0.57241889 10.07015640 + layer.1.v_cache 0.00000556 0.00340114 + layer.2.k_cache 0.00926234 0.97684646 + layer.2.v_cache 0.00002145 0.00988571 + layer.3.k_cache 0.03109839 4.04657913 + layer.3.v_cache 0.00002257 0.01116815 + layer.4.k_cache 0.00070701 0.22157312 + layer.4.v_cache 0.00004920 0.02119938 + layer.4.output 0.01031505 205.74196760 + ------------------------------------------------------------------------------------- + TOTAL 0.04747836 89.59844349 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 259076 +BPFP 0.8852 bits/point +EBPFP 1.7704 equivalent bits/point +MSE 89.598443 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 89.5984 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,352B, BPFP=0.5428 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,436B, BPFP=1.7531 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,088B, BPFP=0.7911 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,332B, BPFP=1.6953 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,772B, BPFP=0.8794 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,592B, BPFP=1.6565 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,328B, BPFP=0.8037 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,824B, BPFP=1.6686 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,656B, BPFP=1.3452 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,332B, BPFP=1.5904 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,092B, BPFP=0.2703 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14357947 64.88712510 + layer.0.v_cache 0.00001420 0.00940476 + layer.1.k_cache 0.63723908 10.10585544 + layer.1.v_cache 0.00000573 0.00371235 + layer.2.k_cache 0.00765126 0.98106763 + layer.2.v_cache 0.00002073 0.01030319 + layer.3.k_cache 0.02843528 4.25118711 + layer.3.v_cache 0.00002112 0.01203405 + layer.4.k_cache 0.00069171 0.22720782 + layer.4.v_cache 0.00005234 0.02247741 + layer.4.output 0.04871205 181.73116910 + ------------------------------------------------------------------------------------- + TOTAL 0.06815855 79.56638580 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 278804 +BPFP 0.8599 bits/point +EBPFP 1.7198 equivalent bits/point +MSE 79.566386 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 79.5664 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,800B, BPFP=0.5533 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,840B, BPFP=1.7336 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,500B, BPFP=0.7941 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,852B, BPFP=1.6830 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,980B, BPFP=0.8699 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,076B, BPFP=1.6432 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,640B, BPFP=0.8012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,396B, BPFP=1.6596 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,172B, BPFP=1.3408 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,908B, BPFP=1.5834 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,352B, BPFP=0.2734 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16857396 61.47384734 + layer.0.v_cache 0.00001478 0.00920118 + layer.1.k_cache 0.58716421 9.89715836 + layer.1.v_cache 0.00000600 0.00377758 + layer.2.k_cache 0.00827454 0.95362359 + layer.2.v_cache 0.00002177 0.01056335 + layer.3.k_cache 0.02531170 3.59513120 + layer.3.v_cache 0.00001983 0.01178400 + layer.4.k_cache 0.00069463 0.22117552 + layer.4.v_cache 0.00005316 0.02269458 + layer.4.output 0.04836898 177.50199063 + ------------------------------------------------------------------------------------- + TOTAL 0.06639514 77.57134654 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 284516 +BPFP 0.8574 bits/point +EBPFP 1.7148 equivalent bits/point +MSE 77.571347 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 77.5713 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,588B, BPFP=0.5552 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,436B, BPFP=1.7531 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,220B, BPFP=0.7980 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,488B, BPFP=1.7034 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,644B, BPFP=0.8727 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,588B, BPFP=1.6562 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,312B, BPFP=0.8029 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,960B, BPFP=1.6758 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,724B, BPFP=1.3488 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,536B, BPFP=1.6011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,672B, BPFP=0.2597 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14028978 64.19666920 + layer.0.v_cache 0.00001397 0.00897370 + layer.1.k_cache 0.59966084 10.30486200 + layer.1.v_cache 0.00000578 0.00373704 + layer.2.k_cache 0.01765286 0.95040607 + layer.2.v_cache 0.00001884 0.01045282 + layer.3.k_cache 0.02653811 3.76928916 + layer.3.v_cache 0.00001935 0.01179290 + layer.4.k_cache 0.00071527 0.23220329 + layer.4.v_cache 0.00005402 0.02289904 + layer.4.output 0.04907703 181.66672160 + ------------------------------------------------------------------------------------- + TOTAL 0.06638282 79.48107861 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 278168 +BPFP 0.8580 bits/point +EBPFP 1.7159 equivalent bits/point +MSE 79.481079 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 79.4811 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,876B, BPFP=0.5673 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,508B, BPFP=1.8100 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,004B, BPFP=0.8045 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,624B, BPFP=1.7592 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,520B, BPFP=0.8915 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,928B, BPFP=1.7192 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,400B, BPFP=0.8272 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,372B, BPFP=1.7447 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,088B, BPFP=1.3837 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,712B, BPFP=1.6494 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,148B, BPFP=0.2802 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13255837 62.70384306 + layer.0.v_cache 0.00001394 0.00847206 + layer.1.k_cache 0.54921044 10.19543008 + layer.1.v_cache 0.00000578 0.00338196 + layer.2.k_cache 0.01292971 0.98997845 + layer.2.v_cache 0.00001901 0.01007435 + layer.3.k_cache 0.07190662 4.00625521 + layer.3.v_cache 0.00001944 0.01113434 + layer.4.k_cache 0.00068159 0.20954042 + layer.4.v_cache 0.00005125 0.02120697 + layer.4.output 0.00790709 203.49015231 + ------------------------------------------------------------------------------------- + TOTAL 0.04839681 88.38766959 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 263180 +BPFP 0.8893 bits/point +EBPFP 1.7786 equivalent bits/point +MSE 88.387670 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.012s, Pack+Encode: 0.251s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 88.3877 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,112B, BPFP=0.5603 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,232B, BPFP=1.7859 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,528B, BPFP=0.8050 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,104B, BPFP=1.7234 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,020B, BPFP=0.8876 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,264B, BPFP=1.6769 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,628B, BPFP=0.8105 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,884B, BPFP=1.7112 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,808B, BPFP=1.3746 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,564B, BPFP=1.6381 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,452B, BPFP=0.2964 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13695534 69.55549091 + layer.0.v_cache 0.00001452 0.00924265 + layer.1.k_cache 0.52272531 9.97035942 + layer.1.v_cache 0.00000574 0.00367600 + layer.2.k_cache 0.00793556 0.99273454 + layer.2.v_cache 0.00001907 0.01056839 + layer.3.k_cache 0.04682703 3.64964998 + layer.3.v_cache 0.00002008 0.01185999 + layer.4.k_cache 0.00068945 0.23033946 + layer.4.v_cache 0.00005221 0.02317283 + layer.4.output 0.00941004 196.10115881 + ------------------------------------------------------------------------------------- + TOTAL 0.04594792 85.71560035 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 271596 +BPFP 0.8852 bits/point +EBPFP 1.7704 equivalent bits/point +MSE 85.715600 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.009s, Pack+Encode: 0.254s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7156 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,088B, BPFP=0.5629 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,020B, BPFP=1.7868 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,208B, BPFP=0.7929 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,884B, BPFP=1.7234 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,608B, BPFP=0.8710 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,976B, BPFP=1.6728 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,572B, BPFP=0.8132 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,480B, BPFP=1.7009 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,152B, BPFP=1.3478 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,044B, BPFP=1.6208 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,828B, BPFP=0.3095 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14340984 63.72208426 + layer.0.v_cache 0.00001397 0.00890906 + layer.1.k_cache 0.51622396 10.01709072 + layer.1.v_cache 0.00000557 0.00354678 + layer.2.k_cache 0.00797994 0.98287310 + layer.2.v_cache 0.00001813 0.00989095 + layer.3.k_cache 0.02756977 3.61326643 + layer.3.v_cache 0.00001890 0.01142432 + layer.4.k_cache 0.00068956 0.22307238 + layer.4.v_cache 0.00005177 0.02211689 + layer.4.output 0.00947078 197.53372130 + ------------------------------------------------------------------------------------- + TOTAL 0.04483982 85.96178376 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 269860 +BPFP 0.8858 bits/point +EBPFP 1.7717 equivalent bits/point +MSE 85.961784 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.009s, Pack+Encode: 0.249s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 85.9618 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,124B, BPFP=0.5381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,948B, BPFP=1.7511 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,948B, BPFP=0.7944 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,016B, BPFP=1.7015 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,352B, BPFP=0.8690 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,220B, BPFP=1.6592 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,168B, BPFP=0.8061 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,564B, BPFP=1.6775 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,532B, BPFP=1.3569 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,196B, BPFP=1.6048 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,644B, BPFP=0.2706 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13622742 63.21275776 + layer.0.v_cache 0.00001380 0.00903229 + layer.1.k_cache 0.61520739 9.95270896 + layer.1.v_cache 0.00000584 0.00375880 + layer.2.k_cache 0.00900942 1.01624449 + layer.2.v_cache 0.00001916 0.01046958 + layer.3.k_cache 0.02667915 3.60603301 + layer.3.v_cache 0.00001881 0.01173152 + layer.4.k_cache 0.00071466 0.22618290 + layer.4.v_cache 0.00005760 0.02276380 + layer.4.output 0.04972813 184.33003827 + ------------------------------------------------------------------------------------- + TOTAL 0.06682648 80.49305594 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 275712 +BPFP 0.8619 bits/point +EBPFP 1.7239 equivalent bits/point +MSE 80.493056 +---------------------- -------------------------------------------------------- +Time: 0.653s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 80.4931 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,624B, BPFP=0.5552 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,500B, BPFP=1.7506 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,268B, BPFP=0.7979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,436B, BPFP=1.6950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,036B, BPFP=0.8903 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,712B, BPFP=1.6572 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,612B, BPFP=0.8158 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,152B, BPFP=1.6802 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,844B, BPFP=1.3505 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,696B, BPFP=1.6041 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,168B, BPFP=0.2775 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11478914 63.15439486 + layer.0.v_cache 0.00001377 0.00851444 + layer.1.k_cache 0.64218915 9.82653115 + layer.1.v_cache 0.00000556 0.00335681 + layer.2.k_cache 0.00898698 0.97525606 + layer.2.v_cache 0.00001980 0.01003817 + layer.3.k_cache 0.05442990 3.82462701 + layer.3.v_cache 0.00001964 0.01161237 + layer.4.k_cache 0.00070461 0.22010196 + layer.4.v_cache 0.00005015 0.02157072 + layer.4.output 0.05017195 179.75964525 + ------------------------------------------------------------------------------------- + TOTAL 0.06896543 78.61020707 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 282048 +BPFP 0.8670 bits/point +EBPFP 1.7340 equivalent bits/point +MSE 78.610207 +---------------------- -------------------------------------------------------- +Time: 0.656s Load: 0.012s, Pack+Encode: 0.257s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 78.6102 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,800B, BPFP=0.5189 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,060B, BPFP=1.8314 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,824B, BPFP=0.8151 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,008B, BPFP=1.7693 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,964B, BPFP=0.8823 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,996B, BPFP=1.7097 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,056B, BPFP=0.8288 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,560B, BPFP=1.7429 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,688B, BPFP=1.3967 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,048B, BPFP=1.6538 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,624B, BPFP=0.2495 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14755792 65.10057488 + layer.0.v_cache 0.00001375 0.00919014 + layer.1.k_cache 0.51940866 10.21212872 + layer.1.v_cache 0.00000586 0.00384270 + layer.2.k_cache 0.00816260 1.03057032 + layer.2.v_cache 0.00001900 0.01075159 + layer.3.k_cache 0.02921419 4.10186837 + layer.3.v_cache 0.00001979 0.01233799 + layer.4.k_cache 0.00071017 0.23304036 + layer.4.v_cache 0.00005507 0.02298548 + layer.4.output 0.00937099 209.12680256 + ------------------------------------------------------------------------------------- + TOTAL 0.04533906 90.86028873 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 252628 +BPFP 0.8762 bits/point +EBPFP 1.7524 equivalent bits/point +MSE 90.860289 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.011s, Pack+Encode: 0.256s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 90.8603 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,300B, BPFP=0.5442 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,268B, BPFP=1.8298 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,120B, BPFP=0.8263 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,536B, BPFP=1.7870 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,192B, BPFP=0.8890 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,820B, BPFP=1.7451 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,232B, BPFP=0.8329 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,932B, BPFP=1.7516 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,988B, BPFP=1.4038 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,556B, BPFP=1.6711 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,488B, BPFP=0.2632 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12840328 64.75630559 + layer.0.v_cache 0.00001372 0.00893602 + layer.1.k_cache 0.60081110 10.01985220 + layer.1.v_cache 0.00000562 0.00352582 + layer.2.k_cache 0.00586274 0.98503770 + layer.2.v_cache 0.00001839 0.01023546 + layer.3.k_cache 0.02491910 4.10367153 + layer.3.v_cache 0.00001957 0.01188751 + layer.4.k_cache 0.00072047 0.23244635 + layer.4.v_cache 0.00005229 0.02319663 + layer.4.output 0.01050496 207.23558721 + ------------------------------------------------------------------------------------- + TOTAL 0.04908006 90.04730619 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 258432 +BPFP 0.8896 bits/point +EBPFP 1.7792 equivalent bits/point +MSE 90.047306 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 90.0473 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,964B, BPFP=0.5501 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,904B, BPFP=1.7615 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,556B, BPFP=0.8037 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,800B, BPFP=1.7005 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,608B, BPFP=0.8617 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,072B, BPFP=1.6603 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,436B, BPFP=0.7970 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,444B, BPFP=1.6809 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,416B, BPFP=1.3481 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,088B, BPFP=1.6060 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,784B, BPFP=0.2822 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14595435 65.75928252 + layer.0.v_cache 0.00001364 0.00850721 + layer.1.k_cache 0.59759904 9.92578470 + layer.1.v_cache 0.00000566 0.00341215 + layer.2.k_cache 0.01409306 0.95738668 + layer.2.v_cache 0.00002047 0.01003431 + layer.3.k_cache 0.06079736 3.80992790 + layer.3.v_cache 0.00002215 0.01160902 + layer.4.k_cache 0.00070776 0.21869527 + layer.4.v_cache 0.00005221 0.02252737 + layer.4.output 0.00769303 195.46815056 + ------------------------------------------------------------------------------------- + TOTAL 0.05135982 85.23554241 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 267072 +BPFP 0.8674 bits/point +EBPFP 1.7348 equivalent bits/point +MSE 85.235542 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.011s, Pack+Encode: 0.256s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2355 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 315, 128) +Output shape: (1, 315, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.output: torch.Size([1, 315, 3584]) -> torch.Size([1, 1, 315, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,924B, BPFP=0.5419 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,148B, BPFP=1.6938 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,616B, BPFP=0.7746 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,252B, BPFP=1.6494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,084B, BPFP=0.8474 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,392B, BPFP=1.6067 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,964B, BPFP=0.7919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,660B, BPFP=1.6200 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,488B, BPFP=1.3139 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,276B, BPFP=1.5514 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,316B, BPFP=0.2857 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15500355 65.59082341 + layer.0.v_cache 0.00001366 0.00905599 + layer.1.k_cache 0.68167124 9.70267702 + layer.1.v_cache 0.00000571 0.00370952 + layer.2.k_cache 0.01007247 0.93743480 + layer.2.v_cache 0.00001936 0.01046548 + layer.3.k_cache 0.03272899 3.42169790 + layer.3.v_cache 0.00002006 0.01171837 + layer.4.k_cache 0.00071175 0.22669835 + layer.4.v_cache 0.00005427 0.02260124 + layer.4.output 0.04584261 171.35172902 + ------------------------------------------------------------------------------------- + TOTAL 0.07065879 75.25876384 + (elements=2,741,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2741760 +Total Bytes 290120 +BPFP 0.8465 bits/point +EBPFP 1.6930 equivalent bits/point +MSE 75.258764 +---------------------- -------------------------------------------------------- +Time: 0.657s Load: 0.012s, Pack+Encode: 0.257s, Decode+Unpack: 0.388s +---------------------- -------------------------------------------------------- +💾 Converting with 75.2588 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,892B, BPFP=0.5703 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,328B, BPFP=1.8063 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,044B, BPFP=0.8097 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,232B, BPFP=1.7431 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,216B, BPFP=0.8773 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,448B, BPFP=1.6979 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,044B, BPFP=0.8097 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,008B, BPFP=1.7302 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,888B, BPFP=1.3773 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,456B, BPFP=1.6407 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,060B, BPFP=0.2723 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12889319 62.55318122 + layer.0.v_cache 0.00001350 0.00908170 + layer.1.k_cache 0.58775566 9.99366045 + layer.1.v_cache 0.00000553 0.00352807 + layer.2.k_cache 0.00723753 0.99393883 + layer.2.v_cache 0.00001924 0.01012363 + layer.3.k_cache 0.01761036 3.97793365 + layer.3.v_cache 0.00001847 0.01146982 + layer.4.k_cache 0.00071814 0.22440707 + layer.4.v_cache 0.00005270 0.02234056 + layer.4.output 0.01035140 204.20690564 + ------------------------------------------------------------------------------------- + TOTAL 0.04792848 88.66164732 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 259616 +BPFP 0.8805 bits/point +EBPFP 1.7610 equivalent bits/point +MSE 88.661647 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6616 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,136B, BPFP=0.5577 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,936B, BPFP=1.7570 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,608B, BPFP=0.8037 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,924B, BPFP=1.7014 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,712B, BPFP=0.8644 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,972B, BPFP=1.6490 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,484B, BPFP=0.7969 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,348B, BPFP=1.6697 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,524B, BPFP=1.3493 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,000B, BPFP=1.5955 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,932B, BPFP=0.2746 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13033964 65.27070037 + layer.0.v_cache 0.00001368 0.00891617 + layer.1.k_cache 0.63029007 9.73128370 + layer.1.v_cache 0.00000551 0.00352790 + layer.2.k_cache 0.00637632 0.97560829 + layer.2.v_cache 0.00001960 0.01000550 + layer.3.k_cache 0.02368569 4.25462449 + layer.3.v_cache 0.00001974 0.01099515 + layer.4.k_cache 0.00068050 0.22547488 + layer.4.v_cache 0.00005209 0.02182850 + layer.4.output 0.00765416 194.71610915 + ------------------------------------------------------------------------------------- + TOTAL 0.04970952 84.91327818 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 266576 +BPFP 0.8627 bits/point +EBPFP 1.7255 equivalent bits/point +MSE 84.913278 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.012s, Pack+Encode: 0.248s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9133 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,028B, BPFP=0.5761 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,696B, BPFP=1.8208 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,948B, BPFP=0.8012 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,552B, BPFP=1.7551 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,392B, BPFP=0.8842 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,928B, BPFP=1.7192 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,228B, BPFP=0.8173 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,484B, BPFP=1.7511 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,208B, BPFP=1.3906 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,920B, BPFP=1.6613 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,564B, BPFP=0.2754 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.396s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15644295 66.23553108 + layer.0.v_cache 0.00001405 0.00876112 + layer.1.k_cache 0.56163917 10.13712804 + layer.1.v_cache 0.00000554 0.00365861 + layer.2.k_cache 0.01059573 0.94111027 + layer.2.v_cache 0.00001807 0.01026854 + layer.3.k_cache 0.04944291 4.02628326 + layer.3.v_cache 0.00001950 0.01178257 + layer.4.k_cache 0.00071267 0.22965278 + layer.4.v_cache 0.00005817 0.02342513 + layer.4.output 0.00956290 203.51705291 + ------------------------------------------------------------------------------------- + TOTAL 0.04975818 88.60276305 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 262948 +BPFP 0.8885 bits/point +EBPFP 1.7771 equivalent bits/point +MSE 88.602763 +---------------------- -------------------------------------------------------- +Time: 0.661s Load: 0.009s, Pack+Encode: 0.255s, Decode+Unpack: 0.396s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6028 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 312, 128) +Output shape: (1, 312, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.output: torch.Size([1, 312, 3584]) -> torch.Size([1, 1, 312, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,972B, BPFP=0.5495 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,252B, BPFP=1.7153 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,624B, BPFP=0.7825 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,256B, BPFP=1.6655 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,172B, BPFP=0.8600 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,448B, BPFP=1.6250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,996B, BPFP=0.8011 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,788B, BPFP=1.6420 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,308B, BPFP=1.3175 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,308B, BPFP=1.5679 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,216B, BPFP=0.2877 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14984696 62.79483423 + layer.0.v_cache 0.00001474 0.00898928 + layer.1.k_cache 0.67859346 10.15568347 + layer.1.v_cache 0.00000581 0.00364932 + layer.2.k_cache 0.01626646 0.99484165 + layer.2.v_cache 0.00001951 0.01029378 + layer.3.k_cache 0.02914289 3.72808603 + layer.3.v_cache 0.00001919 0.01136902 + layer.4.k_cache 0.00071665 0.22658441 + layer.4.v_cache 0.00005096 0.02187705 + layer.4.output 0.04642455 173.59882956 + ------------------------------------------------------------------------------------- + TOTAL 0.07056756 76.06753030 + (elements=2,715,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2715648 +Total Bytes 290340 +BPFP 0.8553 bits/point +EBPFP 1.7106 equivalent bits/point +MSE 76.067530 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 76.0675 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,580B, BPFP=0.5565 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,416B, BPFP=1.8248 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,980B, BPFP=0.8120 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,516B, BPFP=1.7725 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,456B, BPFP=0.8978 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,680B, BPFP=1.7240 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,044B, BPFP=0.8158 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,944B, BPFP=1.7393 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,004B, BPFP=1.3943 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,552B, BPFP=1.6585 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,172B, BPFP=0.2670 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15823836 65.30564591 + layer.0.v_cache 0.00001502 0.00879737 + layer.1.k_cache 0.53668020 10.34926358 + layer.1.v_cache 0.00000576 0.00347757 + layer.2.k_cache 0.01018995 1.02420339 + layer.2.v_cache 0.00002218 0.01020856 + layer.3.k_cache 0.05410197 3.84963070 + layer.3.v_cache 0.00002001 0.01156944 + layer.4.k_cache 0.00069143 0.22502030 + layer.4.v_cache 0.00005304 0.02263680 + layer.4.output 0.01041225 205.89463290 + ------------------------------------------------------------------------------------- + TOTAL 0.04899433 89.53369905 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 259344 +BPFP 0.8861 bits/point +EBPFP 1.7723 equivalent bits/point +MSE 89.533699 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 89.5337 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 329, 128) +Output shape: (1, 329, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.output: torch.Size([1, 329, 3584]) -> torch.Size([1, 1, 329, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,012B, BPFP=0.5230 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,976B, BPFP=1.8036 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,976B, BPFP=0.8062 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,640B, BPFP=1.7401 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,428B, BPFP=0.8752 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,588B, BPFP=1.6902 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,028B, BPFP=0.8087 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,140B, BPFP=1.7164 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,764B, BPFP=1.3661 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,512B, BPFP=1.6391 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,064B, BPFP=0.2990 +⌛️ [2/4] FRONTEND: Frontend time: 0.341s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.436s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11391156 63.04946927 + layer.0.v_cache 0.00001522 0.00892637 + layer.1.k_cache 0.63933222 10.17758522 + layer.1.v_cache 0.00000652 0.00384574 + layer.2.k_cache 0.01643764 0.92529733 + layer.2.v_cache 0.00001873 0.01007304 + layer.3.k_cache 0.02826017 3.68551905 + layer.3.v_cache 0.00001876 0.01141922 + layer.4.k_cache 0.00074607 0.23325327 + layer.4.v_cache 0.00005483 0.02278723 + layer.4.output 3.44632065 163.13337223 + ------------------------------------------------------------------------------------- + TOTAL 1.46606155 71.76834008 + (elements=2,863,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2863616 +Total Bytes 317128 +BPFP 0.8860 bits/point +EBPFP 1.7719 equivalent bits/point +MSE 71.768340 +---------------------- -------------------------------------------------------- +Time: 0.788s Load: 0.011s, Pack+Encode: 0.341s, Decode+Unpack: 0.436s +---------------------- -------------------------------------------------------- +💾 Converting with 71.7683 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,084B, BPFP=0.5647 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,012B, BPFP=1.7928 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,216B, BPFP=0.7961 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,004B, BPFP=1.7363 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,680B, BPFP=0.8781 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,344B, BPFP=1.6994 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,444B, BPFP=0.8089 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,608B, BPFP=1.7142 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,488B, BPFP=1.3714 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,072B, BPFP=1.6281 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,896B, BPFP=0.2872 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13536828 63.24486867 + layer.0.v_cache 0.00001433 0.00884780 + layer.1.k_cache 0.57185238 9.71170580 + layer.1.v_cache 0.00000558 0.00354844 + layer.2.k_cache 0.00766427 0.98717234 + layer.2.v_cache 0.00001891 0.01027126 + layer.3.k_cache 0.04143807 3.95596237 + layer.3.v_cache 0.00001947 0.01180647 + layer.4.k_cache 0.00070404 0.22115515 + layer.4.v_cache 0.00005498 0.02276895 + layer.4.output 0.00948056 198.48991935 + ------------------------------------------------------------------------------------- + TOTAL 0.04844142 86.32985545 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 267848 +BPFP 0.8824 bits/point +EBPFP 1.7648 equivalent bits/point +MSE 86.329855 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.011s, Pack+Encode: 0.253s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 86.3299 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,060B, BPFP=0.5695 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,624B, BPFP=1.7903 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,232B, BPFP=0.8057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,764B, BPFP=1.7416 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,392B, BPFP=0.8714 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,100B, BPFP=1.7040 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,220B, BPFP=0.8050 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,344B, BPFP=1.7178 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,120B, BPFP=1.3655 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,912B, BPFP=1.6368 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,048B, BPFP=0.2915 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15132187 62.56392946 + layer.0.v_cache 0.00001456 0.00892592 + layer.1.k_cache 0.61798344 10.11511850 + layer.1.v_cache 0.00000569 0.00349103 + layer.2.k_cache 0.01686249 0.97367837 + layer.2.v_cache 0.00001898 0.01041582 + layer.3.k_cache 0.04401700 3.63145336 + layer.3.v_cache 0.00001935 0.01137314 + layer.4.k_cache 0.00071189 0.22500812 + layer.4.v_cache 0.00005111 0.02221664 + layer.4.output 0.01058244 200.92729361 + ------------------------------------------------------------------------------------- + TOTAL 0.05324020 87.29745092 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 265816 +BPFP 0.8852 bits/point +EBPFP 1.7704 equivalent bits/point +MSE 87.297451 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 87.2975 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,880B, BPFP=0.5593 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,848B, BPFP=1.8030 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,132B, BPFP=0.8000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,852B, BPFP=1.7466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,664B, BPFP=0.8868 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,040B, BPFP=1.7006 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,340B, BPFP=0.8118 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,584B, BPFP=1.7314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,376B, BPFP=1.3800 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,932B, BPFP=1.6379 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,852B, BPFP=0.2900 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12581717 64.21124887 + layer.0.v_cache 0.00001375 0.00928640 + layer.1.k_cache 0.57814745 9.69330276 + layer.1.v_cache 0.00000592 0.00369452 + layer.2.k_cache 0.00733882 0.91886321 + layer.2.v_cache 0.00002058 0.01075320 + layer.3.k_cache 0.02353939 4.02850740 + layer.3.v_cache 0.00002077 0.01205375 + layer.4.k_cache 0.00069182 0.23172970 + layer.4.v_cache 0.00005027 0.02356599 + layer.4.output 0.01015823 200.86353196 + ------------------------------------------------------------------------------------- + TOTAL 0.04745609 87.36398409 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 266500 +BPFP 0.8875 bits/point +EBPFP 1.7750 equivalent bits/point +MSE 87.363984 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 87.3640 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,028B, BPFP=0.5719 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,564B, BPFP=1.8000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,112B, BPFP=0.8047 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,860B, BPFP=1.7598 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,640B, BPFP=0.8919 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,964B, BPFP=1.7087 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,480B, BPFP=0.8257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,440B, BPFP=1.7359 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,376B, BPFP=1.3901 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,960B, BPFP=1.6515 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,948B, BPFP=0.2929 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13625057 62.75646527 + layer.0.v_cache 0.00001362 0.00855560 + layer.1.k_cache 0.57512581 10.03656763 + layer.1.v_cache 0.00000584 0.00353822 + layer.2.k_cache 0.00872732 1.03877904 + layer.2.v_cache 0.00001851 0.01038089 + layer.3.k_cache 0.04362779 3.85105640 + layer.3.v_cache 0.00001925 0.01194702 + layer.4.k_cache 0.00073619 0.22537216 + layer.4.v_cache 0.00005148 0.02215712 + layer.4.output 0.00783937 202.35893509 + ------------------------------------------------------------------------------------- + TOTAL 0.04820306 87.91043323 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 266372 +BPFP 0.8935 bits/point +EBPFP 1.7871 equivalent bits/point +MSE 87.910433 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 87.9104 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,696B, BPFP=0.5498 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,800B, BPFP=1.7373 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,556B, BPFP=0.7995 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,772B, BPFP=1.6844 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,876B, BPFP=0.8674 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,920B, BPFP=1.6406 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,748B, BPFP=0.8094 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,340B, BPFP=1.6622 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,840B, BPFP=1.3281 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,788B, BPFP=1.5824 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,316B, BPFP=0.2960 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16730670 65.01350483 + layer.0.v_cache 0.00001388 0.00875748 + layer.1.k_cache 0.64158324 10.24600702 + layer.1.v_cache 0.00000571 0.00341745 + layer.2.k_cache 0.02207795 0.91666914 + layer.2.v_cache 0.00001922 0.00972230 + layer.3.k_cache 0.04503629 3.65168039 + layer.3.v_cache 0.00001955 0.01121444 + layer.4.k_cache 0.00070478 0.21265058 + layer.4.v_cache 0.00005260 0.02128469 + layer.4.output 0.04763387 177.85548344 + ------------------------------------------------------------------------------------- + TOTAL 0.07119159 77.94607602 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 286652 +BPFP 0.8667 bits/point +EBPFP 1.7333 equivalent bits/point +MSE 77.946076 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 77.9461 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,020B, BPFP=0.5652 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,632B, BPFP=1.7843 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,380B, BPFP=0.8111 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,684B, BPFP=1.7308 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,272B, BPFP=0.8615 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,736B, BPFP=1.6773 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,184B, BPFP=0.8001 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,272B, BPFP=1.7076 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,984B, BPFP=1.3529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,716B, BPFP=1.6198 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,644B, BPFP=0.3114 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.391s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15225300 64.93266302 + layer.0.v_cache 0.00001427 0.00846409 + layer.1.k_cache 0.52731114 10.05336015 + layer.1.v_cache 0.00000556 0.00349344 + layer.2.k_cache 0.01483519 0.96956097 + layer.2.v_cache 0.00001864 0.01011267 + layer.3.k_cache 0.04699480 3.81019036 + layer.3.v_cache 0.00001968 0.01123002 + layer.4.k_cache 0.00070142 0.22288139 + layer.4.v_cache 0.00005177 0.02191213 + layer.4.output 0.01077793 200.05416774 + ------------------------------------------------------------------------------------- + TOTAL 0.04809712 87.08370838 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 267524 +BPFP 0.8877 bits/point +EBPFP 1.7754 equivalent bits/point +MSE 87.083708 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.391s +---------------------- -------------------------------------------------------- +💾 Converting with 87.0837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,000B, BPFP=0.5641 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,788B, BPFP=1.7931 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,336B, BPFP=0.8087 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,808B, BPFP=1.7378 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,672B, BPFP=0.8840 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,252B, BPFP=1.7065 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,332B, BPFP=0.8084 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,456B, BPFP=1.7180 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,152B, BPFP=1.3624 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,908B, BPFP=1.6306 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,652B, BPFP=0.2954 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15700974 63.32142656 + layer.0.v_cache 0.00001374 0.00851825 + layer.1.k_cache 0.52255304 9.89442549 + layer.1.v_cache 0.00000566 0.00344886 + layer.2.k_cache 0.01360668 0.99421863 + layer.2.v_cache 0.00001926 0.01031415 + layer.3.k_cache 0.08085050 3.81454303 + layer.3.v_cache 0.00001943 0.01134040 + layer.4.k_cache 0.00071029 0.21685832 + layer.4.v_cache 0.00005450 0.02166262 + layer.4.output 0.01014708 199.89749871 + ------------------------------------------------------------------------------------- + TOTAL 0.04975720 86.91642631 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 267356 +BPFP 0.8871 bits/point +EBPFP 1.7742 equivalent bits/point +MSE 86.916426 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9164 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,336B, BPFP=0.5667 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,124B, BPFP=1.7612 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,672B, BPFP=0.8044 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,192B, BPFP=1.7101 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,068B, BPFP=0.8809 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,288B, BPFP=1.6605 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,700B, BPFP=0.8059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,820B, BPFP=1.6897 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,708B, BPFP=1.3546 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,296B, BPFP=1.6061 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,372B, BPFP=0.2770 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16436556 64.46936678 + layer.0.v_cache 0.00001441 0.00883921 + layer.1.k_cache 0.62741763 9.98109152 + layer.1.v_cache 0.00000560 0.00360528 + layer.2.k_cache 0.01487184 0.93607338 + layer.2.v_cache 0.00002167 0.01042667 + layer.3.k_cache 0.02142572 4.12687132 + layer.3.v_cache 0.00001943 0.01156743 + layer.4.k_cache 0.00070776 0.22185586 + layer.4.v_cache 0.00005251 0.02231054 + layer.4.output 0.00894435 194.56087093 + ------------------------------------------------------------------------------------- + TOTAL 0.05244192 84.80694732 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 269576 +BPFP 0.8694 bits/point +EBPFP 1.7388 equivalent bits/point +MSE 84.806947 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 84.8069 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,232B, BPFP=0.5690 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,948B, BPFP=1.7765 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,640B, BPFP=0.8141 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,800B, BPFP=1.7126 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,616B, BPFP=0.8683 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,080B, BPFP=1.6726 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,436B, BPFP=0.8027 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,500B, BPFP=1.6960 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,284B, BPFP=1.3503 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,992B, BPFP=1.6121 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,096B, BPFP=0.2867 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11705062 63.57103536 + layer.0.v_cache 0.00001445 0.00887119 + layer.1.k_cache 0.56620099 9.99104499 + layer.1.v_cache 0.00000566 0.00353224 + layer.2.k_cache 0.01299965 0.99449044 + layer.2.v_cache 0.00001972 0.00989303 + layer.3.k_cache 0.05069733 3.84103926 + layer.3.v_cache 0.00001944 0.01144440 + layer.4.k_cache 0.00068883 0.21934083 + layer.4.v_cache 0.00005144 0.02146857 + layer.4.output 0.00838925 197.09829372 + ------------------------------------------------------------------------------------- + TOTAL 0.04743958 85.78589508 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 267624 +BPFP 0.8754 bits/point +EBPFP 1.7507 equivalent bits/point +MSE 85.785895 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7859 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,068B, BPFP=0.5559 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,984B, BPFP=1.7659 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,592B, BPFP=0.8057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,864B, BPFP=1.7041 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,820B, BPFP=0.8735 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,032B, BPFP=1.6581 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,516B, BPFP=0.8015 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,604B, BPFP=1.6897 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,524B, BPFP=1.3540 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,316B, BPFP=1.6186 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,804B, BPFP=0.2824 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14370025 67.42255825 + layer.0.v_cache 0.00001382 0.00914993 + layer.1.k_cache 0.61448195 10.12867074 + layer.1.v_cache 0.00000570 0.00367018 + layer.2.k_cache 0.01208573 0.96898293 + layer.2.v_cache 0.00001823 0.01046863 + layer.3.k_cache 0.01319265 4.30883013 + layer.3.v_cache 0.00002242 0.01171270 + layer.4.k_cache 0.00071920 0.22539391 + layer.4.v_cache 0.00005145 0.02273617 + layer.4.output 0.00992508 194.39140586 + ------------------------------------------------------------------------------------- + TOTAL 0.05022159 84.93247144 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 268124 +BPFP 0.8708 bits/point +EBPFP 1.7416 equivalent bits/point +MSE 84.932471 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9325 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 259, 128) +Output shape: (1, 259, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.output: torch.Size([1, 259, 3584]) -> torch.Size([1, 1, 259, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,220B, BPFP=0.4959 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,280B, BPFP=1.8267 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,240B, BPFP=0.7987 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,148B, BPFP=1.7584 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,444B, BPFP=0.8714 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,192B, BPFP=1.7008 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,516B, BPFP=0.8154 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,540B, BPFP=1.7218 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,164B, BPFP=1.3974 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,512B, BPFP=1.6597 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,388B, BPFP=0.2705 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13349269 66.04658844 + layer.0.v_cache 0.00001436 0.00911675 + layer.1.k_cache 0.50562990 9.95439378 + layer.1.v_cache 0.00000590 0.00369327 + layer.2.k_cache 0.00820730 0.94047870 + layer.2.v_cache 0.00001794 0.01004052 + layer.3.k_cache 0.03685973 3.84321789 + layer.3.v_cache 0.00002061 0.01144128 + layer.4.k_cache 0.00069539 0.22978864 + layer.4.v_cache 0.00005422 0.02250217 + layer.4.output 0.01018517 214.26037645 + ------------------------------------------------------------------------------------- + TOTAL 0.04448790 92.99375862 + (elements=2,254,336) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2254336 +Total Bytes 247644 +BPFP 0.8788 bits/point +EBPFP 1.7576 equivalent bits/point +MSE 92.993759 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 92.9938 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,664B, BPFP=0.5593 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,468B, BPFP=1.8211 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,796B, BPFP=0.7984 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,304B, BPFP=1.7537 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,204B, BPFP=0.8799 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,556B, BPFP=1.7104 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,104B, BPFP=0.8162 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,900B, BPFP=1.7303 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,888B, BPFP=1.3824 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,588B, BPFP=1.6544 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,860B, BPFP=0.2799 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12649573 65.29564525 + layer.0.v_cache 0.00001373 0.00902297 + layer.1.k_cache 0.51574391 10.05360605 + layer.1.v_cache 0.00000589 0.00364497 + layer.2.k_cache 0.01194659 0.94463970 + layer.2.v_cache 0.00001984 0.01051927 + layer.3.k_cache 0.04962628 3.75044804 + layer.3.v_cache 0.00002059 0.01176615 + layer.4.k_cache 0.00068541 0.23099402 + layer.4.v_cache 0.00005704 0.02358668 + layer.4.output 0.00772497 204.77488426 + ------------------------------------------------------------------------------------- + TOTAL 0.04462881 89.04459194 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 260332 +BPFP 0.8862 bits/point +EBPFP 1.7724 equivalent bits/point +MSE 89.044592 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 89.0446 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,624B, BPFP=0.5590 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,332B, BPFP=1.8199 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,972B, BPFP=0.8116 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,444B, BPFP=1.7684 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,344B, BPFP=0.8913 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,852B, BPFP=1.7340 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,180B, BPFP=0.8237 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,216B, BPFP=1.7551 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,028B, BPFP=1.3957 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,796B, BPFP=1.6726 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,012B, BPFP=0.2739 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12210708 64.67573769 + layer.0.v_cache 0.00001342 0.00878747 + layer.1.k_cache 0.54942129 10.03442746 + layer.1.v_cache 0.00000547 0.00355444 + layer.2.k_cache 0.01053270 0.99370147 + layer.2.v_cache 0.00001820 0.01037180 + layer.3.k_cache 0.02850239 3.68963101 + layer.3.v_cache 0.00002491 0.01197051 + layer.4.k_cache 0.00067987 0.22684608 + layer.4.v_cache 0.00005179 0.02270958 + layer.4.output 0.01126663 205.88241835 + ------------------------------------------------------------------------------------- + TOTAL 0.04648373 89.46203917 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 260800 +BPFP 0.8911 bits/point +EBPFP 1.7822 equivalent bits/point +MSE 89.462039 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 89.4620 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,696B, BPFP=0.5516 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,680B, BPFP=1.7368 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,404B, BPFP=0.7943 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,652B, BPFP=1.6838 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,792B, BPFP=0.8659 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,824B, BPFP=1.6411 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,464B, BPFP=0.7974 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,064B, BPFP=1.6535 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,772B, BPFP=1.3290 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,576B, BPFP=1.5767 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,844B, BPFP=0.2567 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14443027 65.65867368 + layer.0.v_cache 0.00001375 0.00891485 + layer.1.k_cache 0.62837083 9.95692973 + layer.1.v_cache 0.00000570 0.00355439 + layer.2.k_cache 0.01557094 0.95940719 + layer.2.v_cache 0.00001871 0.01004858 + layer.3.k_cache 0.04676386 4.11347503 + layer.3.v_cache 0.00001981 0.01122656 + layer.4.k_cache 0.00070582 0.22819587 + layer.4.v_cache 0.00005437 0.02167377 + layer.4.output 0.04838977 178.43503949 + ------------------------------------------------------------------------------------- + TOTAL 0.06909897 78.23631624 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 279768 +BPFP 0.8486 bits/point +EBPFP 1.6973 equivalent bits/point +MSE 78.236316 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 78.2363 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,912B, BPFP=0.5473 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,072B, BPFP=1.7708 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,476B, BPFP=0.7992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,028B, BPFP=1.7131 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,740B, BPFP=0.8690 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,232B, BPFP=1.6692 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,524B, BPFP=0.8019 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,652B, BPFP=1.6924 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,548B, BPFP=1.3553 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,232B, BPFP=1.6140 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,500B, BPFP=0.2879 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13293209 67.57596483 + layer.0.v_cache 0.00001388 0.00890577 + layer.1.k_cache 0.56267701 9.93768203 + layer.1.v_cache 0.00000566 0.00350989 + layer.2.k_cache 0.01145046 0.94972213 + layer.2.v_cache 0.00001921 0.01044214 + layer.3.k_cache 0.02674564 3.74447621 + layer.3.v_cache 0.00001926 0.01166519 + layer.4.k_cache 0.00069135 0.22850692 + layer.4.v_cache 0.00005493 0.02332591 + layer.4.output 0.00973303 195.21616292 + ------------------------------------------------------------------------------------- + TOTAL 0.04722004 85.23572597 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 268916 +BPFP 0.8734 bits/point +EBPFP 1.7468 equivalent bits/point +MSE 85.235726 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2357 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,100B, BPFP=0.5697 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,820B, BPFP=1.7949 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,372B, BPFP=0.8107 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,096B, BPFP=1.7541 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,504B, BPFP=0.8745 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,100B, BPFP=1.6979 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,308B, BPFP=0.8071 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,564B, BPFP=1.7241 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,648B, BPFP=1.3903 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,128B, BPFP=1.6431 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,864B, BPFP=0.2971 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15182159 61.25269348 + layer.0.v_cache 0.00001420 0.00908877 + layer.1.k_cache 0.56224567 10.01897951 + layer.1.v_cache 0.00000588 0.00369413 + layer.2.k_cache 0.00652889 0.92578389 + layer.2.v_cache 0.00001844 0.01035082 + layer.3.k_cache 0.06011557 4.09386282 + layer.3.v_cache 0.00001921 0.01219582 + layer.4.k_cache 0.00068317 0.23161666 + layer.4.v_cache 0.00005096 0.02267401 + layer.4.output 0.01130176 199.93811243 + ------------------------------------------------------------------------------------- + TOTAL 0.05062446 86.83221923 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 268504 +BPFP 0.8909 bits/point +EBPFP 1.7819 equivalent bits/point +MSE 86.832219 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 86.8322 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,328B, BPFP=0.5662 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,308B, BPFP=1.7713 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,588B, BPFP=0.7998 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,420B, BPFP=1.7226 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,280B, BPFP=0.8925 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,612B, BPFP=1.6783 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,124B, BPFP=0.8292 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,956B, BPFP=1.6971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,096B, BPFP=1.3759 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,744B, BPFP=1.6307 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,024B, BPFP=0.2586 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13053865 63.25690104 + layer.0.v_cache 0.00001414 0.00936713 + layer.1.k_cache 0.56363723 10.11585458 + layer.1.v_cache 0.00000590 0.00380531 + layer.2.k_cache 0.01033439 0.92633721 + layer.2.v_cache 0.00002015 0.01084247 + layer.3.k_cache 0.03655195 3.70622216 + layer.3.v_cache 0.00002026 0.01202771 + layer.4.k_cache 0.00068516 0.23199803 + layer.4.v_cache 0.00005378 0.02336440 + layer.4.output 0.00863618 194.37758459 + ------------------------------------------------------------------------------------- + TOTAL 0.04719499 84.64351836 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 269480 +BPFP 0.8691 bits/point +EBPFP 1.7381 equivalent bits/point +MSE 84.643518 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 84.6435 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,040B, BPFP=0.5663 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,900B, BPFP=1.7994 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,364B, BPFP=0.8102 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,852B, BPFP=1.7403 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,464B, BPFP=0.8723 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,912B, BPFP=1.6873 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,388B, BPFP=0.8116 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,476B, BPFP=1.7191 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,196B, BPFP=1.3648 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,864B, BPFP=1.6282 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,428B, BPFP=0.2935 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15908699 63.93200023 + layer.0.v_cache 0.00001433 0.00896125 + layer.1.k_cache 0.53822481 10.10390607 + layer.1.v_cache 0.00000570 0.00362461 + layer.2.k_cache 0.01139215 0.97142310 + layer.2.v_cache 0.00001845 0.01032573 + layer.3.k_cache 0.02907990 4.00315312 + layer.3.v_cache 0.00001963 0.01167819 + layer.4.k_cache 0.00069750 0.22470751 + layer.4.v_cache 0.00005397 0.02271064 + layer.4.output 0.01069233 199.80877708 + ------------------------------------------------------------------------------------- + TOTAL 0.04784940 86.93846647 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 266884 +BPFP 0.8856 bits/point +EBPFP 1.7711 equivalent bits/point +MSE 86.938466 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9385 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,024B, BPFP=0.5476 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,124B, BPFP=1.7550 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,448B, BPFP=0.7893 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,300B, BPFP=1.7100 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,040B, BPFP=0.8763 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,352B, BPFP=1.6582 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,916B, BPFP=0.8149 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,812B, BPFP=1.6833 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,916B, BPFP=1.3612 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,516B, BPFP=1.6125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,188B, BPFP=0.2512 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16084659 63.10021717 + layer.0.v_cache 0.00001440 0.00883443 + layer.1.k_cache 0.59028764 10.03411396 + layer.1.v_cache 0.00000559 0.00357992 + layer.2.k_cache 0.00929020 0.99231875 + layer.2.v_cache 0.00002205 0.01022118 + layer.3.k_cache 0.04382113 3.74356911 + layer.3.v_cache 0.00002103 0.01149376 + layer.4.k_cache 0.00069271 0.22488862 + layer.4.v_cache 0.00005052 0.02190903 + layer.4.output 0.00984751 193.92090722 + ------------------------------------------------------------------------------------- + TOTAL 0.05141085 84.44691155 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 266636 +BPFP 0.8569 bits/point +EBPFP 1.7138 equivalent bits/point +MSE 84.446912 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 84.4469 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,976B, BPFP=0.5412 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,180B, BPFP=1.7459 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,580B, BPFP=0.7910 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,272B, BPFP=1.6966 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,932B, BPFP=0.8644 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,412B, BPFP=1.6500 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,604B, BPFP=0.7923 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,844B, BPFP=1.6734 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,688B, BPFP=1.3394 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,512B, BPFP=1.6011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,320B, BPFP=0.2737 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14211286 63.02492947 + layer.0.v_cache 0.00001516 0.00864804 + layer.1.k_cache 0.60291523 10.15740458 + layer.1.v_cache 0.00000570 0.00356206 + layer.2.k_cache 0.01041983 0.90520647 + layer.2.v_cache 0.00001835 0.00977741 + layer.3.k_cache 0.03896446 3.70044242 + layer.3.v_cache 0.00001926 0.01135613 + layer.4.k_cache 0.00071470 0.22364865 + layer.4.v_cache 0.00005236 0.02236150 + layer.4.output 0.00754456 192.07519531 + ------------------------------------------------------------------------------------- + TOTAL 0.04988528 83.68198258 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 269320 +BPFP 0.8595 bits/point +EBPFP 1.7190 equivalent bits/point +MSE 83.681983 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.010s, Pack+Encode: 0.249s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 83.6820 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 296, 128) +Output shape: (1, 296, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.output: torch.Size([1, 296, 3584]) -> torch.Size([1, 1, 296, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,392B, BPFP=0.5486 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,028B, BPFP=1.7435 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,248B, BPFP=0.8049 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,028B, BPFP=1.6907 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,608B, BPFP=0.8767 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,292B, BPFP=1.6518 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,236B, BPFP=0.8043 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,736B, BPFP=1.6753 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,656B, BPFP=1.3543 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,280B, BPFP=1.5984 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,812B, BPFP=0.2701 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14616069 66.25253378 + layer.0.v_cache 0.00001366 0.00893522 + layer.1.k_cache 0.61022733 9.72130688 + layer.1.v_cache 0.00000576 0.00353528 + layer.2.k_cache 0.01378242 0.95580106 + layer.2.v_cache 0.00001907 0.01027019 + layer.3.k_cache 0.03184220 4.09651596 + layer.3.v_cache 0.00001934 0.01155356 + layer.4.k_cache 0.00072724 0.22266749 + layer.4.v_cache 0.00005080 0.02222144 + layer.4.output 0.04784788 182.82081021 + ------------------------------------------------------------------------------------- + TOTAL 0.06692845 80.06182425 + (elements=2,576,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2576384 +Total Bytes 277316 +BPFP 0.8611 bits/point +EBPFP 1.7222 equivalent bits/point +MSE 80.061824 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 80.0618 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,536B, BPFP=0.5524 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,480B, BPFP=1.7555 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,280B, BPFP=0.8012 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,568B, BPFP=1.7076 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,772B, BPFP=0.8794 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,712B, BPFP=1.6628 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,492B, BPFP=0.8123 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,128B, BPFP=1.6846 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,716B, BPFP=1.3484 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,684B, BPFP=1.6089 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,132B, BPFP=0.2706 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15387250 64.17627281 + layer.0.v_cache 0.00001382 0.00879810 + layer.1.k_cache 0.63844893 10.30130558 + layer.1.v_cache 0.00000583 0.00363541 + layer.2.k_cache 0.01105453 0.94583335 + layer.2.v_cache 0.00001885 0.01023824 + layer.3.k_cache 0.03893430 3.97230069 + layer.3.v_cache 0.00001923 0.01179786 + layer.4.k_cache 0.00070257 0.22579613 + layer.4.v_cache 0.00005146 0.02189400 + layer.4.output 0.04873301 181.69434624 + ------------------------------------------------------------------------------------- + TOTAL 0.06966195 79.50225270 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 280500 +BPFP 0.8651 bits/point +EBPFP 1.7303 equivalent bits/point +MSE 79.502253 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.250s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 79.5023 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,848B, BPFP=0.5678 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,676B, BPFP=1.8263 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,108B, BPFP=0.8134 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,532B, BPFP=1.7604 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,332B, BPFP=0.8840 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,620B, BPFP=1.7078 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,192B, BPFP=0.8183 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,956B, BPFP=1.7272 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,008B, BPFP=1.3842 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,760B, BPFP=1.6582 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,924B, BPFP=0.2877 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12629252 65.80526407 + layer.0.v_cache 0.00001451 0.00906801 + layer.1.k_cache 0.49097243 10.08915998 + layer.1.v_cache 0.00000577 0.00370077 + layer.2.k_cache 0.01060468 0.98738374 + layer.2.v_cache 0.00001909 0.01031884 + layer.3.k_cache 0.02748298 3.94006348 + layer.3.v_cache 0.00001876 0.01156479 + layer.4.k_cache 0.00069967 0.22840197 + layer.4.v_cache 0.00005414 0.02229616 + layer.4.output 0.01047103 203.81782090 + ------------------------------------------------------------------------------------- + TOTAL 0.04290952 88.69599812 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 262956 +BPFP 0.8918 bits/point +EBPFP 1.7837 equivalent bits/point +MSE 88.695998 +---------------------- -------------------------------------------------------- +Time: 0.634s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6960 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,116B, BPFP=0.5665 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,760B, BPFP=1.7787 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,396B, BPFP=0.8062 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,844B, BPFP=1.7274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,512B, BPFP=0.8687 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,016B, BPFP=1.6810 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,448B, BPFP=0.8091 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,400B, BPFP=1.7025 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,164B, BPFP=1.3533 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,020B, BPFP=1.6252 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,504B, BPFP=0.2841 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15022702 62.68761201 + layer.0.v_cache 0.00001395 0.00861453 + layer.1.k_cache 0.54271635 9.73492716 + layer.1.v_cache 0.00000573 0.00357273 + layer.2.k_cache 0.01332666 0.99480381 + layer.2.v_cache 0.00001832 0.00994387 + layer.3.k_cache 0.02579262 3.99835052 + layer.3.v_cache 0.00001838 0.01100777 + layer.4.k_cache 0.00069325 0.22454386 + layer.4.v_cache 0.00005155 0.02231807 + layer.4.output 0.01006869 197.89120904 + ------------------------------------------------------------------------------------- + TOTAL 0.04725557 86.05495045 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 266180 +BPFP 0.8769 bits/point +EBPFP 1.7538 equivalent bits/point +MSE 86.054950 +---------------------- -------------------------------------------------------- +Time: 0.631s Load: 0.009s, Pack+Encode: 0.248s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 86.0550 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,400B, BPFP=0.5508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,120B, BPFP=1.7542 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,152B, BPFP=0.8025 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,176B, BPFP=1.7042 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,592B, BPFP=0.8788 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,268B, BPFP=1.6561 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,348B, BPFP=0.8129 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,772B, BPFP=1.6828 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,568B, BPFP=1.3542 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,400B, BPFP=1.6102 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,856B, BPFP=0.2789 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12416320 64.76784958 + layer.0.v_cache 0.00001370 0.00859664 + layer.1.k_cache 0.62725913 9.97637546 + layer.1.v_cache 0.00000565 0.00357128 + layer.2.k_cache 0.01102555 0.93186873 + layer.2.v_cache 0.00002033 0.01016582 + layer.3.k_cache 0.04118381 3.83697034 + layer.3.v_cache 0.00001949 0.01164780 + layer.4.k_cache 0.00070768 0.22762588 + layer.4.v_cache 0.00005397 0.02240404 + layer.4.output 0.05105524 183.26584443 + ------------------------------------------------------------------------------------- + TOTAL 0.06834348 80.15635215 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 278652 +BPFP 0.8682 bits/point +EBPFP 1.7364 equivalent bits/point +MSE 80.156352 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 80.1564 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,008B, BPFP=0.5411 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,364B, BPFP=1.7498 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,468B, BPFP=0.7822 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,536B, BPFP=1.7050 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,864B, BPFP=0.8577 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,664B, BPFP=1.6579 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,736B, BPFP=0.7967 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,960B, BPFP=1.6739 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,040B, BPFP=1.3538 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,608B, BPFP=1.6008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,020B, BPFP=0.2628 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15516999 62.82786143 + layer.0.v_cache 0.00001361 0.00921781 + layer.1.k_cache 0.56212566 9.85698681 + layer.1.v_cache 0.00000565 0.00381551 + layer.2.k_cache 0.00839357 0.97797645 + layer.2.v_cache 0.00001980 0.01066878 + layer.3.k_cache 0.02402391 3.48178027 + layer.3.v_cache 0.00001987 0.01177989 + layer.4.k_cache 0.00071212 0.22513377 + layer.4.v_cache 0.00005188 0.02247317 + layer.4.output 0.05073151 187.60107205 + ------------------------------------------------------------------------------------- + TOTAL 0.06503863 81.80207048 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 269268 +BPFP 0.8564 bits/point +EBPFP 1.7127 equivalent bits/point +MSE 81.802070 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 81.8021 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,460B, BPFP=0.5503 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,284B, BPFP=1.7511 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,084B, BPFP=0.7936 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,252B, BPFP=1.6968 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,616B, BPFP=0.8742 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,580B, BPFP=1.6614 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,488B, BPFP=0.8148 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,956B, BPFP=1.6812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,640B, BPFP=1.3489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,408B, BPFP=1.5997 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,348B, BPFP=0.2657 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005571 65.56203967 + layer.0.v_cache 0.00001459 0.00869413 + layer.1.k_cache 0.64038677 10.12015500 + layer.1.v_cache 0.00000545 0.00345376 + layer.2.k_cache 0.01623841 0.99239075 + layer.2.v_cache 0.00002030 0.01047113 + layer.3.k_cache 0.01774169 3.93701788 + layer.3.v_cache 0.00001948 0.01180793 + layer.4.k_cache 0.00069419 0.22729431 + layer.4.v_cache 0.00005057 0.02231500 + layer.4.output 0.05052170 182.41316438 + ------------------------------------------------------------------------------------- + TOTAL 0.06875759 79.86987001 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 278116 +BPFP 0.8607 bits/point +EBPFP 1.7214 equivalent bits/point +MSE 79.869870 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.247s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 79.8699 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,196B, BPFP=0.5710 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,972B, BPFP=1.7905 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,256B, BPFP=0.7984 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,088B, BPFP=1.7410 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,644B, BPFP=0.8761 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,316B, BPFP=1.6978 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,512B, BPFP=0.8127 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,688B, BPFP=1.7186 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,400B, BPFP=1.3665 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,096B, BPFP=1.6295 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,324B, BPFP=0.2906 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14010120 63.75564936 + layer.0.v_cache 0.00001437 0.00893278 + layer.1.k_cache 0.58692467 10.21487438 + layer.1.v_cache 0.00000569 0.00368442 + layer.2.k_cache 0.00806839 0.94407386 + layer.2.v_cache 0.00001934 0.01017630 + layer.3.k_cache 0.03832107 3.60969724 + layer.3.v_cache 0.00001992 0.01181770 + layer.4.k_cache 0.00069804 0.22216689 + layer.4.v_cache 0.00005131 0.02303178 + layer.4.output 0.00759059 198.05949181 + ------------------------------------------------------------------------------------- + TOTAL 0.04866812 86.18944396 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 268492 +BPFP 0.8845 bits/point +EBPFP 1.7690 equivalent bits/point +MSE 86.189444 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 86.1894 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,480B, BPFP=0.5527 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,500B, BPFP=1.8365 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,024B, BPFP=0.8176 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,600B, BPFP=1.7840 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,356B, BPFP=0.8953 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,784B, BPFP=1.7365 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,216B, BPFP=0.8288 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,020B, BPFP=1.7502 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,184B, BPFP=1.4100 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,708B, BPFP=1.6737 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,492B, BPFP=0.2790 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13199043 64.55739855 + layer.0.v_cache 0.00001386 0.00885319 + layer.1.k_cache 0.56989129 10.01912313 + layer.1.v_cache 0.00000577 0.00366644 + layer.2.k_cache 0.00892691 1.00082101 + layer.2.v_cache 0.00001816 0.01011679 + layer.3.k_cache 0.04626510 4.22932184 + layer.3.v_cache 0.00001927 0.01152427 + layer.4.k_cache 0.00069392 0.22863600 + layer.4.v_cache 0.00005195 0.02276773 + layer.4.output 0.00911696 206.57589286 + ------------------------------------------------------------------------------------- + TOTAL 0.04833502 89.77196935 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 261364 +BPFP 0.8964 bits/point +EBPFP 1.7927 equivalent bits/point +MSE 89.771969 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 89.7720 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,924B, BPFP=0.5659 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,604B, BPFP=1.8022 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,248B, BPFP=0.8125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,656B, BPFP=1.7482 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,448B, BPFP=0.8809 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,768B, BPFP=1.6975 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,304B, BPFP=0.8157 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,220B, BPFP=1.7233 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,208B, BPFP=1.3805 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,640B, BPFP=1.6332 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,212B, BPFP=0.2706 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16500157 62.04285470 + layer.0.v_cache 0.00001391 0.00893170 + layer.1.k_cache 0.57858722 9.84497605 + layer.1.v_cache 0.00000575 0.00362331 + layer.2.k_cache 0.01018238 0.99051137 + layer.2.v_cache 0.00001877 0.01048418 + layer.3.k_cache 0.03425542 3.91257003 + layer.3.v_cache 0.00002021 0.01201555 + layer.4.k_cache 0.00069807 0.22576818 + layer.4.v_cache 0.00005194 0.02251394 + layer.4.output 0.01023613 202.26797119 + ------------------------------------------------------------------------------------- + TOTAL 0.05061695 87.82059102 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 262232 +BPFP 0.8796 bits/point +EBPFP 1.7593 equivalent bits/point +MSE 87.820591 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 87.8206 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,668B, BPFP=0.5595 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,316B, BPFP=1.8123 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,908B, BPFP=0.8049 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,104B, BPFP=1.7421 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,984B, BPFP=0.8671 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,460B, BPFP=1.7049 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,948B, BPFP=0.8072 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,784B, BPFP=1.7236 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,708B, BPFP=1.3720 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,528B, BPFP=1.6509 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,700B, BPFP=0.2869 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385483 64.47037760 + layer.0.v_cache 0.00001387 0.00886584 + layer.1.k_cache 0.55871571 10.25213578 + layer.1.v_cache 0.00000565 0.00346022 + layer.2.k_cache 0.00818202 0.94829135 + layer.2.v_cache 0.00001875 0.01008823 + layer.3.k_cache 0.04697755 3.88963397 + layer.3.v_cache 0.00001978 0.01156172 + layer.4.k_cache 0.00071636 0.22334939 + layer.4.v_cache 0.00005146 0.02212165 + layer.4.output 0.01106052 204.81947751 + ------------------------------------------------------------------------------------- + TOTAL 0.04741057 89.03389579 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 260108 +BPFP 0.8854 bits/point +EBPFP 1.7709 equivalent bits/point +MSE 89.033896 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 89.0339 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,108B, BPFP=0.5743 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,488B, BPFP=1.7891 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,284B, BPFP=0.8116 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,556B, BPFP=1.7361 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,436B, BPFP=0.8770 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,788B, BPFP=1.6925 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,156B, BPFP=0.8043 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,352B, BPFP=1.7245 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,208B, BPFP=1.3755 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,784B, BPFP=1.6355 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,472B, BPFP=0.2717 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203608 63.53846591 + layer.0.v_cache 0.00001420 0.00875562 + layer.1.k_cache 0.56673284 9.98483221 + layer.1.v_cache 0.00000552 0.00347336 + layer.2.k_cache 0.00530624 0.95650302 + layer.2.v_cache 0.00001825 0.00999639 + layer.3.k_cache 0.02188354 3.79736861 + layer.3.v_cache 0.00001854 0.01140120 + layer.4.k_cache 0.00069888 0.22446494 + layer.4.v_cache 0.00004918 0.02132249 + layer.4.output 0.00799176 201.42123377 + ------------------------------------------------------------------------------------- + TOTAL 0.04662974 87.55913059 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 262632 +BPFP 0.8778 bits/point +EBPFP 1.7556 equivalent bits/point +MSE 87.559131 +---------------------- -------------------------------------------------------- +Time: 0.628s Load: 0.010s, Pack+Encode: 0.246s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 87.5591 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,956B, BPFP=0.5497 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,884B, BPFP=1.7604 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,516B, BPFP=0.8015 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,800B, BPFP=1.7005 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,628B, BPFP=0.8629 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,032B, BPFP=1.6581 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,484B, BPFP=0.7997 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,448B, BPFP=1.6811 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,372B, BPFP=1.3456 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,072B, BPFP=1.6051 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,712B, BPFP=0.2896 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12732459 66.96273189 + layer.0.v_cache 0.00001480 0.00886083 + layer.1.k_cache 0.57322790 9.91567745 + layer.1.v_cache 0.00000557 0.00358869 + layer.2.k_cache 0.00950604 0.94729884 + layer.2.v_cache 0.00002039 0.01040068 + layer.3.k_cache 0.01510038 4.04618571 + layer.3.v_cache 0.00001916 0.01179262 + layer.4.k_cache 0.00068141 0.22419917 + layer.4.v_cache 0.00005165 0.02214435 + layer.4.output 0.00936729 195.47236244 + ------------------------------------------------------------------------------------- + TOTAL 0.04656017 85.32114220 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 267904 +BPFP 0.8701 bits/point +EBPFP 1.7402 equivalent bits/point +MSE 85.321142 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.011s, Pack+Encode: 0.248s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 85.3211 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,132B, BPFP=0.5736 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,816B, BPFP=1.8012 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,088B, BPFP=0.7976 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,900B, BPFP=1.7493 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,572B, BPFP=0.8816 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,068B, BPFP=1.7022 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,308B, BPFP=0.8100 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,592B, BPFP=1.7319 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,412B, BPFP=1.3820 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,076B, BPFP=1.6461 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,612B, BPFP=0.2961 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13630490 61.25215127 + layer.0.v_cache 0.00001544 0.00895499 + layer.1.k_cache 0.58776402 10.19453939 + layer.1.v_cache 0.00000559 0.00346729 + layer.2.k_cache 0.01000112 0.95078477 + layer.2.v_cache 0.00001940 0.01038680 + layer.3.k_cache 0.04424953 3.70647486 + layer.3.v_cache 0.00002139 0.01179331 + layer.4.k_cache 0.00071093 0.22428156 + layer.4.v_cache 0.00005158 0.02284733 + layer.4.output 0.01101471 200.82084627 + ------------------------------------------------------------------------------------- + TOTAL 0.05036746 87.18421209 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 267576 +BPFP 0.8911 bits/point +EBPFP 1.7821 equivalent bits/point +MSE 87.184212 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.009s, Pack+Encode: 0.247s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 87.1842 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,104B, BPFP=0.5679 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,812B, BPFP=1.7880 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,348B, BPFP=0.8064 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,892B, BPFP=1.7363 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,436B, BPFP=0.8676 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,180B, BPFP=1.6963 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,384B, BPFP=0.8085 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,368B, BPFP=1.7068 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,348B, BPFP=1.3685 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,976B, BPFP=1.6286 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,304B, BPFP=0.2915 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.401s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12958164 62.28239518 + layer.0.v_cache 0.00001430 0.00882769 + layer.1.k_cache 0.54121377 10.01773093 + layer.1.v_cache 0.00000577 0.00356017 + layer.2.k_cache 0.01080061 0.96548352 + layer.2.v_cache 0.00001903 0.01025528 + layer.3.k_cache 0.01783211 3.82369446 + layer.3.v_cache 0.00002028 0.01138906 + layer.4.k_cache 0.00068935 0.22689010 + layer.4.v_cache 0.00005053 0.02231719 + layer.4.output 0.00833506 199.42428379 + ------------------------------------------------------------------------------------- + TOTAL 0.04462193 86.66720765 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 267152 +BPFP 0.8833 bits/point +EBPFP 1.7665 equivalent bits/point +MSE 86.667208 +---------------------- -------------------------------------------------------- +Time: 0.667s Load: 0.009s, Pack+Encode: 0.258s, Decode+Unpack: 0.401s +---------------------- -------------------------------------------------------- +💾 Converting with 86.6672 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,616B, BPFP=0.5119 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,028B, BPFP=1.8434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,740B, BPFP=0.8163 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,032B, BPFP=1.7842 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,180B, BPFP=0.9019 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,200B, BPFP=1.7348 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,928B, BPFP=0.8275 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,556B, BPFP=1.7559 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,968B, BPFP=1.4240 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,328B, BPFP=1.6830 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,540B, BPFP=0.2762 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12547956 65.50316362 + layer.0.v_cache 0.00001492 0.00896889 + layer.1.k_cache 0.50446328 10.10472612 + layer.1.v_cache 0.00000633 0.00359397 + layer.2.k_cache 0.00671430 0.95924650 + layer.2.v_cache 0.00001992 0.01045720 + layer.3.k_cache 0.02142675 3.81485322 + layer.3.v_cache 0.00001986 0.01191032 + layer.4.k_cache 0.00069859 0.22614618 + layer.4.v_cache 0.00005414 0.02317934 + layer.4.output 0.01038299 210.61681831 + ------------------------------------------------------------------------------------- + TOTAL 0.04303403 91.46964550 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 256116 +BPFP 0.8951 bits/point +EBPFP 1.7901 equivalent bits/point +MSE 91.469645 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.011s, Pack+Encode: 0.247s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4696 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,544B, BPFP=0.5473 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,536B, BPFP=1.7409 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,232B, BPFP=0.7907 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,556B, BPFP=1.6900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,672B, BPFP=0.8654 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,676B, BPFP=1.6443 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,492B, BPFP=0.8042 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,012B, BPFP=1.6618 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,688B, BPFP=1.3335 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,528B, BPFP=1.5847 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,720B, BPFP=0.2797 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13970643 63.85438512 + layer.0.v_cache 0.00001441 0.00885154 + layer.1.k_cache 0.60015088 9.90435771 + layer.1.v_cache 0.00000584 0.00356927 + layer.2.k_cache 0.00648033 0.96357104 + layer.2.v_cache 0.00001840 0.01006178 + layer.3.k_cache 0.03288257 4.05360590 + layer.3.v_cache 0.00002024 0.01139812 + layer.4.k_cache 0.00070287 0.22301176 + layer.4.v_cache 0.00005060 0.02198825 + layer.4.output 0.04874230 179.38612364 + ------------------------------------------------------------------------------------- + TOTAL 0.06595463 78.51515682 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 281656 +BPFP 0.8600 bits/point +EBPFP 1.7201 equivalent bits/point +MSE 78.515157 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.010s, Pack+Encode: 0.248s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 78.5152 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,984B, BPFP=0.5248 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,812B, BPFP=1.8068 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,936B, BPFP=0.8093 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,464B, BPFP=1.7424 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,536B, BPFP=0.8857 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,440B, BPFP=1.6934 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,160B, BPFP=0.8200 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,952B, BPFP=1.7179 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,972B, BPFP=1.3844 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,312B, BPFP=1.6395 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,280B, BPFP=0.2750 +⌛️ [2/4] FRONTEND: Frontend time: 0.279s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.435s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13960611 65.43587538 + layer.0.v_cache 0.00001491 0.00916228 + layer.1.k_cache 0.68890666 10.26236905 + layer.1.v_cache 0.00000583 0.00366222 + layer.2.k_cache 0.00940153 0.94348742 + layer.2.v_cache 0.00001882 0.01033571 + layer.3.k_cache 0.04667800 3.72849173 + layer.3.v_cache 0.00001939 0.01181643 + layer.4.k_cache 0.00070827 0.22135067 + layer.4.v_cache 0.00005291 0.02227198 + layer.4.output 0.04427218 165.56658202 + ------------------------------------------------------------------------------------- + TOTAL 0.07031281 72.91852335 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 312848 +BPFP 0.8793 bits/point +EBPFP 1.7587 equivalent bits/point +MSE 72.918523 +---------------------- -------------------------------------------------------- +Time: 0.726s Load: 0.011s, Pack+Encode: 0.279s, Decode+Unpack: 0.435s +---------------------- -------------------------------------------------------- +💾 Converting with 72.9185 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,664B, BPFP=0.5573 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,452B, BPFP=1.7481 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,228B, BPFP=0.7958 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,452B, BPFP=1.6959 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,776B, BPFP=0.8767 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,704B, BPFP=1.6568 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,472B, BPFP=0.8085 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,980B, BPFP=1.6712 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,816B, BPFP=1.3491 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,576B, BPFP=1.5978 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,640B, BPFP=0.2810 +⌛️ [2/4] FRONTEND: Frontend time: 0.279s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.394s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15817952 63.55927310 + layer.0.v_cache 0.00001474 0.00871670 + layer.1.k_cache 0.61609703 10.37140158 + layer.1.v_cache 0.00000566 0.00357708 + layer.2.k_cache 0.00844678 0.93976463 + layer.2.v_cache 0.00001876 0.01011313 + layer.3.k_cache 0.04682518 3.85878629 + layer.3.v_cache 0.00001896 0.01140668 + layer.4.k_cache 0.00069859 0.22091394 + layer.4.v_cache 0.00005216 0.02221299 + layer.4.output 0.04649196 180.63351051 + ------------------------------------------------------------------------------------- + TOTAL 0.06798830 79.02592586 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 281760 +BPFP 0.8661 bits/point +EBPFP 1.7322 equivalent bits/point +MSE 79.025926 +---------------------- -------------------------------------------------------- +Time: 0.686s Load: 0.013s, Pack+Encode: 0.279s, Decode+Unpack: 0.394s +---------------------- -------------------------------------------------------- +💾 Converting with 79.0259 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,972B, BPFP=0.5429 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,312B, BPFP=1.7591 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,324B, BPFP=0.7798 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,332B, BPFP=1.7058 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,940B, BPFP=0.8678 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,444B, BPFP=1.6574 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,652B, BPFP=0.7977 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,736B, BPFP=1.6733 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,968B, BPFP=1.3593 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,480B, BPFP=1.6050 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,000B, BPFP=0.2644 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14002854 62.92581528 + layer.0.v_cache 0.00001601 0.00901887 + layer.1.k_cache 0.54217949 9.94607470 + layer.1.v_cache 0.00000579 0.00374091 + layer.2.k_cache 0.01082110 0.97420435 + layer.2.v_cache 0.00001971 0.01051516 + layer.3.k_cache 0.03102738 3.65720019 + layer.3.v_cache 0.00001907 0.01164818 + layer.4.k_cache 0.00066782 0.22364010 + layer.4.v_cache 0.00005141 0.02249951 + layer.4.output 0.00856273 192.75542870 + ------------------------------------------------------------------------------------- + TOTAL 0.04616326 83.94543283 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 268160 +BPFP 0.8588 bits/point +EBPFP 1.7176 equivalent bits/point +MSE 83.945433 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.011s, Pack+Encode: 0.256s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9454 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,324B, BPFP=0.5431 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,184B, BPFP=1.7458 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,000B, BPFP=0.7891 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,160B, BPFP=1.6919 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,464B, BPFP=0.8662 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,368B, BPFP=1.6503 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,268B, BPFP=0.8032 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,792B, BPFP=1.6726 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,444B, BPFP=1.3386 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,312B, BPFP=1.5947 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,304B, BPFP=0.2728 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.404s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14705232 63.07660590 + layer.0.v_cache 0.00001374 0.00869188 + layer.1.k_cache 0.65666707 10.09623580 + layer.1.v_cache 0.00000579 0.00354578 + layer.2.k_cache 0.01062412 0.97212748 + layer.2.v_cache 0.00001918 0.01027033 + layer.3.k_cache 0.03205616 3.77458258 + layer.3.v_cache 0.00001999 0.01195949 + layer.4.k_cache 0.00071739 0.22233111 + layer.4.v_cache 0.00005111 0.02220930 + layer.4.output 0.04976816 182.41176647 + ------------------------------------------------------------------------------------- + TOTAL 0.07032965 79.71064265 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 277620 +BPFP 0.8591 bits/point +EBPFP 1.7183 equivalent bits/point +MSE 79.710643 +---------------------- -------------------------------------------------------- +Time: 0.668s Load: 0.011s, Pack+Encode: 0.254s, Decode+Unpack: 0.404s +---------------------- -------------------------------------------------------- +💾 Converting with 79.7106 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,276B, BPFP=0.5714 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,904B, BPFP=1.7740 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,652B, BPFP=0.8147 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,736B, BPFP=1.7091 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,628B, BPFP=0.8690 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,976B, BPFP=1.6668 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,564B, BPFP=0.8098 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,388B, BPFP=1.6897 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,104B, BPFP=1.3403 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,968B, BPFP=1.6108 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,004B, BPFP=0.3019 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131647 63.78646992 + layer.0.v_cache 0.00001392 0.00864257 + layer.1.k_cache 0.51967005 10.25908446 + layer.1.v_cache 0.00000557 0.00337318 + layer.2.k_cache 0.01013049 0.97138070 + layer.2.v_cache 0.00001908 0.00981953 + layer.3.k_cache 0.05186962 3.77297224 + layer.3.v_cache 0.00001891 0.01108161 + layer.4.k_cache 0.00069958 0.22077206 + layer.4.v_cache 0.00005017 0.02133355 + layer.4.output 0.01073789 197.22554969 + ------------------------------------------------------------------------------------- + TOTAL 0.04758583 85.86139869 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 269200 +BPFP 0.8805 bits/point +EBPFP 1.7610 equivalent bits/point +MSE 85.861399 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8614 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,100B, BPFP=0.5480 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,172B, BPFP=1.7454 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,572B, BPFP=0.7906 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,244B, BPFP=1.6951 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,944B, BPFP=0.8650 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,536B, BPFP=1.6567 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,772B, BPFP=0.8014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,840B, BPFP=1.6732 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,832B, BPFP=1.3472 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,460B, BPFP=1.5983 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,100B, BPFP=0.2643 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15566496 64.49365234 + layer.0.v_cache 0.00001437 0.00873681 + layer.1.k_cache 0.62643353 9.96091885 + layer.1.v_cache 0.00000573 0.00348969 + layer.2.k_cache 0.01010078 0.99757883 + layer.2.v_cache 0.00001868 0.01016190 + layer.3.k_cache 0.06271097 3.85938390 + layer.3.v_cache 0.00001958 0.01175498 + layer.4.k_cache 0.00073784 0.21662001 + layer.4.v_cache 0.00005294 0.02182933 + layer.4.output 0.00677609 191.56939794 + ------------------------------------------------------------------------------------- + TOTAL 0.05312894 83.56293601 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 268572 +BPFP 0.8571 bits/point +EBPFP 1.7142 equivalent bits/point +MSE 83.562936 +---------------------- -------------------------------------------------------- +Time: 0.637s Load: 0.009s, Pack+Encode: 0.252s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 83.5629 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,172B, BPFP=0.5557 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,060B, BPFP=1.7515 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,428B, BPFP=0.7882 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,116B, BPFP=1.7000 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,824B, BPFP=0.8645 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,384B, BPFP=1.6600 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,600B, BPFP=0.7976 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,612B, BPFP=1.6724 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,780B, BPFP=1.3538 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,440B, BPFP=1.6084 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,328B, BPFP=0.2679 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13918158 62.57780403 + layer.0.v_cache 0.00001566 0.00900383 + layer.1.k_cache 0.57997750 10.05340363 + layer.1.v_cache 0.00000571 0.00371280 + layer.2.k_cache 0.01637170 1.00257063 + layer.2.v_cache 0.00001887 0.01072735 + layer.3.k_cache 0.04393929 3.95704704 + layer.3.v_cache 0.00001903 0.01179698 + layer.4.k_cache 0.00071104 0.22911568 + layer.4.v_cache 0.00005175 0.02344356 + layer.4.output 0.01065216 193.57681381 + ------------------------------------------------------------------------------------- + TOTAL 0.05028572 84.28919542 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 267744 +BPFP 0.8604 bits/point +EBPFP 1.7209 equivalent bits/point +MSE 84.289195 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 84.2892 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,624B, BPFP=0.5355 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,060B, BPFP=1.7167 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,656B, BPFP=0.7891 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,108B, BPFP=1.6687 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,144B, BPFP=0.8641 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,176B, BPFP=1.6218 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,872B, BPFP=0.8000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,540B, BPFP=1.6401 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,320B, BPFP=1.3266 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,040B, BPFP=1.5645 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,744B, BPFP=0.2934 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16496690 62.75876386 + layer.0.v_cache 0.00001433 0.00911899 + layer.1.k_cache 0.65265572 9.82522051 + layer.1.v_cache 0.00000578 0.00365154 + layer.2.k_cache 0.01241827 1.00254851 + layer.2.v_cache 0.00001920 0.01026190 + layer.3.k_cache 0.08444468 3.88310311 + layer.3.v_cache 0.00001975 0.01174154 + layer.4.k_cache 0.00068928 0.22196232 + layer.4.v_cache 0.00005259 0.02259858 + layer.4.output 0.04785494 173.94287154 + ------------------------------------------------------------------------------------- + TOTAL 0.07354536 76.19700422 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 289284 +BPFP 0.8577 bits/point +EBPFP 1.7154 equivalent bits/point +MSE 76.197004 +---------------------- -------------------------------------------------------- +Time: 0.700s Load: 0.009s, Pack+Encode: 0.252s, Decode+Unpack: 0.439s +---------------------- -------------------------------------------------------- +💾 Converting with 76.1970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,120B, BPFP=0.5378 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,020B, BPFP=1.7549 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,008B, BPFP=0.7976 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,024B, BPFP=1.7020 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,440B, BPFP=0.8737 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,248B, BPFP=1.6607 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,152B, BPFP=0.8053 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,572B, BPFP=1.6779 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,460B, BPFP=1.3531 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,224B, BPFP=1.6063 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,716B, BPFP=0.2560 +⌛️ [2/4] FRONTEND: Frontend time: 0.322s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.488s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15001445 63.92226695 + layer.0.v_cache 0.00001441 0.00900752 + layer.1.k_cache 0.57475359 9.92734159 + layer.1.v_cache 0.00000584 0.00373866 + layer.2.k_cache 0.01434258 0.99418609 + layer.2.v_cache 0.00001954 0.01055088 + layer.3.k_cache 0.02649183 3.79292474 + layer.3.v_cache 0.00001945 0.01169858 + layer.4.k_cache 0.00069414 0.23049024 + layer.4.v_cache 0.00005306 0.02330168 + layer.4.output 0.05035525 184.36267007 + ------------------------------------------------------------------------------------- + TOTAL 0.06581739 80.55671749 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 273984 +BPFP 0.8565 bits/point +EBPFP 1.7131 equivalent bits/point +MSE 80.556717 +---------------------- -------------------------------------------------------- +Time: 0.819s Load: 0.009s, Pack+Encode: 0.322s, Decode+Unpack: 0.488s +---------------------- -------------------------------------------------------- +💾 Converting with 80.5567 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,180B, BPFP=0.5641 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,008B, BPFP=1.7735 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,532B, BPFP=0.8052 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,104B, BPFP=1.7234 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,800B, BPFP=0.8754 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,224B, BPFP=1.6746 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,528B, BPFP=0.8050 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,604B, BPFP=1.6957 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,504B, BPFP=1.3577 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,160B, BPFP=1.6157 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,964B, BPFP=0.2847 +⌛️ [2/4] FRONTEND: Frontend time: 0.362s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.502s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11854591 66.64141456 + layer.0.v_cache 0.00001566 0.00879862 + layer.1.k_cache 0.56985614 10.03102577 + layer.1.v_cache 0.00000555 0.00355991 + layer.2.k_cache 0.01267573 0.99273076 + layer.2.v_cache 0.00001902 0.01025756 + layer.3.k_cache 0.04477839 3.70847272 + layer.3.v_cache 0.00001896 0.01171041 + layer.4.k_cache 0.00071754 0.22731094 + layer.4.v_cache 0.00005321 0.02286664 + layer.4.output 0.00761333 196.47088716 + ------------------------------------------------------------------------------------- + TOTAL 0.04705761 85.70319753 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 268608 +BPFP 0.8755 bits/point +EBPFP 1.7509 equivalent bits/point +MSE 85.703198 +---------------------- -------------------------------------------------------- +Time: 0.873s Load: 0.009s, Pack+Encode: 0.362s, Decode+Unpack: 0.502s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7032 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,172B, BPFP=0.5406 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,188B, BPFP=1.7638 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,020B, BPFP=0.7983 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,188B, BPFP=1.7107 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,616B, BPFP=0.8831 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,400B, BPFP=1.6688 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,468B, BPFP=0.8221 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,688B, BPFP=1.6841 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,632B, BPFP=1.3622 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,396B, BPFP=1.6154 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,176B, BPFP=0.2671 +⌛️ [2/4] FRONTEND: Frontend time: 0.353s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.477s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582189 61.87433567 + layer.0.v_cache 0.00001509 0.00913576 + layer.1.k_cache 0.58311327 9.84207506 + layer.1.v_cache 0.00000606 0.00377408 + layer.2.k_cache 0.00657600 0.93000783 + layer.2.v_cache 0.00001909 0.01057751 + layer.3.k_cache 0.02234380 3.57204391 + layer.3.v_cache 0.00001905 0.01208191 + layer.4.k_cache 0.00071828 0.22949403 + layer.4.v_cache 0.00005370 0.02331757 + layer.4.output 0.05044473 184.36974611 + ------------------------------------------------------------------------------------- + TOTAL 0.06363526 80.41735683 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 276944 +BPFP 0.8658 bits/point +EBPFP 1.7316 equivalent bits/point +MSE 80.417357 +---------------------- -------------------------------------------------------- +Time: 0.840s Load: 0.009s, Pack+Encode: 0.353s, Decode+Unpack: 0.477s +---------------------- -------------------------------------------------------- +💾 Converting with 80.4174 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 292, 128) +Output shape: (1, 292, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.output: torch.Size([1, 292, 3584]) -> torch.Size([1, 1, 292, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,984B, BPFP=0.5342 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,796B, BPFP=1.7549 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,888B, BPFP=0.7967 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,896B, BPFP=1.7068 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,448B, BPFP=0.8801 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,016B, BPFP=1.6597 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,148B, BPFP=0.8106 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,600B, BPFP=1.6909 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,332B, BPFP=1.3555 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,112B, BPFP=1.6113 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,784B, BPFP=0.2659 +⌛️ [2/4] FRONTEND: Frontend time: 0.336s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.490s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887579 66.93767391 + layer.0.v_cache 0.00001384 0.00890488 + layer.1.k_cache 0.57277983 9.60772705 + layer.1.v_cache 0.00000568 0.00359328 + layer.2.k_cache 0.01314942 0.98837103 + layer.2.v_cache 0.00002105 0.01057299 + layer.3.k_cache 0.03125376 3.90422748 + layer.3.v_cache 0.00001978 0.01177524 + layer.4.k_cache 0.00068035 0.22354664 + layer.4.v_cache 0.00005123 0.02228566 + layer.4.output 0.04859251 185.70783390 + ------------------------------------------------------------------------------------- + TOTAL 0.06511755 81.27491268 + (elements=2,541,568) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2541568 +Total Bytes 274004 +BPFP 0.8625 bits/point +EBPFP 1.7249 equivalent bits/point +MSE 81.274913 +---------------------- -------------------------------------------------------- +Time: 0.836s Load: 0.010s, Pack+Encode: 0.336s, Decode+Unpack: 0.490s +---------------------- -------------------------------------------------------- +💾 Converting with 81.2749 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 338, 128) +Output shape: (1, 338, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.output: torch.Size([1, 338, 3584]) -> torch.Size([1, 1, 338, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,464B, BPFP=0.5762 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,580B, BPFP=1.7835 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,100B, BPFP=0.7905 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,260B, BPFP=1.7224 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,588B, BPFP=0.8593 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,092B, BPFP=1.6685 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,116B, BPFP=0.7912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,624B, BPFP=1.6930 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,208B, BPFP=1.3502 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,032B, BPFP=1.6195 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,736B, BPFP=0.2888 +⌛️ [2/4] FRONTEND: Frontend time: 0.362s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.565s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13976314 60.84093010 + layer.0.v_cache 0.00001472 0.00917491 + layer.1.k_cache 0.73750278 10.45882402 + layer.1.v_cache 0.00000589 0.00365997 + layer.2.k_cache 0.02011468 0.98502246 + layer.2.v_cache 0.00002041 0.00995385 + layer.3.k_cache 0.02722352 3.84299085 + layer.3.v_cache 0.00002029 0.01116180 + layer.4.k_cache 0.00072609 0.22612255 + layer.4.v_cache 0.00005477 0.02246070 + layer.4.output 0.04270584 160.19995509 + ------------------------------------------------------------------------------------- + TOTAL 0.07202277 70.45941099 + (elements=2,941,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2941952 +Total Bytes 321800 +BPFP 0.8751 bits/point +EBPFP 1.7501 equivalent bits/point +MSE 70.459411 +---------------------- -------------------------------------------------------- +Time: 0.938s Load: 0.012s, Pack+Encode: 0.362s, Decode+Unpack: 0.565s +---------------------- -------------------------------------------------------- +💾 Converting with 70.4594 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,984B, BPFP=0.5361 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,716B, BPFP=1.7567 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,040B, BPFP=0.8076 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,808B, BPFP=1.7079 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,144B, BPFP=0.8668 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,936B, BPFP=1.6611 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,060B, BPFP=0.8086 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,224B, BPFP=1.6765 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,276B, BPFP=1.3572 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,828B, BPFP=1.6016 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,232B, BPFP=0.2703 +⌛️ [2/4] FRONTEND: Frontend time: 0.350s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.458s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150811 65.82262135 + layer.0.v_cache 0.00001374 0.00888646 + layer.1.k_cache 0.66021251 9.76387491 + layer.1.v_cache 0.00000574 0.00356961 + layer.2.k_cache 0.00963001 0.99129095 + layer.2.v_cache 0.00001988 0.01010638 + layer.3.k_cache 0.08055985 3.87624797 + layer.3.v_cache 0.00001942 0.01102444 + layer.4.k_cache 0.00069026 0.22156994 + layer.4.v_cache 0.00005126 0.02178989 + layer.4.output 0.05093874 186.32831676 + ------------------------------------------------------------------------------------- + TOTAL 0.07289894 81.47230584 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 273248 +BPFP 0.8630 bits/point +EBPFP 1.7261 equivalent bits/point +MSE 81.472306 +---------------------- -------------------------------------------------------- +Time: 0.818s Load: 0.010s, Pack+Encode: 0.350s, Decode+Unpack: 0.458s +---------------------- -------------------------------------------------------- +💾 Converting with 81.4723 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,712B, BPFP=0.5524 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,380B, BPFP=1.7213 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,400B, BPFP=0.7941 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,464B, BPFP=1.6741 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,636B, BPFP=0.8579 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,664B, BPFP=1.6328 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,380B, BPFP=0.7931 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,992B, BPFP=1.6498 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,712B, BPFP=1.3259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,612B, BPFP=1.5786 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,056B, BPFP=0.2804 +⌛️ [2/4] FRONTEND: Frontend time: 0.362s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.512s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14893842 63.00684561 + layer.0.v_cache 0.00001415 0.00862768 + layer.1.k_cache 0.65110169 9.78466475 + layer.1.v_cache 0.00000575 0.00360287 + layer.2.k_cache 0.01411552 0.97300463 + layer.2.v_cache 0.00001904 0.01007431 + layer.3.k_cache 0.03939451 3.54732138 + layer.3.v_cache 0.00001876 0.01090685 + layer.4.k_cache 0.00070558 0.22026367 + layer.4.v_cache 0.00005345 0.02186436 + layer.4.output 0.04819407 178.19706212 + ------------------------------------------------------------------------------------- + TOTAL 0.07010149 77.93921241 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 282008 +BPFP 0.8554 bits/point +EBPFP 1.7109 equivalent bits/point +MSE 77.939212 +---------------------- -------------------------------------------------------- +Time: 0.884s Load: 0.010s, Pack+Encode: 0.362s, Decode+Unpack: 0.512s +---------------------- -------------------------------------------------------- +💾 Converting with 77.9392 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,688B, BPFP=0.5162 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,652B, BPFP=1.8211 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,464B, BPFP=0.7999 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,720B, BPFP=1.7657 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,128B, BPFP=0.8988 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,996B, BPFP=1.7227 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,944B, BPFP=0.8284 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,304B, BPFP=1.7410 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,656B, BPFP=1.4054 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,988B, BPFP=1.6628 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,688B, BPFP=0.2520 +⌛️ [2/4] FRONTEND: Frontend time: 0.365s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.531s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14492303 66.65853731 + layer.0.v_cache 0.00001375 0.00880884 + layer.1.k_cache 0.50395557 9.92760627 + layer.1.v_cache 0.00000564 0.00353126 + layer.2.k_cache 0.00856250 0.97513154 + layer.2.v_cache 0.00001902 0.01027916 + layer.3.k_cache 0.02692018 3.54994300 + layer.3.v_cache 0.00002118 0.01158897 + layer.4.k_cache 0.00070790 0.22068890 + layer.4.v_cache 0.00005130 0.02228041 + layer.4.output 0.01131245 210.69057238 + ------------------------------------------------------------------------------------- + TOTAL 0.04496278 91.54249425 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 251228 +BPFP 0.8780 bits/point +EBPFP 1.7560 equivalent bits/point +MSE 91.542494 +---------------------- -------------------------------------------------------- +Time: 0.905s Load: 0.009s, Pack+Encode: 0.365s, Decode+Unpack: 0.531s +---------------------- -------------------------------------------------------- +💾 Converting with 91.5425 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,104B, BPFP=0.5720 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,768B, BPFP=1.7985 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,152B, BPFP=0.8012 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,792B, BPFP=1.7432 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,376B, BPFP=0.8705 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,104B, BPFP=1.7043 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,404B, BPFP=0.8154 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,520B, BPFP=1.7278 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,136B, BPFP=1.3664 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,860B, BPFP=1.6338 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,656B, BPFP=0.2965 +⌛️ [2/4] FRONTEND: Frontend time: 0.289s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14113761 64.61425781 + layer.0.v_cache 0.00001369 0.00884730 + layer.1.k_cache 0.59212339 10.11178279 + layer.1.v_cache 0.00000552 0.00342666 + layer.2.k_cache 0.01129283 0.92840952 + layer.2.v_cache 0.00001853 0.01027634 + layer.3.k_cache 0.01358872 4.03786745 + layer.3.v_cache 0.00001883 0.01130392 + layer.4.k_cache 0.00070763 0.22860696 + layer.4.v_cache 0.00005153 0.02216282 + layer.4.output 0.01019159 200.83545225 + ------------------------------------------------------------------------------------- + TOTAL 0.04884114 87.40147690 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 266872 +BPFP 0.8887 bits/point +EBPFP 1.7774 equivalent bits/point +MSE 87.401477 +---------------------- -------------------------------------------------------- +Time: 0.671s Load: 0.009s, Pack+Encode: 0.289s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 87.4015 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,948B, BPFP=0.5416 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,104B, BPFP=1.7478 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,416B, BPFP=0.7848 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,176B, BPFP=1.6973 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,904B, BPFP=0.8659 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,252B, BPFP=1.6470 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,704B, BPFP=0.8005 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,804B, BPFP=1.6770 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,876B, BPFP=1.3543 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,560B, BPFP=1.6093 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,704B, BPFP=0.2699 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11028925 64.46904943 + layer.0.v_cache 0.00001466 0.00879485 + layer.1.k_cache 0.56032453 9.89993552 + layer.1.v_cache 0.00000595 0.00348849 + layer.2.k_cache 0.00809042 0.92617192 + layer.2.v_cache 0.00001929 0.01007792 + layer.3.k_cache 0.03501172 3.84008747 + layer.3.v_cache 0.00001861 0.01113440 + layer.4.k_cache 0.00070074 0.22441750 + layer.4.v_cache 0.00005246 0.02170320 + layer.4.output 0.00889761 192.86266488 + ------------------------------------------------------------------------------------- + TOTAL 0.04569476 84.08550087 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 268448 +BPFP 0.8597 bits/point +EBPFP 1.7194 equivalent bits/point +MSE 84.085501 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.009s, Pack+Encode: 0.252s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 84.0855 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 311, 128) +Output shape: (1, 311, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.output: torch.Size([1, 311, 3584]) -> torch.Size([1, 1, 311, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,792B, BPFP=0.5422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,096B, BPFP=1.7130 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,564B, BPFP=0.7820 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,148B, BPFP=1.6654 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,032B, BPFP=0.8557 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,412B, BPFP=1.6284 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,940B, BPFP=0.8008 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,784B, BPFP=1.6471 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,328B, BPFP=1.3227 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,272B, BPFP=1.5711 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,960B, BPFP=0.2796 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14227538 60.07711390 + layer.0.v_cache 0.00001500 0.00895141 + layer.1.k_cache 0.64922478 9.88822598 + layer.1.v_cache 0.00000572 0.00359814 + layer.2.k_cache 0.00828172 0.95661843 + layer.2.v_cache 0.00001922 0.00999441 + layer.3.k_cache 0.02636586 3.85163237 + layer.3.v_cache 0.00001996 0.01156916 + layer.4.k_cache 0.00069494 0.22090527 + layer.4.v_cache 0.00005248 0.02199084 + layer.4.output 0.04686997 173.12211472 + ------------------------------------------------------------------------------------- + TOTAL 0.06794382 75.70031782 + (elements=2,706,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2706944 +Total Bytes 288328 +BPFP 0.8521 bits/point +EBPFP 1.7042 equivalent bits/point +MSE 75.700318 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.011s, Pack+Encode: 0.253s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 75.7003 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,524B, BPFP=0.5223 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,304B, BPFP=1.6730 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,304B, BPFP=0.7539 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,560B, BPFP=1.6275 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,884B, BPFP=0.8507 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,012B, BPFP=1.5939 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,720B, BPFP=0.7794 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,316B, BPFP=1.6125 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,376B, BPFP=1.3098 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,196B, BPFP=1.5439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,976B, BPFP=0.3237 +⌛️ [2/4] FRONTEND: Frontend time: 0.283s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.322s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13149661 62.40006127 + layer.0.v_cache 0.00001370 0.00884554 + layer.1.k_cache 0.50214604 9.50972924 + layer.1.v_cache 0.00000625 0.00368364 + layer.2.k_cache 0.00807812 0.92753344 + layer.2.v_cache 0.00001969 0.01042525 + layer.3.k_cache 0.06061829 3.20549125 + layer.3.v_cache 0.00001868 0.01138667 + layer.4.k_cache 0.00068906 0.21369752 + layer.4.v_cache 0.00005044 0.02222539 + layer.4.output 1.20679682 199.99401261 + ------------------------------------------------------------------------------------- + TOTAL 0.53827733 86.83948044 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 237172 +BPFP 0.8549 bits/point +EBPFP 1.7097 equivalent bits/point +MSE 86.839480 +---------------------- -------------------------------------------------------- +Time: 0.614s Load: 0.010s, Pack+Encode: 0.283s, Decode+Unpack: 0.322s +---------------------- -------------------------------------------------------- +💾 Converting with 86.8395 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 300, 128) +Output shape: (1, 300, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.output: torch.Size([1, 300, 3584]) -> torch.Size([1, 1, 300, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,536B, BPFP=0.5487 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,468B, BPFP=1.7431 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,388B, BPFP=0.8015 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,360B, BPFP=1.6854 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,664B, BPFP=0.8679 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,680B, BPFP=1.6500 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,624B, BPFP=0.8137 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,028B, BPFP=1.6681 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,820B, BPFP=1.3448 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,520B, BPFP=1.5896 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,548B, BPFP=0.2719 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13491969 65.41094401 + layer.0.v_cache 0.00001558 0.00867407 + layer.1.k_cache 0.64590566 9.97131755 + layer.1.v_cache 0.00000556 0.00345881 + layer.2.k_cache 0.01395306 0.98243306 + layer.2.v_cache 0.00001972 0.00997085 + layer.3.k_cache 0.02379198 3.86745605 + layer.3.v_cache 0.00001942 0.01145486 + layer.4.k_cache 0.00073123 0.22573011 + layer.4.v_cache 0.00005274 0.02207414 + layer.4.output 0.05011183 179.69627976 + ------------------------------------------------------------------------------------- + TOTAL 0.06883514 78.72867481 + (elements=2,611,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2611200 +Total Bytes 280636 +BPFP 0.8598 bits/point +EBPFP 1.7196 equivalent bits/point +MSE 78.728675 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.250s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 78.7287 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,064B, BPFP=0.5404 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,712B, BPFP=1.7564 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,980B, BPFP=0.8043 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,828B, BPFP=1.7090 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,144B, BPFP=0.8668 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,828B, BPFP=1.6553 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,756B, BPFP=0.7923 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,188B, BPFP=1.6746 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,980B, BPFP=1.3413 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,828B, BPFP=1.6016 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,260B, BPFP=0.2628 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14826349 63.71993798 + layer.0.v_cache 0.00001415 0.00882313 + layer.1.k_cache 0.61440683 10.12850858 + layer.1.v_cache 0.00000562 0.00368959 + layer.2.k_cache 0.00551110 0.96219543 + layer.2.v_cache 0.00001912 0.01020894 + layer.3.k_cache 0.01294636 3.67502618 + layer.3.v_cache 0.00002032 0.01156733 + layer.4.k_cache 0.00070633 0.23139395 + layer.4.v_cache 0.00005074 0.02232602 + layer.4.output 0.04972342 186.30416973 + ------------------------------------------------------------------------------------- + TOTAL 0.06647106 81.34722737 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 271568 +BPFP 0.8577 bits/point +EBPFP 1.7155 equivalent bits/point +MSE 81.347227 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 81.3472 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,012B, BPFP=0.5648 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,776B, BPFP=1.7924 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,396B, BPFP=0.8120 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,776B, BPFP=1.7360 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,412B, BPFP=0.8694 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,908B, BPFP=1.6870 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,388B, BPFP=0.8116 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,396B, BPFP=1.7146 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,164B, BPFP=1.3630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,768B, BPFP=1.6227 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,180B, BPFP=0.2915 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13861307 63.54227775 + layer.0.v_cache 0.00001419 0.00874387 + layer.1.k_cache 0.53867822 10.25641464 + layer.1.v_cache 0.00000573 0.00349794 + layer.2.k_cache 0.00871976 0.92072536 + layer.2.v_cache 0.00001877 0.00997982 + layer.3.k_cache 0.02195022 4.27984068 + layer.3.v_cache 0.00001967 0.01166981 + layer.4.k_cache 0.00069227 0.22182210 + layer.4.v_cache 0.00004961 0.02162166 + layer.4.output 0.00954030 199.75164389 + ------------------------------------------------------------------------------------- + TOTAL 0.04562021 86.91400593 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 266176 +BPFP 0.8832 bits/point +EBPFP 1.7664 equivalent bits/point +MSE 86.914006 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9140 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,116B, BPFP=0.5527 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,840B, BPFP=1.7395 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,468B, BPFP=0.7904 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,988B, BPFP=1.6930 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,852B, BPFP=0.8660 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,296B, BPFP=1.6552 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,736B, BPFP=0.8051 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,832B, BPFP=1.6844 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,824B, BPFP=1.3562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,440B, BPFP=1.6084 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,116B, BPFP=0.2663 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13928204 64.78394067 + layer.0.v_cache 0.00001373 0.00871411 + layer.1.k_cache 0.60860016 10.19892168 + layer.1.v_cache 0.00000555 0.00349351 + layer.2.k_cache 0.01026906 0.96684009 + layer.2.v_cache 0.00001872 0.01021399 + layer.3.k_cache 0.05166252 4.14724966 + layer.3.v_cache 0.00001937 0.01185433 + layer.4.k_cache 0.00073448 0.22500584 + layer.4.v_cache 0.00005026 0.02235901 + layer.4.output 0.00697322 193.86633679 + ------------------------------------------------------------------------------------- + TOTAL 0.05055697 84.55546767 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 267508 +BPFP 0.8597 bits/point +EBPFP 1.7194 equivalent bits/point +MSE 84.555468 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.011s, Pack+Encode: 0.249s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 84.5555 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,436B, BPFP=0.5528 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,100B, BPFP=1.7532 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,092B, BPFP=0.7994 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,048B, BPFP=1.6975 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,416B, BPFP=0.8695 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,212B, BPFP=1.6532 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,192B, BPFP=0.8047 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,724B, BPFP=1.6803 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,668B, BPFP=1.3595 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,380B, BPFP=1.6091 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,948B, BPFP=0.2720 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13820392 64.01295021 + layer.0.v_cache 0.00001454 0.00888620 + layer.1.k_cache 0.61684219 9.90467426 + layer.1.v_cache 0.00000570 0.00355816 + layer.2.k_cache 0.01285210 0.95151719 + layer.2.v_cache 0.00002011 0.01021202 + layer.3.k_cache 0.05980752 3.73453721 + layer.3.v_cache 0.00002263 0.01165118 + layer.4.k_cache 0.00071023 0.22863027 + layer.4.v_cache 0.00005833 0.02281989 + layer.4.output 0.05086143 183.36867433 + ------------------------------------------------------------------------------------- + TOTAL 0.06968043 80.14530335 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 277216 +BPFP 0.8637 bits/point +EBPFP 1.7274 equivalent bits/point +MSE 80.145303 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 80.1453 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,488B, BPFP=0.5444 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,672B, BPFP=1.7479 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,256B, BPFP=0.7919 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,692B, BPFP=1.6971 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,804B, BPFP=0.8723 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,912B, BPFP=1.6566 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,576B, BPFP=0.8086 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,208B, BPFP=1.6719 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,108B, BPFP=1.3553 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,964B, BPFP=1.6074 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,112B, BPFP=0.2678 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14825445 64.06499818 + layer.0.v_cache 0.00001396 0.00917890 + layer.1.k_cache 0.63210902 10.06422277 + layer.1.v_cache 0.00000581 0.00383101 + layer.2.k_cache 0.01184132 0.98177026 + layer.2.v_cache 0.00001890 0.01052285 + layer.3.k_cache 0.01767834 3.77553946 + layer.3.v_cache 0.00002009 0.01190707 + layer.4.k_cache 0.00068272 0.22835944 + layer.4.v_cache 0.00005247 0.02312189 + layer.4.output 0.04867053 179.31776519 + ------------------------------------------------------------------------------------- + TOTAL 0.06772769 78.49398871 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 281792 +BPFP 0.8605 bits/point +EBPFP 1.7209 equivalent bits/point +MSE 78.493989 +---------------------- -------------------------------------------------------- +Time: 0.642s Load: 0.011s, Pack+Encode: 0.253s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 78.4940 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,804B, BPFP=0.5211 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,980B, BPFP=1.8336 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,572B, BPFP=0.8033 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,852B, BPFP=1.7668 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,076B, BPFP=0.8923 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,100B, BPFP=1.7223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,100B, BPFP=0.8345 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,444B, BPFP=1.7427 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,896B, BPFP=1.4143 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,040B, BPFP=1.6596 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,992B, BPFP=0.2536 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14315909 66.34820372 + layer.0.v_cache 0.00001391 0.00920682 + layer.1.k_cache 0.49711875 9.70998498 + layer.1.v_cache 0.00000560 0.00374110 + layer.2.k_cache 0.00646762 1.04092615 + layer.2.v_cache 0.00001905 0.01033408 + layer.3.k_cache 0.04707547 4.00117724 + layer.3.v_cache 0.00001924 0.01186954 + layer.4.k_cache 0.00071833 0.23244161 + layer.4.v_cache 0.00005197 0.02257739 + layer.4.output 0.00997522 210.09818047 + ------------------------------------------------------------------------------------- + TOTAL 0.04496915 91.29868976 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 252856 +BPFP 0.8803 bits/point +EBPFP 1.7606 equivalent bits/point +MSE 91.298690 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 91.2987 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,124B, BPFP=0.5670 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,792B, BPFP=1.7805 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,476B, BPFP=0.8107 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,928B, BPFP=1.7321 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,792B, BPFP=0.8844 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,272B, BPFP=1.6953 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,500B, BPFP=0.8121 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,552B, BPFP=1.7110 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,212B, BPFP=1.3560 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,112B, BPFP=1.6304 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,624B, BPFP=0.2930 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15581964 63.95642921 + layer.0.v_cache 0.00001703 0.00896786 + layer.1.k_cache 0.56298232 9.92849042 + layer.1.v_cache 0.00000550 0.00352019 + layer.2.k_cache 0.00979712 1.02026050 + layer.2.v_cache 0.00002023 0.01052382 + layer.3.k_cache 0.04960944 4.02665158 + layer.3.v_cache 0.00001971 0.01184496 + layer.4.k_cache 0.00068561 0.22065952 + layer.4.v_cache 0.00005245 0.02234361 + layer.4.output 0.01122868 198.47199821 + ------------------------------------------------------------------------------------- + TOTAL 0.05044764 86.38315760 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 268384 +BPFP 0.8841 bits/point +EBPFP 1.7683 equivalent bits/point +MSE 86.383158 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 86.3832 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,036B, BPFP=0.5541 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,912B, BPFP=1.7619 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,484B, BPFP=0.7997 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,912B, BPFP=1.7067 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,644B, BPFP=0.8637 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,952B, BPFP=1.6537 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,604B, BPFP=0.8063 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,508B, BPFP=1.6844 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,316B, BPFP=1.3425 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,980B, BPFP=1.6000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,808B, BPFP=0.2982 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13599485 66.72623813 + layer.0.v_cache 0.00001418 0.00891194 + layer.1.k_cache 0.56373192 10.07573622 + layer.1.v_cache 0.00000588 0.00363948 + layer.2.k_cache 0.01549433 1.01519689 + layer.2.v_cache 0.00001867 0.01016556 + layer.3.k_cache 0.03475572 3.94075190 + layer.3.v_cache 0.00001907 0.01115112 + layer.4.k_cache 0.00072111 0.22483257 + layer.4.v_cache 0.00005295 0.02225394 + layer.4.output 0.00820773 194.46791393 + ------------------------------------------------------------------------------------- + TOTAL 0.04754487 84.90083972 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 269156 +BPFP 0.8742 bits/point +EBPFP 1.7483 equivalent bits/point +MSE 84.900840 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.250s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9008 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,700B, BPFP=0.5428 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,828B, BPFP=1.7161 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,596B, BPFP=0.7912 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,884B, BPFP=1.6682 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,056B, BPFP=0.8653 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,168B, BPFP=1.6319 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,032B, BPFP=0.8133 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,648B, BPFP=1.6562 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,328B, BPFP=1.3356 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,204B, BPFP=1.5830 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,428B, BPFP=0.2785 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13331154 61.23369014 + layer.0.v_cache 0.00001460 0.00870181 + layer.1.k_cache 0.64535681 10.06506189 + layer.1.v_cache 0.00000563 0.00358395 + layer.2.k_cache 0.00762887 0.99326850 + layer.2.v_cache 0.00001910 0.01025900 + layer.3.k_cache 0.03538683 3.62399966 + layer.3.v_cache 0.00002103 0.01186796 + layer.4.k_cache 0.00072257 0.22217768 + layer.4.v_cache 0.00005300 0.02240907 + layer.4.output 0.04522907 175.70596591 + ------------------------------------------------------------------------------------- + TOTAL 0.06700726 76.83157535 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 286872 +BPFP 0.8561 bits/point +EBPFP 1.7121 equivalent bits/point +MSE 76.831575 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.012s, Pack+Encode: 0.252s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 76.8316 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,080B, BPFP=0.5546 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,012B, BPFP=1.7612 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,648B, BPFP=0.8059 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,232B, BPFP=1.7183 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,996B, BPFP=0.8801 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,304B, BPFP=1.6673 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,744B, BPFP=0.8112 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,820B, BPFP=1.6956 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,728B, BPFP=1.3605 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,396B, BPFP=1.6173 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,832B, BPFP=0.2659 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14806804 65.55831178 + layer.0.v_cache 0.00001376 0.00883244 + layer.1.k_cache 0.60662186 10.01257238 + layer.1.v_cache 0.00000571 0.00351969 + layer.2.k_cache 0.01239752 0.95479057 + layer.2.v_cache 0.00002027 0.00986823 + layer.3.k_cache 0.02456893 3.99372326 + layer.3.v_cache 0.00001975 0.01117163 + layer.4.k_cache 0.00069718 0.22248163 + layer.4.v_cache 0.00005379 0.02197056 + layer.4.output 0.00818182 194.80547347 + ------------------------------------------------------------------------------------- + TOTAL 0.04998468 84.96679744 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 267792 +BPFP 0.8667 bits/point +EBPFP 1.7333 equivalent bits/point +MSE 84.966797 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9668 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,912B, BPFP=0.5473 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,136B, BPFP=1.7743 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,488B, BPFP=0.7999 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,196B, BPFP=1.7224 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,944B, BPFP=0.8803 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,204B, BPFP=1.6676 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,800B, BPFP=0.8171 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,792B, BPFP=1.7001 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,668B, BPFP=1.3620 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,292B, BPFP=1.6173 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,684B, BPFP=0.3051 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14592592 67.18048807 + layer.0.v_cache 0.00001391 0.00875922 + layer.1.k_cache 0.53254910 10.12161136 + layer.1.v_cache 0.00000583 0.00371798 + layer.2.k_cache 0.01029040 0.96407251 + layer.2.v_cache 0.00002009 0.01056764 + layer.3.k_cache 0.03848921 3.83145454 + layer.3.v_cache 0.00001972 0.01188719 + layer.4.k_cache 0.00069417 0.22744568 + layer.4.v_cache 0.00005427 0.02300362 + layer.4.output 0.01004909 195.26956083 + ------------------------------------------------------------------------------------- + TOTAL 0.04696507 85.25117257 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 272116 +BPFP 0.8838 bits/point +EBPFP 1.7675 equivalent bits/point +MSE 85.251173 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2512 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,936B, BPFP=0.5708 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,700B, BPFP=1.8210 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,032B, BPFP=0.8061 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,576B, BPFP=1.7564 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,204B, BPFP=0.8734 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,844B, BPFP=1.7144 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,284B, BPFP=0.8205 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,104B, BPFP=1.7293 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,048B, BPFP=1.3814 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,708B, BPFP=1.6491 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,640B, BPFP=0.2843 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15384994 65.93641573 + layer.0.v_cache 0.00001404 0.00901635 + layer.1.k_cache 0.48768111 10.25300688 + layer.1.v_cache 0.00000584 0.00376143 + layer.2.k_cache 0.00538594 1.00268802 + layer.2.v_cache 0.00001948 0.01021854 + layer.3.k_cache 0.02347840 3.85541534 + layer.3.v_cache 0.00001899 0.01105711 + layer.4.k_cache 0.00068094 0.22137457 + layer.4.v_cache 0.00005116 0.02195486 + layer.4.output 0.00909345 203.18559611 + ------------------------------------------------------------------------------------- + TOTAL 0.04322588 88.44847539 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 263076 +BPFP 0.8890 bits/point +EBPFP 1.7779 equivalent bits/point +MSE 88.448475 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.375s +---------------------- -------------------------------------------------------- +💾 Converting with 88.4485 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,156B, BPFP=0.5398 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,988B, BPFP=1.7532 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,928B, BPFP=0.7934 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,148B, BPFP=1.7085 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,504B, BPFP=0.8771 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,284B, BPFP=1.6626 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,428B, BPFP=0.8199 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,568B, BPFP=1.6777 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,656B, BPFP=1.3635 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,132B, BPFP=1.6014 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,280B, BPFP=0.2603 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13453574 61.62496678 + layer.0.v_cache 0.00001386 0.00907778 + layer.1.k_cache 0.70462929 10.08220481 + layer.1.v_cache 0.00000580 0.00366578 + layer.2.k_cache 0.00896267 0.98473550 + layer.2.v_cache 0.00001881 0.01036512 + layer.3.k_cache 0.02468209 4.13880557 + layer.3.v_cache 0.00001940 0.01172603 + layer.4.k_cache 0.00068809 0.22168719 + layer.4.v_cache 0.00005209 0.02160235 + layer.4.output 0.04953035 184.36945760 + ------------------------------------------------------------------------------------- + TOTAL 0.07178355 80.45264942 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 275072 +BPFP 0.8599 bits/point +EBPFP 1.7199 equivalent bits/point +MSE 80.452649 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 80.4526 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,556B, BPFP=0.5571 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,524B, BPFP=1.8379 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,068B, BPFP=0.8202 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,488B, BPFP=1.7775 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,292B, BPFP=0.8916 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,764B, BPFP=1.7353 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,092B, BPFP=0.8216 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,012B, BPFP=1.7498 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,136B, BPFP=1.4072 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,632B, BPFP=1.6693 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,176B, BPFP=0.2763 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150386 66.10911993 + layer.0.v_cache 0.00001401 0.00873392 + layer.1.k_cache 0.57442873 10.21175191 + layer.1.v_cache 0.00000554 0.00348316 + layer.2.k_cache 0.00779028 0.97361471 + layer.2.v_cache 0.00001875 0.01017595 + layer.3.k_cache 0.05176373 3.77009446 + layer.3.v_cache 0.00002027 0.01178826 + layer.4.k_cache 0.00069872 0.22117494 + layer.4.v_cache 0.00005023 0.02247260 + layer.4.output 0.01044900 206.39948694 + ------------------------------------------------------------------------------------- + TOTAL 0.04937866 89.77287167 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 260740 +BPFP 0.8942 bits/point +EBPFP 1.7884 equivalent bits/point +MSE 89.772872 +---------------------- -------------------------------------------------------- +Time: 0.646s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 89.7729 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.8720 bits/point +Avg EBPFP 1.7440 equivalent bits/point +Avg MSE 83.746158 +Avg Time 0.664s +------------------------ ---------------------------- diff --git a/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..b15ab6402703c0acbdc6b537812012ebf280d76e --- /dev/null +++ b/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 599 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.007_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa +Output output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa +---------------- ------------------------------------------------------------------------------------------------------------------------------ +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,016B, BPFP=0.5610 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,876B, BPFP=2.0231 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,948B, BPFP=0.9204 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,516B, BPFP=1.9561 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,340B, BPFP=0.9933 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,116B, BPFP=1.8817 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,188B, BPFP=0.9650 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,460B, BPFP=1.9457 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,340B, BPFP=1.5513 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,164B, BPFP=1.8906 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,288B, BPFP=0.2734 +⌛️ [2/4] FRONTEND: Frontend time: 0.469s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.250s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14073507 71.63523065 + layer.0.v_cache 0.00001426 0.00916135 + layer.1.k_cache 0.05042761 10.98873175 + layer.1.v_cache 0.00000517 0.00316373 + layer.2.k_cache 0.00244098 1.17507226 + layer.2.v_cache 0.00001715 0.00904393 + layer.3.k_cache 0.02726505 4.47053237 + layer.3.v_cache 0.00001796 0.01058505 + layer.4.k_cache 0.00069123 0.24772487 + layer.4.v_cache 0.00005084 0.02287188 + layer.4.output 0.16186065 644.82562713 + ------------------------------------------------------------------------------------- + TOTAL 0.07968764 270.72655928 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 89252 +BPFP 0.9766 bits/point +EBPFP 1.9532 equivalent bits/point +MSE 270.726559 +---------------------- -------------------------------------------------------- +Time: 0.724s Load: 0.005s, Pack+Encode: 0.469s, Decode+Unpack: 0.250s +---------------------- -------------------------------------------------------- +💾 Converting with 270.7266 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,908B, BPFP=0.5474 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,396B, BPFP=1.9571 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,928B, BPFP=0.9277 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,824B, BPFP=1.8494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,196B, BPFP=0.9782 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,468B, BPFP=1.7824 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,000B, BPFP=0.9413 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,708B, BPFP=1.8276 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,812B, BPFP=1.4706 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,232B, BPFP=1.7380 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,096B, BPFP=0.2984 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11392649 72.57998400 + layer.0.v_cache 0.00001670 0.00909088 + layer.1.k_cache 0.05249626 10.89042498 + layer.1.v_cache 0.00000506 0.00301917 + layer.2.k_cache 0.00416583 1.06108167 + layer.2.v_cache 0.00001711 0.00862698 + layer.3.k_cache 0.05931453 4.87221334 + layer.3.v_cache 0.00001789 0.01011291 + layer.4.k_cache 0.00069248 0.22378140 + layer.4.v_cache 0.00004729 0.02120846 + layer.4.output 0.16383441 652.70734725 + ------------------------------------------------------------------------------------- + TOTAL 0.08103179 274.03711615 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 85568 +BPFP 0.9476 bits/point +EBPFP 1.8951 equivalent bits/point +MSE 274.037116 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.005s, Pack+Encode: 0.167s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 274.0371 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,216B, BPFP=0.5346 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,204B, BPFP=1.8624 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,292B, BPFP=0.8797 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,664B, BPFP=1.7726 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,500B, BPFP=0.9142 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,296B, BPFP=1.7114 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,484B, BPFP=0.9116 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,608B, BPFP=1.7633 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,560B, BPFP=1.4229 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,152B, BPFP=1.6875 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,840B, BPFP=0.2574 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17268721 70.47057326 + layer.0.v_cache 0.00001453 0.00911924 + layer.1.k_cache 0.08276152 11.05364081 + layer.1.v_cache 0.00000545 0.00316575 + layer.2.k_cache 0.00739249 1.09020753 + layer.2.v_cache 0.00001625 0.00863721 + layer.3.k_cache 0.05929063 4.41516568 + layer.3.v_cache 0.00001766 0.01036527 + layer.4.k_cache 0.00068371 0.22142110 + layer.4.v_cache 0.00005055 0.02073045 + layer.4.output 0.14466690 576.55808321 + ------------------------------------------------------------------------------------- + TOTAL 0.07856402 242.54174169 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 91816 +BPFP 0.8978 bits/point +EBPFP 1.7955 equivalent bits/point +MSE 242.541742 +---------------------- -------------------------------------------------------- +Time: 0.379s Load: 0.004s, Pack+Encode: 0.164s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 242.5417 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 112, 128) +Output shape: (1, 112, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.output: torch.Size([1, 112, 3584]) -> torch.Size([1, 1, 112, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,912B, BPFP=0.5458 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,904B, BPFP=1.8002 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,380B, BPFP=0.8901 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,316B, BPFP=1.7182 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,640B, BPFP=0.9263 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,828B, BPFP=1.6501 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,308B, BPFP=0.8800 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,120B, BPFP=1.6908 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,840B, BPFP=1.3728 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,724B, BPFP=1.6356 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,888B, BPFP=0.2768 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17321212 65.95033482 + layer.0.v_cache 0.00001450 0.00837112 + layer.1.k_cache 0.08643453 10.48143441 + layer.1.v_cache 0.00000659 0.00327986 + layer.2.k_cache 0.00690480 0.94128404 + layer.2.v_cache 0.00001899 0.00870648 + layer.3.k_cache 0.01205242 4.38879422 + layer.3.v_cache 0.00001923 0.01018514 + layer.4.k_cache 0.00071999 0.21067573 + layer.4.v_cache 0.00004947 0.01969061 + layer.4.output 10.17892331 478.83669483 + ------------------------------------------------------------------------------------- + TOTAL 4.20775858 201.99291884 + (elements=974,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 974848 +Total Bytes 107860 +BPFP 0.8851 bits/point +EBPFP 1.7703 equivalent bits/point +MSE 201.992919 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.006s, Pack+Encode: 0.164s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 201.9929 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,156B, BPFP=0.5604 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,012B, BPFP=1.9553 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,152B, BPFP=0.9148 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,552B, BPFP=1.8736 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,340B, BPFP=0.9482 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,984B, BPFP=1.7727 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,100B, BPFP=0.9055 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,244B, BPFP=1.8189 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,248B, BPFP=1.4645 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,792B, BPFP=1.7386 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,364B, BPFP=0.2629 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11687154 67.06392045 + layer.0.v_cache 0.00001423 0.00872470 + layer.1.k_cache 0.05324238 11.03099754 + layer.1.v_cache 0.00000539 0.00324621 + layer.2.k_cache 0.00397935 1.00125287 + layer.2.v_cache 0.00001766 0.00863148 + layer.3.k_cache 0.01210797 4.61246629 + layer.3.v_cache 0.00001803 0.01077463 + layer.4.k_cache 0.00075037 0.23243130 + layer.4.v_cache 0.00004481 0.02027036 + layer.4.output 0.15448949 615.02693791 + ------------------------------------------------------------------------------------- + TOTAL 0.07461636 258.18713419 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 88944 +BPFP 0.9290 bits/point +EBPFP 1.8580 equivalent bits/point +MSE 258.187134 +---------------------- -------------------------------------------------------- +Time: 0.380s Load: 0.003s, Pack+Encode: 0.165s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 258.1871 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,876B, BPFP=0.5548 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,700B, BPFP=2.0640 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,948B, BPFP=0.9545 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,240B, BPFP=1.9753 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,376B, BPFP=1.0370 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,652B, BPFP=1.8619 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,108B, BPFP=0.9853 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,948B, BPFP=1.9190 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,076B, BPFP=1.5579 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,744B, BPFP=1.8796 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,024B, BPFP=0.3038 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203883 70.41515963 + layer.0.v_cache 0.00001438 0.00959200 + layer.1.k_cache 0.05633916 11.21450843 + layer.1.v_cache 0.00000602 0.00370692 + layer.2.k_cache 0.00244657 1.10809722 + layer.2.v_cache 0.00001847 0.00990499 + layer.3.k_cache 0.05492775 4.64529080 + layer.3.v_cache 0.00001883 0.01126962 + layer.4.k_cache 0.00062984 0.23386416 + layer.4.v_cache 0.00005178 0.02279355 + layer.4.output 0.18538215 668.58454586 + ------------------------------------------------------------------------------------- + TOTAL 0.09142157 280.45682402 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 87692 +BPFP 0.9951 bits/point +EBPFP 1.9901 equivalent bits/point +MSE 280.456824 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 280.4568 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,164B, BPFP=0.5493 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,092B, BPFP=1.9257 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,372B, BPFP=0.9326 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,544B, BPFP=1.8306 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,548B, BPFP=0.9632 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,036B, BPFP=1.7424 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,328B, BPFP=0.9250 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,432B, BPFP=1.8111 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,432B, BPFP=1.4639 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,956B, BPFP=1.7285 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,968B, BPFP=0.2720 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13028670 74.08804253 + layer.0.v_cache 0.00001462 0.00851553 + layer.1.k_cache 0.04946813 10.60405002 + layer.1.v_cache 0.00000624 0.00300670 + layer.2.k_cache 0.00411291 1.05998917 + layer.2.v_cache 0.00001729 0.00860395 + layer.3.k_cache 0.01793825 4.44022149 + layer.3.v_cache 0.00001895 0.01078439 + layer.4.k_cache 0.00077120 0.23524212 + layer.4.v_cache 0.00004532 0.02034572 + layer.4.output 0.15109633 602.01597222 + ------------------------------------------------------------------------------------- + TOTAL 0.07413847 253.21121219 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 90872 +BPFP 0.9280 bits/point +EBPFP 1.8560 equivalent bits/point +MSE 253.211212 +---------------------- -------------------------------------------------------- +Time: 0.382s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 253.2112 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,860B, BPFP=0.5099 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,136B, BPFP=1.9561 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,332B, BPFP=0.9134 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,804B, BPFP=1.8651 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,788B, BPFP=1.0384 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,412B, BPFP=1.7577 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,652B, BPFP=1.0011 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,540B, BPFP=1.7928 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,464B, BPFP=1.4978 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,416B, BPFP=1.7588 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,760B, BPFP=0.3039 +⌛️ [2/4] FRONTEND: Frontend time: 0.199s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.151s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11683119 75.67161801 + layer.0.v_cache 0.00001514 0.01093689 + layer.1.k_cache 0.01218004 11.46147825 + layer.1.v_cache 0.00000521 0.00385937 + layer.2.k_cache 0.00491684 1.08770391 + layer.2.v_cache 0.00001706 0.01053166 + layer.3.k_cache 0.10823302 5.32549192 + layer.3.v_cache 0.00002043 0.01330246 + layer.4.k_cache 0.00062874 0.25242702 + layer.4.v_cache 0.00004794 0.02532923 + layer.4.output 0.23833110 950.26323622 + ------------------------------------------------------------------------------------- + TOTAL 0.11242431 396.80619601 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 59164 +BPFP 0.9540 bits/point +EBPFP 1.9080 equivalent bits/point +MSE 396.806196 +---------------------- -------------------------------------------------------- +Time: 0.354s Load: 0.003s, Pack+Encode: 0.199s, Decode+Unpack: 0.151s +---------------------- -------------------------------------------------------- +💾 Converting with 396.8062 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,832B, BPFP=0.5725 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,552B, BPFP=2.0475 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,948B, BPFP=0.9213 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,156B, BPFP=1.9238 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,372B, BPFP=1.0537 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,888B, BPFP=1.8400 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,268B, BPFP=1.0212 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,148B, BPFP=1.9212 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,052B, BPFP=1.5788 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,016B, BPFP=1.8800 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,732B, BPFP=0.3452 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14598725 68.10476562 + layer.0.v_cache 0.00001449 0.01061882 + layer.1.k_cache 0.01347294 11.21760010 + layer.1.v_cache 0.00000521 0.00383396 + layer.2.k_cache 0.00488209 1.32630905 + layer.2.v_cache 0.00001649 0.01031708 + layer.3.k_cache 0.04421538 5.23240479 + layer.3.v_cache 0.00001905 0.01299545 + layer.4.k_cache 0.00073779 0.25388222 + layer.4.v_cache 0.00005089 0.02369459 + layer.4.output 0.27158585 1082.77401786 + ------------------------------------------------------------------------------------- + TOTAL 0.12414721 450.91850275 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 54964 +BPFP 1.0104 bits/point +EBPFP 2.0207 equivalent bits/point +MSE 450.918503 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 450.9185 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 46, 128) +Output shape: (1, 46, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.output: torch.Size([1, 46, 3584]) -> torch.Size([1, 1, 46, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,760B, BPFP=0.5978 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,048B, BPFP=2.0543 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,816B, BPFP=0.9565 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,708B, BPFP=1.9389 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,128B, BPFP=1.0625 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,328B, BPFP=1.8098 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,024B, BPFP=1.0272 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,668B, BPFP=1.9253 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,672B, BPFP=1.5870 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,612B, BPFP=1.9062 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,564B, BPFP=0.3670 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14129612 78.12264882 + layer.0.v_cache 0.00001532 0.01151861 + layer.1.k_cache 0.01315293 11.19285252 + layer.1.v_cache 0.00000553 0.00411417 + layer.2.k_cache 0.00779452 1.25390982 + layer.2.v_cache 0.00001723 0.01070531 + layer.3.k_cache 0.01828384 5.11944845 + layer.3.v_cache 0.00001944 0.01558181 + layer.4.k_cache 0.00060559 0.25920047 + layer.4.v_cache 0.00005113 0.02634019 + layer.4.output 0.29517000 1177.50242624 + ------------------------------------------------------------------------------------- + TOTAL 0.13220186 490.50195905 + (elements=400,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 400384 +Total Bytes 51328 +BPFP 1.0256 bits/point +EBPFP 2.0512 equivalent bits/point +MSE 490.501959 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 490.5020 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 110, 128) +Output shape: (1, 110, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.output: torch.Size([1, 110, 3584]) -> torch.Size([1, 1, 110, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,940B, BPFP=0.5597 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,880B, BPFP=1.8295 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,316B, BPFP=0.8972 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,200B, BPFP=1.7330 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,612B, BPFP=0.9392 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,704B, BPFP=1.6625 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,360B, BPFP=0.9034 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,028B, BPFP=1.7085 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,848B, BPFP=1.3989 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,496B, BPFP=1.6330 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,060B, BPFP=0.2853 +⌛️ [2/4] FRONTEND: Frontend time: 0.171s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14476244 65.13740234 + layer.0.v_cache 0.00001476 0.00934263 + layer.1.k_cache 0.05665370 10.45439009 + layer.1.v_cache 0.00000525 0.00314113 + layer.2.k_cache 0.00786287 1.07448432 + layer.2.v_cache 0.00001801 0.00873982 + layer.3.k_cache 0.01410385 4.77830478 + layer.3.v_cache 0.00002027 0.01073781 + layer.4.k_cache 0.00066873 0.21594275 + layer.4.v_cache 0.00004626 0.01875356 + layer.4.output 10.36401748 487.73644481 + ------------------------------------------------------------------------------------- + TOTAL 4.28072226 205.63919723 + (elements=957,440) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 957440 +Total Bytes 107444 +BPFP 0.8978 bits/point +EBPFP 1.7955 equivalent bits/point +MSE 205.639197 +---------------------- -------------------------------------------------------- +Time: 0.388s Load: 0.004s, Pack+Encode: 0.171s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 205.6392 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,824B, BPFP=0.5448 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,160B, BPFP=2.1528 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,028B, BPFP=0.9699 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,536B, BPFP=2.0324 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,592B, BPFP=1.0787 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,096B, BPFP=1.9475 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,212B, BPFP=1.0054 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,388B, BPFP=2.0039 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,388B, BPFP=1.6181 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,108B, BPFP=1.9498 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,868B, BPFP=0.3546 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10533771 74.59506414 + layer.0.v_cache 0.00001610 0.01030207 + layer.1.k_cache 0.05553475 11.76230197 + layer.1.v_cache 0.00000748 0.00413228 + layer.2.k_cache 0.02166171 1.13197138 + layer.2.v_cache 0.00002000 0.01062362 + layer.3.k_cache 0.06386927 4.87366175 + layer.3.v_cache 0.00001944 0.01238872 + layer.4.k_cache 0.00064315 0.25815333 + layer.4.v_cache 0.00005385 0.02380293 + layer.4.output 0.16801396 668.52860450 + ------------------------------------------------------------------------------------- + TOTAL 0.08372125 280.72839022 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 92200 +BPFP 1.0462 bits/point +EBPFP 2.0924 equivalent bits/point +MSE 280.728390 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.003s, Pack+Encode: 0.168s, Decode+Unpack: 0.211s +---------------------- -------------------------------------------------------- +💾 Converting with 280.7284 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 49, 128) +Output shape: (1, 49, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.output: torch.Size([1, 49, 3584]) -> torch.Size([1, 1, 49, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,804B, BPFP=0.5753 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,468B, BPFP=2.0625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,984B, BPFP=0.9515 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,168B, BPFP=1.9668 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,432B, BPFP=1.0944 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,804B, BPFP=1.8508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,132B, BPFP=0.9987 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,020B, BPFP=1.9196 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,988B, BPFP=1.5906 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,988B, BPFP=1.9094 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,860B, BPFP=0.4036 +⌛️ [2/4] FRONTEND: Frontend time: 0.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12410851 75.51642219 + layer.0.v_cache 0.00001740 0.01060263 + layer.1.k_cache 0.01401279 11.34368398 + layer.1.v_cache 0.00000579 0.00392012 + layer.2.k_cache 0.00545398 1.14060678 + layer.2.v_cache 0.00001955 0.01123333 + layer.3.k_cache 0.07546502 5.38963224 + layer.3.v_cache 0.00001978 0.01383923 + layer.4.k_cache 0.00063206 0.26618716 + layer.4.v_cache 0.00005205 0.02385987 + layer.4.output 0.27729562 1105.51047741 + ------------------------------------------------------------------------------------- + TOTAL 0.12710919 460.72313702 + (elements=426,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 426496 +Total Bytes 55648 +BPFP 1.0438 bits/point +EBPFP 2.0876 equivalent bits/point +MSE 460.723137 +---------------------- -------------------------------------------------------- +Time: 0.279s Load: 0.002s, Pack+Encode: 0.132s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 460.7231 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,868B, BPFP=0.5121 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,928B, BPFP=1.8991 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,340B, BPFP=0.9156 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,504B, BPFP=1.7829 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,720B, BPFP=1.0197 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,204B, BPFP=1.7007 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,596B, BPFP=0.9857 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,408B, BPFP=1.7566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,244B, BPFP=1.4375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,200B, BPFP=1.6996 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,020B, BPFP=0.3141 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19197876 76.82936712 + layer.0.v_cache 0.00001449 0.01031182 + layer.1.k_cache 0.01466701 11.17222087 + layer.1.v_cache 0.00000519 0.00369685 + layer.2.k_cache 0.00720649 1.16944484 + layer.2.v_cache 0.00001785 0.01017121 + layer.3.k_cache 0.10851213 4.98094418 + layer.3.v_cache 0.00001996 0.01274557 + layer.4.k_cache 0.00062085 0.24909105 + layer.4.v_cache 0.00004770 0.02244488 + layer.4.output 0.23832821 950.19901316 + ------------------------------------------------------------------------------------- + TOTAL 0.11714047 396.81491356 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 58032 +BPFP 0.9358 bits/point +EBPFP 1.8715 equivalent bits/point +MSE 396.814914 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 396.8149 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,812B, BPFP=0.5551 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,528B, BPFP=2.0000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,904B, BPFP=0.8897 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,156B, BPFP=1.8860 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,428B, BPFP=1.0502 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,936B, BPFP=1.8186 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,176B, BPFP=0.9730 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,212B, BPFP=1.9032 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,092B, BPFP=1.5600 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,056B, BPFP=1.8554 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,732B, BPFP=0.3384 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12665685 76.78828221 + layer.0.v_cache 0.00001436 0.01038577 + layer.1.k_cache 0.01497617 11.46590528 + layer.1.v_cache 0.00000537 0.00371932 + layer.2.k_cache 0.00986420 1.16240169 + layer.2.v_cache 0.00001802 0.01036681 + layer.3.k_cache 0.09846177 4.64446603 + layer.3.v_cache 0.00001923 0.01305330 + layer.4.k_cache 0.00062397 0.23735701 + layer.4.v_cache 0.00005014 0.02237644 + layer.4.output 0.26631049 1061.82755602 + ------------------------------------------------------------------------------------- + TOTAL 0.12440374 442.77360035 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 55032 +BPFP 0.9918 bits/point +EBPFP 1.9836 equivalent bits/point +MSE 442.773600 +---------------------- -------------------------------------------------------- +Time: 0.280s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 442.7736 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,788B, BPFP=0.5820 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,576B, BPFP=2.1406 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,900B, BPFP=0.9440 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,196B, BPFP=2.0169 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,380B, BPFP=1.1003 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,900B, BPFP=1.9206 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,212B, BPFP=1.0456 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,184B, BPFP=2.0130 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,060B, BPFP=1.6471 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,056B, BPFP=1.9714 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,408B, BPFP=0.3910 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18531116 77.27261861 + layer.0.v_cache 0.00001874 0.01174228 + layer.1.k_cache 0.01459585 11.52832667 + layer.1.v_cache 0.00000674 0.00483283 + layer.2.k_cache 0.01042949 1.21789598 + layer.2.v_cache 0.00001841 0.01205233 + layer.3.k_cache 0.15260293 5.33760071 + layer.3.v_cache 0.00002390 0.01695926 + layer.4.k_cache 0.00061422 0.27385801 + layer.4.v_cache 0.00004881 0.02657913 + layer.4.output 0.28297247 1127.44140625 + ------------------------------------------------------------------------------------- + TOTAL 0.13791044 469.87013586 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 55660 +BPFP 1.0658 bits/point +EBPFP 2.1316 equivalent bits/point +MSE 469.870136 +---------------------- -------------------------------------------------------- +Time: 0.278s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 469.8701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,848B, BPFP=0.4894 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,088B, BPFP=1.8771 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,496B, BPFP=0.9258 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,856B, BPFP=1.8157 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,024B, BPFP=1.0657 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,552B, BPFP=1.7352 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,792B, BPFP=1.0042 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,772B, BPFP=1.7934 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,604B, BPFP=1.4841 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,572B, BPFP=1.7405 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,824B, BPFP=0.3338 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12052289 83.36357918 + layer.0.v_cache 0.00001546 0.01116356 + layer.1.k_cache 0.01434235 11.30683821 + layer.1.v_cache 0.00000574 0.00396109 + layer.2.k_cache 0.00254285 1.18024315 + layer.2.v_cache 0.00001710 0.01067994 + layer.3.k_cache 0.03737745 4.98301154 + layer.3.v_cache 0.00001946 0.01451161 + layer.4.k_cache 0.00063766 0.26490570 + layer.4.v_cache 0.00005039 0.02516767 + layer.4.output 0.23032015 917.56787228 + ------------------------------------------------------------------------------------- + TOTAL 0.10516308 383.77289221 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 61428 +BPFP 0.9569 bits/point +EBPFP 1.9139 equivalent bits/point +MSE 383.772892 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 383.7729 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,500B, BPFP=0.5279 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,560B, BPFP=2.2297 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,788B, BPFP=1.0110 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,836B, BPFP=2.0769 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,196B, BPFP=1.0971 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,640B, BPFP=2.0355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,780B, BPFP=1.0093 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,872B, BPFP=2.0845 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,896B, BPFP=1.6672 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,584B, BPFP=2.0236 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,108B, BPFP=0.3351 +⌛️ [2/4] FRONTEND: Frontend time: 0.221s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11335979 69.84125581 + layer.0.v_cache 0.00001513 0.01084547 + layer.1.k_cache 0.01655370 10.95706094 + layer.1.v_cache 0.00000667 0.00411882 + layer.2.k_cache 0.00428703 1.16234372 + layer.2.v_cache 0.00001789 0.01067264 + layer.3.k_cache 0.04938591 5.11237974 + layer.3.v_cache 0.00002111 0.01407722 + layer.4.k_cache 0.00066117 0.27569498 + layer.4.v_cache 0.00004716 0.02439991 + layer.4.output 0.18372514 732.34127654 + ------------------------------------------------------------------------------------- + TOTAL 0.08649597 306.69422265 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 85760 +BPFP 1.0652 bits/point +EBPFP 2.1304 equivalent bits/point +MSE 306.694223 +---------------------- -------------------------------------------------------- +Time: 0.438s Load: 0.004s, Pack+Encode: 0.221s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 306.6942 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,692B, BPFP=0.6009 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,016B, BPFP=2.1364 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,812B, BPFP=0.9986 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,660B, BPFP=2.0099 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,216B, BPFP=1.1420 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,436B, BPFP=1.9304 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,012B, BPFP=1.0696 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,680B, BPFP=2.0170 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,648B, BPFP=1.6506 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,516B, BPFP=1.9588 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,636B, BPFP=0.4381 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14777968 83.29393976 + layer.0.v_cache 0.00001531 0.01320410 + layer.1.k_cache 0.01728138 11.75331254 + layer.1.v_cache 0.00000547 0.00455208 + layer.2.k_cache 0.00823456 1.36740424 + layer.2.v_cache 0.00001932 0.01305793 + layer.3.k_cache 0.16770924 6.66814908 + layer.3.v_cache 0.00002046 0.01630474 + layer.4.k_cache 0.00061569 0.28609659 + layer.4.v_cache 0.00005592 0.03015854 + layer.4.output 0.30867824 1229.68881899 + ------------------------------------------------------------------------------------- + TOTAL 0.14720498 512.42752427 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 52324 +BPFP 1.0930 bits/point +EBPFP 2.1860 equivalent bits/point +MSE 512.427524 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.002s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 512.4275 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,656B, BPFP=0.5321 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,140B, BPFP=2.2316 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,956B, BPFP=0.9928 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,660B, BPFP=2.1354 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,344B, BPFP=1.0705 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,968B, BPFP=1.9968 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,080B, BPFP=1.0176 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,272B, BPFP=2.0577 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,136B, BPFP=1.6298 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,840B, BPFP=1.9712 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,120B, BPFP=0.3182 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14408422 75.50801282 + layer.0.v_cache 0.00002097 0.01174505 + layer.1.k_cache 0.05735342 11.17361841 + layer.1.v_cache 0.00000575 0.00455888 + layer.2.k_cache 0.01678689 1.21159402 + layer.2.v_cache 0.00001894 0.01212644 + layer.3.k_cache 0.03029843 5.30332008 + layer.3.v_cache 0.00002125 0.01509750 + layer.4.k_cache 0.00062582 0.28307103 + layer.4.v_cache 0.00005050 0.02692686 + layer.4.output 0.17433185 694.56192766 + ------------------------------------------------------------------------------------- + TOTAL 0.08644642 291.49903322 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 89172 +BPFP 1.0508 bits/point +EBPFP 2.1015 equivalent bits/point +MSE 291.499033 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.003s, Pack+Encode: 0.168s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 291.4990 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,708B, BPFP=0.6065 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,896B, BPFP=2.0938 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,800B, BPFP=0.9943 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 5,560B, BPFP=1.9744 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,068B, BPFP=1.0895 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,280B, BPFP=1.8750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,920B, BPFP=1.0369 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,536B, BPFP=1.9659 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,512B, BPFP=1.6023 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,368B, BPFP=1.9062 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,352B, BPFP=0.3730 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13243668 85.07096169 + layer.0.v_cache 0.00001576 0.01156712 + layer.1.k_cache 0.01981235 11.08774983 + layer.1.v_cache 0.00000577 0.00433150 + layer.2.k_cache 0.01756107 1.18677113 + layer.2.v_cache 0.00001728 0.01206431 + layer.3.k_cache 0.04920139 4.53381694 + layer.3.v_cache 0.00001837 0.01426318 + layer.4.k_cache 0.00061593 0.29296728 + layer.4.v_cache 0.00004907 0.02813931 + layer.4.output 0.30854983 1228.68536932 + ------------------------------------------------------------------------------------- + TOTAL 0.13997544 511.94354221 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 50000 +BPFP 1.0445 bits/point +EBPFP 2.0889 equivalent bits/point +MSE 511.943542 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.002s, Pack+Encode: 0.130s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 511.9435 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,824B, BPFP=0.5700 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,744B, BPFP=2.1075 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,984B, BPFP=0.9325 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,284B, BPFP=1.9638 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,444B, BPFP=1.0762 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,048B, BPFP=1.8900 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,288B, BPFP=1.0275 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,256B, BPFP=1.9550 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,104B, BPFP=1.5950 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,156B, BPFP=1.9238 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,928B, BPFP=0.3539 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20666794 69.14031738 + layer.0.v_cache 0.00002565 0.01123166 + layer.1.k_cache 0.01518770 11.39820557 + layer.1.v_cache 0.00000581 0.00405106 + layer.2.k_cache 0.01031687 1.23865814 + layer.2.v_cache 0.00001873 0.01154703 + layer.3.k_cache 0.09638391 4.74302887 + layer.3.v_cache 0.00001967 0.01490261 + layer.4.k_cache 0.00062172 0.26393038 + layer.4.v_cache 0.00004827 0.02532690 + layer.4.output 0.27162739 1082.86848214 + ------------------------------------------------------------------------------------- + TOTAL 0.13121694 450.99591615 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 56060 +BPFP 1.0305 bits/point +EBPFP 2.0610 equivalent bits/point +MSE 450.995916 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.002s, Pack+Encode: 0.130s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 450.9959 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,100B, BPFP=0.4825 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,072B, BPFP=2.3143 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,100B, BPFP=0.9421 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,472B, BPFP=2.1765 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,808B, BPFP=1.1048 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,888B, BPFP=2.0423 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,564B, BPFP=1.0487 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,108B, BPFP=2.0928 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,396B, BPFP=1.6994 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,056B, BPFP=2.0809 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,816B, BPFP=0.3550 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13431002 78.18803855 + layer.0.v_cache 0.00001720 0.01163468 + layer.1.k_cache 0.01179634 11.06993552 + layer.1.v_cache 0.00000537 0.00392342 + layer.2.k_cache 0.00798416 1.24309843 + layer.2.v_cache 0.00001783 0.01102672 + layer.3.k_cache 0.05795463 4.54612193 + layer.3.v_cache 0.00001974 0.01397621 + layer.4.k_cache 0.00062059 0.26425906 + layer.4.v_cache 0.00005204 0.02615998 + layer.4.output 0.19989206 796.96579569 + ------------------------------------------------------------------------------------- + TOTAL 0.09482484 333.77286731 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 80380 +BPFP 1.0865 bits/point +EBPFP 2.1729 equivalent bits/point +MSE 333.772867 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 333.7729 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,832B, BPFP=0.5725 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,592B, BPFP=2.0600 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,968B, BPFP=0.9275 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,280B, BPFP=1.9625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,420B, BPFP=1.0688 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,040B, BPFP=1.8875 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,292B, BPFP=1.0288 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,292B, BPFP=1.9663 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,096B, BPFP=1.5925 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,228B, BPFP=1.9463 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,996B, BPFP=0.3570 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15467889 66.89653320 + layer.0.v_cache 0.00001741 0.01122018 + layer.1.k_cache 0.01620592 11.12204346 + layer.1.v_cache 0.00000548 0.00397760 + layer.2.k_cache 0.00723796 1.30504990 + layer.2.v_cache 0.00001737 0.01122378 + layer.3.k_cache 0.06661982 5.13412476 + layer.3.v_cache 0.00002040 0.01405685 + layer.4.k_cache 0.00058776 0.25633244 + layer.4.v_cache 0.00004656 0.02532009 + layer.4.output 0.27162074 1082.31357143 + ------------------------------------------------------------------------------------- + TOTAL 0.12628134 450.64558131 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 56036 +BPFP 1.0301 bits/point +EBPFP 2.0601 equivalent bits/point +MSE 450.645581 +---------------------- -------------------------------------------------------- +Time: 0.279s Load: 0.002s, Pack+Encode: 0.131s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 450.6456 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,808B, BPFP=0.5539 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,580B, BPFP=2.0159 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,896B, BPFP=0.8873 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,100B, BPFP=1.8689 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,360B, BPFP=1.0294 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,880B, BPFP=1.8015 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,232B, BPFP=0.9902 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,180B, BPFP=1.8934 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,024B, BPFP=1.5392 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,984B, BPFP=1.8333 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,376B, BPFP=0.3228 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11261671 75.76917701 + layer.0.v_cache 0.00001829 0.01076259 + layer.1.k_cache 0.01190703 11.49882238 + layer.1.v_cache 0.00000528 0.00374267 + layer.2.k_cache 0.00973116 1.36325043 + layer.2.v_cache 0.00001668 0.01017982 + layer.3.k_cache 0.04200072 4.49884422 + layer.3.v_cache 0.00001874 0.01343017 + layer.4.k_cache 0.00062040 0.25132723 + layer.4.v_cache 0.00004776 0.02413161 + layer.4.output 0.26626884 1061.89810924 + ------------------------------------------------------------------------------------- + TOTAL 0.12005086 442.74884899 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 54420 +BPFP 0.9808 bits/point +EBPFP 1.9615 equivalent bits/point +MSE 442.748849 +---------------------- -------------------------------------------------------- +Time: 0.279s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 442.7488 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,796B, BPFP=0.5397 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,820B, BPFP=2.0493 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,952B, BPFP=0.8870 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,456B, BPFP=1.9399 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,516B, BPFP=1.0565 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,196B, BPFP=1.8618 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,452B, BPFP=1.0373 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,436B, BPFP=1.9339 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,348B, BPFP=1.6070 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,220B, BPFP=1.8690 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,324B, BPFP=0.3573 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.152s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12932784 75.26425875 + layer.0.v_cache 0.00001491 0.01163684 + layer.1.k_cache 0.01488123 10.90786156 + layer.1.v_cache 0.00000563 0.00420452 + layer.2.k_cache 0.00939683 1.27592923 + layer.2.v_cache 0.00001981 0.01147636 + layer.3.k_cache 0.07051506 4.55823810 + layer.3.v_cache 0.00001927 0.01417084 + layer.4.k_cache 0.00064413 0.27368824 + layer.4.v_cache 0.00005939 0.02608611 + layer.4.output 0.26127321 1041.21514423 + ------------------------------------------------------------------------------------- + TOTAL 0.12081156 434.16785648 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 57516 +BPFP 1.0166 bits/point +EBPFP 2.0332 equivalent bits/point +MSE 434.167856 +---------------------- -------------------------------------------------------- +Time: 0.287s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.152s +---------------------- -------------------------------------------------------- +💾 Converting with 434.1679 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,828B, BPFP=0.5493 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,892B, BPFP=2.0709 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,204B, BPFP=0.9627 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,660B, BPFP=2.0012 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,812B, BPFP=1.1454 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,352B, BPFP=1.9087 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,584B, BPFP=1.0769 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,468B, BPFP=1.9435 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,336B, BPFP=1.6034 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,296B, BPFP=1.8918 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,136B, BPFP=0.3492 +⌛️ [2/4] FRONTEND: Frontend time: 0.139s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.180s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13190371 76.61653489 + layer.0.v_cache 0.00001552 0.01190870 + layer.1.k_cache 0.01668521 10.66600389 + layer.1.v_cache 0.00000581 0.00440307 + layer.2.k_cache 0.00490881 1.34400148 + layer.2.v_cache 0.00001873 0.01146998 + layer.3.k_cache 0.11824917 5.45749840 + layer.3.v_cache 0.00002115 0.01433655 + layer.4.k_cache 0.00061206 0.25726920 + layer.4.v_cache 0.00004844 0.02368779 + layer.4.output 0.26122982 1040.99476305 + ------------------------------------------------------------------------------------- + TOTAL 0.12359279 434.19826208 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 58568 +BPFP 1.0352 bits/point +EBPFP 2.0704 equivalent bits/point +MSE 434.198262 +---------------------- -------------------------------------------------------- +Time: 0.321s Load: 0.002s, Pack+Encode: 0.139s, Decode+Unpack: 0.180s +---------------------- -------------------------------------------------------- +💾 Converting with 434.1983 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,088B, BPFP=0.5245 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,364B, BPFP=1.9300 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,680B, BPFP=0.9647 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,816B, BPFP=1.8370 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,660B, BPFP=0.9613 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,432B, BPFP=1.7717 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,436B, BPFP=0.9232 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,808B, BPFP=1.8356 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,500B, BPFP=1.4436 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,252B, BPFP=1.7412 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,412B, BPFP=0.3011 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10326390 70.95975395 + layer.0.v_cache 0.00001541 0.00949711 + layer.1.k_cache 0.04976616 10.80632881 + layer.1.v_cache 0.00000534 0.00316750 + layer.2.k_cache 0.01168035 1.10617165 + layer.2.v_cache 0.00001798 0.00888759 + layer.3.k_cache 0.02907468 5.19960586 + layer.3.v_cache 0.00001946 0.01177398 + layer.4.k_cache 0.00073861 0.22980139 + layer.4.v_cache 0.00005077 0.02158855 + layer.4.output 0.14788401 588.39649651 + ------------------------------------------------------------------------------------- + TOTAL 0.07234240 247.47835600 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 94448 +BPFP 0.9436 bits/point +EBPFP 1.8871 equivalent bits/point +MSE 247.478356 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.004s, Pack+Encode: 0.167s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 247.4784 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 64, 128) +Output shape: (1, 64, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.output: torch.Size([1, 64, 3584]) -> torch.Size([1, 1, 64, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,572B, BPFP=0.3838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,812B, BPFP=1.6631 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,812B, BPFP=0.6865 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,500B, BPFP=1.5869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,180B, BPFP=0.7764 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,296B, BPFP=1.5371 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,980B, BPFP=0.7275 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,448B, BPFP=1.5742 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,168B, BPFP=1.2617 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,176B, BPFP=1.5078 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,820B, BPFP=0.3425 +⌛️ [2/4] FRONTEND: Frontend time: 0.134s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.147s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09722494 50.33198929 + layer.0.v_cache 0.00001391 0.00875190 + layer.1.k_cache 0.01806639 9.19944668 + layer.1.v_cache 0.00000557 0.00337588 + layer.2.k_cache 0.00423524 0.74873149 + layer.2.v_cache 0.00001868 0.01002040 + layer.3.k_cache 0.03759564 3.99914074 + layer.3.v_cache 0.00001825 0.01069426 + layer.4.k_cache 0.00061753 0.20643684 + layer.4.v_cache 0.00004894 0.02170564 + layer.4.output 0.21408452 813.45514788 + ------------------------------------------------------------------------------------- + TOTAL 0.09743746 338.74860755 + (elements=557,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 557056 +Total Bytes 57764 +BPFP 0.8296 bits/point +EBPFP 1.6591 equivalent bits/point +MSE 338.748608 +---------------------- -------------------------------------------------------- +Time: 0.284s Load: 0.003s, Pack+Encode: 0.134s, Decode+Unpack: 0.147s +---------------------- -------------------------------------------------------- +💾 Converting with 338.7486 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,232B, BPFP=0.4951 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,316B, BPFP=1.8866 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,080B, BPFP=0.9314 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,700B, BPFP=1.7923 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,392B, BPFP=0.9792 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,320B, BPFP=1.7341 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,956B, BPFP=0.9124 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,604B, BPFP=1.7776 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,700B, BPFP=1.4859 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,148B, BPFP=1.7077 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,128B, BPFP=0.2654 +⌛️ [2/4] FRONTEND: Frontend time: 0.180s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.216s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11648470 67.00045956 + layer.0.v_cache 0.00001454 0.00883395 + layer.1.k_cache 0.02721847 10.10819977 + layer.1.v_cache 0.00000531 0.00319105 + layer.2.k_cache 0.01045542 1.12346634 + layer.2.v_cache 0.00001699 0.00854656 + layer.3.k_cache 0.01744144 3.84373743 + layer.3.v_cache 0.00001839 0.01034748 + layer.4.k_cache 0.00076342 0.22537173 + layer.4.v_cache 0.00004795 0.02073539 + layer.4.output 11.17676328 526.36725315 + ------------------------------------------------------------------------------------- + TOTAL 4.61234174 221.58374478 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 101576 +BPFP 0.9153 bits/point +EBPFP 1.8306 equivalent bits/point +MSE 221.583745 +---------------------- -------------------------------------------------------- +Time: 0.401s Load: 0.006s, Pack+Encode: 0.180s, Decode+Unpack: 0.216s +---------------------- -------------------------------------------------------- +💾 Converting with 221.5837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,776B, BPFP=0.5422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,552B, BPFP=2.0609 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,924B, BPFP=0.9617 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,000B, BPFP=1.9531 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,200B, BPFP=1.0156 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,584B, BPFP=1.8719 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,072B, BPFP=0.9906 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,756B, BPFP=1.9055 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,900B, BPFP=1.5430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,440B, BPFP=1.8438 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,500B, BPFP=0.2930 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11433966 70.50803223 + layer.0.v_cache 0.00001437 0.00853621 + layer.1.k_cache 0.05315235 10.86253204 + layer.1.v_cache 0.00000524 0.00282450 + layer.2.k_cache 0.00678293 1.14107761 + layer.2.v_cache 0.00001704 0.00843285 + layer.3.k_cache 0.01951189 4.89236450 + layer.3.v_cache 0.00001758 0.01024514 + layer.4.k_cache 0.00076624 0.22733111 + layer.4.v_cache 0.00004856 0.02039197 + layer.4.output 0.16992235 677.21941964 + ------------------------------------------------------------------------------------- + TOTAL 0.08141837 284.01280622 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 85704 +BPFP 0.9847 bits/point +EBPFP 1.9693 equivalent bits/point +MSE 284.012806 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 284.0128 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,060B, BPFP=0.5312 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,252B, BPFP=1.9535 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,416B, BPFP=0.9403 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,664B, BPFP=1.8514 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,504B, BPFP=0.9556 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,284B, BPFP=1.7854 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,332B, BPFP=0.9257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,692B, BPFP=1.8562 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,504B, BPFP=1.4764 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,200B, BPFP=1.7708 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,536B, BPFP=0.2861 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11570581 72.05136719 + layer.0.v_cache 0.00001517 0.00923331 + layer.1.k_cache 0.03236232 10.82663303 + layer.1.v_cache 0.00000662 0.00314009 + layer.2.k_cache 0.00641356 1.03905987 + layer.2.v_cache 0.00001676 0.00874276 + layer.3.k_cache 0.02648412 4.27400309 + layer.3.v_cache 0.00001896 0.01123490 + layer.4.k_cache 0.00074524 0.23263715 + layer.4.v_cache 0.00004843 0.02150149 + layer.4.output 0.15113260 601.74241071 + ------------------------------------------------------------------------------------- + TOTAL 0.07292619 252.98084870 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 92444 +BPFP 0.9441 bits/point +EBPFP 1.8882 equivalent bits/point +MSE 252.980849 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 252.9808 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 114, 128) +Output shape: (1, 114, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.output: torch.Size([1, 114, 3584]) -> torch.Size([1, 1, 114, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,864B, BPFP=0.5296 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,868B, BPFP=1.7637 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,432B, BPFP=0.8816 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,112B, BPFP=1.6601 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,596B, BPFP=0.9041 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,660B, BPFP=1.5981 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,232B, BPFP=0.8542 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,944B, BPFP=1.6371 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,900B, BPFP=1.3569 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,468B, BPFP=1.5718 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,856B, BPFP=0.2713 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10081583 62.19887610 + layer.0.v_cache 0.00001633 0.00864409 + layer.1.k_cache 0.08473034 10.17061361 + layer.1.v_cache 0.00000563 0.00306841 + layer.2.k_cache 0.00792928 0.95052177 + layer.2.v_cache 0.00001730 0.00823873 + layer.3.k_cache 0.01019665 4.23072520 + layer.3.v_cache 0.00001810 0.00919645 + layer.4.k_cache 0.00077074 0.21071683 + layer.4.v_cache 0.00005022 0.02007778 + layer.4.output 10.00037814 470.69223841 + ------------------------------------------------------------------------------------- + TOTAL 4.12983514 198.39154987 + (elements=992,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 992256 +Total Bytes 106932 +BPFP 0.8621 bits/point +EBPFP 1.7243 equivalent bits/point +MSE 198.391550 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.006s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 198.3915 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 98, 128) +Output shape: (1, 98, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.output: torch.Size([1, 98, 3584]) -> torch.Size([1, 1, 98, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,992B, BPFP=0.4770 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,592B, BPFP=1.8482 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,788B, BPFP=0.9228 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,116B, BPFP=1.7723 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,820B, BPFP=0.9279 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,736B, BPFP=1.7117 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,432B, BPFP=0.8661 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,000B, BPFP=1.7538 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,972B, BPFP=1.4305 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,588B, BPFP=1.6881 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,080B, BPFP=0.2524 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11008905 73.68257733 + layer.0.v_cache 0.00001718 0.00864757 + layer.1.k_cache 0.06240322 10.19095285 + layer.1.v_cache 0.00000548 0.00301286 + layer.2.k_cache 0.00473963 0.99109899 + layer.2.v_cache 0.00001718 0.00878553 + layer.3.k_cache 0.01141277 4.59521235 + layer.3.v_cache 0.00001769 0.00965208 + layer.4.k_cache 0.00076610 0.23306084 + layer.4.v_cache 0.00004814 0.02063442 + layer.4.output 0.02426045 568.03024781 + ------------------------------------------------------------------------------------- + TOTAL 0.02113762 239.17384526 + (elements=852,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 852992 +Total Bytes 95116 +BPFP 0.8921 bits/point +EBPFP 1.7841 equivalent bits/point +MSE 239.173845 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 239.1738 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,100B, BPFP=0.5265 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,256B, BPFP=1.9117 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,660B, BPFP=0.9613 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,672B, BPFP=1.8125 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,980B, BPFP=1.0156 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,404B, BPFP=1.7670 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,644B, BPFP=0.9586 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,744B, BPFP=1.8247 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,688B, BPFP=1.4755 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,236B, BPFP=1.7385 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,312B, BPFP=0.2987 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08919133 72.32308297 + layer.0.v_cache 0.00001414 0.00854763 + layer.1.k_cache 0.08591949 11.05445265 + layer.1.v_cache 0.00000528 0.00298481 + layer.2.k_cache 0.00389855 1.10929987 + layer.2.v_cache 0.00001738 0.00909671 + layer.3.k_cache 0.07679920 4.55143937 + layer.3.v_cache 0.00002028 0.01120388 + layer.4.k_cache 0.00072407 0.22894018 + layer.4.v_cache 0.00004779 0.02154626 + layer.4.output 0.14786680 588.26470303 + ------------------------------------------------------------------------------------- + TOTAL 0.07598265 247.48079503 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 94696 +BPFP 0.9461 bits/point +EBPFP 1.8921 equivalent bits/point +MSE 247.480795 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 247.4808 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,808B, BPFP=0.5561 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,920B, BPFP=1.8867 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,364B, BPFP=0.9293 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,328B, BPFP=1.8002 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,852B, BPFP=1.0006 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,976B, BPFP=1.7488 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,356B, BPFP=0.9282 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,232B, BPFP=1.7862 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,080B, BPFP=1.4720 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,748B, BPFP=1.7155 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,144B, BPFP=0.2742 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11353527 67.82264895 + layer.0.v_cache 0.00001674 0.00902596 + layer.1.k_cache 0.04276301 10.43324123 + layer.1.v_cache 0.00000542 0.00308646 + layer.2.k_cache 0.00247095 1.05725939 + layer.2.v_cache 0.00001885 0.00905609 + layer.3.k_cache 0.02722364 3.98580419 + layer.3.v_cache 0.00001942 0.01006971 + layer.4.k_cache 0.00068256 0.21857804 + layer.4.v_cache 0.00004843 0.02125455 + layer.4.output 10.65454330 500.87583445 + ------------------------------------------------------------------------------------- + TOTAL 4.39815220 211.15887445 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 107808 +BPFP 0.9261 bits/point +EBPFP 1.8521 equivalent bits/point +MSE 211.158874 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.218s +---------------------- -------------------------------------------------------- +💾 Converting with 211.1589 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,032B, BPFP=0.5640 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,760B, BPFP=2.0015 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,900B, BPFP=0.9115 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,084B, BPFP=1.8757 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,232B, BPFP=0.9732 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,608B, BPFP=1.7872 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,996B, BPFP=0.9293 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,964B, BPFP=1.8534 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,948B, BPFP=1.4784 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,464B, BPFP=1.7604 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,600B, BPFP=0.2817 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.254s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12550225 70.20701381 + layer.0.v_cache 0.00001468 0.00922149 + layer.1.k_cache 0.05234316 10.78762381 + layer.1.v_cache 0.00000508 0.00291750 + layer.2.k_cache 0.00733058 1.11238062 + layer.2.v_cache 0.00001689 0.00850079 + layer.3.k_cache 0.01592888 4.68177577 + layer.3.v_cache 0.00001847 0.01077686 + layer.4.k_cache 0.00069460 0.22833048 + layer.4.v_cache 0.00005205 0.02104850 + layer.4.output 0.16184913 645.02662628 + ------------------------------------------------------------------------------------- + TOTAL 0.07852062 270.72093962 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 86588 +BPFP 0.9474 bits/point +EBPFP 1.8949 equivalent bits/point +MSE 270.720940 +---------------------- -------------------------------------------------------- +Time: 0.418s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.254s +---------------------- -------------------------------------------------------- +💾 Converting with 270.7209 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,680B, BPFP=0.5369 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,304B, BPFP=2.0641 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,728B, BPFP=0.9471 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,716B, BPFP=1.9463 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,932B, BPFP=0.9880 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,080B, BPFP=1.8189 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,812B, BPFP=0.9639 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,464B, BPFP=1.8958 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,796B, BPFP=1.5617 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,332B, BPFP=1.8694 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,440B, BPFP=0.2988 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.251s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11879376 75.65630634 + layer.0.v_cache 0.00001438 0.00917774 + layer.1.k_cache 0.05518859 11.42255891 + layer.1.v_cache 0.00000514 0.00347655 + layer.2.k_cache 0.00243519 1.18659152 + layer.2.v_cache 0.00001742 0.00934994 + layer.3.k_cache 0.07959401 5.02716064 + layer.3.v_cache 0.00001812 0.01105728 + layer.4.k_cache 0.00069358 0.25648005 + layer.4.v_cache 0.00004620 0.02229625 + layer.4.output 0.17426339 694.84615385 + ------------------------------------------------------------------------------------- + TOTAL 0.08686177 291.61926660 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 83284 +BPFP 0.9814 bits/point +EBPFP 1.9628 equivalent bits/point +MSE 291.619267 +---------------------- -------------------------------------------------------- +Time: 0.507s Load: 0.003s, Pack+Encode: 0.253s, Decode+Unpack: 0.251s +---------------------- -------------------------------------------------------- +💾 Converting with 291.6193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,992B, BPFP=0.5701 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,700B, BPFP=2.0389 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,216B, BPFP=0.9939 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,272B, BPFP=1.9573 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,344B, BPFP=1.0183 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,808B, BPFP=1.8689 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,164B, BPFP=0.9840 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,196B, BPFP=1.9428 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,200B, BPFP=1.5625 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,704B, BPFP=1.8491 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,316B, BPFP=0.2808 +⌛️ [2/4] FRONTEND: Frontend time: 0.237s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.303s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11325797 68.27722466 + layer.0.v_cache 0.00001565 0.00904132 + layer.1.k_cache 0.07651102 11.47499419 + layer.1.v_cache 0.00000530 0.00338911 + layer.2.k_cache 0.00872486 1.05323010 + layer.2.v_cache 0.00001696 0.00919529 + layer.3.k_cache 0.12876377 4.99949534 + layer.3.v_cache 0.00001905 0.01125449 + layer.4.k_cache 0.00069190 0.23686107 + layer.4.v_cache 0.00004968 0.02345690 + layer.4.output 0.16578155 660.83013937 + ------------------------------------------------------------------------------------- + TOTAL 0.08756041 277.17112459 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 87912 +BPFP 0.9854 bits/point +EBPFP 1.9708 equivalent bits/point +MSE 277.171125 +---------------------- -------------------------------------------------------- +Time: 0.543s Load: 0.003s, Pack+Encode: 0.237s, Decode+Unpack: 0.303s +---------------------- -------------------------------------------------------- +💾 Converting with 277.1711 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,596B, BPFP=0.5268 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,208B, BPFP=2.0714 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,892B, BPFP=0.9927 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,556B, BPFP=1.9391 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,064B, BPFP=1.0276 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,140B, BPFP=1.8547 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,852B, BPFP=0.9846 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,280B, BPFP=1.8831 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,784B, BPFP=1.5795 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,200B, BPFP=1.8669 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,060B, BPFP=0.2916 +⌛️ [2/4] FRONTEND: Frontend time: 0.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.288s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12835005 76.44376522 + layer.0.v_cache 0.00001616 0.00958045 + layer.1.k_cache 0.05672838 11.13673223 + layer.1.v_cache 0.00000535 0.00387865 + layer.2.k_cache 0.00240846 1.21577553 + layer.2.v_cache 0.00001644 0.01013525 + layer.3.k_cache 0.04570918 4.75531640 + layer.3.v_cache 0.00001733 0.01176010 + layer.4.k_cache 0.00072090 0.25420221 + layer.4.v_cache 0.00004908 0.02418821 + layer.4.output 0.18444509 702.70193646 + ------------------------------------------------------------------------------------- + TOTAL 0.08971394 294.86934644 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 82632 +BPFP 0.9863 bits/point +EBPFP 1.9727 equivalent bits/point +MSE 294.869346 +---------------------- -------------------------------------------------------- +Time: 0.520s Load: 0.004s, Pack+Encode: 0.228s, Decode+Unpack: 0.288s +---------------------- -------------------------------------------------------- +💾 Converting with 294.8693 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,340B, BPFP=0.5150 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,552B, BPFP=2.1021 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,480B, BPFP=0.9859 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,208B, BPFP=2.0264 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,744B, BPFP=1.0440 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,676B, BPFP=1.9093 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,428B, BPFP=0.9745 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,012B, BPFP=1.9833 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,260B, BPFP=1.5977 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,896B, BPFP=1.9577 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,780B, BPFP=0.3075 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10085608 72.00253081 + layer.0.v_cache 0.00001554 0.00898683 + layer.1.k_cache 0.06589249 10.87988195 + layer.1.v_cache 0.00000499 0.00343784 + layer.2.k_cache 0.00422055 1.18296449 + layer.2.v_cache 0.00001579 0.00886631 + layer.3.k_cache 0.05337993 4.70591027 + layer.3.v_cache 0.00001747 0.01141943 + layer.4.k_cache 0.00075412 0.24553428 + layer.4.v_cache 0.00004712 0.02360587 + layer.4.output 0.19135183 763.31476358 + ------------------------------------------------------------------------------------- + TOTAL 0.09203923 319.54567548 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 78376 +BPFP 1.0146 bits/point +EBPFP 2.0292 equivalent bits/point +MSE 319.545675 +---------------------- -------------------------------------------------------- +Time: 0.571s Load: 0.003s, Pack+Encode: 0.247s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 319.5457 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,064B, BPFP=0.5699 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,488B, BPFP=1.9509 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,988B, BPFP=0.9278 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,128B, BPFP=1.8839 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,396B, BPFP=1.0037 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,500B, BPFP=1.7671 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,112B, BPFP=0.9509 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,860B, BPFP=1.8341 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,968B, BPFP=1.4821 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,468B, BPFP=1.7612 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,712B, BPFP=0.2581 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.298s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09057366 70.58025832 + layer.0.v_cache 0.00001789 0.00890713 + layer.1.k_cache 0.03614600 11.13581122 + layer.1.v_cache 0.00000481 0.00298448 + layer.2.k_cache 0.00554357 1.09212957 + layer.2.v_cache 0.00001714 0.00874848 + layer.3.k_cache 0.02742965 4.62051028 + layer.3.v_cache 0.00001729 0.01007072 + layer.4.k_cache 0.00093027 0.23160839 + layer.4.v_cache 0.00004906 0.02085438 + layer.4.output 0.16441821 644.50478316 + ------------------------------------------------------------------------------------- + TOTAL 0.07715628 270.54384501 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 85684 +BPFP 0.9375 bits/point +EBPFP 1.8751 equivalent bits/point +MSE 270.543845 +---------------------- -------------------------------------------------------- +Time: 0.556s Load: 0.004s, Pack+Encode: 0.254s, Decode+Unpack: 0.298s +---------------------- -------------------------------------------------------- +💾 Converting with 270.5438 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,692B, BPFP=0.5463 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,216B, BPFP=2.0731 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,856B, BPFP=0.9854 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,732B, BPFP=1.9748 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,232B, BPFP=1.0617 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,020B, BPFP=1.8304 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,988B, BPFP=1.0122 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,484B, BPFP=1.9245 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,844B, BPFP=1.5917 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,232B, BPFP=1.8734 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,412B, BPFP=0.3018 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09325875 74.48024046 + layer.0.v_cache 0.00001406 0.00920556 + layer.1.k_cache 0.05775181 10.77939675 + layer.1.v_cache 0.00000524 0.00327617 + layer.2.k_cache 0.00413042 1.18055299 + layer.2.v_cache 0.00001743 0.00947634 + layer.3.k_cache 0.04387648 4.93469436 + layer.3.v_cache 0.00002497 0.01235002 + layer.4.k_cache 0.00071512 0.24793330 + layer.4.v_cache 0.00005147 0.02386794 + layer.4.output 0.17652170 703.56893553 + ------------------------------------------------------------------------------------- + TOTAL 0.08444104 295.09785545 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 83708 +BPFP 0.9992 bits/point +EBPFP 1.9984 equivalent bits/point +MSE 295.097855 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 295.0979 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,228B, BPFP=0.5732 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,164B, BPFP=1.8047 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,308B, BPFP=0.9425 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,632B, BPFP=1.7102 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,440B, BPFP=0.9659 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,268B, BPFP=1.6456 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,156B, BPFP=0.9155 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,880B, BPFP=1.7543 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,864B, BPFP=1.3963 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,140B, BPFP=1.6229 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,644B, BPFP=0.2446 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212358 64.52062988 + layer.0.v_cache 0.00001682 0.00759199 + layer.1.k_cache 0.05918076 10.88302196 + layer.1.v_cache 0.00000473 0.00262002 + layer.2.k_cache 0.00260097 1.14136713 + layer.2.v_cache 0.00001544 0.00783158 + layer.3.k_cache 0.02838851 4.62122276 + layer.3.v_cache 0.00001657 0.00942961 + layer.4.k_cache 0.00106615 0.21955824 + layer.4.v_cache 0.00004456 0.01926560 + layer.4.output 0.17173813 614.69891437 + ------------------------------------------------------------------------------------- + TOTAL 0.08091912 257.90146702 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 84724 +BPFP 0.8849 bits/point +EBPFP 1.7698 equivalent bits/point +MSE 257.901467 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 257.9015 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,244B, BPFP=0.5760 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,724B, BPFP=1.9041 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,112B, BPFP=0.9077 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,288B, BPFP=1.8267 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,540B, BPFP=0.9837 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,888B, BPFP=1.7557 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,168B, BPFP=0.9176 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,140B, BPFP=1.8004 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,140B, BPFP=1.4453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,776B, BPFP=1.7358 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,036B, BPFP=0.2799 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10731998 68.04103782 + layer.0.v_cache 0.00001618 0.00825963 + layer.1.k_cache 0.07375650 10.86467535 + layer.1.v_cache 0.00000540 0.00310647 + layer.2.k_cache 0.00246326 1.06141767 + layer.2.v_cache 0.00001746 0.00867277 + layer.3.k_cache 0.01642703 5.07279136 + layer.3.v_cache 0.00001778 0.00990838 + layer.4.k_cache 0.00066607 0.22552802 + layer.4.v_cache 0.00004728 0.02045762 + layer.4.output 0.15452130 615.20312500 + ------------------------------------------------------------------------------------- + TOTAL 0.07543447 258.33751354 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 89056 +BPFP 0.9301 bits/point +EBPFP 1.8603 equivalent bits/point +MSE 258.337514 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 258.3375 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 99, 128) +Output shape: (1, 99, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.output: torch.Size([1, 99, 3584]) -> torch.Size([1, 1, 99, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,912B, BPFP=0.4596 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,640B, BPFP=1.8371 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,920B, BPFP=0.9343 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,184B, BPFP=1.7652 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,024B, BPFP=0.9508 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,800B, BPFP=1.7045 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,632B, BPFP=0.8889 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,152B, BPFP=1.7601 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,992B, BPFP=1.4192 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,648B, BPFP=1.6806 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,372B, BPFP=0.2790 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09061131 72.36127387 + layer.0.v_cache 0.00001461 0.00859552 + layer.1.k_cache 0.05062923 10.75501475 + layer.1.v_cache 0.00000709 0.00304385 + layer.2.k_cache 0.00392854 1.04455073 + layer.2.v_cache 0.00001793 0.00848392 + layer.3.k_cache 0.05082532 4.77739492 + layer.3.v_cache 0.00001820 0.01003161 + layer.4.k_cache 0.00080398 0.23210383 + layer.4.v_cache 0.00005053 0.01949251 + layer.4.output 0.02403493 562.12364719 + ------------------------------------------------------------------------------------- + TOTAL 0.02147948 236.71091270 + (elements=861,696) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 861696 +Total Bytes 97276 +BPFP 0.9031 bits/point +EBPFP 1.8062 equivalent bits/point +MSE 236.710913 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 236.7109 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,308B, BPFP=0.5079 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,112B, BPFP=2.2254 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,480B, BPFP=0.9859 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,692B, BPFP=2.1329 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,884B, BPFP=1.0748 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,100B, BPFP=2.0026 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,488B, BPFP=0.9877 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,304B, BPFP=2.0475 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,404B, BPFP=1.6294 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,196B, BPFP=2.0238 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,828B, BPFP=0.3404 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11136492 70.77325869 + layer.0.v_cache 0.00001955 0.01013942 + layer.1.k_cache 0.06280328 10.54503535 + layer.1.v_cache 0.00000544 0.00362764 + layer.2.k_cache 0.00235174 1.33295387 + layer.2.v_cache 0.00001859 0.00982783 + layer.3.k_cache 0.03326307 5.10616206 + layer.3.v_cache 0.00001926 0.01225204 + layer.4.k_cache 0.00064473 0.23708260 + layer.4.v_cache 0.00005254 0.02244192 + layer.4.output 0.19494371 762.33966298 + ------------------------------------------------------------------------------------- + TOTAL 0.09265583 319.08414249 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 81796 +BPFP 1.0589 bits/point +EBPFP 2.1178 equivalent bits/point +MSE 319.084142 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 319.0841 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,540B, BPFP=0.5363 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,592B, BPFP=2.2365 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,892B, BPFP=1.0329 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,976B, BPFP=2.1064 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,060B, BPFP=1.0684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,436B, BPFP=1.9924 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,736B, BPFP=1.0000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,724B, BPFP=2.0532 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,960B, BPFP=1.6807 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,396B, BPFP=1.9840 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,340B, BPFP=0.3119 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08829598 74.82312342 + layer.0.v_cache 0.00001455 0.01033079 + layer.1.k_cache 0.05950350 10.85598837 + layer.1.v_cache 0.00000529 0.00367934 + layer.2.k_cache 0.00411868 1.13825473 + layer.2.v_cache 0.00001680 0.00987416 + layer.3.k_cache 0.04877868 5.16938576 + layer.3.v_cache 0.00001896 0.01314370 + layer.4.k_cache 0.00063741 0.26079412 + layer.4.v_cache 0.00004764 0.02415995 + layer.4.output 0.18365649 732.21657819 + ------------------------------------------------------------------------------------- + TOTAL 0.08747253 306.93086951 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 84652 +BPFP 1.0514 bits/point +EBPFP 2.1028 equivalent bits/point +MSE 306.930870 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 306.9309 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,588B, BPFP=0.5252 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,200B, BPFP=2.0698 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,968B, BPFP=1.0081 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,964B, BPFP=2.0219 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,080B, BPFP=1.0308 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,492B, BPFP=1.9261 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,860B, BPFP=0.9862 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,716B, BPFP=1.9716 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,908B, BPFP=1.6047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,620B, BPFP=1.9521 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,204B, BPFP=0.2958 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11195674 73.10389610 + layer.0.v_cache 0.00001346 0.00905139 + layer.1.k_cache 0.03755230 10.87628887 + layer.1.v_cache 0.00000535 0.00355459 + layer.2.k_cache 0.00239600 1.21081652 + layer.2.v_cache 0.00001753 0.01046707 + layer.3.k_cache 0.02867049 5.20010970 + layer.3.v_cache 0.00001751 0.01135229 + layer.4.k_cache 0.00071811 0.24713115 + layer.4.v_cache 0.00005486 0.02431421 + layer.4.output 0.19495771 701.40381494 + ------------------------------------------------------------------------------------- + TOTAL 0.09094743 294.14845214 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 84600 +BPFP 1.0098 bits/point +EBPFP 2.0197 equivalent bits/point +MSE 294.148452 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 294.1485 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,240B, BPFP=0.5625 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,808B, BPFP=1.8764 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,288B, BPFP=0.9181 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,368B, BPFP=1.8000 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,300B, BPFP=0.9201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,864B, BPFP=1.7125 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,244B, BPFP=0.9104 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,348B, BPFP=1.7965 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,104B, BPFP=1.4069 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,796B, BPFP=1.7007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,252B, BPFP=0.2543 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08657128 73.60021701 + layer.0.v_cache 0.00001574 0.00906829 + layer.1.k_cache 0.05618976 10.48063829 + layer.1.v_cache 0.00000506 0.00309603 + layer.2.k_cache 0.00392605 1.05050761 + layer.2.v_cache 0.00001570 0.00853367 + layer.3.k_cache 0.02849737 4.61326057 + layer.3.v_cache 0.00001769 0.01128882 + layer.4.k_cache 0.00099738 0.23006927 + layer.4.v_cache 0.00004751 0.02201768 + layer.4.output 0.15107292 601.92018849 + ------------------------------------------------------------------------------------- + TOTAL 0.07257612 253.14529510 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 88612 +BPFP 0.9049 bits/point +EBPFP 1.8099 equivalent bits/point +MSE 253.145295 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 253.1453 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,004B, BPFP=0.5867 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,160B, BPFP=1.9844 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,044B, BPFP=0.9852 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,692B, BPFP=1.8930 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,220B, BPFP=1.0195 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,360B, BPFP=1.8281 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,932B, BPFP=0.9633 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,632B, BPFP=1.8813 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,924B, BPFP=1.5477 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,456B, BPFP=1.8469 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,924B, BPFP=0.2769 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09254595 70.18759155 + layer.0.v_cache 0.00001776 0.00842075 + layer.1.k_cache 0.05775445 11.28441772 + layer.1.v_cache 0.00000504 0.00298355 + layer.2.k_cache 0.00891839 1.13161602 + layer.2.v_cache 0.00001732 0.00907855 + layer.3.k_cache 0.02836140 4.59205093 + layer.3.v_cache 0.00001672 0.01058495 + layer.4.k_cache 0.00083566 0.21832695 + layer.4.v_cache 0.00004876 0.02189821 + layer.4.output 0.16993865 676.73465402 + ------------------------------------------------------------------------------------- + TOTAL 0.08106423 283.80056161 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 84348 +BPFP 0.9691 bits/point +EBPFP 1.9381 equivalent bits/point +MSE 283.800562 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 283.8006 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,056B, BPFP=0.4923 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,012B, BPFP=1.7738 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,700B, BPFP=0.9182 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,560B, BPFP=1.7010 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,376B, BPFP=0.8660 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,988B, BPFP=1.6089 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,224B, BPFP=0.8415 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,304B, BPFP=1.6598 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,412B, BPFP=1.3550 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,088B, BPFP=1.6250 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,676B, BPFP=0.2457 +⌛️ [2/4] FRONTEND: Frontend time: 0.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13093194 68.79441849 + layer.0.v_cache 0.00001635 0.00849288 + layer.1.k_cache 0.05235285 9.90891730 + layer.1.v_cache 0.00000496 0.00300356 + layer.2.k_cache 0.00512381 1.01891382 + layer.2.v_cache 0.00001647 0.00832423 + layer.3.k_cache 0.01138228 4.85435155 + layer.3.v_cache 0.00001850 0.00988658 + layer.4.k_cache 0.00081880 0.21167433 + layer.4.v_cache 0.00004796 0.01896232 + layer.4.output 0.02447439 573.64184462 + ------------------------------------------------------------------------------------- + TOTAL 0.02188439 241.19587397 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 90396 +BPFP 0.8565 bits/point +EBPFP 1.7131 equivalent bits/point +MSE 241.195874 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.005s, Pack+Encode: 0.155s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 241.1959 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,700B, BPFP=0.5340 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,448B, BPFP=2.0665 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,812B, BPFP=0.9517 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,860B, BPFP=1.9502 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,084B, BPFP=1.0055 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,420B, BPFP=1.8631 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,860B, BPFP=0.9612 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,632B, BPFP=1.9051 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,688B, BPFP=1.5206 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,248B, BPFP=1.8291 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,752B, BPFP=0.3038 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510638 76.61931987 + layer.0.v_cache 0.00001399 0.00884686 + layer.1.k_cache 0.05467150 11.42825279 + layer.1.v_cache 0.00000511 0.00323983 + layer.2.k_cache 0.00545186 1.14942430 + layer.2.v_cache 0.00001652 0.00901781 + layer.3.k_cache 0.02986256 5.40044934 + layer.3.v_cache 0.00001747 0.00974467 + layer.4.k_cache 0.00068266 0.22891129 + layer.4.v_cache 0.00005277 0.02159670 + layer.4.output 0.17642128 685.19072107 + ------------------------------------------------------------------------------------- + TOTAL 0.08357822 287.71846182 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 84504 +BPFP 0.9832 bits/point +EBPFP 1.9663 equivalent bits/point +MSE 287.718462 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 287.7185 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,680B, BPFP=0.5438 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,212B, BPFP=2.0722 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,872B, BPFP=0.9886 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,864B, BPFP=2.0016 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,080B, BPFP=1.0308 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,220B, BPFP=1.8709 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,872B, BPFP=0.9886 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,684B, BPFP=1.9651 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,852B, BPFP=1.5933 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,424B, BPFP=1.9123 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,200B, BPFP=0.2957 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08730992 76.72818588 + layer.0.v_cache 0.00001543 0.00951625 + layer.1.k_cache 0.05635426 11.32741002 + layer.1.v_cache 0.00000520 0.00356522 + layer.2.k_cache 0.00390536 1.15822522 + layer.2.v_cache 0.00001661 0.00965413 + layer.3.k_cache 0.06483487 5.25206132 + layer.3.v_cache 0.00002010 0.01163200 + layer.4.k_cache 0.00072652 0.24760754 + layer.4.v_cache 0.00004842 0.02222126 + layer.4.output 0.18845482 702.30026670 + ------------------------------------------------------------------------------------- + TOTAL 0.09014238 294.75717328 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 83960 +BPFP 1.0022 bits/point +EBPFP 2.0044 equivalent bits/point +MSE 294.757173 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 294.7572 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,108B, BPFP=0.4844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,584B, BPFP=2.2022 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,160B, BPFP=0.9559 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,076B, BPFP=2.0855 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,480B, BPFP=1.0294 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,668B, BPFP=1.9917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,276B, BPFP=0.9825 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,972B, BPFP=2.0616 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,016B, BPFP=1.6121 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,456B, BPFP=1.9430 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,568B, BPFP=0.3469 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10784986 75.02103918 + layer.0.v_cache 0.00001387 0.00869396 + layer.1.k_cache 0.01409798 11.05492985 + layer.1.v_cache 0.00000489 0.00298822 + layer.2.k_cache 0.00254294 1.25445377 + layer.2.v_cache 0.00001653 0.00875146 + layer.3.k_cache 0.05709259 5.59098322 + layer.3.v_cache 0.00001736 0.01040252 + layer.4.k_cache 0.00062715 0.21922961 + layer.4.v_cache 0.00004758 0.02107159 + layer.4.output 0.22376562 797.54017857 + ------------------------------------------------------------------------------------- + TOTAL 0.10286295 333.88081138 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 77364 +BPFP 1.0457 bits/point +EBPFP 2.0914 equivalent bits/point +MSE 333.880811 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 333.8808 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,900B, BPFP=0.5736 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,732B, BPFP=1.9248 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,824B, BPFP=0.9541 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,184B, BPFP=1.8165 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,048B, BPFP=0.9984 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,856B, BPFP=1.7516 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,868B, BPFP=0.9628 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,140B, BPFP=1.8078 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,492B, BPFP=1.4818 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,948B, BPFP=1.7698 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,804B, BPFP=0.3053 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14563359 77.09142603 + layer.0.v_cache 0.00001368 0.00806511 + layer.1.k_cache 0.05947259 10.81319379 + layer.1.v_cache 0.00000496 0.00294631 + layer.2.k_cache 0.00553216 1.14527236 + layer.2.v_cache 0.00001557 0.00820890 + layer.3.k_cache 0.04377548 5.01759667 + layer.3.v_cache 0.00001730 0.00995156 + layer.4.k_cache 0.00079035 0.22307678 + layer.4.v_cache 0.00005153 0.02133706 + layer.4.output 0.17347779 684.52797242 + ------------------------------------------------------------------------------------- + TOTAL 0.08645010 287.41393421 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 81796 +BPFP 0.9516 bits/point +EBPFP 1.9033 equivalent bits/point +MSE 287.413934 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 287.4139 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,984B, BPFP=0.4697 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,084B, BPFP=2.1506 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,224B, BPFP=1.0000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,724B, BPFP=2.0653 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,420B, BPFP=1.0464 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,228B, BPFP=1.9479 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,236B, BPFP=1.0028 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,492B, BPFP=2.0104 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,672B, BPFP=1.5795 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,096B, BPFP=1.9167 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,764B, BPFP=0.3302 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12373332 75.48647609 + layer.0.v_cache 0.00001436 0.00972065 + layer.1.k_cache 0.01270404 11.43258944 + layer.1.v_cache 0.00000490 0.00311655 + layer.2.k_cache 0.00252672 1.14454558 + layer.2.v_cache 0.00001629 0.00960622 + layer.3.k_cache 0.05185594 4.89696341 + layer.3.v_cache 0.00001849 0.01112050 + layer.4.k_cache 0.00068458 0.22783239 + layer.4.v_cache 0.00004975 0.02131016 + layer.4.output 0.22283003 821.56310877 + ------------------------------------------------------------------------------------- + TOTAL 0.10302462 343.77559073 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 73924 +BPFP 1.0295 bits/point +EBPFP 2.0589 equivalent bits/point +MSE 343.775591 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.003s, Pack+Encode: 0.157s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 343.7756 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,972B, BPFP=0.5528 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,020B, BPFP=2.0499 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,060B, BPFP=0.9412 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,532B, BPFP=1.9591 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,448B, BPFP=1.0134 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,020B, BPFP=1.8638 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,192B, BPFP=0.9658 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,400B, BPFP=1.9345 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,284B, BPFP=1.5409 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,940B, BPFP=1.8490 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,772B, BPFP=0.2862 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12002775 70.62285505 + layer.0.v_cache 0.00001466 0.00943949 + layer.1.k_cache 0.05271155 11.21678380 + layer.1.v_cache 0.00000524 0.00343450 + layer.2.k_cache 0.00808975 1.08915456 + layer.2.v_cache 0.00001826 0.00949189 + layer.3.k_cache 0.07501186 5.21274894 + layer.3.v_cache 0.00001895 0.01157507 + layer.4.k_cache 0.00071216 0.24624695 + layer.4.v_cache 0.00005170 0.02254906 + layer.4.output 0.17053346 644.91390306 + ------------------------------------------------------------------------------------- + TOTAL 0.08531742 270.75538828 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 89640 +BPFP 0.9808 bits/point +EBPFP 1.9617 equivalent bits/point +MSE 270.755388 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 270.7554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 75, 128) +Output shape: (1, 75, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.output: torch.Size([1, 75, 3584]) -> torch.Size([1, 1, 75, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,512B, BPFP=0.5233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,252B, BPFP=2.1358 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,980B, BPFP=1.0375 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,748B, BPFP=2.0308 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,128B, BPFP=1.0683 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,272B, BPFP=1.9317 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,752B, BPFP=0.9900 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,716B, BPFP=2.0242 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,924B, BPFP=1.6508 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,420B, BPFP=1.9625 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,020B, BPFP=0.3280 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08149677 70.30529948 + layer.0.v_cache 0.00001391 0.00958411 + layer.1.k_cache 0.05841089 10.99744222 + layer.1.v_cache 0.00000541 0.00345284 + layer.2.k_cache 0.00252637 1.22654460 + layer.2.v_cache 0.00001693 0.00980946 + layer.3.k_cache 0.02876982 5.15959757 + layer.3.v_cache 0.00001932 0.01142130 + layer.4.k_cache 0.00075246 0.24612752 + layer.4.v_cache 0.00005041 0.02345987 + layer.4.output 0.19266937 720.63005952 + ------------------------------------------------------------------------------------- + TOTAL 0.08945576 301.90606798 + (elements=652,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 652800 +Total Bytes 84724 +BPFP 1.0383 bits/point +EBPFP 2.0766 equivalent bits/point +MSE 301.906068 +---------------------- -------------------------------------------------------- +Time: 0.359s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 301.9061 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,112B, BPFP=0.4853 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,604B, BPFP=2.2068 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,080B, BPFP=0.9375 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,964B, BPFP=2.0597 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,500B, BPFP=1.0340 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,712B, BPFP=2.0018 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,172B, BPFP=0.9586 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,772B, BPFP=2.0156 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,976B, BPFP=1.6029 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,648B, BPFP=1.9871 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,644B, BPFP=0.3166 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10404026 73.96546128 + layer.0.v_cache 0.00001395 0.00925358 + layer.1.k_cache 0.01458383 11.36135146 + layer.1.v_cache 0.00000499 0.00313671 + layer.2.k_cache 0.00236740 1.17968772 + layer.2.v_cache 0.00001641 0.00957043 + layer.3.k_cache 0.03645969 5.27787377 + layer.3.v_cache 0.00002420 0.01148683 + layer.4.k_cache 0.00069996 0.24316300 + layer.4.v_cache 0.00005119 0.02382591 + layer.4.output 0.20980707 797.67995011 + ------------------------------------------------------------------------------------- + TOTAL 0.09570067 333.87320361 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 76184 +BPFP 1.0297 bits/point +EBPFP 2.0595 equivalent bits/point +MSE 333.873204 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 333.8732 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,276B, BPFP=0.5080 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,744B, BPFP=2.1750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,508B, BPFP=1.0063 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,212B, BPFP=2.0562 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,652B, BPFP=1.0384 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,816B, BPFP=1.9679 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,332B, BPFP=0.9670 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,820B, BPFP=1.9688 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,124B, BPFP=1.5902 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,560B, BPFP=1.9107 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,944B, BPFP=0.3490 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11725240 70.08913225 + layer.0.v_cache 0.00001567 0.00980084 + layer.1.k_cache 0.03693308 11.23692714 + layer.1.v_cache 0.00000533 0.00382131 + layer.2.k_cache 0.00240566 1.29180527 + layer.2.v_cache 0.00001850 0.01116577 + layer.3.k_cache 0.06751672 5.01287144 + layer.3.v_cache 0.00001909 0.01192578 + layer.4.k_cache 0.00067984 0.24240853 + layer.4.v_cache 0.00004901 0.02502005 + layer.4.output 0.20094273 774.81530612 + ------------------------------------------------------------------------------------- + TOTAL 0.09597026 324.21423654 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 78988 +BPFP 1.0371 bits/point +EBPFP 2.0743 equivalent bits/point +MSE 324.214237 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 324.2142 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,248B, BPFP=0.5018 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,564B, BPFP=2.1348 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,308B, BPFP=0.9616 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,120B, BPFP=2.0357 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,704B, BPFP=1.0500 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,644B, BPFP=1.9295 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,388B, BPFP=0.9795 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,852B, BPFP=1.9759 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,068B, BPFP=1.5777 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,476B, BPFP=1.8920 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,948B, BPFP=0.3172 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07685234 70.03192662 + layer.0.v_cache 0.00001390 0.00955184 + layer.1.k_cache 0.01187498 10.78977225 + layer.1.v_cache 0.00000934 0.00360074 + layer.2.k_cache 0.00239603 1.26548331 + layer.2.v_cache 0.00001621 0.00942824 + layer.3.k_cache 0.07241479 5.13617118 + layer.3.v_cache 0.00001716 0.01166532 + layer.4.k_cache 0.00072067 0.23614213 + layer.4.v_cache 0.00004958 0.02224415 + layer.4.output 0.21417375 774.30765306 + ------------------------------------------------------------------------------------- + TOTAL 0.09785772 323.98056219 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 77320 +BPFP 1.0152 bits/point +EBPFP 2.0305 equivalent bits/point +MSE 323.980562 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 323.9806 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,548B, BPFP=0.5238 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,460B, BPFP=2.1505 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,968B, BPFP=1.0214 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,936B, BPFP=2.0428 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,228B, BPFP=1.0748 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,536B, BPFP=1.9605 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,040B, BPFP=1.0362 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,624B, BPFP=1.9786 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,992B, BPFP=1.6431 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,488B, BPFP=1.9507 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,656B, BPFP=0.3130 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09264422 75.19526110 + layer.0.v_cache 0.00001680 0.00980055 + layer.1.k_cache 0.05606840 11.43032194 + layer.1.v_cache 0.00000542 0.00374088 + layer.2.k_cache 0.00241068 1.06266684 + layer.2.v_cache 0.00001702 0.00989115 + layer.3.k_cache 0.15270803 5.17503879 + layer.3.v_cache 0.00001849 0.01291782 + layer.4.k_cache 0.00070970 0.25913379 + layer.4.v_cache 0.00005539 0.02485204 + layer.4.output 0.18838861 710.96551927 + ------------------------------------------------------------------------------------- + TOTAL 0.09549261 298.23189763 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 85476 +BPFP 1.0337 bits/point +EBPFP 2.0674 equivalent bits/point +MSE 298.231898 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 298.2319 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,508B, BPFP=0.5089 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,164B, BPFP=2.0625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,900B, BPFP=0.9943 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,604B, BPFP=1.9489 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,024B, BPFP=1.0195 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,104B, BPFP=1.8474 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,752B, BPFP=0.9643 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,404B, BPFP=1.9083 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,812B, BPFP=1.5852 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,324B, BPFP=1.8920 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,760B, BPFP=0.2829 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10144231 72.36818943 + layer.0.v_cache 0.00001330 0.00899429 + layer.1.k_cache 0.09893422 11.35782772 + layer.1.v_cache 0.00000513 0.00312893 + layer.2.k_cache 0.00368564 1.15120994 + layer.2.v_cache 0.00001777 0.00945417 + layer.3.k_cache 0.06341467 4.79227279 + layer.3.v_cache 0.00001739 0.01074970 + layer.4.k_cache 0.00073693 0.23754209 + layer.4.v_cache 0.00004994 0.02214139 + layer.4.output 0.18261438 703.21405380 + ------------------------------------------------------------------------------------- + TOTAL 0.09097753 294.85058159 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 82356 +BPFP 0.9831 bits/point +EBPFP 1.9661 equivalent bits/point +MSE 294.850582 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 294.8506 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,160B, BPFP=0.5675 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,348B, BPFP=1.8585 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,320B, BPFP=0.9555 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,764B, BPFP=1.7536 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,416B, BPFP=0.9727 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,324B, BPFP=1.6746 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,008B, BPFP=0.8994 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,660B, BPFP=1.7349 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,856B, BPFP=1.4109 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,300B, BPFP=1.6703 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,328B, BPFP=0.2650 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08947531 67.00392309 + layer.0.v_cache 0.00001432 0.00803902 + layer.1.k_cache 0.05320675 10.83796850 + layer.1.v_cache 0.00000497 0.00309832 + layer.2.k_cache 0.00562141 1.01668496 + layer.2.v_cache 0.00001574 0.00796928 + layer.3.k_cache 0.02742601 5.37659709 + layer.3.v_cache 0.00001759 0.00976120 + layer.4.k_cache 0.00114861 0.22474000 + layer.4.v_cache 0.00004734 0.02001154 + layer.4.output 0.15628038 622.26426519 + ------------------------------------------------------------------------------------- + TOTAL 0.07476122 261.19756761 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 85484 +BPFP 0.9031 bits/point +EBPFP 1.8062 equivalent bits/point +MSE 261.197568 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 261.1976 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,100B, BPFP=0.5265 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,708B, BPFP=1.8186 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,336B, BPFP=0.9062 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,088B, BPFP=1.7133 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,408B, BPFP=0.9185 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,728B, BPFP=1.6522 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,392B, BPFP=0.9158 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,160B, BPFP=1.7255 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,000B, BPFP=1.3587 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,596B, BPFP=1.6298 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,168B, BPFP=0.2467 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12311893 69.13318933 + layer.0.v_cache 0.00001357 0.00820629 + layer.1.k_cache 0.03743907 11.10480134 + layer.1.v_cache 0.00000515 0.00310752 + layer.2.k_cache 0.00511221 1.12729545 + layer.2.v_cache 0.00001545 0.00800790 + layer.3.k_cache 0.04225797 4.98773193 + layer.3.v_cache 0.00001799 0.01000471 + layer.4.k_cache 0.00119426 0.22689388 + layer.4.v_cache 0.00004621 0.01910220 + layer.4.output 0.14783020 588.37490295 + ------------------------------------------------------------------------------------- + TOTAL 0.07317837 247.36780360 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 87684 +BPFP 0.8760 bits/point +EBPFP 1.7520 equivalent bits/point +MSE 247.367804 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 247.3678 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,132B, BPFP=0.4798 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,988B, BPFP=1.8364 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,124B, BPFP=0.9381 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,288B, BPFP=1.7292 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,168B, BPFP=0.9449 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,996B, BPFP=1.6844 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,928B, BPFP=0.9081 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,404B, BPFP=1.7469 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,296B, BPFP=1.4240 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,916B, BPFP=1.6722 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,080B, BPFP=0.2862 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09293266 68.60363051 + layer.0.v_cache 0.00001503 0.00883827 + layer.1.k_cache 0.07854857 10.03590243 + layer.1.v_cache 0.00000559 0.00315478 + layer.2.k_cache 0.00650115 1.09464174 + layer.2.v_cache 0.00001763 0.00859865 + layer.3.k_cache 0.03715762 3.80541454 + layer.3.v_cache 0.00001788 0.01037224 + layer.4.k_cache 0.00073365 0.21819500 + layer.4.v_cache 0.00004630 0.01924277 + layer.4.output 11.17680568 526.32593662 + ------------------------------------------------------------------------------------- + TOTAL 4.61491858 221.65232631 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 100320 +BPFP 0.9040 bits/point +EBPFP 1.8080 equivalent bits/point +MSE 221.652326 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 221.6523 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 96, 128) +Output shape: (1, 96, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.output: torch.Size([1, 96, 3584]) -> torch.Size([1, 1, 96, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,884B, BPFP=0.4694 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,608B, BPFP=1.8893 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,596B, BPFP=0.9108 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,136B, BPFP=1.8125 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,732B, BPFP=0.9329 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,744B, BPFP=1.7487 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,564B, BPFP=0.9056 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,040B, BPFP=1.7969 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,024B, BPFP=1.4688 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,556B, BPFP=1.7181 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,356B, BPFP=0.2640 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10158994 66.24422201 + layer.0.v_cache 0.00001467 0.00932538 + layer.1.k_cache 0.06416908 10.83137258 + layer.1.v_cache 0.00000542 0.00328406 + layer.2.k_cache 0.00400108 1.07609852 + layer.2.v_cache 0.00001872 0.00925962 + layer.3.k_cache 0.01491791 4.57206059 + layer.3.v_cache 0.00001877 0.01132593 + layer.4.k_cache 0.00070988 0.23040823 + layer.4.v_cache 0.00004832 0.02199713 + layer.4.output 0.14169503 564.20484561 + ------------------------------------------------------------------------------------- + TOTAL 0.06925641 237.20254549 + (elements=835,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 835584 +Total Bytes 95240 +BPFP 0.9118 bits/point +EBPFP 1.8237 equivalent bits/point +MSE 237.202545 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 237.2025 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,404B, BPFP=0.5217 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,876B, BPFP=2.1432 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,552B, BPFP=0.9878 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,576B, BPFP=2.0781 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,900B, BPFP=1.0634 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,128B, BPFP=1.9809 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,632B, BPFP=1.0052 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,120B, BPFP=1.9792 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,624B, BPFP=1.6545 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,068B, BPFP=1.9679 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,636B, BPFP=0.2987 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10321087 69.26369900 + layer.0.v_cache 0.00001557 0.00939919 + layer.1.k_cache 0.06175087 10.72215949 + layer.1.v_cache 0.00000571 0.00333404 + layer.2.k_cache 0.01329466 1.15897751 + layer.2.v_cache 0.00001784 0.01011498 + layer.3.k_cache 0.05039416 5.08433448 + layer.3.v_cache 0.00001745 0.01128853 + layer.4.k_cache 0.00079786 0.24847783 + layer.4.v_cache 0.00005071 0.02315256 + layer.4.output 0.18870974 752.64676339 + ------------------------------------------------------------------------------------- + TOTAL 0.09120729 315.00366361 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 80516 +BPFP 1.0278 bits/point +EBPFP 2.0557 equivalent bits/point +MSE 315.003664 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 315.0037 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,460B, BPFP=0.5339 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,628B, BPFP=2.0894 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,448B, BPFP=0.9653 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,028B, BPFP=1.9592 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,764B, BPFP=1.0339 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,332B, BPFP=1.8082 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,564B, BPFP=0.9905 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,928B, BPFP=1.9375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,224B, BPFP=1.5677 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,560B, BPFP=1.8576 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,560B, BPFP=0.3274 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09666174 67.81169298 + layer.0.v_cache 0.00001715 0.00871898 + layer.1.k_cache 0.08476152 10.17166222 + layer.1.v_cache 0.00000506 0.00311333 + layer.2.k_cache 0.00238475 1.11910417 + layer.2.v_cache 0.00001567 0.00851806 + layer.3.k_cache 0.07332428 4.46074719 + layer.3.v_cache 0.00001796 0.01089343 + layer.4.k_cache 0.00072290 0.22869669 + layer.4.v_cache 0.00005092 0.02019626 + layer.4.output 0.18875189 752.50620040 + ------------------------------------------------------------------------------------- + TOTAL 0.09289560 314.78745565 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 78496 +BPFP 1.0020 bits/point +EBPFP 2.0041 equivalent bits/point +MSE 314.787456 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 314.7875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,136B, BPFP=0.4908 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,728B, BPFP=2.2353 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,128B, BPFP=0.9485 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,116B, BPFP=2.0947 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,704B, BPFP=1.0809 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,652B, BPFP=1.9881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,188B, BPFP=0.9623 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,972B, BPFP=2.0616 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,980B, BPFP=1.6039 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,772B, BPFP=2.0156 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,596B, BPFP=0.3478 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10601801 78.17115694 + layer.0.v_cache 0.00001817 0.00970754 + layer.1.k_cache 0.01696859 11.12601426 + layer.1.v_cache 0.00000501 0.00354627 + layer.2.k_cache 0.00274148 1.23874373 + layer.2.v_cache 0.00001757 0.00965747 + layer.3.k_cache 0.01706722 5.42096486 + layer.3.v_cache 0.00001843 0.01132383 + layer.4.k_cache 0.00080906 0.23920003 + layer.4.v_cache 0.00005076 0.02216883 + layer.4.output 0.19983917 797.20312500 + ------------------------------------------------------------------------------------- + TOTAL 0.09074050 333.92202110 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 77972 +BPFP 1.0539 bits/point +EBPFP 2.1078 equivalent bits/point +MSE 333.922021 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.004s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 333.9220 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 65, 128) +Output shape: (1, 65, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.output: torch.Size([1, 65, 3584]) -> torch.Size([1, 1, 65, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,968B, BPFP=0.4731 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,076B, BPFP=2.1817 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,012B, BPFP=0.9644 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,508B, BPFP=2.0452 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,488B, BPFP=1.0788 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,284B, BPFP=1.9913 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,400B, BPFP=1.0577 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,548B, BPFP=2.0548 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,880B, BPFP=1.6538 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,508B, BPFP=2.0452 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,996B, BPFP=0.3089 +⌛️ [2/4] FRONTEND: Frontend time: 0.156s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09579474 76.00866887 + layer.0.v_cache 0.00001312 0.00864313 + layer.1.k_cache 0.01648773 10.89791541 + layer.1.v_cache 0.00000712 0.00310448 + layer.2.k_cache 0.00431219 1.14404860 + layer.2.v_cache 0.00001803 0.00956947 + layer.3.k_cache 0.03460605 5.11729783 + layer.3.v_cache 0.00001721 0.01184131 + layer.4.k_cache 0.00068162 0.23565964 + layer.4.v_cache 0.00004727 0.02291593 + layer.4.output 0.20898957 833.87225275 + ------------------------------------------------------------------------------------- + TOTAL 0.09499483 348.85679023 + (elements=565,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 565760 +Total Bytes 73668 +BPFP 1.0417 bits/point +EBPFP 2.0834 equivalent bits/point +MSE 348.856790 +---------------------- -------------------------------------------------------- +Time: 0.360s Load: 0.003s, Pack+Encode: 0.156s, Decode+Unpack: 0.200s +---------------------- -------------------------------------------------------- +💾 Converting with 348.8568 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 41, 128) +Output shape: (1, 41, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.output: torch.Size([1, 41, 3584]) -> torch.Size([1, 1, 41, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,656B, BPFP=0.6311 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 5,316B, BPFP=2.0259 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,608B, BPFP=0.9939 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 4,972B, BPFP=1.8948 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 2,784B, BPFP=1.0610 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 4,868B, BPFP=1.8552 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,568B, BPFP=0.9787 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 4,916B, BPFP=1.8735 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 3,964B, BPFP=1.5107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 4,796B, BPFP=1.8277 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,608B, BPFP=0.4142 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10098721 86.10093131 + layer.0.v_cache 0.00001362 0.01177551 + layer.1.k_cache 0.01708067 11.38698299 + layer.1.v_cache 0.00000533 0.00432378 + layer.2.k_cache 0.00250472 1.48051220 + layer.2.v_cache 0.00001755 0.01323167 + layer.3.k_cache 0.08199174 5.67287240 + layer.3.v_cache 0.00001815 0.01403938 + layer.4.k_cache 0.00059437 0.30684787 + layer.4.v_cache 0.00005232 0.03104505 + layer.4.output 0.35770382 1319.92835366 + ------------------------------------------------------------------------------------- + TOTAL 0.15924661 549.67770810 + (elements=356,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 356864 +Total Bytes 46056 +BPFP 1.0325 bits/point +EBPFP 2.0649 equivalent bits/point +MSE 549.677708 +---------------------- -------------------------------------------------------- +Time: 0.278s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 549.6777 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,824B, BPFP=0.5481 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,724B, BPFP=2.0204 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,028B, BPFP=0.9099 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,308B, BPFP=1.8954 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,460B, BPFP=1.0397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,180B, BPFP=1.8570 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,344B, BPFP=1.0048 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,252B, BPFP=1.8786 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,068B, BPFP=1.5228 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,012B, BPFP=1.8065 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,304B, BPFP=0.3565 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09210638 72.47543100 + layer.0.v_cache 0.00001369 0.00961314 + layer.1.k_cache 0.01682195 10.66076073 + layer.1.v_cache 0.00000520 0.00370922 + layer.2.k_cache 0.00487203 1.29143700 + layer.2.v_cache 0.00001755 0.01055198 + layer.3.k_cache 0.04965206 5.92687167 + layer.3.v_cache 0.00001944 0.01199917 + layer.4.k_cache 0.00060442 0.24481359 + layer.4.v_cache 0.00005207 0.02416953 + layer.4.output 0.26182200 1040.03880495 + ------------------------------------------------------------------------------------- + TOTAL 0.11746581 433.58417598 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 56504 +BPFP 0.9987 bits/point +EBPFP 1.9975 equivalent bits/point +MSE 433.584176 +---------------------- -------------------------------------------------------- +Time: 0.276s Load: 0.003s, Pack+Encode: 0.129s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 433.5842 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 47, 128) +Output shape: (1, 47, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.output: torch.Size([1, 47, 3584]) -> torch.Size([1, 1, 47, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,744B, BPFP=0.5798 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,428B, BPFP=2.1370 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,876B, BPFP=0.9561 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,068B, BPFP=2.0173 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,228B, BPFP=1.0731 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,824B, BPFP=1.9362 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 2,992B, BPFP=0.9947 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 5,860B, BPFP=1.9481 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,832B, BPFP=1.6064 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,780B, BPFP=1.9215 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,084B, BPFP=0.3839 +⌛️ [2/4] FRONTEND: Frontend time: 0.129s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08461995 83.03814827 + layer.0.v_cache 0.00001366 0.01061488 + layer.1.k_cache 0.01838745 11.24823777 + layer.1.v_cache 0.00000609 0.00462650 + layer.2.k_cache 0.00780689 1.19165380 + layer.2.v_cache 0.00001932 0.01205730 + layer.3.k_cache 0.04292001 6.06571993 + layer.3.v_cache 0.00001942 0.01315245 + layer.4.k_cache 0.00059704 0.26382312 + layer.4.v_cache 0.00004850 0.02659265 + layer.4.output 0.29441516 1149.53039514 + ------------------------------------------------------------------------------------- + TOTAL 0.13031438 479.32867015 + (elements=409,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 409088 +Total Bytes 53716 +BPFP 1.0505 bits/point +EBPFP 2.1009 equivalent bits/point +MSE 479.328670 +---------------------- -------------------------------------------------------- +Time: 0.275s Load: 0.002s, Pack+Encode: 0.129s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 479.3287 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,864B, BPFP=0.5110 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,100B, BPFP=1.9463 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,348B, BPFP=0.9178 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,812B, BPFP=1.8673 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,788B, BPFP=1.0384 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,592B, BPFP=1.8070 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,572B, BPFP=0.9792 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,572B, BPFP=1.8015 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,472B, BPFP=1.5000 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,552B, BPFP=1.7961 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,516B, BPFP=0.3335 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10653571 76.67329701 + layer.0.v_cache 0.00001396 0.00944499 + layer.1.k_cache 0.01513677 11.22855310 + layer.1.v_cache 0.00000574 0.00361491 + layer.2.k_cache 0.00241650 1.28753796 + layer.2.v_cache 0.00001997 0.01068014 + layer.3.k_cache 0.04250543 5.65415928 + layer.3.v_cache 0.00001889 0.01201528 + layer.4.k_cache 0.00060099 0.23853031 + layer.4.v_cache 0.00005317 0.02446914 + layer.4.output 0.25650722 948.14317043 + ------------------------------------------------------------------------------------- + TOTAL 0.11546222 396.00849971 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 60188 +BPFP 0.9705 bits/point +EBPFP 1.9410 equivalent bits/point +MSE 396.008500 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 396.0085 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,796B, BPFP=0.5846 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,540B, BPFP=2.1289 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 2,864B, BPFP=0.9323 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,200B, BPFP=2.0182 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,316B, BPFP=1.0794 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,936B, BPFP=1.9323 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,044B, BPFP=0.9909 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,076B, BPFP=1.9779 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 4,992B, BPFP=1.6250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,056B, BPFP=1.9714 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,260B, BPFP=0.3841 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10088900 80.97300212 + layer.0.v_cache 0.00001321 0.00893914 + layer.1.k_cache 0.01601078 10.87465795 + layer.1.v_cache 0.00000529 0.00334299 + layer.2.k_cache 0.00240732 1.28232797 + layer.2.v_cache 0.00001858 0.01073412 + layer.3.k_cache 0.04236705 5.24470520 + layer.3.v_cache 0.00001968 0.01149989 + layer.4.k_cache 0.00061965 0.24369613 + layer.4.v_cache 0.00005235 0.02477581 + layer.4.output 0.30498679 1123.61532738 + ------------------------------------------------------------------------------------- + TOTAL 0.13513591 468.46970429 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 55080 +BPFP 1.0547 bits/point +EBPFP 2.1094 equivalent bits/point +MSE 468.469704 +---------------------- -------------------------------------------------------- +Time: 0.278s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 468.4697 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 53, 128) +Output shape: (1, 53, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.output: torch.Size([1, 53, 3584]) -> torch.Size([1, 1, 53, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,812B, BPFP=0.5342 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,896B, BPFP=2.0330 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,112B, BPFP=0.9175 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,672B, BPFP=1.9670 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,576B, BPFP=1.0542 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,460B, BPFP=1.9045 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,316B, BPFP=0.9776 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,408B, BPFP=1.8892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,424B, BPFP=1.5991 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,332B, BPFP=1.8667 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 8,368B, BPFP=0.3524 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10150213 73.41812997 + layer.0.v_cache 0.00001370 0.01083140 + layer.1.k_cache 0.01753690 10.64482145 + layer.1.v_cache 0.00000536 0.00415852 + layer.2.k_cache 0.00234856 1.20908025 + layer.2.v_cache 0.00001946 0.01256518 + layer.3.k_cache 0.03772523 5.45672550 + layer.3.v_cache 0.00001792 0.01285724 + layer.4.k_cache 0.00061365 0.24954105 + layer.4.v_cache 0.00006478 0.02599876 + layer.4.output 0.25806696 1019.33195755 + ------------------------------------------------------------------------------------- + TOTAL 0.11566567 425.08049483 + (elements=461,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 461312 +Total Bytes 58376 +BPFP 1.0123 bits/point +EBPFP 2.0247 equivalent bits/point +MSE 425.080495 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 425.0805 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,880B, BPFP=0.4896 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,116B, BPFP=1.8531 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,672B, BPFP=0.9563 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,964B, BPFP=1.8135 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,052B, BPFP=1.0552 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,756B, BPFP=1.7594 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,976B, BPFP=1.0354 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,848B, BPFP=1.7833 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,616B, BPFP=1.4625 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,572B, BPFP=1.7115 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,560B, BPFP=0.3557 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11850098 77.63818359 + layer.0.v_cache 0.00001385 0.01003038 + layer.1.k_cache 0.01468890 10.08667806 + layer.1.v_cache 0.00000523 0.00362564 + layer.2.k_cache 0.01477869 1.07692388 + layer.2.v_cache 0.00001915 0.01087028 + layer.3.k_cache 0.14661819 5.43916219 + layer.3.v_cache 0.00001925 0.01248256 + layer.4.k_cache 0.00060183 0.22692092 + layer.4.v_cache 0.00005752 0.02542503 + layer.4.output 0.22966646 900.68459821 + ------------------------------------------------------------------------------------- + TOTAL 0.11193934 376.43073471 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 63012 +BPFP 0.9653 bits/point +EBPFP 1.9305 equivalent bits/point +MSE 376.430735 +---------------------- -------------------------------------------------------- +Time: 0.277s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.144s +---------------------- -------------------------------------------------------- +💾 Converting with 376.4307 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,160B, BPFP=0.5367 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,404B, BPFP=1.9368 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,436B, BPFP=0.9232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,740B, BPFP=1.8240 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,696B, BPFP=0.9674 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,372B, BPFP=1.7615 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,532B, BPFP=0.9395 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,800B, BPFP=1.8342 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,640B, BPFP=1.4674 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,292B, BPFP=1.7480 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,644B, BPFP=0.2825 +⌛️ [2/4] FRONTEND: Frontend time: 0.157s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10836371 72.31056810 + layer.0.v_cache 0.00001445 0.00926716 + layer.1.k_cache 0.06846340 11.17684937 + layer.1.v_cache 0.00000523 0.00329921 + layer.2.k_cache 0.00506382 1.14771279 + layer.2.v_cache 0.00001740 0.00913049 + layer.3.k_cache 0.03973729 4.72065005 + layer.3.v_cache 0.00001777 0.01109847 + layer.4.k_cache 0.00066797 0.23211135 + layer.4.v_cache 0.00005099 0.02267503 + layer.4.output 0.14786712 588.51193711 + ------------------------------------------------------------------------------------- + TOTAL 0.07396893 247.60158364 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 93716 +BPFP 0.9363 bits/point +EBPFP 1.8725 equivalent bits/point +MSE 247.601584 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.005s, Pack+Encode: 0.157s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 247.6016 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,960B, BPFP=0.4640 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,516B, BPFP=2.2528 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,076B, BPFP=0.9650 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,976B, BPFP=2.1250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,416B, BPFP=1.0455 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,476B, BPFP=2.0066 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,300B, BPFP=1.0180 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,712B, BPFP=2.0625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,876B, BPFP=1.6278 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,612B, BPFP=2.0388 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,496B, BPFP=0.3212 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08728886 72.15099728 + layer.0.v_cache 0.00001734 0.01001940 + layer.1.k_cache 0.01499850 11.31911584 + layer.1.v_cache 0.00000528 0.00343193 + layer.2.k_cache 0.00240521 1.19018578 + layer.2.v_cache 0.00001739 0.00939075 + layer.3.k_cache 0.03126823 5.35673708 + layer.3.v_cache 0.00001843 0.01136724 + layer.4.k_cache 0.00061395 0.24614531 + layer.4.v_cache 0.00004830 0.02319025 + layer.4.output 0.21953388 821.03091180 + ------------------------------------------------------------------------------------- + TOTAL 0.09843639 343.38452726 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 75416 +BPFP 1.0502 bits/point +EBPFP 2.1005 equivalent bits/point +MSE 343.384527 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 343.3845 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,980B, BPFP=0.4688 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,392B, BPFP=2.2235 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,096B, BPFP=0.9697 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,896B, BPFP=2.1061 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,508B, BPFP=1.0672 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,464B, BPFP=2.0038 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,320B, BPFP=1.0227 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,816B, BPFP=2.0871 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,892B, BPFP=1.6316 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,516B, BPFP=2.0161 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,456B, BPFP=0.3198 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08223097 74.38491359 + layer.0.v_cache 0.00001442 0.01007734 + layer.1.k_cache 0.01480860 10.69219601 + layer.1.v_cache 0.00000502 0.00324854 + layer.2.k_cache 0.00576455 1.22472335 + layer.2.v_cache 0.00002304 0.00959664 + layer.3.k_cache 0.03215047 4.74472416 + layer.3.v_cache 0.00001933 0.01243371 + layer.4.k_cache 0.00061510 0.24140841 + layer.4.v_cache 0.00004916 0.02291399 + layer.4.output 0.22077460 821.08367154 + ------------------------------------------------------------------------------------- + TOTAL 0.09888840 343.46658450 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 75336 +BPFP 1.0491 bits/point +EBPFP 2.0983 equivalent bits/point +MSE 343.466584 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 343.4666 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,408B, BPFP=0.5154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,208B, BPFP=2.1849 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,516B, BPFP=0.9666 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,464B, BPFP=2.0257 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,844B, BPFP=1.0368 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,880B, BPFP=1.9007 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,496B, BPFP=0.9623 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,328B, BPFP=1.9966 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,524B, BPFP=1.6104 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,980B, BPFP=1.9221 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,188B, BPFP=0.3115 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10114760 72.98782641 + layer.0.v_cache 0.00001420 0.00947705 + layer.1.k_cache 0.03683303 10.63002244 + layer.1.v_cache 0.00000544 0.00365633 + layer.2.k_cache 0.00248627 1.14351226 + layer.2.v_cache 0.00001713 0.01009629 + layer.3.k_cache 0.02909129 4.58850348 + layer.3.v_cache 0.00001766 0.01191013 + layer.4.k_cache 0.00072613 0.26212104 + layer.4.v_cache 0.00005154 0.02378223 + layer.4.output 0.19731793 742.30925881 + ------------------------------------------------------------------------------------- + TOTAL 0.09127152 310.93151290 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 80836 +BPFP 1.0178 bits/point +EBPFP 2.0356 equivalent bits/point +MSE 310.931513 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.003s, Pack+Encode: 0.161s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 310.9315 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,632B, BPFP=0.5411 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,432B, BPFP=2.1447 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,736B, BPFP=0.9737 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,652B, BPFP=1.9844 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,040B, BPFP=1.0362 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,228B, BPFP=1.8972 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,828B, BPFP=0.9926 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,440B, BPFP=1.9408 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,768B, BPFP=1.5970 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,348B, BPFP=1.9219 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,748B, BPFP=0.3157 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10230959 72.67205489 + layer.0.v_cache 0.00001409 0.00939905 + layer.1.k_cache 0.07973289 11.49562314 + layer.1.v_cache 0.00000506 0.00325585 + layer.2.k_cache 0.00553133 1.16725088 + layer.2.v_cache 0.00001928 0.00999027 + layer.3.k_cache 0.03946657 5.35264106 + layer.3.v_cache 0.00001980 0.01158639 + layer.4.k_cache 0.00066525 0.24364896 + layer.4.v_cache 0.00005119 0.02222739 + layer.4.output 0.17985767 711.79158835 + ------------------------------------------------------------------------------------- + TOTAL 0.08745993 298.44287037 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 83852 +BPFP 1.0141 bits/point +EBPFP 2.0282 equivalent bits/point +MSE 298.442870 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 298.4429 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,652B, BPFP=0.5381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,096B, BPFP=2.0487 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,940B, BPFP=1.0024 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,332B, BPFP=1.8937 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,956B, BPFP=1.0057 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,928B, BPFP=1.8117 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,852B, BPFP=0.9846 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,308B, BPFP=1.8888 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,568B, BPFP=1.5357 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,120B, BPFP=1.8506 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,320B, BPFP=0.2992 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633976 72.88214032 + layer.0.v_cache 0.00001461 0.00933519 + layer.1.k_cache 0.07748085 10.84255308 + layer.1.v_cache 0.00000515 0.00322229 + layer.2.k_cache 0.00505460 1.19114180 + layer.2.v_cache 0.00001651 0.00897995 + layer.3.k_cache 0.02793909 4.99653368 + layer.3.v_cache 0.00001809 0.01064856 + layer.4.k_cache 0.00070819 0.23326329 + layer.4.v_cache 0.00004716 0.02130047 + layer.4.output 0.18765952 702.65080009 + ------------------------------------------------------------------------------------- + TOTAL 0.09007298 294.63263055 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 82072 +BPFP 0.9797 bits/point +EBPFP 1.9593 equivalent bits/point +MSE 294.632631 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 294.6326 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,256B, BPFP=0.5036 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,800B, BPFP=2.1875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,604B, BPFP=1.0277 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,348B, BPFP=2.0866 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,704B, BPFP=1.0500 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,984B, BPFP=2.0054 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,472B, BPFP=0.9982 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,116B, BPFP=2.0348 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,148B, BPFP=1.5955 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,808B, BPFP=1.9661 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,664B, BPFP=0.3082 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08399124 70.08126395 + layer.0.v_cache 0.00001671 0.00990048 + layer.1.k_cache 0.03849935 11.08456508 + layer.1.v_cache 0.00000503 0.00334773 + layer.2.k_cache 0.00738276 1.21943556 + layer.2.v_cache 0.00001692 0.00948882 + layer.3.k_cache 0.04914829 5.28219953 + layer.3.v_cache 0.00001713 0.01128334 + layer.4.k_cache 0.00063731 0.23111638 + layer.4.v_cache 0.00004711 0.02231426 + layer.4.output 0.21328995 774.35401786 + ------------------------------------------------------------------------------------- + TOTAL 0.09839950 324.02547295 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 78904 +BPFP 1.0360 bits/point +EBPFP 2.0721 equivalent bits/point +MSE 324.025473 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 324.0255 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 56, 128) +Output shape: (1, 56, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.output: torch.Size([1, 56, 3584]) -> torch.Size([1, 1, 56, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,868B, BPFP=0.5212 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,824B, BPFP=1.9040 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,216B, BPFP=0.8973 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,536B, BPFP=1.8237 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,608B, BPFP=1.0067 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,328B, BPFP=1.7656 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,624B, BPFP=1.0112 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,428B, BPFP=1.7935 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,356B, BPFP=1.4944 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,236B, BPFP=1.7400 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 7,960B, BPFP=0.3173 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13768739 73.38069371 + layer.0.v_cache 0.00001432 0.00970757 + layer.1.k_cache 0.01638686 11.04345158 + layer.1.v_cache 0.00000533 0.00355016 + layer.2.k_cache 0.00251956 1.23281002 + layer.2.v_cache 0.00001745 0.00986684 + layer.3.k_cache 0.08599430 5.52677100 + layer.3.v_cache 0.00002012 0.01295999 + layer.4.k_cache 0.00061995 0.24355893 + layer.4.v_cache 0.00004798 0.02237621 + layer.4.output 0.24256272 966.36567283 + ------------------------------------------------------------------------------------- + TOTAL 0.11419131 403.29679152 + (elements=487,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 487424 +Total Bytes 57984 +BPFP 0.9517 bits/point +EBPFP 1.9034 equivalent bits/point +MSE 403.296792 +---------------------- -------------------------------------------------------- +Time: 0.279s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 403.2968 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,832B, BPFP=0.5531 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,536B, BPFP=2.0578 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,940B, BPFP=0.9648 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,100B, BPFP=1.9727 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,132B, BPFP=1.0023 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,596B, BPFP=1.8742 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,876B, BPFP=0.9523 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,956B, BPFP=1.9445 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,128B, BPFP=1.5875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,536B, BPFP=1.8625 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,512B, BPFP=0.2933 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09278157 71.90894775 + layer.0.v_cache 0.00001504 0.00909077 + layer.1.k_cache 0.07523044 11.13382950 + layer.1.v_cache 0.00000517 0.00332332 + layer.2.k_cache 0.00381728 1.06629658 + layer.2.v_cache 0.00001789 0.00901961 + layer.3.k_cache 0.03327554 4.97824554 + layer.3.v_cache 0.00001939 0.01084512 + layer.4.k_cache 0.00067710 0.23810959 + layer.4.v_cache 0.00005192 0.02186239 + layer.4.output 0.16994183 677.13337054 + ------------------------------------------------------------------------------------- + TOTAL 0.08208730 284.07724494 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 86144 +BPFP 0.9897 bits/point +EBPFP 1.9794 equivalent bits/point +MSE 284.077245 +---------------------- -------------------------------------------------------- +Time: 0.363s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 284.0772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,844B, BPFP=0.4883 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,096B, BPFP=1.8792 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,536B, BPFP=0.9364 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,844B, BPFP=1.8125 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,860B, BPFP=1.0222 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,600B, BPFP=1.7479 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,756B, BPFP=0.9947 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,728B, BPFP=1.7818 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,484B, BPFP=1.4523 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,452B, BPFP=1.7087 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,192B, BPFP=0.3478 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12468526 83.93562963 + layer.0.v_cache 0.00001451 0.01089235 + layer.1.k_cache 0.01313243 10.63034213 + layer.1.v_cache 0.00000580 0.00386046 + layer.2.k_cache 0.00906488 1.20514472 + layer.2.v_cache 0.00002099 0.01089919 + layer.3.k_cache 0.01569995 5.41275904 + layer.3.v_cache 0.00002002 0.01313402 + layer.4.k_cache 0.00064964 0.24221965 + layer.4.v_cache 0.00005478 0.02452881 + layer.4.output 0.23031524 916.54933414 + ------------------------------------------------------------------------------------- + TOTAL 0.10444441 383.37263229 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 61392 +BPFP 0.9564 bits/point +EBPFP 1.9128 equivalent bits/point +MSE 383.372632 +---------------------- -------------------------------------------------------- +Time: 0.279s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 383.3726 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,884B, BPFP=0.4906 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,172B, BPFP=1.8677 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,760B, BPFP=0.9792 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,924B, BPFP=1.8031 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,076B, BPFP=1.0615 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,784B, BPFP=1.7667 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,956B, BPFP=1.0302 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,804B, BPFP=1.7719 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,672B, BPFP=1.4771 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,652B, BPFP=1.7323 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,732B, BPFP=0.3621 +⌛️ [2/4] FRONTEND: Frontend time: 0.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15164433 76.39960124 + layer.0.v_cache 0.00001435 0.01068697 + layer.1.k_cache 0.01704682 10.84427694 + layer.1.v_cache 0.00000573 0.00401945 + layer.2.k_cache 0.00240009 1.10314840 + layer.2.v_cache 0.00001845 0.01067745 + layer.3.k_cache 0.02161660 5.86708781 + layer.3.v_cache 0.00001985 0.01275815 + layer.4.k_cache 0.00061459 0.24088629 + layer.4.v_cache 0.00005275 0.02665225 + layer.4.output 0.22653077 902.15825893 + ------------------------------------------------------------------------------------- + TOTAL 0.10465582 377.03691809 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 63416 +BPFP 0.9714 bits/point +EBPFP 1.9429 equivalent bits/point +MSE 377.036918 +---------------------- -------------------------------------------------------- +Time: 0.278s Load: 0.003s, Pack+Encode: 0.130s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 377.0369 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,392B, BPFP=0.5120 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,392B, BPFP=2.0103 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,512B, BPFP=0.9658 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,872B, BPFP=1.8990 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,660B, BPFP=0.9974 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,408B, BPFP=1.7997 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,408B, BPFP=0.9435 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,860B, BPFP=1.8964 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,388B, BPFP=1.5813 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,540B, BPFP=1.8279 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,776B, BPFP=0.2989 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.199s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08869729 70.38718696 + layer.0.v_cache 0.00001594 0.00872330 + layer.1.k_cache 0.08105310 10.58936550 + layer.1.v_cache 0.00000494 0.00303556 + layer.2.k_cache 0.00256802 1.12047890 + layer.2.v_cache 0.00001661 0.00869424 + layer.3.k_cache 0.01835501 4.69778401 + layer.3.v_cache 0.00001861 0.01068071 + layer.4.k_cache 0.00078324 0.24421935 + layer.4.v_cache 0.00004769 0.02130688 + layer.4.output 0.18611241 742.41224315 + ------------------------------------------------------------------------------------- + TOTAL 0.08790278 310.82218691 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 77208 +BPFP 0.9721 bits/point +EBPFP 1.9442 equivalent bits/point +MSE 310.822187 +---------------------- -------------------------------------------------------- +Time: 0.361s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.199s +---------------------- -------------------------------------------------------- +💾 Converting with 310.8222 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,076B, BPFP=0.4955 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,436B, BPFP=1.8421 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,708B, BPFP=0.9195 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,920B, BPFP=1.7590 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,764B, BPFP=0.9285 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,584B, BPFP=1.7049 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,600B, BPFP=0.9021 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,884B, BPFP=1.7532 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,652B, BPFP=1.3937 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,404B, BPFP=1.6759 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,472B, BPFP=0.2640 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17228427 70.57124879 + layer.0.v_cache 0.00001543 0.00866131 + layer.1.k_cache 0.04621154 10.15174158 + layer.1.v_cache 0.00000545 0.00299000 + layer.2.k_cache 0.00739691 0.95565119 + layer.2.v_cache 0.00001797 0.00843236 + layer.3.k_cache 0.03096496 4.22060528 + layer.3.v_cache 0.00001890 0.01035086 + layer.4.k_cache 0.00071091 0.22602579 + layer.4.v_cache 0.00004900 0.01903590 + layer.4.output 0.02450962 573.96230670 + ------------------------------------------------------------------------------------- + TOTAL 0.02524957 241.40652294 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 94500 +BPFP 0.8954 bits/point +EBPFP 1.7909 equivalent bits/point +MSE 241.406523 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 241.4065 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,044B, BPFP=0.5114 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,496B, BPFP=1.9315 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,548B, BPFP=0.9321 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,088B, BPFP=1.8629 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,840B, BPFP=0.9812 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,564B, BPFP=1.7749 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,596B, BPFP=0.9402 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,876B, BPFP=1.8273 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,928B, BPFP=1.5000 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,404B, BPFP=1.7480 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,300B, BPFP=0.2712 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.201s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16174046 68.55126428 + layer.0.v_cache 0.00001612 0.00945950 + layer.1.k_cache 0.03352992 10.90686232 + layer.1.v_cache 0.00000563 0.00355466 + layer.2.k_cache 0.00496730 1.12008216 + layer.2.v_cache 0.00001740 0.00961929 + layer.3.k_cache 0.01521110 4.94112503 + layer.3.v_cache 0.00001939 0.01165009 + layer.4.k_cache 0.00066431 0.23503400 + layer.4.v_cache 0.00005067 0.02235598 + layer.4.output 0.14624557 582.87207181 + ------------------------------------------------------------------------------------- + TOTAL 0.07293772 245.05385353 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 94684 +BPFP 0.9358 bits/point +EBPFP 1.8715 equivalent bits/point +MSE 245.053854 +---------------------- -------------------------------------------------------- +Time: 0.362s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.201s +---------------------- -------------------------------------------------------- +💾 Converting with 245.0539 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,960B, BPFP=0.5378 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,840B, BPFP=1.9695 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,212B, BPFP=0.9469 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,344B, BPFP=1.8794 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,136B, BPFP=0.9331 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,724B, BPFP=1.7667 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,024B, BPFP=0.9128 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,236B, BPFP=1.8597 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,040B, BPFP=1.4608 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,652B, BPFP=1.7536 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,904B, BPFP=0.2830 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07121617 69.63963708 + layer.0.v_cache 0.00001681 0.00907178 + layer.1.k_cache 0.06877367 10.62159055 + layer.1.v_cache 0.00000530 0.00315120 + layer.2.k_cache 0.00486504 1.15485675 + layer.2.v_cache 0.00001749 0.00863569 + layer.3.k_cache 0.08833314 4.98994872 + layer.3.v_cache 0.00001723 0.01021690 + layer.4.k_cache 0.00077278 0.22414986 + layer.4.v_cache 0.00004595 0.02130180 + layer.4.output 0.15812202 629.80595930 + ------------------------------------------------------------------------------------- + TOTAL 0.07887751 264.43083973 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 88072 +BPFP 0.9413 bits/point +EBPFP 1.8825 equivalent bits/point +MSE 264.430840 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 264.4308 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,000B, BPFP=0.4641 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,148B, BPFP=1.8793 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,724B, BPFP=0.8855 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,616B, BPFP=1.7970 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,948B, BPFP=0.9202 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,980B, BPFP=1.6986 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,792B, BPFP=0.8960 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,456B, BPFP=1.7723 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,356B, BPFP=1.4474 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,008B, BPFP=1.7030 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,284B, BPFP=0.2494 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13964315 70.01398128 + layer.0.v_cache 0.00001432 0.00968549 + layer.1.k_cache 0.04810369 10.55326617 + layer.1.v_cache 0.00000572 0.00378346 + layer.2.k_cache 0.00492308 1.12494659 + layer.2.v_cache 0.00001777 0.00967017 + layer.3.k_cache 0.04255116 4.28660840 + layer.3.v_cache 0.00001880 0.01180985 + layer.4.k_cache 0.00073971 0.24477607 + layer.4.v_cache 0.00004573 0.02334169 + layer.4.output 11.28739116 531.55175919 + ------------------------------------------------------------------------------------- + TOTAL 4.66163537 223.94965785 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 98312 +BPFP 0.8947 bits/point +EBPFP 1.7893 equivalent bits/point +MSE 223.949658 +---------------------- -------------------------------------------------------- +Time: 0.375s Load: 0.005s, Pack+Encode: 0.158s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 223.9497 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,144B, BPFP=0.5582 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,900B, BPFP=1.9354 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,032B, BPFP=0.8935 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,252B, BPFP=1.8203 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,328B, BPFP=0.9460 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,740B, BPFP=1.7294 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,148B, BPFP=0.9141 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,012B, BPFP=1.7777 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,952B, BPFP=1.4119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,664B, BPFP=1.7159 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,924B, BPFP=0.2771 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14145413 66.33931108 + layer.0.v_cache 0.00001516 0.00932236 + layer.1.k_cache 0.10647225 11.23923215 + layer.1.v_cache 0.00000560 0.00326592 + layer.2.k_cache 0.00244946 0.97589085 + layer.2.v_cache 0.00001685 0.00831861 + layer.3.k_cache 0.01581429 4.70087710 + layer.3.v_cache 0.00001841 0.01046149 + layer.4.k_cache 0.00072023 0.21877313 + layer.4.v_cache 0.00004652 0.01942098 + layer.4.output 0.15451367 614.59694602 + ------------------------------------------------------------------------------------- + TOTAL 0.07932992 257.98255858 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 88096 +BPFP 0.9201 bits/point +EBPFP 1.8402 equivalent bits/point +MSE 257.982559 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 257.9826 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,496B, BPFP=0.5270 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,296B, BPFP=2.1740 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,600B, BPFP=0.9713 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,640B, BPFP=2.0355 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,800B, BPFP=1.0135 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,180B, BPFP=1.9383 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,600B, BPFP=0.9713 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,424B, BPFP=1.9899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,640B, BPFP=1.6132 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,396B, BPFP=1.9840 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,488B, BPFP=0.2862 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10237078 72.45551362 + layer.0.v_cache 0.00001391 0.00906915 + layer.1.k_cache 0.05953315 10.83404376 + layer.1.v_cache 0.00000517 0.00331232 + layer.2.k_cache 0.00728044 1.18227531 + layer.2.v_cache 0.00001648 0.00904461 + layer.3.k_cache 0.03101032 5.68008835 + layer.3.v_cache 0.00001666 0.01081878 + layer.4.k_cache 0.00072555 0.23376470 + layer.4.v_cache 0.00004870 0.02173126 + layer.4.output 0.18364459 731.84555985 + ------------------------------------------------------------------------------------- + TOTAL 0.08744314 306.66815181 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 81560 +BPFP 1.0130 bits/point +EBPFP 2.0260 equivalent bits/point +MSE 306.668152 +---------------------- -------------------------------------------------------- +Time: 0.380s Load: 0.003s, Pack+Encode: 0.165s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 306.6682 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 111, 128) +Output shape: (1, 111, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.output: torch.Size([1, 111, 3584]) -> torch.Size([1, 1, 111, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,936B, BPFP=0.5541 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,956B, BPFP=1.8238 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,340B, BPFP=0.8925 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,420B, BPFP=1.7483 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,496B, BPFP=0.9144 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,876B, BPFP=1.6717 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,336B, BPFP=0.8919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,076B, BPFP=1.6999 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,004B, BPFP=1.4082 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,736B, BPFP=1.6520 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,336B, BPFP=0.2481 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11107565 65.40233495 + layer.0.v_cache 0.00001597 0.00858752 + layer.1.k_cache 0.04333284 10.58321566 + layer.1.v_cache 0.00000558 0.00306939 + layer.2.k_cache 0.00477623 1.03470096 + layer.2.v_cache 0.00001758 0.00815767 + layer.3.k_cache 0.01020195 4.65004971 + layer.3.v_cache 0.00001870 0.00962549 + layer.4.k_cache 0.00079488 0.21448787 + layer.4.v_cache 0.00005041 0.01988555 + layer.4.output 10.27057757 483.19538288 + ------------------------------------------------------------------------------------- + TOTAL 4.23907840 203.78245853 + (elements=966,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 966144 +Total Bytes 106512 +BPFP 0.8820 bits/point +EBPFP 1.7639 equivalent bits/point +MSE 203.782459 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 203.7825 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,288B, BPFP=0.5035 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,056B, BPFP=2.2130 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,324B, BPFP=0.9516 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,196B, BPFP=2.0238 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,864B, BPFP=1.0704 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,632B, BPFP=1.8996 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,456B, BPFP=0.9806 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,052B, BPFP=1.9921 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,300B, BPFP=1.6065 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,900B, BPFP=1.9586 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,564B, BPFP=0.3321 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09375787 74.11310932 + layer.0.v_cache 0.00001622 0.00893229 + layer.1.k_cache 0.05984993 10.48173415 + layer.1.v_cache 0.00000512 0.00326842 + layer.2.k_cache 0.00243208 1.13067777 + layer.2.v_cache 0.00001865 0.00980337 + layer.3.k_cache 0.04944374 4.83874426 + layer.3.v_cache 0.00001908 0.01138205 + layer.4.k_cache 0.00084877 0.23335806 + layer.4.v_cache 0.00005139 0.02339292 + layer.4.output 0.19371969 762.86343058 + ------------------------------------------------------------------------------------- + TOTAL 0.09191063 319.46461275 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 79632 +BPFP 1.0309 bits/point +EBPFP 2.0617 equivalent bits/point +MSE 319.464613 +---------------------- -------------------------------------------------------- +Time: 0.380s Load: 0.003s, Pack+Encode: 0.166s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 319.4646 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 109, 128) +Output shape: (1, 109, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.output: torch.Size([1, 109, 3584]) -> torch.Size([1, 1, 109, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,912B, BPFP=0.5608 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,544B, BPFP=1.7982 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,288B, BPFP=0.9014 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,064B, BPFP=1.7294 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,440B, BPFP=0.9232 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,604B, BPFP=1.6634 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,388B, BPFP=0.9157 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,988B, BPFP=1.7185 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,844B, BPFP=1.4111 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,492B, BPFP=1.6474 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,396B, BPFP=0.2743 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860953 68.95482726 + layer.0.v_cache 0.00001500 0.00843292 + layer.1.k_cache 0.05917679 10.40551086 + layer.1.v_cache 0.00000528 0.00329301 + layer.2.k_cache 0.01253135 1.03157813 + layer.2.v_cache 0.00001869 0.00904430 + layer.3.k_cache 0.07406004 4.61461030 + layer.3.v_cache 0.00001854 0.01031319 + layer.4.k_cache 0.00071234 0.21447852 + layer.4.v_cache 0.00004916 0.02072488 + layer.4.output 10.45909253 492.08777031 + ------------------------------------------------------------------------------------- + TOTAL 4.32287320 207.64042386 + (elements=948,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 948736 +Total Bytes 105960 +BPFP 0.8935 bits/point +EBPFP 1.7870 equivalent bits/point +MSE 207.640424 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.006s, Pack+Encode: 0.166s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 207.6404 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst + to output-fixed/kimiaudio/lambda0.007/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 0.9785 bits/point +Avg EBPFP 1.9570 equivalent bits/point +Avg MSE 319.360268 +Avg Time 0.360s +------------------------ ---------------------------- diff --git a/lambda0.01/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.01/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..53f001af33e7714d08f81bf92df29671180fa811 --- /dev/null +++ b/lambda0.01/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 506 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench +Output output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench +---------------- ------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,152B, BPFP=0.6080 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,032B, BPFP=2.8997 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,428B, BPFP=1.2400 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,100B, BPFP=2.9128 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,656B, BPFP=1.2840 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,652B, BPFP=2.6335 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,604B, BPFP=1.2739 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,392B, BPFP=2.5833 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,792B, BPFP=2.0818 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,304B, BPFP=2.5664 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,540B, BPFP=0.4007 +⌛️ [2/4] FRONTEND: Frontend time: 2.187s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.792s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12089630 42.59858760 + layer.0.v_cache 0.00001394 0.00656962 + layer.1.k_cache 0.01174029 8.69279273 + layer.1.v_cache 0.00000579 0.00256329 + layer.2.k_cache 0.00835781 1.09290474 + layer.2.v_cache 0.00001852 0.00747342 + layer.3.k_cache 0.02896961 3.64755927 + layer.3.v_cache 0.00001872 0.00795984 + layer.4.k_cache 0.00063445 0.19709399 + layer.4.v_cache 0.00005063 0.01712965 + layer.4.output 0.17246709 608.11067019 + ------------------------------------------------------------------------------------- + TOTAL 0.08105739 253.70854856 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 118652 +BPFP 1.3464 bits/point +EBPFP 2.6927 equivalent bits/point +MSE 253.708549 +---------------------- -------------------------------------------------------- +Time: 3.986s Load: 0.006s, Pack+Encode: 2.187s, Decode+Unpack: 1.792s +---------------------- -------------------------------------------------------- +💾 Converting with 253.7085 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,144B, BPFP=0.6141 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,036B, BPFP=2.9367 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,484B, BPFP=1.2664 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,388B, BPFP=3.0055 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,944B, BPFP=1.3562 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,552B, BPFP=2.6469 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,596B, BPFP=1.2883 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,880B, BPFP=2.7109 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,456B, BPFP=2.2375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,412B, BPFP=2.8148 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,788B, BPFP=0.3847 +⌛️ [2/4] FRONTEND: Frontend time: 1.986s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.633s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08029544 43.65505981 + layer.0.v_cache 0.00001376 0.00645055 + layer.1.k_cache 0.03632395 8.31900253 + layer.1.v_cache 0.00000563 0.00285937 + layer.2.k_cache 0.00522719 0.83821135 + layer.2.v_cache 0.00001997 0.00881662 + layer.3.k_cache 0.02707962 3.73110123 + layer.3.v_cache 0.00001843 0.00925785 + layer.4.k_cache 0.00063057 0.21487832 + layer.4.v_cache 0.00005085 0.01868598 + layer.4.output 0.18451144 656.58482143 + ------------------------------------------------------------------------------------- + TOTAL 0.08477915 273.69988668 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 120680 +BPFP 1.3865 bits/point +EBPFP 2.7730 equivalent bits/point +MSE 273.699887 +---------------------- -------------------------------------------------------- +Time: 3.623s Load: 0.004s, Pack+Encode: 1.986s, Decode+Unpack: 1.633s +---------------------- -------------------------------------------------------- +💾 Converting with 273.6999 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,208B, BPFP=0.5828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,684B, BPFP=3.0312 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,684B, BPFP=1.2144 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,240B, BPFP=2.9506 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,120B, BPFP=1.2936 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,608B, BPFP=2.6541 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,040B, BPFP=1.2791 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,492B, BPFP=2.8147 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,688B, BPFP=2.1235 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,860B, BPFP=2.8815 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,268B, BPFP=0.3703 +⌛️ [2/4] FRONTEND: Frontend time: 1.874s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10874708 42.66874659 + layer.0.v_cache 0.00001608 0.00649866 + layer.1.k_cache 0.03270756 8.59399272 + layer.1.v_cache 0.00000572 0.00262108 + layer.2.k_cache 0.00777350 0.87253074 + layer.2.v_cache 0.00002080 0.00779407 + layer.3.k_cache 0.05920834 3.41424951 + layer.3.v_cache 0.00001846 0.00842038 + layer.4.k_cache 0.00062493 0.18773367 + layer.4.v_cache 0.00005036 0.01666583 + layer.4.output 0.17137358 581.03862126 + ------------------------------------------------------------------------------------- + TOTAL 0.08286988 242.53232954 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 128892 +BPFP 1.3775 bits/point +EBPFP 2.7550 equivalent bits/point +MSE 242.532330 +---------------------- -------------------------------------------------------- +Time: 3.200s Load: 0.005s, Pack+Encode: 1.874s, Decode+Unpack: 1.321s +---------------------- -------------------------------------------------------- +💾 Converting with 242.5323 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,104B, BPFP=0.6218 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,704B, BPFP=2.9455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,272B, BPFP=1.2564 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,992B, BPFP=2.8029 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,508B, BPFP=1.3037 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,064B, BPFP=2.6170 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,372B, BPFP=1.2764 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,248B, BPFP=2.6538 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,240B, BPFP=2.2516 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,656B, BPFP=2.7356 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,176B, BPFP=0.4057 +⌛️ [2/4] FRONTEND: Frontend time: 1.775s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12073612 40.91750864 + layer.0.v_cache 0.00001348 0.00669224 + layer.1.k_cache 0.03279715 8.25135842 + layer.1.v_cache 0.00000568 0.00286815 + layer.2.k_cache 0.00696802 0.88497749 + layer.2.v_cache 0.00001833 0.00856384 + layer.3.k_cache 0.03227850 4.07926745 + layer.3.v_cache 0.00001834 0.00937652 + layer.4.k_cache 0.00060115 0.20589645 + layer.4.v_cache 0.00005191 0.01799013 + layer.4.output 0.18320683 664.08922848 + ------------------------------------------------------------------------------------- + TOTAL 0.08681979 276.64759404 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 116336 +BPFP 1.3709 bits/point +EBPFP 2.7417 equivalent bits/point +MSE 276.647594 +---------------------- -------------------------------------------------------- +Time: 3.140s Load: 0.003s, Pack+Encode: 1.775s, Decode+Unpack: 1.362s +---------------------- -------------------------------------------------------- +💾 Converting with 276.6476 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.6170 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,740B, BPFP=2.7524 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,232B, BPFP=1.2484 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,748B, BPFP=2.9543 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,516B, BPFP=1.3053 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,880B, BPFP=2.5801 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,472B, BPFP=1.2965 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,684B, BPFP=2.5409 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,144B, BPFP=2.2324 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,888B, BPFP=2.5817 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,400B, BPFP=0.4121 +⌛️ [2/4] FRONTEND: Frontend time: 1.921s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.537s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10595027 41.30607096 + layer.0.v_cache 0.00001376 0.00671910 + layer.1.k_cache 0.03307601 8.58095844 + layer.1.v_cache 0.00000550 0.00270594 + layer.2.k_cache 0.00695119 1.10874000 + layer.2.v_cache 0.00001929 0.00768674 + layer.3.k_cache 0.03450640 3.99826754 + layer.3.v_cache 0.00001844 0.00848035 + layer.4.k_cache 0.00061739 0.18781810 + layer.4.v_cache 0.00005297 0.01678903 + layer.4.output 0.18589628 660.55883700 + ------------------------------------------------------------------------------------- + TOTAL 0.08720501 275.24329972 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 114784 +BPFP 1.3526 bits/point +EBPFP 2.7051 equivalent bits/point +MSE 275.243300 +---------------------- -------------------------------------------------------- +Time: 3.462s Load: 0.004s, Pack+Encode: 1.921s, Decode+Unpack: 1.537s +---------------------- -------------------------------------------------------- +💾 Converting with 275.2433 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,140B, BPFP=0.6210 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,536B, BPFP=2.8750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,368B, BPFP=1.2595 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,644B, BPFP=2.8964 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,832B, BPFP=1.3513 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,188B, BPFP=2.6084 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,468B, BPFP=1.2793 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,884B, BPFP=2.7460 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,408B, BPFP=2.2563 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,548B, BPFP=2.6796 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,728B, BPFP=0.3596 +⌛️ [2/4] FRONTEND: Frontend time: 2.128s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07772056 39.98154420 + layer.0.v_cache 0.00001351 0.00653936 + layer.1.k_cache 0.03518496 8.17891490 + layer.1.v_cache 0.00000556 0.00270859 + layer.2.k_cache 0.00230807 0.88501662 + layer.2.v_cache 0.00001833 0.00793802 + layer.3.k_cache 0.04514034 3.51154723 + layer.3.v_cache 0.00001901 0.00902023 + layer.4.k_cache 0.00062178 0.20504073 + layer.4.v_cache 0.00004990 0.01825860 + layer.4.output 0.18426369 665.35030515 + ------------------------------------------------------------------------------------- + TOTAL 0.08534870 277.07403909 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 116744 +BPFP 1.3582 bits/point +EBPFP 2.7165 equivalent bits/point +MSE 277.074039 +---------------------- -------------------------------------------------------- +Time: 3.479s Load: 0.005s, Pack+Encode: 2.128s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 277.0740 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,076B, BPFP=0.6242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,208B, BPFP=2.8831 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,256B, BPFP=1.2695 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,944B, BPFP=2.8295 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,604B, BPFP=1.3401 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,588B, BPFP=2.5544 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,280B, BPFP=1.2744 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,984B, BPFP=2.8377 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,972B, BPFP=2.2265 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,120B, BPFP=2.6623 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,188B, BPFP=0.4113 +⌛️ [2/4] FRONTEND: Frontend time: 1.801s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.299s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14568528 42.85231078 + layer.0.v_cache 0.00001322 0.00667637 + layer.1.k_cache 0.01379078 8.70953924 + layer.1.v_cache 0.00000548 0.00274908 + layer.2.k_cache 0.00727733 0.88738657 + layer.2.v_cache 0.00001786 0.00857071 + layer.3.k_cache 0.03082054 3.55706391 + layer.3.v_cache 0.00001879 0.00951525 + layer.4.k_cache 0.00061820 0.21071402 + layer.4.v_cache 0.00005377 0.01979592 + layer.4.output 0.19662610 668.37314471 + ------------------------------------------------------------------------------------- + TOTAL 0.09262847 278.52213734 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 115220 +BPFP 1.3753 bits/point +EBPFP 2.7507 equivalent bits/point +MSE 278.522137 +---------------------- -------------------------------------------------------- +Time: 3.106s Load: 0.006s, Pack+Encode: 1.801s, Decode+Unpack: 1.299s +---------------------- -------------------------------------------------------- +💾 Converting with 278.5221 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,120B, BPFP=0.6094 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,000B, BPFP=2.9297 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,492B, BPFP=1.2680 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,540B, BPFP=3.0352 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,804B, BPFP=1.3289 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,008B, BPFP=2.7359 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,620B, BPFP=1.2930 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,476B, BPFP=3.0227 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,252B, BPFP=2.1977 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,640B, BPFP=2.6641 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,276B, BPFP=0.3983 +⌛️ [2/4] FRONTEND: Frontend time: 1.763s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.653s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10149734 43.73874817 + layer.0.v_cache 0.00001421 0.00655205 + layer.1.k_cache 0.01324784 7.59982452 + layer.1.v_cache 0.00000586 0.00275303 + layer.2.k_cache 0.00706157 1.08306789 + layer.2.v_cache 0.00001822 0.00769197 + layer.3.k_cache 0.02712160 4.02731209 + layer.3.v_cache 0.00002016 0.00857844 + layer.4.k_cache 0.00063687 0.21131504 + layer.4.v_cache 0.00005451 0.01863557 + layer.4.output 0.17858309 662.34056920 + ------------------------------------------------------------------------------------- + TOTAL 0.08233881 276.06402724 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 122228 +BPFP 1.4043 bits/point +EBPFP 2.8085 equivalent bits/point +MSE 276.064027 +---------------------- -------------------------------------------------------- +Time: 3.420s Load: 0.003s, Pack+Encode: 1.763s, Decode+Unpack: 1.653s +---------------------- -------------------------------------------------------- +💾 Converting with 276.0640 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,240B, BPFP=0.5819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,176B, BPFP=3.0848 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,924B, BPFP=1.2435 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,656B, BPFP=2.9914 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,388B, BPFP=1.3269 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,332B, BPFP=2.7536 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,256B, BPFP=1.3032 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,324B, BPFP=2.7522 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,192B, BPFP=2.3693 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,828B, BPFP=2.8427 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,528B, BPFP=0.3727 +⌛️ [2/4] FRONTEND: Frontend time: 2.240s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.398s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10772817 41.83729010 + layer.0.v_cache 0.00001348 0.00657285 + layer.1.k_cache 0.03276526 7.82759463 + layer.1.v_cache 0.00000562 0.00269290 + layer.2.k_cache 0.01020141 1.01597455 + layer.2.v_cache 0.00001934 0.00784373 + layer.3.k_cache 0.04694761 3.91766884 + layer.3.v_cache 0.00001910 0.00921234 + layer.4.k_cache 0.00062564 0.20926410 + layer.4.v_cache 0.00005323 0.01702888 + layer.4.output 0.17166949 595.27971059 + ------------------------------------------------------------------------------------- + TOTAL 0.08235678 248.34171277 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 132844 +BPFP 1.4034 bits/point +EBPFP 2.8069 equivalent bits/point +MSE 248.341713 +---------------------- -------------------------------------------------------- +Time: 3.642s Load: 0.004s, Pack+Encode: 2.240s, Decode+Unpack: 1.398s +---------------------- -------------------------------------------------------- +💾 Converting with 248.3417 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,208B, BPFP=0.6039 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,204B, BPFP=3.0505 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,632B, BPFP=1.2485 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,548B, BPFP=2.9270 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,104B, BPFP=1.3373 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,384B, BPFP=2.8961 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,864B, BPFP=1.2922 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,936B, BPFP=2.8117 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,044B, BPFP=2.2673 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,832B, BPFP=2.9804 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,888B, BPFP=0.3735 +⌛️ [2/4] FRONTEND: Frontend time: 2.209s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.407s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12335577 40.56615917 + layer.0.v_cache 0.00001400 0.00646915 + layer.1.k_cache 0.01519604 7.87196001 + layer.1.v_cache 0.00000547 0.00261138 + layer.2.k_cache 0.01098718 1.04230536 + layer.2.v_cache 0.00001779 0.00808482 + layer.3.k_cache 0.07416311 3.13667242 + layer.3.v_cache 0.00001944 0.00926618 + layer.4.k_cache 0.00061719 0.19611184 + layer.4.v_cache 0.00006444 0.01697930 + layer.4.output 0.17504946 638.64656842 + ------------------------------------------------------------------------------------- + TOTAL 0.08528157 266.08132933 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 127644 +BPFP 1.4135 bits/point +EBPFP 2.8270 equivalent bits/point +MSE 266.081329 +---------------------- -------------------------------------------------------- +Time: 3.621s Load: 0.004s, Pack+Encode: 2.209s, Decode+Unpack: 1.407s +---------------------- -------------------------------------------------------- +💾 Converting with 266.0813 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,052B, BPFP=0.6193 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,536B, BPFP=2.9497 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,224B, BPFP=1.2630 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,280B, BPFP=2.6948 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,472B, BPFP=1.3133 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,544B, BPFP=2.9513 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,176B, BPFP=1.2532 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,584B, BPFP=2.9594 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,956B, BPFP=2.2232 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,492B, BPFP=2.7378 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,636B, BPFP=0.3953 +⌛️ [2/4] FRONTEND: Frontend time: 1.822s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13955514 42.40101842 + layer.0.v_cache 0.00001392 0.00700836 + layer.1.k_cache 0.03454666 8.50160356 + layer.1.v_cache 0.00000526 0.00277498 + layer.2.k_cache 0.00236268 1.21877001 + layer.2.v_cache 0.00001660 0.00799286 + layer.3.k_cache 0.02874016 3.53002930 + layer.3.v_cache 0.00001899 0.00974143 + layer.4.k_cache 0.00062191 0.18823351 + layer.4.v_cache 0.00005265 0.01804891 + layer.4.output 0.18728454 680.85053340 + ------------------------------------------------------------------------------------- + TOTAL 0.08923092 283.63758560 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 116952 +BPFP 1.3960 bits/point +EBPFP 2.7920 equivalent bits/point +MSE 283.637586 +---------------------- -------------------------------------------------------- +Time: 3.048s Load: 0.004s, Pack+Encode: 1.822s, Decode+Unpack: 1.222s +---------------------- -------------------------------------------------------- +💾 Converting with 283.6376 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,088B, BPFP=0.6266 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,568B, BPFP=2.7532 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,256B, BPFP=1.2695 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,428B, BPFP=2.9278 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,636B, BPFP=1.3466 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,712B, BPFP=2.5795 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,568B, BPFP=1.3328 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,620B, BPFP=2.7638 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,004B, BPFP=2.2330 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,892B, BPFP=2.6161 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,420B, BPFP=0.4180 +⌛️ [2/4] FRONTEND: Frontend time: 1.742s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.245s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11454884 42.02530185 + layer.0.v_cache 0.00001351 0.00726759 + layer.1.k_cache 0.01665594 8.40029630 + layer.1.v_cache 0.00000570 0.00307403 + layer.2.k_cache 0.00836552 1.03335651 + layer.2.v_cache 0.00001974 0.00900316 + layer.3.k_cache 0.03124422 3.65526908 + layer.3.v_cache 0.00001876 0.00974940 + layer.4.k_cache 0.00062231 0.21191486 + layer.4.v_cache 0.00005487 0.02057755 + layer.4.output 0.19663499 664.62482607 + ------------------------------------------------------------------------------------- + TOTAL 0.09105849 276.92644663 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 115192 +BPFP 1.3750 bits/point +EBPFP 2.7500 equivalent bits/point +MSE 276.926447 +---------------------- -------------------------------------------------------- +Time: 2.990s Load: 0.003s, Pack+Encode: 1.742s, Decode+Unpack: 1.245s +---------------------- -------------------------------------------------------- +💾 Converting with 276.9264 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,208B, BPFP=0.5967 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,264B, BPFP=3.0253 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,768B, BPFP=1.2589 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,240B, BPFP=2.8348 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,004B, BPFP=1.3028 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,488B, BPFP=2.6949 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,896B, BPFP=1.2827 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,040B, BPFP=2.7976 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,616B, BPFP=2.1607 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,040B, BPFP=2.6116 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,408B, BPFP=0.3829 +⌛️ [2/4] FRONTEND: Frontend time: 2.135s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.426s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12290487 42.15727016 + layer.0.v_cache 0.00001422 0.00639944 + layer.1.k_cache 0.01283375 8.32808867 + layer.1.v_cache 0.00000579 0.00274973 + layer.2.k_cache 0.01100613 0.87697819 + layer.2.v_cache 0.00001929 0.00777893 + layer.3.k_cache 0.03011658 3.83323706 + layer.3.v_cache 0.00001858 0.00837021 + layer.4.k_cache 0.00063020 0.19046695 + layer.4.v_cache 0.00005223 0.01949260 + layer.4.output 0.17536174 626.66943027 + ------------------------------------------------------------------------------------- + TOTAL 0.08265493 261.30099081 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 124972 +BPFP 1.3674 bits/point +EBPFP 2.7349 equivalent bits/point +MSE 261.300991 +---------------------- -------------------------------------------------------- +Time: 3.565s Load: 0.003s, Pack+Encode: 2.135s, Decode+Unpack: 1.426s +---------------------- -------------------------------------------------------- +💾 Converting with 261.3010 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,208B, BPFP=0.5897 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,164B, BPFP=2.9713 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,716B, BPFP=1.2346 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,052B, BPFP=2.7669 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,296B, BPFP=1.3412 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,956B, BPFP=2.7493 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,796B, BPFP=1.2493 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,608B, BPFP=2.6853 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,140B, BPFP=2.2316 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,348B, BPFP=2.6375 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,556B, BPFP=0.3822 +⌛️ [2/4] FRONTEND: Frontend time: 1.731s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.444s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11218034 41.41236500 + layer.0.v_cache 0.00001438 0.00663775 + layer.1.k_cache 0.03157643 7.98971019 + layer.1.v_cache 0.00000545 0.00267783 + layer.2.k_cache 0.00802605 0.98481616 + layer.2.v_cache 0.00001826 0.00787451 + layer.3.k_cache 0.04519741 3.43896772 + layer.3.v_cache 0.00001794 0.00842685 + layer.4.k_cache 0.00061137 0.20564665 + layer.4.v_cache 0.00005913 0.01679031 + layer.4.output 0.17010492 602.99721639 + ------------------------------------------------------------------------------------- + TOTAL 0.08167301 251.47378986 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 125840 +BPFP 1.3607 bits/point +EBPFP 2.7215 equivalent bits/point +MSE 251.473790 +---------------------- -------------------------------------------------------- +Time: 3.179s Load: 0.004s, Pack+Encode: 1.731s, Decode+Unpack: 1.444s +---------------------- -------------------------------------------------------- +💾 Converting with 251.4738 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,252B, BPFP=0.5908 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,212B, BPFP=2.9455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,808B, BPFP=1.2369 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,740B, BPFP=2.8597 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,128B, BPFP=1.2951 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,668B, BPFP=2.8467 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,104B, BPFP=1.2907 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,444B, BPFP=2.8060 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,852B, BPFP=2.1533 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,140B, BPFP=2.7507 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,540B, BPFP=0.3774 +⌛️ [2/4] FRONTEND: Frontend time: 2.188s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.515s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10146660 41.58517101 + layer.0.v_cache 0.00001344 0.00651086 + layer.1.k_cache 0.03090366 8.19491861 + layer.1.v_cache 0.00000566 0.00271903 + layer.2.k_cache 0.01188999 0.91429812 + layer.2.v_cache 0.00001845 0.00809208 + layer.3.k_cache 0.02542871 3.58152097 + layer.3.v_cache 0.00001884 0.00899538 + layer.4.k_cache 0.00060243 0.20170480 + layer.4.v_cache 0.00005224 0.01800671 + layer.4.output 0.16839122 590.64856728 + ------------------------------------------------------------------------------------- + TOTAL 0.07936109 246.41540638 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 128888 +BPFP 1.3775 bits/point +EBPFP 2.7550 equivalent bits/point +MSE 246.415406 +---------------------- -------------------------------------------------------- +Time: 3.708s Load: 0.004s, Pack+Encode: 2.188s, Decode+Unpack: 1.515s +---------------------- -------------------------------------------------------- +💾 Converting with 246.4154 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,332B, BPFP=0.5659 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,840B, BPFP=3.0299 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,092B, BPFP=1.2045 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,056B, BPFP=2.8967 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,648B, BPFP=1.2989 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,840B, BPFP=2.8601 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,452B, BPFP=1.2656 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,540B, BPFP=2.8091 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,040B, BPFP=2.3845 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,536B, BPFP=2.8084 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,804B, BPFP=0.3592 +⌛️ [2/4] FRONTEND: Frontend time: 1.907s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.291s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12981709 43.61609417 + layer.0.v_cache 0.00001628 0.00662021 + layer.1.k_cache 0.06761242 8.08946029 + layer.1.v_cache 0.00000547 0.00255787 + layer.2.k_cache 0.00384689 0.86128699 + layer.2.v_cache 0.00001812 0.00809844 + layer.3.k_cache 0.05327560 3.74412172 + layer.3.v_cache 0.00001896 0.00900853 + layer.4.k_cache 0.00060800 0.20100780 + layer.4.v_cache 0.00005116 0.01666583 + layer.4.output 0.15518735 577.57026398 + ------------------------------------------------------------------------------------- + TOTAL 0.07891655 241.14980998 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 139180 +BPFP 1.3905 bits/point +EBPFP 2.7809 equivalent bits/point +MSE 241.149810 +---------------------- -------------------------------------------------------- +Time: 3.203s Load: 0.005s, Pack+Encode: 1.907s, Decode+Unpack: 1.291s +---------------------- -------------------------------------------------------- +💾 Converting with 241.1498 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,164B, BPFP=0.6103 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,444B, BPFP=2.9792 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,448B, BPFP=1.2438 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,692B, BPFP=2.8341 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,780B, BPFP=1.3079 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,112B, BPFP=2.7222 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,508B, BPFP=1.2554 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,188B, BPFP=2.7369 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,724B, BPFP=2.2616 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,404B, BPFP=2.5856 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,784B, BPFP=0.3799 +⌛️ [2/4] FRONTEND: Frontend time: 1.795s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.184s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887507 43.35341315 + layer.0.v_cache 0.00001345 0.00685253 + layer.1.k_cache 0.03259810 8.71726406 + layer.1.v_cache 0.00000523 0.00267448 + layer.2.k_cache 0.00244350 1.05538196 + layer.2.v_cache 0.00001834 0.00826526 + layer.3.k_cache 0.10864470 4.18905866 + layer.3.v_cache 0.00001904 0.00953704 + layer.4.k_cache 0.00060864 0.20144347 + layer.4.v_cache 0.00004940 0.01733388 + layer.4.output 0.18628448 614.45155423 + ------------------------------------------------------------------------------------- + TOTAL 0.09395687 256.39541789 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 120248 +BPFP 1.3645 bits/point +EBPFP 2.7289 equivalent bits/point +MSE 256.395418 +---------------------- -------------------------------------------------------- +Time: 2.984s Load: 0.004s, Pack+Encode: 1.795s, Decode+Unpack: 1.184s +---------------------- -------------------------------------------------------- +💾 Converting with 256.3954 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,264B, BPFP=0.5730 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,904B, BPFP=2.9677 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,944B, BPFP=1.2191 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,416B, BPFP=2.8820 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,336B, BPFP=1.2879 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,276B, BPFP=2.6819 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,232B, BPFP=1.2697 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,984B, BPFP=2.8062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,228B, BPFP=2.1468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,856B, BPFP=2.7837 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,788B, BPFP=0.3960 +⌛️ [2/4] FRONTEND: Frontend time: 1.709s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.195s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422350 43.24858453 + layer.0.v_cache 0.00001493 0.00668473 + layer.1.k_cache 0.01422887 7.64646517 + layer.1.v_cache 0.00000595 0.00272727 + layer.2.k_cache 0.00668612 0.87474934 + layer.2.v_cache 0.00001902 0.00746170 + layer.3.k_cache 0.02497406 3.85554556 + layer.3.v_cache 0.00001869 0.00820482 + layer.4.k_cache 0.00061935 0.18544471 + layer.4.v_cache 0.00005212 0.01713659 + layer.4.output 0.16899524 584.97717697 + ------------------------------------------------------------------------------------- + TOTAL 0.07904761 244.15842607 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 133228 +BPFP 1.3759 bits/point +EBPFP 2.7517 equivalent bits/point +MSE 244.158426 +---------------------- -------------------------------------------------------- +Time: 2.908s Load: 0.004s, Pack+Encode: 1.709s, Decode+Unpack: 1.195s +---------------------- -------------------------------------------------------- +💾 Converting with 244.1584 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 113, 128) +Output shape: (1, 113, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.output: torch.Size([1, 113, 3584]) -> torch.Size([1, 1, 113, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,924B, BPFP=0.5426 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,272B, BPFP=2.5265 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,644B, BPFP=1.1952 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,904B, BPFP=2.4757 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,408B, BPFP=1.4392 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,484B, BPFP=2.4176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,104B, BPFP=1.2588 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,676B, BPFP=2.4441 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,028B, BPFP=2.2163 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,576B, BPFP=2.4303 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,960B, BPFP=0.3350 +⌛️ [2/4] FRONTEND: Frontend time: 1.694s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.169s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10214825 43.06695935 + layer.0.v_cache 0.00001637 0.00648466 + layer.1.k_cache 0.04284562 8.34765409 + layer.1.v_cache 0.00000578 0.00271234 + layer.2.k_cache 0.00913706 1.00570125 + layer.2.v_cache 0.00001928 0.00803065 + layer.3.k_cache 0.05585717 3.65735592 + layer.3.v_cache 0.00001870 0.00839667 + layer.4.k_cache 0.00065968 0.19470704 + layer.4.v_cache 0.00005753 0.01767324 + layer.4.output 10.09706179 435.28634640 + ------------------------------------------------------------------------------------- + TOTAL 4.17001165 182.54824118 + (elements=983,552) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 983552 +Total Bytes 153980 +BPFP 1.2524 bits/point +EBPFP 2.5049 equivalent bits/point +MSE 182.548241 +---------------------- -------------------------------------------------------- +Time: 2.868s Load: 0.004s, Pack+Encode: 1.694s, Decode+Unpack: 1.169s +---------------------- -------------------------------------------------------- +💾 Converting with 182.5482 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,328B, BPFP=0.5714 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,072B, BPFP=3.1030 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,024B, BPFP=1.2060 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,464B, BPFP=2.9986 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,588B, BPFP=1.3029 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,740B, BPFP=2.8743 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,400B, BPFP=1.2706 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,692B, BPFP=2.8661 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,028B, BPFP=2.5804 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,052B, BPFP=2.9279 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,276B, BPFP=0.3992 +⌛️ [2/4] FRONTEND: Frontend time: 2.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.777s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11481400 41.77942029 + layer.0.v_cache 0.00001396 0.00644658 + layer.1.k_cache 0.03135788 7.81988056 + layer.1.v_cache 0.00000545 0.00264867 + layer.2.k_cache 0.00370715 0.94197921 + layer.2.v_cache 0.00001965 0.00773390 + layer.3.k_cache 0.03930106 3.74483750 + layer.3.v_cache 0.00001886 0.00839477 + layer.4.k_cache 0.00060819 0.18628261 + layer.4.v_cache 0.00005437 0.01705282 + layer.4.output 0.16591480 576.82530416 + ------------------------------------------------------------------------------------- + TOTAL 0.07948848 240.72304741 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 142664 +BPFP 1.4409 bits/point +EBPFP 2.8819 equivalent bits/point +MSE 240.723047 +---------------------- -------------------------------------------------------- +Time: 3.945s Load: 0.006s, Pack+Encode: 2.162s, Decode+Unpack: 1.777s +---------------------- -------------------------------------------------------- +💾 Converting with 240.7230 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,300B, BPFP=0.5794 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,428B, BPFP=3.0597 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,008B, BPFP=1.2303 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,068B, BPFP=2.8209 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,328B, BPFP=1.2865 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,100B, BPFP=2.8265 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,828B, BPFP=1.1987 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,776B, BPFP=2.7697 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,476B, BPFP=2.3659 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,632B, BPFP=2.7444 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,792B, BPFP=0.3459 +⌛️ [2/4] FRONTEND: Frontend time: 2.296s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09788724 41.08104646 + layer.0.v_cache 0.00001308 0.00645975 + layer.1.k_cache 0.03299916 7.31865289 + layer.1.v_cache 0.00000544 0.00267689 + layer.2.k_cache 0.00367894 0.94533547 + layer.2.v_cache 0.00002012 0.00836370 + layer.3.k_cache 0.03939044 4.10175795 + layer.3.v_cache 0.00001875 0.00904816 + layer.4.k_cache 0.00061788 0.19346473 + layer.4.v_cache 0.00005010 0.01792854 + layer.4.output 0.17021053 579.52994583 + ------------------------------------------------------------------------------------- + TOTAL 0.08036205 241.78790325 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 132736 +BPFP 1.3708 bits/point +EBPFP 2.7416 equivalent bits/point +MSE 241.787903 +---------------------- -------------------------------------------------------- +Time: 3.686s Load: 0.005s, Pack+Encode: 2.296s, Decode+Unpack: 1.384s +---------------------- -------------------------------------------------------- +💾 Converting with 241.7879 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,920B, BPFP=0.5671 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,168B, BPFP=2.6285 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,276B, BPFP=1.3420 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,840B, BPFP=2.5810 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,172B, BPFP=1.4716 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,504B, BPFP=2.5324 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,780B, BPFP=1.2703 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,068B, BPFP=2.6140 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,016B, BPFP=2.3171 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,604B, BPFP=2.5469 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,136B, BPFP=0.4575 +⌛️ [2/4] FRONTEND: Frontend time: 1.814s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.335s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12672405 42.15594256 + layer.0.v_cache 0.00001401 0.00634152 + layer.1.k_cache 0.12044146 8.84069768 + layer.1.v_cache 0.00000568 0.00258484 + layer.2.k_cache 0.00570256 1.05260277 + layer.2.v_cache 0.00001866 0.00781899 + layer.3.k_cache 0.03611955 4.18499502 + layer.3.v_cache 0.00001983 0.00832747 + layer.4.k_cache 0.00066233 0.21962353 + layer.4.v_cache 0.00005331 0.01680418 + layer.4.output 10.56037249 479.26649306 + ------------------------------------------------------------------------------------- + TOTAL 4.36543346 200.66830529 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 159484 +BPFP 1.3573 bits/point +EBPFP 2.7145 equivalent bits/point +MSE 200.668305 +---------------------- -------------------------------------------------------- +Time: 3.154s Load: 0.004s, Pack+Encode: 1.814s, Decode+Unpack: 1.335s +---------------------- -------------------------------------------------------- +💾 Converting with 200.6683 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,272B, BPFP=0.5810 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,292B, BPFP=3.0703 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,964B, BPFP=1.2365 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,340B, BPFP=2.9013 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,268B, BPFP=1.2905 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,148B, BPFP=2.8672 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,044B, BPFP=1.2507 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,748B, BPFP=2.7962 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,968B, BPFP=2.3026 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,664B, BPFP=2.7812 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,600B, BPFP=0.3450 +⌛️ [2/4] FRONTEND: Frontend time: 1.717s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.255s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13733016 43.12845681 + layer.0.v_cache 0.00001373 0.00658455 + layer.1.k_cache 0.03155638 8.11287204 + layer.1.v_cache 0.00000617 0.00298315 + layer.2.k_cache 0.00751131 1.05882922 + layer.2.v_cache 0.00001754 0.00834326 + layer.3.k_cache 0.02548880 3.99555484 + layer.3.v_cache 0.00001765 0.00897969 + layer.4.k_cache 0.00061011 0.19911814 + layer.4.v_cache 0.00005045 0.01798795 + layer.4.output 0.16239834 581.02901786 + ------------------------------------------------------------------------------------- + TOTAL 0.07878769 242.57310792 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 132308 +BPFP 1.3819 bits/point +EBPFP 2.7638 equivalent bits/point +MSE 242.573108 +---------------------- -------------------------------------------------------- +Time: 2.976s Load: 0.005s, Pack+Encode: 1.717s, Decode+Unpack: 1.255s +---------------------- -------------------------------------------------------- +💾 Converting with 242.5731 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,132B, BPFP=0.6195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,096B, BPFP=2.7880 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,384B, BPFP=1.2627 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,160B, BPFP=2.8006 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,500B, BPFP=1.2856 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,708B, BPFP=2.5134 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,652B, BPFP=1.3157 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,184B, BPFP=2.6076 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,920B, BPFP=2.1598 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,776B, BPFP=2.7247 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,460B, BPFP=0.3803 +⌛️ [2/4] FRONTEND: Frontend time: 1.846s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.301s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10979201 41.97088855 + layer.0.v_cache 0.00001339 0.00648303 + layer.1.k_cache 0.01160004 8.42476915 + layer.1.v_cache 0.00000567 0.00267327 + layer.2.k_cache 0.00714037 0.99395250 + layer.2.v_cache 0.00001829 0.00765681 + layer.3.k_cache 0.04622446 3.97398203 + layer.3.v_cache 0.00001896 0.00877590 + layer.4.k_cache 0.00064378 0.20986031 + layer.4.v_cache 0.00005251 0.01776511 + layer.4.output 0.18415536 661.92828888 + ------------------------------------------------------------------------------------- + TOTAL 0.08615276 275.83028405 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 114972 +BPFP 1.3376 bits/point +EBPFP 2.6753 equivalent bits/point +MSE 275.830284 +---------------------- -------------------------------------------------------- +Time: 3.150s Load: 0.004s, Pack+Encode: 1.846s, Decode+Unpack: 1.301s +---------------------- -------------------------------------------------------- +💾 Converting with 275.8303 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,056B, BPFP=0.6201 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,456B, BPFP=2.9334 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,184B, BPFP=1.2549 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,924B, BPFP=2.8255 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,404B, BPFP=1.2995 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,744B, BPFP=2.5860 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,376B, BPFP=1.2938 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,424B, BPFP=2.7240 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,536B, BPFP=2.1380 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,612B, BPFP=2.5593 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,056B, BPFP=0.4365 +⌛️ [2/4] FRONTEND: Frontend time: 1.867s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11290005 40.96025898 + layer.0.v_cache 0.00001570 0.00687737 + layer.1.k_cache 0.03434501 9.62002999 + layer.1.v_cache 0.00000534 0.00260039 + layer.2.k_cache 0.00238675 1.01647672 + layer.2.v_cache 0.00001838 0.00809717 + layer.3.k_cache 0.02838124 4.14824568 + layer.3.v_cache 0.00002047 0.00949762 + layer.4.k_cache 0.00062757 0.18747855 + layer.4.v_cache 0.00005455 0.01870075 + layer.4.output 0.18420308 669.68842764 + ------------------------------------------------------------------------------------- + TOTAL 0.08636333 279.04689745 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 114772 +BPFP 1.3700 bits/point +EBPFP 2.7400 equivalent bits/point +MSE 279.046897 +---------------------- -------------------------------------------------------- +Time: 3.182s Load: 0.004s, Pack+Encode: 1.867s, Decode+Unpack: 1.311s +---------------------- -------------------------------------------------------- +💾 Converting with 279.0469 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,096B, BPFP=0.6202 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,276B, BPFP=2.8598 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,356B, BPFP=1.2732 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,828B, BPFP=2.7700 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,476B, BPFP=1.2973 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,360B, BPFP=2.6763 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,508B, BPFP=1.3037 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,756B, BPFP=2.5553 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,124B, BPFP=2.2284 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,436B, BPFP=2.6915 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,772B, BPFP=0.4227 +⌛️ [2/4] FRONTEND: Frontend time: 1.806s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.292s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09752406 40.02284280 + layer.0.v_cache 0.00001393 0.00690728 + layer.1.k_cache 0.01493649 8.73879614 + layer.1.v_cache 0.00000588 0.00273242 + layer.2.k_cache 0.01003601 1.06489074 + layer.2.v_cache 0.00001805 0.00825594 + layer.3.k_cache 0.04451986 3.94184758 + layer.3.v_cache 0.00001914 0.00959041 + layer.4.k_cache 0.00061869 0.19247967 + layer.4.v_cache 0.00005251 0.01960443 + layer.4.output 0.18409089 658.49095696 + ------------------------------------------------------------------------------------- + TOTAL 0.08566946 274.32027330 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 115988 +BPFP 1.3668 bits/point +EBPFP 2.7335 equivalent bits/point +MSE 274.320273 +---------------------- -------------------------------------------------------- +Time: 3.101s Load: 0.003s, Pack+Encode: 1.806s, Decode+Unpack: 1.292s +---------------------- -------------------------------------------------------- +💾 Converting with 274.3203 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,212B, BPFP=0.6047 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,840B, BPFP=2.7937 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,548B, BPFP=1.2327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,972B, BPFP=2.8185 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,828B, BPFP=1.2854 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,280B, BPFP=2.6883 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,680B, BPFP=1.2575 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,280B, BPFP=2.6883 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,836B, BPFP=2.2282 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,904B, BPFP=2.6175 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,556B, BPFP=0.3915 +⌛️ [2/4] FRONTEND: Frontend time: 2.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.230s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13658243 41.64017790 + layer.0.v_cache 0.00001379 0.00648933 + layer.1.k_cache 0.01234693 7.95513548 + layer.1.v_cache 0.00000566 0.00264788 + layer.2.k_cache 0.00375735 1.08193924 + layer.2.v_cache 0.00001852 0.00777892 + layer.3.k_cache 0.02700986 4.09984101 + layer.3.v_cache 0.00001930 0.00830341 + layer.4.k_cache 0.00062007 0.19734950 + layer.4.v_cache 0.00005146 0.01716281 + layer.4.output 0.17614663 634.08578959 + ------------------------------------------------------------------------------------- + TOTAL 0.08314423 264.33043251 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 121936 +BPFP 1.3503 bits/point +EBPFP 2.7006 equivalent bits/point +MSE 264.330433 +---------------------- -------------------------------------------------------- +Time: 3.398s Load: 0.004s, Pack+Encode: 2.164s, Decode+Unpack: 1.230s +---------------------- -------------------------------------------------------- +💾 Converting with 264.3304 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,924B, BPFP=0.6259 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,348B, BPFP=2.8570 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,904B, BPFP=1.2637 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,708B, BPFP=2.7200 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,336B, BPFP=1.3562 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,500B, BPFP=2.6755 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,160B, BPFP=1.3185 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,944B, BPFP=2.5565 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,264B, BPFP=2.1969 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,900B, BPFP=2.7611 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,464B, BPFP=0.4423 +⌛️ [2/4] FRONTEND: Frontend time: 34.684s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08738674 41.56747646 + layer.0.v_cache 0.00001368 0.00693687 + layer.1.k_cache 0.01286346 8.53457349 + layer.1.v_cache 0.00000597 0.00289175 + layer.2.k_cache 0.00251921 0.91234024 + layer.2.v_cache 0.00001769 0.00858843 + layer.3.k_cache 0.08324167 3.66974797 + layer.3.v_cache 0.00002021 0.01007993 + layer.4.k_cache 0.00062021 0.21336490 + layer.4.v_cache 0.00005375 0.01858866 + layer.4.output 0.19567458 686.57118395 + ------------------------------------------------------------------------------------- + TOTAL 0.09155674 285.93781626 + (elements=635,392) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 635392 +Total Bytes 109452 +BPFP 1.3781 bits/point +EBPFP 2.7561 equivalent bits/point +MSE 285.937816 +---------------------- --------------------------------------------------------- +Time: 36.061s Load: 0.004s, Pack+Encode: 34.684s, Decode+Unpack: 1.373s +---------------------- --------------------------------------------------------- +💾 Converting with 285.9378 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,208B, BPFP=0.6039 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,536B, BPFP=2.9247 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,656B, BPFP=1.2530 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,580B, BPFP=2.9330 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,828B, BPFP=1.2854 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,332B, BPFP=2.6980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,740B, BPFP=1.2688 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,196B, BPFP=3.0489 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,196B, BPFP=2.1077 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,572B, BPFP=2.7432 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,780B, BPFP=0.3975 +⌛️ [2/4] FRONTEND: Frontend time: 1.886s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.424s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12176956 41.75428276 + layer.0.v_cache 0.00001396 0.00634242 + layer.1.k_cache 0.01120206 8.84992411 + layer.1.v_cache 0.00000562 0.00259638 + layer.2.k_cache 0.00678912 1.05662619 + layer.2.v_cache 0.00001849 0.00745577 + layer.3.k_cache 0.02622183 4.08912052 + layer.3.v_cache 0.00001965 0.00867660 + layer.4.k_cache 0.00062413 0.19757216 + layer.4.v_cache 0.00005248 0.01794534 + layer.4.output 0.17940697 637.13946859 + ------------------------------------------------------------------------------------- + TOTAL 0.08368034 265.64510720 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 125624 +BPFP 1.3911 bits/point +EBPFP 2.7822 equivalent bits/point +MSE 265.645107 +---------------------- -------------------------------------------------------- +Time: 3.315s Load: 0.004s, Pack+Encode: 1.886s, Decode+Unpack: 1.424s +---------------------- -------------------------------------------------------- +💾 Converting with 265.6451 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,832B, BPFP=0.5702 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,112B, BPFP=2.6952 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,264B, BPFP=1.3786 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,680B, BPFP=2.6310 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,724B, BPFP=1.4470 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,588B, BPFP=2.6173 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,192B, BPFP=1.3679 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,084B, BPFP=2.6911 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,020B, BPFP=2.3839 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,688B, BPFP=2.6321 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,724B, BPFP=0.4193 +⌛️ [2/4] FRONTEND: Frontend time: 2.169s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.269s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120816 38.85795666 + layer.0.v_cache 0.00001607 0.00669639 + layer.1.k_cache 0.10678518 8.81765602 + layer.1.v_cache 0.00000585 0.00258251 + layer.2.k_cache 0.00813519 0.98502502 + layer.2.v_cache 0.00001919 0.00778927 + layer.3.k_cache 0.06560528 5.92738095 + layer.3.v_cache 0.00001973 0.00865011 + layer.4.k_cache 0.00062300 0.19855993 + layer.4.v_cache 0.00005287 0.01629231 + layer.4.output 10.86395687 484.29549320 + ------------------------------------------------------------------------------------- + TOTAL 4.49059815 202.64100244 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 156908 +BPFP 1.3735 bits/point +EBPFP 2.7470 equivalent bits/point +MSE 202.641002 +---------------------- -------------------------------------------------------- +Time: 3.443s Load: 0.005s, Pack+Encode: 2.169s, Decode+Unpack: 1.269s +---------------------- -------------------------------------------------------- +💾 Converting with 202.6410 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,228B, BPFP=0.5797 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,532B, BPFP=3.1487 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,876B, BPFP=1.2349 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,684B, BPFP=2.8168 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,188B, BPFP=1.2909 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,152B, BPFP=2.7213 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,948B, BPFP=1.2478 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,160B, BPFP=2.9023 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,544B, BPFP=2.4325 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,176B, BPFP=2.9052 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,092B, BPFP=0.3872 +⌛️ [2/4] FRONTEND: Frontend time: 1.968s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12848519 39.93598745 + layer.0.v_cache 0.00001371 0.00624474 + layer.1.k_cache 0.03007136 8.53898533 + layer.1.v_cache 0.00000570 0.00269626 + layer.2.k_cache 0.00936044 0.96260650 + layer.2.v_cache 0.00001842 0.00777580 + layer.3.k_cache 0.04065019 3.59522502 + layer.3.v_cache 0.00001877 0.00859153 + layer.4.k_cache 0.00062204 0.19189659 + layer.4.v_cache 0.00010837 0.01713471 + layer.4.output 0.16426703 593.63387726 + ------------------------------------------------------------------------------------- + TOTAL 0.07995432 247.57084028 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 133580 +BPFP 1.4112 bits/point +EBPFP 2.8224 equivalent bits/point +MSE 247.570840 +---------------------- -------------------------------------------------------- +Time: 3.186s Load: 0.004s, Pack+Encode: 1.968s, Decode+Unpack: 1.215s +---------------------- -------------------------------------------------------- +💾 Converting with 247.5708 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,136B, BPFP=0.6125 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,976B, BPFP=2.9250 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,392B, BPFP=1.2484 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,176B, BPFP=2.9641 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,852B, BPFP=1.3383 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,384B, BPFP=2.6141 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,744B, BPFP=1.3172 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,148B, BPFP=2.5680 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,584B, BPFP=2.2625 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,320B, BPFP=2.6016 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,088B, BPFP=0.3931 +⌛️ [2/4] FRONTEND: Frontend time: 1.931s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.253s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09257430 40.32271118 + layer.0.v_cache 0.00001329 0.00625005 + layer.1.k_cache 0.01202901 8.31910858 + layer.1.v_cache 0.00000544 0.00252613 + layer.2.k_cache 0.00549439 0.92709742 + layer.2.v_cache 0.00001849 0.00758304 + layer.3.k_cache 0.11014032 4.36787872 + layer.3.v_cache 0.00001916 0.00842365 + layer.4.k_cache 0.00063444 0.20529284 + layer.4.v_cache 0.00005494 0.01655127 + layer.4.output 0.19143520 657.07410714 + ------------------------------------------------------------------------------------- + TOTAL 0.09182531 273.74718664 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 118800 +BPFP 1.3649 bits/point +EBPFP 2.7298 equivalent bits/point +MSE 273.747187 +---------------------- -------------------------------------------------------- +Time: 3.187s Load: 0.003s, Pack+Encode: 1.931s, Decode+Unpack: 1.253s +---------------------- -------------------------------------------------------- +💾 Converting with 273.7472 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,232B, BPFP=0.6012 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,548B, BPFP=3.0781 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,700B, BPFP=1.2463 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,176B, BPFP=3.0089 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,948B, BPFP=1.2924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,228B, BPFP=2.6466 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,828B, BPFP=1.2701 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,412B, BPFP=2.6808 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,064B, BPFP=2.2440 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,424B, BPFP=2.8690 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,544B, BPFP=0.3599 +⌛️ [2/4] FRONTEND: Frontend time: 1.752s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.259s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12365316 44.24245780 + layer.0.v_cache 0.00001332 0.00644698 + layer.1.k_cache 0.03511774 7.70996457 + layer.1.v_cache 0.00000569 0.00277415 + layer.2.k_cache 0.01236608 0.83347075 + layer.2.v_cache 0.00001835 0.00773278 + layer.3.k_cache 0.02851222 3.69619533 + layer.3.v_cache 0.00001844 0.00869501 + layer.4.k_cache 0.00063207 0.18739500 + layer.4.v_cache 0.00005066 0.01692557 + layer.4.output 0.17964420 627.59109269 + ------------------------------------------------------------------------------------- + TOTAL 0.08575865 261.75586510 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 126104 +BPFP 1.3798 bits/point +EBPFP 2.7596 equivalent bits/point +MSE 261.755865 +---------------------- -------------------------------------------------------- +Time: 3.016s Load: 0.006s, Pack+Encode: 1.752s, Decode+Unpack: 1.259s +---------------------- -------------------------------------------------------- +💾 Converting with 261.7559 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,088B, BPFP=0.6186 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,592B, BPFP=2.7228 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,180B, BPFP=1.2380 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,360B, BPFP=2.6763 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,460B, BPFP=1.2941 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,776B, BPFP=2.5593 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,356B, BPFP=1.2732 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,480B, BPFP=2.7003 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,704B, BPFP=2.1442 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,712B, BPFP=2.7468 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,592B, BPFP=0.4176 +⌛️ [2/4] FRONTEND: Frontend time: 2.380s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.303s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12096892 41.26956255 + layer.0.v_cache 0.00001393 0.00663454 + layer.1.k_cache 0.03247837 8.60875448 + layer.1.v_cache 0.00000613 0.00271154 + layer.2.k_cache 0.00394396 0.95133415 + layer.2.v_cache 0.00001920 0.00823916 + layer.3.k_cache 0.02751121 3.82352897 + layer.3.v_cache 0.00001919 0.00929359 + layer.4.k_cache 0.00062481 0.20561962 + layer.4.v_cache 0.00005293 0.02035401 + layer.4.output 0.18564256 653.34489469 + ------------------------------------------------------------------------------------- + TOTAL 0.08736098 272.25413502 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 114300 +BPFP 1.3469 bits/point +EBPFP 2.6937 equivalent bits/point +MSE 272.254135 +---------------------- -------------------------------------------------------- +Time: 3.688s Load: 0.005s, Pack+Encode: 2.380s, Decode+Unpack: 1.303s +---------------------- -------------------------------------------------------- +💾 Converting with 272.2541 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,084B, BPFP=0.6258 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,536B, BPFP=2.7468 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,244B, BPFP=1.2670 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,136B, BPFP=2.6656 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,572B, BPFP=1.3336 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,172B, BPFP=2.6729 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,432B, BPFP=1.3052 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,740B, BPFP=2.5852 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,212B, BPFP=2.2752 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,184B, BPFP=2.6753 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,068B, BPFP=0.4368 +⌛️ [2/4] FRONTEND: Frontend time: 1.976s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08564486 41.80493481 + layer.0.v_cache 0.00001408 0.00697994 + layer.1.k_cache 0.03507188 8.55763066 + layer.1.v_cache 0.00000566 0.00268026 + layer.2.k_cache 0.00239554 0.90757851 + layer.2.v_cache 0.00001938 0.00827152 + layer.3.k_cache 0.03313774 3.84080882 + layer.3.v_cache 0.00001993 0.00990320 + layer.4.k_cache 0.00061714 0.20554998 + layer.4.v_cache 0.00005134 0.01806867 + layer.4.output 0.18226704 671.84577922 + ------------------------------------------------------------------------------------- + TOTAL 0.08428511 279.89899182 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 114380 +BPFP 1.3653 bits/point +EBPFP 2.7306 equivalent bits/point +MSE 279.898992 +---------------------- -------------------------------------------------------- +Time: 3.309s Load: 0.003s, Pack+Encode: 1.976s, Decode+Unpack: 1.330s +---------------------- -------------------------------------------------------- +💾 Converting with 279.8990 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 117, 128) +Output shape: (1, 117, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.output: torch.Size([1, 117, 3584]) -> torch.Size([1, 1, 117, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,048B, BPFP=0.5406 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,460B, BPFP=2.4653 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,468B, BPFP=1.2644 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,968B, BPFP=2.3996 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,368B, BPFP=1.5182 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,676B, BPFP=2.3606 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,884B, BPFP=1.3200 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,892B, BPFP=2.3894 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,168B, BPFP=2.1592 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,528B, BPFP=2.3408 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,300B, BPFP=0.3682 +⌛️ [2/4] FRONTEND: Frontend time: 1.897s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.163s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11900647 42.59265658 + layer.0.v_cache 0.00001428 0.00663204 + layer.1.k_cache 0.12535267 8.58584829 + layer.1.v_cache 0.00000579 0.00274442 + layer.2.k_cache 0.00517510 1.05732701 + layer.2.v_cache 0.00001992 0.00819514 + layer.3.k_cache 0.09315677 4.73058195 + layer.3.v_cache 0.00001978 0.00928347 + layer.4.k_cache 0.00061896 0.20204296 + layer.4.v_cache 0.00005133 0.01717667 + layer.4.output 9.75369281 439.11298077 + ------------------------------------------------------------------------------------- + TOTAL 4.03642769 184.17666788 + (elements=1,018,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1018368 +Total Bytes 159760 +BPFP 1.2550 bits/point +EBPFP 2.5101 equivalent bits/point +MSE 184.176668 +---------------------- -------------------------------------------------------- +Time: 3.065s Load: 0.004s, Pack+Encode: 1.897s, Decode+Unpack: 1.163s +---------------------- -------------------------------------------------------- +💾 Converting with 184.1767 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,792B, BPFP=0.5643 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,144B, BPFP=2.7000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,564B, BPFP=1.2744 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,752B, BPFP=2.6417 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,040B, BPFP=1.3452 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,676B, BPFP=2.6304 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,252B, BPFP=1.3768 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,904B, BPFP=2.6643 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,964B, BPFP=2.3756 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,596B, BPFP=2.6185 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,120B, BPFP=0.3427 +⌛️ [2/4] FRONTEND: Frontend time: 2.563s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14131550 42.53794643 + layer.0.v_cache 0.00001383 0.00637865 + layer.1.k_cache 0.02978072 8.30812349 + layer.1.v_cache 0.00000537 0.00254319 + layer.2.k_cache 0.00578868 0.78150453 + layer.2.v_cache 0.00001872 0.00767802 + layer.3.k_cache 0.04020810 5.54033436 + layer.3.v_cache 0.00001850 0.00849501 + layer.4.k_cache 0.00063358 0.19953909 + layer.4.v_cache 0.00005294 0.01811994 + layer.4.output 10.86925024 468.93618197 + ------------------------------------------------------------------------------------- + TOTAL 4.48838751 196.46846685 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 151804 +BPFP 1.3288 bits/point +EBPFP 2.6576 equivalent bits/point +MSE 196.468467 +---------------------- -------------------------------------------------------- +Time: 3.880s Load: 0.005s, Pack+Encode: 2.563s, Decode+Unpack: 1.312s +---------------------- -------------------------------------------------------- +💾 Converting with 196.4685 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,204B, BPFP=0.5960 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,464B, BPFP=3.0625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,620B, BPFP=1.2314 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,956B, BPFP=2.9680 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,176B, BPFP=1.3348 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,048B, BPFP=2.6131 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,996B, BPFP=1.3013 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,664B, BPFP=2.5417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,564B, BPFP=2.1510 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,816B, BPFP=2.7560 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,772B, BPFP=0.3660 +⌛️ [2/4] FRONTEND: Frontend time: 2.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.485s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09514355 42.72011602 + layer.0.v_cache 0.00001423 0.00650679 + layer.1.k_cache 0.03283414 8.65275210 + layer.1.v_cache 0.00000553 0.00255256 + layer.2.k_cache 0.00489674 0.96288209 + layer.2.v_cache 0.00001883 0.00764801 + layer.3.k_cache 0.07199075 3.72843533 + layer.3.v_cache 0.00001932 0.00829123 + layer.4.k_cache 0.00061592 0.20216774 + layer.4.v_cache 0.00006816 0.01831194 + layer.4.output 0.17253765 637.54820366 + ------------------------------------------------------------------------------------- + TOTAL 0.08313946 265.83218173 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 124280 +BPFP 1.3599 bits/point +EBPFP 2.7197 equivalent bits/point +MSE 265.832182 +---------------------- -------------------------------------------------------- +Time: 3.621s Load: 0.004s, Pack+Encode: 2.132s, Decode+Unpack: 1.485s +---------------------- -------------------------------------------------------- +💾 Converting with 265.8322 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,256B, BPFP=0.5848 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,204B, BPFP=3.0898 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,884B, BPFP=1.2364 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,524B, BPFP=2.9677 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,104B, BPFP=1.2759 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,868B, BPFP=2.6703 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,000B, BPFP=1.2572 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,780B, BPFP=2.8341 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,808B, BPFP=2.3003 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,752B, BPFP=2.8290 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,740B, BPFP=0.3525 +⌛️ [2/4] FRONTEND: Frontend time: 1.932s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.237s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11660385 42.30655083 + layer.0.v_cache 0.00001381 0.00666455 + layer.1.k_cache 0.05175667 8.35016220 + layer.1.v_cache 0.00000569 0.00268797 + layer.2.k_cache 0.00227456 0.92213922 + layer.2.v_cache 0.00001743 0.00769243 + layer.3.k_cache 0.02781237 3.98891335 + layer.3.v_cache 0.00001883 0.00859515 + layer.4.k_cache 0.00061846 0.19958023 + layer.4.v_cache 0.00005344 0.01769804 + layer.4.output 0.17410791 578.56049877 + ------------------------------------------------------------------------------------- + TOTAL 0.08340767 241.51377502 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 130920 +BPFP 1.3831 bits/point +EBPFP 2.7662 equivalent bits/point +MSE 241.513775 +---------------------- -------------------------------------------------------- +Time: 3.173s Load: 0.004s, Pack+Encode: 1.932s, Decode+Unpack: 1.237s +---------------------- -------------------------------------------------------- +💾 Converting with 241.5138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,200B, BPFP=0.6024 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,228B, BPFP=2.8667 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,632B, BPFP=1.2485 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,604B, BPFP=2.7492 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,044B, BPFP=1.3261 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,660B, BPFP=2.5715 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,752B, BPFP=1.2711 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,660B, BPFP=2.5715 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,224B, BPFP=2.1130 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,016B, BPFP=2.8268 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,980B, BPFP=0.3760 +⌛️ [2/4] FRONTEND: Frontend time: 2.152s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212080 42.16750694 + layer.0.v_cache 0.00001360 0.00624343 + layer.1.k_cache 0.03288655 7.94033115 + layer.1.v_cache 0.00000536 0.00256022 + layer.2.k_cache 0.00973156 0.92289219 + layer.2.v_cache 0.00001762 0.00718135 + layer.3.k_cache 0.02906220 4.03276779 + layer.3.v_cache 0.00001891 0.00817230 + layer.4.k_cache 0.00063934 0.18919242 + layer.4.v_cache 0.00004866 0.01652948 + layer.4.output 0.18425958 638.79179217 + ------------------------------------------------------------------------------------- + TOTAL 0.08496245 266.28446603 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 121000 +BPFP 1.3399 bits/point +EBPFP 2.6798 equivalent bits/point +MSE 266.284466 +---------------------- -------------------------------------------------------- +Time: 3.364s Load: 0.004s, Pack+Encode: 2.152s, Decode+Unpack: 1.208s +---------------------- -------------------------------------------------------- +💾 Converting with 266.2845 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,200B, BPFP=0.5882 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,824B, BPFP=2.9088 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,752B, BPFP=1.2412 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,916B, BPFP=2.7419 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,012B, BPFP=1.2890 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,544B, BPFP=2.6735 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,092B, BPFP=1.3037 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,596B, BPFP=2.6831 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,440B, BPFP=2.1029 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,180B, BPFP=2.6066 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,856B, BPFP=0.3901 +⌛️ [2/4] FRONTEND: Frontend time: 1.805s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.409s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10758395 43.56062730 + layer.0.v_cache 0.00001397 0.00677634 + layer.1.k_cache 0.03293324 7.50642233 + layer.1.v_cache 0.00000575 0.00268752 + layer.2.k_cache 0.00823421 1.20090790 + layer.2.v_cache 0.00001745 0.00782764 + layer.3.k_cache 0.04116619 3.72178022 + layer.3.v_cache 0.00001985 0.00908151 + layer.4.k_cache 0.00062779 0.21030715 + layer.4.v_cache 0.00005042 0.01843556 + layer.4.output 0.16954060 581.71045168 + ------------------------------------------------------------------------------------- + TOTAL 0.08102571 242.83635384 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 124412 +BPFP 1.3453 bits/point +EBPFP 2.6906 equivalent bits/point +MSE 242.836354 +---------------------- -------------------------------------------------------- +Time: 3.218s Load: 0.003s, Pack+Encode: 1.805s, Decode+Unpack: 1.409s +---------------------- -------------------------------------------------------- +💾 Converting with 242.8364 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,332B, BPFP=0.5785 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,524B, BPFP=3.0424 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,120B, BPFP=1.2361 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,252B, BPFP=2.8215 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,492B, BPFP=1.3007 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,124B, BPFP=2.7993 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,300B, BPFP=1.2674 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,528B, BPFP=2.8694 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,856B, BPFP=2.4056 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,496B, BPFP=2.8639 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,100B, BPFP=0.3745 +⌛️ [2/4] FRONTEND: Frontend time: 2.310s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.246s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12307767 39.95891656 + layer.0.v_cache 0.00001410 0.00690451 + layer.1.k_cache 0.03088827 8.43660753 + layer.1.v_cache 0.00000584 0.00275244 + layer.2.k_cache 0.00371391 0.95680381 + layer.2.v_cache 0.00001950 0.00794933 + layer.3.k_cache 0.03079298 4.21366747 + layer.3.v_cache 0.00001950 0.00902125 + layer.4.k_cache 0.00063014 0.19703880 + layer.4.v_cache 0.00005282 0.01772734 + layer.4.output 0.15767360 566.23859127 + ------------------------------------------------------------------------------------- + TOTAL 0.07605470 236.32220752 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 137124 +BPFP 1.4004 bits/point +EBPFP 2.8007 equivalent bits/point +MSE 236.322208 +---------------------- -------------------------------------------------------- +Time: 3.562s Load: 0.005s, Pack+Encode: 2.310s, Decode+Unpack: 1.246s +---------------------- -------------------------------------------------------- +💾 Converting with 236.3222 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,156B, BPFP=0.6014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,788B, BPFP=2.8178 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,476B, BPFP=1.2340 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,100B, BPFP=2.6867 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,856B, BPFP=1.3064 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,896B, BPFP=2.6479 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,716B, BPFP=1.2797 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,192B, BPFP=2.7043 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,048B, BPFP=2.1052 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,812B, BPFP=2.6319 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,240B, BPFP=0.3876 +⌛️ [2/4] FRONTEND: Frontend time: 1.801s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.258s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10295349 42.66110304 + layer.0.v_cache 0.00001348 0.00647346 + layer.1.k_cache 0.03334363 8.32265565 + layer.1.v_cache 0.00000550 0.00269080 + layer.2.k_cache 0.01126997 0.90522161 + layer.2.v_cache 0.00001933 0.00798108 + layer.3.k_cache 0.09131574 3.34435068 + layer.3.v_cache 0.00001769 0.00863091 + layer.4.k_cache 0.00062275 0.20644607 + layer.4.v_cache 0.00006541 0.01655504 + layer.4.output 0.17512874 616.59505662 + ------------------------------------------------------------------------------------- + TOTAL 0.08620754 257.15573557 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 119280 +BPFP 1.3370 bits/point +EBPFP 2.6740 equivalent bits/point +MSE 257.155736 +---------------------- -------------------------------------------------------- +Time: 3.063s Load: 0.004s, Pack+Encode: 1.801s, Decode+Unpack: 1.258s +---------------------- -------------------------------------------------------- +💾 Converting with 257.1557 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,260B, BPFP=0.5855 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,728B, BPFP=3.0043 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,908B, BPFP=1.2407 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,716B, BPFP=2.8226 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,224B, BPFP=1.2974 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,432B, BPFP=2.7716 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,076B, BPFP=1.2708 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,628B, BPFP=2.6272 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,952B, BPFP=2.1466 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,772B, BPFP=2.8326 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,792B, BPFP=0.3539 +⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.337s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12314748 41.02591258 + layer.0.v_cache 0.00001636 0.00655600 + layer.1.k_cache 0.01190477 7.54336022 + layer.1.v_cache 0.00000588 0.00265671 + layer.2.k_cache 0.01614088 0.92125553 + layer.2.v_cache 0.00001834 0.00752569 + layer.3.k_cache 0.05604688 4.04145515 + layer.3.v_cache 0.00001882 0.00876838 + layer.4.k_cache 0.00063774 0.19459214 + layer.4.v_cache 0.00005162 0.01671453 + layer.4.output 0.17348555 600.98460591 + ------------------------------------------------------------------------------------- + TOTAL 0.08366986 250.62711990 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 128488 +BPFP 1.3574 bits/point +EBPFP 2.7148 equivalent bits/point +MSE 250.627120 +---------------------- -------------------------------------------------------- +Time: 3.075s Load: 0.004s, Pack+Encode: 1.735s, Decode+Unpack: 1.337s +---------------------- -------------------------------------------------------- +💾 Converting with 250.6271 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,140B, BPFP=0.6057 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,512B, BPFP=2.9923 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,580B, BPFP=1.2693 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,064B, BPFP=2.9059 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,860B, BPFP=1.3233 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,568B, BPFP=2.8102 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,688B, BPFP=1.2901 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,036B, BPFP=3.0934 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,492B, BPFP=2.2168 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,388B, BPFP=2.7755 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,856B, BPFP=0.4094 +⌛️ [2/4] FRONTEND: Frontend time: 257.094s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.526s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11560643 41.93698761 + layer.0.v_cache 0.00001402 0.00696562 + layer.1.k_cache 0.03272506 8.49626555 + layer.1.v_cache 0.00000563 0.00284712 + layer.2.k_cache 0.00645547 0.81623501 + layer.2.v_cache 0.00001830 0.00843396 + layer.3.k_cache 0.02731901 3.74288300 + layer.3.v_cache 0.00001960 0.00935726 + layer.4.k_cache 0.00061302 0.20458881 + layer.4.v_cache 0.00004978 0.01661212 + layer.4.output 0.17251868 639.29348545 + ------------------------------------------------------------------------------------- + TOTAL 0.08179159 266.48797495 + (elements=705,024) +---------------------- ---------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- ---------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- ---------------------------------------------------------- +Total Elements 705024 +Total Bytes 125184 +BPFP 1.4205 bits/point +EBPFP 2.8410 equivalent bits/point +MSE 266.487975 +---------------------- ---------------------------------------------------------- +Time: 258.625s Load: 0.004s, Pack+Encode: 257.094s, Decode+Unpack: 1.526s +---------------------- ---------------------------------------------------------- +💾 Converting with 266.4880 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,260B, BPFP=0.5788 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,080B, BPFP=3.0327 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,916B, BPFP=1.2280 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,368B, BPFP=2.9062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,292B, BPFP=1.2947 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,656B, BPFP=2.7798 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,140B, BPFP=1.2678 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,016B, BPFP=2.8438 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,240B, BPFP=2.1733 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,596B, BPFP=2.7692 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,572B, BPFP=0.3950 +⌛️ [2/4] FRONTEND: Frontend time: 2.140s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11690623 41.41566329 + layer.0.v_cache 0.00001418 0.00623107 + layer.1.k_cache 0.03305150 7.96323187 + layer.1.v_cache 0.00000545 0.00264806 + layer.2.k_cache 0.01095151 0.85494145 + layer.2.v_cache 0.00001854 0.00754545 + layer.3.k_cache 0.04293454 3.49769072 + layer.3.v_cache 0.00002007 0.00818848 + layer.4.k_cache 0.00062489 0.18744959 + layer.4.v_cache 0.00005451 0.01734597 + layer.4.output 0.16672201 586.98173701 + ------------------------------------------------------------------------------------- + TOTAL 0.08068444 244.87253500 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 133136 +BPFP 1.3905 bits/point +EBPFP 2.7811 equivalent bits/point +MSE 244.872535 +---------------------- -------------------------------------------------------- +Time: 3.513s Load: 0.006s, Pack+Encode: 2.140s, Decode+Unpack: 1.367s +---------------------- -------------------------------------------------------- +💾 Converting with 244.8725 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,304B, BPFP=0.5673 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,836B, BPFP=3.0625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,128B, BPFP=1.2239 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,812B, BPFP=2.8867 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,456B, BPFP=1.2802 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,896B, BPFP=2.7294 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,248B, BPFP=1.2445 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,488B, BPFP=2.8310 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,984B, BPFP=2.2294 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,760B, BPFP=2.7060 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,324B, BPFP=0.3514 +⌛️ [2/4] FRONTEND: Frontend time: 1.771s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11067300 39.42471347 + layer.0.v_cache 0.00001409 0.00642640 + layer.1.k_cache 0.05152405 8.00697410 + layer.1.v_cache 0.00000569 0.00270793 + layer.2.k_cache 0.01362713 1.06105436 + layer.2.v_cache 0.00001889 0.00785011 + layer.3.k_cache 0.08227781 3.45182784 + layer.3.v_cache 0.00001889 0.00856649 + layer.4.k_cache 0.00062414 0.19095356 + layer.4.v_cache 0.00005257 0.01667985 + layer.4.output 0.15746453 576.70830063 + ------------------------------------------------------------------------------------- + TOTAL 0.08006400 240.53740344 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 135236 +BPFP 1.3659 bits/point +EBPFP 2.7318 equivalent bits/point +MSE 240.537403 +---------------------- -------------------------------------------------------- +Time: 3.146s Load: 0.005s, Pack+Encode: 1.771s, Decode+Unpack: 1.371s +---------------------- -------------------------------------------------------- +💾 Converting with 240.5374 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 116, 128) +Output shape: (1, 116, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.output: torch.Size([1, 116, 3584]) -> torch.Size([1, 1, 116, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,048B, BPFP=0.5453 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,400B, BPFP=2.4784 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,008B, BPFP=1.2134 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,920B, BPFP=2.4138 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,208B, BPFP=1.3750 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,648B, BPFP=2.3772 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,764B, BPFP=1.4499 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,816B, BPFP=2.3998 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,972B, BPFP=2.1514 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,588B, BPFP=2.3691 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,368B, BPFP=0.3342 +⌛️ [2/4] FRONTEND: Frontend time: 1.761s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.238s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16314862 42.17700616 + layer.0.v_cache 0.00001373 0.00623202 + layer.1.k_cache 0.05481785 8.42046067 + layer.1.v_cache 0.00000557 0.00264176 + layer.2.k_cache 0.00968372 1.04883365 + layer.2.v_cache 0.00001954 0.00833497 + layer.3.k_cache 0.02418142 4.62925562 + layer.3.v_cache 0.00001860 0.00849049 + layer.4.k_cache 0.00062333 0.20169732 + layer.4.v_cache 0.00005467 0.01889138 + layer.4.output 9.83417319 449.51089132 + ------------------------------------------------------------------------------------- + TOTAL 4.06422232 188.41753431 + (elements=1,009,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1009664 +Total Bytes 156740 +BPFP 1.2419 bits/point +EBPFP 2.4838 equivalent bits/point +MSE 188.417534 +---------------------- -------------------------------------------------------- +Time: 3.003s Load: 0.004s, Pack+Encode: 1.761s, Decode+Unpack: 1.238s +---------------------- -------------------------------------------------------- +💾 Converting with 188.4175 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,072B, BPFP=0.6154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,296B, BPFP=2.8638 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,240B, BPFP=1.2500 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,920B, BPFP=2.7885 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,428B, BPFP=1.2877 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,728B, BPFP=2.5497 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,404B, BPFP=1.2829 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,428B, BPFP=2.6899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,416B, BPFP=2.0865 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,212B, BPFP=2.6466 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,476B, BPFP=0.4143 +⌛️ [2/4] FRONTEND: Frontend time: 1.870s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.264s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11749315 40.60647035 + layer.0.v_cache 0.00001389 0.00657322 + layer.1.k_cache 0.03399578 8.46865062 + layer.1.v_cache 0.00000555 0.00272313 + layer.2.k_cache 0.00854440 1.10078127 + layer.2.v_cache 0.00001777 0.00731612 + layer.3.k_cache 0.02740619 3.78328764 + layer.3.v_cache 0.00001935 0.00865920 + layer.4.k_cache 0.00063531 0.19231079 + layer.4.v_cache 0.00005166 0.01826759 + layer.4.output 0.18490290 669.80980998 + ------------------------------------------------------------------------------------- + TOTAL 0.08720608 278.99198293 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 114620 +BPFP 1.3506 bits/point +EBPFP 2.7013 equivalent bits/point +MSE 278.991983 +---------------------- -------------------------------------------------------- +Time: 3.138s Load: 0.004s, Pack+Encode: 1.870s, Decode+Unpack: 1.264s +---------------------- -------------------------------------------------------- +💾 Converting with 278.9920 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,292B, BPFP=0.5472 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,632B, BPFP=2.9309 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,272B, BPFP=1.2088 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,264B, BPFP=2.8697 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,796B, BPFP=1.2959 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,708B, BPFP=2.7773 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,644B, BPFP=1.2706 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,236B, BPFP=2.8650 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,140B, BPFP=2.5166 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,112B, BPFP=2.8444 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,116B, BPFP=0.3589 +⌛️ [2/4] FRONTEND: Frontend time: 1.932s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.537s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11492289 41.20216142 + layer.0.v_cache 0.00001440 0.00652237 + layer.1.k_cache 0.01112092 7.78856253 + layer.1.v_cache 0.00000555 0.00249569 + layer.2.k_cache 0.01096075 0.92322053 + layer.2.v_cache 0.00001882 0.00750188 + layer.3.k_cache 0.03781782 3.89591623 + layer.3.v_cache 0.00001905 0.00832477 + layer.4.k_cache 0.00062716 0.19426232 + layer.4.v_cache 0.00006343 0.01646633 + layer.4.output 0.15224552 548.39988602 + ------------------------------------------------------------------------------------- + TOTAL 0.07301703 228.99086095 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 142212 +BPFP 1.3905 bits/point +EBPFP 2.7811 equivalent bits/point +MSE 228.990861 +---------------------- -------------------------------------------------------- +Time: 3.474s Load: 0.004s, Pack+Encode: 1.932s, Decode+Unpack: 1.537s +---------------------- -------------------------------------------------------- +💾 Converting with 228.9909 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,060B, BPFP=0.6209 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,900B, BPFP=2.8206 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,196B, BPFP=1.2573 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,592B, BPFP=2.7581 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,664B, BPFP=1.3523 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,232B, BPFP=2.6851 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,516B, BPFP=1.3222 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,468B, BPFP=2.5300 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,980B, BPFP=2.2281 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,072B, BPFP=2.4497 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,148B, BPFP=0.4101 +⌛️ [2/4] FRONTEND: Frontend time: 2.080s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.268s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10405910 41.65408127 + layer.0.v_cache 0.00001453 0.00648531 + layer.1.k_cache 0.03397077 8.51709460 + layer.1.v_cache 0.00000547 0.00272402 + layer.2.k_cache 0.00568739 0.88399387 + layer.2.v_cache 0.00001916 0.00829272 + layer.3.k_cache 0.09581309 3.89003011 + layer.3.v_cache 0.00001862 0.00927291 + layer.4.k_cache 0.00061896 0.19407124 + layer.4.v_cache 0.00005653 0.01738904 + layer.4.output 0.19123089 668.65810529 + ------------------------------------------------------------------------------------- + TOTAL 0.09287528 278.57589248 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 112828 +BPFP 1.3468 bits/point +EBPFP 2.6936 equivalent bits/point +MSE 278.575892 +---------------------- -------------------------------------------------------- +Time: 3.353s Load: 0.004s, Pack+Encode: 2.080s, Decode+Unpack: 1.268s +---------------------- -------------------------------------------------------- +💾 Converting with 278.5759 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,216B, BPFP=0.5912 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,592B, BPFP=2.8662 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,824B, BPFP=1.2544 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,276B, BPFP=2.8081 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,156B, BPFP=1.3154 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,568B, BPFP=2.6779 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,016B, BPFP=1.2897 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,204B, BPFP=2.7949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,652B, BPFP=2.1419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,360B, BPFP=2.6397 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,288B, BPFP=0.4015 +⌛️ [2/4] FRONTEND: Frontend time: 1.912s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.220s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09384689 41.77412684 + layer.0.v_cache 0.00001364 0.00651523 + layer.1.k_cache 0.01313462 8.40490220 + layer.1.v_cache 0.00000592 0.00282531 + layer.2.k_cache 0.00813576 0.84395770 + layer.2.v_cache 0.00001762 0.00802969 + layer.3.k_cache 0.02628646 3.41638471 + layer.3.v_cache 0.00002044 0.00919502 + layer.4.k_cache 0.00061596 0.20188453 + layer.4.v_cache 0.00005334 0.01859855 + layer.4.output 0.17343949 587.87410714 + ------------------------------------------------------------------------------------- + TOTAL 0.07977689 245.28265705 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 126152 +BPFP 1.3641 bits/point +EBPFP 2.7282 equivalent bits/point +MSE 245.282657 +---------------------- -------------------------------------------------------- +Time: 3.137s Load: 0.004s, Pack+Encode: 1.912s, Decode+Unpack: 1.220s +---------------------- -------------------------------------------------------- +💾 Converting with 245.2827 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 126, 128) +Output shape: (1, 126, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.output: torch.Size([1, 126, 3584]) -> torch.Size([1, 1, 126, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,200B, BPFP=0.5208 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,520B, BPFP=2.2966 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,960B, BPFP=1.3591 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,236B, BPFP=2.2614 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,656B, BPFP=1.5694 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,760B, BPFP=2.2024 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,120B, BPFP=1.3790 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,040B, BPFP=2.2371 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,064B, BPFP=1.9921 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,596B, BPFP=2.1820 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,248B, BPFP=0.3410 +⌛️ [2/4] FRONTEND: Frontend time: 1.736s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.119s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13141363 41.38978407 + layer.0.v_cache 0.00001429 0.00611524 + layer.1.k_cache 0.06496696 10.28062415 + layer.1.v_cache 0.00000583 0.00268965 + layer.2.k_cache 0.01057190 0.93209718 + layer.2.v_cache 0.00002020 0.00825916 + layer.3.k_cache 0.05250616 5.21192569 + layer.3.v_cache 0.00001853 0.00856315 + layer.4.k_cache 0.00062861 0.24816967 + layer.4.v_cache 0.00004930 0.01688192 + layer.4.output 9.05470577 372.19834184 + ------------------------------------------------------------------------------------- + TOTAL 3.74371387 156.67608840 + (elements=1,096,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1096704 +Total Bytes 164400 +BPFP 1.1992 bits/point +EBPFP 2.3985 equivalent bits/point +MSE 156.676088 +---------------------- -------------------------------------------------------- +Time: 2.862s Load: 0.006s, Pack+Encode: 1.736s, Decode+Unpack: 1.119s +---------------------- -------------------------------------------------------- +💾 Converting with 156.6761 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,252B, BPFP=0.5841 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,144B, BPFP=2.8994 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,800B, BPFP=1.2213 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,544B, BPFP=2.7917 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,140B, BPFP=1.2823 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,488B, BPFP=2.7816 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,988B, BPFP=1.2550 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,244B, BPFP=2.7378 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,544B, BPFP=2.2529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,864B, BPFP=2.8491 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,636B, BPFP=0.3499 +⌛️ [2/4] FRONTEND: Frontend time: 2.191s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.571s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13059476 41.08084422 + layer.0.v_cache 0.00001380 0.00642508 + layer.1.k_cache 0.03034586 7.97979526 + layer.1.v_cache 0.00000531 0.00270377 + layer.2.k_cache 0.00374181 1.02372917 + layer.2.v_cache 0.00001886 0.00822234 + layer.3.k_cache 0.04454902 3.60314626 + layer.3.v_cache 0.00001968 0.00935624 + layer.4.k_cache 0.00060938 0.21095478 + layer.4.v_cache 0.00007316 0.01905057 + layer.4.output 0.16485390 587.13639163 + ------------------------------------------------------------------------------------- + TOTAL 0.08023229 244.93523347 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 128644 +BPFP 1.3591 bits/point +EBPFP 2.7181 equivalent bits/point +MSE 244.935233 +---------------------- -------------------------------------------------------- +Time: 3.768s Load: 0.006s, Pack+Encode: 2.191s, Decode+Unpack: 1.571s +---------------------- -------------------------------------------------------- +💾 Converting with 244.9352 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,356B, BPFP=0.5700 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,748B, BPFP=3.0143 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,100B, BPFP=1.2058 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,548B, BPFP=2.9803 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,488B, BPFP=1.2717 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,244B, BPFP=2.9287 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,428B, BPFP=1.2615 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,692B, BPFP=3.0048 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,936B, BPFP=2.5367 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,280B, BPFP=2.9348 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,796B, BPFP=0.4075 +⌛️ [2/4] FRONTEND: Frontend time: 2.059s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.240s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15436328 41.13433573 + layer.0.v_cache 0.00001586 0.00643493 + layer.1.k_cache 0.01080767 9.08353855 + layer.1.v_cache 0.00000575 0.00262105 + layer.2.k_cache 0.00786338 0.84161443 + layer.2.v_cache 0.00001961 0.00794810 + layer.3.k_cache 0.05652578 3.07862555 + layer.3.v_cache 0.00002042 0.00864183 + layer.4.k_cache 0.00062828 0.19589974 + layer.4.v_cache 0.00005674 0.01842313 + layer.4.output 0.15192759 558.43585016 + ------------------------------------------------------------------------------------- + TOTAL 0.07610588 233.14288436 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 144616 +BPFP 1.4448 bits/point +EBPFP 2.8895 equivalent bits/point +MSE 233.142884 +---------------------- -------------------------------------------------------- +Time: 3.303s Load: 0.005s, Pack+Encode: 2.059s, Decode+Unpack: 1.240s +---------------------- -------------------------------------------------------- +💾 Converting with 233.1429 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,056B, BPFP=0.6122 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,540B, BPFP=2.9127 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,264B, BPFP=1.2548 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,192B, BPFP=2.8429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,840B, BPFP=1.3702 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,640B, BPFP=2.5321 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,308B, BPFP=1.2636 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,688B, BPFP=2.5417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,068B, BPFP=2.2171 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,760B, BPFP=2.7564 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,428B, BPFP=0.3843 +⌛️ [2/4] FRONTEND: Frontend time: 1.773s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10254851 41.42796287 + layer.0.v_cache 0.00001399 0.00694637 + layer.1.k_cache 0.05476840 8.77116511 + layer.1.v_cache 0.00000542 0.00282431 + layer.2.k_cache 0.00244555 0.83264434 + layer.2.v_cache 0.00001844 0.00849429 + layer.3.k_cache 0.02811479 3.68465678 + layer.3.v_cache 0.00001905 0.00974590 + layer.4.k_cache 0.00060111 0.19436628 + layer.4.v_cache 0.00005313 0.01778351 + layer.4.output 0.19635826 664.07474817 + ------------------------------------------------------------------------------------- + TOTAL 0.09194683 276.67528394 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 114784 +BPFP 1.3526 bits/point +EBPFP 2.7051 equivalent bits/point +MSE 276.675284 +---------------------- -------------------------------------------------------- +Time: 3.091s Load: 0.003s, Pack+Encode: 1.773s, Decode+Unpack: 1.315s +---------------------- -------------------------------------------------------- +💾 Converting with 276.6753 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,164B, BPFP=0.6029 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,044B, BPFP=3.0572 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,520B, BPFP=1.2424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,096B, BPFP=2.8765 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,820B, BPFP=1.2995 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,880B, BPFP=2.8354 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,632B, BPFP=1.2637 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,168B, BPFP=2.6997 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,128B, BPFP=2.3110 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,936B, BPFP=2.8460 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,344B, BPFP=0.3905 +⌛️ [2/4] FRONTEND: Frontend time: 1.919s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.255s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11987135 45.09508980 + layer.0.v_cache 0.00001390 0.00654215 + layer.1.k_cache 0.03139273 8.92974556 + layer.1.v_cache 0.00000562 0.00264388 + layer.2.k_cache 0.00816743 1.12035723 + layer.2.v_cache 0.00001857 0.00765665 + layer.3.k_cache 0.02640573 3.58803651 + layer.3.v_cache 0.00001968 0.00858814 + layer.4.k_cache 0.00062091 0.19334344 + layer.4.v_cache 0.00004972 0.01707172 + layer.4.output 0.16983549 626.17356272 + ------------------------------------------------------------------------------------- + TOTAL 0.08090671 261.30494201 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 124732 +BPFP 1.3981 bits/point +EBPFP 2.7962 equivalent bits/point +MSE 261.304942 +---------------------- -------------------------------------------------------- +Time: 3.178s Load: 0.005s, Pack+Encode: 1.919s, Decode+Unpack: 1.255s +---------------------- -------------------------------------------------------- +💾 Converting with 261.3049 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,252B, BPFP=0.5841 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,156B, BPFP=3.0812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,960B, BPFP=1.2500 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,412B, BPFP=2.9476 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,220B, BPFP=1.2967 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,576B, BPFP=2.7974 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,148B, BPFP=1.2838 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,780B, BPFP=2.8341 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,548B, BPFP=2.2536 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,768B, BPFP=2.8319 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,208B, BPFP=0.3645 +⌛️ [2/4] FRONTEND: Frontend time: 1.764s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.244s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14089697 42.44612630 + layer.0.v_cache 0.00001393 0.00636594 + layer.1.k_cache 0.01618495 7.77224907 + layer.1.v_cache 0.00000582 0.00270685 + layer.2.k_cache 0.00661465 1.13435364 + layer.2.v_cache 0.00002083 0.00774827 + layer.3.k_cache 0.07087009 3.54994219 + layer.3.v_cache 0.00001822 0.00859789 + layer.4.k_cache 0.00061891 0.20438161 + layer.4.v_cache 0.00005311 0.01807638 + layer.4.output 0.16425455 599.02729885 + ------------------------------------------------------------------------------------- + TOTAL 0.08147525 249.90244942 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 132028 +BPFP 1.3948 bits/point +EBPFP 2.7896 equivalent bits/point +MSE 249.902449 +---------------------- -------------------------------------------------------- +Time: 3.012s Load: 0.004s, Pack+Encode: 1.764s, Decode+Unpack: 1.244s +---------------------- -------------------------------------------------------- +💾 Converting with 249.9024 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,300B, BPFP=0.5794 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,088B, BPFP=3.0000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,972B, BPFP=1.2240 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,496B, BPFP=2.8961 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,396B, BPFP=1.2985 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,256B, BPFP=2.8539 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,344B, BPFP=1.2893 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,024B, BPFP=2.8132 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,472B, BPFP=2.3652 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,380B, BPFP=2.8757 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,400B, BPFP=0.3862 +⌛️ [2/4] FRONTEND: Frontend time: 1.702s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.498s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12786581 42.71160683 + layer.0.v_cache 0.00001426 0.00689317 + layer.1.k_cache 0.03183105 8.26961972 + layer.1.v_cache 0.00000593 0.00291492 + layer.2.k_cache 0.00241907 0.83453026 + layer.2.v_cache 0.00001785 0.00799706 + layer.3.k_cache 0.02946699 3.55193475 + layer.3.v_cache 0.00001957 0.01001718 + layer.4.k_cache 0.00063747 0.20838384 + layer.4.v_cache 0.00005003 0.01786212 + layer.4.output 0.16568538 575.97802970 + ------------------------------------------------------------------------------------- + TOTAL 0.07953681 240.43929222 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 136128 +BPFP 1.4058 bits/point +EBPFP 2.8116 equivalent bits/point +MSE 240.439292 +---------------------- -------------------------------------------------------- +Time: 3.204s Load: 0.003s, Pack+Encode: 1.702s, Decode+Unpack: 1.498s +---------------------- -------------------------------------------------------- +💾 Converting with 240.4393 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,076B, BPFP=0.6162 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,820B, BPFP=2.9688 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,252B, BPFP=1.2524 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,032B, BPFP=2.8109 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,696B, BPFP=1.3413 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,180B, BPFP=2.8405 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,568B, BPFP=1.3157 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,584B, BPFP=2.7212 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,756B, BPFP=2.1546 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,296B, BPFP=2.6635 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,280B, BPFP=0.4373 +⌛️ [2/4] FRONTEND: Frontend time: 1.929s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.274s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12225728 40.24203100 + layer.0.v_cache 0.00001337 0.00689889 + layer.1.k_cache 0.03480693 8.37472847 + layer.1.v_cache 0.00000545 0.00282484 + layer.2.k_cache 0.00853881 1.03179687 + layer.2.v_cache 0.00001924 0.00855917 + layer.3.k_cache 0.06130744 3.74580579 + layer.3.v_cache 0.00001928 0.00963660 + layer.4.k_cache 0.00058936 0.19837880 + layer.4.v_cache 0.00007197 0.01855192 + layer.4.output 0.18657821 653.20879121 + ------------------------------------------------------------------------------------- + TOTAL 0.09021627 272.12357358 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 118540 +BPFP 1.3968 bits/point +EBPFP 2.7936 equivalent bits/point +MSE 272.123574 +---------------------- -------------------------------------------------------- +Time: 3.207s Load: 0.003s, Pack+Encode: 1.929s, Decode+Unpack: 1.274s +---------------------- -------------------------------------------------------- +💾 Converting with 272.1236 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,040B, BPFP=0.6169 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,792B, BPFP=2.7987 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,212B, BPFP=1.2606 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,072B, BPFP=2.8555 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,432B, BPFP=1.3052 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,556B, BPFP=2.7508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,284B, BPFP=1.2752 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,596B, BPFP=2.5560 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,316B, BPFP=2.2963 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,688B, BPFP=2.5747 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,976B, BPFP=0.4051 +⌛️ [2/4] FRONTEND: Frontend time: 1.800s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.281s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582910 39.83087079 + layer.0.v_cache 0.00001386 0.00679523 + layer.1.k_cache 0.03408930 8.71702437 + layer.1.v_cache 0.00000538 0.00279640 + layer.2.k_cache 0.00228146 0.91301143 + layer.2.v_cache 0.00001859 0.00851285 + layer.3.k_cache 0.06276476 3.68193669 + layer.3.v_cache 0.00001913 0.00992781 + layer.4.k_cache 0.00060520 0.21143571 + layer.4.v_cache 0.00005125 0.01971247 + layer.4.output 0.19356344 680.07699443 + ------------------------------------------------------------------------------------- + TOTAL 0.09238954 283.17299911 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 113964 +BPFP 1.3603 bits/point +EBPFP 2.7207 equivalent bits/point +MSE 283.172999 +---------------------- -------------------------------------------------------- +Time: 3.085s Load: 0.004s, Pack+Encode: 1.800s, Decode+Unpack: 1.281s +---------------------- -------------------------------------------------------- +💾 Converting with 283.1730 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,204B, BPFP=0.6032 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,704B, BPFP=2.9563 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,580B, BPFP=1.2387 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,484B, BPFP=2.7267 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,968B, BPFP=1.3117 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,856B, BPFP=2.7967 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,824B, BPFP=1.2846 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,472B, BPFP=2.7244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,388B, BPFP=2.1438 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,428B, BPFP=2.9044 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,828B, BPFP=0.3988 +⌛️ [2/4] FRONTEND: Frontend time: 1.733s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09700135 39.81043510 + layer.0.v_cache 0.00001355 0.00638883 + layer.1.k_cache 0.01103399 8.50259877 + layer.1.v_cache 0.00000598 0.00268810 + layer.2.k_cache 0.00794396 1.12197104 + layer.2.v_cache 0.00001935 0.00850117 + layer.3.k_cache 0.04207947 3.94132389 + layer.3.v_cache 0.00002010 0.00942038 + layer.4.k_cache 0.00062246 0.19716614 + layer.4.v_cache 0.00005226 0.01729111 + layer.4.output 0.17535226 643.66577022 + ------------------------------------------------------------------------------------- + TOTAL 0.08154460 268.19283389 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 124736 +BPFP 1.3813 bits/point +EBPFP 2.7626 equivalent bits/point +MSE 268.192834 +---------------------- -------------------------------------------------------- +Time: 3.094s Load: 0.003s, Pack+Encode: 1.733s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 268.1928 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,208B, BPFP=0.5897 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,128B, BPFP=2.9647 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,728B, BPFP=1.2368 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,164B, BPFP=2.6037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,172B, BPFP=1.3184 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,372B, BPFP=2.6419 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,100B, BPFP=1.3051 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,900B, BPFP=2.7390 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,528B, BPFP=2.1191 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,040B, BPFP=2.7647 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,828B, BPFP=0.3894 +⌛️ [2/4] FRONTEND: Frontend time: 1.986s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.256s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11944316 43.81154642 + layer.0.v_cache 0.00001525 0.00642920 + layer.1.k_cache 0.01376067 8.01428366 + layer.1.v_cache 0.00000561 0.00265972 + layer.2.k_cache 0.00522648 0.94319655 + layer.2.v_cache 0.00001966 0.00781456 + layer.3.k_cache 0.04600189 3.98602905 + layer.3.v_cache 0.00001912 0.00892286 + layer.4.k_cache 0.00062079 0.20772505 + layer.4.v_cache 0.00005114 0.01697081 + layer.4.output 0.17175110 603.34075630 + ------------------------------------------------------------------------------------- + TOTAL 0.08161303 251.78769835 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 125168 +BPFP 1.3535 bits/point +EBPFP 2.7069 equivalent bits/point +MSE 251.787698 +---------------------- -------------------------------------------------------- +Time: 3.245s Load: 0.004s, Pack+Encode: 1.986s, Decode+Unpack: 1.256s +---------------------- -------------------------------------------------------- +💾 Converting with 251.7877 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.6170 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,840B, BPFP=2.7724 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,216B, BPFP=1.2452 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,224B, BPFP=2.8494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,512B, BPFP=1.3045 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,568B, BPFP=2.5176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,380B, BPFP=1.2780 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,752B, BPFP=2.5545 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,692B, BPFP=2.1418 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,736B, BPFP=2.5513 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,756B, BPFP=0.4223 +⌛️ [2/4] FRONTEND: Frontend time: 1.787s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120220 40.79291867 + layer.0.v_cache 0.00001368 0.00681002 + layer.1.k_cache 0.03273370 7.88632280 + layer.1.v_cache 0.00000573 0.00265329 + layer.2.k_cache 0.00892877 1.11569566 + layer.2.v_cache 0.00001953 0.00787243 + layer.3.k_cache 0.02974504 4.11945245 + layer.3.v_cache 0.00001879 0.00913724 + layer.4.k_cache 0.00063874 0.20728459 + layer.4.v_cache 0.00005200 0.01766904 + layer.4.output 0.17853031 658.93893086 + ------------------------------------------------------------------------------------- + TOTAL 0.08429825 274.51401954 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 113756 +BPFP 1.3405 bits/point +EBPFP 2.6809 equivalent bits/point +MSE 274.514020 +---------------------- -------------------------------------------------------- +Time: 3.110s Load: 0.003s, Pack+Encode: 1.787s, Decode+Unpack: 1.319s +---------------------- -------------------------------------------------------- +💾 Converting with 274.5140 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,200B, BPFP=0.6024 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,212B, BPFP=3.0520 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,820B, BPFP=1.2839 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,304B, BPFP=2.8810 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,048B, BPFP=1.3268 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,268B, BPFP=2.6860 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,112B, BPFP=1.3389 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,536B, BPFP=2.9247 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,460B, BPFP=2.1574 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,380B, BPFP=2.8953 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,268B, BPFP=0.4106 +⌛️ [2/4] FRONTEND: Frontend time: 1.748s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.349s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09802409 42.81890060 + layer.0.v_cache 0.00001402 0.00680554 + layer.1.k_cache 0.03453328 8.41859969 + layer.1.v_cache 0.00000594 0.00284379 + layer.2.k_cache 0.01027090 1.03724321 + layer.2.v_cache 0.00001933 0.00793574 + layer.3.k_cache 0.02613793 4.00606895 + layer.3.v_cache 0.00001931 0.00892764 + layer.4.k_cache 0.00063818 0.20635191 + layer.4.v_cache 0.00005393 0.01710178 + layer.4.output 0.18084440 634.42383821 + ------------------------------------------------------------------------------------- + TOTAL 0.08444869 264.55868508 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 127608 +BPFP 1.4131 bits/point +EBPFP 2.8262 equivalent bits/point +MSE 264.558685 +---------------------- -------------------------------------------------------- +Time: 3.101s Load: 0.004s, Pack+Encode: 1.748s, Decode+Unpack: 1.349s +---------------------- -------------------------------------------------------- +💾 Converting with 264.5587 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,736B, BPFP=0.5667 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,104B, BPFP=2.7464 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,196B, BPFP=1.2433 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,788B, BPFP=2.6984 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,128B, BPFP=1.3847 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,488B, BPFP=2.6529 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,892B, BPFP=1.3489 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,828B, BPFP=2.7045 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,720B, BPFP=2.3847 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,344B, BPFP=2.6311 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,660B, BPFP=0.3610 +⌛️ [2/4] FRONTEND: Frontend time: 2.055s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.288s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13030120 41.71523248 + layer.0.v_cache 0.00001385 0.00637882 + layer.1.k_cache 0.02890817 8.59669272 + layer.1.v_cache 0.00000611 0.00283379 + layer.2.k_cache 0.00721959 1.08731109 + layer.2.v_cache 0.00001830 0.00783317 + layer.3.k_cache 0.02195997 4.95495813 + layer.3.v_cache 0.00001903 0.00845070 + layer.4.k_cache 0.00063766 0.20779671 + layer.4.v_cache 0.00005061 0.01693422 + layer.4.output 11.07892277 499.54802358 + ------------------------------------------------------------------------------------- + TOTAL 4.57303494 209.02591688 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 150884 +BPFP 1.3464 bits/point +EBPFP 2.6928 equivalent bits/point +MSE 209.025917 +---------------------- -------------------------------------------------------- +Time: 3.348s Load: 0.005s, Pack+Encode: 2.055s, Decode+Unpack: 1.288s +---------------------- -------------------------------------------------------- +💾 Converting with 209.0259 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,188B, BPFP=0.6002 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,144B, BPFP=3.0392 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,572B, BPFP=1.2372 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,732B, BPFP=2.9616 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,908B, BPFP=1.3005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,760B, BPFP=2.7786 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,696B, BPFP=1.2605 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,884B, BPFP=2.8020 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,328B, BPFP=2.3208 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,812B, BPFP=2.9767 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,880B, BPFP=0.3733 +⌛️ [2/4] FRONTEND: Frontend time: 1.895s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12633393 42.54879871 + layer.0.v_cache 0.00001372 0.00672346 + layer.1.k_cache 0.01411889 8.07676017 + layer.1.v_cache 0.00000549 0.00284630 + layer.2.k_cache 0.00833277 1.10312073 + layer.2.v_cache 0.00001851 0.00822431 + layer.3.k_cache 0.05797521 4.02180205 + layer.3.v_cache 0.00001879 0.00902413 + layer.4.k_cache 0.00064672 0.19219683 + layer.4.v_cache 0.00005349 0.01841975 + layer.4.output 0.17447430 635.56986876 + ------------------------------------------------------------------------------------- + TOTAL 0.08404927 264.99864693 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 126904 +BPFP 1.4053 bits/point +EBPFP 2.8106 equivalent bits/point +MSE 264.998647 +---------------------- -------------------------------------------------------- +Time: 3.278s Load: 0.005s, Pack+Encode: 1.895s, Decode+Unpack: 1.377s +---------------------- -------------------------------------------------------- +💾 Converting with 264.9986 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,788B, BPFP=0.6223 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,600B, BPFP=3.0357 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,584B, BPFP=1.2464 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,432B, BPFP=2.7750 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,024B, BPFP=1.3446 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,732B, BPFP=2.6187 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,808B, BPFP=1.2964 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,836B, BPFP=2.6420 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,632B, BPFP=2.1500 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,408B, BPFP=2.5464 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,748B, BPFP=0.4065 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.287s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08086312 42.91750140 + layer.0.v_cache 0.00001331 0.00715414 + layer.1.k_cache 0.01551799 7.96870117 + layer.1.v_cache 0.00000513 0.00271781 + layer.2.k_cache 0.00232868 1.01151450 + layer.2.v_cache 0.00002262 0.00906061 + layer.3.k_cache 0.10623371 4.19997210 + layer.3.v_cache 0.00001836 0.00999158 + layer.4.k_cache 0.00061897 0.20636293 + layer.4.v_cache 0.00005922 0.01884223 + layer.4.output 0.20964142 721.42021684 + ------------------------------------------------------------------------------------- + TOTAL 0.09842183 300.37019625 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 103592 +BPFP 1.3602 bits/point +EBPFP 2.7204 equivalent bits/point +MSE 300.370196 +---------------------- -------------------------------------------------------- +Time: 3.132s Load: 0.005s, Pack+Encode: 1.840s, Decode+Unpack: 1.287s +---------------------- -------------------------------------------------------- +💾 Converting with 300.3702 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,340B, BPFP=0.5673 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,528B, BPFP=2.9769 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,056B, BPFP=1.1984 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,204B, BPFP=2.9219 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,480B, BPFP=1.2704 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,044B, BPFP=2.7249 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,148B, BPFP=1.2140 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,664B, BPFP=2.8302 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,976B, BPFP=2.3736 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,540B, BPFP=2.8091 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,288B, BPFP=0.3709 +⌛️ [2/4] FRONTEND: Frontend time: 1.794s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.420s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13670269 42.84261422 + layer.0.v_cache 0.00001568 0.00640012 + layer.1.k_cache 0.03139175 7.85250987 + layer.1.v_cache 0.00000602 0.00273373 + layer.2.k_cache 0.00629326 0.88076749 + layer.2.v_cache 0.00001871 0.00723955 + layer.3.k_cache 0.03837063 3.56976849 + layer.3.v_cache 0.00001860 0.00795932 + layer.4.k_cache 0.00062597 0.18695048 + layer.4.v_cache 0.00005354 0.01696418 + layer.4.output 0.15974679 561.97001165 + ------------------------------------------------------------------------------------- + TOTAL 0.07833673 234.65670523 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 138268 +BPFP 1.3814 bits/point +EBPFP 2.7627 equivalent bits/point +MSE 234.656705 +---------------------- -------------------------------------------------------- +Time: 3.219s Load: 0.005s, Pack+Encode: 1.794s, Decode+Unpack: 1.420s +---------------------- -------------------------------------------------------- +💾 Converting with 234.6567 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,880B, BPFP=0.5666 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,336B, BPFP=2.6776 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,552B, BPFP=1.2488 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,960B, BPFP=2.6227 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,960B, BPFP=1.6005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,720B, BPFP=2.5876 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,656B, BPFP=1.4100 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,984B, BPFP=2.6262 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,052B, BPFP=2.3440 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,412B, BPFP=2.5426 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,092B, BPFP=0.3566 +⌛️ [2/4] FRONTEND: Frontend time: 1.837s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12281527 41.25413442 + layer.0.v_cache 0.00001417 0.00652555 + layer.1.k_cache 0.04442539 8.75498264 + layer.1.v_cache 0.00000572 0.00277585 + layer.2.k_cache 0.01033308 1.07026858 + layer.2.v_cache 0.00001923 0.00822907 + layer.3.k_cache 0.04555132 4.56061846 + layer.3.v_cache 0.00001841 0.00863285 + layer.4.k_cache 0.00060580 0.21447331 + layer.4.v_cache 0.00007068 0.01724598 + layer.4.output 10.66341841 481.67001836 + ------------------------------------------------------------------------------------- + TOTAL 4.40398753 201.62282442 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 155604 +BPFP 1.3366 bits/point +EBPFP 2.6732 equivalent bits/point +MSE 201.622824 +---------------------- -------------------------------------------------------- +Time: 3.114s Load: 0.005s, Pack+Encode: 1.837s, Decode+Unpack: 1.272s +---------------------- -------------------------------------------------------- +💾 Converting with 201.6228 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,208B, BPFP=0.5828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,860B, BPFP=3.0632 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,928B, BPFP=1.2587 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,020B, BPFP=2.9106 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,336B, BPFP=1.3328 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,612B, BPFP=2.6548 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,040B, BPFP=1.2791 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,260B, BPFP=2.5908 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,928B, BPFP=2.1672 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,224B, BPFP=2.7660 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,116B, BPFP=0.3664 +⌛️ [2/4] FRONTEND: Frontend time: 1.848s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11362342 40.31396768 + layer.0.v_cache 0.00001418 0.00668356 + layer.1.k_cache 0.05429294 8.35750988 + layer.1.v_cache 0.00000580 0.00285116 + layer.2.k_cache 0.00513919 0.87623507 + layer.2.v_cache 0.00001816 0.00805780 + layer.3.k_cache 0.05885953 3.81881217 + layer.3.v_cache 0.00001841 0.00867571 + layer.4.k_cache 0.00061763 0.20997464 + layer.4.v_cache 0.00005080 0.01708126 + layer.4.output 0.16727475 578.88377284 + ------------------------------------------------------------------------------------- + TOTAL 0.08256255 241.51801522 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 127532 +BPFP 1.3630 bits/point +EBPFP 2.7260 equivalent bits/point +MSE 241.518015 +---------------------- -------------------------------------------------------- +Time: 3.055s Load: 0.004s, Pack+Encode: 1.848s, Decode+Unpack: 1.203s +---------------------- -------------------------------------------------------- +💾 Converting with 241.5180 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,340B, BPFP=0.5612 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,608B, BPFP=2.9583 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,308B, BPFP=1.2278 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,348B, BPFP=2.7466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,700B, BPFP=1.2937 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,316B, BPFP=2.7413 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,372B, BPFP=1.2386 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,664B, BPFP=2.7997 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,664B, BPFP=2.4637 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,848B, BPFP=2.8306 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,380B, BPFP=0.3691 +⌛️ [2/4] FRONTEND: Frontend time: 2.079s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702597 41.34709446 + layer.0.v_cache 0.00001402 0.00672934 + layer.1.k_cache 0.01332537 8.45120994 + layer.1.v_cache 0.00000581 0.00268691 + layer.2.k_cache 0.00507878 0.92576271 + layer.2.v_cache 0.00001944 0.00767597 + layer.3.k_cache 0.04051446 4.25731962 + layer.3.v_cache 0.00001931 0.00888158 + layer.4.k_cache 0.00062870 0.19061888 + layer.4.v_cache 0.00006514 0.01699804 + layer.4.output 0.16549673 531.15941820 + ------------------------------------------------------------------------------------- + TOTAL 0.07795142 221.96064146 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 139548 +BPFP 1.3792 bits/point +EBPFP 2.7583 equivalent bits/point +MSE 221.960641 +---------------------- -------------------------------------------------------- +Time: 3.398s Load: 0.004s, Pack+Encode: 2.079s, Decode+Unpack: 1.315s +---------------------- -------------------------------------------------------- +💾 Converting with 221.9606 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,212B, BPFP=0.5904 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,904B, BPFP=2.9235 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,636B, BPFP=1.2199 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,340B, BPFP=3.0037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,004B, BPFP=1.2875 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,392B, BPFP=2.6456 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,852B, BPFP=1.2596 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,320B, BPFP=2.8162 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,300B, BPFP=2.0772 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,076B, BPFP=2.7713 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,892B, BPFP=0.3911 +⌛️ [2/4] FRONTEND: Frontend time: 1.948s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.119s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13925676 43.06939913 + layer.0.v_cache 0.00001387 0.00657275 + layer.1.k_cache 0.03290265 8.79715935 + layer.1.v_cache 0.00000589 0.00273568 + layer.2.k_cache 0.00800494 1.00102593 + layer.2.v_cache 0.00001921 0.00760067 + layer.3.k_cache 0.08301921 4.33315430 + layer.3.v_cache 0.00001931 0.00847243 + layer.4.k_cache 0.00065893 0.18500173 + layer.4.v_cache 0.00004891 0.01684769 + layer.4.output 0.17367665 597.99753151 + ------------------------------------------------------------------------------------- + TOTAL 0.08704037 249.61239354 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 126928 +BPFP 1.3725 bits/point +EBPFP 2.7450 equivalent bits/point +MSE 249.612394 +---------------------- -------------------------------------------------------- +Time: 3.073s Load: 0.006s, Pack+Encode: 1.948s, Decode+Unpack: 1.119s +---------------------- -------------------------------------------------------- +💾 Converting with 249.6124 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,736B, BPFP=0.5667 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,180B, BPFP=2.7579 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,628B, BPFP=1.3089 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,856B, BPFP=2.7087 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,352B, BPFP=1.4187 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,600B, BPFP=2.6699 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,400B, BPFP=1.2743 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,704B, BPFP=2.6857 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,828B, BPFP=2.4011 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,464B, BPFP=2.6493 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,604B, BPFP=0.3165 +⌛️ [2/4] FRONTEND: Frontend time: 1.807s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.563s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10583916 40.56218712 + layer.0.v_cache 0.00001406 0.00637922 + layer.1.k_cache 0.01351924 8.88671164 + layer.1.v_cache 0.00000536 0.00265119 + layer.2.k_cache 0.00567923 1.00520465 + layer.2.v_cache 0.00001856 0.00777307 + layer.3.k_cache 0.05293757 3.89689977 + layer.3.v_cache 0.00001841 0.00883532 + layer.4.k_cache 0.00064880 0.19788509 + layer.4.v_cache 0.00004836 0.01731370 + layer.4.output 11.07886250 507.25810506 + ------------------------------------------------------------------------------------- + TOTAL 4.57239802 212.08226919 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 149352 +BPFP 1.3327 bits/point +EBPFP 2.6655 equivalent bits/point +MSE 212.082269 +---------------------- -------------------------------------------------------- +Time: 3.375s Load: 0.005s, Pack+Encode: 1.807s, Decode+Unpack: 1.563s +---------------------- -------------------------------------------------------- +💾 Converting with 212.0823 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,208B, BPFP=0.6039 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,304B, BPFP=2.8810 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,624B, BPFP=1.2470 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,120B, BPFP=2.8464 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,044B, BPFP=1.3261 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,600B, BPFP=2.5602 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,020B, BPFP=1.3215 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,608B, BPFP=2.7500 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,532B, BPFP=2.1709 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,640B, BPFP=2.5678 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,932B, BPFP=0.3747 +⌛️ [2/4] FRONTEND: Frontend time: 2.513s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.415s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860243 43.64367529 + layer.0.v_cache 0.00001403 0.00659215 + layer.1.k_cache 0.01400426 8.46375835 + layer.1.v_cache 0.00000584 0.00274789 + layer.2.k_cache 0.01277778 0.84135180 + layer.2.v_cache 0.00001980 0.00826671 + layer.3.k_cache 0.07376247 3.45627677 + layer.3.v_cache 0.00001877 0.00863180 + layer.4.k_cache 0.00062152 0.19426371 + layer.4.v_cache 0.00005140 0.01776343 + layer.4.output 0.18179212 631.06115534 + ------------------------------------------------------------------------------------- + TOTAL 0.08837783 263.18067149 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 121632 +BPFP 1.3469 bits/point +EBPFP 2.6938 equivalent bits/point +MSE 263.180671 +---------------------- -------------------------------------------------------- +Time: 3.932s Load: 0.004s, Pack+Encode: 2.513s, Decode+Unpack: 1.415s +---------------------- -------------------------------------------------------- +💾 Converting with 263.1807 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,332B, BPFP=0.5785 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,156B, BPFP=2.9785 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,172B, BPFP=1.2451 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,940B, BPFP=2.9410 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,548B, BPFP=1.3104 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,780B, BPFP=2.9132 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,308B, BPFP=1.2688 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,340B, BPFP=2.8368 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,024B, BPFP=2.4347 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,376B, BPFP=2.8431 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,504B, BPFP=0.3845 +⌛️ [2/4] FRONTEND: Frontend time: 1.818s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.160s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13781073 39.49985080 + layer.0.v_cache 0.00001414 0.00685543 + layer.1.k_cache 0.01252146 7.54532335 + layer.1.v_cache 0.00000561 0.00264608 + layer.2.k_cache 0.00238173 0.86362745 + layer.2.v_cache 0.00001849 0.00811237 + layer.3.k_cache 0.04360520 3.52191196 + layer.3.v_cache 0.00001973 0.00939954 + layer.4.k_cache 0.00061668 0.18404728 + layer.4.v_cache 0.00005448 0.01799465 + layer.4.output 0.16772948 567.22678571 + ------------------------------------------------------------------------------------- + TOTAL 0.08065615 236.60278052 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 138480 +BPFP 1.4142 bits/point +EBPFP 2.8284 equivalent bits/point +MSE 236.602781 +---------------------- -------------------------------------------------------- +Time: 2.982s Load: 0.004s, Pack+Encode: 1.818s, Decode+Unpack: 1.160s +---------------------- -------------------------------------------------------- +💾 Converting with 236.6028 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,136B, BPFP=0.6203 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,924B, BPFP=2.9517 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,276B, BPFP=1.2413 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,244B, BPFP=2.8172 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,720B, BPFP=1.3291 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,168B, BPFP=2.6044 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,432B, BPFP=1.2722 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,260B, BPFP=2.6226 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,296B, BPFP=2.2342 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,268B, BPFP=2.6242 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,120B, BPFP=0.3990 +⌛️ [2/4] FRONTEND: Frontend time: 1.830s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07550411 42.39624642 + layer.0.v_cache 0.00001381 0.00652285 + layer.1.k_cache 0.01271453 8.09422437 + layer.1.v_cache 0.00000531 0.00264077 + layer.2.k_cache 0.00681068 0.97069888 + layer.2.v_cache 0.00001904 0.00833700 + layer.3.k_cache 0.02844424 4.45844897 + layer.3.v_cache 0.00001798 0.00910005 + layer.4.k_cache 0.00061127 0.21596271 + layer.4.v_cache 0.00009323 0.01879307 + layer.4.output 0.18255964 665.47406193 + ------------------------------------------------------------------------------------- + TOTAL 0.08247951 277.32349462 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 116844 +BPFP 1.3594 bits/point +EBPFP 2.7188 equivalent bits/point +MSE 277.323495 +---------------------- -------------------------------------------------------- +Time: 3.142s Load: 0.004s, Pack+Encode: 1.830s, Decode+Unpack: 1.309s +---------------------- -------------------------------------------------------- +💾 Converting with 277.3235 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,144B, BPFP=0.6141 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,572B, BPFP=3.0414 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,544B, BPFP=1.2781 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,496B, BPFP=2.8312 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,780B, BPFP=1.3242 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,168B, BPFP=2.7672 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,652B, BPFP=1.2992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,880B, BPFP=2.7109 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,484B, BPFP=2.2430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,916B, BPFP=2.7180 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,488B, BPFP=0.4042 +⌛️ [2/4] FRONTEND: Frontend time: 1.848s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.287s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08590809 44.23999329 + layer.0.v_cache 0.00001388 0.00651561 + layer.1.k_cache 0.01594085 8.91059113 + layer.1.v_cache 0.00000552 0.00269734 + layer.2.k_cache 0.00791870 0.93274021 + layer.2.v_cache 0.00001801 0.00780883 + layer.3.k_cache 0.02757127 3.99528618 + layer.3.v_cache 0.00001862 0.00868731 + layer.4.k_cache 0.00063773 0.19680234 + layer.4.v_cache 0.00005492 0.01762805 + layer.4.output 0.18408036 645.33258929 + ------------------------------------------------------------------------------------- + TOTAL 0.08392059 269.15569855 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 121124 +BPFP 1.3916 bits/point +EBPFP 2.7832 equivalent bits/point +MSE 269.155699 +---------------------- -------------------------------------------------------- +Time: 3.139s Load: 0.004s, Pack+Encode: 1.848s, Decode+Unpack: 1.287s +---------------------- -------------------------------------------------------- +💾 Converting with 269.1557 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,020B, BPFP=0.6209 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,320B, BPFP=2.7385 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,064B, BPFP=1.2467 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,496B, BPFP=2.7747 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,416B, BPFP=1.3191 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,156B, BPFP=2.4992 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,276B, BPFP=1.2903 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,564B, BPFP=2.5831 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,164B, BPFP=2.0896 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,132B, BPFP=2.4942 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,132B, BPFP=0.4151 +⌛️ [2/4] FRONTEND: Frontend time: 1.776s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.329s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08991440 42.40427439 + layer.0.v_cache 0.00001347 0.00631782 + layer.1.k_cache 0.03345199 8.79756486 + layer.1.v_cache 0.00000518 0.00253861 + layer.2.k_cache 0.00398020 0.97469681 + layer.2.v_cache 0.00001815 0.00759729 + layer.3.k_cache 0.04607732 4.33824760 + layer.3.v_cache 0.00001896 0.00868343 + layer.4.k_cache 0.00062119 0.19720193 + layer.4.v_cache 0.00005040 0.01789910 + layer.4.output 0.18612516 690.24236372 + ------------------------------------------------------------------------------------- + TOTAL 0.08688396 287.55597458 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 109740 +BPFP 1.3272 bits/point +EBPFP 2.6543 equivalent bits/point +MSE 287.555975 +---------------------- -------------------------------------------------------- +Time: 3.110s Load: 0.004s, Pack+Encode: 1.776s, Decode+Unpack: 1.329s +---------------------- -------------------------------------------------------- +💾 Converting with 287.5560 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,320B, BPFP=0.5701 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,768B, BPFP=3.0508 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,160B, BPFP=1.2294 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,352B, BPFP=2.8077 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,616B, BPFP=1.3077 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,488B, BPFP=2.8310 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,500B, BPFP=1.2878 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,696B, BPFP=2.8668 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,232B, BPFP=2.4437 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,536B, BPFP=2.8393 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,508B, BPFP=0.3804 +⌛️ [2/4] FRONTEND: Frontend time: 2.082s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11186004 42.38435246 + layer.0.v_cache 0.00001364 0.00646402 + layer.1.k_cache 0.01386243 7.41634428 + layer.1.v_cache 0.00000608 0.00268860 + layer.2.k_cache 0.00506424 1.00333195 + layer.2.v_cache 0.00001872 0.00764947 + layer.3.k_cache 0.02496567 3.75795536 + layer.3.v_cache 0.00001876 0.00861924 + layer.4.k_cache 0.00064518 0.19060785 + layer.4.v_cache 0.00005436 0.01741147 + layer.4.output 0.16682192 578.28787284 + ------------------------------------------------------------------------------------- + TOTAL 0.07789780 241.34179615 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 139176 +BPFP 1.4057 bits/point +EBPFP 2.8114 equivalent bits/point +MSE 241.341796 +---------------------- -------------------------------------------------------- +Time: 3.472s Load: 0.003s, Pack+Encode: 2.082s, Decode+Unpack: 1.387s +---------------------- -------------------------------------------------------- +💾 Converting with 241.3418 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,200B, BPFP=0.6024 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,688B, BPFP=2.9533 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,580B, BPFP=1.2387 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,432B, BPFP=2.9051 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,912B, BPFP=1.3012 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,284B, BPFP=2.6890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,640B, BPFP=1.2500 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,988B, BPFP=2.8215 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,088B, BPFP=2.0873 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,708B, BPFP=2.9571 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,820B, BPFP=0.3986 +⌛️ [2/4] FRONTEND: Frontend time: 1.813s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08542032 42.99557900 + layer.0.v_cache 0.00001402 0.00642904 + layer.1.k_cache 0.01083013 8.91708926 + layer.1.v_cache 0.00000556 0.00262942 + layer.2.k_cache 0.00382075 0.92370403 + layer.2.v_cache 0.00001860 0.00815244 + layer.3.k_cache 0.02610962 3.50269217 + layer.3.v_cache 0.00002051 0.00903445 + layer.4.k_cache 0.00063312 0.20460637 + layer.4.v_cache 0.00005614 0.01899385 + layer.4.output 0.16811641 628.28216437 + ------------------------------------------------------------------------------------- + TOTAL 0.07669080 262.03318004 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 125340 +BPFP 1.3880 bits/point +EBPFP 2.7760 equivalent bits/point +MSE 262.033180 +---------------------- -------------------------------------------------------- +Time: 3.203s Load: 0.004s, Pack+Encode: 1.813s, Decode+Unpack: 1.386s +---------------------- -------------------------------------------------------- +💾 Converting with 262.0332 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,144B, BPFP=0.6065 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,520B, BPFP=2.8009 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,524B, BPFP=1.2585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,060B, BPFP=2.9051 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,916B, BPFP=1.3341 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,916B, BPFP=2.6844 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,816B, BPFP=1.3148 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,464B, BPFP=2.5972 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,112B, BPFP=2.1435 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,104B, BPFP=2.7207 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,884B, BPFP=0.4102 +⌛️ [2/4] FRONTEND: Frontend time: 1.836s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10540552 41.13905165 + layer.0.v_cache 0.00001541 0.00669358 + layer.1.k_cache 0.03402772 8.02103678 + layer.1.v_cache 0.00000555 0.00286403 + layer.2.k_cache 0.00394259 0.92971406 + layer.2.v_cache 0.00001868 0.00822026 + layer.3.k_cache 0.02873192 3.90063552 + layer.3.v_cache 0.00001983 0.00923364 + layer.4.k_cache 0.00063741 0.21126147 + layer.4.v_cache 0.00005806 0.01867263 + layer.4.output 0.17961645 620.49360670 + ------------------------------------------------------------------------------------- + TOTAL 0.08412811 258.68839003 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 120460 +BPFP 1.3669 bits/point +EBPFP 2.7338 equivalent bits/point +MSE 258.688390 +---------------------- -------------------------------------------------------- +Time: 3.160s Load: 0.004s, Pack+Encode: 1.836s, Decode+Unpack: 1.320s +---------------------- -------------------------------------------------------- +💾 Converting with 258.6884 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,132B, BPFP=0.6117 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,296B, BPFP=2.7922 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,416B, BPFP=1.2531 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,072B, BPFP=2.7484 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,932B, BPFP=1.3539 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,768B, BPFP=2.6891 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,668B, BPFP=1.3023 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,652B, BPFP=2.6664 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,856B, BPFP=2.1203 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,020B, BPFP=2.7383 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,160B, BPFP=0.4230 +⌛️ [2/4] FRONTEND: Frontend time: 2.024s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.262s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10022384 41.41996460 + layer.0.v_cache 0.00001350 0.00645192 + layer.1.k_cache 0.03445883 8.03578262 + layer.1.v_cache 0.00000556 0.00276517 + layer.2.k_cache 0.00550650 0.90929461 + layer.2.v_cache 0.00001848 0.00795967 + layer.3.k_cache 0.03016714 3.62084808 + layer.3.v_cache 0.00001937 0.00956415 + layer.4.k_cache 0.00063435 0.21041789 + layer.4.v_cache 0.00006510 0.02121077 + layer.4.output 0.18593751 658.86060268 + ------------------------------------------------------------------------------------- + TOTAL 0.08662796 274.48638107 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 118972 +BPFP 1.3669 bits/point +EBPFP 2.7337 equivalent bits/point +MSE 274.486381 +---------------------- -------------------------------------------------------- +Time: 3.294s Load: 0.008s, Pack+Encode: 2.024s, Decode+Unpack: 1.262s +---------------------- -------------------------------------------------------- +💾 Converting with 274.4864 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,372B, BPFP=0.5432 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,720B, BPFP=2.8544 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,300B, BPFP=1.1759 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,468B, BPFP=2.8138 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,648B, BPFP=1.3930 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,900B, BPFP=2.7223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,920B, BPFP=1.2758 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,296B, BPFP=2.7861 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,436B, BPFP=2.4865 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,920B, BPFP=2.7255 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,108B, BPFP=0.3477 +⌛️ [2/4] FRONTEND: Frontend time: 1.864s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11898830 44.31551526 + layer.0.v_cache 0.00001403 0.00668393 + layer.1.k_cache 0.01243949 7.99355041 + layer.1.v_cache 0.00000551 0.00270704 + layer.2.k_cache 0.00346032 1.03393476 + layer.2.v_cache 0.00001901 0.00843894 + layer.3.k_cache 0.02403709 5.19651873 + layer.3.v_cache 0.00001994 0.00920492 + layer.4.k_cache 0.00063949 0.18609308 + layer.4.v_cache 0.00006591 0.01691340 + layer.4.output 0.03773703 545.02766016 + ------------------------------------------------------------------------------------- + TOTAL 0.02493225 227.88018715 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 144088 +BPFP 1.3653 bits/point +EBPFP 2.7306 equivalent bits/point +MSE 227.880187 +---------------------- -------------------------------------------------------- +Time: 3.141s Load: 0.004s, Pack+Encode: 1.864s, Decode+Unpack: 1.272s +---------------------- -------------------------------------------------------- +💾 Converting with 227.8802 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,144B, BPFP=0.5991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,308B, BPFP=2.9169 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,516B, BPFP=1.2416 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,784B, BPFP=2.6265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,824B, BPFP=1.3003 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,544B, BPFP=2.5808 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,648B, BPFP=1.2668 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,440B, BPFP=2.5610 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,968B, BPFP=2.2805 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,344B, BPFP=2.7332 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,668B, BPFP=0.3993 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10693085 43.29314739 + layer.0.v_cache 0.00001341 0.00684734 + layer.1.k_cache 0.01172146 9.09551406 + layer.1.v_cache 0.00000553 0.00285481 + layer.2.k_cache 0.00236719 0.82329764 + layer.2.v_cache 0.00001835 0.00833210 + layer.3.k_cache 0.02622769 3.57735536 + layer.3.v_cache 0.00001864 0.00934182 + layer.4.k_cache 0.00061713 0.20943290 + layer.4.v_cache 0.00004943 0.01771126 + layer.4.output 0.18273007 618.20268946 + ------------------------------------------------------------------------------------- + TOTAL 0.08394589 257.90956829 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 120188 +BPFP 1.3472 bits/point +EBPFP 2.6943 equivalent bits/point +MSE 257.909568 +---------------------- -------------------------------------------------------- +Time: 3.059s Load: 0.004s, Pack+Encode: 1.834s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 257.9096 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,620B, BPFP=0.5600 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,956B, BPFP=2.7778 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,896B, BPFP=1.2215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,564B, BPFP=2.7172 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,256B, BPFP=1.2772 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,240B, BPFP=2.6671 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,152B, BPFP=1.4158 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,356B, BPFP=2.6850 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,408B, BPFP=2.3837 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,192B, BPFP=2.6597 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,956B, BPFP=0.3305 +⌛️ [2/4] FRONTEND: Frontend time: 2.424s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.614s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13156438 41.83053740 + layer.0.v_cache 0.00001398 0.00676141 + layer.1.k_cache 0.04507210 8.27592604 + layer.1.v_cache 0.00000603 0.00284789 + layer.2.k_cache 0.00848899 1.01222486 + layer.2.v_cache 0.00001983 0.00860208 + layer.3.k_cache 0.06093804 5.06474078 + layer.3.v_cache 0.00001891 0.00895705 + layer.4.k_cache 0.00062755 0.19948823 + layer.4.v_cache 0.00004781 0.01711764 + layer.4.output 11.29576791 499.60930870 + ------------------------------------------------------------------------------------- + TOTAL 4.66571606 209.04072731 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 146596 +BPFP 1.3340 bits/point +EBPFP 2.6681 equivalent bits/point +MSE 209.040727 +---------------------- -------------------------------------------------------- +Time: 4.042s Load: 0.005s, Pack+Encode: 2.424s, Decode+Unpack: 1.614s +---------------------- -------------------------------------------------------- +💾 Converting with 209.0407 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,212B, BPFP=0.6047 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,992B, BPFP=3.0105 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,628B, BPFP=1.2477 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,648B, BPFP=2.9458 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,124B, BPFP=1.3411 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,484B, BPFP=2.7267 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,012B, BPFP=1.3200 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,536B, BPFP=2.7364 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,920B, BPFP=2.2440 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,692B, BPFP=2.7658 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,612B, BPFP=0.3930 +⌛️ [2/4] FRONTEND: Frontend time: 2.046s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.279s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12340720 41.74590550 + layer.0.v_cache 0.00001379 0.00649392 + layer.1.k_cache 0.01365945 8.57208031 + layer.1.v_cache 0.00000586 0.00270180 + layer.2.k_cache 0.00515996 1.10627783 + layer.2.v_cache 0.00002118 0.00841546 + layer.3.k_cache 0.04682968 3.77555866 + layer.3.v_cache 0.00001899 0.00890743 + layer.4.k_cache 0.00062596 0.21123904 + layer.4.v_cache 0.00004975 0.01746355 + layer.4.output 0.18155291 638.81567341 + ------------------------------------------------------------------------------------- + TOTAL 0.08592130 266.30380926 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 125860 +BPFP 1.3937 bits/point +EBPFP 2.7875 equivalent bits/point +MSE 266.303809 +---------------------- -------------------------------------------------------- +Time: 3.329s Load: 0.004s, Pack+Encode: 2.046s, Decode+Unpack: 1.279s +---------------------- -------------------------------------------------------- +💾 Converting with 266.3038 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,204B, BPFP=0.5960 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,636B, BPFP=3.0945 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,636B, BPFP=1.2344 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,788B, BPFP=2.9368 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,892B, BPFP=1.2820 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,068B, BPFP=2.8028 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,948B, BPFP=1.2924 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,504B, BPFP=2.8839 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,544B, BPFP=2.1473 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,632B, BPFP=2.9077 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,916B, BPFP=0.4229 +⌛️ [2/4] FRONTEND: Frontend time: 1.919s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.276s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10400932 42.99704125 + layer.0.v_cache 0.00001381 0.00647033 + layer.1.k_cache 0.01300498 7.81665838 + layer.1.v_cache 0.00000559 0.00267697 + layer.2.k_cache 0.00535372 1.10499736 + layer.2.v_cache 0.00001862 0.00769459 + layer.3.k_cache 0.04415199 3.88760776 + layer.3.v_cache 0.00001841 0.00874198 + layer.4.k_cache 0.00060249 0.19550964 + layer.4.v_cache 0.00005274 0.01791405 + layer.4.output 0.16694923 628.67777423 + ------------------------------------------------------------------------------------- + TOTAL 0.07858096 262.16410188 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 129768 +BPFP 1.4199 bits/point +EBPFP 2.8398 equivalent bits/point +MSE 262.164102 +---------------------- -------------------------------------------------------- +Time: 3.200s Load: 0.005s, Pack+Encode: 1.919s, Decode+Unpack: 1.276s +---------------------- -------------------------------------------------------- +💾 Converting with 262.1641 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,128B, BPFP=0.6109 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,284B, BPFP=2.9852 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,360B, BPFP=1.2422 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,832B, BPFP=2.7016 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,900B, BPFP=1.3477 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,692B, BPFP=2.6742 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,576B, BPFP=1.2844 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,116B, BPFP=2.9523 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,240B, BPFP=2.1953 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,848B, BPFP=3.0953 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,476B, BPFP=0.3760 +⌛️ [2/4] FRONTEND: Frontend time: 1.904s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.276s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09321334 42.59488220 + layer.0.v_cache 0.00001862 0.00685701 + layer.1.k_cache 0.03257937 8.63594131 + layer.1.v_cache 0.00000532 0.00267999 + layer.2.k_cache 0.01442847 1.05609093 + layer.2.v_cache 0.00001865 0.00845844 + layer.3.k_cache 0.04385926 3.78235016 + layer.3.v_cache 0.00001828 0.00937005 + layer.4.k_cache 0.00061064 0.23110783 + layer.4.v_cache 0.00005353 0.01964566 + layer.4.output 0.18500509 653.45245536 + ------------------------------------------------------------------------------------- + TOTAL 0.08704948 272.38321006 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 121452 +BPFP 1.3954 bits/point +EBPFP 2.7907 equivalent bits/point +MSE 272.383210 +---------------------- -------------------------------------------------------- +Time: 3.185s Load: 0.004s, Pack+Encode: 1.904s, Decode+Unpack: 1.276s +---------------------- -------------------------------------------------------- +💾 Converting with 272.3832 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,712B, BPFP=0.5631 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,080B, BPFP=2.7427 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,132B, BPFP=1.2336 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,568B, BPFP=2.6650 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,848B, BPFP=1.3422 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,500B, BPFP=2.6547 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,824B, BPFP=1.3386 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,632B, BPFP=2.6748 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,956B, BPFP=2.4205 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,396B, BPFP=2.6390 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,632B, BPFP=0.3388 +⌛️ [2/4] FRONTEND: Frontend time: 1.963s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.275s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11527186 40.60390435 + layer.0.v_cache 0.00001510 0.00667457 + layer.1.k_cache 0.02920971 8.58644474 + layer.1.v_cache 0.00000551 0.00271206 + layer.2.k_cache 0.00857845 0.94764969 + layer.2.v_cache 0.00001837 0.00791298 + layer.3.k_cache 0.03550095 5.26634142 + layer.3.v_cache 0.00001950 0.00847258 + layer.4.k_cache 0.00063242 0.20149487 + layer.4.v_cache 0.00005150 0.01733418 + layer.4.output 11.08123415 493.99267510 + ------------------------------------------------------------------------------------- + TOTAL 4.57399661 206.68221572 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 149280 +BPFP 1.3321 bits/point +EBPFP 2.6642 equivalent bits/point +MSE 206.682216 +---------------------- -------------------------------------------------------- +Time: 3.242s Load: 0.004s, Pack+Encode: 1.963s, Decode+Unpack: 1.275s +---------------------- -------------------------------------------------------- +💾 Converting with 206.6822 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,364B, BPFP=0.5713 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,244B, BPFP=2.9287 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,192B, BPFP=1.2215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,364B, BPFP=2.9490 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,484B, BPFP=1.2711 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,712B, BPFP=2.8383 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,404B, BPFP=1.2575 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,952B, BPFP=2.8791 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,324B, BPFP=2.6026 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,592B, BPFP=2.8179 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,360B, BPFP=0.3484 +⌛️ [2/4] FRONTEND: Frontend time: 2.100s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.418s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09673090 42.70353102 + layer.0.v_cache 0.00001344 0.00618463 + layer.1.k_cache 0.04918879 8.64888465 + layer.1.v_cache 0.00000572 0.00254722 + layer.2.k_cache 0.00880511 0.89965323 + layer.2.v_cache 0.00001944 0.00780106 + layer.3.k_cache 0.06586690 3.84743997 + layer.3.v_cache 0.00001823 0.00847856 + layer.4.k_cache 0.00062484 0.19010558 + layer.4.v_cache 0.00006325 0.01727796 + layer.4.output 0.15294319 574.88092003 + ------------------------------------------------------------------------------------- + TOTAL 0.07599641 240.02931436 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 139992 +BPFP 1.3986 bits/point +EBPFP 2.7972 equivalent bits/point +MSE 240.029314 +---------------------- -------------------------------------------------------- +Time: 3.523s Load: 0.004s, Pack+Encode: 2.100s, Decode+Unpack: 1.418s +---------------------- -------------------------------------------------------- +💾 Converting with 240.0293 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,344B, BPFP=0.5679 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,828B, BPFP=3.0279 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,224B, BPFP=1.2269 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,292B, BPFP=2.9368 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,740B, BPFP=1.3145 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,240B, BPFP=2.9280 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,488B, BPFP=1.2717 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,560B, BPFP=2.8125 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,724B, BPFP=2.5007 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,756B, BPFP=2.8458 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,352B, BPFP=0.3725 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12676482 41.51425569 + layer.0.v_cache 0.00001339 0.00634812 + layer.1.k_cache 0.03133728 8.25828221 + layer.1.v_cache 0.00000590 0.00262737 + layer.2.k_cache 0.00372887 0.85861729 + layer.2.v_cache 0.00001820 0.00768708 + layer.3.k_cache 0.02418065 3.38599296 + layer.3.v_cache 0.00002034 0.00876616 + layer.4.k_cache 0.00062996 0.19579589 + layer.4.v_cache 0.00006053 0.01702707 + layer.4.output 0.15389095 566.17857143 + ------------------------------------------------------------------------------------- + TOTAL 0.07435274 236.32384705 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 141548 +BPFP 1.4141 bits/point +EBPFP 2.8282 equivalent bits/point +MSE 236.323847 +---------------------- -------------------------------------------------------- +Time: 3.175s Load: 0.005s, Pack+Encode: 1.855s, Decode+Unpack: 1.315s +---------------------- -------------------------------------------------------- +💾 Converting with 236.3238 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,340B, BPFP=0.5799 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,340B, BPFP=3.0104 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,048B, BPFP=1.2236 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,020B, BPFP=2.9549 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,344B, BPFP=1.2750 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,108B, BPFP=2.7965 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,256B, BPFP=1.2597 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,500B, BPFP=2.8646 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,824B, BPFP=2.2264 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,472B, BPFP=2.8597 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,288B, BPFP=0.3792 +⌛️ [2/4] FRONTEND: Frontend time: 1.944s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12863920 40.61123047 + layer.0.v_cache 0.00001425 0.00661674 + layer.1.k_cache 0.01389826 7.87736138 + layer.1.v_cache 0.00000571 0.00272419 + layer.2.k_cache 0.00659207 0.97839237 + layer.2.v_cache 0.00001883 0.00763897 + layer.3.k_cache 0.05540012 3.50797933 + layer.3.v_cache 0.00001922 0.00834421 + layer.4.k_cache 0.00063279 0.20313038 + layer.4.v_cache 0.00005416 0.01691712 + layer.4.output 0.16196846 578.77465278 + ------------------------------------------------------------------------------------- + TOTAL 0.07876787 241.44958262 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 136540 +BPFP 1.3944 bits/point +EBPFP 2.7888 equivalent bits/point +MSE 241.449583 +---------------------- -------------------------------------------------------- +Time: 3.293s Load: 0.004s, Pack+Encode: 1.944s, Decode+Unpack: 1.345s +---------------------- -------------------------------------------------------- +💾 Converting with 241.4496 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,904B, BPFP=0.5648 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,256B, BPFP=2.6412 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,372B, BPFP=1.3559 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,772B, BPFP=2.5712 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,196B, BPFP=1.4751 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,636B, BPFP=2.5515 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,552B, BPFP=1.3819 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,912B, BPFP=2.5914 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,020B, BPFP=2.3177 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,400B, BPFP=2.5174 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,084B, BPFP=0.3531 +⌛️ [2/4] FRONTEND: Frontend time: 1.869s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.245s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13160179 43.52072483 + layer.0.v_cache 0.00001577 0.00659283 + layer.1.k_cache 0.05775581 9.26262524 + layer.1.v_cache 0.00000549 0.00267830 + layer.2.k_cache 0.01027601 0.91621173 + layer.2.v_cache 0.00002197 0.00810572 + layer.3.k_cache 0.07329086 4.36653250 + layer.3.v_cache 0.00001983 0.00910653 + layer.4.k_cache 0.00064477 0.21493675 + layer.4.v_cache 0.00005260 0.01745362 + layer.4.output 10.56448284 481.32113922 + ------------------------------------------------------------------------------------- + TOTAL 4.36618028 201.62193780 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 155104 +BPFP 1.3200 bits/point +EBPFP 2.6400 equivalent bits/point +MSE 201.621938 +---------------------- -------------------------------------------------------- +Time: 3.117s Load: 0.004s, Pack+Encode: 1.869s, Decode+Unpack: 1.245s +---------------------- -------------------------------------------------------- +💾 Converting with 201.6219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,196B, BPFP=0.6017 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,400B, BPFP=2.8991 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,584B, BPFP=1.2395 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,100B, BPFP=2.6544 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,056B, BPFP=1.3283 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,892B, BPFP=2.6152 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,700B, BPFP=1.2613 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,160B, BPFP=2.6657 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,544B, BPFP=2.1732 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,060B, BPFP=2.6468 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,032B, BPFP=0.3774 +⌛️ [2/4] FRONTEND: Frontend time: 1.718s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10139442 42.33614046 + layer.0.v_cache 0.00001346 0.00617913 + layer.1.k_cache 0.03266963 7.82201266 + layer.1.v_cache 0.00000526 0.00258066 + layer.2.k_cache 0.01283502 0.92626585 + layer.2.v_cache 0.00001884 0.00779103 + layer.3.k_cache 0.02666297 3.73302965 + layer.3.v_cache 0.00001780 0.00817679 + layer.4.k_cache 0.00061512 0.20173498 + layer.4.v_cache 0.00004947 0.01697747 + layer.4.output 0.18393325 641.09079174 + ------------------------------------------------------------------------------------- + TOTAL 0.08598910 267.21743711 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 120724 +BPFP 1.3369 bits/point +EBPFP 2.6737 equivalent bits/point +MSE 267.217437 +---------------------- -------------------------------------------------------- +Time: 3.048s Load: 0.004s, Pack+Encode: 1.718s, Decode+Unpack: 1.327s +---------------------- -------------------------------------------------------- +💾 Converting with 267.2174 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,624B, BPFP=0.5606 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,056B, BPFP=2.7933 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,136B, BPFP=1.2587 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,720B, BPFP=2.7413 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,540B, BPFP=1.3212 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,260B, BPFP=2.6702 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,788B, BPFP=1.3595 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,700B, BPFP=2.7382 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,328B, BPFP=2.3713 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,156B, BPFP=2.6541 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,764B, BPFP=0.3484 +⌛️ [2/4] FRONTEND: Frontend time: 1.904s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.279s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11149573 39.65437906 + layer.0.v_cache 0.00001357 0.00658034 + layer.1.k_cache 0.08092193 8.57558570 + layer.1.v_cache 0.00000571 0.00264368 + layer.2.k_cache 0.01109533 0.85529244 + layer.2.v_cache 0.00001920 0.00798762 + layer.3.k_cache 0.03617269 4.02628742 + layer.3.v_cache 0.00001858 0.00838883 + layer.4.k_cache 0.00062559 0.18285360 + layer.4.v_cache 0.00005684 0.01741697 + layer.4.output 11.30219499 494.03925035 + ------------------------------------------------------------------------------------- + TOTAL 4.66798765 206.56542166 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 148072 +BPFP 1.3475 bits/point +EBPFP 2.6950 equivalent bits/point +MSE 206.565422 +---------------------- -------------------------------------------------------- +Time: 3.188s Load: 0.005s, Pack+Encode: 1.904s, Decode+Unpack: 1.279s +---------------------- -------------------------------------------------------- +💾 Converting with 206.5654 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,088B, BPFP=0.6186 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,348B, BPFP=2.8742 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,344B, BPFP=1.2708 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,524B, BPFP=2.9095 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,632B, BPFP=1.3285 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,912B, BPFP=2.5865 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,376B, BPFP=1.2772 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,400B, BPFP=2.6843 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,604B, BPFP=2.1242 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,840B, BPFP=2.5721 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,952B, BPFP=0.3993 +⌛️ [2/4] FRONTEND: Frontend time: 2.180s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.268s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10167113 40.09376252 + layer.0.v_cache 0.00001386 0.00706233 + layer.1.k_cache 0.03545215 8.09193538 + layer.1.v_cache 0.00000569 0.00280757 + layer.2.k_cache 0.01047806 0.87978725 + layer.2.v_cache 0.00001755 0.00799178 + layer.3.k_cache 0.06687863 3.74737197 + layer.3.v_cache 0.00001960 0.00953279 + layer.4.k_cache 0.00059181 0.19862043 + layer.4.v_cache 0.00004882 0.01768207 + layer.4.output 0.19633365 669.94602793 + ------------------------------------------------------------------------------------- + TOTAL 0.09350076 278.98110292 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 115020 +BPFP 1.3553 bits/point +EBPFP 2.7107 equivalent bits/point +MSE 278.981103 +---------------------- -------------------------------------------------------- +Time: 3.452s Load: 0.005s, Pack+Encode: 2.180s, Decode+Unpack: 1.268s +---------------------- -------------------------------------------------------- +💾 Converting with 278.9811 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,264B, BPFP=0.5862 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,976B, BPFP=3.0489 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,760B, BPFP=1.2141 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,512B, BPFP=2.9655 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,216B, BPFP=1.2960 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,264B, BPFP=2.5618 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,040B, BPFP=1.2644 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,900B, BPFP=2.8556 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,556B, BPFP=2.2550 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,756B, BPFP=2.8297 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,012B, BPFP=0.3852 +⌛️ [2/4] FRONTEND: Frontend time: 1.783s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.181s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11445767 41.09184458 + layer.0.v_cache 0.00001583 0.00637293 + layer.1.k_cache 0.03092440 7.47083291 + layer.1.v_cache 0.00000551 0.00264327 + layer.2.k_cache 0.00639992 0.89293793 + layer.2.v_cache 0.00001897 0.00793760 + layer.3.k_cache 0.05613507 3.70566813 + layer.3.v_cache 0.00002026 0.00908743 + layer.4.k_cache 0.00063356 0.22138008 + layer.4.v_cache 0.00005123 0.01819640 + layer.4.output 0.16055210 596.72095649 + ------------------------------------------------------------------------------------- + TOTAL 0.07838395 248.85138804 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 131256 +BPFP 1.3867 bits/point +EBPFP 2.7733 equivalent bits/point +MSE 248.851388 +---------------------- -------------------------------------------------------- +Time: 2.967s Load: 0.003s, Pack+Encode: 1.783s, Decode+Unpack: 1.181s +---------------------- -------------------------------------------------------- +💾 Converting with 248.8514 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,188B, BPFP=0.5930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,932B, BPFP=2.9635 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,724B, BPFP=1.2507 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,464B, BPFP=3.0625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,164B, BPFP=1.3326 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,988B, BPFP=2.9740 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,984B, BPFP=1.2991 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,204B, BPFP=2.8281 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,488B, BPFP=2.3229 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,532B, BPFP=2.8891 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,860B, BPFP=0.3683 +⌛️ [2/4] FRONTEND: Frontend time: 1.761s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.157s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13956541 40.15977841 + layer.0.v_cache 0.00001349 0.00663230 + layer.1.k_cache 0.03336124 8.65053813 + layer.1.v_cache 0.00000558 0.00288525 + layer.2.k_cache 0.00698305 0.88601367 + layer.2.v_cache 0.00001831 0.00798990 + layer.3.k_cache 0.07289400 3.68189385 + layer.3.v_cache 0.00001855 0.00914288 + layer.4.k_cache 0.00061006 0.20922243 + layer.4.v_cache 0.00005429 0.01850081 + layer.4.output 0.17891866 619.82334184 + ------------------------------------------------------------------------------------- + TOTAL 0.08858556 258.37623473 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 129528 +BPFP 1.4173 bits/point +EBPFP 2.8346 equivalent bits/point +MSE 258.376235 +---------------------- -------------------------------------------------------- +Time: 2.922s Load: 0.004s, Pack+Encode: 1.761s, Decode+Unpack: 1.157s +---------------------- -------------------------------------------------------- +💾 Converting with 258.3762 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,144B, BPFP=0.6065 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,768B, BPFP=2.8488 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,380B, BPFP=1.2307 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,988B, BPFP=2.6983 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,752B, BPFP=1.3025 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,128B, BPFP=2.5324 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,580B, BPFP=1.2693 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,412B, BPFP=2.5872 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,892B, BPFP=2.1011 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,240B, BPFP=2.5540 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,308B, BPFP=0.3943 +⌛️ [2/4] FRONTEND: Frontend time: 1.814s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10837562 42.37031612 + layer.0.v_cache 0.00001644 0.00663286 + layer.1.k_cache 0.01241452 9.09593898 + layer.1.v_cache 0.00000532 0.00272891 + layer.2.k_cache 0.01724591 0.88334345 + layer.2.v_cache 0.00001958 0.00812395 + layer.3.k_cache 0.04641535 3.76963599 + layer.3.v_cache 0.00001847 0.00842876 + layer.4.k_cache 0.00063678 0.19367783 + layer.4.v_cache 0.00004974 0.01687495 + layer.4.output 0.18014439 608.19835758 + ------------------------------------------------------------------------------------- + TOTAL 0.08507109 253.74965911 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 116592 +BPFP 1.3230 bits/point +EBPFP 2.6460 equivalent bits/point +MSE 253.749659 +---------------------- -------------------------------------------------------- +Time: 3.163s Load: 0.003s, Pack+Encode: 1.814s, Decode+Unpack: 1.346s +---------------------- -------------------------------------------------------- +💾 Converting with 253.7497 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.3674 bits/point +Avg EBPFP 2.7348 equivalent bits/point +Avg MSE 250.531977 +Avg Time 6.151s +------------------------ ---------------------------- diff --git a/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..ae93f30f351aa6f19054f33f76deb4950d76df15 --- /dev/null +++ b/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 506 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean +Output output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean +---------------- ------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,996B, BPFP=0.4991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,776B, BPFP=2.6239 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,052B, BPFP=1.2880 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,760B, BPFP=2.5514 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,048B, BPFP=1.6444 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,584B, BPFP=2.5388 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,604B, BPFP=1.3987 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,052B, BPFP=2.5722 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,784B, BPFP=2.2677 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,244B, BPFP=2.5146 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,864B, BPFP=0.3553 +⌛️ [2/4] FRONTEND: Frontend time: 2.716s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 29.768s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12261864 40.32814284 + layer.0.v_cache 0.00001806 0.00706480 + layer.1.k_cache 0.30198812 8.51977539 + layer.1.v_cache 0.00000607 0.00267064 + layer.2.k_cache 0.01966049 0.81833095 + layer.2.v_cache 0.00002021 0.00779298 + layer.3.k_cache 0.01413035 5.61267146 + layer.3.v_cache 0.00002054 0.00880282 + layer.4.k_cache 0.00067868 0.20943319 + layer.4.v_cache 0.00004987 0.01519420 + layer.4.output 1.39792533 240.13676207 + ------------------------------------------------------------------------------------- + TOTAL 0.60262755 102.14630669 + (elements=1,906,176) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 1906176 +Total Bytes 313764 +BPFP 1.3168 bits/point +EBPFP 2.6337 equivalent bits/point +MSE 102.146307 +---------------------- --------------------------------------------------------- +Time: 32.492s Load: 0.008s, Pack+Encode: 2.716s, Decode+Unpack: 29.768s +---------------------- --------------------------------------------------------- +💾 Converting with 102.1463 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,976B, BPFP=0.5046 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,332B, BPFP=2.6282 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,744B, BPFP=1.2836 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,440B, BPFP=2.5637 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,204B, BPFP=1.5339 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,340B, BPFP=2.5564 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,604B, BPFP=1.4181 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,872B, BPFP=2.5949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,896B, BPFP=2.3073 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,828B, BPFP=2.5194 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,492B, BPFP=0.3151 +⌛️ [2/4] FRONTEND: Frontend time: 2.052s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.454s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11787021 40.88027615 + layer.0.v_cache 0.00001738 0.00673779 + layer.1.k_cache 0.25265321 8.65803980 + layer.1.v_cache 0.00000586 0.00266454 + layer.2.k_cache 0.01010228 0.93295627 + layer.2.v_cache 0.00002017 0.00790546 + layer.3.k_cache 0.01159736 3.79698238 + layer.3.v_cache 0.00002197 0.00922430 + layer.4.k_cache 0.00067244 0.25162850 + layer.4.v_cache 0.00004845 0.01551785 + layer.4.output 1.41728832 238.44186095 + ------------------------------------------------------------------------------------- + TOTAL 0.60670750 101.39146821 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 305728 +BPFP 1.3009 bits/point +EBPFP 2.6019 equivalent bits/point +MSE 101.391468 +---------------------- -------------------------------------------------------- +Time: 3.514s Load: 0.008s, Pack+Encode: 2.052s, Decode+Unpack: 1.454s +---------------------- -------------------------------------------------------- +💾 Converting with 101.3915 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,984B, BPFP=0.4893 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,300B, BPFP=2.5434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,032B, BPFP=1.3335 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,372B, BPFP=2.4784 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,520B, BPFP=1.5779 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,024B, BPFP=2.4540 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,796B, BPFP=1.4571 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,680B, BPFP=2.5000 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,776B, BPFP=2.2265 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,848B, BPFP=2.4417 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,208B, BPFP=0.2824 +⌛️ [2/4] FRONTEND: Frontend time: 1.971s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.547s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13311632 39.61321118 + layer.0.v_cache 0.00001526 0.00673339 + layer.1.k_cache 0.36474110 8.74039093 + layer.1.v_cache 0.00000573 0.00267052 + layer.2.k_cache 0.01296863 1.01201017 + layer.2.v_cache 0.00001984 0.00796693 + layer.3.k_cache 0.01490937 6.06519868 + layer.3.v_cache 0.00002017 0.00923645 + layer.4.k_cache 0.00067370 0.23001458 + layer.4.v_cache 0.00005228 0.01530941 + layer.4.output 1.37278975 240.59533152 + ------------------------------------------------------------------------------------- + TOTAL 0.59623828 102.34529782 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 306540 +BPFP 1.2634 bits/point +EBPFP 2.5269 equivalent bits/point +MSE 102.345298 +---------------------- -------------------------------------------------------- +Time: 3.525s Load: 0.007s, Pack+Encode: 1.971s, Decode+Unpack: 1.547s +---------------------- -------------------------------------------------------- +💾 Converting with 102.3453 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,672B, BPFP=0.4913 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,916B, BPFP=2.3640 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,080B, BPFP=1.2859 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,832B, BPFP=2.2946 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,004B, BPFP=1.5371 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,552B, BPFP=2.2766 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,996B, BPFP=1.4726 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,292B, BPFP=2.3240 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,940B, BPFP=2.0453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,188B, BPFP=2.2533 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,656B, BPFP=0.3170 +⌛️ [2/4] FRONTEND: Frontend time: 2.189s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.511s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11263794 40.53433578 + layer.0.v_cache 0.00001639 0.00677697 + layer.1.k_cache 0.42101050 8.90817911 + layer.1.v_cache 0.00000577 0.00261380 + layer.2.k_cache 0.00499437 0.87100745 + layer.2.v_cache 0.00001973 0.00759806 + layer.3.k_cache 0.03029772 6.40315722 + layer.3.v_cache 0.00002073 0.00903694 + layer.4.k_cache 0.00068836 0.23520920 + layer.4.v_cache 0.00005112 0.01558836 + layer.4.output 1.25471843 217.43266979 + ------------------------------------------------------------------------------------- + TOTAL 0.55016304 92.88365832 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 321128 +BPFP 1.2096 bits/point +EBPFP 2.4193 equivalent bits/point +MSE 92.883658 +---------------------- -------------------------------------------------------- +Time: 3.710s Load: 0.009s, Pack+Encode: 2.189s, Decode+Unpack: 1.511s +---------------------- -------------------------------------------------------- +💾 Converting with 92.8837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,996B, BPFP=0.5014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,776B, BPFP=2.6359 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,516B, BPFP=1.2554 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,780B, BPFP=2.5645 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,216B, BPFP=1.5923 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,444B, BPFP=2.5404 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,676B, BPFP=1.3386 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,240B, BPFP=2.5975 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,544B, BPFP=2.2609 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,332B, BPFP=2.5324 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,652B, BPFP=0.3343 +⌛️ [2/4] FRONTEND: Frontend time: 1.966s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.548s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10233516 39.46031931 + layer.0.v_cache 0.00001822 0.00704593 + layer.1.k_cache 0.31275415 8.47915117 + layer.1.v_cache 0.00000587 0.00264553 + layer.2.k_cache 0.01225604 0.93419612 + layer.2.v_cache 0.00002018 0.00789404 + layer.3.k_cache 0.02847941 4.96817381 + layer.3.v_cache 0.00002101 0.00902576 + layer.4.k_cache 0.00067306 0.22591309 + layer.4.v_cache 0.00005947 0.01622932 + layer.4.output 1.40430263 242.68405963 + ------------------------------------------------------------------------------------- + TOTAL 0.60510241 103.11170656 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 309172 +BPFP 1.3035 bits/point +EBPFP 2.6070 equivalent bits/point +MSE 103.111707 +---------------------- -------------------------------------------------------- +Time: 3.523s Load: 0.009s, Pack+Encode: 1.966s, Decode+Unpack: 1.548s +---------------------- -------------------------------------------------------- +💾 Converting with 103.1117 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,280B, BPFP=0.4932 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,012B, BPFP=2.4454 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,144B, BPFP=1.3363 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,796B, BPFP=2.3807 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,604B, BPFP=1.5733 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,148B, BPFP=2.3463 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,840B, BPFP=1.4264 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,236B, BPFP=2.4041 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,564B, BPFP=2.1027 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,716B, BPFP=2.3233 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,464B, BPFP=0.3376 +⌛️ [2/4] FRONTEND: Frontend time: 2.831s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.546s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16298727 41.32295320 + layer.0.v_cache 0.00001770 0.00675896 + layer.1.k_cache 0.61736183 8.66258437 + layer.1.v_cache 0.00000612 0.00266198 + layer.2.k_cache 0.02593859 0.89516911 + layer.2.v_cache 0.00002024 0.00750299 + layer.3.k_cache 0.01326758 6.29899047 + layer.3.v_cache 0.00002270 0.00861027 + layer.4.k_cache 0.00075427 0.22662094 + layer.4.v_cache 0.00005043 0.01467412 + layer.4.output 0.04538776 178.21855260 + ------------------------------------------------------------------------------------- + TOTAL 0.06694947 76.76331733 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 398804 +BPFP 1.2468 bits/point +EBPFP 2.4935 equivalent bits/point +MSE 76.763317 +---------------------- -------------------------------------------------------- +Time: 4.387s Load: 0.011s, Pack+Encode: 2.831s, Decode+Unpack: 1.546s +---------------------- -------------------------------------------------------- +💾 Converting with 76.7633 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 239, 128) +Output shape: (1, 239, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.output: torch.Size([1, 239, 3584]) -> torch.Size([1, 1, 239, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,640B, BPFP=0.4995 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,792B, BPFP=2.4053 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,572B, BPFP=1.2796 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,892B, BPFP=2.3465 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,228B, BPFP=1.5839 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,576B, BPFP=2.3258 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,576B, BPFP=1.4106 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,232B, BPFP=2.3687 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,784B, BPFP=2.0779 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,076B, BPFP=2.2931 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,444B, BPFP=0.3590 +⌛️ [2/4] FRONTEND: Frontend time: 2.110s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.530s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11863919 39.89238608 + layer.0.v_cache 0.00001702 0.00696476 + layer.1.k_cache 0.43747612 8.51277754 + layer.1.v_cache 0.00000797 0.00283120 + layer.2.k_cache 0.02286501 0.80816510 + layer.2.v_cache 0.00002174 0.00745194 + layer.3.k_cache 0.01903599 5.96094618 + layer.3.v_cache 0.00002123 0.00862541 + layer.4.k_cache 0.00068597 0.22791274 + layer.4.v_cache 0.00006921 0.01536842 + layer.4.output 1.28101462 223.52121937 + ------------------------------------------------------------------------------------- + TOTAL 0.56270246 95.29952735 + (elements=2,080,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2080256 +Total Bytes 322812 +BPFP 1.2414 bits/point +EBPFP 2.4829 equivalent bits/point +MSE 95.299527 +---------------------- -------------------------------------------------------- +Time: 3.648s Load: 0.009s, Pack+Encode: 2.110s, Decode+Unpack: 1.530s +---------------------- -------------------------------------------------------- +💾 Converting with 95.2995 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,360B, BPFP=0.4967 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,060B, BPFP=2.6176 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,952B, BPFP=1.3042 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,400B, BPFP=2.5784 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,944B, BPFP=1.4225 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,672B, BPFP=2.5352 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,976B, BPFP=1.3650 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,452B, BPFP=2.5815 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,388B, BPFP=2.2212 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,048B, BPFP=2.4981 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,184B, BPFP=0.3580 +⌛️ [2/4] FRONTEND: Frontend time: 2.209s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.590s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15151563 41.88047321 + layer.0.v_cache 0.00001631 0.00669854 + layer.1.k_cache 0.46755480 8.64402069 + layer.1.v_cache 0.00000667 0.00267109 + layer.2.k_cache 0.02171374 0.80931700 + layer.2.v_cache 0.00002115 0.00745751 + layer.3.k_cache 0.01385514 5.52556422 + layer.3.v_cache 0.00002107 0.00851471 + layer.4.k_cache 0.00069081 0.21920158 + layer.4.v_cache 0.00005638 0.01511590 + layer.4.output 0.00504555 202.36800652 + ------------------------------------------------------------------------------------- + TOTAL 0.04063356 86.68794589 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 372436 +BPFP 1.3016 bits/point +EBPFP 2.6031 equivalent bits/point +MSE 86.687946 +---------------------- -------------------------------------------------------- +Time: 3.810s Load: 0.011s, Pack+Encode: 2.209s, Decode+Unpack: 1.590s +---------------------- -------------------------------------------------------- +💾 Converting with 86.6879 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,996B, BPFP=0.4858 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,456B, BPFP=2.5317 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,976B, BPFP=1.3872 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,596B, BPFP=2.4719 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,268B, BPFP=1.6158 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,184B, BPFP=2.4433 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,492B, BPFP=1.4231 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,808B, BPFP=2.4867 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,364B, BPFP=2.1781 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,852B, BPFP=2.4203 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,840B, BPFP=0.3456 +⌛️ [2/4] FRONTEND: Frontend time: 2.264s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.504s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14529029 41.94795573 + layer.0.v_cache 0.00001764 0.00703258 + layer.1.k_cache 0.31089128 8.80428982 + layer.1.v_cache 0.00000630 0.00273554 + layer.2.k_cache 0.01355181 0.84432380 + layer.2.v_cache 0.00002032 0.00786556 + layer.3.k_cache 0.02426839 6.53002713 + layer.3.v_cache 0.00002043 0.00857333 + layer.4.k_cache 0.00068925 0.21602298 + layer.4.v_cache 0.00005165 0.01537228 + layer.4.output 1.36066019 234.21595238 + ------------------------------------------------------------------------------------- + TOTAL 0.58937816 99.87622738 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 314832 +BPFP 1.2861 bits/point +EBPFP 2.5722 equivalent bits/point +MSE 99.876227 +---------------------- -------------------------------------------------------- +Time: 3.778s Load: 0.011s, Pack+Encode: 2.264s, Decode+Unpack: 1.504s +---------------------- -------------------------------------------------------- +💾 Converting with 99.8762 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 262, 128) +Output shape: (1, 262, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.output: torch.Size([1, 262, 3584]) -> torch.Size([1, 1, 262, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,340B, BPFP=0.4974 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,396B, BPFP=2.6477 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,612B, BPFP=1.2889 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,304B, BPFP=2.5825 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,964B, BPFP=1.4292 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,488B, BPFP=2.5339 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,136B, BPFP=1.3798 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,776B, BPFP=2.6107 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,792B, BPFP=2.2538 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,168B, BPFP=2.5148 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,844B, BPFP=0.3565 +⌛️ [2/4] FRONTEND: Frontend time: 2.090s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.630s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12917162 41.65759184 + layer.0.v_cache 0.00001830 0.00673069 + layer.1.k_cache 0.40634074 8.57022211 + layer.1.v_cache 0.00000633 0.00264200 + layer.2.k_cache 0.02363680 0.82850979 + layer.2.v_cache 0.00002248 0.00759100 + layer.3.k_cache 0.02343226 5.24761171 + layer.3.v_cache 0.00002042 0.00825783 + layer.4.k_cache 0.00071170 0.22658778 + layer.4.v_cache 0.00005093 0.01523536 + layer.4.output 0.00505940 201.07103667 + ------------------------------------------------------------------------------------- + TOTAL 0.03640161 86.12166099 + (elements=2,280,448) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2280448 +Total Bytes 372820 +BPFP 1.3079 bits/point +EBPFP 2.6158 equivalent bits/point +MSE 86.121661 +---------------------- -------------------------------------------------------- +Time: 3.733s Load: 0.013s, Pack+Encode: 2.090s, Decode+Unpack: 1.630s +---------------------- -------------------------------------------------------- +💾 Converting with 86.1217 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,024B, BPFP=0.4878 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,552B, BPFP=2.5383 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,640B, BPFP=1.3639 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,660B, BPFP=2.4764 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,552B, BPFP=1.6356 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,220B, BPFP=2.4458 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,984B, BPFP=1.4572 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,920B, BPFP=2.4944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,716B, BPFP=2.2025 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,032B, BPFP=2.4328 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,616B, BPFP=0.3533 +⌛️ [2/4] FRONTEND: Frontend time: 2.079s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.481s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18074451 42.09462240 + layer.0.v_cache 0.00001764 0.00700689 + layer.1.k_cache 0.33421082 9.00691515 + layer.1.v_cache 0.00000655 0.00272141 + layer.2.k_cache 0.01487562 1.03325819 + layer.2.v_cache 0.00002082 0.00793043 + layer.3.k_cache 0.01319740 6.19849175 + layer.3.v_cache 0.00002096 0.00887429 + layer.4.k_cache 0.00067848 0.23117059 + layer.4.v_cache 0.00005239 0.01603365 + layer.4.output 1.36066546 234.66438492 + ------------------------------------------------------------------------------------- + TOTAL 0.59226373 100.07398348 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 316916 +BPFP 1.2946 bits/point +EBPFP 2.5892 equivalent bits/point +MSE 100.073983 +---------------------- -------------------------------------------------------- +Time: 3.567s Load: 0.008s, Pack+Encode: 2.079s, Decode+Unpack: 1.481s +---------------------- -------------------------------------------------------- +💾 Converting with 100.0740 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 232, 128) +Output shape: (1, 232, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.output: torch.Size([1, 232, 3584]) -> torch.Size([1, 1, 232, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,464B, BPFP=0.5027 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,692B, BPFP=2.4712 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,088B, BPFP=1.2856 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,640B, BPFP=2.4003 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,844B, BPFP=1.6059 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,576B, BPFP=2.3960 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,560B, BPFP=1.4520 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,012B, BPFP=2.4254 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,532B, BPFP=2.1237 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,256B, BPFP=2.3745 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,216B, BPFP=0.3677 +⌛️ [2/4] FRONTEND: Frontend time: 2.017s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11978169 42.15074000 + layer.0.v_cache 0.00001746 0.00694417 + layer.1.k_cache 0.32299256 8.90118829 + layer.1.v_cache 0.00000596 0.00262158 + layer.2.k_cache 0.02091339 0.91456052 + layer.2.v_cache 0.00002092 0.00760626 + layer.3.k_cache 0.01769928 6.82727103 + layer.3.v_cache 0.00002180 0.00883232 + layer.4.k_cache 0.00071208 0.23010793 + layer.4.v_cache 0.00005763 0.01551773 + layer.4.output 1.31963615 224.96243842 + ------------------------------------------------------------------------------------- + TOTAL 0.57174564 96.10602699 + (elements=2,019,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2019328 +Total Bytes 320880 +BPFP 1.2712 bits/point +EBPFP 2.5425 equivalent bits/point +MSE 96.106027 +---------------------- -------------------------------------------------------- +Time: 3.399s Load: 0.010s, Pack+Encode: 2.017s, Decode+Unpack: 1.373s +---------------------- -------------------------------------------------------- +💾 Converting with 96.1060 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,988B, BPFP=0.4963 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,708B, BPFP=2.6071 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,508B, BPFP=1.3145 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,696B, BPFP=2.5352 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,640B, BPFP=1.5369 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,128B, BPFP=2.4949 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,592B, BPFP=1.4625 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,016B, BPFP=2.5580 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,772B, BPFP=2.2565 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,252B, BPFP=2.5037 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,288B, BPFP=0.3986 +⌛️ [2/4] FRONTEND: Frontend time: 2.134s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.406s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15639258 40.89597834 + layer.0.v_cache 0.00001700 0.00706690 + layer.1.k_cache 0.33150226 9.20855602 + layer.1.v_cache 0.00000613 0.00275162 + layer.2.k_cache 0.01694729 0.91393842 + layer.2.v_cache 0.00002208 0.00762091 + layer.3.k_cache 0.03858858 5.28467685 + layer.3.v_cache 0.00002090 0.00900214 + layer.4.k_cache 0.00067984 0.22535491 + layer.4.v_cache 0.00005217 0.01573838 + layer.4.output 1.39161012 243.60825893 + ------------------------------------------------------------------------------------- + TOTAL 0.60502939 103.63697041 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 317588 +BPFP 1.3268 bits/point +EBPFP 2.6536 equivalent bits/point +MSE 103.636970 +---------------------- -------------------------------------------------------- +Time: 3.549s Load: 0.009s, Pack+Encode: 2.134s, Decode+Unpack: 1.406s +---------------------- -------------------------------------------------------- +💾 Converting with 103.6370 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,588B, BPFP=0.4915 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,296B, BPFP=2.5925 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,544B, BPFP=1.2331 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,068B, BPFP=2.5222 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,540B, BPFP=1.4618 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,628B, BPFP=2.4970 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,812B, BPFP=1.4201 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,308B, BPFP=2.5359 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,440B, BPFP=2.2001 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,324B, BPFP=2.4796 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,632B, BPFP=0.3649 +⌛️ [2/4] FRONTEND: Frontend time: 2.199s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.582s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13605902 40.62735377 + layer.0.v_cache 0.00001757 0.00664669 + layer.1.k_cache 0.58683352 8.72564407 + layer.1.v_cache 0.00000623 0.00267381 + layer.2.k_cache 0.01500385 0.75775806 + layer.2.v_cache 0.00002229 0.00774567 + layer.3.k_cache 0.00978726 5.20806281 + layer.3.v_cache 0.00001998 0.00827618 + layer.4.k_cache 0.00067570 0.24035250 + layer.4.v_cache 0.00005472 0.01534099 + layer.4.output 0.00490606 191.99630429 + ------------------------------------------------------------------------------------- + TOTAL 0.04604839 82.32788145 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 384180 +BPFP 1.2934 bits/point +EBPFP 2.5869 equivalent bits/point +MSE 82.327881 +---------------------- -------------------------------------------------------- +Time: 3.790s Load: 0.009s, Pack+Encode: 2.199s, Decode+Unpack: 1.582s +---------------------- -------------------------------------------------------- +💾 Converting with 82.3279 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,756B, BPFP=0.4752 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,156B, BPFP=2.2767 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,704B, BPFP=1.3299 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,212B, BPFP=2.2189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,232B, BPFP=1.4235 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,664B, BPFP=2.1853 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,448B, BPFP=1.4368 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,700B, BPFP=2.2488 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,580B, BPFP=1.9350 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,388B, BPFP=2.1684 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,804B, BPFP=0.3397 +⌛️ [2/4] FRONTEND: Frontend time: 1.988s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.558s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12967140 40.70065487 + layer.0.v_cache 0.00001601 0.00646068 + layer.1.k_cache 0.49353925 10.07780810 + layer.1.v_cache 0.00000620 0.00260330 + layer.2.k_cache 0.03189284 0.70867884 + layer.2.v_cache 0.00002040 0.00740314 + layer.3.k_cache 0.04285583 5.97262561 + layer.3.v_cache 0.00002221 0.00840211 + layer.4.k_cache 0.00071132 0.23433427 + layer.4.v_cache 0.00005107 0.01426419 + layer.4.output 1.20063613 204.21955532 + ------------------------------------------------------------------------------------- + TOTAL 0.53548467 87.48647779 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 327644 +BPFP 1.1810 bits/point +EBPFP 2.3619 equivalent bits/point +MSE 87.486478 +---------------------- -------------------------------------------------------- +Time: 3.554s Load: 0.008s, Pack+Encode: 1.988s, Decode+Unpack: 1.558s +---------------------- -------------------------------------------------------- +💾 Converting with 87.4865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,676B, BPFP=0.4817 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,824B, BPFP=2.3107 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,044B, BPFP=1.3205 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,924B, BPFP=2.2543 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,712B, BPFP=1.5507 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,432B, BPFP=2.2234 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,460B, BPFP=1.4094 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,116B, BPFP=2.2663 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,836B, BPFP=1.9977 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,960B, BPFP=2.1938 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,796B, BPFP=0.3030 +⌛️ [2/4] FRONTEND: Frontend time: 2.120s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.578s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16529650 39.83904760 + layer.0.v_cache 0.00001613 0.00678410 + layer.1.k_cache 0.54577502 9.16082102 + layer.1.v_cache 0.00000662 0.00271375 + layer.2.k_cache 0.01453520 0.86003511 + layer.2.v_cache 0.00002096 0.00776774 + layer.3.k_cache 0.02409829 5.75895525 + layer.3.v_cache 0.00002071 0.00881297 + layer.4.k_cache 0.00068699 0.24415996 + layer.4.v_cache 0.00004971 0.01554309 + layer.4.output 1.22953407 209.45611015 + ------------------------------------------------------------------------------------- + TOTAL 0.55042615 89.53514186 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 320780 +BPFP 1.1841 bits/point +EBPFP 2.3681 equivalent bits/point +MSE 89.535142 +---------------------- -------------------------------------------------------- +Time: 3.706s Load: 0.008s, Pack+Encode: 2.120s, Decode+Unpack: 1.578s +---------------------- -------------------------------------------------------- +💾 Converting with 89.5351 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,324B, BPFP=0.4933 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 53,696B, BPFP=2.5657 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,064B, BPFP=1.2932 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 52,596B, BPFP=2.5132 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,692B, BPFP=1.4666 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 51,840B, BPFP=2.4771 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,140B, BPFP=1.3924 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 52,372B, BPFP=2.5025 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,356B, BPFP=2.1672 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 51,364B, BPFP=2.4543 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 50,928B, BPFP=0.3476 +⌛️ [2/4] FRONTEND: Frontend time: 2.767s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.786s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19043710 40.16835699 + layer.0.v_cache 0.00001603 0.00669767 + layer.1.k_cache 0.64205214 8.64585051 + layer.1.v_cache 0.00000630 0.00270016 + layer.2.k_cache 0.01513053 0.79551487 + layer.2.v_cache 0.00002191 0.00772105 + layer.3.k_cache 0.02775778 5.05909808 + layer.3.v_cache 0.00002125 0.00848727 + layer.4.k_cache 0.00071623 0.24648332 + layer.4.v_cache 0.00005146 0.01476038 + layer.4.output 0.04089398 159.31678681 + ------------------------------------------------------------------------------------- + TOTAL 0.06838051 68.83371635 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 455372 +BPFP 1.2799 bits/point +EBPFP 2.5599 equivalent bits/point +MSE 68.833716 +---------------------- -------------------------------------------------------- +Time: 4.565s Load: 0.012s, Pack+Encode: 2.767s, Decode+Unpack: 1.786s +---------------------- -------------------------------------------------------- +💾 Converting with 68.8337 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,804B, BPFP=0.4861 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,932B, BPFP=2.5360 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,448B, BPFP=1.2946 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,856B, BPFP=2.4766 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,416B, BPFP=1.5689 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,280B, BPFP=2.4448 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,288B, BPFP=1.4514 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,984B, BPFP=2.4837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,448B, BPFP=2.1780 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,892B, BPFP=2.4234 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,100B, BPFP=0.3478 +⌛️ [2/4] FRONTEND: Frontend time: 2.065s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 28.598s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14392723 40.71369465 + layer.0.v_cache 0.00001682 0.00687159 + layer.1.k_cache 0.53349789 8.80640425 + layer.1.v_cache 0.00000638 0.00270916 + layer.2.k_cache 0.01523476 0.76494833 + layer.2.v_cache 0.00002109 0.00759960 + layer.3.k_cache 0.03867486 5.86759167 + layer.3.v_cache 0.00001994 0.00847181 + layer.4.k_cache 0.00070525 0.22043038 + layer.4.v_cache 0.00005078 0.01442261 + layer.4.output 0.00472849 193.05913995 + ------------------------------------------------------------------------------------- + TOTAL 0.04501497 82.81336022 + (elements=2,463,232) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2463232 +Total Bytes 394448 +BPFP 1.2811 bits/point +EBPFP 2.5621 equivalent bits/point +MSE 82.813360 +---------------------- --------------------------------------------------------- +Time: 30.672s Load: 0.009s, Pack+Encode: 2.065s, Decode+Unpack: 28.598s +---------------------- --------------------------------------------------------- +💾 Converting with 82.8134 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,992B, BPFP=0.4856 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,332B, BPFP=2.5231 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,320B, BPFP=1.3417 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,572B, BPFP=2.4703 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,948B, BPFP=1.6631 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,144B, BPFP=2.4406 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,800B, BPFP=1.4444 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,844B, BPFP=2.4892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,696B, BPFP=2.2011 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,724B, BPFP=2.4114 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,336B, BPFP=0.3307 +⌛️ [2/4] FRONTEND: Frontend time: 2.386s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.393s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14673459 41.68026476 + layer.0.v_cache 0.00001655 0.00692236 + layer.1.k_cache 0.37581726 8.91157552 + layer.1.v_cache 0.00000603 0.00270446 + layer.2.k_cache 0.01062915 0.92658169 + layer.2.v_cache 0.00002006 0.00802213 + layer.3.k_cache 0.04027582 5.90467285 + layer.3.v_cache 0.00002066 0.00904269 + layer.4.k_cache 0.00066373 0.22436273 + layer.4.v_cache 0.00004861 0.01588241 + layer.4.output 1.36064577 236.27091270 + ------------------------------------------------------------------------------------- + TOTAL 0.59404429 100.68155415 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 313708 +BPFP 1.2815 bits/point +EBPFP 2.5630 equivalent bits/point +MSE 100.681554 +---------------------- -------------------------------------------------------- +Time: 3.790s Load: 0.011s, Pack+Encode: 2.386s, Decode+Unpack: 1.393s +---------------------- -------------------------------------------------------- +💾 Converting with 100.6816 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,540B, BPFP=0.5035 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,060B, BPFP=2.4746 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,904B, BPFP=1.3291 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,960B, BPFP=2.4012 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,452B, BPFP=1.6327 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,568B, BPFP=2.3750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,196B, BPFP=1.4821 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,252B, BPFP=2.4207 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,908B, BPFP=2.1306 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,040B, BPFP=2.3397 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,244B, BPFP=0.3744 +⌛️ [2/4] FRONTEND: Frontend time: 2.035s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.453s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12385076 40.95820396 + layer.0.v_cache 0.00002004 0.00743283 + layer.1.k_cache 0.38583452 8.40534334 + layer.1.v_cache 0.00000646 0.00285433 + layer.2.k_cache 0.02500911 0.77622999 + layer.2.v_cache 0.00002014 0.00768034 + layer.3.k_cache 0.03425702 6.36547278 + layer.3.v_cache 0.00002057 0.00903320 + layer.4.k_cache 0.00068928 0.24081908 + layer.4.v_cache 0.00004882 0.01580128 + layer.4.output 1.30836928 224.25129731 + ------------------------------------------------------------------------------------- + TOTAL 0.57225539 95.67929131 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 325124 +BPFP 1.2770 bits/point +EBPFP 2.5541 equivalent bits/point +MSE 95.679291 +---------------------- -------------------------------------------------------- +Time: 3.495s Load: 0.008s, Pack+Encode: 2.035s, Decode+Unpack: 1.453s +---------------------- -------------------------------------------------------- +💾 Converting with 95.6793 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 247, 128) +Output shape: (1, 247, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.output: torch.Size([1, 247, 3584]) -> torch.Size([1, 1, 247, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,748B, BPFP=0.4901 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,112B, BPFP=2.3477 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,336B, BPFP=1.3497 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,056B, BPFP=2.2809 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,856B, BPFP=1.5091 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,488B, BPFP=2.2449 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,328B, BPFP=1.4124 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,144B, BPFP=2.2864 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,456B, BPFP=1.9899 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,060B, BPFP=2.2179 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,420B, BPFP=0.3291 +⌛️ [2/4] FRONTEND: Frontend time: 2.104s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.448s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13907220 41.20090698 + layer.0.v_cache 0.00001735 0.00677899 + layer.1.k_cache 0.47208408 8.96008350 + layer.1.v_cache 0.00000611 0.00261328 + layer.2.k_cache 0.03408500 0.98671114 + layer.2.v_cache 0.00001989 0.00717334 + layer.3.k_cache 0.02387486 6.08208710 + layer.3.v_cache 0.00002031 0.00831316 + layer.4.k_cache 0.00069623 0.21479280 + layer.4.v_cache 0.00004922 0.01410366 + layer.4.output 1.23951819 209.11988505 + ------------------------------------------------------------------------------------- + TOTAL 0.54979721 89.48957408 + (elements=2,149,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2149888 +Total Bytes 323004 +BPFP 1.2019 bits/point +EBPFP 2.4039 equivalent bits/point +MSE 89.489574 +---------------------- -------------------------------------------------------- +Time: 3.560s Load: 0.008s, Pack+Encode: 2.104s, Decode+Unpack: 1.448s +---------------------- -------------------------------------------------------- +💾 Converting with 89.4896 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 427, 128) +Output shape: (1, 427, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.output: torch.Size([1, 427, 3584]) -> torch.Size([1, 1, 427, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,268B, BPFP=0.4855 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 64,104B, BPFP=2.3457 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 35,428B, BPFP=1.2964 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 62,644B, BPFP=2.2923 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 40,648B, BPFP=1.4874 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 61,532B, BPFP=2.2516 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 36,812B, BPFP=1.3470 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 62,664B, BPFP=2.2930 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 54,076B, BPFP=1.9788 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 60,960B, BPFP=2.2307 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,604B, BPFP=0.3429 +⌛️ [2/4] FRONTEND: Frontend time: 2.913s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.762s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15738957 39.49866437 + layer.0.v_cache 0.00001693 0.00627546 + layer.1.k_cache 1.00627419 9.01679024 + layer.1.v_cache 0.00000663 0.00249717 + layer.2.k_cache 0.03590691 0.74233830 + layer.2.v_cache 0.00002108 0.00658225 + layer.3.k_cache 0.02176077 6.06270297 + layer.3.v_cache 0.00002004 0.00741476 + layer.4.k_cache 0.00078997 0.20665603 + layer.4.v_cache 0.00005212 0.01365327 + layer.4.output 0.00621442 127.06948394 + ------------------------------------------------------------------------------------- + TOTAL 0.07445524 55.59117426 + (elements=3,716,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3716608 +Total Bytes 557740 +BPFP 1.2005 bits/point +EBPFP 2.4011 equivalent bits/point +MSE 55.591174 +---------------------- -------------------------------------------------------- +Time: 4.689s Load: 0.013s, Pack+Encode: 2.913s, Decode+Unpack: 1.762s +---------------------- -------------------------------------------------------- +💾 Converting with 55.5912 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,612B, BPFP=0.4965 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,612B, BPFP=2.5722 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,412B, BPFP=1.2345 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,688B, BPFP=2.5189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,096B, BPFP=1.4470 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,828B, BPFP=2.4693 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,500B, BPFP=1.4126 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,892B, BPFP=2.5307 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,484B, BPFP=2.2189 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,652B, BPFP=2.4592 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,256B, BPFP=0.3233 +⌛️ [2/4] FRONTEND: Frontend time: 2.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.604s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15504150 40.73388492 + layer.0.v_cache 0.00001776 0.00647046 + layer.1.k_cache 0.59434689 8.28732108 + layer.1.v_cache 0.00000612 0.00252821 + layer.2.k_cache 0.01687181 0.83631340 + layer.2.v_cache 0.00002031 0.00720780 + layer.3.k_cache 0.01346795 5.53269998 + layer.3.v_cache 0.00002023 0.00800541 + layer.4.k_cache 0.00071492 0.21804948 + layer.4.v_cache 0.00006004 0.01411099 + layer.4.output 0.00488892 200.45028334 + ------------------------------------------------------------------------------------- + TOTAL 0.04792882 85.81168089 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 375032 +BPFP 1.2720 bits/point +EBPFP 2.5439 equivalent bits/point +MSE 85.811681 +---------------------- -------------------------------------------------------- +Time: 3.768s Load: 0.009s, Pack+Encode: 2.155s, Decode+Unpack: 1.604s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8117 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 361, 128) +Output shape: (1, 361, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.output: torch.Size([1, 361, 3584]) -> torch.Size([1, 1, 361, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,200B, BPFP=0.4848 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 54,648B, BPFP=2.3653 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,336B, BPFP=1.3130 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,284B, BPFP=2.3063 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 34,720B, BPFP=1.5028 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,584B, BPFP=2.2760 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,844B, BPFP=1.3783 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,592B, BPFP=2.3196 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,656B, BPFP=2.0194 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 51,972B, BPFP=2.2495 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 46,268B, BPFP=0.2861 +⌛️ [2/4] FRONTEND: Frontend time: 2.296s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.681s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21057716 38.92264868 + layer.0.v_cache 0.00001554 0.00597834 + layer.1.k_cache 0.82271198 8.83551482 + layer.1.v_cache 0.00000592 0.00239000 + layer.2.k_cache 0.01184860 1.01462198 + layer.2.v_cache 0.00001987 0.00703634 + layer.3.k_cache 0.01345186 6.15514224 + layer.3.v_cache 0.00002111 0.00755296 + layer.4.k_cache 0.00076976 0.22720865 + layer.4.v_cache 0.00005077 0.01398171 + layer.4.output 0.03703259 147.02788633 + ------------------------------------------------------------------------------------- + TOTAL 0.07757063 63.78748706 + (elements=3,142,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3142144 +Total Bytes 467104 +BPFP 1.1893 bits/point +EBPFP 2.3785 equivalent bits/point +MSE 63.787487 +---------------------- -------------------------------------------------------- +Time: 3.989s Load: 0.011s, Pack+Encode: 2.296s, Decode+Unpack: 1.681s +---------------------- -------------------------------------------------------- +💾 Converting with 63.7875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,856B, BPFP=0.4907 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,424B, BPFP=2.5168 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,984B, BPFP=1.2735 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,440B, BPFP=2.4623 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,924B, BPFP=1.4364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,704B, BPFP=2.4215 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,448B, BPFP=1.3546 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,292B, BPFP=2.4541 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,048B, BPFP=2.1636 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,260B, BPFP=2.3969 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,604B, BPFP=0.3135 +⌛️ [2/4] FRONTEND: Frontend time: 2.139s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.754s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21757852 39.87270404 + layer.0.v_cache 0.00001544 0.00577064 + layer.1.k_cache 0.59705023 8.54855368 + layer.1.v_cache 0.00000603 0.00235709 + layer.2.k_cache 0.00787334 0.85505579 + layer.2.v_cache 0.00001970 0.00707345 + layer.3.k_cache 0.01914395 5.12722973 + layer.3.v_cache 0.00001916 0.00733227 + layer.4.k_cache 0.00071861 0.21296900 + layer.4.v_cache 0.00005003 0.01396797 + layer.4.output 0.00471288 191.18770580 + ------------------------------------------------------------------------------------- + TOTAL 0.05149795 81.93923260 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 381984 +BPFP 1.2450 bits/point +EBPFP 2.4900 equivalent bits/point +MSE 81.939233 +---------------------- -------------------------------------------------------- +Time: 3.903s Load: 0.010s, Pack+Encode: 2.139s, Decode+Unpack: 1.754s +---------------------- -------------------------------------------------------- +💾 Converting with 81.9392 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 372, 128) +Output shape: (1, 372, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.output: torch.Size([1, 372, 3584]) -> torch.Size([1, 1, 372, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,408B, BPFP=0.4792 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,156B, BPFP=2.3167 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,104B, BPFP=1.3065 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,920B, BPFP=2.2648 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 35,580B, BPFP=1.4945 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,884B, BPFP=2.2213 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 33,536B, BPFP=1.4086 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,968B, BPFP=2.2668 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,684B, BPFP=1.9609 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,412B, BPFP=2.2014 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 58,748B, BPFP=0.3525 +⌛️ [2/4] FRONTEND: Frontend time: 2.201s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.937s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18192377 39.68880208 + layer.0.v_cache 0.00001786 0.00622527 + layer.1.k_cache 0.82695811 8.75096081 + layer.1.v_cache 0.00000687 0.00250052 + layer.2.k_cache 0.01216373 0.84776741 + layer.2.v_cache 0.00002073 0.00691855 + layer.3.k_cache 0.01743993 6.03525076 + layer.3.v_cache 0.00002103 0.00755359 + layer.4.k_cache 0.00069676 0.20987593 + layer.4.v_cache 0.00005544 0.01398238 + layer.4.output 0.03606073 143.99780386 + ------------------------------------------------------------------------------------- + TOTAL 0.07598408 62.56202731 + (elements=3,237,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3237888 +Total Bytes 485400 +BPFP 1.1993 bits/point +EBPFP 2.3986 equivalent bits/point +MSE 62.562027 +---------------------- -------------------------------------------------------- +Time: 4.150s Load: 0.012s, Pack+Encode: 2.201s, Decode+Unpack: 1.937s +---------------------- -------------------------------------------------------- +💾 Converting with 62.5620 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,956B, BPFP=0.4816 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 52,620B, BPFP=2.5455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,844B, BPFP=1.2986 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 51,232B, BPFP=2.4783 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,632B, BPFP=1.3851 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 51,252B, BPFP=2.4793 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,252B, BPFP=1.3667 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 51,588B, BPFP=2.4955 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,292B, BPFP=2.0459 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 49,544B, BPFP=2.3967 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 48,904B, BPFP=0.3380 +⌛️ [2/4] FRONTEND: Frontend time: 2.388s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.096s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16470559 40.30296415 + layer.0.v_cache 0.00001558 0.00630257 + layer.1.k_cache 0.70577214 8.55000197 + layer.1.v_cache 0.00000663 0.00253498 + layer.2.k_cache 0.01487649 0.88578574 + layer.2.v_cache 0.00002061 0.00717188 + layer.3.k_cache 0.03012228 5.14677916 + layer.3.v_cache 0.00002030 0.00782637 + layer.4.k_cache 0.00074575 0.21642707 + layer.4.v_cache 0.00005065 0.01350714 + layer.4.output 0.04139163 163.99729102 + ------------------------------------------------------------------------------------- + TOTAL 0.07094573 70.77178460 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 441116 +BPFP 1.2552 bits/point +EBPFP 2.5104 equivalent bits/point +MSE 70.771785 +---------------------- -------------------------------------------------------- +Time: 4.498s Load: 0.014s, Pack+Encode: 2.388s, Decode+Unpack: 2.096s +---------------------- -------------------------------------------------------- +💾 Converting with 70.7718 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,412B, BPFP=0.4960 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,340B, BPFP=2.6144 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,424B, BPFP=1.2632 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,888B, BPFP=2.5288 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,564B, BPFP=1.4483 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,264B, BPFP=2.5509 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,844B, BPFP=1.3469 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,568B, BPFP=2.5689 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,080B, BPFP=2.2453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,116B, BPFP=2.4833 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,728B, BPFP=0.3599 +⌛️ [2/4] FRONTEND: Frontend time: 2.273s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.726s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16585089 40.26142394 + layer.0.v_cache 0.00001549 0.00628979 + layer.1.k_cache 0.52552974 8.22906287 + layer.1.v_cache 0.00000676 0.00252341 + layer.2.k_cache 0.01473994 0.86167873 + layer.2.v_cache 0.00002188 0.00743803 + layer.3.k_cache 0.05129043 5.97872982 + layer.3.v_cache 0.00002012 0.00785520 + layer.4.k_cache 0.00075129 0.23428227 + layer.4.v_cache 0.00005717 0.01448742 + layer.4.output 0.00501330 205.90096024 + ------------------------------------------------------------------------------------- + TOTAL 0.04666922 88.05355842 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 374228 +BPFP 1.2980 bits/point +EBPFP 2.5959 equivalent bits/point +MSE 88.053558 +---------------------- -------------------------------------------------------- +Time: 4.010s Load: 0.010s, Pack+Encode: 2.273s, Decode+Unpack: 1.726s +---------------------- -------------------------------------------------------- +💾 Converting with 88.0536 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 368, 128) +Output shape: (1, 368, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.output: torch.Size([1, 368, 3584]) -> torch.Size([1, 1, 368, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,408B, BPFP=0.4844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,144B, BPFP=2.3414 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,484B, BPFP=1.2943 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,816B, BPFP=2.2850 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 35,516B, BPFP=1.5080 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,852B, BPFP=2.2441 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 32,956B, BPFP=1.3993 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,904B, BPFP=2.2887 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,480B, BPFP=1.9735 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,432B, BPFP=2.2262 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 53,656B, BPFP=0.3255 +⌛️ [2/4] FRONTEND: Frontend time: 2.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.827s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14147143 41.17792544 + layer.0.v_cache 0.00001611 0.00603792 + layer.1.k_cache 0.89062865 9.14029130 + layer.1.v_cache 0.00000637 0.00238017 + layer.2.k_cache 0.02201829 0.81626511 + layer.2.v_cache 0.00002088 0.00687455 + layer.3.k_cache 0.00965744 6.56760307 + layer.3.v_cache 0.00002101 0.00757697 + layer.4.k_cache 0.00077426 0.19988417 + layer.4.v_cache 0.00005366 0.01402033 + layer.4.output 0.03638816 144.83110928 + ------------------------------------------------------------------------------------- + TOTAL 0.07761090 63.04450729 + (elements=3,203,072) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3203072 +Total Bytes 478648 +BPFP 1.1955 bits/point +EBPFP 2.3909 equivalent bits/point +MSE 63.044507 +---------------------- -------------------------------------------------------- +Time: 4.099s Load: 0.012s, Pack+Encode: 2.260s, Decode+Unpack: 1.827s +---------------------- -------------------------------------------------------- +💾 Converting with 63.0445 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 453, 128) +Output shape: (1, 453, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.output: torch.Size([1, 453, 3584]) -> torch.Size([1, 1, 453, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,848B, BPFP=0.4776 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 71,032B, BPFP=2.4501 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 36,564B, BPFP=1.2612 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 69,192B, BPFP=2.3866 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 39,764B, BPFP=1.3716 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 68,156B, BPFP=2.3509 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 39,464B, BPFP=1.3612 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 69,484B, BPFP=2.3967 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 59,428B, BPFP=2.0498 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 67,724B, BPFP=2.3360 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,740B, BPFP=0.3141 +⌛️ [2/4] FRONTEND: Frontend time: 2.764s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.965s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13665763 39.48002897 + layer.0.v_cache 0.00001897 0.00625991 + layer.1.k_cache 1.07282356 8.86942842 + layer.1.v_cache 0.00000630 0.00242208 + layer.2.k_cache 0.02241544 0.75490700 + layer.2.v_cache 0.00001979 0.00682192 + layer.3.k_cache 0.01806638 6.53025565 + layer.3.v_cache 0.00002079 0.00781680 + layer.4.k_cache 0.00076567 0.21461003 + layer.4.v_cache 0.00005104 0.01370100 + layer.4.output 0.00584071 114.98013245 + ------------------------------------------------------------------------------------- + TOTAL 0.07598415 50.63218700 + (elements=3,942,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3942912 +Total Bytes 598396 +BPFP 1.2141 bits/point +EBPFP 2.4282 equivalent bits/point +MSE 50.632187 +---------------------- -------------------------------------------------------- +Time: 4.745s Load: 0.016s, Pack+Encode: 2.764s, Decode+Unpack: 1.965s +---------------------- -------------------------------------------------------- +💾 Converting with 50.6322 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 402, 128) +Output shape: (1, 402, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.output: torch.Size([1, 402, 3584]) -> torch.Size([1, 1, 402, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,500B, BPFP=0.4859 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 63,636B, BPFP=2.4734 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 32,504B, BPFP=1.2634 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 61,832B, BPFP=2.4033 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 36,784B, BPFP=1.4297 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 60,772B, BPFP=2.3621 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 34,404B, BPFP=1.3372 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 62,036B, BPFP=2.4112 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 53,128B, BPFP=2.0650 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 60,072B, BPFP=2.3349 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 53,624B, BPFP=0.2978 +⌛️ [2/4] FRONTEND: Frontend time: 2.402s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18062694 40.85374883 + layer.0.v_cache 0.00001752 0.00612939 + layer.1.k_cache 0.89169403 8.84942232 + layer.1.v_cache 0.00000593 0.00227562 + layer.2.k_cache 0.03854401 0.73946673 + layer.2.v_cache 0.00001967 0.00672424 + layer.3.k_cache 0.02336847 5.95436757 + layer.3.v_cache 0.00001924 0.00745869 + layer.4.k_cache 0.00085529 0.19827948 + layer.4.v_cache 0.00004959 0.01328170 + layer.4.output 0.00648942 130.55728056 + ------------------------------------------------------------------------------------- + TOTAL 0.06944863 57.09012462 + (elements=3,499,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3499008 +Total Bytes 531292 +BPFP 1.2147 bits/point +EBPFP 2.4295 equivalent bits/point +MSE 57.090125 +---------------------- -------------------------------------------------------- +Time: 4.426s Load: 0.014s, Pack+Encode: 2.402s, Decode+Unpack: 2.010s +---------------------- -------------------------------------------------------- +💾 Converting with 57.0901 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,812B, BPFP=0.4935 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,772B, BPFP=2.5634 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,504B, BPFP=1.2603 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,656B, BPFP=2.5009 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,508B, BPFP=1.4285 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,880B, BPFP=2.4574 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,360B, BPFP=1.4203 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,892B, BPFP=2.5141 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,452B, BPFP=2.2095 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,704B, BPFP=2.4476 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,592B, BPFP=0.3488 +⌛️ [2/4] FRONTEND: Frontend time: 2.110s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.848s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13958113 40.89034498 + layer.0.v_cache 0.00001688 0.00682518 + layer.1.k_cache 0.56864935 8.32460640 + layer.1.v_cache 0.00000621 0.00273111 + layer.2.k_cache 0.01018704 0.93407252 + layer.2.v_cache 0.00002046 0.00783412 + layer.3.k_cache 0.04374025 5.69109561 + layer.3.v_cache 0.00002178 0.00908441 + layer.4.k_cache 0.00068741 0.25400647 + layer.4.v_cache 0.00005272 0.01527852 + layer.4.output 0.00479460 189.17809140 + ------------------------------------------------------------------------------------- + TOTAL 0.04685444 81.19897171 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 388132 +BPFP 1.2786 bits/point +EBPFP 2.5573 equivalent bits/point +MSE 81.198972 +---------------------- -------------------------------------------------------- +Time: 3.967s Load: 0.009s, Pack+Encode: 2.110s, Decode+Unpack: 1.848s +---------------------- -------------------------------------------------------- +💾 Converting with 81.1990 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 381, 128) +Output shape: (1, 381, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.output: torch.Size([1, 381, 3584]) -> torch.Size([1, 1, 381, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,588B, BPFP=0.4752 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,076B, BPFP=2.2587 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 32,368B, BPFP=1.3274 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,672B, BPFP=2.2011 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 35,964B, BPFP=1.4749 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,184B, BPFP=2.1811 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 34,252B, BPFP=1.4047 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 54,168B, BPFP=2.2215 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,928B, BPFP=1.9245 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,432B, BPFP=2.1503 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 59,004B, BPFP=0.3457 +⌛️ [2/4] FRONTEND: Frontend time: 2.540s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.817s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15713709 42.27111784 + layer.0.v_cache 0.00001719 0.00643938 + layer.1.k_cache 0.86767426 9.97779922 + layer.1.v_cache 0.00000622 0.00259213 + layer.2.k_cache 0.01635739 0.73934744 + layer.2.v_cache 0.00002098 0.00728482 + layer.3.k_cache 0.01709203 6.01840474 + layer.3.v_cache 0.00002073 0.00812173 + layer.4.k_cache 0.00073405 0.23908760 + layer.4.v_cache 0.00005083 0.01421183 + layer.4.output 0.03519400 135.61764154 + ------------------------------------------------------------------------------------- + TOTAL 0.07679228 59.32987633 + (elements=3,316,224) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3316224 +Total Bytes 488636 +BPFP 1.1788 bits/point +EBPFP 2.3576 equivalent bits/point +MSE 59.329876 +---------------------- -------------------------------------------------------- +Time: 4.371s Load: 0.014s, Pack+Encode: 2.540s, Decode+Unpack: 1.817s +---------------------- -------------------------------------------------------- +💾 Converting with 59.3299 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,484B, BPFP=0.4997 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,736B, BPFP=2.4530 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,252B, BPFP=1.3523 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,848B, BPFP=2.3937 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,288B, BPFP=1.6218 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,464B, BPFP=2.3681 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,536B, BPFP=1.5048 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,968B, BPFP=2.4017 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,808B, BPFP=2.1239 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,996B, BPFP=2.3368 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,884B, BPFP=0.3614 +⌛️ [2/4] FRONTEND: Frontend time: 2.107s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.738s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422791 40.92836038 + layer.0.v_cache 0.00001930 0.00711168 + layer.1.k_cache 0.38928963 8.84312922 + layer.1.v_cache 0.00000631 0.00280205 + layer.2.k_cache 0.01623082 0.71776431 + layer.2.v_cache 0.00002095 0.00810668 + layer.3.k_cache 0.01313608 6.10108100 + layer.3.v_cache 0.00002100 0.00909199 + layer.4.k_cache 0.00066994 0.24808893 + layer.4.v_cache 0.00005661 0.01641719 + layer.4.output 1.30836332 224.93505800 + ------------------------------------------------------------------------------------- + TOTAL 0.57013069 95.96631526 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 323264 +BPFP 1.2697 bits/point +EBPFP 2.5395 equivalent bits/point +MSE 95.966315 +---------------------- -------------------------------------------------------- +Time: 3.856s Load: 0.011s, Pack+Encode: 2.107s, Decode+Unpack: 1.738s +---------------------- -------------------------------------------------------- +💾 Converting with 95.9663 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,568B, BPFP=0.5011 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,688B, BPFP=2.4290 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,660B, BPFP=1.3016 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,760B, BPFP=2.3676 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,604B, BPFP=1.6290 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,300B, BPFP=2.3371 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,468B, BPFP=1.4213 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,040B, BPFP=2.3861 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,992B, BPFP=2.1181 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,908B, BPFP=2.3112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,144B, BPFP=0.3324 +⌛️ [2/4] FRONTEND: Frontend time: 27.271s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13591524 41.44661596 + layer.0.v_cache 0.00001656 0.00699101 + layer.1.k_cache 0.34520291 8.82500083 + layer.1.v_cache 0.00000598 0.00272813 + layer.2.k_cache 0.01250870 0.81165275 + layer.2.v_cache 0.00001982 0.00765148 + layer.3.k_cache 0.02633334 5.37415378 + layer.3.v_cache 0.00002177 0.00934052 + layer.4.k_cache 0.00067446 0.22584265 + layer.4.v_cache 0.00005295 0.01552466 + layer.4.output 1.29725540 225.62325969 + ------------------------------------------------------------------------------------- + TOTAL 0.56479644 96.24048939 + (elements=2,054,144) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2054144 +Total Bytes 319132 +BPFP 1.2429 bits/point +EBPFP 2.4858 equivalent bits/point +MSE 96.240489 +---------------------- --------------------------------------------------------- +Time: 28.638s Load: 0.009s, Pack+Encode: 27.271s, Decode+Unpack: 1.357s +---------------------- --------------------------------------------------------- +💾 Converting with 96.2405 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,968B, BPFP=0.5017 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,504B, BPFP=2.6285 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,516B, BPFP=1.3332 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,596B, BPFP=2.5631 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,016B, BPFP=1.5132 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,052B, BPFP=2.5239 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,072B, BPFP=1.3733 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,716B, BPFP=2.5717 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,384B, BPFP=2.2598 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,860B, BPFP=2.5101 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,184B, BPFP=0.3208 +⌛️ [2/4] FRONTEND: Frontend time: 2.045s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150146 42.38342454 + layer.0.v_cache 0.00001564 0.00675603 + layer.1.k_cache 0.36060091 8.84924710 + layer.1.v_cache 0.00000598 0.00263856 + layer.2.k_cache 0.01419523 0.89392568 + layer.2.v_cache 0.00002081 0.00767334 + layer.3.k_cache 0.03680107 5.30501321 + layer.3.v_cache 0.00002092 0.00879704 + layer.4.k_cache 0.00067164 0.22511010 + layer.4.v_cache 0.00004838 0.01472945 + layer.4.output 1.41076765 243.50384710 + ------------------------------------------------------------------------------------- + TOTAL 0.61289739 103.66024969 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 305868 +BPFP 1.2955 bits/point +EBPFP 2.5910 equivalent bits/point +MSE 103.660250 +---------------------- -------------------------------------------------------- +Time: 3.407s Load: 0.007s, Pack+Encode: 2.045s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 103.6602 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 343, 128) +Output shape: (1, 343, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.output: torch.Size([1, 343, 3584]) -> torch.Size([1, 1, 343, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,608B, BPFP=0.4832 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 54,864B, BPFP=2.4993 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,020B, BPFP=1.2764 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,552B, BPFP=2.4395 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,312B, BPFP=1.4264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,888B, BPFP=2.4093 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,936B, BPFP=1.4093 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,908B, BPFP=2.4557 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,104B, BPFP=2.1458 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,232B, BPFP=2.3794 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 54,684B, BPFP=0.3559 +⌛️ [2/4] FRONTEND: Frontend time: 2.578s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.661s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16677478 39.20447454 + layer.0.v_cache 0.00001698 0.00660021 + layer.1.k_cache 0.72144124 8.83134723 + layer.1.v_cache 0.00000659 0.00267592 + layer.2.k_cache 0.01351720 0.75856912 + layer.2.v_cache 0.00002117 0.00735563 + layer.3.k_cache 0.02347133 6.04251108 + layer.3.v_cache 0.00002006 0.00822118 + layer.4.k_cache 0.00069292 0.23392249 + layer.4.v_cache 0.00005144 0.01454240 + layer.4.output 0.03905215 148.95180394 + ------------------------------------------------------------------------------------- + TOTAL 0.07055169 64.57487337 + (elements=2,985,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2985472 +Total Bytes 470108 +BPFP 1.2597 bits/point +EBPFP 2.5194 equivalent bits/point +MSE 64.574873 +---------------------- -------------------------------------------------------- +Time: 4.251s Load: 0.012s, Pack+Encode: 2.578s, Decode+Unpack: 1.661s +---------------------- -------------------------------------------------------- +💾 Converting with 64.5749 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,660B, BPFP=0.4920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,272B, BPFP=2.5723 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,104B, BPFP=1.3127 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,068B, BPFP=2.5039 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,824B, BPFP=1.4673 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,436B, BPFP=2.4680 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,216B, BPFP=1.3759 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,876B, BPFP=2.4930 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,300B, BPFP=2.1761 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,380B, BPFP=2.4648 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,792B, BPFP=0.3392 +⌛️ [2/4] FRONTEND: Frontend time: 2.079s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.554s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13317820 41.64084162 + layer.0.v_cache 0.00001784 0.00665024 + layer.1.k_cache 0.51782770 8.38291371 + layer.1.v_cache 0.00000612 0.00264362 + layer.2.k_cache 0.01694398 0.79732638 + layer.2.v_cache 0.00002016 0.00733883 + layer.3.k_cache 0.01920085 5.74428489 + layer.3.v_cache 0.00002016 0.00809341 + layer.4.k_cache 0.00071488 0.21993265 + layer.4.v_cache 0.00005349 0.01459313 + layer.4.output 0.00484358 198.91371753 + ------------------------------------------------------------------------------------- + TOTAL 0.04246403 85.24827301 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 381928 +BPFP 1.2765 bits/point +EBPFP 2.5530 equivalent bits/point +MSE 85.248273 +---------------------- -------------------------------------------------------- +Time: 3.642s Load: 0.009s, Pack+Encode: 2.079s, Decode+Unpack: 1.554s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2483 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 374, 128) +Output shape: (1, 374, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.output: torch.Size([1, 374, 3584]) -> torch.Size([1, 1, 374, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,472B, BPFP=0.4793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,156B, BPFP=2.3043 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,120B, BPFP=1.3001 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,604B, BPFP=2.2395 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 34,992B, BPFP=1.4619 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,608B, BPFP=2.1979 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 32,892B, BPFP=1.3742 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,768B, BPFP=2.2463 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,568B, BPFP=1.9455 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,292B, BPFP=2.1847 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 48,628B, BPFP=0.2902 +⌛️ [2/4] FRONTEND: Frontend time: 2.487s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.770s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18781370 39.84415995 + layer.0.v_cache 0.00001670 0.00627685 + layer.1.k_cache 0.85217587 8.79848150 + layer.1.v_cache 0.00000605 0.00250016 + layer.2.k_cache 0.02919085 0.78453399 + layer.2.v_cache 0.00002159 0.00682387 + layer.3.k_cache 0.02976600 5.72554612 + layer.3.v_cache 0.00002022 0.00788051 + layer.4.k_cache 0.00072409 0.21426779 + layer.4.v_cache 0.00004949 0.01364648 + layer.4.output 0.03578218 137.63921648 + ------------------------------------------------------------------------------------- + TOTAL 0.07942705 59.93403721 + (elements=3,255,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3255296 +Total Bytes 473100 +BPFP 1.1627 bits/point +EBPFP 2.3253 equivalent bits/point +MSE 59.934037 +---------------------- -------------------------------------------------------- +Time: 4.272s Load: 0.014s, Pack+Encode: 2.487s, Decode+Unpack: 1.770s +---------------------- -------------------------------------------------------- +💾 Converting with 59.9340 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,584B, BPFP=0.4917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,744B, BPFP=2.3823 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,096B, BPFP=1.3029 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,936B, BPFP=2.3299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,412B, BPFP=1.5827 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,364B, BPFP=2.2928 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,476B, BPFP=1.4572 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,088B, BPFP=2.3397 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,584B, BPFP=2.0477 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,108B, BPFP=2.2762 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,888B, BPFP=0.3231 +⌛️ [2/4] FRONTEND: Frontend time: 2.140s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.486s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13979621 42.59804525 + layer.0.v_cache 0.00001743 0.00654497 + layer.1.k_cache 0.46959303 8.49340010 + layer.1.v_cache 0.00000602 0.00264597 + layer.2.k_cache 0.01642122 0.95151491 + layer.2.v_cache 0.00002053 0.00765036 + layer.3.k_cache 0.03876302 5.31892769 + layer.3.v_cache 0.00002036 0.00857176 + layer.4.k_cache 0.00069112 0.21328466 + layer.4.v_cache 0.00005399 0.01599065 + layer.4.output 1.27033806 205.66554905 + ------------------------------------------------------------------------------------- + TOTAL 0.56222055 88.07502469 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 320280 +BPFP 1.2215 bits/point +EBPFP 2.4429 equivalent bits/point +MSE 88.075025 +---------------------- -------------------------------------------------------- +Time: 3.633s Load: 0.008s, Pack+Encode: 2.140s, Decode+Unpack: 1.486s +---------------------- -------------------------------------------------------- +💾 Converting with 88.0750 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,128B, BPFP=0.4906 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,548B, BPFP=2.5157 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,540B, BPFP=1.3450 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,680B, BPFP=2.4559 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,056B, BPFP=1.6558 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,196B, BPFP=2.4226 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,820B, BPFP=1.5019 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,916B, BPFP=2.4722 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,900B, BPFP=2.1958 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,856B, BPFP=2.3992 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,236B, BPFP=0.3563 +⌛️ [2/4] FRONTEND: Frontend time: 1.998s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.507s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13893990 41.08559334 + layer.0.v_cache 0.00001711 0.00699159 + layer.1.k_cache 0.36820191 9.01671019 + layer.1.v_cache 0.00000637 0.00279261 + layer.2.k_cache 0.01566778 0.71739963 + layer.2.v_cache 0.00002161 0.00814011 + layer.3.k_cache 0.01884081 5.91291816 + layer.3.v_cache 0.00002076 0.00912525 + layer.4.k_cache 0.00070054 0.23245745 + layer.4.v_cache 0.00005210 0.01595104 + layer.4.output 1.34866800 236.13286658 + ------------------------------------------------------------------------------------- + TOTAL 0.58724382 100.58459679 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 318876 +BPFP 1.2911 bits/point +EBPFP 2.5822 equivalent bits/point +MSE 100.584597 +---------------------- -------------------------------------------------------- +Time: 3.514s Load: 0.009s, Pack+Encode: 1.998s, Decode+Unpack: 1.507s +---------------------- -------------------------------------------------------- +💾 Converting with 100.5846 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,568B, BPFP=0.5014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,580B, BPFP=2.6088 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,696B, BPFP=1.2697 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,628B, BPFP=2.5531 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,392B, BPFP=1.4860 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,204B, BPFP=2.5283 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,512B, BPFP=1.4345 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,776B, BPFP=2.5618 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,984B, BPFP=2.2814 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,804B, BPFP=2.5049 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,600B, BPFP=0.3311 +⌛️ [2/4] FRONTEND: Frontend time: 2.299s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.816s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11943251 40.42209079 + layer.0.v_cache 0.00001561 0.00649505 + layer.1.k_cache 0.49619702 8.75940353 + layer.1.v_cache 0.00000608 0.00263590 + layer.2.k_cache 0.01349212 0.86516574 + layer.2.v_cache 0.00002006 0.00761208 + layer.3.k_cache 0.01080775 5.60829968 + layer.3.v_cache 0.00002073 0.00849008 + layer.4.k_cache 0.00068886 0.23912760 + layer.4.v_cache 0.00007406 0.01535850 + layer.4.output 0.00494179 204.71214553 + ------------------------------------------------------------------------------------- + TOTAL 0.03972631 87.58351163 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 376744 +BPFP 1.2969 bits/point +EBPFP 2.5938 equivalent bits/point +MSE 87.583512 +---------------------- -------------------------------------------------------- +Time: 4.125s Load: 0.010s, Pack+Encode: 2.299s, Decode+Unpack: 1.816s +---------------------- -------------------------------------------------------- +💾 Converting with 87.5835 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,984B, BPFP=0.4938 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,536B, BPFP=2.5831 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,112B, BPFP=1.3512 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,676B, BPFP=2.5223 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,876B, BPFP=1.6174 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,356B, BPFP=2.4997 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,608B, BPFP=1.4570 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,972B, BPFP=2.5433 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,940B, BPFP=2.2582 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,132B, BPFP=2.4839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,580B, BPFP=0.3291 +⌛️ [2/4] FRONTEND: Frontend time: 2.079s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.568s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11815372 41.13622826 + layer.0.v_cache 0.00001627 0.00686001 + layer.1.k_cache 0.37623358 8.56014476 + layer.1.v_cache 0.00000616 0.00275276 + layer.2.k_cache 0.01429399 0.90561134 + layer.2.v_cache 0.00002051 0.00813967 + layer.3.k_cache 0.01320818 5.75557878 + layer.3.v_cache 0.00002004 0.00918446 + layer.4.k_cache 0.00067885 0.23326031 + layer.4.v_cache 0.00005318 0.01654514 + layer.4.output 1.38524649 236.60989011 + ------------------------------------------------------------------------------------- + TOTAL 0.60114176 100.75903154 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 312772 +BPFP 1.3008 bits/point +EBPFP 2.6016 equivalent bits/point +MSE 100.759032 +---------------------- -------------------------------------------------------- +Time: 3.656s Load: 0.009s, Pack+Encode: 2.079s, Decode+Unpack: 1.568s +---------------------- -------------------------------------------------------- +💾 Converting with 100.7590 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,740B, BPFP=0.4912 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,684B, BPFP=2.5677 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,184B, BPFP=1.3031 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,332B, BPFP=2.4917 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,088B, BPFP=1.4663 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,832B, BPFP=2.4636 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,716B, BPFP=1.3330 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,716B, BPFP=2.5133 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,676B, BPFP=2.1738 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,608B, BPFP=2.4510 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,080B, BPFP=0.2977 +⌛️ [2/4] FRONTEND: Frontend time: 2.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.728s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09820673 40.09988337 + layer.0.v_cache 0.00001666 0.00622041 + layer.1.k_cache 0.50544519 9.15767972 + layer.1.v_cache 0.00000613 0.00257232 + layer.2.k_cache 0.01930715 0.73472107 + layer.2.v_cache 0.00002033 0.00709219 + layer.3.k_cache 0.01337509 6.17116190 + layer.3.v_cache 0.00001987 0.00799227 + layer.4.k_cache 0.00074230 0.21477230 + layer.4.v_cache 0.00004869 0.01446449 + layer.4.output 0.00473525 190.42894078 + ------------------------------------------------------------------------------------- + TOTAL 0.03943146 81.73053797 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 379656 +BPFP 1.2552 bits/point +EBPFP 2.5104 equivalent bits/point +MSE 81.730538 +---------------------- -------------------------------------------------------- +Time: 4.062s Load: 0.010s, Pack+Encode: 2.324s, Decode+Unpack: 1.728s +---------------------- -------------------------------------------------------- +💾 Converting with 81.7305 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,024B, BPFP=0.5034 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,524B, BPFP=2.6178 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,568B, BPFP=1.2592 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,568B, BPFP=2.5493 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,116B, BPFP=1.5135 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,112B, BPFP=2.5166 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,536B, BPFP=1.4002 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,820B, BPFP=2.5674 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,748B, BPFP=2.2755 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,056B, BPFP=2.5126 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,584B, BPFP=0.3029 +⌛️ [2/4] FRONTEND: Frontend time: 2.512s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.475s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13853801 40.19670119 + layer.0.v_cache 0.00001803 0.00698452 + layer.1.k_cache 0.23740530 9.09214741 + layer.1.v_cache 0.00000581 0.00264789 + layer.2.k_cache 0.01896443 0.71678267 + layer.2.v_cache 0.00001977 0.00810878 + layer.3.k_cache 0.00876656 5.64790834 + layer.3.v_cache 0.00002034 0.00957941 + layer.4.k_cache 0.00073522 0.23327178 + layer.4.v_cache 0.00004914 0.01561538 + layer.4.output 1.40427464 241.39678899 + ------------------------------------------------------------------------------------- + TOTAL 0.60202618 102.68866296 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 304656 +BPFP 1.2845 bits/point +EBPFP 2.5689 equivalent bits/point +MSE 102.688663 +---------------------- -------------------------------------------------------- +Time: 3.995s Load: 0.009s, Pack+Encode: 2.512s, Decode+Unpack: 1.475s +---------------------- -------------------------------------------------------- +💾 Converting with 102.6887 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,744B, BPFP=0.4950 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,224B, BPFP=2.5602 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,868B, BPFP=1.2946 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,184B, BPFP=2.5014 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,492B, BPFP=1.4998 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,684B, BPFP=2.4731 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,416B, BPFP=1.3256 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,516B, BPFP=2.5202 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,408B, BPFP=2.1744 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,528B, BPFP=2.4642 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,128B, BPFP=0.3326 +⌛️ [2/4] FRONTEND: Frontend time: 2.118s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.974s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10827675 41.60679560 + layer.0.v_cache 0.00001706 0.00640736 + layer.1.k_cache 0.56919524 8.65073384 + layer.1.v_cache 0.00000592 0.00252824 + layer.2.k_cache 0.01911232 0.74099654 + layer.2.v_cache 0.00002018 0.00749626 + layer.3.k_cache 0.01342056 5.88796245 + layer.3.v_cache 0.00002333 0.00870799 + layer.4.k_cache 0.00074072 0.20674479 + layer.4.v_cache 0.00005062 0.01462450 + layer.4.output 0.00480139 198.26240619 + ------------------------------------------------------------------------------------- + TOTAL 0.04379250 84.99822593 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 382192 +BPFP 1.2728 bits/point +EBPFP 2.5455 equivalent bits/point +MSE 84.998226 +---------------------- -------------------------------------------------------- +Time: 4.101s Load: 0.009s, Pack+Encode: 2.118s, Decode+Unpack: 1.974s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9982 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,572B, BPFP=0.5016 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,768B, BPFP=2.6199 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,524B, BPFP=1.2596 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,592B, BPFP=2.5510 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,108B, BPFP=1.4693 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,048B, BPFP=2.5192 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,048B, BPFP=1.4658 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,932B, BPFP=2.5709 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,308B, BPFP=2.2418 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,324B, BPFP=2.4768 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,620B, BPFP=0.3396 +⌛️ [2/4] FRONTEND: Frontend time: 2.717s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.765s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15963619 41.14493724 + layer.0.v_cache 0.00001721 0.00707514 + layer.1.k_cache 0.49929055 8.60117663 + layer.1.v_cache 0.00000624 0.00269278 + layer.2.k_cache 0.03270629 0.77780214 + layer.2.v_cache 0.00002062 0.00728992 + layer.3.k_cache 0.00946706 5.87842025 + layer.3.v_cache 0.00002056 0.00842216 + layer.4.k_cache 0.00068536 0.24186238 + layer.4.v_cache 0.00004941 0.01439371 + layer.4.output 0.00495045 202.23618914 + ------------------------------------------------------------------------------------- + TOTAL 0.04332662 86.60808214 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 376844 +BPFP 1.2972 bits/point +EBPFP 2.5945 equivalent bits/point +MSE 86.608082 +---------------------- -------------------------------------------------------- +Time: 4.491s Load: 0.009s, Pack+Encode: 2.717s, Decode+Unpack: 1.765s +---------------------- -------------------------------------------------------- +💾 Converting with 86.6081 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,548B, BPFP=0.4997 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,728B, BPFP=2.4317 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,116B, BPFP=1.3318 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,880B, BPFP=2.3755 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,000B, BPFP=1.5890 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,136B, BPFP=2.3263 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,084B, BPFP=1.4621 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,956B, BPFP=2.3806 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,848B, BPFP=2.1086 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,032B, BPFP=2.3194 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,444B, BPFP=0.3163 +⌛️ [2/4] FRONTEND: Frontend time: 2.293s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.562s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14874478 40.24514202 + layer.0.v_cache 0.00001714 0.00679210 + layer.1.k_cache 0.45078779 8.87915970 + layer.1.v_cache 0.00000615 0.00268684 + layer.2.k_cache 0.01724745 0.98750208 + layer.2.v_cache 0.00001987 0.00746280 + layer.3.k_cache 0.02849353 5.61144140 + layer.3.v_cache 0.00002056 0.00862197 + layer.4.k_cache 0.00066912 0.22972854 + layer.4.v_cache 0.00005068 0.01506134 + layer.4.output 1.29722614 222.53085275 + ------------------------------------------------------------------------------------- + TOTAL 0.57215530 94.92409224 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 317772 +BPFP 1.2376 bits/point +EBPFP 2.4752 equivalent bits/point +MSE 94.924092 +---------------------- -------------------------------------------------------- +Time: 3.864s Load: 0.009s, Pack+Encode: 2.293s, Decode+Unpack: 1.562s +---------------------- -------------------------------------------------------- +💾 Converting with 94.9241 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 341, 128) +Output shape: (1, 341, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.output: torch.Size([1, 341, 3584]) -> torch.Size([1, 1, 341, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,584B, BPFP=0.4850 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 54,720B, BPFP=2.5073 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,628B, BPFP=1.2659 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,316B, BPFP=2.4430 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,940B, BPFP=1.4177 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,508B, BPFP=2.4060 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,532B, BPFP=1.3990 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,636B, BPFP=2.4577 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,152B, BPFP=2.1147 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,040B, BPFP=2.3845 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 46,960B, BPFP=0.3074 +⌛️ [2/4] FRONTEND: Frontend time: 2.447s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.979s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17009405 40.21315982 + layer.0.v_cache 0.00001631 0.00654995 + layer.1.k_cache 0.79816426 8.41117219 + layer.1.v_cache 0.00000611 0.00254946 + layer.2.k_cache 0.02878924 0.81136620 + layer.2.v_cache 0.00002016 0.00717726 + layer.3.k_cache 0.01035325 5.37421675 + layer.3.v_cache 0.00002082 0.00806022 + layer.4.k_cache 0.00072160 0.21875436 + layer.4.v_cache 0.00005837 0.01420765 + layer.4.output 0.03919861 147.67777283 + ------------------------------------------------------------------------------------- + TOTAL 0.07544909 64.04774257 + (elements=2,968,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2968064 +Total Bytes 459016 +BPFP 1.2372 bits/point +EBPFP 2.4744 equivalent bits/point +MSE 64.047743 +---------------------- -------------------------------------------------------- +Time: 4.438s Load: 0.012s, Pack+Encode: 2.447s, Decode+Unpack: 1.979s +---------------------- -------------------------------------------------------- +💾 Converting with 64.0477 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 335, 128) +Output shape: (1, 335, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.output: torch.Size([1, 335, 3584]) -> torch.Size([1, 1, 335, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,488B, BPFP=0.4892 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 54,048B, BPFP=2.5209 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,708B, BPFP=1.2457 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 52,828B, BPFP=2.4640 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,388B, BPFP=1.4640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 51,940B, BPFP=2.4226 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,892B, BPFP=1.4409 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 52,968B, BPFP=2.4705 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,140B, BPFP=2.1521 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 51,392B, BPFP=2.3970 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 47,180B, BPFP=0.3144 +⌛️ [2/4] FRONTEND: Frontend time: 2.510s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.744s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15195324 39.71372726 + layer.0.v_cache 0.00001666 0.00650434 + layer.1.k_cache 0.76788093 8.77286322 + layer.1.v_cache 0.00000624 0.00257495 + layer.2.k_cache 0.02686723 0.95342052 + layer.2.v_cache 0.00001960 0.00688712 + layer.3.k_cache 0.01080669 5.64621108 + layer.3.v_cache 0.00002024 0.00786499 + layer.4.k_cache 0.00071790 0.22608303 + layer.4.v_cache 0.00005106 0.01400640 + layer.4.output 3.38143948 158.39339019 + ------------------------------------------------------------------------------------- + TOTAL 1.44873036 68.47669849 + (elements=2,915,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2915840 +Total Bytes 455972 +BPFP 1.2510 bits/point +EBPFP 2.5020 equivalent bits/point +MSE 68.476698 +---------------------- -------------------------------------------------------- +Time: 4.268s Load: 0.013s, Pack+Encode: 2.510s, Decode+Unpack: 1.744s +---------------------- -------------------------------------------------------- +💾 Converting with 68.4767 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,432B, BPFP=0.4972 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,528B, BPFP=2.6255 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,656B, BPFP=1.2769 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,888B, BPFP=2.5877 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,012B, BPFP=1.4748 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,108B, BPFP=2.5417 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,704B, BPFP=1.3976 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,992B, BPFP=2.5939 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,056B, BPFP=2.1849 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,756B, BPFP=2.5210 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,492B, BPFP=0.3748 +⌛️ [2/4] FRONTEND: Frontend time: 2.320s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.615s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17469128 40.21429835 + layer.0.v_cache 0.00001775 0.00673441 + layer.1.k_cache 0.51265420 8.34379791 + layer.1.v_cache 0.00000620 0.00264628 + layer.2.k_cache 0.03428403 0.87630857 + layer.2.v_cache 0.00002024 0.00713337 + layer.3.k_cache 0.01443097 6.19534474 + layer.3.v_cache 0.00002182 0.00844177 + layer.4.k_cache 0.00070119 0.20226129 + layer.4.v_cache 0.00005097 0.01463236 + layer.4.output 0.00503371 199.52543801 + ------------------------------------------------------------------------------------- + TOTAL 0.04541851 85.44409795 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 378624 +BPFP 1.3132 bits/point +EBPFP 2.6264 equivalent bits/point +MSE 85.444098 +---------------------- -------------------------------------------------------- +Time: 3.943s Load: 0.008s, Pack+Encode: 2.320s, Decode+Unpack: 1.615s +---------------------- -------------------------------------------------------- +💾 Converting with 85.4441 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,420B, BPFP=0.4965 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,180B, BPFP=2.6050 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,152B, BPFP=1.2472 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,292B, BPFP=2.5526 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,956B, BPFP=1.4125 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,156B, BPFP=2.4856 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,188B, BPFP=1.3672 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,500B, BPFP=2.5649 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,856B, BPFP=2.2321 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,648B, BPFP=2.4557 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,984B, BPFP=0.3284 +⌛️ [2/4] FRONTEND: Frontend time: 2.115s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.655s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18043976 40.14568838 + layer.0.v_cache 0.00001814 0.00654131 + layer.1.k_cache 0.58196952 8.77039449 + layer.1.v_cache 0.00000608 0.00260789 + layer.2.k_cache 0.02409397 0.81851680 + layer.2.v_cache 0.00002020 0.00725709 + layer.3.k_cache 0.01875559 4.84112894 + layer.3.v_cache 0.00002085 0.00812341 + layer.4.k_cache 0.00069073 0.23785981 + layer.4.v_cache 0.00005389 0.01540972 + layer.4.output 0.00498411 205.21202830 + ------------------------------------------------------------------------------------- + TOTAL 0.04946809 87.72574859 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 368332 +BPFP 1.2775 bits/point +EBPFP 2.5550 equivalent bits/point +MSE 87.725749 +---------------------- -------------------------------------------------------- +Time: 3.778s Load: 0.009s, Pack+Encode: 2.115s, Decode+Unpack: 1.655s +---------------------- -------------------------------------------------------- +💾 Converting with 87.7257 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 386, 128) +Output shape: (1, 386, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.output: torch.Size([1, 386, 3584]) -> torch.Size([1, 1, 386, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,752B, BPFP=0.4757 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 61,148B, BPFP=2.4752 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,768B, BPFP=1.2859 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 60,620B, BPFP=2.4539 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 35,036B, BPFP=1.4182 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 59,712B, BPFP=2.4171 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 35,072B, BPFP=1.4197 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 60,404B, BPFP=2.4451 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 52,152B, BPFP=2.1111 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 58,580B, BPFP=2.3713 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 54,840B, BPFP=0.3171 +⌛️ [2/4] FRONTEND: Frontend time: 2.362s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.972s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16808204 41.62699613 + layer.0.v_cache 0.00001827 0.00679233 + layer.1.k_cache 0.89383947 8.68594724 + layer.1.v_cache 0.00000605 0.00266755 + layer.2.k_cache 0.02643975 0.87639748 + layer.2.v_cache 0.00002094 0.00757677 + layer.3.k_cache 0.04368651 5.86732538 + layer.3.v_cache 0.00002149 0.00875873 + layer.4.k_cache 0.00075087 0.24825441 + layer.4.v_cache 0.00005208 0.01519080 + layer.4.output 0.03469837 137.72271697 + ------------------------------------------------------------------------------------- + TOTAL 0.08092977 60.08264268 + (elements=3,359,744) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3359744 +Total Bytes 521084 +BPFP 1.2408 bits/point +EBPFP 2.4815 equivalent bits/point +MSE 60.082643 +---------------------- -------------------------------------------------------- +Time: 4.348s Load: 0.014s, Pack+Encode: 2.362s, Decode+Unpack: 1.972s +---------------------- -------------------------------------------------------- +💾 Converting with 60.0826 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,812B, BPFP=0.4781 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,748B, BPFP=2.4820 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,724B, BPFP=1.3414 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,856B, BPFP=2.4336 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,044B, BPFP=1.5757 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,284B, BPFP=2.4026 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,292B, BPFP=1.4264 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,996B, BPFP=2.4412 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,668B, BPFP=2.1521 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,836B, BPFP=2.3783 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,716B, BPFP=0.3311 +⌛️ [2/4] FRONTEND: Frontend time: 2.092s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.847s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14364293 40.18555705 + layer.0.v_cache 0.00001600 0.00677194 + layer.1.k_cache 0.62366660 9.03734250 + layer.1.v_cache 0.00000628 0.00268813 + layer.2.k_cache 0.01981349 0.84875520 + layer.2.v_cache 0.00002090 0.00782986 + layer.3.k_cache 0.02158341 6.14494705 + layer.3.v_cache 0.00002097 0.00862394 + layer.4.k_cache 0.00071696 0.23722098 + layer.4.v_cache 0.00005883 0.01542149 + layer.4.output 0.00463133 190.23794023 + ------------------------------------------------------------------------------------- + TOTAL 0.04952739 81.65651410 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 394976 +BPFP 1.2605 bits/point +EBPFP 2.5210 equivalent bits/point +MSE 81.656514 +---------------------- -------------------------------------------------------- +Time: 3.949s Load: 0.010s, Pack+Encode: 2.092s, Decode+Unpack: 1.847s +---------------------- -------------------------------------------------------- +💾 Converting with 81.6565 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,692B, BPFP=0.4903 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,504B, BPFP=2.5668 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,460B, BPFP=1.2669 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,620B, BPFP=2.5169 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,632B, BPFP=1.4458 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,148B, BPFP=2.4903 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,680B, BPFP=1.3921 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,004B, BPFP=2.5386 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,212B, BPFP=2.1555 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,668B, BPFP=2.4632 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 47,176B, BPFP=0.3802 +⌛️ [2/4] FRONTEND: Frontend time: 2.138s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.593s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12300906 39.24809271 + layer.0.v_cache 0.00001720 0.00619759 + layer.1.k_cache 0.58060381 8.17597163 + layer.1.v_cache 0.00000644 0.00239489 + layer.2.k_cache 0.02305070 0.81135917 + layer.2.v_cache 0.00002053 0.00702981 + layer.3.k_cache 0.05279791 5.93818218 + layer.3.v_cache 0.00002021 0.00772632 + layer.4.k_cache 0.00074809 0.19787854 + layer.4.v_cache 0.00005348 0.01446541 + layer.4.output 0.00483522 188.63953713 + ------------------------------------------------------------------------------------- + TOTAL 0.04789259 80.87565048 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 389796 +BPFP 1.2934 bits/point +EBPFP 2.5868 equivalent bits/point +MSE 80.875650 +---------------------- -------------------------------------------------------- +Time: 3.742s Load: 0.011s, Pack+Encode: 2.138s, Decode+Unpack: 1.593s +---------------------- -------------------------------------------------------- +💾 Converting with 80.8757 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,924B, BPFP=0.4851 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,920B, BPFP=2.5869 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,112B, BPFP=1.3391 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,688B, BPFP=2.5006 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,812B, BPFP=1.6684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,316B, BPFP=2.4745 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,852B, BPFP=1.4610 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,904B, BPFP=2.5157 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,912B, BPFP=2.2360 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,040B, BPFP=2.4552 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,300B, BPFP=0.3433 +⌛️ [2/4] FRONTEND: Frontend time: 1.985s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.511s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16177035 41.04801797 + layer.0.v_cache 0.00001782 0.00715571 + layer.1.k_cache 0.31952496 8.44387338 + layer.1.v_cache 0.00000626 0.00275485 + layer.2.k_cache 0.01271096 0.77080067 + layer.2.v_cache 0.00002040 0.00785656 + layer.3.k_cache 0.01363577 7.09219100 + layer.3.v_cache 0.00002033 0.00916056 + layer.4.k_cache 0.00066992 0.25396403 + layer.4.v_cache 0.00005030 0.01522244 + layer.4.output 1.37284199 239.16383728 + ------------------------------------------------------------------------------------- + TOTAL 0.59519535 101.87046225 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 315780 +BPFP 1.3015 bits/point +EBPFP 2.6030 equivalent bits/point +MSE 101.870462 +---------------------- -------------------------------------------------------- +Time: 3.504s Load: 0.008s, Pack+Encode: 1.985s, Decode+Unpack: 1.511s +---------------------- -------------------------------------------------------- +💾 Converting with 101.8705 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,912B, BPFP=0.5048 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,832B, BPFP=2.3764 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,024B, BPFP=1.2828 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,068B, BPFP=2.3111 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,008B, BPFP=1.5376 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,896B, BPFP=2.2964 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,948B, BPFP=1.4471 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,460B, BPFP=2.3446 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,288B, BPFP=2.0738 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,724B, BPFP=2.2818 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,760B, BPFP=0.3752 +⌛️ [2/4] FRONTEND: Frontend time: 2.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.286s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09962382 39.58933145 + layer.0.v_cache 0.00001686 0.00720730 + layer.1.k_cache 0.10988800 8.91347149 + layer.1.v_cache 0.00000590 0.00277742 + layer.2.k_cache 0.01214521 0.91431314 + layer.2.v_cache 0.00002216 0.00838708 + layer.3.k_cache 0.02260447 4.60773506 + layer.3.v_cache 0.00002006 0.01006377 + layer.4.k_cache 0.00069723 0.22093643 + layer.4.v_cache 0.00005242 0.01701350 + layer.4.output 0.00882482 287.55562061 + ------------------------------------------------------------------------------------- + TOTAL 0.01804999 121.59885770 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 246920 +BPFP 1.2402 bits/point +EBPFP 2.4803 equivalent bits/point +MSE 121.598858 +---------------------- -------------------------------------------------------- +Time: 3.450s Load: 0.006s, Pack+Encode: 2.158s, Decode+Unpack: 1.286s +---------------------- -------------------------------------------------------- +💾 Converting with 121.5989 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,824B, BPFP=0.4907 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,740B, BPFP=2.5434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,160B, BPFP=1.2878 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,596B, BPFP=2.4798 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,872B, BPFP=1.4942 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,800B, BPFP=2.4355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,408B, BPFP=1.4128 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,652B, BPFP=2.4829 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,044B, BPFP=2.1710 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,512B, BPFP=2.4195 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,628B, BPFP=0.3307 +⌛️ [2/4] FRONTEND: Frontend time: 2.283s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.557s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11148910 41.04438668 + layer.0.v_cache 0.00001684 0.00663412 + layer.1.k_cache 0.60671574 8.58631370 + layer.1.v_cache 0.00000630 0.00259950 + layer.2.k_cache 0.04149003 0.77921977 + layer.2.v_cache 0.00001976 0.00712539 + layer.3.k_cache 0.02200458 6.32432171 + layer.3.v_cache 0.00001954 0.00817929 + layer.4.k_cache 0.00078895 0.21690257 + layer.4.v_cache 0.00004864 0.01429274 + layer.4.output 0.00472744 191.69109049 + ------------------------------------------------------------------------------------- + TOTAL 0.04798185 82.28397700 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 387236 +BPFP 1.2666 bits/point +EBPFP 2.5332 equivalent bits/point +MSE 82.283977 +---------------------- -------------------------------------------------------- +Time: 3.852s Load: 0.013s, Pack+Encode: 2.283s, Decode+Unpack: 1.557s +---------------------- -------------------------------------------------------- +💾 Converting with 82.2840 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,568B, BPFP=0.5131 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,356B, BPFP=2.7622 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,300B, BPFP=1.2734 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,592B, BPFP=2.7025 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,420B, BPFP=1.4391 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,432B, BPFP=2.6900 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,560B, BPFP=1.3719 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,876B, BPFP=2.7247 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,688B, BPFP=2.3194 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,864B, BPFP=2.6456 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,200B, BPFP=0.3259 +⌛️ [2/4] FRONTEND: Frontend time: 2.082s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.487s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14565937 41.63491699 + layer.0.v_cache 0.00001550 0.00663926 + layer.1.k_cache 0.22559690 8.52740845 + layer.1.v_cache 0.00000590 0.00271157 + layer.2.k_cache 0.00679684 0.86516426 + layer.2.v_cache 0.00002314 0.00803035 + layer.3.k_cache 0.06866968 5.20657898 + layer.3.v_cache 0.00002119 0.00905968 + layer.4.k_cache 0.00065507 0.22122257 + layer.4.v_cache 0.00005191 0.01573160 + layer.4.output 1.53062134 257.71194196 + ------------------------------------------------------------------------------------- + TOTAL 0.65657911 109.44006220 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 290856 +BPFP 1.3367 bits/point +EBPFP 2.6733 equivalent bits/point +MSE 109.440062 +---------------------- -------------------------------------------------------- +Time: 3.575s Load: 0.007s, Pack+Encode: 2.082s, Decode+Unpack: 1.487s +---------------------- -------------------------------------------------------- +💾 Converting with 109.4401 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,560B, BPFP=0.5100 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,352B, BPFP=2.7481 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,420B, BPFP=1.2764 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,624B, BPFP=2.6915 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,956B, BPFP=1.5513 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,004B, BPFP=2.6433 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,028B, BPFP=1.4014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,796B, BPFP=2.7049 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,684B, BPFP=2.3853 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,280B, BPFP=2.6648 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,004B, BPFP=0.3332 +⌛️ [2/4] FRONTEND: Frontend time: 2.017s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.519s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10663509 40.69429707 + layer.0.v_cache 0.00001613 0.00701181 + layer.1.k_cache 0.26620468 8.77928351 + layer.1.v_cache 0.00000602 0.00277942 + layer.2.k_cache 0.00649704 0.81429348 + layer.2.v_cache 0.00002002 0.00830412 + layer.3.k_cache 0.02565914 5.31823123 + layer.3.v_cache 0.00002186 0.00959613 + layer.4.k_cache 0.00070101 0.25683313 + layer.4.v_cache 0.00005421 0.01721912 + layer.4.output 1.52303164 264.33357765 + ------------------------------------------------------------------------------------- + TOTAL 0.65100216 112.13193486 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 294708 +BPFP 1.3476 bits/point +EBPFP 2.6952 equivalent bits/point +MSE 112.131935 +---------------------- -------------------------------------------------------- +Time: 3.543s Load: 0.008s, Pack+Encode: 2.017s, Decode+Unpack: 1.519s +---------------------- -------------------------------------------------------- +💾 Converting with 112.1319 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,552B, BPFP=0.5005 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,452B, BPFP=2.6014 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,132B, BPFP=1.2952 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,552B, BPFP=2.5487 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,424B, BPFP=1.4293 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,076B, BPFP=2.5208 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,648B, BPFP=1.3839 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,052B, BPFP=2.5779 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,676B, BPFP=2.2633 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,960B, BPFP=2.5140 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,980B, BPFP=0.3593 +⌛️ [2/4] FRONTEND: Frontend time: 2.238s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.534s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258070 39.99836508 + layer.0.v_cache 0.00001713 0.00705338 + layer.1.k_cache 0.49703808 9.00627908 + layer.1.v_cache 0.00000670 0.00292062 + layer.2.k_cache 0.01458295 0.81350182 + layer.2.v_cache 0.00002181 0.00774107 + layer.3.k_cache 0.02940617 6.07390237 + layer.3.v_cache 0.00002120 0.00903952 + layer.4.k_cache 0.00070157 0.24399375 + layer.4.v_cache 0.00005095 0.01516853 + layer.4.output 0.00497170 201.48938269 + ------------------------------------------------------------------------------------- + TOTAL 0.04171936 86.27080259 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 378504 +BPFP 1.3030 bits/point +EBPFP 2.6059 equivalent bits/point +MSE 86.270803 +---------------------- -------------------------------------------------------- +Time: 3.782s Load: 0.010s, Pack+Encode: 2.238s, Decode+Unpack: 1.534s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2708 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,040B, BPFP=0.4867 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,232B, BPFP=2.5050 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,800B, BPFP=1.2998 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,376B, BPFP=2.4458 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,144B, BPFP=1.6692 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,076B, BPFP=2.4251 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,980B, BPFP=1.5196 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,952B, BPFP=2.4856 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,112B, BPFP=2.2201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,724B, BPFP=2.4007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,220B, BPFP=0.3281 +⌛️ [2/4] FRONTEND: Frontend time: 2.105s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.431s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15433978 40.85343352 + layer.0.v_cache 0.00001575 0.00698820 + layer.1.k_cache 0.29362690 8.79798606 + layer.1.v_cache 0.00000607 0.00288382 + layer.2.k_cache 0.00940348 0.97076659 + layer.2.v_cache 0.00002159 0.00838078 + layer.3.k_cache 0.02320663 6.24493894 + layer.3.v_cache 0.00002067 0.00976482 + layer.4.k_cache 0.00067785 0.25176065 + layer.4.v_cache 0.00005016 0.01669893 + layer.4.output 1.35461534 235.68135667 + ------------------------------------------------------------------------------------- + TOTAL 0.58609860 100.40782935 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 314656 +BPFP 1.2797 bits/point +EBPFP 2.5593 equivalent bits/point +MSE 100.407829 +---------------------- -------------------------------------------------------- +Time: 3.543s Load: 0.007s, Pack+Encode: 2.105s, Decode+Unpack: 1.431s +---------------------- -------------------------------------------------------- +💾 Converting with 100.4078 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,872B, BPFP=0.4864 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,924B, BPFP=2.5178 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,344B, BPFP=1.2798 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,828B, BPFP=2.4577 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,568B, BPFP=1.6211 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,448B, BPFP=2.4368 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,440B, BPFP=1.4496 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,236B, BPFP=2.4800 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,104B, BPFP=2.1439 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,932B, BPFP=2.4086 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 47,472B, BPFP=0.3718 +⌛️ [2/4] FRONTEND: Frontend time: 2.363s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.611s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14751465 41.29850603 + layer.0.v_cache 0.00001719 0.00677724 + layer.1.k_cache 0.52245269 8.56933422 + layer.1.v_cache 0.00000618 0.00260763 + layer.2.k_cache 0.02707210 0.73254705 + layer.2.v_cache 0.00002227 0.00740973 + layer.3.k_cache 0.03364062 5.83943599 + layer.3.v_cache 0.00002083 0.00838438 + layer.4.k_cache 0.00069748 0.22017253 + layer.4.v_cache 0.00005305 0.01440340 + layer.4.output 0.00471531 186.65758145 + ------------------------------------------------------------------------------------- + TOTAL 0.04497084 80.19427343 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 399168 +BPFP 1.2873 bits/point +EBPFP 2.5746 equivalent bits/point +MSE 80.194273 +---------------------- -------------------------------------------------------- +Time: 3.984s Load: 0.010s, Pack+Encode: 2.363s, Decode+Unpack: 1.611s +---------------------- -------------------------------------------------------- +💾 Converting with 80.1943 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,740B, BPFP=0.4743 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,836B, BPFP=2.2571 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,476B, BPFP=1.3159 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,092B, BPFP=2.2115 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,152B, BPFP=1.4186 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,492B, BPFP=2.1748 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,544B, BPFP=1.3814 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,260B, BPFP=2.2218 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,160B, BPFP=1.9093 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,216B, BPFP=2.1578 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,120B, BPFP=0.3862 +⌛️ [2/4] FRONTEND: Frontend time: 2.069s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.701s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17432048 40.27840074 + layer.0.v_cache 0.00001704 0.00601732 + layer.1.k_cache 0.39057378 9.55921607 + layer.1.v_cache 0.00000668 0.00252321 + layer.2.k_cache 0.01299327 0.95851177 + layer.2.v_cache 0.00002114 0.00682789 + layer.3.k_cache 0.02499930 6.96633779 + layer.3.v_cache 0.00002096 0.00759123 + layer.4.k_cache 0.00069121 0.21682087 + layer.4.v_cache 0.00005416 0.01348322 + layer.4.output 1.20070616 203.31939776 + ------------------------------------------------------------------------------------- + TOTAL 0.52992007 87.13244202 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 330088 +BPFP 1.1898 bits/point +EBPFP 2.3795 equivalent bits/point +MSE 87.132442 +---------------------- -------------------------------------------------------- +Time: 3.779s Load: 0.008s, Pack+Encode: 2.069s, Decode+Unpack: 1.701s +---------------------- -------------------------------------------------------- +💾 Converting with 87.1324 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,272B, BPFP=0.4962 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,508B, BPFP=2.4910 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,336B, BPFP=1.3876 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,840B, BPFP=2.4454 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,920B, BPFP=1.6321 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,168B, BPFP=2.3996 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,800B, BPFP=1.4192 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,896B, BPFP=2.4492 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,540B, BPFP=2.1520 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,724B, BPFP=2.3693 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,224B, BPFP=0.3628 +⌛️ [2/4] FRONTEND: Frontend time: 2.417s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.563s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16239141 41.76975300 + layer.0.v_cache 0.00001645 0.00671565 + layer.1.k_cache 0.33751179 8.80028870 + layer.1.v_cache 0.00000668 0.00273362 + layer.2.k_cache 0.00660169 0.77206914 + layer.2.v_cache 0.00002045 0.00770088 + layer.3.k_cache 0.02213682 6.09527881 + layer.3.v_cache 0.00002049 0.00877889 + layer.4.k_cache 0.00067824 0.23839208 + layer.4.v_cache 0.00005183 0.01574757 + layer.4.output 1.33690934 226.08458749 + ------------------------------------------------------------------------------------- + TOTAL 0.58163537 96.48879828 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 319228 +BPFP 1.2813 bits/point +EBPFP 2.5625 equivalent bits/point +MSE 96.488798 +---------------------- -------------------------------------------------------- +Time: 3.987s Load: 0.008s, Pack+Encode: 2.417s, Decode+Unpack: 1.563s +---------------------- -------------------------------------------------------- +💾 Converting with 96.4888 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,560B, BPFP=0.4991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,144B, BPFP=2.5737 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,108B, BPFP=1.2889 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,608B, BPFP=2.5424 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,284B, BPFP=1.4158 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,320B, BPFP=2.4674 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,932B, BPFP=1.3953 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,284B, BPFP=2.5236 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,988B, BPFP=2.2148 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,456B, BPFP=2.4753 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,388B, BPFP=0.3114 +⌛️ [2/4] FRONTEND: Frontend time: 2.079s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.748s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15796844 40.18062398 + layer.0.v_cache 0.00001527 0.00636096 + layer.1.k_cache 0.50235566 8.26697962 + layer.1.v_cache 0.00000626 0.00270149 + layer.2.k_cache 0.02709456 1.14553719 + layer.2.v_cache 0.00002063 0.00696833 + layer.3.k_cache 0.00856834 5.22762743 + layer.3.v_cache 0.00001945 0.00813584 + layer.4.k_cache 0.00071483 0.22390595 + layer.4.v_cache 0.00005464 0.01468960 + layer.4.output 0.00491706 203.52581956 + ------------------------------------------------------------------------------------- + TOTAL 0.04301397 87.04495690 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 370072 +BPFP 1.2692 bits/point +EBPFP 2.5384 equivalent bits/point +MSE 87.044957 +---------------------- -------------------------------------------------------- +Time: 3.835s Load: 0.008s, Pack+Encode: 2.079s, Decode+Unpack: 1.748s +---------------------- -------------------------------------------------------- +💾 Converting with 87.0450 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,696B, BPFP=0.4849 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,128B, BPFP=2.3392 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,688B, BPFP=1.3034 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,384B, BPFP=2.2923 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,352B, BPFP=1.5343 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,660B, BPFP=2.2467 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,984B, BPFP=1.4481 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,508B, BPFP=2.3002 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,644B, BPFP=1.9937 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,288B, BPFP=2.2233 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,740B, BPFP=0.3847 +⌛️ [2/4] FRONTEND: Frontend time: 2.357s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.605s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13184587 40.36543126 + layer.0.v_cache 0.00001706 0.00682125 + layer.1.k_cache 0.45546043 9.07988616 + layer.1.v_cache 0.00000678 0.00278021 + layer.2.k_cache 0.02272066 0.88377602 + layer.2.v_cache 0.00001996 0.00728512 + layer.3.k_cache 0.01643563 6.09764936 + layer.3.v_cache 0.00002069 0.00866105 + layer.4.k_cache 0.00071781 0.22318305 + layer.4.v_cache 0.00005262 0.01490310 + layer.4.output 1.23455937 209.40068764 + ------------------------------------------------------------------------------------- + TOTAL 0.54524783 89.55854059 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 331072 +BPFP 1.2270 bits/point +EBPFP 2.4540 equivalent bits/point +MSE 89.558541 +---------------------- -------------------------------------------------------- +Time: 3.971s Load: 0.010s, Pack+Encode: 2.357s, Decode+Unpack: 1.605s +---------------------- -------------------------------------------------------- +💾 Converting with 89.5585 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,440B, BPFP=0.4995 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,232B, BPFP=2.6179 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,988B, BPFP=1.3014 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,248B, BPFP=2.5597 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,372B, BPFP=1.3833 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,592B, BPFP=2.5208 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,312B, BPFP=1.3797 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,792B, BPFP=2.5919 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,164B, BPFP=2.1996 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,464B, BPFP=2.5133 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,424B, BPFP=0.3249 +⌛️ [2/4] FRONTEND: Frontend time: 2.069s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.719s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17705467 40.50932173 + layer.0.v_cache 0.00001633 0.00626656 + layer.1.k_cache 0.50406676 8.06025742 + layer.1.v_cache 0.00000599 0.00247843 + layer.2.k_cache 0.00944760 0.82782000 + layer.2.v_cache 0.00001883 0.00677501 + layer.3.k_cache 0.04120483 5.96680982 + layer.3.v_cache 0.00002392 0.00812373 + layer.4.k_cache 0.00070100 0.21183622 + layer.4.v_cache 0.00005002 0.01430799 + layer.4.output 0.00499618 201.30769751 + ------------------------------------------------------------------------------------- + TOTAL 0.04515078 86.16281644 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 369028 +BPFP 1.2848 bits/point +EBPFP 2.5695 equivalent bits/point +MSE 86.162816 +---------------------- -------------------------------------------------------- +Time: 3.797s Load: 0.009s, Pack+Encode: 2.069s, Decode+Unpack: 1.719s +---------------------- -------------------------------------------------------- +💾 Converting with 86.1628 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 28.682s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,664B, BPFP=0.5014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,592B, BPFP=2.5806 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,152B, BPFP=1.2819 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,676B, BPFP=2.5275 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,664B, BPFP=1.4273 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,764B, BPFP=2.4748 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,396B, BPFP=1.4118 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,680B, BPFP=2.5278 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,936B, BPFP=2.2532 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,732B, BPFP=2.4729 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,144B, BPFP=0.2988 +⌛️ [2/4] FRONTEND: Frontend time: 2.238s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.699s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20501094 40.17383174 + layer.0.v_cache 0.00001516 0.00630644 + layer.1.k_cache 0.57085758 8.36137334 + layer.1.v_cache 0.00000574 0.00252445 + layer.2.k_cache 0.01259345 1.03214812 + layer.2.v_cache 0.00002099 0.00711596 + layer.3.k_cache 0.01648921 6.14296875 + layer.3.v_cache 0.00001970 0.00818277 + layer.4.k_cache 0.00070444 0.23458410 + layer.4.v_cache 0.00005121 0.01492524 + layer.4.output 0.00487421 201.34161706 + ------------------------------------------------------------------------------------- + TOTAL 0.04940517 86.19854590 + (elements=2,350,080) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2350080 +Total Bytes 372400 +BPFP 1.2677 bits/point +EBPFP 2.5354 equivalent bits/point +MSE 86.198546 +---------------------- --------------------------------------------------------- +Time: 32.619s Load: 28.682s, Pack+Encode: 2.238s, Decode+Unpack: 1.699s +---------------------- --------------------------------------------------------- +💾 Converting with 86.1985 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,596B, BPFP=0.4956 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,896B, BPFP=2.5886 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,120B, BPFP=1.2754 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,700B, BPFP=2.5196 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,808B, BPFP=1.4880 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,052B, BPFP=2.4822 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,792B, BPFP=1.3718 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,604B, BPFP=2.5141 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,992B, BPFP=2.1905 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,708B, BPFP=2.4624 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,760B, BPFP=0.3440 +⌛️ [2/4] FRONTEND: Frontend time: 2.047s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.652s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14466163 40.68536670 + layer.0.v_cache 0.00001631 0.00666638 + layer.1.k_cache 0.53789613 8.38949348 + layer.1.v_cache 0.00000653 0.00266042 + layer.2.k_cache 0.02095415 0.84981748 + layer.2.v_cache 0.00002169 0.00733801 + layer.3.k_cache 0.03083320 5.36461547 + layer.3.v_cache 0.00002083 0.00842695 + layer.4.k_cache 0.00069737 0.22777439 + layer.4.v_cache 0.00004943 0.01450705 + layer.4.output 0.00490546 198.95418753 + ------------------------------------------------------------------------------------- + TOTAL 0.04526444 85.19035171 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 378028 +BPFP 1.2821 bits/point +EBPFP 2.5642 equivalent bits/point +MSE 85.190352 +---------------------- -------------------------------------------------------- +Time: 3.709s Load: 0.010s, Pack+Encode: 2.047s, Decode+Unpack: 1.652s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,416B, BPFP=0.4921 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,800B, BPFP=2.3934 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,484B, BPFP=1.3317 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,740B, BPFP=2.3380 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,360B, BPFP=1.5343 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,840B, BPFP=2.2910 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,424B, BPFP=1.4331 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,856B, BPFP=2.3441 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,812B, BPFP=2.0282 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,408B, BPFP=2.2684 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,204B, BPFP=0.3225 +⌛️ [2/4] FRONTEND: Frontend time: 2.208s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.616s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14240686 40.24498328 + layer.0.v_cache 0.00001673 0.00633565 + layer.1.k_cache 0.61438330 9.09189812 + layer.1.v_cache 0.00000638 0.00248439 + layer.2.k_cache 0.02027981 0.88367148 + layer.2.v_cache 0.00002012 0.00693984 + layer.3.k_cache 0.02199794 6.69938859 + layer.3.v_cache 0.00002003 0.00779425 + layer.4.k_cache 0.00078937 0.20354878 + layer.4.v_cache 0.00005038 0.01391980 + layer.4.output 0.04464151 176.89830686 + ------------------------------------------------------------------------------------- + TOTAL 0.06543891 76.20288895 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 396344 +BPFP 1.2184 bits/point +EBPFP 2.4367 equivalent bits/point +MSE 76.202889 +---------------------- -------------------------------------------------------- +Time: 3.835s Load: 0.011s, Pack+Encode: 2.208s, Decode+Unpack: 1.616s +---------------------- -------------------------------------------------------- +💾 Converting with 76.2029 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 399, 128) +Output shape: (1, 399, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.output: torch.Size([1, 399, 3584]) -> torch.Size([1, 1, 399, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,308B, BPFP=0.4820 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 62,724B, BPFP=2.4563 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,544B, BPFP=1.2353 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 61,040B, BPFP=2.3904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 37,332B, BPFP=1.4619 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 60,612B, BPFP=2.3736 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 35,704B, BPFP=1.3982 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 61,432B, BPFP=2.4057 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 53,800B, BPFP=2.1068 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 59,824B, BPFP=2.3427 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 56,848B, BPFP=0.3180 +⌛️ [2/4] FRONTEND: Frontend time: 2.366s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.925s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17889012 40.52157249 + layer.0.v_cache 0.00001635 0.00610311 + layer.1.k_cache 0.94984295 8.49470172 + layer.1.v_cache 0.00000590 0.00238267 + layer.2.k_cache 0.02305104 0.93167313 + layer.2.v_cache 0.00002114 0.00725804 + layer.3.k_cache 0.01620252 5.12106071 + layer.3.v_cache 0.00002049 0.00811008 + layer.4.k_cache 0.00073175 0.21684510 + layer.4.v_cache 0.00005255 0.01395179 + layer.4.output 0.00656733 137.35117929 + ------------------------------------------------------------------------------------- + TOTAL 0.07145918 59.81070081 + (elements=3,472,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3472896 +Total Bytes 533168 +BPFP 1.2282 bits/point +EBPFP 2.4564 equivalent bits/point +MSE 59.810701 +---------------------- -------------------------------------------------------- +Time: 4.305s Load: 0.014s, Pack+Encode: 2.366s, Decode+Unpack: 1.925s +---------------------- -------------------------------------------------------- +💾 Converting with 59.8107 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,104B, BPFP=0.4912 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,560B, BPFP=2.5277 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,480B, BPFP=1.3468 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,692B, BPFP=2.4676 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,908B, BPFP=1.5838 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,404B, BPFP=2.4477 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,556B, BPFP=1.4903 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,144B, BPFP=2.4989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,812B, BPFP=2.1994 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,064B, BPFP=2.4242 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,504B, BPFP=0.3507 +⌛️ [2/4] FRONTEND: Frontend time: 2.107s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.575s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10670873 42.93719752 + layer.0.v_cache 0.00001672 0.00726977 + layer.1.k_cache 0.29393762 9.17720612 + layer.1.v_cache 0.00000627 0.00277235 + layer.2.k_cache 0.01243114 0.81283387 + layer.2.v_cache 0.00002105 0.00815094 + layer.3.k_cache 0.00919804 6.29551980 + layer.3.v_cache 0.00002002 0.00917253 + layer.4.k_cache 0.00070102 0.24454063 + layer.4.v_cache 0.00005024 0.01649773 + layer.4.output 1.35463108 230.89234355 + ------------------------------------------------------------------------------------- + TOTAL 0.58267697 98.57397448 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 317228 +BPFP 1.2901 bits/point +EBPFP 2.5803 equivalent bits/point +MSE 98.573974 +---------------------- -------------------------------------------------------- +Time: 3.691s Load: 0.009s, Pack+Encode: 2.107s, Decode+Unpack: 1.575s +---------------------- -------------------------------------------------------- +💾 Converting with 98.5740 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,840B, BPFP=0.4823 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,300B, BPFP=2.2945 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,468B, BPFP=1.3206 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,416B, BPFP=2.2402 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,864B, BPFP=1.4680 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,056B, BPFP=2.2180 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,796B, BPFP=1.4023 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,060B, BPFP=2.2798 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,576B, BPFP=1.9424 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,764B, BPFP=2.2000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,876B, BPFP=0.3329 +⌛️ [2/4] FRONTEND: Frontend time: 2.377s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.690s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633087 39.88895639 + layer.0.v_cache 0.00001891 0.00657312 + layer.1.k_cache 0.36012592 9.77302804 + layer.1.v_cache 0.00000581 0.00244099 + layer.2.k_cache 0.01407751 0.88512048 + layer.2.v_cache 0.00002023 0.00729482 + layer.3.k_cache 0.01230825 6.79569659 + layer.3.v_cache 0.00002000 0.00837229 + layer.4.k_cache 0.00071033 0.21604268 + layer.4.v_cache 0.00004856 0.01430479 + layer.4.output 1.20534739 205.96692210 + ------------------------------------------------------------------------------------- + TOTAL 0.52535871 88.19801676 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 328016 +BPFP 1.1869 bits/point +EBPFP 2.3739 equivalent bits/point +MSE 88.198017 +---------------------- -------------------------------------------------------- +Time: 4.075s Load: 0.009s, Pack+Encode: 2.377s, Decode+Unpack: 1.690s +---------------------- -------------------------------------------------------- +💾 Converting with 88.1980 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,592B, BPFP=0.4922 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,544B, BPFP=2.3693 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,664B, BPFP=1.3397 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,760B, BPFP=2.3185 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,144B, BPFP=1.5654 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,408B, BPFP=2.2956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,952B, BPFP=1.4232 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,932B, BPFP=2.3296 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,828B, BPFP=2.0635 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,220B, BPFP=2.2835 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,008B, BPFP=0.3242 +⌛️ [2/4] FRONTEND: Frontend time: 2.452s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.551s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11142051 41.04243792 + layer.0.v_cache 0.00001650 0.00656307 + layer.1.k_cache 0.42032433 8.69775593 + layer.1.v_cache 0.00000619 0.00262744 + layer.2.k_cache 0.01443043 0.81898062 + layer.2.v_cache 0.00002030 0.00760930 + layer.3.k_cache 0.02895402 5.34050323 + layer.3.v_cache 0.00001914 0.00857143 + layer.4.k_cache 0.00067703 0.22573538 + layer.4.v_cache 0.00005693 0.01573312 + layer.4.output 1.27033357 202.61299644 + ------------------------------------------------------------------------------------- + TOTAL 0.55695649 86.73279368 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 320052 +BPFP 1.2206 bits/point +EBPFP 2.4412 equivalent bits/point +MSE 86.732794 +---------------------- -------------------------------------------------------- +Time: 4.014s Load: 0.010s, Pack+Encode: 2.452s, Decode+Unpack: 1.551s +---------------------- -------------------------------------------------------- +💾 Converting with 86.7328 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 302, 128) +Output shape: (1, 302, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.output: torch.Size([1, 302, 3584]) -> torch.Size([1, 1, 302, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,440B, BPFP=0.4884 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,396B, BPFP=2.3487 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,720B, BPFP=1.3307 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,228B, BPFP=2.2883 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,404B, BPFP=1.5213 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,756B, BPFP=2.2639 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,380B, BPFP=1.4166 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,464B, BPFP=2.3005 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,468B, BPFP=2.0420 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,352B, BPFP=2.2430 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,548B, BPFP=0.2627 +⌛️ [2/4] FRONTEND: Frontend time: 2.352s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.713s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15012774 40.91464456 + layer.0.v_cache 0.00001504 0.00626230 + layer.1.k_cache 0.60100308 8.74700806 + layer.1.v_cache 0.00000575 0.00253281 + layer.2.k_cache 0.00814273 0.88620409 + layer.2.v_cache 0.00002032 0.00754820 + layer.3.k_cache 0.01797006 6.07666299 + layer.3.v_cache 0.00001899 0.00835660 + layer.4.k_cache 0.00070977 0.24889591 + layer.4.v_cache 0.00004868 0.01494665 + layer.4.output 0.04414053 176.37479305 + ------------------------------------------------------------------------------------- + TOTAL 0.06394387 75.97274197 + (elements=2,628,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2628608 +Total Bytes 388156 +BPFP 1.1813 bits/point +EBPFP 2.3627 equivalent bits/point +MSE 75.972742 +---------------------- -------------------------------------------------------- +Time: 4.077s Load: 0.012s, Pack+Encode: 2.352s, Decode+Unpack: 1.713s +---------------------- -------------------------------------------------------- +💾 Converting with 75.9727 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,616B, BPFP=0.5023 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,612B, BPFP=2.6010 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,596B, BPFP=1.2591 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,396B, BPFP=2.5301 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,312B, BPFP=1.4174 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,792B, BPFP=2.4949 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,228B, BPFP=1.3542 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,640B, BPFP=2.5443 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,912B, BPFP=2.2104 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,668B, BPFP=2.4876 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,016B, BPFP=0.3583 +⌛️ [2/4] FRONTEND: Frontend time: 2.099s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 26.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14350806 39.31641718 + layer.0.v_cache 0.00001655 0.00620449 + layer.1.k_cache 0.46812200 8.79830796 + layer.1.v_cache 0.00000598 0.00247711 + layer.2.k_cache 0.01096032 0.89073625 + layer.2.v_cache 0.00002019 0.00701909 + layer.3.k_cache 0.00658789 5.57003238 + layer.3.v_cache 0.00002014 0.00803755 + layer.4.k_cache 0.00067528 0.21375976 + layer.4.v_cache 0.00005785 0.01491028 + layer.4.output 0.00496594 201.72051572 + ------------------------------------------------------------------------------------- + TOTAL 0.03910211 86.28655954 + (elements=2,332,672) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2332672 +Total Bytes 375788 +BPFP 1.2888 bits/point +EBPFP 2.5776 equivalent bits/point +MSE 86.286560 +---------------------- --------------------------------------------------------- +Time: 28.128s Load: 0.009s, Pack+Encode: 2.099s, Decode+Unpack: 26.020s +---------------------- --------------------------------------------------------- +💾 Converting with 86.2866 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 260, 128) +Output shape: (1, 260, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.output: torch.Size([1, 260, 3584]) -> torch.Size([1, 1, 260, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,208B, BPFP=0.4933 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,568B, BPFP=2.6183 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,904B, BPFP=1.3163 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,460B, BPFP=2.5517 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,192B, BPFP=1.3938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,868B, BPFP=2.5161 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,036B, BPFP=1.3243 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,372B, BPFP=2.5464 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,084B, BPFP=2.2286 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,348B, BPFP=2.4849 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,612B, BPFP=0.3057 +⌛️ [2/4] FRONTEND: Frontend time: 2.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.666s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14548651 41.01817157 + layer.0.v_cache 0.00001584 0.00659040 + layer.1.k_cache 0.47062912 8.66211125 + layer.1.v_cache 0.00000604 0.00262415 + layer.2.k_cache 0.00619981 0.94913841 + layer.2.v_cache 0.00002003 0.00777556 + layer.3.k_cache 0.01299995 4.90039485 + layer.3.v_cache 0.00001981 0.00879961 + layer.4.k_cache 0.00067777 0.26575849 + layer.4.v_cache 0.00006192 0.01574745 + layer.4.output 0.00504520 207.16644918 + ------------------------------------------------------------------------------------- + TOTAL 0.03949607 88.58836800 + (elements=2,263,040) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2263040 +Total Bytes 359652 +BPFP 1.2714 bits/point +EBPFP 2.5428 equivalent bits/point +MSE 88.588368 +---------------------- -------------------------------------------------------- +Time: 3.904s Load: 0.008s, Pack+Encode: 2.230s, Decode+Unpack: 1.666s +---------------------- -------------------------------------------------------- +💾 Converting with 88.5884 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,664B, BPFP=0.4977 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,756B, BPFP=2.5710 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,032B, BPFP=1.2656 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,928B, BPFP=2.5234 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,192B, BPFP=1.5046 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,208B, BPFP=2.4821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,532B, BPFP=1.4092 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,708B, BPFP=2.5108 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,244B, BPFP=2.1969 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,992B, BPFP=2.4697 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,304B, BPFP=0.3308 +⌛️ [2/4] FRONTEND: Frontend time: 2.075s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.733s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14555165 41.01823156 + layer.0.v_cache 0.00001661 0.00669734 + layer.1.k_cache 0.55468144 8.62746115 + layer.1.v_cache 0.00000621 0.00263068 + layer.2.k_cache 0.02243149 0.86330436 + layer.2.v_cache 0.00002103 0.00734603 + layer.3.k_cache 0.02028967 6.14267596 + layer.3.v_cache 0.00002036 0.00832606 + layer.4.k_cache 0.00072975 0.21834756 + layer.4.v_cache 0.00005124 0.01483800 + layer.4.output 0.00486695 199.38332130 + ------------------------------------------------------------------------------------- + TOTAL 0.04575695 85.44665340 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 378560 +BPFP 1.2792 bits/point +EBPFP 2.5584 equivalent bits/point +MSE 85.446653 +---------------------- -------------------------------------------------------- +Time: 3.817s Load: 0.009s, Pack+Encode: 2.075s, Decode+Unpack: 1.733s +---------------------- -------------------------------------------------------- +💾 Converting with 85.4467 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 356, 128) +Output shape: (1, 356, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.output: torch.Size([1, 356, 3584]) -> torch.Size([1, 1, 356, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,016B, BPFP=0.4835 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,020B, BPFP=2.4149 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,936B, BPFP=1.3578 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,676B, BPFP=2.3559 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 35,104B, BPFP=1.5407 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,828B, BPFP=2.3186 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,952B, BPFP=1.4024 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,872B, BPFP=2.3645 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,352B, BPFP=2.0344 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,116B, BPFP=2.2874 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 51,376B, BPFP=0.3221 +⌛️ [2/4] FRONTEND: Frontend time: 2.460s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.825s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15223843 40.05422116 + layer.0.v_cache 0.00001740 0.00619939 + layer.1.k_cache 0.82972649 8.36588107 + layer.1.v_cache 0.00000631 0.00250847 + layer.2.k_cache 0.03086275 0.97733058 + layer.2.v_cache 0.00001960 0.00682314 + layer.3.k_cache 0.01100341 6.73659970 + layer.3.v_cache 0.00002062 0.00738413 + layer.4.k_cache 0.00073857 0.20064429 + layer.4.v_cache 0.00005279 0.01348202 + layer.4.output 0.03758925 150.48366021 + ------------------------------------------------------------------------------------- + TOTAL 0.07575360 65.27980561 + (elements=3,098,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3098624 +Total Bytes 474248 +BPFP 1.2244 bits/point +EBPFP 2.4488 equivalent bits/point +MSE 65.279806 +---------------------- -------------------------------------------------------- +Time: 4.298s Load: 0.014s, Pack+Encode: 2.460s, Decode+Unpack: 1.825s +---------------------- -------------------------------------------------------- +💾 Converting with 65.2798 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,556B, BPFP=0.5026 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,272B, BPFP=2.5418 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,080B, BPFP=1.2383 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,152B, BPFP=2.5348 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,996B, BPFP=1.4095 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,932B, BPFP=2.5219 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,680B, BPFP=1.3322 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,240B, BPFP=2.5399 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,184B, BPFP=2.2430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,348B, BPFP=2.4875 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,328B, BPFP=0.3720 +⌛️ [2/4] FRONTEND: Frontend time: 2.170s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.646s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12347319 40.67789224 + layer.0.v_cache 0.00001696 0.00665138 + layer.1.k_cache 0.44284580 9.19793288 + layer.1.v_cache 0.00000660 0.00263071 + layer.2.k_cache 0.00882101 0.86030028 + layer.2.v_cache 0.00002210 0.00783840 + layer.3.k_cache 0.01145097 4.20909451 + layer.3.v_cache 0.00001984 0.00837923 + layer.4.k_cache 0.00068754 0.21073673 + layer.4.v_cache 0.00006483 0.01529796 + layer.4.output 0.00500964 200.47900443 + ------------------------------------------------------------------------------------- + TOTAL 0.03661626 85.79704620 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 373768 +BPFP 1.2915 bits/point +EBPFP 2.5830 equivalent bits/point +MSE 85.797046 +---------------------- -------------------------------------------------------- +Time: 3.827s Load: 0.010s, Pack+Encode: 2.170s, Decode+Unpack: 1.646s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,820B, BPFP=0.4836 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,912B, BPFP=2.5171 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,752B, BPFP=1.2474 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,564B, BPFP=2.4432 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,164B, BPFP=1.5989 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,068B, BPFP=2.4160 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,564B, BPFP=1.4015 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,824B, BPFP=2.4575 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,536B, BPFP=2.1675 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,884B, BPFP=2.4059 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,420B, BPFP=0.3244 +⌛️ [2/4] FRONTEND: Frontend time: 2.119s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.820s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13002123 39.86102659 + layer.0.v_cache 0.00001833 0.00685784 + layer.1.k_cache 0.50397591 8.32622070 + layer.1.v_cache 0.00000608 0.00257882 + layer.2.k_cache 0.01926375 0.91682750 + layer.2.v_cache 0.00002060 0.00751055 + layer.3.k_cache 0.04992905 5.40693274 + layer.3.v_cache 0.00002096 0.00879789 + layer.4.k_cache 0.00069540 0.24857263 + layer.4.v_cache 0.00005467 0.01559149 + layer.4.output 0.00466491 185.39705514 + ------------------------------------------------------------------------------------- + TOTAL 0.04333296 79.56354722 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 390508 +BPFP 1.2594 bits/point +EBPFP 2.5188 equivalent bits/point +MSE 79.563547 +---------------------- -------------------------------------------------------- +Time: 3.949s Load: 0.010s, Pack+Encode: 2.119s, Decode+Unpack: 1.820s +---------------------- -------------------------------------------------------- +💾 Converting with 79.5635 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,908B, BPFP=0.4971 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,684B, BPFP=2.5493 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,328B, BPFP=1.2460 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,596B, BPFP=2.4886 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,492B, BPFP=1.4225 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,756B, BPFP=2.4417 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,208B, BPFP=1.3509 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,892B, BPFP=2.5051 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,692B, BPFP=2.1592 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,784B, BPFP=2.4433 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 48,620B, BPFP=0.3876 +⌛️ [2/4] FRONTEND: Frontend time: 2.750s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.836s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15974760 40.07991071 + layer.0.v_cache 0.00001739 0.00632706 + layer.1.k_cache 0.45644542 8.55712280 + layer.1.v_cache 0.00000643 0.00249107 + layer.2.k_cache 0.01631587 0.82129386 + layer.2.v_cache 0.00002005 0.00668516 + layer.3.k_cache 0.01511352 5.64049683 + layer.3.v_cache 0.00002063 0.00791099 + layer.4.k_cache 0.00073517 0.19540747 + layer.4.v_cache 0.00005091 0.01420514 + layer.4.output 0.00481773 189.44813457 + ------------------------------------------------------------------------------------- + TOTAL 0.04012924 81.26287018 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 390960 +BPFP 1.2834 bits/point +EBPFP 2.5667 equivalent bits/point +MSE 81.262870 +---------------------- -------------------------------------------------------- +Time: 4.596s Load: 0.011s, Pack+Encode: 2.750s, Decode+Unpack: 1.836s +---------------------- -------------------------------------------------------- +💾 Converting with 81.2629 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,544B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,304B, BPFP=2.5927 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,876B, BPFP=1.2802 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,632B, BPFP=2.5534 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,340B, BPFP=1.4244 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,072B, BPFP=2.5206 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,600B, BPFP=1.3811 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,804B, BPFP=2.5634 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,816B, BPFP=2.2130 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,612B, BPFP=2.4937 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,144B, BPFP=0.3105 +⌛️ [2/4] FRONTEND: Frontend time: 2.210s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.729s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14148096 41.15703271 + layer.0.v_cache 0.00001588 0.00636422 + layer.1.k_cache 0.55266568 8.79967484 + layer.1.v_cache 0.00000632 0.00256701 + layer.2.k_cache 0.00687984 0.99129369 + layer.2.v_cache 0.00001885 0.00730715 + layer.3.k_cache 0.01675770 5.68340447 + layer.3.v_cache 0.00001964 0.00836693 + layer.4.k_cache 0.00068991 0.22183243 + layer.4.v_cache 0.00004859 0.01459446 + layer.4.output 0.00492290 204.60814607 + ------------------------------------------------------------------------------------- + TOTAL 0.04429669 87.59702708 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 370744 +BPFP 1.2762 bits/point +EBPFP 2.5525 equivalent bits/point +MSE 87.597027 +---------------------- -------------------------------------------------------- +Time: 3.948s Load: 0.010s, Pack+Encode: 2.210s, Decode+Unpack: 1.729s +---------------------- -------------------------------------------------------- +💾 Converting with 87.5970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,932B, BPFP=0.4805 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 52,840B, BPFP=2.5561 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,552B, BPFP=1.2844 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 51,240B, BPFP=2.4787 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,940B, BPFP=1.4000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 49,564B, BPFP=2.3976 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,628B, BPFP=1.3365 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 51,164B, BPFP=2.4750 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,868B, BPFP=2.0737 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 48,660B, BPFP=2.3539 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 45,092B, BPFP=0.3116 +⌛️ [2/4] FRONTEND: Frontend time: 2.282s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.850s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13144862 41.15717516 + layer.0.v_cache 0.00001769 0.00669927 + layer.1.k_cache 0.64513603 8.64286913 + layer.1.v_cache 0.00000607 0.00251766 + layer.2.k_cache 0.02726801 0.80205760 + layer.2.v_cache 0.00001956 0.00702521 + layer.3.k_cache 0.02271994 4.14485060 + layer.3.v_cache 0.00002018 0.00791849 + layer.4.k_cache 0.00073057 0.20163665 + layer.4.v_cache 0.00004940 0.01360759 + layer.4.output 0.04135688 163.77808492 + ------------------------------------------------------------------------------------- + TOTAL 0.06570084 70.67252658 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 434480 +BPFP 1.2363 bits/point +EBPFP 2.4727 equivalent bits/point +MSE 70.672527 +---------------------- -------------------------------------------------------- +Time: 4.142s Load: 0.010s, Pack+Encode: 2.282s, Decode+Unpack: 1.850s +---------------------- -------------------------------------------------------- +💾 Converting with 70.6725 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,960B, BPFP=0.5058 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,396B, BPFP=2.6451 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,552B, BPFP=1.2756 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,348B, BPFP=2.5689 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,952B, BPFP=1.4500 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,716B, BPFP=2.5230 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,700B, BPFP=1.4317 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,716B, BPFP=2.5956 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,464B, BPFP=2.2866 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,984B, BPFP=2.5424 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,756B, BPFP=0.2985 +⌛️ [2/4] FRONTEND: Frontend time: 2.382s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.446s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12396691 41.47867006 + layer.0.v_cache 0.00001598 0.00697253 + layer.1.k_cache 0.24854917 8.45738554 + layer.1.v_cache 0.00000584 0.00261670 + layer.2.k_cache 0.02290044 0.88908436 + layer.2.v_cache 0.00002299 0.00760515 + layer.3.k_cache 0.01294550 5.88470260 + layer.3.v_cache 0.00002077 0.00888863 + layer.4.k_cache 0.00065144 0.23406209 + layer.4.v_cache 0.00005054 0.01484262 + layer.4.output 1.42385976 241.98133306 + ------------------------------------------------------------------------------------- + TOTAL 0.61036164 102.99142127 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 301544 +BPFP 1.2891 bits/point +EBPFP 2.5782 equivalent bits/point +MSE 102.991421 +---------------------- -------------------------------------------------------- +Time: 3.839s Load: 0.011s, Pack+Encode: 2.382s, Decode+Unpack: 1.446s +---------------------- -------------------------------------------------------- +💾 Converting with 102.9914 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,828B, BPFP=0.4773 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,648B, BPFP=2.4680 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,480B, BPFP=1.3235 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,460B, BPFP=2.4038 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,812B, BPFP=1.5577 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,004B, BPFP=2.3791 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,156B, BPFP=1.4141 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,900B, BPFP=2.4276 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,428B, BPFP=2.1317 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,508B, BPFP=2.3523 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,656B, BPFP=0.3140 +⌛️ [2/4] FRONTEND: Frontend time: 2.304s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.580s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14783513 41.97127419 + layer.0.v_cache 0.00001600 0.00684697 + layer.1.k_cache 0.54442752 8.59573354 + layer.1.v_cache 0.00000628 0.00266642 + layer.2.k_cache 0.01848349 0.79511269 + layer.2.v_cache 0.00002061 0.00753417 + layer.3.k_cache 0.01777397 6.44542654 + layer.3.v_cache 0.00002092 0.00875323 + layer.4.k_cache 0.00070447 0.22735271 + layer.4.v_cache 0.00005049 0.01490637 + layer.4.output 0.04615090 175.49485603 + ------------------------------------------------------------------------------------- + TOTAL 0.06190560 75.67879994 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 390880 +BPFP 1.2431 bits/point +EBPFP 2.4863 equivalent bits/point +MSE 75.678800 +---------------------- -------------------------------------------------------- +Time: 3.896s Load: 0.012s, Pack+Encode: 2.304s, Decode+Unpack: 1.580s +---------------------- -------------------------------------------------------- +💾 Converting with 75.6788 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,348B, BPFP=0.4901 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,888B, BPFP=2.4060 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,916B, BPFP=1.3064 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,652B, BPFP=2.3412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,684B, BPFP=1.5564 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,968B, BPFP=2.3054 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,396B, BPFP=1.3840 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,808B, BPFP=2.3494 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,148B, BPFP=2.0526 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,456B, BPFP=2.2785 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,296B, BPFP=0.3018 +⌛️ [2/4] FRONTEND: Frontend time: 2.290s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.630s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16274841 40.51521537 + layer.0.v_cache 0.00001794 0.00671629 + layer.1.k_cache 0.56737488 9.14303527 + layer.1.v_cache 0.00000586 0.00248348 + layer.2.k_cache 0.02027899 0.81000985 + layer.2.v_cache 0.00002117 0.00723729 + layer.3.k_cache 0.01891868 5.53504688 + layer.3.v_cache 0.00002140 0.00802005 + layer.4.k_cache 0.00069658 0.21610967 + layer.4.v_cache 0.00005005 0.01466116 + layer.4.output 0.04476916 178.63355405 + ------------------------------------------------------------------------------------- + TOTAL 0.06373636 76.86431845 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 392560 +BPFP 1.2108 bits/point +EBPFP 2.4215 equivalent bits/point +MSE 76.864318 +---------------------- -------------------------------------------------------- +Time: 3.931s Load: 0.010s, Pack+Encode: 2.290s, Decode+Unpack: 1.630s +---------------------- -------------------------------------------------------- +💾 Converting with 76.8643 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,616B, BPFP=0.4931 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,096B, BPFP=2.5810 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,124B, BPFP=1.2663 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,964B, BPFP=2.5163 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,404B, BPFP=1.4540 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,220B, BPFP=2.4737 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,104B, BPFP=1.3796 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,992B, BPFP=2.5179 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,040B, BPFP=2.1772 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,088B, BPFP=2.4661 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,360B, BPFP=0.3136 +⌛️ [2/4] FRONTEND: Frontend time: 2.551s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.584s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13366390 41.98673234 + layer.0.v_cache 0.00001719 0.00662961 + layer.1.k_cache 0.57618574 8.24609017 + layer.1.v_cache 0.00000658 0.00251168 + layer.2.k_cache 0.02858028 0.84021667 + layer.2.v_cache 0.00002021 0.00728933 + layer.3.k_cache 0.00873641 5.48761635 + layer.3.v_cache 0.00002023 0.00829222 + layer.4.k_cache 0.00067337 0.20597686 + layer.4.v_cache 0.00005011 0.01421752 + layer.4.output 0.00484872 192.59836146 + ------------------------------------------------------------------------------------- + TOTAL 0.04599383 82.64671194 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 376008 +BPFP 1.2659 bits/point +EBPFP 2.5318 equivalent bits/point +MSE 82.646712 +---------------------- -------------------------------------------------------- +Time: 4.145s Load: 0.011s, Pack+Encode: 2.551s, Decode+Unpack: 1.584s +---------------------- -------------------------------------------------------- +💾 Converting with 82.6467 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,960B, BPFP=0.4921 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,748B, BPFP=2.5981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,980B, BPFP=1.3419 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,740B, BPFP=2.5269 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,888B, BPFP=1.6182 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,400B, BPFP=2.5028 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,076B, BPFP=1.4901 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,152B, BPFP=2.5560 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,928B, BPFP=2.2574 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,132B, BPFP=2.4839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,948B, BPFP=0.3833 +⌛️ [2/4] FRONTEND: Frontend time: 2.138s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.560s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13057878 42.20772059 + layer.0.v_cache 0.00001797 0.00735973 + layer.1.k_cache 0.31658528 9.36463120 + layer.1.v_cache 0.00000647 0.00274885 + layer.2.k_cache 0.02849985 0.78578580 + layer.2.v_cache 0.00001991 0.00790516 + layer.3.k_cache 0.03599559 5.92251297 + layer.3.v_cache 0.00002239 0.00974667 + layer.4.k_cache 0.00078811 0.23882528 + layer.4.v_cache 0.00005540 0.01623343 + layer.4.output 1.38530015 238.50979719 + ------------------------------------------------------------------------------------- + TOTAL 0.60056887 101.65482647 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 318952 +BPFP 1.3265 bits/point +EBPFP 2.6530 equivalent bits/point +MSE 101.654826 +---------------------- -------------------------------------------------------- +Time: 3.706s Load: 0.008s, Pack+Encode: 2.138s, Decode+Unpack: 1.560s +---------------------- -------------------------------------------------------- +💾 Converting with 101.6548 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,964B, BPFP=0.5014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,940B, BPFP=2.6599 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,892B, BPFP=1.2883 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,912B, BPFP=2.5858 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,288B, BPFP=1.6048 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,444B, BPFP=2.5521 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,324B, BPFP=1.3914 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,252B, BPFP=2.6103 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,860B, BPFP=2.2941 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,224B, BPFP=2.5363 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,752B, BPFP=0.3678 +⌛️ [2/4] FRONTEND: Frontend time: 2.025s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.533s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12725546 40.90178571 + layer.0.v_cache 0.00001736 0.00733582 + layer.1.k_cache 0.32262681 8.51575382 + layer.1.v_cache 0.00000608 0.00270872 + layer.2.k_cache 0.02157613 0.79811553 + layer.2.v_cache 0.00002208 0.00777687 + layer.3.k_cache 0.01760322 5.42543079 + layer.3.v_cache 0.00002192 0.00934303 + layer.4.k_cache 0.00069785 0.23256128 + layer.4.v_cache 0.00005172 0.01570242 + layer.4.output 1.41081570 246.20490454 + ------------------------------------------------------------------------------------- + TOTAL 0.60974050 104.66769681 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 313852 +BPFP 1.3293 bits/point +EBPFP 2.6587 equivalent bits/point +MSE 104.667697 +---------------------- -------------------------------------------------------- +Time: 3.565s Load: 0.007s, Pack+Encode: 2.025s, Decode+Unpack: 1.533s +---------------------- -------------------------------------------------------- +💾 Converting with 104.6677 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,832B, BPFP=0.5012 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,732B, BPFP=2.6945 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,668B, BPFP=1.2961 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,748B, BPFP=2.6224 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,316B, BPFP=1.5637 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,188B, BPFP=2.5813 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,060B, BPFP=1.3982 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,168B, BPFP=2.6532 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,632B, BPFP=2.3204 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,968B, BPFP=2.5651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,616B, BPFP=0.3837 +⌛️ [2/4] FRONTEND: Frontend time: 2.032s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.600s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13788442 41.76352516 + layer.0.v_cache 0.00001717 0.00758101 + layer.1.k_cache 0.30723858 8.81415741 + layer.1.v_cache 0.00000664 0.00303775 + layer.2.k_cache 0.01337219 0.77323534 + layer.2.v_cache 0.00002159 0.00839266 + layer.3.k_cache 0.03522867 5.30960470 + layer.3.v_cache 0.00002126 0.00964969 + layer.4.k_cache 0.00073009 0.23604218 + layer.4.v_cache 0.00005119 0.01662491 + layer.4.output 1.43730973 238.41163649 + ------------------------------------------------------------------------------------- + TOTAL 0.62092588 101.51901801 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 311928 +BPFP 1.3460 bits/point +EBPFP 2.6920 equivalent bits/point +MSE 101.519018 +---------------------- -------------------------------------------------------- +Time: 3.639s Load: 0.007s, Pack+Encode: 2.032s, Decode+Unpack: 1.600s +---------------------- -------------------------------------------------------- +💾 Converting with 101.5190 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,736B, BPFP=0.5159 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,396B, BPFP=2.7111 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,580B, BPFP=1.2699 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,836B, BPFP=2.6682 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,984B, BPFP=1.5306 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,648B, BPFP=2.6538 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,516B, BPFP=1.4182 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,864B, BPFP=2.6703 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,820B, BPFP=2.3606 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,260B, BPFP=2.6241 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,544B, BPFP=0.4108 +⌛️ [2/4] FRONTEND: Frontend time: 1.982s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.494s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11546621 41.92207606 + layer.0.v_cache 0.00001823 0.00725474 + layer.1.k_cache 0.26924565 9.57586849 + layer.1.v_cache 0.00000737 0.00284028 + layer.2.k_cache 0.01761903 0.74198839 + layer.2.v_cache 0.00002145 0.00787657 + layer.3.k_cache 0.02374693 5.09160061 + layer.3.v_cache 0.00002104 0.00914027 + layer.4.k_cache 0.00067272 0.23097536 + layer.4.v_cache 0.00005153 0.01549128 + layer.4.output 1.50071327 258.54199492 + ------------------------------------------------------------------------------------- + TOTAL 0.64305077 109.84700450 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 304184 +BPFP 1.3705 bits/point +EBPFP 2.7410 equivalent bits/point +MSE 109.847005 +---------------------- -------------------------------------------------------- +Time: 3.483s Load: 0.007s, Pack+Encode: 1.982s, Decode+Unpack: 1.494s +---------------------- -------------------------------------------------------- +💾 Converting with 109.8470 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,524B, BPFP=0.4832 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,112B, BPFP=2.3393 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,892B, BPFP=1.3135 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,804B, BPFP=2.2729 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,560B, BPFP=1.5503 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,352B, BPFP=2.2500 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,924B, BPFP=1.4166 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,116B, BPFP=2.2888 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,792B, BPFP=2.0187 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,008B, BPFP=2.2325 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 45,680B, BPFP=0.3311 +⌛️ [2/4] FRONTEND: Frontend time: 30.559s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.896s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18206066 41.23943854 + layer.0.v_cache 0.00001652 0.00678872 + layer.1.k_cache 0.62405802 8.43606686 + layer.1.v_cache 0.00000616 0.00266950 + layer.2.k_cache 0.02036299 0.91098013 + layer.2.v_cache 0.00002087 0.00777931 + layer.3.k_cache 0.02546444 5.95904224 + layer.3.v_cache 0.00002064 0.00861371 + layer.4.k_cache 0.00069880 0.23359799 + layer.4.v_cache 0.00007146 0.01536708 + layer.4.output 0.04336799 172.68718112 + ------------------------------------------------------------------------------------- + TOTAL 0.06802097 74.44885953 + (elements=2,680,832) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2680832 +Total Bytes 403764 +BPFP 1.2049 bits/point +EBPFP 2.4098 equivalent bits/point +MSE 74.448860 +---------------------- --------------------------------------------------------- +Time: 32.465s Load: 0.010s, Pack+Encode: 30.559s, Decode+Unpack: 1.896s +---------------------- --------------------------------------------------------- +💾 Converting with 74.4489 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,312B, BPFP=0.4989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,720B, BPFP=2.5055 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,236B, BPFP=1.3125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,904B, BPFP=2.4498 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,812B, BPFP=1.6247 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,212B, BPFP=2.4026 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,360B, BPFP=1.4574 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,000B, BPFP=2.4563 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,244B, BPFP=2.1318 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,836B, BPFP=2.3769 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,868B, BPFP=0.3691 +⌛️ [2/4] FRONTEND: Frontend time: 2.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.584s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14018573 41.20046824 + layer.0.v_cache 0.00001653 0.00695607 + layer.1.k_cache 0.40671090 9.05025672 + layer.1.v_cache 0.00000662 0.00275828 + layer.2.k_cache 0.03036383 0.69982524 + layer.2.v_cache 0.00002088 0.00778199 + layer.3.k_cache 0.05940984 5.33319545 + layer.3.v_cache 0.00002025 0.00852921 + layer.4.k_cache 0.00071161 0.22707160 + layer.4.v_cache 0.00005094 0.01499004 + layer.4.output 1.33692960 226.94880692 + ------------------------------------------------------------------------------------- + TOTAL 0.58800025 96.77608714 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 319504 +BPFP 1.2824 bits/point +EBPFP 2.5647 equivalent bits/point +MSE 96.776087 +---------------------- -------------------------------------------------------- +Time: 3.842s Load: 0.009s, Pack+Encode: 2.249s, Decode+Unpack: 1.584s +---------------------- -------------------------------------------------------- +💾 Converting with 96.7761 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,000B, BPFP=0.4994 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,828B, BPFP=2.6276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,520B, BPFP=1.3213 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,820B, BPFP=2.5557 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,440B, BPFP=1.6724 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,416B, BPFP=2.5268 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,940B, BPFP=1.4227 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,144B, BPFP=2.5788 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,952B, BPFP=2.2797 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,260B, BPFP=2.5157 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,656B, BPFP=0.3736 +⌛️ [2/4] FRONTEND: Frontend time: 1.997s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.547s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351684 40.93623805 + layer.0.v_cache 0.00001872 0.00737009 + layer.1.k_cache 0.28219827 8.83861023 + layer.1.v_cache 0.00000627 0.00285327 + layer.2.k_cache 0.01639774 0.73939723 + layer.2.v_cache 0.00002172 0.00833012 + layer.3.k_cache 0.02752649 4.82992317 + layer.3.v_cache 0.00002090 0.00953271 + layer.4.k_cache 0.00067264 0.23892755 + layer.4.v_cache 0.00005058 0.01601686 + layer.4.output 1.39794763 242.85229126 + ------------------------------------------------------------------------------------- + TOTAL 0.60212139 103.27019048 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 316976 +BPFP 1.3303 bits/point +EBPFP 2.6606 equivalent bits/point +MSE 103.270190 +---------------------- -------------------------------------------------------- +Time: 3.552s Load: 0.008s, Pack+Encode: 1.997s, Decode+Unpack: 1.547s +---------------------- -------------------------------------------------------- +💾 Converting with 103.2702 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,556B, BPFP=0.4988 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,248B, BPFP=2.5798 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,368B, BPFP=1.3041 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,408B, BPFP=2.5308 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,560B, BPFP=1.4319 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,840B, BPFP=2.4977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,140B, BPFP=1.4657 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,388B, BPFP=2.5296 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,484B, BPFP=2.2437 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,040B, BPFP=2.4510 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,272B, BPFP=0.3604 +⌛️ [2/4] FRONTEND: Frontend time: 2.535s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.948s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15646637 41.28224114 + layer.0.v_cache 0.00001727 0.00682933 + layer.1.k_cache 0.53079696 8.70799961 + layer.1.v_cache 0.00000635 0.00279389 + layer.2.k_cache 0.02340859 0.78023461 + layer.2.v_cache 0.00002061 0.00762848 + layer.3.k_cache 0.01050839 6.22706490 + layer.3.v_cache 0.00002156 0.00839935 + layer.4.k_cache 0.00068596 0.23505413 + layer.4.v_cache 0.00006072 0.01489996 + layer.4.output 0.00497423 201.72389725 + ------------------------------------------------------------------------------------- + TOTAL 0.04451838 86.43178978 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 378304 +BPFP 1.2974 bits/point +EBPFP 2.5948 equivalent bits/point +MSE 86.431790 +---------------------- -------------------------------------------------------- +Time: 4.493s Load: 0.010s, Pack+Encode: 2.535s, Decode+Unpack: 1.948s +---------------------- -------------------------------------------------------- +💾 Converting with 86.4318 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,544B, BPFP=0.4842 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,080B, BPFP=2.3377 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,448B, BPFP=1.3417 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,128B, BPFP=2.2894 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,496B, BPFP=1.4963 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,400B, BPFP=2.2524 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,104B, BPFP=1.4257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,188B, BPFP=2.2924 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,532B, BPFP=2.0055 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,896B, BPFP=2.2269 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 49,548B, BPFP=0.3591 +⌛️ [2/4] FRONTEND: Frontend time: 2.472s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.655s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15292851 40.31985593 + layer.0.v_cache 0.00001614 0.00694105 + layer.1.k_cache 0.61736927 9.36423165 + layer.1.v_cache 0.00000669 0.00276081 + layer.2.k_cache 0.02003183 0.88782154 + layer.2.v_cache 0.00002344 0.00766808 + layer.3.k_cache 0.01028479 5.80363365 + layer.3.v_cache 0.00002212 0.00822492 + layer.4.k_cache 0.00070434 0.25151060 + layer.4.v_cache 0.00006490 0.01487357 + layer.4.output 0.04342050 173.84488057 + ------------------------------------------------------------------------------------- + TOTAL 0.06502327 74.91656975 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 407364 +BPFP 1.2156 bits/point +EBPFP 2.4313 equivalent bits/point +MSE 74.916570 +---------------------- -------------------------------------------------------- +Time: 4.139s Load: 0.011s, Pack+Encode: 2.472s, Decode+Unpack: 1.655s +---------------------- -------------------------------------------------------- +💾 Converting with 74.9166 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,552B, BPFP=0.4967 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,952B, BPFP=2.6111 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,220B, BPFP=1.2907 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,512B, BPFP=2.5274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,900B, BPFP=1.5044 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,368B, BPFP=2.5191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,144B, BPFP=1.4024 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,972B, BPFP=2.5541 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,380B, BPFP=2.2293 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,524B, BPFP=2.4700 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,952B, BPFP=0.3066 +⌛️ [2/4] FRONTEND: Frontend time: 2.388s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.726s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12950333 40.82161216 + layer.0.v_cache 0.00001690 0.00686020 + layer.1.k_cache 0.52417032 8.70467062 + layer.1.v_cache 0.00000602 0.00269085 + layer.2.k_cache 0.01578268 0.87987782 + layer.2.v_cache 0.00002268 0.00810451 + layer.3.k_cache 0.03345619 4.90917424 + layer.3.v_cache 0.00002007 0.00906346 + layer.4.k_cache 0.00067301 0.23959697 + layer.4.v_cache 0.00005077 0.01542295 + layer.4.output 0.00489311 202.36293481 + ------------------------------------------------------------------------------------- + TOTAL 0.04340905 86.59633044 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 374476 +BPFP 1.2795 bits/point +EBPFP 2.5590 equivalent bits/point +MSE 86.596330 +---------------------- -------------------------------------------------------- +Time: 4.123s Load: 0.009s, Pack+Encode: 2.388s, Decode+Unpack: 1.726s +---------------------- -------------------------------------------------------- +💾 Converting with 86.5963 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,072B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,788B, BPFP=2.6010 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,192B, BPFP=1.3569 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,764B, BPFP=2.5286 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,668B, BPFP=1.5320 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,468B, BPFP=2.5076 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,768B, BPFP=1.4683 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,840B, BPFP=2.5339 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,588B, BPFP=2.2333 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,160B, BPFP=2.4859 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,696B, BPFP=0.3302 +⌛️ [2/4] FRONTEND: Frontend time: 1.973s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.644s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13888381 40.91833109 + layer.0.v_cache 0.00001731 0.00682936 + layer.1.k_cache 0.27925331 8.43094797 + layer.1.v_cache 0.00000589 0.00260042 + layer.2.k_cache 0.03172064 0.88873857 + layer.2.v_cache 0.00001980 0.00744885 + layer.3.k_cache 0.01738505 6.18987236 + layer.3.v_cache 0.00001961 0.00858124 + layer.4.k_cache 0.00068340 0.21434736 + layer.4.v_cache 0.00005063 0.01548145 + layer.4.output 1.38525280 234.99012201 + ------------------------------------------------------------------------------------- + TOTAL 0.59792994 100.09494310 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 312004 +BPFP 1.2976 bits/point +EBPFP 2.5952 equivalent bits/point +MSE 100.094943 +---------------------- -------------------------------------------------------- +Time: 3.626s Load: 0.008s, Pack+Encode: 1.973s, Decode+Unpack: 1.644s +---------------------- -------------------------------------------------------- +💾 Converting with 100.0949 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.2642 bits/point +Avg EBPFP 2.5284 equivalent bits/point +Avg MSE 86.313535 +Avg Time 5.523s +------------------------ ---------------------------- diff --git a/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..baee8292feb9f3e0b286271f8fd5e3e0d8492fd0 --- /dev/null +++ b/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 506 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other +Output output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other +---------------- ------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,620B, BPFP=0.5146 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,688B, BPFP=2.7743 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,512B, BPFP=1.2836 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,728B, BPFP=2.6996 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,624B, BPFP=1.5255 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,652B, BPFP=2.6937 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,476B, BPFP=1.4363 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,280B, BPFP=2.7425 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,872B, BPFP=2.3999 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,256B, BPFP=2.6629 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,064B, BPFP=0.4116 +⌛️ [2/4] FRONTEND: Frontend time: 2.756s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.494s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09475397 40.57013516 + layer.0.v_cache 0.00001927 0.00742089 + layer.1.k_cache 0.17401673 9.19385130 + layer.1.v_cache 0.00000670 0.00288606 + layer.2.k_cache 0.00814445 0.85652062 + layer.2.v_cache 0.00002157 0.00816923 + layer.3.k_cache 0.01691350 6.18382210 + layer.3.v_cache 0.00002161 0.01019491 + layer.4.k_cache 0.00069969 0.23607324 + layer.4.v_cache 0.00005477 0.01724896 + layer.4.output 1.52309445 263.98978323 + ------------------------------------------------------------------------------------- + TOTAL 0.64448903 112.05969442 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 303772 +BPFP 1.3891 bits/point +EBPFP 2.7781 equivalent bits/point +MSE 112.059694 +---------------------- -------------------------------------------------------- +Time: 4.260s Load: 0.010s, Pack+Encode: 2.756s, Decode+Unpack: 1.494s +---------------------- -------------------------------------------------------- +💾 Converting with 112.0597 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 21.800s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 212, 128) +Output shape: (1, 212, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.output: torch.Size([1, 212, 3584]) -> torch.Size([1, 1, 212, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,872B, BPFP=0.5065 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,356B, BPFP=2.6795 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,488B, BPFP=1.2889 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,320B, BPFP=2.6032 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,292B, BPFP=1.6430 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,076B, BPFP=2.5852 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,292B, BPFP=1.4956 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,784B, BPFP=2.6374 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,476B, BPFP=2.3199 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,864B, BPFP=2.5696 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,884B, BPFP=0.3462 +⌛️ [2/4] FRONTEND: Frontend time: 2.392s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.572s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09403566 41.55108067 + layer.0.v_cache 0.00001799 0.00751789 + layer.1.k_cache 0.24957257 8.71578346 + layer.1.v_cache 0.00000621 0.00285631 + layer.2.k_cache 0.01169950 0.71168878 + layer.2.v_cache 0.00002031 0.00827973 + layer.3.k_cache 0.03645871 6.01505899 + layer.3.v_cache 0.00002058 0.01023323 + layer.4.k_cache 0.00069161 0.25295464 + layer.4.v_cache 0.00005130 0.01658639 + layer.4.output 1.44404146 252.26078167 + ------------------------------------------------------------------------------------- + TOTAL 0.61769792 107.24220658 + (elements=1,845,248) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 1845248 +Total Bytes 308704 +BPFP 1.3384 bits/point +EBPFP 2.6767 equivalent bits/point +MSE 107.242207 +---------------------- --------------------------------------------------------- +Time: 25.764s Load: 21.800s, Pack+Encode: 2.392s, Decode+Unpack: 1.572s +---------------------- --------------------------------------------------------- +💾 Converting with 107.2422 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 250, 128) +Output shape: (1, 250, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.output: torch.Size([1, 250, 3584]) -> torch.Size([1, 1, 250, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,740B, BPFP=0.4838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,788B, BPFP=2.2992 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,228B, BPFP=1.3268 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,800B, BPFP=2.2375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,240B, BPFP=1.5150 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,388B, BPFP=2.2117 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,956B, BPFP=1.3722 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,924B, BPFP=2.2452 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,484B, BPFP=1.9678 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,096B, BPFP=2.1935 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,224B, BPFP=0.3324 +⌛️ [2/4] FRONTEND: Frontend time: 1.919s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.590s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09121272 39.76755469 + layer.0.v_cache 0.00001793 0.00684292 + layer.1.k_cache 0.34710925 9.21348242 + layer.1.v_cache 0.00000615 0.00267167 + layer.2.k_cache 0.02211576 0.85001605 + layer.2.v_cache 0.00002179 0.00776740 + layer.3.k_cache 0.02174057 5.31809326 + layer.3.v_cache 0.00001938 0.00832228 + layer.4.k_cache 0.00073173 0.23085135 + layer.4.v_cache 0.00005145 0.01557222 + layer.4.output 1.22463198 210.16669643 + ------------------------------------------------------------------------------------- + TOTAL 0.53267356 89.79929702 + (elements=2,176,000) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2176000 +Total Bytes 322868 +BPFP 1.1870 bits/point +EBPFP 2.3740 equivalent bits/point +MSE 89.799297 +---------------------- -------------------------------------------------------- +Time: 3.518s Load: 0.008s, Pack+Encode: 1.919s, Decode+Unpack: 1.590s +---------------------- -------------------------------------------------------- +💾 Converting with 89.7993 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,184B, BPFP=0.4945 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,640B, BPFP=2.5220 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,592B, BPFP=1.3486 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,760B, BPFP=2.4615 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,656B, BPFP=1.6283 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,156B, BPFP=2.4199 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,960B, BPFP=1.4427 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,736B, BPFP=2.4598 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,732B, BPFP=2.1842 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,776B, BPFP=2.3937 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,144B, BPFP=0.3357 +⌛️ [2/4] FRONTEND: Frontend time: 2.045s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.527s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12780563 41.09351339 + layer.0.v_cache 0.00001798 0.00716976 + layer.1.k_cache 0.36413978 8.81878205 + layer.1.v_cache 0.00000619 0.00274728 + layer.2.k_cache 0.01529831 0.91810655 + layer.2.v_cache 0.00002052 0.00793027 + layer.3.k_cache 0.01269315 5.62677620 + layer.3.v_cache 0.00002023 0.00898037 + layer.4.k_cache 0.00069261 0.23019060 + layer.4.v_cache 0.00005545 0.01550334 + layer.4.output 1.34865724 235.92251416 + ------------------------------------------------------------------------------------- + TOTAL 0.58596180 100.48160582 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 315336 +BPFP 1.2768 bits/point +EBPFP 2.5536 equivalent bits/point +MSE 100.481606 +---------------------- -------------------------------------------------------- +Time: 3.582s Load: 0.010s, Pack+Encode: 2.045s, Decode+Unpack: 1.527s +---------------------- -------------------------------------------------------- +💾 Converting with 100.4816 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 214, 128) +Output shape: (1, 214, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.output: torch.Size([1, 214, 3584]) -> torch.Size([1, 1, 214, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,872B, BPFP=0.5018 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,508B, BPFP=2.6656 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,744B, BPFP=1.2956 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,596B, BPFP=2.5990 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,380B, BPFP=1.6341 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,088B, BPFP=2.5619 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,580B, BPFP=1.3566 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,856B, BPFP=2.6180 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,092B, BPFP=2.2702 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,872B, BPFP=2.5461 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,072B, BPFP=0.3137 +⌛️ [2/4] FRONTEND: Frontend time: 2.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.469s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12820032 42.41255202 + layer.0.v_cache 0.00001693 0.00679544 + layer.1.k_cache 0.24875343 8.43832540 + layer.1.v_cache 0.00000587 0.00254827 + layer.2.k_cache 0.01453762 0.86266712 + layer.2.v_cache 0.00001988 0.00767235 + layer.3.k_cache 0.02276520 5.86141711 + layer.3.v_cache 0.00001932 0.00856510 + layer.4.k_cache 0.00071399 0.22413375 + layer.4.v_cache 0.00004895 0.01526162 + layer.4.output 1.43051836 238.18124166 + ------------------------------------------------------------------------------------- + TOTAL 0.61345353 101.47697822 + (elements=1,862,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1862656 +Total Bytes 304660 +BPFP 1.3085 bits/point +EBPFP 2.6170 equivalent bits/point +MSE 101.476978 +---------------------- -------------------------------------------------------- +Time: 3.732s Load: 0.008s, Pack+Encode: 2.255s, Decode+Unpack: 1.469s +---------------------- -------------------------------------------------------- +💾 Converting with 101.4770 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,696B, BPFP=0.4905 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,656B, BPFP=2.5754 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,788B, BPFP=1.2290 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,436B, BPFP=2.5065 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,888B, BPFP=1.5167 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,912B, BPFP=2.4770 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,804B, BPFP=1.3427 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,016B, BPFP=2.5393 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,328B, BPFP=2.1620 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,620B, BPFP=2.4605 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,908B, BPFP=0.3538 +⌛️ [2/4] FRONTEND: Frontend time: 2.629s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.542s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11875103 41.09137382 + layer.0.v_cache 0.00001689 0.00659165 + layer.1.k_cache 0.49171178 8.47343186 + layer.1.v_cache 0.00000616 0.00256827 + layer.2.k_cache 0.02532451 0.94702358 + layer.2.v_cache 0.00002086 0.00756600 + layer.3.k_cache 0.02370731 5.88631949 + layer.3.v_cache 0.00002129 0.00826040 + layer.4.k_cache 0.00073268 0.21666189 + layer.4.v_cache 0.00004879 0.01496328 + layer.4.output 0.00480151 188.49521338 + ------------------------------------------------------------------------------------- + TOTAL 0.04082070 80.94830905 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 386052 +BPFP 1.2810 bits/point +EBPFP 2.5619 equivalent bits/point +MSE 80.948309 +---------------------- -------------------------------------------------------- +Time: 4.181s Load: 0.009s, Pack+Encode: 2.629s, Decode+Unpack: 1.542s +---------------------- -------------------------------------------------------- +💾 Converting with 80.9483 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,652B, BPFP=0.4880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,860B, BPFP=2.3508 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,844B, BPFP=1.3293 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,840B, BPFP=2.2857 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,544B, BPFP=1.5653 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,504B, BPFP=2.2643 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,336B, BPFP=1.4245 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,128B, BPFP=2.3041 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,080B, BPFP=2.0459 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,184B, BPFP=2.2439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,004B, BPFP=0.3280 +⌛️ [2/4] FRONTEND: Frontend time: 2.528s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.456s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10657067 40.19531250 + layer.0.v_cache 0.00001604 0.00683301 + layer.1.k_cache 0.39270948 8.70795201 + layer.1.v_cache 0.00000611 0.00270344 + layer.2.k_cache 0.01541229 0.84996699 + layer.2.v_cache 0.00002285 0.00803633 + layer.3.k_cache 0.01321528 5.85941287 + layer.3.v_cache 0.00001985 0.00869931 + layer.4.k_cache 0.00076706 0.22847268 + layer.4.v_cache 0.00005205 0.01549552 + layer.4.output 1.24960368 207.66991618 + ------------------------------------------------------------------------------------- + TOTAL 0.54564808 88.79837047 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 322976 +BPFP 1.2116 bits/point +EBPFP 2.4233 equivalent bits/point +MSE 88.798370 +---------------------- -------------------------------------------------------- +Time: 3.995s Load: 0.011s, Pack+Encode: 2.528s, Decode+Unpack: 1.456s +---------------------- -------------------------------------------------------- +💾 Converting with 88.7984 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 197, 128) +Output shape: (1, 197, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.output: torch.Size([1, 197, 3584]) -> torch.Size([1, 1, 197, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,448B, BPFP=0.5114 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,360B, BPFP=2.7253 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,980B, BPFP=1.2674 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,088B, BPFP=2.7037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,176B, BPFP=1.4416 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,764B, BPFP=2.6780 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,464B, BPFP=1.3852 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,948B, BPFP=2.6926 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,940B, BPFP=2.3747 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,068B, BPFP=2.6228 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,260B, BPFP=0.3995 +⌛️ [2/4] FRONTEND: Frontend time: 2.021s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.471s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09219642 43.47700369 + layer.0.v_cache 0.00001621 0.00692654 + layer.1.k_cache 0.17840836 9.45485133 + layer.1.v_cache 0.00000639 0.00271036 + layer.2.k_cache 0.00922235 0.85837458 + layer.2.v_cache 0.00002142 0.00817737 + layer.3.k_cache 0.01957878 5.28338530 + layer.3.v_cache 0.00002124 0.00917545 + layer.4.k_cache 0.00068177 0.23728503 + layer.4.v_cache 0.00005086 0.01656575 + layer.4.output 1.55401793 260.61761240 + ------------------------------------------------------------------------------------- + TOTAL 0.65754878 110.80457307 + (elements=1,714,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1714688 +Total Bytes 292496 +BPFP 1.3647 bits/point +EBPFP 2.7293 equivalent bits/point +MSE 110.804573 +---------------------- -------------------------------------------------------- +Time: 3.498s Load: 0.007s, Pack+Encode: 2.021s, Decode+Unpack: 1.471s +---------------------- -------------------------------------------------------- +💾 Converting with 110.8046 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,628B, BPFP=0.4853 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,968B, BPFP=2.3169 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,984B, BPFP=1.2593 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,580B, BPFP=2.2470 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,892B, BPFP=1.4563 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,352B, BPFP=2.2355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,492B, BPFP=1.4361 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,792B, BPFP=2.2577 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,352B, BPFP=1.9835 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,744B, BPFP=2.2048 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,128B, BPFP=0.3105 +⌛️ [2/4] FRONTEND: Frontend time: 2.456s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.607s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11169702 39.62225302 + layer.0.v_cache 0.00001595 0.00637859 + layer.1.k_cache 0.54955656 9.18716135 + layer.1.v_cache 0.00000603 0.00254790 + layer.2.k_cache 0.01751511 0.78081695 + layer.2.v_cache 0.00002011 0.00758043 + layer.3.k_cache 0.01650532 6.51005781 + layer.3.v_cache 0.00001967 0.00805385 + layer.4.k_cache 0.00077751 0.21635779 + layer.4.v_cache 0.00005008 0.01547519 + layer.4.output 0.04306356 164.57629608 + ------------------------------------------------------------------------------------- + TOTAL 0.05868284 71.08180915 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 397912 +BPFP 1.1798 bits/point +EBPFP 2.3595 equivalent bits/point +MSE 71.081809 +---------------------- -------------------------------------------------------- +Time: 4.074s Load: 0.012s, Pack+Encode: 2.456s, Decode+Unpack: 1.607s +---------------------- -------------------------------------------------------- +💾 Converting with 71.0818 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,996B, BPFP=0.5014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,684B, BPFP=2.6293 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,176B, BPFP=1.3028 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,852B, BPFP=2.5697 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,436B, BPFP=1.5364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,412B, BPFP=2.5381 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,124B, BPFP=1.3707 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,940B, BPFP=2.5760 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,420B, BPFP=2.2520 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,152B, BPFP=2.5195 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,168B, BPFP=0.3806 +⌛️ [2/4] FRONTEND: Frontend time: 2.036s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 27.244s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09832749 41.23032092 + layer.0.v_cache 0.00001658 0.00717467 + layer.1.k_cache 0.33493462 8.53972773 + layer.1.v_cache 0.00000656 0.00273243 + layer.2.k_cache 0.02762392 0.80911773 + layer.2.v_cache 0.00002012 0.00765547 + layer.3.k_cache 0.01647770 5.69971935 + layer.3.v_cache 0.00001951 0.00903187 + layer.4.k_cache 0.00067889 0.21856490 + layer.4.v_cache 0.00004874 0.01552646 + layer.4.output 1.40435175 243.61609191 + ------------------------------------------------------------------------------------- + TOTAL 0.60638920 103.63836558 + (elements=1,897,472) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 1897472 +Total Bytes 313360 +BPFP 1.3212 bits/point +EBPFP 2.6423 equivalent bits/point +MSE 103.638366 +---------------------- --------------------------------------------------------- +Time: 29.287s Load: 0.007s, Pack+Encode: 2.036s, Decode+Unpack: 27.244s +---------------------- --------------------------------------------------------- +💾 Converting with 103.6384 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,152B, BPFP=0.4945 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,464B, BPFP=2.5210 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,012B, BPFP=1.3836 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,564B, BPFP=2.4588 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,820B, BPFP=1.7160 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,112B, BPFP=2.4275 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,456B, BPFP=1.4143 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,704B, BPFP=2.4685 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,544B, BPFP=2.1809 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,912B, BPFP=2.4137 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,996B, BPFP=0.3555 +⌛️ [2/4] FRONTEND: Frontend time: 1.984s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.488s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102250 40.45828436 + layer.0.v_cache 0.00001721 0.00670733 + layer.1.k_cache 0.37371607 8.35798578 + layer.1.v_cache 0.00000658 0.00259879 + layer.2.k_cache 0.01127678 0.89817081 + layer.2.v_cache 0.00001998 0.00769751 + layer.3.k_cache 0.02136800 6.30667303 + layer.3.v_cache 0.00001963 0.00853196 + layer.4.k_cache 0.00069450 0.21137845 + layer.4.v_cache 0.00005147 0.01603988 + layer.4.output 1.35464589 230.39659055 + ------------------------------------------------------------------------------------- + TOTAL 0.58945376 98.17942363 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 317736 +BPFP 1.2922 bits/point +EBPFP 2.5844 equivalent bits/point +MSE 98.179424 +---------------------- -------------------------------------------------------- +Time: 3.482s Load: 0.010s, Pack+Encode: 1.984s, Decode+Unpack: 1.488s +---------------------- -------------------------------------------------------- +💾 Converting with 98.1794 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,760B, BPFP=0.4831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,916B, BPFP=2.2981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,300B, BPFP=1.3259 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,944B, BPFP=2.2375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,908B, BPFP=1.4883 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,440B, BPFP=2.2062 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,704B, BPFP=1.4133 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,260B, BPFP=2.2572 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,216B, BPFP=1.9432 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,128B, BPFP=2.1868 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,660B, BPFP=0.3260 +⌛️ [2/4] FRONTEND: Frontend time: 2.369s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.755s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11871604 40.20750980 + layer.0.v_cache 0.00001755 0.00645841 + layer.1.k_cache 0.46315337 9.49261256 + layer.1.v_cache 0.00000588 0.00245894 + layer.2.k_cache 0.02202094 0.96914320 + layer.2.v_cache 0.00001989 0.00715826 + layer.3.k_cache 0.01278766 5.69294939 + layer.3.v_cache 0.00002092 0.00830610 + layer.4.k_cache 0.00074254 0.20689739 + layer.4.v_cache 0.00004925 0.01430560 + layer.4.output 1.21975157 212.36653386 + ------------------------------------------------------------------------------------- + TOTAL 0.53857618 90.77491392 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 323236 +BPFP 1.1836 bits/point +EBPFP 2.3673 equivalent bits/point +MSE 90.774914 +---------------------- -------------------------------------------------------- +Time: 4.134s Load: 0.011s, Pack+Encode: 2.369s, Decode+Unpack: 1.755s +---------------------- -------------------------------------------------------- +💾 Converting with 90.7749 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 195, 128) +Output shape: (1, 195, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.output: torch.Size([1, 195, 3584]) -> torch.Size([1, 1, 195, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,284B, BPFP=0.5035 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,300B, BPFP=2.7484 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,052B, BPFP=1.2862 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,892B, BPFP=2.7157 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,900B, BPFP=1.4343 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,948B, BPFP=2.7202 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,140B, BPFP=1.3734 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,052B, BPFP=2.7285 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,484B, BPFP=2.2824 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,112B, BPFP=2.6532 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,488B, BPFP=0.3490 +⌛️ [2/4] FRONTEND: Frontend time: 1.978s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.444s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605874 41.61122045 + layer.0.v_cache 0.00001676 0.00698741 + layer.1.k_cache 0.13872500 8.39267077 + layer.1.v_cache 0.00000593 0.00264168 + layer.2.k_cache 0.01222654 0.84276366 + layer.2.v_cache 0.00001968 0.00749899 + layer.3.k_cache 0.01415822 5.09008163 + layer.3.v_cache 0.00002127 0.00921277 + layer.4.k_cache 0.00071127 0.21874253 + layer.4.v_cache 0.00005419 0.01638845 + layer.4.output 1.56987251 270.63365385 + ------------------------------------------------------------------------------------- + TOTAL 0.66300619 114.74316384 + (elements=1,697,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1697280 +Total Bytes 285652 +BPFP 1.3464 bits/point +EBPFP 2.6928 equivalent bits/point +MSE 114.743164 +---------------------- -------------------------------------------------------- +Time: 3.434s Load: 0.012s, Pack+Encode: 1.978s, Decode+Unpack: 1.444s +---------------------- -------------------------------------------------------- +💾 Converting with 114.7432 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,900B, BPFP=0.5065 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,908B, BPFP=2.3959 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,924B, BPFP=1.2812 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,100B, BPFP=2.3266 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,672B, BPFP=1.5172 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,864B, BPFP=2.3063 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,664B, BPFP=1.4306 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,276B, BPFP=2.3417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,352B, BPFP=2.0907 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,604B, BPFP=2.2840 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,188B, BPFP=0.3702 +⌛️ [2/4] FRONTEND: Frontend time: 2.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11176265 42.35521656 + layer.0.v_cache 0.00001802 0.00741145 + layer.1.k_cache 0.18491925 8.70712046 + layer.1.v_cache 0.00000608 0.00289551 + layer.2.k_cache 0.01999094 0.96996416 + layer.2.v_cache 0.00002176 0.00823138 + layer.3.k_cache 0.03910406 4.70808394 + layer.3.v_cache 0.00002004 0.00920525 + layer.4.k_cache 0.00067358 0.21056544 + layer.4.v_cache 0.00005212 0.01619418 + layer.4.output 0.00886795 285.01591935 + ------------------------------------------------------------------------------------- + TOTAL 0.02462613 120.71213669 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 245452 +BPFP 1.2396 bits/point +EBPFP 2.4791 equivalent bits/point +MSE 120.712137 +---------------------- -------------------------------------------------------- +Time: 3.603s Load: 0.007s, Pack+Encode: 2.252s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 120.7121 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,836B, BPFP=0.5015 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,516B, BPFP=2.6787 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,168B, BPFP=1.2594 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,564B, BPFP=2.6089 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,524B, BPFP=1.5056 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,156B, BPFP=2.5789 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,252B, BPFP=1.4123 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,744B, BPFP=2.6221 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,520B, BPFP=2.3122 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,976B, BPFP=2.5657 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,988B, BPFP=0.3352 +⌛️ [2/4] FRONTEND: Frontend time: 2.079s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15089922 43.16042217 + layer.0.v_cache 0.00001686 0.00700669 + layer.1.k_cache 0.31217133 9.41768048 + layer.1.v_cache 0.00000593 0.00270261 + layer.2.k_cache 0.01305961 0.82449635 + layer.2.v_cache 0.00001999 0.00792056 + layer.3.k_cache 0.01178671 5.81705500 + layer.3.v_cache 0.00002039 0.00918832 + layer.4.k_cache 0.00067606 0.22127769 + layer.4.v_cache 0.00005285 0.01576381 + layer.4.output 1.43724915 235.79814302 + ------------------------------------------------------------------------------------- + TOTAL 0.62055606 100.59238323 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 305244 +BPFP 1.3172 bits/point +EBPFP 2.6343 equivalent bits/point +MSE 100.592383 +---------------------- -------------------------------------------------------- +Time: 3.476s Load: 0.008s, Pack+Encode: 2.079s, Decode+Unpack: 1.390s +---------------------- -------------------------------------------------------- +💾 Converting with 100.5924 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 187, 128) +Output shape: (1, 187, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.output: torch.Size([1, 187, 3584]) -> torch.Size([1, 1, 187, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,916B, BPFP=0.4943 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,936B, BPFP=2.3342 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,308B, BPFP=1.2791 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,060B, BPFP=2.2610 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,136B, BPFP=1.5989 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,944B, BPFP=2.2513 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,164B, BPFP=1.4342 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,624B, BPFP=2.3082 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,024B, BPFP=2.0074 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,800B, BPFP=2.2393 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,540B, BPFP=0.4004 +⌛️ [2/4] FRONTEND: Frontend time: 2.607s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.328s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08258281 41.89939840 + layer.0.v_cache 0.00001733 0.00758777 + layer.1.k_cache 0.10049684 9.35482902 + layer.1.v_cache 0.00000647 0.00291607 + layer.2.k_cache 0.01472783 0.74846376 + layer.2.v_cache 0.00002140 0.00868241 + layer.3.k_cache 0.02125903 5.48190879 + layer.3.v_cache 0.00002180 0.01009744 + layer.4.k_cache 0.00070831 0.25419003 + layer.4.v_cache 0.00005280 0.01730598 + layer.4.output 0.00866398 282.21213713 + ------------------------------------------------------------------------------------- + TOTAL 0.01650250 119.60413762 + (elements=1,627,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1627648 +Total Bytes 251452 +BPFP 1.2359 bits/point +EBPFP 2.4718 equivalent bits/point +MSE 119.604138 +---------------------- -------------------------------------------------------- +Time: 3.941s Load: 0.007s, Pack+Encode: 2.607s, Decode+Unpack: 1.328s +---------------------- -------------------------------------------------------- +💾 Converting with 119.6041 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,840B, BPFP=0.5065 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,148B, BPFP=2.6768 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,696B, BPFP=1.3104 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,316B, BPFP=2.6152 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,392B, BPFP=1.5101 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,968B, BPFP=2.5895 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,880B, BPFP=1.4722 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,496B, BPFP=2.6286 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,076B, BPFP=2.3012 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,924B, BPFP=2.5862 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,140B, BPFP=0.3823 +⌛️ [2/4] FRONTEND: Frontend time: 1.971s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.489s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11071491 41.72846934 + layer.0.v_cache 0.00001804 0.00702517 + layer.1.k_cache 0.26784508 8.39228539 + layer.1.v_cache 0.00000610 0.00274859 + layer.2.k_cache 0.01369454 0.71445411 + layer.2.v_cache 0.00002102 0.00768045 + layer.3.k_cache 0.01444715 5.53779783 + layer.3.v_cache 0.00002041 0.00884809 + layer.4.k_cache 0.00067564 0.21990884 + layer.4.v_cache 0.00005248 0.01536822 + layer.4.output 1.45091435 254.11594448 + ------------------------------------------------------------------------------------- + TOTAL 0.62140564 107.96742338 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 308876 +BPFP 1.3455 bits/point +EBPFP 2.6909 equivalent bits/point +MSE 107.967423 +---------------------- -------------------------------------------------------- +Time: 3.467s Load: 0.007s, Pack+Encode: 1.971s, Decode+Unpack: 1.489s +---------------------- -------------------------------------------------------- +💾 Converting with 107.9674 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,976B, BPFP=0.4888 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,608B, BPFP=2.5650 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,956B, BPFP=1.3282 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,616B, BPFP=2.4955 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,676B, BPFP=1.6589 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,236B, BPFP=2.4689 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,376B, BPFP=1.4277 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,868B, BPFP=2.5132 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,684B, BPFP=2.2200 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,128B, BPFP=2.4613 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,236B, BPFP=0.3427 +⌛️ [2/4] FRONTEND: Frontend time: 2.105s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.867s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13389712 41.77134424 + layer.0.v_cache 0.00001748 0.00686924 + layer.1.k_cache 0.36644789 8.33422304 + layer.1.v_cache 0.00000587 0.00266120 + layer.2.k_cache 0.02260907 0.83401845 + layer.2.v_cache 0.00001979 0.00752920 + layer.3.k_cache 0.02550721 5.71285504 + layer.3.v_cache 0.00001968 0.00857095 + layer.4.k_cache 0.00067529 0.22392500 + layer.4.v_cache 0.00006705 0.01567535 + layer.4.output 1.37284074 239.79300128 + ------------------------------------------------------------------------------------- + TOTAL 0.59759716 102.08639298 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 314360 +BPFP 1.2957 bits/point +EBPFP 2.5913 equivalent bits/point +MSE 102.086393 +---------------------- -------------------------------------------------------- +Time: 3.982s Load: 0.009s, Pack+Encode: 2.105s, Decode+Unpack: 1.867s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0864 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 406, 128) +Output shape: (1, 406, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.output: torch.Size([1, 406, 3584]) -> torch.Size([1, 1, 406, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,504B, BPFP=0.4812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 64,160B, BPFP=2.4692 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 33,168B, BPFP=1.2765 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 62,312B, BPFP=2.3981 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 38,656B, BPFP=1.4877 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 62,256B, BPFP=2.3959 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 36,684B, BPFP=1.4118 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 63,076B, BPFP=2.4275 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 53,816B, BPFP=2.0711 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 61,444B, BPFP=2.3647 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,836B, BPFP=0.3675 +⌛️ [2/4] FRONTEND: Frontend time: 3.360s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.938s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10474718 40.17621421 + layer.0.v_cache 0.00001775 0.00659959 + layer.1.k_cache 0.89873824 9.45868491 + layer.1.v_cache 0.00000687 0.00261116 + layer.2.k_cache 0.02602503 0.68742453 + layer.2.v_cache 0.00002164 0.00742583 + layer.3.k_cache 0.01558245 6.07906729 + layer.3.v_cache 0.00002201 0.00817161 + layer.4.k_cache 0.00072958 0.22485055 + layer.4.v_cache 0.00005327 0.01451433 + layer.4.output 0.00651461 128.39263723 + ------------------------------------------------------------------------------------- + TOTAL 0.06420861 56.20082498 + (elements=3,533,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3533824 +Total Bytes 554912 +BPFP 1.2562 bits/point +EBPFP 2.5125 equivalent bits/point +MSE 56.200825 +---------------------- -------------------------------------------------------- +Time: 5.314s Load: 0.016s, Pack+Encode: 3.360s, Decode+Unpack: 1.938s +---------------------- -------------------------------------------------------- +💾 Converting with 56.2008 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,740B, BPFP=0.4936 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,076B, BPFP=2.3645 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,496B, BPFP=1.3071 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,236B, BPFP=2.3110 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,620B, BPFP=1.5702 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,704B, BPFP=2.2770 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,908B, BPFP=1.3972 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,448B, BPFP=2.3245 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,852B, BPFP=2.0314 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,268B, BPFP=2.2492 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,508B, BPFP=0.3599 +⌛️ [2/4] FRONTEND: Frontend time: 2.081s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.417s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14414167 41.15707908 + layer.0.v_cache 0.00001717 0.00692972 + layer.1.k_cache 0.44179790 8.52358897 + layer.1.v_cache 0.00000667 0.00276348 + layer.2.k_cache 0.03016641 0.76132893 + layer.2.v_cache 0.00002043 0.00800494 + layer.3.k_cache 0.03726396 6.12194575 + layer.3.v_cache 0.00002099 0.00887872 + layer.4.k_cache 0.00070653 0.22524827 + layer.4.v_cache 0.00004910 0.01528677 + layer.4.output 1.24963187 207.71446793 + ------------------------------------------------------------------------------------- + TOTAL 0.55303611 88.87249001 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 326856 +BPFP 1.2262 bits/point +EBPFP 2.4524 equivalent bits/point +MSE 88.872490 +---------------------- -------------------------------------------------------- +Time: 3.509s Load: 0.011s, Pack+Encode: 2.081s, Decode+Unpack: 1.417s +---------------------- -------------------------------------------------------- +💾 Converting with 88.8725 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,536B, BPFP=0.5132 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,636B, BPFP=2.7981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,028B, BPFP=1.2585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,520B, BPFP=2.7104 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,728B, BPFP=1.4705 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,160B, BPFP=2.6822 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,320B, BPFP=1.3599 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,968B, BPFP=2.7456 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,440B, BPFP=2.3901 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,168B, BPFP=2.6828 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,088B, BPFP=0.3375 +⌛️ [2/4] FRONTEND: Frontend time: 2.347s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.715s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15985105 41.31237241 + layer.0.v_cache 0.00001905 0.00741888 + layer.1.k_cache 0.22745673 8.75476810 + layer.1.v_cache 0.00000590 0.00277889 + layer.2.k_cache 0.01549998 0.89042648 + layer.2.v_cache 0.00002003 0.00791384 + layer.3.k_cache 0.01333464 5.06857192 + layer.3.v_cache 0.00002034 0.00940407 + layer.4.k_cache 0.00067076 0.24984442 + layer.4.v_cache 0.00004949 0.01585434 + layer.4.output 1.53832061 260.59020549 + ------------------------------------------------------------------------------------- + TOTAL 0.65795131 110.61475246 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 292592 +BPFP 1.3514 bits/point +EBPFP 2.7028 equivalent bits/point +MSE 110.614752 +---------------------- -------------------------------------------------------- +Time: 4.070s Load: 0.008s, Pack+Encode: 2.347s, Decode+Unpack: 1.715s +---------------------- -------------------------------------------------------- +💾 Converting with 110.6148 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 237, 128) +Output shape: (1, 237, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.output: torch.Size([1, 237, 3584]) -> torch.Size([1, 1, 237, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,532B, BPFP=0.4966 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,864B, BPFP=2.4304 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,244B, BPFP=1.2687 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,092B, BPFP=2.3795 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,912B, BPFP=1.5765 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,748B, BPFP=2.3568 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,172B, BPFP=1.4618 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,216B, BPFP=2.3877 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,776B, BPFP=2.0949 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,252B, BPFP=2.3241 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,072B, BPFP=0.3492 +⌛️ [2/4] FRONTEND: Frontend time: 2.214s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.405s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351868 40.77248154 + layer.0.v_cache 0.00001655 0.00696831 + layer.1.k_cache 0.34375901 8.83669877 + layer.1.v_cache 0.00000619 0.00269065 + layer.2.k_cache 0.01282014 0.71980002 + layer.2.v_cache 0.00002104 0.00791998 + layer.3.k_cache 0.03241006 5.57000037 + layer.3.v_cache 0.00001985 0.00857719 + layer.4.k_cache 0.00067667 0.21654939 + layer.4.v_cache 0.00005279 0.01572257 + layer.4.output 1.29178675 225.87040386 + ------------------------------------------------------------------------------------- + TOTAL 0.56210637 96.30883740 + (elements=2,062,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2062848 +Total Bytes 321880 +BPFP 1.2483 bits/point +EBPFP 2.4966 equivalent bits/point +MSE 96.308837 +---------------------- -------------------------------------------------------- +Time: 3.629s Load: 0.010s, Pack+Encode: 2.214s, Decode+Unpack: 1.405s +---------------------- -------------------------------------------------------- +💾 Converting with 96.3088 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,896B, BPFP=0.5012 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,236B, BPFP=2.6334 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,092B, BPFP=1.2422 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,356B, BPFP=2.5695 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,924B, BPFP=1.5206 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,224B, BPFP=2.5599 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,592B, BPFP=1.4238 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,700B, BPFP=2.5945 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,220B, BPFP=2.2689 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,888B, BPFP=2.5355 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,388B, BPFP=0.3363 +⌛️ [2/4] FRONTEND: Frontend time: 2.048s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.514s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11023432 40.31351744 + layer.0.v_cache 0.00001677 0.00667812 + layer.1.k_cache 0.30291741 8.43778161 + layer.1.v_cache 0.00000597 0.00262633 + layer.2.k_cache 0.03464115 0.85783989 + layer.2.v_cache 0.00001998 0.00744178 + layer.3.k_cache 0.01330193 5.58425861 + layer.3.v_cache 0.00002112 0.00842029 + layer.4.k_cache 0.00069379 0.20622245 + layer.4.v_cache 0.00005021 0.01477994 + layer.4.output 1.42388847 244.79574336 + ------------------------------------------------------------------------------------- + TOTAL 0.61347776 104.05939823 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 305516 +BPFP 1.3061 bits/point +EBPFP 2.6121 equivalent bits/point +MSE 104.059398 +---------------------- -------------------------------------------------------- +Time: 3.569s Load: 0.007s, Pack+Encode: 2.048s, Decode+Unpack: 1.514s +---------------------- -------------------------------------------------------- +💾 Converting with 104.0594 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 231, 128) +Output shape: (1, 231, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.output: torch.Size([1, 231, 3584]) -> torch.Size([1, 1, 231, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,452B, BPFP=0.5041 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,056B, BPFP=2.5065 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,272B, BPFP=1.3712 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,084B, BPFP=2.4407 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,316B, BPFP=1.6448 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,772B, BPFP=2.4196 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,900B, BPFP=1.4813 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,356B, BPFP=2.4591 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,764B, BPFP=2.1485 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,128B, BPFP=2.3761 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,516B, BPFP=0.3915 +⌛️ [2/4] FRONTEND: Frontend time: 2.113s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.654s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13257201 41.69075098 + layer.0.v_cache 0.00001701 0.00706054 + layer.1.k_cache 0.40462213 8.24143447 + layer.1.v_cache 0.00000658 0.00280142 + layer.2.k_cache 0.02649644 0.83444002 + layer.2.v_cache 0.00002056 0.00743336 + layer.3.k_cache 0.06147716 6.09727049 + layer.3.v_cache 0.00002105 0.00906027 + layer.4.k_cache 0.00069437 0.22239452 + layer.4.v_cache 0.00005136 0.01539752 + layer.4.output 1.32536713 221.43958720 + ------------------------------------------------------------------------------------- + TOTAL 0.58256168 94.54147965 + (elements=2,010,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2010624 +Total Bytes 326616 +BPFP 1.2996 bits/point +EBPFP 2.5991 equivalent bits/point +MSE 94.541480 +---------------------- -------------------------------------------------------- +Time: 3.776s Load: 0.008s, Pack+Encode: 2.113s, Decode+Unpack: 1.654s +---------------------- -------------------------------------------------------- +💾 Converting with 94.5415 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,764B, BPFP=0.4757 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,120B, BPFP=2.2745 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,964B, BPFP=1.2846 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,228B, BPFP=2.2199 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,516B, BPFP=1.4409 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,020B, BPFP=2.2071 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,620B, BPFP=1.3860 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,800B, BPFP=2.2549 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,140B, BPFP=1.9694 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,528B, BPFP=2.1770 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,252B, BPFP=0.3523 +⌛️ [2/4] FRONTEND: Frontend time: 1.941s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.533s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12841698 41.88120404 + layer.0.v_cache 0.00001718 0.00681121 + layer.1.k_cache 0.45450200 9.12797947 + layer.1.v_cache 0.00000627 0.00269777 + layer.2.k_cache 0.01434461 0.74841715 + layer.2.v_cache 0.00002215 0.00806841 + layer.3.k_cache 0.03530993 6.52222876 + layer.3.v_cache 0.00002139 0.00882295 + layer.4.k_cache 0.00068344 0.25693605 + layer.4.v_cache 0.00005073 0.01508655 + layer.4.output 1.20064817 201.52951681 + ------------------------------------------------------------------------------------- + TOTAL 0.53164187 86.42852177 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 328952 +BPFP 1.1857 bits/point +EBPFP 2.3713 equivalent bits/point +MSE 86.428522 +---------------------- -------------------------------------------------------- +Time: 3.482s Load: 0.008s, Pack+Encode: 1.941s, Decode+Unpack: 1.533s +---------------------- -------------------------------------------------------- +💾 Converting with 86.4285 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,860B, BPFP=0.4909 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,896B, BPFP=2.5430 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,840B, BPFP=1.2655 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,748B, BPFP=2.4794 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,640B, BPFP=1.4761 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,460B, BPFP=2.4634 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,144B, BPFP=1.3932 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,088B, BPFP=2.4982 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,660B, BPFP=2.1975 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,052B, BPFP=2.4408 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,528B, BPFP=0.3366 +⌛️ [2/4] FRONTEND: Frontend time: 2.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.719s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13731835 39.46585494 + layer.0.v_cache 0.00001709 0.00664731 + layer.1.k_cache 0.55879569 8.93080604 + layer.1.v_cache 0.00000637 0.00255657 + layer.2.k_cache 0.01348169 0.92008972 + layer.2.v_cache 0.00002375 0.00758881 + layer.3.k_cache 0.01306405 5.63891948 + layer.3.v_cache 0.00002005 0.00821974 + layer.4.k_cache 0.00072533 0.23747881 + layer.4.v_cache 0.00005384 0.01523153 + layer.4.output 0.00472113 191.49893934 + ------------------------------------------------------------------------------------- + TOTAL 0.04450318 82.10152755 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 389916 +BPFP 1.2708 bits/point +EBPFP 2.5417 equivalent bits/point +MSE 82.101528 +---------------------- -------------------------------------------------------- +Time: 3.861s Load: 0.012s, Pack+Encode: 2.131s, Decode+Unpack: 1.719s +---------------------- -------------------------------------------------------- +💾 Converting with 82.1015 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,448B, BPFP=0.4904 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,996B, BPFP=2.3877 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,360B, BPFP=1.3164 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,880B, BPFP=2.3297 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,852B, BPFP=1.6015 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,392B, BPFP=2.3044 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,864B, BPFP=1.4983 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,212B, BPFP=2.3470 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,028B, BPFP=2.0779 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,808B, BPFP=2.2741 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,124B, BPFP=0.3272 +⌛️ [2/4] FRONTEND: Frontend time: 2.283s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 26.004s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13972237 41.56189654 + layer.0.v_cache 0.00001689 0.00702989 + layer.1.k_cache 0.51719427 9.24070887 + layer.1.v_cache 0.00000613 0.00281712 + layer.2.k_cache 0.02019965 0.79332944 + layer.2.v_cache 0.00002053 0.00806288 + layer.3.k_cache 0.01759115 6.42551683 + layer.3.v_cache 0.00002125 0.00928142 + layer.4.k_cache 0.00069604 0.25083895 + layer.4.v_cache 0.00005872 0.01604855 + layer.4.output 0.04434862 176.97622508 + ------------------------------------------------------------------------------------- + TOTAL 0.05917455 76.30288859 + (elements=2,619,904) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2619904 +Total Bytes 402964 +BPFP 1.2305 bits/point +EBPFP 2.4609 equivalent bits/point +MSE 76.302889 +---------------------- --------------------------------------------------------- +Time: 28.300s Load: 0.012s, Pack+Encode: 2.283s, Decode+Unpack: 26.004s +---------------------- --------------------------------------------------------- +💾 Converting with 76.3029 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,540B, BPFP=0.5135 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,188B, BPFP=2.7629 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,948B, BPFP=1.2522 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,528B, BPFP=2.7111 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,684B, BPFP=1.4670 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,276B, BPFP=2.6128 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,528B, BPFP=1.3763 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,700B, BPFP=2.7246 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,596B, BPFP=2.4023 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,676B, BPFP=2.6442 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,936B, BPFP=0.3470 +⌛️ [2/4] FRONTEND: Frontend time: 2.076s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.502s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13044639 39.03503111 + layer.0.v_cache 0.00001737 0.00722873 + layer.1.k_cache 0.22565776 9.60679496 + layer.1.v_cache 0.00000620 0.00276014 + layer.2.k_cache 0.01998205 0.83496945 + layer.2.v_cache 0.00001989 0.00802720 + layer.3.k_cache 0.00629339 5.25897247 + layer.3.v_cache 0.00002050 0.00931192 + layer.4.k_cache 0.00068118 0.23011218 + layer.4.v_cache 0.00005692 0.01566532 + layer.4.output 1.53833376 260.96471195 + ------------------------------------------------------------------------------------- + TOTAL 0.65597165 110.69187395 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 291600 +BPFP 1.3468 bits/point +EBPFP 2.6936 equivalent bits/point +MSE 110.691874 +---------------------- -------------------------------------------------------- +Time: 3.585s Load: 0.007s, Pack+Encode: 2.076s, Decode+Unpack: 1.502s +---------------------- -------------------------------------------------------- +💾 Converting with 110.6919 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 246, 128) +Output shape: (1, 246, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.output: torch.Size([1, 246, 3584]) -> torch.Size([1, 1, 246, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,684B, BPFP=0.4881 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,884B, BPFP=2.3427 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,196B, BPFP=1.3463 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,972B, BPFP=2.2848 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,300B, BPFP=1.5434 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,584B, BPFP=2.2602 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,660B, BPFP=1.4393 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,252B, BPFP=2.3026 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,024B, BPFP=2.0340 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,036B, BPFP=2.2254 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,628B, BPFP=0.3233 +⌛️ [2/4] FRONTEND: Frontend time: 1.944s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.938s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131485 40.32981215 + layer.0.v_cache 0.00001613 0.00690552 + layer.1.k_cache 0.49701033 9.14104381 + layer.1.v_cache 0.00000620 0.00269128 + layer.2.k_cache 0.01298508 0.96792057 + layer.2.v_cache 0.00002120 0.00767751 + layer.3.k_cache 0.02762915 6.09221966 + layer.3.v_cache 0.00002136 0.00893525 + layer.4.k_cache 0.00068392 0.25614526 + layer.4.v_cache 0.00006372 0.01505368 + layer.4.output 1.24453647 205.53616797 + ------------------------------------------------------------------------------------- + TOTAL 0.55302984 87.97538709 + (elements=2,141,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2141184 +Total Bytes 323220 +BPFP 1.2076 bits/point +EBPFP 2.4153 equivalent bits/point +MSE 87.975387 +---------------------- -------------------------------------------------------- +Time: 3.892s Load: 0.011s, Pack+Encode: 1.944s, Decode+Unpack: 1.938s +---------------------- -------------------------------------------------------- +💾 Converting with 87.9754 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,592B, BPFP=0.5150 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,128B, BPFP=2.7444 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,288B, BPFP=1.2725 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,708B, BPFP=2.7116 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,848B, BPFP=1.4725 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,580B, BPFP=2.7016 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,560B, BPFP=1.3719 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,760B, BPFP=2.7156 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,056B, BPFP=2.3481 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,952B, BPFP=2.6525 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,500B, BPFP=0.3181 +⌛️ [2/4] FRONTEND: Frontend time: 2.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.511s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11599884 41.20846191 + layer.0.v_cache 0.00001671 0.00678310 + layer.1.k_cache 0.22559120 8.70484131 + layer.1.v_cache 0.00000595 0.00267443 + layer.2.k_cache 0.01146694 0.79505035 + layer.2.v_cache 0.00002201 0.00805906 + layer.3.k_cache 0.03668072 4.77154663 + layer.3.v_cache 0.00002131 0.00910327 + layer.4.k_cache 0.00066762 0.22729679 + layer.4.v_cache 0.00005146 0.01594924 + layer.4.output 1.53061454 257.52292411 + ------------------------------------------------------------------------------------- + TOTAL 0.65322497 109.31824911 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 290972 +BPFP 1.3372 bits/point +EBPFP 2.6744 equivalent bits/point +MSE 109.318249 +---------------------- -------------------------------------------------------- +Time: 3.749s Load: 0.010s, Pack+Encode: 2.228s, Decode+Unpack: 1.511s +---------------------- -------------------------------------------------------- +💾 Converting with 109.3182 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,356B, BPFP=0.5019 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,768B, BPFP=2.5087 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,872B, BPFP=1.3559 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,900B, BPFP=2.4495 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,560B, BPFP=1.6075 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,472B, BPFP=2.4203 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,900B, BPFP=1.4943 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,172B, BPFP=2.4681 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,828B, BPFP=2.1717 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,976B, BPFP=2.3865 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,368B, BPFP=0.4032 +⌛️ [2/4] FRONTEND: Frontend time: 2.237s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.457s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605523 42.29151883 + layer.0.v_cache 0.00001866 0.00720017 + layer.1.k_cache 0.31881074 8.70658561 + layer.1.v_cache 0.00000653 0.00283132 + layer.2.k_cache 0.02091764 0.78934769 + layer.2.v_cache 0.00002113 0.00792161 + layer.3.k_cache 0.03221333 6.03929491 + layer.3.v_cache 0.00002185 0.00889895 + layer.4.k_cache 0.00068277 0.23305278 + layer.4.v_cache 0.00005596 0.01624863 + layer.4.output 1.33695214 221.49668590 + ------------------------------------------------------------------------------------- + TOTAL 0.57926287 94.62233540 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 325172 +BPFP 1.3051 bits/point +EBPFP 2.6102 equivalent bits/point +MSE 94.622335 +---------------------- -------------------------------------------------------- +Time: 3.706s Load: 0.012s, Pack+Encode: 2.237s, Decode+Unpack: 1.457s +---------------------- -------------------------------------------------------- +💾 Converting with 94.6223 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 192, 128) +Output shape: (1, 192, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.output: torch.Size([1, 192, 3584]) -> torch.Size([1, 1, 192, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,644B, BPFP=0.4593 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,048B, BPFP=2.2012 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,432B, BPFP=1.2559 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,428B, BPFP=2.1507 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,232B, BPFP=1.4023 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,200B, BPFP=2.1322 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,444B, BPFP=1.3382 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,640B, BPFP=2.1680 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,620B, BPFP=1.9222 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,948B, BPFP=2.1117 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,440B, BPFP=0.3190 +⌛️ [2/4] FRONTEND: Frontend time: 1.999s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.323s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12444629 41.82856496 + layer.0.v_cache 0.00001540 0.00602646 + layer.1.k_cache 0.18411521 8.83156522 + layer.1.v_cache 0.00000585 0.00246707 + layer.2.k_cache 0.01124177 0.84355140 + layer.2.v_cache 0.00002179 0.00759477 + layer.3.k_cache 0.01616090 6.19691531 + layer.3.v_cache 0.00002117 0.00842399 + layer.4.k_cache 0.00068507 0.22581406 + layer.4.v_cache 0.00005275 0.01400724 + layer.4.output 0.00839530 275.21347191 + ------------------------------------------------------------------------------------- + TOTAL 0.02326666 116.73289611 + (elements=1,671,168) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1671168 +Total Bytes 238076 +BPFP 1.1397 bits/point +EBPFP 2.2794 equivalent bits/point +MSE 116.732896 +---------------------- -------------------------------------------------------- +Time: 3.330s Load: 0.008s, Pack+Encode: 1.999s, Decode+Unpack: 1.323s +---------------------- -------------------------------------------------------- +💾 Converting with 116.7329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,576B, BPFP=0.5248 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,352B, BPFP=2.5745 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,280B, BPFP=1.2500 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,664B, BPFP=2.5098 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,592B, BPFP=1.5617 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,376B, BPFP=2.4827 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,312B, BPFP=1.4413 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,804B, BPFP=2.5230 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,084B, BPFP=2.2669 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,116B, BPFP=2.4582 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,236B, BPFP=0.3124 +⌛️ [2/4] FRONTEND: Frontend time: 1.919s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.298s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09999813 41.21663215 + layer.0.v_cache 0.00001899 0.00722732 + layer.1.k_cache 0.12989934 8.79699192 + layer.1.v_cache 0.00000610 0.00279820 + layer.2.k_cache 0.00966061 0.92549317 + layer.2.v_cache 0.00001994 0.00822707 + layer.3.k_cache 0.03384685 4.12698769 + layer.3.v_cache 0.00002040 0.00948015 + layer.4.k_cache 0.00066502 0.23615826 + layer.4.v_cache 0.00005200 0.01686301 + layer.4.output 0.00959846 310.05962242 + ------------------------------------------------------------------------------------- + TOTAL 0.02008098 130.92730682 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 231392 +BPFP 1.2812 bits/point +EBPFP 2.5624 equivalent bits/point +MSE 130.927307 +---------------------- -------------------------------------------------------- +Time: 3.224s Load: 0.006s, Pack+Encode: 1.919s, Decode+Unpack: 1.298s +---------------------- -------------------------------------------------------- +💾 Converting with 130.9273 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 428, 128) +Output shape: (1, 428, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.output: torch.Size([1, 428, 3584]) -> torch.Size([1, 1, 428, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,292B, BPFP=0.4853 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 64,240B, BPFP=2.3452 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 35,644B, BPFP=1.3013 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 62,492B, BPFP=2.2814 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 41,580B, BPFP=1.5180 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 61,980B, BPFP=2.2627 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 37,720B, BPFP=1.3770 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 62,848B, BPFP=2.2944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 55,000B, BPFP=2.0079 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 61,112B, BPFP=2.2310 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,840B, BPFP=0.3277 +⌛️ [2/4] FRONTEND: Frontend time: 2.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.858s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13760449 40.03068414 + layer.0.v_cache 0.00001624 0.00638532 + layer.1.k_cache 0.89392254 8.87760683 + layer.1.v_cache 0.00000615 0.00245275 + layer.2.k_cache 0.03917003 0.82869257 + layer.2.v_cache 0.00002113 0.00743824 + layer.3.k_cache 0.03749864 5.84261942 + layer.3.v_cache 0.00002059 0.00778875 + layer.4.k_cache 0.00086919 0.24257947 + layer.4.v_cache 0.00005177 0.01452557 + layer.4.output 0.00617707 125.39283628 + ------------------------------------------------------------------------------------- + TOTAL 0.06778943 54.91827218 + (elements=3,725,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3725312 +Total Bytes 558748 +BPFP 1.1999 bits/point +EBPFP 2.3998 equivalent bits/point +MSE 54.918272 +---------------------- -------------------------------------------------------- +Time: 4.353s Load: 0.014s, Pack+Encode: 2.481s, Decode+Unpack: 1.858s +---------------------- -------------------------------------------------------- +💾 Converting with 54.9183 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,872B, BPFP=0.4916 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,964B, BPFP=2.5468 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,772B, BPFP=1.2617 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,960B, BPFP=2.4911 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,596B, BPFP=1.5844 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,496B, BPFP=2.4654 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,568B, BPFP=1.4167 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,164B, BPFP=2.5024 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,896B, BPFP=2.1551 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,004B, BPFP=2.4382 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 45,640B, BPFP=0.3613 +⌛️ [2/4] FRONTEND: Frontend time: 2.104s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.619s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385377 40.13894891 + layer.0.v_cache 0.00001705 0.00657043 + layer.1.k_cache 0.49372572 8.90493406 + layer.1.v_cache 0.00000626 0.00255461 + layer.2.k_cache 0.02422354 0.81594118 + layer.2.v_cache 0.00002046 0.00714684 + layer.3.k_cache 0.04153839 6.20356177 + layer.3.v_cache 0.00002075 0.00793706 + layer.4.k_cache 0.00074884 0.21333145 + layer.4.v_cache 0.00004954 0.01442969 + layer.4.output 0.00474379 191.46697695 + ------------------------------------------------------------------------------------- + TOTAL 0.04161240 82.15201145 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 394932 +BPFP 1.2872 bits/point +EBPFP 2.5744 equivalent bits/point +MSE 82.152011 +---------------------- -------------------------------------------------------- +Time: 3.733s Load: 0.009s, Pack+Encode: 2.104s, Decode+Unpack: 1.619s +---------------------- -------------------------------------------------------- +💾 Converting with 82.1520 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,000B, BPFP=0.4837 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 52,692B, BPFP=2.5490 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,920B, BPFP=1.3022 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 51,524B, BPFP=2.4925 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,948B, BPFP=1.4003 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 50,196B, BPFP=2.4282 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,444B, BPFP=1.3276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 51,412B, BPFP=2.4870 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,676B, BPFP=2.1128 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 49,628B, BPFP=2.4007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 52,876B, BPFP=0.3654 +⌛️ [2/4] FRONTEND: Frontend time: 2.208s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.520s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12985682 40.61500460 + layer.0.v_cache 0.00001645 0.00646155 + layer.1.k_cache 0.61935935 8.88018062 + layer.1.v_cache 0.00000656 0.00252952 + layer.2.k_cache 0.01492825 0.81170872 + layer.2.v_cache 0.00002178 0.00706742 + layer.3.k_cache 0.02579401 4.75294216 + layer.3.v_cache 0.00002068 0.00770986 + layer.4.k_cache 0.00071613 0.21204364 + layer.4.v_cache 0.00005057 0.01402088 + layer.4.output 0.04141478 164.92720312 + ------------------------------------------------------------------------------------- + TOTAL 0.06356906 71.16471122 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 445316 +BPFP 1.2672 bits/point +EBPFP 2.5344 equivalent bits/point +MSE 71.164711 +---------------------- -------------------------------------------------------- +Time: 3.738s Load: 0.010s, Pack+Encode: 2.208s, Decode+Unpack: 1.520s +---------------------- -------------------------------------------------------- +💾 Converting with 71.1647 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,832B, BPFP=0.4825 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,092B, BPFP=2.5181 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,132B, BPFP=1.3730 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,896B, BPFP=2.4528 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,408B, BPFP=1.6066 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,468B, BPFP=2.4294 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,280B, BPFP=1.4904 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,272B, BPFP=2.4733 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,528B, BPFP=2.1595 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,032B, BPFP=2.4056 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 46,052B, BPFP=0.3594 +⌛️ [2/4] FRONTEND: Frontend time: 1.878s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.350s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12359551 40.20828439 + layer.0.v_cache 0.00001766 0.00689942 + layer.1.k_cache 0.52692803 8.77917907 + layer.1.v_cache 0.00000637 0.00264476 + layer.2.k_cache 0.02102186 0.79813887 + layer.2.v_cache 0.00002187 0.00773365 + layer.3.k_cache 0.03868174 6.14213637 + layer.3.v_cache 0.00002097 0.00868605 + layer.4.k_cache 0.00070799 0.22660526 + layer.4.v_cache 0.00005204 0.01512946 + layer.4.output 0.00469152 185.65820117 + ------------------------------------------------------------------------------------- + TOTAL 0.04375851 79.75310856 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 400992 +BPFP 1.2887 bits/point +EBPFP 2.5773 equivalent bits/point +MSE 79.753109 +---------------------- -------------------------------------------------------- +Time: 3.237s Load: 0.009s, Pack+Encode: 1.878s, Decode+Unpack: 1.350s +---------------------- -------------------------------------------------------- +💾 Converting with 79.7531 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,576B, BPFP=0.5163 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,264B, BPFP=2.7688 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,040B, BPFP=1.2594 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,832B, BPFP=2.6564 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,904B, BPFP=1.4843 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,260B, BPFP=2.6900 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,920B, BPFP=1.4070 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,720B, BPFP=2.7261 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,560B, BPFP=2.3995 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,724B, BPFP=2.6479 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,876B, BPFP=0.3688 +⌛️ [2/4] FRONTEND: Frontend time: 1.728s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.250s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10671347 40.47364753 + layer.0.v_cache 0.00001788 0.00728679 + layer.1.k_cache 0.15257932 9.46376413 + layer.1.v_cache 0.00000642 0.00271701 + layer.2.k_cache 0.02197224 0.84272804 + layer.2.v_cache 0.00002094 0.00772667 + layer.3.k_cache 0.03325213 5.14601450 + layer.3.v_cache 0.00002126 0.00926366 + layer.4.k_cache 0.00067229 0.22770591 + layer.4.v_cache 0.00005023 0.01520135 + layer.4.output 1.53834953 264.48505922 + ------------------------------------------------------------------------------------- + TOTAL 0.65198546 112.21126295 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 294676 +BPFP 1.3610 bits/point +EBPFP 2.7220 equivalent bits/point +MSE 112.211263 +---------------------- -------------------------------------------------------- +Time: 2.986s Load: 0.008s, Pack+Encode: 1.728s, Decode+Unpack: 1.250s +---------------------- -------------------------------------------------------- +💾 Converting with 112.2113 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,744B, BPFP=0.4821 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,988B, BPFP=2.3025 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,816B, BPFP=1.3581 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,124B, BPFP=2.2488 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,724B, BPFP=1.5391 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,772B, BPFP=2.2268 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,552B, BPFP=1.4661 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,504B, BPFP=2.2724 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,988B, BPFP=1.9913 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,344B, BPFP=2.2002 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,844B, BPFP=0.3632 +⌛️ [2/4] FRONTEND: Frontend time: 1.727s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.253s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15214830 40.22400632 + layer.0.v_cache 0.00001722 0.00715363 + layer.1.k_cache 0.49754258 9.41609149 + layer.1.v_cache 0.00000639 0.00287852 + layer.2.k_cache 0.01862443 0.82959331 + layer.2.v_cache 0.00002219 0.00832686 + layer.3.k_cache 0.01200796 5.28042080 + layer.3.v_cache 0.00002210 0.00954861 + layer.4.k_cache 0.00075226 0.24195658 + layer.4.v_cache 0.00005409 0.01653543 + layer.4.output 1.21979868 212.20420105 + ------------------------------------------------------------------------------------- + TOTAL 0.54234049 90.67446582 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 331400 +BPFP 1.2135 bits/point +EBPFP 2.4271 equivalent bits/point +MSE 90.674466 +---------------------- -------------------------------------------------------- +Time: 2.991s Load: 0.010s, Pack+Encode: 1.727s, Decode+Unpack: 1.253s +---------------------- -------------------------------------------------------- +💾 Converting with 90.6745 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,460B, BPFP=0.4846 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,932B, BPFP=2.3531 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,264B, BPFP=1.3455 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,836B, BPFP=2.2969 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,440B, BPFP=1.5594 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,344B, BPFP=2.2717 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,712B, BPFP=1.4197 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,296B, BPFP=2.3205 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,844B, BPFP=2.0412 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,960B, BPFP=2.2520 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,120B, BPFP=0.3083 +⌛️ [2/4] FRONTEND: Frontend time: 1.854s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.410s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18432382 40.00262551 + layer.0.v_cache 0.00001629 0.00679286 + layer.1.k_cache 0.53789903 9.14249168 + layer.1.v_cache 0.00000607 0.00268830 + layer.2.k_cache 0.01514517 0.80563805 + layer.2.v_cache 0.00002151 0.00786778 + layer.3.k_cache 0.04182221 5.14364674 + layer.3.v_cache 0.00002017 0.00883553 + layer.4.k_cache 0.00068627 0.23949705 + layer.4.v_cache 0.00005063 0.01570858 + layer.4.output 0.04375917 162.60930913 + ------------------------------------------------------------------------------------- + TOTAL 0.06390032 70.21417388 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 400208 +BPFP 1.2060 bits/point +EBPFP 2.4121 equivalent bits/point +MSE 70.214174 +---------------------- -------------------------------------------------------- +Time: 3.274s Load: 0.011s, Pack+Encode: 1.854s, Decode+Unpack: 1.410s +---------------------- -------------------------------------------------------- +💾 Converting with 70.2142 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 371, 128) +Output shape: (1, 371, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.output: torch.Size([1, 371, 3584]) -> torch.Size([1, 1, 371, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,368B, BPFP=0.4788 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,312B, BPFP=2.3295 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,248B, BPFP=1.3160 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,904B, BPFP=2.2702 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 36,164B, BPFP=1.5231 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,472B, BPFP=2.2520 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 33,820B, BPFP=1.4244 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 54,432B, BPFP=2.2925 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,588B, BPFP=2.0042 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,792B, BPFP=2.2234 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 56,264B, BPFP=0.3385 +⌛️ [2/4] FRONTEND: Frontend time: 2.034s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.480s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16952868 40.19597056 + layer.0.v_cache 0.00001698 0.00683488 + layer.1.k_cache 0.82091874 9.55189469 + layer.1.v_cache 0.00000666 0.00270922 + layer.2.k_cache 0.02418445 0.83259801 + layer.2.v_cache 0.00002238 0.00791796 + layer.3.k_cache 0.03055777 6.47602947 + layer.3.v_cache 0.00002117 0.00861315 + layer.4.k_cache 0.00071103 0.24256839 + layer.4.v_cache 0.00005915 0.01550142 + layer.4.output 0.03610923 144.28198402 + ------------------------------------------------------------------------------------- + TOTAL 0.07639951 62.78320740 + (elements=3,229,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3229184 +Total Bytes 486364 +BPFP 1.2049 bits/point +EBPFP 2.4098 equivalent bits/point +MSE 62.783207 +---------------------- -------------------------------------------------------- +Time: 3.526s Load: 0.012s, Pack+Encode: 2.034s, Decode+Unpack: 1.480s +---------------------- -------------------------------------------------------- +💾 Converting with 62.7832 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,988B, BPFP=0.5009 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,960B, BPFP=2.6491 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,288B, BPFP=1.3108 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,960B, BPFP=2.5774 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,676B, BPFP=1.6970 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,676B, BPFP=2.5571 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,288B, BPFP=1.4541 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,304B, BPFP=2.6021 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,764B, BPFP=2.2767 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,352B, BPFP=2.5338 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,924B, BPFP=0.3678 +⌛️ [2/4] FRONTEND: Frontend time: 1.740s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.237s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13674593 38.80640768 + layer.0.v_cache 0.00001768 0.00739076 + layer.1.k_cache 0.26206151 8.70785830 + layer.1.v_cache 0.00000624 0.00276542 + layer.2.k_cache 0.02938290 0.84484030 + layer.2.v_cache 0.00002158 0.00784730 + layer.3.k_cache 0.05653211 5.98124877 + layer.3.v_cache 0.00002129 0.00932850 + layer.4.k_cache 0.00070776 0.23523150 + layer.4.v_cache 0.00005193 0.01639791 + layer.4.output 1.40434551 241.54122297 + ------------------------------------------------------------------------------------- + TOTAL 0.60682162 102.67105160 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 317180 +BPFP 1.3373 bits/point +EBPFP 2.6745 equivalent bits/point +MSE 102.671052 +---------------------- -------------------------------------------------------- +Time: 2.986s Load: 0.009s, Pack+Encode: 1.740s, Decode+Unpack: 1.237s +---------------------- -------------------------------------------------------- +💾 Converting with 102.6711 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 193, 128) +Output shape: (1, 193, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.output: torch.Size([1, 193, 3584]) -> torch.Size([1, 1, 193, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,112B, BPFP=0.4948 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,296B, BPFP=2.6956 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,348B, BPFP=1.2426 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,236B, BPFP=2.6907 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,432B, BPFP=1.4922 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,528B, BPFP=2.7144 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,172B, BPFP=1.3902 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,440B, BPFP=2.7073 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,424B, BPFP=2.3821 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,880B, BPFP=2.6619 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,212B, BPFP=0.3610 +⌛️ [2/4] FRONTEND: Frontend time: 1.773s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.243s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14305212 40.79567580 + layer.0.v_cache 0.00001770 0.00733032 + layer.1.k_cache 0.14179654 8.38761119 + layer.1.v_cache 0.00000608 0.00277378 + layer.2.k_cache 0.01555968 0.73316703 + layer.2.v_cache 0.00002047 0.00810321 + layer.3.k_cache 0.02655400 5.06526935 + layer.3.v_cache 0.00002031 0.00959823 + layer.4.k_cache 0.00069483 0.23217105 + layer.4.v_cache 0.00005004 0.01634191 + layer.4.output 1.58615237 275.86755181 + ------------------------------------------------------------------------------------- + TOTAL 0.67240225 116.84299439 + (elements=1,679,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1679872 +Total Bytes 284080 +BPFP 1.3529 bits/point +EBPFP 2.7057 equivalent bits/point +MSE 116.842994 +---------------------- -------------------------------------------------------- +Time: 3.022s Load: 0.007s, Pack+Encode: 1.773s, Decode+Unpack: 1.243s +---------------------- -------------------------------------------------------- +💾 Converting with 116.8430 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,736B, BPFP=0.4854 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,968B, BPFP=2.3198 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,768B, BPFP=1.2405 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,988B, BPFP=2.2583 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,244B, BPFP=1.5213 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,696B, BPFP=2.2400 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,392B, BPFP=1.4051 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,436B, BPFP=2.2864 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,292B, BPFP=1.9636 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,388B, BPFP=2.2206 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,436B, BPFP=0.3715 +⌛️ [2/4] FRONTEND: Frontend time: 35.455s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.235s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10821181 41.12260762 + layer.0.v_cache 0.00001942 0.00661799 + layer.1.k_cache 0.30751975 9.14541074 + layer.1.v_cache 0.00000599 0.00244180 + layer.2.k_cache 0.01421910 0.75725283 + layer.2.v_cache 0.00002020 0.00729512 + layer.3.k_cache 0.01730646 6.05758728 + layer.3.v_cache 0.00002068 0.00838387 + layer.4.k_cache 0.00075739 0.20239431 + layer.4.v_cache 0.00005209 0.01544035 + layer.4.output 1.22958295 212.06746629 + ------------------------------------------------------------------------------------- + TOTAL 0.53265962 90.69398212 + (elements=2,167,296) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2167296 +Total Bytes 327344 +BPFP 1.2083 bits/point +EBPFP 2.4166 equivalent bits/point +MSE 90.693982 +---------------------- --------------------------------------------------------- +Time: 36.699s Load: 0.009s, Pack+Encode: 35.455s, Decode+Unpack: 1.235s +---------------------- --------------------------------------------------------- +💾 Converting with 90.6940 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 202, 128) +Output shape: (1, 202, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.output: torch.Size([1, 202, 3584]) -> torch.Size([1, 1, 202, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,652B, BPFP=0.5145 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,720B, BPFP=2.7630 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,296B, BPFP=1.2605 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,892B, BPFP=2.6989 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,656B, BPFP=1.5204 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,732B, BPFP=2.6866 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,744B, BPFP=1.3725 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,280B, BPFP=2.7290 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,312B, BPFP=2.3447 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,088B, BPFP=2.6368 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,148B, BPFP=0.3884 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.481s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15282421 40.88369721 + layer.0.v_cache 0.00001801 0.00763646 + layer.1.k_cache 0.08597872 9.57233935 + layer.1.v_cache 0.00000649 0.00290778 + layer.2.k_cache 0.01425038 0.88813329 + layer.2.v_cache 0.00002047 0.00802911 + layer.3.k_cache 0.05704349 5.23323104 + layer.3.v_cache 0.00002256 0.00978799 + layer.4.k_cache 0.00069767 0.23944301 + layer.4.v_cache 0.00005056 0.01652748 + layer.4.output 1.51553867 260.89721093 + ------------------------------------------------------------------------------------- + TOTAL 0.64233431 110.77307113 + (elements=1,758,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1758208 +Total Bytes 300520 +BPFP 1.3674 bits/point +EBPFP 2.7348 equivalent bits/point +MSE 110.773071 +---------------------- -------------------------------------------------------- +Time: 3.330s Load: 0.009s, Pack+Encode: 1.840s, Decode+Unpack: 1.481s +---------------------- -------------------------------------------------------- +💾 Converting with 110.7731 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,900B, BPFP=0.4914 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,428B, BPFP=2.5082 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,376B, BPFP=1.2354 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,388B, BPFP=2.4508 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,392B, BPFP=1.6228 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,108B, BPFP=2.4353 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,968B, BPFP=1.4337 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,828B, BPFP=2.4750 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,716B, BPFP=2.1376 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,820B, BPFP=2.4194 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,548B, BPFP=0.3277 +⌛️ [2/4] FRONTEND: Frontend time: 2.022s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.469s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14837397 38.95333205 + layer.0.v_cache 0.00001603 0.00622163 + layer.1.k_cache 0.60911565 8.35129679 + layer.1.v_cache 0.00000597 0.00239865 + layer.2.k_cache 0.02477714 0.91684521 + layer.2.v_cache 0.00001938 0.00697274 + layer.3.k_cache 0.01052417 5.85910929 + layer.3.v_cache 0.00002039 0.00781645 + layer.4.k_cache 0.00073933 0.20922172 + layer.4.v_cache 0.00005963 0.01527100 + layer.4.output 0.00470031 194.53650303 + ------------------------------------------------------------------------------------- + TOTAL 0.04862081 83.29905922 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 389472 +BPFP 1.2649 bits/point +EBPFP 2.5298 equivalent bits/point +MSE 83.299059 +---------------------- -------------------------------------------------------- +Time: 3.501s Load: 0.009s, Pack+Encode: 2.022s, Decode+Unpack: 1.469s +---------------------- -------------------------------------------------------- +💾 Converting with 83.2991 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 210, 128) +Output shape: (1, 210, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.output: torch.Size([1, 210, 3584]) -> torch.Size([1, 1, 210, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,820B, BPFP=0.5074 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,272B, BPFP=2.6988 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,856B, BPFP=1.2542 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,292B, BPFP=2.6259 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,824B, BPFP=1.5494 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,080B, BPFP=2.6101 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,792B, BPFP=1.4726 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,576B, BPFP=2.6470 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,024B, BPFP=2.3083 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,684B, BPFP=2.5807 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,196B, BPFP=0.3847 +⌛️ [2/4] FRONTEND: Frontend time: 1.851s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.336s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11415643 41.03927641 + layer.0.v_cache 0.00001882 0.00720051 + layer.1.k_cache 0.27145578 8.76988584 + layer.1.v_cache 0.00000623 0.00279507 + layer.2.k_cache 0.01201470 0.83929225 + layer.2.v_cache 0.00002208 0.00821101 + layer.3.k_cache 0.00887682 6.36412644 + layer.3.v_cache 0.00002165 0.00959449 + layer.4.k_cache 0.00072611 0.23874461 + layer.4.v_cache 0.00005233 0.01639424 + layer.4.output 1.45781997 237.27402211 + ------------------------------------------------------------------------------------- + TOTAL 0.62424063 101.07139268 + (elements=1,827,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1827840 +Total Bytes 308416 +BPFP 1.3499 bits/point +EBPFP 2.6997 equivalent bits/point +MSE 101.071393 +---------------------- -------------------------------------------------------- +Time: 3.194s Load: 0.006s, Pack+Encode: 1.851s, Decode+Unpack: 1.336s +---------------------- -------------------------------------------------------- +💾 Converting with 101.0714 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 165, 128) +Output shape: (1, 165, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.output: torch.Size([1, 165, 3584]) -> torch.Size([1, 1, 165, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,516B, BPFP=0.5223 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,708B, BPFP=2.6239 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,472B, BPFP=1.2758 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,948B, BPFP=2.5519 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,828B, BPFP=1.6883 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,584B, BPFP=2.5174 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,704B, BPFP=1.4871 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,172B, BPFP=2.5731 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,372B, BPFP=2.3080 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,380B, BPFP=2.4981 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,756B, BPFP=0.3890 +⌛️ [2/4] FRONTEND: Frontend time: 2.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.420s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10245823 40.40217803 + layer.0.v_cache 0.00002070 0.00785253 + layer.1.k_cache 0.13127365 9.57213838 + layer.1.v_cache 0.00000615 0.00289778 + layer.2.k_cache 0.01458293 0.97469584 + layer.2.v_cache 0.00002074 0.00860483 + layer.3.k_cache 0.02375244 4.29311745 + layer.3.v_cache 0.00002214 0.01024745 + layer.4.k_cache 0.00067489 0.24780449 + layer.4.v_cache 0.00005196 0.01785826 + layer.4.output 0.00973386 315.90343615 + ------------------------------------------------------------------------------------- + TOTAL 0.02005887 133.34479106 + (elements=1,436,160) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1436160 +Total Bytes 240440 +BPFP 1.3393 bits/point +EBPFP 2.6787 equivalent bits/point +MSE 133.344791 +---------------------- -------------------------------------------------------- +Time: 3.593s Load: 0.008s, Pack+Encode: 2.166s, Decode+Unpack: 1.420s +---------------------- -------------------------------------------------------- +💾 Converting with 133.3448 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,956B, BPFP=0.5085 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,904B, BPFP=2.3825 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,432B, BPFP=1.3176 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,152B, BPFP=2.3183 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,720B, BPFP=1.5984 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,872B, BPFP=2.2944 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,312B, BPFP=1.4781 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,400B, BPFP=2.3395 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,600B, BPFP=2.1004 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,640B, BPFP=2.2746 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,088B, BPFP=0.3914 +⌛️ [2/4] FRONTEND: Frontend time: 2.174s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.703s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12273688 40.45501676 + layer.0.v_cache 0.00001959 0.00767622 + layer.1.k_cache 0.11126613 9.22604820 + layer.1.v_cache 0.00000674 0.00315585 + layer.2.k_cache 0.01416154 0.78832024 + layer.2.v_cache 0.00002164 0.00875821 + layer.3.k_cache 0.04491856 4.98738774 + layer.3.v_cache 0.00002114 0.01047066 + layer.4.k_cache 0.00068813 0.23056383 + layer.4.v_cache 0.00006151 0.01719704 + layer.4.output 0.00884247 286.66444672 + ------------------------------------------------------------------------------------- + TOTAL 0.02092936 121.31680716 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 250076 +BPFP 1.2560 bits/point +EBPFP 2.5120 equivalent bits/point +MSE 121.316807 +---------------------- -------------------------------------------------------- +Time: 3.884s Load: 0.007s, Pack+Encode: 2.174s, Decode+Unpack: 1.703s +---------------------- -------------------------------------------------------- +💾 Converting with 121.3168 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 257, 128) +Output shape: (1, 257, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.output: torch.Size([1, 257, 3584]) -> torch.Size([1, 1, 257, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,940B, BPFP=0.4827 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,404B, BPFP=2.5781 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,116B, BPFP=1.2230 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,492B, BPFP=2.5226 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,264B, BPFP=1.4144 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,984B, BPFP=2.4917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,124B, BPFP=1.3451 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,512B, BPFP=2.5846 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,828B, BPFP=2.1783 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,736B, BPFP=2.5375 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,508B, BPFP=0.3518 +⌛️ [2/4] FRONTEND: Frontend time: 2.889s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.799s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12694686 40.95904517 + layer.0.v_cache 0.00002055 0.00699557 + layer.1.k_cache 0.54844232 8.75460067 + layer.1.v_cache 0.00000676 0.00271516 + layer.2.k_cache 0.01835482 0.71810889 + layer.2.v_cache 0.00002099 0.00776943 + layer.3.k_cache 0.02141366 5.00892396 + layer.3.v_cache 0.00002047 0.00865940 + layer.4.k_cache 0.00070767 0.21764596 + layer.4.v_cache 0.00005046 0.01502643 + layer.4.output 0.00513101 209.28010353 + ------------------------------------------------------------------------------------- + TOTAL 0.04422951 89.45060090 + (elements=2,236,928) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2236928 +Total Bytes 358908 +BPFP 1.2836 bits/point +EBPFP 2.5671 equivalent bits/point +MSE 89.450601 +---------------------- -------------------------------------------------------- +Time: 4.695s Load: 0.008s, Pack+Encode: 2.889s, Decode+Unpack: 1.799s +---------------------- -------------------------------------------------------- +💾 Converting with 89.4506 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,900B, BPFP=0.5065 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,000B, BPFP=2.4038 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,536B, BPFP=1.3338 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,940B, BPFP=2.3128 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,940B, BPFP=1.5402 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,708B, BPFP=2.2929 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,460B, BPFP=1.4131 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,376B, BPFP=2.3503 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,220B, BPFP=2.0793 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,540B, BPFP=2.2785 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,048B, BPFP=0.3440 +⌛️ [2/4] FRONTEND: Frontend time: 1.872s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12731336 41.81987788 + layer.0.v_cache 0.00001750 0.00751724 + layer.1.k_cache 0.10052182 9.54146500 + layer.1.v_cache 0.00000572 0.00268598 + layer.2.k_cache 0.01141533 0.86974066 + layer.2.v_cache 0.00001926 0.00790519 + layer.3.k_cache 0.03745206 5.71463918 + layer.3.v_cache 0.00002083 0.00979597 + layer.4.k_cache 0.00066429 0.23247195 + layer.4.v_cache 0.00005101 0.01634840 + layer.4.output 0.00884527 284.23692602 + ------------------------------------------------------------------------------------- + TOTAL 0.01996459 120.46358409 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 243668 +BPFP 1.2305 bits/point +EBPFP 2.4611 equivalent bits/point +MSE 120.463584 +---------------------- -------------------------------------------------------- +Time: 3.206s Load: 0.007s, Pack+Encode: 1.872s, Decode+Unpack: 1.327s +---------------------- -------------------------------------------------------- +💾 Converting with 120.4636 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,988B, BPFP=0.4918 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,144B, BPFP=2.6143 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,096B, BPFP=1.3440 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,780B, BPFP=2.5183 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,852B, BPFP=1.6788 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,468B, BPFP=2.4963 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,704B, BPFP=1.4572 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,316B, BPFP=2.5560 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,488B, BPFP=2.2162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,180B, BPFP=2.4761 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,932B, BPFP=0.3713 +⌛️ [2/4] FRONTEND: Frontend time: 2.105s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.488s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12752426 41.29156989 + layer.0.v_cache 0.00002004 0.00724960 + layer.1.k_cache 0.30660921 8.86972403 + layer.1.v_cache 0.00000613 0.00257440 + layer.2.k_cache 0.01808885 0.81538419 + layer.2.v_cache 0.00002107 0.00753739 + layer.3.k_cache 0.01678786 6.19662833 + layer.3.v_cache 0.00002050 0.00843737 + layer.4.k_cache 0.00068290 0.20972786 + layer.4.v_cache 0.00004851 0.01473144 + layer.4.output 1.37905047 237.88471284 + ------------------------------------------------------------------------------------- + TOTAL 0.59548016 101.33038555 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 318948 +BPFP 1.3205 bits/point +EBPFP 2.6410 equivalent bits/point +MSE 101.330386 +---------------------- -------------------------------------------------------- +Time: 3.601s Load: 0.008s, Pack+Encode: 2.105s, Decode+Unpack: 1.488s +---------------------- -------------------------------------------------------- +💾 Converting with 101.3304 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,984B, BPFP=0.5029 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,816B, BPFP=2.6509 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,524B, BPFP=1.3338 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,624B, BPFP=2.5651 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,796B, BPFP=1.6414 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,300B, BPFP=2.5418 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,404B, BPFP=1.3972 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,260B, BPFP=2.6109 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,444B, BPFP=2.2641 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,244B, BPFP=2.5377 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,496B, BPFP=0.4063 +⌛️ [2/4] FRONTEND: Frontend time: 2.172s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.566s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12766143 40.57808900 + layer.0.v_cache 0.00001653 0.00698212 + layer.1.k_cache 0.34631467 8.88463849 + layer.1.v_cache 0.00000659 0.00270735 + layer.2.k_cache 0.01786991 0.81473671 + layer.2.v_cache 0.00002041 0.00750732 + layer.3.k_cache 0.06874437 5.54655098 + layer.3.v_cache 0.00002276 0.00924210 + layer.4.k_cache 0.00067882 0.22204625 + layer.4.v_cache 0.00005621 0.01606592 + layer.4.output 1.41085444 243.13506007 + ------------------------------------------------------------------------------------- + TOTAL 0.61396310 103.41376393 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 317892 +BPFP 1.3465 bits/point +EBPFP 2.6929 equivalent bits/point +MSE 103.413764 +---------------------- -------------------------------------------------------- +Time: 3.747s Load: 0.008s, Pack+Encode: 2.172s, Decode+Unpack: 1.566s +---------------------- -------------------------------------------------------- +💾 Converting with 103.4138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,132B, BPFP=0.4931 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,596B, BPFP=2.5301 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,224B, BPFP=1.3291 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,908B, BPFP=2.4826 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,432B, BPFP=1.6892 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,472B, BPFP=2.4524 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,696B, BPFP=1.5000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,248B, BPFP=2.5061 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,628B, BPFP=2.1867 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,120B, BPFP=2.4281 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,876B, BPFP=0.4334 +⌛️ [2/4] FRONTEND: Frontend time: 2.004s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.590s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12070838 41.65990563 + layer.0.v_cache 0.00001763 0.00732627 + layer.1.k_cache 0.31015487 8.91056243 + layer.1.v_cache 0.00000677 0.00290292 + layer.2.k_cache 0.03544329 0.72623882 + layer.2.v_cache 0.00002212 0.00813293 + layer.3.k_cache 0.05682086 5.21640366 + layer.3.v_cache 0.00002136 0.00954322 + layer.4.k_cache 0.00073636 0.22965939 + layer.4.v_cache 0.00005031 0.01621162 + layer.4.output 1.35470685 229.56903840 + ------------------------------------------------------------------------------------- + TOTAL 0.58864294 97.86883269 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 327332 +BPFP 1.3312 bits/point +EBPFP 2.6624 equivalent bits/point +MSE 97.868833 +---------------------- -------------------------------------------------------- +Time: 3.602s Load: 0.008s, Pack+Encode: 2.004s, Decode+Unpack: 1.590s +---------------------- -------------------------------------------------------- +💾 Converting with 97.8688 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,800B, BPFP=0.4988 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,592B, BPFP=2.6843 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,104B, BPFP=1.2547 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,808B, BPFP=2.6268 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,260B, BPFP=1.6329 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,404B, BPFP=2.5971 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,080B, BPFP=1.4730 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,144B, BPFP=2.6514 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,452B, BPFP=2.3072 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,032B, BPFP=2.5698 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,552B, BPFP=0.4250 +⌛️ [2/4] FRONTEND: Frontend time: 2.112s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.510s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13227103 40.92476801 + layer.0.v_cache 0.00002050 0.00719971 + layer.1.k_cache 0.22864332 8.27753746 + layer.1.v_cache 0.00000662 0.00289296 + layer.2.k_cache 0.02950178 0.82858577 + layer.2.v_cache 0.00002095 0.00805107 + layer.3.k_cache 0.01835468 5.31806251 + layer.3.v_cache 0.00002123 0.00923344 + layer.4.k_cache 0.00070999 0.22221351 + layer.4.v_cache 0.00005457 0.01578120 + layer.4.output 1.43734575 239.01750084 + ------------------------------------------------------------------------------------- + TOTAL 0.61594264 101.69040185 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 317228 +BPFP 1.3689 bits/point +EBPFP 2.7377 equivalent bits/point +MSE 101.690402 +---------------------- -------------------------------------------------------- +Time: 3.630s Load: 0.008s, Pack+Encode: 2.112s, Decode+Unpack: 1.510s +---------------------- -------------------------------------------------------- +💾 Converting with 101.6904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,000B, BPFP=0.4927 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,804B, BPFP=2.5904 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,048B, BPFP=1.3407 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,804B, BPFP=2.5200 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,644B, BPFP=1.6641 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,444B, BPFP=2.4947 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,524B, BPFP=1.4445 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,068B, BPFP=2.5386 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,912B, BPFP=2.2461 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,080B, BPFP=2.4690 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,760B, BPFP=0.3696 +⌛️ [2/4] FRONTEND: Frontend time: 2.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.507s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12518616 41.84489812 + layer.0.v_cache 0.00001676 0.00707476 + layer.1.k_cache 0.33007166 9.18043422 + layer.1.v_cache 0.00000630 0.00284996 + layer.2.k_cache 0.01131002 0.77418394 + layer.2.v_cache 0.00002125 0.00822830 + layer.3.k_cache 0.02232737 5.85110666 + layer.3.v_cache 0.00001983 0.00875933 + layer.4.k_cache 0.00071526 0.23076413 + layer.4.v_cache 0.00005006 0.01505666 + layer.4.output 1.37904434 230.08015605 + ------------------------------------------------------------------------------------- + TOTAL 0.59664912 98.14614402 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 318088 +BPFP 1.3169 bits/point +EBPFP 2.6339 equivalent bits/point +MSE 98.146144 +---------------------- -------------------------------------------------------- +Time: 3.840s Load: 0.009s, Pack+Encode: 2.324s, Decode+Unpack: 1.507s +---------------------- -------------------------------------------------------- +💾 Converting with 98.1461 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,700B, BPFP=0.4931 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,904B, BPFP=2.3632 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,176B, BPFP=1.2920 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,932B, BPFP=2.3010 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,648B, BPFP=1.5784 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,488B, BPFP=2.2725 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,692B, BPFP=1.4531 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,084B, BPFP=2.3107 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,432B, BPFP=2.0128 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,140B, BPFP=2.2503 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,220B, BPFP=0.3771 +⌛️ [2/4] FRONTEND: Frontend time: 2.031s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.618s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14655781 41.17063028 + layer.0.v_cache 0.00001709 0.00670042 + layer.1.k_cache 0.36683986 9.52276211 + layer.1.v_cache 0.00000651 0.00267190 + layer.2.k_cache 0.01755252 0.82057159 + layer.2.v_cache 0.00002060 0.00752352 + layer.3.k_cache 0.02713263 5.85379178 + layer.3.v_cache 0.00001997 0.00846811 + layer.4.k_cache 0.00071102 0.19891017 + layer.4.v_cache 0.00005162 0.01497101 + layer.4.output 1.25479176 219.50620243 + ------------------------------------------------------------------------------------- + TOTAL 0.54955600 93.77355399 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 327416 +BPFP 1.2333 bits/point +EBPFP 2.4667 equivalent bits/point +MSE 93.773554 +---------------------- -------------------------------------------------------- +Time: 3.658s Load: 0.009s, Pack+Encode: 2.031s, Decode+Unpack: 1.618s +---------------------- -------------------------------------------------------- +💾 Converting with 93.7736 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 155, 128) +Output shape: (1, 155, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.output: torch.Size([1, 155, 3584]) -> torch.Size([1, 1, 155, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,192B, BPFP=0.5234 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,632B, BPFP=2.7855 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,984B, BPFP=1.3089 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,916B, BPFP=2.7133 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,532B, BPFP=1.6665 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,848B, BPFP=2.7065 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,640B, BPFP=1.5766 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,324B, BPFP=2.7544 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,472B, BPFP=2.4669 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,628B, BPFP=2.6843 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,688B, BPFP=0.3555 +⌛️ [2/4] FRONTEND: Frontend time: 2.266s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.404s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102877 40.85737462 + layer.0.v_cache 0.00001822 0.00737386 + layer.1.k_cache 0.12773433 9.34505891 + layer.1.v_cache 0.00000585 0.00278548 + layer.2.k_cache 0.00931232 0.96186612 + layer.2.v_cache 0.00002110 0.00841744 + layer.3.k_cache 0.01537714 4.91870117 + layer.3.v_cache 0.00002091 0.01019821 + layer.4.k_cache 0.00066471 0.24485540 + layer.4.v_cache 0.00005180 0.01734540 + layer.4.output 0.01742802 349.40671083 + ------------------------------------------------------------------------------------- + TOTAL 0.02389596 147.18946779 + (elements=1,349,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1349120 +Total Bytes 234856 +BPFP 1.3926 bits/point +EBPFP 2.7853 equivalent bits/point +MSE 147.189468 +---------------------- -------------------------------------------------------- +Time: 3.678s Load: 0.008s, Pack+Encode: 2.266s, Decode+Unpack: 1.404s +---------------------- -------------------------------------------------------- +💾 Converting with 147.1895 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,940B, BPFP=0.5044 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,768B, BPFP=2.6721 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,576B, BPFP=1.2773 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,724B, BPFP=2.5962 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,252B, BPFP=1.5445 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,256B, BPFP=2.5622 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,856B, BPFP=1.5157 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,196B, BPFP=2.6305 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,800B, BPFP=2.3110 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,020B, BPFP=2.5451 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,008B, BPFP=0.3635 +⌛️ [2/4] FRONTEND: Frontend time: 2.814s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.883s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11659910 40.06235011 + layer.0.v_cache 0.00001673 0.00717433 + layer.1.k_cache 0.29571861 8.31489201 + layer.1.v_cache 0.00000615 0.00276197 + layer.2.k_cache 0.04707494 0.80228030 + layer.2.v_cache 0.00002002 0.00763562 + layer.3.k_cache 0.03274950 5.75465854 + layer.3.v_cache 0.00002078 0.00885645 + layer.4.k_cache 0.00068608 0.24222847 + layer.4.v_cache 0.00005136 0.01524104 + layer.4.output 1.42391751 239.78162375 + ------------------------------------------------------------------------------------- + TOTAL 0.61531564 101.98173207 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 312396 +BPFP 1.3355 bits/point +EBPFP 2.6710 equivalent bits/point +MSE 101.981732 +---------------------- -------------------------------------------------------- +Time: 4.706s Load: 0.009s, Pack+Encode: 2.814s, Decode+Unpack: 1.883s +---------------------- -------------------------------------------------------- +💾 Converting with 101.9817 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,728B, BPFP=0.4869 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,164B, BPFP=2.3415 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,480B, BPFP=1.3533 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,144B, BPFP=2.2772 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,264B, BPFP=1.5287 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,628B, BPFP=2.2447 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,680B, BPFP=1.4289 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,484B, BPFP=2.2986 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,988B, BPFP=2.0154 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,424B, BPFP=2.2319 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,412B, BPFP=0.3907 +⌛️ [2/4] FRONTEND: Frontend time: 2.280s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.600s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10631144 41.14544481 + layer.0.v_cache 0.00001780 0.00723387 + layer.1.k_cache 0.46192560 9.53516117 + layer.1.v_cache 0.00000619 0.00283515 + layer.2.k_cache 0.02890189 0.71277274 + layer.2.v_cache 0.00002115 0.00807634 + layer.3.k_cache 0.04994748 5.66089655 + layer.3.v_cache 0.00002140 0.00975725 + layer.4.k_cache 0.00068365 0.25717177 + layer.4.v_cache 0.00005040 0.01565553 + layer.4.output 1.23456514 208.73626512 + ------------------------------------------------------------------------------------- + TOTAL 0.54646135 89.32405065 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 332396 +BPFP 1.2319 bits/point +EBPFP 2.4638 equivalent bits/point +MSE 89.324051 +---------------------- -------------------------------------------------------- +Time: 3.893s Load: 0.013s, Pack+Encode: 2.280s, Decode+Unpack: 1.600s +---------------------- -------------------------------------------------------- +💾 Converting with 89.3241 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 28.291s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 194, 128) +Output shape: (1, 194, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.output: torch.Size([1, 194, 3584]) -> torch.Size([1, 1, 194, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,136B, BPFP=0.4942 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,236B, BPFP=2.7574 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,200B, BPFP=1.3048 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,868B, BPFP=2.7278 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,528B, BPFP=1.4923 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,604B, BPFP=2.7065 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,288B, BPFP=1.3924 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,240B, BPFP=2.7577 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,124B, BPFP=2.3457 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,996B, BPFP=2.6575 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,212B, BPFP=0.3246 +⌛️ [2/4] FRONTEND: Frontend time: 2.155s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.500s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15027005 41.17184480 + layer.0.v_cache 0.00001782 0.00758181 + layer.1.k_cache 0.21402168 9.28112793 + layer.1.v_cache 0.00000588 0.00286117 + layer.2.k_cache 0.02177384 0.89838378 + layer.2.v_cache 0.00002019 0.00859263 + layer.3.k_cache 0.04024272 3.89845260 + layer.3.v_cache 0.00002076 0.01016865 + layer.4.k_cache 0.00069591 0.24720387 + layer.4.v_cache 0.00005073 0.01715940 + layer.4.output 1.57794378 270.66402798 + ------------------------------------------------------------------------------------- + TOTAL 0.67486624 114.71715132 + (elements=1,688,576) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 1688576 +Total Bytes 284432 +BPFP 1.3476 bits/point +EBPFP 2.6951 equivalent bits/point +MSE 114.717151 +---------------------- --------------------------------------------------------- +Time: 31.946s Load: 28.291s, Pack+Encode: 2.155s, Decode+Unpack: 1.500s +---------------------- --------------------------------------------------------- +💾 Converting with 114.7172 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 258, 128) +Output shape: (1, 258, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.output: torch.Size([1, 258, 3584]) -> torch.Size([1, 1, 258, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,076B, BPFP=0.4891 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,268B, BPFP=2.6204 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,160B, BPFP=1.2209 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,892B, BPFP=2.5371 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,108B, BPFP=1.3995 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,652B, BPFP=2.5225 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,228B, BPFP=1.3462 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,500B, BPFP=2.5133 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,180B, BPFP=2.1911 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,856B, BPFP=2.5349 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,900B, BPFP=0.3019 +⌛️ [2/4] FRONTEND: Frontend time: 2.237s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.662s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09035230 41.06379073 + layer.0.v_cache 0.00001726 0.00688537 + layer.1.k_cache 0.42455262 8.34962925 + layer.1.v_cache 0.00000600 0.00262061 + layer.2.k_cache 0.01891134 0.77475697 + layer.2.v_cache 0.00001996 0.00748852 + layer.3.k_cache 0.02093011 5.16611215 + layer.3.v_cache 0.00001928 0.00866066 + layer.4.k_cache 0.00072082 0.22652731 + layer.4.v_cache 0.00004951 0.01567026 + layer.4.output 0.00505962 211.14813469 + ------------------------------------------------------------------------------------- + TOTAL 0.03476450 90.21524027 + (elements=2,245,632) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2245632 +Total Bytes 354820 +BPFP 1.2640 bits/point +EBPFP 2.5281 equivalent bits/point +MSE 90.215240 +---------------------- -------------------------------------------------------- +Time: 3.910s Load: 0.010s, Pack+Encode: 2.237s, Decode+Unpack: 1.662s +---------------------- -------------------------------------------------------- +💾 Converting with 90.2152 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,380B, BPFP=0.5014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,696B, BPFP=2.4929 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,240B, BPFP=1.3071 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,840B, BPFP=2.4348 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,740B, BPFP=1.6807 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,064B, BPFP=2.3821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,408B, BPFP=1.4543 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,124B, BPFP=2.4541 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,336B, BPFP=2.1288 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,812B, BPFP=2.3649 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,668B, BPFP=0.4238 +⌛️ [2/4] FRONTEND: Frontend time: 1.997s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.486s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15008528 41.47936481 + layer.0.v_cache 0.00001847 0.00708871 + layer.1.k_cache 0.25612781 9.33296960 + layer.1.v_cache 0.00000678 0.00266237 + layer.2.k_cache 0.03931687 0.70872610 + layer.2.v_cache 0.00002107 0.00771510 + layer.3.k_cache 0.03786138 6.65157630 + layer.3.v_cache 0.00002177 0.00915080 + layer.4.k_cache 0.00069674 0.21343744 + layer.4.v_cache 0.00005115 0.01569875 + layer.4.output 1.33117570 228.18553960 + ------------------------------------------------------------------------------------- + TOTAL 0.57661395 97.39571571 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 326308 +BPFP 1.3040 bits/point +EBPFP 2.6080 equivalent bits/point +MSE 97.395716 +---------------------- -------------------------------------------------------- +Time: 3.492s Load: 0.009s, Pack+Encode: 1.997s, Decode+Unpack: 1.486s +---------------------- -------------------------------------------------------- +💾 Converting with 97.3957 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,840B, BPFP=0.5069 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,036B, BPFP=2.4337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,936B, BPFP=1.2965 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,236B, BPFP=2.3642 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,788B, BPFP=1.6309 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,952B, BPFP=2.3396 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,808B, BPFP=1.4590 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,568B, BPFP=2.3931 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,400B, BPFP=2.1181 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,776B, BPFP=2.3243 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,392B, BPFP=0.4141 +⌛️ [2/4] FRONTEND: Frontend time: 1.829s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15161699 40.07671984 + layer.0.v_cache 0.00001786 0.00749990 + layer.1.k_cache 0.08662784 8.73593818 + layer.1.v_cache 0.00000652 0.00288475 + layer.2.k_cache 0.02063684 0.87397334 + layer.2.v_cache 0.00002327 0.00842244 + layer.3.k_cache 0.08087930 5.98264296 + layer.3.v_cache 0.00002116 0.00976251 + layer.4.k_cache 0.00074161 0.25743332 + layer.4.v_cache 0.00006096 0.01721098 + layer.4.output 0.00898438 299.48194444 + ------------------------------------------------------------------------------------- + TOTAL 0.02373665 126.60859408 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 250732 +BPFP 1.2803 bits/point +EBPFP 2.5606 equivalent bits/point +MSE 126.608594 +---------------------- -------------------------------------------------------- +Time: 3.186s Load: 0.006s, Pack+Encode: 1.829s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 126.6086 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,000B, BPFP=0.4972 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,768B, BPFP=2.6114 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,624B, BPFP=1.3227 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,820B, BPFP=2.5440 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,520B, BPFP=1.7415 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,476B, BPFP=2.5196 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,988B, BPFP=1.4196 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,072B, BPFP=2.5619 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,672B, BPFP=2.2494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,416B, BPFP=2.5153 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,044B, BPFP=0.3759 +⌛️ [2/4] FRONTEND: Frontend time: 2.118s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.508s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14327769 40.35616122 + layer.0.v_cache 0.00001763 0.00729670 + layer.1.k_cache 0.34925704 9.07558316 + layer.1.v_cache 0.00000642 0.00288687 + layer.2.k_cache 0.02352169 0.88451885 + layer.2.v_cache 0.00002271 0.00825964 + layer.3.k_cache 0.00845094 5.71814575 + layer.3.v_cache 0.00002026 0.00961479 + layer.4.k_cache 0.00069169 0.23419694 + layer.4.v_cache 0.00005086 0.01642503 + layer.4.output 1.39158238 243.70746753 + ------------------------------------------------------------------------------------- + TOTAL 0.60390550 103.66266833 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 318400 +BPFP 1.3302 bits/point +EBPFP 2.6604 equivalent bits/point +MSE 103.662668 +---------------------- -------------------------------------------------------- +Time: 3.633s Load: 0.008s, Pack+Encode: 2.118s, Decode+Unpack: 1.508s +---------------------- -------------------------------------------------------- +💾 Converting with 103.6627 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,604B, BPFP=0.4889 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,244B, BPFP=2.5707 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,712B, BPFP=1.2905 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,984B, BPFP=2.4991 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,784B, BPFP=1.4650 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,744B, BPFP=2.4855 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,560B, BPFP=1.3955 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,552B, BPFP=2.5314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,984B, BPFP=2.2150 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,348B, BPFP=2.4630 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,064B, BPFP=0.3333 +⌛️ [2/4] FRONTEND: Frontend time: 2.390s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.662s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12716211 41.22813920 + layer.0.v_cache 0.00001677 0.00685360 + layer.1.k_cache 0.51313377 8.41870117 + layer.1.v_cache 0.00000586 0.00262721 + layer.2.k_cache 0.02995576 0.81898288 + layer.2.v_cache 0.00002133 0.00792376 + layer.3.k_cache 0.03339291 6.31742631 + layer.3.v_cache 0.00002012 0.00874773 + layer.4.k_cache 0.00068476 0.23392825 + layer.4.v_cache 0.00005875 0.01485935 + layer.4.output 0.00481428 198.80113636 + ------------------------------------------------------------------------------------- + TOTAL 0.04342071 85.21565553 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 382580 +BPFP 1.2787 bits/point +EBPFP 2.5574 equivalent bits/point +MSE 85.215656 +---------------------- -------------------------------------------------------- +Time: 4.062s Load: 0.010s, Pack+Encode: 2.390s, Decode+Unpack: 1.662s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2157 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,836B, BPFP=0.4820 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,212B, BPFP=2.2891 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,288B, BPFP=1.3095 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,492B, BPFP=2.2448 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,936B, BPFP=1.4724 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,280B, BPFP=2.2318 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,036B, BPFP=1.4171 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,976B, BPFP=2.2746 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,736B, BPFP=1.9523 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,788B, BPFP=2.2015 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,240B, BPFP=0.3448 +⌛️ [2/4] FRONTEND: Frontend time: 2.234s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.599s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13396681 42.12972902 + layer.0.v_cache 0.00001641 0.00647496 + layer.1.k_cache 0.47111181 9.99457124 + layer.1.v_cache 0.00000608 0.00257693 + layer.2.k_cache 0.02933135 0.79881040 + layer.2.v_cache 0.00002255 0.00749253 + layer.3.k_cache 0.03744583 6.17245171 + layer.3.v_cache 0.00002260 0.00815326 + layer.4.k_cache 0.00071777 0.24313841 + layer.4.v_cache 0.00004917 0.01435668 + layer.4.output 1.20538323 203.11076350 + ------------------------------------------------------------------------------------- + TOTAL 0.53590429 87.12665292 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 329820 +BPFP 1.1935 bits/point +EBPFP 2.3870 equivalent bits/point +MSE 87.126653 +---------------------- -------------------------------------------------------- +Time: 3.842s Load: 0.010s, Pack+Encode: 2.234s, Decode+Unpack: 1.599s +---------------------- -------------------------------------------------------- +💾 Converting with 87.1267 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,840B, BPFP=0.4796 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,876B, BPFP=2.4889 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,412B, BPFP=1.3244 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,768B, BPFP=2.4288 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,368B, BPFP=1.5933 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,320B, BPFP=2.4045 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,144B, BPFP=1.4184 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,156B, BPFP=2.4499 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,576B, BPFP=2.1471 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,844B, BPFP=2.3787 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,344B, BPFP=0.3359 +⌛️ [2/4] FRONTEND: Frontend time: 2.264s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.700s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14060297 41.13810900 + layer.0.v_cache 0.00001598 0.00652140 + layer.1.k_cache 0.58610762 8.86348046 + layer.1.v_cache 0.00000669 0.00255282 + layer.2.k_cache 0.02834670 0.78673448 + layer.2.v_cache 0.00002256 0.00747184 + layer.3.k_cache 0.01962393 6.96139230 + layer.3.v_cache 0.00002003 0.00821937 + layer.4.k_cache 0.00067797 0.22706432 + layer.4.v_cache 0.00005035 0.01461843 + layer.4.output 0.00463834 188.39098152 + ------------------------------------------------------------------------------------- + TOTAL 0.04752607 80.98547265 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 395648 +BPFP 1.2627 bits/point +EBPFP 2.5253 equivalent bits/point +MSE 80.985473 +---------------------- -------------------------------------------------------- +Time: 3.975s Load: 0.012s, Pack+Encode: 2.264s, Decode+Unpack: 1.700s +---------------------- -------------------------------------------------------- +💾 Converting with 80.9855 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,480B, BPFP=0.4873 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,112B, BPFP=2.3701 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,684B, BPFP=1.3201 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,832B, BPFP=2.3043 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,668B, BPFP=1.5249 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,576B, BPFP=2.2911 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,544B, BPFP=1.4157 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,376B, BPFP=2.3322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,432B, BPFP=2.0267 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,868B, BPFP=2.2547 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,776B, BPFP=0.3141 +⌛️ [2/4] FRONTEND: Frontend time: 2.270s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.870s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12404394 40.89998265 + layer.0.v_cache 0.00001645 0.00645829 + layer.1.k_cache 0.57408182 8.75859391 + layer.1.v_cache 0.00000616 0.00250794 + layer.2.k_cache 0.03522489 0.87980843 + layer.2.v_cache 0.00002326 0.00734712 + layer.3.k_cache 0.02908128 6.54598999 + layer.3.v_cache 0.00002078 0.00809100 + layer.4.k_cache 0.00072493 0.22239916 + layer.4.v_cache 0.00004967 0.01429265 + layer.4.output 0.04390477 175.87201891 + ------------------------------------------------------------------------------------- + TOTAL 0.06297686 75.79115315 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 399348 +BPFP 1.2074 bits/point +EBPFP 2.4148 equivalent bits/point +MSE 75.791153 +---------------------- -------------------------------------------------------- +Time: 4.152s Load: 0.012s, Pack+Encode: 2.270s, Decode+Unpack: 1.870s +---------------------- -------------------------------------------------------- +💾 Converting with 75.7912 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 242, 128) +Output shape: (1, 242, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.output: torch.Size([1, 242, 3584]) -> torch.Size([1, 1, 242, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,588B, BPFP=0.4899 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,044B, BPFP=2.3918 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,484B, BPFP=1.2580 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,916B, BPFP=2.3190 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,708B, BPFP=1.5953 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,600B, BPFP=2.2986 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,544B, BPFP=1.4556 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,412B, BPFP=2.3510 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,756B, BPFP=2.0504 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,296B, BPFP=2.2789 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,980B, BPFP=0.3688 +⌛️ [2/4] FRONTEND: Frontend time: 2.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.550s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11466674 40.02755763 + layer.0.v_cache 0.00001794 0.00701805 + layer.1.k_cache 0.39062487 9.51993681 + layer.1.v_cache 0.00000616 0.00266957 + layer.2.k_cache 0.02414889 0.79621295 + layer.2.v_cache 0.00002046 0.00762309 + layer.3.k_cache 0.03495786 6.18515040 + layer.3.v_cache 0.00002053 0.00876736 + layer.4.k_cache 0.00069125 0.21859368 + layer.4.v_cache 0.00005034 0.01581117 + layer.4.output 1.26514320 212.00826446 + ------------------------------------------------------------------------------------- + TOTAL 0.55418868 90.63807011 + (elements=2,106,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2106368 +Total Bytes 326328 +BPFP 1.2394 bits/point +EBPFP 2.4788 equivalent bits/point +MSE 90.638070 +---------------------- -------------------------------------------------------- +Time: 3.785s Load: 0.010s, Pack+Encode: 2.226s, Decode+Unpack: 1.550s +---------------------- -------------------------------------------------------- +💾 Converting with 90.6381 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 189, 128) +Output shape: (1, 189, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.output: torch.Size([1, 189, 3584]) -> torch.Size([1, 1, 189, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,988B, BPFP=0.4950 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,796B, BPFP=2.2979 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,168B, BPFP=1.3366 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,184B, BPFP=2.2474 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,236B, BPFP=1.5076 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,976B, BPFP=2.2302 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,140B, BPFP=1.4170 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,500B, BPFP=2.2735 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,172B, BPFP=1.9983 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,604B, BPFP=2.1994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,604B, BPFP=0.4087 +⌛️ [2/4] FRONTEND: Frontend time: 2.090s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.583s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14361043 42.47705336 + layer.0.v_cache 0.00001648 0.00706700 + layer.1.k_cache 0.18932698 10.02135778 + layer.1.v_cache 0.00000663 0.00291068 + layer.2.k_cache 0.00768016 0.77467217 + layer.2.v_cache 0.00002453 0.00821958 + layer.3.k_cache 0.01421528 5.86004542 + layer.3.v_cache 0.00002024 0.00904487 + layer.4.k_cache 0.00072804 0.22763935 + layer.4.v_cache 0.00005086 0.01562329 + layer.4.output 0.00858909 281.10744992 + ------------------------------------------------------------------------------------- + TOTAL 0.02445902 119.24445782 + (elements=1,645,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1645056 +Total Bytes 252368 +BPFP 1.2273 bits/point +EBPFP 2.4546 equivalent bits/point +MSE 119.244458 +---------------------- -------------------------------------------------------- +Time: 3.680s Load: 0.006s, Pack+Encode: 2.090s, Decode+Unpack: 1.583s +---------------------- -------------------------------------------------------- +💾 Converting with 119.2445 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,860B, BPFP=0.5032 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,512B, BPFP=2.6784 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,768B, BPFP=1.3034 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,756B, BPFP=2.6229 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,220B, BPFP=1.5566 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,372B, BPFP=2.5948 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,764B, BPFP=1.3765 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,008B, BPFP=2.6414 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,436B, BPFP=2.3060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,028B, BPFP=2.5695 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,672B, BPFP=0.3529 +⌛️ [2/4] FRONTEND: Frontend time: 2.137s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.543s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16894778 40.99089000 + layer.0.v_cache 0.00001779 0.00717245 + layer.1.k_cache 0.37628758 9.26189068 + layer.1.v_cache 0.00000652 0.00275386 + layer.2.k_cache 0.02181128 0.79687894 + layer.2.v_cache 0.00002030 0.00783644 + layer.3.k_cache 0.01815432 5.94546666 + layer.3.v_cache 0.00002063 0.00920110 + layer.4.k_cache 0.00071780 0.21006123 + layer.4.v_cache 0.00005059 0.01542261 + layer.4.output 1.43727796 232.41555584 + ------------------------------------------------------------------------------------- + TOTAL 0.62629296 99.06802734 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 308396 +BPFP 1.3308 bits/point +EBPFP 2.6615 equivalent bits/point +MSE 99.068027 +---------------------- -------------------------------------------------------- +Time: 3.687s Load: 0.007s, Pack+Encode: 2.137s, Decode+Unpack: 1.543s +---------------------- -------------------------------------------------------- +💾 Converting with 99.0680 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,028B, BPFP=0.5084 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,632B, BPFP=2.6499 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,208B, BPFP=1.2448 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,764B, BPFP=2.5871 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,120B, BPFP=1.6001 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,296B, BPFP=2.5532 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,764B, BPFP=1.5020 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,976B, BPFP=2.6024 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,824B, BPFP=2.2297 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,932B, BPFP=2.5269 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,256B, BPFP=0.3333 +⌛️ [2/4] FRONTEND: Frontend time: 2.173s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.526s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09877178 40.79060420 + layer.0.v_cache 0.00001652 0.00675416 + layer.1.k_cache 0.28824160 8.28302115 + layer.1.v_cache 0.00000619 0.00266906 + layer.2.k_cache 0.02137197 0.80876365 + layer.2.v_cache 0.00001995 0.00744328 + layer.3.k_cache 0.03811082 5.67573491 + layer.3.v_cache 0.00002058 0.00856220 + layer.4.k_cache 0.00070031 0.20818251 + layer.4.v_cache 0.00004695 0.01392110 + layer.4.output 1.41729720 242.14242312 + ------------------------------------------------------------------------------------- + TOTAL 0.60990512 102.98838930 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 308800 +BPFP 1.3140 bits/point +EBPFP 2.6280 equivalent bits/point +MSE 102.988389 +---------------------- -------------------------------------------------------- +Time: 3.707s Load: 0.008s, Pack+Encode: 2.173s, Decode+Unpack: 1.526s +---------------------- -------------------------------------------------------- +💾 Converting with 102.9884 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,052B, BPFP=0.4876 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,800B, BPFP=2.5442 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,372B, BPFP=1.3393 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,884B, BPFP=2.4809 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,828B, BPFP=1.6474 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,448B, BPFP=2.4508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,032B, BPFP=1.4541 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,960B, BPFP=2.4862 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,604B, BPFP=2.1850 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,924B, BPFP=2.4145 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,064B, BPFP=0.3661 +⌛️ [2/4] FRONTEND: Frontend time: 2.174s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.507s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12161526 40.78513897 + layer.0.v_cache 0.00001916 0.00709255 + layer.1.k_cache 0.26028682 8.39297836 + layer.1.v_cache 0.00000628 0.00269569 + layer.2.k_cache 0.01296707 0.75969953 + layer.2.v_cache 0.00002040 0.00766959 + layer.3.k_cache 0.04595671 6.55714727 + layer.3.v_cache 0.00002073 0.00872970 + layer.4.k_cache 0.00067352 0.21494086 + layer.4.v_cache 0.00005047 0.01574544 + layer.4.output 1.35465500 234.03766988 + ------------------------------------------------------------------------------------- + TOTAL 0.58377655 99.70679571 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 318968 +BPFP 1.2972 bits/point +EBPFP 2.5944 equivalent bits/point +MSE 99.706796 +---------------------- -------------------------------------------------------- +Time: 3.688s Load: 0.007s, Pack+Encode: 2.174s, Decode+Unpack: 1.507s +---------------------- -------------------------------------------------------- +💾 Converting with 99.7068 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,092B, BPFP=0.4903 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,608B, BPFP=2.5310 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,232B, BPFP=1.3988 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,748B, BPFP=2.4715 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,692B, BPFP=1.6380 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,460B, BPFP=2.4516 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,376B, BPFP=1.4779 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,988B, BPFP=2.4881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,496B, BPFP=2.1775 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,004B, BPFP=2.4201 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,916B, BPFP=0.3449 +⌛️ [2/4] FRONTEND: Frontend time: 2.142s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.536s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10229862 43.11144912 + layer.0.v_cache 0.00001727 0.00689423 + layer.1.k_cache 0.34669285 8.65973224 + layer.1.v_cache 0.00000643 0.00267584 + layer.2.k_cache 0.01639456 0.78402656 + layer.2.v_cache 0.00002164 0.00778705 + layer.3.k_cache 0.03343437 5.60763901 + layer.3.v_cache 0.00002043 0.00874681 + layer.4.k_cache 0.00071218 0.21071010 + layer.4.v_cache 0.00005088 0.01569702 + layer.4.output 1.35462105 232.88373104 + ------------------------------------------------------------------------------------- + TOTAL 0.58717627 99.32949854 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 317612 +BPFP 1.2917 bits/point +EBPFP 2.5834 equivalent bits/point +MSE 99.329499 +---------------------- -------------------------------------------------------- +Time: 3.686s Load: 0.008s, Pack+Encode: 2.142s, Decode+Unpack: 1.536s +---------------------- -------------------------------------------------------- +💾 Converting with 99.3295 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,156B, BPFP=0.4926 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,544B, BPFP=2.5154 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,284B, BPFP=1.3274 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,528B, BPFP=2.4455 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,560B, BPFP=1.6217 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,244B, BPFP=2.4259 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,572B, BPFP=1.4849 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,840B, BPFP=2.4670 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,448B, BPFP=2.1646 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,880B, BPFP=2.4009 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,832B, BPFP=0.3327 +⌛️ [2/4] FRONTEND: Frontend time: 2.018s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.474s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11800324 41.93652774 + layer.0.v_cache 0.00001819 0.00701503 + layer.1.k_cache 0.35011090 8.93707786 + layer.1.v_cache 0.00000612 0.00264710 + layer.2.k_cache 0.02017111 0.69474154 + layer.2.v_cache 0.00002006 0.00767255 + layer.3.k_cache 0.03026835 5.85679055 + layer.3.v_cache 0.00001975 0.00900182 + layer.4.k_cache 0.00074701 0.20839784 + layer.4.v_cache 0.00005118 0.01536563 + layer.4.output 1.34863711 236.39423379 + ------------------------------------------------------------------------------------- + TOTAL 0.58587504 100.73146319 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 314888 +BPFP 1.2750 bits/point +EBPFP 2.5499 equivalent bits/point +MSE 100.731463 +---------------------- -------------------------------------------------------- +Time: 3.499s Load: 0.008s, Pack+Encode: 2.018s, Decode+Unpack: 1.474s +---------------------- -------------------------------------------------------- +💾 Converting with 100.7315 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,344B, BPFP=0.5011 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,904B, BPFP=2.5180 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,248B, BPFP=1.3133 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,856B, BPFP=2.4465 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,136B, BPFP=1.6468 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,652B, BPFP=2.4326 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,692B, BPFP=1.4801 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,564B, BPFP=2.4948 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,028B, BPFP=2.1853 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,308B, BPFP=2.4091 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,908B, BPFP=0.3598 +⌛️ [2/4] FRONTEND: Frontend time: 2.465s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.629s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10957497 42.55359580 + layer.0.v_cache 0.00001697 0.00698899 + layer.1.k_cache 0.31262030 8.58503237 + layer.1.v_cache 0.00000629 0.00280829 + layer.2.k_cache 0.01311818 0.76253526 + layer.2.v_cache 0.00002149 0.00817045 + layer.3.k_cache 0.06227440 5.62111827 + layer.3.v_cache 0.00002139 0.00953111 + layer.4.k_cache 0.00068925 0.23392497 + layer.4.v_cache 0.00005302 0.01669226 + layer.4.output 1.33688878 228.01029320 + ------------------------------------------------------------------------------------- + TOTAL 0.57980104 97.28661472 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 321640 +BPFP 1.2909 bits/point +EBPFP 2.5819 equivalent bits/point +MSE 97.286615 +---------------------- -------------------------------------------------------- +Time: 4.105s Load: 0.010s, Pack+Encode: 2.465s, Decode+Unpack: 1.629s +---------------------- -------------------------------------------------------- +💾 Converting with 97.2866 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,820B, BPFP=0.5003 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,732B, BPFP=2.6945 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,836B, BPFP=1.2350 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,544B, BPFP=2.6074 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,108B, BPFP=1.4751 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,972B, BPFP=2.5654 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,868B, BPFP=1.3841 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,848B, BPFP=2.6297 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,228B, BPFP=2.2908 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,756B, BPFP=2.5496 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,520B, BPFP=0.3513 +⌛️ [2/4] FRONTEND: Frontend time: 2.336s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.737s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13987277 41.35392826 + layer.0.v_cache 0.00001887 0.00713248 + layer.1.k_cache 0.23498858 8.76986479 + layer.1.v_cache 0.00000603 0.00263337 + layer.2.k_cache 0.02028183 0.82843677 + layer.2.v_cache 0.00001992 0.00784282 + layer.3.k_cache 0.05552547 5.74295603 + layer.3.v_cache 0.00002044 0.00905659 + layer.4.k_cache 0.00067773 0.21248297 + layer.4.v_cache 0.00004777 0.01530310 + layer.4.output 1.43727485 235.69462609 + ------------------------------------------------------------------------------------- + TOTAL 0.61837549 100.40070705 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 305232 +BPFP 1.3171 bits/point +EBPFP 2.6342 equivalent bits/point +MSE 100.400707 +---------------------- -------------------------------------------------------- +Time: 4.084s Load: 0.011s, Pack+Encode: 2.336s, Decode+Unpack: 1.737s +---------------------- -------------------------------------------------------- +💾 Converting with 100.4007 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,804B, BPFP=0.5039 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,184B, BPFP=2.6795 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,428B, BPFP=1.2906 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,212B, BPFP=2.6075 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,032B, BPFP=1.4834 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,800B, BPFP=2.5770 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,596B, BPFP=1.4511 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,604B, BPFP=2.6366 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,204B, BPFP=2.3107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,584B, BPFP=2.5610 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,236B, BPFP=0.3410 +⌛️ [2/4] FRONTEND: Frontend time: 2.148s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.643s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17004660 42.65536137 + layer.0.v_cache 0.00001740 0.00695327 + layer.1.k_cache 0.16740718 8.49083778 + layer.1.v_cache 0.00000654 0.00275739 + layer.2.k_cache 0.01191964 0.79103638 + layer.2.v_cache 0.00001991 0.00776835 + layer.3.k_cache 0.00660476 4.93374243 + layer.3.v_cache 0.00002076 0.00884322 + layer.4.k_cache 0.00070551 0.23008489 + layer.4.v_cache 0.00005179 0.01593321 + layer.4.output 1.45088344 252.51305433 + ------------------------------------------------------------------------------------- + TOTAL 0.61841083 107.33733521 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 303684 +BPFP 1.3228 bits/point +EBPFP 2.6457 equivalent bits/point +MSE 107.337335 +---------------------- -------------------------------------------------------- +Time: 3.799s Load: 0.008s, Pack+Encode: 2.148s, Decode+Unpack: 1.643s +---------------------- -------------------------------------------------------- +💾 Converting with 107.3373 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,356B, BPFP=0.4997 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,568B, BPFP=2.4842 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,392B, BPFP=1.3174 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,640B, BPFP=2.4212 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,412B, BPFP=1.5905 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,364B, BPFP=2.4024 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,596B, BPFP=1.3992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,932B, BPFP=2.4410 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,648B, BPFP=2.1500 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,848B, BPFP=2.3674 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,112B, BPFP=0.3214 +⌛️ [2/4] FRONTEND: Frontend time: 2.090s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.571s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14712874 41.44341033 + layer.0.v_cache 0.00001826 0.00664235 + layer.1.k_cache 0.36788522 8.74260360 + layer.1.v_cache 0.00000598 0.00261096 + layer.2.k_cache 0.01156470 0.76424693 + layer.2.v_cache 0.00002030 0.00787805 + layer.3.k_cache 0.01315050 5.44407853 + layer.3.v_cache 0.00002017 0.00855486 + layer.4.k_cache 0.00068370 0.22075806 + layer.4.v_cache 0.00004817 0.01503674 + layer.4.output 1.33105469 219.70737578 + ------------------------------------------------------------------------------------- + TOTAL 0.57987697 93.80043828 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 313868 +BPFP 1.2543 bits/point +EBPFP 2.5085 equivalent bits/point +MSE 93.800438 +---------------------- -------------------------------------------------------- +Time: 3.669s Load: 0.008s, Pack+Encode: 2.090s, Decode+Unpack: 1.571s +---------------------- -------------------------------------------------------- +💾 Converting with 93.8004 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 188, 128) +Output shape: (1, 188, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.output: torch.Size([1, 188, 3584]) -> torch.Size([1, 1, 188, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,928B, BPFP=0.4927 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,720B, BPFP=2.3039 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,944B, BPFP=1.3251 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,132B, BPFP=2.2550 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,256B, BPFP=1.5173 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,868B, BPFP=2.2330 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,448B, BPFP=1.3670 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,288B, BPFP=2.2680 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,128B, BPFP=2.0053 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,636B, BPFP=2.2138 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,768B, BPFP=0.3534 +⌛️ [2/4] FRONTEND: Frontend time: 2.117s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.596s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005002 40.28161361 + layer.0.v_cache 0.00001663 0.00686858 + layer.1.k_cache 0.14149688 9.31110139 + layer.1.v_cache 0.00000584 0.00264017 + layer.2.k_cache 0.00881058 0.84968421 + layer.2.v_cache 0.00002285 0.00823056 + layer.3.k_cache 0.03192479 4.14490753 + layer.3.v_cache 0.00001958 0.00898922 + layer.4.k_cache 0.00066612 0.22073675 + layer.4.v_cache 0.00005162 0.01638441 + layer.4.output 0.00857985 286.62599734 + ------------------------------------------------------------------------------------- + TOTAL 0.02253670 121.24900811 + (elements=1,636,352) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1636352 +Total Bytes 246116 +BPFP 1.2032 bits/point +EBPFP 2.4065 equivalent bits/point +MSE 121.249008 +---------------------- -------------------------------------------------------- +Time: 3.721s Load: 0.007s, Pack+Encode: 2.117s, Decode+Unpack: 1.596s +---------------------- -------------------------------------------------------- +💾 Converting with 121.2490 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,536B, BPFP=0.5132 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,528B, BPFP=2.7896 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,276B, BPFP=1.2780 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,900B, BPFP=2.7403 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,812B, BPFP=1.5556 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,468B, BPFP=2.7063 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,316B, BPFP=1.3596 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,984B, BPFP=2.7469 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,052B, BPFP=2.3596 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,192B, BPFP=2.6847 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,612B, BPFP=0.3546 +⌛️ [2/4] FRONTEND: Frontend time: 2.747s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.696s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11435634 41.18259265 + layer.0.v_cache 0.00001764 0.00727622 + layer.1.k_cache 0.25025516 9.08730481 + layer.1.v_cache 0.00000690 0.00280136 + layer.2.k_cache 0.01528106 0.86663105 + layer.2.v_cache 0.00002105 0.00794731 + layer.3.k_cache 0.02533426 4.14879451 + layer.3.v_cache 0.00002127 0.00950571 + layer.4.k_cache 0.00070809 0.21784797 + layer.4.v_cache 0.00005072 0.01580635 + layer.4.output 1.53833902 263.65577889 + ------------------------------------------------------------------------------------- + TOTAL 0.65731915 111.83158589 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 295676 +BPFP 1.3656 bits/point +EBPFP 2.7313 equivalent bits/point +MSE 111.831586 +---------------------- -------------------------------------------------------- +Time: 4.452s Load: 0.009s, Pack+Encode: 2.747s, Decode+Unpack: 1.696s +---------------------- -------------------------------------------------------- +💾 Converting with 111.8316 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,860B, BPFP=0.5087 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,988B, BPFP=2.4295 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,284B, BPFP=1.2399 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,020B, BPFP=2.3455 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,752B, BPFP=1.6278 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,996B, BPFP=2.3434 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,880B, BPFP=1.4653 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,600B, BPFP=2.3958 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,372B, BPFP=2.1156 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,756B, BPFP=2.3226 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,764B, BPFP=0.3939 +⌛️ [2/4] FRONTEND: Frontend time: 1.983s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.664s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258550 40.64502496 + layer.0.v_cache 0.00001767 0.00747193 + layer.1.k_cache 0.14790291 9.18350423 + layer.1.v_cache 0.00000629 0.00281316 + layer.2.k_cache 0.00999225 0.85226746 + layer.2.v_cache 0.00002110 0.00829074 + layer.3.k_cache 0.02117847 5.01164347 + layer.3.v_cache 0.00002045 0.01006456 + layer.4.k_cache 0.00067024 0.22266973 + layer.4.v_cache 0.00006105 0.01692247 + layer.4.output 0.00897616 301.09809028 + ------------------------------------------------------------------------------------- + TOTAL 0.02207583 127.27337086 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 248272 +BPFP 1.2677 bits/point +EBPFP 2.5355 equivalent bits/point +MSE 127.273371 +---------------------- -------------------------------------------------------- +Time: 3.654s Load: 0.007s, Pack+Encode: 1.983s, Decode+Unpack: 1.664s +---------------------- -------------------------------------------------------- +💾 Converting with 127.2734 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,732B, BPFP=0.5156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,720B, BPFP=2.7359 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,344B, BPFP=1.2518 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,032B, BPFP=2.6832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,996B, BPFP=1.5316 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,692B, BPFP=2.6572 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,912B, BPFP=1.3719 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,288B, BPFP=2.7028 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,608B, BPFP=2.3444 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,496B, BPFP=2.6422 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,576B, BPFP=0.3783 +⌛️ [2/4] FRONTEND: Frontend time: 2.636s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.547s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09364704 40.95813706 + layer.0.v_cache 0.00001853 0.00713961 + layer.1.k_cache 0.25163280 9.10641061 + layer.1.v_cache 0.00000592 0.00273078 + layer.2.k_cache 0.01595125 0.74847397 + layer.2.v_cache 0.00002050 0.00785121 + layer.3.k_cache 0.04325477 5.00585608 + layer.3.v_cache 0.00002016 0.00927155 + layer.4.k_cache 0.00069210 0.22000687 + layer.4.v_cache 0.00005026 0.01543795 + layer.4.output 1.50065788 260.60121236 + ------------------------------------------------------------------------------------- + TOTAL 0.64175873 110.60528248 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 301396 +BPFP 1.3579 bits/point +EBPFP 2.7159 equivalent bits/point +MSE 110.605282 +---------------------- -------------------------------------------------------- +Time: 4.193s Load: 0.010s, Pack+Encode: 2.636s, Decode+Unpack: 1.547s +---------------------- -------------------------------------------------------- +💾 Converting with 110.6053 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,120B, BPFP=0.4923 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,508B, BPFP=2.5241 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,296B, BPFP=1.3341 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,672B, BPFP=2.4663 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,152B, BPFP=1.6007 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,200B, BPFP=2.4336 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,748B, BPFP=1.4345 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,972B, BPFP=2.4870 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,196B, BPFP=2.1568 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,068B, BPFP=2.4245 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,328B, BPFP=0.3687 +⌛️ [2/4] FRONTEND: Frontend time: 2.295s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.485s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510114 42.14560720 + layer.0.v_cache 0.00001717 0.00685087 + layer.1.k_cache 0.31797088 8.86347678 + layer.1.v_cache 0.00000673 0.00267868 + layer.2.k_cache 0.01711024 0.86051245 + layer.2.v_cache 0.00002022 0.00784190 + layer.3.k_cache 0.01307247 6.10745752 + layer.3.v_cache 0.00002089 0.00868081 + layer.4.k_cache 0.00069413 0.21494212 + layer.4.v_cache 0.00005461 0.01635331 + layer.4.output 1.35466395 231.63013590 + ------------------------------------------------------------------------------------- + TOTAL 0.58392448 98.80266782 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 317260 +BPFP 1.2903 bits/point +EBPFP 2.5805 equivalent bits/point +MSE 98.802668 +---------------------- -------------------------------------------------------- +Time: 3.789s Load: 0.009s, Pack+Encode: 2.295s, Decode+Unpack: 1.485s +---------------------- -------------------------------------------------------- +💾 Converting with 98.8027 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,764B, BPFP=0.4892 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,068B, BPFP=2.3354 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,008B, BPFP=1.2606 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,340B, BPFP=2.2896 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,452B, BPFP=1.5406 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,920B, BPFP=2.2631 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,164B, BPFP=1.4594 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,496B, BPFP=2.2994 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,644B, BPFP=1.9937 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,348B, BPFP=2.2271 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,032B, BPFP=0.3783 +⌛️ [2/4] FRONTEND: Frontend time: 2.551s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 26.528s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14860459 41.67078030 + layer.0.v_cache 0.00001629 0.00656971 + layer.1.k_cache 0.38266554 8.65433133 + layer.1.v_cache 0.00000669 0.00263971 + layer.2.k_cache 0.01628031 0.84892150 + layer.2.v_cache 0.00002106 0.00746366 + layer.3.k_cache 0.01842327 6.22181702 + layer.3.v_cache 0.00002289 0.00903789 + layer.4.k_cache 0.00072690 0.21378231 + layer.4.v_cache 0.00004923 0.01500594 + layer.4.output 1.23454798 209.29836910 + ------------------------------------------------------------------------------------- + TOTAL 0.54168545 89.57287841 + (elements=2,158,592) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2158592 +Total Bytes 330236 +BPFP 1.2239 bits/point +EBPFP 2.4478 equivalent bits/point +MSE 89.572878 +---------------------- --------------------------------------------------------- +Time: 29.090s Load: 0.010s, Pack+Encode: 2.551s, Decode+Unpack: 26.528s +---------------------- --------------------------------------------------------- +💾 Converting with 89.5729 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.018s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 435, 128) +Output shape: (1, 435, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.output: torch.Size([1, 435, 3584]) -> torch.Size([1, 1, 435, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,236B, BPFP=0.4754 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 64,476B, BPFP=2.3159 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 36,384B, BPFP=1.3069 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 62,756B, BPFP=2.2542 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 41,380B, BPFP=1.4864 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 62,256B, BPFP=2.2362 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 39,616B, BPFP=1.4230 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 63,164B, BPFP=2.2688 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 54,416B, BPFP=1.9546 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 61,340B, BPFP=2.2033 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,144B, BPFP=0.3291 +⌛️ [2/4] FRONTEND: Frontend time: 2.590s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.195s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15864419 40.49625539 + layer.0.v_cache 0.00001607 0.00640263 + layer.1.k_cache 0.94329666 9.08826554 + layer.1.v_cache 0.00000639 0.00251185 + layer.2.k_cache 0.02505282 0.77779527 + layer.2.v_cache 0.00002131 0.00703429 + layer.3.k_cache 0.01724711 6.30999686 + layer.3.v_cache 0.00002032 0.00778238 + layer.4.k_cache 0.00074513 0.23592889 + layer.4.v_cache 0.00005173 0.01399608 + layer.4.output 0.00608059 125.86654351 + ------------------------------------------------------------------------------------- + TOTAL 0.06986270 55.17716316 + (elements=3,786,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3786240 +Total Bytes 563168 +BPFP 1.1899 bits/point +EBPFP 2.3799 equivalent bits/point +MSE 55.177163 +---------------------- -------------------------------------------------------- +Time: 4.803s Load: 0.018s, Pack+Encode: 2.590s, Decode+Unpack: 2.195s +---------------------- -------------------------------------------------------- +💾 Converting with 55.1772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 437, 128) +Output shape: (1, 437, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.output: torch.Size([1, 437, 3584]) -> torch.Size([1, 1, 437, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,288B, BPFP=0.4751 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 64,316B, BPFP=2.2996 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 36,548B, BPFP=1.3068 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 62,896B, BPFP=2.2489 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 41,952B, BPFP=1.5000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 62,424B, BPFP=2.2320 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 39,764B, BPFP=1.4218 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 63,436B, BPFP=2.2682 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 54,908B, BPFP=1.9632 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 61,708B, BPFP=2.2064 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,220B, BPFP=0.3689 +⌛️ [2/4] FRONTEND: Frontend time: 2.761s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.036s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14454122 40.01303722 + layer.0.v_cache 0.00001702 0.00642888 + layer.1.k_cache 0.94267751 9.19223588 + layer.1.v_cache 0.00000668 0.00260005 + layer.2.k_cache 0.01808156 0.78973354 + layer.2.v_cache 0.00002154 0.00728774 + layer.3.k_cache 0.03034366 5.96978466 + layer.3.v_cache 0.00002210 0.00839103 + layer.4.k_cache 0.00074160 0.24040723 + layer.4.v_cache 0.00005552 0.01504181 + layer.4.output 0.00611241 117.13763689 + ------------------------------------------------------------------------------------- + TOTAL 0.06937031 51.54167096 + (elements=3,803,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3803648 +Total Bytes 573460 +BPFP 1.2061 bits/point +EBPFP 2.4123 equivalent bits/point +MSE 51.541671 +---------------------- -------------------------------------------------------- +Time: 4.812s Load: 0.015s, Pack+Encode: 2.761s, Decode+Unpack: 2.036s +---------------------- -------------------------------------------------------- +💾 Converting with 51.5417 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,600B, BPFP=0.4870 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,264B, BPFP=2.3470 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,644B, BPFP=1.2502 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,988B, BPFP=2.2823 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,844B, BPFP=1.5140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,584B, BPFP=2.2618 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,036B, BPFP=1.4223 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,588B, BPFP=2.3127 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,168B, BPFP=1.9870 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,232B, BPFP=2.2439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 51,724B, BPFP=0.3749 +⌛️ [2/4] FRONTEND: Frontend time: 2.554s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.707s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12827574 39.57867352 + layer.0.v_cache 0.00001866 0.00676728 + layer.1.k_cache 0.53653341 8.61683556 + layer.1.v_cache 0.00000626 0.00253620 + layer.2.k_cache 0.01950284 0.94156617 + layer.2.v_cache 0.00002083 0.00710230 + layer.3.k_cache 0.04340698 5.99200836 + layer.3.v_cache 0.00002181 0.00844441 + layer.4.k_cache 0.00072091 0.20473483 + layer.4.v_cache 0.00005393 0.01478731 + layer.4.output 0.04342448 174.27720605 + ------------------------------------------------------------------------------------- + TOTAL 0.06073722 75.01846461 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 408672 +BPFP 1.2195 bits/point +EBPFP 2.4391 equivalent bits/point +MSE 75.018465 +---------------------- -------------------------------------------------------- +Time: 4.271s Load: 0.011s, Pack+Encode: 2.554s, Decode+Unpack: 1.707s +---------------------- -------------------------------------------------------- +💾 Converting with 75.0185 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,472B, BPFP=0.4977 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,708B, BPFP=2.6262 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,544B, BPFP=1.2655 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,580B, BPFP=2.5599 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,936B, BPFP=1.5235 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,420B, BPFP=2.5505 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,248B, BPFP=1.4243 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,024B, BPFP=2.5860 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,628B, BPFP=2.2690 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,912B, BPFP=2.5207 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 45,108B, BPFP=0.3785 +⌛️ [2/4] FRONTEND: Frontend time: 2.518s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.951s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702504 41.02685180 + layer.0.v_cache 0.00001641 0.00708043 + layer.1.k_cache 0.51130361 8.49859390 + layer.1.v_cache 0.00000656 0.00281841 + layer.2.k_cache 0.01810413 0.74822493 + layer.2.v_cache 0.00002122 0.00799972 + layer.3.k_cache 0.02476560 5.83378658 + layer.3.v_cache 0.00002197 0.00909881 + layer.4.k_cache 0.00071356 0.24525922 + layer.4.v_cache 0.00005290 0.01614395 + layer.4.output 0.00501064 197.91472543 + ------------------------------------------------------------------------------------- + TOTAL 0.04100620 84.81170210 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 382580 +BPFP 1.3219 bits/point +EBPFP 2.6439 equivalent bits/point +MSE 84.811702 +---------------------- -------------------------------------------------------- +Time: 4.482s Load: 0.013s, Pack+Encode: 2.518s, Decode+Unpack: 1.951s +---------------------- -------------------------------------------------------- +💾 Converting with 84.8117 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,496B, BPFP=0.5005 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,940B, BPFP=2.4666 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,504B, BPFP=1.3024 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,984B, BPFP=2.4028 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,240B, BPFP=1.6186 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,736B, BPFP=2.3862 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,508B, BPFP=1.5029 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,496B, BPFP=2.4370 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,768B, BPFP=2.1213 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,308B, BPFP=2.3576 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,456B, BPFP=0.3859 +⌛️ [2/4] FRONTEND: Frontend time: 2.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.659s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11892304 41.44487847 + layer.0.v_cache 0.00001803 0.00724659 + layer.1.k_cache 0.31174104 9.08630058 + layer.1.v_cache 0.00000657 0.00275027 + layer.2.k_cache 0.01452858 1.04552016 + layer.2.v_cache 0.00002161 0.00776582 + layer.3.k_cache 0.02263702 6.37142501 + layer.3.v_cache 0.00002141 0.00905559 + layer.4.k_cache 0.00068312 0.22139225 + layer.4.v_cache 0.00005251 0.01579748 + layer.4.output 1.30838121 222.60721917 + ------------------------------------------------------------------------------------- + TOTAL 0.56631185 95.08603920 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 326436 +BPFP 1.2822 bits/point +EBPFP 2.5644 equivalent bits/point +MSE 95.086039 +---------------------- -------------------------------------------------------- +Time: 3.806s Load: 0.015s, Pack+Encode: 2.132s, Decode+Unpack: 1.659s +---------------------- -------------------------------------------------------- +💾 Converting with 95.0860 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 233, 128) +Output shape: (1, 233, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.output: torch.Size([1, 233, 3584]) -> torch.Size([1, 1, 233, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,520B, BPFP=0.5043 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,012B, BPFP=2.4820 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,172B, BPFP=1.2857 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,320B, BPFP=2.4356 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,152B, BPFP=1.6196 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,748B, BPFP=2.3973 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,660B, BPFP=1.4525 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,300B, BPFP=2.4343 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,840B, BPFP=2.1352 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,148B, BPFP=2.3570 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,352B, BPFP=0.3578 +⌛️ [2/4] FRONTEND: Frontend time: 2.170s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.722s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13365312 42.04923049 + layer.0.v_cache 0.00001735 0.00701942 + layer.1.k_cache 0.40872042 9.18407994 + layer.1.v_cache 0.00000792 0.00289600 + layer.2.k_cache 0.01308113 0.86446914 + layer.2.v_cache 0.00002137 0.00822543 + layer.3.k_cache 0.00601081 6.04936041 + layer.3.v_cache 0.00002094 0.00898020 + layer.4.k_cache 0.00068546 0.22941660 + layer.4.v_cache 0.00005338 0.01646664 + layer.4.output 1.31399239 228.24360055 + ------------------------------------------------------------------------------------- + TOTAL 0.57413051 97.41913813 + (elements=2,028,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2028032 +Total Bytes 322224 +BPFP 1.2711 bits/point +EBPFP 2.5422 equivalent bits/point +MSE 97.419138 +---------------------- -------------------------------------------------------- +Time: 3.900s Load: 0.008s, Pack+Encode: 2.170s, Decode+Unpack: 1.722s +---------------------- -------------------------------------------------------- +💾 Converting with 97.4191 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 148, 128) +Output shape: (1, 148, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.output: torch.Size([1, 148, 3584]) -> torch.Size([1, 1, 148, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,016B, BPFP=0.5296 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,216B, BPFP=2.8733 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,056B, BPFP=1.2728 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,524B, BPFP=2.8003 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,568B, BPFP=1.5380 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,392B, BPFP=2.7863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,964B, BPFP=1.4742 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,756B, BPFP=2.8247 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,744B, BPFP=2.5068 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,048B, BPFP=2.7500 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,884B, BPFP=0.3602 +⌛️ [2/4] FRONTEND: Frontend time: 2.008s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.483s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08608804 42.66316512 + layer.0.v_cache 0.00001767 0.00746957 + layer.1.k_cache 0.12276106 9.02849847 + layer.1.v_cache 0.00000594 0.00286755 + layer.2.k_cache 0.00886028 0.91552559 + layer.2.v_cache 0.00002088 0.00846154 + layer.3.k_cache 0.01513371 5.23503773 + layer.3.v_cache 0.00002085 0.01002306 + layer.4.k_cache 0.00068430 0.23072248 + layer.4.v_cache 0.00005331 0.01676191 + layer.4.output 0.08957551 362.93385014 + ------------------------------------------------------------------------------------- + TOTAL 0.05062792 152.86208730 + (elements=1,288,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1288192 +Total Bytes 226168 +BPFP 1.4046 bits/point +EBPFP 2.8091 equivalent bits/point +MSE 152.862087 +---------------------- -------------------------------------------------------- +Time: 3.496s Load: 0.006s, Pack+Encode: 2.008s, Decode+Unpack: 1.483s +---------------------- -------------------------------------------------------- +💾 Converting with 152.8621 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 170, 128) +Output shape: (1, 170, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.output: torch.Size([1, 170, 3584]) -> torch.Size([1, 1, 170, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,700B, BPFP=0.5239 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,748B, BPFP=2.5504 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,780B, BPFP=1.2665 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,984B, BPFP=2.4801 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,152B, BPFP=1.6684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,508B, BPFP=2.4364 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,216B, BPFP=1.4904 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,196B, BPFP=2.4996 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,036B, BPFP=2.2092 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,232B, BPFP=2.4110 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,832B, BPFP=0.3786 +⌛️ [2/4] FRONTEND: Frontend time: 2.195s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.399s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13519069 41.02545381 + layer.0.v_cache 0.00001779 0.00781283 + layer.1.k_cache 0.14803200 9.41913775 + layer.1.v_cache 0.00000678 0.00305557 + layer.2.k_cache 0.02267574 0.96087799 + layer.2.v_cache 0.00002005 0.00838756 + layer.3.k_cache 0.04618364 5.62776094 + layer.3.v_cache 0.00002282 0.01008976 + layer.4.k_cache 0.00077641 0.23559950 + layer.4.v_cache 0.00005264 0.01776087 + layer.4.output 0.00945594 315.02481618 + ------------------------------------------------------------------------------------- + TOTAL 0.02465707 133.08762646 + (elements=1,479,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1479680 +Total Bytes 241384 +BPFP 1.3051 bits/point +EBPFP 2.6101 equivalent bits/point +MSE 133.087626 +---------------------- -------------------------------------------------------- +Time: 3.601s Load: 0.007s, Pack+Encode: 2.195s, Decode+Unpack: 1.399s +---------------------- -------------------------------------------------------- +💾 Converting with 133.0876 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,620B, BPFP=0.5290 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,836B, BPFP=2.6201 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,900B, BPFP=1.3084 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,016B, BPFP=2.5429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,732B, BPFP=1.6691 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,688B, BPFP=2.5120 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,436B, BPFP=1.4529 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,148B, BPFP=2.5553 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,188B, BPFP=2.2767 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,256B, BPFP=2.4714 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,604B, BPFP=0.3846 +⌛️ [2/4] FRONTEND: Frontend time: 2.422s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12520085 41.68554688 + layer.0.v_cache 0.00001716 0.00783128 + layer.1.k_cache 0.15152623 9.24230075 + layer.1.v_cache 0.00000633 0.00294971 + layer.2.k_cache 0.01605446 0.67815307 + layer.2.v_cache 0.00002200 0.00899702 + layer.3.k_cache 0.04332644 5.82826766 + layer.3.v_cache 0.00002349 0.01041255 + layer.4.k_cache 0.00069336 0.23839650 + layer.4.v_cache 0.00005614 0.01756265 + layer.4.output 0.00966471 311.03649419 + ------------------------------------------------------------------------------------- + TOTAL 0.02379879 131.46916926 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 240424 +BPFP 1.3312 bits/point +EBPFP 2.6624 equivalent bits/point +MSE 131.469169 +---------------------- -------------------------------------------------------- +Time: 3.900s Load: 0.008s, Pack+Encode: 2.422s, Decode+Unpack: 1.470s +---------------------- -------------------------------------------------------- +💾 Converting with 131.4692 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 181, 128) +Output shape: (1, 181, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.output: torch.Size([1, 181, 3584]) -> torch.Size([1, 1, 181, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,816B, BPFP=0.5021 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,832B, BPFP=2.4026 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,324B, BPFP=1.3229 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,160B, BPFP=2.3446 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,124B, BPFP=1.5646 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,832B, BPFP=2.3163 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,636B, BPFP=1.4361 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,484B, BPFP=2.3726 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,360B, BPFP=2.1029 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,556B, BPFP=2.2925 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,908B, BPFP=0.3565 +⌛️ [2/4] FRONTEND: Frontend time: 2.147s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.441s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13566811 40.84006766 + layer.0.v_cache 0.00002016 0.00757660 + layer.1.k_cache 0.14782082 9.16937829 + layer.1.v_cache 0.00000634 0.00302160 + layer.2.k_cache 0.01546242 0.86593451 + layer.2.v_cache 0.00002121 0.00838046 + layer.3.k_cache 0.02895848 5.69336747 + layer.3.v_cache 0.00002197 0.00974395 + layer.4.k_cache 0.00067433 0.26277298 + layer.4.v_cache 0.00005229 0.01715588 + layer.4.output 0.00892717 284.21929262 + ------------------------------------------------------------------------------------- + TOTAL 0.02301155 120.37720281 + (elements=1,575,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1575424 +Total Bytes 245032 +BPFP 1.2443 bits/point +EBPFP 2.4885 equivalent bits/point +MSE 120.377203 +---------------------- -------------------------------------------------------- +Time: 3.597s Load: 0.010s, Pack+Encode: 2.147s, Decode+Unpack: 1.441s +---------------------- -------------------------------------------------------- +💾 Converting with 120.3772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 186, 128) +Output shape: (1, 186, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.output: torch.Size([1, 186, 3584]) -> torch.Size([1, 1, 186, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,848B, BPFP=0.4913 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,920B, BPFP=2.3454 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,488B, BPFP=1.3011 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,056B, BPFP=2.2728 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,368B, BPFP=1.5430 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,696B, BPFP=2.2426 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,720B, BPFP=1.4046 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,320B, BPFP=2.2950 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,920B, BPFP=2.0094 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,540B, BPFP=2.2295 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,676B, BPFP=0.3801 +⌛️ [2/4] FRONTEND: Frontend time: 1.962s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.415s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09795722 41.65394248 + layer.0.v_cache 0.00001815 0.00767109 + layer.1.k_cache 0.13430934 9.15485604 + layer.1.v_cache 0.00000632 0.00290809 + layer.2.k_cache 0.01830538 0.84419062 + layer.2.v_cache 0.00002120 0.00828310 + layer.3.k_cache 0.04892764 6.11444551 + layer.3.v_cache 0.00002113 0.00980609 + layer.4.k_cache 0.00067235 0.24089719 + layer.4.v_cache 0.00008605 0.01702528 + layer.4.output 0.00869856 286.06060388 + ------------------------------------------------------------------------------------- + TOTAL 0.02124792 121.20460310 + (elements=1,618,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1618944 +Total Bytes 247552 +BPFP 1.2233 bits/point +EBPFP 2.4466 equivalent bits/point +MSE 121.204603 +---------------------- -------------------------------------------------------- +Time: 3.387s Load: 0.010s, Pack+Encode: 1.962s, Decode+Unpack: 1.415s +---------------------- -------------------------------------------------------- +💾 Converting with 121.2046 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 163, 128) +Output shape: (1, 163, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.output: torch.Size([1, 163, 3584]) -> torch.Size([1, 1, 163, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,332B, BPFP=0.5111 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,496B, BPFP=2.6357 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,200B, BPFP=1.3612 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,792B, BPFP=2.5683 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,160B, BPFP=1.6449 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,464B, BPFP=2.5368 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,644B, BPFP=1.4996 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,084B, BPFP=2.5962 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,908B, BPFP=2.2918 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,232B, BPFP=2.5146 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,064B, BPFP=0.3706 +⌛️ [2/4] FRONTEND: Frontend time: 2.122s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.600s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11530171 41.50742906 + layer.0.v_cache 0.00001813 0.00782658 + layer.1.k_cache 0.14319220 9.34459625 + layer.1.v_cache 0.00000612 0.00292086 + layer.2.k_cache 0.02368733 0.90761131 + layer.2.v_cache 0.00002106 0.00831335 + layer.3.k_cache 0.02525310 5.53994395 + layer.3.v_cache 0.00002161 0.00991165 + layer.4.k_cache 0.00069984 0.21661428 + layer.4.v_cache 0.00005709 0.01699214 + layer.4.output 0.00982084 331.60949825 + ------------------------------------------------------------------------------------- + TOTAL 0.02217671 139.93109689 + (elements=1,418,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1418752 +Total Bytes 237376 +BPFP 1.3385 bits/point +EBPFP 2.6770 equivalent bits/point +MSE 139.931097 +---------------------- -------------------------------------------------------- +Time: 3.729s Load: 0.007s, Pack+Encode: 2.122s, Decode+Unpack: 1.600s +---------------------- -------------------------------------------------------- +💾 Converting with 139.9311 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 191, 128) +Output shape: (1, 191, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.output: torch.Size([1, 191, 3584]) -> torch.Size([1, 1, 191, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,924B, BPFP=0.4846 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,048B, BPFP=2.2945 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,004B, BPFP=1.3092 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,248B, BPFP=2.2291 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,944B, BPFP=1.4679 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,924B, BPFP=2.2026 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,052B, BPFP=1.3950 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,584B, BPFP=2.2565 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,440B, BPFP=1.9993 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,916B, BPFP=2.2019 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,152B, BPFP=0.3524 +⌛️ [2/4] FRONTEND: Frontend time: 2.003s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.448s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15112510 42.14326325 + layer.0.v_cache 0.00002069 0.00719760 + layer.1.k_cache 0.18491925 10.21660898 + layer.1.v_cache 0.00000629 0.00286013 + layer.2.k_cache 0.02122488 0.88569729 + layer.2.v_cache 0.00002327 0.00867096 + layer.3.k_cache 0.01259250 5.71804059 + layer.3.v_cache 0.00002182 0.00980700 + layer.4.k_cache 0.00068006 0.25484930 + layer.4.v_cache 0.00005885 0.01644992 + layer.4.output 0.00845948 268.00703534 + ------------------------------------------------------------------------------------- + TOTAL 0.02528759 113.84192308 + (elements=1,662,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1662464 +Total Bytes 248236 +BPFP 1.1945 bits/point +EBPFP 2.3891 equivalent bits/point +MSE 113.841923 +---------------------- -------------------------------------------------------- +Time: 3.458s Load: 0.008s, Pack+Encode: 2.003s, Decode+Unpack: 1.448s +---------------------- -------------------------------------------------------- +💾 Converting with 113.8419 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 174, 128) +Output shape: (1, 174, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.output: torch.Size([1, 174, 3584]) -> torch.Size([1, 1, 174, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,764B, BPFP=0.5176 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,856B, BPFP=2.5014 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,324B, BPFP=1.2863 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,092B, BPFP=2.4328 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,276B, BPFP=1.6412 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,908B, BPFP=2.4163 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,820B, BPFP=1.5104 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,392B, BPFP=2.4598 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,476B, BPFP=2.1979 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,592B, BPFP=2.3879 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,616B, BPFP=0.3671 +⌛️ [2/4] FRONTEND: Frontend time: 1.975s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.695s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14146894 40.58722275 + layer.0.v_cache 0.00001664 0.00757365 + layer.1.k_cache 0.16292955 9.00375051 + layer.1.v_cache 0.00000602 0.00285038 + layer.2.k_cache 0.01834387 0.82011545 + layer.2.v_cache 0.00002001 0.00829766 + layer.3.k_cache 0.03015787 5.89658470 + layer.3.v_cache 0.00002180 0.01005521 + layer.4.k_cache 0.00067479 0.23625714 + layer.4.v_cache 0.00005130 0.01751176 + layer.4.output 0.00923280 310.91345956 + ------------------------------------------------------------------------------------- + TOTAL 0.02460708 131.35202566 + (elements=1,514,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1514496 +Total Bytes 244116 +BPFP 1.2895 bits/point +EBPFP 2.5790 equivalent bits/point +MSE 131.352026 +---------------------- -------------------------------------------------------- +Time: 3.677s Load: 0.007s, Pack+Encode: 1.975s, Decode+Unpack: 1.695s +---------------------- -------------------------------------------------------- +💾 Converting with 131.3520 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.2816 bits/point +Avg EBPFP 2.5632 equivalent bits/point +Avg MSE 100.299716 +Avg Time 5.341s +------------------------ ---------------------------- diff --git a/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..80dc8d9082313f6a4478d581a2cf3ac3fb8595df --- /dev/null +++ b/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 506 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa +Output output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa +---------------- ------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,848B, BPFP=0.4800 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,104B, BPFP=2.4470 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,236B, BPFP=1.3149 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,100B, BPFP=2.3926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,228B, BPFP=1.4772 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,616B, BPFP=2.3663 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,780B, BPFP=1.3444 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,492B, BPFP=2.4138 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,500B, BPFP=2.0888 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,616B, BPFP=2.3663 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,140B, BPFP=0.2724 +⌛️ [2/4] FRONTEND: Frontend time: 3.239s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14134069 40.54581706 + layer.0.v_cache 0.00001405 0.00529245 + layer.1.k_cache 0.62039534 8.68234677 + layer.1.v_cache 0.00000582 0.00219784 + layer.2.k_cache 0.00676800 0.99780549 + layer.2.v_cache 0.00001893 0.00677987 + layer.3.k_cache 0.03896382 6.57959747 + layer.3.v_cache 0.00001942 0.00728648 + layer.4.k_cache 0.00069280 0.19882348 + layer.4.v_cache 0.00005238 0.01467677 + layer.4.output 0.00804578 183.66266741 + ------------------------------------------------------------------------------------- + TOTAL 0.05085834 78.98113503 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 379660 +BPFP 1.2116 bits/point +EBPFP 2.4233 equivalent bits/point +MSE 78.981135 +---------------------- -------------------------------------------------------- +Time: 5.266s Load: 0.014s, Pack+Encode: 3.239s, Decode+Unpack: 2.012s +---------------------- -------------------------------------------------------- +💾 Converting with 78.9811 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,212B, BPFP=0.4913 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,120B, BPFP=2.4061 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,576B, BPFP=1.3639 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,004B, BPFP=2.3466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,100B, BPFP=1.4452 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,608B, BPFP=2.3255 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,300B, BPFP=1.4025 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,700B, BPFP=2.3837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,680B, BPFP=2.0627 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,772B, BPFP=2.3343 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,388B, BPFP=0.2924 +⌛️ [2/4] FRONTEND: Frontend time: 3.000s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.163s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15682776 40.39125160 + layer.0.v_cache 0.00001501 0.00562325 + layer.1.k_cache 0.54447479 8.37613738 + layer.1.v_cache 0.00000580 0.00234482 + layer.2.k_cache 0.00853899 0.96428152 + layer.2.v_cache 0.00001972 0.00683603 + layer.3.k_cache 0.02401569 5.57636602 + layer.3.v_cache 0.00001939 0.00758123 + layer.4.k_cache 0.00071059 0.20539790 + layer.4.v_cache 0.00005329 0.01456976 + layer.4.output 0.05121508 170.32610617 + ------------------------------------------------------------------------------------- + TOTAL 0.06430510 73.40194898 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 386460 +BPFP 1.2123 bits/point +EBPFP 2.4246 equivalent bits/point +MSE 73.401949 +---------------------- -------------------------------------------------------- +Time: 5.174s Load: 0.011s, Pack+Encode: 3.000s, Decode+Unpack: 2.163s +---------------------- -------------------------------------------------------- +💾 Converting with 73.4019 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,264B, BPFP=0.4923 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,912B, BPFP=2.3869 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,440B, BPFP=1.2989 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,748B, BPFP=2.3250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,968B, BPFP=1.5395 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,384B, BPFP=2.3057 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,256B, BPFP=1.3954 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,412B, BPFP=2.3603 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,444B, BPFP=2.0432 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,604B, BPFP=2.3174 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,660B, BPFP=0.3011 +⌛️ [2/4] FRONTEND: Frontend time: 2.508s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.652s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15363285 38.94798642 + layer.0.v_cache 0.00001412 0.00545765 + layer.1.k_cache 0.60852461 8.19189204 + layer.1.v_cache 0.00000580 0.00224771 + layer.2.k_cache 0.01182756 0.99208671 + layer.2.v_cache 0.00001863 0.00657664 + layer.3.k_cache 0.03467803 5.53786588 + layer.3.v_cache 0.00001919 0.00743260 + layer.4.k_cache 0.00072319 0.18734592 + layer.4.v_cache 0.00005301 0.01413386 + layer.4.output 0.05035121 173.36078717 + ------------------------------------------------------------------------------------- + TOTAL 0.06835032 74.55403151 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 387092 +BPFP 1.2101 bits/point +EBPFP 2.4203 equivalent bits/point +MSE 74.554032 +---------------------- -------------------------------------------------------- +Time: 4.174s Load: 0.014s, Pack+Encode: 2.508s, Decode+Unpack: 1.652s +---------------------- -------------------------------------------------------- +💾 Converting with 74.5540 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,860B, BPFP=0.4892 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,924B, BPFP=2.4803 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,460B, BPFP=1.2401 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,584B, BPFP=2.4064 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,260B, BPFP=1.5051 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,576B, BPFP=2.4059 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,288B, BPFP=1.3962 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,268B, BPFP=2.4441 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,104B, BPFP=2.1038 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,592B, BPFP=2.4068 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,900B, BPFP=0.2910 +⌛️ [2/4] FRONTEND: Frontend time: 2.194s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.590s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12197856 39.16005618 + layer.0.v_cache 0.00001365 0.00535298 + layer.1.k_cache 0.62524759 8.46067524 + layer.1.v_cache 0.00000551 0.00218540 + layer.2.k_cache 0.01486346 1.08788038 + layer.2.v_cache 0.00001981 0.00704914 + layer.3.k_cache 0.07387851 5.60090384 + layer.3.v_cache 0.00001943 0.00767492 + layer.4.k_cache 0.00069560 0.20168019 + layer.4.v_cache 0.00005405 0.01466456 + layer.4.output 0.01098318 190.76470217 + ------------------------------------------------------------------------------------- + TOTAL 0.05374461 81.75888459 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 378816 +BPFP 1.2303 bits/point +EBPFP 2.4606 equivalent bits/point +MSE 81.758885 +---------------------- -------------------------------------------------------- +Time: 3.793s Load: 0.009s, Pack+Encode: 2.194s, Decode+Unpack: 1.590s +---------------------- -------------------------------------------------------- +💾 Converting with 81.7589 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,848B, BPFP=0.4851 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,100B, BPFP=2.4726 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,136B, BPFP=1.3232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,884B, BPFP=2.4059 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,024B, BPFP=1.4268 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,524B, BPFP=2.3862 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,184B, BPFP=1.3807 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,448B, BPFP=2.4368 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,996B, BPFP=2.0831 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,648B, BPFP=2.3930 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,108B, BPFP=0.2985 +⌛️ [2/4] FRONTEND: Frontend time: 2.275s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.942s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15362121 41.15607182 + layer.0.v_cache 0.00001420 0.00542977 + layer.1.k_cache 0.52952420 8.22345977 + layer.1.v_cache 0.00000567 0.00220993 + layer.2.k_cache 0.00939641 0.88888828 + layer.2.v_cache 0.00001891 0.00653564 + layer.3.k_cache 0.04339437 5.73327808 + layer.3.v_cache 0.00002732 0.00735856 + layer.4.k_cache 0.00068326 0.19594122 + layer.4.v_cache 0.00005156 0.01396992 + layer.4.output 0.00776138 180.41499060 + ------------------------------------------------------------------------------------- + TOTAL 0.04653334 77.59635748 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 380900 +BPFP 1.2284 bits/point +EBPFP 2.4568 equivalent bits/point +MSE 77.596357 +---------------------- -------------------------------------------------------- +Time: 4.227s Load: 0.010s, Pack+Encode: 2.275s, Decode+Unpack: 1.942s +---------------------- -------------------------------------------------------- +💾 Converting with 77.5964 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,348B, BPFP=0.4968 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,924B, BPFP=2.3875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,356B, BPFP=1.3476 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,988B, BPFP=2.3378 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,612B, BPFP=1.5206 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,488B, BPFP=2.3112 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,160B, BPFP=1.3903 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,328B, BPFP=2.3559 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,748B, BPFP=2.0593 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,500B, BPFP=2.3119 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,256B, BPFP=0.2905 +⌛️ [2/4] FRONTEND: Frontend time: 2.315s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.640s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14921317 40.16785249 + layer.0.v_cache 0.00001395 0.00556390 + layer.1.k_cache 0.64083395 8.82446621 + layer.1.v_cache 0.00000574 0.00232214 + layer.2.k_cache 0.00578804 1.01639759 + layer.2.v_cache 0.00001912 0.00710677 + layer.3.k_cache 0.02848778 5.71951688 + layer.3.v_cache 0.00001956 0.00783809 + layer.4.k_cache 0.00070939 0.20986477 + layer.4.v_cache 0.00005206 0.01456412 + layer.4.output 0.04916091 171.15910471 + ------------------------------------------------------------------------------------- + TOTAL 0.06878054 73.76995447 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 386708 +BPFP 1.2089 bits/point +EBPFP 2.4179 equivalent bits/point +MSE 73.769954 +---------------------- -------------------------------------------------------- +Time: 3.966s Load: 0.011s, Pack+Encode: 2.315s, Decode+Unpack: 1.640s +---------------------- -------------------------------------------------------- +💾 Converting with 73.7700 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,200B, BPFP=0.4906 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,112B, BPFP=2.4057 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,004B, BPFP=1.3334 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,932B, BPFP=2.3428 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,452B, BPFP=1.4640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,504B, BPFP=2.3200 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,104B, BPFP=1.3387 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,512B, BPFP=2.3737 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,216B, BPFP=2.0380 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,416B, BPFP=2.3153 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,016B, BPFP=0.2972 +⌛️ [2/4] FRONTEND: Frontend time: 2.295s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.784s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12806672 39.41801208 + layer.0.v_cache 0.00001388 0.00539756 + layer.1.k_cache 0.56701790 8.52634636 + layer.1.v_cache 0.00000557 0.00218674 + layer.2.k_cache 0.01020011 0.96999726 + layer.2.v_cache 0.00001881 0.00655731 + layer.3.k_cache 0.04301871 5.61943018 + layer.3.v_cache 0.00002051 0.00741212 + layer.4.k_cache 0.00071755 0.19102533 + layer.4.v_cache 0.00005284 0.01394962 + layer.4.output 0.04940383 170.86582765 + ------------------------------------------------------------------------------------- + TOTAL 0.06440938 73.57771224 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 384468 +BPFP 1.2060 bits/point +EBPFP 2.4121 equivalent bits/point +MSE 73.577712 +---------------------- -------------------------------------------------------- +Time: 4.090s Load: 0.012s, Pack+Encode: 2.295s, Decode+Unpack: 1.784s +---------------------- -------------------------------------------------------- +💾 Converting with 73.5777 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,876B, BPFP=0.4866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,148B, BPFP=2.4752 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,164B, BPFP=1.2700 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,168B, BPFP=2.4215 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,880B, BPFP=1.4737 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,720B, BPFP=2.3969 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,644B, BPFP=1.4059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,524B, BPFP=2.4410 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,312B, BPFP=2.1004 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,732B, BPFP=2.3976 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,520B, BPFP=0.2860 +⌛️ [2/4] FRONTEND: Frontend time: 2.629s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.021s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12903612 40.49424342 + layer.0.v_cache 0.00001458 0.00540728 + layer.1.k_cache 0.59860808 8.31529263 + layer.1.v_cache 0.00000581 0.00224300 + layer.2.k_cache 0.00940504 0.95995259 + layer.2.v_cache 0.00001954 0.00691794 + layer.3.k_cache 0.02644065 6.06792849 + layer.3.v_cache 0.00002029 0.00754030 + layer.4.k_cache 0.00068693 0.20441548 + layer.4.v_cache 0.00005376 0.01415950 + layer.4.output 0.00862506 176.71588346 + ------------------------------------------------------------------------------------- + TOTAL 0.04850978 76.06407558 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 380688 +BPFP 1.2277 bits/point +EBPFP 2.4554 equivalent bits/point +MSE 76.064076 +---------------------- -------------------------------------------------------- +Time: 4.660s Load: 0.010s, Pack+Encode: 2.629s, Decode+Unpack: 2.021s +---------------------- -------------------------------------------------------- +💾 Converting with 76.0641 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,556B, BPFP=0.4848 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,104B, BPFP=2.2881 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,992B, BPFP=1.3186 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,880B, BPFP=2.2261 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,908B, BPFP=1.4665 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,568B, BPFP=2.2102 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,980B, BPFP=1.3180 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,460B, BPFP=2.2555 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,456B, BPFP=1.9509 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,500B, BPFP=2.2068 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,208B, BPFP=0.2769 +⌛️ [2/4] FRONTEND: Frontend time: 2.215s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.633s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16469599 40.85395952 + layer.0.v_cache 0.00001357 0.00525296 + layer.1.k_cache 0.70445866 8.72348617 + layer.1.v_cache 0.00000558 0.00224758 + layer.2.k_cache 0.01447187 0.82260697 + layer.2.v_cache 0.00001924 0.00664554 + layer.3.k_cache 0.02617162 5.40190778 + layer.3.v_cache 0.00001974 0.00741397 + layer.4.k_cache 0.00070375 0.19502548 + layer.4.v_cache 0.00005076 0.01431149 + layer.4.output 0.04814868 167.88479824 + ------------------------------------------------------------------------------------- + TOTAL 0.07339127 72.42508501 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 387612 +BPFP 1.1567 bits/point +EBPFP 2.3134 equivalent bits/point +MSE 72.425085 +---------------------- -------------------------------------------------------- +Time: 3.858s Load: 0.010s, Pack+Encode: 2.215s, Decode+Unpack: 1.633s +---------------------- -------------------------------------------------------- +💾 Converting with 72.4251 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,632B, BPFP=0.5014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,688B, BPFP=2.5376 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,768B, BPFP=1.2644 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,920B, BPFP=2.4930 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,652B, BPFP=1.3738 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,624B, BPFP=2.4758 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,512B, BPFP=1.3076 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,820B, BPFP=2.4872 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,528B, BPFP=2.1798 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,988B, BPFP=2.4389 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,308B, BPFP=0.2764 +⌛️ [2/4] FRONTEND: Frontend time: 2.569s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.624s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12132761 40.83616984 + layer.0.v_cache 0.00001366 0.00539933 + layer.1.k_cache 0.57241889 8.61357168 + layer.1.v_cache 0.00000556 0.00217148 + layer.2.k_cache 0.00926234 0.91426756 + layer.2.v_cache 0.00002145 0.00673735 + layer.3.k_cache 0.03109839 4.82864686 + layer.3.v_cache 0.00002257 0.00734362 + layer.4.k_cache 0.00070701 0.19160595 + layer.4.v_cache 0.00004920 0.01377877 + layer.4.output 0.01031505 198.46305762 + ------------------------------------------------------------------------------------- + TOTAL 0.04747836 84.98006446 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 361440 +BPFP 1.2350 bits/point +EBPFP 2.4699 equivalent bits/point +MSE 84.980064 +---------------------- -------------------------------------------------------- +Time: 4.203s Load: 0.010s, Pack+Encode: 2.569s, Decode+Unpack: 1.624s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,404B, BPFP=0.4931 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,064B, BPFP=2.3628 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,928B, BPFP=1.3070 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,836B, BPFP=2.2984 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,424B, BPFP=1.4904 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,636B, BPFP=2.2880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,724B, BPFP=1.4012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,168B, BPFP=2.3159 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,700B, BPFP=2.0292 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,424B, BPFP=2.2768 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,260B, BPFP=0.2791 +⌛️ [2/4] FRONTEND: Frontend time: 2.272s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.589s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14357947 40.72071623 + layer.0.v_cache 0.00001420 0.00560176 + layer.1.k_cache 0.63723908 8.74471575 + layer.1.v_cache 0.00000573 0.00223346 + layer.2.k_cache 0.00765126 0.88415886 + layer.2.v_cache 0.00002073 0.00693260 + layer.3.k_cache 0.02843528 6.00418849 + layer.3.v_cache 0.00002112 0.00783424 + layer.4.k_cache 0.00069171 0.20791083 + layer.4.v_cache 0.00005234 0.01430345 + layer.4.output 0.04871205 169.70276546 + ------------------------------------------------------------------------------------- + TOTAL 0.06815855 73.20693846 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 385568 +BPFP 1.1892 bits/point +EBPFP 2.3784 equivalent bits/point +MSE 73.206938 +---------------------- -------------------------------------------------------- +Time: 3.872s Load: 0.011s, Pack+Encode: 2.272s, Decode+Unpack: 1.589s +---------------------- -------------------------------------------------------- +💾 Converting with 73.2069 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,480B, BPFP=0.4857 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,136B, BPFP=2.3123 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,124B, BPFP=1.2871 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,908B, BPFP=2.2494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,104B, BPFP=1.4398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,680B, BPFP=2.2377 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,092B, BPFP=1.2855 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,428B, BPFP=2.2760 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,784B, BPFP=1.9869 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,592B, BPFP=2.2332 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,688B, BPFP=0.2758 +⌛️ [2/4] FRONTEND: Frontend time: 2.142s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.685s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16857396 40.37749744 + layer.0.v_cache 0.00001478 0.00546745 + layer.1.k_cache 0.58716421 9.09461210 + layer.1.v_cache 0.00000600 0.00229149 + layer.2.k_cache 0.00827454 1.00475294 + layer.2.v_cache 0.00002177 0.00676920 + layer.3.k_cache 0.02531170 5.54398654 + layer.3.v_cache 0.00001983 0.00743598 + layer.4.k_cache 0.00069463 0.20345967 + layer.4.v_cache 0.00005316 0.01469039 + layer.4.output 0.04836898 161.75216628 + ------------------------------------------------------------------------------------- + TOTAL 0.06639514 69.91330160 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 385016 +BPFP 1.1602 bits/point +EBPFP 2.3205 equivalent bits/point +MSE 69.913302 +---------------------- -------------------------------------------------------- +Time: 3.838s Load: 0.011s, Pack+Encode: 2.142s, Decode+Unpack: 1.685s +---------------------- -------------------------------------------------------- +💾 Converting with 69.9133 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,384B, BPFP=0.4920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,120B, BPFP=2.3658 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,820B, BPFP=1.3014 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,056B, BPFP=2.3100 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,284B, BPFP=1.4306 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,892B, BPFP=2.3014 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,688B, BPFP=1.3993 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,696B, BPFP=2.3435 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,644B, BPFP=2.0262 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,576B, BPFP=2.2848 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,000B, BPFP=0.2921 +⌛️ [2/4] FRONTEND: Frontend time: 2.584s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.749s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14028978 40.18820784 + layer.0.v_cache 0.00001397 0.00544982 + layer.1.k_cache 0.59966084 8.89067498 + layer.1.v_cache 0.00000578 0.00225542 + layer.2.k_cache 0.01765286 0.85037595 + layer.2.v_cache 0.00001884 0.00681705 + layer.3.k_cache 0.02653811 5.20473649 + layer.3.v_cache 0.00001935 0.00749299 + layer.4.k_cache 0.00071527 0.20268512 + layer.4.v_cache 0.00005402 0.01438457 + layer.4.output 0.04907703 165.17572507 + ------------------------------------------------------------------------------------- + TOTAL 0.06638282 71.27077387 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 387160 +BPFP 1.1941 bits/point +EBPFP 2.3882 equivalent bits/point +MSE 71.270774 +---------------------- -------------------------------------------------------- +Time: 4.344s Load: 0.011s, Pack+Encode: 2.584s, Decode+Unpack: 1.749s +---------------------- -------------------------------------------------------- +💾 Converting with 71.2708 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,660B, BPFP=0.4975 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,832B, BPFP=2.5179 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,576B, BPFP=1.2969 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,368B, BPFP=2.4338 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,188B, BPFP=1.3895 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,060B, BPFP=2.4161 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,780B, BPFP=1.4235 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,632B, BPFP=2.4490 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,340B, BPFP=2.1450 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,444B, BPFP=2.4382 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,328B, BPFP=0.2981 +⌛️ [2/4] FRONTEND: Frontend time: 2.293s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.821s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13255837 41.41934384 + layer.0.v_cache 0.00001394 0.00516304 + layer.1.k_cache 0.54921044 8.13066550 + layer.1.v_cache 0.00000578 0.00207885 + layer.2.k_cache 0.01292971 0.91799276 + layer.2.v_cache 0.00001901 0.00653129 + layer.3.k_cache 0.07190662 4.58465038 + layer.3.v_cache 0.00001944 0.00694811 + layer.4.k_cache 0.00068159 0.17827366 + layer.4.v_cache 0.00005125 0.01370781 + layer.4.output 0.00790709 198.25620404 + ------------------------------------------------------------------------------------- + TOTAL 0.04839681 84.88581080 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 367208 +BPFP 1.2408 bits/point +EBPFP 2.4817 equivalent bits/point +MSE 84.885811 +---------------------- -------------------------------------------------------- +Time: 4.127s Load: 0.013s, Pack+Encode: 2.293s, Decode+Unpack: 1.821s +---------------------- -------------------------------------------------------- +💾 Converting with 84.8858 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,840B, BPFP=0.4898 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,048B, BPFP=2.4960 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,364B, BPFP=1.2945 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,884B, BPFP=2.4315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,632B, BPFP=1.4756 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,332B, BPFP=2.4009 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,380B, BPFP=1.3508 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,272B, BPFP=2.4530 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,164B, BPFP=2.1146 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,580B, BPFP=2.4147 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,904B, BPFP=0.3079 +⌛️ [2/4] FRONTEND: Frontend time: 2.242s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.590s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13695534 40.41697141 + layer.0.v_cache 0.00001452 0.00560591 + layer.1.k_cache 0.52272531 8.03799503 + layer.1.v_cache 0.00000574 0.00230480 + layer.2.k_cache 0.00793556 0.84960716 + layer.2.v_cache 0.00001907 0.00687034 + layer.3.k_cache 0.04682703 5.06162560 + layer.3.v_cache 0.00002008 0.00777976 + layer.4.k_cache 0.00068945 0.21228831 + layer.4.v_cache 0.00005221 0.01464916 + layer.4.output 0.00941004 182.66085676 + ------------------------------------------------------------------------------------- + TOTAL 0.04594792 78.42598205 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 380400 +BPFP 1.2398 bits/point +EBPFP 2.4797 equivalent bits/point +MSE 78.425982 +---------------------- -------------------------------------------------------- +Time: 3.843s Load: 0.010s, Pack+Encode: 2.242s, Decode+Unpack: 1.590s +---------------------- -------------------------------------------------------- +💾 Converting with 78.4260 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,832B, BPFP=0.4929 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,856B, BPFP=2.5031 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,732B, BPFP=1.2685 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,592B, BPFP=2.4326 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,272B, BPFP=1.4103 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,296B, BPFP=2.4161 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,252B, BPFP=1.2975 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,084B, BPFP=2.4600 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,716B, BPFP=2.1047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,488B, BPFP=2.4268 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,960B, BPFP=0.2867 +⌛️ [2/4] FRONTEND: Frontend time: 2.130s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.542s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14340984 39.86740025 + layer.0.v_cache 0.00001397 0.00531156 + layer.1.k_cache 0.51622396 8.38114362 + layer.1.v_cache 0.00000557 0.00220435 + layer.2.k_cache 0.00797994 0.92655313 + layer.2.v_cache 0.00001813 0.00635859 + layer.3.k_cache 0.02756977 5.48503592 + layer.3.v_cache 0.00001890 0.00703757 + layer.4.k_cache 0.00068956 0.18848111 + layer.4.v_cache 0.00005177 0.01415370 + layer.4.output 0.00947078 185.34209184 + ------------------------------------------------------------------------------------- + TOTAL 0.04483982 79.54578369 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 373080 +BPFP 1.2247 bits/point +EBPFP 2.4493 equivalent bits/point +MSE 79.545784 +---------------------- -------------------------------------------------------- +Time: 3.682s Load: 0.009s, Pack+Encode: 2.130s, Decode+Unpack: 1.542s +---------------------- -------------------------------------------------------- +💾 Converting with 79.5458 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,288B, BPFP=0.4936 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,924B, BPFP=2.3875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,664B, BPFP=1.3108 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,684B, BPFP=2.3216 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,340B, BPFP=1.5062 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,340B, BPFP=2.3034 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,392B, BPFP=1.4026 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,076B, BPFP=2.3425 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,676B, BPFP=2.0555 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,388B, BPFP=2.3059 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,144B, BPFP=0.2896 +⌛️ [2/4] FRONTEND: Frontend time: 2.146s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.806s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13622742 39.52854951 + layer.0.v_cache 0.00001380 0.00545122 + layer.1.k_cache 0.61520739 8.31162226 + layer.1.v_cache 0.00000584 0.00230907 + layer.2.k_cache 0.00900942 0.98964478 + layer.2.v_cache 0.00001916 0.00693051 + layer.3.k_cache 0.02667915 4.65662784 + layer.3.v_cache 0.00001881 0.00751670 + layer.4.k_cache 0.00071466 0.20538958 + layer.4.v_cache 0.00005760 0.01430093 + layer.4.output 0.04972813 171.03970785 + ------------------------------------------------------------------------------------- + TOTAL 0.06682648 73.58860573 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 384916 +BPFP 1.2033 bits/point +EBPFP 2.4067 equivalent bits/point +MSE 73.588606 +---------------------- -------------------------------------------------------- +Time: 3.962s Load: 0.010s, Pack+Encode: 2.146s, Decode+Unpack: 1.806s +---------------------- -------------------------------------------------------- +💾 Converting with 73.5886 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,436B, BPFP=0.4931 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,248B, BPFP=2.3645 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,132B, BPFP=1.2611 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,872B, BPFP=2.2926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,996B, BPFP=1.5153 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,624B, BPFP=2.2797 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,084B, BPFP=1.4676 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,612B, BPFP=2.3313 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,700B, BPFP=2.0224 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,612B, BPFP=2.2791 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,936B, BPFP=0.2832 +⌛️ [2/4] FRONTEND: Frontend time: 2.571s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11478914 40.13771491 + layer.0.v_cache 0.00001377 0.00522119 + layer.1.k_cache 0.64218915 8.34757705 + layer.1.v_cache 0.00000556 0.00211037 + layer.2.k_cache 0.00898698 0.89141968 + layer.2.v_cache 0.00001980 0.00647850 + layer.3.k_cache 0.05442990 5.90160561 + layer.3.v_cache 0.00001964 0.00728266 + layer.4.k_cache 0.00070461 0.19159209 + layer.4.v_cache 0.00005015 0.01413278 + layer.4.output 0.05017195 173.74750657 + ------------------------------------------------------------------------------------- + TOTAL 0.06896543 74.80809887 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 388252 +BPFP 1.1935 bits/point +EBPFP 2.3870 equivalent bits/point +MSE 74.808099 +---------------------- -------------------------------------------------------- +Time: 4.786s Load: 0.013s, Pack+Encode: 2.571s, Decode+Unpack: 2.202s +---------------------- -------------------------------------------------------- +💾 Converting with 74.8081 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,492B, BPFP=0.5007 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,448B, BPFP=2.5618 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,020B, BPFP=1.2394 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,544B, BPFP=2.5085 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,516B, BPFP=1.4455 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,812B, BPFP=2.5243 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,108B, BPFP=1.3035 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,512B, BPFP=2.5066 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,188B, BPFP=2.1927 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,828B, BPFP=2.4663 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,312B, BPFP=0.2722 +⌛️ [2/4] FRONTEND: Frontend time: 2.276s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.638s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14755792 40.64230542 + layer.0.v_cache 0.00001375 0.00549129 + layer.1.k_cache 0.51940866 8.03616690 + layer.1.v_cache 0.00000586 0.00235028 + layer.2.k_cache 0.00816260 0.99049095 + layer.2.v_cache 0.00001900 0.00698695 + layer.3.k_cache 0.02921419 5.33706515 + layer.3.v_cache 0.00001979 0.00774331 + layer.4.k_cache 0.00071017 0.21483143 + layer.4.v_cache 0.00005507 0.01472218 + layer.4.output 0.00937099 186.67345013 + ------------------------------------------------------------------------------------- + TOTAL 0.04533906 80.11601793 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 358780 +BPFP 1.2444 bits/point +EBPFP 2.4888 equivalent bits/point +MSE 80.116018 +---------------------- -------------------------------------------------------- +Time: 3.927s Load: 0.012s, Pack+Encode: 2.276s, Decode+Unpack: 1.638s +---------------------- -------------------------------------------------------- +💾 Converting with 80.1160 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,604B, BPFP=0.5035 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,476B, BPFP=2.5442 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,780B, BPFP=1.2746 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,676B, BPFP=2.4974 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,208B, BPFP=1.3581 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,228B, BPFP=2.4712 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,932B, BPFP=1.4590 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,580B, BPFP=2.4918 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,760B, BPFP=2.2097 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,372B, BPFP=2.4796 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,940B, BPFP=0.2754 +⌛️ [2/4] FRONTEND: Frontend time: 2.290s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.040s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12840328 40.77328388 + layer.0.v_cache 0.00001372 0.00537625 + layer.1.k_cache 0.60081110 8.35726094 + layer.1.v_cache 0.00000562 0.00222339 + layer.2.k_cache 0.00586274 0.83940142 + layer.2.v_cache 0.00001839 0.00682110 + layer.3.k_cache 0.02491910 4.79210283 + layer.3.v_cache 0.00001957 0.00763855 + layer.4.k_cache 0.00072047 0.20675368 + layer.4.v_cache 0.00005229 0.01545706 + layer.4.output 0.01050496 193.59545546 + ------------------------------------------------------------------------------------- + TOTAL 0.04908006 82.95144161 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 362556 +BPFP 1.2481 bits/point +EBPFP 2.4961 equivalent bits/point +MSE 82.951442 +---------------------- -------------------------------------------------------- +Time: 4.339s Load: 0.008s, Pack+Encode: 2.290s, Decode+Unpack: 2.040s +---------------------- -------------------------------------------------------- +💾 Converting with 82.9514 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,844B, BPFP=0.4883 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,072B, BPFP=2.4885 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,292B, BPFP=1.2860 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,568B, BPFP=2.4055 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,652B, BPFP=1.4715 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,324B, BPFP=2.3920 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,552B, BPFP=1.3556 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,120B, BPFP=2.4360 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,996B, BPFP=2.0978 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,452B, BPFP=2.3991 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,244B, BPFP=0.2780 +⌛️ [2/4] FRONTEND: Frontend time: 2.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.708s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14595435 40.51595282 + layer.0.v_cache 0.00001364 0.00521387 + layer.1.k_cache 0.59759904 8.42851079 + layer.1.v_cache 0.00000566 0.00215728 + layer.2.k_cache 0.01409306 0.99331751 + layer.2.v_cache 0.00002047 0.00645921 + layer.3.k_cache 0.06079736 5.87579885 + layer.3.v_cache 0.00002215 0.00736447 + layer.4.k_cache 0.00070776 0.19307671 + layer.4.v_cache 0.00005221 0.01403791 + layer.4.output 0.00769303 191.32429960 + ------------------------------------------------------------------------------------- + TOTAL 0.05135982 82.07717568 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 376116 +BPFP 1.2215 bits/point +EBPFP 2.4431 equivalent bits/point +MSE 82.077176 +---------------------- -------------------------------------------------------- +Time: 3.971s Load: 0.013s, Pack+Encode: 2.250s, Decode+Unpack: 1.708s +---------------------- -------------------------------------------------------- +💾 Converting with 82.0772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 315, 128) +Output shape: (1, 315, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.output: torch.Size([1, 315, 3584]) -> torch.Size([1, 1, 315, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,612B, BPFP=0.4768 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,052B, BPFP=2.2347 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,536B, BPFP=1.3163 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,048B, BPFP=2.1849 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,432B, BPFP=1.4599 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,816B, BPFP=2.1734 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,424B, BPFP=1.3603 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,588B, BPFP=2.2117 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,636B, BPFP=1.9165 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,472B, BPFP=2.1563 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,148B, BPFP=0.2703 +⌛️ [2/4] FRONTEND: Frontend time: 2.408s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.065s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15500355 40.57660280 + layer.0.v_cache 0.00001366 0.00543740 + layer.1.k_cache 0.68167124 8.85009611 + layer.1.v_cache 0.00000571 0.00230290 + layer.2.k_cache 0.01007247 0.90307259 + layer.2.v_cache 0.00001936 0.00706830 + layer.3.k_cache 0.03272899 5.61934136 + layer.3.v_cache 0.00002006 0.00771139 + layer.4.k_cache 0.00071175 0.23062475 + layer.4.v_cache 0.00005427 0.01483882 + layer.4.output 0.04584261 163.80185658 + ------------------------------------------------------------------------------------- + TOTAL 0.07065879 70.75471132 + (elements=2,741,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2741760 +Total Bytes 390764 +BPFP 1.1402 bits/point +EBPFP 2.2804 equivalent bits/point +MSE 70.754711 +---------------------- -------------------------------------------------------- +Time: 4.484s Load: 0.011s, Pack+Encode: 2.408s, Decode+Unpack: 2.065s +---------------------- -------------------------------------------------------- +💾 Converting with 70.7547 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,652B, BPFP=0.4988 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,544B, BPFP=2.5106 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,580B, BPFP=1.2442 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,740B, BPFP=2.4066 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,300B, BPFP=1.3434 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,604B, BPFP=2.3988 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,276B, BPFP=1.3420 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,352B, BPFP=2.4419 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,072B, BPFP=2.1375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,640B, BPFP=2.4008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,380B, BPFP=0.2667 +⌛️ [2/4] FRONTEND: Frontend time: 2.480s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.853s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12889319 41.31320269 + layer.0.v_cache 0.00001350 0.00534339 + layer.1.k_cache 0.58775566 8.10482462 + layer.1.v_cache 0.00000553 0.00218618 + layer.2.k_cache 0.00723753 0.98438986 + layer.2.v_cache 0.00001924 0.00660922 + layer.3.k_cache 0.01761036 5.84263380 + layer.3.v_cache 0.00001847 0.00742458 + layer.4.k_cache 0.00071814 0.19200313 + layer.4.v_cache 0.00005270 0.01478178 + layer.4.output 0.01035140 194.17249275 + ------------------------------------------------------------------------------------- + TOTAL 0.04792848 83.27534403 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 357140 +BPFP 1.2113 bits/point +EBPFP 2.4225 equivalent bits/point +MSE 83.275344 +---------------------- -------------------------------------------------------- +Time: 4.345s Load: 0.012s, Pack+Encode: 2.480s, Decode+Unpack: 1.853s +---------------------- -------------------------------------------------------- +💾 Converting with 83.2753 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,916B, BPFP=0.4905 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,036B, BPFP=2.4778 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,592B, BPFP=1.2430 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,644B, BPFP=2.4012 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,132B, BPFP=1.4377 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,500B, BPFP=2.3933 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,044B, BPFP=1.3228 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,320B, BPFP=2.4384 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,424B, BPFP=2.1140 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,408B, BPFP=2.3882 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,948B, BPFP=0.2668 +⌛️ [2/4] FRONTEND: Frontend time: 2.788s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.024s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13033964 40.74747263 + layer.0.v_cache 0.00001368 0.00523914 + layer.1.k_cache 0.63029007 8.07369264 + layer.1.v_cache 0.00000551 0.00211781 + layer.2.k_cache 0.00637632 0.97501610 + layer.2.v_cache 0.00001960 0.00663840 + layer.3.k_cache 0.02368569 4.26318359 + layer.3.v_cache 0.00001974 0.00688758 + layer.4.k_cache 0.00068050 0.19874649 + layer.4.v_cache 0.00005209 0.01382910 + layer.4.output 0.00765416 187.79626195 + ------------------------------------------------------------------------------------- + TOTAL 0.04970952 80.52156807 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 373964 +BPFP 1.2103 bits/point +EBPFP 2.4205 equivalent bits/point +MSE 80.521568 +---------------------- -------------------------------------------------------- +Time: 4.823s Load: 0.010s, Pack+Encode: 2.788s, Decode+Unpack: 2.024s +---------------------- -------------------------------------------------------- +💾 Converting with 80.5216 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,692B, BPFP=0.4993 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,784B, BPFP=2.5152 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,660B, BPFP=1.2443 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,544B, BPFP=2.4439 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,160B, BPFP=1.3879 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,032B, BPFP=2.4145 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,408B, BPFP=1.3447 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,900B, BPFP=2.4644 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,444B, BPFP=2.1510 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,524B, BPFP=2.4428 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,836B, BPFP=0.2777 +⌛️ [2/4] FRONTEND: Frontend time: 2.220s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.629s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15644295 40.03524960 + layer.0.v_cache 0.00001405 0.00524626 + layer.1.k_cache 0.56163917 7.85013715 + layer.1.v_cache 0.00000554 0.00219347 + layer.2.k_cache 0.01059573 0.98518450 + layer.2.v_cache 0.00001807 0.00671143 + layer.3.k_cache 0.04944291 4.76057928 + layer.3.v_cache 0.00001950 0.00729897 + layer.4.k_cache 0.00071267 0.20652558 + layer.4.v_cache 0.00005817 0.01473004 + layer.4.output 0.00956290 196.23626247 + ------------------------------------------------------------------------------------- + TOTAL 0.04975818 83.97221727 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 362984 +BPFP 1.2266 bits/point +EBPFP 2.4531 equivalent bits/point +MSE 83.972217 +---------------------- -------------------------------------------------------- +Time: 3.863s Load: 0.013s, Pack+Encode: 2.220s, Decode+Unpack: 1.629s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9722 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 312, 128) +Output shape: (1, 312, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.output: torch.Size([1, 312, 3584]) -> torch.Size([1, 1, 312, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,636B, BPFP=0.4826 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,372B, BPFP=2.2722 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,304B, BPFP=1.2672 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,260B, BPFP=2.2165 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,440B, BPFP=1.4243 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,888B, BPFP=2.1979 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,004B, BPFP=1.4024 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,740B, BPFP=2.2406 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,600B, BPFP=1.9331 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,652B, BPFP=2.1861 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,468B, BPFP=0.2824 +⌛️ [2/4] FRONTEND: Frontend time: 2.395s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.089s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14984696 39.83851663 + layer.0.v_cache 0.00001474 0.00537043 + layer.1.k_cache 0.67859346 8.63684239 + layer.1.v_cache 0.00000581 0.00230437 + layer.2.k_cache 0.01626646 0.93289644 + layer.2.v_cache 0.00001951 0.00674437 + layer.3.k_cache 0.02914289 6.28409479 + layer.3.v_cache 0.00001919 0.00724712 + layer.4.k_cache 0.00071665 0.21289566 + layer.4.v_cache 0.00005096 0.01423230 + layer.4.output 0.04642455 167.46867846 + ------------------------------------------------------------------------------------- + TOTAL 0.07056756 72.24834669 + (elements=2,715,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2715648 +Total Bytes 391364 +BPFP 1.1529 bits/point +EBPFP 2.3058 equivalent bits/point +MSE 72.248347 +---------------------- -------------------------------------------------------- +Time: 4.497s Load: 0.013s, Pack+Encode: 2.395s, Decode+Unpack: 2.089s +---------------------- -------------------------------------------------------- +💾 Converting with 72.2483 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,612B, BPFP=0.5002 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,788B, BPFP=2.5434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,984B, BPFP=1.2770 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,544B, BPFP=2.4712 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,076B, BPFP=1.4566 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,984B, BPFP=2.4387 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,712B, BPFP=1.3773 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,220B, BPFP=2.5105 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,744B, BPFP=2.1343 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,052B, BPFP=2.4426 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,736B, BPFP=0.2882 +⌛️ [2/4] FRONTEND: Frontend time: 2.187s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.919s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15823836 40.46929092 + layer.0.v_cache 0.00001502 0.00537246 + layer.1.k_cache 0.53668020 8.61393108 + layer.1.v_cache 0.00000576 0.00221648 + layer.2.k_cache 0.01018995 0.97066557 + layer.2.v_cache 0.00002218 0.00672931 + layer.3.k_cache 0.05410197 6.19772986 + layer.3.v_cache 0.00002001 0.00746854 + layer.4.k_cache 0.00069143 0.19318915 + layer.4.v_cache 0.00005304 0.01431632 + layer.4.output 0.01041225 195.45797929 + ------------------------------------------------------------------------------------- + TOTAL 0.04899433 83.80510381 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 364452 +BPFP 1.2453 bits/point +EBPFP 2.4905 equivalent bits/point +MSE 83.805104 +---------------------- -------------------------------------------------------- +Time: 4.115s Load: 0.010s, Pack+Encode: 2.187s, Decode+Unpack: 1.919s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8051 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 329, 128) +Output shape: (1, 329, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.output: torch.Size([1, 329, 3584]) -> torch.Size([1, 1, 329, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,348B, BPFP=0.4915 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 52,288B, BPFP=2.4833 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,136B, BPFP=1.2413 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,524B, BPFP=2.3995 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,060B, BPFP=1.4276 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 50,712B, BPFP=2.4084 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,892B, BPFP=1.4196 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 51,300B, BPFP=2.4364 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,600B, BPFP=2.0707 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 50,600B, BPFP=2.4031 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 50,504B, BPFP=0.3427 +⌛️ [2/4] FRONTEND: Frontend time: 2.774s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.716s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11391156 39.96772891 + layer.0.v_cache 0.00001522 0.00557201 + layer.1.k_cache 0.63933222 8.62172006 + layer.1.v_cache 0.00000652 0.00243130 + layer.2.k_cache 0.01643764 0.93942530 + layer.2.v_cache 0.00001873 0.00668260 + layer.3.k_cache 0.02826017 4.46273961 + layer.3.v_cache 0.00001876 0.00726794 + layer.4.k_cache 0.00074607 0.19616224 + layer.4.v_cache 0.00005483 0.01462696 + layer.4.output 3.44632065 148.49136995 + ------------------------------------------------------------------------------------- + TOTAL 1.46606155 64.33317333 + (elements=2,863,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2863616 +Total Bytes 445964 +BPFP 1.2459 bits/point +EBPFP 2.4918 equivalent bits/point +MSE 64.333173 +---------------------- -------------------------------------------------------- +Time: 4.501s Load: 0.011s, Pack+Encode: 2.774s, Decode+Unpack: 1.716s +---------------------- -------------------------------------------------------- +💾 Converting with 64.3332 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,672B, BPFP=0.4857 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,780B, BPFP=2.5078 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,068B, BPFP=1.2359 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,260B, BPFP=2.4227 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,688B, BPFP=1.4386 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,000B, BPFP=2.4082 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,460B, BPFP=1.3138 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,796B, BPFP=2.4527 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,568B, BPFP=2.1039 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,112B, BPFP=2.4144 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,564B, BPFP=0.2605 +⌛️ [2/4] FRONTEND: Frontend time: 2.214s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.836s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13536828 40.32595486 + layer.0.v_cache 0.00001433 0.00534591 + layer.1.k_cache 0.57185238 8.11911787 + layer.1.v_cache 0.00000558 0.00218254 + layer.2.k_cache 0.00766427 0.91046657 + layer.2.v_cache 0.00001891 0.00678249 + layer.3.k_cache 0.04143807 3.81207341 + layer.3.v_cache 0.00001947 0.00768061 + layer.4.k_cache 0.00070404 0.20908084 + layer.4.v_cache 0.00005498 0.01529446 + layer.4.output 0.00948056 189.71700589 + ------------------------------------------------------------------------------------- + TOTAL 0.04844142 81.26076593 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 367968 +BPFP 1.2122 bits/point +EBPFP 2.4244 equivalent bits/point +MSE 81.260766 +---------------------- -------------------------------------------------------- +Time: 4.061s Load: 0.010s, Pack+Encode: 2.214s, Decode+Unpack: 1.836s +---------------------- -------------------------------------------------------- +💾 Converting with 81.2608 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,672B, BPFP=0.4909 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,540B, BPFP=2.5215 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,336B, BPFP=1.2645 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,192B, BPFP=2.4452 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,964B, BPFP=1.4133 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,128B, BPFP=2.4416 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,996B, BPFP=1.3019 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,600B, BPFP=2.4683 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,512B, BPFP=2.1236 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,672B, BPFP=2.4158 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,968B, BPFP=0.2828 +⌛️ [2/4] FRONTEND: Frontend time: 2.831s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.950s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15132187 41.75963471 + layer.0.v_cache 0.00001456 0.00523690 + layer.1.k_cache 0.61798344 8.15584841 + layer.1.v_cache 0.00000569 0.00214368 + layer.2.k_cache 0.01686249 0.96173383 + layer.2.v_cache 0.00001898 0.00670357 + layer.3.k_cache 0.04401700 4.70380169 + layer.3.v_cache 0.00001935 0.00722479 + layer.4.k_cache 0.00071189 0.19673964 + layer.4.v_cache 0.00005111 0.01370697 + layer.4.output 0.01058244 194.04323564 + ------------------------------------------------------------------------------------- + TOTAL 0.05324020 83.18326021 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 368580 +BPFP 1.2274 bits/point +EBPFP 2.4548 equivalent bits/point +MSE 83.183260 +---------------------- -------------------------------------------------------- +Time: 4.791s Load: 0.009s, Pack+Encode: 2.831s, Decode+Unpack: 1.950s +---------------------- -------------------------------------------------------- +💾 Converting with 83.1833 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,688B, BPFP=0.4918 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,416B, BPFP=2.5145 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,540B, BPFP=1.2194 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,136B, BPFP=2.4420 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,580B, BPFP=1.4481 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,952B, BPFP=2.4316 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,628B, BPFP=1.2810 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,436B, BPFP=2.4590 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,176B, BPFP=2.1612 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,776B, BPFP=2.4216 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,428B, BPFP=0.2784 +⌛️ [2/4] FRONTEND: Frontend time: 2.230s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.699s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12581717 41.28475643 + layer.0.v_cache 0.00001375 0.00556318 + layer.1.k_cache 0.57814745 8.30020717 + layer.1.v_cache 0.00000592 0.00227871 + layer.2.k_cache 0.00733882 0.79698004 + layer.2.v_cache 0.00002058 0.00708838 + layer.3.k_cache 0.02353939 4.96966376 + layer.3.v_cache 0.00002077 0.00795506 + layer.4.k_cache 0.00069182 0.21235307 + layer.4.v_cache 0.00005027 0.01421272 + layer.4.output 0.01015823 194.89660973 + ------------------------------------------------------------------------------------- + TOTAL 0.04745609 83.52219568 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 367756 +BPFP 1.2247 bits/point +EBPFP 2.4494 equivalent bits/point +MSE 83.522196 +---------------------- -------------------------------------------------------- +Time: 3.940s Load: 0.010s, Pack+Encode: 2.230s, Decode+Unpack: 1.699s +---------------------- -------------------------------------------------------- +💾 Converting with 83.5222 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,648B, BPFP=0.4932 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,384B, BPFP=2.5310 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,844B, BPFP=1.2457 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,800B, BPFP=2.4407 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,620B, BPFP=1.4610 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,024B, BPFP=2.4535 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,616B, BPFP=1.2897 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,928B, BPFP=2.5050 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,716B, BPFP=2.1508 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,124B, BPFP=2.4592 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,844B, BPFP=0.3001 +⌛️ [2/4] FRONTEND: Frontend time: 2.174s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.574s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13625057 42.17130474 + layer.0.v_cache 0.00001362 0.00542586 + layer.1.k_cache 0.57512581 8.34450648 + layer.1.v_cache 0.00000584 0.00233294 + layer.2.k_cache 0.00872732 0.85404979 + layer.2.v_cache 0.00001851 0.00697477 + layer.3.k_cache 0.04362779 5.22379899 + layer.3.v_cache 0.00001925 0.00795949 + layer.4.k_cache 0.00073619 0.21379805 + layer.4.v_cache 0.00005148 0.01469400 + layer.4.output 0.00783937 187.07494786 + ------------------------------------------------------------------------------------- + TOTAL 0.04820306 80.37467530 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 370548 +BPFP 1.2430 bits/point +EBPFP 2.4860 equivalent bits/point +MSE 80.374675 +---------------------- -------------------------------------------------------- +Time: 3.758s Load: 0.010s, Pack+Encode: 2.174s, Decode+Unpack: 1.574s +---------------------- -------------------------------------------------------- +💾 Converting with 80.3747 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,404B, BPFP=0.4833 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,068B, BPFP=2.3164 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,904B, BPFP=1.2800 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,824B, BPFP=2.2525 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,908B, BPFP=1.4344 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,312B, BPFP=2.2262 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,100B, BPFP=1.3929 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,520B, BPFP=2.2882 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,232B, BPFP=1.9650 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,436B, BPFP=2.2325 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,676B, BPFP=0.3060 +⌛️ [2/4] FRONTEND: Frontend time: 2.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.777s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16730670 40.37794254 + layer.0.v_cache 0.00001388 0.00518469 + layer.1.k_cache 0.64158324 8.86808215 + layer.1.v_cache 0.00000571 0.00217454 + layer.2.k_cache 0.02207795 0.97838763 + layer.2.v_cache 0.00001922 0.00650369 + layer.3.k_cache 0.04503629 6.11968512 + layer.3.v_cache 0.00001955 0.00718574 + layer.4.k_cache 0.00070478 0.18587228 + layer.4.v_cache 0.00005260 0.01389945 + layer.4.output 0.04763387 168.03901844 + ------------------------------------------------------------------------------------- + TOTAL 0.07119159 72.51988511 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 389384 +BPFP 1.1773 bits/point +EBPFP 2.3545 equivalent bits/point +MSE 72.519885 +---------------------- -------------------------------------------------------- +Time: 4.287s Load: 0.010s, Pack+Encode: 2.500s, Decode+Unpack: 1.777s +---------------------- -------------------------------------------------------- +💾 Converting with 72.5199 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,684B, BPFP=0.4898 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,584B, BPFP=2.5149 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,484B, BPFP=1.2119 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,236B, BPFP=2.4389 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,404B, BPFP=1.3766 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,200B, BPFP=2.4368 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,844B, BPFP=1.2886 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,732B, BPFP=2.4668 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,216B, BPFP=2.0993 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,868B, BPFP=2.4181 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,500B, BPFP=0.2941 +⌛️ [2/4] FRONTEND: Frontend time: 2.196s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.613s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15225300 40.72730991 + layer.0.v_cache 0.00001427 0.00514278 + layer.1.k_cache 0.52731114 7.92576803 + layer.1.v_cache 0.00000556 0.00214154 + layer.2.k_cache 0.01483519 0.96451511 + layer.2.v_cache 0.00001864 0.00653031 + layer.3.k_cache 0.04699480 3.69966032 + layer.3.v_cache 0.00001968 0.00703093 + layer.4.k_cache 0.00070142 0.18835557 + layer.4.v_cache 0.00005177 0.01407638 + layer.4.output 0.01077793 191.20227888 + ------------------------------------------------------------------------------------- + TOTAL 0.04809712 81.87979312 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 368752 +BPFP 1.2236 bits/point +EBPFP 2.4471 equivalent bits/point +MSE 81.879793 +---------------------- -------------------------------------------------------- +Time: 3.820s Load: 0.011s, Pack+Encode: 2.196s, Decode+Unpack: 1.613s +---------------------- -------------------------------------------------------- +💾 Converting with 81.8798 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,748B, BPFP=0.4935 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,652B, BPFP=2.5187 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,360B, BPFP=1.3177 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,328B, BPFP=2.4440 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,588B, BPFP=1.3870 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,164B, BPFP=2.4348 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,612B, BPFP=1.3319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,756B, BPFP=2.4682 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,556B, BPFP=2.0620 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,304B, BPFP=2.3863 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,404B, BPFP=0.3014 +⌛️ [2/4] FRONTEND: Frontend time: 2.428s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.852s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15700974 40.15253413 + layer.0.v_cache 0.00001374 0.00523365 + layer.1.k_cache 0.52255304 8.16824638 + layer.1.v_cache 0.00000566 0.00218010 + layer.2.k_cache 0.01360668 0.99336281 + layer.2.v_cache 0.00001926 0.00669836 + layer.3.k_cache 0.08085050 3.83219430 + layer.3.v_cache 0.00001943 0.00720818 + layer.4.k_cache 0.00071029 0.18601763 + layer.4.v_cache 0.00005450 0.01373414 + layer.4.output 0.01014708 187.72717896 + ------------------------------------------------------------------------------------- + TOTAL 0.04975720 80.43868602 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 371472 +BPFP 1.2326 bits/point +EBPFP 2.4652 equivalent bits/point +MSE 80.438686 +---------------------- -------------------------------------------------------- +Time: 4.290s Load: 0.010s, Pack+Encode: 2.428s, Decode+Unpack: 1.852s +---------------------- -------------------------------------------------------- +💾 Converting with 80.4387 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,880B, BPFP=0.4868 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,788B, BPFP=2.4555 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,368B, BPFP=1.2811 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,876B, BPFP=2.4055 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,152B, BPFP=1.4886 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,632B, BPFP=2.3921 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,956B, BPFP=1.3682 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,324B, BPFP=2.4300 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,516B, BPFP=2.1116 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,420B, BPFP=2.3805 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,780B, BPFP=0.2802 +⌛️ [2/4] FRONTEND: Frontend time: 2.202s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.666s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16436556 41.49641927 + layer.0.v_cache 0.00001441 0.00544310 + layer.1.k_cache 0.62741763 8.46061369 + layer.1.v_cache 0.00000560 0.00226052 + layer.2.k_cache 0.01487184 0.99775733 + layer.2.v_cache 0.00002167 0.00704492 + layer.3.k_cache 0.02142572 5.98404948 + layer.3.v_cache 0.00001943 0.00779285 + layer.4.k_cache 0.00070776 0.19592853 + layer.4.v_cache 0.00005251 0.01468043 + layer.4.output 0.00894435 183.74857456 + ------------------------------------------------------------------------------------- + TOTAL 0.05244192 79.02423600 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 378692 +BPFP 1.2213 bits/point +EBPFP 2.4425 equivalent bits/point +MSE 79.024236 +---------------------- -------------------------------------------------------- +Time: 3.878s Load: 0.010s, Pack+Encode: 2.202s, Decode+Unpack: 1.666s +---------------------- -------------------------------------------------------- +💾 Converting with 79.0242 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,860B, BPFP=0.4927 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,656B, BPFP=2.4831 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,472B, BPFP=1.2496 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,276B, BPFP=2.4064 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,320B, BPFP=1.4079 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,200B, BPFP=2.4021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,736B, BPFP=1.3754 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,724B, BPFP=2.4313 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,776B, BPFP=2.1005 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,892B, BPFP=2.3850 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,980B, BPFP=0.2699 +⌛️ [2/4] FRONTEND: Frontend time: 3.035s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.710s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11705062 40.54987767 + layer.0.v_cache 0.00001445 0.00535681 + layer.1.k_cache 0.56620099 8.37627196 + layer.1.v_cache 0.00000566 0.00215934 + layer.2.k_cache 0.01299965 1.00793012 + layer.2.v_cache 0.00001972 0.00651425 + layer.3.k_cache 0.05069733 5.11037936 + layer.3.v_cache 0.00001944 0.00734885 + layer.4.k_cache 0.00068883 0.18105037 + layer.4.v_cache 0.00005144 0.01323137 + layer.4.output 0.00838925 186.02400547 + ------------------------------------------------------------------------------------- + TOTAL 0.04743958 79.84871520 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 370892 +BPFP 1.2131 bits/point +EBPFP 2.4263 equivalent bits/point +MSE 79.848715 +---------------------- -------------------------------------------------------- +Time: 4.756s Load: 0.011s, Pack+Encode: 3.035s, Decode+Unpack: 1.710s +---------------------- -------------------------------------------------------- +💾 Converting with 79.8487 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,840B, BPFP=0.4881 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,908B, BPFP=2.4795 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,180B, BPFP=1.2798 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,792B, BPFP=2.4178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,780B, BPFP=1.4786 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,380B, BPFP=2.3951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,036B, BPFP=1.3271 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,320B, BPFP=2.4470 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,392B, BPFP=2.1197 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,280B, BPFP=2.3896 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,304B, BPFP=0.2785 +⌛️ [2/4] FRONTEND: Frontend time: 2.243s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.643s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14370025 40.16910750 + layer.0.v_cache 0.00001382 0.00531423 + layer.1.k_cache 0.61448195 7.99330468 + layer.1.v_cache 0.00000570 0.00221877 + layer.2.k_cache 0.01208573 0.90069472 + layer.2.v_cache 0.00001823 0.00660105 + layer.3.k_cache 0.01319265 5.72266358 + layer.3.v_cache 0.00002242 0.00739307 + layer.4.k_cache 0.00071920 0.20232691 + layer.4.v_cache 0.00005145 0.01413585 + layer.4.output 0.00992508 188.02796883 + ------------------------------------------------------------------------------------- + TOTAL 0.05022159 80.65997307 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 376212 +BPFP 1.2218 bits/point +EBPFP 2.4437 equivalent bits/point +MSE 80.659973 +---------------------- -------------------------------------------------------- +Time: 3.898s Load: 0.011s, Pack+Encode: 2.243s, Decode+Unpack: 1.643s +---------------------- -------------------------------------------------------- +💾 Converting with 80.6600 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 259, 128) +Output shape: (1, 259, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.output: torch.Size([1, 259, 3584]) -> torch.Size([1, 1, 259, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,164B, BPFP=0.4925 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,996B, BPFP=2.5939 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,048B, BPFP=1.2698 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,448B, BPFP=2.4402 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,044B, BPFP=1.3299 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,632B, BPFP=2.4513 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,364B, BPFP=1.3492 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,200B, BPFP=2.4855 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,768B, BPFP=2.0975 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 40,860B, BPFP=2.4650 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,544B, BPFP=0.2977 +⌛️ [2/4] FRONTEND: Frontend time: 2.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.754s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13349269 42.11159960 + layer.0.v_cache 0.00001436 0.00548878 + layer.1.k_cache 0.50562990 8.24052659 + layer.1.v_cache 0.00000590 0.00226870 + layer.2.k_cache 0.00820730 0.89816367 + layer.2.v_cache 0.00001794 0.00690469 + layer.3.k_cache 0.03685973 4.81208524 + layer.3.v_cache 0.00002061 0.00746804 + layer.4.k_cache 0.00069539 0.19717444 + layer.4.v_cache 0.00005422 0.01449857 + layer.4.output 0.01018517 209.67820946 + ------------------------------------------------------------------------------------- + TOTAL 0.04448790 89.64962615 + (elements=2,254,336) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2254336 +Total Bytes 349068 +BPFP 1.2387 bits/point +EBPFP 2.4775 equivalent bits/point +MSE 89.649626 +---------------------- -------------------------------------------------------- +Time: 4.012s Load: 0.009s, Pack+Encode: 2.248s, Decode+Unpack: 1.754s +---------------------- -------------------------------------------------------- +💾 Converting with 89.6496 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,576B, BPFP=0.4963 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,784B, BPFP=2.5338 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,760B, BPFP=1.2593 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,148B, BPFP=2.4391 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,924B, BPFP=1.3845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,300B, BPFP=2.4479 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,316B, BPFP=1.3493 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,684B, BPFP=2.4701 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,800B, BPFP=2.1296 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,284B, BPFP=2.4470 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,288B, BPFP=0.2917 +⌛️ [2/4] FRONTEND: Frontend time: 2.185s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.619s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12649573 40.22951389 + layer.0.v_cache 0.00001373 0.00546770 + layer.1.k_cache 0.51574391 8.27654622 + layer.1.v_cache 0.00000589 0.00227911 + layer.2.k_cache 0.01194659 1.00345629 + layer.2.v_cache 0.00001984 0.00692234 + layer.3.k_cache 0.04962628 5.76955566 + layer.3.v_cache 0.00002059 0.00754375 + layer.4.k_cache 0.00068541 0.19659233 + layer.4.v_cache 0.00005704 0.01453341 + layer.4.output 0.00772497 197.42397487 + ------------------------------------------------------------------------------------- + TOTAL 0.04462881 84.55766087 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 362864 +BPFP 1.2352 bits/point +EBPFP 2.4705 equivalent bits/point +MSE 84.557661 +---------------------- -------------------------------------------------------- +Time: 3.816s Load: 0.012s, Pack+Encode: 2.185s, Decode+Unpack: 1.619s +---------------------- -------------------------------------------------------- +💾 Converting with 84.5577 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,632B, BPFP=0.5014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,448B, BPFP=2.5237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,768B, BPFP=1.2644 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,828B, BPFP=2.4296 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,580B, BPFP=1.3697 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,160B, BPFP=2.4489 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,352B, BPFP=1.4145 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,892B, BPFP=2.4914 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,228B, BPFP=2.1624 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,332B, BPFP=2.4589 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,036B, BPFP=0.2741 +⌛️ [2/4] FRONTEND: Frontend time: 2.243s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.747s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12210708 40.41063546 + layer.0.v_cache 0.00001342 0.00534623 + layer.1.k_cache 0.54942129 8.39380881 + layer.1.v_cache 0.00000547 0.00218359 + layer.2.k_cache 0.01053270 0.99522088 + layer.2.v_cache 0.00001820 0.00666535 + layer.3.k_cache 0.02850239 5.89042442 + layer.3.v_cache 0.00002491 0.00764692 + layer.4.k_cache 0.00067987 0.18670237 + layer.4.v_cache 0.00005179 0.01414261 + layer.4.output 0.01126663 196.70021243 + ------------------------------------------------------------------------------------- + TOTAL 0.04648373 84.28319198 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 361256 +BPFP 1.2343 bits/point +EBPFP 2.4687 equivalent bits/point +MSE 84.283192 +---------------------- -------------------------------------------------------- +Time: 4.000s Load: 0.010s, Pack+Encode: 2.243s, Decode+Unpack: 1.747s +---------------------- -------------------------------------------------------- +💾 Converting with 84.2832 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,492B, BPFP=0.4895 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,928B, BPFP=2.3168 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,180B, BPFP=1.2469 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,552B, BPFP=2.2459 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,216B, BPFP=1.4550 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,396B, BPFP=2.2378 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,948B, BPFP=1.4412 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,032B, BPFP=2.2706 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,436B, BPFP=1.9821 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,172B, BPFP=2.2263 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,056B, BPFP=0.2656 +⌛️ [2/4] FRONTEND: Frontend time: 2.376s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.825s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14443027 39.68197904 + layer.0.v_cache 0.00001375 0.00538754 + layer.1.k_cache 0.62837083 8.20037983 + layer.1.v_cache 0.00000570 0.00225680 + layer.2.k_cache 0.01557094 0.91886252 + layer.2.v_cache 0.00001871 0.00709207 + layer.3.k_cache 0.04676386 6.65104328 + layer.3.v_cache 0.00001981 0.00741292 + layer.4.k_cache 0.00070582 0.20604360 + layer.4.v_cache 0.00005437 0.01397124 + layer.4.output 0.04838977 170.99655233 + ------------------------------------------------------------------------------------- + TOTAL 0.06909897 73.68648795 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 383408 +BPFP 1.1630 bits/point +EBPFP 2.3261 equivalent bits/point +MSE 73.686488 +---------------------- -------------------------------------------------------- +Time: 4.215s Load: 0.014s, Pack+Encode: 2.376s, Decode+Unpack: 1.825s +---------------------- -------------------------------------------------------- +💾 Converting with 73.6865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,848B, BPFP=0.4885 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,812B, BPFP=2.4742 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,716B, BPFP=1.3094 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,584B, BPFP=2.4064 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,408B, BPFP=1.4580 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,576B, BPFP=2.4059 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,064B, BPFP=1.3286 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,044B, BPFP=2.4318 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,520B, BPFP=2.1268 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,348B, BPFP=2.3933 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,440B, BPFP=0.2716 +⌛️ [2/4] FRONTEND: Frontend time: 2.266s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.636s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13293209 39.96033017 + layer.0.v_cache 0.00001388 0.00541143 + layer.1.k_cache 0.56267701 8.19633478 + layer.1.v_cache 0.00000566 0.00220122 + layer.2.k_cache 0.01145046 0.91960678 + layer.2.v_cache 0.00001921 0.00709573 + layer.3.k_cache 0.02674564 4.42443977 + layer.3.v_cache 0.00001926 0.00758632 + layer.4.k_cache 0.00069135 0.22389294 + layer.4.v_cache 0.00005493 0.01497197 + layer.4.output 0.00973303 185.06925164 + ------------------------------------------------------------------------------------- + TOTAL 0.04722004 79.36744898 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 375360 +BPFP 1.2191 bits/point +EBPFP 2.4382 equivalent bits/point +MSE 79.367449 +---------------------- -------------------------------------------------------- +Time: 3.912s Load: 0.011s, Pack+Encode: 2.266s, Decode+Unpack: 1.636s +---------------------- -------------------------------------------------------- +💾 Converting with 79.3674 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,692B, BPFP=0.4903 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,604B, BPFP=2.5160 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,724B, BPFP=1.2254 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,424B, BPFP=2.4495 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,616B, BPFP=1.3885 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,180B, BPFP=2.4357 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,728B, BPFP=1.2820 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,788B, BPFP=2.4700 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,776B, BPFP=2.1309 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,844B, BPFP=2.4167 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,272B, BPFP=0.2681 +⌛️ [2/4] FRONTEND: Frontend time: 2.149s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.655s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15182159 40.67413484 + layer.0.v_cache 0.00001420 0.00542908 + layer.1.k_cache 0.56224567 7.95895000 + layer.1.v_cache 0.00000588 0.00229770 + layer.2.k_cache 0.00652889 0.84857965 + layer.2.v_cache 0.00001844 0.00679039 + layer.3.k_cache 0.06011557 3.71148836 + layer.3.v_cache 0.00001921 0.00778387 + layer.4.k_cache 0.00068317 0.20786710 + layer.4.v_cache 0.00005096 0.01428352 + layer.4.output 0.01130176 170.82608626 + ------------------------------------------------------------------------------------- + TOTAL 0.05062446 73.48354166 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 366648 +BPFP 1.2166 bits/point +EBPFP 2.4332 equivalent bits/point +MSE 73.483542 +---------------------- -------------------------------------------------------- +Time: 3.814s Load: 0.010s, Pack+Encode: 2.149s, Decode+Unpack: 1.655s +---------------------- -------------------------------------------------------- +💾 Converting with 73.4835 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,868B, BPFP=0.4862 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,184B, BPFP=2.4772 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,988B, BPFP=1.2603 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,096B, BPFP=2.4175 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,248B, BPFP=1.5487 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,776B, BPFP=2.4000 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,280B, BPFP=1.3860 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,404B, BPFP=2.4344 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,704B, BPFP=2.1219 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,724B, BPFP=2.3971 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,904B, BPFP=0.2969 +⌛️ [2/4] FRONTEND: Frontend time: 2.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.602s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13053865 40.97881373 + layer.0.v_cache 0.00001414 0.00563313 + layer.1.k_cache 0.56363723 8.25851665 + layer.1.v_cache 0.00000590 0.00230954 + layer.2.k_cache 0.01033439 0.94542151 + layer.2.v_cache 0.00002015 0.00690710 + layer.3.k_cache 0.03655195 5.98529717 + layer.3.v_cache 0.00002026 0.00760195 + layer.4.k_cache 0.00068516 0.21248821 + layer.4.v_cache 0.00005378 0.01466153 + layer.4.output 0.00863618 181.20722118 + ------------------------------------------------------------------------------------- + TOTAL 0.04719499 77.93342346 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 383176 +BPFP 1.2357 bits/point +EBPFP 2.4715 equivalent bits/point +MSE 77.933423 +---------------------- -------------------------------------------------------- +Time: 3.863s Load: 0.010s, Pack+Encode: 2.251s, Decode+Unpack: 1.602s +---------------------- -------------------------------------------------------- +💾 Converting with 77.9334 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,696B, BPFP=0.4905 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,640B, BPFP=2.5181 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,020B, BPFP=1.2421 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,452B, BPFP=2.4510 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,164B, BPFP=1.4194 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,076B, BPFP=2.4298 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,232B, BPFP=1.3105 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,688B, BPFP=2.4644 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,924B, BPFP=2.1392 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,056B, BPFP=2.4287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,940B, BPFP=0.2896 +⌛️ [2/4] FRONTEND: Frontend time: 2.464s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.916s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15908699 42.38275609 + layer.0.v_cache 0.00001433 0.00545494 + layer.1.k_cache 0.53822481 8.22051099 + layer.1.v_cache 0.00000570 0.00226862 + layer.2.k_cache 0.01139215 0.89477291 + layer.2.v_cache 0.00001845 0.00678830 + layer.3.k_cache 0.02907990 5.31199674 + layer.3.v_cache 0.00001963 0.00749211 + layer.4.k_cache 0.00069750 0.19182529 + layer.4.v_cache 0.00005397 0.01441277 + layer.4.output 0.01069233 154.91471119 + ------------------------------------------------------------------------------------- + TOTAL 0.04784940 67.14360336 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 370888 +BPFP 1.2306 bits/point +EBPFP 2.4613 equivalent bits/point +MSE 67.143603 +---------------------- -------------------------------------------------------- +Time: 4.390s Load: 0.011s, Pack+Encode: 2.464s, Decode+Unpack: 1.916s +---------------------- -------------------------------------------------------- +💾 Converting with 67.1436 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,808B, BPFP=0.4812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,024B, BPFP=2.4598 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,036B, BPFP=1.2585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,972B, BPFP=2.4023 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,172B, BPFP=1.4845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,632B, BPFP=2.3837 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,464B, BPFP=1.3912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,580B, BPFP=2.4355 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,488B, BPFP=2.1027 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,512B, BPFP=2.3772 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,916B, BPFP=0.2959 +⌛️ [2/4] FRONTEND: Frontend time: 2.587s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.868s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16084659 41.39549757 + layer.0.v_cache 0.00001440 0.00540791 + layer.1.k_cache 0.59028764 8.25382942 + layer.1.v_cache 0.00000559 0.00228029 + layer.2.k_cache 0.00929020 0.90642457 + layer.2.v_cache 0.00002205 0.00681253 + layer.3.k_cache 0.04382113 6.08522471 + layer.3.v_cache 0.00002103 0.00743378 + layer.4.k_cache 0.00069271 0.19463787 + layer.4.v_cache 0.00005052 0.01399167 + layer.4.output 0.00984751 184.89622877 + ------------------------------------------------------------------------------------- + TOTAL 0.05141085 79.47912598 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 381604 +BPFP 1.2264 bits/point +EBPFP 2.4527 equivalent bits/point +MSE 79.479126 +---------------------- -------------------------------------------------------- +Time: 4.469s Load: 0.013s, Pack+Encode: 2.587s, Decode+Unpack: 1.868s +---------------------- -------------------------------------------------------- +💾 Converting with 79.4791 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,828B, BPFP=0.4789 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,032B, BPFP=2.4431 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,188B, BPFP=1.2580 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,124B, BPFP=2.3939 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,088B, BPFP=1.5239 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,460B, BPFP=2.3579 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,592B, BPFP=1.3885 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,556B, BPFP=2.4173 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,344B, BPFP=2.0803 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,636B, BPFP=2.3674 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,680B, BPFP=0.3075 +⌛️ [2/4] FRONTEND: Frontend time: 2.241s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.697s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14211286 38.63860067 + layer.0.v_cache 0.00001516 0.00540078 + layer.1.k_cache 0.60291523 8.57789188 + layer.1.v_cache 0.00000570 0.00227724 + layer.2.k_cache 0.01041983 0.92435349 + layer.2.v_cache 0.00001835 0.00660443 + layer.3.k_cache 0.03896446 5.61787754 + layer.3.v_cache 0.00001926 0.00744805 + layer.4.k_cache 0.00071470 0.19254249 + layer.4.v_cache 0.00005236 0.01465044 + layer.4.output 0.00754456 180.30574157 + ------------------------------------------------------------------------------------- + TOTAL 0.04988528 77.41928459 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 384528 +BPFP 1.2272 bits/point +EBPFP 2.4544 equivalent bits/point +MSE 77.419285 +---------------------- -------------------------------------------------------- +Time: 3.948s Load: 0.010s, Pack+Encode: 2.241s, Decode+Unpack: 1.697s +---------------------- -------------------------------------------------------- +💾 Converting with 77.4193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 296, 128) +Output shape: (1, 296, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.output: torch.Size([1, 296, 3584]) -> torch.Size([1, 1, 296, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,352B, BPFP=0.4937 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,992B, BPFP=2.3750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,452B, BPFP=1.2908 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,800B, BPFP=2.3121 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,092B, BPFP=1.4829 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,716B, BPFP=2.3076 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,308B, BPFP=1.3359 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,524B, BPFP=2.3503 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,416B, BPFP=2.0279 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,424B, BPFP=2.2922 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,840B, BPFP=0.2778 +⌛️ [2/4] FRONTEND: Frontend time: 2.240s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.822s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14616069 39.55759739 + layer.0.v_cache 0.00001366 0.00526035 + layer.1.k_cache 0.61022733 8.36192281 + layer.1.v_cache 0.00000576 0.00213041 + layer.2.k_cache 0.01378242 1.02775409 + layer.2.v_cache 0.00001907 0.00668683 + layer.3.k_cache 0.03184220 5.16747057 + layer.3.v_cache 0.00001934 0.00714557 + layer.4.k_cache 0.00072724 0.18830003 + layer.4.v_cache 0.00005080 0.01384074 + layer.4.output 0.04784788 177.85809303 + ------------------------------------------------------------------------------------- + TOTAL 0.06692845 76.43204470 + (elements=2,576,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2576384 +Total Bytes 382916 +BPFP 1.1890 bits/point +EBPFP 2.3780 equivalent bits/point +MSE 76.432045 +---------------------- -------------------------------------------------------- +Time: 4.072s Load: 0.010s, Pack+Encode: 2.240s, Decode+Unpack: 1.822s +---------------------- -------------------------------------------------------- +💾 Converting with 76.4320 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,356B, BPFP=0.4906 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,052B, BPFP=2.3622 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,856B, BPFP=1.3033 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,832B, BPFP=2.2982 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,300B, BPFP=1.5363 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,616B, BPFP=2.2869 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,132B, BPFP=1.4226 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,488B, BPFP=2.3326 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,612B, BPFP=2.0245 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,352B, BPFP=2.2731 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,672B, BPFP=0.2897 +⌛️ [2/4] FRONTEND: Frontend time: 2.284s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.661s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15387250 41.38877019 + layer.0.v_cache 0.00001382 0.00550799 + layer.1.k_cache 0.63844893 8.68744921 + layer.1.v_cache 0.00000583 0.00231221 + layer.2.k_cache 0.01105453 0.94422995 + layer.2.v_cache 0.00001885 0.00688294 + layer.3.k_cache 0.03893430 5.91046429 + layer.3.v_cache 0.00001923 0.00767345 + layer.4.k_cache 0.00070257 0.20736580 + layer.4.v_cache 0.00005146 0.01457082 + layer.4.output 0.04873301 169.22117689 + ------------------------------------------------------------------------------------- + TOTAL 0.06966195 73.04255677 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 388268 +BPFP 1.1975 bits/point +EBPFP 2.3951 equivalent bits/point +MSE 73.042557 +---------------------- -------------------------------------------------------- +Time: 3.956s Load: 0.012s, Pack+Encode: 2.284s, Decode+Unpack: 1.661s +---------------------- -------------------------------------------------------- +💾 Converting with 73.0426 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,644B, BPFP=0.4984 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,108B, BPFP=2.5431 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,600B, BPFP=1.2454 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,704B, BPFP=2.4622 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,388B, BPFP=1.4061 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,472B, BPFP=2.4488 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,464B, BPFP=1.4105 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,368B, BPFP=2.4428 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,220B, BPFP=2.1460 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,292B, BPFP=2.4384 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,504B, BPFP=0.3089 +⌛️ [2/4] FRONTEND: Frontend time: 2.101s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.731s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12629252 40.52752393 + layer.0.v_cache 0.00001451 0.00547859 + layer.1.k_cache 0.49097243 8.19920614 + layer.1.v_cache 0.00000577 0.00233132 + layer.2.k_cache 0.01060468 0.86710462 + layer.2.v_cache 0.00001909 0.00698816 + layer.3.k_cache 0.02748298 5.45177637 + layer.3.v_cache 0.00001876 0.00734617 + layer.4.k_cache 0.00069967 0.20148167 + layer.4.v_cache 0.00005414 0.01447923 + layer.4.output 0.01047103 201.11765287 + ------------------------------------------------------------------------------------- + TOTAL 0.04290952 86.06513449 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 367764 +BPFP 1.2473 bits/point +EBPFP 2.4946 equivalent bits/point +MSE 86.065134 +---------------------- -------------------------------------------------------- +Time: 3.843s Load: 0.010s, Pack+Encode: 2.101s, Decode+Unpack: 1.731s +---------------------- -------------------------------------------------------- +💾 Converting with 86.0651 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,780B, BPFP=0.4917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,772B, BPFP=2.5074 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,868B, BPFP=1.2247 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,156B, BPFP=2.4169 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,312B, BPFP=1.3616 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,612B, BPFP=2.3864 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,068B, BPFP=1.2919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,260B, BPFP=2.4227 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,736B, BPFP=2.1134 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,144B, BPFP=2.4162 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,564B, BPFP=0.2845 +⌛️ [2/4] FRONTEND: Frontend time: 2.228s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.708s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15022702 39.87001568 + layer.0.v_cache 0.00001395 0.00522265 + layer.1.k_cache 0.54271635 8.19077446 + layer.1.v_cache 0.00000573 0.00217783 + layer.2.k_cache 0.01332666 0.96067750 + layer.2.v_cache 0.00001832 0.00664000 + layer.3.k_cache 0.02579262 3.92690920 + layer.3.v_cache 0.00001838 0.00722839 + layer.4.k_cache 0.00069325 0.18549426 + layer.4.v_cache 0.00005155 0.01443100 + layer.4.output 0.01006869 185.18485983 + ------------------------------------------------------------------------------------- + TOTAL 0.04725557 79.38021116 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 368272 +BPFP 1.2132 bits/point +EBPFP 2.4264 equivalent bits/point +MSE 79.380211 +---------------------- -------------------------------------------------------- +Time: 3.947s Load: 0.011s, Pack+Encode: 2.228s, Decode+Unpack: 1.708s +---------------------- -------------------------------------------------------- +💾 Converting with 79.3802 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,316B, BPFP=0.4934 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,916B, BPFP=2.3790 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,168B, BPFP=1.3331 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,708B, BPFP=2.3150 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,672B, BPFP=1.5186 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,440B, BPFP=2.3008 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,720B, BPFP=1.4153 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,384B, BPFP=2.3508 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,412B, BPFP=2.0345 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,300B, BPFP=2.2934 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,388B, BPFP=0.2980 +⌛️ [2/4] FRONTEND: Frontend time: 2.448s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.599s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12416320 39.88829449 + layer.0.v_cache 0.00001370 0.00532051 + layer.1.k_cache 0.62725913 8.34275358 + layer.1.v_cache 0.00000565 0.00223332 + layer.2.k_cache 0.01102555 1.00149495 + layer.2.v_cache 0.00002033 0.00682126 + layer.3.k_cache 0.04118381 5.73250257 + layer.3.v_cache 0.00001949 0.00743001 + layer.4.k_cache 0.00070768 0.20737274 + layer.4.v_cache 0.00005397 0.01440327 + layer.4.output 0.05105524 173.82952482 + ------------------------------------------------------------------------------------- + TOTAL 0.06834348 74.82442944 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 387424 +BPFP 1.2071 bits/point +EBPFP 2.4142 equivalent bits/point +MSE 74.824429 +---------------------- -------------------------------------------------------- +Time: 4.057s Load: 0.010s, Pack+Encode: 2.448s, Decode+Unpack: 1.599s +---------------------- -------------------------------------------------------- +💾 Converting with 74.8244 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,880B, BPFP=0.4801 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,932B, BPFP=2.4293 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,032B, BPFP=1.3534 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,012B, BPFP=2.3795 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,060B, BPFP=1.5171 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,480B, BPFP=2.3508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,296B, BPFP=1.4217 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,464B, BPFP=2.4040 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,596B, BPFP=2.0867 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,524B, BPFP=2.3532 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,684B, BPFP=0.2756 +⌛️ [2/4] FRONTEND: Frontend time: 2.199s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.741s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15516999 40.90917631 + layer.0.v_cache 0.00001361 0.00567753 + layer.1.k_cache 0.56212566 8.68737835 + layer.1.v_cache 0.00000565 0.00236453 + layer.2.k_cache 0.00839357 0.92960756 + layer.2.v_cache 0.00001980 0.00717260 + layer.3.k_cache 0.02402391 6.36193721 + layer.3.v_cache 0.00001987 0.00779378 + layer.4.k_cache 0.00071212 0.21163988 + layer.4.v_cache 0.00005188 0.01523491 + layer.4.output 0.05073151 169.96342066 + ------------------------------------------------------------------------------------- + TOTAL 0.06503863 73.34599572 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 382960 +BPFP 1.2179 bits/point +EBPFP 2.4359 equivalent bits/point +MSE 73.345996 +---------------------- -------------------------------------------------------- +Time: 3.952s Load: 0.012s, Pack+Encode: 2.199s, Decode+Unpack: 1.741s +---------------------- -------------------------------------------------------- +💾 Converting with 73.3460 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,356B, BPFP=0.4922 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,072B, BPFP=2.3712 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,928B, BPFP=1.3641 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,820B, BPFP=2.3053 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,508B, BPFP=1.4998 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,636B, BPFP=2.2957 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,316B, BPFP=1.3845 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,676B, BPFP=2.3504 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,612B, BPFP=2.0314 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,584B, BPFP=2.2929 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,352B, BPFP=0.2882 +⌛️ [2/4] FRONTEND: Frontend time: 2.180s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.594s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005571 41.52881681 + layer.0.v_cache 0.00001459 0.00544763 + layer.1.k_cache 0.64038677 8.54774388 + layer.1.v_cache 0.00000545 0.00220849 + layer.2.k_cache 0.01623841 0.89583600 + layer.2.v_cache 0.00002030 0.00697357 + layer.3.k_cache 0.01774169 5.83948083 + layer.3.v_cache 0.00001948 0.00752494 + layer.4.k_cache 0.00069419 0.19884912 + layer.4.v_cache 0.00005057 0.01430331 + layer.4.output 0.05052170 170.40145503 + ------------------------------------------------------------------------------------- + TOTAL 0.06875759 73.52102175 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 387860 +BPFP 1.2003 bits/point +EBPFP 2.4006 equivalent bits/point +MSE 73.521022 +---------------------- -------------------------------------------------------- +Time: 3.785s Load: 0.011s, Pack+Encode: 2.180s, Decode+Unpack: 1.594s +---------------------- -------------------------------------------------------- +💾 Converting with 73.5210 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,724B, BPFP=0.4886 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,896B, BPFP=2.5143 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,260B, BPFP=1.2466 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,384B, BPFP=2.4297 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,944B, BPFP=1.3970 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,020B, BPFP=2.4093 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,076B, BPFP=1.3483 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,912B, BPFP=2.4592 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,668B, BPFP=2.1095 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,048B, BPFP=2.4108 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,564B, BPFP=0.2765 +⌛️ [2/4] FRONTEND: Frontend time: 2.378s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.678s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14010120 39.61760053 + layer.0.v_cache 0.00001437 0.00535315 + layer.1.k_cache 0.58692467 7.98052913 + layer.1.v_cache 0.00000569 0.00224348 + layer.2.k_cache 0.00806839 0.99740863 + layer.2.v_cache 0.00001934 0.00669521 + layer.3.k_cache 0.03832107 4.15136981 + layer.3.v_cache 0.00001992 0.00744420 + layer.4.k_cache 0.00069804 0.19390101 + layer.4.v_cache 0.00005131 0.01413534 + layer.4.output 0.00759059 187.51657706 + ------------------------------------------------------------------------------------- + TOTAL 0.04866812 80.32898352 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 370496 +BPFP 1.2205 bits/point +EBPFP 2.4411 equivalent bits/point +MSE 80.328984 +---------------------- -------------------------------------------------------- +Time: 4.065s Load: 0.009s, Pack+Encode: 2.378s, Decode+Unpack: 1.678s +---------------------- -------------------------------------------------------- +💾 Converting with 80.3290 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,584B, BPFP=0.5005 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,884B, BPFP=2.5585 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,608B, BPFP=1.2598 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,948B, BPFP=2.5040 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,112B, BPFP=1.4058 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,812B, BPFP=2.4960 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,732B, BPFP=1.3836 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,508B, BPFP=2.4783 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,452B, BPFP=2.1835 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,088B, BPFP=2.4538 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,612B, BPFP=0.2883 +⌛️ [2/4] FRONTEND: Frontend time: 2.204s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.623s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13199043 42.22207323 + layer.0.v_cache 0.00001386 0.00532144 + layer.1.k_cache 0.56989129 8.22463124 + layer.1.v_cache 0.00000577 0.00220838 + layer.2.k_cache 0.00892691 0.96718677 + layer.2.v_cache 0.00001816 0.00646951 + layer.3.k_cache 0.04626510 5.58549568 + layer.3.v_cache 0.00001927 0.00708541 + layer.4.k_cache 0.00069392 0.20544661 + layer.4.v_cache 0.00005195 0.01427216 + layer.4.output 0.00911696 189.56871335 + ------------------------------------------------------------------------------------- + TOTAL 0.04833502 81.42477552 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 364340 +BPFP 1.2495 bits/point +EBPFP 2.4990 equivalent bits/point +MSE 81.424776 +---------------------- -------------------------------------------------------- +Time: 3.838s Load: 0.011s, Pack+Encode: 2.204s, Decode+Unpack: 1.623s +---------------------- -------------------------------------------------------- +💾 Converting with 81.4248 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,648B, BPFP=0.4932 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,140B, BPFP=2.5171 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,792B, BPFP=1.2427 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,912B, BPFP=2.4471 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,432B, BPFP=1.4503 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,764B, BPFP=2.4386 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,340B, BPFP=1.2740 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,280B, BPFP=2.4681 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,696B, BPFP=2.1496 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,876B, BPFP=2.4450 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,148B, BPFP=0.2782 +⌛️ [2/4] FRONTEND: Frontend time: 2.406s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.694s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16500157 42.20090813 + layer.0.v_cache 0.00001391 0.00549240 + layer.1.k_cache 0.57858722 8.08941428 + layer.1.v_cache 0.00000575 0.00229431 + layer.2.k_cache 0.01018238 0.98748623 + layer.2.v_cache 0.00001877 0.00707294 + layer.3.k_cache 0.03425542 4.88771001 + layer.3.v_cache 0.00002021 0.00803770 + layer.4.k_cache 0.00069807 0.20325272 + layer.4.v_cache 0.00005194 0.01433483 + layer.4.output 0.01023613 186.52235401 + ------------------------------------------------------------------------------------- + TOTAL 0.05061695 80.12132245 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 366028 +BPFP 1.2278 bits/point +EBPFP 2.4556 equivalent bits/point +MSE 80.121322 +---------------------- -------------------------------------------------------- +Time: 4.110s Load: 0.010s, Pack+Encode: 2.406s, Decode+Unpack: 1.694s +---------------------- -------------------------------------------------------- +💾 Converting with 80.1213 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,588B, BPFP=0.4970 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,148B, BPFP=2.4970 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,320B, BPFP=1.2338 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,976B, BPFP=2.4292 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,584B, BPFP=1.3648 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,344B, BPFP=2.4505 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,348B, BPFP=1.3512 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,556B, BPFP=2.4627 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,732B, BPFP=2.1257 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,436B, BPFP=2.4558 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,652B, BPFP=0.3030 +⌛️ [2/4] FRONTEND: Frontend time: 2.135s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.797s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385483 39.92359664 + layer.0.v_cache 0.00001387 0.00537781 + layer.1.k_cache 0.55871571 8.18388310 + layer.1.v_cache 0.00000565 0.00214087 + layer.2.k_cache 0.00818202 0.83547499 + layer.2.v_cache 0.00001875 0.00674247 + layer.3.k_cache 0.04697755 4.58917146 + layer.3.v_cache 0.00001978 0.00766624 + layer.4.k_cache 0.00071636 0.17568494 + layer.4.v_cache 0.00005146 0.01409639 + layer.4.output 0.01106052 191.96111111 + ------------------------------------------------------------------------------------- + TOTAL 0.04741057 82.20421251 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 362684 +BPFP 1.2346 bits/point +EBPFP 2.4693 equivalent bits/point +MSE 82.204213 +---------------------- -------------------------------------------------------- +Time: 3.942s Load: 0.009s, Pack+Encode: 2.135s, Decode+Unpack: 1.797s +---------------------- -------------------------------------------------------- +💾 Converting with 82.2042 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,656B, BPFP=0.4918 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,412B, BPFP=2.5234 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,332B, BPFP=1.2689 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,088B, BPFP=2.4482 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,204B, BPFP=1.4320 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,188B, BPFP=2.4539 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,904B, BPFP=1.3014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,808B, BPFP=2.4891 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,972B, BPFP=2.1575 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,872B, BPFP=2.4359 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,628B, BPFP=0.2730 +⌛️ [2/4] FRONTEND: Frontend time: 2.243s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.794s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203608 39.50447088 + layer.0.v_cache 0.00001420 0.00521994 + layer.1.k_cache 0.56673284 7.86614080 + layer.1.v_cache 0.00000552 0.00216443 + layer.2.k_cache 0.00530624 0.91840032 + layer.2.v_cache 0.00001825 0.00685652 + layer.3.k_cache 0.02188354 5.21073908 + layer.3.v_cache 0.00001854 0.00742491 + layer.4.k_cache 0.00069888 0.20504200 + layer.4.v_cache 0.00004918 0.01417359 + layer.4.output 0.00799176 192.93391234 + ------------------------------------------------------------------------------------- + TOTAL 0.04662974 82.60458934 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 368064 +BPFP 1.2302 bits/point +EBPFP 2.4603 equivalent bits/point +MSE 82.604589 +---------------------- -------------------------------------------------------- +Time: 4.046s Load: 0.010s, Pack+Encode: 2.243s, Decode+Unpack: 1.794s +---------------------- -------------------------------------------------------- +💾 Converting with 82.6046 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,824B, BPFP=0.4872 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,640B, BPFP=2.4647 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,208B, BPFP=1.2814 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,388B, BPFP=2.3955 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,892B, BPFP=1.4848 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,172B, BPFP=2.3836 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,312B, BPFP=1.3423 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,132B, BPFP=2.4366 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,460B, BPFP=2.0682 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,168B, BPFP=2.3834 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,912B, BPFP=0.2911 +⌛️ [2/4] FRONTEND: Frontend time: 2.329s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.644s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12732459 40.99100044 + layer.0.v_cache 0.00001480 0.00523822 + layer.1.k_cache 0.57322790 7.78253972 + layer.1.v_cache 0.00000557 0.00222432 + layer.2.k_cache 0.00950604 0.96014210 + layer.2.v_cache 0.00002039 0.00683361 + layer.3.k_cache 0.01510038 5.79325775 + layer.3.v_cache 0.00001916 0.00763892 + layer.4.k_cache 0.00068141 0.18841643 + layer.4.v_cache 0.00005165 0.01398596 + layer.4.output 0.00936729 190.28039816 + ------------------------------------------------------------------------------------- + TOTAL 0.04656017 81.63023909 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 376108 +BPFP 1.2215 bits/point +EBPFP 2.4430 equivalent bits/point +MSE 81.630239 +---------------------- -------------------------------------------------------- +Time: 3.984s Load: 0.011s, Pack+Encode: 2.329s, Decode+Unpack: 1.644s +---------------------- -------------------------------------------------------- +💾 Converting with 81.6302 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,696B, BPFP=0.4923 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,476B, BPFP=2.5179 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,008B, BPFP=1.2459 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,268B, BPFP=2.4495 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,700B, BPFP=1.3983 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,988B, BPFP=2.4337 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,888B, BPFP=1.2957 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,464B, BPFP=2.4606 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,848B, BPFP=2.1427 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,016B, BPFP=2.4352 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,388B, BPFP=0.2943 +⌛️ [2/4] FRONTEND: Frontend time: 2.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.701s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13630490 40.57463273 + layer.0.v_cache 0.00001544 0.00532522 + layer.1.k_cache 0.58776402 8.24684917 + layer.1.v_cache 0.00000559 0.00215393 + layer.2.k_cache 0.01000112 1.07688240 + layer.2.v_cache 0.00001940 0.00683611 + layer.3.k_cache 0.04424953 4.10070624 + layer.3.v_cache 0.00002139 0.00749314 + layer.4.k_cache 0.00071093 0.19985970 + layer.4.v_cache 0.00005158 0.01416670 + layer.4.output 0.01101471 193.92412332 + ------------------------------------------------------------------------------------- + TOTAL 0.05036746 83.04139815 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 369740 +BPFP 1.2313 bits/point +EBPFP 2.4626 equivalent bits/point +MSE 83.041398 +---------------------- -------------------------------------------------------- +Time: 4.036s Load: 0.010s, Pack+Encode: 2.324s, Decode+Unpack: 1.701s +---------------------- -------------------------------------------------------- +💾 Converting with 83.0414 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,712B, BPFP=0.4897 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,780B, BPFP=2.5169 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,020B, BPFP=1.2376 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,268B, BPFP=2.4319 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,112B, BPFP=1.4114 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,104B, BPFP=2.4227 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,680B, BPFP=1.3309 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,468B, BPFP=2.4431 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,444B, BPFP=2.1045 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,212B, BPFP=2.4287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,120B, BPFP=0.2900 +⌛️ [2/4] FRONTEND: Frontend time: 2.548s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.993s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12958164 40.98652836 + layer.0.v_cache 0.00001430 0.00542416 + layer.1.k_cache 0.54121377 7.98511269 + layer.1.v_cache 0.00000577 0.00223080 + layer.2.k_cache 0.01080061 0.83967771 + layer.2.v_cache 0.00001903 0.00669104 + layer.3.k_cache 0.01783211 4.91296738 + layer.3.v_cache 0.00002028 0.00739528 + layer.4.k_cache 0.00068935 0.19424167 + layer.4.v_cache 0.00005053 0.01445494 + layer.4.output 0.00833506 186.79146326 + ------------------------------------------------------------------------------------- + TOTAL 0.04462193 80.14676276 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 370920 +BPFP 1.2263 bits/point +EBPFP 2.4527 equivalent bits/point +MSE 80.146763 +---------------------- -------------------------------------------------------- +Time: 4.551s Load: 0.010s, Pack+Encode: 2.548s, Decode+Unpack: 1.993s +---------------------- -------------------------------------------------------- +💾 Converting with 80.1468 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,408B, BPFP=0.4995 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,880B, BPFP=2.6069 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,500B, BPFP=1.2773 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,448B, BPFP=2.5219 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,288B, BPFP=1.3836 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,492B, BPFP=2.5245 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,044B, BPFP=1.3691 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,540B, BPFP=2.5273 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,620B, BPFP=2.1756 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,100B, BPFP=2.5012 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,408B, BPFP=0.3175 +⌛️ [2/4] FRONTEND: Frontend time: 2.223s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.937s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12547956 41.39410795 + layer.0.v_cache 0.00001492 0.00546161 + layer.1.k_cache 0.50446328 8.11194637 + layer.1.v_cache 0.00000633 0.00225169 + layer.2.k_cache 0.00671430 0.95522804 + layer.2.v_cache 0.00001992 0.00706131 + layer.3.k_cache 0.02142675 3.90846026 + layer.3.v_cache 0.00001986 0.00760318 + layer.4.k_cache 0.00069859 0.20254279 + layer.4.v_cache 0.00005414 0.01484387 + layer.4.output 0.01038299 202.42933528 + ------------------------------------------------------------------------------------- + TOTAL 0.04303403 86.56557965 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 363728 +BPFP 1.2711 bits/point +EBPFP 2.5423 equivalent bits/point +MSE 86.565580 +---------------------- -------------------------------------------------------- +Time: 4.174s Load: 0.014s, Pack+Encode: 2.223s, Decode+Unpack: 1.937s +---------------------- -------------------------------------------------------- +💾 Converting with 86.5656 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,400B, BPFP=0.4880 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,152B, BPFP=2.3439 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,056B, BPFP=1.2488 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,124B, BPFP=2.2905 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,164B, BPFP=1.5139 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,784B, BPFP=2.2728 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,500B, BPFP=1.3756 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,664B, BPFP=2.3185 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,760B, BPFP=2.0120 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,576B, BPFP=2.2620 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,588B, BPFP=0.2936 +⌛️ [2/4] FRONTEND: Frontend time: 3.084s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.915s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13970643 41.25898048 + layer.0.v_cache 0.00001441 0.00541188 + layer.1.k_cache 0.60015088 8.33004649 + layer.1.v_cache 0.00000584 0.00226475 + layer.2.k_cache 0.00648033 0.86191242 + layer.2.v_cache 0.00001840 0.00694473 + layer.3.k_cache 0.03288257 5.07750410 + layer.3.v_cache 0.00002024 0.00748257 + layer.4.k_cache 0.00070287 0.19885734 + layer.4.v_cache 0.00005060 0.01434538 + layer.4.output 0.04874230 170.26699692 + ------------------------------------------------------------------------------------- + TOTAL 0.06595463 73.39016050 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 388768 +BPFP 1.1871 bits/point +EBPFP 2.3742 equivalent bits/point +MSE 73.390161 +---------------------- -------------------------------------------------------- +Time: 5.010s Load: 0.011s, Pack+Encode: 3.084s, Decode+Unpack: 1.915s +---------------------- -------------------------------------------------------- +💾 Converting with 73.3902 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,264B, BPFP=0.4904 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 52,264B, BPFP=2.4973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,264B, BPFP=1.3028 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,692B, BPFP=2.4222 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,100B, BPFP=1.3905 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 50,552B, BPFP=2.4155 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,172B, BPFP=1.3461 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 51,652B, BPFP=2.4681 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,816B, BPFP=2.0937 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 50,844B, BPFP=2.4295 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,464B, BPFP=0.2830 +⌛️ [2/4] FRONTEND: Frontend time: 2.270s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.798s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13960611 40.51686138 + layer.0.v_cache 0.00001491 0.00547515 + layer.1.k_cache 0.68890666 8.69795997 + layer.1.v_cache 0.00000583 0.00224794 + layer.2.k_cache 0.00940153 0.86341321 + layer.2.v_cache 0.00001882 0.00689020 + layer.3.k_cache 0.04667800 5.93524226 + layer.3.v_cache 0.00001939 0.00758523 + layer.4.k_cache 0.00070827 0.19814662 + layer.4.v_cache 0.00005291 0.01410689 + layer.4.output 0.04427218 159.94211446 + ------------------------------------------------------------------------------------- + TOTAL 0.07031281 69.16721942 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 436084 +BPFP 1.2257 bits/point +EBPFP 2.4515 equivalent bits/point +MSE 69.167219 +---------------------- -------------------------------------------------------- +Time: 4.078s Load: 0.010s, Pack+Encode: 2.270s, Decode+Unpack: 1.798s +---------------------- -------------------------------------------------------- +💾 Converting with 69.1672 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,460B, BPFP=0.4944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,924B, BPFP=2.3476 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,508B, BPFP=1.2807 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,776B, BPFP=2.2876 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,964B, BPFP=1.5136 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,436B, BPFP=2.2699 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,000B, BPFP=1.4110 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,160B, BPFP=2.3077 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,520B, BPFP=2.0130 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,320B, BPFP=2.2638 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,364B, BPFP=0.3013 +⌛️ [2/4] FRONTEND: Frontend time: 2.400s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.643s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15817952 40.21471311 + layer.0.v_cache 0.00001474 0.00532769 + layer.1.k_cache 0.61609703 8.49902507 + layer.1.v_cache 0.00000566 0.00218018 + layer.2.k_cache 0.00844678 0.97650841 + layer.2.v_cache 0.00001876 0.00665913 + layer.3.k_cache 0.04682518 5.53505173 + layer.3.v_cache 0.00001896 0.00749939 + layer.4.k_cache 0.00069859 0.20183640 + layer.4.v_cache 0.00005216 0.01475290 + layer.4.output 0.04649196 174.41482023 + ------------------------------------------------------------------------------------- + TOTAL 0.06798830 75.08042916 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 388432 +BPFP 1.1940 bits/point +EBPFP 2.3881 equivalent bits/point +MSE 75.080429 +---------------------- -------------------------------------------------------- +Time: 4.056s Load: 0.013s, Pack+Encode: 2.400s, Decode+Unpack: 1.643s +---------------------- -------------------------------------------------------- +💾 Converting with 75.0804 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,864B, BPFP=0.4826 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,176B, BPFP=2.4595 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,260B, BPFP=1.3208 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,348B, BPFP=2.4144 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,824B, BPFP=1.5693 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,652B, BPFP=2.3765 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,484B, BPFP=1.3330 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,584B, BPFP=2.4273 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,684B, BPFP=2.1061 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,584B, BPFP=2.3728 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,268B, BPFP=0.2821 +⌛️ [2/4] FRONTEND: Frontend time: 2.234s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.871s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14002854 39.89107755 + layer.0.v_cache 0.00001601 0.00547585 + layer.1.k_cache 0.54217949 8.59032438 + layer.1.v_cache 0.00000579 0.00224293 + layer.2.k_cache 0.01082110 1.05173740 + layer.2.v_cache 0.00001971 0.00681238 + layer.3.k_cache 0.03102738 6.07608341 + layer.3.v_cache 0.00001907 0.00747452 + layer.4.k_cache 0.00066782 0.20326026 + layer.4.v_cache 0.00005141 0.01460840 + layer.4.output 0.00856273 179.62095570 + ------------------------------------------------------------------------------------- + TOTAL 0.04616326 77.24681100 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 382728 +BPFP 1.2257 bits/point +EBPFP 2.4514 equivalent bits/point +MSE 77.246811 +---------------------- -------------------------------------------------------- +Time: 4.116s Load: 0.010s, Pack+Encode: 2.234s, Decode+Unpack: 1.871s +---------------------- -------------------------------------------------------- +💾 Converting with 77.2468 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,344B, BPFP=0.4916 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,888B, BPFP=2.3615 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,664B, BPFP=1.2976 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,632B, BPFP=2.2955 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,152B, BPFP=1.5337 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,484B, BPFP=2.2877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,216B, BPFP=1.3792 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,376B, BPFP=2.3346 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,528B, BPFP=2.0269 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,296B, BPFP=2.2778 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,888B, BPFP=0.2923 +⌛️ [2/4] FRONTEND: Frontend time: 2.241s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.976s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14705232 41.70003748 + layer.0.v_cache 0.00001374 0.00541010 + layer.1.k_cache 0.65666707 8.84293620 + layer.1.v_cache 0.00000579 0.00223261 + layer.2.k_cache 0.01062412 0.90319660 + layer.2.v_cache 0.00001918 0.00678055 + layer.3.k_cache 0.03205616 5.62239296 + layer.3.v_cache 0.00001999 0.00766109 + layer.4.k_cache 0.00071739 0.20383929 + layer.4.v_cache 0.00005111 0.01371333 + layer.4.output 0.04976816 165.99410774 + ------------------------------------------------------------------------------------- + TOTAL 0.07032965 71.72158555 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 386468 +BPFP 1.1960 bits/point +EBPFP 2.3920 equivalent bits/point +MSE 71.721586 +---------------------- -------------------------------------------------------- +Time: 4.227s Load: 0.010s, Pack+Encode: 2.241s, Decode+Unpack: 1.976s +---------------------- -------------------------------------------------------- +💾 Converting with 71.7216 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,812B, BPFP=0.4900 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,884B, BPFP=2.4958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,140B, BPFP=1.2867 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,668B, BPFP=2.4282 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,124B, BPFP=1.4526 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,392B, BPFP=2.4128 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,940B, BPFP=1.3312 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,068B, BPFP=2.4504 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,444B, BPFP=2.0821 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,476B, BPFP=2.4175 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,108B, BPFP=0.3027 +⌛️ [2/4] FRONTEND: Frontend time: 2.639s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.664s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131647 40.01236516 + layer.0.v_cache 0.00001392 0.00521386 + layer.1.k_cache 0.51967005 8.06441403 + layer.1.v_cache 0.00000557 0.00209770 + layer.2.k_cache 0.01013049 0.92684339 + layer.2.v_cache 0.00001908 0.00641916 + layer.3.k_cache 0.05186962 5.92503884 + layer.3.v_cache 0.00001891 0.00682996 + layer.4.k_cache 0.00069958 0.18118792 + layer.4.v_cache 0.00005017 0.01353746 + layer.4.output 0.01073789 186.92934990 + ------------------------------------------------------------------------------------- + TOTAL 0.04758583 80.21467040 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 377056 +BPFP 1.2333 bits/point +EBPFP 2.4666 equivalent bits/point +MSE 80.214670 +---------------------- -------------------------------------------------------- +Time: 4.315s Load: 0.012s, Pack+Encode: 2.639s, Decode+Unpack: 1.664s +---------------------- -------------------------------------------------------- +💾 Converting with 80.2147 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,824B, BPFP=0.4787 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,892B, BPFP=2.4355 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,248B, BPFP=1.2613 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,932B, BPFP=2.3835 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,424B, BPFP=1.4878 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,316B, BPFP=2.3500 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,492B, BPFP=1.3830 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,444B, BPFP=2.4112 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,384B, BPFP=2.0825 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,628B, BPFP=2.3670 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,200B, BPFP=0.2961 +⌛️ [2/4] FRONTEND: Frontend time: 2.201s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.642s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15566496 41.49299113 + layer.0.v_cache 0.00001437 0.00525412 + layer.1.k_cache 0.62643353 8.77038744 + layer.1.v_cache 0.00000573 0.00218201 + layer.2.k_cache 0.01010078 1.01072905 + layer.2.v_cache 0.00001868 0.00684401 + layer.3.k_cache 0.06271097 5.74306700 + layer.3.v_cache 0.00001958 0.00742172 + layer.4.k_cache 0.00073784 0.19517083 + layer.4.v_cache 0.00005294 0.01421635 + layer.4.output 0.00677609 183.80371094 + ------------------------------------------------------------------------------------- + TOTAL 0.05312894 79.05142589 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 381784 +BPFP 1.2184 bits/point +EBPFP 2.4368 equivalent bits/point +MSE 79.051426 +---------------------- -------------------------------------------------------- +Time: 3.853s Load: 0.011s, Pack+Encode: 2.201s, Decode+Unpack: 1.642s +---------------------- -------------------------------------------------------- +💾 Converting with 79.0514 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,832B, BPFP=0.4825 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,932B, BPFP=2.4548 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,712B, BPFP=1.2955 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,952B, BPFP=2.4012 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,716B, BPFP=1.4596 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,560B, BPFP=2.3798 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,048B, BPFP=1.3684 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,316B, BPFP=2.4211 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,656B, BPFP=2.1119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,576B, BPFP=2.3807 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,324B, BPFP=0.2835 +⌛️ [2/4] FRONTEND: Frontend time: 2.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.537s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13918158 39.68223476 + layer.0.v_cache 0.00001566 0.00549050 + layer.1.k_cache 0.57997750 8.72969450 + layer.1.v_cache 0.00000571 0.00227736 + layer.2.k_cache 0.01637170 1.11617327 + layer.2.v_cache 0.00001887 0.00704106 + layer.3.k_cache 0.04393929 4.96432773 + layer.3.v_cache 0.00001903 0.00781385 + layer.4.k_cache 0.00071104 0.19531006 + layer.4.v_cache 0.00005175 0.01460594 + layer.4.output 0.01065216 178.73239261 + ------------------------------------------------------------------------------------- + TOTAL 0.05028572 76.81480690 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 379624 +BPFP 1.2200 bits/point +EBPFP 2.4400 equivalent bits/point +MSE 76.814807 +---------------------- -------------------------------------------------------- +Time: 3.775s Load: 0.011s, Pack+Encode: 2.227s, Decode+Unpack: 1.537s +---------------------- -------------------------------------------------------- +💾 Converting with 76.8148 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,624B, BPFP=0.4851 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,104B, BPFP=2.2734 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,432B, BPFP=1.2819 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,920B, BPFP=2.2137 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,448B, BPFP=1.4339 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,584B, BPFP=2.1968 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,936B, BPFP=1.3577 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,276B, BPFP=2.2317 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,868B, BPFP=1.9591 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,532B, BPFP=2.1942 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,328B, BPFP=0.2832 +⌛️ [2/4] FRONTEND: Frontend time: 2.390s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.167s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16496690 40.82666961 + layer.0.v_cache 0.00001433 0.00544779 + layer.1.k_cache 0.65265572 8.12556546 + layer.1.v_cache 0.00000578 0.00229018 + layer.2.k_cache 0.01241827 0.98081734 + layer.2.v_cache 0.00001920 0.00672813 + layer.3.k_cache 0.08444468 4.96433854 + layer.3.v_cache 0.00001975 0.00791596 + layer.4.k_cache 0.00068928 0.20560896 + layer.4.v_cache 0.00005259 0.01475437 + layer.4.output 0.04785494 161.91624424 + ------------------------------------------------------------------------------------- + TOTAL 0.07354536 69.91493212 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 389052 +BPFP 1.1535 bits/point +EBPFP 2.3070 equivalent bits/point +MSE 69.914932 +---------------------- -------------------------------------------------------- +Time: 4.569s Load: 0.012s, Pack+Encode: 2.390s, Decode+Unpack: 2.167s +---------------------- -------------------------------------------------------- +💾 Converting with 69.9149 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,252B, BPFP=0.4917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,972B, BPFP=2.3901 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,756B, BPFP=1.3688 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,724B, BPFP=2.3238 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,684B, BPFP=1.5244 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,308B, BPFP=2.3017 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,760B, BPFP=1.3690 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,104B, BPFP=2.3440 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,352B, BPFP=2.0383 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,284B, BPFP=2.3004 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,848B, BPFP=0.2949 +⌛️ [2/4] FRONTEND: Frontend time: 2.147s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.649s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15001445 40.51426977 + layer.0.v_cache 0.00001441 0.00559316 + layer.1.k_cache 0.57475359 8.51113082 + layer.1.v_cache 0.00000584 0.00230798 + layer.2.k_cache 0.01434258 0.98831654 + layer.2.v_cache 0.00001954 0.00692242 + layer.3.k_cache 0.02649183 5.28347384 + layer.3.v_cache 0.00001945 0.00761492 + layer.4.k_cache 0.00069414 0.20104124 + layer.4.v_cache 0.00005306 0.01464220 + layer.4.output 0.05035525 174.34625547 + ------------------------------------------------------------------------------------- + TOTAL 0.06581739 75.05641771 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 386044 +BPFP 1.2069 bits/point +EBPFP 2.4137 equivalent bits/point +MSE 75.056418 +---------------------- -------------------------------------------------------- +Time: 3.807s Load: 0.011s, Pack+Encode: 2.147s, Decode+Unpack: 1.649s +---------------------- -------------------------------------------------------- +💾 Converting with 75.0564 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,860B, BPFP=0.4909 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,888B, BPFP=2.4871 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,612B, BPFP=1.3083 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,832B, BPFP=2.4286 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,516B, BPFP=1.4692 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,464B, BPFP=2.4082 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,872B, BPFP=1.3227 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,172B, BPFP=2.4475 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,720B, BPFP=2.1454 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,644B, BPFP=2.4182 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,624B, BPFP=0.2741 +⌛️ [2/4] FRONTEND: Frontend time: 2.360s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.634s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11854591 40.39123795 + layer.0.v_cache 0.00001566 0.00535486 + layer.1.k_cache 0.56985614 8.51814259 + layer.1.v_cache 0.00000555 0.00220794 + layer.2.k_cache 0.01267573 0.94183891 + layer.2.v_cache 0.00001902 0.00684181 + layer.3.k_cache 0.04477839 5.23364777 + layer.3.v_cache 0.00001896 0.00764192 + layer.4.k_cache 0.00071754 0.21006851 + layer.4.v_cache 0.00005321 0.01448760 + layer.4.output 0.00761333 185.51283878 + ------------------------------------------------------------------------------------- + TOTAL 0.04705761 79.64243184 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 376204 +BPFP 1.2262 bits/point +EBPFP 2.4523 equivalent bits/point +MSE 79.642432 +---------------------- -------------------------------------------------------- +Time: 4.004s Load: 0.010s, Pack+Encode: 2.360s, Decode+Unpack: 1.634s +---------------------- -------------------------------------------------------- +💾 Converting with 79.6424 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,292B, BPFP=0.4938 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,132B, BPFP=2.3986 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,796B, BPFP=1.3178 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,960B, BPFP=2.3363 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,196B, BPFP=1.4985 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,820B, BPFP=2.3289 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,328B, BPFP=1.3992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,416B, BPFP=2.3605 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,776B, BPFP=2.0608 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,676B, BPFP=2.3212 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,464B, BPFP=0.2844 +⌛️ [2/4] FRONTEND: Frontend time: 2.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.678s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582189 40.08720637 + layer.0.v_cache 0.00001509 0.00562755 + layer.1.k_cache 0.58311327 8.47407791 + layer.1.v_cache 0.00000606 0.00225218 + layer.2.k_cache 0.00657600 0.90057383 + layer.2.v_cache 0.00001909 0.00706624 + layer.3.k_cache 0.02234380 6.05887194 + layer.3.v_cache 0.00001905 0.00753200 + layer.4.k_cache 0.00071828 0.21908780 + layer.4.v_cache 0.00005370 0.01406083 + layer.4.output 0.05044473 169.74547498 + ------------------------------------------------------------------------------------- + TOTAL 0.06363526 73.17615773 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 385856 +BPFP 1.2063 bits/point +EBPFP 2.4126 equivalent bits/point +MSE 73.176158 +---------------------- -------------------------------------------------------- +Time: 3.945s Load: 0.010s, Pack+Encode: 2.257s, Decode+Unpack: 1.678s +---------------------- -------------------------------------------------------- +💾 Converting with 73.1762 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 292, 128) +Output shape: (1, 292, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.output: torch.Size([1, 292, 3584]) -> torch.Size([1, 1, 292, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,116B, BPFP=0.4878 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,980B, BPFP=2.4069 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,436B, BPFP=1.3611 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,964B, BPFP=2.3525 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,256B, BPFP=1.5120 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,552B, BPFP=2.3305 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,116B, BPFP=1.3440 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,548B, BPFP=2.3838 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,416B, BPFP=2.0557 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,556B, BPFP=2.3307 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,444B, BPFP=0.2786 +⌛️ [2/4] FRONTEND: Frontend time: 2.206s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.758s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887579 40.75246147 + layer.0.v_cache 0.00001384 0.00539846 + layer.1.k_cache 0.57277983 8.69003421 + layer.1.v_cache 0.00000568 0.00222425 + layer.2.k_cache 0.01314942 0.92952634 + layer.2.v_cache 0.00002105 0.00683066 + layer.3.k_cache 0.03125376 5.47902899 + layer.3.v_cache 0.00001978 0.00739011 + layer.4.k_cache 0.00068035 0.19485462 + layer.4.v_cache 0.00005123 0.01417261 + layer.4.output 0.04859251 180.09923863 + ------------------------------------------------------------------------------------- + TOTAL 0.06511755 77.45744659 + (elements=2,541,568) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2541568 +Total Bytes 383384 +BPFP 1.2068 bits/point +EBPFP 2.4135 equivalent bits/point +MSE 77.457447 +---------------------- -------------------------------------------------------- +Time: 3.975s Load: 0.011s, Pack+Encode: 2.206s, Decode+Unpack: 1.758s +---------------------- -------------------------------------------------------- +💾 Converting with 77.4574 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.024s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 338, 128) +Output shape: (1, 338, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.output: torch.Size([1, 338, 3584]) -> torch.Size([1, 1, 338, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,520B, BPFP=0.4863 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 53,464B, BPFP=2.4715 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,372B, BPFP=1.2653 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 52,296B, BPFP=2.4175 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,336B, BPFP=1.4486 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 51,668B, BPFP=2.3885 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,732B, BPFP=1.2820 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 52,692B, BPFP=2.4358 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,580B, BPFP=2.1071 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 51,644B, BPFP=2.3874 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 45,148B, BPFP=0.2982 +⌛️ [2/4] FRONTEND: Frontend time: 3.098s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.951s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13976314 40.44742164 + layer.0.v_cache 0.00001472 0.00546617 + layer.1.k_cache 0.73750278 9.12305915 + layer.1.v_cache 0.00000589 0.00224459 + layer.2.k_cache 0.02011468 0.86673073 + layer.2.v_cache 0.00002041 0.00655566 + layer.3.k_cache 0.02722352 5.40003994 + layer.3.v_cache 0.00002029 0.00717290 + layer.4.k_cache 0.00072609 0.20551657 + layer.4.v_cache 0.00005477 0.01411689 + layer.4.output 0.04270584 147.93921703 + ------------------------------------------------------------------------------------- + TOTAL 0.07202277 64.21487315 + (elements=2,941,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2941952 +Total Bytes 449452 +BPFP 1.2222 bits/point +EBPFP 2.4444 equivalent bits/point +MSE 64.214873 +---------------------- -------------------------------------------------------- +Time: 5.073s Load: 0.024s, Pack+Encode: 3.098s, Decode+Unpack: 1.951s +---------------------- -------------------------------------------------------- +💾 Converting with 64.2149 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,080B, BPFP=0.4875 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,812B, BPFP=2.4061 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,408B, BPFP=1.3643 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,724B, BPFP=2.3477 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,984B, BPFP=1.5563 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,084B, BPFP=2.3134 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,520B, BPFP=1.3166 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,020B, BPFP=2.3636 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,332B, BPFP=2.0582 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,196B, BPFP=2.3194 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,560B, BPFP=0.2804 +⌛️ [2/4] FRONTEND: Frontend time: 2.522s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.659s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150811 40.38847052 + layer.0.v_cache 0.00001374 0.00540674 + layer.1.k_cache 0.66021251 8.56883172 + layer.1.v_cache 0.00000574 0.00228873 + layer.2.k_cache 0.00963001 0.87854161 + layer.2.v_cache 0.00001988 0.00674605 + layer.3.k_cache 0.08055985 5.96383363 + layer.3.v_cache 0.00001942 0.00742965 + layer.4.k_cache 0.00069026 0.19870887 + layer.4.v_cache 0.00005126 0.01512264 + layer.4.output 0.05093874 181.87426362 + ------------------------------------------------------------------------------------- + TOTAL 0.07289894 78.18560150 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 381720 +BPFP 1.2057 bits/point +EBPFP 2.4113 equivalent bits/point +MSE 78.185602 +---------------------- -------------------------------------------------------- +Time: 4.193s Load: 0.012s, Pack+Encode: 2.522s, Decode+Unpack: 1.659s +---------------------- -------------------------------------------------------- +💾 Converting with 78.1856 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,456B, BPFP=0.4876 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,800B, BPFP=2.3102 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,540B, BPFP=1.2655 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,696B, BPFP=2.2533 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,796B, BPFP=1.4334 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,224B, BPFP=2.2290 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,036B, BPFP=1.3426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,244B, BPFP=2.2816 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,604B, BPFP=1.9907 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,264B, BPFP=2.2310 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,424B, BPFP=0.2904 +⌛️ [2/4] FRONTEND: Frontend time: 2.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.678s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14893842 41.33740396 + layer.0.v_cache 0.00001415 0.00531021 + layer.1.k_cache 0.65110169 8.82213993 + layer.1.v_cache 0.00000575 0.00225745 + layer.2.k_cache 0.01411552 0.90908320 + layer.2.v_cache 0.00001904 0.00681800 + layer.3.k_cache 0.03939451 5.93411104 + layer.3.v_cache 0.00001876 0.00733390 + layer.4.k_cache 0.00070558 0.20142298 + layer.4.v_cache 0.00005345 0.01478187 + layer.4.output 0.04819407 172.99848244 + ------------------------------------------------------------------------------------- + TOTAL 0.07010149 74.60176703 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 385084 +BPFP 1.1681 bits/point +EBPFP 2.3362 equivalent bits/point +MSE 74.601767 +---------------------- -------------------------------------------------------- +Time: 3.854s Load: 0.011s, Pack+Encode: 2.165s, Decode+Unpack: 1.678s +---------------------- -------------------------------------------------------- +💾 Converting with 74.6018 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,368B, BPFP=0.4971 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,560B, BPFP=2.5879 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,580B, BPFP=1.2821 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,080B, BPFP=2.5000 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,372B, BPFP=1.3885 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,152B, BPFP=2.5043 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,000B, BPFP=1.3664 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,636B, BPFP=2.5330 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,192B, BPFP=2.1502 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,700B, BPFP=2.4774 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,084B, BPFP=0.2808 +⌛️ [2/4] FRONTEND: Frontend time: 2.212s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.697s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14492303 41.47340631 + layer.0.v_cache 0.00001375 0.00535924 + layer.1.k_cache 0.50395557 8.49395682 + layer.1.v_cache 0.00000564 0.00221591 + layer.2.k_cache 0.00856250 0.93236179 + layer.2.v_cache 0.00001902 0.00707224 + layer.3.k_cache 0.02692018 4.28362086 + layer.3.v_cache 0.00002118 0.00768997 + layer.4.k_cache 0.00070790 0.19582258 + layer.4.v_cache 0.00005130 0.01450729 + layer.4.output 0.01131245 204.59784085 + ------------------------------------------------------------------------------------- + TOTAL 0.04496278 87.50593523 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 357724 +BPFP 1.2502 bits/point +EBPFP 2.5003 equivalent bits/point +MSE 87.505935 +---------------------- -------------------------------------------------------- +Time: 3.922s Load: 0.013s, Pack+Encode: 2.212s, Decode+Unpack: 1.697s +---------------------- -------------------------------------------------------- +💾 Converting with 87.5059 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,692B, BPFP=0.4921 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,560B, BPFP=2.5226 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,256B, BPFP=1.2600 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,280B, BPFP=2.4502 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,384B, BPFP=1.4937 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,024B, BPFP=2.4357 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,788B, BPFP=1.3467 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,556B, BPFP=2.4658 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,968B, BPFP=2.1495 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,856B, BPFP=2.4262 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,456B, BPFP=0.2948 +⌛️ [2/4] FRONTEND: Frontend time: 2.231s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.634s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14113761 41.39172894 + layer.0.v_cache 0.00001369 0.00523474 + layer.1.k_cache 0.59212339 8.06414264 + layer.1.v_cache 0.00000552 0.00211430 + layer.2.k_cache 0.01129283 0.98628721 + layer.2.v_cache 0.00001853 0.00666900 + layer.3.k_cache 0.01358872 4.05420895 + layer.3.v_cache 0.00001883 0.00721958 + layer.4.k_cache 0.00070763 0.18983918 + layer.4.v_cache 0.00005153 0.01362136 + layer.4.output 0.01019159 198.37700569 + ------------------------------------------------------------------------------------- + TOTAL 0.04884114 84.90353563 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 372820 +BPFP 1.2415 bits/point +EBPFP 2.4831 equivalent bits/point +MSE 84.903536 +---------------------- -------------------------------------------------------- +Time: 3.874s Load: 0.009s, Pack+Encode: 2.231s, Decode+Unpack: 1.634s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9035 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,848B, BPFP=0.4817 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,020B, BPFP=2.4510 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,540B, BPFP=1.2816 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,924B, BPFP=2.3913 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,136B, BPFP=1.5318 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,532B, BPFP=2.3700 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,224B, BPFP=1.3733 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,328B, BPFP=2.4133 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,456B, BPFP=2.0936 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,512B, BPFP=2.3689 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,708B, BPFP=0.3011 +⌛️ [2/4] FRONTEND: Frontend time: 2.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.767s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11028925 40.14428626 + layer.0.v_cache 0.00001466 0.00540179 + layer.1.k_cache 0.56032453 8.25575134 + layer.1.v_cache 0.00000595 0.00219894 + layer.2.k_cache 0.00809042 0.96571164 + layer.2.v_cache 0.00001929 0.00682096 + layer.3.k_cache 0.03501172 5.76436432 + layer.3.v_cache 0.00001861 0.00716861 + layer.4.k_cache 0.00070074 0.20223949 + layer.4.v_cache 0.00005246 0.01389388 + layer.4.output 0.00889761 183.96532790 + ------------------------------------------------------------------------------------- + TOTAL 0.04569476 79.00736074 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 383228 +BPFP 1.2273 bits/point +EBPFP 2.4546 equivalent bits/point +MSE 79.007361 +---------------------- -------------------------------------------------------- +Time: 3.940s Load: 0.009s, Pack+Encode: 2.164s, Decode+Unpack: 1.767s +---------------------- -------------------------------------------------------- +💾 Converting with 79.0074 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 311, 128) +Output shape: (1, 311, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.output: torch.Size([1, 311, 3584]) -> torch.Size([1, 1, 311, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,600B, BPFP=0.4823 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,076B, BPFP=2.2647 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,620B, BPFP=1.2872 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,884B, BPFP=2.2048 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,796B, BPFP=1.4467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,556B, BPFP=2.1883 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,760B, BPFP=1.3947 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,700B, BPFP=2.2458 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,740B, BPFP=1.9463 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,556B, BPFP=2.1883 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,408B, BPFP=0.2828 +⌛️ [2/4] FRONTEND: Frontend time: 2.219s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.629s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14227538 41.04162480 + layer.0.v_cache 0.00001500 0.00531455 + layer.1.k_cache 0.64922478 8.59554612 + layer.1.v_cache 0.00000572 0.00224668 + layer.2.k_cache 0.00828172 0.98584877 + layer.2.v_cache 0.00001922 0.00665826 + layer.3.k_cache 0.02636586 5.71278504 + layer.3.v_cache 0.00001996 0.00734462 + layer.4.k_cache 0.00069494 0.20486621 + layer.4.v_cache 0.00005248 0.01496039 + layer.4.output 0.04686997 165.76220142 + ------------------------------------------------------------------------------------- + TOTAL 0.06794382 71.58309444 + (elements=2,706,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2706944 +Total Bytes 390696 +BPFP 1.1546 bits/point +EBPFP 2.3093 equivalent bits/point +MSE 71.583094 +---------------------- -------------------------------------------------------- +Time: 3.858s Load: 0.010s, Pack+Encode: 2.219s, Decode+Unpack: 1.629s +---------------------- -------------------------------------------------------- +💾 Converting with 71.5831 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,744B, BPFP=0.4745 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,088B, BPFP=2.2113 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,440B, BPFP=1.3137 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,208B, BPFP=2.1574 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,308B, BPFP=1.4282 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,956B, BPFP=2.1419 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,300B, BPFP=1.4277 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,684B, BPFP=2.1865 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,896B, BPFP=1.8931 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,188B, BPFP=2.1561 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,456B, BPFP=0.2929 +⌛️ [2/4] FRONTEND: Frontend time: 2.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.584s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13149661 41.43736213 + layer.0.v_cache 0.00001370 0.00519956 + layer.1.k_cache 0.50214604 8.98849571 + layer.1.v_cache 0.00000625 0.00222616 + layer.2.k_cache 0.00807812 0.99204885 + layer.2.v_cache 0.00001969 0.00682032 + layer.3.k_cache 0.06061829 4.65175398 + layer.3.v_cache 0.00001868 0.00768792 + layer.4.k_cache 0.00068906 0.20994976 + layer.4.v_cache 0.00005044 0.01354661 + layer.4.output 1.20679682 200.25614496 + ------------------------------------------------------------------------------------- + TOTAL 0.53827733 85.77106504 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 317268 +BPFP 1.1436 bits/point +EBPFP 2.2871 equivalent bits/point +MSE 85.771065 +---------------------- -------------------------------------------------------- +Time: 4.099s Load: 0.011s, Pack+Encode: 2.503s, Decode+Unpack: 1.584s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7711 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 300, 128) +Output shape: (1, 300, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.output: torch.Size([1, 300, 3584]) -> torch.Size([1, 1, 300, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,476B, BPFP=0.4935 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,092B, BPFP=2.3485 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,304B, BPFP=1.2658 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,896B, BPFP=2.2862 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,064B, BPFP=1.5137 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,668B, BPFP=2.2744 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,084B, BPFP=1.3585 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,744B, BPFP=2.3304 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,556B, BPFP=2.0081 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,580B, BPFP=2.2698 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,824B, BPFP=0.2889 +⌛️ [2/4] FRONTEND: Frontend time: 2.188s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.754s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13491969 41.18836263 + layer.0.v_cache 0.00001558 0.00517324 + layer.1.k_cache 0.64590566 8.59165771 + layer.1.v_cache 0.00000556 0.00210500 + layer.2.k_cache 0.01395306 0.91604980 + layer.2.v_cache 0.00001972 0.00648059 + layer.3.k_cache 0.02379198 6.26305461 + layer.3.v_cache 0.00001942 0.00723220 + layer.4.k_cache 0.00073123 0.18985149 + layer.4.v_cache 0.00005274 0.01405963 + layer.4.output 0.05011183 171.83976190 + ------------------------------------------------------------------------------------- + TOTAL 0.06883514 74.12131531 + (elements=2,611,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2611200 +Total Bytes 387288 +BPFP 1.1865 bits/point +EBPFP 2.3731 equivalent bits/point +MSE 74.121315 +---------------------- -------------------------------------------------------- +Time: 3.952s Load: 0.010s, Pack+Encode: 2.188s, Decode+Unpack: 1.754s +---------------------- -------------------------------------------------------- +💾 Converting with 74.1213 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,068B, BPFP=0.4869 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,800B, BPFP=2.4055 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,188B, BPFP=1.3524 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,820B, BPFP=2.3529 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,056B, BPFP=1.5064 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,240B, BPFP=2.3217 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,000B, BPFP=1.3424 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,064B, BPFP=2.3660 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,456B, BPFP=2.0649 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,220B, BPFP=2.3207 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,500B, BPFP=0.2800 +⌛️ [2/4] FRONTEND: Frontend time: 2.462s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.743s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14826349 40.23448910 + layer.0.v_cache 0.00001415 0.00540679 + layer.1.k_cache 0.61440683 8.57006081 + layer.1.v_cache 0.00000562 0.00228854 + layer.2.k_cache 0.00551110 0.94217622 + layer.2.v_cache 0.00001912 0.00675697 + layer.3.k_cache 0.01294636 5.81543850 + layer.3.v_cache 0.00002032 0.00744953 + layer.4.k_cache 0.00070633 0.20070887 + layer.4.v_cache 0.00005074 0.01445817 + layer.4.output 0.04972342 182.45064740 + ------------------------------------------------------------------------------------- + TOTAL 0.06647106 78.40904502 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 381412 +BPFP 1.2047 bits/point +EBPFP 2.4094 equivalent bits/point +MSE 78.409045 +---------------------- -------------------------------------------------------- +Time: 4.214s Load: 0.010s, Pack+Encode: 2.462s, Decode+Unpack: 1.743s +---------------------- -------------------------------------------------------- +💾 Converting with 78.4090 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,668B, BPFP=0.4889 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,564B, BPFP=2.5138 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,084B, BPFP=1.2457 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,504B, BPFP=2.4540 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,176B, BPFP=1.4201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,072B, BPFP=2.4296 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,612B, BPFP=1.3319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,060B, BPFP=2.4853 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,192B, BPFP=2.0979 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,908B, BPFP=2.4204 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,852B, BPFP=0.2808 +⌛️ [2/4] FRONTEND: Frontend time: 2.673s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.824s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13861307 41.84731075 + layer.0.v_cache 0.00001419 0.00534844 + layer.1.k_cache 0.53867822 8.26585592 + layer.1.v_cache 0.00000573 0.00219140 + layer.2.k_cache 0.00871976 0.93713676 + layer.2.v_cache 0.00001877 0.00690051 + layer.3.k_cache 0.02195022 4.94744146 + layer.3.v_cache 0.00001967 0.00744517 + layer.4.k_cache 0.00069227 0.18517522 + layer.4.v_cache 0.00004961 0.01357039 + layer.4.output 0.00954030 183.35222731 + ------------------------------------------------------------------------------------- + TOTAL 0.04562021 78.80493925 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 369692 +BPFP 1.2267 bits/point +EBPFP 2.4534 equivalent bits/point +MSE 78.804939 +---------------------- -------------------------------------------------------- +Time: 4.508s Load: 0.011s, Pack+Encode: 2.673s, Decode+Unpack: 1.824s +---------------------- -------------------------------------------------------- +💾 Converting with 78.8049 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,784B, BPFP=0.4799 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,904B, BPFP=2.4532 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,832B, BPFP=1.2474 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,936B, BPFP=2.4003 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,296B, BPFP=1.4366 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,520B, BPFP=2.3776 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,620B, BPFP=1.3451 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,492B, BPFP=2.4307 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,160B, BPFP=2.0848 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,560B, BPFP=2.3798 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,920B, BPFP=0.2803 +⌛️ [2/4] FRONTEND: Frontend time: 2.269s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.744s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13928204 39.60005327 + layer.0.v_cache 0.00001373 0.00536771 + layer.1.k_cache 0.60860016 8.33417644 + layer.1.v_cache 0.00000555 0.00219036 + layer.2.k_cache 0.01026906 0.95660027 + layer.2.v_cache 0.00001872 0.00684576 + layer.3.k_cache 0.05166252 4.51622169 + layer.3.v_cache 0.00001937 0.00770825 + layer.4.k_cache 0.00073448 0.19257492 + layer.4.v_cache 0.00005026 0.01443575 + layer.4.output 0.00697322 181.81070492 + ------------------------------------------------------------------------------------- + TOTAL 0.05055697 78.01830052 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 377024 +BPFP 1.2116 bits/point +EBPFP 2.4233 equivalent bits/point +MSE 78.018301 +---------------------- -------------------------------------------------------- +Time: 4.023s Load: 0.010s, Pack+Encode: 2.269s, Decode+Unpack: 1.744s +---------------------- -------------------------------------------------------- +💾 Converting with 78.0183 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,320B, BPFP=0.4936 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,960B, BPFP=2.3814 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,508B, BPFP=1.2981 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,604B, BPFP=2.3095 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,500B, BPFP=1.5095 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,340B, BPFP=2.2956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,752B, BPFP=1.3640 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,324B, BPFP=2.3477 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,544B, BPFP=2.0415 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,324B, BPFP=2.2947 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,564B, BPFP=0.2842 +⌛️ [2/4] FRONTEND: Frontend time: 2.268s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.638s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13820392 40.36713784 + layer.0.v_cache 0.00001454 0.00539367 + layer.1.k_cache 0.61684219 8.49012762 + layer.1.v_cache 0.00000570 0.00215510 + layer.2.k_cache 0.01285210 1.01037887 + layer.2.v_cache 0.00002011 0.00664633 + layer.3.k_cache 0.05980752 4.14319965 + layer.3.v_cache 0.00002263 0.00734448 + layer.4.k_cache 0.00071023 0.18899907 + layer.4.v_cache 0.00005833 0.01435438 + layer.4.output 0.05086143 177.23999697 + ------------------------------------------------------------------------------------- + TOTAL 0.06968043 76.17151270 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 383740 +BPFP 1.1956 bits/point +EBPFP 2.3912 equivalent bits/point +MSE 76.171513 +---------------------- -------------------------------------------------------- +Time: 3.918s Load: 0.012s, Pack+Encode: 2.268s, Decode+Unpack: 1.638s +---------------------- -------------------------------------------------------- +💾 Converting with 76.1715 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,452B, BPFP=0.4907 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,168B, BPFP=2.3447 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,348B, BPFP=1.2639 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,128B, BPFP=2.2907 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,912B, BPFP=1.5008 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,864B, BPFP=2.2770 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,176B, BPFP=1.3588 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,852B, BPFP=2.3283 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,900B, BPFP=2.0193 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,948B, BPFP=2.2814 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,040B, BPFP=0.2895 +⌛️ [2/4] FRONTEND: Frontend time: 2.201s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.696s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14825445 40.96680661 + layer.0.v_cache 0.00001396 0.00560025 + layer.1.k_cache 0.63210902 8.77709555 + layer.1.v_cache 0.00000581 0.00231687 + layer.2.k_cache 0.01184132 0.93705724 + layer.2.v_cache 0.00001890 0.00676401 + layer.3.k_cache 0.01767834 5.42338892 + layer.3.v_cache 0.00002009 0.00743201 + layer.4.k_cache 0.00068272 0.22063738 + layer.4.v_cache 0.00005247 0.01487082 + layer.4.output 0.04867053 167.94844566 + ------------------------------------------------------------------------------------- + TOTAL 0.06772769 72.47065231 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 388788 +BPFP 1.1872 bits/point +EBPFP 2.3744 equivalent bits/point +MSE 72.470652 +---------------------- -------------------------------------------------------- +Time: 3.908s Load: 0.012s, Pack+Encode: 2.201s, Decode+Unpack: 1.696s +---------------------- -------------------------------------------------------- +💾 Converting with 72.4707 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,444B, BPFP=0.4998 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,204B, BPFP=2.5571 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,088B, BPFP=1.2481 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,036B, BPFP=2.4879 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,184B, BPFP=1.4313 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,512B, BPFP=2.5161 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,784B, BPFP=1.3485 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,204B, BPFP=2.4979 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,400B, BPFP=2.1544 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,908B, BPFP=2.4804 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,468B, BPFP=0.2661 +⌛️ [2/4] FRONTEND: Frontend time: 2.202s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.615s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14315909 42.86105809 + layer.0.v_cache 0.00001391 0.00538914 + layer.1.k_cache 0.49711875 8.59485881 + layer.1.v_cache 0.00000560 0.00222789 + layer.2.k_cache 0.00646762 0.98085276 + layer.2.v_cache 0.00001905 0.00671355 + layer.3.k_cache 0.04707547 5.66546631 + layer.3.v_cache 0.00001924 0.00744198 + layer.4.k_cache 0.00071833 0.20383862 + layer.4.v_cache 0.00005197 0.01409508 + layer.4.output 0.00997522 199.36126894 + ------------------------------------------------------------------------------------- + TOTAL 0.04496915 85.52181322 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 356232 +BPFP 1.2402 bits/point +EBPFP 2.4804 equivalent bits/point +MSE 85.521813 +---------------------- -------------------------------------------------------- +Time: 3.828s Load: 0.011s, Pack+Encode: 2.202s, Decode+Unpack: 1.615s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5218 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,768B, BPFP=0.4910 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,668B, BPFP=2.5016 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,784B, BPFP=1.2200 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,104B, BPFP=2.4140 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,512B, BPFP=1.3728 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,224B, BPFP=2.4207 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,788B, BPFP=1.3322 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,472B, BPFP=2.4346 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,548B, BPFP=2.1028 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,144B, BPFP=2.4162 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,028B, BPFP=0.2882 +⌛️ [2/4] FRONTEND: Frontend time: 2.582s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.710s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15581964 40.26130922 + layer.0.v_cache 0.00001703 0.00539976 + layer.1.k_cache 0.56298232 8.21292125 + layer.1.v_cache 0.00000550 0.00220124 + layer.2.k_cache 0.00979712 0.97401007 + layer.2.v_cache 0.00002023 0.00684721 + layer.3.k_cache 0.04960944 5.12293224 + layer.3.v_cache 0.00001971 0.00745233 + layer.4.k_cache 0.00068561 0.19249857 + layer.4.v_cache 0.00005245 0.01394113 + layer.4.output 0.01122868 187.61930364 + ------------------------------------------------------------------------------------- + TOTAL 0.05044764 80.47850815 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 370040 +BPFP 1.2190 bits/point +EBPFP 2.4381 equivalent bits/point +MSE 80.478508 +---------------------- -------------------------------------------------------- +Time: 4.303s Load: 0.011s, Pack+Encode: 2.582s, Decode+Unpack: 1.710s +---------------------- -------------------------------------------------------- +💾 Converting with 80.4785 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,840B, BPFP=0.4881 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,740B, BPFP=2.4702 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,592B, BPFP=1.2473 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,672B, BPFP=2.4112 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,700B, BPFP=1.4189 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,136B, BPFP=2.3816 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,568B, BPFP=1.3012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,096B, BPFP=2.4346 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,892B, BPFP=2.0921 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,176B, BPFP=2.3838 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,000B, BPFP=0.2997 +⌛️ [2/4] FRONTEND: Frontend time: 2.579s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.740s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13599485 39.70188618 + layer.0.v_cache 0.00001418 0.00533609 + layer.1.k_cache 0.56373192 8.33700033 + layer.1.v_cache 0.00000588 0.00225561 + layer.2.k_cache 0.01549433 0.92465566 + layer.2.v_cache 0.00001867 0.00664219 + layer.3.k_cache 0.03475572 5.09681038 + layer.3.v_cache 0.00001907 0.00727572 + layer.4.k_cache 0.00072111 0.19690040 + layer.4.v_cache 0.00005295 0.01400190 + layer.4.output 0.00820773 190.13585310 + ------------------------------------------------------------------------------------- + TOTAL 0.04754487 81.48492566 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 375412 +BPFP 1.2193 bits/point +EBPFP 2.4385 equivalent bits/point +MSE 81.484926 +---------------------- -------------------------------------------------------- +Time: 4.328s Load: 0.009s, Pack+Encode: 2.579s, Decode+Unpack: 1.740s +---------------------- -------------------------------------------------------- +💾 Converting with 81.4849 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,544B, BPFP=0.4842 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,232B, BPFP=2.2946 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,920B, BPFP=1.3149 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,044B, BPFP=2.2344 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,208B, BPFP=1.4817 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,844B, BPFP=2.2242 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,208B, BPFP=1.3803 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,888B, BPFP=2.2772 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,732B, BPFP=1.9649 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,736B, BPFP=2.2188 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,768B, BPFP=0.2737 +⌛️ [2/4] FRONTEND: Frontend time: 2.367s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.911s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13331154 39.24104289 + layer.0.v_cache 0.00001460 0.00516958 + layer.1.k_cache 0.64535681 8.43339321 + layer.1.v_cache 0.00000563 0.00215322 + layer.2.k_cache 0.00762887 0.86159872 + layer.2.v_cache 0.00001910 0.00660786 + layer.3.k_cache 0.03538683 6.35611239 + layer.3.v_cache 0.00002103 0.00736884 + layer.4.k_cache 0.00072257 0.20287885 + layer.4.v_cache 0.00005300 0.01446809 + layer.4.output 0.04522907 169.70654569 + ------------------------------------------------------------------------------------- + TOTAL 0.06700726 73.12215373 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 390124 +BPFP 1.1642 bits/point +EBPFP 2.3284 equivalent bits/point +MSE 73.122154 +---------------------- -------------------------------------------------------- +Time: 4.290s Load: 0.012s, Pack+Encode: 2.367s, Decode+Unpack: 1.911s +---------------------- -------------------------------------------------------- +💾 Converting with 73.1222 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,860B, BPFP=0.4875 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,028B, BPFP=2.4773 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,672B, BPFP=1.3024 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,944B, BPFP=2.4177 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,332B, BPFP=1.4487 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,540B, BPFP=2.3955 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,220B, BPFP=1.3875 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,508B, BPFP=2.4487 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,148B, BPFP=2.0988 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,512B, BPFP=2.3939 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,100B, BPFP=0.2995 +⌛️ [2/4] FRONTEND: Frontend time: 2.540s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.702s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14806804 40.30532708 + layer.0.v_cache 0.00001376 0.00524247 + layer.1.k_cache 0.60662186 8.00293485 + layer.1.v_cache 0.00000571 0.00221754 + layer.2.k_cache 0.01239752 1.00382458 + layer.2.v_cache 0.00002027 0.00647214 + layer.3.k_cache 0.02456893 6.12185755 + layer.3.v_cache 0.00001975 0.00706557 + layer.4.k_cache 0.00069718 0.18733876 + layer.4.v_cache 0.00005379 0.01415022 + layer.4.output 0.00818182 191.09445737 + ------------------------------------------------------------------------------------- + TOTAL 0.04998468 81.95986073 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 380864 +BPFP 1.2326 bits/point +EBPFP 2.4652 equivalent bits/point +MSE 81.959861 +---------------------- -------------------------------------------------------- +Time: 4.254s Load: 0.012s, Pack+Encode: 2.540s, Decode+Unpack: 1.702s +---------------------- -------------------------------------------------------- +💾 Converting with 81.9599 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,900B, BPFP=0.4914 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,188B, BPFP=2.4949 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,184B, BPFP=1.2800 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,100B, BPFP=2.4348 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,068B, BPFP=1.4393 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,756B, BPFP=2.4159 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,320B, BPFP=1.3428 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,480B, BPFP=2.4558 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,504B, BPFP=2.1259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,592B, BPFP=2.4068 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,640B, BPFP=0.3048 +⌛️ [2/4] FRONTEND: Frontend time: 2.294s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.709s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14592592 38.99572452 + layer.0.v_cache 0.00001391 0.00544781 + layer.1.k_cache 0.53254910 8.03157696 + layer.1.v_cache 0.00000583 0.00230974 + layer.2.k_cache 0.01029040 0.87402759 + layer.2.v_cache 0.00002009 0.00695896 + layer.3.k_cache 0.03848921 5.58573994 + layer.3.v_cache 0.00001972 0.00776436 + layer.4.k_cache 0.00069417 0.21439607 + layer.4.v_cache 0.00005427 0.01488508 + layer.4.output 0.01004909 190.77027070 + ------------------------------------------------------------------------------------- + TOTAL 0.04696507 81.71357211 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 380732 +BPFP 1.2365 bits/point +EBPFP 2.4731 equivalent bits/point +MSE 81.713572 +---------------------- -------------------------------------------------------- +Time: 4.013s Load: 0.010s, Pack+Encode: 2.294s, Decode+Unpack: 1.709s +---------------------- -------------------------------------------------------- +💾 Converting with 81.7136 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,676B, BPFP=0.4984 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,088B, BPFP=2.5326 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,892B, BPFP=1.2576 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,620B, BPFP=2.4483 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,668B, BPFP=1.3596 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,556B, BPFP=2.4446 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,724B, BPFP=1.3628 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,804B, BPFP=2.4589 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,348B, BPFP=2.1455 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,760B, BPFP=2.4563 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,180B, BPFP=0.2887 +⌛️ [2/4] FRONTEND: Frontend time: 2.354s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.742s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15384994 41.51361443 + layer.0.v_cache 0.00001404 0.00540688 + layer.1.k_cache 0.48768111 8.40150452 + layer.1.v_cache 0.00000584 0.00231641 + layer.2.k_cache 0.00538594 0.92972385 + layer.2.v_cache 0.00001948 0.00665802 + layer.3.k_cache 0.02347840 5.71813202 + layer.3.v_cache 0.00001899 0.00704585 + layer.4.k_cache 0.00068094 0.19133439 + layer.4.v_cache 0.00005116 0.01457327 + layer.4.output 0.00909345 196.75579372 + ------------------------------------------------------------------------------------- + TOTAL 0.04322588 84.35769798 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 365316 +BPFP 1.2344 bits/point +EBPFP 2.4689 equivalent bits/point +MSE 84.357698 +---------------------- -------------------------------------------------------- +Time: 4.106s Load: 0.009s, Pack+Encode: 2.354s, Decode+Unpack: 1.742s +---------------------- -------------------------------------------------------- +💾 Converting with 84.3577 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,328B, BPFP=0.4957 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,896B, BPFP=2.3861 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,628B, BPFP=1.3620 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,840B, BPFP=2.3299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,612B, BPFP=1.5206 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,552B, BPFP=2.3146 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,404B, BPFP=1.3501 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,364B, BPFP=2.3578 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,568B, BPFP=2.0497 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,316B, BPFP=2.3021 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,656B, BPFP=0.2707 +⌛️ [2/4] FRONTEND: Frontend time: 2.461s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.786s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13453574 40.12468112 + layer.0.v_cache 0.00001386 0.00548178 + layer.1.k_cache 0.70462929 8.82064134 + layer.1.v_cache 0.00000580 0.00222608 + layer.2.k_cache 0.00896267 0.95304455 + layer.2.v_cache 0.00001881 0.00678021 + layer.3.k_cache 0.02468209 6.15783401 + layer.3.v_cache 0.00001940 0.00724499 + layer.4.k_cache 0.00068809 0.19908152 + layer.4.v_cache 0.00005209 0.01414262 + layer.4.output 0.04953035 166.17079689 + ------------------------------------------------------------------------------------- + TOTAL 0.07178355 71.73451391 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 383164 +BPFP 1.1979 bits/point +EBPFP 2.3957 equivalent bits/point +MSE 71.734514 +---------------------- -------------------------------------------------------- +Time: 4.258s Load: 0.012s, Pack+Encode: 2.461s, Decode+Unpack: 1.786s +---------------------- -------------------------------------------------------- +💾 Converting with 71.7345 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,620B, BPFP=0.5026 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,920B, BPFP=2.5606 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,572B, BPFP=1.2577 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,940B, BPFP=2.5035 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,392B, BPFP=1.4221 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,664B, BPFP=2.4874 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,524B, BPFP=1.4298 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,824B, BPFP=2.4967 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,600B, BPFP=2.1922 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,240B, BPFP=2.4627 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,220B, BPFP=0.2933 +⌛️ [2/4] FRONTEND: Frontend time: 2.941s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.912s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150386 40.32084086 + layer.0.v_cache 0.00001401 0.00536212 + layer.1.k_cache 0.57442873 8.40787882 + layer.1.v_cache 0.00000554 0.00221103 + layer.2.k_cache 0.00779028 0.85027467 + layer.2.v_cache 0.00001875 0.00672126 + layer.3.k_cache 0.05176373 5.66520577 + layer.3.v_cache 0.00002027 0.00731655 + layer.4.k_cache 0.00069872 0.19823097 + layer.4.v_cache 0.00005023 0.01473398 + layer.4.output 0.01044900 198.78346549 + ------------------------------------------------------------------------------------- + TOTAL 0.04937866 85.11547261 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 366516 +BPFP 1.2570 bits/point +EBPFP 2.5140 equivalent bits/point +MSE 85.115473 +---------------------- -------------------------------------------------------- +Time: 4.866s Load: 0.012s, Pack+Encode: 2.941s, Decode+Unpack: 1.912s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1155 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.2145 bits/point +Avg EBPFP 2.4291 equivalent bits/point +Avg MSE 78.083965 +Avg Time 4.143s +------------------------ ---------------------------- diff --git a/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..4858ff083f8de51fb7bf1b9692aef7dc84f5b568 --- /dev/null +++ b/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 506 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.01_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa +Output output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa +---------------- ------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,212B, BPFP=0.5975 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,248B, BPFP=3.0223 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,464B, BPFP=1.2024 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,636B, BPFP=2.5365 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,964B, BPFP=1.2954 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,024B, BPFP=2.4226 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,756B, BPFP=1.2567 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,716B, BPFP=2.5513 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,060B, BPFP=2.0573 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,568B, BPFP=2.5238 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,368B, BPFP=0.3818 +⌛️ [2/4] FRONTEND: Frontend time: 2.451s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.260s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14073507 41.07988049 + layer.0.v_cache 0.00001426 0.00639551 + layer.1.k_cache 0.05042761 8.42514765 + layer.1.v_cache 0.00000517 0.00223432 + layer.2.k_cache 0.00244098 1.03848430 + layer.2.v_cache 0.00001715 0.00688687 + layer.3.k_cache 0.02726505 3.38090588 + layer.3.v_cache 0.00001796 0.00757038 + layer.4.k_cache 0.00069123 0.18743631 + layer.4.v_cache 0.00005084 0.01779644 + layer.4.output 0.16186065 629.95487883 + ------------------------------------------------------------------------------------- + TOTAL 0.07968764 262.57864058 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 119016 +BPFP 1.3023 bits/point +EBPFP 2.6045 equivalent bits/point +MSE 262.578641 +---------------------- -------------------------------------------------------- +Time: 3.716s Load: 0.005s, Pack+Encode: 2.451s, Decode+Unpack: 1.260s +---------------------- -------------------------------------------------------- +💾 Converting with 262.5786 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,192B, BPFP=0.6009 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,300B, BPFP=2.5038 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,328B, BPFP=1.1913 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,544B, BPFP=2.3614 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,716B, BPFP=1.2643 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,140B, BPFP=2.2854 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,532B, BPFP=1.2297 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,928B, BPFP=2.4337 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,308B, BPFP=1.9405 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,260B, BPFP=2.3080 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,836B, BPFP=0.3990 +⌛️ [2/4] FRONTEND: Frontend time: 1.941s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.300s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11392649 41.50688300 + layer.0.v_cache 0.00001670 0.00625171 + layer.1.k_cache 0.05249626 8.45975062 + layer.1.v_cache 0.00000506 0.00222641 + layer.2.k_cache 0.00416583 0.75865495 + layer.2.v_cache 0.00001711 0.00666124 + layer.3.k_cache 0.05931453 3.70222050 + layer.3.v_cache 0.00001789 0.00776170 + layer.4.k_cache 0.00069248 0.17312806 + layer.4.v_cache 0.00004729 0.01608881 + layer.4.output 0.16383441 634.92491394 + ------------------------------------------------------------------------------------- + TOTAL 0.08103179 264.65376615 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 111084 +BPFP 1.2301 bits/point +EBPFP 2.4602 equivalent bits/point +MSE 264.653766 +---------------------- -------------------------------------------------------- +Time: 3.245s Load: 0.004s, Pack+Encode: 1.941s, Decode+Unpack: 1.300s +---------------------- -------------------------------------------------------- +💾 Converting with 264.6538 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,344B, BPFP=0.5559 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,488B, BPFP=2.9069 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,004B, BPFP=1.1642 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,524B, BPFP=2.7467 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,196B, BPFP=1.1961 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,468B, BPFP=2.7374 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,080B, BPFP=1.1769 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,264B, BPFP=2.8697 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,652B, BPFP=2.1031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,880B, BPFP=2.8059 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,316B, BPFP=0.3400 +⌛️ [2/4] FRONTEND: Frontend time: 1.821s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.188s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17268721 39.78134090 + layer.0.v_cache 0.00001453 0.00637116 + layer.1.k_cache 0.08276152 8.13030876 + layer.1.v_cache 0.00000545 0.00220508 + layer.2.k_cache 0.00739249 0.82098389 + layer.2.v_cache 0.00001625 0.00623088 + layer.3.k_cache 0.05929063 3.88981856 + layer.3.v_cache 0.00001766 0.00764275 + layer.4.k_cache 0.00068371 0.17545205 + layer.4.v_cache 0.00005055 0.01548725 + layer.4.output 0.14466690 558.07513298 + ------------------------------------------------------------------------------------- + TOTAL 0.07856402 232.90363365 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 136216 +BPFP 1.3319 bits/point +EBPFP 2.6638 equivalent bits/point +MSE 232.903634 +---------------------- -------------------------------------------------------- +Time: 3.013s Load: 0.004s, Pack+Encode: 1.821s, Decode+Unpack: 1.188s +---------------------- -------------------------------------------------------- +💾 Converting with 232.9036 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 112, 128) +Output shape: (1, 112, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.output: torch.Size([1, 112, 3584]) -> torch.Size([1, 1, 112, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,988B, BPFP=0.5564 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,500B, BPFP=2.5809 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,420B, BPFP=1.1747 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,452B, BPFP=2.4347 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,924B, BPFP=1.3845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,128B, BPFP=2.3895 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,268B, BPFP=1.5720 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,816B, BPFP=2.4855 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,400B, BPFP=2.1484 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,280B, BPFP=2.4107 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,668B, BPFP=0.3920 +⌛️ [2/4] FRONTEND: Frontend time: 1.703s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.291s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17321212 39.70253209 + layer.0.v_cache 0.00001450 0.00591597 + layer.1.k_cache 0.08643453 9.15949249 + layer.1.v_cache 0.00000659 0.00225503 + layer.2.k_cache 0.00690480 0.83522415 + layer.2.v_cache 0.00001899 0.00638169 + layer.3.k_cache 0.01205242 5.98859624 + layer.3.v_cache 0.00001923 0.00733493 + layer.4.k_cache 0.00071999 0.17942401 + layer.4.v_cache 0.00004947 0.01467990 + layer.4.output 10.17892331 459.49426020 + ------------------------------------------------------------------------------------- + TOTAL 4.20775858 192.49186223 + (elements=974,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 974848 +Total Bytes 156844 +BPFP 1.2871 bits/point +EBPFP 2.5743 equivalent bits/point +MSE 192.491862 +---------------------- -------------------------------------------------------- +Time: 2.998s Load: 0.004s, Pack+Encode: 1.703s, Decode+Unpack: 1.291s +---------------------- -------------------------------------------------------- +💾 Converting with 192.4919 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,284B, BPFP=0.5831 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,552B, BPFP=2.7614 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,820B, BPFP=1.2109 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,176B, BPFP=2.6946 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,972B, BPFP=1.2379 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,752B, BPFP=2.4418 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,856B, BPFP=1.2173 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,172B, BPFP=2.5163 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,788B, BPFP=2.0930 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,392B, BPFP=2.5554 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,692B, BPFP=0.3473 +⌛️ [2/4] FRONTEND: Frontend time: 1.894s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.287s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11687154 43.63887440 + layer.0.v_cache 0.00001423 0.00638804 + layer.1.k_cache 0.05324238 8.00352964 + layer.1.v_cache 0.00000539 0.00227755 + layer.2.k_cache 0.00397935 0.68851449 + layer.2.v_cache 0.00001766 0.00621002 + layer.3.k_cache 0.01210797 4.07151760 + layer.3.v_cache 0.00001803 0.00770185 + layer.4.k_cache 0.00075037 0.19228205 + layer.4.v_cache 0.00004481 0.01526156 + layer.4.output 0.15448949 587.27709010 + ------------------------------------------------------------------------------------- + TOTAL 0.07461636 245.15130517 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 122456 +BPFP 1.2790 bits/point +EBPFP 2.5580 equivalent bits/point +MSE 245.151305 +---------------------- -------------------------------------------------------- +Time: 3.186s Load: 0.004s, Pack+Encode: 1.894s, Decode+Unpack: 1.287s +---------------------- -------------------------------------------------------- +💾 Converting with 245.1513 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,148B, BPFP=0.6073 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,820B, BPFP=2.8588 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,372B, BPFP=1.2292 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,224B, BPFP=2.5509 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,692B, BPFP=1.2909 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,760B, BPFP=2.4614 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,600B, BPFP=1.2731 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,300B, BPFP=2.5656 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,560B, BPFP=2.0370 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,656B, BPFP=2.4414 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,476B, BPFP=0.4265 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203883 41.97442853 + layer.0.v_cache 0.00001438 0.00650768 + layer.1.k_cache 0.05633916 8.93966110 + layer.1.v_cache 0.00000602 0.00251619 + layer.2.k_cache 0.00244657 0.91877652 + layer.2.v_cache 0.00001847 0.00693277 + layer.3.k_cache 0.05492775 4.15714933 + layer.3.v_cache 0.00001883 0.00774776 + layer.4.k_cache 0.00062984 0.17886324 + layer.4.v_cache 0.00005178 0.01763245 + layer.4.output 0.18538215 625.30456349 + ------------------------------------------------------------------------------------- + TOTAL 0.09142157 260.78483294 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 115608 +BPFP 1.3118 bits/point +EBPFP 2.6236 equivalent bits/point +MSE 260.784833 +---------------------- -------------------------------------------------------- +Time: 3.057s Load: 0.005s, Pack+Encode: 1.838s, Decode+Unpack: 1.215s +---------------------- -------------------------------------------------------- +💾 Converting with 260.7848 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,348B, BPFP=0.5813 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,516B, BPFP=2.8674 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,780B, BPFP=1.1771 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,956B, BPFP=2.7701 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,332B, BPFP=1.2729 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,008B, BPFP=2.6056 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,136B, BPFP=1.2389 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,136B, BPFP=2.8014 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,288B, BPFP=2.1333 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,252B, BPFP=2.8215 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,172B, BPFP=0.3763 +⌛️ [2/4] FRONTEND: Frontend time: 1.999s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.332s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13028670 40.65171170 + layer.0.v_cache 0.00001462 0.00619427 + layer.1.k_cache 0.04946813 8.19225260 + layer.1.v_cache 0.00000624 0.00217546 + layer.2.k_cache 0.00411291 0.81480899 + layer.2.v_cache 0.00001729 0.00615768 + layer.3.k_cache 0.01793825 4.09349331 + layer.3.v_cache 0.00001895 0.00774200 + layer.4.k_cache 0.00077120 0.17549917 + layer.4.v_cache 0.00004532 0.01469975 + layer.4.output 0.15109633 577.92876984 + ------------------------------------------------------------------------------------- + TOTAL 0.07413847 241.14506611 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 131924 +BPFP 1.3473 bits/point +EBPFP 2.6945 equivalent bits/point +MSE 241.145066 +---------------------- -------------------------------------------------------- +Time: 3.338s Load: 0.006s, Pack+Encode: 1.999s, Decode+Unpack: 1.332s +---------------------- -------------------------------------------------------- +💾 Converting with 241.1451 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,388B, BPFP=0.6546 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,296B, BPFP=2.5482 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,052B, BPFP=1.6590 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,900B, BPFP=2.4397 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,608B, BPFP=1.8114 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,712B, BPFP=2.3882 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,460B, BPFP=1.7708 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,008B, BPFP=2.4693 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,004B, BPFP=2.1941 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,792B, BPFP=2.4101 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,520B, BPFP=0.4511 +⌛️ [2/4] FRONTEND: Frontend time: 2.085s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.169s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11683119 42.68959875 + layer.0.v_cache 0.00001514 0.00826908 + layer.1.k_cache 0.01218004 9.06874807 + layer.1.v_cache 0.00000521 0.00285449 + layer.2.k_cache 0.00491684 1.01135615 + layer.2.v_cache 0.00001706 0.00814026 + layer.3.k_cache 0.10823302 4.60935894 + layer.3.v_cache 0.00002043 0.01057662 + layer.4.k_cache 0.00062874 0.20751155 + layer.4.v_cache 0.00004794 0.01935700 + layer.4.output 0.23833110 935.59421992 + ------------------------------------------------------------------------------------- + TOTAL 0.11242431 388.63501826 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 85740 +BPFP 1.3825 bits/point +EBPFP 2.7651 equivalent bits/point +MSE 388.635018 +---------------------- -------------------------------------------------------- +Time: 3.258s Load: 0.003s, Pack+Encode: 2.085s, Decode+Unpack: 1.169s +---------------------- -------------------------------------------------------- +💾 Converting with 388.6350 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,200B, BPFP=0.6875 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,988B, BPFP=2.8087 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,648B, BPFP=1.4525 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,564B, BPFP=2.6763 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,564B, BPFP=2.0513 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,212B, BPFP=2.5663 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,160B, BPFP=1.6125 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,396B, BPFP=2.6237 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,540B, BPFP=2.3563 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,400B, BPFP=2.6250 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,456B, BPFP=0.4668 +⌛️ [2/4] FRONTEND: Frontend time: 1.772s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.184s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14598725 41.84382324 + layer.0.v_cache 0.00001449 0.00892478 + layer.1.k_cache 0.01347294 8.66471497 + layer.1.v_cache 0.00000521 0.00296985 + layer.2.k_cache 0.00488209 1.04606361 + layer.2.v_cache 0.00001649 0.00819431 + layer.3.k_cache 0.04421538 5.27764648 + layer.3.v_cache 0.00001905 0.01069496 + layer.4.k_cache 0.00073779 0.21248919 + layer.4.v_cache 0.00005089 0.02079260 + layer.4.output 0.27158585 1018.67553571 + ------------------------------------------------------------------------------------- + TOTAL 0.12414721 422.81323906 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 79128 +BPFP 1.4546 bits/point +EBPFP 2.9091 equivalent bits/point +MSE 422.813239 +---------------------- -------------------------------------------------------- +Time: 2.958s Load: 0.002s, Pack+Encode: 1.772s, Decode+Unpack: 1.184s +---------------------- -------------------------------------------------------- +💾 Converting with 422.8132 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 46, 128) +Output shape: (1, 46, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.output: torch.Size([1, 46, 3584]) -> torch.Size([1, 1, 46, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,164B, BPFP=0.7351 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,052B, BPFP=3.0747 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,892B, BPFP=1.3220 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,568B, BPFP=2.9103 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,788B, BPFP=1.9660 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,168B, BPFP=2.7745 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,092B, BPFP=1.3899 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,580B, BPFP=2.9144 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,616B, BPFP=2.2473 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,568B, BPFP=2.9103 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 9,980B, BPFP=0.4843 +⌛️ [2/4] FRONTEND: Frontend time: 1.767s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.129s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14129612 37.41382632 + layer.0.v_cache 0.00001532 0.00878352 + layer.1.k_cache 0.01315293 9.27805959 + layer.1.v_cache 0.00000553 0.00307669 + layer.2.k_cache 0.00779452 1.30984829 + layer.2.v_cache 0.00001723 0.00814531 + layer.3.k_cache 0.01828384 4.96878914 + layer.3.v_cache 0.00001944 0.01132233 + layer.4.k_cache 0.00060559 0.20564701 + layer.4.v_cache 0.00005113 0.02191348 + layer.4.output 0.29517000 1151.87732919 + ------------------------------------------------------------------------------------- + TOTAL 0.13220186 477.43357153 + (elements=400,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 400384 +Total Bytes 75468 +BPFP 1.5079 bits/point +EBPFP 3.0158 equivalent bits/point +MSE 477.433572 +---------------------- -------------------------------------------------------- +Time: 2.899s Load: 0.002s, Pack+Encode: 1.767s, Decode+Unpack: 1.129s +---------------------- -------------------------------------------------------- +💾 Converting with 477.4336 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 110, 128) +Output shape: (1, 110, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.output: torch.Size([1, 110, 3584]) -> torch.Size([1, 1, 110, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,980B, BPFP=0.5653 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,516B, BPFP=2.6301 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,876B, BPFP=1.2608 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,860B, BPFP=2.5369 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,924B, BPFP=1.6938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,168B, BPFP=2.4386 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,912B, BPFP=1.4080 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,576B, BPFP=2.4966 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,892B, BPFP=2.1153 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,220B, BPFP=2.4460 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,408B, BPFP=0.3938 +⌛️ [2/4] FRONTEND: Frontend time: 2.123s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.465s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14476244 42.13398437 + layer.0.v_cache 0.00001476 0.00636436 + layer.1.k_cache 0.05665370 8.71079989 + layer.1.v_cache 0.00000525 0.00215380 + layer.2.k_cache 0.00786287 0.99274202 + layer.2.v_cache 0.00001801 0.00607808 + layer.3.k_cache 0.01410385 5.23952248 + layer.3.v_cache 0.00002027 0.00716817 + layer.4.k_cache 0.00066873 0.19129666 + layer.4.v_cache 0.00004626 0.01433214 + layer.4.output 10.36401748 463.29147727 + ------------------------------------------------------------------------------------- + TOTAL 4.28072226 194.13792841 + (elements=957,440) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 957440 +Total Bytes 157332 +BPFP 1.3146 bits/point +EBPFP 2.6292 equivalent bits/point +MSE 194.137928 +---------------------- -------------------------------------------------------- +Time: 3.592s Load: 0.005s, Pack+Encode: 2.123s, Decode+Unpack: 1.465s +---------------------- -------------------------------------------------------- +💾 Converting with 194.1379 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,184B, BPFP=0.6142 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,320B, BPFP=2.9552 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,404B, BPFP=1.2353 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,752B, BPFP=2.8457 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,056B, BPFP=1.3611 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,844B, BPFP=2.8634 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,912B, BPFP=1.3333 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,016B, BPFP=2.8966 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,140B, BPFP=2.1489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,460B, BPFP=2.7894 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,488B, BPFP=0.5370 +⌛️ [2/4] FRONTEND: Frontend time: 2.027s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.532s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10533771 41.79807460 + layer.0.v_cache 0.00001610 0.00711136 + layer.1.k_cache 0.05553475 8.60803901 + layer.1.v_cache 0.00000748 0.00284956 + layer.2.k_cache 0.02166171 0.80142447 + layer.2.v_cache 0.00002000 0.00778201 + layer.3.k_cache 0.06386927 3.86536285 + layer.3.v_cache 0.00001944 0.00890715 + layer.4.k_cache 0.00064315 0.18956170 + layer.4.v_cache 0.00005385 0.01744017 + layer.4.output 0.16801396 602.79177690 + ------------------------------------------------------------------------------------- + TOTAL 0.08372125 251.46170536 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 128576 +BPFP 1.4590 bits/point +EBPFP 2.9179 equivalent bits/point +MSE 251.461705 +---------------------- -------------------------------------------------------- +Time: 3.563s Load: 0.004s, Pack+Encode: 2.027s, Decode+Unpack: 1.532s +---------------------- -------------------------------------------------------- +💾 Converting with 251.4617 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 49, 128) +Output shape: (1, 49, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.output: torch.Size([1, 49, 3584]) -> torch.Size([1, 1, 49, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,196B, BPFP=0.7003 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,064B, BPFP=2.8903 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,556B, BPFP=1.4528 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,532B, BPFP=2.7207 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,008B, BPFP=1.9158 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,456B, BPFP=2.6964 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,228B, BPFP=1.3482 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,588B, BPFP=2.7385 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,168B, BPFP=2.2857 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,444B, BPFP=2.6926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,324B, BPFP=0.5614 +⌛️ [2/4] FRONTEND: Frontend time: 1.872s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.125s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12410851 42.30633171 + layer.0.v_cache 0.00001740 0.00856881 + layer.1.k_cache 0.01401279 8.90319700 + layer.1.v_cache 0.00000579 0.00307102 + layer.2.k_cache 0.00545398 0.98810219 + layer.2.v_cache 0.00001955 0.00892103 + layer.3.k_cache 0.07546502 4.77008243 + layer.3.v_cache 0.00001978 0.01090188 + layer.4.k_cache 0.00063206 0.22400064 + layer.4.v_cache 0.00005205 0.02034838 + layer.4.output 0.27729562 1040.26785714 + ------------------------------------------------------------------------------------- + TOTAL 0.12710919 431.71285442 + (elements=426,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 426496 +Total Bytes 79564 +BPFP 1.4924 bits/point +EBPFP 2.9848 equivalent bits/point +MSE 431.712854 +---------------------- -------------------------------------------------------- +Time: 2.999s Load: 0.003s, Pack+Encode: 1.872s, Decode+Unpack: 1.125s +---------------------- -------------------------------------------------------- +💾 Converting with 431.7129 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,392B, BPFP=0.6557 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,328B, BPFP=2.5570 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,440B, BPFP=1.4912 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,824B, BPFP=2.4189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,528B, BPFP=1.7895 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,548B, BPFP=2.3432 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,448B, BPFP=1.7675 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,740B, BPFP=2.3958 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,668B, BPFP=2.1020 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,652B, BPFP=2.3717 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,968B, BPFP=0.4687 +⌛️ [2/4] FRONTEND: Frontend time: 1.810s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.200s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19197876 40.48446923 + layer.0.v_cache 0.00001449 0.00790453 + layer.1.k_cache 0.01466701 9.07294022 + layer.1.v_cache 0.00000519 0.00284338 + layer.2.k_cache 0.00720649 0.81967772 + layer.2.v_cache 0.00001785 0.00780613 + layer.3.k_cache 0.10851213 5.58794631 + layer.3.v_cache 0.00001996 0.00993256 + layer.4.k_cache 0.00062085 0.18962549 + layer.4.v_cache 0.00004770 0.01884577 + layer.4.output 0.23832821 902.13024749 + ------------------------------------------------------------------------------------- + TOTAL 0.11714047 374.77139552 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 84536 +BPFP 1.3631 bits/point +EBPFP 2.7263 equivalent bits/point +MSE 374.771396 +---------------------- -------------------------------------------------------- +Time: 3.013s Load: 0.003s, Pack+Encode: 1.810s, Decode+Unpack: 1.200s +---------------------- -------------------------------------------------------- +💾 Converting with 374.7714 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,240B, BPFP=0.6863 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,152B, BPFP=2.8039 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,892B, BPFP=1.4988 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,620B, BPFP=2.6409 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,716B, BPFP=1.7512 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,392B, BPFP=2.5711 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,820B, BPFP=1.4767 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,712B, BPFP=2.6691 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,672B, BPFP=2.3505 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,624B, BPFP=2.6422 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,660B, BPFP=0.4666 +⌛️ [2/4] FRONTEND: Frontend time: 1.892s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.179s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12665685 39.45423799 + layer.0.v_cache 0.00001436 0.00856476 + layer.1.k_cache 0.01497617 8.48598705 + layer.1.v_cache 0.00000537 0.00294700 + layer.2.k_cache 0.00986420 0.96169782 + layer.2.v_cache 0.00001802 0.00871760 + layer.3.k_cache 0.09846177 4.97620646 + layer.3.v_cache 0.00001923 0.01072519 + layer.4.k_cache 0.00062397 0.22131681 + layer.4.v_cache 0.00005014 0.01993073 + layer.4.output 0.26631049 1052.10810574 + ------------------------------------------------------------------------------------- + TOTAL 0.12440374 436.40629833 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 79500 +BPFP 1.4327 bits/point +EBPFP 2.8655 equivalent bits/point +MSE 436.406298 +---------------------- -------------------------------------------------------- +Time: 3.074s Load: 0.003s, Pack+Encode: 1.892s, Decode+Unpack: 1.179s +---------------------- -------------------------------------------------------- +💾 Converting with 436.4063 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,196B, BPFP=0.7148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,988B, BPFP=2.9258 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,452B, BPFP=1.4492 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,660B, BPFP=2.8190 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,388B, BPFP=1.7539 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,456B, BPFP=2.7526 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,276B, BPFP=1.3919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,820B, BPFP=2.8711 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,360B, BPFP=2.3958 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,460B, BPFP=2.7539 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,284B, BPFP=0.5712 +⌛️ [2/4] FRONTEND: Frontend time: 1.834s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.179s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18531116 44.96631877 + layer.0.v_cache 0.00001874 0.00842153 + layer.1.k_cache 0.01459585 8.66974767 + layer.1.v_cache 0.00000674 0.00343210 + layer.2.k_cache 0.01042949 0.97245391 + layer.2.v_cache 0.00001841 0.00900405 + layer.3.k_cache 0.15260293 4.25282160 + layer.3.v_cache 0.00002390 0.01182977 + layer.4.k_cache 0.00061422 0.21745028 + layer.4.v_cache 0.00004881 0.02017027 + layer.4.output 0.28297247 1098.90476190 + ------------------------------------------------------------------------------------- + TOTAL 0.13791044 455.96852843 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 79340 +BPFP 1.5192 bits/point +EBPFP 3.0384 equivalent bits/point +MSE 455.968528 +---------------------- -------------------------------------------------------- +Time: 3.016s Load: 0.003s, Pack+Encode: 1.834s, Decode+Unpack: 1.179s +---------------------- -------------------------------------------------------- +💾 Converting with 455.9685 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,408B, BPFP=0.6377 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,536B, BPFP=2.5254 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,484B, BPFP=1.7172 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,952B, BPFP=2.3708 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,412B, BPFP=1.6981 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,676B, BPFP=2.2977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,416B, BPFP=1.6992 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,156B, BPFP=2.4248 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,772B, BPFP=2.0583 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,840B, BPFP=2.3411 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,092B, BPFP=0.4953 +⌛️ [2/4] FRONTEND: Frontend time: 1.700s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.160s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12052289 38.78402244 + layer.0.v_cache 0.00001546 0.00829127 + layer.1.k_cache 0.01434235 8.68321306 + layer.1.v_cache 0.00000574 0.00316369 + layer.2.k_cache 0.00254285 0.77350810 + layer.2.v_cache 0.00001710 0.00872878 + layer.3.k_cache 0.03737745 4.75501730 + layer.3.v_cache 0.00001946 0.01222449 + layer.4.k_cache 0.00063766 0.22053405 + layer.4.v_cache 0.00005039 0.02283529 + layer.4.output 0.23032015 872.81204600 + ------------------------------------------------------------------------------------- + TOTAL 0.10516308 362.52681532 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 87744 +BPFP 1.3669 bits/point +EBPFP 2.7338 equivalent bits/point +MSE 362.526815 +---------------------- -------------------------------------------------------- +Time: 2.864s Load: 0.003s, Pack+Encode: 1.700s, Decode+Unpack: 1.160s +---------------------- -------------------------------------------------------- +💾 Converting with 362.5268 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,004B, BPFP=0.6343 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,708B, BPFP=3.1056 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,804B, BPFP=1.2255 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,712B, BPFP=2.8953 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,328B, BPFP=1.3361 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,364B, BPFP=2.8218 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,204B, BPFP=1.3100 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,480B, BPFP=2.8463 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,112B, BPFP=2.1351 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,376B, BPFP=2.8243 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,240B, BPFP=0.4597 +⌛️ [2/4] FRONTEND: Frontend time: 1.876s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.352s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11335979 39.06251320 + layer.0.v_cache 0.00001513 0.00757168 + layer.1.k_cache 0.01655370 8.40151483 + layer.1.v_cache 0.00000667 0.00272914 + layer.2.k_cache 0.00428703 0.85896894 + layer.2.v_cache 0.00001789 0.00736214 + layer.3.k_cache 0.04938591 4.07658840 + layer.3.v_cache 0.00002111 0.00981968 + layer.4.k_cache 0.00066117 0.19616208 + layer.4.v_cache 0.00004716 0.01733442 + layer.4.output 0.18372514 692.67181467 + ------------------------------------------------------------------------------------- + TOTAL 0.08649597 288.31430984 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 115332 +BPFP 1.4325 bits/point +EBPFP 2.8650 equivalent bits/point +MSE 288.314310 +---------------------- -------------------------------------------------------- +Time: 3.232s Load: 0.003s, Pack+Encode: 1.876s, Decode+Unpack: 1.352s +---------------------- -------------------------------------------------------- +💾 Converting with 288.3143 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,188B, BPFP=0.7770 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,908B, BPFP=3.1634 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,948B, BPFP=1.4020 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,528B, BPFP=3.0284 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,044B, BPFP=1.4361 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,408B, BPFP=2.9858 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,304B, BPFP=1.5284 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,624B, BPFP=3.0625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,956B, BPFP=2.4702 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,512B, BPFP=3.0227 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,348B, BPFP=0.6264 +⌛️ [2/4] FRONTEND: Frontend time: 1.796s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.243s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14777968 42.35502208 + layer.0.v_cache 0.00001531 0.00885659 + layer.1.k_cache 0.01728138 9.14341597 + layer.1.v_cache 0.00000547 0.00322531 + layer.2.k_cache 0.00823456 1.04751968 + layer.2.v_cache 0.00001932 0.00932936 + layer.3.k_cache 0.16770924 5.34390744 + layer.3.v_cache 0.00002046 0.01157243 + layer.4.k_cache 0.00061569 0.21426825 + layer.4.v_cache 0.00005592 0.02276528 + layer.4.output 0.30867824 1158.75588474 + ------------------------------------------------------------------------------------- + TOTAL 0.14720498 480.55594562 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 76768 +BPFP 1.6036 bits/point +EBPFP 3.2072 equivalent bits/point +MSE 480.555946 +---------------------- -------------------------------------------------------- +Time: 3.042s Load: 0.003s, Pack+Encode: 1.796s, Decode+Unpack: 1.243s +---------------------- -------------------------------------------------------- +💾 Converting with 480.5559 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,092B, BPFP=0.6194 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,540B, BPFP=3.1130 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,304B, BPFP=1.2628 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,392B, BPFP=3.0833 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,784B, BPFP=1.3590 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,544B, BPFP=2.9135 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,584B, BPFP=1.3189 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,904B, BPFP=2.9856 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,972B, BPFP=2.1979 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,388B, BPFP=2.8822 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,364B, BPFP=0.4397 +⌛️ [2/4] FRONTEND: Frontend time: 2.045s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14408422 43.18982247 + layer.0.v_cache 0.00002097 0.00776757 + layer.1.k_cache 0.05735342 9.36712412 + layer.1.v_cache 0.00000575 0.00291827 + layer.2.k_cache 0.01678689 0.81948427 + layer.2.v_cache 0.00001894 0.00806114 + layer.3.k_cache 0.03029843 3.57888207 + layer.3.v_cache 0.00002125 0.01014750 + layer.4.k_cache 0.00062582 0.21203674 + layer.4.v_cache 0.00005050 0.01846854 + layer.4.output 0.17433185 682.87620192 + ------------------------------------------------------------------------------------- + TOTAL 0.08644642 284.54988977 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 123868 +BPFP 1.4596 bits/point +EBPFP 2.9192 equivalent bits/point +MSE 284.549890 +---------------------- -------------------------------------------------------- +Time: 3.430s Load: 0.005s, Pack+Encode: 2.045s, Decode+Unpack: 1.380s +---------------------- -------------------------------------------------------- +💾 Converting with 284.5499 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,184B, BPFP=0.7756 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,828B, BPFP=3.1349 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,876B, BPFP=1.3764 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,740B, BPFP=3.1037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,372B, BPFP=1.5526 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,612B, BPFP=3.0582 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,128B, BPFP=1.4659 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,744B, BPFP=3.1051 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,488B, BPFP=2.6591 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,568B, BPFP=3.0426 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,080B, BPFP=0.5114 +⌛️ [2/4] FRONTEND: Frontend time: 1.817s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.122s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13243668 43.81263317 + layer.0.v_cache 0.00001576 0.00854381 + layer.1.k_cache 0.01981235 9.28330924 + layer.1.v_cache 0.00000577 0.00318601 + layer.2.k_cache 0.01756107 1.09446118 + layer.2.v_cache 0.00001728 0.00851985 + layer.3.k_cache 0.04920139 4.85959140 + layer.3.v_cache 0.00001837 0.01009415 + layer.4.k_cache 0.00061593 0.22742271 + layer.4.v_cache 0.00004907 0.02147853 + layer.4.output 0.30854983 1123.67887581 + ------------------------------------------------------------------------------------- + TOTAL 0.13997544 466.18125710 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 75620 +BPFP 1.5796 bits/point +EBPFP 3.1593 equivalent bits/point +MSE 466.181257 +---------------------- -------------------------------------------------------- +Time: 2.942s Load: 0.003s, Pack+Encode: 1.817s, Decode+Unpack: 1.122s +---------------------- -------------------------------------------------------- +💾 Converting with 466.1813 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,196B, BPFP=0.6863 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,084B, BPFP=2.8388 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,212B, BPFP=1.3162 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,616B, BPFP=2.6925 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,964B, BPFP=1.8638 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,604B, BPFP=2.6888 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,084B, BPFP=1.2763 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,792B, BPFP=2.7475 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,904B, BPFP=2.4700 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,512B, BPFP=2.6600 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,964B, BPFP=0.4895 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.125s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20666794 39.21616211 + layer.0.v_cache 0.00002565 0.00922439 + layer.1.k_cache 0.01518770 8.52504822 + layer.1.v_cache 0.00000581 0.00319182 + layer.2.k_cache 0.01031687 0.93721786 + layer.2.v_cache 0.00001873 0.00905705 + layer.3.k_cache 0.09638391 4.29981842 + layer.3.v_cache 0.00001967 0.01232076 + layer.4.k_cache 0.00062172 0.22059851 + layer.4.v_cache 0.00004827 0.02181848 + layer.4.output 0.27162739 986.19348214 + ------------------------------------------------------------------------------------- + TOTAL 0.13121694 409.21228427 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 78932 +BPFP 1.4510 bits/point +EBPFP 2.9019 equivalent bits/point +MSE 409.212284 +---------------------- -------------------------------------------------------- +Time: 2.966s Load: 0.003s, Pack+Encode: 1.838s, Decode+Unpack: 1.125s +---------------------- -------------------------------------------------------- +💾 Converting with 409.2123 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,728B, BPFP=0.6268 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,196B, BPFP=2.8024 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,352B, BPFP=1.2298 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,312B, BPFP=2.5993 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,852B, BPFP=1.3447 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,080B, BPFP=2.5460 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,704B, BPFP=1.3107 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,644B, BPFP=2.6756 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,288B, BPFP=2.1342 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,288B, BPFP=2.5938 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,208B, BPFP=0.4992 +⌛️ [2/4] FRONTEND: Frontend time: 2.025s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13431002 41.29576201 + layer.0.v_cache 0.00001720 0.00795325 + layer.1.k_cache 0.01179634 8.26033110 + layer.1.v_cache 0.00000537 0.00286346 + layer.2.k_cache 0.00798416 0.80232205 + layer.2.v_cache 0.00001783 0.00820348 + layer.3.k_cache 0.05795463 4.07876138 + layer.3.v_cache 0.00001974 0.01013887 + layer.4.k_cache 0.00062059 0.20143994 + layer.4.v_cache 0.00005204 0.01862793 + layer.4.output 0.19989206 780.52934611 + ------------------------------------------------------------------------------------- + TOTAL 0.09482484 324.61128390 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 101652 +BPFP 1.3740 bits/point +EBPFP 2.7479 equivalent bits/point +MSE 324.611284 +---------------------- -------------------------------------------------------- +Time: 3.301s Load: 0.004s, Pack+Encode: 2.025s, Decode+Unpack: 1.272s +---------------------- -------------------------------------------------------- +💾 Converting with 324.6113 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,208B, BPFP=0.6900 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,024B, BPFP=2.8200 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,552B, BPFP=1.4225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,656B, BPFP=2.7050 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,536B, BPFP=1.7300 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,536B, BPFP=2.6675 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,276B, BPFP=1.6487 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,784B, BPFP=2.7450 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,564B, BPFP=2.3638 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,620B, BPFP=2.6938 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,968B, BPFP=0.4896 +⌛️ [2/4] FRONTEND: Frontend time: 1.703s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.181s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15467889 38.93844727 + layer.0.v_cache 0.00001741 0.00925566 + layer.1.k_cache 0.01620592 8.49069458 + layer.1.v_cache 0.00000548 0.00317213 + layer.2.k_cache 0.00723796 1.09138382 + layer.2.v_cache 0.00001737 0.00897685 + layer.3.k_cache 0.06661982 5.51653809 + layer.3.v_cache 0.00002040 0.01209866 + layer.4.k_cache 0.00058776 0.21994242 + layer.4.v_cache 0.00004656 0.02193662 + layer.4.output 0.27162074 933.16750000 + ------------------------------------------------------------------------------------- + TOTAL 0.12628134 387.44029095 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 79724 +BPFP 1.4655 bits/point +EBPFP 2.9310 equivalent bits/point +MSE 387.440291 +---------------------- -------------------------------------------------------- +Time: 2.886s Load: 0.002s, Pack+Encode: 1.703s, Decode+Unpack: 1.181s +---------------------- -------------------------------------------------------- +💾 Converting with 387.4403 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,232B, BPFP=0.6838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,012B, BPFP=2.7610 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,672B, BPFP=1.4314 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,672B, BPFP=2.6569 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,324B, BPFP=1.9375 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,368B, BPFP=2.5637 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,780B, BPFP=1.7708 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,732B, BPFP=2.6752 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,624B, BPFP=2.3358 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,568B, BPFP=2.6250 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,056B, BPFP=0.4401 +⌛️ [2/4] FRONTEND: Frontend time: 1.770s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11261671 36.99169443 + layer.0.v_cache 0.00001829 0.00808652 + layer.1.k_cache 0.01190703 9.26247391 + layer.1.v_cache 0.00000528 0.00299221 + layer.2.k_cache 0.00973116 1.08785637 + layer.2.v_cache 0.00001668 0.00807569 + layer.3.k_cache 0.04200072 5.43674963 + layer.3.v_cache 0.00001874 0.01080826 + layer.4.k_cache 0.00062040 0.21131091 + layer.4.v_cache 0.00004776 0.02048177 + layer.4.output 0.26626884 1025.27179622 + ------------------------------------------------------------------------------------- + TOTAL 0.12005086 425.29077078 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 80040 +BPFP 1.4425 bits/point +EBPFP 2.8849 equivalent bits/point +MSE 425.290771 +---------------------- -------------------------------------------------------- +Time: 3.101s Load: 0.003s, Pack+Encode: 1.770s, Decode+Unpack: 1.327s +---------------------- -------------------------------------------------------- +💾 Converting with 425.2908 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,268B, BPFP=0.6815 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,268B, BPFP=2.7849 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,948B, BPFP=1.4868 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,848B, BPFP=2.6587 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,204B, BPFP=1.8642 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,616B, BPFP=2.5889 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,200B, BPFP=1.8630 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,884B, BPFP=2.6695 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,388B, BPFP=2.2200 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,680B, BPFP=2.6082 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,656B, BPFP=0.5433 +⌛️ [2/4] FRONTEND: Frontend time: 1.746s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12932784 41.32878230 + layer.0.v_cache 0.00001491 0.00873794 + layer.1.k_cache 0.01488123 9.24044858 + layer.1.v_cache 0.00000563 0.00320716 + layer.2.k_cache 0.00939683 1.08754129 + layer.2.v_cache 0.00001981 0.00914789 + layer.3.k_cache 0.07051506 4.19145027 + layer.3.v_cache 0.00001927 0.01106160 + layer.4.k_cache 0.00064413 0.22676224 + layer.4.v_cache 0.00005939 0.02107520 + layer.4.output 0.26127321 1014.96583104 + ------------------------------------------------------------------------------------- + TOTAL 0.12081156 421.22876658 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 83960 +BPFP 1.4840 bits/point +EBPFP 2.9680 equivalent bits/point +MSE 421.228767 +---------------------- -------------------------------------------------------- +Time: 2.960s Load: 0.002s, Pack+Encode: 1.746s, Decode+Unpack: 1.211s +---------------------- -------------------------------------------------------- +💾 Converting with 421.2288 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,288B, BPFP=0.6875 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,284B, BPFP=2.7897 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,612B, BPFP=1.6863 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,028B, BPFP=2.7127 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,972B, BPFP=2.0950 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,708B, BPFP=2.6166 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,748B, BPFP=2.0276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,020B, BPFP=2.7103 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,976B, BPFP=2.3966 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,744B, BPFP=2.6274 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,256B, BPFP=0.5261 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13190371 41.39474722 + layer.0.v_cache 0.00001552 0.00857069 + layer.1.k_cache 0.01668521 8.65149923 + layer.1.v_cache 0.00000581 0.00323899 + layer.2.k_cache 0.00490881 1.04949100 + layer.2.v_cache 0.00001873 0.00895019 + layer.3.k_cache 0.11824917 5.45010845 + layer.3.v_cache 0.00002115 0.01114158 + layer.4.k_cache 0.00061206 0.21980656 + layer.4.v_cache 0.00004844 0.01984625 + layer.4.output 0.26122982 1024.77034684 + ------------------------------------------------------------------------------------- + TOTAL 0.12359279 425.30646047 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 86636 +BPFP 1.5313 bits/point +EBPFP 3.0626 equivalent bits/point +MSE 425.306460 +---------------------- -------------------------------------------------------- +Time: 2.950s Load: 0.003s, Pack+Encode: 1.730s, Decode+Unpack: 1.217s +---------------------- -------------------------------------------------------- +💾 Converting with 425.3065 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,372B, BPFP=0.5727 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,356B, BPFP=2.9477 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,884B, BPFP=1.1692 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,496B, BPFP=2.8016 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,220B, BPFP=1.2262 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,616B, BPFP=2.4823 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,164B, BPFP=1.2167 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,632B, BPFP=2.6549 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,288B, BPFP=2.2568 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,632B, BPFP=2.8247 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,048B, BPFP=0.4136 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.226s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10326390 43.83174465 + layer.0.v_cache 0.00001541 0.00673632 + layer.1.k_cache 0.04976616 8.29056715 + layer.1.v_cache 0.00000534 0.00235856 + layer.2.k_cache 0.01168035 0.91821364 + layer.2.v_cache 0.00001798 0.00654197 + layer.3.k_cache 0.02907468 3.67804154 + layer.3.v_cache 0.00001946 0.00810599 + layer.4.k_cache 0.00073861 0.18361185 + layer.4.v_cache 0.00005077 0.01641405 + layer.4.output 0.14788401 573.22792120 + ------------------------------------------------------------------------------------- + TOTAL 0.07234240 239.38457553 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 135708 +BPFP 1.3558 bits/point +EBPFP 2.7116 equivalent bits/point +MSE 239.384576 +---------------------- -------------------------------------------------------- +Time: 3.071s Load: 0.004s, Pack+Encode: 1.841s, Decode+Unpack: 1.226s +---------------------- -------------------------------------------------------- +💾 Converting with 239.3846 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 64, 128) +Output shape: (1, 64, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.output: torch.Size([1, 64, 3584]) -> torch.Size([1, 1, 64, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,072B, BPFP=0.5059 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,124B, BPFP=2.2275 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,272B, BPFP=1.2871 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,804B, BPFP=2.1494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,480B, BPFP=1.3379 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,712B, BPFP=2.1270 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,244B, BPFP=1.2803 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,056B, BPFP=2.2109 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,484B, BPFP=1.8271 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,712B, BPFP=2.1270 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,756B, BPFP=0.3751 +⌛️ [2/4] FRONTEND: Frontend time: 1.788s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.227s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09722494 47.60351181 + layer.0.v_cache 0.00001391 0.00533193 + layer.1.k_cache 0.01806639 10.37802696 + layer.1.v_cache 0.00000557 0.00200564 + layer.2.k_cache 0.00423524 1.09083509 + layer.2.v_cache 0.00001868 0.00622276 + layer.3.k_cache 0.03759564 5.74566126 + layer.3.v_cache 0.00001825 0.00680461 + layer.4.k_cache 0.00061753 0.17817059 + layer.4.v_cache 0.00004894 0.01277922 + layer.4.output 0.21408452 780.14843750 + ------------------------------------------------------------------------------------- + TOTAL 0.09743746 325.06284779 + (elements=557,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 557056 +Total Bytes 80716 +BPFP 1.1592 bits/point +EBPFP 2.3184 equivalent bits/point +MSE 325.062848 +---------------------- -------------------------------------------------------- +Time: 3.019s Load: 0.003s, Pack+Encode: 1.788s, Decode+Unpack: 1.227s +---------------------- -------------------------------------------------------- +💾 Converting with 325.0628 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,732B, BPFP=0.5717 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,220B, BPFP=2.7911 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,668B, BPFP=1.1746 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,268B, BPFP=2.6452 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,248B, BPFP=1.2635 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,824B, BPFP=2.5772 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,920B, BPFP=1.2132 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,292B, BPFP=2.6489 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,340B, BPFP=2.3499 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,044B, BPFP=2.6109 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,924B, BPFP=0.3485 +⌛️ [2/4] FRONTEND: Frontend time: 1.928s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.339s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11648470 42.61809704 + layer.0.v_cache 0.00001454 0.00598961 + layer.1.k_cache 0.02721847 8.43316890 + layer.1.v_cache 0.00000531 0.00215789 + layer.2.k_cache 0.01045542 0.86827746 + layer.2.v_cache 0.00001699 0.00614776 + layer.3.k_cache 0.01744144 4.54821179 + layer.3.v_cache 0.00001839 0.00743656 + layer.4.k_cache 0.00076342 0.19238670 + layer.4.v_cache 0.00004795 0.01559254 + layer.4.output 11.17676328 499.65121674 + ------------------------------------------------------------------------------------- + TOTAL 4.61234174 209.07388138 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 145480 +BPFP 1.3109 bits/point +EBPFP 2.6218 equivalent bits/point +MSE 209.073881 +---------------------- -------------------------------------------------------- +Time: 3.271s Load: 0.004s, Pack+Encode: 1.928s, Decode+Unpack: 1.339s +---------------------- -------------------------------------------------------- +💾 Converting with 209.0739 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,148B, BPFP=0.6148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,428B, BPFP=2.6227 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,148B, BPFP=1.2008 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,632B, BPFP=2.4672 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,488B, BPFP=1.2672 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,840B, BPFP=2.3125 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,364B, BPFP=1.2430 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,076B, BPFP=2.5539 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,240B, BPFP=2.0000 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,648B, BPFP=2.4703 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,512B, BPFP=0.4049 +⌛️ [2/4] FRONTEND: Frontend time: 2.194s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.490s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11433966 44.31491394 + layer.0.v_cache 0.00001437 0.00619915 + layer.1.k_cache 0.05315235 9.55365143 + layer.1.v_cache 0.00000524 0.00217052 + layer.2.k_cache 0.00678293 0.84852161 + layer.2.v_cache 0.00001704 0.00643572 + layer.3.k_cache 0.01951189 3.88142509 + layer.3.v_cache 0.00001758 0.00737514 + layer.4.k_cache 0.00076624 0.17712969 + layer.4.v_cache 0.00004856 0.01562271 + layer.4.output 0.16992235 651.71065848 + ------------------------------------------------------------------------------------- + TOTAL 0.08141837 271.81106202 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 110524 +BPFP 1.2698 bits/point +EBPFP 2.5396 equivalent bits/point +MSE 271.811062 +---------------------- -------------------------------------------------------- +Time: 3.688s Load: 0.004s, Pack+Encode: 2.194s, Decode+Unpack: 1.490s +---------------------- -------------------------------------------------------- +💾 Converting with 271.8111 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,316B, BPFP=0.5757 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,640B, BPFP=3.0625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,716B, BPFP=1.1660 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,068B, BPFP=2.9632 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,148B, BPFP=1.2410 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,148B, BPFP=2.8035 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,100B, BPFP=1.2326 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,056B, BPFP=2.9611 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,372B, BPFP=2.3215 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,864B, BPFP=2.9278 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,852B, BPFP=0.3932 +⌛️ [2/4] FRONTEND: Frontend time: 1.771s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11570581 41.96143392 + layer.0.v_cache 0.00001517 0.00639070 + layer.1.k_cache 0.03236232 8.60039537 + layer.1.v_cache 0.00000662 0.00220102 + layer.2.k_cache 0.00641356 0.75663537 + layer.2.v_cache 0.00001676 0.00644535 + layer.3.k_cache 0.02648412 4.24502563 + layer.3.v_cache 0.00001896 0.00796093 + layer.4.k_cache 0.00074524 0.18786320 + layer.4.v_cache 0.00004843 0.01546847 + layer.4.output 0.15113260 581.56031746 + ------------------------------------------------------------------------------------- + TOTAL 0.07292619 242.74776719 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 138280 +BPFP 1.4122 bits/point +EBPFP 2.8243 equivalent bits/point +MSE 242.747767 +---------------------- -------------------------------------------------------- +Time: 3.155s Load: 0.004s, Pack+Encode: 1.771s, Decode+Unpack: 1.380s +---------------------- -------------------------------------------------------- +💾 Converting with 242.7478 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 114, 128) +Output shape: (1, 114, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.output: torch.Size([1, 114, 3584]) -> torch.Size([1, 1, 114, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,996B, BPFP=0.5477 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,272B, BPFP=2.5044 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,268B, BPFP=1.1332 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,184B, BPFP=2.3553 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,020B, BPFP=1.2363 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,384B, BPFP=2.2456 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,772B, BPFP=1.2023 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,276B, BPFP=2.3679 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,468B, BPFP=2.1201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,300B, BPFP=2.3712 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,676B, BPFP=0.3657 +⌛️ [2/4] FRONTEND: Frontend time: 1.964s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.301s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10081583 43.33295641 + layer.0.v_cache 0.00001633 0.00615668 + layer.1.k_cache 0.08473034 9.19558876 + layer.1.v_cache 0.00000563 0.00213265 + layer.2.k_cache 0.00792928 0.88161991 + layer.2.v_cache 0.00001730 0.00631158 + layer.3.k_cache 0.01019665 5.11226346 + layer.3.v_cache 0.00001810 0.00690542 + layer.4.k_cache 0.00077074 0.18921446 + layer.4.v_cache 0.00005022 0.01564505 + layer.4.output 10.00037814 421.18190006 + ------------------------------------------------------------------------------------- + TOTAL 4.12983514 176.88365264 + (elements=992,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 992256 +Total Bytes 150616 +BPFP 1.2143 bits/point +EBPFP 2.4287 equivalent bits/point +MSE 176.883653 +---------------------- -------------------------------------------------------- +Time: 3.272s Load: 0.007s, Pack+Encode: 1.964s, Decode+Unpack: 1.301s +---------------------- -------------------------------------------------------- +💾 Converting with 176.8837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 98, 128) +Output shape: (1, 98, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.output: torch.Size([1, 98, 3584]) -> torch.Size([1, 1, 98, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,416B, BPFP=0.5446 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,984B, BPFP=2.8673 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,228B, BPFP=1.1524 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,324B, BPFP=2.7621 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,224B, BPFP=1.3112 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,092B, BPFP=2.5657 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,916B, BPFP=1.2621 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,188B, BPFP=2.7404 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,908B, BPFP=2.3769 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,880B, BPFP=2.6913 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,456B, BPFP=0.3520 +⌛️ [2/4] FRONTEND: Frontend time: 1.891s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11008905 43.59146803 + layer.0.v_cache 0.00001718 0.00620230 + layer.1.k_cache 0.06240322 8.93126071 + layer.1.v_cache 0.00000548 0.00213093 + layer.2.k_cache 0.00473963 0.76242610 + layer.2.v_cache 0.00001718 0.00631945 + layer.3.k_cache 0.01141277 5.46705098 + layer.3.v_cache 0.00001769 0.00707151 + layer.4.k_cache 0.00076610 0.18393053 + layer.4.v_cache 0.00004814 0.01505729 + layer.4.output 0.02426045 555.18727223 + ------------------------------------------------------------------------------------- + TOTAL 0.02113762 232.07551903 + (elements=852,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 852992 +Total Bytes 142616 +BPFP 1.3376 bits/point +EBPFP 2.6751 equivalent bits/point +MSE 232.075519 +---------------------- -------------------------------------------------------- +Time: 3.220s Load: 0.004s, Pack+Encode: 1.891s, Decode+Unpack: 1.325s +---------------------- -------------------------------------------------------- +💾 Converting with 232.0755 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,356B, BPFP=0.5700 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,300B, BPFP=2.9382 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,080B, BPFP=1.2024 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,416B, BPFP=2.7880 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,412B, BPFP=1.2588 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,356B, BPFP=2.6080 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,300B, BPFP=1.2398 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,448B, BPFP=2.7935 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,352B, BPFP=2.2677 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,160B, BPFP=2.7446 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,180B, BPFP=0.4168 +⌛️ [2/4] FRONTEND: Frontend time: 1.836s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.302s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08919133 42.91490372 + layer.0.v_cache 0.00001414 0.00611254 + layer.1.k_cache 0.08591949 8.63000886 + layer.1.v_cache 0.00000528 0.00219060 + layer.2.k_cache 0.00389855 0.92635395 + layer.2.v_cache 0.00001738 0.00628164 + layer.3.k_cache 0.07679920 4.21657396 + layer.3.v_cache 0.00002028 0.00764679 + layer.4.k_cache 0.00072407 0.17186393 + layer.4.v_cache 0.00004779 0.01507576 + layer.4.output 0.14786680 573.99849573 + ------------------------------------------------------------------------------------- + TOTAL 0.07598265 239.69920482 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 137360 +BPFP 1.3723 bits/point +EBPFP 2.7446 equivalent bits/point +MSE 239.699205 +---------------------- -------------------------------------------------------- +Time: 3.142s Load: 0.004s, Pack+Encode: 1.836s, Decode+Unpack: 1.302s +---------------------- -------------------------------------------------------- +💾 Converting with 239.6992 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,880B, BPFP=0.5666 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,712B, BPFP=2.7325 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,764B, BPFP=1.2798 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,944B, BPFP=2.6203 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,628B, BPFP=1.4060 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,464B, BPFP=2.5502 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,816B, BPFP=1.2874 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,936B, BPFP=2.6192 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,012B, BPFP=2.1922 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,620B, BPFP=2.5730 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,472B, BPFP=0.3853 +⌛️ [2/4] FRONTEND: Frontend time: 1.768s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.292s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11353527 41.54809342 + layer.0.v_cache 0.00001674 0.00625366 + layer.1.k_cache 0.04276301 9.47495277 + layer.1.v_cache 0.00000542 0.00214271 + layer.2.k_cache 0.00247095 0.99740601 + layer.2.v_cache 0.00001885 0.00670863 + layer.3.k_cache 0.02722364 4.32039150 + layer.3.v_cache 0.00001942 0.00710252 + layer.4.k_cache 0.00068256 0.18940949 + layer.4.v_cache 0.00004843 0.01554901 + layer.4.output 10.65454330 475.19730474 + ------------------------------------------------------------------------------------- + TOTAL 4.39815220 198.99700841 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 154248 +BPFP 1.3250 bits/point +EBPFP 2.6499 equivalent bits/point +MSE 198.997008 +---------------------- -------------------------------------------------------- +Time: 3.065s Load: 0.005s, Pack+Encode: 1.768s, Decode+Unpack: 1.292s +---------------------- -------------------------------------------------------- +💾 Converting with 198.9970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,252B, BPFP=0.6049 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,560B, BPFP=2.7083 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,396B, BPFP=1.1897 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,068B, BPFP=2.4308 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,736B, BPFP=1.2530 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,524B, BPFP=2.3296 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,448B, BPFP=1.1994 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,288B, BPFP=2.4717 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,696B, BPFP=1.9896 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,008B, BPFP=2.4196 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,656B, BPFP=0.3895 +⌛️ [2/4] FRONTEND: Frontend time: 1.867s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.287s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12550225 39.91749791 + layer.0.v_cache 0.00001468 0.00634556 + layer.1.k_cache 0.05234316 8.76235090 + layer.1.v_cache 0.00000508 0.00211575 + layer.2.k_cache 0.00733058 0.91788846 + layer.2.v_cache 0.00001689 0.00661184 + layer.3.k_cache 0.01592888 3.55066427 + layer.3.v_cache 0.00001847 0.00811091 + layer.4.k_cache 0.00069460 0.16879536 + layer.4.v_cache 0.00005205 0.01539646 + layer.4.output 0.16184913 627.20801446 + ------------------------------------------------------------------------------------- + TOTAL 0.07852062 261.40069874 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 114632 +BPFP 1.2543 bits/point +EBPFP 2.5086 equivalent bits/point +MSE 261.400699 +---------------------- -------------------------------------------------------- +Time: 3.157s Load: 0.004s, Pack+Encode: 1.867s, Decode+Unpack: 1.287s +---------------------- -------------------------------------------------------- +💾 Converting with 261.4007 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,128B, BPFP=0.6266 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,860B, BPFP=2.5761 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,988B, BPFP=1.1995 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,336B, BPFP=2.4712 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,352B, BPFP=1.2724 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,176B, BPFP=2.2388 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,304B, BPFP=1.2628 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,068B, BPFP=2.4175 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,176B, BPFP=2.0385 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,652B, BPFP=2.5345 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,736B, BPFP=0.3931 +⌛️ [2/4] FRONTEND: Frontend time: 1.789s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.172s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11879376 38.84210988 + layer.0.v_cache 0.00001438 0.00672815 + layer.1.k_cache 0.05518859 8.63139695 + layer.1.v_cache 0.00000514 0.00232459 + layer.2.k_cache 0.00243519 0.89521144 + layer.2.v_cache 0.00001742 0.00648359 + layer.3.k_cache 0.07959401 4.06143345 + layer.3.v_cache 0.00001812 0.00771743 + layer.4.k_cache 0.00069358 0.17512811 + layer.4.v_cache 0.00004620 0.01545175 + layer.4.output 0.17426339 676.70947802 + ------------------------------------------------------------------------------------- + TOTAL 0.08686177 281.74178421 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 106776 +BPFP 1.2582 bits/point +EBPFP 2.5164 equivalent bits/point +MSE 281.741784 +---------------------- -------------------------------------------------------- +Time: 2.965s Load: 0.003s, Pack+Encode: 1.789s, Decode+Unpack: 1.172s +---------------------- -------------------------------------------------------- +💾 Converting with 281.7418 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,176B, BPFP=0.6052 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,084B, BPFP=2.6837 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,388B, BPFP=1.2172 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,608B, BPFP=2.5930 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,624B, BPFP=1.2622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,760B, BPFP=2.4314 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,492B, BPFP=1.2370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,696B, BPFP=2.6098 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,660B, BPFP=2.0312 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,680B, BPFP=2.6067 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,868B, BPFP=0.3775 +⌛️ [2/4] FRONTEND: Frontend time: 1.788s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.268s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11325797 40.86600252 + layer.0.v_cache 0.00001565 0.00658052 + layer.1.k_cache 0.07651102 8.10753873 + layer.1.v_cache 0.00000530 0.00244049 + layer.2.k_cache 0.00872486 0.91994997 + layer.2.v_cache 0.00001696 0.00683246 + layer.3.k_cache 0.12876377 3.64872035 + layer.3.v_cache 0.00001905 0.00853984 + layer.4.k_cache 0.00069190 0.19340585 + layer.4.v_cache 0.00004968 0.01683488 + layer.4.output 0.16578155 590.65837326 + ------------------------------------------------------------------------------------- + TOTAL 0.08756041 246.37561520 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 115036 +BPFP 1.2894 bits/point +EBPFP 2.5788 equivalent bits/point +MSE 246.375615 +---------------------- -------------------------------------------------------- +Time: 3.059s Load: 0.003s, Pack+Encode: 1.788s, Decode+Unpack: 1.268s +---------------------- -------------------------------------------------------- +💾 Converting with 246.3756 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.6250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,424B, BPFP=2.5211 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,000B, BPFP=1.2175 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,704B, BPFP=2.5779 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,288B, BPFP=1.2760 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,064B, BPFP=2.2451 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,088B, BPFP=1.2354 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,688B, BPFP=2.3718 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,020B, BPFP=2.0333 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,820B, BPFP=2.6015 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,308B, BPFP=0.3858 +⌛️ [2/4] FRONTEND: Frontend time: 1.865s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.273s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12835005 42.99421672 + layer.0.v_cache 0.00001616 0.00645798 + layer.1.k_cache 0.05672838 8.70932998 + layer.1.v_cache 0.00000535 0.00258504 + layer.2.k_cache 0.00240846 0.98459615 + layer.2.v_cache 0.00001644 0.00682924 + layer.3.k_cache 0.04570918 3.96242573 + layer.3.v_cache 0.00001733 0.00823830 + layer.4.k_cache 0.00072090 0.19416180 + layer.4.v_cache 0.00004908 0.01660555 + layer.4.output 0.18444509 665.52382885 + ------------------------------------------------------------------------------------- + TOTAL 0.08971394 277.38542638 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 105484 +BPFP 1.2591 bits/point +EBPFP 2.5182 equivalent bits/point +MSE 277.385426 +---------------------- -------------------------------------------------------- +Time: 3.144s Load: 0.006s, Pack+Encode: 1.865s, Decode+Unpack: 1.273s +---------------------- -------------------------------------------------------- +💾 Converting with 277.3854 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,892B, BPFP=0.6364 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,984B, BPFP=2.6373 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,628B, BPFP=1.2386 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,024B, BPFP=2.4261 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,660B, BPFP=1.2456 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,572B, BPFP=2.3266 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,368B, BPFP=1.1813 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,324B, BPFP=2.4921 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,264B, BPFP=2.0387 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,704B, BPFP=2.5757 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,964B, BPFP=0.4076 +⌛️ [2/4] FRONTEND: Frontend time: 1.804s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10085608 42.53790713 + layer.0.v_cache 0.00001554 0.00655087 + layer.1.k_cache 0.06589249 8.76310107 + layer.1.v_cache 0.00000499 0.00243163 + layer.2.k_cache 0.00422055 0.91451575 + layer.2.v_cache 0.00001579 0.00662791 + layer.3.k_cache 0.05337993 3.97389114 + layer.3.v_cache 0.00001747 0.00832501 + layer.4.k_cache 0.00075412 0.19451119 + layer.4.v_cache 0.00004712 0.01614660 + layer.4.output 0.19135183 739.24126006 + ------------------------------------------------------------------------------------- + TOTAL 0.09203923 307.71251934 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 98384 +BPFP 1.2736 bits/point +EBPFP 2.5472 equivalent bits/point +MSE 307.712519 +---------------------- -------------------------------------------------------- +Time: 3.178s Load: 0.004s, Pack+Encode: 1.804s, Decode+Unpack: 1.369s +---------------------- -------------------------------------------------------- +💾 Converting with 307.7125 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,240B, BPFP=0.6027 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,272B, BPFP=2.6548 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,476B, BPFP=1.2046 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,944B, BPFP=2.4077 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,784B, BPFP=1.2619 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,344B, BPFP=2.2961 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,788B, BPFP=1.2626 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,476B, BPFP=2.5067 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,728B, BPFP=1.9955 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,064B, BPFP=2.4301 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,620B, BPFP=0.3619 +⌛️ [2/4] FRONTEND: Frontend time: 1.970s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.287s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09057366 40.22925967 + layer.0.v_cache 0.00001789 0.00623844 + layer.1.k_cache 0.03614600 7.48392305 + layer.1.v_cache 0.00000481 0.00222000 + layer.2.k_cache 0.00554357 0.87313770 + layer.2.v_cache 0.00001714 0.00652239 + layer.3.k_cache 0.02742965 3.80582682 + layer.3.v_cache 0.00001729 0.00760519 + layer.4.k_cache 0.00093027 0.17830989 + layer.4.v_cache 0.00004906 0.01482408 + layer.4.output 0.16441821 613.90056335 + ------------------------------------------------------------------------------------- + TOTAL 0.07715628 255.87716533 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 113736 +BPFP 1.2445 bits/point +EBPFP 2.4890 equivalent bits/point +MSE 255.877165 +---------------------- -------------------------------------------------------- +Time: 3.261s Load: 0.004s, Pack+Encode: 1.970s, Decode+Unpack: 1.287s +---------------------- -------------------------------------------------------- +💾 Converting with 255.8772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,088B, BPFP=0.6266 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,580B, BPFP=2.5528 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,972B, BPFP=1.2119 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,764B, BPFP=2.3872 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,268B, BPFP=1.2719 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,920B, BPFP=2.4188 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,180B, BPFP=1.2541 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,188B, BPFP=2.4732 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,092B, BPFP=2.0479 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,524B, BPFP=2.5414 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,048B, BPFP=0.4072 +⌛️ [2/4] FRONTEND: Frontend time: 1.774s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.266s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09325875 42.11668654 + layer.0.v_cache 0.00001406 0.00651456 + layer.1.k_cache 0.05775181 8.66429317 + layer.1.v_cache 0.00000524 0.00232987 + layer.2.k_cache 0.00413042 0.93041774 + layer.2.v_cache 0.00001743 0.00666793 + layer.3.k_cache 0.04387648 3.78120482 + layer.3.v_cache 0.00002497 0.00882393 + layer.4.k_cache 0.00071512 0.19017880 + layer.4.v_cache 0.00005147 0.01779656 + layer.4.output 0.17652170 676.35992579 + ------------------------------------------------------------------------------------- + TOTAL 0.08444104 281.77908203 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 106624 +BPFP 1.2727 bits/point +EBPFP 2.5455 equivalent bits/point +MSE 281.779082 +---------------------- -------------------------------------------------------- +Time: 3.045s Load: 0.005s, Pack+Encode: 1.774s, Decode+Unpack: 1.266s +---------------------- -------------------------------------------------------- +💾 Converting with 281.7791 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,328B, BPFP=0.5909 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,224B, BPFP=2.5256 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,528B, BPFP=1.1591 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,768B, BPFP=2.4446 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,672B, BPFP=1.1847 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,120B, BPFP=2.1520 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,508B, BPFP=1.1555 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,988B, BPFP=2.4837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,788B, BPFP=1.9155 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,696B, BPFP=2.6094 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,332B, BPFP=0.3128 +⌛️ [2/4] FRONTEND: Frontend time: 1.935s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212358 43.01338335 + layer.0.v_cache 0.00001682 0.00575366 + layer.1.k_cache 0.05918076 8.54229875 + layer.1.v_cache 0.00000473 0.00198228 + layer.2.k_cache 0.00260097 0.85228643 + layer.2.v_cache 0.00001544 0.00590251 + layer.3.k_cache 0.02838851 4.06435880 + layer.3.v_cache 0.00001657 0.00695589 + layer.4.k_cache 0.00106615 0.17879174 + layer.4.v_cache 0.00004456 0.01523751 + layer.4.output 0.17173813 584.71986607 + ------------------------------------------------------------------------------------- + TOTAL 0.08091912 244.10153020 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 114952 +BPFP 1.2006 bits/point +EBPFP 2.4012 equivalent bits/point +MSE 244.101530 +---------------------- -------------------------------------------------------- +Time: 3.247s Load: 0.003s, Pack+Encode: 1.935s, Decode+Unpack: 1.309s +---------------------- -------------------------------------------------------- +💾 Converting with 244.1015 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,300B, BPFP=0.5859 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,160B, BPFP=2.8693 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,724B, BPFP=1.1939 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,716B, BPFP=2.4354 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,088B, BPFP=1.2585 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,864B, BPFP=2.4616 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,848B, BPFP=1.2159 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,100B, BPFP=2.5036 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,560B, BPFP=2.0526 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,632B, BPFP=2.4205 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,812B, BPFP=0.3757 +⌛️ [2/4] FRONTEND: Frontend time: 1.738s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.255s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10731998 42.51646839 + layer.0.v_cache 0.00001618 0.00599402 + layer.1.k_cache 0.07375650 8.88167364 + layer.1.v_cache 0.00000540 0.00222403 + layer.2.k_cache 0.00246326 0.81664623 + layer.2.v_cache 0.00001746 0.00670099 + layer.3.k_cache 0.01642703 3.97793475 + layer.3.v_cache 0.00001778 0.00759101 + layer.4.k_cache 0.00066607 0.17242833 + layer.4.v_cache 0.00004728 0.01537673 + layer.4.output 0.15452130 567.87870333 + ------------------------------------------------------------------------------------- + TOTAL 0.07543447 237.15023302 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 121804 +BPFP 1.2722 bits/point +EBPFP 2.5444 equivalent bits/point +MSE 237.150233 +---------------------- -------------------------------------------------------- +Time: 2.996s Load: 0.003s, Pack+Encode: 1.738s, Decode+Unpack: 1.255s +---------------------- -------------------------------------------------------- +💾 Converting with 237.1502 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 99, 128) +Output shape: (1, 99, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.output: torch.Size([1, 99, 3584]) -> torch.Size([1, 1, 99, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,516B, BPFP=0.5549 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,572B, BPFP=2.7734 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,292B, BPFP=1.1509 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,672B, BPFP=2.6313 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,884B, BPFP=1.2443 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,136B, BPFP=2.5467 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,392B, BPFP=1.3245 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,784B, BPFP=2.6490 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,908B, BPFP=2.3529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,644B, BPFP=2.6269 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,744B, BPFP=0.3775 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.160s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09061131 42.92513021 + layer.0.v_cache 0.00001461 0.00587256 + layer.1.k_cache 0.05062923 8.50453016 + layer.1.v_cache 0.00000709 0.00215739 + layer.2.k_cache 0.00392854 0.84715795 + layer.2.v_cache 0.00001793 0.00601203 + layer.3.k_cache 0.05082532 5.11873958 + layer.3.v_cache 0.00001820 0.00730795 + layer.4.k_cache 0.00080398 0.19869629 + layer.4.v_cache 0.00005053 0.01518219 + layer.4.output 0.02403493 558.47537879 + ------------------------------------------------------------------------------------- + TOTAL 0.02147948 233.35049634 + (elements=861,696) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 861696 +Total Bytes 142544 +BPFP 1.3234 bits/point +EBPFP 2.6468 equivalent bits/point +MSE 233.350496 +---------------------- -------------------------------------------------------- +Time: 3.005s Load: 0.004s, Pack+Encode: 1.841s, Decode+Unpack: 1.160s +---------------------- -------------------------------------------------------- +💾 Converting with 233.3505 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,884B, BPFP=0.6347 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,104B, BPFP=2.8838 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,628B, BPFP=1.2386 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,256B, BPFP=2.6972 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,800B, BPFP=1.2764 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,756B, BPFP=2.5871 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,676B, BPFP=1.2491 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,320B, BPFP=2.7113 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,360B, BPFP=2.0599 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,924B, BPFP=2.6241 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,944B, BPFP=0.4384 +⌛️ [2/4] FRONTEND: Frontend time: 1.784s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.251s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11136492 42.24375894 + layer.0.v_cache 0.00001955 0.00672188 + layer.1.k_cache 0.06280328 7.84111259 + layer.1.v_cache 0.00000544 0.00257851 + layer.2.k_cache 0.00235174 0.90599554 + layer.2.v_cache 0.00001859 0.00746612 + layer.3.k_cache 0.03326307 4.71337848 + layer.3.v_cache 0.00001926 0.00937986 + layer.4.k_cache 0.00064473 0.18506493 + layer.4.v_cache 0.00005254 0.01789031 + layer.4.output 0.19494371 733.66486419 + ------------------------------------------------------------------------------------- + TOTAL 0.09265583 305.38749391 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 104652 +BPFP 1.3548 bits/point +EBPFP 2.7095 equivalent bits/point +MSE 305.387494 +---------------------- -------------------------------------------------------- +Time: 3.038s Load: 0.003s, Pack+Encode: 1.784s, Decode+Unpack: 1.251s +---------------------- -------------------------------------------------------- +💾 Converting with 305.3875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,976B, BPFP=0.6284 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,864B, BPFP=2.9274 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,708B, BPFP=1.2052 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,612B, BPFP=2.6630 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,088B, BPFP=1.2855 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,888B, BPFP=2.5101 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,852B, BPFP=1.2356 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,264B, BPFP=2.5895 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,824B, BPFP=2.0743 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,260B, BPFP=2.5887 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,096B, BPFP=0.3950 +⌛️ [2/4] FRONTEND: Frontend time: 1.784s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.274s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08829598 40.69330329 + layer.0.v_cache 0.00001455 0.00709570 + layer.1.k_cache 0.05950350 8.22829540 + layer.1.v_cache 0.00000529 0.00246293 + layer.2.k_cache 0.00411868 0.89286577 + layer.2.v_cache 0.00001680 0.00728366 + layer.3.k_cache 0.04877868 4.81603097 + layer.3.v_cache 0.00001896 0.00932221 + layer.4.k_cache 0.00063741 0.18634960 + layer.4.v_cache 0.00004764 0.01793827 + layer.4.output 0.18365649 697.55007239 + ------------------------------------------------------------------------------------- + TOTAL 0.08747253 290.45361497 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 106432 +BPFP 1.3219 bits/point +EBPFP 2.6439 equivalent bits/point +MSE 290.453615 +---------------------- -------------------------------------------------------- +Time: 3.061s Load: 0.003s, Pack+Encode: 1.784s, Decode+Unpack: 1.274s +---------------------- -------------------------------------------------------- +💾 Converting with 290.4536 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,052B, BPFP=0.6193 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,668B, BPFP=2.7735 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,112B, BPFP=1.2403 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,208B, BPFP=2.6802 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,220B, BPFP=1.2622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,144B, BPFP=2.4643 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,976B, BPFP=1.2127 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,988B, BPFP=2.6356 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,032B, BPFP=2.0357 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,160B, BPFP=2.6705 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,052B, BPFP=0.3784 +⌛️ [2/4] FRONTEND: Frontend time: 1.756s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.396s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11195674 40.29731889 + layer.0.v_cache 0.00001346 0.00604155 + layer.1.k_cache 0.03755230 8.24426904 + layer.1.v_cache 0.00000535 0.00233434 + layer.2.k_cache 0.00239600 0.97842269 + layer.2.v_cache 0.00001753 0.00735763 + layer.3.k_cache 0.02867049 3.93021660 + layer.3.v_cache 0.00001751 0.00838732 + layer.4.k_cache 0.00071811 0.19222993 + layer.4.v_cache 0.00005486 0.01881566 + layer.4.output 0.19495771 666.14465445 + ------------------------------------------------------------------------------------- + TOTAL 0.09094743 277.45282205 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 109612 +BPFP 1.3084 bits/point +EBPFP 2.6168 equivalent bits/point +MSE 277.452822 +---------------------- -------------------------------------------------------- +Time: 3.155s Load: 0.004s, Pack+Encode: 1.756s, Decode+Unpack: 1.396s +---------------------- -------------------------------------------------------- +💾 Converting with 277.4528 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,368B, BPFP=0.5847 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,952B, BPFP=2.9431 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,680B, BPFP=1.1597 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,012B, BPFP=2.7799 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,892B, BPFP=1.1965 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,396B, BPFP=2.4993 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,848B, BPFP=1.1889 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,388B, BPFP=2.8451 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,616B, BPFP=2.0167 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,756B, BPFP=2.5618 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,964B, BPFP=0.3463 +⌛️ [2/4] FRONTEND: Frontend time: 1.895s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.457s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08657128 39.72529297 + layer.0.v_cache 0.00001574 0.00640599 + layer.1.k_cache 0.05618976 8.48082614 + layer.1.v_cache 0.00000506 0.00222951 + layer.2.k_cache 0.00392605 0.70657980 + layer.2.v_cache 0.00001570 0.00613281 + layer.3.k_cache 0.02849737 4.20623711 + layer.3.v_cache 0.00001769 0.00799913 + layer.4.k_cache 0.00099738 0.19209281 + layer.4.v_cache 0.00004751 0.01618392 + layer.4.output 0.15107292 566.19141865 + ------------------------------------------------------------------------------------- + TOTAL 0.07257612 236.27587710 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 127872 +BPFP 1.3059 bits/point +EBPFP 2.6118 equivalent bits/point +MSE 236.275877 +---------------------- -------------------------------------------------------- +Time: 3.356s Load: 0.004s, Pack+Encode: 1.895s, Decode+Unpack: 1.457s +---------------------- -------------------------------------------------------- +💾 Converting with 236.2759 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,156B, BPFP=0.6164 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,736B, BPFP=2.6828 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,204B, BPFP=1.2117 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,884B, BPFP=2.5164 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,392B, BPFP=1.2484 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,480B, BPFP=2.4375 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,152B, BPFP=1.2016 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,820B, BPFP=2.8945 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,444B, BPFP=2.0398 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,472B, BPFP=2.6313 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,992B, BPFP=0.3625 +⌛️ [2/4] FRONTEND: Frontend time: 1.694s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.508s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09254595 43.26074219 + layer.0.v_cache 0.00001776 0.00595925 + layer.1.k_cache 0.05775445 8.12187500 + layer.1.v_cache 0.00000504 0.00220921 + layer.2.k_cache 0.00891839 0.82397652 + layer.2.v_cache 0.00001732 0.00684602 + layer.3.k_cache 0.02836140 3.88663559 + layer.3.v_cache 0.00001672 0.00791132 + layer.4.k_cache 0.00083566 0.17939482 + layer.4.v_cache 0.00004876 0.01789458 + layer.4.output 0.16993865 650.38777902 + ------------------------------------------------------------------------------------- + TOTAL 0.08106423 271.11928809 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 112732 +BPFP 1.2952 bits/point +EBPFP 2.5903 equivalent bits/point +MSE 271.119288 +---------------------- -------------------------------------------------------- +Time: 3.205s Load: 0.003s, Pack+Encode: 1.694s, Decode+Unpack: 1.508s +---------------------- -------------------------------------------------------- +💾 Converting with 271.1193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,400B, BPFP=0.5477 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,108B, BPFP=2.7558 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,084B, BPFP=1.1411 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,996B, BPFP=2.5767 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,240B, BPFP=1.1662 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,372B, BPFP=2.4762 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,100B, BPFP=1.1437 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,224B, BPFP=2.6134 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,736B, BPFP=2.2126 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,864B, BPFP=2.5554 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,060B, BPFP=0.3466 +⌛️ [2/4] FRONTEND: Frontend time: 2.440s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.713s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13093194 44.42512685 + layer.0.v_cache 0.00001635 0.00630398 + layer.1.k_cache 0.05235285 8.31790885 + layer.1.v_cache 0.00000496 0.00223724 + layer.2.k_cache 0.00512381 0.82295384 + layer.2.v_cache 0.00001647 0.00614328 + layer.3.k_cache 0.01138228 4.60212849 + layer.3.v_cache 0.00001850 0.00714207 + layer.4.k_cache 0.00081880 0.18106905 + layer.4.v_cache 0.00004796 0.01501850 + layer.4.output 0.02447439 558.91296944 + ------------------------------------------------------------------------------------- + TOTAL 0.02188439 233.57510695 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 134184 +BPFP 1.2715 bits/point +EBPFP 2.5429 equivalent bits/point +MSE 233.575107 +---------------------- -------------------------------------------------------- +Time: 4.158s Load: 0.005s, Pack+Encode: 2.440s, Decode+Unpack: 1.713s +---------------------- -------------------------------------------------------- +💾 Converting with 233.5751 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,156B, BPFP=0.6242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,124B, BPFP=2.7935 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,124B, BPFP=1.2112 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,268B, BPFP=2.4264 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,408B, BPFP=1.2674 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,692B, BPFP=2.3125 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,076B, BPFP=1.2017 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,348B, BPFP=2.4422 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,044B, BPFP=1.9866 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,660B, BPFP=2.5040 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,172B, BPFP=0.4004 +⌛️ [2/4] FRONTEND: Frontend time: 2.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.586s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510638 39.84975153 + layer.0.v_cache 0.00001399 0.00642994 + layer.1.k_cache 0.05467150 8.68665787 + layer.1.v_cache 0.00000511 0.00218393 + layer.2.k_cache 0.00545186 0.78875457 + layer.2.v_cache 0.00001652 0.00684245 + layer.3.k_cache 0.02986256 3.94199806 + layer.3.v_cache 0.00001747 0.00729409 + layer.4.k_cache 0.00068266 0.18407911 + layer.4.v_cache 0.00005277 0.01647125 + layer.4.output 0.17642128 663.74547920 + ------------------------------------------------------------------------------------- + TOTAL 0.08357822 276.45345984 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 109072 +BPFP 1.2690 bits/point +EBPFP 2.5380 equivalent bits/point +MSE 276.453460 +---------------------- -------------------------------------------------------- +Time: 3.753s Load: 0.006s, Pack+Encode: 2.160s, Decode+Unpack: 1.586s +---------------------- -------------------------------------------------------- +💾 Converting with 276.4535 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,088B, BPFP=0.6266 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,444B, BPFP=2.7281 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,032B, BPFP=1.2240 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,148B, BPFP=2.6680 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,256B, BPFP=1.2695 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,572B, BPFP=2.3482 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,036B, BPFP=1.2248 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,388B, BPFP=2.5138 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,904B, BPFP=2.0097 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,684B, BPFP=2.5739 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,600B, BPFP=0.3942 +⌛️ [2/4] FRONTEND: Frontend time: 2.074s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.548s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08730992 42.09645457 + layer.0.v_cache 0.00001543 0.00663930 + layer.1.k_cache 0.05635426 8.90522429 + layer.1.v_cache 0.00000520 0.00251029 + layer.2.k_cache 0.00390536 0.85360975 + layer.2.v_cache 0.00001661 0.00664930 + layer.3.k_cache 0.06483487 4.30570964 + layer.3.v_cache 0.00002010 0.00813685 + layer.4.k_cache 0.00072652 0.18525321 + layer.4.v_cache 0.00004842 0.01681285 + layer.4.output 0.18845482 661.11178108 + ------------------------------------------------------------------------------------- + TOTAL 0.09014238 275.53938045 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 108152 +BPFP 1.2910 bits/point +EBPFP 2.5819 equivalent bits/point +MSE 275.539380 +---------------------- -------------------------------------------------------- +Time: 3.627s Load: 0.005s, Pack+Encode: 2.074s, Decode+Unpack: 1.548s +---------------------- -------------------------------------------------------- +💾 Converting with 275.5394 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,716B, BPFP=0.6241 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,176B, BPFP=2.5680 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,292B, BPFP=1.2160 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,796B, BPFP=2.4807 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,304B, BPFP=1.2188 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,492B, BPFP=2.4108 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,296B, BPFP=1.2169 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,748B, BPFP=2.4697 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,608B, BPFP=1.9779 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,684B, BPFP=2.4550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,736B, BPFP=0.4509 +⌛️ [2/4] FRONTEND: Frontend time: 1.808s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10784986 42.88011977 + layer.0.v_cache 0.00001387 0.00615930 + layer.1.k_cache 0.01409798 8.28844856 + layer.1.v_cache 0.00000489 0.00222151 + layer.2.k_cache 0.00254294 0.98217291 + layer.2.v_cache 0.00001653 0.00692339 + layer.3.k_cache 0.05709259 3.86927167 + layer.3.v_cache 0.00001736 0.00806715 + layer.4.k_cache 0.00062715 0.17546062 + layer.4.v_cache 0.00004758 0.01669211 + layer.4.output 0.22376562 776.94918592 + ------------------------------------------------------------------------------------- + TOTAL 0.10286295 323.22822579 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 94848 +BPFP 1.2820 bits/point +EBPFP 2.5640 equivalent bits/point +MSE 323.228226 +---------------------- -------------------------------------------------------- +Time: 3.168s Load: 0.005s, Pack+Encode: 1.808s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 323.2282 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,144B, BPFP=0.6218 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,968B, BPFP=2.5649 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,060B, BPFP=1.1986 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,228B, BPFP=2.4185 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,300B, BPFP=1.2460 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,752B, BPFP=2.3244 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,112B, BPFP=1.2089 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,952B, BPFP=2.3639 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,840B, BPFP=1.9462 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,676B, BPFP=2.3093 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,496B, BPFP=0.4096 +⌛️ [2/4] FRONTEND: Frontend time: 1.861s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.279s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14563359 41.62986427 + layer.0.v_cache 0.00001368 0.00562033 + layer.1.k_cache 0.05947259 8.85595394 + layer.1.v_cache 0.00000496 0.00205409 + layer.2.k_cache 0.00553216 0.78272189 + layer.2.v_cache 0.00001557 0.00642669 + layer.3.k_cache 0.04377548 3.98132788 + layer.3.v_cache 0.00001730 0.00720441 + layer.4.k_cache 0.00079035 0.18379484 + layer.4.v_cache 0.00005153 0.01663299 + layer.4.output 0.17347779 664.45083635 + ------------------------------------------------------------------------------------- + TOTAL 0.08645010 276.86043857 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 106528 +BPFP 1.2394 bits/point +EBPFP 2.4788 equivalent bits/point +MSE 276.860439 +---------------------- -------------------------------------------------------- +Time: 3.144s Load: 0.003s, Pack+Encode: 1.861s, Decode+Unpack: 1.279s +---------------------- -------------------------------------------------------- +💾 Converting with 276.8604 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,600B, BPFP=0.6155 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,084B, BPFP=2.6241 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,876B, BPFP=1.1544 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,092B, BPFP=2.3892 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,128B, BPFP=1.2140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,960B, BPFP=2.3580 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,932B, BPFP=1.1676 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,612B, BPFP=2.5123 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,236B, BPFP=1.9498 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,028B, BPFP=2.3741 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,992B, BPFP=0.4394 +⌛️ [2/4] FRONTEND: Frontend time: 1.837s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.358s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12373332 40.28309215 + layer.0.v_cache 0.00001436 0.00666487 + layer.1.k_cache 0.01270404 8.05129080 + layer.1.v_cache 0.00000490 0.00228097 + layer.2.k_cache 0.00252672 0.90811527 + layer.2.v_cache 0.00001629 0.00673633 + layer.3.k_cache 0.05185594 3.86826625 + layer.3.v_cache 0.00001849 0.00805019 + layer.4.k_cache 0.00068458 0.17845191 + layer.4.v_cache 0.00004975 0.01634537 + layer.4.output 0.22283003 784.60876623 + ------------------------------------------------------------------------------------- + TOTAL 0.10302462 326.21121516 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 90540 +BPFP 1.2609 bits/point +EBPFP 2.5217 equivalent bits/point +MSE 326.211215 +---------------------- -------------------------------------------------------- +Time: 3.198s Load: 0.003s, Pack+Encode: 1.837s, Decode+Unpack: 1.358s +---------------------- -------------------------------------------------------- +💾 Converting with 326.2112 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,212B, BPFP=0.5975 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,280B, BPFP=2.6562 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,484B, BPFP=1.2061 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,792B, BPFP=2.5655 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,084B, BPFP=1.3177 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,252B, BPFP=2.4650 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,952B, BPFP=1.2932 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,628B, BPFP=2.5350 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,024B, BPFP=2.0506 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,880B, BPFP=2.5818 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,744B, BPFP=0.4184 +⌛️ [2/4] FRONTEND: Frontend time: 1.989s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.611s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12002775 42.34913853 + layer.0.v_cache 0.00001466 0.00641379 + layer.1.k_cache 0.05271155 7.93531145 + layer.1.v_cache 0.00000524 0.00232662 + layer.2.k_cache 0.00808975 0.83429282 + layer.2.v_cache 0.00001826 0.00690391 + layer.3.k_cache 0.07501186 3.55516815 + layer.3.v_cache 0.00001895 0.00806727 + layer.4.k_cache 0.00071216 0.18377968 + layer.4.v_cache 0.00005170 0.01735370 + layer.4.output 0.17053346 630.08599065 + ------------------------------------------------------------------------------------- + TOTAL 0.08531742 262.67651120 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 119332 +BPFP 1.3057 bits/point +EBPFP 2.6114 equivalent bits/point +MSE 262.676511 +---------------------- -------------------------------------------------------- +Time: 3.604s Load: 0.004s, Pack+Encode: 1.989s, Decode+Unpack: 1.611s +---------------------- -------------------------------------------------------- +💾 Converting with 262.6765 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 75, 128) +Output shape: (1, 75, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.output: torch.Size([1, 75, 3584]) -> torch.Size([1, 1, 75, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,012B, BPFP=0.6275 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,780B, BPFP=2.6625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,868B, BPFP=1.2225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,176B, BPFP=2.5367 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,240B, BPFP=1.3000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,160B, BPFP=2.3250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,916B, BPFP=1.2325 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,848B, BPFP=2.4683 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,708B, BPFP=2.0225 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,128B, BPFP=2.5267 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,664B, BPFP=0.4364 +⌛️ [2/4] FRONTEND: Frontend time: 1.970s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.263s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08149677 42.02102539 + layer.0.v_cache 0.00001391 0.00655388 + layer.1.k_cache 0.05841089 8.79023275 + layer.1.v_cache 0.00000541 0.00238785 + layer.2.k_cache 0.00252637 0.73845540 + layer.2.v_cache 0.00001693 0.00694370 + layer.3.k_cache 0.02876982 3.58000448 + layer.3.v_cache 0.00001932 0.00837082 + layer.4.k_cache 0.00075246 0.19010231 + layer.4.v_cache 0.00005041 0.01618490 + layer.4.output 0.19266937 691.10327381 + ------------------------------------------------------------------------------------- + TOTAL 0.08945576 287.82842224 + (elements=652,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 652800 +Total Bytes 105500 +BPFP 1.2929 bits/point +EBPFP 2.5858 equivalent bits/point +MSE 287.828422 +---------------------- -------------------------------------------------------- +Time: 3.235s Load: 0.003s, Pack+Encode: 1.970s, Decode+Unpack: 1.263s +---------------------- -------------------------------------------------------- +💾 Converting with 287.8284 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,692B, BPFP=0.6186 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,360B, BPFP=2.6103 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,200B, BPFP=1.1949 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,184B, BPFP=2.3401 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,372B, BPFP=1.2344 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,092B, BPFP=2.3189 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,136B, BPFP=1.1801 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,536B, BPFP=2.4210 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,696B, BPFP=1.9982 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,596B, BPFP=2.4347 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,248B, BPFP=0.4020 +⌛️ [2/4] FRONTEND: Frontend time: 1.837s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.251s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10404026 41.18429027 + layer.0.v_cache 0.00001395 0.00651795 + layer.1.k_cache 0.01458383 8.40196767 + layer.1.v_cache 0.00000499 0.00230160 + layer.2.k_cache 0.00236740 0.83793786 + layer.2.v_cache 0.00001641 0.00705661 + layer.3.k_cache 0.03645969 4.26910266 + layer.3.v_cache 0.00002420 0.00832238 + layer.4.k_cache 0.00069996 0.18596802 + layer.4.v_cache 0.00005119 0.01688555 + layer.4.output 0.20980707 780.68178834 + ------------------------------------------------------------------------------------- + TOTAL 0.09570067 324.68781582 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 92112 +BPFP 1.2450 bits/point +EBPFP 2.4901 equivalent bits/point +MSE 324.687816 +---------------------- -------------------------------------------------------- +Time: 3.091s Load: 0.003s, Pack+Encode: 1.837s, Decode+Unpack: 1.251s +---------------------- -------------------------------------------------------- +💾 Converting with 324.6878 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,812B, BPFP=0.6277 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,764B, BPFP=2.6259 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,496B, BPFP=1.2268 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,908B, BPFP=2.4348 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,664B, BPFP=1.2643 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,424B, BPFP=2.3268 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,548B, BPFP=1.2384 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,032B, BPFP=2.4625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,144B, BPFP=2.0411 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,096B, BPFP=2.4768 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,140B, BPFP=0.4509 +⌛️ [2/4] FRONTEND: Frontend time: 1.918s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.457s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11725240 41.53664551 + layer.0.v_cache 0.00001567 0.00630857 + layer.1.k_cache 0.03693308 9.03017404 + layer.1.v_cache 0.00000533 0.00236656 + layer.2.k_cache 0.00240566 0.94220232 + layer.2.v_cache 0.00001850 0.00734475 + layer.3.k_cache 0.06751672 3.79581604 + layer.3.v_cache 0.00001909 0.00791963 + layer.4.k_cache 0.00067984 0.17916817 + layer.4.v_cache 0.00004901 0.01625433 + layer.4.output 0.20094273 714.49821429 + ------------------------------------------------------------------------------------- + TOTAL 0.09597026 297.47127647 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 98028 +BPFP 1.2871 bits/point +EBPFP 2.5743 equivalent bits/point +MSE 297.471276 +---------------------- -------------------------------------------------------- +Time: 3.378s Load: 0.003s, Pack+Encode: 1.918s, Decode+Unpack: 1.457s +---------------------- -------------------------------------------------------- +💾 Converting with 297.4713 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,836B, BPFP=0.6330 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,292B, BPFP=2.5205 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,480B, BPFP=1.2232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,828B, BPFP=2.4170 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,672B, BPFP=1.2661 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,092B, BPFP=2.2527 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,372B, BPFP=1.1991 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,700B, BPFP=2.3884 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,976B, BPFP=2.0036 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,432B, BPFP=2.3286 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,588B, BPFP=0.4014 +⌛️ [2/4] FRONTEND: Frontend time: 2.081s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07685234 39.58151158 + layer.0.v_cache 0.00001390 0.00669000 + layer.1.k_cache 0.01187498 8.86328910 + layer.1.v_cache 0.00000934 0.00246448 + layer.2.k_cache 0.00239603 0.87958587 + layer.2.v_cache 0.00001621 0.00663371 + layer.3.k_cache 0.07241479 3.68126439 + layer.3.v_cache 0.00001716 0.00856252 + layer.4.k_cache 0.00072067 0.18245408 + layer.4.v_cache 0.00004958 0.01627817 + layer.4.output 0.21417375 716.02110969 + ------------------------------------------------------------------------------------- + TOTAL 0.09785772 297.96332363 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 94268 +BPFP 1.2378 bits/point +EBPFP 2.4755 equivalent bits/point +MSE 297.963324 +---------------------- -------------------------------------------------------- +Time: 3.453s Load: 0.005s, Pack+Encode: 2.081s, Decode+Unpack: 1.367s +---------------------- -------------------------------------------------------- +💾 Converting with 297.9633 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,048B, BPFP=0.6266 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,248B, BPFP=2.7237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,044B, BPFP=1.2426 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,256B, BPFP=2.5197 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,300B, BPFP=1.2952 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,620B, BPFP=2.3890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,128B, BPFP=1.2599 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,064B, BPFP=2.4803 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,152B, BPFP=2.0872 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,776B, BPFP=2.4211 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,872B, BPFP=0.4368 +⌛️ [2/4] FRONTEND: Frontend time: 1.964s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09264422 41.61605353 + layer.0.v_cache 0.00001680 0.00677846 + layer.1.k_cache 0.05606840 8.66128460 + layer.1.v_cache 0.00000542 0.00258749 + layer.2.k_cache 0.00241068 0.92630808 + layer.2.v_cache 0.00001702 0.00709484 + layer.3.k_cache 0.15270803 4.08185698 + layer.3.v_cache 0.00001849 0.00908306 + layer.4.k_cache 0.00070970 0.19258189 + layer.4.v_cache 0.00005539 0.01711821 + layer.4.output 0.18838861 690.88639568 + ------------------------------------------------------------------------------------- + TOTAL 0.09549261 287.74855982 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 107508 +BPFP 1.3002 bits/point +EBPFP 2.6003 equivalent bits/point +MSE 287.748560 +---------------------- -------------------------------------------------------- +Time: 3.111s Load: 0.005s, Pack+Encode: 1.964s, Decode+Unpack: 1.142s +---------------------- -------------------------------------------------------- +💾 Converting with 287.7486 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.6250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,476B, BPFP=2.5317 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,932B, BPFP=1.2037 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,268B, BPFP=2.4894 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,276B, BPFP=1.2735 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,560B, BPFP=2.3458 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,096B, BPFP=1.2370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,288B, BPFP=2.4935 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,056B, BPFP=2.0406 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,168B, BPFP=2.4692 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,664B, BPFP=0.3671 +⌛️ [2/4] FRONTEND: Frontend time: 2.608s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.264s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10144231 40.77822836 + layer.0.v_cache 0.00001330 0.00608444 + layer.1.k_cache 0.09893422 8.63896992 + layer.1.v_cache 0.00000513 0.00215705 + layer.2.k_cache 0.00368564 0.87064659 + layer.2.v_cache 0.00001777 0.00689089 + layer.3.k_cache 0.06341467 4.39886316 + layer.3.v_cache 0.00001739 0.00740555 + layer.4.k_cache 0.00073693 0.18873469 + layer.4.v_cache 0.00004994 0.01622257 + layer.4.output 0.18261438 669.23921614 + ------------------------------------------------------------------------------------- + TOTAL 0.09097753 278.79933625 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 104864 +BPFP 1.2517 bits/point +EBPFP 2.5034 equivalent bits/point +MSE 278.799336 +---------------------- -------------------------------------------------------- +Time: 3.876s Load: 0.004s, Pack+Encode: 2.608s, Decode+Unpack: 1.264s +---------------------- -------------------------------------------------------- +💾 Converting with 278.7993 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,300B, BPFP=0.5927 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,396B, BPFP=2.7651 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,508B, BPFP=1.1688 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,840B, BPFP=2.6652 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,536B, BPFP=1.1739 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,036B, BPFP=2.3412 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,496B, BPFP=1.1667 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,292B, BPFP=2.7464 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,844B, BPFP=1.9476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,612B, BPFP=2.4447 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,732B, BPFP=0.3523 +⌛️ [2/4] FRONTEND: Frontend time: 2.469s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.595s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08947531 42.84539164 + layer.0.v_cache 0.00001432 0.00608343 + layer.1.k_cache 0.05320675 8.22588340 + layer.1.v_cache 0.00000497 0.00212947 + layer.2.k_cache 0.00562141 0.98699767 + layer.2.v_cache 0.00001574 0.00599813 + layer.3.k_cache 0.02742601 3.88232001 + layer.3.v_cache 0.00001759 0.00712197 + layer.4.k_cache 0.00114861 0.17077893 + layer.4.v_cache 0.00004734 0.01439073 + layer.4.output 0.15628038 599.99610016 + ------------------------------------------------------------------------------------- + TOTAL 0.07476122 250.35998803 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 119592 +BPFP 1.2634 bits/point +EBPFP 2.5269 equivalent bits/point +MSE 250.359988 +---------------------- -------------------------------------------------------- +Time: 4.069s Load: 0.005s, Pack+Encode: 2.469s, Decode+Unpack: 1.595s +---------------------- -------------------------------------------------------- +💾 Converting with 250.3600 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,396B, BPFP=0.5768 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,432B, BPFP=2.7908 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,836B, BPFP=1.1610 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,376B, BPFP=2.6114 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,880B, BPFP=1.1685 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,704B, BPFP=2.4973 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,876B, BPFP=1.1678 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,368B, BPFP=2.6101 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,044B, BPFP=2.0455 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,104B, BPFP=2.7351 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,104B, BPFP=0.3179 +⌛️ [2/4] FRONTEND: Frontend time: 2.039s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.403s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12311893 40.37231976 + layer.0.v_cache 0.00001357 0.00575279 + layer.1.k_cache 0.03743907 8.71335038 + layer.1.v_cache 0.00000515 0.00213777 + layer.2.k_cache 0.00511221 0.78352754 + layer.2.v_cache 0.00001545 0.00610391 + layer.3.k_cache 0.04225797 3.36171424 + layer.3.v_cache 0.00001799 0.00737351 + layer.4.k_cache 0.00119426 0.18852221 + layer.4.v_cache 0.00004621 0.01470435 + layer.4.output 0.14783020 580.97777562 + ------------------------------------------------------------------------------------- + TOTAL 0.07317837 242.37058446 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 127120 +BPFP 1.2700 bits/point +EBPFP 2.5400 equivalent bits/point +MSE 242.370584 +---------------------- -------------------------------------------------------- +Time: 3.447s Load: 0.006s, Pack+Encode: 2.039s, Decode+Unpack: 1.403s +---------------------- -------------------------------------------------------- +💾 Converting with 242.3706 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,740B, BPFP=0.5729 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,288B, BPFP=2.8015 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,628B, BPFP=1.1685 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,124B, BPFP=2.6232 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,136B, BPFP=1.2463 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,916B, BPFP=2.5913 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,064B, BPFP=1.2353 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,536B, BPFP=2.6863 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,940B, BPFP=2.2886 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,116B, BPFP=2.6219 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,572B, BPFP=0.3845 +⌛️ [2/4] FRONTEND: Frontend time: 1.870s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09293266 42.84929343 + layer.0.v_cache 0.00001503 0.00617022 + layer.1.k_cache 0.07854857 8.76111019 + layer.1.v_cache 0.00000559 0.00220301 + layer.2.k_cache 0.00650115 0.79641559 + layer.2.v_cache 0.00001763 0.00625690 + layer.3.k_cache 0.03715762 4.94171651 + layer.3.v_cache 0.00001788 0.00763006 + layer.4.k_cache 0.00073365 0.17714381 + layer.4.v_cache 0.00004630 0.01492135 + layer.4.output 11.17680568 482.15896359 + ------------------------------------------------------------------------------------- + TOTAL 4.61491858 201.92209448 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 147060 +BPFP 1.3252 bits/point +EBPFP 2.6503 equivalent bits/point +MSE 201.922094 +---------------------- -------------------------------------------------------- +Time: 3.088s Load: 0.005s, Pack+Encode: 1.870s, Decode+Unpack: 1.213s +---------------------- -------------------------------------------------------- +💾 Converting with 201.9221 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 96, 128) +Output shape: (1, 96, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.output: torch.Size([1, 96, 3584]) -> torch.Size([1, 1, 96, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,328B, BPFP=0.5417 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,888B, BPFP=2.9115 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,196B, BPFP=1.1712 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,956B, BPFP=2.7598 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,824B, BPFP=1.2734 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,724B, BPFP=2.7220 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,284B, BPFP=1.3483 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,160B, BPFP=2.7930 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,384B, BPFP=2.5039 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,188B, BPFP=2.7975 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,968B, BPFP=0.3713 +⌛️ [2/4] FRONTEND: Frontend time: 2.013s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.261s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10158994 41.19872538 + layer.0.v_cache 0.00001467 0.00637651 + layer.1.k_cache 0.06416908 8.22466405 + layer.1.v_cache 0.00000542 0.00233904 + layer.2.k_cache 0.00400108 0.84611686 + layer.2.v_cache 0.00001872 0.00684845 + layer.3.k_cache 0.01491791 5.72645760 + layer.3.v_cache 0.00001877 0.00813564 + layer.4.k_cache 0.00070988 0.19388835 + layer.4.v_cache 0.00004832 0.01637758 + layer.4.output 0.14169503 535.75618490 + ------------------------------------------------------------------------------------- + TOTAL 0.06925641 223.91313081 + (elements=835,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 835584 +Total Bytes 143900 +BPFP 1.3777 bits/point +EBPFP 2.7554 equivalent bits/point +MSE 223.913131 +---------------------- -------------------------------------------------------- +Time: 3.279s Load: 0.005s, Pack+Encode: 2.013s, Decode+Unpack: 1.261s +---------------------- -------------------------------------------------------- +💾 Converting with 223.9131 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,932B, BPFP=0.6363 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,872B, BPFP=2.5764 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,720B, BPFP=1.2413 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,492B, BPFP=2.4939 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,904B, BPFP=1.2812 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,968B, BPFP=2.3802 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,720B, BPFP=1.2413 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,252B, BPFP=2.4418 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,424B, BPFP=2.0451 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,248B, BPFP=2.4410 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,256B, BPFP=0.3800 +⌛️ [2/4] FRONTEND: Frontend time: 1.899s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.344s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10321087 42.90809801 + layer.0.v_cache 0.00001557 0.00681095 + layer.1.k_cache 0.06175087 8.68544515 + layer.1.v_cache 0.00000571 0.00238047 + layer.2.k_cache 0.01329466 0.80469068 + layer.2.v_cache 0.00001784 0.00738829 + layer.3.k_cache 0.05039416 4.07582940 + layer.3.v_cache 0.00001745 0.00840712 + layer.4.k_cache 0.00079786 0.19551537 + layer.4.v_cache 0.00005071 0.01735947 + layer.4.output 0.18870974 720.03546627 + ------------------------------------------------------------------------------------- + TOTAL 0.09120729 299.82118758 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 98788 +BPFP 1.2611 bits/point +EBPFP 2.5222 equivalent bits/point +MSE 299.821188 +---------------------- -------------------------------------------------------- +Time: 3.246s Load: 0.003s, Pack+Encode: 1.899s, Decode+Unpack: 1.344s +---------------------- -------------------------------------------------------- +💾 Converting with 299.8212 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,908B, BPFP=0.6311 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,392B, BPFP=2.4722 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,696B, BPFP=1.2361 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,536B, BPFP=2.2865 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,644B, BPFP=1.2248 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,364B, BPFP=2.2491 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,624B, BPFP=1.2205 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,856B, BPFP=2.3559 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,964B, BPFP=1.9453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,656B, BPFP=2.3125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,792B, BPFP=0.4276 +⌛️ [2/4] FRONTEND: Frontend time: 2.092s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.270s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09666174 41.60135905 + layer.0.v_cache 0.00001715 0.00638011 + layer.1.k_cache 0.08476152 8.99242147 + layer.1.v_cache 0.00000506 0.00223477 + layer.2.k_cache 0.00238475 0.96580866 + layer.2.v_cache 0.00001567 0.00634229 + layer.3.k_cache 0.07332428 3.86078305 + layer.3.v_cache 0.00001796 0.00815311 + layer.4.k_cache 0.00072290 0.16742471 + layer.4.v_cache 0.00005092 0.01540511 + layer.4.output 0.18875189 737.02244544 + ------------------------------------------------------------------------------------- + TOTAL 0.09289560 306.75196649 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 96432 +BPFP 1.2310 bits/point +EBPFP 2.4620 equivalent bits/point +MSE 306.751966 +---------------------- -------------------------------------------------------- +Time: 3.365s Load: 0.004s, Pack+Encode: 2.092s, Decode+Unpack: 1.270s +---------------------- -------------------------------------------------------- +💾 Converting with 306.7520 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,716B, BPFP=0.6241 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,468B, BPFP=2.6351 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,260B, BPFP=1.2086 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,824B, BPFP=2.4871 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,504B, BPFP=1.2647 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,264B, BPFP=2.3585 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,304B, BPFP=1.2188 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,756B, BPFP=2.4715 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,668B, BPFP=1.9917 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,744B, BPFP=2.4688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,856B, BPFP=0.4548 +⌛️ [2/4] FRONTEND: Frontend time: 1.984s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10601801 39.68014706 + layer.0.v_cache 0.00001817 0.00675609 + layer.1.k_cache 0.01696859 8.22422611 + layer.1.v_cache 0.00000501 0.00244435 + layer.2.k_cache 0.00274148 0.81991386 + layer.2.v_cache 0.00001757 0.00730235 + layer.3.k_cache 0.01706722 4.61037714 + layer.3.v_cache 0.00001843 0.00858698 + layer.4.k_cache 0.00080906 0.18274133 + layer.4.v_cache 0.00005076 0.01607033 + layer.4.output 0.19983917 788.33849790 + ------------------------------------------------------------------------------------- + TOTAL 0.09074050 327.76047358 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 95364 +BPFP 1.2890 bits/point +EBPFP 2.5780 equivalent bits/point +MSE 327.760474 +---------------------- -------------------------------------------------------- +Time: 3.372s Load: 0.003s, Pack+Encode: 1.984s, Decode+Unpack: 1.384s +---------------------- -------------------------------------------------------- +💾 Converting with 327.7605 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 65, 128) +Output shape: (1, 65, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.output: torch.Size([1, 65, 3584]) -> torch.Size([1, 1, 65, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,480B, BPFP=0.5962 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,976B, BPFP=2.6385 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,672B, BPFP=1.1231 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,432B, BPFP=2.5077 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,036B, BPFP=1.2106 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,720B, BPFP=2.3365 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,908B, BPFP=1.1798 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,824B, BPFP=2.6019 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,368B, BPFP=2.0115 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,400B, BPFP=2.5000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,048B, BPFP=0.4137 +⌛️ [2/4] FRONTEND: Frontend time: 1.917s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.211s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09579474 42.27696064 + layer.0.v_cache 0.00001312 0.00654163 + layer.1.k_cache 0.01648773 8.76585599 + layer.1.v_cache 0.00000712 0.00241080 + layer.2.k_cache 0.00431219 0.82714938 + layer.2.v_cache 0.00001803 0.00741711 + layer.3.k_cache 0.03460605 4.59968637 + layer.3.v_cache 0.00001721 0.00872854 + layer.4.k_cache 0.00068162 0.19001409 + layer.4.v_cache 0.00004727 0.01780116 + layer.4.output 0.20898957 815.56270604 + ------------------------------------------------------------------------------------- + TOTAL 0.09499483 339.15538283 + (elements=565,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 565760 +Total Bytes 89864 +BPFP 1.2707 bits/point +EBPFP 2.5414 equivalent bits/point +MSE 339.155383 +---------------------- -------------------------------------------------------- +Time: 3.132s Load: 0.005s, Pack+Encode: 1.917s, Decode+Unpack: 1.211s +---------------------- -------------------------------------------------------- +💾 Converting with 339.1554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 41, 128) +Output shape: (1, 41, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.output: torch.Size([1, 41, 3584]) -> torch.Size([1, 1, 41, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,000B, BPFP=0.7622 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,828B, BPFP=3.3643 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,328B, BPFP=1.2683 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,596B, BPFP=3.2759 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,652B, BPFP=1.3918 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,628B, BPFP=3.2881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,456B, BPFP=1.3171 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,660B, BPFP=3.3003 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,484B, BPFP=2.4710 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,472B, BPFP=3.2287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,396B, BPFP=0.5660 +⌛️ [2/4] FRONTEND: Frontend time: 1.728s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.162s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10098721 46.19185880 + layer.0.v_cache 0.00001362 0.00856070 + layer.1.k_cache 0.01708067 9.10827339 + layer.1.v_cache 0.00000533 0.00315758 + layer.2.k_cache 0.00250472 1.10269351 + layer.2.v_cache 0.00001755 0.01006607 + layer.3.k_cache 0.08199174 4.48769807 + layer.3.v_cache 0.00001815 0.01093646 + layer.4.k_cache 0.00059437 0.22481472 + layer.4.v_cache 0.00005232 0.02317250 + layer.4.output 0.35770382 1236.24912892 + ------------------------------------------------------------------------------------- + TOTAL 0.15924661 512.64206672 + (elements=356,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 356864 +Total Bytes 72500 +BPFP 1.6253 bits/point +EBPFP 3.2505 equivalent bits/point +MSE 512.642067 +---------------------- -------------------------------------------------------- +Time: 2.893s Load: 0.002s, Pack+Encode: 1.728s, Decode+Unpack: 1.162s +---------------------- -------------------------------------------------------- +💾 Converting with 512.6421 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,276B, BPFP=0.6839 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,092B, BPFP=2.7320 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,852B, BPFP=1.4579 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,784B, BPFP=2.6394 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,320B, BPFP=1.5986 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,500B, BPFP=2.5541 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,472B, BPFP=1.6442 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,488B, BPFP=2.5505 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,492B, BPFP=2.2512 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,388B, BPFP=2.5204 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,900B, BPFP=0.5108 +⌛️ [2/4] FRONTEND: Frontend time: 1.883s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.150s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09210638 40.84254338 + layer.0.v_cache 0.00001369 0.00708244 + layer.1.k_cache 0.01682195 8.42163086 + layer.1.v_cache 0.00000520 0.00277035 + layer.2.k_cache 0.00487203 1.21064597 + layer.2.v_cache 0.00001755 0.00849550 + layer.3.k_cache 0.04965206 5.02415525 + layer.3.v_cache 0.00001944 0.00913107 + layer.4.k_cache 0.00060442 0.20465348 + layer.4.v_cache 0.00005207 0.01943337 + layer.4.output 0.26182200 1010.09443681 + ------------------------------------------------------------------------------------- + TOTAL 0.11746581 419.20068232 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 80564 +BPFP 1.4240 bits/point +EBPFP 2.8480 equivalent bits/point +MSE 419.200682 +---------------------- -------------------------------------------------------- +Time: 3.036s Load: 0.003s, Pack+Encode: 1.883s, Decode+Unpack: 1.150s +---------------------- -------------------------------------------------------- +💾 Converting with 419.2007 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 47, 128) +Output shape: (1, 47, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.output: torch.Size([1, 47, 3584]) -> torch.Size([1, 1, 47, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,160B, BPFP=0.7181 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,836B, BPFP=2.9375 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,020B, BPFP=1.3364 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,600B, BPFP=2.8590 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,248B, BPFP=1.4122 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,312B, BPFP=2.7633 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,740B, BPFP=1.2434 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,640B, BPFP=2.8723 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,464B, BPFP=2.4814 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,436B, BPFP=2.8045 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 10,748B, BPFP=0.5104 +⌛️ [2/4] FRONTEND: Frontend time: 1.836s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08461995 39.05618871 + layer.0.v_cache 0.00001366 0.00785868 + layer.1.k_cache 0.01838745 8.75321668 + layer.1.v_cache 0.00000609 0.00320357 + layer.2.k_cache 0.00780689 0.87155297 + layer.2.v_cache 0.00001932 0.00940291 + layer.3.k_cache 0.04292001 3.98929742 + layer.3.v_cache 0.00001942 0.01034941 + layer.4.k_cache 0.00059704 0.21068230 + layer.4.v_cache 0.00004850 0.02171282 + layer.4.output 0.29441516 1113.13668313 + ------------------------------------------------------------------------------------- + TOTAL 0.13031438 461.46413220 + (elements=409,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 409088 +Total Bytes 75204 +BPFP 1.4707 bits/point +EBPFP 2.9413 equivalent bits/point +MSE 461.464132 +---------------------- -------------------------------------------------------- +Time: 3.041s Load: 0.003s, Pack+Encode: 1.836s, Decode+Unpack: 1.202s +---------------------- -------------------------------------------------------- +💾 Converting with 461.4641 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,384B, BPFP=0.6535 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,192B, BPFP=2.5197 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,564B, BPFP=1.5252 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,900B, BPFP=2.4397 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,240B, BPFP=1.7105 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,764B, BPFP=2.4024 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,636B, BPFP=1.5450 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,928B, BPFP=2.4474 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,164B, BPFP=1.9638 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,732B, BPFP=2.3936 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,876B, BPFP=0.5042 +⌛️ [2/4] FRONTEND: Frontend time: 1.794s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.183s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10653571 42.26373184 + layer.0.v_cache 0.00001396 0.00744883 + layer.1.k_cache 0.01513677 9.36193848 + layer.1.v_cache 0.00000574 0.00284412 + layer.2.k_cache 0.00241650 0.98216174 + layer.2.v_cache 0.00001997 0.00846891 + layer.3.k_cache 0.04250543 4.75601785 + layer.3.v_cache 0.00001889 0.00921812 + layer.4.k_cache 0.00060099 0.19775411 + layer.4.v_cache 0.00005317 0.02000520 + layer.4.output 0.25650722 891.25164474 + ------------------------------------------------------------------------------------- + TOTAL 0.11546222 370.37477073 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 84380 +BPFP 1.3606 bits/point +EBPFP 2.7212 equivalent bits/point +MSE 370.374771 +---------------------- -------------------------------------------------------- +Time: 2.980s Load: 0.003s, Pack+Encode: 1.794s, Decode+Unpack: 1.183s +---------------------- -------------------------------------------------------- +💾 Converting with 370.3748 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,172B, BPFP=0.7070 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,792B, BPFP=2.8620 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,472B, BPFP=1.4557 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,452B, BPFP=2.7513 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,812B, BPFP=1.5664 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,320B, BPFP=2.7083 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,892B, BPFP=1.2669 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,304B, BPFP=2.7031 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,252B, BPFP=2.3607 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,444B, BPFP=2.7487 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,132B, BPFP=0.5177 +⌛️ [2/4] FRONTEND: Frontend time: 1.878s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.165s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10088900 42.01306661 + layer.0.v_cache 0.00001321 0.00697635 + layer.1.k_cache 0.01601078 8.93947538 + layer.1.v_cache 0.00000529 0.00282866 + layer.2.k_cache 0.00240732 1.02965689 + layer.2.v_cache 0.00001858 0.00925544 + layer.3.k_cache 0.04236705 4.55202103 + layer.3.v_cache 0.00001968 0.00966841 + layer.4.k_cache 0.00061965 0.21726725 + layer.4.v_cache 0.00005235 0.02196266 + layer.4.output 0.30498679 1063.78385417 + ------------------------------------------------------------------------------------- + TOTAL 0.13513591 441.36995046 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 76044 +BPFP 1.4561 bits/point +EBPFP 2.9122 equivalent bits/point +MSE 441.369950 +---------------------- -------------------------------------------------------- +Time: 3.045s Load: 0.002s, Pack+Encode: 1.878s, Decode+Unpack: 1.165s +---------------------- -------------------------------------------------------- +💾 Converting with 441.3700 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 53, 128) +Output shape: (1, 53, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.output: torch.Size([1, 53, 3584]) -> torch.Size([1, 1, 53, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,280B, BPFP=0.6722 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,336B, BPFP=2.7524 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,072B, BPFP=1.4953 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,096B, BPFP=2.6816 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,072B, BPFP=1.7901 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,876B, BPFP=2.6167 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,860B, BPFP=1.7276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,916B, BPFP=2.6285 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,860B, BPFP=2.3172 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,832B, BPFP=2.6038 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,968B, BPFP=0.5040 +⌛️ [2/4] FRONTEND: Frontend time: 1.850s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.153s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10150213 39.50011977 + layer.0.v_cache 0.00001370 0.00795817 + layer.1.k_cache 0.01753690 8.29203652 + layer.1.v_cache 0.00000536 0.00326401 + layer.2.k_cache 0.00234856 1.07019417 + layer.2.v_cache 0.00001946 0.01049084 + layer.3.k_cache 0.03772523 5.12114097 + layer.3.v_cache 0.00001792 0.01044369 + layer.4.k_cache 0.00061365 0.22238379 + layer.4.v_cache 0.00006478 0.02296808 + layer.4.output 0.25806696 900.33928571 + ------------------------------------------------------------------------------------- + TOTAL 0.11566567 373.91976471 + (elements=461,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 461312 +Total Bytes 84168 +BPFP 1.4596 bits/point +EBPFP 2.9193 equivalent bits/point +MSE 373.919765 +---------------------- -------------------------------------------------------- +Time: 3.006s Load: 0.003s, Pack+Encode: 1.850s, Decode+Unpack: 1.153s +---------------------- -------------------------------------------------------- +💾 Converting with 373.9198 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,432B, BPFP=0.6333 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,312B, BPFP=2.4250 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,524B, BPFP=1.6990 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,084B, BPFP=2.3656 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,396B, BPFP=1.6656 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,984B, BPFP=2.3396 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,492B, BPFP=1.6906 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,288B, BPFP=2.4188 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,824B, BPFP=2.0375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,036B, BPFP=2.3531 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,512B, BPFP=0.5027 +⌛️ [2/4] FRONTEND: Frontend time: 1.905s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.148s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11850098 40.78011068 + layer.0.v_cache 0.00001385 0.00737243 + layer.1.k_cache 0.01468890 9.84709473 + layer.1.v_cache 0.00000523 0.00304294 + layer.2.k_cache 0.01477869 1.10943820 + layer.2.v_cache 0.00001915 0.00900855 + layer.3.k_cache 0.14661819 5.82587382 + layer.3.v_cache 0.00001925 0.00971552 + layer.4.k_cache 0.00060183 0.22177517 + layer.4.v_cache 0.00005752 0.02160022 + layer.4.output 0.22966646 859.27946429 + ------------------------------------------------------------------------------------- + TOTAL 0.11193934 357.22301660 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 88884 +BPFP 1.3616 bits/point +EBPFP 2.7232 equivalent bits/point +MSE 357.223017 +---------------------- -------------------------------------------------------- +Time: 3.058s Load: 0.005s, Pack+Encode: 1.905s, Decode+Unpack: 1.148s +---------------------- -------------------------------------------------------- +💾 Converting with 357.2230 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,384B, BPFP=0.5747 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,552B, BPFP=2.9810 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,828B, BPFP=1.1596 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,124B, BPFP=2.7385 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,396B, BPFP=1.2561 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,152B, BPFP=2.7432 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,968B, BPFP=1.1834 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,460B, BPFP=2.7955 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,100B, BPFP=2.3947 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,420B, BPFP=2.7887 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,480B, BPFP=0.3756 +⌛️ [2/4] FRONTEND: Frontend time: 1.926s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.329s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10836371 43.67252781 + layer.0.v_cache 0.00001445 0.00648468 + layer.1.k_cache 0.06846340 8.19802724 + layer.1.v_cache 0.00000523 0.00227584 + layer.2.k_cache 0.00506382 0.90481227 + layer.2.v_cache 0.00001740 0.00664914 + layer.3.k_cache 0.03973729 4.24772113 + layer.3.v_cache 0.00001777 0.00762827 + layer.4.k_cache 0.00066797 0.18204880 + layer.4.v_cache 0.00005099 0.01557569 + layer.4.output 0.14786712 580.11291731 + ------------------------------------------------------------------------------------- + TOTAL 0.07396893 242.23730424 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 136864 +BPFP 1.3673 bits/point +EBPFP 2.7347 equivalent bits/point +MSE 242.237304 +---------------------- -------------------------------------------------------- +Time: 3.259s Load: 0.004s, Pack+Encode: 1.926s, Decode+Unpack: 1.329s +---------------------- -------------------------------------------------------- +💾 Converting with 242.2373 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,584B, BPFP=0.6117 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,668B, BPFP=2.7623 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,004B, BPFP=1.1847 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,932B, BPFP=2.5881 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,280B, BPFP=1.2500 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,900B, BPFP=2.5805 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,068B, BPFP=1.1998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,088B, BPFP=2.6250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,640B, BPFP=2.0455 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,832B, BPFP=2.5644 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,764B, BPFP=0.4317 +⌛️ [2/4] FRONTEND: Frontend time: 1.909s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.340s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08728886 41.05472449 + layer.0.v_cache 0.00001734 0.00719923 + layer.1.k_cache 0.01499850 7.95406550 + layer.1.v_cache 0.00000528 0.00247457 + layer.2.k_cache 0.00240521 0.79156165 + layer.2.v_cache 0.00001739 0.00769532 + layer.3.k_cache 0.03126823 4.20206197 + layer.3.v_cache 0.00001843 0.00896886 + layer.4.k_cache 0.00061395 0.19374449 + layer.4.v_cache 0.00004830 0.01772795 + layer.4.output 0.21953388 792.48349567 + ------------------------------------------------------------------------------------- + TOTAL 0.09843639 329.50733493 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 94760 +BPFP 1.3196 bits/point +EBPFP 2.6393 equivalent bits/point +MSE 329.507335 +---------------------- -------------------------------------------------------- +Time: 3.253s Load: 0.003s, Pack+Encode: 1.909s, Decode+Unpack: 1.340s +---------------------- -------------------------------------------------------- +💾 Converting with 329.5073 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,568B, BPFP=0.6080 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,796B, BPFP=2.7926 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,024B, BPFP=1.1894 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,188B, BPFP=2.6487 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,368B, BPFP=1.2708 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,936B, BPFP=2.5890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,204B, BPFP=1.2320 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,228B, BPFP=2.6581 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,652B, BPFP=2.0483 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,824B, BPFP=2.5625 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,820B, BPFP=0.4336 +⌛️ [2/4] FRONTEND: Frontend time: 2.277s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.279s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08223097 40.46950832 + layer.0.v_cache 0.00001442 0.00712535 + layer.1.k_cache 0.01480860 7.88457512 + layer.1.v_cache 0.00000502 0.00232898 + layer.2.k_cache 0.00576455 0.85993992 + layer.2.v_cache 0.00002304 0.00716017 + layer.3.k_cache 0.03215047 3.93395210 + layer.3.v_cache 0.00001933 0.00848082 + layer.4.k_cache 0.00061510 0.18511658 + layer.4.v_cache 0.00004916 0.01743080 + layer.4.output 0.22077460 787.85355790 + ------------------------------------------------------------------------------------- + TOTAL 0.09888840 327.55003079 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 95608 +BPFP 1.3314 bits/point +EBPFP 2.6629 equivalent bits/point +MSE 327.550031 +---------------------- -------------------------------------------------------- +Time: 3.560s Load: 0.004s, Pack+Encode: 2.277s, Decode+Unpack: 1.279s +---------------------- -------------------------------------------------------- +💾 Converting with 327.5500 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,916B, BPFP=0.6241 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,116B, BPFP=2.5933 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,768B, BPFP=1.2346 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,652B, BPFP=2.4940 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,100B, BPFP=1.3057 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,320B, BPFP=2.4229 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,704B, BPFP=1.2209 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,996B, BPFP=2.5676 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,296B, BPFP=1.9897 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,204B, BPFP=2.3981 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,428B, BPFP=0.4106 +⌛️ [2/4] FRONTEND: Frontend time: 2.472s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.293s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10114760 42.88620371 + layer.0.v_cache 0.00001420 0.00666403 + layer.1.k_cache 0.03683303 9.02031049 + layer.1.v_cache 0.00000544 0.00248677 + layer.2.k_cache 0.00248627 0.88913737 + layer.2.v_cache 0.00001713 0.00699072 + layer.3.k_cache 0.02909129 4.47125788 + layer.3.v_cache 0.00001766 0.00842191 + layer.4.k_cache 0.00072613 0.19590706 + layer.4.v_cache 0.00005154 0.01737100 + layer.4.output 0.19731793 678.30925881 + ------------------------------------------------------------------------------------- + TOTAL 0.09127152 282.68644486 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 101500 +BPFP 1.2780 bits/point +EBPFP 2.5559 equivalent bits/point +MSE 282.686445 +---------------------- -------------------------------------------------------- +Time: 3.769s Load: 0.005s, Pack+Encode: 2.472s, Decode+Unpack: 1.293s +---------------------- -------------------------------------------------------- +💾 Converting with 282.6864 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,020B, BPFP=0.6209 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,484B, BPFP=2.5666 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,040B, BPFP=1.2418 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,616B, BPFP=2.3882 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,304B, BPFP=1.2961 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,680B, BPFP=2.4013 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,076B, BPFP=1.2492 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,908B, BPFP=2.4482 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,864B, BPFP=2.0280 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,692B, BPFP=2.4038 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,720B, BPFP=0.4323 +⌛️ [2/4] FRONTEND: Frontend time: 1.822s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10230959 41.70454808 + layer.0.v_cache 0.00001409 0.00666833 + layer.1.k_cache 0.07973289 8.82035346 + layer.1.v_cache 0.00000506 0.00232384 + layer.2.k_cache 0.00553133 0.88983837 + layer.2.v_cache 0.00001928 0.00687247 + layer.3.k_cache 0.03946657 4.32882329 + layer.3.v_cache 0.00001980 0.00801949 + layer.4.k_cache 0.00066525 0.18329666 + layer.4.v_cache 0.00005119 0.01608961 + layer.4.output 0.17985767 700.81185385 + ------------------------------------------------------------------------------------- + TOTAL 0.08745993 291.86175356 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 105404 +BPFP 1.2747 bits/point +EBPFP 2.5494 equivalent bits/point +MSE 291.861754 +---------------------- -------------------------------------------------------- +Time: 3.200s Load: 0.004s, Pack+Encode: 1.822s, Decode+Unpack: 1.375s +---------------------- -------------------------------------------------------- +💾 Converting with 291.8618 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,096B, BPFP=0.6282 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,260B, BPFP=2.6907 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,884B, BPFP=1.1940 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,100B, BPFP=2.6583 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,052B, BPFP=1.2281 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,864B, BPFP=2.2045 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,092B, BPFP=1.2362 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,124B, BPFP=2.4602 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,604B, BPFP=1.9489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,940B, BPFP=2.4229 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 13,660B, BPFP=0.3960 +⌛️ [2/4] FRONTEND: Frontend time: 1.819s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.407s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633976 43.87201324 + layer.0.v_cache 0.00001461 0.00671405 + layer.1.k_cache 0.07748085 9.42905099 + layer.1.v_cache 0.00000515 0.00224722 + layer.2.k_cache 0.00505460 0.86549853 + layer.2.v_cache 0.00001651 0.00651095 + layer.3.k_cache 0.02793909 3.99713927 + layer.3.v_cache 0.00001809 0.00785886 + layer.4.k_cache 0.00070819 0.17691859 + layer.4.v_cache 0.00004716 0.01515930 + layer.4.output 0.18765952 677.75852273 + ------------------------------------------------------------------------------------- + TOTAL 0.09007298 282.51110412 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 105676 +BPFP 1.2614 bits/point +EBPFP 2.5228 equivalent bits/point +MSE 282.511104 +---------------------- -------------------------------------------------------- +Time: 3.230s Load: 0.004s, Pack+Encode: 1.819s, Decode+Unpack: 1.407s +---------------------- -------------------------------------------------------- +💾 Converting with 282.5111 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,832B, BPFP=0.6321 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,164B, BPFP=2.7152 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,472B, BPFP=1.2214 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,900B, BPFP=2.4330 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,724B, BPFP=1.2777 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,592B, BPFP=2.3643 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,536B, BPFP=1.2357 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,452B, BPFP=2.5562 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,980B, BPFP=2.0045 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,764B, BPFP=2.4027 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,932B, BPFP=0.3805 +⌛️ [2/4] FRONTEND: Frontend time: 2.527s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.429s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08399124 39.52564174 + layer.0.v_cache 0.00001671 0.00665100 + layer.1.k_cache 0.03849935 8.09471697 + layer.1.v_cache 0.00000503 0.00225465 + layer.2.k_cache 0.00738276 0.96568037 + layer.2.v_cache 0.00001692 0.00666895 + layer.3.k_cache 0.04914829 3.89575936 + layer.3.v_cache 0.00001713 0.00790469 + layer.4.k_cache 0.00063731 0.17863204 + layer.4.v_cache 0.00004711 0.01588341 + layer.4.output 0.21328995 715.77908163 + ------------------------------------------------------------------------------------- + TOTAL 0.09839950 297.83255086 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 96348 +BPFP 1.2651 bits/point +EBPFP 2.5301 equivalent bits/point +MSE 297.832551 +---------------------- -------------------------------------------------------- +Time: 3.961s Load: 0.004s, Pack+Encode: 2.527s, Decode+Unpack: 1.429s +---------------------- -------------------------------------------------------- +💾 Converting with 297.8326 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 56, 128) +Output shape: (1, 56, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.output: torch.Size([1, 56, 3584]) -> torch.Size([1, 1, 56, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,416B, BPFP=0.6741 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,220B, BPFP=2.5725 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,664B, BPFP=1.3013 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,816B, BPFP=2.4598 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,620B, BPFP=1.5681 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,656B, BPFP=2.4152 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,388B, BPFP=1.5033 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,008B, BPFP=2.5134 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,652B, BPFP=2.1350 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,712B, BPFP=2.4308 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,112B, BPFP=0.4828 +⌛️ [2/4] FRONTEND: Frontend time: 1.777s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.180s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13768739 39.51930891 + layer.0.v_cache 0.00001432 0.00768120 + layer.1.k_cache 0.01638686 8.44557408 + layer.1.v_cache 0.00000533 0.00275936 + layer.2.k_cache 0.00251956 0.98944882 + layer.2.v_cache 0.00001745 0.00807955 + layer.3.k_cache 0.08599430 5.68843460 + layer.3.v_cache 0.00002012 0.00959009 + layer.4.k_cache 0.00061995 0.19805816 + layer.4.v_cache 0.00004798 0.01966790 + layer.4.output 0.24256272 874.04073661 + ------------------------------------------------------------------------------------- + TOTAL 0.11419131 363.12786817 + (elements=487,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 487424 +Total Bytes 82264 +BPFP 1.3502 bits/point +EBPFP 2.7004 equivalent bits/point +MSE 363.127868 +---------------------- -------------------------------------------------------- +Time: 2.960s Load: 0.003s, Pack+Encode: 1.777s, Decode+Unpack: 1.180s +---------------------- -------------------------------------------------------- +💾 Converting with 363.1279 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,176B, BPFP=0.6203 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,104B, BPFP=2.5594 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,228B, BPFP=1.2164 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,148B, BPFP=2.5680 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,484B, BPFP=1.2664 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,896B, BPFP=2.3234 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,300B, BPFP=1.2305 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,464B, BPFP=2.4344 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,352B, BPFP=2.0219 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,980B, BPFP=2.3398 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,896B, BPFP=0.4156 +⌛️ [2/4] FRONTEND: Frontend time: 1.771s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09278157 43.53057861 + layer.0.v_cache 0.00001504 0.00659393 + layer.1.k_cache 0.07523044 9.16480255 + layer.1.v_cache 0.00000517 0.00236133 + layer.2.k_cache 0.00381728 0.89414654 + layer.2.v_cache 0.00001789 0.00677962 + layer.3.k_cache 0.03327554 3.68048706 + layer.3.v_cache 0.00001939 0.00842386 + layer.4.k_cache 0.00067710 0.18377473 + layer.4.v_cache 0.00005192 0.01654801 + layer.4.output 0.16994183 661.21757813 + ------------------------------------------------------------------------------------- + TOTAL 0.08208730 275.64809077 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 110028 +BPFP 1.2641 bits/point +EBPFP 2.5282 equivalent bits/point +MSE 275.648091 +---------------------- -------------------------------------------------------- +Time: 3.100s Load: 0.004s, Pack+Encode: 1.771s, Decode+Unpack: 1.325s +---------------------- -------------------------------------------------------- +💾 Converting with 275.6481 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,388B, BPFP=0.6324 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,360B, BPFP=2.4788 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,444B, BPFP=1.7066 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,952B, BPFP=2.3708 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,572B, BPFP=1.7405 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,812B, BPFP=2.3337 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,116B, BPFP=1.6197 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,048B, BPFP=2.3962 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,268B, BPFP=1.9248 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,848B, BPFP=2.3432 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,616B, BPFP=0.4773 +⌛️ [2/4] FRONTEND: Frontend time: 1.912s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.163s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12468526 39.52244852 + layer.0.v_cache 0.00001451 0.00794342 + layer.1.k_cache 0.01313243 8.83949719 + layer.1.v_cache 0.00000580 0.00302121 + layer.2.k_cache 0.00906488 0.94656029 + layer.2.v_cache 0.00002099 0.00854267 + layer.3.k_cache 0.01569995 5.03345761 + layer.3.v_cache 0.00002002 0.00995040 + layer.4.k_cache 0.00064964 0.20782244 + layer.4.v_cache 0.00005478 0.02131502 + layer.4.output 0.23031524 890.94582324 + ------------------------------------------------------------------------------------- + TOTAL 0.10444441 370.07184244 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 86424 +BPFP 1.3463 bits/point +EBPFP 2.6927 equivalent bits/point +MSE 370.071842 +---------------------- -------------------------------------------------------- +Time: 3.078s Load: 0.004s, Pack+Encode: 1.912s, Decode+Unpack: 1.163s +---------------------- -------------------------------------------------------- +💾 Converting with 370.0718 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,432B, BPFP=0.6333 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,512B, BPFP=2.4771 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,536B, BPFP=1.7021 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,180B, BPFP=2.3906 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,872B, BPFP=1.7896 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,016B, BPFP=2.3479 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,376B, BPFP=1.6604 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,268B, BPFP=2.4135 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,892B, BPFP=2.0552 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,012B, BPFP=2.3469 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,064B, BPFP=0.5232 +⌛️ [2/4] FRONTEND: Frontend time: 1.713s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15164433 40.87945150 + layer.0.v_cache 0.00001435 0.00792163 + layer.1.k_cache 0.01704682 9.61057027 + layer.1.v_cache 0.00000573 0.00311772 + layer.2.k_cache 0.00240009 0.99761645 + layer.2.v_cache 0.00001845 0.00887890 + layer.3.k_cache 0.02161660 4.64056040 + layer.3.v_cache 0.00001985 0.01009719 + layer.4.k_cache 0.00061459 0.21897319 + layer.4.v_cache 0.00005275 0.02159684 + layer.4.output 0.22653077 873.78519345 + ------------------------------------------------------------------------------------- + TOTAL 0.10465582 363.11147872 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 90160 +BPFP 1.3811 bits/point +EBPFP 2.7623 equivalent bits/point +MSE 363.111479 +---------------------- -------------------------------------------------------- +Time: 3.083s Load: 0.003s, Pack+Encode: 1.713s, Decode+Unpack: 1.367s +---------------------- -------------------------------------------------------- +💾 Converting with 363.1115 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,936B, BPFP=0.6284 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,460B, BPFP=2.4529 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,656B, BPFP=1.2106 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,944B, BPFP=2.3425 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,844B, BPFP=1.2509 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,412B, BPFP=2.2286 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,680B, BPFP=1.2158 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,392B, BPFP=2.4384 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,188B, BPFP=1.9666 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,864B, BPFP=2.3253 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 12,584B, BPFP=0.3848 +⌛️ [2/4] FRONTEND: Frontend time: 2.518s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08869729 41.52550433 + layer.0.v_cache 0.00001594 0.00625207 + layer.1.k_cache 0.08105310 8.85568112 + layer.1.v_cache 0.00000494 0.00214397 + layer.2.k_cache 0.00256802 0.77747721 + layer.2.v_cache 0.00001661 0.00655214 + layer.3.k_cache 0.01835501 4.59066647 + layer.3.v_cache 0.00001861 0.00738059 + layer.4.k_cache 0.00078324 0.17957314 + layer.4.v_cache 0.00004769 0.01518511 + layer.4.output 0.18611241 704.59717466 + ------------------------------------------------------------------------------------- + TOTAL 0.08790278 293.42039051 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 96960 +BPFP 1.2208 bits/point +EBPFP 2.4416 equivalent bits/point +MSE 293.420391 +---------------------- -------------------------------------------------------- +Time: 3.871s Load: 0.006s, Pack+Encode: 2.518s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 293.4204 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,356B, BPFP=0.5406 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,876B, BPFP=2.8795 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,056B, BPFP=1.1366 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,020B, BPFP=2.7416 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,524B, BPFP=1.2120 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,796B, BPFP=2.7055 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,020B, BPFP=1.2919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,172B, BPFP=2.7661 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,304B, BPFP=2.3041 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,992B, BPFP=2.7371 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,820B, BPFP=0.3871 +⌛️ [2/4] FRONTEND: Frontend time: 1.766s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17228427 41.72441306 + layer.0.v_cache 0.00001543 0.00627572 + layer.1.k_cache 0.04621154 9.21810001 + layer.1.v_cache 0.00000545 0.00214769 + layer.2.k_cache 0.00739691 0.76088502 + layer.2.v_cache 0.00001797 0.00620029 + layer.3.k_cache 0.03096496 5.37493519 + layer.3.v_cache 0.00001890 0.00736801 + layer.4.k_cache 0.00071091 0.19308900 + layer.4.v_cache 0.00004900 0.01592724 + layer.4.output 0.02450962 551.72813881 + ------------------------------------------------------------------------------------- + TOTAL 0.02524957 230.55331252 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 142936 +BPFP 1.3544 bits/point +EBPFP 2.7088 equivalent bits/point +MSE 230.553313 +---------------------- -------------------------------------------------------- +Time: 3.086s Load: 0.005s, Pack+Encode: 1.766s, Decode+Unpack: 1.315s +---------------------- -------------------------------------------------------- +💾 Converting with 230.5533 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,356B, BPFP=0.5638 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,632B, BPFP=2.9624 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,088B, BPFP=1.1909 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,688B, BPFP=2.6358 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,476B, BPFP=1.2560 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,132B, BPFP=2.5423 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,404B, BPFP=1.2440 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,212B, BPFP=2.7238 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,248B, BPFP=2.3938 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,892B, BPFP=2.6700 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,308B, BPFP=0.3674 +⌛️ [2/4] FRONTEND: Frontend time: 1.909s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.288s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16174046 41.63537204 + layer.0.v_cache 0.00001612 0.00665006 + layer.1.k_cache 0.03352992 8.31422670 + layer.1.v_cache 0.00000563 0.00238638 + layer.2.k_cache 0.00496730 1.02330033 + layer.2.v_cache 0.00001740 0.00694672 + layer.3.k_cache 0.01521110 4.33429940 + layer.3.v_cache 0.00001939 0.00794031 + layer.4.k_cache 0.00066431 0.19529012 + layer.4.v_cache 0.00005067 0.01556424 + layer.4.output 0.14624557 548.53273810 + ------------------------------------------------------------------------------------- + TOTAL 0.07293772 229.13359665 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 135436 +BPFP 1.3385 bits/point +EBPFP 2.6770 equivalent bits/point +MSE 229.133597 +---------------------- -------------------------------------------------------- +Time: 3.202s Load: 0.005s, Pack+Encode: 1.909s, Decode+Unpack: 1.288s +---------------------- -------------------------------------------------------- +💾 Converting with 229.1336 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,272B, BPFP=0.5945 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,912B, BPFP=2.7093 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,484B, BPFP=1.1781 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,156B, BPFP=2.5719 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,844B, BPFP=1.2435 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,696B, BPFP=2.3067 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,648B, BPFP=1.2078 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,360B, BPFP=2.7907 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,660B, BPFP=1.9368 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,056B, BPFP=2.5538 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,344B, BPFP=0.3983 +⌛️ [2/4] FRONTEND: Frontend time: 1.973s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.290s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07121617 41.51965616 + layer.0.v_cache 0.00001681 0.00637869 + layer.1.k_cache 0.06877367 8.65702642 + layer.1.v_cache 0.00000530 0.00227298 + layer.2.k_cache 0.00486504 0.87336589 + layer.2.v_cache 0.00001749 0.00647549 + layer.3.k_cache 0.08833314 4.25655347 + layer.3.v_cache 0.00001723 0.00739497 + layer.4.k_cache 0.00077278 0.17822388 + layer.4.v_cache 0.00004595 0.01562235 + layer.4.output 0.15812202 599.48577658 + ------------------------------------------------------------------------------------- + TOTAL 0.07887751 250.11314155 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 120432 +BPFP 1.2871 bits/point +EBPFP 2.5742 equivalent bits/point +MSE 250.113142 +---------------------- -------------------------------------------------------- +Time: 3.267s Load: 0.004s, Pack+Encode: 1.973s, Decode+Unpack: 1.290s +---------------------- -------------------------------------------------------- +💾 Converting with 250.1131 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,620B, BPFP=0.5600 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,116B, BPFP=2.8026 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,720B, BPFP=1.1943 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,200B, BPFP=2.6609 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,084B, BPFP=1.2506 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,276B, BPFP=2.5179 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,364B, BPFP=1.2939 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,008B, BPFP=2.6312 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,588B, BPFP=2.2568 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,144B, BPFP=2.4975 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,816B, BPFP=0.3274 +⌛️ [2/4] FRONTEND: Frontend time: 1.871s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.244s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13964315 43.90318978 + layer.0.v_cache 0.00001432 0.00655018 + layer.1.k_cache 0.04810369 8.54696323 + layer.1.v_cache 0.00000572 0.00241473 + layer.2.k_cache 0.00492308 0.91662794 + layer.2.v_cache 0.00001777 0.00662751 + layer.3.k_cache 0.04255116 5.21136233 + layer.3.v_cache 0.00001880 0.00797997 + layer.4.k_cache 0.00073971 0.19985539 + layer.4.v_cache 0.00004573 0.01692509 + layer.4.output 11.28739116 505.58632426 + ------------------------------------------------------------------------------------- + TOTAL 4.66163537 211.64251564 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 141936 +BPFP 1.2916 bits/point +EBPFP 2.5833 equivalent bits/point +MSE 211.642516 +---------------------- -------------------------------------------------------- +Time: 3.119s Load: 0.005s, Pack+Encode: 1.871s, Decode+Unpack: 1.244s +---------------------- -------------------------------------------------------- +💾 Converting with 211.6425 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,300B, BPFP=0.5859 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,756B, BPFP=2.7976 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,688B, BPFP=1.1875 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,352B, BPFP=2.7259 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,956B, BPFP=1.2351 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,736B, BPFP=2.4389 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,840B, BPFP=1.2145 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,152B, BPFP=2.6903 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,508B, BPFP=2.0433 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,132B, BPFP=2.5092 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,512B, BPFP=0.3681 +⌛️ [2/4] FRONTEND: Frontend time: 1.900s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.299s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14145413 42.85371815 + layer.0.v_cache 0.00001516 0.00617111 + layer.1.k_cache 0.10647225 8.37528437 + layer.1.v_cache 0.00000560 0.00222338 + layer.2.k_cache 0.00244946 0.87187836 + layer.2.v_cache 0.00001685 0.00601876 + layer.3.k_cache 0.01581429 4.12121582 + layer.3.v_cache 0.00001841 0.00739931 + layer.4.k_cache 0.00072023 0.17676269 + layer.4.v_cache 0.00004652 0.01442493 + layer.4.output 0.15451367 598.00071023 + ------------------------------------------------------------------------------------- + TOTAL 0.07932992 249.55529815 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 123932 +BPFP 1.2944 bits/point +EBPFP 2.5888 equivalent bits/point +MSE 249.555298 +---------------------- -------------------------------------------------------- +Time: 3.204s Load: 0.005s, Pack+Encode: 1.900s, Decode+Unpack: 1.299s +---------------------- -------------------------------------------------------- +💾 Converting with 249.5553 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,964B, BPFP=0.6258 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,324B, BPFP=2.6022 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,908B, BPFP=1.2475 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,548B, BPFP=2.4383 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,004B, BPFP=1.2677 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,104B, BPFP=2.3446 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,680B, BPFP=1.1993 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,676B, BPFP=2.4654 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,404B, BPFP=1.9856 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,828B, BPFP=2.4975 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 11,788B, BPFP=0.3556 +⌛️ [2/4] FRONTEND: Frontend time: 2.028s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.256s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10237078 40.75990419 + layer.0.v_cache 0.00001391 0.00648805 + layer.1.k_cache 0.05953315 8.18872895 + layer.1.v_cache 0.00000517 0.00234387 + layer.2.k_cache 0.00728044 0.73389559 + layer.2.v_cache 0.00001648 0.00686056 + layer.3.k_cache 0.03101032 4.23899140 + layer.3.v_cache 0.00001666 0.00776997 + layer.4.k_cache 0.00072555 0.17726195 + layer.4.v_cache 0.00004870 0.01748303 + layer.4.output 0.18364459 671.84863658 + ------------------------------------------------------------------------------------- + TOTAL 0.08744314 279.82824610 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 100228 +BPFP 1.2449 bits/point +EBPFP 2.4898 equivalent bits/point +MSE 279.828246 +---------------------- -------------------------------------------------------- +Time: 3.288s Load: 0.003s, Pack+Encode: 2.028s, Decode+Unpack: 1.256s +---------------------- -------------------------------------------------------- +💾 Converting with 279.8282 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 111, 128) +Output shape: (1, 111, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.output: torch.Size([1, 111, 3584]) -> torch.Size([1, 1, 111, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,956B, BPFP=0.5569 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,404B, BPFP=2.5907 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,368B, BPFP=1.1779 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,588B, BPFP=2.4758 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,228B, BPFP=1.4398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,964B, BPFP=2.3880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,648B, BPFP=1.2173 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,524B, BPFP=2.4668 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,020B, BPFP=2.1143 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,300B, BPFP=2.4352 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,868B, BPFP=0.3392 +⌛️ [2/4] FRONTEND: Frontend time: 1.926s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.458s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11107565 42.52761648 + layer.0.v_cache 0.00001597 0.00614439 + layer.1.k_cache 0.04333284 9.77673890 + layer.1.v_cache 0.00000558 0.00219211 + layer.2.k_cache 0.00477623 0.98759536 + layer.2.v_cache 0.00001758 0.00610410 + layer.3.k_cache 0.01020195 4.34234812 + layer.3.v_cache 0.00001870 0.00707111 + layer.4.k_cache 0.00079488 0.18384579 + layer.4.v_cache 0.00005041 0.01517146 + layer.4.output 10.27057757 457.42495174 + ------------------------------------------------------------------------------------- + TOTAL 4.23907840 191.75467588 + (elements=966,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 966144 +Total Bytes 150868 +BPFP 1.2492 bits/point +EBPFP 2.4985 equivalent bits/point +MSE 191.754676 +---------------------- -------------------------------------------------------- +Time: 3.389s Load: 0.005s, Pack+Encode: 1.926s, Decode+Unpack: 1.458s +---------------------- -------------------------------------------------------- +💾 Converting with 191.7547 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,896B, BPFP=0.6373 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,700B, BPFP=2.5748 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,616B, BPFP=1.2359 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,324B, BPFP=2.4921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,816B, BPFP=1.2799 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,772B, BPFP=2.3706 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,580B, BPFP=1.2280 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,192B, BPFP=2.4630 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,360B, BPFP=2.0599 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,916B, BPFP=2.4023 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,220B, BPFP=0.4471 +⌛️ [2/4] FRONTEND: Frontend time: 1.989s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.184s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09375787 40.38665686 + layer.0.v_cache 0.00001622 0.00640248 + layer.1.k_cache 0.05984993 9.70054433 + layer.1.v_cache 0.00000512 0.00226884 + layer.2.k_cache 0.00243208 0.92755310 + layer.2.v_cache 0.00001865 0.00687256 + layer.3.k_cache 0.04944374 3.96672209 + layer.3.v_cache 0.00001908 0.00811204 + layer.4.k_cache 0.00084877 0.19179222 + layer.4.v_cache 0.00005139 0.01696223 + layer.4.output 0.19371969 726.80394869 + ------------------------------------------------------------------------------------- + TOTAL 0.09191063 302.52008986 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 99392 +BPFP 1.2867 bits/point +EBPFP 2.5733 equivalent bits/point +MSE 302.520090 +---------------------- -------------------------------------------------------- +Time: 3.177s Load: 0.003s, Pack+Encode: 1.989s, Decode+Unpack: 1.184s +---------------------- -------------------------------------------------------- +💾 Converting with 302.5201 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 109, 128) +Output shape: (1, 109, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.output: torch.Size([1, 109, 3584]) -> torch.Size([1, 1, 109, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,916B, BPFP=0.5614 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,604B, BPFP=2.6669 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,724B, BPFP=1.2506 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,752B, BPFP=2.5447 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,700B, BPFP=1.5338 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,356B, BPFP=2.4880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,616B, BPFP=1.3784 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,688B, BPFP=2.5356 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,504B, BPFP=2.2225 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,464B, BPFP=2.5034 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,452B, BPFP=0.3779 +⌛️ [2/4] FRONTEND: Frontend time: 1.746s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860953 40.98178577 + layer.0.v_cache 0.00001500 0.00613885 + layer.1.k_cache 0.05917679 7.86917030 + layer.1.v_cache 0.00000528 0.00220820 + layer.2.k_cache 0.01253135 1.06935778 + layer.2.v_cache 0.00001869 0.00635090 + layer.3.k_cache 0.07406004 5.52656709 + layer.3.v_cache 0.00001854 0.00742263 + layer.4.k_cache 0.00071234 0.19529414 + layer.4.v_cache 0.00004916 0.01490557 + layer.4.output 10.45909253 475.48275721 + ------------------------------------------------------------------------------------- + TOTAL 4.32287320 199.06226481 + (elements=948,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 948736 +Total Bytes 155776 +BPFP 1.3135 bits/point +EBPFP 2.6271 equivalent bits/point +MSE 199.062265 +---------------------- -------------------------------------------------------- +Time: 2.976s Load: 0.004s, Pack+Encode: 1.746s, Decode+Unpack: 1.225s +---------------------- -------------------------------------------------------- +💾 Converting with 199.0623 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst + to output-fixed/kimiaudio/lambda0.01/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.3350 bits/point +Avg EBPFP 2.6700 equivalent bits/point +Avg MSE 302.925885 +Avg Time 3.223s +------------------------ ---------------------------- diff --git a/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..fe261b12ccf92be72720fed1bbf30d0551b0cd8c --- /dev/null +++ b/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 599 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench +Output output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,188B, BPFP=0.6150 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,700B, BPFP=2.2569 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,364B, BPFP=1.0347 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,284B, BPFP=2.1767 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,472B, BPFP=1.2485 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,944B, BPFP=2.1111 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,104B, BPFP=1.1775 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,152B, BPFP=2.1512 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,256B, BPFP=1.7855 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,728B, BPFP=2.0694 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,896B, BPFP=0.6310 +⌛️ [2/4] FRONTEND: Frontend time: 0.398s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.251s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12089630 55.38466917 + layer.0.v_cache 0.00001394 0.00848349 + layer.1.k_cache 0.01174029 5.37840214 + layer.1.v_cache 0.00000579 0.00344826 + layer.2.k_cache 0.00835781 0.85938969 + layer.2.v_cache 0.00001852 0.00916578 + layer.3.k_cache 0.02896961 3.86757368 + layer.3.v_cache 0.00001872 0.00973807 + layer.4.k_cache 0.00063445 0.19996707 + layer.4.v_cache 0.00005063 0.02037806 + layer.4.output 0.17246709 664.52414021 + ------------------------------------------------------------------------------------- + TOTAL 0.08105739 277.49471747 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 109088 +BPFP 1.2378 bits/point +EBPFP 2.4757 equivalent bits/point +MSE 277.494717 +---------------------- -------------------------------------------------------- +Time: 0.653s Load: 0.005s, Pack+Encode: 0.398s, Decode+Unpack: 0.251s +---------------------- -------------------------------------------------------- +💾 Converting with 277.4947 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,152B, BPFP=0.6156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,724B, BPFP=2.2898 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,372B, BPFP=1.0492 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,352B, BPFP=2.2172 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,672B, BPFP=1.3031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,076B, BPFP=2.1633 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,176B, BPFP=1.2063 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,304B, BPFP=2.2078 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,328B, BPFP=1.8219 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,964B, BPFP=2.1414 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,096B, BPFP=0.6165 +⌛️ [2/4] FRONTEND: Frontend time: 0.171s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08029544 59.50405273 + layer.0.v_cache 0.00001376 0.00877180 + layer.1.k_cache 0.03632395 5.40148163 + layer.1.v_cache 0.00000563 0.00400649 + layer.2.k_cache 0.00522719 0.89164219 + layer.2.v_cache 0.00001997 0.01117315 + layer.3.k_cache 0.02707962 3.43539772 + layer.3.v_cache 0.00001843 0.01140180 + layer.4.k_cache 0.00063057 0.20555124 + layer.4.v_cache 0.00005085 0.02266395 + layer.4.output 0.18451144 674.22449777 + ------------------------------------------------------------------------------------- + TOTAL 0.08477915 281.70986042 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 109216 +BPFP 1.2548 bits/point +EBPFP 2.5096 equivalent bits/point +MSE 281.709860 +---------------------- -------------------------------------------------------- +Time: 0.389s Load: 0.004s, Pack+Encode: 0.171s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 281.7099 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,260B, BPFP=0.5923 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,408B, BPFP=2.2544 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,608B, BPFP=1.0189 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,892B, BPFP=2.1606 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,808B, BPFP=1.2369 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,660B, BPFP=2.1185 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,484B, BPFP=1.1781 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,832B, BPFP=2.1497 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,812B, BPFP=1.7827 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,348B, BPFP=2.0618 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,056B, BPFP=0.5984 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10874708 57.23976313 + layer.0.v_cache 0.00001608 0.00845219 + layer.1.k_cache 0.03270756 5.43575198 + layer.1.v_cache 0.00000572 0.00340783 + layer.2.k_cache 0.00777350 0.82014873 + layer.2.v_cache 0.00002080 0.00926701 + layer.3.k_cache 0.05920834 3.71012275 + layer.3.v_cache 0.00001846 0.01007315 + layer.4.k_cache 0.00062493 0.19428417 + layer.4.v_cache 0.00005036 0.01971213 + layer.4.output 0.17137358 626.43313953 + ------------------------------------------------------------------------------------- + TOTAL 0.08286988 261.91076234 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 114168 +BPFP 1.2202 bits/point +EBPFP 2.4403 equivalent bits/point +MSE 261.910762 +---------------------- -------------------------------------------------------- +Time: 0.385s Load: 0.003s, Pack+Encode: 0.168s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 261.9108 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,072B, BPFP=0.6154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,312B, BPFP=2.2660 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,204B, BPFP=1.0425 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,964B, BPFP=2.1963 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,316B, BPFP=1.2652 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,636B, BPFP=2.1306 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,912B, BPFP=1.1843 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,820B, BPFP=2.1675 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,952B, BPFP=1.7933 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,500B, BPFP=2.1034 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,644B, BPFP=0.6194 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12073612 60.30096905 + layer.0.v_cache 0.00001348 0.00867126 + layer.1.k_cache 0.03279715 5.42561066 + layer.1.v_cache 0.00000568 0.00417210 + layer.2.k_cache 0.00696802 0.93628702 + layer.2.v_cache 0.00001833 0.01042740 + layer.3.k_cache 0.03227850 4.15429570 + layer.3.v_cache 0.00001834 0.01106773 + layer.4.k_cache 0.00060115 0.21817467 + layer.4.v_cache 0.00005191 0.02330005 + layer.4.output 0.18320683 677.13278388 + ------------------------------------------------------------------------------------- + TOTAL 0.08681979 283.00132134 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 105332 +BPFP 1.2412 bits/point +EBPFP 2.4824 equivalent bits/point +MSE 283.001321 +---------------------- -------------------------------------------------------- +Time: 0.367s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 283.0013 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,084B, BPFP=0.6178 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,472B, BPFP=2.2981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,268B, BPFP=1.0553 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,036B, BPFP=2.2107 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,348B, BPFP=1.2716 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,692B, BPFP=2.1418 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,012B, BPFP=1.2043 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,800B, BPFP=2.1635 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,072B, BPFP=1.8173 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,480B, BPFP=2.0994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,308B, BPFP=0.6384 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10595027 59.22807742 + layer.0.v_cache 0.00001376 0.00821085 + layer.1.k_cache 0.03307601 5.53540274 + layer.1.v_cache 0.00000550 0.00343330 + layer.2.k_cache 0.00695119 0.80421634 + layer.2.v_cache 0.00001929 0.00884532 + layer.3.k_cache 0.03450640 4.03435732 + layer.3.v_cache 0.00001844 0.00972724 + layer.4.k_cache 0.00061739 0.18756260 + layer.4.v_cache 0.00005297 0.01969946 + layer.4.output 0.18589628 687.28428342 + ------------------------------------------------------------------------------------- + TOTAL 0.08720501 287.10761862 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 106572 +BPFP 1.2558 bits/point +EBPFP 2.5116 equivalent bits/point +MSE 287.107619 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 287.1076 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,104B, BPFP=0.6139 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,512B, BPFP=2.2769 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,312B, BPFP=1.0506 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,212B, BPFP=2.2176 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,520B, BPFP=1.2896 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,884B, BPFP=2.1527 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,212B, BPFP=1.2286 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,140B, BPFP=2.2033 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,328B, BPFP=1.8449 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,860B, BPFP=2.1479 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,332B, BPFP=0.6027 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07772056 60.76035898 + layer.0.v_cache 0.00001351 0.00872719 + layer.1.k_cache 0.03518496 5.54195742 + layer.1.v_cache 0.00000556 0.00370946 + layer.2.k_cache 0.00230807 0.82540430 + layer.2.v_cache 0.00001833 0.00967826 + layer.3.k_cache 0.04514034 4.22783419 + layer.3.v_cache 0.00001901 0.01066314 + layer.4.k_cache 0.00062178 0.20525495 + layer.4.v_cache 0.00004990 0.02274901 + layer.4.output 0.18426369 681.36115506 + ------------------------------------------------------------------------------------- + TOTAL 0.08534870 284.77320131 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 107416 +BPFP 1.2497 bits/point +EBPFP 2.4994 equivalent bits/point +MSE 284.773201 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 284.7732 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.6250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,220B, BPFP=2.2768 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,188B, BPFP=1.0528 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,860B, BPFP=2.2037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,360B, BPFP=1.2906 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,580B, BPFP=2.1469 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,008B, BPFP=1.2192 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,792B, BPFP=2.1899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,948B, BPFP=1.8157 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,452B, BPFP=2.1209 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,392B, BPFP=0.6491 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14568528 59.87458147 + layer.0.v_cache 0.00001322 0.00832938 + layer.1.k_cache 0.01379078 5.44884620 + layer.1.v_cache 0.00000548 0.00356563 + layer.2.k_cache 0.00727733 0.84321921 + layer.2.v_cache 0.00001786 0.01013214 + layer.3.k_cache 0.03082054 3.82336386 + layer.3.v_cache 0.00001879 0.01110677 + layer.4.k_cache 0.00061820 0.20065295 + layer.4.v_cache 0.00005377 0.02179125 + layer.4.output 0.19662610 694.71596707 + ------------------------------------------------------------------------------------- + TOTAL 0.09262847 290.19160931 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 105880 +BPFP 1.2638 bits/point +EBPFP 2.5277 equivalent bits/point +MSE 290.191609 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.003s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1916 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,152B, BPFP=0.6156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,732B, BPFP=2.2914 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,424B, BPFP=1.0594 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,428B, BPFP=2.2320 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,492B, BPFP=1.2680 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,004B, BPFP=2.1492 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,132B, BPFP=1.1977 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,108B, BPFP=2.1695 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,252B, BPFP=1.8070 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,776B, BPFP=2.1047 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,708B, BPFP=0.6336 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10149734 59.19224243 + layer.0.v_cache 0.00001421 0.00896953 + layer.1.k_cache 0.01324784 5.72554016 + layer.1.v_cache 0.00000586 0.00404661 + layer.2.k_cache 0.00706157 0.84203186 + layer.2.v_cache 0.00001822 0.01015723 + layer.3.k_cache 0.02712160 3.74711723 + layer.3.v_cache 0.00002016 0.01051934 + layer.4.k_cache 0.00063687 0.20447955 + layer.4.v_cache 0.00005451 0.02153644 + layer.4.output 0.17858309 674.44882812 + ------------------------------------------------------------------------------------- + TOTAL 0.08233881 281.81814337 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 109208 +BPFP 1.2547 bits/point +EBPFP 2.5094 equivalent bits/point +MSE 281.818143 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 281.8181 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,304B, BPFP=0.5934 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,368B, BPFP=2.2213 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,888B, BPFP=1.0575 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,116B, BPFP=2.1760 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,112B, BPFP=1.2773 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,712B, BPFP=2.1034 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,772B, BPFP=1.2162 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,980B, BPFP=2.1516 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,184B, BPFP=1.8290 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,616B, BPFP=2.0862 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,900B, BPFP=0.6389 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10772817 62.69944325 + layer.0.v_cache 0.00001348 0.00872007 + layer.1.k_cache 0.03276526 5.42905961 + layer.1.v_cache 0.00000562 0.00367118 + layer.2.k_cache 0.01020141 0.82166729 + layer.2.v_cache 0.00001934 0.00950193 + layer.3.k_cache 0.04694761 3.75748908 + layer.3.v_cache 0.00001910 0.01066482 + layer.4.k_cache 0.00062564 0.19469811 + layer.4.v_cache 0.00005323 0.02075015 + layer.4.output 0.17166949 616.51493227 + ------------------------------------------------------------------------------------- + TOTAL 0.08235678 258.15059949 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 117952 +BPFP 1.2461 bits/point +EBPFP 2.4922 equivalent bits/point +MSE 258.150599 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.003s, Pack+Encode: 0.163s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 258.1506 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,228B, BPFP=0.6077 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,984B, BPFP=2.2560 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,384B, BPFP=1.0136 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,480B, BPFP=2.1611 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,716B, BPFP=1.2643 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,172B, BPFP=2.1032 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,256B, BPFP=1.1777 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,308B, BPFP=2.1288 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,460B, BPFP=1.7809 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,020B, BPFP=2.0745 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,676B, BPFP=0.6098 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12335577 57.20989034 + layer.0.v_cache 0.00001400 0.00895269 + layer.1.k_cache 0.01519604 5.17830355 + layer.1.v_cache 0.00000547 0.00376416 + layer.2.k_cache 0.01098718 0.83920279 + layer.2.v_cache 0.00001779 0.01071418 + layer.3.k_cache 0.07416311 3.69162539 + layer.3.v_cache 0.00001944 0.01142473 + layer.4.k_cache 0.00061719 0.21059606 + layer.4.v_cache 0.00006444 0.02290645 + layer.4.output 0.17504946 650.07067556 + ------------------------------------------------------------------------------------- + TOTAL 0.08528157 271.62835937 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 110684 +BPFP 1.2257 bits/point +EBPFP 2.4514 equivalent bits/point +MSE 271.628359 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.003s, Pack+Encode: 0.161s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 271.6284 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,056B, BPFP=0.6201 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,176B, BPFP=2.2679 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,264B, BPFP=1.0682 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,808B, BPFP=2.1932 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,256B, BPFP=1.2695 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,424B, BPFP=2.1153 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,928B, BPFP=1.2029 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,604B, BPFP=2.1518 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,928B, BPFP=1.8117 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,264B, BPFP=2.0828 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,264B, BPFP=0.6164 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13955514 60.09707919 + layer.0.v_cache 0.00001392 0.00867125 + layer.1.k_cache 0.03454666 5.40779015 + layer.1.v_cache 0.00000526 0.00390175 + layer.2.k_cache 0.00236268 0.83467786 + layer.2.v_cache 0.00001660 0.00939223 + layer.3.k_cache 0.02874016 4.22651335 + layer.3.v_cache 0.00001899 0.01079227 + layer.4.k_cache 0.00062191 0.21200029 + layer.4.v_cache 0.00005265 0.02175965 + layer.4.output 0.18728454 696.34589518 + ------------------------------------------------------------------------------------- + TOTAL 0.08923092 290.89728495 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 103972 +BPFP 1.2411 bits/point +EBPFP 2.4821 equivalent bits/point +MSE 290.897285 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 290.8973 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,056B, BPFP=0.6201 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,360B, BPFP=2.3052 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,256B, BPFP=1.0666 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,076B, BPFP=2.2476 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,368B, BPFP=1.2922 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,684B, BPFP=2.1680 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,012B, BPFP=1.2200 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,844B, BPFP=2.2005 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,056B, BPFP=1.8377 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,496B, BPFP=2.1299 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,288B, BPFP=0.6461 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11454884 61.13390955 + layer.0.v_cache 0.00001351 0.00904735 + layer.1.k_cache 0.01665594 5.28779067 + layer.1.v_cache 0.00000570 0.00414151 + layer.2.k_cache 0.00836552 0.94896193 + layer.2.v_cache 0.00001974 0.01064584 + layer.3.k_cache 0.03124422 3.94223775 + layer.3.v_cache 0.00001876 0.01141462 + layer.4.k_cache 0.00062231 0.20426230 + layer.4.v_cache 0.00005487 0.02300177 + layer.4.output 0.19663499 696.18216605 + ------------------------------------------------------------------------------------- + TOTAL 0.09105849 290.87356327 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 106496 +BPFP 1.2712 bits/point +EBPFP 2.5424 equivalent bits/point +MSE 290.873563 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 290.8736 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,220B, BPFP=0.5990 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,060B, BPFP=2.2433 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,508B, BPFP=1.0246 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,552B, BPFP=2.1488 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,556B, BPFP=1.2195 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,124B, BPFP=2.0692 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,248B, BPFP=1.1622 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,376B, BPFP=2.1161 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,400B, BPFP=1.7485 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,980B, BPFP=2.0424 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,884B, BPFP=0.6081 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12290487 55.76752000 + layer.0.v_cache 0.00001422 0.00918198 + layer.1.k_cache 0.01283375 4.98764474 + layer.1.v_cache 0.00000579 0.00435595 + layer.2.k_cache 0.01100613 0.92607898 + layer.2.v_cache 0.00001929 0.01055457 + layer.3.k_cache 0.03011658 3.66979690 + layer.3.v_cache 0.00001858 0.01118132 + layer.4.k_cache 0.00063020 0.22220248 + layer.4.v_cache 0.00005223 0.02390734 + layer.4.output 0.17536174 642.00499575 + ------------------------------------------------------------------------------------- + TOTAL 0.08265493 268.21572909 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 110908 +BPFP 1.2135 bits/point +EBPFP 2.4271 equivalent bits/point +MSE 268.215729 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 268.2157 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,236B, BPFP=0.5949 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,212B, BPFP=2.2449 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,588B, BPFP=1.0272 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,880B, BPFP=2.1838 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,688B, BPFP=1.2294 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,356B, BPFP=2.0875 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,356B, BPFP=1.1684 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,504B, BPFP=2.1147 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,668B, BPFP=1.7772 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,124B, BPFP=2.0449 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,356B, BPFP=0.6133 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11218034 56.26287339 + layer.0.v_cache 0.00001438 0.00872116 + layer.1.k_cache 0.03157643 5.44308220 + layer.1.v_cache 0.00000545 0.00378786 + layer.2.k_cache 0.00802605 0.81254326 + layer.2.v_cache 0.00001826 0.01005807 + layer.3.k_cache 0.04519741 3.60520738 + layer.3.v_cache 0.00001794 0.01023807 + layer.4.k_cache 0.00061137 0.19585701 + layer.4.v_cache 0.00005913 0.02145903 + layer.4.output 0.17010492 634.15766807 + ------------------------------------------------------------------------------------- + TOTAL 0.08167301 265.02808847 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 112968 +BPFP 1.2215 bits/point +EBPFP 2.4431 equivalent bits/point +MSE 265.028088 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 265.0281 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,260B, BPFP=0.5923 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,300B, BPFP=2.2347 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,632B, BPFP=1.0233 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,864B, BPFP=2.1555 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,816B, BPFP=1.2384 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,548B, BPFP=2.0981 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,452B, BPFP=1.1722 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,744B, BPFP=2.1337 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,884B, BPFP=1.7958 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,404B, BPFP=2.0719 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,264B, BPFP=0.6038 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10146660 58.64721112 + layer.0.v_cache 0.00001344 0.00847791 + layer.1.k_cache 0.03090366 5.31783809 + layer.1.v_cache 0.00000566 0.00386485 + layer.2.k_cache 0.01188999 0.78403012 + layer.2.v_cache 0.00001845 0.01003254 + layer.3.k_cache 0.02542871 3.69548780 + layer.3.v_cache 0.00001884 0.01065116 + layer.4.k_cache 0.00060243 0.20624300 + layer.4.v_cache 0.00005224 0.02132510 + layer.4.output 0.16839122 625.85465116 + ------------------------------------------------------------------------------------- + TOTAL 0.07936109 261.74633646 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 114168 +BPFP 1.2202 bits/point +EBPFP 2.4403 equivalent bits/point +MSE 261.746336 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 261.7463 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,280B, BPFP=0.5571 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,000B, BPFP=2.2079 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,068B, BPFP=1.0306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,776B, BPFP=2.1698 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,408B, BPFP=1.2582 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,248B, BPFP=2.0802 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,044B, BPFP=1.1963 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,456B, BPFP=2.1155 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,708B, BPFP=1.8186 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,148B, BPFP=2.0632 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,876B, BPFP=0.6036 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12981709 59.86587126 + layer.0.v_cache 0.00001628 0.00863856 + layer.1.k_cache 0.06761242 4.85760664 + layer.1.v_cache 0.00000547 0.00339045 + layer.2.k_cache 0.00384689 0.82988092 + layer.2.v_cache 0.00001812 0.00950194 + layer.3.k_cache 0.05327560 3.42079030 + layer.3.v_cache 0.00001896 0.01017058 + layer.4.k_cache 0.00060800 0.18779512 + layer.4.v_cache 0.00005116 0.02042126 + layer.4.output 0.15518735 585.90227096 + ------------------------------------------------------------------------------------- + TOTAL 0.07891655 245.32529199 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 122012 +BPFP 1.2189 bits/point +EBPFP 2.4379 equivalent bits/point +MSE 245.325292 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 245.3253 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,152B, BPFP=0.6080 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,800B, BPFP=2.2762 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,424B, BPFP=1.0463 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,388B, BPFP=2.1968 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,664B, BPFP=1.2855 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,148B, BPFP=2.1505 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,320B, BPFP=1.2191 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,340B, BPFP=2.1875 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,460B, BPFP=1.8248 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,028B, BPFP=2.1273 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,948B, BPFP=0.6324 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887507 57.82106602 + layer.0.v_cache 0.00001345 0.00857680 + layer.1.k_cache 0.03259810 5.33954159 + layer.1.v_cache 0.00000523 0.00349900 + layer.2.k_cache 0.00244350 0.92088148 + layer.2.v_cache 0.00001834 0.00942862 + layer.3.k_cache 0.10864470 4.10385998 + layer.3.v_cache 0.00001904 0.01060366 + layer.4.k_cache 0.00060864 0.19946155 + layer.4.v_cache 0.00004940 0.02102918 + layer.4.output 0.18628448 660.53301367 + ------------------------------------------------------------------------------------- + TOTAL 0.09395687 276.00994374 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 110672 +BPFP 1.2558 bits/point +EBPFP 2.5116 equivalent bits/point +MSE 276.009944 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 276.0099 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,380B, BPFP=0.5934 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,532B, BPFP=2.2001 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,948B, BPFP=1.0442 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,160B, BPFP=2.1348 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,080B, BPFP=1.2430 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,776B, BPFP=2.0674 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,704B, BPFP=1.1770 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,896B, BPFP=2.0885 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,972B, BPFP=1.7507 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,520B, BPFP=2.0225 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,424B, BPFP=0.6376 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422350 58.60778397 + layer.0.v_cache 0.00001493 0.00862923 + layer.1.k_cache 0.01422887 5.01166526 + layer.1.v_cache 0.00000595 0.00351893 + layer.2.k_cache 0.00668612 0.74360537 + layer.2.v_cache 0.00001902 0.00934028 + layer.3.k_cache 0.02497406 3.54451726 + layer.3.v_cache 0.00001869 0.00988068 + layer.4.k_cache 0.00061935 0.18419411 + layer.4.v_cache 0.00005212 0.02064011 + layer.4.output 0.16899524 602.42245185 + ------------------------------------------------------------------------------------- + TOTAL 0.07904761 252.06476107 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 118392 +BPFP 1.2227 bits/point +EBPFP 2.4453 equivalent bits/point +MSE 252.064761 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 252.0648 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 113, 128) +Output shape: (1, 113, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.output: torch.Size([1, 113, 3584]) -> torch.Size([1, 1, 113, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,808B, BPFP=0.5265 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,100B, BPFP=2.0879 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,736B, BPFP=0.9314 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,672B, BPFP=2.0288 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,304B, BPFP=1.1482 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,128B, BPFP=1.9535 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,808B, BPFP=1.0796 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,276B, BPFP=1.9740 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,052B, BPFP=1.6665 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,776B, BPFP=1.9049 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,380B, BPFP=0.5409 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10214825 56.56283704 + layer.0.v_cache 0.00001637 0.00850995 + layer.1.k_cache 0.04284562 4.33724138 + layer.1.v_cache 0.00000578 0.00381519 + layer.2.k_cache 0.00913706 0.85248005 + layer.2.v_cache 0.00001928 0.00964975 + layer.3.k_cache 0.05585717 3.39180411 + layer.3.v_cache 0.00001870 0.01046529 + layer.4.k_cache 0.00065968 0.20563402 + layer.4.v_cache 0.00005753 0.02228926 + layer.4.output 10.09706179 470.43125790 + ------------------------------------------------------------------------------------- + TOTAL 4.17001165 197.55432537 + (elements=983,552) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 983552 +Total Bytes 138040 +BPFP 1.1228 bits/point +EBPFP 2.2456 equivalent bits/point +MSE 197.554325 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.005s, Pack+Encode: 0.162s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 197.5543 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,356B, BPFP=0.5762 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,832B, BPFP=2.2033 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,892B, BPFP=1.0117 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,396B, BPFP=2.1284 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,312B, BPFP=1.2555 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,036B, BPFP=2.0666 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,844B, BPFP=1.1751 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,264B, BPFP=2.1058 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,376B, BPFP=1.7816 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,908B, BPFP=2.0446 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,344B, BPFP=0.6462 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11481400 60.34486607 + layer.0.v_cache 0.00001396 0.00903576 + layer.1.k_cache 0.03135788 4.85789842 + layer.1.v_cache 0.00000545 0.00368886 + layer.2.k_cache 0.00370715 0.77767324 + layer.2.v_cache 0.00001965 0.00975739 + layer.3.k_cache 0.03930106 3.45499630 + layer.3.v_cache 0.00001886 0.01030309 + layer.4.k_cache 0.00060819 0.19608282 + layer.4.v_cache 0.00005437 0.02113226 + layer.4.output 0.16591480 593.44304356 + ------------------------------------------------------------------------------------- + TOTAL 0.07948848 248.45804348 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 121560 +BPFP 1.2278 bits/point +EBPFP 2.4556 equivalent bits/point +MSE 248.458043 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 248.4580 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,376B, BPFP=0.5927 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,432B, BPFP=2.1826 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,824B, BPFP=1.0225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,000B, BPFP=2.1067 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,100B, BPFP=1.2465 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,676B, BPFP=2.0499 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,584B, BPFP=1.1559 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,788B, BPFP=2.0695 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,028B, BPFP=1.7605 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,416B, BPFP=2.0042 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,084B, BPFP=0.5790 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09788724 58.89014221 + layer.0.v_cache 0.00001308 0.00869673 + layer.1.k_cache 0.03299916 5.07529132 + layer.1.v_cache 0.00000544 0.00390191 + layer.2.k_cache 0.00367894 0.89007294 + layer.2.v_cache 0.00002012 0.01056866 + layer.3.k_cache 0.03939044 3.52784438 + layer.3.v_cache 0.00001875 0.01022432 + layer.4.k_cache 0.00061788 0.20184579 + layer.4.v_cache 0.00005010 0.02200479 + layer.4.output 0.17021053 604.50185594 + ------------------------------------------------------------------------------------- + TOTAL 0.08036205 252.95021086 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 115308 +BPFP 1.1908 bits/point +EBPFP 2.3816 equivalent bits/point +MSE 252.950211 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 252.9502 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,720B, BPFP=0.5382 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,728B, BPFP=2.1308 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,588B, BPFP=0.9531 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,316B, BPFP=2.0712 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,144B, BPFP=1.1782 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,852B, BPFP=2.0041 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,712B, BPFP=1.1157 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,176B, BPFP=2.0509 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,772B, BPFP=1.7031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,656B, BPFP=1.9757 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,812B, BPFP=0.6782 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12672405 55.84657118 + layer.0.v_cache 0.00001401 0.00821266 + layer.1.k_cache 0.12044146 5.08360177 + layer.1.v_cache 0.00000568 0.00368351 + layer.2.k_cache 0.00570256 0.78902139 + layer.2.v_cache 0.00001866 0.00922898 + layer.3.k_cache 0.03611955 3.33698103 + layer.3.v_cache 0.00001983 0.01008844 + layer.4.k_cache 0.00066233 0.19178431 + layer.4.v_cache 0.00005331 0.02013113 + layer.4.output 10.56037249 492.67439649 + ------------------------------------------------------------------------------------- + TOTAL 4.36543346 206.70706352 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 141476 +BPFP 1.2040 bits/point +EBPFP 2.4080 equivalent bits/point +MSE 206.707064 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.006s, Pack+Encode: 0.161s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 206.7071 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,336B, BPFP=0.5923 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,448B, BPFP=2.2102 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,760B, BPFP=1.0227 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,988B, BPFP=2.1286 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,960B, BPFP=1.2358 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,664B, BPFP=2.0710 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,524B, BPFP=1.1584 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,812B, BPFP=2.0973 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,060B, BPFP=1.7862 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,420B, BPFP=2.0277 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,136B, BPFP=0.5869 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13733016 62.57936790 + layer.0.v_cache 0.00001373 0.00936514 + layer.1.k_cache 0.03155638 4.94892883 + layer.1.v_cache 0.00000617 0.00437695 + layer.2.k_cache 0.00751131 0.76465563 + layer.2.v_cache 0.00001754 0.01088251 + layer.3.k_cache 0.02548880 3.76743872 + layer.3.v_cache 0.00001765 0.01214036 + layer.4.k_cache 0.00061011 0.21567141 + layer.4.v_cache 0.00005045 0.02441705 + layer.4.output 0.16239834 609.07213880 + ------------------------------------------------------------------------------------- + TOTAL 0.07878769 255.04954212 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 115108 +BPFP 1.2022 bits/point +EBPFP 2.4045 equivalent bits/point +MSE 255.049542 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.005s, Pack+Encode: 0.163s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 255.0495 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,096B, BPFP=0.6123 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,492B, BPFP=2.2729 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,312B, BPFP=1.0506 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,120B, BPFP=2.1994 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,420B, BPFP=1.2698 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,808B, BPFP=2.1377 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,112B, BPFP=1.2089 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,024B, BPFP=2.1804 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,168B, BPFP=1.8133 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,656B, BPFP=2.1076 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,740B, BPFP=0.6143 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10979201 57.56233312 + layer.0.v_cache 0.00001339 0.00868121 + layer.1.k_cache 0.01160004 5.71052416 + layer.1.v_cache 0.00000567 0.00384519 + layer.2.k_cache 0.00714037 0.89800523 + layer.2.v_cache 0.00001829 0.00908813 + layer.3.k_cache 0.04622446 3.81145622 + layer.3.v_cache 0.00001896 0.01017168 + layer.4.k_cache 0.00064378 0.19871394 + layer.4.v_cache 0.00005251 0.02098080 + layer.4.output 0.18415536 682.66964286 + ------------------------------------------------------------------------------------- + TOTAL 0.08615276 285.11301763 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 106948 +BPFP 1.2443 bits/point +EBPFP 2.4886 equivalent bits/point +MSE 285.113018 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.003s, Pack+Encode: 0.163s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 285.1130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,036B, BPFP=0.6161 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,316B, BPFP=2.2963 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,268B, BPFP=1.0690 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,984B, BPFP=2.2289 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,304B, BPFP=1.2792 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,560B, BPFP=2.1429 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,976B, BPFP=1.2127 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,724B, BPFP=2.1761 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,892B, BPFP=1.8044 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,408B, BPFP=2.1120 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,328B, BPFP=0.6763 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11290005 59.43342253 + layer.0.v_cache 0.00001570 0.00821633 + layer.1.k_cache 0.03434501 5.28109702 + layer.1.v_cache 0.00000534 0.00332069 + layer.2.k_cache 0.00238675 0.86814236 + layer.2.v_cache 0.00001838 0.00898153 + layer.3.k_cache 0.02838124 3.78957218 + layer.3.v_cache 0.00002047 0.00969016 + layer.4.k_cache 0.00062757 0.19205495 + layer.4.v_cache 0.00005455 0.02000179 + layer.4.output 0.18420308 695.41941095 + ------------------------------------------------------------------------------------- + TOTAL 0.08636333 290.44413977 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 106796 +BPFP 1.2748 bits/point +EBPFP 2.5496 equivalent bits/point +MSE 290.444140 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 290.4441 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,084B, BPFP=0.6178 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,436B, BPFP=2.2909 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,304B, BPFP=1.0625 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,132B, BPFP=2.2300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,384B, BPFP=1.2788 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,800B, BPFP=2.1635 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,048B, BPFP=1.2115 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,952B, BPFP=2.1939 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,076B, BPFP=1.8181 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,624B, BPFP=2.1282 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,608B, BPFP=0.6470 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09752406 58.93320563 + layer.0.v_cache 0.00001393 0.00890417 + layer.1.k_cache 0.01493649 5.48466922 + layer.1.v_cache 0.00000588 0.00371775 + layer.2.k_cache 0.01003601 0.90608294 + layer.2.v_cache 0.00001805 0.00954599 + layer.3.k_cache 0.04451986 4.01546772 + layer.3.v_cache 0.00001914 0.01096466 + layer.4.k_cache 0.00061869 0.20293644 + layer.4.v_cache 0.00005251 0.02193988 + layer.4.output 0.18409089 687.38850733 + ------------------------------------------------------------------------------------- + TOTAL 0.08566946 287.13629328 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 107448 +BPFP 1.2661 bits/point +EBPFP 2.5322 equivalent bits/point +MSE 287.136293 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 287.1363 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,240B, BPFP=0.6099 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,900B, BPFP=2.2402 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,420B, BPFP=1.0203 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,404B, BPFP=2.1468 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,448B, BPFP=1.2139 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,052B, BPFP=2.0806 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,208B, BPFP=1.1687 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,324B, BPFP=2.1318 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,404B, BPFP=1.7703 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,004B, BPFP=2.0715 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,980B, BPFP=0.6180 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13658243 56.82417169 + layer.0.v_cache 0.00001379 0.00872869 + layer.1.k_cache 0.01234693 5.02539173 + layer.1.v_cache 0.00000566 0.00383380 + layer.2.k_cache 0.00375735 0.85822462 + layer.2.v_cache 0.00001852 0.00971316 + layer.3.k_cache 0.02700986 3.56252316 + layer.3.v_cache 0.00001930 0.01081979 + layer.4.k_cache 0.00062007 0.20399096 + layer.4.v_cache 0.00005146 0.02147071 + layer.4.output 0.17614663 649.04233003 + ------------------------------------------------------------------------------------- + TOTAL 0.08314423 271.16618697 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 110384 +BPFP 1.2224 bits/point +EBPFP 2.4447 equivalent bits/point +MSE 271.166187 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.003s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 271.1662 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,972B, BPFP=0.6361 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,692B, BPFP=2.2885 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,068B, BPFP=1.0848 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,148B, BPFP=2.1721 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,056B, BPFP=1.2962 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,812B, BPFP=2.1002 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,708B, BPFP=1.2217 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,080B, BPFP=2.1575 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,368B, BPFP=1.7911 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,828B, BPFP=2.1036 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,472B, BPFP=0.6871 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08738674 61.84655929 + layer.0.v_cache 0.00001368 0.00912106 + layer.1.k_cache 0.01286346 5.74192078 + layer.1.v_cache 0.00000597 0.00407549 + layer.2.k_cache 0.00251921 0.89933840 + layer.2.v_cache 0.00001769 0.01032957 + layer.3.k_cache 0.08324167 3.49173768 + layer.3.v_cache 0.00002021 0.01089546 + layer.4.k_cache 0.00062021 0.21460936 + layer.4.v_cache 0.00005375 0.02309243 + layer.4.output 0.19567458 735.14573141 + ------------------------------------------------------------------------------------- + TOTAL 0.09155674 306.95716467 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 101204 +BPFP 1.2742 bits/point +EBPFP 2.5484 equivalent bits/point +MSE 306.957165 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 306.9572 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,256B, BPFP=0.6130 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,840B, BPFP=2.2289 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,464B, BPFP=1.0286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,412B, BPFP=2.1483 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,576B, BPFP=1.2380 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,064B, BPFP=2.0828 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,216B, BPFP=1.1702 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,276B, BPFP=2.1227 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,376B, BPFP=1.7651 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,884B, BPFP=2.0489 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,512B, BPFP=0.6323 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12176956 57.94044145 + layer.0.v_cache 0.00001396 0.00864186 + layer.1.k_cache 0.01120206 5.20894945 + layer.1.v_cache 0.00000562 0.00382417 + layer.2.k_cache 0.00678912 0.77959442 + layer.2.v_cache 0.00001849 0.00957866 + layer.3.k_cache 0.02622183 3.95255399 + layer.3.v_cache 0.00001965 0.01072537 + layer.4.k_cache 0.00062413 0.20558991 + layer.4.v_cache 0.00005248 0.02082752 + layer.4.output 0.17940697 648.88516566 + ------------------------------------------------------------------------------------- + TOTAL 0.08368034 271.19628744 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 110876 +BPFP 1.2278 bits/point +EBPFP 2.4556 equivalent bits/point +MSE 271.196287 +---------------------- -------------------------------------------------------- +Time: 0.367s Load: 0.003s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 271.1963 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,660B, BPFP=0.5446 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,500B, BPFP=2.1577 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,568B, BPFP=0.9774 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,040B, BPFP=2.0893 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,072B, BPFP=1.2012 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,536B, BPFP=2.0143 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,724B, BPFP=1.1494 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,840B, BPFP=2.0595 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,688B, BPFP=1.7393 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,504B, BPFP=2.0095 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,672B, BPFP=0.6520 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120816 58.31476004 + layer.0.v_cache 0.00001607 0.00844370 + layer.1.k_cache 0.10678518 4.61043701 + layer.1.v_cache 0.00000585 0.00339804 + layer.2.k_cache 0.00813519 0.82236633 + layer.2.v_cache 0.00001919 0.00958129 + layer.3.k_cache 0.06560528 3.07530866 + layer.3.v_cache 0.00001973 0.01063224 + layer.4.k_cache 0.00062300 0.19108107 + layer.4.v_cache 0.00005287 0.02024261 + layer.4.output 10.86395687 507.83431122 + ------------------------------------------------------------------------------------- + TOTAL 4.49059815 213.05331939 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 137804 +BPFP 1.2063 bits/point +EBPFP 2.4125 equivalent bits/point +MSE 213.053319 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 213.0533 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,272B, BPFP=0.5876 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,332B, BPFP=2.2148 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,648B, BPFP=1.0144 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,940B, BPFP=2.1444 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,864B, BPFP=1.2328 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,568B, BPFP=2.0776 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,444B, BPFP=1.1573 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,784B, BPFP=2.1164 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,900B, BPFP=1.7780 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,388B, BPFP=2.0453 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,480B, BPFP=0.6281 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12848519 59.88274515 + layer.0.v_cache 0.00001371 0.00829078 + layer.1.k_cache 0.03007136 5.03823186 + layer.1.v_cache 0.00000570 0.00372294 + layer.2.k_cache 0.00936044 0.77527083 + layer.2.v_cache 0.00001842 0.00960126 + layer.3.k_cache 0.04065019 3.36129094 + layer.3.v_cache 0.00001877 0.01099343 + layer.4.k_cache 0.00062204 0.20106820 + layer.4.v_cache 0.00010837 0.02104239 + layer.4.output 0.16426703 612.14295977 + ------------------------------------------------------------------------------------- + TOTAL 0.07995432 256.13605742 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 115620 +BPFP 1.2215 bits/point +EBPFP 2.4430 equivalent bits/point +MSE 256.136057 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 256.1361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,144B, BPFP=0.6141 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,752B, BPFP=2.2953 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,428B, BPFP=1.0602 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,396B, BPFP=2.2258 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,664B, BPFP=1.3016 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,076B, BPFP=2.1633 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,224B, BPFP=1.2156 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,184B, BPFP=2.1844 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,436B, BPFP=1.8430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,892B, BPFP=2.1273 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,000B, BPFP=0.6417 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09257430 59.46525879 + layer.0.v_cache 0.00001329 0.00825752 + layer.1.k_cache 0.01202901 5.55698166 + layer.1.v_cache 0.00000544 0.00320456 + layer.2.k_cache 0.00549439 0.86886139 + layer.2.v_cache 0.00001849 0.00887978 + layer.3.k_cache 0.11014032 3.67804337 + layer.3.v_cache 0.00001916 0.00938014 + layer.4.k_cache 0.00063444 0.18760152 + layer.4.v_cache 0.00005494 0.01852674 + layer.4.output 0.19143520 670.84157366 + ------------------------------------------------------------------------------------- + TOTAL 0.09182531 280.33505948 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 110196 +BPFP 1.2660 bits/point +EBPFP 2.5321 equivalent bits/point +MSE 280.335059 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 280.3351 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,228B, BPFP=0.6004 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,028B, BPFP=2.2374 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,612B, BPFP=1.0439 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,604B, BPFP=2.1585 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,640B, BPFP=1.2351 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,188B, BPFP=2.0811 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,344B, BPFP=1.1801 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,364B, BPFP=2.1138 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,440B, BPFP=1.7560 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,984B, BPFP=2.0432 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,248B, BPFP=0.5912 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12365316 56.77267020 + layer.0.v_cache 0.00001332 0.00912025 + layer.1.k_cache 0.03511774 4.51793126 + layer.1.v_cache 0.00000569 0.00436432 + layer.2.k_cache 0.01236608 0.90175093 + layer.2.v_cache 0.00001835 0.01014738 + layer.3.k_cache 0.02851222 3.63545590 + layer.3.v_cache 0.00001844 0.01156155 + layer.4.k_cache 0.00063207 0.21559777 + layer.4.v_cache 0.00005066 0.02361521 + layer.4.output 0.17964420 639.38350340 + ------------------------------------------------------------------------------------- + TOTAL 0.08575865 267.16392580 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 110680 +BPFP 1.2110 bits/point +EBPFP 2.4221 equivalent bits/point +MSE 267.163926 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 267.1639 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,072B, BPFP=0.6154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,504B, BPFP=2.3045 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,292B, BPFP=1.0601 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,132B, BPFP=2.2300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,360B, BPFP=1.2740 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,788B, BPFP=2.1611 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,992B, BPFP=1.2003 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,844B, BPFP=2.1723 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,032B, BPFP=1.8093 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,564B, BPFP=2.1162 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,704B, BPFP=0.6497 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12096892 59.13748247 + layer.0.v_cache 0.00001393 0.00855000 + layer.1.k_cache 0.03247837 5.29702954 + layer.1.v_cache 0.00000613 0.00357824 + layer.2.k_cache 0.00394396 0.82364498 + layer.2.v_cache 0.00001920 0.00959389 + layer.3.k_cache 0.02751121 4.24960249 + layer.3.v_cache 0.00001919 0.01053086 + layer.4.k_cache 0.00062481 0.20482472 + layer.4.v_cache 0.00005293 0.02071720 + layer.4.output 0.18564256 685.43835852 + ------------------------------------------------------------------------------------- + TOTAL 0.08736098 286.34318024 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 107284 +BPFP 1.2642 bits/point +EBPFP 2.5284 equivalent bits/point +MSE 286.343180 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 286.3432 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,036B, BPFP=0.6161 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,272B, BPFP=2.2873 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,220B, BPFP=1.0593 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,796B, BPFP=2.1907 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,304B, BPFP=1.2792 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,492B, BPFP=2.1291 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,988B, BPFP=1.2151 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,664B, BPFP=2.1640 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,948B, BPFP=1.8157 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,468B, BPFP=2.1242 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,080B, BPFP=0.6691 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08564486 61.06293755 + layer.0.v_cache 0.00001408 0.00891635 + layer.1.k_cache 0.03507188 5.33919773 + layer.1.v_cache 0.00000566 0.00388528 + layer.2.k_cache 0.00239554 0.80464172 + layer.2.v_cache 0.00001938 0.01043248 + layer.3.k_cache 0.03313774 3.84304849 + layer.3.v_cache 0.00001993 0.01178434 + layer.4.k_cache 0.00061714 0.21349421 + layer.4.v_cache 0.00005134 0.02273629 + layer.4.output 0.18226704 694.42596243 + ------------------------------------------------------------------------------------- + TOTAL 0.08428511 290.13545950 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 106268 +BPFP 1.2685 bits/point +EBPFP 2.5370 equivalent bits/point +MSE 290.135459 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 290.1355 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 117, 128) +Output shape: (1, 117, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.output: torch.Size([1, 117, 3584]) -> torch.Size([1, 1, 117, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,844B, BPFP=0.5134 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,380B, BPFP=2.0540 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,972B, BPFP=0.9311 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,116B, BPFP=2.0187 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,752B, BPFP=1.1688 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,624B, BPFP=1.9530 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,340B, BPFP=1.1138 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,832B, BPFP=1.9808 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,576B, BPFP=1.6795 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,240B, BPFP=1.9017 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,936B, BPFP=0.5902 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11900647 53.97263956 + layer.0.v_cache 0.00001428 0.00886461 + layer.1.k_cache 0.12535267 4.80116833 + layer.1.v_cache 0.00000579 0.00363707 + layer.2.k_cache 0.00517510 0.77224966 + layer.2.v_cache 0.00001992 0.00994405 + layer.3.k_cache 0.09315677 3.24614148 + layer.3.v_cache 0.00001978 0.01052402 + layer.4.k_cache 0.00061896 0.18565763 + layer.4.v_cache 0.00005133 0.01938253 + layer.4.output 9.75369281 456.12023046 + ------------------------------------------------------------------------------------- + TOTAL 4.03642769 191.52187189 + (elements=1,018,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1018368 +Total Bytes 145612 +BPFP 1.1439 bits/point +EBPFP 2.2878 equivalent bits/point +MSE 191.521872 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 191.5219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,668B, BPFP=0.5458 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,572B, BPFP=2.1685 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,780B, BPFP=1.0089 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,200B, BPFP=2.1131 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,212B, BPFP=1.2220 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,660B, BPFP=2.0327 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,740B, BPFP=1.1518 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,832B, BPFP=2.0583 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,892B, BPFP=1.7696 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,468B, BPFP=2.0042 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,912B, BPFP=0.5721 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14131550 57.80955171 + layer.0.v_cache 0.00001383 0.00835596 + layer.1.k_cache 0.02978072 5.22351016 + layer.1.v_cache 0.00000537 0.00337556 + layer.2.k_cache 0.00578868 0.83729125 + layer.2.v_cache 0.00001872 0.00905298 + layer.3.k_cache 0.04020810 3.99723307 + layer.3.v_cache 0.00001850 0.01026447 + layer.4.k_cache 0.00063358 0.19418867 + layer.4.v_cache 0.00005294 0.02193319 + layer.4.output 10.86925024 507.47002551 + ------------------------------------------------------------------------------------- + TOTAL 4.48838751 212.96499621 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 134936 +BPFP 1.1812 bits/point +EBPFP 2.3623 equivalent bits/point +MSE 212.964996 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.006s, Pack+Encode: 0.161s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 212.9650 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,236B, BPFP=0.6019 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,104B, BPFP=2.2515 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,484B, BPFP=1.0201 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,740B, BPFP=2.1838 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,652B, BPFP=1.2374 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,304B, BPFP=2.1027 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,348B, BPFP=1.1808 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,552B, BPFP=2.1488 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,624B, BPFP=1.7902 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,068B, BPFP=2.0588 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,780B, BPFP=0.6053 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09514355 55.77197266 + layer.0.v_cache 0.00001423 0.00883046 + layer.1.k_cache 0.03283414 4.95627521 + layer.1.v_cache 0.00000553 0.00352548 + layer.2.k_cache 0.00489674 0.80480348 + layer.2.v_cache 0.00001883 0.00920556 + layer.3.k_cache 0.07199075 3.72370039 + layer.3.v_cache 0.00001932 0.01038009 + layer.4.k_cache 0.00061592 0.19557669 + layer.4.v_cache 0.00006816 0.02059060 + layer.4.output 0.17253765 642.11936650 + ------------------------------------------------------------------------------------- + TOTAL 0.08313946 268.25531918 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 111892 +BPFP 1.2243 bits/point +EBPFP 2.4486 equivalent bits/point +MSE 268.255319 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.003s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 268.2553 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,300B, BPFP=0.5927 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,336B, BPFP=2.2155 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,676B, BPFP=1.0194 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,868B, BPFP=2.1315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,816B, BPFP=1.2241 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,532B, BPFP=2.0711 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,544B, BPFP=1.1753 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,704B, BPFP=2.1020 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,968B, BPFP=1.7902 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,548B, BPFP=2.0740 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,508B, BPFP=0.6031 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11660385 61.06617614 + layer.0.v_cache 0.00001381 0.00892934 + layer.1.k_cache 0.05175667 5.24101029 + layer.1.v_cache 0.00000569 0.00399247 + layer.2.k_cache 0.00227456 0.80662212 + layer.2.v_cache 0.00001743 0.01000328 + layer.3.k_cache 0.02781237 3.46263315 + layer.3.v_cache 0.00001883 0.01073686 + layer.4.k_cache 0.00061846 0.20533187 + layer.4.v_cache 0.00005344 0.02183734 + layer.4.output 0.17410791 617.99758826 + ------------------------------------------------------------------------------------- + TOTAL 0.08340767 258.63649357 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 114800 +BPFP 1.2128 bits/point +EBPFP 2.4256 equivalent bits/point +MSE 258.636494 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 258.6365 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,252B, BPFP=0.6122 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,856B, BPFP=2.2319 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,504B, BPFP=1.0361 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,428B, BPFP=2.1514 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,580B, BPFP=1.2387 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,132B, BPFP=2.0956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,280B, BPFP=1.1822 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,204B, BPFP=2.1092 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,304B, BPFP=1.7515 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,772B, BPFP=2.0279 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,664B, BPFP=0.6095 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212080 55.80846433 + layer.0.v_cache 0.00001360 0.00847495 + layer.1.k_cache 0.03288655 5.03828154 + layer.1.v_cache 0.00000536 0.00379285 + layer.2.k_cache 0.00973156 0.79851477 + layer.2.v_cache 0.00001762 0.00953237 + layer.3.k_cache 0.02906220 3.95910130 + layer.3.v_cache 0.00001891 0.01023970 + layer.4.k_cache 0.00063934 0.20537330 + layer.4.v_cache 0.00004866 0.02179458 + layer.4.output 0.18425958 650.39280336 + ------------------------------------------------------------------------------------- + TOTAL 0.08496245 271.68312901 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 109976 +BPFP 1.2178 bits/point +EBPFP 2.4357 equivalent bits/point +MSE 271.683129 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 271.6831 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,240B, BPFP=0.5956 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,272B, BPFP=2.2559 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,576B, BPFP=1.0250 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,892B, BPFP=2.1860 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,752B, BPFP=1.2412 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,480B, BPFP=2.1103 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,556B, BPFP=1.2051 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,616B, BPFP=2.1353 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,700B, BPFP=1.7831 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,260B, BPFP=2.0699 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,184B, BPFP=0.6351 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10758395 55.27501149 + layer.0.v_cache 0.00001397 0.00838165 + layer.1.k_cache 0.03293324 5.20099092 + layer.1.v_cache 0.00000575 0.00336438 + layer.2.k_cache 0.00823421 0.85803312 + layer.2.v_cache 0.00001745 0.00901330 + layer.3.k_cache 0.04116619 3.61140891 + layer.3.v_cache 0.00001985 0.01043815 + layer.4.k_cache 0.00062779 0.19349260 + layer.4.v_cache 0.00005042 0.02041321 + layer.4.output 0.16954060 633.69936975 + ------------------------------------------------------------------------------------- + TOTAL 0.08102571 264.76977270 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 114528 +BPFP 1.2384 bits/point +EBPFP 2.4768 equivalent bits/point +MSE 264.769773 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 264.7698 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,372B, BPFP=0.5854 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,616B, BPFP=2.1903 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,884B, BPFP=1.0215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,400B, BPFP=2.1528 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,140B, BPFP=1.2396 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,860B, BPFP=2.0590 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,696B, BPFP=1.1625 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,068B, BPFP=2.0951 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,224B, BPFP=1.7750 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,752B, BPFP=2.0403 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,648B, BPFP=0.6113 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12307767 59.26816949 + layer.0.v_cache 0.00001410 0.00878669 + layer.1.k_cache 0.03088827 4.78137444 + layer.1.v_cache 0.00000584 0.00375514 + layer.2.k_cache 0.00371391 0.79847022 + layer.2.v_cache 0.00001950 0.01023021 + layer.3.k_cache 0.03079298 3.48496263 + layer.3.v_cache 0.00001950 0.01055816 + layer.4.k_cache 0.00063014 0.20367908 + layer.4.v_cache 0.00005282 0.02235289 + layer.4.output 0.15767360 598.33849206 + ------------------------------------------------------------------------------------- + TOTAL 0.07605470 250.40951667 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 118660 +BPFP 1.2118 bits/point +EBPFP 2.4236 equivalent bits/point +MSE 250.409517 +---------------------- -------------------------------------------------------- +Time: 0.372s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 250.4095 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,168B, BPFP=0.6037 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,752B, BPFP=2.2393 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,344B, BPFP=1.0183 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,340B, BPFP=2.1608 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,568B, BPFP=1.2515 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,112B, BPFP=2.1174 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,172B, BPFP=1.1761 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,108B, BPFP=2.1166 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,328B, BPFP=1.7774 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,804B, BPFP=2.0587 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,492B, BPFP=0.6123 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10295349 57.84048685 + layer.0.v_cache 0.00001348 0.00831300 + layer.1.k_cache 0.03334363 5.18265180 + layer.1.v_cache 0.00000550 0.00366396 + layer.2.k_cache 0.01126997 0.80840357 + layer.2.v_cache 0.00001933 0.01011182 + layer.3.k_cache 0.09131574 4.15108732 + layer.3.v_cache 0.00001769 0.01024050 + layer.4.k_cache 0.00062275 0.20095041 + layer.4.v_cache 0.00006541 0.02073379 + layer.4.output 0.17512874 657.41060540 + ------------------------------------------------------------------------------------- + TOTAL 0.08620754 274.71240475 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 109188 +BPFP 1.2239 bits/point +EBPFP 2.4477 equivalent bits/point +MSE 274.712405 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 274.7124 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,308B, BPFP=0.5941 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,388B, BPFP=2.2249 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,796B, BPFP=1.0409 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,088B, BPFP=2.1710 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,920B, BPFP=1.2428 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,692B, BPFP=2.0999 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,620B, BPFP=1.1889 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,924B, BPFP=2.1415 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,972B, BPFP=1.7909 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,516B, BPFP=2.0682 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,528B, BPFP=0.6037 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12314748 59.70893948 + layer.0.v_cache 0.00001636 0.00868134 + layer.1.k_cache 0.01190477 5.84927964 + layer.1.v_cache 0.00000588 0.00354266 + layer.2.k_cache 0.01614088 0.81123291 + layer.2.v_cache 0.00001834 0.00925730 + layer.3.k_cache 0.05604688 3.61366991 + layer.3.v_cache 0.00001882 0.01055856 + layer.4.k_cache 0.00063774 0.18924143 + layer.4.v_cache 0.00005162 0.01932250 + layer.4.output 0.17348555 619.37838670 + ------------------------------------------------------------------------------------- + TOTAL 0.08366986 259.16896663 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 115752 +BPFP 1.2229 bits/point +EBPFP 2.4457 equivalent bits/point +MSE 259.168967 +---------------------- -------------------------------------------------------- +Time: 0.367s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 259.1690 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,168B, BPFP=0.6111 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,824B, BPFP=2.2809 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,376B, BPFP=1.0370 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,308B, BPFP=2.1813 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,544B, BPFP=1.2623 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,892B, BPFP=2.1011 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,164B, BPFP=1.1890 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,076B, BPFP=2.1366 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,332B, BPFP=1.8002 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,840B, BPFP=2.0910 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,084B, BPFP=0.6361 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11560643 56.72840109 + layer.0.v_cache 0.00001402 0.00931744 + layer.1.k_cache 0.03272506 5.63230312 + layer.1.v_cache 0.00000563 0.00407431 + layer.2.k_cache 0.00645547 0.83001182 + layer.2.v_cache 0.00001830 0.01046771 + layer.3.k_cache 0.02731901 4.07449454 + layer.3.v_cache 0.00001960 0.01129292 + layer.4.k_cache 0.00061302 0.20713309 + layer.4.v_cache 0.00004978 0.02198600 + layer.4.output 0.17251868 665.50049603 + ------------------------------------------------------------------------------------- + TOTAL 0.08179159 278.00193849 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 109608 +BPFP 1.2437 bits/point +EBPFP 2.4875 equivalent bits/point +MSE 278.001938 +---------------------- -------------------------------------------------------- +Time: 0.375s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 278.0019 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,356B, BPFP=0.5959 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,516B, BPFP=2.2223 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,836B, BPFP=1.0362 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,076B, BPFP=2.1442 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,008B, BPFP=1.2443 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,700B, BPFP=2.0774 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,624B, BPFP=1.1761 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,016B, BPFP=2.1335 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,020B, BPFP=1.7791 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,444B, BPFP=2.0320 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,320B, BPFP=0.6422 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11690623 62.36922940 + layer.0.v_cache 0.00001418 0.00882683 + layer.1.k_cache 0.03305150 5.16128332 + layer.1.v_cache 0.00000545 0.00386069 + layer.2.k_cache 0.01095151 0.78078738 + layer.2.v_cache 0.00001854 0.00941068 + layer.3.k_cache 0.04293454 3.51316001 + layer.3.v_cache 0.00002007 0.01075100 + layer.4.k_cache 0.00062489 0.19298991 + layer.4.v_cache 0.00005451 0.02051354 + layer.4.output 0.16672201 610.24157873 + ------------------------------------------------------------------------------------- + TOTAL 0.08068444 255.51540376 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 117916 +BPFP 1.2316 bits/point +EBPFP 2.4632 equivalent bits/point +MSE 255.515404 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 255.5154 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,364B, BPFP=0.5776 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,692B, BPFP=2.1793 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,856B, BPFP=1.0055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,332B, BPFP=2.1174 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,180B, BPFP=1.2328 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,848B, BPFP=2.0343 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,792B, BPFP=1.1662 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,100B, BPFP=2.0776 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,296B, BPFP=1.7679 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,724B, BPFP=2.0130 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,036B, BPFP=0.5896 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11067300 59.71009508 + layer.0.v_cache 0.00001409 0.00900583 + layer.1.k_cache 0.05152405 4.74977933 + layer.1.v_cache 0.00000569 0.00382772 + layer.2.k_cache 0.01362713 0.75818978 + layer.2.v_cache 0.00001889 0.01016550 + layer.3.k_cache 0.08227781 3.65149496 + layer.3.v_cache 0.00001889 0.01038021 + layer.4.k_cache 0.00062414 0.20163913 + layer.4.v_cache 0.00005257 0.02022608 + layer.4.output 0.15746453 592.73165228 + ------------------------------------------------------------------------------------- + TOTAL 0.08006400 248.13213939 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 118220 +BPFP 1.1940 bits/point +EBPFP 2.3881 equivalent bits/point +MSE 248.132139 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 248.1321 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 116, 128) +Output shape: (1, 116, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.output: torch.Size([1, 116, 3584]) -> torch.Size([1, 1, 116, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,896B, BPFP=0.5248 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,236B, BPFP=2.0523 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,872B, BPFP=0.9256 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,944B, BPFP=2.0129 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,592B, BPFP=1.1573 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,532B, BPFP=1.9574 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,140B, BPFP=1.0964 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,608B, BPFP=1.9677 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,380B, BPFP=1.6676 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,060B, BPFP=1.8939 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,600B, BPFP=0.5503 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16314862 53.07925731 + layer.0.v_cache 0.00001373 0.00854065 + layer.1.k_cache 0.05481785 4.93260877 + layer.1.v_cache 0.00000557 0.00397739 + layer.2.k_cache 0.00968372 0.82479326 + layer.2.v_cache 0.00001954 0.01067268 + layer.3.k_cache 0.02418142 3.54708178 + layer.3.v_cache 0.00001860 0.01088349 + layer.4.k_cache 0.00062333 0.20630776 + layer.4.v_cache 0.00005467 0.02273208 + layer.4.output 9.83417319 458.67491533 + ------------------------------------------------------------------------------------- + TOTAL 4.06422232 192.55125073 + (elements=1,009,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1009664 +Total Bytes 141860 +BPFP 1.1240 bits/point +EBPFP 2.2480 equivalent bits/point +MSE 192.551251 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.006s, Pack+Encode: 0.160s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 192.5513 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,084B, BPFP=0.6178 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,400B, BPFP=2.2837 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,264B, BPFP=1.0545 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,956B, BPFP=2.1947 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,296B, BPFP=1.2612 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,604B, BPFP=2.1242 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,888B, BPFP=1.1795 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,784B, BPFP=2.1603 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,880B, BPFP=1.7788 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,408B, BPFP=2.0849 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,992B, BPFP=0.6293 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11749315 58.19779773 + layer.0.v_cache 0.00001389 0.00846894 + layer.1.k_cache 0.03399578 5.36520542 + layer.1.v_cache 0.00000555 0.00365550 + layer.2.k_cache 0.00854440 0.90659900 + layer.2.v_cache 0.00001777 0.00916272 + layer.3.k_cache 0.02740619 4.19371502 + layer.3.v_cache 0.00001935 0.01081034 + layer.4.k_cache 0.00063531 0.19841179 + layer.4.v_cache 0.00005166 0.02044910 + layer.4.output 0.18490290 688.41300366 + ------------------------------------------------------------------------------------- + TOTAL 0.08720608 287.51795889 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 105556 +BPFP 1.2438 bits/point +EBPFP 2.4877 equivalent bits/point +MSE 287.517959 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 287.5180 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,160B, BPFP=0.5253 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,968B, BPFP=2.1556 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,024B, BPFP=1.0013 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,592B, BPFP=2.0931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,404B, BPFP=1.2307 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,192B, BPFP=2.0266 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,948B, BPFP=1.1549 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,404B, BPFP=2.0618 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,652B, BPFP=1.7706 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,020B, BPFP=1.9980 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,260B, BPFP=0.5761 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11492289 57.55600690 + layer.0.v_cache 0.00001440 0.00878468 + layer.1.k_cache 0.01112092 5.25308552 + layer.1.v_cache 0.00000555 0.00351850 + layer.2.k_cache 0.01096075 0.83836308 + layer.2.v_cache 0.00001882 0.00955516 + layer.3.k_cache 0.03781782 3.66817556 + layer.3.v_cache 0.00001905 0.01014875 + layer.4.k_cache 0.00062716 0.18884070 + layer.4.v_cache 0.00006343 0.02020065 + layer.4.output 0.15224552 573.16361132 + ------------------------------------------------------------------------------------- + TOTAL 0.07301703 239.98246816 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 120624 +BPFP 1.1794 bits/point +EBPFP 2.3589 equivalent bits/point +MSE 239.982468 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.003s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 239.9825 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,040B, BPFP=0.6169 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,300B, BPFP=2.2930 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,240B, BPFP=1.0633 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,880B, BPFP=2.2078 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,388B, BPFP=1.2963 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,596B, BPFP=2.1502 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,032B, BPFP=1.2240 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,736B, BPFP=2.1786 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,972B, BPFP=1.8206 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,484B, BPFP=2.1274 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,216B, BPFP=0.6440 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10405910 59.09800502 + layer.0.v_cache 0.00001453 0.00831998 + layer.1.k_cache 0.03397077 5.38884151 + layer.1.v_cache 0.00000547 0.00350201 + layer.2.k_cache 0.00568739 0.81936953 + layer.2.v_cache 0.00001916 0.00969531 + layer.3.k_cache 0.09581309 3.84239217 + layer.3.v_cache 0.00001862 0.01017287 + layer.4.k_cache 0.00061896 0.19802383 + layer.4.v_cache 0.00005653 0.02051244 + layer.4.output 0.19123089 693.74756494 + ------------------------------------------------------------------------------------- + TOTAL 0.09287528 289.74304642 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 105884 +BPFP 1.2639 bits/point +EBPFP 2.5278 equivalent bits/point +MSE 289.743046 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 289.7430 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,232B, BPFP=0.5941 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,080B, BPFP=2.2206 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,576B, BPFP=1.0250 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,640B, BPFP=2.1397 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,684B, BPFP=1.2287 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,284B, BPFP=2.0743 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,300B, BPFP=1.1581 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,432B, BPFP=2.1015 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,540B, BPFP=1.7537 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,144B, BPFP=2.0485 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,616B, BPFP=0.6202 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09384689 54.11572840 + layer.0.v_cache 0.00001364 0.00906896 + layer.1.k_cache 0.01313462 5.21819314 + layer.1.v_cache 0.00000592 0.00427919 + layer.2.k_cache 0.00813576 0.88837406 + layer.2.v_cache 0.00001762 0.01070573 + layer.3.k_cache 0.02628646 3.27586131 + layer.3.v_cache 0.00002044 0.01192946 + layer.4.k_cache 0.00061596 0.21010661 + layer.4.v_cache 0.00005334 0.02205866 + layer.4.output 0.17343949 633.22536765 + ------------------------------------------------------------------------------------- + TOTAL 0.07977689 264.49081642 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 112528 +BPFP 1.2168 bits/point +EBPFP 2.4336 equivalent bits/point +MSE 264.490816 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.003s, Pack+Encode: 0.161s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 264.4908 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 126, 128) +Output shape: (1, 126, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.output: torch.Size([1, 126, 3584]) -> torch.Size([1, 1, 126, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,420B, BPFP=0.5481 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,744B, BPFP=1.9524 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,100B, BPFP=1.0045 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,400B, BPFP=1.9097 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,648B, BPFP=1.1964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,960B, BPFP=1.8552 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,196B, BPFP=1.1404 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,072B, BPFP=1.8690 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,020B, BPFP=1.6146 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,484B, BPFP=1.7961 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,820B, BPFP=0.5991 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13141363 53.02160838 + layer.0.v_cache 0.00001429 0.00863235 + layer.1.k_cache 0.06496696 4.56976318 + layer.1.v_cache 0.00000583 0.00409418 + layer.2.k_cache 0.01057190 0.72258528 + layer.2.v_cache 0.00002020 0.01011171 + layer.3.k_cache 0.05250616 3.13498215 + layer.3.v_cache 0.00001853 0.01093419 + layer.4.k_cache 0.00062861 0.20718767 + layer.4.v_cache 0.00004930 0.02194597 + layer.4.output 9.05470577 410.82529053 + ------------------------------------------------------------------------------------- + TOTAL 3.74371387 172.79346346 + (elements=1,096,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1096704 +Total Bytes 153864 +BPFP 1.1224 bits/point +EBPFP 2.2447 equivalent bits/point +MSE 172.793463 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 172.7935 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,276B, BPFP=0.5884 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,436B, BPFP=2.2335 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,780B, BPFP=1.0381 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,076B, BPFP=2.1688 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,016B, BPFP=1.2601 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,716B, BPFP=2.1042 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,704B, BPFP=1.2040 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,008B, BPFP=2.1566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,076B, BPFP=1.8096 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,584B, BPFP=2.0805 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,580B, BPFP=0.6050 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13059476 60.98917924 + layer.0.v_cache 0.00001380 0.00824369 + layer.1.k_cache 0.03034586 5.20685726 + layer.1.v_cache 0.00000531 0.00361035 + layer.2.k_cache 0.00374181 0.81399887 + layer.2.v_cache 0.00001886 0.00988498 + layer.3.k_cache 0.04454902 3.78420144 + layer.3.v_cache 0.00001968 0.01082549 + layer.4.k_cache 0.00060938 0.19442837 + layer.4.v_cache 0.00007316 0.02103781 + layer.4.output 0.16485390 615.80747126 + ------------------------------------------------------------------------------------- + TOTAL 0.08023229 257.74673920 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 116252 +BPFP 1.2282 bits/point +EBPFP 2.4563 equivalent bits/point +MSE 257.746739 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 257.7467 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,264B, BPFP=0.5543 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,880B, BPFP=2.1875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,924B, BPFP=1.0061 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,476B, BPFP=2.1189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,260B, BPFP=1.2330 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,136B, BPFP=2.0611 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,880B, BPFP=1.1685 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,268B, BPFP=2.0836 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,460B, BPFP=1.7765 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,100B, BPFP=2.0550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,108B, BPFP=0.6334 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15436328 58.60569166 + layer.0.v_cache 0.00001586 0.00864480 + layer.1.k_cache 0.01080767 4.71148085 + layer.1.v_cache 0.00000575 0.00365220 + layer.2.k_cache 0.00786338 0.79608685 + layer.2.v_cache 0.00001961 0.00957709 + layer.3.k_cache 0.05652578 3.40757353 + layer.3.v_cache 0.00002042 0.01083316 + layer.4.k_cache 0.00062828 0.20235934 + layer.4.v_cache 0.00005674 0.02246411 + layer.4.output 0.15192759 585.57201087 + ------------------------------------------------------------------------------------- + TOTAL 0.07610588 245.10484939 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 121756 +BPFP 1.2164 bits/point +EBPFP 2.4328 equivalent bits/point +MSE 245.104849 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 245.1048 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,100B, BPFP=0.6210 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,428B, BPFP=2.2893 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,300B, BPFP=1.0617 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,948B, BPFP=2.1931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,440B, BPFP=1.2901 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,700B, BPFP=2.1434 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,036B, BPFP=1.2091 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,844B, BPFP=2.1723 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,064B, BPFP=1.8157 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,708B, BPFP=2.1450 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,788B, BPFP=0.6235 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10254851 60.15024038 + layer.0.v_cache 0.00001399 0.00859143 + layer.1.k_cache 0.05476840 5.33576144 + layer.1.v_cache 0.00000542 0.00353726 + layer.2.k_cache 0.00244555 0.85598902 + layer.2.v_cache 0.00001844 0.00994836 + layer.3.k_cache 0.02811479 3.56580920 + layer.3.v_cache 0.00001905 0.01054224 + layer.4.k_cache 0.00060111 0.19867901 + layer.4.v_cache 0.00005313 0.02110655 + layer.4.output 0.19635826 687.28039148 + ------------------------------------------------------------------------------------- + TOTAL 0.09194683 287.12487913 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 106356 +BPFP 1.2533 bits/point +EBPFP 2.5065 equivalent bits/point +MSE 287.124879 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 287.1249 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,192B, BPFP=0.6082 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,680B, BPFP=2.2256 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,432B, BPFP=1.0351 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,384B, BPFP=2.1692 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,524B, BPFP=1.2431 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,980B, BPFP=2.0922 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,236B, BPFP=1.1883 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,128B, BPFP=2.1204 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,420B, BPFP=1.7950 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,812B, BPFP=2.0602 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,928B, BPFP=0.6241 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11987135 55.79585080 + layer.0.v_cache 0.00001390 0.00820290 + layer.1.k_cache 0.03139273 5.66819428 + layer.1.v_cache 0.00000562 0.00340640 + layer.2.k_cache 0.00816743 0.79695222 + layer.2.v_cache 0.00001857 0.00918963 + layer.3.k_cache 0.02640573 4.18656959 + layer.3.v_cache 0.00001968 0.00984258 + layer.4.k_cache 0.00062091 0.19518987 + layer.4.v_cache 0.00004972 0.02005539 + layer.4.output 0.16983549 656.78413545 + ------------------------------------------------------------------------------------- + TOTAL 0.08090671 274.36367070 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 109716 +BPFP 1.2298 bits/point +EBPFP 2.4596 equivalent bits/point +MSE 274.363671 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 274.3637 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,280B, BPFP=0.5891 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,400B, BPFP=2.2270 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,840B, BPFP=1.0489 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,068B, BPFP=2.1674 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,032B, BPFP=1.2629 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,608B, BPFP=2.0848 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,628B, BPFP=1.1904 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,808B, BPFP=2.1207 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,092B, BPFP=1.8125 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,500B, BPFP=2.0654 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,992B, BPFP=0.6156 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14089697 61.67512460 + layer.0.v_cache 0.00001393 0.00853663 + layer.1.k_cache 0.01618495 5.19641745 + layer.1.v_cache 0.00000582 0.00372246 + layer.2.k_cache 0.00661465 0.80948876 + layer.2.v_cache 0.00002083 0.00965540 + layer.3.k_cache 0.07087009 3.43716045 + layer.3.v_cache 0.00001822 0.01077292 + layer.4.k_cache 0.00061891 0.20487253 + layer.4.v_cache 0.00005311 0.02134271 + layer.4.output 0.16425455 616.70089286 + ------------------------------------------------------------------------------------- + TOTAL 0.08147525 258.13431435 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 116248 +BPFP 1.2281 bits/point +EBPFP 2.4562 equivalent bits/point +MSE 258.134314 +---------------------- -------------------------------------------------------- +Time: 0.372s Load: 0.003s, Pack+Encode: 0.165s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 258.1343 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,372B, BPFP=0.5920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,604B, BPFP=2.2128 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,892B, BPFP=1.0344 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,264B, BPFP=2.1531 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,108B, BPFP=1.2479 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,780B, BPFP=2.0681 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,836B, BPFP=1.2001 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,988B, BPFP=2.1046 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,176B, BPFP=1.7865 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,608B, BPFP=2.0379 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,392B, BPFP=0.6368 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12786581 58.38387684 + layer.0.v_cache 0.00001426 0.00905741 + layer.1.k_cache 0.03183105 4.68054405 + layer.1.v_cache 0.00000593 0.00416537 + layer.2.k_cache 0.00241907 0.73009380 + layer.2.v_cache 0.00001785 0.00943994 + layer.3.k_cache 0.02946699 3.17517879 + layer.3.v_cache 0.00001957 0.01131000 + layer.4.k_cache 0.00063747 0.19636600 + layer.4.v_cache 0.00005003 0.02159320 + layer.4.output 0.16568538 605.29730136 + ------------------------------------------------------------------------------------- + TOTAL 0.07953681 253.19427853 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 119020 +BPFP 1.2291 bits/point +EBPFP 2.4583 equivalent bits/point +MSE 253.194279 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 253.1943 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,072B, BPFP=0.6154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,540B, BPFP=2.3117 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,308B, BPFP=1.0633 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,264B, BPFP=2.2564 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,500B, BPFP=1.3021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,848B, BPFP=2.1731 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,164B, BPFP=1.2348 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,128B, BPFP=2.2292 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,132B, BPFP=1.8293 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,772B, BPFP=2.1579 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,628B, BPFP=0.6762 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12225728 61.04991111 + layer.0.v_cache 0.00001337 0.00845090 + layer.1.k_cache 0.03480693 5.11811711 + layer.1.v_cache 0.00000545 0.00362748 + layer.2.k_cache 0.00853881 0.89072047 + layer.2.v_cache 0.00001924 0.00988662 + layer.3.k_cache 0.06130744 3.87149752 + layer.3.v_cache 0.00001928 0.01111450 + layer.4.k_cache 0.00058936 0.19323721 + layer.4.v_cache 0.00007197 0.02070693 + layer.4.output 0.18657821 688.99713828 + ------------------------------------------------------------------------------------- + TOTAL 0.09021627 287.89160222 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 109356 +BPFP 1.2886 bits/point +EBPFP 2.5772 equivalent bits/point +MSE 287.891602 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 287.8916 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,060B, BPFP=0.6209 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,176B, BPFP=2.2679 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,192B, BPFP=1.0536 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,736B, BPFP=2.1786 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,224B, BPFP=1.2630 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,460B, BPFP=2.1226 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,880B, BPFP=1.1932 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,612B, BPFP=2.1534 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,948B, BPFP=1.8157 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,420B, BPFP=2.1144 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,524B, BPFP=0.6240 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582910 59.52244825 + layer.0.v_cache 0.00001386 0.00914496 + layer.1.k_cache 0.03408930 5.26673949 + layer.1.v_cache 0.00000538 0.00393231 + layer.2.k_cache 0.00228146 0.84988483 + layer.2.v_cache 0.00001859 0.01039423 + layer.3.k_cache 0.06276476 4.03366842 + layer.3.v_cache 0.00001913 0.01192210 + layer.4.k_cache 0.00060520 0.21746970 + layer.4.v_cache 0.00005125 0.02435037 + layer.4.output 0.19356344 695.19805195 + ------------------------------------------------------------------------------------- + TOTAL 0.09238954 290.37272461 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 104232 +BPFP 1.2442 bits/point +EBPFP 2.4883 equivalent bits/point +MSE 290.372725 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 290.3727 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,240B, BPFP=0.6099 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,948B, BPFP=2.2492 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,440B, BPFP=1.0241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,468B, BPFP=2.1589 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,676B, BPFP=1.2568 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,172B, BPFP=2.1032 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,332B, BPFP=1.1920 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,448B, BPFP=2.1551 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,368B, BPFP=1.7636 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,000B, BPFP=2.0708 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,664B, BPFP=0.6364 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09700135 55.53607398 + layer.0.v_cache 0.00001355 0.00865695 + layer.1.k_cache 0.01103399 4.99852486 + layer.1.v_cache 0.00000598 0.00380285 + layer.2.k_cache 0.00794396 0.88262020 + layer.2.v_cache 0.00001935 0.01074786 + layer.3.k_cache 0.04207947 3.50729113 + layer.3.v_cache 0.00002010 0.01182601 + layer.4.k_cache 0.00062246 0.20569597 + layer.4.v_cache 0.00005226 0.02318638 + layer.4.output 0.17535226 649.50505594 + ------------------------------------------------------------------------------------- + TOTAL 0.08154460 271.27787163 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 111756 +BPFP 1.2376 bits/point +EBPFP 2.4751 equivalent bits/point +MSE 271.277872 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 271.2779 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,224B, BPFP=0.5926 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,088B, BPFP=2.2221 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,676B, BPFP=1.0434 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,772B, BPFP=2.1640 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,864B, BPFP=1.2618 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,432B, BPFP=2.1015 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,528B, BPFP=1.2000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,628B, BPFP=2.1375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,684B, BPFP=1.7801 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,344B, BPFP=2.0853 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,784B, BPFP=0.6246 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11944316 53.82400620 + layer.0.v_cache 0.00001525 0.00831620 + layer.1.k_cache 0.01376067 5.02877700 + layer.1.v_cache 0.00000561 0.00353080 + layer.2.k_cache 0.00522648 0.84813161 + layer.2.v_cache 0.00001966 0.00945922 + layer.3.k_cache 0.04600189 3.59161628 + layer.3.v_cache 0.00001912 0.01075174 + layer.4.k_cache 0.00062079 0.19191607 + layer.4.v_cache 0.00005114 0.02087854 + layer.4.output 0.17175110 633.53240546 + ------------------------------------------------------------------------------------- + TOTAL 0.08161303 264.60377776 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 114024 +BPFP 1.2330 bits/point +EBPFP 2.4659 equivalent bits/point +MSE 264.603778 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 264.6038 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,104B, BPFP=0.6218 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,340B, BPFP=2.2716 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,244B, BPFP=1.0505 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,856B, BPFP=2.1747 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,332B, BPFP=1.2684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,604B, BPFP=2.1242 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,992B, BPFP=1.2003 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,844B, BPFP=2.1723 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,948B, BPFP=1.7925 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,512B, BPFP=2.1058 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,464B, BPFP=0.6429 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120220 59.82334235 + layer.0.v_cache 0.00001368 0.00851449 + layer.1.k_cache 0.03273370 5.41733336 + layer.1.v_cache 0.00000573 0.00374975 + layer.2.k_cache 0.00892877 0.90877797 + layer.2.v_cache 0.00001953 0.00930311 + layer.3.k_cache 0.02974504 4.08042399 + layer.3.v_cache 0.00001879 0.01042572 + layer.4.k_cache 0.00063874 0.20490077 + layer.4.v_cache 0.00005200 0.02141386 + layer.4.output 0.17853031 686.38885073 + ------------------------------------------------------------------------------------- + TOTAL 0.08429825 286.77706709 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 106240 +BPFP 1.2519 bits/point +EBPFP 2.5038 equivalent bits/point +MSE 286.777067 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 286.7771 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,224B, BPFP=0.6069 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,072B, BPFP=2.2726 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,636B, BPFP=1.0610 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,700B, BPFP=2.2026 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,768B, BPFP=1.2741 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,184B, BPFP=2.1054 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,512B, BPFP=1.2259 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,516B, BPFP=2.1679 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,604B, BPFP=1.8080 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,108B, BPFP=2.0911 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,540B, BPFP=0.6600 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09802409 57.21134930 + layer.0.v_cache 0.00001402 0.00869259 + layer.1.k_cache 0.03453328 5.27825376 + layer.1.v_cache 0.00000594 0.00370489 + layer.2.k_cache 0.01027090 0.86705293 + layer.2.v_cache 0.00001933 0.00932277 + layer.3.k_cache 0.02613793 3.81535247 + layer.3.v_cache 0.00001931 0.01025475 + layer.4.k_cache 0.00063818 0.19843759 + layer.4.v_cache 0.00005393 0.02082786 + layer.4.output 0.18084440 649.89371773 + ------------------------------------------------------------------------------------- + TOTAL 0.08444869 271.56936900 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 113864 +BPFP 1.2609 bits/point +EBPFP 2.5218 equivalent bits/point +MSE 271.569369 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.003s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 271.5694 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,544B, BPFP=0.5376 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,324B, BPFP=2.1729 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,588B, BPFP=0.9994 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,984B, BPFP=2.1214 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,944B, BPFP=1.2051 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,440B, BPFP=2.0388 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,496B, BPFP=1.1371 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,584B, BPFP=2.0607 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,476B, BPFP=1.7409 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,104B, BPFP=1.9879 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,668B, BPFP=0.5996 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13030120 57.60454433 + layer.0.v_cache 0.00001385 0.00873874 + layer.1.k_cache 0.02890817 5.21649348 + layer.1.v_cache 0.00000611 0.00424341 + layer.2.k_cache 0.00721959 0.84732396 + layer.2.v_cache 0.00001830 0.01014884 + layer.3.k_cache 0.02195997 3.54921508 + layer.3.v_cache 0.00001903 0.01139717 + layer.4.k_cache 0.00063766 0.20859724 + layer.4.v_cache 0.00005061 0.02191979 + layer.4.output 11.07892277 518.88462205 + ------------------------------------------------------------------------------------- + TOTAL 4.57303494 217.62793979 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 133152 +BPFP 1.1882 bits/point +EBPFP 2.3764 equivalent bits/point +MSE 217.627940 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 217.6279 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,212B, BPFP=0.6047 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,936B, BPFP=2.2470 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,432B, BPFP=1.0226 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,544B, BPFP=2.1732 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,580B, BPFP=1.2387 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,140B, BPFP=2.0971 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,244B, BPFP=1.1755 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,360B, BPFP=2.1386 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,492B, BPFP=1.7869 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,104B, BPFP=2.0904 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,732B, BPFP=0.6113 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12633393 56.62191736 + layer.0.v_cache 0.00001372 0.00909044 + layer.1.k_cache 0.01411889 4.99304677 + layer.1.v_cache 0.00000549 0.00395797 + layer.2.k_cache 0.00833277 0.81517167 + layer.2.v_cache 0.00001851 0.01035299 + layer.3.k_cache 0.05797521 4.17981067 + layer.3.v_cache 0.00001879 0.01134085 + layer.4.k_cache 0.00064672 0.22149564 + layer.4.v_cache 0.00005349 0.02415118 + layer.4.output 0.17447430 649.57051420 + ------------------------------------------------------------------------------------- + TOTAL 0.08404927 271.40493735 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 110776 +BPFP 1.2267 bits/point +EBPFP 2.4534 equivalent bits/point +MSE 271.404937 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 271.4049 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,848B, BPFP=0.6357 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,332B, BPFP=2.3062 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,836B, BPFP=1.0795 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,792B, BPFP=2.1857 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,916B, BPFP=1.3205 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,584B, BPFP=2.1393 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,576B, BPFP=1.2446 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,772B, BPFP=2.1812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,072B, BPFP=1.8018 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,616B, BPFP=2.1464 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,736B, BPFP=0.6612 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08086312 63.19727958 + layer.0.v_cache 0.00001331 0.00920332 + layer.1.k_cache 0.01551799 5.76908395 + layer.1.v_cache 0.00000513 0.00358645 + layer.2.k_cache 0.00232868 0.90612613 + layer.2.v_cache 0.00002262 0.01056543 + layer.3.k_cache 0.10623371 3.98383963 + layer.3.v_cache 0.00001836 0.01101165 + layer.4.k_cache 0.00061897 0.21897449 + layer.4.v_cache 0.00005922 0.02303315 + layer.4.output 0.20964142 770.31536990 + ------------------------------------------------------------------------------------- + TOTAL 0.09842183 321.54942900 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 97080 +BPFP 1.2747 bits/point +EBPFP 2.5494 equivalent bits/point +MSE 321.549429 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 321.5494 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,312B, BPFP=0.5625 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,820B, BPFP=2.1773 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,980B, BPFP=1.0156 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,364B, BPFP=2.0999 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,100B, BPFP=1.2058 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,840B, BPFP=2.0109 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,612B, BPFP=1.1230 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,048B, BPFP=2.0462 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,268B, BPFP=1.7439 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,728B, BPFP=1.9918 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,240B, BPFP=0.5881 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13670269 60.75572138 + layer.0.v_cache 0.00001568 0.00898365 + layer.1.k_cache 0.03139175 4.85524153 + layer.1.v_cache 0.00000602 0.00418063 + layer.2.k_cache 0.00629326 0.79890923 + layer.2.v_cache 0.00001871 0.00963460 + layer.3.k_cache 0.03837063 3.54336282 + layer.3.v_cache 0.00001860 0.01041136 + layer.4.k_cache 0.00062597 0.20348441 + layer.4.v_cache 0.00005354 0.02179434 + layer.4.output 0.15974679 585.11417896 + ------------------------------------------------------------------------------------- + TOTAL 0.07833673 245.05946922 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 118312 +BPFP 1.1820 bits/point +EBPFP 2.3640 equivalent bits/point +MSE 245.059469 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 245.0595 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,736B, BPFP=0.5456 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,752B, BPFP=2.1542 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,788B, BPFP=0.9912 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,368B, BPFP=2.0981 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,352B, BPFP=1.2196 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,040B, BPFP=2.0502 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,916B, BPFP=1.1560 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,128B, BPFP=2.0631 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,892B, BPFP=1.7366 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,552B, BPFP=1.9790 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,304B, BPFP=0.5905 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12281527 54.74537730 + layer.0.v_cache 0.00001417 0.00864170 + layer.1.k_cache 0.04442539 4.99229588 + layer.1.v_cache 0.00000572 0.00383624 + layer.2.k_cache 0.01033308 0.75052308 + layer.2.v_cache 0.00001923 0.00996797 + layer.3.k_cache 0.04555132 3.49177851 + layer.3.v_cache 0.00001841 0.01062090 + layer.4.k_cache 0.00060580 0.19399942 + layer.4.v_cache 0.00007068 0.02079389 + layer.4.output 10.66341841 499.08761682 + ------------------------------------------------------------------------------------- + TOTAL 4.40398753 209.28477369 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 137828 +BPFP 1.1839 bits/point +EBPFP 2.3679 equivalent bits/point +MSE 209.284774 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 209.2848 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,272B, BPFP=0.5945 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,276B, BPFP=2.2304 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,764B, BPFP=1.0472 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,036B, BPFP=2.1868 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,956B, BPFP=1.2638 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,584B, BPFP=2.1047 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,540B, BPFP=1.1882 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,716B, BPFP=2.1286 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,824B, BPFP=1.7849 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,328B, BPFP=2.0581 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,872B, BPFP=0.5936 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11362342 56.28148846 + layer.0.v_cache 0.00001418 0.00874024 + layer.1.k_cache 0.05429294 5.21454745 + layer.1.v_cache 0.00000580 0.00398504 + layer.2.k_cache 0.00513919 0.81620434 + layer.2.v_cache 0.00001816 0.00970153 + layer.3.k_cache 0.05885953 3.75870212 + layer.3.v_cache 0.00001841 0.01034086 + layer.4.k_cache 0.00061763 0.20019642 + layer.4.v_cache 0.00005080 0.02043292 + layer.4.output 0.16727475 623.37458472 + ------------------------------------------------------------------------------------- + TOTAL 0.08256255 260.58508426 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 114168 +BPFP 1.2202 bits/point +EBPFP 2.4403 equivalent bits/point +MSE 260.585084 +---------------------- -------------------------------------------------------- +Time: 0.367s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 260.5851 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,288B, BPFP=0.5524 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,004B, BPFP=2.1848 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,140B, BPFP=1.0316 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,560B, BPFP=2.1102 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,432B, BPFP=1.2487 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,204B, BPFP=2.0504 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,024B, BPFP=1.1801 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,444B, BPFP=2.0907 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,612B, BPFP=1.7829 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,132B, BPFP=2.0383 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,560B, BPFP=0.5895 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702597 59.55580057 + layer.0.v_cache 0.00001402 0.00881870 + layer.1.k_cache 0.01332537 5.06349691 + layer.1.v_cache 0.00000581 0.00379477 + layer.2.k_cache 0.00507878 0.83803206 + layer.2.v_cache 0.00001944 0.00928094 + layer.3.k_cache 0.04051446 3.08470400 + layer.3.v_cache 0.00001931 0.00961291 + layer.4.k_cache 0.00062870 0.19267189 + layer.4.v_cache 0.00006514 0.02008169 + layer.4.output 0.16549673 577.48708717 + ------------------------------------------------------------------------------------- + TOTAL 0.07795142 241.83505321 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 121400 +BPFP 1.1998 bits/point +EBPFP 2.3996 equivalent bits/point +MSE 241.835053 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 241.8351 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,248B, BPFP=0.5971 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,052B, BPFP=2.2154 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,488B, BPFP=1.0088 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,552B, BPFP=2.1235 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,580B, BPFP=1.2096 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,216B, BPFP=2.0618 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,288B, BPFP=1.1559 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,408B, BPFP=2.0971 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,472B, BPFP=1.7412 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,972B, BPFP=2.0169 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,156B, BPFP=0.6081 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13925676 54.57804458 + layer.0.v_cache 0.00001387 0.00884386 + layer.1.k_cache 0.03290265 5.00151044 + layer.1.v_cache 0.00000589 0.00372646 + layer.2.k_cache 0.00800494 0.80503666 + layer.2.v_cache 0.00001921 0.00954611 + layer.3.k_cache 0.08301921 3.60469289 + layer.3.v_cache 0.00001931 0.01016945 + layer.4.k_cache 0.00065893 0.19715267 + layer.4.v_cache 0.00004891 0.02058756 + layer.4.output 0.17367665 631.04201681 + ------------------------------------------------------------------------------------- + TOTAL 0.08704037 263.61961343 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 111432 +BPFP 1.2049 bits/point +EBPFP 2.4099 equivalent bits/point +MSE 263.619613 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 263.6196 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,580B, BPFP=0.5431 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,308B, BPFP=2.1705 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,524B, BPFP=0.9897 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,788B, BPFP=2.0916 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,908B, BPFP=1.1996 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,452B, BPFP=2.0407 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,396B, BPFP=1.1220 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,644B, BPFP=2.0698 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,480B, BPFP=1.7415 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,096B, BPFP=1.9867 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,844B, BPFP=0.5601 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.209s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10583916 58.87813827 + layer.0.v_cache 0.00001406 0.00882298 + layer.1.k_cache 0.01351924 5.30631471 + layer.1.v_cache 0.00000536 0.00389833 + layer.2.k_cache 0.00567923 0.74908632 + layer.2.v_cache 0.00001856 0.01013039 + layer.3.k_cache 0.05293757 3.35390080 + layer.3.v_cache 0.00001841 0.01085183 + layer.4.k_cache 0.00064880 0.20536219 + layer.4.v_cache 0.00004836 0.02151784 + layer.4.output 11.07886250 518.03844487 + ------------------------------------------------------------------------------------- + TOTAL 4.57239802 217.34218457 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 131020 +BPFP 1.1692 bits/point +EBPFP 2.3383 equivalent bits/point +MSE 217.342185 +---------------------- -------------------------------------------------------- +Time: 0.375s Load: 0.005s, Pack+Encode: 0.162s, Decode+Unpack: 0.209s +---------------------- -------------------------------------------------------- +💾 Converting with 217.3422 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,232B, BPFP=0.6084 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,004B, BPFP=2.2598 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,504B, BPFP=1.0361 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,628B, BPFP=2.1890 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,644B, BPFP=1.2508 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,356B, BPFP=2.1378 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,240B, BPFP=1.1747 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,360B, BPFP=2.1386 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,452B, BPFP=1.7794 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,048B, BPFP=2.0798 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,012B, BPFP=0.6189 +⌛️ [2/4] FRONTEND: Frontend time: 0.176s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860243 55.99185217 + layer.0.v_cache 0.00001403 0.00892450 + layer.1.k_cache 0.01400426 5.41528835 + layer.1.v_cache 0.00000584 0.00402740 + layer.2.k_cache 0.01277778 0.85023958 + layer.2.v_cache 0.00001980 0.01053278 + layer.3.k_cache 0.07376247 3.72261066 + layer.3.v_cache 0.00001877 0.01071669 + layer.4.k_cache 0.00062152 0.20113796 + layer.4.v_cache 0.00005140 0.02181762 + layer.4.output 0.18179212 649.55088210 + ------------------------------------------------------------------------------------- + TOTAL 0.08837783 271.35843073 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 111480 +BPFP 1.2345 bits/point +EBPFP 2.4690 equivalent bits/point +MSE 271.358431 +---------------------- -------------------------------------------------------- +Time: 0.400s Load: 0.004s, Pack+Encode: 0.176s, Decode+Unpack: 0.219s +---------------------- -------------------------------------------------------- +💾 Converting with 271.3584 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,368B, BPFP=0.5847 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,688B, BPFP=2.2028 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,916B, BPFP=1.0271 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,352B, BPFP=2.1444 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,264B, BPFP=1.2611 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,936B, BPFP=2.0722 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,820B, BPFP=1.1840 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,272B, BPFP=2.1306 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,344B, BPFP=1.7958 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,856B, BPFP=2.0583 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,572B, BPFP=0.6342 +⌛️ [2/4] FRONTEND: Frontend time: 0.172s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13781073 60.78524848 + layer.0.v_cache 0.00001414 0.00878265 + layer.1.k_cache 0.01252146 4.74712084 + layer.1.v_cache 0.00000561 0.00373593 + layer.2.k_cache 0.00238173 0.74425074 + layer.2.v_cache 0.00001849 0.00992106 + layer.3.k_cache 0.04360520 3.51222670 + layer.3.v_cache 0.00001973 0.01060690 + layer.4.k_cache 0.00061668 0.18772299 + layer.4.v_cache 0.00005448 0.02172038 + layer.4.output 0.16772948 597.12346230 + ------------------------------------------------------------------------------------- + TOTAL 0.08065615 249.99385722 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 120388 +BPFP 1.2295 bits/point +EBPFP 2.4589 equivalent bits/point +MSE 249.993857 +---------------------- -------------------------------------------------------- +Time: 0.386s Load: 0.004s, Pack+Encode: 0.172s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 249.9939 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,128B, BPFP=0.6187 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,568B, BPFP=2.2880 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,232B, BPFP=1.0348 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,052B, BPFP=2.1859 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,416B, BPFP=1.2690 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,772B, BPFP=2.1305 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,048B, BPFP=1.1962 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,968B, BPFP=2.1693 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,184B, BPFP=1.8165 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,592B, BPFP=2.0949 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,360B, BPFP=0.6318 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07550411 58.84862663 + layer.0.v_cache 0.00001381 0.00869208 + layer.1.k_cache 0.01271453 5.36412879 + layer.1.v_cache 0.00000531 0.00373430 + layer.2.k_cache 0.00681068 0.86086476 + layer.2.v_cache 0.00001904 0.01055377 + layer.3.k_cache 0.02844424 3.77049004 + layer.3.v_cache 0.00001798 0.01074488 + layer.4.k_cache 0.00061127 0.21405459 + layer.4.v_cache 0.00009323 0.02286906 + layer.4.output 0.18255964 679.90557188 + ------------------------------------------------------------------------------------- + TOTAL 0.08247951 284.02669189 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 107320 +BPFP 1.2486 bits/point +EBPFP 2.4972 equivalent bits/point +MSE 284.026692 +---------------------- -------------------------------------------------------- +Time: 0.376s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 284.0267 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,168B, BPFP=0.6188 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,788B, BPFP=2.3023 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,428B, BPFP=1.0602 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,392B, BPFP=2.2250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,536B, BPFP=1.2766 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,052B, BPFP=2.1586 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,196B, BPFP=1.2102 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,288B, BPFP=2.2047 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,320B, BPFP=1.8203 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,804B, BPFP=2.1102 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,312B, BPFP=0.6504 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08590809 58.36546631 + layer.0.v_cache 0.00001388 0.00844289 + layer.1.k_cache 0.01594085 5.51166992 + layer.1.v_cache 0.00000552 0.00348309 + layer.2.k_cache 0.00791870 0.82831984 + layer.2.v_cache 0.00001801 0.00905717 + layer.3.k_cache 0.02757127 4.08710060 + layer.3.v_cache 0.00001862 0.01041112 + layer.4.k_cache 0.00063773 0.19660797 + layer.4.v_cache 0.00005492 0.02162522 + layer.4.output 0.18408036 670.84988839 + ------------------------------------------------------------------------------------- + TOTAL 0.08392059 280.29361193 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 110284 +BPFP 1.2670 bits/point +EBPFP 2.5341 equivalent bits/point +MSE 280.293612 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.003s, Pack+Encode: 0.164s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 280.2936 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,044B, BPFP=0.6258 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,092B, BPFP=2.2804 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,124B, BPFP=1.0535 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,732B, BPFP=2.2064 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,192B, BPFP=1.2730 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,380B, BPFP=2.1340 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,732B, BPFP=1.1785 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,608B, BPFP=2.1809 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,588B, BPFP=1.7656 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,120B, BPFP=2.0806 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,536B, BPFP=0.6325 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08991440 57.67518375 + layer.0.v_cache 0.00001347 0.00844009 + layer.1.k_cache 0.03345199 5.48170230 + layer.1.v_cache 0.00000518 0.00369524 + layer.2.k_cache 0.00398020 0.84197777 + layer.2.v_cache 0.00001815 0.00937459 + layer.3.k_cache 0.04607732 3.99085035 + layer.3.v_cache 0.00001896 0.00998523 + layer.4.k_cache 0.00062119 0.19689370 + layer.4.v_cache 0.00005040 0.02053837 + layer.4.output 0.18612516 710.03389333 + ------------------------------------------------------------------------------------- + TOTAL 0.08688396 296.38093498 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 103148 +BPFP 1.2474 bits/point +EBPFP 2.4949 equivalent bits/point +MSE 296.380935 +---------------------- -------------------------------------------------------- +Time: 0.372s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 296.3809 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,356B, BPFP=0.5762 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,788B, BPFP=2.1957 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,944B, BPFP=1.0206 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,396B, BPFP=2.1284 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,276B, BPFP=1.2493 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,028B, BPFP=2.0652 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,884B, BPFP=1.1820 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,256B, BPFP=2.1044 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,388B, BPFP=1.7837 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,880B, BPFP=2.0398 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,300B, BPFP=0.6206 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11186004 59.68337912 + layer.0.v_cache 0.00001364 0.00872042 + layer.1.k_cache 0.01386243 4.50669307 + layer.1.v_cache 0.00000608 0.00374984 + layer.2.k_cache 0.00506424 0.87950939 + layer.2.v_cache 0.00001872 0.00999970 + layer.3.k_cache 0.02496567 3.63676503 + layer.3.v_cache 0.00001876 0.01099639 + layer.4.k_cache 0.00064518 0.19663572 + layer.4.v_cache 0.00005436 0.02151992 + layer.4.output 0.16682192 593.03718603 + ------------------------------------------------------------------------------------- + TOTAL 0.07789780 248.24813358 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 120496 +BPFP 1.2170 bits/point +EBPFP 2.4341 equivalent bits/point +MSE 248.248134 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.003s, Pack+Encode: 0.161s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 248.2481 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,228B, BPFP=0.6077 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,888B, BPFP=2.2380 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,428B, BPFP=1.0218 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,376B, BPFP=2.1416 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,520B, BPFP=1.2274 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,108B, BPFP=2.0911 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,244B, BPFP=1.1755 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,224B, BPFP=2.1130 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,420B, BPFP=1.7733 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,944B, BPFP=2.0602 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,552B, BPFP=0.6334 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08542032 56.98739293 + layer.0.v_cache 0.00001402 0.00855354 + layer.1.k_cache 0.01083013 5.46320021 + layer.1.v_cache 0.00000556 0.00374945 + layer.2.k_cache 0.00382075 0.78060472 + layer.2.v_cache 0.00001860 0.01013368 + layer.3.k_cache 0.02610962 3.64070037 + layer.3.v_cache 0.00002051 0.01153080 + layer.4.k_cache 0.00063312 0.20593583 + layer.4.v_cache 0.00005614 0.02364200 + layer.4.output 0.16811641 649.40716437 + ------------------------------------------------------------------------------------- + TOTAL 0.07669080 271.35209377 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 110932 +BPFP 1.2284 bits/point +EBPFP 2.4569 equivalent bits/point +MSE 271.352094 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 271.3521 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,172B, BPFP=0.6119 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,764B, BPFP=2.2693 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,420B, BPFP=1.0455 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,284B, BPFP=2.1767 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,620B, BPFP=1.2770 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,968B, BPFP=2.1157 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,148B, BPFP=1.1860 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,264B, BPFP=2.1728 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,412B, BPFP=1.8156 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,860B, BPFP=2.0949 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,952B, BPFP=0.6325 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10540552 57.88337312 + layer.0.v_cache 0.00001541 0.00891264 + layer.1.k_cache 0.03402772 5.41202498 + layer.1.v_cache 0.00000555 0.00404664 + layer.2.k_cache 0.00394259 0.83961393 + layer.2.v_cache 0.00001868 0.01038348 + layer.3.k_cache 0.02873192 3.78749744 + layer.3.v_cache 0.00001983 0.01100619 + layer.4.k_cache 0.00063741 0.21476096 + layer.4.v_cache 0.00005806 0.02409682 + layer.4.output 0.17961645 665.50529101 + ------------------------------------------------------------------------------------- + TOTAL 0.08412811 278.04310313 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 109864 +BPFP 1.2466 bits/point +EBPFP 2.4933 equivalent bits/point +MSE 278.043103 +---------------------- -------------------------------------------------------- +Time: 0.367s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 278.0431 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,148B, BPFP=0.6148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,768B, BPFP=2.2984 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,400B, BPFP=1.0547 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,424B, BPFP=2.2313 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,620B, BPFP=1.2930 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,044B, BPFP=2.1570 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,184B, BPFP=1.2078 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,284B, BPFP=2.2039 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,444B, BPFP=1.8445 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,928B, BPFP=2.1344 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,876B, BPFP=0.6662 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10022384 59.02933960 + layer.0.v_cache 0.00001350 0.00878409 + layer.1.k_cache 0.03445883 5.65025291 + layer.1.v_cache 0.00000556 0.00385998 + layer.2.k_cache 0.00550650 0.85441742 + layer.2.v_cache 0.00001848 0.00978352 + layer.3.k_cache 0.03016714 3.93536720 + layer.3.v_cache 0.00001937 0.01121766 + layer.4.k_cache 0.00063435 0.19955907 + layer.4.v_cache 0.00006510 0.02141205 + layer.4.output 0.18593751 669.76462054 + ------------------------------------------------------------------------------------- + TOTAL 0.08662796 279.88684337 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 111120 +BPFP 1.2767 bits/point +EBPFP 2.5533 equivalent bits/point +MSE 279.886843 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 279.8868 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,124B, BPFP=0.5032 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,116B, BPFP=2.1128 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,744B, BPFP=0.9253 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,652B, BPFP=2.0380 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,332B, BPFP=1.1811 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,204B, BPFP=1.9659 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,812B, BPFP=1.0973 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,480B, BPFP=2.0103 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,492B, BPFP=1.6901 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,036B, BPFP=1.9388 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,484B, BPFP=0.5404 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11898830 58.50596005 + layer.0.v_cache 0.00001403 0.00871104 + layer.1.k_cache 0.01243949 5.13919980 + layer.1.v_cache 0.00000551 0.00365223 + layer.2.k_cache 0.00346032 0.80110939 + layer.2.v_cache 0.00001901 0.01013430 + layer.3.k_cache 0.02403709 3.92203986 + layer.3.v_cache 0.00001994 0.01079122 + layer.4.k_cache 0.00063949 0.19903989 + layer.4.v_cache 0.00006591 0.02198737 + layer.4.output 0.03773703 570.75883652 + ------------------------------------------------------------------------------------- + TOTAL 0.02493225 239.05496946 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 119476 +BPFP 1.1321 bits/point +EBPFP 2.2642 equivalent bits/point +MSE 239.054969 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 239.0550 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,180B, BPFP=0.6059 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,796B, BPFP=2.2477 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,360B, BPFP=1.0213 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,252B, BPFP=2.1441 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,568B, BPFP=1.2515 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,888B, BPFP=2.0747 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,152B, BPFP=1.1723 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,056B, BPFP=2.1067 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,276B, BPFP=1.7675 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,672B, BPFP=2.0335 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,852B, BPFP=0.6221 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10693085 58.11830697 + layer.0.v_cache 0.00001341 0.00889876 + layer.1.k_cache 0.01172146 5.59758889 + layer.1.v_cache 0.00000553 0.00403229 + layer.2.k_cache 0.00236719 0.79173893 + layer.2.v_cache 0.00001835 0.01037748 + layer.3.k_cache 0.02622769 3.75786609 + layer.3.v_cache 0.00001864 0.01083367 + layer.4.k_cache 0.00061713 0.20669130 + layer.4.v_cache 0.00004943 0.02243987 + layer.4.output 0.18273007 656.56990418 + ------------------------------------------------------------------------------------- + TOTAL 0.08394589 274.38341785 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 109052 +BPFP 1.2223 bits/point +EBPFP 2.4447 equivalent bits/point +MSE 274.383418 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.003s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 274.3834 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,408B, BPFP=0.5272 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,916B, BPFP=2.1528 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,220B, BPFP=0.9623 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,488B, BPFP=2.0866 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,620B, BPFP=1.1788 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,040B, BPFP=2.0173 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,052B, BPFP=1.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,192B, BPFP=2.0408 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,124B, BPFP=1.7209 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,860B, BPFP=1.9895 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,332B, BPFP=0.5598 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13156438 56.45986908 + layer.0.v_cache 0.00001398 0.00903263 + layer.1.k_cache 0.04507210 4.95087667 + layer.1.v_cache 0.00000603 0.00404630 + layer.2.k_cache 0.00848899 0.85166667 + layer.2.v_cache 0.00001983 0.01084037 + layer.3.k_cache 0.06093804 3.79747145 + layer.3.v_cache 0.00001891 0.01158683 + layer.4.k_cache 0.00062755 0.20898343 + layer.4.v_cache 0.00004781 0.02190159 + layer.4.output 11.29576791 528.83442362 + ------------------------------------------------------------------------------------- + TOTAL 4.66571606 221.65689649 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 127252 +BPFP 1.1580 bits/point +EBPFP 2.3160 equivalent bits/point +MSE 221.656896 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 221.6569 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,244B, BPFP=0.6107 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,964B, BPFP=2.2523 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,500B, BPFP=1.0354 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,608B, BPFP=2.1852 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,812B, BPFP=1.2824 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,292B, BPFP=2.1258 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,424B, BPFP=1.2093 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,520B, BPFP=2.1687 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,480B, BPFP=1.7846 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,128B, BPFP=2.0949 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,852B, BPFP=0.6415 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12340720 55.85222139 + layer.0.v_cache 0.00001379 0.00868103 + layer.1.k_cache 0.01365945 4.62171403 + layer.1.v_cache 0.00000586 0.00385489 + layer.2.k_cache 0.00515996 0.84473934 + layer.2.v_cache 0.00002118 0.00991773 + layer.3.k_cache 0.04682968 4.04283418 + layer.3.v_cache 0.00001899 0.01035852 + layer.4.k_cache 0.00062596 0.20639888 + layer.4.v_cache 0.00004975 0.02268808 + layer.4.output 0.18155291 649.41813683 + ------------------------------------------------------------------------------------- + TOTAL 0.08592130 271.26766858 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 112824 +BPFP 1.2494 bits/point +EBPFP 2.4988 equivalent bits/point +MSE 271.267669 +---------------------- -------------------------------------------------------- +Time: 0.367s Load: 0.005s, Pack+Encode: 0.159s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 271.2677 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,264B, BPFP=0.6071 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,100B, BPFP=2.2507 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,520B, BPFP=1.0268 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,736B, BPFP=2.1830 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,604B, BPFP=1.2284 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,260B, BPFP=2.0945 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,376B, BPFP=1.1860 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,484B, BPFP=2.1362 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,472B, BPFP=1.7619 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,160B, BPFP=2.0759 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,780B, BPFP=0.6585 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10400932 54.65159970 + layer.0.v_cache 0.00001381 0.00838364 + layer.1.k_cache 0.01300498 5.36908613 + layer.1.v_cache 0.00000559 0.00359838 + layer.2.k_cache 0.00535372 0.81231408 + layer.2.v_cache 0.00001862 0.00950000 + layer.3.k_cache 0.04415199 3.59101432 + layer.3.v_cache 0.00001841 0.01059668 + layer.4.k_cache 0.00060249 0.19732587 + layer.4.v_cache 0.00005274 0.02199164 + layer.4.output 0.16694923 641.82366071 + ------------------------------------------------------------------------------------- + TOTAL 0.07858096 268.08476679 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 113756 +BPFP 1.2447 bits/point +EBPFP 2.4894 equivalent bits/point +MSE 268.084767 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 268.0848 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,136B, BPFP=0.6125 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,808B, BPFP=2.3062 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,316B, BPFP=1.0383 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,404B, BPFP=2.2273 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,564B, BPFP=1.2820 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,048B, BPFP=2.1578 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,144B, BPFP=1.2000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,280B, BPFP=2.2031 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,348B, BPFP=1.8258 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,916B, BPFP=2.1320 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,084B, BPFP=0.6162 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09321334 58.48198242 + layer.0.v_cache 0.00001862 0.00922667 + layer.1.k_cache 0.03257937 5.39741974 + layer.1.v_cache 0.00000532 0.00372734 + layer.2.k_cache 0.01442847 0.82858124 + layer.2.v_cache 0.00001865 0.01023608 + layer.3.k_cache 0.04385926 3.67368088 + layer.3.v_cache 0.00001828 0.01111403 + layer.4.k_cache 0.00061064 0.20630510 + layer.4.v_cache 0.00005353 0.02316311 + layer.4.output 0.18500509 674.13018973 + ------------------------------------------------------------------------------------- + TOTAL 0.08704948 281.62098616 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 109048 +BPFP 1.2528 bits/point +EBPFP 2.5057 equivalent bits/point +MSE 281.620986 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 281.6210 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,616B, BPFP=0.5485 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,376B, BPFP=2.1808 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,560B, BPFP=0.9951 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,972B, BPFP=2.1195 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,056B, BPFP=1.2221 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,516B, BPFP=2.0504 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,688B, BPFP=1.1663 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,700B, BPFP=2.0783 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,644B, BPFP=1.7664 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,308B, BPFP=2.0188 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,140B, BPFP=0.5882 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11527186 56.72207316 + layer.0.v_cache 0.00001510 0.00887196 + layer.1.k_cache 0.02920971 4.87194943 + layer.1.v_cache 0.00000551 0.00388689 + layer.2.k_cache 0.00857845 0.73260994 + layer.2.v_cache 0.00001837 0.01017558 + layer.3.k_cache 0.03550095 3.39819780 + layer.3.v_cache 0.00001950 0.01091686 + layer.4.k_cache 0.00063242 0.20041727 + layer.4.v_cache 0.00005150 0.02227641 + layer.4.output 11.08123415 518.85142164 + ------------------------------------------------------------------------------------- + TOTAL 4.57399661 217.52596040 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 133576 +BPFP 1.1920 bits/point +EBPFP 2.3839 equivalent bits/point +MSE 217.525960 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 217.5260 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,308B, BPFP=0.5618 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,820B, BPFP=2.1773 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,016B, BPFP=1.0217 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,480B, BPFP=2.1196 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,268B, BPFP=1.2344 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,084B, BPFP=2.0523 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,860B, BPFP=1.1651 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,316B, BPFP=2.0917 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,608B, BPFP=1.8016 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,992B, BPFP=2.0367 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,148B, BPFP=0.5859 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09673090 61.07108738 + layer.0.v_cache 0.00001344 0.00832057 + layer.1.k_cache 0.04918879 5.19748422 + layer.1.v_cache 0.00000572 0.00393416 + layer.2.k_cache 0.00880511 0.71844217 + layer.2.v_cache 0.00001944 0.00958997 + layer.3.k_cache 0.06586690 3.54057113 + layer.3.v_cache 0.00001823 0.01051899 + layer.4.k_cache 0.00062484 0.20390121 + layer.4.v_cache 0.00006325 0.02102003 + layer.4.output 0.15294319 585.68890722 + ------------------------------------------------------------------------------------- + TOTAL 0.07599641 245.32983649 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 119900 +BPFP 1.1979 bits/point +EBPFP 2.3957 equivalent bits/point +MSE 245.329836 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 245.3298 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,300B, BPFP=0.5605 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,920B, BPFP=2.1943 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,052B, BPFP=1.0279 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,624B, BPFP=2.1440 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,404B, BPFP=1.2575 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,216B, BPFP=2.0747 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,888B, BPFP=1.1698 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,376B, BPFP=2.1019 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,424B, BPFP=1.7704 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,932B, BPFP=2.0265 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,940B, BPFP=0.6051 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12676482 61.54594090 + layer.0.v_cache 0.00001339 0.00865978 + layer.1.k_cache 0.03133728 5.20227250 + layer.1.v_cache 0.00000590 0.00388044 + layer.2.k_cache 0.00372887 0.76054482 + layer.2.v_cache 0.00001820 0.00969955 + layer.3.k_cache 0.02418065 3.44832080 + layer.3.v_cache 0.00002034 0.01072244 + layer.4.k_cache 0.00062996 0.20115740 + layer.4.v_cache 0.00006053 0.02113786 + layer.4.output 0.15389095 584.63465644 + ------------------------------------------------------------------------------------- + TOTAL 0.07435274 244.92087833 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 121076 +BPFP 1.2096 bits/point +EBPFP 2.4192 equivalent bits/point +MSE 244.920878 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 244.9209 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,360B, BPFP=0.5833 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,652B, BPFP=2.1965 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,844B, BPFP=1.0146 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,252B, BPFP=2.1271 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,112B, BPFP=1.2347 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,744B, BPFP=2.0389 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,692B, BPFP=1.1618 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,880B, BPFP=2.0625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,140B, BPFP=1.7604 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,656B, BPFP=2.0236 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,628B, BPFP=0.6108 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12863920 60.56795790 + layer.0.v_cache 0.00001425 0.00853188 + layer.1.k_cache 0.01389826 4.69260763 + layer.1.v_cache 0.00000571 0.00377700 + layer.2.k_cache 0.00659207 0.76124963 + layer.2.v_cache 0.00001883 0.00954515 + layer.3.k_cache 0.05540012 3.79337836 + layer.3.v_cache 0.00001922 0.00990673 + layer.4.k_cache 0.00063279 0.19089764 + layer.4.v_cache 0.00005416 0.01991635 + layer.4.output 0.16196846 597.24985119 + ------------------------------------------------------------------------------------- + TOTAL 0.07876787 250.04745451 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 117960 +BPFP 1.2047 bits/point +EBPFP 2.4093 equivalent bits/point +MSE 250.047455 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 250.0475 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,724B, BPFP=0.5388 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,744B, BPFP=2.1331 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,716B, BPFP=0.9716 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,376B, BPFP=2.0799 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,276B, BPFP=1.1973 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,956B, BPFP=2.0191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,824B, BPFP=1.1319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,192B, BPFP=2.0532 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,976B, BPFP=1.7326 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,684B, BPFP=1.9797 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,188B, BPFP=0.5826 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13160179 55.59020996 + layer.0.v_cache 0.00001577 0.00870140 + layer.1.k_cache 0.05775581 5.26570299 + layer.1.v_cache 0.00000549 0.00379186 + layer.2.k_cache 0.01027601 0.76948251 + layer.2.v_cache 0.00002197 0.01020080 + layer.3.k_cache 0.07329086 3.20874899 + layer.3.v_cache 0.00001983 0.01129116 + layer.4.k_cache 0.00064477 0.20106918 + layer.4.v_cache 0.00005260 0.02150931 + layer.4.output 10.56448284 494.92596726 + ------------------------------------------------------------------------------------- + TOTAL 4.36618028 207.62191053 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 137656 +BPFP 1.1715 bits/point +EBPFP 2.3430 equivalent bits/point +MSE 207.621911 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 207.6219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,220B, BPFP=0.6062 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,848B, BPFP=2.2304 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,440B, BPFP=1.0241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,468B, BPFP=2.1589 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,668B, BPFP=1.2553 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,144B, BPFP=2.0979 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,276B, BPFP=1.1815 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,352B, BPFP=2.1370 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,464B, BPFP=1.7816 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,968B, BPFP=2.0648 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,872B, BPFP=0.6151 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10139442 55.95387213 + layer.0.v_cache 0.00001346 0.00843063 + layer.1.k_cache 0.03266963 5.11837548 + layer.1.v_cache 0.00000526 0.00354213 + layer.2.k_cache 0.01283502 0.81314684 + layer.2.v_cache 0.00001884 0.00981945 + layer.3.k_cache 0.02666297 3.75590019 + layer.3.v_cache 0.00001780 0.00975037 + layer.4.k_cache 0.00061512 0.18896448 + layer.4.v_cache 0.00004947 0.02058430 + layer.4.output 0.18393325 650.39420181 + ------------------------------------------------------------------------------------- + TOTAL 0.08598910 271.68481169 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 110720 +BPFP 1.2261 bits/point +EBPFP 2.4522 equivalent bits/point +MSE 271.684812 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 271.6848 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,436B, BPFP=0.5316 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,132B, BPFP=2.1863 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,376B, BPFP=0.9864 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,664B, BPFP=2.1139 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,880B, BPFP=1.2191 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,308B, BPFP=2.0588 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,440B, BPFP=1.1510 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,540B, BPFP=2.0947 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,360B, BPFP=1.7574 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,120B, BPFP=2.0297 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,980B, BPFP=0.5963 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11149573 58.18677483 + layer.0.v_cache 0.00001357 0.00852846 + layer.1.k_cache 0.08092193 4.86455513 + layer.1.v_cache 0.00000571 0.00366761 + layer.2.k_cache 0.01109533 0.88551376 + layer.2.v_cache 0.00001920 0.00976140 + layer.3.k_cache 0.03617269 3.54638007 + layer.3.v_cache 0.00001858 0.01011655 + layer.4.k_cache 0.00062559 0.19496597 + layer.4.v_cache 0.00005684 0.02045540 + layer.4.output 11.30219499 529.18250530 + ------------------------------------------------------------------------------------- + TOTAL 4.66798765 221.88283861 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 131236 +BPFP 1.1943 bits/point +EBPFP 2.3885 equivalent bits/point +MSE 221.882839 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 221.8828 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.6170 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,540B, BPFP=2.3117 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,380B, BPFP=1.0777 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,312B, BPFP=2.2660 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,504B, BPFP=1.3029 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,816B, BPFP=2.1667 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,064B, BPFP=1.2147 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,944B, BPFP=2.1923 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,272B, BPFP=1.8574 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,572B, BPFP=2.1178 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,068B, BPFP=0.6315 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10167113 60.42119892 + layer.0.v_cache 0.00001386 0.00870916 + layer.1.k_cache 0.03545215 5.46288671 + layer.1.v_cache 0.00000569 0.00375114 + layer.2.k_cache 0.01047806 0.87750939 + layer.2.v_cache 0.00001755 0.00927880 + layer.3.k_cache 0.06687863 3.43355697 + layer.3.v_cache 0.00001960 0.01072509 + layer.4.k_cache 0.00059181 0.19170958 + layer.4.v_cache 0.00004882 0.02012242 + layer.4.output 0.19633365 687.56284341 + ------------------------------------------------------------------------------------- + TOTAL 0.09350076 287.25760894 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 107552 +BPFP 1.2673 bits/point +EBPFP 2.5347 equivalent bits/point +MSE 287.257609 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 287.2576 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,284B, BPFP=0.5898 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,376B, BPFP=2.2227 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,736B, BPFP=1.0302 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,968B, BPFP=2.1494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,912B, BPFP=1.2414 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,652B, BPFP=2.0927 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,572B, BPFP=1.1803 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,712B, BPFP=2.1034 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,816B, BPFP=1.7629 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,396B, BPFP=2.0467 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,832B, BPFP=0.6371 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11445767 63.20649246 + layer.0.v_cache 0.00001583 0.00857444 + layer.1.k_cache 0.03092440 5.21917234 + layer.1.v_cache 0.00000551 0.00372823 + layer.2.k_cache 0.00639992 0.81912582 + layer.2.v_cache 0.00001897 0.01007191 + layer.3.k_cache 0.05613507 3.32188468 + layer.3.v_cache 0.00002026 0.01098196 + layer.4.k_cache 0.00063356 0.20315012 + layer.4.v_cache 0.00005123 0.02228667 + layer.4.output 0.16055210 616.68457512 + ------------------------------------------------------------------------------------- + TOTAL 0.07838395 258.21279379 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 116256 +BPFP 1.2282 bits/point +EBPFP 2.4564 equivalent bits/point +MSE 258.212794 +---------------------- -------------------------------------------------------- +Time: 0.382s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.217s +---------------------- -------------------------------------------------------- +💾 Converting with 258.2128 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,208B, BPFP=0.5967 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,084B, BPFP=2.2478 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,540B, BPFP=1.0305 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,684B, BPFP=2.1734 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,696B, BPFP=1.2455 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,264B, BPFP=2.0952 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,324B, BPFP=1.1763 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,368B, BPFP=2.1146 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,548B, BPFP=1.7760 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,032B, BPFP=2.0521 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,784B, BPFP=0.6054 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13956541 55.25978306 + layer.0.v_cache 0.00001349 0.00898472 + layer.1.k_cache 0.03336124 4.94042133 + layer.1.v_cache 0.00000558 0.00410783 + layer.2.k_cache 0.00698305 0.86613710 + layer.2.v_cache 0.00001831 0.01032487 + layer.3.k_cache 0.07289400 3.55422901 + layer.3.v_cache 0.00001855 0.01110179 + layer.4.k_cache 0.00061006 0.21100557 + layer.4.v_cache 0.00005429 0.02391295 + layer.4.output 0.17891866 641.79947917 + ------------------------------------------------------------------------------------- + TOTAL 0.08858556 268.08743308 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 111532 +BPFP 1.2204 bits/point +EBPFP 2.4407 equivalent bits/point +MSE 268.087433 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 268.0874 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,148B, BPFP=0.6073 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,680B, BPFP=2.2531 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,376B, BPFP=1.0370 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,316B, BPFP=2.1829 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,492B, BPFP=1.2523 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,860B, BPFP=2.0949 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,156B, BPFP=1.1875 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,036B, BPFP=2.1289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,152B, BPFP=1.7654 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,752B, BPFP=2.0741 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,464B, BPFP=0.6190 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10837562 57.83648003 + layer.0.v_cache 0.00001644 0.00864222 + layer.1.k_cache 0.01241452 5.27717571 + layer.1.v_cache 0.00000532 0.00367695 + layer.2.k_cache 0.01724591 0.84193562 + layer.2.v_cache 0.00001958 0.00910558 + layer.3.k_cache 0.04641535 3.75951772 + layer.3.v_cache 0.00001847 0.00942143 + layer.4.k_cache 0.00063678 0.18800788 + layer.4.v_cache 0.00004974 0.02004780 + layer.4.output 0.18014439 665.15950176 + ------------------------------------------------------------------------------------- + TOTAL 0.08507109 277.88650137 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 108432 +BPFP 1.2304 bits/point +EBPFP 2.4608 equivalent bits/point +MSE 277.886501 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.003s, Pack+Encode: 0.161s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 277.8865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.2245 bits/point +Avg EBPFP 2.4491 equivalent bits/point +Avg MSE 261.372775 +Avg Time 0.373s +------------------------ ---------------------------- diff --git a/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..0772b86d4e339dd27c29512197a467cb89df9696 --- /dev/null +++ b/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 599 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean +Output output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,040B, BPFP=0.5023 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,440B, BPFP=2.1005 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,916B, BPFP=0.9215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,408B, BPFP=2.0268 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,704B, BPFP=1.1918 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,804B, BPFP=1.9837 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,596B, BPFP=1.1127 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,200B, BPFP=2.0120 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,588B, BPFP=1.6829 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,920B, BPFP=1.9207 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,160B, BPFP=0.6234 +⌛️ [2/4] FRONTEND: Frontend time: 0.755s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.463s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12261864 53.83860410 + layer.0.v_cache 0.00001806 0.01023912 + layer.1.k_cache 0.30198812 4.47900781 + layer.1.v_cache 0.00000607 0.00387669 + layer.2.k_cache 0.01966049 0.75180249 + layer.2.v_cache 0.00002021 0.01002775 + layer.3.k_cache 0.01413035 2.85454291 + layer.3.v_cache 0.00002054 0.01100692 + layer.4.k_cache 0.00067868 0.20545928 + layer.4.v_cache 0.00004987 0.02073427 + layer.4.output 1.39792533 245.91996902 + ------------------------------------------------------------------------------------- + TOTAL 0.60262755 104.91912261 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 277776 +BPFP 1.1658 bits/point +EBPFP 2.3316 equivalent bits/point +MSE 104.919123 +---------------------- -------------------------------------------------------- +Time: 1.226s Load: 0.008s, Pack+Encode: 0.755s, Decode+Unpack: 0.463s +---------------------- -------------------------------------------------------- +💾 Converting with 104.9191 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,052B, BPFP=0.5101 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,880B, BPFP=2.0891 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,868B, BPFP=0.9308 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,148B, BPFP=2.0362 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,604B, BPFP=1.2011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,600B, BPFP=1.9965 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,612B, BPFP=1.1293 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,084B, BPFP=2.0315 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,600B, BPFP=1.7072 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,768B, BPFP=1.9363 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 57,996B, BPFP=0.5993 +⌛️ [2/4] FRONTEND: Frontend time: 0.328s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.421s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11787021 52.58145707 + layer.0.v_cache 0.00001738 0.00957648 + layer.1.k_cache 0.25265321 4.53552981 + layer.1.v_cache 0.00000586 0.00398587 + layer.2.k_cache 0.01010228 0.72874422 + layer.2.v_cache 0.00002017 0.01080051 + layer.3.k_cache 0.01159736 3.03581407 + layer.3.v_cache 0.00002197 0.01180105 + layer.4.k_cache 0.00067244 0.21498461 + layer.4.v_cache 0.00004845 0.02123217 + layer.4.output 1.41728832 249.65457589 + ------------------------------------------------------------------------------------- + TOTAL 0.60670750 106.39623277 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 273212 +BPFP 1.1626 bits/point +EBPFP 2.3251 equivalent bits/point +MSE 106.396233 +---------------------- -------------------------------------------------------- +Time: 0.756s Load: 0.007s, Pack+Encode: 0.328s, Decode+Unpack: 0.421s +---------------------- -------------------------------------------------------- +💾 Converting with 106.3962 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,876B, BPFP=0.4818 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,424B, BPFP=2.0617 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,220B, BPFP=0.9263 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,716B, BPFP=2.0121 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,932B, BPFP=1.1864 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,036B, BPFP=1.9644 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,024B, BPFP=1.1228 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,592B, BPFP=2.0034 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,980B, BPFP=1.6802 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,240B, BPFP=1.9086 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 55,688B, BPFP=0.5574 +⌛️ [2/4] FRONTEND: Frontend time: 0.311s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.446s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13311632 52.74990366 + layer.0.v_cache 0.00001526 0.00961892 + layer.1.k_cache 0.36474110 4.39570761 + layer.1.v_cache 0.00000573 0.00397578 + layer.2.k_cache 0.01296863 0.72436585 + layer.2.v_cache 0.00001984 0.01052061 + layer.3.k_cache 0.01490937 2.75931840 + layer.3.v_cache 0.00002017 0.01137276 + layer.4.k_cache 0.00067370 0.20965474 + layer.4.v_cache 0.00005228 0.02076569 + layer.4.output 1.37278975 241.40028427 + ------------------------------------------------------------------------------------- + TOTAL 0.59623828 102.98218788 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 274728 +BPFP 1.1323 bits/point +EBPFP 2.2646 equivalent bits/point +MSE 102.982188 +---------------------- -------------------------------------------------------- +Time: 0.766s Load: 0.009s, Pack+Encode: 0.311s, Decode+Unpack: 0.446s +---------------------- -------------------------------------------------------- +💾 Converting with 102.9822 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,780B, BPFP=0.4982 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,552B, BPFP=2.0205 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,896B, BPFP=0.8899 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,832B, BPFP=1.9744 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,080B, BPFP=1.1578 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,184B, BPFP=1.9329 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,048B, BPFP=1.0917 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,684B, BPFP=1.9649 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,692B, BPFP=1.6452 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,348B, BPFP=1.8794 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,880B, BPFP=0.5844 +⌛️ [2/4] FRONTEND: Frontend time: 0.331s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.453s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11263794 50.52988521 + layer.0.v_cache 0.00001639 0.00983403 + layer.1.k_cache 0.42101050 4.83397637 + layer.1.v_cache 0.00000577 0.00386405 + layer.2.k_cache 0.00499437 0.70430074 + layer.2.v_cache 0.00001973 0.01029775 + layer.3.k_cache 0.03029772 3.14875418 + layer.3.v_cache 0.00002073 0.01123656 + layer.4.k_cache 0.00068836 0.20723202 + layer.4.v_cache 0.00005112 0.02123558 + layer.4.output 1.25471843 220.86266833 + ------------------------------------------------------------------------------------- + TOTAL 0.55016304 94.44231146 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 298976 +BPFP 1.1262 bits/point +EBPFP 2.2524 equivalent bits/point +MSE 94.442311 +---------------------- -------------------------------------------------------- +Time: 0.793s Load: 0.009s, Pack+Encode: 0.331s, Decode+Unpack: 0.453s +---------------------- -------------------------------------------------------- +💾 Converting with 94.4423 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,992B, BPFP=0.5011 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,320B, BPFP=2.1015 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,868B, BPFP=0.9223 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,356B, BPFP=2.0324 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,688B, BPFP=1.1961 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,736B, BPFP=1.9880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,628B, BPFP=1.1201 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,244B, BPFP=2.0244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,612B, BPFP=1.6924 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,028B, BPFP=1.9372 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 58,744B, BPFP=0.6015 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.401s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10233516 53.28629408 + layer.0.v_cache 0.00001822 0.00987704 + layer.1.k_cache 0.31275415 4.34530444 + layer.1.v_cache 0.00000587 0.00377551 + layer.2.k_cache 0.01225604 0.71763009 + layer.2.v_cache 0.00002018 0.00992106 + layer.3.k_cache 0.02847941 2.83274043 + layer.3.v_cache 0.00002101 0.01113515 + layer.4.k_cache 0.00067306 0.20643539 + layer.4.v_cache 0.00005947 0.02059547 + layer.4.output 1.40430263 246.62805128 + ------------------------------------------------------------------------------------- + TOTAL 0.60510241 105.16706280 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 275216 +BPFP 1.1603 bits/point +EBPFP 2.3207 equivalent bits/point +MSE 105.167063 +---------------------- -------------------------------------------------------- +Time: 0.752s Load: 0.008s, Pack+Encode: 0.342s, Decode+Unpack: 0.401s +---------------------- -------------------------------------------------------- +💾 Converting with 105.1671 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,060B, BPFP=0.4815 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,804B, BPFP=2.0623 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,392B, BPFP=0.9243 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,692B, BPFP=2.0032 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,920B, BPFP=1.1650 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,648B, BPFP=1.9477 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,528B, BPFP=1.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,280B, BPFP=1.9813 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,172B, BPFP=1.6567 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,764B, BPFP=1.9007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,732B, BPFP=0.5978 +⌛️ [2/4] FRONTEND: Frontend time: 0.430s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.497s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16298727 51.44031343 + layer.0.v_cache 0.00001770 0.00953822 + layer.1.k_cache 0.61736183 4.59529872 + layer.1.v_cache 0.00000612 0.00379669 + layer.2.k_cache 0.02593859 0.78464160 + layer.2.v_cache 0.00002024 0.00989988 + layer.3.k_cache 0.01326758 2.82251973 + layer.3.v_cache 0.00002270 0.01122931 + layer.4.k_cache 0.00075427 0.20511811 + layer.4.v_cache 0.00005043 0.01913111 + layer.4.output 0.04538776 183.92431973 + ------------------------------------------------------------------------------------- + TOTAL 0.06694947 79.25716029 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 364992 +BPFP 1.1411 bits/point +EBPFP 2.2821 equivalent bits/point +MSE 79.257160 +---------------------- -------------------------------------------------------- +Time: 0.937s Load: 0.010s, Pack+Encode: 0.430s, Decode+Unpack: 0.497s +---------------------- -------------------------------------------------------- +💾 Converting with 79.2572 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 239, 128) +Output shape: (1, 239, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.output: torch.Size([1, 239, 3584]) -> torch.Size([1, 1, 239, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,436B, BPFP=0.4861 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,456B, BPFP=2.0565 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,892B, BPFP=0.9082 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,532B, BPFP=1.9961 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,608B, BPFP=1.1512 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,792B, BPFP=1.9477 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,660B, BPFP=1.0892 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,308B, BPFP=1.9814 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,276B, BPFP=1.6525 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,924B, BPFP=1.8910 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,488B, BPFP=0.6303 +⌛️ [2/4] FRONTEND: Frontend time: 0.278s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.399s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11863919 53.27485127 + layer.0.v_cache 0.00001702 0.00990044 + layer.1.k_cache 0.43747612 4.57755908 + layer.1.v_cache 0.00000797 0.00400439 + layer.2.k_cache 0.02286501 0.73841615 + layer.2.v_cache 0.00002174 0.00998055 + layer.3.k_cache 0.01903599 2.79839705 + layer.3.v_cache 0.00002123 0.01098946 + layer.4.k_cache 0.00068597 0.20534343 + layer.4.v_cache 0.00006921 0.02022131 + layer.4.output 1.28101462 225.75341826 + ------------------------------------------------------------------------------------- + TOTAL 0.56270246 96.58374064 + (elements=2,080,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2080256 +Total Bytes 299372 +BPFP 1.1513 bits/point +EBPFP 2.3026 equivalent bits/point +MSE 96.583741 +---------------------- -------------------------------------------------------- +Time: 0.685s Load: 0.008s, Pack+Encode: 0.278s, Decode+Unpack: 0.399s +---------------------- -------------------------------------------------------- +💾 Converting with 96.5837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,252B, BPFP=0.4903 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,200B, BPFP=2.0913 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,676B, BPFP=0.9313 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,196B, BPFP=2.0316 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,904B, BPFP=1.1825 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,408B, BPFP=1.9848 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,760B, BPFP=1.1145 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,980B, BPFP=2.0188 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,532B, BPFP=1.6951 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,668B, BPFP=1.9408 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,988B, BPFP=0.6195 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.460s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15151563 52.46898393 + layer.0.v_cache 0.00001631 0.00989214 + layer.1.k_cache 0.46755480 4.56984603 + layer.1.v_cache 0.00000667 0.00389832 + layer.2.k_cache 0.02171374 0.73246568 + layer.2.v_cache 0.00002115 0.01035590 + layer.3.k_cache 0.01385514 3.08769499 + layer.3.v_cache 0.00002107 0.01125085 + layer.4.k_cache 0.00069081 0.20752413 + layer.4.v_cache 0.00005638 0.02142021 + layer.4.output 0.00504555 210.33247895 + ------------------------------------------------------------------------------------- + TOTAL 0.04063356 90.20298146 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 333564 +BPFP 1.1657 bits/point +EBPFP 2.3314 equivalent bits/point +MSE 90.202981 +---------------------- -------------------------------------------------------- +Time: 0.810s Load: 0.009s, Pack+Encode: 0.342s, Decode+Unpack: 0.460s +---------------------- -------------------------------------------------------- +💾 Converting with 90.2030 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,812B, BPFP=0.4731 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,668B, BPFP=2.0603 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,080B, BPFP=0.9083 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,864B, BPFP=2.0044 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,420B, BPFP=1.1403 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,948B, BPFP=1.9408 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,532B, BPFP=1.0786 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,416B, BPFP=1.9733 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,568B, BPFP=1.6367 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,060B, BPFP=1.8792 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,424B, BPFP=0.6094 +⌛️ [2/4] FRONTEND: Frontend time: 0.299s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.403s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14529029 54.25254340 + layer.0.v_cache 0.00001764 0.00980396 + layer.1.k_cache 0.31089128 4.55579481 + layer.1.v_cache 0.00000630 0.00401613 + layer.2.k_cache 0.01355181 0.75750699 + layer.2.v_cache 0.00002032 0.01026230 + layer.3.k_cache 0.02426839 3.09876519 + layer.3.v_cache 0.00002043 0.01141216 + layer.4.k_cache 0.00068925 0.19960256 + layer.4.v_cache 0.00005165 0.02032527 + layer.4.output 1.36066019 239.97341270 + ------------------------------------------------------------------------------------- + TOTAL 0.58937816 102.51376010 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 278792 +BPFP 1.1389 bits/point +EBPFP 2.2777 equivalent bits/point +MSE 102.513760 +---------------------- -------------------------------------------------------- +Time: 0.711s Load: 0.008s, Pack+Encode: 0.299s, Decode+Unpack: 0.403s +---------------------- -------------------------------------------------------- +💾 Converting with 102.5138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 262, 128) +Output shape: (1, 262, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.output: torch.Size([1, 262, 3584]) -> torch.Size([1, 1, 262, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,256B, BPFP=0.4924 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,008B, BPFP=2.0878 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,480B, BPFP=0.9232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,036B, BPFP=2.0298 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,000B, BPFP=1.1927 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,376B, BPFP=1.9905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,792B, BPFP=1.1207 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,932B, BPFP=2.0236 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,384B, BPFP=1.6927 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,640B, BPFP=1.9466 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,888B, BPFP=0.6210 +⌛️ [2/4] FRONTEND: Frontend time: 0.337s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.480s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12917162 51.82033114 + layer.0.v_cache 0.00001830 0.00952505 + layer.1.k_cache 0.40634074 4.51189766 + layer.1.v_cache 0.00000633 0.00389127 + layer.2.k_cache 0.02363680 0.77598024 + layer.2.v_cache 0.00002248 0.01009964 + layer.3.k_cache 0.02343226 3.02106528 + layer.3.v_cache 0.00002042 0.01126224 + layer.4.k_cache 0.00071170 0.20515405 + layer.4.v_cache 0.00005093 0.02071223 + layer.4.output 0.00505940 211.16577495 + ------------------------------------------------------------------------------------- + TOTAL 0.03640161 90.50296138 + (elements=2,280,448) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2280448 +Total Bytes 332792 +BPFP 1.1675 bits/point +EBPFP 2.3349 equivalent bits/point +MSE 90.502961 +---------------------- -------------------------------------------------------- +Time: 0.826s Load: 0.009s, Pack+Encode: 0.337s, Decode+Unpack: 0.480s +---------------------- -------------------------------------------------------- +💾 Converting with 90.5030 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,768B, BPFP=0.4700 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,736B, BPFP=2.0650 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,092B, BPFP=0.9092 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,752B, BPFP=1.9967 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,752B, BPFP=1.1633 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,088B, BPFP=1.9506 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,796B, BPFP=1.0969 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,584B, BPFP=1.9850 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,896B, BPFP=1.6594 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,320B, BPFP=1.8972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,976B, BPFP=0.6148 +⌛️ [2/4] FRONTEND: Frontend time: 0.314s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.440s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18074451 53.27500000 + layer.0.v_cache 0.00001764 0.00977275 + layer.1.k_cache 0.33421082 5.06800564 + layer.1.v_cache 0.00000655 0.00400995 + layer.2.k_cache 0.01487562 0.73156731 + layer.2.v_cache 0.00002082 0.01018522 + layer.3.k_cache 0.01319740 2.86681505 + layer.3.v_cache 0.00002096 0.01137372 + layer.4.k_cache 0.00067848 0.20851283 + layer.4.v_cache 0.00005239 0.02095617 + layer.4.output 1.36066546 240.07619048 + ------------------------------------------------------------------------------------- + TOTAL 0.59226373 102.51409012 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 280760 +BPFP 1.1469 bits/point +EBPFP 2.2938 equivalent bits/point +MSE 102.514090 +---------------------- -------------------------------------------------------- +Time: 0.763s Load: 0.008s, Pack+Encode: 0.314s, Decode+Unpack: 0.440s +---------------------- -------------------------------------------------------- +💾 Converting with 102.5141 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 232, 128) +Output shape: (1, 232, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.output: torch.Size([1, 232, 3584]) -> torch.Size([1, 1, 232, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,320B, BPFP=0.4930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,824B, BPFP=2.0760 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,868B, BPFP=0.9340 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,836B, BPFP=2.0094 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,408B, BPFP=1.1724 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,344B, BPFP=1.9763 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,316B, BPFP=1.0989 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,684B, BPFP=1.9992 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,100B, BPFP=1.6905 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,640B, BPFP=1.9289 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,076B, BPFP=0.6357 +⌛️ [2/4] FRONTEND: Frontend time: 0.310s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11978169 52.67056169 + layer.0.v_cache 0.00001746 0.00965471 + layer.1.k_cache 0.32299256 4.42787486 + layer.1.v_cache 0.00000596 0.00362401 + layer.2.k_cache 0.02091339 0.78000601 + layer.2.v_cache 0.00002092 0.01005848 + layer.3.k_cache 0.01769928 3.01712457 + layer.3.v_cache 0.00002180 0.01101512 + layer.4.k_cache 0.00071208 0.20538192 + layer.4.v_cache 0.00005763 0.02042419 + layer.4.output 1.31963615 232.40113147 + ------------------------------------------------------------------------------------- + TOTAL 0.57174564 99.29197917 + (elements=2,019,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2019328 +Total Bytes 294416 +BPFP 1.1664 bits/point +EBPFP 2.3328 equivalent bits/point +MSE 99.291979 +---------------------- -------------------------------------------------------- +Time: 0.731s Load: 0.010s, Pack+Encode: 0.310s, Decode+Unpack: 0.411s +---------------------- -------------------------------------------------------- +💾 Converting with 99.2920 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,024B, BPFP=0.4989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,628B, BPFP=2.1043 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,296B, BPFP=0.9443 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,756B, BPFP=2.0423 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,924B, BPFP=1.2020 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,852B, BPFP=1.9781 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,868B, BPFP=1.1270 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,472B, BPFP=2.0222 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,012B, BPFP=1.7054 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,344B, BPFP=1.9420 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,836B, BPFP=0.6680 +⌛️ [2/4] FRONTEND: Frontend time: 0.308s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.424s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15639258 53.20035067 + layer.0.v_cache 0.00001700 0.01015191 + layer.1.k_cache 0.33150226 4.73968950 + layer.1.v_cache 0.00000613 0.00378037 + layer.2.k_cache 0.01694729 0.73316096 + layer.2.v_cache 0.00002208 0.01004233 + layer.3.k_cache 0.03858858 3.22444902 + layer.3.v_cache 0.00002090 0.01157236 + layer.4.k_cache 0.00067984 0.20737721 + layer.4.v_cache 0.00005217 0.02136279 + layer.4.output 1.39161012 244.96948052 + ------------------------------------------------------------------------------------- + TOTAL 0.60502939 104.52637063 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 285012 +BPFP 1.1907 bits/point +EBPFP 2.3815 equivalent bits/point +MSE 104.526371 +---------------------- -------------------------------------------------------- +Time: 0.740s Load: 0.008s, Pack+Encode: 0.308s, Decode+Unpack: 0.424s +---------------------- -------------------------------------------------------- +💾 Converting with 104.5264 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,644B, BPFP=0.4947 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,240B, BPFP=2.0742 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,896B, BPFP=0.9098 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,272B, BPFP=2.0188 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,344B, BPFP=1.1644 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,444B, BPFP=1.9714 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,316B, BPFP=1.1055 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,824B, BPFP=1.9931 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,108B, BPFP=1.6660 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,392B, BPFP=1.9112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 77,668B, BPFP=0.6350 +⌛️ [2/4] FRONTEND: Frontend time: 0.359s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.494s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13605902 52.01293498 + layer.0.v_cache 0.00001757 0.00943949 + layer.1.k_cache 0.58683352 4.24767888 + layer.1.v_cache 0.00000623 0.00387717 + layer.2.k_cache 0.01500385 0.72373946 + layer.2.v_cache 0.00002229 0.01021706 + layer.3.k_cache 0.00978726 3.14356095 + layer.3.v_cache 0.00001998 0.01068168 + layer.4.k_cache 0.00067570 0.20327748 + layer.4.v_cache 0.00005472 0.02039543 + layer.4.output 0.00490606 202.31869767 + ------------------------------------------------------------------------------------- + TOTAL 0.04604839 86.85980507 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 345148 +BPFP 1.1620 bits/point +EBPFP 2.3240 equivalent bits/point +MSE 86.859805 +---------------------- -------------------------------------------------------- +Time: 0.861s Load: 0.008s, Pack+Encode: 0.359s, Decode+Unpack: 0.494s +---------------------- -------------------------------------------------------- +💾 Converting with 86.8598 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,092B, BPFP=0.4958 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,040B, BPFP=1.9632 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,576B, BPFP=0.8931 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,180B, BPFP=1.9105 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,592B, BPFP=1.1392 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,372B, BPFP=1.8610 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,716B, BPFP=1.0855 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,920B, BPFP=1.8946 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,936B, BPFP=1.5892 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,464B, BPFP=1.8054 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,948B, BPFP=0.6210 +⌛️ [2/4] FRONTEND: Frontend time: 0.293s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.416s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12967140 50.02746630 + layer.0.v_cache 0.00001601 0.00926890 + layer.1.k_cache 0.49353925 4.52021389 + layer.1.v_cache 0.00000620 0.00370766 + layer.2.k_cache 0.03189284 0.75906893 + layer.2.v_cache 0.00002040 0.00975464 + layer.3.k_cache 0.04285583 2.68861108 + layer.3.v_cache 0.00002221 0.01076448 + layer.4.k_cache 0.00071132 0.19841306 + layer.4.v_cache 0.00005107 0.01974545 + layer.4.output 1.20063613 202.43939076 + ------------------------------------------------------------------------------------- + TOTAL 0.53548467 86.78369116 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 309836 +BPFP 1.1168 bits/point +EBPFP 2.2335 equivalent bits/point +MSE 86.783691 +---------------------- -------------------------------------------------------- +Time: 0.718s Load: 0.008s, Pack+Encode: 0.293s, Decode+Unpack: 0.416s +---------------------- -------------------------------------------------------- +💾 Converting with 86.7837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,744B, BPFP=0.4859 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,796B, BPFP=1.9952 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,184B, BPFP=0.8901 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,032B, BPFP=1.9473 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,136B, BPFP=1.1381 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,312B, BPFP=1.9021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,168B, BPFP=1.0773 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,704B, BPFP=1.9267 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,912B, BPFP=1.6260 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,300B, BPFP=1.8386 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,772B, BPFP=0.5627 +⌛️ [2/4] FRONTEND: Frontend time: 0.358s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.461s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16529650 52.61433233 + layer.0.v_cache 0.00001613 0.00976082 + layer.1.k_cache 0.54577502 4.51915278 + layer.1.v_cache 0.00000662 0.00409166 + layer.2.k_cache 0.01453520 0.74394575 + layer.2.v_cache 0.00002096 0.01077857 + layer.3.k_cache 0.02409829 2.81131705 + layer.3.v_cache 0.00002071 0.01172115 + layer.4.k_cache 0.00068699 0.21695273 + layer.4.v_cache 0.00004971 0.02139975 + layer.4.output 1.22953407 214.68167312 + ------------------------------------------------------------------------------------- + TOTAL 0.55042615 91.98442144 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 299060 +BPFP 1.1039 bits/point +EBPFP 2.2078 equivalent bits/point +MSE 91.984421 +---------------------- -------------------------------------------------------- +Time: 0.826s Load: 0.008s, Pack+Encode: 0.358s, Decode+Unpack: 0.461s +---------------------- -------------------------------------------------------- +💾 Converting with 91.9844 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,184B, BPFP=0.4866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,196B, BPFP=2.0640 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,384B, BPFP=0.9262 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,204B, BPFP=2.0166 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,740B, BPFP=1.1821 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,200B, BPFP=1.9687 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,128B, BPFP=1.1051 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,808B, BPFP=1.9977 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,828B, BPFP=1.6642 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,980B, BPFP=1.9104 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,184B, BPFP=0.6156 +⌛️ [2/4] FRONTEND: Frontend time: 0.472s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.560s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19043710 51.14457306 + layer.0.v_cache 0.00001603 0.00971766 + layer.1.k_cache 0.64205214 4.35937948 + layer.1.v_cache 0.00000630 0.00393595 + layer.2.k_cache 0.01513053 0.74831360 + layer.2.v_cache 0.00002191 0.01059697 + layer.3.k_cache 0.02775778 3.00613338 + layer.3.v_cache 0.00002125 0.01118230 + layer.4.k_cache 0.00071623 0.20026829 + layer.4.v_cache 0.00005146 0.02055456 + layer.4.output 0.04089398 165.37138215 + ------------------------------------------------------------------------------------- + TOTAL 0.06838051 71.59496061 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 410836 +BPFP 1.1548 bits/point +EBPFP 2.3095 equivalent bits/point +MSE 71.594961 +---------------------- -------------------------------------------------------- +Time: 1.042s Load: 0.011s, Pack+Encode: 0.472s, Decode+Unpack: 0.560s +---------------------- -------------------------------------------------------- +💾 Converting with 71.5950 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,912B, BPFP=0.4920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,460B, BPFP=2.0682 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,616B, BPFP=0.9174 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,484B, BPFP=2.0144 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,312B, BPFP=1.1767 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,492B, BPFP=1.9596 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,920B, BPFP=1.0998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,944B, BPFP=1.9845 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,072B, BPFP=1.6603 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,364B, BPFP=1.8973 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 79,192B, BPFP=0.6246 +⌛️ [2/4] FRONTEND: Frontend time: 0.364s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.528s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14392723 52.59329450 + layer.0.v_cache 0.00001682 0.00961068 + layer.1.k_cache 0.53349789 4.54494776 + layer.1.v_cache 0.00000638 0.00377590 + layer.2.k_cache 0.01523476 0.75788637 + layer.2.v_cache 0.00002109 0.00974982 + layer.3.k_cache 0.03867486 3.13320114 + layer.3.v_cache 0.00001994 0.01073634 + layer.4.k_cache 0.00070525 0.19736413 + layer.4.v_cache 0.00005078 0.01896963 + layer.4.output 0.00472849 195.13427562 + ------------------------------------------------------------------------------------- + TOTAL 0.04501497 83.95408621 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 355768 +BPFP 1.1555 bits/point +EBPFP 2.3109 equivalent bits/point +MSE 83.954086 +---------------------- -------------------------------------------------------- +Time: 0.900s Load: 0.009s, Pack+Encode: 0.364s, Decode+Unpack: 0.528s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9541 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,836B, BPFP=0.4747 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,732B, BPFP=2.0647 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,144B, BPFP=0.9128 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,956B, BPFP=2.0108 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,780B, BPFP=1.1653 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,212B, BPFP=1.9592 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,848B, BPFP=1.1006 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,660B, BPFP=1.9903 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,048B, BPFP=1.6700 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,312B, BPFP=1.8967 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 60,000B, BPFP=0.5952 +⌛️ [2/4] FRONTEND: Frontend time: 0.323s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.416s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14673459 52.28097222 + layer.0.v_cache 0.00001655 0.00989685 + layer.1.k_cache 0.37581726 4.51931478 + layer.1.v_cache 0.00000603 0.00385305 + layer.2.k_cache 0.01062915 0.68972080 + layer.2.v_cache 0.00002006 0.01048982 + layer.3.k_cache 0.04027582 3.14145779 + layer.3.v_cache 0.00002066 0.01145238 + layer.4.k_cache 0.00066373 0.20622840 + layer.4.v_cache 0.00004861 0.02089714 + layer.4.output 1.36064577 239.91164683 + ------------------------------------------------------------------------------------- + TOTAL 0.59404429 102.36916535 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 279528 +BPFP 1.1419 bits/point +EBPFP 2.2837 equivalent bits/point +MSE 102.369165 +---------------------- -------------------------------------------------------- +Time: 0.746s Load: 0.007s, Pack+Encode: 0.323s, Decode+Unpack: 0.416s +---------------------- -------------------------------------------------------- +💾 Converting with 102.3692 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,380B, BPFP=0.4928 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,444B, BPFP=2.0996 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,840B, BPFP=0.9241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,576B, BPFP=2.0417 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,956B, BPFP=1.1990 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,836B, BPFP=1.9923 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,836B, BPFP=1.1242 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,368B, BPFP=2.0278 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,440B, BPFP=1.6987 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,956B, BPFP=1.9335 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,476B, BPFP=0.6532 +⌛️ [2/4] FRONTEND: Frontend time: 0.326s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12385076 52.05849359 + layer.0.v_cache 0.00002004 0.01063664 + layer.1.k_cache 0.38583452 4.76599330 + layer.1.v_cache 0.00000646 0.00415669 + layer.2.k_cache 0.02500911 0.75522542 + layer.2.v_cache 0.00002014 0.01013268 + layer.3.k_cache 0.03425702 3.01417215 + layer.3.v_cache 0.00002057 0.01163046 + layer.4.k_cache 0.00068928 0.21555795 + layer.4.v_cache 0.00004882 0.02072385 + layer.4.output 1.30836928 230.70678800 + ------------------------------------------------------------------------------------- + TOTAL 0.57225539 98.57730816 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 301108 +BPFP 1.1827 bits/point +EBPFP 2.3654 equivalent bits/point +MSE 98.577308 +---------------------- -------------------------------------------------------- +Time: 0.744s Load: 0.007s, Pack+Encode: 0.326s, Decode+Unpack: 0.411s +---------------------- -------------------------------------------------------- +💾 Converting with 98.5773 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 247, 128) +Output shape: (1, 247, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.output: torch.Size([1, 247, 3584]) -> torch.Size([1, 1, 247, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,704B, BPFP=0.4873 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,972B, BPFP=2.0225 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,104B, BPFP=0.8922 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,960B, BPFP=1.9585 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,060B, BPFP=1.1425 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,164B, BPFP=1.9081 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,896B, BPFP=1.0688 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,592B, BPFP=1.9352 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,708B, BPFP=1.6263 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,168B, BPFP=1.8451 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,292B, BPFP=0.5900 +⌛️ [2/4] FRONTEND: Frontend time: 0.297s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.447s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13907220 52.80518962 + layer.0.v_cache 0.00001735 0.00933332 + layer.1.k_cache 0.47208408 4.28052845 + layer.1.v_cache 0.00000611 0.00367998 + layer.2.k_cache 0.03408500 0.70878947 + layer.2.v_cache 0.00001989 0.00920585 + layer.3.k_cache 0.02387486 3.29198576 + layer.3.v_cache 0.00002031 0.01033379 + layer.4.k_cache 0.00069623 0.19796648 + layer.4.v_cache 0.00004922 0.01826656 + layer.4.output 1.23951819 217.83413462 + ------------------------------------------------------------------------------------- + TOTAL 0.54979721 93.30436598 + (elements=2,149,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2149888 +Total Bytes 300620 +BPFP 1.1186 bits/point +EBPFP 2.2373 equivalent bits/point +MSE 93.304366 +---------------------- -------------------------------------------------------- +Time: 0.753s Load: 0.009s, Pack+Encode: 0.297s, Decode+Unpack: 0.447s +---------------------- -------------------------------------------------------- +💾 Converting with 93.3044 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 427, 128) +Output shape: (1, 427, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.output: torch.Size([1, 427, 3584]) -> torch.Size([1, 1, 427, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,060B, BPFP=0.4779 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 54,824B, BPFP=2.0061 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,952B, BPFP=0.8765 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,264B, BPFP=1.9491 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,388B, BPFP=1.1120 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 51,544B, BPFP=1.8861 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,252B, BPFP=1.0338 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 52,164B, BPFP=1.9088 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,756B, BPFP=1.6011 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 50,228B, BPFP=1.8380 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 112,184B, BPFP=0.5864 +⌛️ [2/4] FRONTEND: Frontend time: 0.529s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.656s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15738957 51.12509148 + layer.0.v_cache 0.00001693 0.00881802 + layer.1.k_cache 1.00627419 4.33593550 + layer.1.v_cache 0.00000663 0.00342740 + layer.2.k_cache 0.03590691 0.76921428 + layer.2.v_cache 0.00002108 0.00870100 + layer.3.k_cache 0.02176077 2.74768467 + layer.3.v_cache 0.00002004 0.00953056 + layer.4.k_cache 0.00078997 0.18787786 + layer.4.v_cache 0.00005212 0.01779824 + layer.4.output 0.00621442 129.22007779 + ------------------------------------------------------------------------------------- + TOTAL 0.07445524 56.69144844 + (elements=3,716,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3716608 +Total Bytes 513616 +BPFP 1.1056 bits/point +EBPFP 2.2111 equivalent bits/point +MSE 56.691448 +---------------------- -------------------------------------------------------- +Time: 1.199s Load: 0.014s, Pack+Encode: 0.529s, Decode+Unpack: 0.656s +---------------------- -------------------------------------------------------- +💾 Converting with 56.6914 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,680B, BPFP=0.5005 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,980B, BPFP=2.0745 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,068B, BPFP=0.9264 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,016B, BPFP=2.0189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,096B, BPFP=1.1587 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,052B, BPFP=1.9633 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,956B, BPFP=1.0929 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,516B, BPFP=1.9901 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,784B, BPFP=1.6596 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,948B, BPFP=1.8997 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,804B, BPFP=0.5750 +⌛️ [2/4] FRONTEND: Frontend time: 0.374s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.478s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15504150 52.95844384 + layer.0.v_cache 0.00001776 0.00876130 + layer.1.k_cache 0.59434689 4.21797298 + layer.1.v_cache 0.00000612 0.00343592 + layer.2.k_cache 0.01687181 0.77208425 + layer.2.v_cache 0.00002031 0.00882166 + layer.3.k_cache 0.01346795 3.04904637 + layer.3.v_cache 0.00002023 0.00986705 + layer.4.k_cache 0.00071492 0.18821691 + layer.4.v_cache 0.00006004 0.01795326 + layer.4.output 0.00488892 204.01612744 + ------------------------------------------------------------------------------------- + TOTAL 0.04792882 87.60867621 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 334900 +BPFP 1.1358 bits/point +EBPFP 2.2717 equivalent bits/point +MSE 87.608676 +---------------------- -------------------------------------------------------- +Time: 0.860s Load: 0.008s, Pack+Encode: 0.374s, Decode+Unpack: 0.478s +---------------------- -------------------------------------------------------- +💾 Converting with 87.6087 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 361, 128) +Output shape: (1, 361, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.output: torch.Size([1, 361, 3584]) -> torch.Size([1, 1, 361, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,240B, BPFP=0.4865 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,160B, BPFP=1.9979 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,468B, BPFP=0.8859 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,984B, BPFP=1.9470 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,464B, BPFP=1.1021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,824B, BPFP=1.8968 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,164B, BPFP=1.0459 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,204B, BPFP=1.9133 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,416B, BPFP=1.6195 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,604B, BPFP=1.8440 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,580B, BPFP=0.5353 +⌛️ [2/4] FRONTEND: Frontend time: 0.367s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.508s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21057716 52.49298823 + layer.0.v_cache 0.00001554 0.00831807 + layer.1.k_cache 0.82271198 4.24222402 + layer.1.v_cache 0.00000592 0.00333393 + layer.2.k_cache 0.01184860 0.76247670 + layer.2.v_cache 0.00001987 0.00914849 + layer.3.k_cache 0.01345186 2.90412307 + layer.3.v_cache 0.00002111 0.00957082 + layer.4.k_cache 0.00076976 0.19168355 + layer.4.v_cache 0.00005077 0.01801947 + layer.4.output 0.03703259 149.69668827 + ------------------------------------------------------------------------------------- + TOTAL 0.07757063 65.20698260 + (elements=3,142,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3142144 +Total Bytes 427108 +BPFP 1.0874 bits/point +EBPFP 2.1749 equivalent bits/point +MSE 65.206983 +---------------------- -------------------------------------------------------- +Time: 0.886s Load: 0.011s, Pack+Encode: 0.367s, Decode+Unpack: 0.508s +---------------------- -------------------------------------------------------- +💾 Converting with 65.2070 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,984B, BPFP=0.4978 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,456B, BPFP=2.0199 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,272B, BPFP=0.9016 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,732B, BPFP=1.9798 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,568B, BPFP=1.1396 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,712B, BPFP=1.9233 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,492B, BPFP=1.0800 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,992B, BPFP=1.9388 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,616B, BPFP=1.6410 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,644B, BPFP=1.8641 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,716B, BPFP=0.5677 +⌛️ [2/4] FRONTEND: Frontend time: 0.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.482s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21757852 52.96142924 + layer.0.v_cache 0.00001544 0.00792930 + layer.1.k_cache 0.59705023 4.06620367 + layer.1.v_cache 0.00000603 0.00324347 + layer.2.k_cache 0.00787334 0.71711179 + layer.2.v_cache 0.00001970 0.00858059 + layer.3.k_cache 0.01914395 2.72636619 + layer.3.v_cache 0.00001916 0.00913733 + layer.4.k_cache 0.00071861 0.18694696 + layer.4.v_cache 0.00005003 0.01821539 + layer.4.output 0.00471288 195.65015514 + ------------------------------------------------------------------------------------- + TOTAL 0.05149795 84.13272058 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 342184 +BPFP 1.1153 bits/point +EBPFP 2.2305 equivalent bits/point +MSE 84.132721 +---------------------- -------------------------------------------------------- +Time: 0.816s Load: 0.010s, Pack+Encode: 0.324s, Decode+Unpack: 0.482s +---------------------- -------------------------------------------------------- +💾 Converting with 84.1327 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 372, 128) +Output shape: (1, 372, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.output: torch.Size([1, 372, 3584]) -> torch.Size([1, 1, 372, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,200B, BPFP=0.4704 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,176B, BPFP=1.9815 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,592B, BPFP=0.8649 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,036B, BPFP=1.9336 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,356B, BPFP=1.1070 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,664B, BPFP=1.8760 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,844B, BPFP=1.0435 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,116B, BPFP=1.8950 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,780B, BPFP=1.5869 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,344B, BPFP=1.8206 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,328B, BPFP=0.5960 +⌛️ [2/4] FRONTEND: Frontend time: 0.373s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.513s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18192377 50.09349798 + layer.0.v_cache 0.00001786 0.00862737 + layer.1.k_cache 0.82695811 3.95223704 + layer.1.v_cache 0.00000687 0.00342527 + layer.2.k_cache 0.01216373 0.70085357 + layer.2.v_cache 0.00002073 0.00900554 + layer.3.k_cache 0.01743993 2.98203237 + layer.3.v_cache 0.00002103 0.00971249 + layer.4.k_cache 0.00069676 0.18619726 + layer.4.v_cache 0.00005544 0.01841802 + layer.4.output 0.03606073 145.12863623 + ------------------------------------------------------------------------------------- + TOTAL 0.07598408 63.16849768 + (elements=3,237,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3237888 +Total Bytes 446436 +BPFP 1.1030 bits/point +EBPFP 2.2061 equivalent bits/point +MSE 63.168498 +---------------------- -------------------------------------------------------- +Time: 0.898s Load: 0.013s, Pack+Encode: 0.373s, Decode+Unpack: 0.513s +---------------------- -------------------------------------------------------- +💾 Converting with 63.1685 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,160B, BPFP=0.4915 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,296B, BPFP=2.0461 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,804B, BPFP=0.9096 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,096B, BPFP=1.9880 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,284B, BPFP=1.1264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,088B, BPFP=1.9392 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,124B, BPFP=1.0702 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,608B, BPFP=1.9644 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,708B, BPFP=1.6306 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,828B, BPFP=1.8783 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,764B, BPFP=0.5858 +⌛️ [2/4] FRONTEND: Frontend time: 0.372s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.571s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16470559 52.58724361 + layer.0.v_cache 0.00001558 0.00891273 + layer.1.k_cache 0.70577214 4.43747468 + layer.1.v_cache 0.00000663 0.00352577 + layer.2.k_cache 0.01487649 0.70502701 + layer.2.v_cache 0.00002061 0.00908560 + layer.3.k_cache 0.03012228 2.93014252 + layer.3.v_cache 0.00002030 0.00987811 + layer.4.k_cache 0.00074575 0.19266523 + layer.4.v_cache 0.00005065 0.01780923 + layer.4.output 0.04139163 167.35107530 + ------------------------------------------------------------------------------------- + TOTAL 0.07094573 72.49172303 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 395760 +BPFP 1.1262 bits/point +EBPFP 2.2523 equivalent bits/point +MSE 72.491723 +---------------------- -------------------------------------------------------- +Time: 0.952s Load: 0.010s, Pack+Encode: 0.372s, Decode+Unpack: 0.571s +---------------------- -------------------------------------------------------- +💾 Converting with 72.4917 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,504B, BPFP=0.5014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,244B, BPFP=2.0781 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,768B, BPFP=0.9297 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,304B, BPFP=2.0226 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,748B, BPFP=1.1644 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,420B, BPFP=1.9705 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,612B, BPFP=1.0974 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,944B, BPFP=2.0014 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,316B, BPFP=1.6696 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,476B, BPFP=1.9149 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,680B, BPFP=0.6122 +⌛️ [2/4] FRONTEND: Frontend time: 0.344s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.464s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16585089 52.45952978 + layer.0.v_cache 0.00001549 0.00870218 + layer.1.k_cache 0.52552974 4.31026773 + layer.1.v_cache 0.00000676 0.00357784 + layer.2.k_cache 0.01473994 0.75700966 + layer.2.v_cache 0.00002188 0.00933846 + layer.3.k_cache 0.05129043 3.14128925 + layer.3.v_cache 0.00002012 0.01010132 + layer.4.k_cache 0.00075129 0.19875435 + layer.4.v_cache 0.00005717 0.01897315 + layer.4.output 0.00501330 208.68795485 + ------------------------------------------------------------------------------------- + TOTAL 0.04666922 89.51371927 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 333016 +BPFP 1.1550 bits/point +EBPFP 2.3100 equivalent bits/point +MSE 89.513719 +---------------------- -------------------------------------------------------- +Time: 0.817s Load: 0.009s, Pack+Encode: 0.344s, Decode+Unpack: 0.464s +---------------------- -------------------------------------------------------- +💾 Converting with 89.5137 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 368, 128) +Output shape: (1, 368, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.output: torch.Size([1, 368, 3584]) -> torch.Size([1, 1, 368, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,248B, BPFP=0.4776 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,036B, BPFP=1.9971 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,444B, BPFP=0.8680 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,928B, BPFP=1.9501 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,196B, BPFP=1.1123 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,572B, BPFP=1.8925 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,696B, BPFP=1.0486 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,288B, BPFP=1.9229 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,828B, BPFP=1.6061 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,504B, BPFP=1.8471 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,940B, BPFP=0.5759 +⌛️ [2/4] FRONTEND: Frontend time: 0.374s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.561s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14147143 51.35215162 + layer.0.v_cache 0.00001611 0.00819047 + layer.1.k_cache 0.89062865 4.42520407 + layer.1.v_cache 0.00000637 0.00318205 + layer.2.k_cache 0.02201829 0.72193336 + layer.2.v_cache 0.00002088 0.00900610 + layer.3.k_cache 0.00965744 2.69133079 + layer.3.v_cache 0.00002101 0.00952187 + layer.4.k_cache 0.00077426 0.18320507 + layer.4.v_cache 0.00005366 0.01830176 + layer.4.output 0.03638816 146.84543624 + ------------------------------------------------------------------------------------- + TOTAL 0.07761090 63.96118122 + (elements=3,203,072) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3203072 +Total Bytes 441680 +BPFP 1.1031 bits/point +EBPFP 2.2063 equivalent bits/point +MSE 63.961181 +---------------------- -------------------------------------------------------- +Time: 0.947s Load: 0.012s, Pack+Encode: 0.374s, Decode+Unpack: 0.561s +---------------------- -------------------------------------------------------- +💾 Converting with 63.9612 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 453, 128) +Output shape: (1, 453, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.output: torch.Size([1, 453, 3584]) -> torch.Size([1, 1, 453, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,968B, BPFP=0.4818 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,700B, BPFP=2.0247 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,424B, BPFP=0.8769 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,804B, BPFP=1.9593 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,444B, BPFP=1.1191 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,472B, BPFP=1.9134 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,864B, BPFP=1.0646 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,408B, BPFP=1.9456 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,800B, BPFP=1.6142 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,052B, BPFP=1.8644 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 114,772B, BPFP=0.5655 +⌛️ [2/4] FRONTEND: Frontend time: 0.554s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.699s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13665763 51.67407388 + layer.0.v_cache 0.00001897 0.00897505 + layer.1.k_cache 1.07282356 4.38360205 + layer.1.v_cache 0.00000630 0.00347442 + layer.2.k_cache 0.02241544 0.73796031 + layer.2.v_cache 0.00001979 0.00908822 + layer.3.k_cache 0.01806638 2.82053778 + layer.3.v_cache 0.00002079 0.01032977 + layer.4.k_cache 0.00076567 0.19350610 + layer.4.v_cache 0.00005104 0.01844078 + layer.4.output 0.00584071 121.99934957 + ------------------------------------------------------------------------------------- + TOTAL 0.07598415 53.75620208 + (elements=3,942,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3942912 +Total Bytes 545708 +BPFP 1.1072 bits/point +EBPFP 2.2144 equivalent bits/point +MSE 53.756202 +---------------------- -------------------------------------------------------- +Time: 1.268s Load: 0.015s, Pack+Encode: 0.554s, Decode+Unpack: 0.699s +---------------------- -------------------------------------------------------- +💾 Converting with 53.7562 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 402, 128) +Output shape: (1, 402, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.output: torch.Size([1, 402, 3584]) -> torch.Size([1, 1, 402, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,488B, BPFP=0.4854 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 51,812B, BPFP=2.0138 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,320B, BPFP=0.8675 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 49,964B, BPFP=1.9420 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,404B, BPFP=1.1040 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 48,644B, BPFP=1.8907 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,716B, BPFP=1.0384 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 49,572B, BPFP=1.9268 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,764B, BPFP=1.5844 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 47,180B, BPFP=1.8338 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 97,348B, BPFP=0.5405 +⌛️ [2/4] FRONTEND: Frontend time: 0.432s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.622s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18062694 51.23549246 + layer.0.v_cache 0.00001752 0.00850392 + layer.1.k_cache 0.89169403 3.99570264 + layer.1.v_cache 0.00000593 0.00310403 + layer.2.k_cache 0.03854401 0.71091457 + layer.2.v_cache 0.00001967 0.00844141 + layer.3.k_cache 0.02336847 2.73254577 + layer.3.v_cache 0.00001924 0.00955037 + layer.4.k_cache 0.00085529 0.18607607 + layer.4.v_cache 0.00004959 0.01735405 + layer.4.output 0.00648942 137.19971571 + ------------------------------------------------------------------------------------- + TOTAL 0.06944863 59.95915854 + (elements=3,499,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3499008 +Total Bytes 475212 +BPFP 1.0865 bits/point +EBPFP 2.1730 equivalent bits/point +MSE 59.959159 +---------------------- -------------------------------------------------------- +Time: 1.068s Load: 0.014s, Pack+Encode: 0.432s, Decode+Unpack: 0.622s +---------------------- -------------------------------------------------------- +💾 Converting with 59.9592 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,804B, BPFP=0.4931 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,020B, BPFP=2.0733 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,408B, BPFP=0.9189 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,028B, BPFP=2.0177 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,152B, BPFP=1.1846 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,060B, BPFP=1.9635 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,848B, BPFP=1.1116 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,680B, BPFP=1.9982 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,976B, BPFP=1.6788 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,196B, BPFP=1.9151 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,240B, BPFP=0.6260 +⌛️ [2/4] FRONTEND: Frontend time: 0.339s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.477s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13958113 52.58736909 + layer.0.v_cache 0.00001688 0.00999702 + layer.1.k_cache 0.56864935 4.19108467 + layer.1.v_cache 0.00000621 0.00414673 + layer.2.k_cache 0.01018704 0.73347818 + layer.2.v_cache 0.00002046 0.01072926 + layer.3.k_cache 0.04374025 2.98905699 + layer.3.v_cache 0.00002178 0.01187703 + layer.4.k_cache 0.00068741 0.20957039 + layer.4.v_cache 0.00005272 0.02128215 + layer.4.output 0.00479460 198.15850614 + ------------------------------------------------------------------------------------- + TOTAL 0.04685444 85.16930203 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 352412 +BPFP 1.1610 bits/point +EBPFP 2.3219 equivalent bits/point +MSE 85.169302 +---------------------- -------------------------------------------------------- +Time: 0.825s Load: 0.008s, Pack+Encode: 0.339s, Decode+Unpack: 0.477s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1693 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 381, 128) +Output shape: (1, 381, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.output: torch.Size([1, 381, 3584]) -> torch.Size([1, 1, 381, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,952B, BPFP=0.4902 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,780B, BPFP=1.9595 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,520B, BPFP=0.8825 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,528B, BPFP=1.9081 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,676B, BPFP=1.1350 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,492B, BPFP=1.8656 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,172B, BPFP=1.0733 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,100B, BPFP=1.8906 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,620B, BPFP=1.5838 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,944B, BPFP=1.8022 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,580B, BPFP=0.6010 +⌛️ [2/4] FRONTEND: Frontend time: 0.377s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.561s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15713709 49.93440883 + layer.0.v_cache 0.00001719 0.00946987 + layer.1.k_cache 0.86767426 3.96388449 + layer.1.v_cache 0.00000622 0.00379989 + layer.2.k_cache 0.01635739 0.69945512 + layer.2.v_cache 0.00002098 0.00992784 + layer.3.k_cache 0.01709203 2.80904294 + layer.3.v_cache 0.00002073 0.01077438 + layer.4.k_cache 0.00073405 0.19743794 + layer.4.v_cache 0.00005083 0.01983615 + layer.4.output 0.03519400 139.44387420 + ------------------------------------------------------------------------------------- + TOTAL 0.07679228 60.80971511 + (elements=3,316,224) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3316224 +Total Bytes 458364 +BPFP 1.1057 bits/point +EBPFP 2.2115 equivalent bits/point +MSE 60.809715 +---------------------- -------------------------------------------------------- +Time: 0.951s Load: 0.013s, Pack+Encode: 0.377s, Decode+Unpack: 0.561s +---------------------- -------------------------------------------------------- +💾 Converting with 60.8097 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,244B, BPFP=0.4837 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,220B, BPFP=2.0847 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,928B, BPFP=0.9300 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,416B, BPFP=2.0310 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,796B, BPFP=1.1883 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,712B, BPFP=1.9840 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,736B, BPFP=1.1175 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,172B, BPFP=2.0147 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,332B, BPFP=1.6915 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,976B, BPFP=1.9348 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,892B, BPFP=0.6285 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.416s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422791 53.65897102 + layer.0.v_cache 0.00001930 0.00983367 + layer.1.k_cache 0.38928963 4.54589531 + layer.1.v_cache 0.00000631 0.00378832 + layer.2.k_cache 0.01623082 0.77764658 + layer.2.v_cache 0.00002095 0.01023400 + layer.3.k_cache 0.01313608 3.10464399 + layer.3.v_cache 0.00002100 0.01133586 + layer.4.k_cache 0.00066994 0.20785565 + layer.4.v_cache 0.00005661 0.02110979 + layer.4.output 1.30836332 230.64955357 + ------------------------------------------------------------------------------------- + TOTAL 0.57013069 98.64106995 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 297424 +BPFP 1.1682 bits/point +EBPFP 2.3365 equivalent bits/point +MSE 98.641070 +---------------------- -------------------------------------------------------- +Time: 0.765s Load: 0.008s, Pack+Encode: 0.342s, Decode+Unpack: 0.416s +---------------------- -------------------------------------------------------- +💾 Converting with 98.6411 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,428B, BPFP=0.4918 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,212B, BPFP=2.0665 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,676B, BPFP=0.9055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,444B, BPFP=2.0156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,620B, BPFP=1.1666 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,688B, BPFP=1.9656 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,804B, BPFP=1.1126 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,192B, BPFP=1.9989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,268B, BPFP=1.6729 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,848B, BPFP=1.9100 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,536B, BPFP=0.6009 +⌛️ [2/4] FRONTEND: Frontend time: 0.308s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.403s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13591524 53.66180316 + layer.0.v_cache 0.00001656 0.00993520 + layer.1.k_cache 0.34520291 4.40155598 + layer.1.v_cache 0.00000598 0.00406922 + layer.2.k_cache 0.01250870 0.75337711 + layer.2.v_cache 0.00001982 0.01028134 + layer.3.k_cache 0.02633334 3.11028432 + layer.3.v_cache 0.00002177 0.01190921 + layer.4.k_cache 0.00067446 0.20812979 + layer.4.v_cache 0.00005295 0.02089072 + layer.4.output 1.29725540 228.60317040 + ------------------------------------------------------------------------------------- + TOTAL 0.56479644 97.78908405 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 294716 +BPFP 1.1478 bits/point +EBPFP 2.2956 equivalent bits/point +MSE 97.789084 +---------------------- -------------------------------------------------------- +Time: 0.719s Load: 0.008s, Pack+Encode: 0.308s, Decode+Unpack: 0.403s +---------------------- -------------------------------------------------------- +💾 Converting with 97.7891 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,084B, BPFP=0.5101 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,904B, BPFP=2.0812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,080B, BPFP=0.9418 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,140B, BPFP=2.0262 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,404B, BPFP=1.1812 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,576B, BPFP=1.9856 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,384B, BPFP=1.1077 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,904B, BPFP=2.0092 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,416B, BPFP=1.6861 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,636B, BPFP=1.9179 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 57,408B, BPFP=0.5905 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.423s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150146 53.33781412 + layer.0.v_cache 0.00001564 0.00944119 + layer.1.k_cache 0.36060091 4.28000792 + layer.1.v_cache 0.00000598 0.00377757 + layer.2.k_cache 0.01419523 0.70268299 + layer.2.v_cache 0.00002081 0.01008784 + layer.3.k_cache 0.03680107 2.91054665 + layer.3.v_cache 0.00002092 0.01093014 + layer.4.k_cache 0.00067164 0.20415701 + layer.4.v_cache 0.00004838 0.01979035 + layer.4.output 1.41076765 247.80466178 + ------------------------------------------------------------------------------------- + TOTAL 0.61289739 105.65422755 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 271936 +BPFP 1.1518 bits/point +EBPFP 2.3036 equivalent bits/point +MSE 105.654228 +---------------------- -------------------------------------------------------- +Time: 0.773s Load: 0.008s, Pack+Encode: 0.342s, Decode+Unpack: 0.423s +---------------------- -------------------------------------------------------- +💾 Converting with 105.6542 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 343, 128) +Output shape: (1, 343, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.output: torch.Size([1, 343, 3584]) -> torch.Size([1, 1, 343, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,700B, BPFP=0.4874 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,892B, BPFP=2.0450 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,688B, BPFP=0.8969 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,768B, BPFP=1.9938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,212B, BPFP=1.1485 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,480B, BPFP=1.9351 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,796B, BPFP=1.0840 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,208B, BPFP=1.9683 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,988B, BPFP=1.6394 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,232B, BPFP=1.8783 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,388B, BPFP=0.6273 +⌛️ [2/4] FRONTEND: Frontend time: 0.373s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.562s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16677478 51.76040907 + layer.0.v_cache 0.00001698 0.00933722 + layer.1.k_cache 0.72144124 4.05513592 + layer.1.v_cache 0.00000659 0.00383652 + layer.2.k_cache 0.01351720 0.72704157 + layer.2.v_cache 0.00002117 0.00984036 + layer.3.k_cache 0.02347133 2.80516866 + layer.3.v_cache 0.00002006 0.01073538 + layer.4.k_cache 0.00069292 0.19499151 + layer.4.v_cache 0.00005144 0.01992679 + layer.4.output 0.03905215 157.42130883 + ------------------------------------------------------------------------------------- + TOTAL 0.07055169 68.32621087 + (elements=2,985,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2985472 +Total Bytes 427352 +BPFP 1.1452 bits/point +EBPFP 2.2903 equivalent bits/point +MSE 68.326211 +---------------------- -------------------------------------------------------- +Time: 0.948s Load: 0.012s, Pack+Encode: 0.373s, Decode+Unpack: 0.562s +---------------------- -------------------------------------------------------- +💾 Converting with 68.3262 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,728B, BPFP=0.4959 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,184B, BPFP=2.0559 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,820B, BPFP=0.8989 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,096B, BPFP=1.9941 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,140B, BPFP=1.1443 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,180B, BPFP=1.9420 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,924B, BPFP=1.0752 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,688B, BPFP=1.9709 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,908B, BPFP=1.6425 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,200B, BPFP=1.8864 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 73,884B, BPFP=0.5997 +⌛️ [2/4] FRONTEND: Frontend time: 0.347s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.461s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13317820 51.40082031 + layer.0.v_cache 0.00001784 0.00934637 + layer.1.k_cache 0.51782770 4.20929288 + layer.1.v_cache 0.00000612 0.00377398 + layer.2.k_cache 0.01694398 0.72202193 + layer.2.v_cache 0.00002016 0.00967085 + layer.3.k_cache 0.01920085 2.78606800 + layer.3.v_cache 0.00002016 0.01041013 + layer.4.k_cache 0.00071488 0.20493908 + layer.4.v_cache 0.00005349 0.01972373 + layer.4.output 0.00484358 201.05547078 + ------------------------------------------------------------------------------------- + TOTAL 0.04246403 86.28025663 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 339752 +BPFP 1.1355 bits/point +EBPFP 2.2711 equivalent bits/point +MSE 86.280257 +---------------------- -------------------------------------------------------- +Time: 0.816s Load: 0.009s, Pack+Encode: 0.347s, Decode+Unpack: 0.461s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2803 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 374, 128) +Output shape: (1, 374, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.output: torch.Size([1, 374, 3584]) -> torch.Size([1, 1, 374, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,396B, BPFP=0.4761 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,288B, BPFP=1.9756 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,768B, BPFP=0.8676 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,996B, BPFP=1.9216 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,180B, BPFP=1.0938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,564B, BPFP=1.8618 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,540B, BPFP=1.0252 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,208B, BPFP=1.8887 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,588B, BPFP=1.5704 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,080B, BPFP=1.7998 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,500B, BPFP=0.5342 +⌛️ [2/4] FRONTEND: Frontend time: 0.375s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.508s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18781370 51.33947819 + layer.0.v_cache 0.00001670 0.00906359 + layer.1.k_cache 0.85217587 4.38798221 + layer.1.v_cache 0.00000605 0.00371419 + layer.2.k_cache 0.02919085 0.69540014 + layer.2.v_cache 0.00002159 0.00913448 + layer.3.k_cache 0.02976600 2.69515338 + layer.3.v_cache 0.00002022 0.01021184 + layer.4.k_cache 0.00072409 0.19220322 + layer.4.v_cache 0.00004949 0.01845314 + layer.4.output 0.03578218 144.10639085 + ------------------------------------------------------------------------------------- + TOTAL 0.07942705 62.82973708 + (elements=3,255,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3255296 +Total Bytes 436108 +BPFP 1.0718 bits/point +EBPFP 2.1435 equivalent bits/point +MSE 62.829737 +---------------------- -------------------------------------------------------- +Time: 0.895s Load: 0.012s, Pack+Encode: 0.375s, Decode+Unpack: 0.508s +---------------------- -------------------------------------------------------- +💾 Converting with 62.8297 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,588B, BPFP=0.4920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,172B, BPFP=2.0210 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,800B, BPFP=0.8947 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,460B, BPFP=1.9748 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,612B, BPFP=1.1419 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,724B, BPFP=1.9271 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,584B, BPFP=1.0752 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,148B, BPFP=1.9546 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,216B, BPFP=1.6349 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,900B, BPFP=1.8737 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,760B, BPFP=0.5813 +⌛️ [2/4] FRONTEND: Frontend time: 0.336s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.436s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13979621 51.88382148 + layer.0.v_cache 0.00001743 0.00907217 + layer.1.k_cache 0.46959303 4.60477178 + layer.1.v_cache 0.00000602 0.00380590 + layer.2.k_cache 0.01642122 0.68775813 + layer.2.v_cache 0.00002053 0.00974954 + layer.3.k_cache 0.03876302 3.11036017 + layer.3.v_cache 0.00002036 0.01090618 + layer.4.k_cache 0.00069112 0.20344977 + layer.4.v_cache 0.00005399 0.02022450 + layer.4.output 1.27033806 223.61642338 + ------------------------------------------------------------------------------------- + TOTAL 0.56222055 95.63875784 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 293964 +BPFP 1.1211 bits/point +EBPFP 2.2422 equivalent bits/point +MSE 95.638758 +---------------------- -------------------------------------------------------- +Time: 0.780s Load: 0.008s, Pack+Encode: 0.336s, Decode+Unpack: 0.436s +---------------------- -------------------------------------------------------- +💾 Converting with 95.6388 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,904B, BPFP=0.4752 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,204B, BPFP=2.0790 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,328B, BPFP=0.9174 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,348B, BPFP=2.0201 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,132B, BPFP=1.1792 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,752B, BPFP=1.9791 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,144B, BPFP=1.1112 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,224B, BPFP=2.0116 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,360B, BPFP=1.6768 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,984B, BPFP=1.9262 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,368B, BPFP=0.6231 +⌛️ [2/4] FRONTEND: Frontend time: 0.301s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.437s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13893990 53.44547598 + layer.0.v_cache 0.00001711 0.00971850 + layer.1.k_cache 0.36820191 4.68660464 + layer.1.v_cache 0.00000637 0.00404399 + layer.2.k_cache 0.01566778 0.71434660 + layer.2.v_cache 0.00002161 0.01055733 + layer.3.k_cache 0.01884081 3.06711770 + layer.3.v_cache 0.00002076 0.01167887 + layer.4.k_cache 0.00070054 0.21851665 + layer.4.v_cache 0.00005210 0.02109744 + layer.4.output 1.34866800 237.81948159 + ------------------------------------------------------------------------------------- + TOTAL 0.58724382 101.58385464 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 286748 +BPFP 1.1610 bits/point +EBPFP 2.3221 equivalent bits/point +MSE 101.583855 +---------------------- -------------------------------------------------------- +Time: 0.746s Load: 0.008s, Pack+Encode: 0.301s, Decode+Unpack: 0.437s +---------------------- -------------------------------------------------------- +💾 Converting with 101.5839 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,628B, BPFP=0.5049 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,340B, BPFP=2.0681 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,688B, BPFP=0.9181 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,464B, BPFP=2.0169 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,896B, BPFP=1.1643 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,756B, BPFP=1.9754 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,792B, BPFP=1.0997 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,368B, BPFP=2.0112 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,768B, BPFP=1.6835 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,920B, BPFP=1.9265 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,300B, BPFP=0.5877 +⌛️ [2/4] FRONTEND: Frontend time: 0.374s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.506s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11943251 52.97661005 + layer.0.v_cache 0.00001561 0.00931815 + layer.1.k_cache 0.49619702 4.46908558 + layer.1.v_cache 0.00000608 0.00383715 + layer.2.k_cache 0.01349212 0.75250953 + layer.2.v_cache 0.00002006 0.00973546 + layer.3.k_cache 0.01080775 2.68119601 + layer.3.v_cache 0.00002073 0.01052668 + layer.4.k_cache 0.00068886 0.20425465 + layer.4.v_cache 0.00007406 0.02041262 + layer.4.output 0.00494179 207.15400950 + ------------------------------------------------------------------------------------- + TOTAL 0.03972631 88.89503249 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 332920 +BPFP 1.1460 bits/point +EBPFP 2.2921 equivalent bits/point +MSE 88.895032 +---------------------- -------------------------------------------------------- +Time: 0.891s Load: 0.012s, Pack+Encode: 0.374s, Decode+Unpack: 0.506s +---------------------- -------------------------------------------------------- +💾 Converting with 88.8950 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,020B, BPFP=0.4963 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,324B, BPFP=2.0732 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,204B, BPFP=0.9335 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,660B, BPFP=2.0263 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,796B, BPFP=1.1875 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,968B, BPFP=1.9774 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,700B, BPFP=1.1100 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,532B, BPFP=2.0173 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,880B, BPFP=1.6883 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,276B, BPFP=1.9285 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 59,680B, BPFP=0.6028 +⌛️ [2/4] FRONTEND: Frontend time: 0.318s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.394s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11815372 54.49249240 + layer.0.v_cache 0.00001627 0.01018748 + layer.1.k_cache 0.37623358 4.59823954 + layer.1.v_cache 0.00000616 0.00414287 + layer.2.k_cache 0.01429399 0.75359534 + layer.2.v_cache 0.00002051 0.01057789 + layer.3.k_cache 0.01320818 3.25320863 + layer.3.v_cache 0.00002004 0.01193251 + layer.4.k_cache 0.00067885 0.21199768 + layer.4.v_cache 0.00005318 0.02191316 + layer.4.output 1.38524649 244.25997899 + ------------------------------------------------------------------------------------- + TOTAL 0.60114176 104.30518473 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 278040 +BPFP 1.1563 bits/point +EBPFP 2.3127 equivalent bits/point +MSE 104.305185 +---------------------- -------------------------------------------------------- +Time: 0.719s Load: 0.007s, Pack+Encode: 0.318s, Decode+Unpack: 0.394s +---------------------- -------------------------------------------------------- +💾 Converting with 104.3052 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,980B, BPFP=0.5047 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,436B, BPFP=2.0479 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,976B, BPFP=0.8979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,508B, BPFP=1.9957 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,484B, BPFP=1.1513 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,564B, BPFP=1.9427 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,180B, BPFP=1.0780 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,160B, BPFP=1.9762 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,560B, BPFP=1.6614 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,864B, BPFP=1.9033 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,480B, BPFP=0.5579 +⌛️ [2/4] FRONTEND: Frontend time: 0.330s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.520s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09820673 51.30902231 + layer.0.v_cache 0.00001666 0.00888859 + layer.1.k_cache 0.50544519 4.18520394 + layer.1.v_cache 0.00000613 0.00360636 + layer.2.k_cache 0.01930715 0.71202428 + layer.2.v_cache 0.00002033 0.00941388 + layer.3.k_cache 0.01337509 2.94819279 + layer.3.v_cache 0.00001987 0.01047948 + layer.4.k_cache 0.00074230 0.19725565 + layer.4.v_cache 0.00004869 0.01925781 + layer.4.output 0.00473525 198.99091084 + ------------------------------------------------------------------------------------- + TOTAL 0.03943146 85.43174829 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 339192 +BPFP 1.1214 bits/point +EBPFP 2.2429 equivalent bits/point +MSE 85.431748 +---------------------- -------------------------------------------------------- +Time: 0.860s Load: 0.010s, Pack+Encode: 0.330s, Decode+Unpack: 0.520s +---------------------- -------------------------------------------------------- +💾 Converting with 85.4317 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,948B, BPFP=0.4980 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,952B, BPFP=2.0751 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,748B, BPFP=0.9137 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,152B, BPFP=2.0178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,556B, BPFP=1.1866 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,472B, BPFP=1.9690 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,440B, BPFP=1.1067 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,224B, BPFP=2.0229 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,636B, BPFP=1.6941 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,788B, BPFP=1.9200 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 54,044B, BPFP=0.5534 +⌛️ [2/4] FRONTEND: Frontend time: 0.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.400s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13853801 55.05872814 + layer.0.v_cache 0.00001803 0.00991607 + layer.1.k_cache 0.23740530 4.36667290 + layer.1.v_cache 0.00000581 0.00401595 + layer.2.k_cache 0.01896443 0.81819986 + layer.2.v_cache 0.00001977 0.01049204 + layer.3.k_cache 0.00876656 2.91112889 + layer.3.v_cache 0.00002034 0.01220701 + layer.4.k_cache 0.00073522 0.21593659 + layer.4.v_cache 0.00004914 0.02160452 + layer.4.output 1.40427464 246.86867218 + ------------------------------------------------------------------------------------- + TOTAL 0.60202618 105.38291807 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 268960 +BPFP 1.1340 bits/point +EBPFP 2.2679 equivalent bits/point +MSE 105.382918 +---------------------- -------------------------------------------------------- +Time: 0.743s Load: 0.010s, Pack+Encode: 0.334s, Decode+Unpack: 0.400s +---------------------- -------------------------------------------------------- +💾 Converting with 105.3829 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,692B, BPFP=0.4921 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,128B, BPFP=2.0453 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,880B, BPFP=0.8990 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,088B, BPFP=1.9864 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,484B, BPFP=1.1596 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,384B, BPFP=1.9466 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,228B, BPFP=1.0885 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,132B, BPFP=1.9889 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,332B, BPFP=1.6606 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,652B, BPFP=1.9051 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,996B, BPFP=0.5904 +⌛️ [2/4] FRONTEND: Frontend time: 0.366s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.492s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10827675 51.22416355 + layer.0.v_cache 0.00001706 0.00897934 + layer.1.k_cache 0.56919524 4.28169450 + layer.1.v_cache 0.00000592 0.00358231 + layer.2.k_cache 0.01911232 0.71871335 + layer.2.v_cache 0.00002018 0.00998969 + layer.3.k_cache 0.01342056 3.01141910 + layer.3.v_cache 0.00002333 0.01114975 + layer.4.k_cache 0.00074072 0.20004358 + layer.4.v_cache 0.00005062 0.01973866 + layer.4.output 0.00480139 200.43455616 + ------------------------------------------------------------------------------------- + TOTAL 0.04379250 86.03125688 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 340996 +BPFP 1.1356 bits/point +EBPFP 2.2711 equivalent bits/point +MSE 86.031257 +---------------------- -------------------------------------------------------- +Time: 0.868s Load: 0.010s, Pack+Encode: 0.366s, Decode+Unpack: 0.492s +---------------------- -------------------------------------------------------- +💾 Converting with 86.0313 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,508B, BPFP=0.4979 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,088B, BPFP=2.1119 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,884B, BPFP=0.9295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,936B, BPFP=2.0445 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,164B, BPFP=1.1800 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,888B, BPFP=1.9831 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,960B, BPFP=1.1096 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,572B, BPFP=2.0232 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,696B, BPFP=1.6793 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,000B, BPFP=1.9312 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,028B, BPFP=0.6022 +⌛️ [2/4] FRONTEND: Frontend time: 0.359s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.529s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15963619 53.47804015 + layer.0.v_cache 0.00001721 0.00983182 + layer.1.k_cache 0.49929055 4.55619441 + layer.1.v_cache 0.00000624 0.00374407 + layer.2.k_cache 0.03270629 0.76283070 + layer.2.v_cache 0.00002062 0.00931840 + layer.3.k_cache 0.00946706 2.95883167 + layer.3.v_cache 0.00002056 0.01048278 + layer.4.k_cache 0.00068536 0.19657648 + layer.4.v_cache 0.00004941 0.01898893 + layer.4.output 0.00495045 207.14220506 + ------------------------------------------------------------------------------------- + TOTAL 0.04332662 88.94119264 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 336724 +BPFP 1.1591 bits/point +EBPFP 2.3183 equivalent bits/point +MSE 88.941193 +---------------------- -------------------------------------------------------- +Time: 0.896s Load: 0.009s, Pack+Encode: 0.359s, Decode+Unpack: 0.529s +---------------------- -------------------------------------------------------- +💾 Converting with 88.9412 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,524B, BPFP=0.4981 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,404B, BPFP=2.0792 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,736B, BPFP=0.9094 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,600B, BPFP=2.0260 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,580B, BPFP=1.1639 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,656B, BPFP=1.9635 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,840B, BPFP=1.1149 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,148B, BPFP=1.9960 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,364B, BPFP=1.6793 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,916B, BPFP=1.9145 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,496B, BPFP=0.5816 +⌛️ [2/4] FRONTEND: Frontend time: 0.326s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.471s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14874478 52.63192697 + layer.0.v_cache 0.00001714 0.00917719 + layer.1.k_cache 0.45078779 4.77397595 + layer.1.v_cache 0.00000615 0.00362289 + layer.2.k_cache 0.01724745 0.73570484 + layer.2.v_cache 0.00001987 0.00936952 + layer.3.k_cache 0.02849353 3.08809183 + layer.3.v_cache 0.00002056 0.01069955 + layer.4.k_cache 0.00066912 0.19806765 + layer.4.v_cache 0.00005068 0.01936675 + layer.4.output 1.29722614 228.67567721 + ------------------------------------------------------------------------------------- + TOTAL 0.57215530 97.77704374 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 293264 +BPFP 1.1421 bits/point +EBPFP 2.2843 equivalent bits/point +MSE 97.777044 +---------------------- -------------------------------------------------------- +Time: 0.804s Load: 0.007s, Pack+Encode: 0.326s, Decode+Unpack: 0.471s +---------------------- -------------------------------------------------------- +💾 Converting with 97.7770 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 341, 128) +Output shape: (1, 341, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.output: torch.Size([1, 341, 3584]) -> torch.Size([1, 1, 341, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,624B, BPFP=0.4868 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,692B, BPFP=2.0478 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,340B, BPFP=0.8862 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,548B, BPFP=1.9954 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,936B, BPFP=1.1426 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 42,240B, BPFP=1.9355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,772B, BPFP=1.0893 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,884B, BPFP=1.9650 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,612B, BPFP=1.6318 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 40,792B, BPFP=1.8691 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,260B, BPFP=0.5646 +⌛️ [2/4] FRONTEND: Frontend time: 0.417s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.540s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17009405 50.39502955 + layer.0.v_cache 0.00001631 0.00923669 + layer.1.k_cache 0.79816426 4.36029858 + layer.1.v_cache 0.00000611 0.00365720 + layer.2.k_cache 0.02878924 0.77377516 + layer.2.v_cache 0.00002016 0.00942803 + layer.3.k_cache 0.01035325 3.15123545 + layer.3.v_cache 0.00002082 0.01029017 + layer.4.k_cache 0.00072160 0.19512611 + layer.4.v_cache 0.00005837 0.01864002 + layer.4.output 0.03919861 158.46676005 + ------------------------------------------------------------------------------------- + TOTAL 0.07544909 68.71729631 + (elements=2,968,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2968064 +Total Bytes 414700 +BPFP 1.1178 bits/point +EBPFP 2.2355 equivalent bits/point +MSE 68.717296 +---------------------- -------------------------------------------------------- +Time: 0.969s Load: 0.012s, Pack+Encode: 0.417s, Decode+Unpack: 0.540s +---------------------- -------------------------------------------------------- +💾 Converting with 68.7173 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 335, 128) +Output shape: (1, 335, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.output: torch.Size([1, 335, 3584]) -> torch.Size([1, 1, 335, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,580B, BPFP=0.4935 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,072B, BPFP=2.0556 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,300B, BPFP=0.9002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 42,688B, BPFP=1.9910 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,376B, BPFP=1.1369 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 41,336B, BPFP=1.9280 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,120B, BPFP=1.0784 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 42,036B, BPFP=1.9606 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,144B, BPFP=1.6392 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 40,260B, BPFP=1.8778 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,304B, BPFP=0.5751 +⌛️ [2/4] FRONTEND: Frontend time: 0.361s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.545s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15195324 51.25745103 + layer.0.v_cache 0.00001666 0.00937727 + layer.1.k_cache 0.76788093 4.27199197 + layer.1.v_cache 0.00000624 0.00373923 + layer.2.k_cache 0.02686723 0.72116963 + layer.2.v_cache 0.00001960 0.00939717 + layer.3.k_cache 0.01080669 2.96162146 + layer.3.v_cache 0.00002024 0.01044880 + layer.4.k_cache 0.00071790 0.20021355 + layer.4.v_cache 0.00005106 0.01946162 + layer.4.output 3.38143948 159.80387793 + ------------------------------------------------------------------------------------- + TOTAL 1.44873036 69.29953043 + (elements=2,915,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2915840 +Total Bytes 409216 +BPFP 1.1227 bits/point +EBPFP 2.2455 equivalent bits/point +MSE 69.299530 +---------------------- -------------------------------------------------------- +Time: 0.917s Load: 0.010s, Pack+Encode: 0.361s, Decode+Unpack: 0.545s +---------------------- -------------------------------------------------------- +💾 Converting with 69.2995 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,392B, BPFP=0.4948 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,564B, BPFP=2.0969 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,852B, BPFP=0.9347 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,576B, BPFP=2.0387 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,880B, BPFP=1.1722 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,392B, BPFP=1.9689 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,844B, BPFP=1.1111 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,272B, BPFP=2.0208 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,224B, BPFP=1.6642 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,520B, BPFP=1.9175 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,432B, BPFP=0.6270 +⌛️ [2/4] FRONTEND: Frontend time: 0.339s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.477s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17469128 52.64248231 + layer.0.v_cache 0.00001775 0.00948512 + layer.1.k_cache 0.51265420 4.55313329 + layer.1.v_cache 0.00000620 0.00368079 + layer.2.k_cache 0.03428403 0.73511353 + layer.2.v_cache 0.00002024 0.00919249 + layer.3.k_cache 0.01443097 3.03795281 + layer.3.v_cache 0.00002182 0.01110230 + layer.4.k_cache 0.00070119 0.19127740 + layer.4.v_cache 0.00005097 0.01848934 + layer.4.output 0.00503371 208.55625000 + ------------------------------------------------------------------------------------- + TOTAL 0.04541851 89.47680349 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 335948 +BPFP 1.1652 bits/point +EBPFP 2.3304 equivalent bits/point +MSE 89.476803 +---------------------- -------------------------------------------------------- +Time: 0.825s Load: 0.009s, Pack+Encode: 0.339s, Decode+Unpack: 0.477s +---------------------- -------------------------------------------------------- +💾 Converting with 89.4768 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,472B, BPFP=0.4995 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,296B, BPFP=2.0811 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,668B, BPFP=0.9238 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,224B, BPFP=2.0179 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,744B, BPFP=1.1642 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,116B, BPFP=1.9526 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,728B, BPFP=1.1042 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,912B, BPFP=1.9995 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,456B, BPFP=1.6778 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,520B, BPFP=1.9175 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,556B, BPFP=0.5859 +⌛️ [2/4] FRONTEND: Frontend time: 0.331s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.460s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18043976 52.40495283 + layer.0.v_cache 0.00001814 0.00928075 + layer.1.k_cache 0.58196952 4.54366432 + layer.1.v_cache 0.00000608 0.00388880 + layer.2.k_cache 0.02409397 0.73043155 + layer.2.v_cache 0.00002020 0.00939784 + layer.3.k_cache 0.01875559 2.81540573 + layer.3.v_cache 0.00002085 0.01077707 + layer.4.k_cache 0.00069073 0.20473067 + layer.4.v_cache 0.00005389 0.02085445 + layer.4.output 0.00498411 208.52764488 + ------------------------------------------------------------------------------------- + TOTAL 0.04946809 89.43805283 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 329692 +BPFP 1.1435 bits/point +EBPFP 2.2870 equivalent bits/point +MSE 89.438053 +---------------------- -------------------------------------------------------- +Time: 0.800s Load: 0.009s, Pack+Encode: 0.331s, Decode+Unpack: 0.460s +---------------------- -------------------------------------------------------- +💾 Converting with 89.4381 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 386, 128) +Output shape: (1, 386, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.output: torch.Size([1, 386, 3584]) -> torch.Size([1, 1, 386, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,868B, BPFP=0.4804 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 50,784B, BPFP=2.0557 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,048B, BPFP=0.8925 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 49,320B, BPFP=1.9964 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,900B, BPFP=1.1699 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 48,360B, BPFP=1.9576 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,176B, BPFP=1.1001 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 49,184B, BPFP=1.9909 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,820B, BPFP=1.6524 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,936B, BPFP=1.8999 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,800B, BPFP=0.5887 +⌛️ [2/4] FRONTEND: Frontend time: 0.415s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.614s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16808204 51.46043657 + layer.0.v_cache 0.00001827 0.00979541 + layer.1.k_cache 0.89383947 4.31603213 + layer.1.v_cache 0.00000605 0.00382752 + layer.2.k_cache 0.02643975 0.78682491 + layer.2.v_cache 0.00002094 0.01019556 + layer.3.k_cache 0.04368651 2.90752193 + layer.3.v_cache 0.00002149 0.01132745 + layer.4.k_cache 0.00075087 0.20808294 + layer.4.v_cache 0.00005208 0.02053426 + layer.4.output 0.03469837 140.09048853 + ------------------------------------------------------------------------------------- + TOTAL 0.08092977 61.19811755 + (elements=3,359,744) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3359744 +Total Bytes 477196 +BPFP 1.1363 bits/point +EBPFP 2.2725 equivalent bits/point +MSE 61.198118 +---------------------- -------------------------------------------------------- +Time: 1.041s Load: 0.012s, Pack+Encode: 0.415s, Decode+Unpack: 0.614s +---------------------- -------------------------------------------------------- +💾 Converting with 61.1981 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,644B, BPFP=0.4690 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,808B, BPFP=2.0512 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,660B, BPFP=0.9039 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,976B, BPFP=2.0061 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,464B, BPFP=1.1645 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,988B, BPFP=1.9525 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,112B, BPFP=1.0911 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,388B, BPFP=1.9742 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,672B, BPFP=1.6641 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,956B, BPFP=1.8965 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 77,764B, BPFP=0.6027 +⌛️ [2/4] FRONTEND: Frontend time: 0.350s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.520s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14364293 52.24926080 + layer.0.v_cache 0.00001600 0.00989602 + layer.1.k_cache 0.62366660 4.30236562 + layer.1.v_cache 0.00000628 0.00408342 + layer.2.k_cache 0.01981349 0.79492214 + layer.2.v_cache 0.00002090 0.01046415 + layer.3.k_cache 0.02158341 3.09652562 + layer.3.v_cache 0.00002097 0.01137309 + layer.4.k_cache 0.00071696 0.20999025 + layer.4.v_cache 0.00005883 0.02111212 + layer.4.output 0.00463133 191.59311446 + ------------------------------------------------------------------------------------- + TOTAL 0.04952739 82.46245850 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 357432 +BPFP 1.1407 bits/point +EBPFP 2.2814 equivalent bits/point +MSE 82.462458 +---------------------- -------------------------------------------------------- +Time: 0.878s Load: 0.009s, Pack+Encode: 0.350s, Decode+Unpack: 0.520s +---------------------- -------------------------------------------------------- +💾 Converting with 82.4625 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,756B, BPFP=0.4939 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,392B, BPFP=2.0528 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,932B, BPFP=0.8987 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,408B, BPFP=1.9973 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,340B, BPFP=1.1473 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,656B, BPFP=1.9549 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,340B, BPFP=1.0909 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,116B, BPFP=1.9808 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,208B, BPFP=1.6476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,624B, BPFP=1.8967 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,700B, BPFP=0.6342 +⌛️ [2/4] FRONTEND: Frontend time: 0.338s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.487s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12300906 50.36665162 + layer.0.v_cache 0.00001720 0.00832158 + layer.1.k_cache 0.58060381 4.51648081 + layer.1.v_cache 0.00000644 0.00313488 + layer.2.k_cache 0.02305070 0.73962424 + layer.2.v_cache 0.00002053 0.00888617 + layer.3.k_cache 0.05279791 2.87554469 + layer.3.v_cache 0.00002021 0.01006001 + layer.4.k_cache 0.00074809 0.18909578 + layer.4.v_cache 0.00005348 0.01907533 + layer.4.output 0.00483522 199.76213577 + ------------------------------------------------------------------------------------- + TOTAL 0.04789259 85.71010738 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 347472 +BPFP 1.1530 bits/point +EBPFP 2.3059 equivalent bits/point +MSE 85.710107 +---------------------- -------------------------------------------------------- +Time: 0.834s Load: 0.009s, Pack+Encode: 0.338s, Decode+Unpack: 0.487s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7101 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,836B, BPFP=0.4790 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,932B, BPFP=2.0973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,396B, BPFP=0.9386 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,088B, BPFP=2.0381 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,300B, BPFP=1.2122 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,396B, BPFP=1.9896 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,160B, BPFP=1.1323 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,740B, BPFP=2.0137 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,356B, BPFP=1.7066 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,564B, BPFP=1.9313 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,960B, BPFP=0.6202 +⌛️ [2/4] FRONTEND: Frontend time: 0.317s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.402s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16177035 52.31515380 + layer.0.v_cache 0.00001782 0.01026476 + layer.1.k_cache 0.31952496 4.80742505 + layer.1.v_cache 0.00000626 0.00385569 + layer.2.k_cache 0.01271096 0.77503413 + layer.2.v_cache 0.00002040 0.01020847 + layer.3.k_cache 0.01363577 3.03585200 + layer.3.v_cache 0.00002033 0.01144660 + layer.4.k_cache 0.00066992 0.20414968 + layer.4.v_cache 0.00005030 0.01974758 + layer.4.output 1.37284199 241.29656470 + ------------------------------------------------------------------------------------- + TOTAL 0.59519535 102.95700533 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 283728 +BPFP 1.1694 bits/point +EBPFP 2.3388 equivalent bits/point +MSE 102.957005 +---------------------- -------------------------------------------------------- +Time: 0.726s Load: 0.007s, Pack+Encode: 0.317s, Decode+Unpack: 0.402s +---------------------- -------------------------------------------------------- +💾 Converting with 102.9570 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,824B, BPFP=0.4973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,932B, BPFP=2.0434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,748B, BPFP=0.9177 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,316B, BPFP=1.9908 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,024B, BPFP=1.1974 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,976B, BPFP=1.9617 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,220B, BPFP=1.1288 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,412B, BPFP=1.9990 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,704B, BPFP=1.6824 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,332B, BPFP=1.9068 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 52,264B, BPFP=0.6375 +⌛️ [2/4] FRONTEND: Frontend time: 0.348s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.334s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09962382 53.65008111 + layer.0.v_cache 0.00001686 0.01025018 + layer.1.k_cache 0.10988800 4.73201431 + layer.1.v_cache 0.00000590 0.00400458 + layer.2.k_cache 0.01214521 0.67357377 + layer.2.v_cache 0.00002216 0.01078292 + layer.3.k_cache 0.02260447 3.33779123 + layer.3.v_cache 0.00002006 0.01212924 + layer.4.k_cache 0.00069723 0.21713657 + layer.4.v_cache 0.00005242 0.02172091 + layer.4.output 0.00882482 298.09562842 + ------------------------------------------------------------------------------------- + TOTAL 0.01804999 126.43169904 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 231752 +BPFP 1.1640 bits/point +EBPFP 2.3279 equivalent bits/point +MSE 126.431699 +---------------------- -------------------------------------------------------- +Time: 0.689s Load: 0.008s, Pack+Encode: 0.348s, Decode+Unpack: 0.334s +---------------------- -------------------------------------------------------- +💾 Converting with 126.4317 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,004B, BPFP=0.5007 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,936B, BPFP=2.0538 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,216B, BPFP=0.9017 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,828B, BPFP=1.9922 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,560B, BPFP=1.1432 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,720B, BPFP=1.9306 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,624B, BPFP=1.0912 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,424B, BPFP=1.9698 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,652B, BPFP=1.6488 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,884B, BPFP=1.8841 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,164B, BPFP=0.5971 +⌛️ [2/4] FRONTEND: Frontend time: 0.339s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.465s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11148910 53.14786407 + layer.0.v_cache 0.00001684 0.00928374 + layer.1.k_cache 0.60671574 4.39644224 + layer.1.v_cache 0.00000630 0.00364737 + layer.2.k_cache 0.04149003 0.72985796 + layer.2.v_cache 0.00001976 0.00921684 + layer.3.k_cache 0.02200458 2.78156452 + layer.3.v_cache 0.00001954 0.01024486 + layer.4.k_cache 0.00078895 0.19729026 + layer.4.v_cache 0.00004864 0.01868022 + layer.4.output 0.00472744 195.92545755 + ------------------------------------------------------------------------------------- + TOTAL 0.04798185 84.28131147 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 347012 +BPFP 1.1350 bits/point +EBPFP 2.2701 equivalent bits/point +MSE 84.281311 +---------------------- -------------------------------------------------------- +Time: 0.813s Load: 0.009s, Pack+Encode: 0.339s, Decode+Unpack: 0.465s +---------------------- -------------------------------------------------------- +💾 Converting with 84.2813 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,400B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,236B, BPFP=2.1278 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,412B, BPFP=0.9697 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,640B, BPFP=2.0812 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,628B, BPFP=1.2209 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,088B, BPFP=2.0381 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,668B, BPFP=1.1459 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,480B, BPFP=2.0688 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,332B, BPFP=1.7447 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,464B, BPFP=1.9894 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 53,644B, BPFP=0.5987 +⌛️ [2/4] FRONTEND: Frontend time: 0.321s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.461s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14565937 53.38940430 + layer.0.v_cache 0.00001550 0.00947690 + layer.1.k_cache 0.22559690 4.87157776 + layer.1.v_cache 0.00000590 0.00399014 + layer.2.k_cache 0.00679684 0.70883415 + layer.2.v_cache 0.00002314 0.01098266 + layer.3.k_cache 0.06866968 3.17413269 + layer.3.v_cache 0.00002119 0.01179001 + layer.4.k_cache 0.00065507 0.20805540 + layer.4.v_cache 0.00005191 0.02146446 + layer.4.output 1.53062134 269.98147321 + ------------------------------------------------------------------------------------- + TOTAL 0.65657911 114.84000123 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 256992 +BPFP 1.1810 bits/point +EBPFP 2.3621 equivalent bits/point +MSE 114.840001 +---------------------- -------------------------------------------------------- +Time: 0.789s Load: 0.007s, Pack+Encode: 0.321s, Decode+Unpack: 0.461s +---------------------- -------------------------------------------------------- +💾 Converting with 114.8400 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,484B, BPFP=0.5040 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,424B, BPFP=2.1318 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,376B, BPFP=0.9621 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,644B, BPFP=2.0712 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,688B, BPFP=1.2195 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,952B, BPFP=2.0174 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,616B, BPFP=1.1362 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,544B, BPFP=2.0634 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,240B, BPFP=1.7289 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,364B, BPFP=1.9717 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 53,940B, BPFP=0.5990 +⌛️ [2/4] FRONTEND: Frontend time: 0.292s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.425s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10663509 53.59976971 + layer.0.v_cache 0.00001613 0.01016584 + layer.1.k_cache 0.26620468 4.63086120 + layer.1.v_cache 0.00000602 0.00430373 + layer.2.k_cache 0.00649704 0.72677407 + layer.2.v_cache 0.00002002 0.01122467 + layer.3.k_cache 0.02565914 3.13663919 + layer.3.v_cache 0.00002186 0.01235919 + layer.4.k_cache 0.00070101 0.22635358 + layer.4.v_cache 0.00005421 0.02327554 + layer.4.output 1.52303164 268.32544865 + ------------------------------------------------------------------------------------- + TOTAL 0.65100216 114.15646278 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 257272 +BPFP 1.1764 bits/point +EBPFP 2.3529 equivalent bits/point +MSE 114.156463 +---------------------- -------------------------------------------------------- +Time: 0.723s Load: 0.007s, Pack+Encode: 0.292s, Decode+Unpack: 0.425s +---------------------- -------------------------------------------------------- +💾 Converting with 114.1565 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,544B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,704B, BPFP=2.0894 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,960B, BPFP=0.9340 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,828B, BPFP=2.0382 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,548B, BPFP=1.2025 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,056B, BPFP=1.9930 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,540B, BPFP=1.1435 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,712B, BPFP=2.0314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,012B, BPFP=1.6978 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,116B, BPFP=1.9380 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,204B, BPFP=0.6287 +⌛️ [2/4] FRONTEND: Frontend time: 0.338s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.537s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258070 51.81568937 + layer.0.v_cache 0.00001713 0.01030587 + layer.1.k_cache 0.49703808 4.96456989 + layer.1.v_cache 0.00000670 0.00421106 + layer.2.k_cache 0.01458295 0.74764917 + layer.2.v_cache 0.00002181 0.01042486 + layer.3.k_cache 0.02940617 3.09679079 + layer.3.v_cache 0.00002120 0.01174993 + layer.4.k_cache 0.00070157 0.21216115 + layer.4.v_cache 0.00005095 0.02067717 + layer.4.output 0.00497170 207.15039794 + ------------------------------------------------------------------------------------- + TOTAL 0.04171936 88.87923617 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 341224 +BPFP 1.1746 bits/point +EBPFP 2.3493 equivalent bits/point +MSE 88.879236 +---------------------- -------------------------------------------------------- +Time: 0.884s Load: 0.009s, Pack+Encode: 0.338s, Decode+Unpack: 0.537s +---------------------- -------------------------------------------------------- +💾 Converting with 88.8792 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,892B, BPFP=0.4765 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,920B, BPFP=2.0686 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,248B, BPFP=0.9159 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,104B, BPFP=2.0122 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,208B, BPFP=1.1897 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,508B, BPFP=1.9710 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,240B, BPFP=1.1228 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,168B, BPFP=2.0166 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,352B, BPFP=1.6836 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,724B, BPFP=1.9168 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 60,752B, BPFP=0.6000 +⌛️ [2/4] FRONTEND: Frontend time: 0.313s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.415s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15433978 54.02264242 + layer.0.v_cache 0.00001575 0.00994124 + layer.1.k_cache 0.29362690 4.76626074 + layer.1.v_cache 0.00000607 0.00412194 + layer.2.k_cache 0.00940348 0.77033808 + layer.2.v_cache 0.00002159 0.01114423 + layer.3.k_cache 0.02320663 3.32349605 + layer.3.v_cache 0.00002067 0.01228069 + layer.4.k_cache 0.00067785 0.21970416 + layer.4.v_cache 0.00005016 0.02187226 + layer.4.output 1.35461534 239.05937895 + ------------------------------------------------------------------------------------- + TOTAL 0.58609860 102.15161497 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 283116 +BPFP 1.1514 bits/point +EBPFP 2.3028 equivalent bits/point +MSE 102.151615 +---------------------- -------------------------------------------------------- +Time: 0.737s Load: 0.008s, Pack+Encode: 0.313s, Decode+Unpack: 0.415s +---------------------- -------------------------------------------------------- +💾 Converting with 102.1516 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,964B, BPFP=0.4914 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,568B, BPFP=2.0596 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,836B, BPFP=0.9230 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,496B, BPFP=2.0009 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,312B, BPFP=1.1684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,536B, BPFP=1.9482 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,992B, BPFP=1.0961 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,224B, BPFP=1.9860 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,072B, BPFP=1.6487 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,444B, BPFP=1.8884 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 82,252B, BPFP=0.6442 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.494s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14751465 51.83189419 + layer.0.v_cache 0.00001719 0.00933599 + layer.1.k_cache 0.52245269 4.45433371 + layer.1.v_cache 0.00000618 0.00352710 + layer.2.k_cache 0.02707210 0.71736252 + layer.2.v_cache 0.00002227 0.00934072 + layer.3.k_cache 0.03364062 2.87764850 + layer.3.v_cache 0.00002083 0.01067022 + layer.4.k_cache 0.00069748 0.19579804 + layer.4.v_cache 0.00005305 0.01876126 + layer.4.output 0.00471531 194.17142857 + ------------------------------------------------------------------------------------- + TOTAL 0.04497084 83.48992190 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 359696 +BPFP 1.1600 bits/point +EBPFP 2.3200 equivalent bits/point +MSE 83.489922 +---------------------- -------------------------------------------------------- +Time: 0.845s Load: 0.010s, Pack+Encode: 0.342s, Decode+Unpack: 0.494s +---------------------- -------------------------------------------------------- +💾 Converting with 83.4899 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,088B, BPFP=0.4956 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,632B, BPFP=1.9382 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,632B, BPFP=0.8966 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,880B, BPFP=1.8922 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,136B, BPFP=1.1113 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,896B, BPFP=1.8319 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,276B, BPFP=1.0586 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,264B, BPFP=1.8544 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,656B, BPFP=1.5721 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,080B, BPFP=1.7819 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,824B, BPFP=0.6550 +⌛️ [2/4] FRONTEND: Frontend time: 0.299s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.399s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17432048 49.37187117 + layer.0.v_cache 0.00001704 0.00884626 + layer.1.k_cache 0.39057378 3.96936179 + layer.1.v_cache 0.00000668 0.00364894 + layer.2.k_cache 0.01299327 0.73701022 + layer.2.v_cache 0.00002114 0.00900788 + layer.3.k_cache 0.02499930 2.69982264 + layer.3.v_cache 0.00002096 0.01026265 + layer.4.k_cache 0.00069121 0.19408348 + layer.4.v_cache 0.00005416 0.01864366 + layer.4.output 1.20070616 198.54504552 + ------------------------------------------------------------------------------------- + TOTAL 0.52992007 85.10811043 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 310364 +BPFP 1.1187 bits/point +EBPFP 2.2373 equivalent bits/point +MSE 85.108110 +---------------------- -------------------------------------------------------- +Time: 0.707s Load: 0.008s, Pack+Encode: 0.299s, Decode+Unpack: 0.399s +---------------------- -------------------------------------------------------- +💾 Converting with 85.1081 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,044B, BPFP=0.4806 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,476B, BPFP=2.0794 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,480B, BPFP=0.9198 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,640B, BPFP=2.0224 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,876B, BPFP=1.1515 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,780B, BPFP=1.9637 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,948B, BPFP=1.0882 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,348B, BPFP=2.0025 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,668B, BPFP=1.6831 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,220B, BPFP=1.9255 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,128B, BPFP=0.6251 +⌛️ [2/4] FRONTEND: Frontend time: 0.282s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.400s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16239141 52.72604224 + layer.0.v_cache 0.00001645 0.00960750 + layer.1.k_cache 0.33751179 4.22513337 + layer.1.v_cache 0.00000668 0.00397847 + layer.2.k_cache 0.00660169 0.74691759 + layer.2.v_cache 0.00002045 0.00981481 + layer.3.k_cache 0.02213682 3.05655747 + layer.3.v_cache 0.00002049 0.01150548 + layer.4.k_cache 0.00067824 0.20898414 + layer.4.v_cache 0.00005183 0.02041829 + layer.4.output 1.33690934 235.77584997 + ------------------------------------------------------------------------------------- + TOTAL 0.58163537 100.67352407 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 288608 +BPFP 1.1584 bits/point +EBPFP 2.3167 equivalent bits/point +MSE 100.673524 +---------------------- -------------------------------------------------------- +Time: 0.690s Load: 0.008s, Pack+Encode: 0.282s, Decode+Unpack: 0.400s +---------------------- -------------------------------------------------------- +💾 Converting with 100.6735 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,652B, BPFP=0.5044 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,380B, BPFP=2.0627 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,736B, BPFP=0.9174 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,660B, BPFP=2.0208 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,648B, BPFP=1.1455 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,316B, BPFP=1.9424 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,320B, BPFP=1.0681 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,792B, BPFP=1.9701 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,676B, BPFP=1.6719 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,548B, BPFP=1.8976 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,872B, BPFP=0.5570 +⌛️ [2/4] FRONTEND: Frontend time: 0.344s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.477s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15796844 53.54546846 + layer.0.v_cache 0.00001527 0.00878248 + layer.1.k_cache 0.50235566 4.51272264 + layer.1.v_cache 0.00000626 0.00392806 + layer.2.k_cache 0.02709456 0.73572916 + layer.2.v_cache 0.00002063 0.00920071 + layer.3.k_cache 0.00856834 2.95966237 + layer.3.v_cache 0.00001945 0.01043110 + layer.4.k_cache 0.00071483 0.19957605 + layer.4.v_cache 0.00005464 0.01979807 + layer.4.output 0.00491706 206.35799240 + ------------------------------------------------------------------------------------- + TOTAL 0.04301397 88.61830858 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 327600 +BPFP 1.1235 bits/point +EBPFP 2.2470 equivalent bits/point +MSE 88.618309 +---------------------- -------------------------------------------------------- +Time: 0.829s Load: 0.008s, Pack+Encode: 0.344s, Decode+Unpack: 0.477s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6183 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,744B, BPFP=0.4879 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,016B, BPFP=2.0171 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,300B, BPFP=0.9010 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,272B, BPFP=1.9703 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,240B, BPFP=1.1492 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,344B, BPFP=1.9118 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,352B, BPFP=1.0932 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,832B, BPFP=1.9425 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,936B, BPFP=1.6341 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,460B, BPFP=1.8561 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,224B, BPFP=0.6411 +⌛️ [2/4] FRONTEND: Frontend time: 0.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13184587 53.69868715 + layer.0.v_cache 0.00001706 0.00963315 + layer.1.k_cache 0.45546043 4.65022179 + layer.1.v_cache 0.00000678 0.00388201 + layer.2.k_cache 0.02272066 0.72282065 + layer.2.v_cache 0.00001996 0.00947300 + layer.3.k_cache 0.01643563 2.99111102 + layer.3.v_cache 0.00002069 0.01101019 + layer.4.k_cache 0.00071781 0.20205676 + layer.4.v_cache 0.00005262 0.01940232 + layer.4.output 1.23455937 217.12928427 + ------------------------------------------------------------------------------------- + TOTAL 0.54524783 93.07195812 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 308720 +BPFP 1.1442 bits/point +EBPFP 2.2883 equivalent bits/point +MSE 93.071958 +---------------------- -------------------------------------------------------- +Time: 0.706s Load: 0.008s, Pack+Encode: 0.309s, Decode+Unpack: 0.390s +---------------------- -------------------------------------------------------- +💾 Converting with 93.0720 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,292B, BPFP=0.4908 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,004B, BPFP=2.0717 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,704B, BPFP=0.9295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,144B, BPFP=2.0208 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,704B, BPFP=1.1662 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,204B, BPFP=1.9652 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,924B, BPFP=1.1200 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,788B, BPFP=1.9998 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,584B, BPFP=1.6918 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,440B, BPFP=1.9200 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,660B, BPFP=0.5805 +⌛️ [2/4] FRONTEND: Frontend time: 0.359s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.478s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17705467 53.82248387 + layer.0.v_cache 0.00001633 0.00883175 + layer.1.k_cache 0.50406676 4.48575615 + layer.1.v_cache 0.00000599 0.00355767 + layer.2.k_cache 0.00944760 0.69767790 + layer.2.v_cache 0.00001883 0.00909989 + layer.3.k_cache 0.04120483 3.01171644 + layer.3.v_cache 0.00002392 0.01078360 + layer.4.k_cache 0.00070100 0.19849449 + layer.4.v_cache 0.00005002 0.01870260 + layer.4.output 0.00499618 209.53654288 + ------------------------------------------------------------------------------------- + TOTAL 0.04515078 89.94252380 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 328448 +BPFP 1.1435 bits/point +EBPFP 2.2870 equivalent bits/point +MSE 89.942524 +---------------------- -------------------------------------------------------- +Time: 0.846s Load: 0.009s, Pack+Encode: 0.359s, Decode+Unpack: 0.478s +---------------------- -------------------------------------------------------- +💾 Converting with 89.9425 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,512B, BPFP=0.4926 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,720B, BPFP=2.0671 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,872B, BPFP=0.9185 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,804B, BPFP=2.0141 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,032B, BPFP=1.1593 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,864B, BPFP=1.9597 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,988B, BPFP=1.0988 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,400B, BPFP=1.9907 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,016B, BPFP=1.6792 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,964B, BPFP=1.9076 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,820B, BPFP=0.5524 +⌛️ [2/4] FRONTEND: Frontend time: 0.369s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.525s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20501094 53.77290582 + layer.0.v_cache 0.00001516 0.00893767 + layer.1.k_cache 0.57085758 4.43100586 + layer.1.v_cache 0.00000574 0.00371071 + layer.2.k_cache 0.01259345 0.73304392 + layer.2.v_cache 0.00002099 0.00964003 + layer.3.k_cache 0.01648921 3.11357784 + layer.3.v_cache 0.00001970 0.01054581 + layer.4.k_cache 0.00070444 0.19844421 + layer.4.v_cache 0.00005121 0.01997543 + layer.4.output 0.00487421 204.88063823 + ------------------------------------------------------------------------------------- + TOTAL 0.04940517 88.02742676 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 330992 +BPFP 1.1267 bits/point +EBPFP 2.2535 equivalent bits/point +MSE 88.027427 +---------------------- -------------------------------------------------------- +Time: 0.903s Load: 0.009s, Pack+Encode: 0.369s, Decode+Unpack: 0.525s +---------------------- -------------------------------------------------------- +💾 Converting with 88.0274 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,580B, BPFP=0.4947 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,156B, BPFP=2.0846 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,900B, BPFP=0.9167 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,084B, BPFP=2.0228 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,316B, BPFP=1.1714 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,220B, BPFP=1.9730 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,280B, BPFP=1.1116 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,900B, BPFP=2.0122 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,176B, BPFP=1.6822 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,308B, BPFP=1.9204 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,800B, BPFP=0.6161 +⌛️ [2/4] FRONTEND: Frontend time: 0.331s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.481s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14466163 52.37986479 + layer.0.v_cache 0.00001631 0.00936771 + layer.1.k_cache 0.53789613 4.31203514 + layer.1.v_cache 0.00000653 0.00388440 + layer.2.k_cache 0.02095415 0.75044025 + layer.2.v_cache 0.00002169 0.01010494 + layer.3.k_cache 0.03083320 3.03211260 + layer.3.v_cache 0.00002083 0.01097286 + layer.4.k_cache 0.00069737 0.20387320 + layer.4.v_cache 0.00004943 0.02011075 + layer.4.output 0.00490546 204.06907288 + ------------------------------------------------------------------------------------- + TOTAL 0.04526444 87.60095746 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 341720 +BPFP 1.1590 bits/point +EBPFP 2.3179 equivalent bits/point +MSE 87.600957 +---------------------- -------------------------------------------------------- +Time: 0.821s Load: 0.009s, Pack+Encode: 0.331s, Decode+Unpack: 0.481s +---------------------- -------------------------------------------------------- +💾 Converting with 87.6010 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,260B, BPFP=0.4839 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,616B, BPFP=2.0180 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,000B, BPFP=0.8884 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,528B, BPFP=1.9611 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,292B, BPFP=1.1127 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,368B, BPFP=1.9005 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,040B, BPFP=1.0472 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,028B, BPFP=1.9350 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,928B, BPFP=1.6162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,560B, BPFP=1.8583 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,168B, BPFP=0.5836 +⌛️ [2/4] FRONTEND: Frontend time: 0.396s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.520s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14240686 52.26071932 + layer.0.v_cache 0.00001673 0.00841407 + layer.1.k_cache 0.61438330 4.22402148 + layer.1.v_cache 0.00000638 0.00345908 + layer.2.k_cache 0.02027981 0.71774328 + layer.2.v_cache 0.00002012 0.00896625 + layer.3.k_cache 0.02199794 3.00828527 + layer.3.v_cache 0.00002003 0.00974749 + layer.4.k_cache 0.00078937 0.19034007 + layer.4.v_cache 0.00005038 0.01802447 + layer.4.output 0.04464151 180.69636885 + ------------------------------------------------------------------------------------- + TOTAL 0.06543891 77.96025310 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 361788 +BPFP 1.1121 bits/point +EBPFP 2.2243 equivalent bits/point +MSE 77.960253 +---------------------- -------------------------------------------------------- +Time: 0.928s Load: 0.011s, Pack+Encode: 0.396s, Decode+Unpack: 0.520s +---------------------- -------------------------------------------------------- +💾 Converting with 77.9603 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 399, 128) +Output shape: (1, 399, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.output: torch.Size([1, 399, 3584]) -> torch.Size([1, 1, 399, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,276B, BPFP=0.4807 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 51,544B, BPFP=2.0185 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,548B, BPFP=0.8830 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,092B, BPFP=1.9616 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,160B, BPFP=1.1419 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 49,240B, BPFP=1.9283 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,448B, BPFP=1.0749 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 49,968B, BPFP=1.9568 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,616B, BPFP=1.6297 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 47,848B, BPFP=1.8737 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 103,420B, BPFP=0.5786 +⌛️ [2/4] FRONTEND: Frontend time: 0.442s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.581s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17889012 50.58025924 + layer.0.v_cache 0.00001635 0.00875044 + layer.1.k_cache 0.94984295 4.06320489 + layer.1.v_cache 0.00000590 0.00343653 + layer.2.k_cache 0.02305104 0.76113219 + layer.2.v_cache 0.00002114 0.00963679 + layer.3.k_cache 0.01620252 2.83739745 + layer.3.v_cache 0.00002049 0.01043827 + layer.4.k_cache 0.00073175 0.19382867 + layer.4.v_cache 0.00005255 0.01881009 + layer.4.output 0.00656733 138.41871420 + ------------------------------------------------------------------------------------- + TOTAL 0.07145918 60.43634670 + (elements=3,472,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3472896 +Total Bytes 485160 +BPFP 1.1176 bits/point +EBPFP 2.2352 equivalent bits/point +MSE 60.436347 +---------------------- -------------------------------------------------------- +Time: 1.037s Load: 0.013s, Pack+Encode: 0.442s, Decode+Unpack: 0.581s +---------------------- -------------------------------------------------------- +💾 Converting with 60.4363 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,792B, BPFP=0.4696 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,928B, BPFP=2.0691 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,220B, BPFP=0.9140 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,916B, BPFP=1.9992 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,096B, BPFP=1.1820 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,584B, BPFP=1.9762 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,908B, BPFP=1.0998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,168B, BPFP=2.0166 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,140B, BPFP=1.6690 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,836B, BPFP=1.9245 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,092B, BPFP=0.6231 +⌛️ [2/4] FRONTEND: Frontend time: 0.297s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.419s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10670873 53.38018529 + layer.0.v_cache 0.00001672 0.01022248 + layer.1.k_cache 0.29393762 4.53967825 + layer.1.v_cache 0.00000627 0.00414740 + layer.2.k_cache 0.01243114 0.75065896 + layer.2.v_cache 0.00002105 0.01083754 + layer.3.k_cache 0.00919804 3.10785209 + layer.3.v_cache 0.00002002 0.01203742 + layer.4.k_cache 0.00070102 0.21023440 + layer.4.v_cache 0.00005024 0.02221576 + layer.4.output 1.35463108 239.00517541 + ------------------------------------------------------------------------------------- + TOTAL 0.58267697 102.06378220 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 284680 +BPFP 1.1578 bits/point +EBPFP 2.3155 equivalent bits/point +MSE 102.063782 +---------------------- -------------------------------------------------------- +Time: 0.724s Load: 0.007s, Pack+Encode: 0.297s, Decode+Unpack: 0.419s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0638 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,392B, BPFP=0.5162 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,836B, BPFP=1.9584 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,788B, BPFP=0.9097 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,848B, BPFP=1.8976 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,720B, BPFP=1.1516 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,292B, BPFP=1.8634 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,716B, BPFP=1.0898 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,800B, BPFP=1.8947 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,884B, BPFP=1.5923 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,532B, BPFP=1.8167 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,036B, BPFP=0.6067 +⌛️ [2/4] FRONTEND: Frontend time: 0.302s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633087 51.88574219 + layer.0.v_cache 0.00001891 0.00971072 + layer.1.k_cache 0.36012592 4.13206470 + layer.1.v_cache 0.00000581 0.00358043 + layer.2.k_cache 0.01407751 0.74496953 + layer.2.v_cache 0.00002023 0.00958544 + layer.3.k_cache 0.01230825 3.02335436 + layer.3.v_cache 0.00002000 0.01080517 + layer.4.k_cache 0.00071033 0.19553531 + layer.4.v_cache 0.00004856 0.01918305 + layer.4.output 1.20534739 210.66275309 + ------------------------------------------------------------------------------------- + TOTAL 0.52535871 90.27492956 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 307844 +BPFP 1.1140 bits/point +EBPFP 2.2279 equivalent bits/point +MSE 90.274930 +---------------------- -------------------------------------------------------- +Time: 0.721s Load: 0.008s, Pack+Encode: 0.302s, Decode+Unpack: 0.411s +---------------------- -------------------------------------------------------- +💾 Converting with 90.2749 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,576B, BPFP=0.4912 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,180B, BPFP=2.0215 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,724B, BPFP=0.8898 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,448B, BPFP=1.9741 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,560B, BPFP=1.1385 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,776B, BPFP=1.9305 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,684B, BPFP=1.0817 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,108B, BPFP=1.9520 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,392B, BPFP=1.6463 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,036B, BPFP=1.8825 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,276B, BPFP=0.5861 +⌛️ [2/4] FRONTEND: Frontend time: 0.315s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.404s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11142051 52.17798155 + layer.0.v_cache 0.00001650 0.00933748 + layer.1.k_cache 0.42032433 4.54955346 + layer.1.v_cache 0.00000619 0.00384270 + layer.2.k_cache 0.01443043 0.74781882 + layer.2.v_cache 0.00002030 0.00988997 + layer.3.k_cache 0.02895402 3.07304845 + layer.3.v_cache 0.00001914 0.01063949 + layer.4.k_cache 0.00067703 0.20262353 + layer.4.v_cache 0.00005693 0.02062521 + layer.4.output 1.27033357 223.85175237 + ------------------------------------------------------------------------------------- + TOTAL 0.55695649 95.75103690 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 294760 +BPFP 1.1241 bits/point +EBPFP 2.2483 equivalent bits/point +MSE 95.751037 +---------------------- -------------------------------------------------------- +Time: 0.727s Load: 0.009s, Pack+Encode: 0.315s, Decode+Unpack: 0.404s +---------------------- -------------------------------------------------------- +💾 Converting with 95.7510 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 302, 128) +Output shape: (1, 302, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.output: torch.Size([1, 302, 3584]) -> torch.Size([1, 1, 302, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,384B, BPFP=0.4855 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,784B, BPFP=2.0066 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,028B, BPFP=0.8810 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,844B, BPFP=1.9580 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,992B, BPFP=1.1378 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,076B, BPFP=1.9183 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,696B, BPFP=1.0708 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,516B, BPFP=1.9410 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,736B, BPFP=1.6420 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,120B, BPFP=1.8688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,324B, BPFP=0.5272 +⌛️ [2/4] FRONTEND: Frontend time: 0.352s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.531s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15012774 52.31764150 + layer.0.v_cache 0.00001504 0.00896570 + layer.1.k_cache 0.60100308 4.47363766 + layer.1.v_cache 0.00000575 0.00375936 + layer.2.k_cache 0.00814273 0.71433627 + layer.2.v_cache 0.00002032 0.00985645 + layer.3.k_cache 0.01797006 3.11536918 + layer.3.v_cache 0.00001899 0.01053763 + layer.4.k_cache 0.00070977 0.19756305 + layer.4.v_cache 0.00004868 0.01937798 + layer.4.output 0.04414053 178.95086329 + ------------------------------------------------------------------------------------- + TOTAL 0.06394387 77.26629928 + (elements=2,628,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2628608 +Total Bytes 359500 +BPFP 1.0941 bits/point +EBPFP 2.1882 equivalent bits/point +MSE 77.266299 +---------------------- -------------------------------------------------------- +Time: 0.894s Load: 0.010s, Pack+Encode: 0.352s, Decode+Unpack: 0.531s +---------------------- -------------------------------------------------------- +💾 Converting with 77.2663 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,552B, BPFP=0.4986 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,412B, BPFP=2.0646 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,692B, BPFP=0.9149 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,500B, BPFP=2.0114 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,852B, BPFP=1.1574 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,712B, BPFP=1.9655 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,744B, BPFP=1.0928 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,328B, BPFP=2.0014 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,668B, BPFP=1.6714 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,000B, BPFP=1.9240 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 73,236B, BPFP=0.6100 +⌛️ [2/4] FRONTEND: Frontend time: 0.338s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.440s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14350806 52.08967977 + layer.0.v_cache 0.00001655 0.00889551 + layer.1.k_cache 0.46812200 4.43970638 + layer.1.v_cache 0.00000598 0.00345605 + layer.2.k_cache 0.01096032 0.70918382 + layer.2.v_cache 0.00002019 0.00926571 + layer.3.k_cache 0.00658789 2.94126277 + layer.3.v_cache 0.00002014 0.01016210 + layer.4.k_cache 0.00067528 0.19263282 + layer.4.v_cache 0.00005785 0.01953785 + layer.4.output 0.00496594 206.42157516 + ------------------------------------------------------------------------------------- + TOTAL 0.03910211 88.55145935 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 335696 +BPFP 1.1513 bits/point +EBPFP 2.3026 equivalent bits/point +MSE 88.551459 +---------------------- -------------------------------------------------------- +Time: 0.788s Load: 0.009s, Pack+Encode: 0.338s, Decode+Unpack: 0.440s +---------------------- -------------------------------------------------------- +💾 Converting with 88.5515 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 260, 128) +Output shape: (1, 260, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.output: torch.Size([1, 260, 3584]) -> torch.Size([1, 1, 260, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,260B, BPFP=0.4964 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,632B, BPFP=2.0812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,464B, BPFP=0.9293 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,620B, BPFP=2.0204 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,744B, BPFP=1.1865 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,876B, BPFP=1.9757 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,512B, BPFP=1.1125 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,368B, BPFP=2.0053 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,184B, BPFP=1.6938 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,120B, BPFP=1.9303 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,284B, BPFP=0.5691 +⌛️ [2/4] FRONTEND: Frontend time: 0.341s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.460s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14548651 52.80588191 + layer.0.v_cache 0.00001584 0.00957349 + layer.1.k_cache 0.47062912 4.42347177 + layer.1.v_cache 0.00000604 0.00394606 + layer.2.k_cache 0.00619981 0.71811430 + layer.2.v_cache 0.00002003 0.01084392 + layer.3.k_cache 0.01299995 3.19315303 + layer.3.v_cache 0.00001981 0.01135663 + layer.4.k_cache 0.00067777 0.21244448 + layer.4.v_cache 0.00006192 0.02184570 + layer.4.output 0.00504520 212.81866415 + ------------------------------------------------------------------------------------- + TOTAL 0.03949607 91.24360472 + (elements=2,263,040) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2263040 +Total Bytes 323064 +BPFP 1.1421 bits/point +EBPFP 2.2841 equivalent bits/point +MSE 91.243605 +---------------------- -------------------------------------------------------- +Time: 0.810s Load: 0.009s, Pack+Encode: 0.341s, Decode+Unpack: 0.460s +---------------------- -------------------------------------------------------- +💾 Converting with 91.2436 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,684B, BPFP=0.4989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,284B, BPFP=2.0843 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,016B, BPFP=0.9200 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,244B, BPFP=2.0246 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,340B, BPFP=1.1684 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,260B, BPFP=1.9681 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,112B, BPFP=1.0979 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,800B, BPFP=1.9991 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,968B, BPFP=1.6641 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,400B, BPFP=1.9187 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,848B, BPFP=0.5978 +⌛️ [2/4] FRONTEND: Frontend time: 0.339s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.488s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14555165 53.82028378 + layer.0.v_cache 0.00001661 0.00925570 + layer.1.k_cache 0.55468144 4.82554043 + layer.1.v_cache 0.00000621 0.00365397 + layer.2.k_cache 0.02243149 0.73590021 + layer.2.v_cache 0.00002103 0.00937877 + layer.3.k_cache 0.02028967 3.05058962 + layer.3.v_cache 0.00002036 0.01044518 + layer.4.k_cache 0.00072975 0.19540355 + layer.4.v_cache 0.00005124 0.01923719 + layer.4.output 0.00486695 203.34122243 + ------------------------------------------------------------------------------------- + TOTAL 0.04575695 87.41577914 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 339956 +BPFP 1.1487 bits/point +EBPFP 2.2975 equivalent bits/point +MSE 87.415779 +---------------------- -------------------------------------------------------- +Time: 0.837s Load: 0.010s, Pack+Encode: 0.339s, Decode+Unpack: 0.488s +---------------------- -------------------------------------------------------- +💾 Converting with 87.4158 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 356, 128) +Output shape: (1, 356, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.output: torch.Size([1, 356, 3584]) -> torch.Size([1, 1, 356, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,804B, BPFP=0.4742 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,928B, BPFP=2.0158 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,220B, BPFP=0.8875 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,644B, BPFP=1.9594 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,280B, BPFP=1.1096 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,212B, BPFP=1.8966 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,736B, BPFP=1.0418 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 43,708B, BPFP=1.9184 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,340B, BPFP=1.5950 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 41,756B, BPFP=1.8327 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,276B, BPFP=0.5723 +⌛️ [2/4] FRONTEND: Frontend time: 0.418s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.624s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15223843 52.34218640 + layer.0.v_cache 0.00001740 0.00870158 + layer.1.k_cache 0.82972649 4.28429764 + layer.1.v_cache 0.00000631 0.00350371 + layer.2.k_cache 0.03086275 0.74612461 + layer.2.v_cache 0.00001960 0.00897403 + layer.3.k_cache 0.01100341 2.96813005 + layer.3.v_cache 0.00002062 0.00976657 + layer.4.k_cache 0.00073857 0.18451811 + layer.4.v_cache 0.00005279 0.01803066 + layer.4.output 0.03758925 151.88236106 + ------------------------------------------------------------------------------------- + TOTAL 0.07575360 66.10298593 + (elements=3,098,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3098624 +Total Bytes 426904 +BPFP 1.1022 bits/point +EBPFP 2.2044 equivalent bits/point +MSE 66.102986 +---------------------- -------------------------------------------------------- +Time: 1.054s Load: 0.012s, Pack+Encode: 0.418s, Decode+Unpack: 0.624s +---------------------- -------------------------------------------------------- +💾 Converting with 66.1030 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,476B, BPFP=0.4979 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,472B, BPFP=2.0836 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,800B, BPFP=0.9281 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,580B, BPFP=2.0312 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,244B, BPFP=1.1891 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,852B, BPFP=1.9885 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,808B, BPFP=1.1048 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,280B, BPFP=2.0136 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,764B, BPFP=1.6896 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,956B, BPFP=1.9359 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,784B, BPFP=0.6359 +⌛️ [2/4] FRONTEND: Frontend time: 0.336s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.507s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12347319 53.15938161 + layer.0.v_cache 0.00001696 0.00908187 + layer.1.k_cache 0.44284580 4.71814883 + layer.1.v_cache 0.00000660 0.00370754 + layer.2.k_cache 0.00882101 0.74309121 + layer.2.v_cache 0.00002210 0.01008997 + layer.3.k_cache 0.01145097 3.09306691 + layer.3.v_cache 0.00001984 0.01027578 + layer.4.k_cache 0.00068754 0.20283328 + layer.4.v_cache 0.00006483 0.02014897 + layer.4.output 0.00500964 207.89312567 + ------------------------------------------------------------------------------------- + TOTAL 0.03661626 89.24833563 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 339016 +BPFP 1.1714 bits/point +EBPFP 2.3428 equivalent bits/point +MSE 89.248336 +---------------------- -------------------------------------------------------- +Time: 0.851s Load: 0.008s, Pack+Encode: 0.336s, Decode+Unpack: 0.507s +---------------------- -------------------------------------------------------- +💾 Converting with 89.2483 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,016B, BPFP=0.4943 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,692B, BPFP=2.0664 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,616B, BPFP=0.9110 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,424B, BPFP=1.9969 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,432B, BPFP=1.1750 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,716B, BPFP=1.9581 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,124B, BPFP=1.1033 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,236B, BPFP=1.9866 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,548B, BPFP=1.6748 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,700B, BPFP=1.9024 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 77,068B, BPFP=0.6036 +⌛️ [2/4] FRONTEND: Frontend time: 0.352s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.541s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13002123 51.72559279 + layer.0.v_cache 0.00001833 0.00967870 + layer.1.k_cache 0.50397591 4.55255320 + layer.1.v_cache 0.00000608 0.00374133 + layer.2.k_cache 0.01926375 0.76995566 + layer.2.v_cache 0.00002060 0.00999594 + layer.3.k_cache 0.04992905 3.23879437 + layer.3.v_cache 0.00002096 0.01151942 + layer.4.k_cache 0.00069540 0.20365159 + layer.4.v_cache 0.00005467 0.02090400 + layer.4.output 0.00466491 194.14418860 + ------------------------------------------------------------------------------------- + TOTAL 0.04333296 83.50327689 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 355572 +BPFP 1.1467 bits/point +EBPFP 2.2934 equivalent bits/point +MSE 83.503277 +---------------------- -------------------------------------------------------- +Time: 0.902s Load: 0.010s, Pack+Encode: 0.352s, Decode+Unpack: 0.541s +---------------------- -------------------------------------------------------- +💾 Converting with 83.5033 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,024B, BPFP=0.5036 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,816B, BPFP=2.0545 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,476B, BPFP=0.9194 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,748B, BPFP=1.9949 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,424B, BPFP=1.1397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,540B, BPFP=1.9275 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,332B, BPFP=1.0788 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,292B, BPFP=1.9694 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,456B, BPFP=1.6438 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,808B, BPFP=1.8866 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,948B, BPFP=0.6533 +⌛️ [2/4] FRONTEND: Frontend time: 0.371s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.537s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15974760 51.98928571 + layer.0.v_cache 0.00001739 0.00854860 + layer.1.k_cache 0.45644542 4.16685137 + layer.1.v_cache 0.00000643 0.00340668 + layer.2.k_cache 0.01631587 0.69764927 + layer.2.v_cache 0.00002005 0.00874565 + layer.3.k_cache 0.01511352 2.91003069 + layer.3.v_cache 0.00002063 0.01005157 + layer.4.k_cache 0.00073517 0.18645713 + layer.4.v_cache 0.00005091 0.01839716 + layer.4.output 0.00481773 197.41364796 + ------------------------------------------------------------------------------------- + TOTAL 0.04012924 84.81735056 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 352864 +BPFP 1.1583 bits/point +EBPFP 2.3166 equivalent bits/point +MSE 84.817351 +---------------------- -------------------------------------------------------- +Time: 0.921s Load: 0.012s, Pack+Encode: 0.371s, Decode+Unpack: 0.537s +---------------------- -------------------------------------------------------- +💾 Converting with 84.8174 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,496B, BPFP=0.4972 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,332B, BPFP=2.0676 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,772B, BPFP=0.9230 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,520B, BPFP=2.0201 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,844B, BPFP=1.1613 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,548B, BPFP=1.9632 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,740B, BPFP=1.0967 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,168B, BPFP=1.9995 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,720B, BPFP=1.6807 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,816B, BPFP=1.9204 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,828B, BPFP=0.5670 +⌛️ [2/4] FRONTEND: Frontend time: 0.347s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.480s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14148096 53.78540496 + layer.0.v_cache 0.00001588 0.00897245 + layer.1.k_cache 0.55266568 4.69109856 + layer.1.v_cache 0.00000632 0.00374204 + layer.2.k_cache 0.00687984 0.69378216 + layer.2.v_cache 0.00001885 0.00925982 + layer.3.k_cache 0.01675770 3.16304662 + layer.3.v_cache 0.00001964 0.01036763 + layer.4.k_cache 0.00068991 0.19444303 + layer.4.v_cache 0.00004859 0.01914537 + layer.4.output 0.00492290 207.09095773 + ------------------------------------------------------------------------------------- + TOTAL 0.04429669 88.95388040 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 329784 +BPFP 1.1352 bits/point +EBPFP 2.2705 equivalent bits/point +MSE 88.953880 +---------------------- -------------------------------------------------------- +Time: 0.839s Load: 0.011s, Pack+Encode: 0.347s, Decode+Unpack: 0.480s +---------------------- -------------------------------------------------------- +💾 Converting with 88.9539 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,060B, BPFP=0.4866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,620B, BPFP=2.0617 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,608B, BPFP=0.9002 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,068B, BPFP=1.9866 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,544B, BPFP=1.1389 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,972B, BPFP=1.9336 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,160B, BPFP=1.0720 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,720B, BPFP=1.9698 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,556B, BPFP=1.6233 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,736B, BPFP=1.8738 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 82,828B, BPFP=0.5724 +⌛️ [2/4] FRONTEND: Frontend time: 0.393s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.604s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13144862 53.03254402 + layer.0.v_cache 0.00001769 0.00965591 + layer.1.k_cache 0.64513603 4.17257227 + layer.1.v_cache 0.00000607 0.00368974 + layer.2.k_cache 0.02726801 0.69418307 + layer.2.v_cache 0.00001956 0.00947473 + layer.3.k_cache 0.02271994 2.76513067 + layer.3.v_cache 0.00002018 0.01021351 + layer.4.k_cache 0.00073057 0.20180901 + layer.4.v_cache 0.00004940 0.01867260 + layer.4.output 0.04135688 167.39562141 + ------------------------------------------------------------------------------------- + TOTAL 0.06570084 72.51101737 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 393872 +BPFP 1.1208 bits/point +EBPFP 2.2416 equivalent bits/point +MSE 72.511017 +---------------------- -------------------------------------------------------- +Time: 1.011s Load: 0.014s, Pack+Encode: 0.393s, Decode+Unpack: 0.604s +---------------------- -------------------------------------------------------- +💾 Converting with 72.5110 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,872B, BPFP=0.4994 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,912B, BPFP=2.1012 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,832B, BPFP=0.9326 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,072B, BPFP=2.0401 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,392B, BPFP=1.1913 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,280B, BPFP=1.9826 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,536B, BPFP=1.1291 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,652B, BPFP=2.0096 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,508B, BPFP=1.7084 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,608B, BPFP=1.9337 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 54,600B, BPFP=0.5669 +⌛️ [2/4] FRONTEND: Frontend time: 0.323s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.421s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12396691 54.37556323 + layer.0.v_cache 0.00001598 0.00959524 + layer.1.k_cache 0.24854917 4.46879826 + layer.1.v_cache 0.00000584 0.00372229 + layer.2.k_cache 0.02290044 0.70710932 + layer.2.v_cache 0.00002299 0.00977321 + layer.3.k_cache 0.01294550 3.09575706 + layer.3.v_cache 0.00002077 0.01089202 + layer.4.k_cache 0.00065144 0.19594815 + layer.4.v_cache 0.00005054 0.01968510 + layer.4.output 1.42385976 250.53239203 + ------------------------------------------------------------------------------------- + TOTAL 0.61036164 106.86021106 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 268264 +BPFP 1.1468 bits/point +EBPFP 2.2936 equivalent bits/point +MSE 106.860211 +---------------------- -------------------------------------------------------- +Time: 0.753s Load: 0.009s, Pack+Encode: 0.323s, Decode+Unpack: 0.421s +---------------------- -------------------------------------------------------- +💾 Converting with 106.8602 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,808B, BPFP=0.4762 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,600B, BPFP=2.0329 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,372B, BPFP=0.8852 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,332B, BPFP=1.9643 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,112B, BPFP=1.1414 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,572B, BPFP=1.9232 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,960B, BPFP=1.0792 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,316B, BPFP=1.9635 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,064B, BPFP=1.6254 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,432B, BPFP=1.8616 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,616B, BPFP=0.5840 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.460s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14783513 53.61913387 + layer.0.v_cache 0.00001600 0.00959713 + layer.1.k_cache 0.54442752 4.23789476 + layer.1.v_cache 0.00000628 0.00379973 + layer.2.k_cache 0.01848349 0.77675948 + layer.2.v_cache 0.00002061 0.00996851 + layer.3.k_cache 0.01777397 2.86682847 + layer.3.v_cache 0.00002092 0.01121696 + layer.4.k_cache 0.00070447 0.20535431 + layer.4.v_cache 0.00005049 0.02015856 + layer.4.output 0.04615090 187.16028176 + ------------------------------------------------------------------------------------- + TOTAL 0.06190560 80.69898142 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 352184 +BPFP 1.1201 bits/point +EBPFP 2.2401 equivalent bits/point +MSE 80.698981 +---------------------- -------------------------------------------------------- +Time: 0.813s Load: 0.010s, Pack+Encode: 0.342s, Decode+Unpack: 0.460s +---------------------- -------------------------------------------------------- +💾 Converting with 80.6990 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,232B, BPFP=0.4841 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,896B, BPFP=2.0394 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,848B, BPFP=0.8834 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,612B, BPFP=1.9721 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,644B, BPFP=1.1349 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,668B, BPFP=1.9226 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,340B, BPFP=1.0665 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,252B, BPFP=1.9532 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,172B, BPFP=1.6344 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,628B, BPFP=1.8681 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,432B, BPFP=0.5575 +⌛️ [2/4] FRONTEND: Frontend time: 0.359s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.467s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16274841 52.35727113 + layer.0.v_cache 0.00001794 0.00950649 + layer.1.k_cache 0.56737488 4.54061091 + layer.1.v_cache 0.00000586 0.00358399 + layer.2.k_cache 0.02027899 0.72715237 + layer.2.v_cache 0.00002117 0.00933191 + layer.3.k_cache 0.01891868 2.81030949 + layer.3.v_cache 0.00002140 0.01051902 + layer.4.k_cache 0.00069658 0.19471540 + layer.4.v_cache 0.00005005 0.01878788 + layer.4.output 0.04476916 181.39931388 + ------------------------------------------------------------------------------------- + TOTAL 0.06373636 78.26335210 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 359724 +BPFP 1.1095 bits/point +EBPFP 2.2190 equivalent bits/point +MSE 78.263352 +---------------------- -------------------------------------------------------- +Time: 0.837s Load: 0.011s, Pack+Encode: 0.359s, Decode+Unpack: 0.467s +---------------------- -------------------------------------------------------- +💾 Converting with 78.2634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,752B, BPFP=0.5009 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,140B, BPFP=2.0685 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,844B, BPFP=0.9068 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,916B, BPFP=1.9984 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,992B, BPFP=1.1442 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,008B, BPFP=1.9464 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,768B, BPFP=1.0742 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,524B, BPFP=1.9760 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,784B, BPFP=1.6474 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,996B, BPFP=1.8885 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,920B, BPFP=0.5717 +⌛️ [2/4] FRONTEND: Frontend time: 0.333s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.515s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13366390 52.84318839 + layer.0.v_cache 0.00001719 0.00945866 + layer.1.k_cache 0.57618574 4.50358162 + layer.1.v_cache 0.00000658 0.00370829 + layer.2.k_cache 0.02858028 0.72539562 + layer.2.v_cache 0.00002021 0.00967195 + layer.3.k_cache 0.00873641 2.80310685 + layer.3.v_cache 0.00002023 0.01088395 + layer.4.k_cache 0.00067337 0.20000219 + layer.4.v_cache 0.00005011 0.01938439 + layer.4.output 0.00484872 202.21034471 + ------------------------------------------------------------------------------------- + TOTAL 0.04599383 86.85887029 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 334644 +BPFP 1.1267 bits/point +EBPFP 2.2533 equivalent bits/point +MSE 86.858870 +---------------------- -------------------------------------------------------- +Time: 0.858s Load: 0.009s, Pack+Encode: 0.333s, Decode+Unpack: 0.515s +---------------------- -------------------------------------------------------- +💾 Converting with 86.8589 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,084B, BPFP=0.5008 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,780B, BPFP=2.1055 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,364B, BPFP=0.9449 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,772B, BPFP=2.0342 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,016B, BPFP=1.2031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,152B, BPFP=1.9904 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,116B, BPFP=1.1394 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,692B, BPFP=2.0286 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,240B, BPFP=1.7138 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,508B, BPFP=1.9449 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,188B, BPFP=0.6685 +⌛️ [2/4] FRONTEND: Frontend time: 0.315s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.409s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13057878 53.64000636 + layer.0.v_cache 0.00001797 0.01075656 + layer.1.k_cache 0.31658528 4.65178808 + layer.1.v_cache 0.00000647 0.00394744 + layer.2.k_cache 0.02849985 0.77653987 + layer.2.v_cache 0.00001991 0.01041415 + layer.3.k_cache 0.03599559 3.31034665 + layer.3.v_cache 0.00002239 0.01224711 + layer.4.k_cache 0.00078811 0.21751364 + layer.4.v_cache 0.00005540 0.02195074 + layer.4.output 1.38530015 244.17277392 + ------------------------------------------------------------------------------------- + TOTAL 0.60056887 104.22734871 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 286912 +BPFP 1.1932 bits/point +EBPFP 2.3865 equivalent bits/point +MSE 104.227349 +---------------------- -------------------------------------------------------- +Time: 0.731s Load: 0.007s, Pack+Encode: 0.315s, Decode+Unpack: 0.409s +---------------------- -------------------------------------------------------- +💾 Converting with 104.2273 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,960B, BPFP=0.5012 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,304B, BPFP=2.1100 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,984B, BPFP=0.9349 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,284B, BPFP=2.0366 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,868B, BPFP=1.2146 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,704B, BPFP=1.9948 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,796B, BPFP=1.1374 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,392B, BPFP=2.0444 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,672B, BPFP=1.7045 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,100B, BPFP=1.9513 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,152B, BPFP=0.6496 +⌛️ [2/4] FRONTEND: Frontend time: 0.327s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.453s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12725546 53.20244996 + layer.0.v_cache 0.00001736 0.01042844 + layer.1.k_cache 0.32262681 4.39847744 + layer.1.v_cache 0.00000608 0.00392703 + layer.2.k_cache 0.02157613 0.70762655 + layer.2.v_cache 0.00002208 0.01018816 + layer.3.k_cache 0.01760322 2.92653111 + layer.3.v_cache 0.00002192 0.01211119 + layer.4.k_cache 0.00069785 0.21090719 + layer.4.v_cache 0.00005172 0.02156593 + layer.4.output 1.41081570 247.93542215 + ------------------------------------------------------------------------------------- + TOTAL 0.60974050 105.70895106 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 280216 +BPFP 1.1869 bits/point +EBPFP 2.3737 equivalent bits/point +MSE 105.708951 +---------------------- -------------------------------------------------------- +Time: 0.787s Load: 0.007s, Pack+Encode: 0.327s, Decode+Unpack: 0.453s +---------------------- -------------------------------------------------------- +💾 Converting with 105.7090 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,868B, BPFP=0.5038 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,032B, BPFP=2.1297 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,876B, BPFP=0.9445 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,332B, BPFP=2.0783 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,832B, BPFP=1.2347 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,576B, BPFP=2.0229 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,708B, BPFP=1.1523 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,168B, BPFP=2.0663 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,452B, BPFP=1.7204 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,840B, BPFP=1.9689 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,320B, BPFP=0.6636 +⌛️ [2/4] FRONTEND: Frontend time: 0.335s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.424s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13788442 51.31447146 + layer.0.v_cache 0.00001717 0.01089745 + layer.1.k_cache 0.30723858 4.67264123 + layer.1.v_cache 0.00000664 0.00416685 + layer.2.k_cache 0.01337219 0.75038469 + layer.2.v_cache 0.00002159 0.01072181 + layer.3.k_cache 0.03522867 3.24225513 + layer.3.v_cache 0.00002126 0.01231470 + layer.4.k_cache 0.00073009 0.21069583 + layer.4.v_cache 0.00005119 0.02173137 + layer.4.output 1.43730973 253.27276157 + ------------------------------------------------------------------------------------- + TOTAL 0.62092588 107.83291832 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 279004 +BPFP 1.2039 bits/point +EBPFP 2.4079 equivalent bits/point +MSE 107.832918 +---------------------- -------------------------------------------------------- +Time: 0.767s Load: 0.008s, Pack+Encode: 0.335s, Decode+Unpack: 0.424s +---------------------- -------------------------------------------------------- +💾 Converting with 107.8329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,664B, BPFP=0.5104 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,084B, BPFP=2.1510 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,568B, BPFP=0.9626 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,204B, BPFP=2.0836 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,140B, BPFP=1.2362 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,624B, BPFP=2.0392 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,144B, BPFP=1.1599 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,276B, BPFP=2.0892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,700B, BPFP=1.7387 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,000B, BPFP=1.9914 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,868B, BPFP=0.6770 +⌛️ [2/4] FRONTEND: Frontend time: 0.264s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11546621 52.71031997 + layer.0.v_cache 0.00001823 0.01020284 + layer.1.k_cache 0.26924565 4.96476416 + layer.1.v_cache 0.00000737 0.00413976 + layer.2.k_cache 0.01761903 0.77975509 + layer.2.v_cache 0.00002145 0.01028434 + layer.3.k_cache 0.02374693 2.99670829 + layer.3.v_cache 0.00002104 0.01160669 + layer.4.k_cache 0.00067272 0.20601643 + layer.4.v_cache 0.00005153 0.02055965 + layer.4.output 1.50071327 264.46124387 + ------------------------------------------------------------------------------------- + TOTAL 0.64305077 112.52606261 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 270272 +BPFP 1.2177 bits/point +EBPFP 2.4354 equivalent bits/point +MSE 112.526063 +---------------------- -------------------------------------------------------- +Time: 0.710s Load: 0.007s, Pack+Encode: 0.264s, Decode+Unpack: 0.439s +---------------------- -------------------------------------------------------- +💾 Converting with 112.5261 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,488B, BPFP=0.4813 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,800B, BPFP=2.0191 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,484B, BPFP=0.8870 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,764B, BPFP=1.9665 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,708B, BPFP=1.1520 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,820B, BPFP=1.9186 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,528B, BPFP=1.0921 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,316B, BPFP=1.9438 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,184B, BPFP=1.6327 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,680B, BPFP=1.8608 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 82,228B, BPFP=0.5959 +⌛️ [2/4] FRONTEND: Frontend time: 0.390s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.521s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18206066 50.71514179 + layer.0.v_cache 0.00001652 0.00959053 + layer.1.k_cache 0.62405802 4.56616964 + layer.1.v_cache 0.00000616 0.00376303 + layer.2.k_cache 0.02036299 0.69935326 + layer.2.v_cache 0.00002087 0.01005712 + layer.3.k_cache 0.02546444 2.85145826 + layer.3.v_cache 0.00002064 0.01066187 + layer.4.k_cache 0.00069880 0.19867691 + layer.4.v_cache 0.00007146 0.02042679 + layer.4.output 0.04336799 175.30246985 + ------------------------------------------------------------------------------------- + TOTAL 0.06802097 75.65897577 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 377000 +BPFP 1.1250 bits/point +EBPFP 2.2500 equivalent bits/point +MSE 75.658976 +---------------------- -------------------------------------------------------- +Time: 0.921s Load: 0.011s, Pack+Encode: 0.390s, Decode+Unpack: 0.521s +---------------------- -------------------------------------------------------- +💾 Converting with 75.6590 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,088B, BPFP=0.4836 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,856B, BPFP=2.1053 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,588B, BPFP=0.9271 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,956B, BPFP=2.0439 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,304B, BPFP=1.1807 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,988B, BPFP=1.9779 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,104B, BPFP=1.0988 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,592B, BPFP=2.0191 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,568B, BPFP=1.6763 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,216B, BPFP=1.9252 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,500B, BPFP=0.6385 +⌛️ [2/4] FRONTEND: Frontend time: 0.302s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.412s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14018573 53.34496111 + layer.0.v_cache 0.00001653 0.00987945 + layer.1.k_cache 0.40671090 4.43972765 + layer.1.v_cache 0.00000662 0.00386720 + layer.2.k_cache 0.03036383 0.77179488 + layer.2.v_cache 0.00002088 0.00991648 + layer.3.k_cache 0.05940984 3.23304049 + layer.3.v_cache 0.00002025 0.01055578 + layer.4.k_cache 0.00071161 0.19871929 + layer.4.v_cache 0.00005094 0.01898483 + layer.4.output 1.33692960 235.79524719 + ------------------------------------------------------------------------------------- + TOTAL 0.58800025 100.74165750 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 291760 +BPFP 1.1710 bits/point +EBPFP 2.3420 equivalent bits/point +MSE 100.741658 +---------------------- -------------------------------------------------------- +Time: 0.721s Load: 0.008s, Pack+Encode: 0.302s, Decode+Unpack: 0.412s +---------------------- -------------------------------------------------------- +💾 Converting with 100.7417 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,044B, BPFP=0.5026 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,588B, BPFP=2.1110 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,128B, BPFP=0.9366 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,680B, BPFP=2.0462 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,104B, BPFP=1.2203 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,192B, BPFP=2.0114 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,076B, BPFP=1.1470 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,616B, BPFP=2.0417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,008B, BPFP=1.7129 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,456B, BPFP=1.9589 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,588B, BPFP=0.6583 +⌛️ [2/4] FRONTEND: Frontend time: 0.336s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.433s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351684 53.25235445 + layer.0.v_cache 0.00001872 0.01087096 + layer.1.k_cache 0.28219827 4.73280146 + layer.1.v_cache 0.00000627 0.00428182 + layer.2.k_cache 0.01639774 0.74065361 + layer.2.v_cache 0.00002172 0.01164210 + layer.3.k_cache 0.02752649 3.29474490 + layer.3.v_cache 0.00002090 0.01222252 + layer.4.k_cache 0.00067264 0.21594352 + layer.4.v_cache 0.00005058 0.02265339 + layer.4.output 1.39794763 245.95592792 + ------------------------------------------------------------------------------------- + TOTAL 0.60212139 104.94056848 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 284480 +BPFP 1.1939 bits/point +EBPFP 2.3879 equivalent bits/point +MSE 104.940568 +---------------------- -------------------------------------------------------- +Time: 0.778s Load: 0.009s, Pack+Encode: 0.336s, Decode+Unpack: 0.433s +---------------------- -------------------------------------------------------- +💾 Converting with 104.9406 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,656B, BPFP=0.5047 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,696B, BPFP=2.0812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,880B, BPFP=0.9258 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,780B, BPFP=2.0278 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,076B, BPFP=1.1705 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,724B, BPFP=1.9662 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,784B, BPFP=1.0951 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,244B, BPFP=1.9965 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,484B, BPFP=1.6607 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,792B, BPFP=1.9118 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 73,896B, BPFP=0.6155 +⌛️ [2/4] FRONTEND: Frontend time: 0.358s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.487s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15646637 53.84546627 + layer.0.v_cache 0.00001727 0.00952251 + layer.1.k_cache 0.53079696 4.25712335 + layer.1.v_cache 0.00000635 0.00403269 + layer.2.k_cache 0.02340859 0.74920205 + layer.2.v_cache 0.00002061 0.01004482 + layer.3.k_cache 0.01050839 3.05731383 + layer.3.v_cache 0.00002156 0.01045846 + layer.4.k_cache 0.00068596 0.19727056 + layer.4.v_cache 0.00006072 0.02001495 + layer.4.output 0.00497423 206.37893124 + ------------------------------------------------------------------------------------- + TOTAL 0.04451838 88.63605695 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 337012 +BPFP 1.1558 bits/point +EBPFP 2.3116 equivalent bits/point +MSE 88.636057 +---------------------- -------------------------------------------------------- +Time: 0.854s Load: 0.008s, Pack+Encode: 0.358s, Decode+Unpack: 0.487s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,484B, BPFP=0.4811 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,648B, BPFP=2.0114 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,324B, BPFP=0.8789 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,680B, BPFP=1.9623 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,520B, BPFP=1.1425 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,540B, BPFP=1.9044 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,912B, BPFP=1.0609 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,892B, BPFP=1.9223 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,828B, BPFP=1.6147 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,160B, BPFP=1.8344 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,624B, BPFP=0.6133 +⌛️ [2/4] FRONTEND: Frontend time: 0.336s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15292851 50.32734502 + layer.0.v_cache 0.00001614 0.01010347 + layer.1.k_cache 0.61736927 4.57809448 + layer.1.v_cache 0.00000669 0.00415800 + layer.2.k_cache 0.02003183 0.74415172 + layer.2.v_cache 0.00002344 0.01045921 + layer.3.k_cache 0.01028479 2.88311034 + layer.3.v_cache 0.00002212 0.01086626 + layer.4.k_cache 0.00070434 0.21043976 + layer.4.v_cache 0.00006490 0.02026974 + layer.4.output 0.04342050 175.32154453 + ------------------------------------------------------------------------------------- + TOTAL 0.06502327 75.64998881 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 376612 +BPFP 1.1239 bits/point +EBPFP 2.2477 equivalent bits/point +MSE 75.649989 +---------------------- -------------------------------------------------------- +Time: 0.816s Load: 0.010s, Pack+Encode: 0.336s, Decode+Unpack: 0.470s +---------------------- -------------------------------------------------------- +💾 Converting with 75.6500 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,424B, BPFP=0.4893 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,988B, BPFP=2.0904 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,832B, BPFP=0.9196 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,960B, BPFP=2.0307 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,352B, BPFP=1.1822 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,088B, BPFP=1.9800 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,108B, BPFP=1.1099 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,864B, BPFP=2.0251 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,932B, BPFP=1.6805 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,096B, BPFP=1.9224 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,652B, BPFP=0.5697 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.482s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12950333 52.73647334 + layer.0.v_cache 0.00001690 0.00986989 + layer.1.k_cache 0.52417032 4.11660846 + layer.1.v_cache 0.00000602 0.00400718 + layer.2.k_cache 0.01578268 0.74994243 + layer.2.v_cache 0.00002268 0.01054040 + layer.3.k_cache 0.03345619 3.00322148 + layer.3.v_cache 0.00002007 0.01151115 + layer.4.k_cache 0.00067301 0.20176850 + layer.4.v_cache 0.00005077 0.02023407 + layer.4.output 0.00489311 205.63100770 + ------------------------------------------------------------------------------------- + TOTAL 0.04340905 88.25183711 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 334296 +BPFP 1.1422 bits/point +EBPFP 2.2844 equivalent bits/point +MSE 88.251837 +---------------------- -------------------------------------------------------- +Time: 0.810s Load: 0.008s, Pack+Encode: 0.319s, Decode+Unpack: 0.482s +---------------------- -------------------------------------------------------- +💾 Converting with 88.2518 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,092B, BPFP=0.5014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,436B, BPFP=2.0812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,064B, BPFP=0.9236 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,456B, BPFP=2.0119 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,672B, BPFP=1.1787 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,784B, BPFP=1.9644 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,680B, BPFP=1.1086 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,424B, BPFP=2.0096 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,816B, BPFP=1.6838 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,192B, BPFP=1.9225 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 59,540B, BPFP=0.6014 +⌛️ [2/4] FRONTEND: Frontend time: 0.306s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.428s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13888381 52.99592583 + layer.0.v_cache 0.00001731 0.00950417 + layer.1.k_cache 0.27925331 4.57409889 + layer.1.v_cache 0.00000589 0.00374753 + layer.2.k_cache 0.03172064 0.78680013 + layer.2.v_cache 0.00001980 0.00972413 + layer.3.k_cache 0.01738505 3.08682458 + layer.3.v_cache 0.00001961 0.01084625 + layer.4.k_cache 0.00068340 0.20230745 + layer.4.v_cache 0.00005063 0.02030489 + layer.4.output 1.38525280 244.32864011 + ------------------------------------------------------------------------------------- + TOTAL 0.59792994 104.23532733 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 277156 +BPFP 1.1527 bits/point +EBPFP 2.3053 equivalent bits/point +MSE 104.235327 +---------------------- -------------------------------------------------------- +Time: 0.742s Load: 0.008s, Pack+Encode: 0.306s, Decode+Unpack: 0.428s +---------------------- -------------------------------------------------------- +💾 Converting with 104.2353 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.1427 bits/point +Avg EBPFP 2.2854 equivalent bits/point +Avg MSE 88.844285 +Avg Time 0.837s +------------------------ ---------------------------- diff --git a/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..dd87746c205a0d3916487bbec4c0b9858f6219b2 --- /dev/null +++ b/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 599 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other +Output output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,540B, BPFP=0.5084 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,660B, BPFP=2.1502 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,264B, BPFP=0.9534 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,884B, BPFP=2.0899 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,776B, BPFP=1.2264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,352B, BPFP=2.0485 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,860B, BPFP=1.1552 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,012B, BPFP=2.0998 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,500B, BPFP=1.7491 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,860B, BPFP=2.0103 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,932B, BPFP=0.6878 +⌛️ [2/4] FRONTEND: Frontend time: 0.895s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.503s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09475397 52.79920223 + layer.0.v_cache 0.00001927 0.01030736 + layer.1.k_cache 0.17401673 4.61821113 + layer.1.v_cache 0.00000670 0.00410179 + layer.2.k_cache 0.00814445 0.78204376 + layer.2.v_cache 0.00002157 0.01092330 + layer.3.k_cache 0.01691350 3.13868402 + layer.3.v_cache 0.00002161 0.01305342 + layer.4.k_cache 0.00069969 0.22756768 + layer.4.v_cache 0.00005477 0.02255981 + layer.4.output 1.52309445 268.51645789 + ------------------------------------------------------------------------------------- + TOTAL 0.64448903 114.19069763 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 267640 +BPFP 1.2238 bits/point +EBPFP 2.4477 equivalent bits/point +MSE 114.190698 +---------------------- -------------------------------------------------------- +Time: 1.406s Load: 0.008s, Pack+Encode: 0.895s, Decode+Unpack: 0.503s +---------------------- -------------------------------------------------------- +💾 Converting with 114.1907 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 212, 128) +Output shape: (1, 212, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.output: torch.Size([1, 212, 3584]) -> torch.Size([1, 1, 212, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,892B, BPFP=0.5080 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,020B, BPFP=2.1389 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,756B, BPFP=0.9402 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,124B, BPFP=2.0728 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,660B, BPFP=1.2279 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,528B, BPFP=2.0289 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,676B, BPFP=1.1554 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,320B, BPFP=2.0873 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,544B, BPFP=1.7353 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,868B, BPFP=1.9802 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 60,768B, BPFP=0.6398 +⌛️ [2/4] FRONTEND: Frontend time: 0.350s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.450s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09403566 50.87670438 + layer.0.v_cache 0.00001799 0.01090452 + layer.1.k_cache 0.24957257 4.41672674 + layer.1.v_cache 0.00000621 0.00424117 + layer.2.k_cache 0.01169950 0.77739025 + layer.2.v_cache 0.00002031 0.01116993 + layer.3.k_cache 0.03645871 3.24416942 + layer.3.v_cache 0.00002058 0.01270630 + layer.4.k_cache 0.00069161 0.22048018 + layer.4.v_cache 0.00005130 0.02216238 + layer.4.output 1.44404146 254.55767773 + ------------------------------------------------------------------------------------- + TOTAL 0.61769792 108.32355290 + (elements=1,845,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1845248 +Total Bytes 276156 +BPFP 1.1973 bits/point +EBPFP 2.3945 equivalent bits/point +MSE 108.323553 +---------------------- -------------------------------------------------------- +Time: 0.810s Load: 0.010s, Pack+Encode: 0.350s, Decode+Unpack: 0.450s +---------------------- -------------------------------------------------------- +💾 Converting with 108.3236 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 250, 128) +Output shape: (1, 250, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.output: torch.Size([1, 250, 3584]) -> torch.Size([1, 1, 250, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,764B, BPFP=0.4853 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,816B, BPFP=1.9885 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,376B, BPFP=0.8985 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,952B, BPFP=1.9345 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,412B, BPFP=1.1507 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,288B, BPFP=1.8930 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,268B, BPFP=1.0793 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,688B, BPFP=1.9180 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,920B, BPFP=1.6200 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,468B, BPFP=1.8417 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,996B, BPFP=0.5893 +⌛️ [2/4] FRONTEND: Frontend time: 0.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.437s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09121272 52.10854688 + layer.0.v_cache 0.00001793 0.00971594 + layer.1.k_cache 0.34710925 4.18262158 + layer.1.v_cache 0.00000615 0.00385698 + layer.2.k_cache 0.02211576 0.77846436 + layer.2.v_cache 0.00002179 0.00992364 + layer.3.k_cache 0.02174057 3.02483545 + layer.3.v_cache 0.00001938 0.01054313 + layer.4.k_cache 0.00073173 0.20007437 + layer.4.v_cache 0.00005145 0.01972987 + layer.4.output 1.22463198 213.99826786 + ------------------------------------------------------------------------------------- + TOTAL 0.53267356 91.66683454 + (elements=2,176,000) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2176000 +Total Bytes 302948 +BPFP 1.1138 bits/point +EBPFP 2.2276 equivalent bits/point +MSE 91.666835 +---------------------- -------------------------------------------------------- +Time: 0.780s Load: 0.010s, Pack+Encode: 0.334s, Decode+Unpack: 0.437s +---------------------- -------------------------------------------------------- +💾 Converting with 91.6668 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,904B, BPFP=0.4752 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,148B, BPFP=2.0752 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,132B, BPFP=0.9039 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,404B, BPFP=2.0240 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,036B, BPFP=1.1726 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,708B, BPFP=1.9760 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,892B, BPFP=1.0939 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,084B, BPFP=2.0019 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,432B, BPFP=1.6817 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,840B, BPFP=1.9163 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 60,548B, BPFP=0.5954 +⌛️ [2/4] FRONTEND: Frontend time: 0.294s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.389s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12780563 51.89763302 + layer.0.v_cache 0.00001798 0.01015412 + layer.1.k_cache 0.36413978 4.54509073 + layer.1.v_cache 0.00000619 0.00388811 + layer.2.k_cache 0.01529831 0.78054319 + layer.2.v_cache 0.00002052 0.01021017 + layer.3.k_cache 0.01269315 3.09478356 + layer.3.v_cache 0.00002023 0.01138713 + layer.4.k_cache 0.00069261 0.20778589 + layer.4.v_cache 0.00005545 0.02098868 + layer.4.output 1.34865724 237.92861076 + ------------------------------------------------------------------------------------- + TOTAL 0.58596180 101.53427882 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 283128 +BPFP 1.1464 bits/point +EBPFP 2.2928 equivalent bits/point +MSE 101.534279 +---------------------- -------------------------------------------------------- +Time: 0.690s Load: 0.007s, Pack+Encode: 0.294s, Decode+Unpack: 0.389s +---------------------- -------------------------------------------------------- +💾 Converting with 101.5343 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 214, 128) +Output shape: (1, 214, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.output: torch.Size([1, 214, 3584]) -> torch.Size([1, 1, 214, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,100B, BPFP=0.5184 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,816B, BPFP=2.1040 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,676B, BPFP=0.9255 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,992B, BPFP=2.0438 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,336B, BPFP=1.1928 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,364B, BPFP=1.9980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,360B, BPFP=1.1215 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,752B, BPFP=2.0263 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,236B, BPFP=1.6966 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,596B, BPFP=1.9419 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 55,416B, BPFP=0.5780 +⌛️ [2/4] FRONTEND: Frontend time: 0.298s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.399s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12820032 52.25259656 + layer.0.v_cache 0.00001693 0.00946171 + layer.1.k_cache 0.24875343 4.27842470 + layer.1.v_cache 0.00000587 0.00353584 + layer.2.k_cache 0.01453762 0.72910929 + layer.2.v_cache 0.00001988 0.00950391 + layer.3.k_cache 0.02276520 2.72207499 + layer.3.v_cache 0.00001932 0.01065314 + layer.4.k_cache 0.00071399 0.20694647 + layer.4.v_cache 0.00004895 0.01962931 + layer.4.output 1.43051836 251.97648949 + ------------------------------------------------------------------------------------- + TOTAL 0.61345353 107.29866837 + (elements=1,862,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1862656 +Total Bytes 268644 +BPFP 1.1538 bits/point +EBPFP 2.3076 equivalent bits/point +MSE 107.298668 +---------------------- -------------------------------------------------------- +Time: 0.704s Load: 0.008s, Pack+Encode: 0.298s, Decode+Unpack: 0.399s +---------------------- -------------------------------------------------------- +💾 Converting with 107.2987 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,840B, BPFP=0.4986 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,500B, BPFP=2.0589 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,976B, BPFP=0.9012 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,508B, BPFP=2.0029 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,636B, BPFP=1.1640 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,788B, BPFP=1.9623 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,376B, BPFP=1.0930 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,344B, BPFP=1.9937 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,368B, BPFP=1.6566 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,900B, BPFP=1.9122 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 76,884B, BPFP=0.6196 +⌛️ [2/4] FRONTEND: Frontend time: 0.414s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.529s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11875103 51.35088913 + layer.0.v_cache 0.00001689 0.00939079 + layer.1.k_cache 0.49171178 4.57153585 + layer.1.v_cache 0.00000616 0.00368617 + layer.2.k_cache 0.02532451 0.74803288 + layer.2.v_cache 0.00002086 0.01004311 + layer.3.k_cache 0.02370731 3.21920633 + layer.3.v_cache 0.00002129 0.01093288 + layer.4.k_cache 0.00073268 0.19886587 + layer.4.v_cache 0.00004879 0.01988695 + layer.4.output 0.00480151 199.62292096 + ------------------------------------------------------------------------------------- + TOTAL 0.04082070 85.73546569 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 347120 +BPFP 1.1518 bits/point +EBPFP 2.3036 equivalent bits/point +MSE 85.735466 +---------------------- -------------------------------------------------------- +Time: 0.952s Load: 0.009s, Pack+Encode: 0.414s, Decode+Unpack: 0.529s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7355 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,632B, BPFP=0.4867 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,612B, BPFP=2.0161 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,864B, BPFP=0.8842 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,804B, BPFP=1.9645 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,104B, BPFP=1.1546 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,224B, BPFP=1.9276 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,148B, BPFP=1.0936 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,608B, BPFP=1.9520 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,848B, BPFP=1.6485 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,356B, BPFP=1.8722 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,200B, BPFP=0.6031 +⌛️ [2/4] FRONTEND: Frontend time: 0.292s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.444s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10657067 50.47553811 + layer.0.v_cache 0.00001604 0.00959304 + layer.1.k_cache 0.39270948 4.67349530 + layer.1.v_cache 0.00000611 0.00378986 + layer.2.k_cache 0.01541229 0.71921580 + layer.2.v_cache 0.00002285 0.01048623 + layer.3.k_cache 0.01321528 3.07501669 + layer.3.v_cache 0.00001985 0.01118850 + layer.4.k_cache 0.00076706 0.20645389 + layer.4.v_cache 0.00005205 0.01990660 + layer.4.output 1.24960368 219.76406706 + ------------------------------------------------------------------------------------- + TOTAL 0.54564808 93.97371491 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 301400 +BPFP 1.1307 bits/point +EBPFP 2.2614 equivalent bits/point +MSE 93.973715 +---------------------- -------------------------------------------------------- +Time: 0.745s Load: 0.009s, Pack+Encode: 0.292s, Decode+Unpack: 0.444s +---------------------- -------------------------------------------------------- +💾 Converting with 93.9737 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 197, 128) +Output shape: (1, 197, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.output: torch.Size([1, 197, 3584]) -> torch.Size([1, 1, 197, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,316B, BPFP=0.5010 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,944B, BPFP=2.1371 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,988B, BPFP=0.9508 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,100B, BPFP=2.0701 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,692B, BPFP=1.2446 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,716B, BPFP=2.0397 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,596B, BPFP=1.1577 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,184B, BPFP=2.0768 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,936B, BPFP=1.7398 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,340B, BPFP=2.0098 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 59,072B, BPFP=0.6693 +⌛️ [2/4] FRONTEND: Frontend time: 0.345s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09219642 52.61525420 + layer.0.v_cache 0.00001621 0.00977164 + layer.1.k_cache 0.17840836 4.40676059 + layer.1.v_cache 0.00000639 0.00387955 + layer.2.k_cache 0.00922235 0.74949514 + layer.2.v_cache 0.00002142 0.01041693 + layer.3.k_cache 0.01957878 2.97488822 + layer.3.v_cache 0.00002124 0.01143246 + layer.4.k_cache 0.00068177 0.21296508 + layer.4.v_cache 0.00005086 0.02160321 + layer.4.output 1.55401793 274.13039340 + ------------------------------------------------------------------------------------- + TOTAL 0.65754878 116.46642475 + (elements=1,714,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1714688 +Total Bytes 259884 +BPFP 1.2125 bits/point +EBPFP 2.4250 equivalent bits/point +MSE 116.466425 +---------------------- -------------------------------------------------------- +Time: 0.821s Load: 0.007s, Pack+Encode: 0.345s, Decode+Unpack: 0.470s +---------------------- -------------------------------------------------------- +💾 Converting with 116.4664 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,652B, BPFP=0.4865 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,120B, BPFP=1.9718 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,124B, BPFP=0.8631 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,920B, BPFP=1.9113 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,168B, BPFP=1.1173 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,300B, BPFP=1.8800 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,608B, BPFP=1.0387 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,736B, BPFP=1.9020 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,632B, BPFP=1.5944 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,380B, BPFP=1.8337 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,524B, BPFP=0.5654 +⌛️ [2/4] FRONTEND: Frontend time: 0.386s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.517s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11169702 49.68211946 + layer.0.v_cache 0.00001595 0.00953419 + layer.1.k_cache 0.54955656 4.18206984 + layer.1.v_cache 0.00000603 0.00384412 + layer.2.k_cache 0.01751511 0.70751899 + layer.2.v_cache 0.00002011 0.01014074 + layer.3.k_cache 0.01650532 3.03076172 + layer.3.v_cache 0.00001967 0.01049629 + layer.4.k_cache 0.00077751 0.20476394 + layer.4.v_cache 0.00005008 0.02053181 + layer.4.output 0.04306356 174.16507776 + ------------------------------------------------------------------------------------- + TOTAL 0.05868284 75.11866620 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 368164 +BPFP 1.0916 bits/point +EBPFP 2.1831 equivalent bits/point +MSE 75.118666 +---------------------- -------------------------------------------------------- +Time: 0.912s Load: 0.010s, Pack+Encode: 0.386s, Decode+Unpack: 0.517s +---------------------- -------------------------------------------------------- +💾 Converting with 75.1187 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,976B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,304B, BPFP=2.1003 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,016B, BPFP=0.9329 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,416B, BPFP=2.0367 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,520B, BPFP=1.1841 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,600B, BPFP=1.9782 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,476B, BPFP=1.1092 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,240B, BPFP=2.0241 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,436B, BPFP=1.6798 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,916B, BPFP=1.9292 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,404B, BPFP=0.6492 +⌛️ [2/4] FRONTEND: Frontend time: 0.322s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.429s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09832749 53.12497760 + layer.0.v_cache 0.00001658 0.01010478 + layer.1.k_cache 0.33493462 4.42280957 + layer.1.v_cache 0.00000656 0.00388690 + layer.2.k_cache 0.02762392 0.74327815 + layer.2.v_cache 0.00002012 0.00975934 + layer.3.k_cache 0.01647770 2.78048090 + layer.3.v_cache 0.00001951 0.01087812 + layer.4.k_cache 0.00067889 0.20545615 + layer.4.v_cache 0.00004874 0.01968148 + layer.4.output 1.40435175 246.60073313 + ------------------------------------------------------------------------------------- + TOTAL 0.60638920 105.14920264 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 279304 +BPFP 1.1776 bits/point +EBPFP 2.3552 equivalent bits/point +MSE 105.149203 +---------------------- -------------------------------------------------------- +Time: 0.757s Load: 0.007s, Pack+Encode: 0.322s, Decode+Unpack: 0.429s +---------------------- -------------------------------------------------------- +💾 Converting with 105.1492 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,904B, BPFP=0.4773 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,664B, BPFP=2.0509 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,216B, BPFP=0.9137 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,808B, BPFP=1.9917 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,852B, BPFP=1.1651 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,160B, BPFP=1.9469 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,656B, BPFP=1.0824 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,656B, BPFP=1.9812 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,888B, BPFP=1.6515 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,480B, BPFP=1.8999 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,236B, BPFP=0.6147 +⌛️ [2/4] FRONTEND: Frontend time: 0.338s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.460s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102250 54.08536712 + layer.0.v_cache 0.00001721 0.00941540 + layer.1.k_cache 0.37371607 4.41077308 + layer.1.v_cache 0.00000658 0.00371814 + layer.2.k_cache 0.01127678 0.72197156 + layer.2.v_cache 0.00001998 0.01001810 + layer.3.k_cache 0.02136800 3.01287869 + layer.3.v_cache 0.00001963 0.01077812 + layer.4.k_cache 0.00069450 0.20403816 + layer.4.v_cache 0.00005147 0.02047871 + layer.4.output 1.35464589 238.95591024 + ------------------------------------------------------------------------------------- + TOTAL 0.58945376 102.06945934 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 281520 +BPFP 1.1449 bits/point +EBPFP 2.2898 equivalent bits/point +MSE 102.069459 +---------------------- -------------------------------------------------------- +Time: 0.808s Load: 0.010s, Pack+Encode: 0.338s, Decode+Unpack: 0.460s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0695 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,868B, BPFP=0.4898 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,744B, BPFP=1.9761 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,392B, BPFP=0.8959 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,900B, BPFP=1.9236 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,252B, BPFP=1.1362 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,224B, BPFP=1.8815 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,264B, BPFP=1.0747 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,748B, BPFP=1.9141 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,732B, BPFP=1.6018 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,292B, BPFP=1.8235 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,440B, BPFP=0.5820 +⌛️ [2/4] FRONTEND: Frontend time: 0.303s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.410s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11871604 53.15537070 + layer.0.v_cache 0.00001755 0.00881046 + layer.1.k_cache 0.46315337 3.97837313 + layer.1.v_cache 0.00000588 0.00340782 + layer.2.k_cache 0.02202094 0.71987246 + layer.2.v_cache 0.00001989 0.00928432 + layer.3.k_cache 0.01278766 2.73828782 + layer.3.v_cache 0.00002092 0.01038678 + layer.4.k_cache 0.00074254 0.19262154 + layer.4.v_cache 0.00004925 0.01838102 + layer.4.output 1.21975157 213.91359562 + ------------------------------------------------------------------------------------- + TOTAL 0.53857618 91.66058620 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 301856 +BPFP 1.1053 bits/point +EBPFP 2.2107 equivalent bits/point +MSE 91.660586 +---------------------- -------------------------------------------------------- +Time: 0.724s Load: 0.010s, Pack+Encode: 0.303s, Decode+Unpack: 0.410s +---------------------- -------------------------------------------------------- +💾 Converting with 91.6606 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 195, 128) +Output shape: (1, 195, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.output: torch.Size([1, 195, 3584]) -> torch.Size([1, 1, 195, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,184B, BPFP=0.4955 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,844B, BPFP=2.1510 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,928B, BPFP=0.9558 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,116B, BPFP=2.0926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,088B, BPFP=1.2090 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,392B, BPFP=2.0346 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,500B, BPFP=1.1619 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,000B, BPFP=2.0833 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,728B, BPFP=1.7410 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,968B, BPFP=2.0006 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 54,428B, BPFP=0.6230 +⌛️ [2/4] FRONTEND: Frontend time: 0.323s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.462s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605874 53.17071815 + layer.0.v_cache 0.00001676 0.00965697 + layer.1.k_cache 0.13872500 4.95088642 + layer.1.v_cache 0.00000593 0.00368330 + layer.2.k_cache 0.01222654 0.72942145 + layer.2.v_cache 0.00001968 0.00949328 + layer.3.k_cache 0.01415822 3.24882906 + layer.3.v_cache 0.00002127 0.01159932 + layer.4.k_cache 0.00071127 0.20687512 + layer.4.v_cache 0.00005419 0.02128612 + layer.4.output 1.56987251 276.83514194 + ------------------------------------------------------------------------------------- + TOTAL 0.66300619 117.65932016 + (elements=1,697,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1697280 +Total Bytes 253176 +BPFP 1.1933 bits/point +EBPFP 2.3867 equivalent bits/point +MSE 117.659320 +---------------------- -------------------------------------------------------- +Time: 0.792s Load: 0.007s, Pack+Encode: 0.323s, Decode+Unpack: 0.462s +---------------------- -------------------------------------------------------- +💾 Converting with 117.6593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,876B, BPFP=0.5045 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,908B, BPFP=2.0525 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,684B, BPFP=0.9172 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,268B, BPFP=1.9976 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,728B, BPFP=1.1786 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,768B, BPFP=1.9547 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,940B, BPFP=1.1109 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,144B, BPFP=1.9870 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,400B, BPFP=1.6655 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,092B, BPFP=1.8966 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 51,024B, BPFP=0.6258 +⌛️ [2/4] FRONTEND: Frontend time: 0.367s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.358s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11176265 52.49808444 + layer.0.v_cache 0.00001802 0.01031226 + layer.1.k_cache 0.18491925 4.29727709 + layer.1.v_cache 0.00000608 0.00397425 + layer.2.k_cache 0.01999094 0.71444493 + layer.2.v_cache 0.00002176 0.01046291 + layer.3.k_cache 0.03910406 2.85885419 + layer.3.v_cache 0.00002004 0.01129010 + layer.4.k_cache 0.00067358 0.20982738 + layer.4.v_cache 0.00005212 0.02059055 + layer.4.output 0.00886795 300.49568289 + ------------------------------------------------------------------------------------- + TOTAL 0.02462613 127.30028814 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 228832 +BPFP 1.1556 bits/point +EBPFP 2.3112 equivalent bits/point +MSE 127.300288 +---------------------- -------------------------------------------------------- +Time: 0.732s Load: 0.007s, Pack+Encode: 0.367s, Decode+Unpack: 0.358s +---------------------- -------------------------------------------------------- +💾 Converting with 127.3003 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,892B, BPFP=0.5056 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,708B, BPFP=2.1059 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,624B, BPFP=0.9261 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,788B, BPFP=2.0384 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,232B, BPFP=1.1907 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,256B, BPFP=1.9994 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,296B, BPFP=1.1221 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,844B, BPFP=2.0425 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,148B, BPFP=1.6981 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,464B, BPFP=1.9413 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 57,948B, BPFP=0.6073 +⌛️ [2/4] FRONTEND: Frontend time: 0.300s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.443s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15089922 50.76023327 + layer.0.v_cache 0.00001686 0.01006826 + layer.1.k_cache 0.31217133 4.18509585 + layer.1.v_cache 0.00000593 0.00417199 + layer.2.k_cache 0.01305961 0.76002825 + layer.2.v_cache 0.00001999 0.01054185 + layer.3.k_cache 0.01178671 3.16661623 + layer.3.v_cache 0.00002039 0.01197033 + layer.4.k_cache 0.00067606 0.21043865 + layer.4.v_cache 0.00005285 0.02129279 + layer.4.output 1.43724915 253.30464034 + ------------------------------------------------------------------------------------- + TOTAL 0.62055606 107.78076117 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 270200 +BPFP 1.1659 bits/point +EBPFP 2.3319 equivalent bits/point +MSE 107.780761 +---------------------- -------------------------------------------------------- +Time: 0.750s Load: 0.008s, Pack+Encode: 0.300s, Decode+Unpack: 0.443s +---------------------- -------------------------------------------------------- +💾 Converting with 107.7808 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 187, 128) +Output shape: (1, 187, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.output: torch.Size([1, 187, 3584]) -> torch.Size([1, 1, 187, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,024B, BPFP=0.5033 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,224B, BPFP=2.0241 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,092B, BPFP=0.9268 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,432B, BPFP=1.9579 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,516B, BPFP=1.2129 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,240B, BPFP=1.9418 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,648B, BPFP=1.1404 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,712B, BPFP=1.9813 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,840B, BPFP=1.6578 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,636B, BPFP=1.8914 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 57,168B, BPFP=0.6824 +⌛️ [2/4] FRONTEND: Frontend time: 0.280s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.361s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08258281 52.90594711 + layer.0.v_cache 0.00001733 0.01070529 + layer.1.k_cache 0.10049684 4.31648393 + layer.1.v_cache 0.00000647 0.00392865 + layer.2.k_cache 0.01472783 0.73626317 + layer.2.v_cache 0.00002140 0.01127408 + layer.3.k_cache 0.02125903 3.22499713 + layer.3.v_cache 0.00002180 0.01219059 + layer.4.k_cache 0.00070831 0.21771897 + layer.4.v_cache 0.00005280 0.02192669 + layer.4.output 0.00866398 291.25272154 + ------------------------------------------------------------------------------------- + TOTAL 0.01650250 123.54296979 + (elements=1,627,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1627648 +Total Bytes 239532 +BPFP 1.1773 bits/point +EBPFP 2.3546 equivalent bits/point +MSE 123.542970 +---------------------- -------------------------------------------------------- +Time: 0.649s Load: 0.008s, Pack+Encode: 0.280s, Decode+Unpack: 0.361s +---------------------- -------------------------------------------------------- +💾 Converting with 123.5430 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,852B, BPFP=0.5074 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,720B, BPFP=2.1268 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,592B, BPFP=0.9325 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,824B, BPFP=2.0604 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,376B, BPFP=1.2127 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,300B, BPFP=2.0216 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,444B, BPFP=1.1437 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,808B, BPFP=2.0592 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,176B, BPFP=1.7162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,648B, BPFP=1.9733 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,328B, BPFP=0.6594 +⌛️ [2/4] FRONTEND: Frontend time: 0.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.438s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11071491 51.49386293 + layer.0.v_cache 0.00001804 0.00989076 + layer.1.k_cache 0.26784508 4.37601648 + layer.1.v_cache 0.00000610 0.00385021 + layer.2.k_cache 0.01369454 0.75429329 + layer.2.v_cache 0.00002102 0.00992232 + layer.3.k_cache 0.01444715 3.15114908 + layer.3.v_cache 0.00002041 0.01128481 + layer.4.k_cache 0.00067564 0.20437510 + layer.4.v_cache 0.00005248 0.02045119 + layer.4.output 1.45091435 255.63234174 + ------------------------------------------------------------------------------------- + TOTAL 0.62140564 108.79185226 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 275068 +BPFP 1.1982 bits/point +EBPFP 2.3964 equivalent bits/point +MSE 108.791852 +---------------------- -------------------------------------------------------- +Time: 0.779s Load: 0.007s, Pack+Encode: 0.334s, Decode+Unpack: 0.438s +---------------------- -------------------------------------------------------- +💾 Converting with 108.7919 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,868B, BPFP=0.4812 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,612B, BPFP=2.0748 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,104B, BPFP=0.9182 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,760B, BPFP=2.0151 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,784B, BPFP=1.1760 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,996B, BPFP=1.9616 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,780B, BPFP=1.1057 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,600B, BPFP=2.0039 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,972B, BPFP=1.6797 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,332B, BPFP=1.9151 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 60,484B, BPFP=0.6054 +⌛️ [2/4] FRONTEND: Frontend time: 0.301s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.394s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13389712 52.97768796 + layer.0.v_cache 0.00001748 0.00961721 + layer.1.k_cache 0.36644789 4.66732501 + layer.1.v_cache 0.00000587 0.00370455 + layer.2.k_cache 0.02260907 0.75354154 + layer.2.v_cache 0.00001979 0.00992563 + layer.3.k_cache 0.02550721 3.03467426 + layer.3.v_cache 0.00001968 0.01089733 + layer.4.k_cache 0.00067529 0.19922395 + layer.4.v_cache 0.00006705 0.02034246 + layer.4.output 1.37284074 241.30537316 + ------------------------------------------------------------------------------------- + TOTAL 0.59759716 102.98967953 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 279292 +BPFP 1.1511 bits/point +EBPFP 2.3023 equivalent bits/point +MSE 102.989680 +---------------------- -------------------------------------------------------- +Time: 0.702s Load: 0.007s, Pack+Encode: 0.301s, Decode+Unpack: 0.394s +---------------------- -------------------------------------------------------- +💾 Converting with 102.9897 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 406, 128) +Output shape: (1, 406, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.output: torch.Size([1, 406, 3584]) -> torch.Size([1, 1, 406, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,656B, BPFP=0.4871 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 52,816B, BPFP=2.0326 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,788B, BPFP=0.8770 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,996B, BPFP=1.9626 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,636B, BPFP=1.1405 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 50,196B, BPFP=1.9318 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,572B, BPFP=1.0611 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 50,876B, BPFP=1.9580 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,888B, BPFP=1.6121 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 48,560B, BPFP=1.8688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 112,804B, BPFP=0.6202 +⌛️ [2/4] FRONTEND: Frontend time: 0.579s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.630s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10474718 51.32662408 + layer.0.v_cache 0.00001775 0.00980678 + layer.1.k_cache 0.89873824 3.91413586 + layer.1.v_cache 0.00000687 0.00381773 + layer.2.k_cache 0.02602503 0.70626072 + layer.2.v_cache 0.00002164 0.01027669 + layer.3.k_cache 0.01558245 2.74637818 + layer.3.v_cache 0.00002201 0.01080622 + layer.4.k_cache 0.00072958 0.19842822 + layer.4.v_cache 0.00005327 0.01973681 + layer.4.output 0.00651461 135.97048733 + ------------------------------------------------------------------------------------- + TOTAL 0.06420861 59.45527545 + (elements=3,533,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3533824 +Total Bytes 500788 +BPFP 1.1337 bits/point +EBPFP 2.2674 equivalent bits/point +MSE 59.455275 +---------------------- -------------------------------------------------------- +Time: 1.222s Load: 0.014s, Pack+Encode: 0.579s, Decode+Unpack: 0.630s +---------------------- -------------------------------------------------------- +💾 Converting with 59.4553 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,752B, BPFP=0.4944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,652B, BPFP=2.0186 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,076B, BPFP=0.8977 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,000B, BPFP=1.9770 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,104B, BPFP=1.1546 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,212B, BPFP=1.9268 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,968B, BPFP=1.0821 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,660B, BPFP=1.9554 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,748B, BPFP=1.6421 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,376B, BPFP=1.8735 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,532B, BPFP=0.6153 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.430s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14414167 50.37386001 + layer.0.v_cache 0.00001717 0.00964189 + layer.1.k_cache 0.44179790 4.16437291 + layer.1.v_cache 0.00000667 0.00378180 + layer.2.k_cache 0.03016641 0.72703808 + layer.2.v_cache 0.00002043 0.01018536 + layer.3.k_cache 0.03726396 3.14881392 + layer.3.v_cache 0.00002099 0.01129157 + layer.4.k_cache 0.00070653 0.20265403 + layer.4.v_cache 0.00004910 0.01993052 + layer.4.output 1.24963187 219.80034621 + ------------------------------------------------------------------------------------- + TOTAL 0.55303611 93.95729374 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 303080 +BPFP 1.1370 bits/point +EBPFP 2.2740 equivalent bits/point +MSE 93.957294 +---------------------- -------------------------------------------------------- +Time: 0.758s Load: 0.009s, Pack+Encode: 0.319s, Decode+Unpack: 0.430s +---------------------- -------------------------------------------------------- +💾 Converting with 93.9573 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,376B, BPFP=0.5006 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,480B, BPFP=2.1577 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,292B, BPFP=0.9651 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,556B, BPFP=2.0851 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,904B, BPFP=1.2487 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,144B, BPFP=2.0528 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,884B, BPFP=1.1687 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,684B, BPFP=2.0952 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,228B, BPFP=1.7453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,520B, BPFP=2.0038 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 54,468B, BPFP=0.6110 +⌛️ [2/4] FRONTEND: Frontend time: 0.306s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.408s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15985105 52.88198316 + layer.0.v_cache 0.00001905 0.01085501 + layer.1.k_cache 0.22745673 4.64472724 + layer.1.v_cache 0.00000590 0.00401684 + layer.2.k_cache 0.01549998 0.76403747 + layer.2.v_cache 0.00002003 0.01066008 + layer.3.k_cache 0.01333464 3.25108084 + layer.3.v_cache 0.00002034 0.01223321 + layer.4.k_cache 0.00067076 0.21889824 + layer.4.v_cache 0.00004949 0.02188470 + layer.4.output 1.53832061 271.32351490 + ------------------------------------------------------------------------------------- + TOTAL 0.65795131 115.35794006 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 258536 +BPFP 1.1941 bits/point +EBPFP 2.3882 equivalent bits/point +MSE 115.357940 +---------------------- -------------------------------------------------------- +Time: 0.721s Load: 0.007s, Pack+Encode: 0.306s, Decode+Unpack: 0.408s +---------------------- -------------------------------------------------------- +💾 Converting with 115.3579 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 237, 128) +Output shape: (1, 237, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.output: torch.Size([1, 237, 3584]) -> torch.Size([1, 1, 237, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,528B, BPFP=0.4963 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,408B, BPFP=2.0707 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,896B, BPFP=0.9161 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,540B, BPFP=2.0134 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,700B, BPFP=1.1669 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,928B, BPFP=1.9731 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,604B, BPFP=1.0947 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,152B, BPFP=1.9879 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,284B, BPFP=1.6669 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,056B, BPFP=1.9156 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,600B, BPFP=0.6084 +⌛️ [2/4] FRONTEND: Frontend time: 0.321s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.461s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351868 52.74461036 + layer.0.v_cache 0.00001655 0.00964788 + layer.1.k_cache 0.34375901 4.45688085 + layer.1.v_cache 0.00000619 0.00375384 + layer.2.k_cache 0.01282014 0.72405381 + layer.2.v_cache 0.00002104 0.00999389 + layer.3.k_cache 0.03241006 3.03132082 + layer.3.v_cache 0.00001985 0.01098927 + layer.4.k_cache 0.00067667 0.20284191 + layer.4.v_cache 0.00005279 0.01995473 + layer.4.output 1.29178675 227.58331450 + ------------------------------------------------------------------------------------- + TOTAL 0.56210637 97.31160287 + (elements=2,062,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2062848 +Total Bytes 296696 +BPFP 1.1506 bits/point +EBPFP 2.3013 equivalent bits/point +MSE 97.311603 +---------------------- -------------------------------------------------------- +Time: 0.792s Load: 0.009s, Pack+Encode: 0.321s, Decode+Unpack: 0.461s +---------------------- -------------------------------------------------------- +💾 Converting with 97.3116 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,860B, BPFP=0.4985 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,920B, BPFP=2.1017 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,848B, BPFP=0.9337 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,140B, BPFP=2.0451 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,288B, BPFP=1.1837 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,480B, BPFP=1.9971 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,460B, BPFP=1.1235 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,812B, BPFP=2.0212 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,368B, BPFP=1.6983 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,576B, BPFP=1.9314 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 58,052B, BPFP=0.6027 +⌛️ [2/4] FRONTEND: Frontend time: 0.314s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11023432 52.98267169 + layer.0.v_cache 0.00001677 0.00916094 + layer.1.k_cache 0.30291741 4.51192712 + layer.1.v_cache 0.00000597 0.00362457 + layer.2.k_cache 0.03464115 0.76274464 + layer.2.v_cache 0.00001998 0.00978026 + layer.3.k_cache 0.01330193 3.00678995 + layer.3.v_cache 0.00002112 0.01069998 + layer.4.k_cache 0.00069379 0.20308192 + layer.4.v_cache 0.00005021 0.01983296 + layer.4.output 1.42388847 250.80836794 + ------------------------------------------------------------------------------------- + TOTAL 0.61347776 106.89287586 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 271804 +BPFP 1.1620 bits/point +EBPFP 2.3239 equivalent bits/point +MSE 106.892876 +---------------------- -------------------------------------------------------- +Time: 0.763s Load: 0.010s, Pack+Encode: 0.314s, Decode+Unpack: 0.439s +---------------------- -------------------------------------------------------- +💾 Converting with 106.8929 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 231, 128) +Output shape: (1, 231, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.output: torch.Size([1, 231, 3584]) -> torch.Size([1, 1, 231, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,336B, BPFP=0.4962 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,924B, BPFP=2.0917 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,804B, BPFP=0.9337 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,076B, BPFP=2.0344 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,468B, BPFP=1.1815 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,500B, BPFP=1.9954 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,624B, BPFP=1.1245 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,860B, BPFP=2.0198 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,180B, BPFP=1.7032 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,708B, BPFP=1.9418 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,216B, BPFP=0.6592 +⌛️ [2/4] FRONTEND: Frontend time: 0.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.435s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13257201 52.64839522 + layer.0.v_cache 0.00001701 0.00968734 + layer.1.k_cache 0.40462213 4.46153834 + layer.1.v_cache 0.00000658 0.00386365 + layer.2.k_cache 0.02649644 0.76245249 + layer.2.v_cache 0.00002056 0.00985585 + layer.3.k_cache 0.06147716 3.08552796 + layer.3.v_cache 0.00002105 0.01158507 + layer.4.k_cache 0.00069437 0.20510861 + layer.4.v_cache 0.00005136 0.02007384 + layer.4.output 1.32536713 233.33240569 + ------------------------------------------------------------------------------------- + TOTAL 0.58256168 99.67911342 + (elements=2,010,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2010624 +Total Bytes 297696 +BPFP 1.1845 bits/point +EBPFP 2.3690 equivalent bits/point +MSE 99.679113 +---------------------- -------------------------------------------------------- +Time: 0.767s Load: 0.008s, Pack+Encode: 0.324s, Decode+Unpack: 0.435s +---------------------- -------------------------------------------------------- +💾 Converting with 99.6791 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,220B, BPFP=0.5037 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,224B, BPFP=1.9745 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,744B, BPFP=0.9034 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,328B, BPFP=1.9196 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,300B, BPFP=1.1826 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,980B, BPFP=1.8983 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,152B, BPFP=1.1123 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,348B, BPFP=1.9208 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,400B, BPFP=1.6176 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,980B, BPFP=1.8370 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,844B, BPFP=0.6376 +⌛️ [2/4] FRONTEND: Frontend time: 0.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.431s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12841698 47.80433517 + layer.0.v_cache 0.00001718 0.01037556 + layer.1.k_cache 0.45450200 4.52682770 + layer.1.v_cache 0.00000627 0.00401112 + layer.2.k_cache 0.01434461 0.70668742 + layer.2.v_cache 0.00002215 0.01079742 + layer.3.k_cache 0.03530993 3.12371324 + layer.3.v_cache 0.00002139 0.01200013 + layer.4.k_cache 0.00068344 0.20933296 + layer.4.v_cache 0.00005073 0.02169537 + layer.4.output 1.20064817 196.44833683 + ------------------------------------------------------------------------------------- + TOTAL 0.53164187 84.20989023 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 315520 +BPFP 1.1373 bits/point +EBPFP 2.2745 equivalent bits/point +MSE 84.209890 +---------------------- -------------------------------------------------------- +Time: 0.749s Load: 0.008s, Pack+Encode: 0.309s, Decode+Unpack: 0.431s +---------------------- -------------------------------------------------------- +💾 Converting with 84.2099 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,844B, BPFP=0.4900 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,176B, BPFP=2.0598 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,416B, BPFP=0.9096 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,144B, BPFP=2.0027 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,240B, BPFP=1.1769 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,344B, BPFP=1.9583 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,740B, BPFP=1.0938 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,740B, BPFP=1.9803 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,168B, BPFP=1.6715 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,444B, BPFP=1.9085 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 76,096B, BPFP=0.6023 +⌛️ [2/4] FRONTEND: Frontend time: 0.354s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.491s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13731835 53.39158078 + layer.0.v_cache 0.00001709 0.00924566 + layer.1.k_cache 0.55879569 4.33738633 + layer.1.v_cache 0.00000637 0.00359690 + layer.2.k_cache 0.01348169 0.76308284 + layer.2.v_cache 0.00002375 0.01000814 + layer.3.k_cache 0.01306405 2.85312820 + layer.3.v_cache 0.00002005 0.01038602 + layer.4.k_cache 0.00072533 0.20124015 + layer.4.v_cache 0.00005384 0.02016739 + layer.4.output 0.00472113 195.43275076 + ------------------------------------------------------------------------------------- + TOTAL 0.04450318 84.09582810 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 351352 +BPFP 1.1452 bits/point +EBPFP 2.2903 equivalent bits/point +MSE 84.095828 +---------------------- -------------------------------------------------------- +Time: 0.854s Load: 0.010s, Pack+Encode: 0.354s, Decode+Unpack: 0.491s +---------------------- -------------------------------------------------------- +💾 Converting with 84.0958 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,360B, BPFP=0.4859 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,436B, BPFP=2.0471 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,340B, BPFP=0.9001 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,536B, BPFP=2.0004 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,524B, BPFP=1.1692 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,680B, BPFP=1.9560 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,360B, BPFP=1.1088 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,272B, BPFP=1.9867 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,984B, BPFP=1.6603 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,500B, BPFP=1.8947 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 80,260B, BPFP=0.5952 +⌛️ [2/4] FRONTEND: Frontend time: 0.346s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.530s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13972237 51.57501038 + layer.0.v_cache 0.00001689 0.00984500 + layer.1.k_cache 0.51719427 4.39640813 + layer.1.v_cache 0.00000613 0.00407235 + layer.2.k_cache 0.02019965 0.75641837 + layer.2.v_cache 0.00002053 0.01057111 + layer.3.k_cache 0.01759115 3.10980995 + layer.3.v_cache 0.00002125 0.01161777 + layer.4.k_cache 0.00069604 0.20749408 + layer.4.v_cache 0.00005872 0.02088289 + layer.4.output 0.04434862 179.56595574 + ------------------------------------------------------------------------------------- + TOTAL 0.05917455 77.47434237 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 373252 +BPFP 1.1397 bits/point +EBPFP 2.2795 equivalent bits/point +MSE 77.474342 +---------------------- -------------------------------------------------------- +Time: 0.885s Load: 0.010s, Pack+Encode: 0.346s, Decode+Unpack: 0.530s +---------------------- -------------------------------------------------------- +💾 Converting with 77.4743 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,396B, BPFP=0.5022 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,420B, BPFP=2.1530 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,324B, BPFP=0.9677 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,640B, BPFP=2.0917 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,628B, BPFP=1.2271 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,908B, BPFP=2.0342 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,664B, BPFP=1.1514 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,472B, BPFP=2.0785 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,112B, BPFP=1.7362 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,364B, BPFP=1.9915 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 55,056B, BPFP=0.6176 +⌛️ [2/4] FRONTEND: Frontend time: 0.322s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.472s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13044639 52.49376767 + layer.0.v_cache 0.00001737 0.00988007 + layer.1.k_cache 0.22565776 4.30267794 + layer.1.v_cache 0.00000620 0.00400884 + layer.2.k_cache 0.01998205 0.76992008 + layer.2.v_cache 0.00001989 0.01020325 + layer.3.k_cache 0.00629339 3.34958545 + layer.3.v_cache 0.00002050 0.01169584 + layer.4.k_cache 0.00068118 0.21015069 + layer.4.v_cache 0.00005692 0.02091492 + layer.4.output 1.53833376 271.17749462 + ------------------------------------------------------------------------------------- + TOTAL 0.65597165 115.26030983 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 257984 +BPFP 1.1915 bits/point +EBPFP 2.3831 equivalent bits/point +MSE 115.260310 +---------------------- -------------------------------------------------------- +Time: 0.801s Load: 0.007s, Pack+Encode: 0.322s, Decode+Unpack: 0.472s +---------------------- -------------------------------------------------------- +💾 Converting with 115.2603 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 246, 128) +Output shape: (1, 246, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.output: torch.Size([1, 246, 3584]) -> torch.Size([1, 1, 246, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,728B, BPFP=0.4909 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,708B, BPFP=2.0140 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,064B, BPFP=0.8933 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,860B, BPFP=1.9601 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,160B, BPFP=1.1535 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,208B, BPFP=1.9187 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,160B, BPFP=1.0899 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,620B, BPFP=1.9449 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,948B, BPFP=1.6481 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,252B, BPFP=1.8580 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,372B, BPFP=0.5841 +⌛️ [2/4] FRONTEND: Frontend time: 0.325s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.445s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131485 51.32998682 + layer.0.v_cache 0.00001613 0.01015588 + layer.1.k_cache 0.49701033 4.46973152 + layer.1.v_cache 0.00000620 0.00403026 + layer.2.k_cache 0.01298508 0.78125043 + layer.2.v_cache 0.00002120 0.00995445 + layer.3.k_cache 0.02762915 3.07810168 + layer.3.v_cache 0.00002136 0.01169615 + layer.4.k_cache 0.00068392 0.20654362 + layer.4.v_cache 0.00006372 0.01978090 + layer.4.output 1.24453647 219.23274172 + ------------------------------------------------------------------------------------- + TOTAL 0.55302984 93.79708375 + (elements=2,141,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2141184 +Total Bytes 300080 +BPFP 1.1212 bits/point +EBPFP 2.2423 equivalent bits/point +MSE 93.797084 +---------------------- -------------------------------------------------------- +Time: 0.781s Load: 0.011s, Pack+Encode: 0.325s, Decode+Unpack: 0.445s +---------------------- -------------------------------------------------------- +💾 Converting with 93.7971 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,400B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,272B, BPFP=2.1306 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,204B, BPFP=0.9534 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,524B, BPFP=2.0722 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,652B, BPFP=1.2228 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,072B, BPFP=2.0369 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,876B, BPFP=1.1622 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,488B, BPFP=2.0694 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,464B, BPFP=1.7550 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,564B, BPFP=1.9972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 52,472B, BPFP=0.5856 +⌛️ [2/4] FRONTEND: Frontend time: 0.263s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.408s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11599884 53.25801270 + layer.0.v_cache 0.00001671 0.00966967 + layer.1.k_cache 0.22559120 4.69839661 + layer.1.v_cache 0.00000595 0.00400897 + layer.2.k_cache 0.01146694 0.71060753 + layer.2.v_cache 0.00002201 0.01101702 + layer.3.k_cache 0.03668072 3.17014343 + layer.3.v_cache 0.00002131 0.01163533 + layer.4.k_cache 0.00066762 0.20884674 + layer.4.v_cache 0.00005146 0.02138695 + layer.4.output 1.53061454 269.89308036 + ------------------------------------------------------------------------------------- + TOTAL 0.65322497 114.78560514 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 255988 +BPFP 1.1764 bits/point +EBPFP 2.3528 equivalent bits/point +MSE 114.785605 +---------------------- -------------------------------------------------------- +Time: 0.679s Load: 0.008s, Pack+Encode: 0.263s, Decode+Unpack: 0.408s +---------------------- -------------------------------------------------------- +💾 Converting with 114.7856 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,000B, BPFP=0.4776 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,660B, BPFP=2.0920 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,492B, BPFP=0.9206 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,820B, BPFP=2.0347 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,272B, BPFP=1.1785 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,060B, BPFP=1.9828 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,172B, BPFP=1.1034 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,556B, BPFP=2.0166 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,804B, BPFP=1.6924 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,432B, BPFP=1.9400 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,052B, BPFP=0.6731 +⌛️ [2/4] FRONTEND: Frontend time: 0.303s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.418s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605523 51.93665564 + layer.0.v_cache 0.00001866 0.01013266 + layer.1.k_cache 0.31881074 4.69669525 + layer.1.v_cache 0.00000653 0.00391265 + layer.2.k_cache 0.02091764 0.76361031 + layer.2.v_cache 0.00002113 0.01071233 + layer.3.k_cache 0.03221333 3.24539704 + layer.3.v_cache 0.00002185 0.01181910 + layer.4.k_cache 0.00068277 0.21410047 + layer.4.v_cache 0.00005596 0.02107321 + layer.4.output 1.33695214 235.78290705 + ------------------------------------------------------------------------------------- + TOTAL 0.57926287 100.67026224 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 295320 +BPFP 1.1853 bits/point +EBPFP 2.3706 equivalent bits/point +MSE 100.670262 +---------------------- -------------------------------------------------------- +Time: 0.729s Load: 0.008s, Pack+Encode: 0.303s, Decode+Unpack: 0.418s +---------------------- -------------------------------------------------------- +💾 Converting with 100.6703 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 192, 128) +Output shape: (1, 192, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.output: torch.Size([1, 192, 3584]) -> torch.Size([1, 1, 192, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,680B, BPFP=0.4622 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,628B, BPFP=1.9229 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,336B, BPFP=0.8411 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,084B, BPFP=1.8786 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,608B, BPFP=1.1074 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,704B, BPFP=1.8477 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,848B, BPFP=1.0456 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,028B, BPFP=1.8740 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,568B, BPFP=1.5924 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,120B, BPFP=1.8001 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 50,380B, BPFP=0.5857 +⌛️ [2/4] FRONTEND: Frontend time: 0.276s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12444629 47.07043457 + layer.0.v_cache 0.00001540 0.00924497 + layer.1.k_cache 0.18411521 4.34900093 + layer.1.v_cache 0.00000585 0.00374853 + layer.2.k_cache 0.01124177 0.61975543 + layer.2.v_cache 0.00002179 0.01053931 + layer.3.k_cache 0.01616090 2.56215827 + layer.3.v_cache 0.00002117 0.01145280 + layer.4.k_cache 0.00068507 0.20503521 + layer.4.v_cache 0.00005275 0.02056262 + layer.4.output 0.00839530 266.55457124 + ------------------------------------------------------------------------------------- + TOTAL 0.02326666 112.98493714 + (elements=1,671,168) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1671168 +Total Bytes 226984 +BPFP 1.0866 bits/point +EBPFP 2.1732 equivalent bits/point +MSE 112.984937 +---------------------- -------------------------------------------------------- +Time: 0.633s Load: 0.006s, Pack+Encode: 0.276s, Decode+Unpack: 0.351s +---------------------- -------------------------------------------------------- +💾 Converting with 112.9849 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,460B, BPFP=0.5139 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,804B, BPFP=2.1465 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,292B, BPFP=0.9688 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,144B, BPFP=2.0843 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,164B, BPFP=1.2391 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,652B, BPFP=2.0380 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,324B, BPFP=1.1600 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,996B, BPFP=2.0704 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,880B, BPFP=1.7771 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,260B, BPFP=2.0011 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,132B, BPFP=0.5934 +⌛️ [2/4] FRONTEND: Frontend time: 0.275s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09999813 52.85346856 + layer.0.v_cache 0.00001899 0.01021889 + layer.1.k_cache 0.12989934 4.56469028 + layer.1.v_cache 0.00000610 0.00406550 + layer.2.k_cache 0.00966061 0.79759419 + layer.2.v_cache 0.00001994 0.01110495 + layer.3.k_cache 0.03384685 3.22148849 + layer.3.v_cache 0.00002040 0.01208016 + layer.4.k_cache 0.00066502 0.21448411 + layer.4.v_cache 0.00005200 0.02169259 + layer.4.output 0.00959846 329.81091330 + ------------------------------------------------------------------------------------- + TOTAL 0.02008098 139.43454593 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 214108 +BPFP 1.1855 bits/point +EBPFP 2.3710 equivalent bits/point +MSE 139.434546 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.006s, Pack+Encode: 0.275s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 139.4345 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 428, 128) +Output shape: (1, 428, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.output: torch.Size([1, 428, 3584]) -> torch.Size([1, 1, 428, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,004B, BPFP=0.4747 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,212B, BPFP=2.0156 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,976B, BPFP=0.8753 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,436B, BPFP=1.9508 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,244B, BPFP=1.1406 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,472B, BPFP=1.9156 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,160B, BPFP=1.0645 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,136B, BPFP=1.9398 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,448B, BPFP=1.6227 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 50,980B, BPFP=1.8611 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 112,900B, BPFP=0.5888 +⌛️ [2/4] FRONTEND: Frontend time: 0.422s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.596s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13760449 50.92742863 + layer.0.v_cache 0.00001624 0.00922640 + layer.1.k_cache 0.89392254 4.17160975 + layer.1.v_cache 0.00000615 0.00353847 + layer.2.k_cache 0.03917003 0.72132831 + layer.2.v_cache 0.00002113 0.00990567 + layer.3.k_cache 0.03749864 2.75814106 + layer.3.v_cache 0.00002059 0.01050655 + layer.4.k_cache 0.00086919 0.20414582 + layer.4.v_cache 0.00005177 0.02022692 + layer.4.output 0.00617707 128.94279873 + ------------------------------------------------------------------------------------- + TOTAL 0.06778943 56.55503816 + (elements=3,725,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3725312 +Total Bytes 519968 +BPFP 1.1166 bits/point +EBPFP 2.2332 equivalent bits/point +MSE 56.555038 +---------------------- -------------------------------------------------------- +Time: 1.031s Load: 0.013s, Pack+Encode: 0.422s, Decode+Unpack: 0.596s +---------------------- -------------------------------------------------------- +💾 Converting with 56.5550 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,940B, BPFP=0.4953 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,092B, BPFP=2.0552 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,252B, BPFP=0.9005 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,184B, BPFP=2.0049 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,040B, BPFP=1.1658 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,272B, BPFP=1.9543 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,772B, BPFP=1.0955 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,836B, BPFP=1.9856 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,824B, BPFP=1.6525 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,312B, BPFP=1.9012 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 79,484B, BPFP=0.6291 +⌛️ [2/4] FRONTEND: Frontend time: 0.393s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.465s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385377 53.39048648 + layer.0.v_cache 0.00001705 0.00906039 + layer.1.k_cache 0.49372572 3.98742806 + layer.1.v_cache 0.00000626 0.00347976 + layer.2.k_cache 0.02422354 0.69047633 + layer.2.v_cache 0.00002046 0.00940762 + layer.3.k_cache 0.04153839 2.53448053 + layer.3.v_cache 0.00002075 0.01025803 + layer.4.k_cache 0.00074884 0.19997225 + layer.4.v_cache 0.00004954 0.01907380 + layer.4.output 0.00474379 194.72835930 + ------------------------------------------------------------------------------------- + TOTAL 0.04161240 83.76191990 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 354008 +BPFP 1.1538 bits/point +EBPFP 2.3076 equivalent bits/point +MSE 83.761920 +---------------------- -------------------------------------------------------- +Time: 0.866s Load: 0.009s, Pack+Encode: 0.393s, Decode+Unpack: 0.465s +---------------------- -------------------------------------------------------- +💾 Converting with 83.7619 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,192B, BPFP=0.4930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,628B, BPFP=2.0621 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,732B, BPFP=0.9062 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,328B, BPFP=1.9992 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,872B, BPFP=1.1548 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,388B, BPFP=1.9538 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,716B, BPFP=1.0989 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,048B, BPFP=1.9857 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,148B, BPFP=1.6519 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,396B, BPFP=1.9058 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,236B, BPFP=0.6305 +⌛️ [2/4] FRONTEND: Frontend time: 0.466s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.550s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12985682 51.53230215 + layer.0.v_cache 0.00001645 0.00901890 + layer.1.k_cache 0.61935935 4.15355916 + layer.1.v_cache 0.00000656 0.00359914 + layer.2.k_cache 0.01492825 0.68258223 + layer.2.v_cache 0.00002178 0.00924415 + layer.3.k_cache 0.02579401 3.03329345 + layer.3.v_cache 0.00002068 0.00997998 + layer.4.k_cache 0.00071613 0.20005770 + layer.4.v_cache 0.00005057 0.01926694 + layer.4.output 0.04141478 167.38954003 + ------------------------------------------------------------------------------------- + TOTAL 0.06356906 72.43409906 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 405684 +BPFP 1.1544 bits/point +EBPFP 2.3088 equivalent bits/point +MSE 72.434099 +---------------------- -------------------------------------------------------- +Time: 1.027s Load: 0.010s, Pack+Encode: 0.466s, Decode+Unpack: 0.550s +---------------------- -------------------------------------------------------- +💾 Converting with 72.4341 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,852B, BPFP=0.4836 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,716B, BPFP=2.0605 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,864B, BPFP=0.9213 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,712B, BPFP=2.0057 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,748B, BPFP=1.1882 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,900B, BPFP=1.9613 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,360B, BPFP=1.1123 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,524B, BPFP=1.9954 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,476B, BPFP=1.6650 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,900B, BPFP=1.9067 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,088B, BPFP=0.6329 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.526s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12359551 51.55467384 + layer.0.v_cache 0.00001766 0.00965559 + layer.1.k_cache 0.52692803 4.40522948 + layer.1.v_cache 0.00000637 0.00381005 + layer.2.k_cache 0.02102186 0.74763382 + layer.2.v_cache 0.00002187 0.01008322 + layer.3.k_cache 0.03868174 3.18114146 + layer.3.v_cache 0.00002097 0.01153267 + layer.4.k_cache 0.00070799 0.20499718 + layer.4.v_cache 0.00005204 0.02045267 + layer.4.output 0.00469152 193.46843781 + ------------------------------------------------------------------------------------- + TOTAL 0.04375851 83.20166322 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 361140 +BPFP 1.1606 bits/point +EBPFP 2.3212 equivalent bits/point +MSE 83.201663 +---------------------- -------------------------------------------------------- +Time: 0.877s Load: 0.009s, Pack+Encode: 0.342s, Decode+Unpack: 0.526s +---------------------- -------------------------------------------------------- +💾 Converting with 83.2017 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,316B, BPFP=0.4959 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,468B, BPFP=2.1567 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,340B, BPFP=0.9689 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,700B, BPFP=2.0964 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,788B, BPFP=1.2396 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,152B, BPFP=2.0534 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,868B, BPFP=1.1674 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,624B, BPFP=2.0905 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,124B, BPFP=1.7371 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,416B, BPFP=1.9956 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 57,776B, BPFP=0.6481 +⌛️ [2/4] FRONTEND: Frontend time: 0.312s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.416s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10671347 52.09481489 + layer.0.v_cache 0.00001788 0.01023033 + layer.1.k_cache 0.15257932 4.25159213 + layer.1.v_cache 0.00000642 0.00383828 + layer.2.k_cache 0.02197224 0.74680984 + layer.2.v_cache 0.00002094 0.01003935 + layer.3.k_cache 0.03325213 3.03817396 + layer.3.v_cache 0.00002126 0.01178975 + layer.4.k_cache 0.00067229 0.20579345 + layer.4.v_cache 0.00005023 0.01984105 + layer.4.output 1.53834953 271.21058417 + ------------------------------------------------------------------------------------- + TOTAL 0.65198546 115.22747131 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 261572 +BPFP 1.2081 bits/point +EBPFP 2.4162 equivalent bits/point +MSE 115.227471 +---------------------- -------------------------------------------------------- +Time: 0.735s Load: 0.007s, Pack+Encode: 0.312s, Decode+Unpack: 0.416s +---------------------- -------------------------------------------------------- +💾 Converting with 115.2275 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,772B, BPFP=0.4838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,188B, BPFP=2.0037 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,684B, BPFP=0.9141 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,444B, BPFP=1.9574 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,028B, BPFP=1.1845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,876B, BPFP=1.9221 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,052B, BPFP=1.1238 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,296B, BPFP=1.9482 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,304B, BPFP=1.6375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,816B, BPFP=1.8561 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,024B, BPFP=0.6405 +⌛️ [2/4] FRONTEND: Frontend time: 0.297s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.405s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15214830 51.74615600 + layer.0.v_cache 0.00001722 0.01051135 + layer.1.k_cache 0.49754258 4.10224677 + layer.1.v_cache 0.00000639 0.00421919 + layer.2.k_cache 0.01862443 0.73044261 + layer.2.v_cache 0.00002219 0.01065995 + layer.3.k_cache 0.01200796 2.80110661 + layer.3.v_cache 0.00002210 0.01188570 + layer.4.k_cache 0.00075226 0.21376998 + layer.4.v_cache 0.00005409 0.02160729 + layer.4.output 1.21979868 214.05181061 + ------------------------------------------------------------------------------------- + TOTAL 0.54234049 91.64795763 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 313484 +BPFP 1.1479 bits/point +EBPFP 2.2958 equivalent bits/point +MSE 91.647958 +---------------------- -------------------------------------------------------- +Time: 0.714s Load: 0.012s, Pack+Encode: 0.297s, Decode+Unpack: 0.405s +---------------------- -------------------------------------------------------- +💾 Converting with 91.6480 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,504B, BPFP=0.4869 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,508B, BPFP=2.0240 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,408B, BPFP=0.8918 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,504B, BPFP=1.9725 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,712B, BPFP=1.1635 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,720B, BPFP=1.9324 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,332B, BPFP=1.0928 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,300B, BPFP=1.9621 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,124B, BPFP=1.6457 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,648B, BPFP=1.8775 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 79,352B, BPFP=0.5807 +⌛️ [2/4] FRONTEND: Frontend time: 0.370s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.493s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18432382 52.57617828 + layer.0.v_cache 0.00001629 0.00988802 + layer.1.k_cache 0.53789903 4.27008397 + layer.1.v_cache 0.00000607 0.00397790 + layer.2.k_cache 0.01514517 0.76362235 + layer.2.v_cache 0.00002151 0.01060672 + layer.3.k_cache 0.04182221 3.08525150 + layer.3.v_cache 0.00002017 0.01144158 + layer.4.k_cache 0.00068627 0.20877339 + layer.4.v_cache 0.00005063 0.02105868 + layer.4.output 0.04375917 177.03116218 + ------------------------------------------------------------------------------------- + TOTAL 0.06390032 76.48111868 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 373112 +BPFP 1.1244 bits/point +EBPFP 2.2487 equivalent bits/point +MSE 76.481119 +---------------------- -------------------------------------------------------- +Time: 0.873s Load: 0.010s, Pack+Encode: 0.370s, Decode+Unpack: 0.493s +---------------------- -------------------------------------------------------- +💾 Converting with 76.4811 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 371, 128) +Output shape: (1, 371, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.output: torch.Size([1, 371, 3584]) -> torch.Size([1, 1, 371, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,272B, BPFP=0.4747 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,640B, BPFP=2.0064 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,756B, BPFP=0.8742 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,380B, BPFP=1.9533 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,192B, BPFP=1.1452 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,532B, BPFP=1.9176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,632B, BPFP=1.0795 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,104B, BPFP=1.9417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,476B, BPFP=1.6205 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,112B, BPFP=1.8578 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,868B, BPFP=0.6009 +⌛️ [2/4] FRONTEND: Frontend time: 0.387s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.547s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16952868 50.99200324 + layer.0.v_cache 0.00001698 0.00984035 + layer.1.k_cache 0.82091874 4.31217656 + layer.1.v_cache 0.00000666 0.00403597 + layer.2.k_cache 0.02418445 0.78662554 + layer.2.v_cache 0.00002238 0.01065910 + layer.3.k_cache 0.03055777 2.82756269 + layer.3.v_cache 0.00002117 0.01135850 + layer.4.k_cache 0.00071103 0.20817434 + layer.4.v_cache 0.00005915 0.02128058 + layer.4.output 0.03610923 145.48428475 + ------------------------------------------------------------------------------------- + TOTAL 0.07639951 63.38668883 + (elements=3,229,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3229184 +Total Bytes 452964 +BPFP 1.1222 bits/point +EBPFP 2.2444 equivalent bits/point +MSE 63.386689 +---------------------- -------------------------------------------------------- +Time: 0.946s Load: 0.012s, Pack+Encode: 0.387s, Decode+Unpack: 0.547s +---------------------- -------------------------------------------------------- +💾 Converting with 63.3867 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,076B, BPFP=0.5072 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,404B, BPFP=2.1075 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,104B, BPFP=0.9392 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,588B, BPFP=2.0490 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,720B, BPFP=1.1984 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,864B, BPFP=1.9971 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,916B, BPFP=1.1408 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,460B, BPFP=2.0399 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,692B, BPFP=1.6981 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,008B, BPFP=1.9358 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,936B, BPFP=0.6444 +⌛️ [2/4] FRONTEND: Frontend time: 0.288s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13674593 52.35569273 + layer.0.v_cache 0.00001768 0.01063973 + layer.1.k_cache 0.26206151 4.36790172 + layer.1.v_cache 0.00000624 0.00410735 + layer.2.k_cache 0.02938290 0.80092173 + layer.2.v_cache 0.00002158 0.01076685 + layer.3.k_cache 0.05653211 3.41444523 + layer.3.v_cache 0.00002129 0.01193724 + layer.4.k_cache 0.00070776 0.21208674 + layer.4.v_cache 0.00005193 0.02175395 + layer.4.output 1.40434551 246.76445773 + ------------------------------------------------------------------------------------- + TOTAL 0.60682162 105.20949749 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 280768 +BPFP 1.1838 bits/point +EBPFP 2.3675 equivalent bits/point +MSE 105.209497 +---------------------- -------------------------------------------------------- +Time: 0.682s Load: 0.007s, Pack+Encode: 0.288s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 105.2095 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 193, 128) +Output shape: (1, 193, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.output: torch.Size([1, 193, 3584]) -> torch.Size([1, 1, 193, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,084B, BPFP=0.4926 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,628B, BPFP=2.1558 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,640B, BPFP=0.9424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,884B, BPFP=2.0955 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,224B, BPFP=1.2325 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,480B, BPFP=2.0628 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,360B, BPFP=1.1626 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 25,856B, BPFP=2.0933 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,652B, BPFP=1.7529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,696B, BPFP=1.9994 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 55,496B, BPFP=0.6418 +⌛️ [2/4] FRONTEND: Frontend time: 0.329s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.462s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14305212 54.05776392 + layer.0.v_cache 0.00001770 0.01042212 + layer.1.k_cache 0.14179654 4.94748478 + layer.1.v_cache 0.00000608 0.00406100 + layer.2.k_cache 0.01555968 0.80605723 + layer.2.v_cache 0.00002047 0.01092282 + layer.3.k_cache 0.02655400 3.08651243 + layer.3.v_cache 0.00002031 0.01227883 + layer.4.k_cache 0.00069483 0.21609029 + layer.4.v_cache 0.00005004 0.02225718 + layer.4.output 1.58615237 279.76216691 + ------------------------------------------------------------------------------------- + TOTAL 0.67240225 118.91229524 + (elements=1,679,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1679872 +Total Bytes 253000 +BPFP 1.2049 bits/point +EBPFP 2.4097 equivalent bits/point +MSE 118.912295 +---------------------- -------------------------------------------------------- +Time: 0.799s Load: 0.007s, Pack+Encode: 0.329s, Decode+Unpack: 0.462s +---------------------- -------------------------------------------------------- +💾 Converting with 118.9123 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,780B, BPFP=0.4882 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,724B, BPFP=1.9907 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,092B, BPFP=0.8843 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,848B, BPFP=1.9357 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,336B, BPFP=1.1506 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,468B, BPFP=1.9119 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,296B, BPFP=1.0853 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,864B, BPFP=1.9367 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,740B, BPFP=1.6152 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,436B, BPFP=1.8471 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,652B, BPFP=0.6244 +⌛️ [2/4] FRONTEND: Frontend time: 0.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.427s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10821181 52.01563677 + layer.0.v_cache 0.00001942 0.00907890 + layer.1.k_cache 0.30751975 4.37235759 + layer.1.v_cache 0.00000599 0.00328409 + layer.2.k_cache 0.01421910 0.70600634 + layer.2.v_cache 0.00002020 0.00938657 + layer.3.k_cache 0.01730646 2.83115323 + layer.3.v_cache 0.00002068 0.01071302 + layer.4.k_cache 0.00075739 0.19419692 + layer.4.v_cache 0.00005209 0.01945797 + layer.4.output 1.22958295 214.90463640 + ------------------------------------------------------------------------------------- + TOTAL 0.53265962 92.02963095 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 306236 +BPFP 1.1304 bits/point +EBPFP 2.2608 equivalent bits/point +MSE 92.029631 +---------------------- -------------------------------------------------------- +Time: 0.772s Load: 0.011s, Pack+Encode: 0.334s, Decode+Unpack: 0.427s +---------------------- -------------------------------------------------------- +💾 Converting with 92.0296 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 202, 128) +Output shape: (1, 202, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.output: torch.Size([1, 202, 3584]) -> torch.Size([1, 1, 202, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,468B, BPFP=0.5003 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,896B, BPFP=2.1578 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,240B, BPFP=0.9468 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,980B, BPFP=2.0869 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,908B, BPFP=1.2305 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,476B, BPFP=2.0480 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,044B, BPFP=1.1637 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,136B, BPFP=2.0990 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,488B, BPFP=1.7395 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,712B, BPFP=1.9889 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 59,924B, BPFP=0.6622 +⌛️ [2/4] FRONTEND: Frontend time: 0.310s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15282421 53.60172687 + layer.0.v_cache 0.00001801 0.01062023 + layer.1.k_cache 0.08597872 4.37176816 + layer.1.v_cache 0.00000649 0.00419244 + layer.2.k_cache 0.01425038 0.75039106 + layer.2.v_cache 0.00002047 0.01063905 + layer.3.k_cache 0.05704349 3.15979125 + layer.3.v_cache 0.00002256 0.01232947 + layer.4.k_cache 0.00069767 0.21879957 + layer.4.v_cache 0.00005056 0.02133847 + layer.4.output 1.51553867 267.17121199 + ------------------------------------------------------------------------------------- + TOTAL 0.64233431 113.66824003 + (elements=1,758,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1758208 +Total Bytes 266272 +BPFP 1.2116 bits/point +EBPFP 2.4231 equivalent bits/point +MSE 113.668240 +---------------------- -------------------------------------------------------- +Time: 0.729s Load: 0.007s, Pack+Encode: 0.310s, Decode+Unpack: 0.411s +---------------------- -------------------------------------------------------- +💾 Converting with 113.6682 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,984B, BPFP=0.4960 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,788B, BPFP=2.0311 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,292B, BPFP=0.8995 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,716B, BPFP=1.9720 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,692B, BPFP=1.1424 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,764B, BPFP=1.9194 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,588B, BPFP=1.0815 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,344B, BPFP=1.9514 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,468B, BPFP=1.6270 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,868B, BPFP=1.8699 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,648B, BPFP=0.5888 +⌛️ [2/4] FRONTEND: Frontend time: 0.343s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.481s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14837397 53.32439129 + layer.0.v_cache 0.00001603 0.00861022 + layer.1.k_cache 0.60911565 4.43932157 + layer.1.v_cache 0.00000597 0.00325297 + layer.2.k_cache 0.02477714 0.72038824 + layer.2.v_cache 0.00001938 0.00869467 + layer.3.k_cache 0.01052417 3.03228091 + layer.3.v_cache 0.00002039 0.00982773 + layer.4.k_cache 0.00073933 0.18501307 + layer.4.v_cache 0.00005963 0.01904485 + layer.4.output 0.00470031 195.08572060 + ------------------------------------------------------------------------------------- + TOTAL 0.04862081 83.96181586 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 346152 +BPFP 1.1242 bits/point +EBPFP 2.2484 equivalent bits/point +MSE 83.961816 +---------------------- -------------------------------------------------------- +Time: 0.833s Load: 0.009s, Pack+Encode: 0.343s, Decode+Unpack: 0.481s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9618 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 210, 128) +Output shape: (1, 210, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.output: torch.Size([1, 210, 3584]) -> torch.Size([1, 1, 210, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,800B, BPFP=0.5060 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,516B, BPFP=2.1217 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,464B, BPFP=0.9274 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,692B, BPFP=2.0604 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,472B, BPFP=1.2256 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,108B, BPFP=2.0170 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,344B, BPFP=1.1417 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,636B, BPFP=2.0562 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,104B, BPFP=1.7190 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,496B, BPFP=1.9714 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,824B, BPFP=0.6678 +⌛️ [2/4] FRONTEND: Frontend time: 0.305s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.425s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11415643 50.76237909 + layer.0.v_cache 0.00001882 0.00998363 + layer.1.k_cache 0.27145578 4.45185779 + layer.1.v_cache 0.00000623 0.00390586 + layer.2.k_cache 0.01201470 0.74101795 + layer.2.v_cache 0.00002208 0.01074540 + layer.3.k_cache 0.00887682 3.24201689 + layer.3.v_cache 0.00002165 0.01195807 + layer.4.k_cache 0.00072611 0.21422657 + layer.4.v_cache 0.00005233 0.02162198 + layer.4.output 1.45781997 256.62170493 + ------------------------------------------------------------------------------------- + TOTAL 0.62424063 109.16597928 + (elements=1,827,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1827840 +Total Bytes 274456 +BPFP 1.2012 bits/point +EBPFP 2.4025 equivalent bits/point +MSE 109.165979 +---------------------- -------------------------------------------------------- +Time: 0.737s Load: 0.007s, Pack+Encode: 0.305s, Decode+Unpack: 0.425s +---------------------- -------------------------------------------------------- +💾 Converting with 109.1660 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 165, 128) +Output shape: (1, 165, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.output: torch.Size([1, 165, 3584]) -> torch.Size([1, 1, 165, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,388B, BPFP=0.5102 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,000B, BPFP=2.1780 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,132B, BPFP=0.9595 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,416B, BPFP=2.1227 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,296B, BPFP=1.2591 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,796B, BPFP=2.0640 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,536B, BPFP=1.1871 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,300B, BPFP=2.1117 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,920B, BPFP=1.7917 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,328B, BPFP=2.0197 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 49,796B, BPFP=0.6736 +⌛️ [2/4] FRONTEND: Frontend time: 0.240s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.340s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10245823 53.82520123 + layer.0.v_cache 0.00002070 0.01078336 + layer.1.k_cache 0.13127365 5.15836663 + layer.1.v_cache 0.00000615 0.00415753 + layer.2.k_cache 0.01458293 0.80891835 + layer.2.v_cache 0.00002074 0.01091963 + layer.3.k_cache 0.02375244 3.30042688 + layer.3.v_cache 0.00002214 0.01301603 + layer.4.k_cache 0.00067489 0.22387596 + layer.4.v_cache 0.00005196 0.02241654 + layer.4.output 0.00973386 331.77746212 + ------------------------------------------------------------------------------------- + TOTAL 0.02005887 140.34237159 + (elements=1,436,160) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1436160 +Total Bytes 220908 +BPFP 1.2305 bits/point +EBPFP 2.4611 equivalent bits/point +MSE 140.342372 +---------------------- -------------------------------------------------------- +Time: 0.587s Load: 0.006s, Pack+Encode: 0.240s, Decode+Unpack: 0.340s +---------------------- -------------------------------------------------------- +💾 Converting with 140.3424 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,900B, BPFP=0.5038 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,088B, BPFP=2.0567 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,880B, BPFP=0.9290 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,460B, BPFP=2.0031 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,224B, BPFP=1.2145 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,972B, BPFP=1.9614 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,312B, BPFP=1.1366 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,368B, BPFP=1.9952 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,784B, BPFP=1.6892 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,384B, BPFP=1.9112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 54,328B, BPFP=0.6627 +⌛️ [2/4] FRONTEND: Frontend time: 0.303s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12273688 52.83275700 + layer.0.v_cache 0.00001959 0.01111499 + layer.1.k_cache 0.11126613 4.67005962 + layer.1.v_cache 0.00000674 0.00463898 + layer.2.k_cache 0.01416154 0.75518098 + layer.2.v_cache 0.00002164 0.01167239 + layer.3.k_cache 0.04491856 3.08549458 + layer.3.v_cache 0.00002114 0.01298377 + layer.4.k_cache 0.00068813 0.23090056 + layer.4.v_cache 0.00006151 0.02338483 + layer.4.output 0.00884247 297.80971897 + ------------------------------------------------------------------------------------- + TOTAL 0.02092936 126.25330709 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 234700 +BPFP 1.1788 bits/point +EBPFP 2.3576 equivalent bits/point +MSE 126.253307 +---------------------- -------------------------------------------------------- +Time: 0.680s Load: 0.007s, Pack+Encode: 0.303s, Decode+Unpack: 0.370s +---------------------- -------------------------------------------------------- +💾 Converting with 126.2533 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 257, 128) +Output shape: (1, 257, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.output: torch.Size([1, 257, 3584]) -> torch.Size([1, 1, 257, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,988B, BPFP=0.4857 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,628B, BPFP=2.1053 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,104B, BPFP=0.9183 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,452B, BPFP=2.0338 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,332B, BPFP=1.1753 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,700B, BPFP=1.9881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,068B, BPFP=1.0985 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,312B, BPFP=2.0253 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,432B, BPFP=1.6678 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,604B, BPFP=1.9214 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,220B, BPFP=0.6186 +⌛️ [2/4] FRONTEND: Frontend time: 0.366s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.540s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12694686 51.72416479 + layer.0.v_cache 0.00002055 0.00973376 + layer.1.k_cache 0.54844232 4.70399707 + layer.1.v_cache 0.00000676 0.00369568 + layer.2.k_cache 0.01835482 0.72606567 + layer.2.v_cache 0.00002099 0.00992542 + layer.3.k_cache 0.02141366 2.73181960 + layer.3.v_cache 0.00002047 0.01113645 + layer.4.k_cache 0.00070767 0.20240131 + layer.4.v_cache 0.00005046 0.01983312 + layer.4.output 0.00513101 215.45978669 + ------------------------------------------------------------------------------------- + TOTAL 0.04422951 92.25654586 + (elements=2,236,928) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2236928 +Total Bytes 324840 +BPFP 1.1617 bits/point +EBPFP 2.3235 equivalent bits/point +MSE 92.256546 +---------------------- -------------------------------------------------------- +Time: 0.917s Load: 0.011s, Pack+Encode: 0.366s, Decode+Unpack: 0.540s +---------------------- -------------------------------------------------------- +💾 Converting with 92.2565 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,892B, BPFP=0.5058 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,996B, BPFP=2.0601 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,548B, BPFP=0.9056 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,180B, BPFP=1.9900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,636B, BPFP=1.1707 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,720B, BPFP=1.9505 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,024B, BPFP=1.1181 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,232B, BPFP=1.9945 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,460B, BPFP=1.6707 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,104B, BPFP=1.8977 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 49,852B, BPFP=0.6114 +⌛️ [2/4] FRONTEND: Frontend time: 0.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12731336 51.71735491 + layer.0.v_cache 0.00001750 0.01041717 + layer.1.k_cache 0.10052182 4.94583365 + layer.1.v_cache 0.00000572 0.00372983 + layer.2.k_cache 0.01141533 0.76880981 + layer.2.v_cache 0.00001926 0.01011320 + layer.3.k_cache 0.03745206 3.26709722 + layer.3.v_cache 0.00002083 0.01220349 + layer.4.k_cache 0.00066429 0.20962126 + layer.4.v_cache 0.00005101 0.02057136 + layer.4.output 0.00884527 300.32206633 + ------------------------------------------------------------------------------------- + TOTAL 0.01996459 127.24824801 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 227644 +BPFP 1.1496 bits/point +EBPFP 2.2992 equivalent bits/point +MSE 127.248248 +---------------------- -------------------------------------------------------- +Time: 0.687s Load: 0.007s, Pack+Encode: 0.309s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 127.2482 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,972B, BPFP=0.4907 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,952B, BPFP=2.1081 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,216B, BPFP=0.9302 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,960B, BPFP=2.0383 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,992B, BPFP=1.1959 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,252B, BPFP=1.9885 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,844B, BPFP=1.1151 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,704B, BPFP=2.0203 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,924B, BPFP=1.6838 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,420B, BPFP=1.9299 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,132B, BPFP=0.6549 +⌛️ [2/4] FRONTEND: Frontend time: 0.321s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.444s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12752426 52.56197653 + layer.0.v_cache 0.00002004 0.01016702 + layer.1.k_cache 0.30660921 4.57047169 + layer.1.v_cache 0.00000613 0.00352283 + layer.2.k_cache 0.01808885 0.73146751 + layer.2.v_cache 0.00002107 0.00973072 + layer.3.k_cache 0.01678786 3.14644478 + layer.3.v_cache 0.00002050 0.01112400 + layer.4.k_cache 0.00068290 0.19513092 + layer.4.v_cache 0.00004851 0.01921335 + layer.4.output 1.37905047 243.30544160 + ------------------------------------------------------------------------------------- + TOTAL 0.59548016 103.78807886 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 285368 +BPFP 1.1815 bits/point +EBPFP 2.3629 equivalent bits/point +MSE 103.788079 +---------------------- -------------------------------------------------------- +Time: 0.774s Load: 0.010s, Pack+Encode: 0.321s, Decode+Unpack: 0.444s +---------------------- -------------------------------------------------------- +💾 Converting with 103.7881 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,944B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,332B, BPFP=2.1120 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,084B, BPFP=0.9421 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,500B, BPFP=2.0521 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,692B, BPFP=1.2019 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,756B, BPFP=1.9986 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,956B, BPFP=1.1489 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,424B, BPFP=2.0467 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,592B, BPFP=1.6987 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,100B, BPFP=1.9513 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,520B, BPFP=0.6842 +⌛️ [2/4] FRONTEND: Frontend time: 0.312s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.410s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12766143 53.73282240 + layer.0.v_cache 0.00001653 0.00961539 + layer.1.k_cache 0.34631467 4.64066550 + layer.1.v_cache 0.00000659 0.00373523 + layer.2.k_cache 0.01786991 0.75829727 + layer.2.v_cache 0.00002041 0.00967846 + layer.3.k_cache 0.06874437 2.86689976 + layer.3.v_cache 0.00002276 0.01185216 + layer.4.k_cache 0.00067882 0.20014222 + layer.4.v_cache 0.00005621 0.02104302 + layer.4.output 1.41085444 247.71761850 + ------------------------------------------------------------------------------------- + TOTAL 0.61396310 105.66341652 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 283900 +BPFP 1.2025 bits/point +EBPFP 2.4050 equivalent bits/point +MSE 105.663417 +---------------------- -------------------------------------------------------- +Time: 0.730s Load: 0.008s, Pack+Encode: 0.312s, Decode+Unpack: 0.410s +---------------------- -------------------------------------------------------- +💾 Converting with 105.6634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,928B, BPFP=0.4790 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,120B, BPFP=2.0824 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,252B, BPFP=0.9162 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,288B, BPFP=2.0249 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,140B, BPFP=1.1850 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,612B, BPFP=1.9782 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,972B, BPFP=1.1043 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,164B, BPFP=2.0163 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,176B, BPFP=1.6715 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,760B, BPFP=1.9192 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,628B, BPFP=0.7075 +⌛️ [2/4] FRONTEND: Frontend time: 0.315s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.395s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12070838 52.17839118 + layer.0.v_cache 0.00001763 0.01012997 + layer.1.k_cache 0.31015487 4.56370990 + layer.1.v_cache 0.00000677 0.00385417 + layer.2.k_cache 0.03544329 0.73975318 + layer.2.v_cache 0.00002212 0.01034935 + layer.3.k_cache 0.05682086 3.32353764 + layer.3.v_cache 0.00002136 0.01199593 + layer.4.k_cache 0.00073636 0.21269626 + layer.4.v_cache 0.00005031 0.02094332 + layer.4.output 1.35470685 239.01432127 + ------------------------------------------------------------------------------------- + TOTAL 0.58864294 102.01032999 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 294040 +BPFP 1.1958 bits/point +EBPFP 2.3917 equivalent bits/point +MSE 102.010330 +---------------------- -------------------------------------------------------- +Time: 0.719s Load: 0.008s, Pack+Encode: 0.315s, Decode+Unpack: 0.395s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0103 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,824B, BPFP=0.5006 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,956B, BPFP=2.1241 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,764B, BPFP=0.9363 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,320B, BPFP=2.0775 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,616B, BPFP=1.2189 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,604B, BPFP=2.0249 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,672B, BPFP=1.1496 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,116B, BPFP=2.0625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,380B, BPFP=1.7151 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,848B, BPFP=1.9695 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,172B, BPFP=0.7039 +⌛️ [2/4] FRONTEND: Frontend time: 0.301s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.421s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13227103 51.57293042 + layer.0.v_cache 0.00002050 0.01008245 + layer.1.k_cache 0.22864332 3.98654189 + layer.1.v_cache 0.00000662 0.00385803 + layer.2.k_cache 0.02950178 0.75550534 + layer.2.v_cache 0.00002095 0.01016315 + layer.3.k_cache 0.01835468 3.27176347 + layer.3.v_cache 0.00002123 0.01208236 + layer.4.k_cache 0.00070999 0.20713512 + layer.4.v_cache 0.00005457 0.02085045 + layer.4.output 1.43734575 253.16369048 + ------------------------------------------------------------------------------------- + TOTAL 0.61594264 107.76451447 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 282272 +BPFP 1.2180 bits/point +EBPFP 2.4361 equivalent bits/point +MSE 107.764514 +---------------------- -------------------------------------------------------- +Time: 0.728s Load: 0.007s, Pack+Encode: 0.301s, Decode+Unpack: 0.421s +---------------------- -------------------------------------------------------- +💾 Converting with 107.7645 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,988B, BPFP=0.4918 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,732B, BPFP=2.0926 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,456B, BPFP=0.9471 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,856B, BPFP=2.0310 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,228B, BPFP=1.2126 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,304B, BPFP=1.9921 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,144B, BPFP=1.1363 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,628B, BPFP=2.0149 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,228B, BPFP=1.7052 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,476B, BPFP=1.9338 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,132B, BPFP=0.6549 +⌛️ [2/4] FRONTEND: Frontend time: 0.287s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.397s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12518616 53.89368226 + layer.0.v_cache 0.00001676 0.00990847 + layer.1.k_cache 0.33007166 4.36329830 + layer.1.v_cache 0.00000630 0.00388000 + layer.2.k_cache 0.01131002 0.76256685 + layer.2.v_cache 0.00002125 0.01047782 + layer.3.k_cache 0.02232737 2.87034428 + layer.3.v_cache 0.00001983 0.01105900 + layer.4.k_cache 0.00071526 0.21021214 + layer.4.v_cache 0.00005006 0.01950955 + layer.4.output 1.37904434 243.18920930 + ------------------------------------------------------------------------------------- + TOTAL 0.59664912 103.79290610 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 286172 +BPFP 1.1848 bits/point +EBPFP 2.3696 equivalent bits/point +MSE 103.792906 +---------------------- -------------------------------------------------------- +Time: 0.693s Load: 0.009s, Pack+Encode: 0.287s, Decode+Unpack: 0.397s +---------------------- -------------------------------------------------------- +💾 Converting with 103.7929 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,596B, BPFP=0.4864 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,364B, BPFP=2.0085 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,676B, BPFP=0.8758 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,364B, BPFP=1.9444 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,676B, BPFP=1.1319 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,836B, BPFP=1.9106 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,536B, BPFP=1.0589 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,248B, BPFP=1.9370 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,280B, BPFP=1.6189 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,020B, BPFP=1.8584 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,724B, BPFP=0.6287 +⌛️ [2/4] FRONTEND: Frontend time: 0.333s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.422s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14655781 50.12552030 + layer.0.v_cache 0.00001709 0.00940080 + layer.1.k_cache 0.36683986 4.14002365 + layer.1.v_cache 0.00000651 0.00385994 + layer.2.k_cache 0.01755252 0.71513854 + layer.2.v_cache 0.00002060 0.00996506 + layer.3.k_cache 0.02713263 2.91155230 + layer.3.v_cache 0.00001997 0.01068018 + layer.4.k_cache 0.00071102 0.20113518 + layer.4.v_cache 0.00005162 0.01939009 + layer.4.output 1.25479176 221.04949136 + ------------------------------------------------------------------------------------- + TOTAL 0.54955600 94.44077092 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 300320 +BPFP 1.1313 bits/point +EBPFP 2.2625 equivalent bits/point +MSE 94.440771 +---------------------- -------------------------------------------------------- +Time: 0.764s Load: 0.009s, Pack+Encode: 0.333s, Decode+Unpack: 0.422s +---------------------- -------------------------------------------------------- +💾 Converting with 94.4408 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 155, 128) +Output shape: (1, 155, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.output: torch.Size([1, 155, 3584]) -> torch.Size([1, 1, 155, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,276B, BPFP=0.5319 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,584B, BPFP=2.1758 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,704B, BPFP=0.9782 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,824B, BPFP=2.0992 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,676B, BPFP=1.2778 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,556B, BPFP=2.0722 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,948B, BPFP=1.2044 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,960B, BPFP=2.1129 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,756B, BPFP=1.7899 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,028B, BPFP=2.0190 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 45,916B, BPFP=0.6612 +⌛️ [2/4] FRONTEND: Frontend time: 0.296s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102877 53.00042843 + layer.0.v_cache 0.00001822 0.01044380 + layer.1.k_cache 0.12773433 4.56999748 + layer.1.v_cache 0.00000585 0.00405868 + layer.2.k_cache 0.00931232 0.73780749 + layer.2.v_cache 0.00002110 0.01150184 + layer.3.k_cache 0.01537714 3.33837339 + layer.3.v_cache 0.00002091 0.01314592 + layer.4.k_cache 0.00066471 0.22902229 + layer.4.v_cache 0.00005180 0.02359250 + layer.4.output 0.01742802 355.94061060 + ------------------------------------------------------------------------------------- + TOTAL 0.02389596 150.20721447 + (elements=1,349,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1349120 +Total Bytes 207228 +BPFP 1.2288 bits/point +EBPFP 2.4576 equivalent bits/point +MSE 150.207214 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.005s, Pack+Encode: 0.296s, Decode+Unpack: 0.354s +---------------------- -------------------------------------------------------- +💾 Converting with 150.2072 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,904B, BPFP=0.5017 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,128B, BPFP=2.1169 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,888B, BPFP=0.9366 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,372B, BPFP=2.0619 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,704B, BPFP=1.2140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,604B, BPFP=2.0061 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,768B, BPFP=1.1459 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,168B, BPFP=2.0471 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,476B, BPFP=1.7061 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,788B, BPFP=1.9468 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,468B, BPFP=0.6382 +⌛️ [2/4] FRONTEND: Frontend time: 0.322s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11659910 53.07673510 + layer.0.v_cache 0.00001673 0.01013980 + layer.1.k_cache 0.29571861 4.63225154 + layer.1.v_cache 0.00000615 0.00391791 + layer.2.k_cache 0.04707494 0.75685553 + layer.2.v_cache 0.00002002 0.00990815 + layer.3.k_cache 0.03274950 2.97829448 + layer.3.v_cache 0.00002078 0.01119709 + layer.4.k_cache 0.00068608 0.20771257 + layer.4.v_cache 0.00005136 0.02016641 + layer.4.output 1.42391751 250.76146179 + ------------------------------------------------------------------------------------- + TOTAL 0.61531564 106.88455360 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 277268 +BPFP 1.1853 bits/point +EBPFP 2.3706 equivalent bits/point +MSE 106.884554 +---------------------- -------------------------------------------------------- +Time: 0.715s Load: 0.009s, Pack+Encode: 0.322s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 106.8846 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,864B, BPFP=0.4955 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,180B, BPFP=2.0275 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,464B, BPFP=0.9113 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,300B, BPFP=1.9720 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,736B, BPFP=1.1804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,588B, BPFP=1.9272 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,608B, BPFP=1.1094 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,156B, BPFP=1.9630 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,160B, BPFP=1.6482 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,740B, BPFP=1.8737 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,592B, BPFP=0.6534 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.437s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10631144 51.47103783 + layer.0.v_cache 0.00001780 0.01009015 + layer.1.k_cache 0.46192560 4.42929274 + layer.1.v_cache 0.00000619 0.00400062 + layer.2.k_cache 0.02890189 0.74270255 + layer.2.v_cache 0.00002115 0.01028905 + layer.3.k_cache 0.04994748 3.27093457 + layer.3.v_cache 0.00002140 0.01196053 + layer.4.k_cache 0.00068365 0.20701765 + layer.4.v_cache 0.00005040 0.02058477 + layer.4.output 1.23456514 216.85008641 + ------------------------------------------------------------------------------------- + TOTAL 0.54646135 92.83108913 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 312388 +BPFP 1.1577 bits/point +EBPFP 2.3155 equivalent bits/point +MSE 92.831089 +---------------------- -------------------------------------------------------- +Time: 0.764s Load: 0.008s, Pack+Encode: 0.319s, Decode+Unpack: 0.437s +---------------------- -------------------------------------------------------- +💾 Converting with 92.8311 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 194, 128) +Output shape: (1, 194, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.output: torch.Size([1, 194, 3584]) -> torch.Size([1, 1, 194, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,060B, BPFP=0.4881 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 26,808B, BPFP=2.1591 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,640B, BPFP=0.9375 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 25,864B, BPFP=2.0831 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,200B, BPFP=1.2242 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 25,560B, BPFP=2.0586 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,272B, BPFP=1.1495 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,056B, BPFP=2.0986 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 21,600B, BPFP=1.7397 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 24,744B, BPFP=1.9929 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 52,132B, BPFP=0.5998 +⌛️ [2/4] FRONTEND: Frontend time: 0.305s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.418s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15027005 52.56067775 + layer.0.v_cache 0.00001782 0.01110051 + layer.1.k_cache 0.21402168 4.67672761 + layer.1.v_cache 0.00000588 0.00428561 + layer.2.k_cache 0.02177384 0.84972909 + layer.2.v_cache 0.00002019 0.01146914 + layer.3.k_cache 0.04024272 3.20950978 + layer.3.v_cache 0.00002076 0.01285910 + layer.4.k_cache 0.00069591 0.22558250 + layer.4.v_cache 0.00005073 0.02295245 + layer.4.output 1.57794378 278.16547772 + ------------------------------------------------------------------------------------- + TOTAL 0.67486624 118.16136692 + (elements=1,688,576) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1688576 +Total Bytes 249936 +BPFP 1.1841 bits/point +EBPFP 2.3683 equivalent bits/point +MSE 118.161367 +---------------------- -------------------------------------------------------- +Time: 0.730s Load: 0.007s, Pack+Encode: 0.305s, Decode+Unpack: 0.418s +---------------------- -------------------------------------------------------- +💾 Converting with 118.1614 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 258, 128) +Output shape: (1, 258, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.output: torch.Size([1, 258, 3584]) -> torch.Size([1, 1, 258, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,948B, BPFP=0.4813 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,580B, BPFP=2.0942 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,028B, BPFP=0.9101 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,376B, BPFP=2.0213 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,448B, BPFP=1.1778 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,828B, BPFP=1.9881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,236B, BPFP=1.1044 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,424B, BPFP=2.0242 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,692B, BPFP=1.6771 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,116B, BPFP=1.9450 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,656B, BPFP=0.5680 +⌛️ [2/4] FRONTEND: Frontend time: 0.361s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.505s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09035230 53.09951853 + layer.0.v_cache 0.00001726 0.00977983 + layer.1.k_cache 0.42455262 4.14810015 + layer.1.v_cache 0.00000600 0.00369189 + layer.2.k_cache 0.01891134 0.75379506 + layer.2.v_cache 0.00001996 0.00976998 + layer.3.k_cache 0.02093011 3.06621415 + layer.3.v_cache 0.00001928 0.01079449 + layer.4.k_cache 0.00072082 0.20165236 + layer.4.v_cache 0.00004951 0.02053691 + layer.4.output 0.00505962 214.21883652 + ------------------------------------------------------------------------------------- + TOTAL 0.03476450 91.81504171 + (elements=2,245,632) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2245632 +Total Bytes 320332 +BPFP 1.1412 bits/point +EBPFP 2.2823 equivalent bits/point +MSE 91.815042 +---------------------- -------------------------------------------------------- +Time: 0.874s Load: 0.008s, Pack+Encode: 0.361s, Decode+Unpack: 0.505s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8150 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,156B, BPFP=0.4861 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,852B, BPFP=2.0959 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,720B, BPFP=0.9321 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,804B, BPFP=2.0247 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,300B, BPFP=1.1753 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,072B, BPFP=1.9750 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,236B, BPFP=1.1030 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,692B, BPFP=2.0171 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,584B, BPFP=1.6701 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,252B, BPFP=1.9193 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,768B, BPFP=0.6771 +⌛️ [2/4] FRONTEND: Frontend time: 0.339s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.445s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15008528 52.72824389 + layer.0.v_cache 0.00001847 0.00957832 + layer.1.k_cache 0.25612781 4.50146750 + layer.1.v_cache 0.00000678 0.00358040 + layer.2.k_cache 0.03931687 0.74022476 + layer.2.v_cache 0.00002107 0.00951723 + layer.3.k_cache 0.03786138 3.22251720 + layer.3.v_cache 0.00002177 0.01149541 + layer.4.k_cache 0.00069674 0.20039550 + layer.4.v_cache 0.00005115 0.01965995 + layer.4.output 1.33117570 234.74497283 + ------------------------------------------------------------------------------------- + TOTAL 0.57661395 100.27420529 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 296436 +BPFP 1.1846 bits/point +EBPFP 2.3692 equivalent bits/point +MSE 100.274205 +---------------------- -------------------------------------------------------- +Time: 0.792s Load: 0.008s, Pack+Encode: 0.339s, Decode+Unpack: 0.445s +---------------------- -------------------------------------------------------- +💾 Converting with 100.2742 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,812B, BPFP=0.5045 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,868B, BPFP=2.0719 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,624B, BPFP=0.9222 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,232B, BPFP=2.0167 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,788B, BPFP=1.1969 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,792B, BPFP=1.9785 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,940B, BPFP=1.1233 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,124B, BPFP=2.0073 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,440B, BPFP=1.6875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,192B, BPFP=1.9264 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 55,476B, BPFP=0.6879 +⌛️ [2/4] FRONTEND: Frontend time: 0.270s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15161699 51.12962240 + layer.0.v_cache 0.00001786 0.01068448 + layer.1.k_cache 0.08662784 4.49427117 + layer.1.v_cache 0.00000652 0.00393319 + layer.2.k_cache 0.02063684 0.76501965 + layer.2.v_cache 0.00002327 0.01088772 + layer.3.k_cache 0.08087930 3.30659620 + layer.3.v_cache 0.00002116 0.01238902 + layer.4.k_cache 0.00074161 0.21315787 + layer.4.v_cache 0.00006096 0.02171893 + layer.4.output 0.00898438 303.76071429 + ------------------------------------------------------------------------------------- + TOTAL 0.02373665 128.60548710 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 233288 +BPFP 1.1912 bits/point +EBPFP 2.3824 equivalent bits/point +MSE 128.605487 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.006s, Pack+Encode: 0.270s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 128.6055 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,124B, BPFP=0.5060 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,752B, BPFP=2.1131 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,308B, BPFP=0.9452 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,896B, BPFP=2.0523 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,308B, BPFP=1.2293 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,248B, BPFP=2.0063 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,212B, BPFP=1.1514 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,812B, BPFP=2.0463 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,164B, BPFP=1.7162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,484B, BPFP=1.9520 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,644B, BPFP=0.6559 +⌛️ [2/4] FRONTEND: Frontend time: 0.322s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.425s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14327769 53.16583807 + layer.0.v_cache 0.00001763 0.01023993 + layer.1.k_cache 0.34925704 4.76284901 + layer.1.v_cache 0.00000642 0.00411985 + layer.2.k_cache 0.02352169 0.77313496 + layer.2.v_cache 0.00002271 0.01096010 + layer.3.k_cache 0.00845094 3.03597606 + layer.3.v_cache 0.00002026 0.01214141 + layer.4.k_cache 0.00069169 0.21169905 + layer.4.v_cache 0.00005086 0.02173236 + layer.4.output 1.39158238 244.93478084 + ------------------------------------------------------------------------------------- + TOTAL 0.60390550 104.50306804 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 285952 +BPFP 1.1947 bits/point +EBPFP 2.3893 equivalent bits/point +MSE 104.503068 +---------------------- -------------------------------------------------------- +Time: 0.754s Load: 0.007s, Pack+Encode: 0.322s, Decode+Unpack: 0.425s +---------------------- -------------------------------------------------------- +💾 Converting with 104.5031 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,712B, BPFP=0.4950 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,776B, BPFP=2.0895 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,288B, BPFP=0.9255 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,576B, BPFP=2.0214 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,124B, BPFP=1.2002 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,972B, BPFP=1.9870 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,880B, BPFP=1.1295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,472B, BPFP=2.0155 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,596B, BPFP=1.6816 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,900B, BPFP=1.9261 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,584B, BPFP=0.6054 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12716211 52.18318182 + layer.0.v_cache 0.00001677 0.00950292 + layer.1.k_cache 0.51313377 4.45834295 + layer.1.v_cache 0.00000586 0.00367262 + layer.2.k_cache 0.02995576 0.70920055 + layer.2.v_cache 0.00002133 0.01032489 + layer.3.k_cache 0.03339291 3.12459384 + layer.3.v_cache 0.00002012 0.01086291 + layer.4.k_cache 0.00068476 0.19996615 + layer.4.v_cache 0.00005875 0.01987104 + layer.4.output 0.00481428 201.07590909 + ------------------------------------------------------------------------------------- + TOTAL 0.04342071 86.36828725 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 346880 +BPFP 1.1594 bits/point +EBPFP 2.3187 equivalent bits/point +MSE 86.368287 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 86.3683 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,432B, BPFP=0.5187 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 32,028B, BPFP=1.9702 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,052B, BPFP=0.9259 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,264B, BPFP=1.9232 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,896B, BPFP=1.1624 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,592B, BPFP=1.8819 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,104B, BPFP=1.1137 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,996B, BPFP=1.9067 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,012B, BPFP=1.6001 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,616B, BPFP=1.8219 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,052B, BPFP=0.6332 +⌛️ [2/4] FRONTEND: Frontend time: 0.271s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.465s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13396681 51.74603992 + layer.0.v_cache 0.00001641 0.00946331 + layer.1.k_cache 0.47111181 4.35963836 + layer.1.v_cache 0.00000608 0.00371538 + layer.2.k_cache 0.02933135 0.69972734 + layer.2.v_cache 0.00002255 0.00980673 + layer.3.k_cache 0.03744583 2.93938584 + layer.3.v_cache 0.00002260 0.01079190 + layer.4.k_cache 0.00071777 0.19776568 + layer.4.v_cache 0.00004917 0.01980572 + layer.4.output 1.20538323 210.56149817 + ------------------------------------------------------------------------------------- + TOTAL 0.53590429 90.23097808 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 313044 +BPFP 1.1328 bits/point +EBPFP 2.2655 equivalent bits/point +MSE 90.230978 +---------------------- -------------------------------------------------------- +Time: 0.746s Load: 0.010s, Pack+Encode: 0.271s, Decode+Unpack: 0.465s +---------------------- -------------------------------------------------------- +💾 Converting with 90.2310 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,628B, BPFP=0.4681 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,712B, BPFP=2.0460 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,484B, BPFP=0.8943 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,564B, BPFP=1.9837 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,504B, BPFP=1.1667 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,836B, BPFP=1.9442 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,060B, BPFP=1.0883 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,260B, BPFP=1.9672 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,448B, BPFP=1.6519 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,768B, BPFP=1.8863 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,252B, BPFP=0.6065 +⌛️ [2/4] FRONTEND: Frontend time: 0.354s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.497s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14060297 52.56159804 + layer.0.v_cache 0.00001598 0.00920644 + layer.1.k_cache 0.58610762 4.32000520 + layer.1.v_cache 0.00000669 0.00362863 + layer.2.k_cache 0.02834670 0.71730979 + layer.2.v_cache 0.00002256 0.01006127 + layer.3.k_cache 0.01962393 3.02142461 + layer.3.v_cache 0.00002003 0.01074885 + layer.4.k_cache 0.00067797 0.20058036 + layer.4.v_cache 0.00005035 0.01960112 + layer.4.output 0.00463834 191.67162698 + ------------------------------------------------------------------------------------- + TOTAL 0.04752607 82.50444431 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 356516 +BPFP 1.1378 bits/point +EBPFP 2.2756 equivalent bits/point +MSE 82.504444 +---------------------- -------------------------------------------------------- +Time: 0.860s Load: 0.010s, Pack+Encode: 0.354s, Decode+Unpack: 0.497s +---------------------- -------------------------------------------------------- +💾 Converting with 82.5044 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,364B, BPFP=0.4813 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,464B, BPFP=2.0284 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,292B, BPFP=0.8888 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,448B, BPFP=1.9762 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,408B, BPFP=1.1517 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,672B, BPFP=1.9363 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,060B, BPFP=1.0824 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,120B, BPFP=1.9593 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,708B, BPFP=1.6297 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,452B, BPFP=1.8736 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,928B, BPFP=0.5795 +⌛️ [2/4] FRONTEND: Frontend time: 0.338s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.484s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12404394 51.52722489 + layer.0.v_cache 0.00001645 0.00885685 + layer.1.k_cache 0.57408182 4.50804058 + layer.1.v_cache 0.00000616 0.00344289 + layer.2.k_cache 0.03522489 0.75802396 + layer.2.v_cache 0.00002326 0.00943797 + layer.3.k_cache 0.02908128 2.62330547 + layer.3.v_cache 0.00002078 0.01037106 + layer.4.k_cache 0.00072493 0.19203621 + layer.4.v_cache 0.00004967 0.01876773 + layer.4.output 0.04390477 177.77669760 + ------------------------------------------------------------------------------------- + TOTAL 0.06297686 76.71155240 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 370916 +BPFP 1.1214 bits/point +EBPFP 2.2429 equivalent bits/point +MSE 76.711552 +---------------------- -------------------------------------------------------- +Time: 0.832s Load: 0.010s, Pack+Encode: 0.338s, Decode+Unpack: 0.484s +---------------------- -------------------------------------------------------- +💾 Converting with 76.7116 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 242, 128) +Output shape: (1, 242, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.output: torch.Size([1, 242, 3584]) -> torch.Size([1, 1, 242, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,624B, BPFP=0.4923 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,656B, BPFP=2.0439 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,840B, BPFP=0.8936 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,724B, BPFP=1.9837 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,956B, BPFP=1.1593 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,112B, BPFP=1.9442 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,892B, BPFP=1.0907 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,612B, BPFP=1.9765 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,480B, BPFP=1.6451 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,220B, BPFP=1.8866 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,964B, BPFP=0.6361 +⌛️ [2/4] FRONTEND: Frontend time: 0.287s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.406s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11466674 51.03784381 + layer.0.v_cache 0.00001794 0.00943709 + layer.1.k_cache 0.39062487 4.37420755 + layer.1.v_cache 0.00000616 0.00370070 + layer.2.k_cache 0.02414889 0.70639322 + layer.2.v_cache 0.00002046 0.01006136 + layer.3.k_cache 0.03495786 2.81383521 + layer.3.v_cache 0.00002053 0.01092314 + layer.4.k_cache 0.00069125 0.19877806 + layer.4.v_cache 0.00005034 0.02045452 + layer.4.output 1.26514320 222.69375369 + ------------------------------------------------------------------------------------- + TOTAL 0.55418868 95.17893591 + (elements=2,106,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2106368 +Total Bytes 303080 +BPFP 1.1511 bits/point +EBPFP 2.3022 equivalent bits/point +MSE 95.178936 +---------------------- -------------------------------------------------------- +Time: 0.701s Load: 0.008s, Pack+Encode: 0.287s, Decode+Unpack: 0.406s +---------------------- -------------------------------------------------------- +💾 Converting with 95.1789 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 189, 128) +Output shape: (1, 189, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.output: torch.Size([1, 189, 3584]) -> torch.Size([1, 1, 189, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,280B, BPFP=0.5192 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,216B, BPFP=2.0020 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,548B, BPFP=0.9547 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,612B, BPFP=1.9521 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,716B, BPFP=1.2166 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,208B, BPFP=1.9187 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,016B, BPFP=1.1587 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,584B, BPFP=1.9497 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,908B, BPFP=1.6458 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,540B, BPFP=1.8634 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 58,468B, BPFP=0.6905 +⌛️ [2/4] FRONTEND: Frontend time: 0.280s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14361043 51.67430349 + layer.0.v_cache 0.00001648 0.01022590 + layer.1.k_cache 0.18932698 4.43313970 + layer.1.v_cache 0.00000663 0.00427283 + layer.2.k_cache 0.00768016 0.72229650 + layer.2.v_cache 0.00002453 0.01103784 + layer.3.k_cache 0.01421528 2.96555970 + layer.3.v_cache 0.00002024 0.01206110 + layer.4.k_cache 0.00072804 0.22001099 + layer.4.v_cache 0.00005086 0.02145436 + layer.4.output 0.00858909 286.23776455 + ------------------------------------------------------------------------------------- + TOTAL 0.02445902 121.39639496 + (elements=1,645,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1645056 +Total Bytes 242096 +BPFP 1.1773 bits/point +EBPFP 2.3547 equivalent bits/point +MSE 121.396395 +---------------------- -------------------------------------------------------- +Time: 0.653s Load: 0.007s, Pack+Encode: 0.280s, Decode+Unpack: 0.366s +---------------------- -------------------------------------------------------- +💾 Converting with 121.3964 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,888B, BPFP=0.5053 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,816B, BPFP=2.1138 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,712B, BPFP=0.9325 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,052B, BPFP=2.0578 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,248B, BPFP=1.1919 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,384B, BPFP=2.0088 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,400B, BPFP=1.1297 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,948B, BPFP=2.0502 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,224B, BPFP=1.7036 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,680B, BPFP=1.9572 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 60,336B, BPFP=0.6323 +⌛️ [2/4] FRONTEND: Frontend time: 0.302s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.452s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16894778 50.90110127 + layer.0.v_cache 0.00001779 0.00996384 + layer.1.k_cache 0.37628758 4.64970978 + layer.1.v_cache 0.00000652 0.00389417 + layer.2.k_cache 0.02181128 0.74371073 + layer.2.v_cache 0.00002030 0.01005576 + layer.3.k_cache 0.01815432 3.05173202 + layer.3.v_cache 0.00002063 0.01166502 + layer.4.k_cache 0.00071780 0.20939282 + layer.4.v_cache 0.00005059 0.02044351 + layer.4.output 1.43727796 253.30482897 + ------------------------------------------------------------------------------------- + TOTAL 0.62629296 107.80855716 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 273688 +BPFP 1.1810 bits/point +EBPFP 2.3620 equivalent bits/point +MSE 107.808557 +---------------------- -------------------------------------------------------- +Time: 0.763s Load: 0.009s, Pack+Encode: 0.302s, Decode+Unpack: 0.452s +---------------------- -------------------------------------------------------- +💾 Converting with 107.8086 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,056B, BPFP=0.5104 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,060B, BPFP=2.1021 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,972B, BPFP=0.9384 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,256B, BPFP=2.0440 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,416B, BPFP=1.1875 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,544B, BPFP=1.9925 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,480B, BPFP=1.1198 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,068B, BPFP=2.0304 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,276B, BPFP=1.6837 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,584B, BPFP=1.9230 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 58,700B, BPFP=0.6066 +⌛️ [2/4] FRONTEND: Frontend time: 0.317s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.412s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09877178 53.98146792 + layer.0.v_cache 0.00001652 0.00957280 + layer.1.k_cache 0.28824160 4.02536689 + layer.1.v_cache 0.00000619 0.00376757 + layer.2.k_cache 0.02137197 0.73973649 + layer.2.v_cache 0.00001995 0.00972068 + layer.3.k_cache 0.03811082 3.07711396 + layer.3.v_cache 0.00002058 0.01073531 + layer.4.k_cache 0.00070031 0.20152975 + layer.4.v_cache 0.00004695 0.01838223 + layer.4.output 1.41729720 249.67197834 + ------------------------------------------------------------------------------------- + TOTAL 0.60990512 106.45772012 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 273412 +BPFP 1.1634 bits/point +EBPFP 2.3268 equivalent bits/point +MSE 106.457720 +---------------------- -------------------------------------------------------- +Time: 0.737s Load: 0.008s, Pack+Encode: 0.317s, Decode+Unpack: 0.412s +---------------------- -------------------------------------------------------- +💾 Converting with 106.4577 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,864B, BPFP=0.4746 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,972B, BPFP=2.0722 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,188B, BPFP=0.9118 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,104B, BPFP=2.0122 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,668B, BPFP=1.1524 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,376B, BPFP=1.9618 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,780B, BPFP=1.0910 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,740B, BPFP=1.9870 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,908B, BPFP=1.6529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,564B, BPFP=1.9057 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,784B, BPFP=0.6300 +⌛️ [2/4] FRONTEND: Frontend time: 0.320s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.477s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12161526 53.93118259 + layer.0.v_cache 0.00001916 0.00985748 + layer.1.k_cache 0.26028682 4.74737819 + layer.1.v_cache 0.00000628 0.00379091 + layer.2.k_cache 0.01296707 0.75154370 + layer.2.v_cache 0.00002040 0.00979958 + layer.3.k_cache 0.04595671 3.05262527 + layer.3.v_cache 0.00002073 0.01094830 + layer.4.k_cache 0.00067352 0.20309693 + layer.4.v_cache 0.00005047 0.01990398 + layer.4.output 1.35465500 238.93295670 + ------------------------------------------------------------------------------------- + TOTAL 0.58377655 102.07475434 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 283948 +BPFP 1.1548 bits/point +EBPFP 2.3096 equivalent bits/point +MSE 102.074754 +---------------------- -------------------------------------------------------- +Time: 0.806s Load: 0.008s, Pack+Encode: 0.320s, Decode+Unpack: 0.477s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0748 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,812B, BPFP=0.4710 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,820B, BPFP=2.0617 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,056B, BPFP=0.9027 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,832B, BPFP=1.9934 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,648B, BPFP=1.1510 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,176B, BPFP=1.9480 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,584B, BPFP=1.0774 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,680B, BPFP=1.9829 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,748B, BPFP=1.6419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,324B, BPFP=1.8891 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,320B, BPFP=0.6056 +⌛️ [2/4] FRONTEND: Frontend time: 0.337s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10229862 54.59575498 + layer.0.v_cache 0.00001727 0.00945322 + layer.1.k_cache 0.34669285 4.81024008 + layer.1.v_cache 0.00000643 0.00366976 + layer.2.k_cache 0.01639456 0.74721838 + layer.2.v_cache 0.00002164 0.00996222 + layer.3.k_cache 0.03343437 3.08132475 + layer.3.v_cache 0.00002043 0.01067998 + layer.4.k_cache 0.00071218 0.19857658 + layer.4.v_cache 0.00005088 0.01928030 + layer.4.output 1.35462105 238.74770860 + ------------------------------------------------------------------------------------- + TOTAL 0.58717627 102.04236003 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 280000 +BPFP 1.1387 bits/point +EBPFP 2.2775 equivalent bits/point +MSE 102.042360 +---------------------- -------------------------------------------------------- +Time: 0.736s Load: 0.010s, Pack+Encode: 0.337s, Decode+Unpack: 0.390s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0424 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,884B, BPFP=0.4738 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,284B, BPFP=2.0845 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,136B, BPFP=0.9042 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,392B, BPFP=2.0231 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,104B, BPFP=1.1773 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,656B, BPFP=1.9725 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,204B, BPFP=1.1154 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,248B, BPFP=2.0132 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,336B, BPFP=1.6751 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,912B, BPFP=1.9213 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,304B, BPFP=0.6028 +⌛️ [2/4] FRONTEND: Frontend time: 0.299s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11800324 54.88890849 + layer.0.v_cache 0.00001819 0.00956370 + layer.1.k_cache 0.35011090 4.31673697 + layer.1.v_cache 0.00000612 0.00354950 + layer.2.k_cache 0.02017111 0.77894599 + layer.2.v_cache 0.00002006 0.00950984 + layer.3.k_cache 0.03026835 3.00184181 + layer.3.v_cache 0.00001975 0.01088453 + layer.4.k_cache 0.00074701 0.20144660 + layer.4.v_cache 0.00005118 0.01996868 + layer.4.output 1.34863711 237.80878304 + ------------------------------------------------------------------------------------- + TOTAL 0.58587504 101.64134338 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 284460 +BPFP 1.1518 bits/point +EBPFP 2.3035 equivalent bits/point +MSE 101.641343 +---------------------- -------------------------------------------------------- +Time: 0.719s Load: 0.008s, Pack+Encode: 0.299s, Decode+Unpack: 0.411s +---------------------- -------------------------------------------------------- +💾 Converting with 101.6413 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,112B, BPFP=0.4853 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,688B, BPFP=2.0939 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,672B, BPFP=0.9329 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,928B, BPFP=2.0420 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,596B, BPFP=1.2006 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,328B, BPFP=2.0011 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,420B, BPFP=1.1204 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,968B, BPFP=2.0448 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,912B, BPFP=1.6998 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,756B, BPFP=1.9621 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,100B, BPFP=0.6346 +⌛️ [2/4] FRONTEND: Frontend time: 0.349s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10957497 52.21918924 + layer.0.v_cache 0.00001697 0.00996765 + layer.1.k_cache 0.31262030 4.40428888 + layer.1.v_cache 0.00000629 0.00405388 + layer.2.k_cache 0.01311818 0.74947334 + layer.2.v_cache 0.00002149 0.01067532 + layer.3.k_cache 0.06227440 3.14995405 + layer.3.v_cache 0.00002139 0.01261024 + layer.4.k_cache 0.00068925 0.21428477 + layer.4.v_cache 0.00005302 0.02223622 + layer.4.output 1.33688878 235.81639894 + ------------------------------------------------------------------------------------- + TOTAL 0.57980104 100.67714860 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 293480 +BPFP 1.1779 bits/point +EBPFP 2.3558 equivalent bits/point +MSE 100.677149 +---------------------- -------------------------------------------------------- +Time: 0.828s Load: 0.009s, Pack+Encode: 0.349s, Decode+Unpack: 0.470s +---------------------- -------------------------------------------------------- +💾 Converting with 100.6771 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,892B, BPFP=0.5056 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,876B, BPFP=2.1183 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,608B, BPFP=0.9249 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,860B, BPFP=2.0437 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,128B, BPFP=1.1831 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,180B, BPFP=1.9938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,124B, BPFP=1.1094 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,516B, BPFP=2.0185 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,008B, BPFP=1.6878 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,344B, BPFP=1.9325 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 59,700B, BPFP=0.6256 +⌛️ [2/4] FRONTEND: Frontend time: 0.332s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.445s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13987277 51.68904049 + layer.0.v_cache 0.00001887 0.00996293 + layer.1.k_cache 0.23498858 4.18511447 + layer.1.v_cache 0.00000603 0.00372817 + layer.2.k_cache 0.02028183 0.76138119 + layer.2.v_cache 0.00001992 0.01004696 + layer.3.k_cache 0.05552547 3.24719181 + layer.3.v_cache 0.00002044 0.01164538 + layer.4.k_cache 0.00067773 0.20315423 + layer.4.v_cache 0.00004777 0.01993628 + layer.4.output 1.43727485 253.43085597 + ------------------------------------------------------------------------------------- + TOTAL 0.61837549 107.89159963 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 271236 +BPFP 1.1704 bits/point +EBPFP 2.3408 equivalent bits/point +MSE 107.891600 +---------------------- -------------------------------------------------------- +Time: 0.785s Load: 0.008s, Pack+Encode: 0.332s, Decode+Unpack: 0.445s +---------------------- -------------------------------------------------------- +💾 Converting with 107.8916 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,944B, BPFP=0.5142 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,368B, BPFP=2.1007 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,552B, BPFP=0.9295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,720B, BPFP=2.0527 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,124B, BPFP=1.1940 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,904B, BPFP=1.9923 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,332B, BPFP=1.1354 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,476B, BPFP=2.0347 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,988B, BPFP=1.7023 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,264B, BPFP=1.9449 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 58,248B, BPFP=0.6162 +⌛️ [2/4] FRONTEND: Frontend time: 0.311s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.417s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17004660 50.89953903 + layer.0.v_cache 0.00001740 0.01000327 + layer.1.k_cache 0.16740718 4.53568069 + layer.1.v_cache 0.00000654 0.00407397 + layer.2.k_cache 0.01191964 0.73286286 + layer.2.v_cache 0.00001991 0.01056019 + layer.3.k_cache 0.00660476 3.28136137 + layer.3.v_cache 0.00002076 0.01163521 + layer.4.k_cache 0.00070551 0.21498634 + layer.4.v_cache 0.00005179 0.02075699 + layer.4.output 1.45088344 255.74407583 + ------------------------------------------------------------------------------------- + TOTAL 0.61841083 108.81941122 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 268920 +BPFP 1.1714 bits/point +EBPFP 2.3428 equivalent bits/point +MSE 108.819411 +---------------------- -------------------------------------------------------- +Time: 0.736s Load: 0.008s, Pack+Encode: 0.311s, Decode+Unpack: 0.417s +---------------------- -------------------------------------------------------- +💾 Converting with 108.8194 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,140B, BPFP=0.4851 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,440B, BPFP=2.0679 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,676B, BPFP=0.9291 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,628B, BPFP=2.0128 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,368B, BPFP=1.1799 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,100B, BPFP=1.9769 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,216B, BPFP=1.1016 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,292B, BPFP=1.9899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,832B, BPFP=1.6870 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,260B, BPFP=1.9198 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 60,340B, BPFP=0.5856 +⌛️ [2/4] FRONTEND: Frontend time: 0.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.406s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14712874 53.41652514 + layer.0.v_cache 0.00001826 0.00936698 + layer.1.k_cache 0.36788522 4.35064803 + layer.1.v_cache 0.00000598 0.00389168 + layer.2.k_cache 0.01156470 0.70913729 + layer.2.v_cache 0.00002030 0.01016046 + layer.3.k_cache 0.01315050 3.08639659 + layer.3.v_cache 0.00002017 0.01102393 + layer.4.k_cache 0.00068370 0.20800539 + layer.4.v_cache 0.00004817 0.01937289 + layer.4.output 1.33105469 234.77688276 + ------------------------------------------------------------------------------------- + TOTAL 0.57987697 100.30957104 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 286292 +BPFP 1.1441 bits/point +EBPFP 2.2881 equivalent bits/point +MSE 100.309571 +---------------------- -------------------------------------------------------- +Time: 0.723s Load: 0.009s, Pack+Encode: 0.309s, Decode+Unpack: 0.406s +---------------------- -------------------------------------------------------- +💾 Converting with 100.3096 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 188, 128) +Output shape: (1, 188, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.output: torch.Size([1, 188, 3584]) -> torch.Size([1, 1, 188, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,168B, BPFP=0.5126 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,032B, BPFP=1.9973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,324B, BPFP=0.9412 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,440B, BPFP=1.9481 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,516B, BPFP=1.2064 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,072B, BPFP=1.9176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,824B, BPFP=1.1489 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,320B, BPFP=1.9382 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,952B, BPFP=1.6582 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,396B, BPFP=1.8614 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 53,244B, BPFP=0.6322 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005002 52.39757002 + layer.0.v_cache 0.00001663 0.00961656 + layer.1.k_cache 0.14149688 4.54617732 + layer.1.v_cache 0.00000584 0.00367008 + layer.2.k_cache 0.00881058 0.68496461 + layer.2.v_cache 0.00002285 0.01045855 + layer.3.k_cache 0.03192479 3.03181750 + layer.3.v_cache 0.00001958 0.01127590 + layer.4.k_cache 0.00066612 0.20859172 + layer.4.v_cache 0.00005162 0.02113846 + layer.4.output 0.00857985 289.90995441 + ------------------------------------------------------------------------------------- + TOTAL 0.02253670 122.95852715 + (elements=1,636,352) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1636352 +Total Bytes 235288 +BPFP 1.1503 bits/point +EBPFP 2.3006 equivalent bits/point +MSE 122.958527 +---------------------- -------------------------------------------------------- +Time: 0.609s Load: 0.006s, Pack+Encode: 0.262s, Decode+Unpack: 0.341s +---------------------- -------------------------------------------------------- +💾 Converting with 122.9585 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,324B, BPFP=0.4965 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,416B, BPFP=2.1526 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,016B, BPFP=0.9435 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,472B, BPFP=2.0785 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,684B, BPFP=1.2315 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,048B, BPFP=2.0452 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,460B, BPFP=1.1354 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,572B, BPFP=2.0864 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,108B, BPFP=1.7359 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,408B, BPFP=1.9950 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 55,476B, BPFP=0.6223 +⌛️ [2/4] FRONTEND: Frontend time: 0.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.422s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11435634 53.94443408 + layer.0.v_cache 0.00001764 0.01045831 + layer.1.k_cache 0.25025516 4.33401045 + layer.1.v_cache 0.00000690 0.00411037 + layer.2.k_cache 0.01528106 0.72184500 + layer.2.v_cache 0.00002105 0.01068785 + layer.3.k_cache 0.02533426 3.02082250 + layer.3.v_cache 0.00002127 0.01225697 + layer.4.k_cache 0.00070809 0.21412260 + layer.4.v_cache 0.00005072 0.02108407 + layer.4.output 1.53833902 271.23225503 + ------------------------------------------------------------------------------------- + TOTAL 0.65731915 115.34821279 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 257984 +BPFP 1.1915 bits/point +EBPFP 2.3831 equivalent bits/point +MSE 115.348213 +---------------------- -------------------------------------------------------- +Time: 0.737s Load: 0.006s, Pack+Encode: 0.309s, Decode+Unpack: 0.422s +---------------------- -------------------------------------------------------- +💾 Converting with 115.3482 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,784B, BPFP=0.5021 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,860B, BPFP=2.0712 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,320B, BPFP=0.8958 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,076B, BPFP=2.0031 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,636B, BPFP=1.1837 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,764B, BPFP=1.9760 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,892B, BPFP=1.1191 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,232B, BPFP=2.0167 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,488B, BPFP=1.6917 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,168B, BPFP=1.9243 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 53,168B, BPFP=0.6593 +⌛️ [2/4] FRONTEND: Frontend time: 0.284s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258550 49.89121094 + layer.0.v_cache 0.00001767 0.01051517 + layer.1.k_cache 0.14790291 4.42862752 + layer.1.v_cache 0.00000629 0.00392828 + layer.2.k_cache 0.00999225 0.76221364 + layer.2.v_cache 0.00002110 0.01088555 + layer.3.k_cache 0.02117847 3.32149760 + layer.3.v_cache 0.00002045 0.01250041 + layer.4.k_cache 0.00067024 0.21719905 + layer.4.v_cache 0.00006105 0.02166665 + layer.4.output 0.00897616 303.85290179 + ------------------------------------------------------------------------------------- + TOTAL 0.02207583 128.56767984 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 230388 +BPFP 1.1764 bits/point +EBPFP 2.3528 equivalent bits/point +MSE 128.567680 +---------------------- -------------------------------------------------------- +Time: 0.657s Load: 0.006s, Pack+Encode: 0.284s, Decode+Unpack: 0.367s +---------------------- -------------------------------------------------------- +💾 Converting with 128.5677 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,660B, BPFP=0.5101 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,804B, BPFP=2.1296 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 12,348B, BPFP=0.9458 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,016B, BPFP=2.0692 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 15,916B, BPFP=1.2191 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,372B, BPFP=2.0199 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 14,928B, BPFP=1.1434 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,924B, BPFP=2.0622 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 22,588B, BPFP=1.7301 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,844B, BPFP=1.9795 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 58,400B, BPFP=0.6390 +⌛️ [2/4] FRONTEND: Frontend time: 0.306s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.415s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09364704 51.58972886 + layer.0.v_cache 0.00001853 0.00984405 + layer.1.k_cache 0.25163280 4.39818139 + layer.1.v_cache 0.00000592 0.00380481 + layer.2.k_cache 0.01595125 0.77159956 + layer.2.v_cache 0.00002050 0.01012373 + layer.3.k_cache 0.04325477 3.24450534 + layer.3.v_cache 0.00002016 0.01171408 + layer.4.k_cache 0.00069210 0.20054438 + layer.4.v_cache 0.00005026 0.02051762 + layer.4.output 1.50065788 264.47378326 + ------------------------------------------------------------------------------------- + TOTAL 0.64175873 112.44570863 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 264800 +BPFP 1.1931 bits/point +EBPFP 2.3861 equivalent bits/point +MSE 112.445709 +---------------------- -------------------------------------------------------- +Time: 0.728s Load: 0.008s, Pack+Encode: 0.306s, Decode+Unpack: 0.415s +---------------------- -------------------------------------------------------- +💾 Converting with 112.4457 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,780B, BPFP=0.4688 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,780B, BPFP=2.0589 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,048B, BPFP=0.9021 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,824B, BPFP=1.9928 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 16,352B, BPFP=1.1305 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,116B, BPFP=1.9439 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 15,692B, BPFP=1.0849 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 28,692B, BPFP=1.9837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,648B, BPFP=1.6350 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,332B, BPFP=1.8897 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,204B, BPFP=0.6242 +⌛️ [2/4] FRONTEND: Frontend time: 0.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510114 52.94828540 + layer.0.v_cache 0.00001717 0.00927688 + layer.1.k_cache 0.31797088 4.30465212 + layer.1.v_cache 0.00000673 0.00373139 + layer.2.k_cache 0.01711024 0.71944596 + layer.2.v_cache 0.00002022 0.00996550 + layer.3.k_cache 0.01307247 3.36514741 + layer.3.v_cache 0.00002089 0.01081311 + layer.4.k_cache 0.00069413 0.20351731 + layer.4.v_cache 0.00005461 0.02055622 + layer.4.output 1.35466395 238.92406764 + ------------------------------------------------------------------------------------- + TOTAL 0.58392448 102.00375675 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 281468 +BPFP 1.1447 bits/point +EBPFP 2.2894 equivalent bits/point +MSE 102.003757 +---------------------- -------------------------------------------------------- +Time: 0.743s Load: 0.008s, Pack+Encode: 0.324s, Decode+Unpack: 0.411s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0038 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,752B, BPFP=0.4884 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,856B, BPFP=2.0071 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,240B, BPFP=0.8972 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 31,132B, BPFP=1.9614 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,256B, BPFP=1.1502 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,436B, BPFP=1.9176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,296B, BPFP=1.0897 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,788B, BPFP=1.9398 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,860B, BPFP=1.6293 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,568B, BPFP=1.8629 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,156B, BPFP=0.6404 +⌛️ [2/4] FRONTEND: Frontend time: 0.339s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.475s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14860459 52.32595530 + layer.0.v_cache 0.00001629 0.00903492 + layer.1.k_cache 0.38266554 4.27379337 + layer.1.v_cache 0.00000669 0.00353824 + layer.2.k_cache 0.01628031 0.67407024 + layer.2.v_cache 0.00002106 0.00963392 + layer.3.k_cache 0.01842327 2.69646282 + layer.3.v_cache 0.00002289 0.01121288 + layer.4.k_cache 0.00072690 0.19714205 + layer.4.v_cache 0.00004923 0.01951048 + layer.4.output 1.23454798 216.33262529 + ------------------------------------------------------------------------------------- + TOTAL 0.54168545 92.62051360 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 308340 +BPFP 1.1427 bits/point +EBPFP 2.2855 equivalent bits/point +MSE 92.620514 +---------------------- -------------------------------------------------------- +Time: 0.822s Load: 0.008s, Pack+Encode: 0.339s, Decode+Unpack: 0.475s +---------------------- -------------------------------------------------------- +💾 Converting with 92.6205 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 435, 128) +Output shape: (1, 435, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.output: torch.Size([1, 435, 3584]) -> torch.Size([1, 1, 435, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,264B, BPFP=0.4764 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,444B, BPFP=1.9915 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,964B, BPFP=0.8608 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,908B, BPFP=1.9364 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,896B, BPFP=1.1098 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,680B, BPFP=1.8922 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,164B, BPFP=1.0476 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,192B, BPFP=1.9106 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,100B, BPFP=1.5841 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 50,868B, BPFP=1.8272 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 113,208B, BPFP=0.5809 +⌛️ [2/4] FRONTEND: Frontend time: 0.424s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.628s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15864419 50.11407148 + layer.0.v_cache 0.00001607 0.00926399 + layer.1.k_cache 0.94329666 4.40582626 + layer.1.v_cache 0.00000639 0.00360415 + layer.2.k_cache 0.02505282 0.71081248 + layer.2.v_cache 0.00002131 0.00960655 + layer.3.k_cache 0.01724711 2.99905066 + layer.3.v_cache 0.00002032 0.01039830 + layer.4.k_cache 0.00074513 0.19627505 + layer.4.v_cache 0.00005173 0.01920335 + layer.4.output 0.00608059 126.85967775 + ------------------------------------------------------------------------------------- + TOTAL 0.06986270 55.67622685 + (elements=3,786,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3786240 +Total Bytes 520688 +BPFP 1.1002 bits/point +EBPFP 2.2003 equivalent bits/point +MSE 55.676227 +---------------------- -------------------------------------------------------- +Time: 1.066s Load: 0.014s, Pack+Encode: 0.424s, Decode+Unpack: 0.628s +---------------------- -------------------------------------------------------- +💾 Converting with 55.6762 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 437, 128) +Output shape: (1, 437, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.output: torch.Size([1, 437, 3584]) -> torch.Size([1, 1, 437, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,036B, BPFP=0.4661 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,424B, BPFP=1.9817 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,124B, BPFP=0.8626 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,916B, BPFP=1.9278 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,192B, BPFP=1.1153 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,860B, BPFP=1.8900 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,280B, BPFP=1.0469 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,576B, BPFP=1.9156 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,284B, BPFP=1.5834 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 51,180B, BPFP=1.8299 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 122,296B, BPFP=0.6247 +⌛️ [2/4] FRONTEND: Frontend time: 0.440s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.620s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14454122 49.81968232 + layer.0.v_cache 0.00001702 0.00924594 + layer.1.k_cache 0.94267751 4.16611004 + layer.1.v_cache 0.00000668 0.00373654 + layer.2.k_cache 0.01808156 0.71378897 + layer.2.v_cache 0.00002154 0.00979458 + layer.3.k_cache 0.03034366 2.91373316 + layer.3.v_cache 0.00002210 0.01105322 + layer.4.k_cache 0.00074160 0.20045667 + layer.4.v_cache 0.00005552 0.01998302 + layer.4.output 0.00611241 126.21781015 + ------------------------------------------------------------------------------------- + TOTAL 0.06937031 55.37601503 + (elements=3,803,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3803648 +Total Bytes 531168 +BPFP 1.1172 bits/point +EBPFP 2.2344 equivalent bits/point +MSE 55.376015 +---------------------- -------------------------------------------------------- +Time: 1.075s Load: 0.015s, Pack+Encode: 0.440s, Decode+Unpack: 0.620s +---------------------- -------------------------------------------------------- +💾 Converting with 55.3760 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,428B, BPFP=0.4783 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,676B, BPFP=2.0128 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,264B, BPFP=0.8758 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,448B, BPFP=1.9505 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,244B, BPFP=1.1284 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,668B, BPFP=1.9109 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,992B, BPFP=1.0649 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,228B, BPFP=1.9393 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,692B, BPFP=1.6078 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,460B, BPFP=1.8496 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,420B, BPFP=0.6336 +⌛️ [2/4] FRONTEND: Frontend time: 0.355s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.472s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12827574 50.09894671 + layer.0.v_cache 0.00001866 0.00954379 + layer.1.k_cache 0.53653341 4.09421490 + layer.1.v_cache 0.00000626 0.00354448 + layer.2.k_cache 0.01950284 0.72091174 + layer.2.v_cache 0.00002083 0.00955652 + layer.3.k_cache 0.04340698 3.05990997 + layer.3.v_cache 0.00002181 0.01069711 + layer.4.k_cache 0.00072091 0.19585920 + layer.4.v_cache 0.00005393 0.01979343 + layer.4.output 0.04342448 175.21767741 + ------------------------------------------------------------------------------------- + TOTAL 0.06073722 75.57333646 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 379520 +BPFP 1.1325 bits/point +EBPFP 2.2651 equivalent bits/point +MSE 75.573336 +---------------------- -------------------------------------------------------- +Time: 0.838s Load: 0.011s, Pack+Encode: 0.355s, Decode+Unpack: 0.472s +---------------------- -------------------------------------------------------- +💾 Converting with 75.5733 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,520B, BPFP=0.5005 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,784B, BPFP=2.1020 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,808B, BPFP=0.9286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,888B, BPFP=2.0493 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,300B, BPFP=1.1924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,164B, BPFP=2.0068 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,164B, BPFP=1.1257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,568B, BPFP=2.0305 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,920B, BPFP=1.6988 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,136B, BPFP=1.9464 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,952B, BPFP=0.6374 +⌛️ [2/4] FRONTEND: Frontend time: 0.385s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.481s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702504 52.31672932 + layer.0.v_cache 0.00001641 0.00964725 + layer.1.k_cache 0.51130361 4.64610164 + layer.1.v_cache 0.00000656 0.00395284 + layer.2.k_cache 0.01810413 0.73897237 + layer.2.v_cache 0.00002122 0.01047166 + layer.3.k_cache 0.02476560 2.87955240 + layer.3.v_cache 0.00002197 0.01134790 + layer.4.k_cache 0.00071356 0.20813606 + layer.4.v_cache 0.00005290 0.02138316 + layer.4.output 0.00501064 207.99457908 + ------------------------------------------------------------------------------------- + TOTAL 0.04100620 89.22402048 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 341204 +BPFP 1.1790 bits/point +EBPFP 2.3579 equivalent bits/point +MSE 89.224020 +---------------------- -------------------------------------------------------- +Time: 0.875s Load: 0.009s, Pack+Encode: 0.385s, Decode+Unpack: 0.481s +---------------------- -------------------------------------------------------- +💾 Converting with 89.2240 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,400B, BPFP=0.4941 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,236B, BPFP=2.0857 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,888B, BPFP=0.9274 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,292B, BPFP=2.0227 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,668B, BPFP=1.1798 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,848B, BPFP=1.9931 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,624B, BPFP=1.1100 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,292B, BPFP=2.0227 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,112B, BPFP=1.6768 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,904B, BPFP=1.9300 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,376B, BPFP=0.6522 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.399s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11892304 53.43805505 + layer.0.v_cache 0.00001803 0.00990114 + layer.1.k_cache 0.31174104 4.49237008 + layer.1.v_cache 0.00000657 0.00383135 + layer.2.k_cache 0.01452858 0.76328832 + layer.2.v_cache 0.00002161 0.01030096 + layer.3.k_cache 0.02263702 3.10578801 + layer.3.v_cache 0.00002141 0.01151914 + layer.4.k_cache 0.00068312 0.20551799 + layer.4.v_cache 0.00005251 0.02074210 + layer.4.output 1.30838121 230.67948718 + ------------------------------------------------------------------------------------- + TOTAL 0.56631185 98.63633673 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 299640 +BPFP 1.1769 bits/point +EBPFP 2.3539 equivalent bits/point +MSE 98.636337 +---------------------- -------------------------------------------------------- +Time: 0.669s Load: 0.008s, Pack+Encode: 0.262s, Decode+Unpack: 0.399s +---------------------- -------------------------------------------------------- +💾 Converting with 98.6363 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 233, 128) +Output shape: (1, 233, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.output: torch.Size([1, 233, 3584]) -> torch.Size([1, 1, 233, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,300B, BPFP=0.4895 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,080B, BPFP=2.0842 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 13,876B, BPFP=0.9305 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,272B, BPFP=2.0300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,564B, BPFP=1.1778 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,560B, BPFP=1.9823 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,508B, BPFP=1.1070 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,948B, BPFP=2.0083 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,020B, BPFP=1.6778 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,712B, BPFP=1.9254 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,312B, BPFP=0.6353 +⌛️ [2/4] FRONTEND: Frontend time: 0.283s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13365312 53.82676703 + layer.0.v_cache 0.00001735 0.00991202 + layer.1.k_cache 0.40872042 4.76785789 + layer.1.v_cache 0.00000792 0.00423216 + layer.2.k_cache 0.01308113 0.77081260 + layer.2.v_cache 0.00002137 0.01061387 + layer.3.k_cache 0.00601081 2.90896908 + layer.3.v_cache 0.00002094 0.01128320 + layer.4.k_cache 0.00068546 0.20844544 + layer.4.v_cache 0.00005338 0.02106501 + layer.4.output 1.31399239 231.74047747 + ------------------------------------------------------------------------------------- + TOTAL 0.57413051 99.10137062 + (elements=2,028,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2028032 +Total Bytes 296152 +BPFP 1.1682 bits/point +EBPFP 2.3365 equivalent bits/point +MSE 99.101371 +---------------------- -------------------------------------------------------- +Time: 0.702s Load: 0.008s, Pack+Encode: 0.283s, Decode+Unpack: 0.411s +---------------------- -------------------------------------------------------- +💾 Converting with 99.1014 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 148, 128) +Output shape: (1, 148, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.output: torch.Size([1, 148, 3584]) -> torch.Size([1, 1, 148, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,064B, BPFP=0.5346 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,876B, BPFP=2.2040 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,408B, BPFP=0.9932 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,248B, BPFP=2.1377 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,920B, BPFP=1.2584 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,816B, BPFP=2.0921 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,240B, BPFP=1.1867 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,124B, BPFP=2.1246 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,960B, BPFP=1.7905 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,256B, BPFP=2.0329 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,484B, BPFP=0.6407 +⌛️ [2/4] FRONTEND: Frontend time: 0.284s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08608804 52.44859850 + layer.0.v_cache 0.00001767 0.01046920 + layer.1.k_cache 0.12276106 4.69259726 + layer.1.v_cache 0.00000594 0.00411723 + layer.2.k_cache 0.00886028 0.73459986 + layer.2.v_cache 0.00002088 0.01088886 + layer.3.k_cache 0.01513371 3.50998667 + layer.3.v_cache 0.00002085 0.01238878 + layer.4.k_cache 0.00068430 0.21856880 + layer.4.v_cache 0.00005331 0.02221044 + layer.4.output 0.08957551 365.19489020 + ------------------------------------------------------------------------------------- + TOTAL 0.05062792 154.00168571 + (elements=1,288,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1288192 +Total Bytes 197396 +BPFP 1.2259 bits/point +EBPFP 2.4518 equivalent bits/point +MSE 154.001686 +---------------------- -------------------------------------------------------- +Time: 0.664s Load: 0.006s, Pack+Encode: 0.284s, Decode+Unpack: 0.374s +---------------------- -------------------------------------------------------- +💾 Converting with 154.0017 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 170, 128) +Output shape: (1, 170, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.output: torch.Size([1, 170, 3584]) -> torch.Size([1, 1, 170, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,600B, BPFP=0.5147 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,276B, BPFP=2.1393 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,340B, BPFP=0.9504 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,616B, BPFP=2.0787 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,180B, BPFP=1.2114 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,844B, BPFP=2.0077 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,356B, BPFP=1.1357 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,372B, BPFP=2.0562 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,620B, BPFP=1.7114 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,276B, BPFP=1.9555 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 48,732B, BPFP=0.6399 +⌛️ [2/4] FRONTEND: Frontend time: 0.310s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.394s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13519069 53.81096048 + layer.0.v_cache 0.00001779 0.01109164 + layer.1.k_cache 0.14803200 4.33407018 + layer.1.v_cache 0.00000678 0.00439421 + layer.2.k_cache 0.02267574 0.81972692 + layer.2.v_cache 0.00002005 0.01111274 + layer.3.k_cache 0.04618364 2.97054659 + layer.3.v_cache 0.00002282 0.01274250 + layer.4.k_cache 0.00077641 0.21926781 + layer.4.v_cache 0.00005264 0.02216441 + layer.4.output 0.00945594 321.68211660 + ------------------------------------------------------------------------------------- + TOTAL 0.02465707 136.11711139 + (elements=1,479,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1479680 +Total Bytes 220212 +BPFP 1.1906 bits/point +EBPFP 2.3812 equivalent bits/point +MSE 136.117111 +---------------------- -------------------------------------------------------- +Time: 0.710s Load: 0.007s, Pack+Encode: 0.310s, Decode+Unpack: 0.394s +---------------------- -------------------------------------------------------- +💾 Converting with 136.1171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,492B, BPFP=0.5169 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,108B, BPFP=2.1751 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,252B, BPFP=0.9650 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,348B, BPFP=2.1035 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,316B, BPFP=1.2534 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,992B, BPFP=2.0700 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,404B, BPFP=1.1675 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,368B, BPFP=2.1054 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,700B, BPFP=1.7602 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,340B, BPFP=2.0087 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 49,644B, BPFP=0.6675 +⌛️ [2/4] FRONTEND: Frontend time: 0.326s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.352s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12520085 54.33314900 + layer.0.v_cache 0.00001716 0.01117970 + layer.1.k_cache 0.15152623 5.33080246 + layer.1.v_cache 0.00000633 0.00425649 + layer.2.k_cache 0.01605446 0.81293313 + layer.2.v_cache 0.00002200 0.01158017 + layer.3.k_cache 0.04332644 2.98250129 + layer.3.v_cache 0.00002349 0.01316133 + layer.4.k_cache 0.00069336 0.22796668 + layer.4.v_cache 0.00005614 0.02248019 + layer.4.output 0.00966471 329.94425022 + ------------------------------------------------------------------------------------- + TOTAL 0.02379879 139.60939776 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 220964 +BPFP 1.2234 bits/point +EBPFP 2.4469 equivalent bits/point +MSE 139.609398 +---------------------- -------------------------------------------------------- +Time: 0.687s Load: 0.008s, Pack+Encode: 0.326s, Decode+Unpack: 0.352s +---------------------- -------------------------------------------------------- +💾 Converting with 139.6094 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 181, 128) +Output shape: (1, 181, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.output: torch.Size([1, 181, 3584]) -> torch.Size([1, 1, 181, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,696B, BPFP=0.4917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,828B, BPFP=2.0570 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,588B, BPFP=0.9140 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,068B, BPFP=1.9914 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,612B, BPFP=1.1751 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,632B, BPFP=1.9537 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,872B, BPFP=1.1112 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,060B, BPFP=1.9907 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,264B, BPFP=1.6630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,944B, BPFP=1.8943 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 49,920B, BPFP=0.6156 +⌛️ [2/4] FRONTEND: Frontend time: 0.284s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.397s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13566811 50.25502849 + layer.0.v_cache 0.00002016 0.01098426 + layer.1.k_cache 0.14782082 4.38972330 + layer.1.v_cache 0.00000634 0.00457339 + layer.2.k_cache 0.01546242 0.75782784 + layer.2.v_cache 0.00002121 0.01128183 + layer.3.k_cache 0.02895848 3.17162277 + layer.3.v_cache 0.00002197 0.01262902 + layer.4.k_cache 0.00067433 0.22567254 + layer.4.v_cache 0.00005229 0.02248542 + layer.4.output 0.00892717 301.71100533 + ------------------------------------------------------------------------------------- + TOTAL 0.02301155 127.69640389 + (elements=1,575,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1575424 +Total Bytes 226484 +BPFP 1.1501 bits/point +EBPFP 2.3002 equivalent bits/point +MSE 127.696404 +---------------------- -------------------------------------------------------- +Time: 0.688s Load: 0.007s, Pack+Encode: 0.284s, Decode+Unpack: 0.397s +---------------------- -------------------------------------------------------- +💾 Converting with 127.6964 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 186, 128) +Output shape: (1, 186, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.output: torch.Size([1, 186, 3584]) -> torch.Size([1, 1, 186, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,904B, BPFP=0.4960 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,128B, BPFP=2.0269 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,016B, BPFP=0.9254 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,388B, BPFP=1.9647 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,036B, BPFP=1.1791 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,812B, BPFP=1.9163 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,168B, BPFP=1.1062 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,392B, BPFP=1.9651 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,628B, BPFP=1.6489 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,248B, BPFP=1.8690 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 53,996B, BPFP=0.6480 +⌛️ [2/4] FRONTEND: Frontend time: 0.306s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.337s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09795722 53.63434560 + layer.0.v_cache 0.00001815 0.01066508 + layer.1.k_cache 0.13430934 4.39435602 + layer.1.v_cache 0.00000632 0.00423306 + layer.2.k_cache 0.01830538 0.78939179 + layer.2.v_cache 0.00002120 0.01068722 + layer.3.k_cache 0.04892764 3.02700117 + layer.3.v_cache 0.00002113 0.01213763 + layer.4.k_cache 0.00067235 0.21326309 + layer.4.v_cache 0.00008605 0.02120491 + layer.4.output 0.00869856 291.19501248 + ------------------------------------------------------------------------------------- + TOTAL 0.02124792 123.55778664 + (elements=1,618,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1618944 +Total Bytes 233716 +BPFP 1.1549 bits/point +EBPFP 2.3098 equivalent bits/point +MSE 123.557787 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.009s, Pack+Encode: 0.306s, Decode+Unpack: 0.337s +---------------------- -------------------------------------------------------- +💾 Converting with 123.5578 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 163, 128) +Output shape: (1, 163, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.output: torch.Size([1, 163, 3584]) -> torch.Size([1, 1, 163, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,268B, BPFP=0.5050 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,468B, BPFP=2.1538 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,028B, BPFP=0.9613 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,632B, BPFP=2.0736 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,808B, BPFP=1.2278 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,144B, BPFP=2.0268 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,232B, BPFP=1.1725 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,692B, BPFP=2.0794 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,184B, BPFP=1.7431 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,628B, BPFP=1.9774 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 47,584B, BPFP=0.6516 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.340s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11530171 54.14433953 + layer.0.v_cache 0.00001813 0.01070459 + layer.1.k_cache 0.14319220 4.92698847 + layer.1.v_cache 0.00000612 0.00421272 + layer.2.k_cache 0.02368733 0.77150104 + layer.2.v_cache 0.00002106 0.01064574 + layer.3.k_cache 0.02525310 3.41559175 + layer.3.v_cache 0.00002161 0.01251094 + layer.4.k_cache 0.00069984 0.22168168 + layer.4.v_cache 0.00005709 0.02145663 + layer.4.output 0.00982084 335.99950701 + ------------------------------------------------------------------------------------- + TOTAL 0.02217671 142.09036366 + (elements=1,418,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1418752 +Total Bytes 213668 +BPFP 1.2048 bits/point +EBPFP 2.4096 equivalent bits/point +MSE 142.090364 +---------------------- -------------------------------------------------------- +Time: 0.609s Load: 0.006s, Pack+Encode: 0.262s, Decode+Unpack: 0.340s +---------------------- -------------------------------------------------------- +💾 Converting with 142.0904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 191, 128) +Output shape: (1, 191, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.output: torch.Size([1, 191, 3584]) -> torch.Size([1, 1, 191, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,320B, BPFP=0.5170 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 24,204B, BPFP=1.9800 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,372B, BPFP=0.9303 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,604B, BPFP=1.9310 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 14,636B, BPFP=1.1973 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 23,172B, BPFP=1.8956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 13,852B, BPFP=1.1332 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,576B, BPFP=1.9287 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 20,064B, BPFP=1.6414 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,676B, BPFP=1.8550 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 55,932B, BPFP=0.6537 +⌛️ [2/4] FRONTEND: Frontend time: 0.325s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15112510 49.68559800 + layer.0.v_cache 0.00002069 0.01087726 + layer.1.k_cache 0.18491925 4.75408968 + layer.1.v_cache 0.00000629 0.00430988 + layer.2.k_cache 0.02122488 0.73976031 + layer.2.v_cache 0.00002327 0.01159655 + layer.3.k_cache 0.01259250 3.01839044 + layer.3.v_cache 0.00002182 0.01243054 + layer.4.k_cache 0.00068006 0.21754755 + layer.4.v_cache 0.00005885 0.02209837 + layer.4.output 0.00845948 267.18305909 + ------------------------------------------------------------------------------------- + TOTAL 0.02528759 113.45635954 + (elements=1,662,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1662464 +Total Bytes 239408 +BPFP 1.1521 bits/point +EBPFP 2.3041 equivalent bits/point +MSE 113.456360 +---------------------- -------------------------------------------------------- +Time: 0.701s Load: 0.007s, Pack+Encode: 0.325s, Decode+Unpack: 0.369s +---------------------- -------------------------------------------------------- +💾 Converting with 113.4564 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 174, 128) +Output shape: (1, 174, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.output: torch.Size([1, 174, 3584]) -> torch.Size([1, 1, 174, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,632B, BPFP=0.5057 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,608B, BPFP=2.1200 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,484B, BPFP=0.9415 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,892B, BPFP=2.0557 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,616B, BPFP=1.2227 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,580B, BPFP=2.0277 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,824B, BPFP=1.1516 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,936B, BPFP=2.0596 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,312B, BPFP=1.7342 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,988B, BPFP=1.9745 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 50,316B, BPFP=0.6455 +⌛️ [2/4] FRONTEND: Frontend time: 0.290s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.416s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14146894 52.54283405 + layer.0.v_cache 0.00001664 0.01041862 + layer.1.k_cache 0.16292955 4.54745589 + layer.1.v_cache 0.00000602 0.00404669 + layer.2.k_cache 0.01834387 0.78170049 + layer.2.v_cache 0.00002001 0.01099969 + layer.3.k_cache 0.03015787 3.24404416 + layer.3.v_cache 0.00002180 0.01237687 + layer.4.k_cache 0.00067479 0.21703450 + layer.4.v_cache 0.00005130 0.02209066 + layer.4.output 0.00923280 314.50638855 + ------------------------------------------------------------------------------------- + TOTAL 0.02460708 133.11398361 + (elements=1,514,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1514496 +Total Bytes 226188 +BPFP 1.1948 bits/point +EBPFP 2.3896 equivalent bits/point +MSE 133.113984 +---------------------- -------------------------------------------------------- +Time: 0.715s Load: 0.009s, Pack+Encode: 0.290s, Decode+Unpack: 0.416s +---------------------- -------------------------------------------------------- +💾 Converting with 133.1140 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.1647 bits/point +Avg EBPFP 2.3294 equivalent bits/point +Avg MSE 103.463931 +Avg Time 0.773s +------------------------ ---------------------------- diff --git a/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..b09ebeb190f85753c461e3bef7688abf09ad1387 --- /dev/null +++ b/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 599 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa +Output output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,780B, BPFP=0.4763 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,484B, BPFP=1.9794 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,944B, BPFP=0.8650 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,348B, BPFP=1.9178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,132B, BPFP=1.0922 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,580B, BPFP=1.8761 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,032B, BPFP=1.0326 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,052B, BPFP=1.9017 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,412B, BPFP=1.5957 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,776B, BPFP=1.8325 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,300B, BPFP=0.5139 +⌛️ [2/4] FRONTEND: Frontend time: 0.506s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.432s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14134069 53.17666626 + layer.0.v_cache 0.00001405 0.00737295 + layer.1.k_cache 0.62039534 4.33691237 + layer.1.v_cache 0.00000582 0.00314553 + layer.2.k_cache 0.00676800 0.73605061 + layer.2.v_cache 0.00001893 0.00875241 + layer.3.k_cache 0.03896382 2.96712282 + layer.3.v_cache 0.00001942 0.00974739 + layer.4.k_cache 0.00069280 0.17957040 + layer.4.v_cache 0.00005238 0.01943978 + layer.4.output 0.00804578 189.28476873 + ------------------------------------------------------------------------------------- + TOTAL 0.05085834 81.55518598 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 334840 +BPFP 1.0686 bits/point +EBPFP 2.1372 equivalent bits/point +MSE 81.555186 +---------------------- -------------------------------------------------------- +Time: 0.949s Load: 0.012s, Pack+Encode: 0.506s, Decode+Unpack: 0.432s +---------------------- -------------------------------------------------------- +💾 Converting with 81.5552 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,916B, BPFP=0.4755 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,616B, BPFP=2.0060 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,464B, BPFP=0.8780 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,636B, BPFP=1.9537 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,792B, BPFP=1.1088 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,844B, BPFP=1.9115 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,664B, BPFP=1.0486 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,356B, BPFP=1.9388 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,332B, BPFP=1.6175 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,012B, BPFP=1.8671 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,964B, BPFP=0.5406 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15682776 53.23892785 + layer.0.v_cache 0.00001501 0.00774476 + layer.1.k_cache 0.54447479 4.29198802 + layer.1.v_cache 0.00000580 0.00329164 + layer.2.k_cache 0.00853899 0.71596498 + layer.2.v_cache 0.00001972 0.00865761 + layer.3.k_cache 0.02401569 2.84532358 + layer.3.v_cache 0.00001939 0.00961293 + layer.4.k_cache 0.00071059 0.18124923 + layer.4.v_cache 0.00005329 0.01936196 + layer.4.output 0.05121508 184.23051255 + ------------------------------------------------------------------------------------- + TOTAL 0.06430510 79.46680650 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 348596 +BPFP 1.0935 bits/point +EBPFP 2.1870 equivalent bits/point +MSE 79.466806 +---------------------- -------------------------------------------------------- +Time: 0.663s Load: 0.011s, Pack+Encode: 0.262s, Decode+Unpack: 0.390s +---------------------- -------------------------------------------------------- +💾 Converting with 79.4668 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,084B, BPFP=0.4828 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,612B, BPFP=1.9989 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,640B, BPFP=0.8844 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,728B, BPFP=1.9520 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,644B, BPFP=1.0972 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,748B, BPFP=1.8999 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,700B, BPFP=1.0470 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,260B, BPFP=1.9271 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,172B, BPFP=1.6035 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,796B, BPFP=1.8493 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,704B, BPFP=0.5520 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.392s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15363285 52.57433169 + layer.0.v_cache 0.00001412 0.00731584 + layer.1.k_cache 0.60852461 4.55378401 + layer.1.v_cache 0.00000580 0.00299656 + layer.2.k_cache 0.01182756 0.71504393 + layer.2.v_cache 0.00001863 0.00804362 + layer.3.k_cache 0.03467803 2.95328921 + layer.3.v_cache 0.00001919 0.00927024 + layer.4.k_cache 0.00072319 0.17133191 + layer.4.v_cache 0.00005301 0.01759062 + layer.4.output 0.05035121 183.46387573 + ------------------------------------------------------------------------------------- + TOTAL 0.06835032 79.13294869 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 350088 +BPFP 1.0945 bits/point +EBPFP 2.1889 equivalent bits/point +MSE 79.132949 +---------------------- -------------------------------------------------------- +Time: 0.664s Load: 0.010s, Pack+Encode: 0.261s, Decode+Unpack: 0.392s +---------------------- -------------------------------------------------------- +💾 Converting with 79.1329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,936B, BPFP=0.4934 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,200B, BPFP=1.9987 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,060B, BPFP=0.8867 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,188B, BPFP=1.9428 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,248B, BPFP=1.1179 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,592B, BPFP=1.9099 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,056B, BPFP=1.0521 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,044B, BPFP=1.9348 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,320B, BPFP=1.6188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,848B, BPFP=1.8688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,252B, BPFP=0.5462 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.391s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12197856 52.04470103 + layer.0.v_cache 0.00001365 0.00737548 + layer.1.k_cache 0.62524759 4.43570733 + layer.1.v_cache 0.00000551 0.00289633 + layer.2.k_cache 0.01486346 0.76511914 + layer.2.v_cache 0.00001981 0.00861816 + layer.3.k_cache 0.07387851 3.15366086 + layer.3.v_cache 0.00001943 0.00920612 + layer.4.k_cache 0.00069560 0.17420980 + layer.4.v_cache 0.00005405 0.01893897 + layer.4.output 0.01098318 195.18759465 + ------------------------------------------------------------------------------------- + TOTAL 0.05374461 83.93727034 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 337744 +BPFP 1.0969 bits/point +EBPFP 2.1938 equivalent bits/point +MSE 83.937270 +---------------------- -------------------------------------------------------- +Time: 0.662s Load: 0.011s, Pack+Encode: 0.260s, Decode+Unpack: 0.391s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9373 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,984B, BPFP=0.4925 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,252B, BPFP=1.9875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,336B, BPFP=0.8956 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,132B, BPFP=1.9261 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,072B, BPFP=1.1004 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,432B, BPFP=1.8877 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,132B, BPFP=1.0489 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,100B, BPFP=1.9243 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,216B, BPFP=1.6018 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,712B, BPFP=1.8482 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,416B, BPFP=0.5515 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.392s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15362121 51.78549548 + layer.0.v_cache 0.00001420 0.00757171 + layer.1.k_cache 0.52952420 4.52153449 + layer.1.v_cache 0.00000567 0.00316155 + layer.2.k_cache 0.00939641 0.69729422 + layer.2.v_cache 0.00001891 0.00835261 + layer.3.k_cache 0.04339437 3.06569824 + layer.3.v_cache 0.00002732 0.00939268 + layer.4.k_cache 0.00068326 0.17589301 + layer.4.v_cache 0.00005156 0.01849114 + layer.4.output 0.00776138 193.23128133 + ------------------------------------------------------------------------------------- + TOTAL 0.04653334 83.11246202 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 338784 +BPFP 1.0926 bits/point +EBPFP 2.1851 equivalent bits/point +MSE 83.112462 +---------------------- -------------------------------------------------------- +Time: 0.663s Load: 0.010s, Pack+Encode: 0.261s, Decode+Unpack: 0.392s +---------------------- -------------------------------------------------------- +💾 Converting with 83.1125 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,148B, BPFP=0.4862 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,520B, BPFP=1.9940 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,724B, BPFP=0.8888 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,696B, BPFP=1.9503 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,904B, BPFP=1.1110 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,996B, BPFP=1.9131 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,712B, BPFP=1.0476 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,380B, BPFP=1.9335 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,580B, BPFP=1.6252 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,968B, BPFP=1.8584 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,312B, BPFP=0.5338 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14921317 52.80955703 + layer.0.v_cache 0.00001395 0.00767009 + layer.1.k_cache 0.64083395 4.47194874 + layer.1.v_cache 0.00000574 0.00331220 + layer.2.k_cache 0.00578804 0.73529691 + layer.2.v_cache 0.00001912 0.00899184 + layer.3.k_cache 0.02848778 2.97635801 + layer.3.v_cache 0.00001956 0.00956578 + layer.4.k_cache 0.00070939 0.17726283 + layer.4.v_cache 0.00005206 0.01894346 + layer.4.output 0.04916091 183.62269193 + ------------------------------------------------------------------------------------- + TOTAL 0.06878054 79.21045591 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 348940 +BPFP 1.0909 bits/point +EBPFP 2.1817 equivalent bits/point +MSE 79.210456 +---------------------- -------------------------------------------------------- +Time: 0.657s Load: 0.011s, Pack+Encode: 0.259s, Decode+Unpack: 0.388s +---------------------- -------------------------------------------------------- +💾 Converting with 79.2105 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,964B, BPFP=0.4780 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,500B, BPFP=1.9998 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,372B, BPFP=0.8731 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,408B, BPFP=1.9416 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,640B, BPFP=1.1007 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,620B, BPFP=1.8995 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,568B, BPFP=1.0435 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,264B, BPFP=1.9339 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,164B, BPFP=1.6086 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,756B, BPFP=1.8535 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,504B, BPFP=0.5447 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12806672 53.36935394 + layer.0.v_cache 0.00001388 0.00727146 + layer.1.k_cache 0.56701790 4.43206808 + layer.1.v_cache 0.00000557 0.00292241 + layer.2.k_cache 0.01020011 0.69588817 + layer.2.v_cache 0.00001881 0.00815610 + layer.3.k_cache 0.04301871 3.01114923 + layer.3.v_cache 0.00002051 0.00933538 + layer.4.k_cache 0.00071755 0.17141261 + layer.4.v_cache 0.00005284 0.01768595 + layer.4.output 0.04940383 184.48934971 + ------------------------------------------------------------------------------------- + TOTAL 0.06440938 79.59709949 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 347760 +BPFP 1.0909 bits/point +EBPFP 2.1818 equivalent bits/point +MSE 79.597099 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.011s, Pack+Encode: 0.257s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 79.5971 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,972B, BPFP=0.4919 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,356B, BPFP=1.9932 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,484B, BPFP=0.9037 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,384B, BPFP=1.9399 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,368B, BPFP=1.1167 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,720B, BPFP=1.9035 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,284B, BPFP=1.0572 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,260B, BPFP=1.9331 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,504B, BPFP=1.6175 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,912B, BPFP=1.8592 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,712B, BPFP=0.5382 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12903612 52.62133703 + layer.0.v_cache 0.00001458 0.00738053 + layer.1.k_cache 0.59860808 4.67332657 + layer.1.v_cache 0.00000581 0.00307374 + layer.2.k_cache 0.00940504 0.72224325 + layer.2.v_cache 0.00001954 0.00856774 + layer.3.k_cache 0.02644065 3.11614755 + layer.3.v_cache 0.00002029 0.00948219 + layer.4.k_cache 0.00068693 0.17671669 + layer.4.v_cache 0.00005376 0.01897941 + layer.4.output 0.00862506 193.97984023 + ------------------------------------------------------------------------------------- + TOTAL 0.04850978 83.48330213 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 338956 +BPFP 1.0931 bits/point +EBPFP 2.1862 equivalent bits/point +MSE 83.483302 +---------------------- -------------------------------------------------------- +Time: 0.642s Load: 0.009s, Pack+Encode: 0.254s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 83.4833 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,336B, BPFP=0.4736 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,320B, BPFP=1.9440 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,860B, BPFP=0.8553 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,308B, BPFP=1.8927 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,168B, BPFP=1.0739 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,612B, BPFP=1.8573 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,108B, BPFP=1.0201 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,104B, BPFP=1.8823 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,824B, BPFP=1.5637 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,544B, BPFP=1.8032 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,796B, BPFP=0.5131 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16469599 51.57918400 + layer.0.v_cache 0.00001357 0.00730363 + layer.1.k_cache 0.70445866 4.58841497 + layer.1.v_cache 0.00000558 0.00316413 + layer.2.k_cache 0.01447187 0.72922902 + layer.2.v_cache 0.00001924 0.00843915 + layer.3.k_cache 0.02617162 3.08605779 + layer.3.v_cache 0.00001974 0.00938667 + layer.4.k_cache 0.00070375 0.17406119 + layer.4.v_cache 0.00005076 0.01822453 + layer.4.output 0.04814868 174.59969272 + ------------------------------------------------------------------------------------- + TOTAL 0.07339127 75.43537142 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 353980 +BPFP 1.0563 bits/point +EBPFP 2.1127 equivalent bits/point +MSE 75.435371 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.250s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 75.4354 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,536B, BPFP=0.4958 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,740B, BPFP=2.0179 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,480B, BPFP=0.8992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,836B, BPFP=1.9654 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,340B, BPFP=1.1234 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,944B, BPFP=1.9136 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,288B, BPFP=1.0623 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,516B, BPFP=1.9468 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,044B, BPFP=1.6289 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,228B, BPFP=1.8720 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,384B, BPFP=0.5177 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12132761 52.96679688 + layer.0.v_cache 0.00001366 0.00721066 + layer.1.k_cache 0.57241889 4.50436050 + layer.1.v_cache 0.00000556 0.00281565 + layer.2.k_cache 0.00926234 0.73652762 + layer.2.v_cache 0.00002145 0.00802807 + layer.3.k_cache 0.03109839 3.07734130 + layer.3.v_cache 0.00002257 0.00872674 + layer.4.k_cache 0.00070701 0.16865761 + layer.4.v_cache 0.00004920 0.01741542 + layer.4.output 0.01031505 199.62353956 + ------------------------------------------------------------------------------------- + TOTAL 0.04747836 85.81545043 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 319336 +BPFP 1.0911 bits/point +EBPFP 2.1822 equivalent bits/point +MSE 85.815450 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.010s, Pack+Encode: 0.260s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8155 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,188B, BPFP=0.4818 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,888B, BPFP=1.9866 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,676B, BPFP=0.8744 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,712B, BPFP=1.9249 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,924B, BPFP=1.0971 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,032B, BPFP=1.8893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,632B, BPFP=1.0294 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,428B, BPFP=1.9100 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,404B, BPFP=1.5942 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,996B, BPFP=1.8349 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,520B, BPFP=0.5282 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14357947 52.51467138 + layer.0.v_cache 0.00001420 0.00777326 + layer.1.k_cache 0.63723908 4.52678993 + layer.1.v_cache 0.00000573 0.00322947 + layer.2.k_cache 0.00765126 0.67989590 + layer.2.v_cache 0.00002073 0.00857530 + layer.3.k_cache 0.02843528 2.89835312 + layer.3.v_cache 0.00002112 0.00973535 + layer.4.k_cache 0.00069171 0.17407204 + layer.4.v_cache 0.00005234 0.01878664 + layer.4.output 0.04871205 180.87874521 + ------------------------------------------------------------------------------------- + TOTAL 0.06815855 78.05841758 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 349400 +BPFP 1.0777 bits/point +EBPFP 2.1553 equivalent bits/point +MSE 78.058418 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 78.0584 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,488B, BPFP=0.4861 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,324B, BPFP=1.9633 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,020B, BPFP=0.8719 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,180B, BPFP=1.9047 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,376B, BPFP=1.0951 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,592B, BPFP=1.8746 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,252B, BPFP=1.0375 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,964B, BPFP=1.8936 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,960B, BPFP=1.5861 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,488B, BPFP=1.8180 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,204B, BPFP=0.5138 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.396s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16857396 52.08157659 + layer.0.v_cache 0.00001478 0.00771490 + layer.1.k_cache 0.58716421 4.43609399 + layer.1.v_cache 0.00000600 0.00336161 + layer.2.k_cache 0.00827454 0.69017329 + layer.2.v_cache 0.00002177 0.00884673 + layer.3.k_cache 0.02531170 3.12760690 + layer.3.v_cache 0.00001983 0.00962784 + layer.4.k_cache 0.00069463 0.17624080 + layer.4.v_cache 0.00005316 0.01951931 + layer.4.output 0.04836898 176.14735070 + ------------------------------------------------------------------------------------- + TOTAL 0.06639514 76.09365982 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 353848 +BPFP 1.0663 bits/point +EBPFP 2.1326 equivalent bits/point +MSE 76.093660 +---------------------- -------------------------------------------------------- +Time: 0.666s Load: 0.010s, Pack+Encode: 0.260s, Decode+Unpack: 0.396s +---------------------- -------------------------------------------------------- +💾 Converting with 76.0937 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,256B, BPFP=0.4853 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,856B, BPFP=1.9849 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,696B, BPFP=0.8754 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,864B, BPFP=1.9329 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,832B, BPFP=1.0923 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,144B, BPFP=1.8951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,792B, BPFP=1.0378 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,680B, BPFP=1.9232 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,524B, BPFP=1.6005 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,140B, BPFP=1.8425 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,040B, BPFP=0.5321 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14028978 52.11317311 + layer.0.v_cache 0.00001397 0.00763416 + layer.1.k_cache 0.59966084 4.34128361 + layer.1.v_cache 0.00000578 0.00323313 + layer.2.k_cache 0.01765286 0.74021277 + layer.2.v_cache 0.00001884 0.00863692 + layer.3.k_cache 0.02653811 3.27037673 + layer.3.v_cache 0.00001935 0.00964727 + layer.4.k_cache 0.00071527 0.18469666 + layer.4.v_cache 0.00005402 0.01946576 + layer.4.output 0.04907703 180.27296560 + ------------------------------------------------------------------------------------- + TOTAL 0.06638282 77.80053643 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 350824 +BPFP 1.0820 bits/point +EBPFP 2.1641 equivalent bits/point +MSE 77.800536 +---------------------- -------------------------------------------------------- +Time: 0.659s Load: 0.010s, Pack+Encode: 0.260s, Decode+Unpack: 0.390s +---------------------- -------------------------------------------------------- +💾 Converting with 77.8005 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,736B, BPFP=0.5018 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,336B, BPFP=2.0299 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,684B, BPFP=0.9010 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,284B, BPFP=1.9694 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,656B, BPFP=1.1291 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,640B, BPFP=1.9324 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,508B, BPFP=1.0632 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,160B, BPFP=1.9623 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,196B, BPFP=1.6197 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,724B, BPFP=1.8798 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,912B, BPFP=0.5573 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13255837 53.47189870 + layer.0.v_cache 0.00001394 0.00706129 + layer.1.k_cache 0.54921044 4.48317853 + layer.1.v_cache 0.00000578 0.00290559 + layer.2.k_cache 0.01292971 0.70622657 + layer.2.v_cache 0.00001901 0.00821347 + layer.3.k_cache 0.07190662 3.05631458 + layer.3.v_cache 0.00001944 0.00872324 + layer.4.k_cache 0.00068159 0.16661609 + layer.4.v_cache 0.00005125 0.01763303 + layer.4.output 0.00790709 202.29776129 + ------------------------------------------------------------------------------------- + TOTAL 0.04839681 86.94194707 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 328836 +BPFP 1.1112 bits/point +EBPFP 2.2223 equivalent bits/point +MSE 86.941947 +---------------------- -------------------------------------------------------- +Time: 0.653s Load: 0.011s, Pack+Encode: 0.262s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9419 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,916B, BPFP=0.4940 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,208B, BPFP=2.0062 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,032B, BPFP=0.8883 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,100B, BPFP=1.9448 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,452B, BPFP=1.1332 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,408B, BPFP=1.9065 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,320B, BPFP=1.0705 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,028B, BPFP=1.9408 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,352B, BPFP=1.6263 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,808B, BPFP=1.8732 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,952B, BPFP=0.5616 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13695534 52.59230247 + layer.0.v_cache 0.00001452 0.00769310 + layer.1.k_cache 0.52272531 4.39903833 + layer.1.v_cache 0.00000574 0.00321078 + layer.2.k_cache 0.00793556 0.75505840 + layer.2.v_cache 0.00001907 0.00885213 + layer.3.k_cache 0.04682703 3.09218819 + layer.3.v_cache 0.00002008 0.00967639 + layer.4.k_cache 0.00068945 0.17870159 + layer.4.v_cache 0.00005221 0.01982900 + layer.4.output 0.00941004 195.14855623 + ------------------------------------------------------------------------------------- + TOTAL 0.04594792 83.94743788 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 339576 +BPFP 1.1068 bits/point +EBPFP 2.2135 equivalent bits/point +MSE 83.947438 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.009s, Pack+Encode: 0.252s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9474 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,912B, BPFP=0.4973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,948B, BPFP=2.0060 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,928B, BPFP=0.8888 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,844B, BPFP=1.9444 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,048B, BPFP=1.1187 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,064B, BPFP=1.9009 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,024B, BPFP=1.0616 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,720B, BPFP=1.9375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,800B, BPFP=1.6071 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,380B, BPFP=1.8627 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,472B, BPFP=0.5379 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14340984 52.59239676 + layer.0.v_cache 0.00001397 0.00736377 + layer.1.k_cache 0.51622396 4.22721122 + layer.1.v_cache 0.00000557 0.00303506 + layer.2.k_cache 0.00797994 0.71791447 + layer.2.v_cache 0.00001813 0.00818234 + layer.3.k_cache 0.02756977 3.07335074 + layer.3.v_cache 0.00001890 0.00901769 + layer.4.k_cache 0.00068956 0.17233682 + layer.4.v_cache 0.00005177 0.01809171 + layer.4.output 0.00947078 196.49094388 + ------------------------------------------------------------------------------------- + TOTAL 0.04483982 84.48620634 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 333140 +BPFP 1.0936 bits/point +EBPFP 2.1871 equivalent bits/point +MSE 84.486206 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 84.4862 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,068B, BPFP=0.4819 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,468B, BPFP=1.9913 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,768B, BPFP=0.8912 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,464B, BPFP=1.9379 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,700B, BPFP=1.1001 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,784B, BPFP=1.9018 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,568B, BPFP=1.0400 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,196B, BPFP=1.9237 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,360B, BPFP=1.6135 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,928B, BPFP=1.8563 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,836B, BPFP=0.5302 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13622742 52.64490327 + layer.0.v_cache 0.00001380 0.00759803 + layer.1.k_cache 0.61520739 4.50818079 + layer.1.v_cache 0.00000584 0.00326835 + layer.2.k_cache 0.00900942 0.67973219 + layer.2.v_cache 0.00001916 0.00872524 + layer.3.k_cache 0.02667915 2.91284636 + layer.3.v_cache 0.00001881 0.00933731 + layer.4.k_cache 0.00071466 0.17718772 + layer.4.v_cache 0.00005760 0.01897405 + layer.4.output 0.04972813 183.42851069 + ------------------------------------------------------------------------------------- + TOTAL 0.06682648 79.11590166 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 347140 +BPFP 1.0852 bits/point +EBPFP 2.1705 equivalent bits/point +MSE 79.115902 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.009s, Pack+Encode: 0.253s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 79.1159 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,120B, BPFP=0.4766 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,016B, BPFP=1.9866 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,720B, BPFP=0.8737 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,892B, BPFP=1.9279 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,112B, BPFP=1.1033 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,236B, BPFP=1.8936 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,912B, BPFP=1.0406 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,868B, BPFP=1.9266 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,616B, BPFP=1.5999 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,264B, BPFP=1.8428 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,084B, BPFP=0.5307 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11478914 51.24407530 + layer.0.v_cache 0.00001377 0.00707337 + layer.1.k_cache 0.64218915 4.54239147 + layer.1.v_cache 0.00000556 0.00283034 + layer.2.k_cache 0.00898698 0.70187460 + layer.2.v_cache 0.00001980 0.00819629 + layer.3.k_cache 0.05442990 2.79752597 + layer.3.v_cache 0.00001964 0.00903439 + layer.4.k_cache 0.00070461 0.17278795 + layer.4.v_cache 0.00005015 0.01776568 + layer.4.output 0.05017195 179.33865862 + ------------------------------------------------------------------------------------- + TOTAL 0.06896543 77.34553916 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 351840 +BPFP 1.0815 bits/point +EBPFP 2.1631 equivalent bits/point +MSE 77.345539 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 77.3455 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,404B, BPFP=0.4955 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,316B, BPFP=2.0233 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,300B, BPFP=0.9021 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,484B, BPFP=1.9743 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,300B, BPFP=1.1380 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,680B, BPFP=1.9269 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,184B, BPFP=1.0722 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,232B, BPFP=1.9594 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,768B, BPFP=1.6373 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,872B, BPFP=1.8792 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,036B, BPFP=0.5141 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14755792 55.09454599 + layer.0.v_cache 0.00001375 0.00767853 + layer.1.k_cache 0.51940866 4.02978331 + layer.1.v_cache 0.00000586 0.00330693 + layer.2.k_cache 0.00816260 0.72523406 + layer.2.v_cache 0.00001900 0.00889404 + layer.3.k_cache 0.02921419 2.86282429 + layer.3.v_cache 0.00001979 0.00983233 + layer.4.k_cache 0.00071017 0.17901889 + layer.4.v_cache 0.00005507 0.01907497 + layer.4.output 0.00937099 207.28810647 + ------------------------------------------------------------------------------------- + TOTAL 0.04533906 89.05629051 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 315576 +BPFP 1.0945 bits/point +EBPFP 2.1891 equivalent bits/point +MSE 89.056291 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 89.0563 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,628B, BPFP=0.5049 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,608B, BPFP=2.0253 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,380B, BPFP=0.9000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,720B, BPFP=1.9733 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,320B, BPFP=1.1306 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,948B, BPFP=1.9281 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,360B, BPFP=1.0744 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,440B, BPFP=1.9569 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,076B, BPFP=1.6430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,192B, BPFP=1.8839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,660B, BPFP=0.5155 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12840328 52.37039516 + layer.0.v_cache 0.00001372 0.00730185 + layer.1.k_cache 0.60081110 4.48564371 + layer.1.v_cache 0.00000562 0.00298537 + layer.2.k_cache 0.00586274 0.70433507 + layer.2.v_cache 0.00001839 0.00830682 + layer.3.k_cache 0.02491910 3.06291033 + layer.3.v_cache 0.00001957 0.00949194 + layer.4.k_cache 0.00072047 0.17995271 + layer.4.v_cache 0.00005229 0.01912211 + layer.4.output 0.01050496 206.58722913 + ------------------------------------------------------------------------------------- + TOTAL 0.04908006 88.64476759 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 318332 +BPFP 1.0958 bits/point +EBPFP 2.1916 equivalent bits/point +MSE 88.644768 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6448 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,912B, BPFP=0.4920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,932B, BPFP=1.9839 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,860B, BPFP=0.8757 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,852B, BPFP=1.9242 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,056B, BPFP=1.1073 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,244B, BPFP=1.8907 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,924B, BPFP=1.0448 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,712B, BPFP=1.9165 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,984B, BPFP=1.6003 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,472B, BPFP=1.8481 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,164B, BPFP=0.5219 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14595435 52.96651736 + layer.0.v_cache 0.00001364 0.00715905 + layer.1.k_cache 0.59759904 4.38141667 + layer.1.v_cache 0.00000566 0.00294760 + layer.2.k_cache 0.01409306 0.69380921 + layer.2.v_cache 0.00002047 0.00815694 + layer.3.k_cache 0.06079736 2.97113792 + layer.3.v_cache 0.00002215 0.00900025 + layer.4.k_cache 0.00070776 0.17120905 + layer.4.v_cache 0.00005221 0.01825011 + layer.4.output 0.00769303 194.93884086 + ------------------------------------------------------------------------------------- + TOTAL 0.05135982 83.87067589 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 332112 +BPFP 1.0786 bits/point +EBPFP 2.1572 equivalent bits/point +MSE 83.870676 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.012s, Pack+Encode: 0.250s, Decode+Unpack: 0.376s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8707 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 315, 128) +Output shape: (1, 315, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.output: torch.Size([1, 315, 3584]) -> torch.Size([1, 1, 315, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,620B, BPFP=0.4772 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,800B, BPFP=1.9246 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,548B, BPFP=0.8704 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,800B, BPFP=1.8750 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,912B, BPFP=1.0869 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,960B, BPFP=1.8333 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,824B, BPFP=1.0329 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,532B, BPFP=1.8617 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,488B, BPFP=1.5619 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,972B, BPFP=1.7843 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,732B, BPFP=0.5083 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15500355 52.24064980 + layer.0.v_cache 0.00001366 0.00749284 + layer.1.k_cache 0.68167124 4.12451094 + layer.1.v_cache 0.00000571 0.00333971 + layer.2.k_cache 0.01007247 0.71446659 + layer.2.v_cache 0.00001936 0.00888204 + layer.3.k_cache 0.03272899 2.74070386 + layer.3.v_cache 0.00002006 0.00952792 + layer.4.k_cache 0.00071175 0.18031485 + layer.4.v_cache 0.00005427 0.01897810 + layer.4.output 0.04584261 170.32118764 + ------------------------------------------------------------------------------------- + TOTAL 0.07065879 73.66454001 + (elements=2,741,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2741760 +Total Bytes 360188 +BPFP 1.0510 bits/point +EBPFP 2.1019 equivalent bits/point +MSE 73.664540 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 73.6645 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,716B, BPFP=0.5025 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,868B, BPFP=2.0104 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,388B, BPFP=0.8872 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,748B, BPFP=1.9458 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,300B, BPFP=1.1128 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,208B, BPFP=1.9147 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,344B, BPFP=1.0577 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,708B, BPFP=1.9435 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,016B, BPFP=1.6153 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,256B, BPFP=1.8598 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,424B, BPFP=0.5059 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12889319 51.58124207 + layer.0.v_cache 0.00001350 0.00770049 + layer.1.k_cache 0.58775566 4.54202327 + layer.1.v_cache 0.00000553 0.00318392 + layer.2.k_cache 0.00723753 0.75961315 + layer.2.v_cache 0.00001924 0.00854196 + layer.3.k_cache 0.01761036 3.07420186 + layer.3.v_cache 0.00001847 0.00943664 + layer.4.k_cache 0.00071814 0.17698619 + layer.4.v_cache 0.00005270 0.01893629 + layer.4.output 0.01035140 202.62745453 + ------------------------------------------------------------------------------------- + TOTAL 0.04792848 86.97494397 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 318976 +BPFP 1.0818 bits/point +EBPFP 2.1637 equivalent bits/point +MSE 86.974944 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.009s, Pack+Encode: 0.255s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9749 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,820B, BPFP=0.4853 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,180B, BPFP=1.9905 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,152B, BPFP=0.8886 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,080B, BPFP=1.9300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,020B, BPFP=1.1015 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,288B, BPFP=1.8864 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,896B, BPFP=1.0396 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,720B, BPFP=1.9102 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,232B, BPFP=1.6083 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,512B, BPFP=1.8438 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,088B, BPFP=0.5116 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13033964 53.04201970 + layer.0.v_cache 0.00001368 0.00742161 + layer.1.k_cache 0.63029007 4.55629376 + layer.1.v_cache 0.00000551 0.00308386 + layer.2.k_cache 0.00637632 0.69405505 + layer.2.v_cache 0.00001960 0.00831054 + layer.3.k_cache 0.02368569 3.26636344 + layer.3.v_cache 0.00001974 0.00887095 + layer.4.k_cache 0.00068050 0.17352490 + layer.4.v_cache 0.00005209 0.01818696 + layer.4.output 0.00765416 193.92769115 + ------------------------------------------------------------------------------------- + TOTAL 0.04970952 83.48658640 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 331988 +BPFP 1.0744 bits/point +EBPFP 2.1488 equivalent bits/point +MSE 83.486586 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 83.4866 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,820B, BPFP=0.5067 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,208B, BPFP=2.0225 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,528B, BPFP=0.8920 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,104B, BPFP=1.9591 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,592B, BPFP=1.1255 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,500B, BPFP=1.9244 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,588B, BPFP=1.0678 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,160B, BPFP=1.9623 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,336B, BPFP=1.6278 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,784B, BPFP=1.8833 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,896B, BPFP=0.5244 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15644295 52.48624196 + layer.0.v_cache 0.00001405 0.00745756 + layer.1.k_cache 0.56163917 4.53749354 + layer.1.v_cache 0.00000554 0.00314591 + layer.2.k_cache 0.01059573 0.71411155 + layer.2.v_cache 0.00001807 0.00850654 + layer.3.k_cache 0.04944291 3.12130019 + layer.3.v_cache 0.00001950 0.00938704 + layer.4.k_cache 0.00071267 0.17642582 + layer.4.v_cache 0.00005817 0.01977343 + layer.4.output 0.00956290 202.88991925 + ------------------------------------------------------------------------------------- + TOTAL 0.04975818 87.13607519 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 324516 +BPFP 1.0966 bits/point +EBPFP 2.1931 equivalent bits/point +MSE 87.136075 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.009s, Pack+Encode: 0.250s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 87.1361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 312, 128) +Output shape: (1, 312, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.output: torch.Size([1, 312, 3584]) -> torch.Size([1, 1, 312, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,684B, BPFP=0.4850 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,736B, BPFP=1.9399 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,308B, BPFP=0.8668 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,676B, BPFP=1.8868 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,584B, BPFP=1.0809 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,972B, BPFP=1.8516 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,372B, BPFP=1.0202 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,452B, BPFP=1.8756 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,200B, BPFP=1.5625 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,832B, BPFP=1.7945 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,856B, BPFP=0.5212 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14984696 52.28076172 + layer.0.v_cache 0.00001474 0.00759146 + layer.1.k_cache 0.67859346 4.26252629 + layer.1.v_cache 0.00000581 0.00319335 + layer.2.k_cache 0.01626646 0.74306825 + layer.2.v_cache 0.00001951 0.00853298 + layer.3.k_cache 0.02914289 2.98672114 + layer.3.v_cache 0.00001919 0.00925266 + layer.4.k_cache 0.00071665 0.17613371 + layer.4.v_cache 0.00005096 0.01876283 + layer.4.output 0.04642455 172.48013965 + ------------------------------------------------------------------------------------- + TOTAL 0.07056756 74.57985423 + (elements=2,715,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2715648 +Total Bytes 359672 +BPFP 1.0596 bits/point +EBPFP 2.1191 equivalent bits/point +MSE 74.579854 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 74.5799 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,464B, BPFP=0.4916 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,752B, BPFP=2.0186 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,404B, BPFP=0.8947 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,940B, BPFP=1.9714 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,444B, BPFP=1.1294 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,144B, BPFP=1.9252 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,224B, BPFP=1.0586 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,572B, BPFP=1.9500 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,100B, BPFP=1.6322 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,432B, BPFP=1.8838 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,812B, BPFP=0.5295 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15823836 52.35476446 + layer.0.v_cache 0.00001502 0.00736265 + layer.1.k_cache 0.53668020 4.47226690 + layer.1.v_cache 0.00000576 0.00299133 + layer.2.k_cache 0.01018995 0.72962526 + layer.2.v_cache 0.00002218 0.00822478 + layer.3.k_cache 0.05410197 3.19337889 + layer.3.v_cache 0.00002001 0.00924478 + layer.4.k_cache 0.00069143 0.17291572 + layer.4.v_cache 0.00005304 0.01875733 + layer.4.output 0.01041225 204.22752921 + ------------------------------------------------------------------------------------- + TOTAL 0.04899433 87.68013156 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 321288 +BPFP 1.0978 bits/point +EBPFP 2.1955 equivalent bits/point +MSE 87.680132 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 87.6801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 329, 128) +Output shape: (1, 329, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.output: torch.Size([1, 329, 3584]) -> torch.Size([1, 1, 329, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,320B, BPFP=0.4901 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,176B, BPFP=2.0030 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,832B, BPFP=0.8944 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,068B, BPFP=1.9504 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,336B, BPFP=1.1083 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,108B, BPFP=1.9048 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,084B, BPFP=1.0488 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,696B, BPFP=1.9328 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,716B, BPFP=1.6013 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,248B, BPFP=1.8640 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,460B, BPFP=0.5866 +⌛️ [2/4] FRONTEND: Frontend time: 0.341s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.445s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11391156 52.99461555 + layer.0.v_cache 0.00001522 0.00744313 + layer.1.k_cache 0.63933222 4.50645934 + layer.1.v_cache 0.00000652 0.00322567 + layer.2.k_cache 0.01643764 0.70351227 + layer.2.v_cache 0.00001873 0.00820313 + layer.3.k_cache 0.02826017 2.95096298 + layer.3.v_cache 0.00001876 0.00910211 + layer.4.k_cache 0.00074607 0.17667532 + layer.4.v_cache 0.00005483 0.01893224 + layer.4.output 3.44632065 161.95737896 + ------------------------------------------------------------------------------------- + TOTAL 1.46606155 70.29886967 + (elements=2,863,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2863616 +Total Bytes 398044 +BPFP 1.1120 bits/point +EBPFP 2.2240 equivalent bits/point +MSE 70.298870 +---------------------- -------------------------------------------------------- +Time: 0.796s Load: 0.011s, Pack+Encode: 0.341s, Decode+Unpack: 0.445s +---------------------- -------------------------------------------------------- +💾 Converting with 70.2989 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,820B, BPFP=0.4940 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,716B, BPFP=2.0002 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,700B, BPFP=0.8793 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,740B, BPFP=1.9456 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,108B, BPFP=1.1261 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,172B, BPFP=1.9138 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,048B, BPFP=1.0668 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,596B, BPFP=1.9375 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,032B, BPFP=1.6259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,148B, BPFP=1.8564 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,896B, BPFP=0.5032 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13536828 52.61783854 + layer.0.v_cache 0.00001433 0.00752814 + layer.1.k_cache 0.57185238 4.44050057 + layer.1.v_cache 0.00000558 0.00316695 + layer.2.k_cache 0.00766427 0.73581026 + layer.2.v_cache 0.00001891 0.00866516 + layer.3.k_cache 0.04143807 3.21900595 + layer.3.v_cache 0.00001947 0.00944934 + layer.4.k_cache 0.00070404 0.17467830 + layer.4.v_cache 0.00005498 0.01879993 + layer.4.output 0.00948056 196.75968062 + ------------------------------------------------------------------------------------- + TOTAL 0.04844142 84.62077691 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 327976 +BPFP 1.0805 bits/point +EBPFP 2.1609 equivalent bits/point +MSE 84.620777 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 84.6208 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,880B, BPFP=0.5027 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,360B, BPFP=2.0018 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,668B, BPFP=0.8870 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,324B, BPFP=1.9432 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,760B, BPFP=1.1187 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,820B, BPFP=1.9146 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,628B, BPFP=1.0546 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,140B, BPFP=1.9327 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,520B, BPFP=1.6146 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,888B, BPFP=1.8619 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,352B, BPFP=0.5285 +⌛️ [2/4] FRONTEND: Frontend time: 0.272s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.400s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15132187 49.33136606 + layer.0.v_cache 0.00001456 0.00728737 + layer.1.k_cache 0.61798344 4.33622454 + layer.1.v_cache 0.00000569 0.00309852 + layer.2.k_cache 0.01686249 0.72737470 + layer.2.v_cache 0.00001898 0.00851229 + layer.3.k_cache 0.04401700 3.21182737 + layer.3.v_cache 0.00001935 0.00922869 + layer.4.k_cache 0.00071189 0.17476242 + layer.4.v_cache 0.00005111 0.01865944 + layer.4.output 0.01058244 200.03642598 + ------------------------------------------------------------------------------------- + TOTAL 0.05324020 85.76960725 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 327340 +BPFP 1.0901 bits/point +EBPFP 2.1802 equivalent bits/point +MSE 85.769607 +---------------------- -------------------------------------------------------- +Time: 0.681s Load: 0.010s, Pack+Encode: 0.272s, Decode+Unpack: 0.400s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7696 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,832B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,544B, BPFP=2.0122 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,624B, BPFP=0.8845 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,544B, BPFP=1.9556 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,900B, BPFP=1.1266 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,884B, BPFP=1.9183 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,644B, BPFP=1.0555 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,392B, BPFP=1.9470 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,592B, BPFP=1.6187 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,972B, BPFP=1.8666 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,268B, BPFP=0.5359 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12581717 51.10462664 + layer.0.v_cache 0.00001375 0.00776843 + layer.1.k_cache 0.57814745 4.75931096 + layer.1.v_cache 0.00000592 0.00333363 + layer.2.k_cache 0.00733882 0.73984196 + layer.2.v_cache 0.00002058 0.00911468 + layer.3.k_cache 0.02353939 2.95286096 + layer.3.v_cache 0.00002077 0.00985925 + layer.4.k_cache 0.00069182 0.18022074 + layer.4.v_cache 0.00005027 0.01941504 + layer.4.output 0.01015823 200.03097503 + ------------------------------------------------------------------------------------- + TOTAL 0.04745609 85.88253985 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 329196 +BPFP 1.0963 bits/point +EBPFP 2.1925 equivalent bits/point +MSE 85.882540 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.009s, Pack+Encode: 0.252s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8825 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,796B, BPFP=0.5016 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,256B, BPFP=2.0105 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,664B, BPFP=0.8932 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,424B, BPFP=1.9630 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,808B, BPFP=1.1296 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,728B, BPFP=1.9234 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,752B, BPFP=1.0693 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,252B, BPFP=1.9532 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,476B, BPFP=1.6239 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,992B, BPFP=1.8814 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,112B, BPFP=0.5467 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13625057 51.90030152 + layer.0.v_cache 0.00001362 0.00718411 + layer.1.k_cache 0.57512581 4.38642026 + layer.1.v_cache 0.00000584 0.00305610 + layer.2.k_cache 0.00872732 0.74915976 + layer.2.v_cache 0.00001851 0.00861208 + layer.3.k_cache 0.04362779 3.03146541 + layer.3.v_cache 0.00001925 0.00955219 + layer.4.k_cache 0.00073619 0.17422638 + layer.4.v_cache 0.00005148 0.01841283 + layer.4.output 0.00783937 201.28810936 + ------------------------------------------------------------------------------------- + TOTAL 0.04820306 86.42971507 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 329260 +BPFP 1.1045 bits/point +EBPFP 2.2090 equivalent bits/point +MSE 86.429715 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.009s, Pack+Encode: 0.250s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 86.4297 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,424B, BPFP=0.4844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,272B, BPFP=1.9671 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,840B, BPFP=0.8655 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,208B, BPFP=1.9124 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,160B, BPFP=1.0876 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,508B, BPFP=1.8764 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,080B, BPFP=1.0321 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,092B, BPFP=1.9065 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,632B, BPFP=1.5744 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,516B, BPFP=1.8255 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,384B, BPFP=0.5462 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16730670 51.00349828 + layer.0.v_cache 0.00001388 0.00719276 + layer.1.k_cache 0.64158324 4.24260150 + layer.1.v_cache 0.00000571 0.00295959 + layer.2.k_cache 0.02207795 0.73600965 + layer.2.v_cache 0.00001922 0.00798620 + layer.3.k_cache 0.04503629 2.99001995 + layer.3.v_cache 0.00001955 0.00911961 + layer.4.k_cache 0.00070478 0.16460662 + layer.4.v_cache 0.00005260 0.01773002 + layer.4.output 0.04763387 176.44895442 + ------------------------------------------------------------------------------------- + TOTAL 0.07119159 76.13672971 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 357116 +BPFP 1.0797 bits/point +EBPFP 2.1594 equivalent bits/point +MSE 76.136730 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.010s, Pack+Encode: 0.262s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 76.1367 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,848B, BPFP=0.4991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,416B, BPFP=1.9977 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,568B, BPFP=0.8782 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,340B, BPFP=1.9370 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,516B, BPFP=1.1009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,664B, BPFP=1.8989 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,684B, BPFP=1.0539 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,116B, BPFP=1.9244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,248B, BPFP=1.5934 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,752B, BPFP=1.8475 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,664B, BPFP=0.5372 +⌛️ [2/4] FRONTEND: Frontend time: 0.263s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15225300 52.11520617 + layer.0.v_cache 0.00001427 0.00694791 + layer.1.k_cache 0.52731114 4.43551118 + layer.1.v_cache 0.00000556 0.00301426 + layer.2.k_cache 0.01483519 0.69957433 + layer.2.v_cache 0.00001864 0.00830821 + layer.3.k_cache 0.04699480 2.86703326 + layer.3.v_cache 0.00001968 0.00882347 + layer.4.k_cache 0.00070142 0.17213889 + layer.4.v_cache 0.00005177 0.01836002 + layer.4.output 0.01077793 198.49329551 + ------------------------------------------------------------------------------------- + TOTAL 0.04809712 85.28164625 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 327816 +BPFP 1.0877 bits/point +EBPFP 2.1755 equivalent bits/point +MSE 85.281646 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.012s, Pack+Encode: 0.263s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2816 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,800B, BPFP=0.4964 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,608B, BPFP=2.0086 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,652B, BPFP=0.8829 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,692B, BPFP=1.9569 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,868B, BPFP=1.1207 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,004B, BPFP=1.9181 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,896B, BPFP=1.0659 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,432B, BPFP=1.9422 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,532B, BPFP=1.6094 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,972B, BPFP=1.8599 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,060B, BPFP=0.5484 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15700974 51.85540177 + layer.0.v_cache 0.00001374 0.00691708 + layer.1.k_cache 0.52255304 4.23388540 + layer.1.v_cache 0.00000566 0.00291952 + layer.2.k_cache 0.01360668 0.76517594 + layer.2.v_cache 0.00001926 0.00838806 + layer.3.k_cache 0.08085050 3.03267783 + layer.3.v_cache 0.00001943 0.00882240 + layer.4.k_cache 0.00071029 0.16895294 + layer.4.v_cache 0.00005450 0.01801431 + layer.4.output 0.01014708 198.89698298 + ------------------------------------------------------------------------------------- + TOTAL 0.04975720 85.43411977 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 331516 +BPFP 1.1000 bits/point +EBPFP 2.2000 equivalent bits/point +MSE 85.434120 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.009s, Pack+Encode: 0.254s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 85.4341 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,992B, BPFP=0.4930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,212B, BPFP=1.9853 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,396B, BPFP=0.8989 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,340B, BPFP=1.9375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,336B, BPFP=1.1149 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,604B, BPFP=1.8971 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,992B, BPFP=1.0412 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,132B, BPFP=1.9261 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,212B, BPFP=1.6015 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,628B, BPFP=1.8436 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,444B, BPFP=0.5282 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16436556 52.34468887 + layer.0.v_cache 0.00001441 0.00734948 + layer.1.k_cache 0.62741763 4.54605948 + layer.1.v_cache 0.00000560 0.00311494 + layer.2.k_cache 0.01487184 0.76860170 + layer.2.v_cache 0.00002167 0.00848614 + layer.3.k_cache 0.02142572 2.93081804 + layer.3.v_cache 0.00001943 0.00921979 + layer.4.k_cache 0.00070776 0.17541592 + layer.4.v_cache 0.00005251 0.01818629 + layer.4.output 0.00894435 193.09260652 + ------------------------------------------------------------------------------------- + TOTAL 0.05244192 83.08589331 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 336288 +BPFP 1.0845 bits/point +EBPFP 2.1690 equivalent bits/point +MSE 83.085893 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 83.0859 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,980B, BPFP=0.4993 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,728B, BPFP=1.9867 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,940B, BPFP=0.8863 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,664B, BPFP=1.9275 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,856B, BPFP=1.1041 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,004B, BPFP=1.8908 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,716B, BPFP=1.0407 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,424B, BPFP=1.9141 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,856B, BPFP=1.6045 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,152B, BPFP=1.8434 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,252B, BPFP=0.5183 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11705062 51.71533724 + layer.0.v_cache 0.00001445 0.00747355 + layer.1.k_cache 0.56620099 4.42694461 + layer.1.v_cache 0.00000566 0.00304160 + layer.2.k_cache 0.01299965 0.72995060 + layer.2.v_cache 0.00001972 0.00808340 + layer.3.k_cache 0.05069733 2.91029657 + layer.3.v_cache 0.00001944 0.00922699 + layer.4.k_cache 0.00068883 0.16956889 + layer.4.v_cache 0.00005144 0.01768322 + layer.4.output 0.00838925 196.02014489 + ------------------------------------------------------------------------------------- + TOTAL 0.04743958 84.24344829 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 329572 +BPFP 1.0780 bits/point +EBPFP 2.1560 equivalent bits/point +MSE 84.243448 +---------------------- -------------------------------------------------------- +Time: 0.646s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 84.2434 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,000B, BPFP=0.4969 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,016B, BPFP=1.9885 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,024B, BPFP=0.8847 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,956B, BPFP=1.9300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,008B, BPFP=1.1047 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,224B, BPFP=1.8896 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,948B, BPFP=1.0462 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,736B, BPFP=1.9178 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,020B, BPFP=1.6023 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,476B, BPFP=1.8483 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,640B, BPFP=0.5256 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14370025 52.70411882 + layer.0.v_cache 0.00001382 0.00755413 + layer.1.k_cache 0.61448195 4.39962887 + layer.1.v_cache 0.00000570 0.00315018 + layer.2.k_cache 0.01208573 0.71210116 + layer.2.v_cache 0.00001823 0.00870326 + layer.3.k_cache 0.01319265 2.97589392 + layer.3.v_cache 0.00002242 0.00949192 + layer.4.k_cache 0.00071920 0.17527106 + layer.4.v_cache 0.00005145 0.01929399 + layer.4.output 0.00992508 195.13039500 + ------------------------------------------------------------------------------------- + TOTAL 0.05022159 83.93693955 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 333048 +BPFP 1.0817 bits/point +EBPFP 2.1633 equivalent bits/point +MSE 83.936940 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9369 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 259, 128) +Output shape: (1, 259, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.output: torch.Size([1, 259, 3584]) -> torch.Size([1, 1, 259, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,120B, BPFP=0.4899 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 33,700B, BPFP=2.0331 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,076B, BPFP=0.9095 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 32,720B, BPFP=1.9739 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,784B, BPFP=1.1332 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 31,920B, BPFP=1.9257 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,592B, BPFP=1.0613 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 32,400B, BPFP=1.9546 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,080B, BPFP=1.6337 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,164B, BPFP=1.8801 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,868B, BPFP=0.5418 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13349269 52.75818201 + layer.0.v_cache 0.00001436 0.00769255 + layer.1.k_cache 0.50562990 4.72878899 + layer.1.v_cache 0.00000590 0.00320791 + layer.2.k_cache 0.00820730 0.74621570 + layer.2.v_cache 0.00001794 0.00819252 + layer.3.k_cache 0.03685973 2.99932060 + layer.3.v_cache 0.00002061 0.00947507 + layer.4.k_cache 0.00069539 0.18249205 + layer.4.v_cache 0.00005422 0.01880576 + layer.4.output 0.01018517 213.39661473 + ------------------------------------------------------------------------------------- + TOTAL 0.04448790 91.48462801 + (elements=2,254,336) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2254336 +Total Bytes 311424 +BPFP 1.1052 bits/point +EBPFP 2.2103 equivalent bits/point +MSE 91.484628 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4846 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,544B, BPFP=0.4944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,004B, BPFP=2.0257 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,392B, BPFP=0.8907 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,908B, BPFP=1.9623 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,352B, BPFP=1.1199 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,280B, BPFP=1.9259 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,372B, BPFP=1.0632 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,700B, BPFP=1.9502 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,132B, BPFP=1.6280 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,508B, BPFP=1.8813 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,388B, BPFP=0.5323 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12649573 52.38973163 + layer.0.v_cache 0.00001373 0.00755752 + layer.1.k_cache 0.51574391 4.61804335 + layer.1.v_cache 0.00000589 0.00324315 + layer.2.k_cache 0.01194659 0.75549271 + layer.2.v_cache 0.00001984 0.00859061 + layer.3.k_cache 0.04962628 3.18598768 + layer.3.v_cache 0.00002059 0.00958194 + layer.4.k_cache 0.00068541 0.17753396 + layer.4.v_cache 0.00005704 0.01971902 + layer.4.output 0.00772497 202.76005291 + ------------------------------------------------------------------------------------- + TOTAL 0.04462881 87.08799129 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 322580 +BPFP 1.0981 bits/point +EBPFP 2.1962 equivalent bits/point +MSE 87.087991 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 87.0880 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,568B, BPFP=0.4977 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,712B, BPFP=2.0163 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,512B, BPFP=0.9010 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,704B, BPFP=1.9577 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,524B, BPFP=1.1341 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,156B, BPFP=1.9259 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,376B, BPFP=1.0674 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,680B, BPFP=1.9563 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,076B, BPFP=1.6308 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,392B, BPFP=1.8815 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,400B, BPFP=0.5095 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12210708 52.24388287 + layer.0.v_cache 0.00001342 0.00727042 + layer.1.k_cache 0.54942129 4.38400870 + layer.1.v_cache 0.00000547 0.00306815 + layer.2.k_cache 0.01053270 0.72777719 + layer.2.v_cache 0.00001820 0.00851093 + layer.3.k_cache 0.02850239 3.32506825 + layer.3.v_cache 0.00002491 0.00939578 + layer.4.k_cache 0.00067987 0.17756807 + layer.4.v_cache 0.00005179 0.01896226 + layer.4.output 0.01126663 204.25049788 + ------------------------------------------------------------------------------------- + TOTAL 0.04648373 87.68582340 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 319100 +BPFP 1.0903 bits/point +EBPFP 2.1806 equivalent bits/point +MSE 87.685823 +---------------------- -------------------------------------------------------- +Time: 0.649s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 87.6858 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,384B, BPFP=0.4839 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,216B, BPFP=1.9707 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,744B, BPFP=0.8634 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,132B, BPFP=1.9148 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,052B, BPFP=1.0856 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,292B, BPFP=1.8715 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,944B, BPFP=1.0285 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,756B, BPFP=1.8954 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,756B, BPFP=1.5860 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,252B, BPFP=1.8179 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,136B, BPFP=0.5019 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.393s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14443027 51.50106036 + layer.0.v_cache 0.00001375 0.00748895 + layer.1.k_cache 0.62837083 3.89424239 + layer.1.v_cache 0.00000570 0.00315367 + layer.2.k_cache 0.01557094 0.72147050 + layer.2.v_cache 0.00001871 0.00830099 + layer.3.k_cache 0.04676386 2.96008663 + layer.3.v_cache 0.00001981 0.00923184 + layer.4.k_cache 0.00070582 0.17493582 + layer.4.v_cache 0.00005437 0.01813814 + layer.4.output 0.04838977 177.83552864 + ------------------------------------------------------------------------------------- + TOTAL 0.06909897 76.71451822 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 349664 +BPFP 1.0607 bits/point +EBPFP 2.1213 equivalent bits/point +MSE 76.714518 +---------------------- -------------------------------------------------------- +Time: 0.662s Load: 0.010s, Pack+Encode: 0.259s, Decode+Unpack: 0.393s +---------------------- -------------------------------------------------------- +💾 Converting with 76.7145 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,968B, BPFP=0.4951 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,116B, BPFP=1.9940 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,964B, BPFP=0.8814 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,092B, BPFP=1.9375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,172B, BPFP=1.1137 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,412B, BPFP=1.9000 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,192B, BPFP=1.0596 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,836B, BPFP=1.9234 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,088B, BPFP=1.6060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,512B, BPFP=1.8503 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,664B, BPFP=0.5179 +⌛️ [2/4] FRONTEND: Frontend time: 0.264s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13293209 53.06190647 + layer.0.v_cache 0.00001388 0.00740673 + layer.1.k_cache 0.56267701 4.45020351 + layer.1.v_cache 0.00000566 0.00318948 + layer.2.k_cache 0.01145046 0.73949106 + layer.2.v_cache 0.00001921 0.00867316 + layer.3.k_cache 0.02674564 3.15685086 + layer.3.v_cache 0.00001926 0.00944686 + layer.4.k_cache 0.00069135 0.18194939 + layer.4.v_cache 0.00005493 0.01986713 + layer.4.output 0.00973303 194.78456272 + ------------------------------------------------------------------------------------- + TOTAL 0.04722004 83.83123081 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 333016 +BPFP 1.0816 bits/point +EBPFP 2.1631 equivalent bits/point +MSE 83.831231 +---------------------- -------------------------------------------------------- +Time: 0.661s Load: 0.009s, Pack+Encode: 0.264s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8312 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,896B, BPFP=0.5018 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,568B, BPFP=2.0063 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,776B, BPFP=0.8899 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,708B, BPFP=1.9578 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,864B, BPFP=1.1205 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,848B, BPFP=1.9093 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,644B, BPFP=1.0517 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,428B, BPFP=1.9420 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,792B, BPFP=1.6241 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,064B, BPFP=1.8651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,732B, BPFP=0.5216 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.393s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15182159 52.02930040 + layer.0.v_cache 0.00001420 0.00764492 + layer.1.k_cache 0.56224567 4.53510249 + layer.1.v_cache 0.00000588 0.00329231 + layer.2.k_cache 0.00652889 0.70082952 + layer.2.v_cache 0.00001844 0.00872787 + layer.3.k_cache 0.06011557 2.91414667 + layer.3.v_cache 0.00001921 0.00964367 + layer.4.k_cache 0.00068317 0.17830458 + layer.4.v_cache 0.00005096 0.01913593 + layer.4.output 0.01130176 198.76895307 + ------------------------------------------------------------------------------------- + TOTAL 0.05062446 85.39934117 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 328320 +BPFP 1.0894 bits/point +EBPFP 2.1788 equivalent bits/point +MSE 85.399341 +---------------------- -------------------------------------------------------- +Time: 0.664s Load: 0.010s, Pack+Encode: 0.261s, Decode+Unpack: 0.393s +---------------------- -------------------------------------------------------- +💾 Converting with 85.3993 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,016B, BPFP=0.4943 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,528B, BPFP=2.0026 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,580B, BPFP=0.9090 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,608B, BPFP=1.9522 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,600B, BPFP=1.1294 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,856B, BPFP=1.9110 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,436B, BPFP=1.0656 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,332B, BPFP=1.9371 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,680B, BPFP=1.6272 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,064B, BPFP=1.8675 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,684B, BPFP=0.5536 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.389s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13053865 52.75844298 + layer.0.v_cache 0.00001414 0.00798086 + layer.1.k_cache 0.56363723 4.53050816 + layer.1.v_cache 0.00000590 0.00331064 + layer.2.k_cache 0.01033439 0.77339702 + layer.2.v_cache 0.00002015 0.00881473 + layer.3.k_cache 0.03655195 2.87351824 + layer.3.v_cache 0.00002026 0.00956182 + layer.4.k_cache 0.00068516 0.18042900 + layer.4.v_cache 0.00005378 0.01943007 + layer.4.output 0.00863618 193.55307018 + ------------------------------------------------------------------------------------- + TOTAL 0.04719499 83.29628734 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 342384 +BPFP 1.1042 bits/point +EBPFP 2.2084 equivalent bits/point +MSE 83.296287 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.009s, Pack+Encode: 0.256s, Decode+Unpack: 0.389s +---------------------- -------------------------------------------------------- +💾 Converting with 83.2963 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,812B, BPFP=0.4971 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,568B, BPFP=2.0063 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,648B, BPFP=0.8827 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,540B, BPFP=1.9483 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,672B, BPFP=1.1097 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,740B, BPFP=1.9032 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,712B, BPFP=1.0555 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,400B, BPFP=1.9404 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,536B, BPFP=1.6097 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,964B, BPFP=1.8594 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,720B, BPFP=0.5457 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.393s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15908699 52.02013764 + layer.0.v_cache 0.00001433 0.00740280 + layer.1.k_cache 0.53822481 4.65969397 + layer.1.v_cache 0.00000570 0.00313062 + layer.2.k_cache 0.01139215 0.71244168 + layer.2.v_cache 0.00001845 0.00844104 + layer.3.k_cache 0.02907990 2.75764328 + layer.3.v_cache 0.00001963 0.00927402 + layer.4.k_cache 0.00069750 0.17495305 + layer.4.v_cache 0.00005397 0.01868405 + layer.4.output 0.01069233 198.38931150 + ------------------------------------------------------------------------------------- + TOTAL 0.04784940 85.24099898 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 330312 +BPFP 1.0960 bits/point +EBPFP 2.1920 equivalent bits/point +MSE 85.240999 +---------------------- -------------------------------------------------------- +Time: 0.659s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.393s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2410 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,832B, BPFP=0.4825 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,528B, BPFP=1.9956 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,256B, BPFP=0.8881 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,532B, BPFP=1.9412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,392B, BPFP=1.1141 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,772B, BPFP=1.8997 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,328B, BPFP=1.0559 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,248B, BPFP=1.9257 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,456B, BPFP=1.6093 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,812B, BPFP=1.8472 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,528B, BPFP=0.5426 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16084659 53.23602082 + layer.0.v_cache 0.00001440 0.00736826 + layer.1.k_cache 0.59028764 4.45495008 + layer.1.v_cache 0.00000559 0.00305396 + layer.2.k_cache 0.00929020 0.75329440 + layer.2.v_cache 0.00002205 0.00831731 + layer.3.k_cache 0.04382113 2.95189192 + layer.3.v_cache 0.00002103 0.00918480 + layer.4.k_cache 0.00069271 0.17355689 + layer.4.v_cache 0.00005052 0.01816744 + layer.4.output 0.00984751 192.65269106 + ------------------------------------------------------------------------------------- + TOTAL 0.05141085 82.95203784 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 339684 +BPFP 1.0916 bits/point +EBPFP 2.1833 equivalent bits/point +MSE 82.952038 +---------------------- -------------------------------------------------------- +Time: 0.657s Load: 0.012s, Pack+Encode: 0.258s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 82.9520 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,696B, BPFP=0.4718 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,492B, BPFP=1.9798 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,936B, BPFP=0.8646 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,500B, BPFP=1.9260 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,088B, BPFP=1.0898 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,700B, BPFP=1.8826 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,064B, BPFP=1.0343 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,272B, BPFP=1.9136 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,392B, BPFP=1.5946 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,968B, BPFP=1.8429 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,064B, BPFP=0.5585 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.395s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14211286 52.66151259 + layer.0.v_cache 0.00001516 0.00733293 + layer.1.k_cache 0.60291523 4.39320034 + layer.1.v_cache 0.00000570 0.00309678 + layer.2.k_cache 0.01041983 0.73696645 + layer.2.v_cache 0.00001835 0.00795045 + layer.3.k_cache 0.03896446 2.93572913 + layer.3.v_cache 0.00001926 0.00919192 + layer.4.k_cache 0.00071470 0.17413172 + layer.4.v_cache 0.00005236 0.01849328 + layer.4.output 0.00754456 191.50280568 + ------------------------------------------------------------------------------------- + TOTAL 0.04988528 82.43924973 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 341172 +BPFP 1.0888 bits/point +EBPFP 2.1776 equivalent bits/point +MSE 82.439250 +---------------------- -------------------------------------------------------- +Time: 0.664s Load: 0.011s, Pack+Encode: 0.258s, Decode+Unpack: 0.395s +---------------------- -------------------------------------------------------- +💾 Converting with 82.4392 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 296, 128) +Output shape: (1, 296, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.output: torch.Size([1, 296, 3584]) -> torch.Size([1, 1, 296, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,184B, BPFP=0.4848 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,664B, BPFP=1.9882 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,696B, BPFP=0.8813 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,588B, BPFP=1.9314 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,852B, BPFP=1.1007 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,908B, BPFP=1.8955 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,744B, BPFP=1.0422 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,520B, BPFP=1.9278 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,448B, BPFP=1.6073 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,992B, BPFP=1.8471 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,656B, BPFP=0.5253 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.398s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14616069 53.92043985 + layer.0.v_cache 0.00001366 0.00735924 + layer.1.k_cache 0.61022733 4.50671016 + layer.1.v_cache 0.00000576 0.00301101 + layer.2.k_cache 0.01378242 0.73684847 + layer.2.v_cache 0.00001907 0.00831933 + layer.3.k_cache 0.03184220 2.90741276 + layer.3.v_cache 0.00001934 0.00907120 + layer.4.k_cache 0.00072724 0.17198614 + layer.4.v_cache 0.00005080 0.01785391 + layer.4.output 0.04784788 182.15864805 + ------------------------------------------------------------------------------------- + TOTAL 0.06692845 78.67056167 + (elements=2,576,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2576384 +Total Bytes 348252 +BPFP 1.0814 bits/point +EBPFP 2.1627 equivalent bits/point +MSE 78.670562 +---------------------- -------------------------------------------------------- +Time: 0.670s Load: 0.012s, Pack+Encode: 0.260s, Decode+Unpack: 0.398s +---------------------- -------------------------------------------------------- +💾 Converting with 78.6706 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,196B, BPFP=0.4822 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,988B, BPFP=1.9918 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,760B, BPFP=0.8788 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,980B, BPFP=1.9390 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,188B, BPFP=1.1109 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,204B, BPFP=1.8983 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,136B, BPFP=1.0558 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,824B, BPFP=1.9308 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,508B, BPFP=1.5996 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,288B, BPFP=1.8503 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,212B, BPFP=0.5409 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15387250 53.03557571 + layer.0.v_cache 0.00001382 0.00736122 + layer.1.k_cache 0.63844893 4.47757224 + layer.1.v_cache 0.00000583 0.00315354 + layer.2.k_cache 0.01105453 0.68588748 + layer.2.v_cache 0.00001885 0.00850748 + layer.3.k_cache 0.03893430 2.80216929 + layer.3.v_cache 0.00001923 0.00954982 + layer.4.k_cache 0.00070257 0.17152868 + layer.4.v_cache 0.00005146 0.01855032 + layer.4.output 0.04873301 180.76817174 + ------------------------------------------------------------------------------------- + TOTAL 0.06966195 78.03512106 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 353284 +BPFP 1.0896 bits/point +EBPFP 2.1793 equivalent bits/point +MSE 78.035121 +---------------------- -------------------------------------------------------- +Time: 0.656s Load: 0.010s, Pack+Encode: 0.262s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 78.0351 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,704B, BPFP=0.5018 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,196B, BPFP=2.0293 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,676B, BPFP=0.9038 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,328B, BPFP=1.9792 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,644B, BPFP=1.1326 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,360B, BPFP=1.9234 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,668B, BPFP=1.0763 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,936B, BPFP=1.9566 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,288B, BPFP=1.6310 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,584B, BPFP=1.8787 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,024B, BPFP=0.5603 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12629252 52.37479099 + layer.0.v_cache 0.00001451 0.00747891 + layer.1.k_cache 0.49097243 4.44772958 + layer.1.v_cache 0.00000577 0.00320019 + layer.2.k_cache 0.01060468 0.74440171 + layer.2.v_cache 0.00001909 0.00829898 + layer.3.k_cache 0.02748298 3.08742194 + layer.3.v_cache 0.00001876 0.00931181 + layer.4.k_cache 0.00069967 0.17679680 + layer.4.v_cache 0.00005414 0.01896451 + layer.4.output 0.01047103 202.43469953 + ------------------------------------------------------------------------------------- + TOTAL 0.04290952 86.93654659 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 328408 +BPFP 1.1138 bits/point +EBPFP 2.2276 equivalent bits/point +MSE 86.936547 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.010s, Pack+Encode: 0.256s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9365 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,772B, BPFP=0.4913 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,676B, BPFP=1.9980 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,796B, BPFP=0.8846 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,692B, BPFP=1.9429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,820B, BPFP=1.1100 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,856B, BPFP=1.8961 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,696B, BPFP=1.0470 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,428B, BPFP=1.9281 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,828B, BPFP=1.6145 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,132B, BPFP=1.8555 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,292B, BPFP=0.5384 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15022702 52.66681718 + layer.0.v_cache 0.00001395 0.00717063 + layer.1.k_cache 0.54271635 4.22465225 + layer.1.v_cache 0.00000573 0.00307650 + layer.2.k_cache 0.01332666 0.69058807 + layer.2.v_cache 0.00001832 0.00833758 + layer.3.k_cache 0.02579262 2.92855496 + layer.3.v_cache 0.00001838 0.00882208 + layer.4.k_cache 0.00069325 0.17160771 + layer.4.v_cache 0.00005155 0.01831538 + layer.4.output 0.01006869 195.21122952 + ------------------------------------------------------------------------------------- + TOTAL 0.04725557 83.95332641 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 330988 +BPFP 1.0904 bits/point +EBPFP 2.1808 equivalent bits/point +MSE 83.953326 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9533 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,168B, BPFP=0.4856 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,652B, BPFP=1.9943 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,752B, BPFP=0.8873 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,744B, BPFP=1.9462 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,772B, BPFP=1.1002 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,940B, BPFP=1.9036 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,704B, BPFP=1.0436 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,464B, BPFP=1.9314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,304B, BPFP=1.6051 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,076B, BPFP=1.8578 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,796B, BPFP=0.5357 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12416320 52.37143141 + layer.0.v_cache 0.00001370 0.00718106 + layer.1.k_cache 0.62725913 4.23505156 + layer.1.v_cache 0.00000565 0.00301639 + layer.2.k_cache 0.01102555 0.69102245 + layer.2.v_cache 0.00002033 0.00832641 + layer.3.k_cache 0.04118381 3.05306045 + layer.3.v_cache 0.00001949 0.00925387 + layer.4.k_cache 0.00070768 0.17555402 + layer.4.v_cache 0.00005397 0.01852139 + layer.4.output 0.05105524 183.05845944 + ------------------------------------------------------------------------------------- + TOTAL 0.06834348 78.94009618 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 349372 +BPFP 1.0885 bits/point +EBPFP 2.1770 equivalent bits/point +MSE 78.940096 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.011s, Pack+Encode: 0.253s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 78.9401 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,656B, BPFP=0.4680 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,556B, BPFP=1.9764 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,972B, BPFP=0.8635 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,692B, BPFP=1.9297 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,368B, BPFP=1.1012 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,828B, BPFP=1.8830 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,304B, BPFP=1.0437 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,316B, BPFP=1.9094 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,384B, BPFP=1.5887 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,776B, BPFP=1.8261 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,616B, BPFP=0.5222 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15516999 53.75356158 + layer.0.v_cache 0.00001361 0.00789227 + layer.1.k_cache 0.56212566 4.13348579 + layer.1.v_cache 0.00000565 0.00339501 + layer.2.k_cache 0.00839357 0.71716921 + layer.2.v_cache 0.00001980 0.00881678 + layer.3.k_cache 0.02402391 2.88358822 + layer.3.v_cache 0.00001987 0.00966571 + layer.4.k_cache 0.00071212 0.17876073 + layer.4.v_cache 0.00005188 0.01927192 + layer.4.output 0.05073151 186.91687778 + ------------------------------------------------------------------------------------- + TOTAL 0.06503863 80.59610304 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 337468 +BPFP 1.0733 bits/point +EBPFP 2.1465 equivalent bits/point +MSE 80.596103 +---------------------- -------------------------------------------------------- +Time: 0.642s Load: 0.010s, Pack+Encode: 0.251s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 80.5961 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,300B, BPFP=0.4893 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,896B, BPFP=1.9937 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,836B, BPFP=0.8857 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,808B, BPFP=1.9364 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,928B, BPFP=1.1010 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,176B, BPFP=1.9032 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,884B, BPFP=1.0461 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,724B, BPFP=1.9320 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,376B, BPFP=1.5981 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,152B, BPFP=1.8493 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,568B, BPFP=0.5379 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005571 54.13998711 + layer.0.v_cache 0.00001459 0.00729888 + layer.1.k_cache 0.64038677 4.45155987 + layer.1.v_cache 0.00000545 0.00294810 + layer.2.k_cache 0.01623841 0.74520401 + layer.2.v_cache 0.00002030 0.00847373 + layer.3.k_cache 0.01774169 2.96436616 + layer.3.v_cache 0.00001948 0.00911547 + layer.4.k_cache 0.00069419 0.17381071 + layer.4.v_cache 0.00005057 0.01842924 + layer.4.output 0.05052170 180.34047318 + ------------------------------------------------------------------------------------- + TOTAL 0.06875759 77.93555915 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 351648 +BPFP 1.0882 bits/point +EBPFP 2.1765 equivalent bits/point +MSE 77.935559 +---------------------- -------------------------------------------------------- +Time: 0.642s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 77.9356 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,928B, BPFP=0.5000 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,808B, BPFP=2.0054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,856B, BPFP=0.8880 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,808B, BPFP=1.9494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,996B, BPFP=1.1198 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,184B, BPFP=1.9144 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,908B, BPFP=1.0589 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,648B, BPFP=1.9404 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,848B, BPFP=1.6156 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,272B, BPFP=1.8634 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,076B, BPFP=0.5286 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14010120 52.28458921 + layer.0.v_cache 0.00001437 0.00758277 + layer.1.k_cache 0.58692467 4.51399040 + layer.1.v_cache 0.00000569 0.00324772 + layer.2.k_cache 0.00806839 0.71287236 + layer.2.v_cache 0.00001934 0.00864933 + layer.3.k_cache 0.03832107 2.77306779 + layer.3.v_cache 0.00001992 0.00958665 + layer.4.k_cache 0.00069804 0.17305414 + layer.4.v_cache 0.00005131 0.01931613 + layer.4.output 0.00759059 197.92292307 + ------------------------------------------------------------------------------------- + TOTAL 0.04866812 85.05684811 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 331332 +BPFP 1.0915 bits/point +EBPFP 2.1830 equivalent bits/point +MSE 85.056848 +---------------------- -------------------------------------------------------- +Time: 0.649s Load: 0.010s, Pack+Encode: 0.255s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 85.0568 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,604B, BPFP=0.5016 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,832B, BPFP=2.0308 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,572B, BPFP=0.9079 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,976B, BPFP=1.9809 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,520B, BPFP=1.1381 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,168B, BPFP=1.9338 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,448B, BPFP=1.0756 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,712B, BPFP=1.9655 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,164B, BPFP=1.6420 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,368B, BPFP=1.8871 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,640B, BPFP=0.5301 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13199043 52.42549703 + layer.0.v_cache 0.00001386 0.00735592 + layer.1.k_cache 0.56989129 4.70325800 + layer.1.v_cache 0.00000577 0.00310464 + layer.2.k_cache 0.00892691 0.69211499 + layer.2.v_cache 0.00001816 0.00842138 + layer.3.k_cache 0.04626510 2.86933466 + layer.3.v_cache 0.00001927 0.00916574 + layer.4.k_cache 0.00069392 0.17633037 + layer.4.v_cache 0.00005195 0.01871683 + layer.4.output 0.00911696 205.91529518 + ------------------------------------------------------------------------------------- + TOTAL 0.04833502 88.37178622 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 322004 +BPFP 1.1043 bits/point +EBPFP 2.2087 equivalent bits/point +MSE 88.371786 +---------------------- -------------------------------------------------------- +Time: 0.647s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 88.3718 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,736B, BPFP=0.4982 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,224B, BPFP=2.0087 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,756B, BPFP=0.8985 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,244B, BPFP=1.9528 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,652B, BPFP=1.1207 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,480B, BPFP=1.9092 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,696B, BPFP=1.0661 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,996B, BPFP=1.9386 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,404B, BPFP=1.6198 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,668B, BPFP=1.8629 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,128B, BPFP=0.5224 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16500157 50.58235202 + layer.0.v_cache 0.00001391 0.00747596 + layer.1.k_cache 0.57858722 4.45970967 + layer.1.v_cache 0.00000575 0.00311722 + layer.2.k_cache 0.01018238 0.74298207 + layer.2.v_cache 0.00001877 0.00879644 + layer.3.k_cache 0.03425542 3.09903065 + layer.3.v_cache 0.00002021 0.00957885 + layer.4.k_cache 0.00069807 0.17606983 + layer.4.v_cache 0.00005194 0.01869715 + layer.4.output 0.01023613 201.11933003 + ------------------------------------------------------------------------------------- + TOTAL 0.05061695 86.29077177 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 324984 +BPFP 1.0901 bits/point +EBPFP 2.1803 equivalent bits/point +MSE 86.290772 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.011s, Pack+Encode: 0.253s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2908 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,572B, BPFP=0.4961 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,916B, BPFP=2.0206 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,332B, BPFP=0.8873 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,784B, BPFP=1.9551 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,360B, BPFP=1.1204 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,148B, BPFP=1.9183 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,444B, BPFP=1.0674 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,700B, BPFP=1.9502 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,992B, BPFP=1.6199 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,328B, BPFP=1.8708 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,352B, BPFP=0.5403 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385483 51.58002749 + layer.0.v_cache 0.00001387 0.00724504 + layer.1.k_cache 0.55871571 4.31790726 + layer.1.v_cache 0.00000565 0.00296235 + layer.2.k_cache 0.00818202 0.68943402 + layer.2.v_cache 0.00001875 0.00836962 + layer.3.k_cache 0.04697755 3.23121383 + layer.3.v_cache 0.00001978 0.00913415 + layer.4.k_cache 0.00071636 0.17226540 + layer.4.v_cache 0.00005146 0.01823945 + layer.4.output 0.01106052 204.28392857 + ------------------------------------------------------------------------------------- + TOTAL 0.04741057 87.64848815 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 322928 +BPFP 1.0993 bits/point +EBPFP 2.1986 equivalent bits/point +MSE 87.648488 +---------------------- -------------------------------------------------------- +Time: 0.638s Load: 0.009s, Pack+Encode: 0.247s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 87.6485 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,768B, BPFP=0.4982 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,236B, BPFP=2.0020 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,608B, BPFP=0.8868 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,172B, BPFP=1.9416 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,644B, BPFP=1.1161 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,564B, BPFP=1.9070 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,548B, BPFP=1.0539 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,104B, BPFP=1.9377 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,432B, BPFP=1.6155 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,824B, BPFP=1.8650 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,616B, BPFP=0.5245 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203608 51.89210937 + layer.0.v_cache 0.00001420 0.00721710 + layer.1.k_cache 0.56673284 4.74749467 + layer.1.v_cache 0.00000552 0.00294864 + layer.2.k_cache 0.00530624 0.67579379 + layer.2.v_cache 0.00001825 0.00825930 + layer.3.k_cache 0.02188354 3.16063299 + layer.3.v_cache 0.00001854 0.00921378 + layer.4.k_cache 0.00069888 0.17236797 + layer.4.v_cache 0.00004918 0.01814183 + layer.4.output 0.00799176 200.78566558 + ------------------------------------------------------------------------------------- + TOTAL 0.04662974 86.24669638 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 325516 +BPFP 1.0880 bits/point +EBPFP 2.1759 equivalent bits/point +MSE 86.246696 +---------------------- -------------------------------------------------------- +Time: 0.642s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 86.2467 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,856B, BPFP=0.4890 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,980B, BPFP=1.9865 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,920B, BPFP=0.8790 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,936B, BPFP=1.9289 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,056B, BPFP=1.1073 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,220B, BPFP=1.8894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,884B, BPFP=1.0426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,712B, BPFP=1.9165 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,876B, BPFP=1.5943 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,380B, BPFP=1.8430 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,140B, BPFP=0.5374 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12732459 51.96376712 + layer.0.v_cache 0.00001480 0.00738972 + layer.1.k_cache 0.57322790 4.39296720 + layer.1.v_cache 0.00000557 0.00319455 + layer.2.k_cache 0.00950604 0.70221470 + layer.2.v_cache 0.00002039 0.00893594 + layer.3.k_cache 0.01510038 3.08406460 + layer.3.v_cache 0.00001916 0.00951883 + layer.4.k_cache 0.00068141 0.17418601 + layer.4.v_cache 0.00005165 0.01852167 + layer.4.output 0.00936729 194.86091305 + ------------------------------------------------------------------------------------- + TOTAL 0.04656017 83.78771480 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 333960 +BPFP 1.0846 bits/point +EBPFP 2.1692 equivalent bits/point +MSE 83.787715 +---------------------- -------------------------------------------------------- +Time: 0.642s Load: 0.009s, Pack+Encode: 0.250s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 83.7877 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,888B, BPFP=0.5032 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,468B, BPFP=2.0079 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,752B, BPFP=0.8918 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,548B, BPFP=1.9558 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,732B, BPFP=1.1171 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,828B, BPFP=1.9151 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,852B, BPFP=1.0673 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,364B, BPFP=1.9454 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,692B, BPFP=1.6243 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,120B, BPFP=1.8750 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,592B, BPFP=0.5386 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13630490 51.25420700 + layer.0.v_cache 0.00001544 0.00747711 + layer.1.k_cache 0.58776402 4.32285298 + layer.1.v_cache 0.00000559 0.00299997 + layer.2.k_cache 0.01000112 0.73520716 + layer.2.v_cache 0.00001940 0.00854478 + layer.3.k_cache 0.04424953 2.99719548 + layer.3.v_cache 0.00002139 0.00936309 + layer.4.k_cache 0.00071093 0.17393183 + layer.4.v_cache 0.00005158 0.01827476 + layer.4.output 0.01101471 200.02733566 + ------------------------------------------------------------------------------------- + TOTAL 0.05036746 85.86596493 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 329836 +BPFP 1.0984 bits/point +EBPFP 2.1968 equivalent bits/point +MSE 85.865965 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8660 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,876B, BPFP=0.4989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,640B, BPFP=2.0031 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,692B, BPFP=0.8820 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,708B, BPFP=1.9508 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,788B, BPFP=1.1122 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,976B, BPFP=1.9096 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,668B, BPFP=1.0492 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,212B, BPFP=1.9229 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,824B, BPFP=1.6201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,120B, BPFP=1.8615 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,788B, BPFP=0.5363 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12958164 51.95679589 + layer.0.v_cache 0.00001430 0.00760550 + layer.1.k_cache 0.54121377 4.20307319 + layer.1.v_cache 0.00000577 0.00314697 + layer.2.k_cache 0.01080061 0.72776937 + layer.2.v_cache 0.00001903 0.00867355 + layer.3.k_cache 0.01783211 2.99214535 + layer.3.v_cache 0.00002028 0.00934674 + layer.4.k_cache 0.00068935 0.17658890 + layer.4.v_cache 0.00005053 0.01883380 + layer.4.output 0.00833506 197.31380075 + ------------------------------------------------------------------------------------- + TOTAL 0.04462193 84.78238732 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 330292 +BPFP 1.0920 bits/point +EBPFP 2.1840 equivalent bits/point +MSE 84.782387 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 84.7824 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,276B, BPFP=0.4917 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,372B, BPFP=2.0421 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,216B, BPFP=0.9040 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,488B, BPFP=1.9895 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,300B, BPFP=1.1466 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,836B, BPFP=1.9508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,248B, BPFP=1.0841 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,168B, BPFP=1.9705 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,792B, BPFP=1.6511 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,992B, BPFP=1.9007 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,016B, BPFP=0.5688 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12547956 53.01969463 + layer.0.v_cache 0.00001492 0.00749010 + layer.1.k_cache 0.50446328 4.21414103 + layer.1.v_cache 0.00000633 0.00308732 + layer.2.k_cache 0.00671430 0.70139809 + layer.2.v_cache 0.00001992 0.00834113 + layer.3.k_cache 0.02142675 2.91367587 + layer.3.v_cache 0.00001986 0.00949215 + layer.4.k_cache 0.00069859 0.17862118 + layer.4.v_cache 0.00005414 0.01922781 + layer.4.output 0.01038299 209.79583107 + ------------------------------------------------------------------------------------- + TOTAL 0.04303403 89.97917569 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 321704 +BPFP 1.1243 bits/point +EBPFP 2.2485 equivalent bits/point +MSE 89.979176 +---------------------- -------------------------------------------------------- +Time: 0.646s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 89.9792 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,256B, BPFP=0.4805 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,128B, BPFP=1.9792 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,672B, BPFP=0.8654 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,084B, BPFP=1.9250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,912B, BPFP=1.0855 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,276B, BPFP=1.8831 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,840B, BPFP=1.0299 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,832B, BPFP=1.9120 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,480B, BPFP=1.5822 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,268B, BPFP=1.8308 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,276B, BPFP=0.5286 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13970643 52.60668540 + layer.0.v_cache 0.00001441 0.00739870 + layer.1.k_cache 0.60015088 4.49618094 + layer.1.v_cache 0.00000584 0.00312052 + layer.2.k_cache 0.00648033 0.72059631 + layer.2.v_cache 0.00001840 0.00828666 + layer.3.k_cache 0.03288257 3.10699929 + layer.3.v_cache 0.00002024 0.00905460 + layer.4.k_cache 0.00070287 0.17324842 + layer.4.v_cache 0.00005060 0.01841504 + layer.4.output 0.04874230 177.94104473 + ------------------------------------------------------------------------------------- + TOTAL 0.06595463 76.86689994 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 352024 +BPFP 1.0749 bits/point +EBPFP 2.1498 equivalent bits/point +MSE 76.866900 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.385s +---------------------- -------------------------------------------------------- +💾 Converting with 76.8669 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,124B, BPFP=0.4838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,136B, BPFP=2.0134 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,464B, BPFP=0.8823 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,836B, BPFP=1.9513 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,612B, BPFP=1.1282 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,996B, BPFP=1.9111 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,356B, BPFP=1.0682 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,636B, BPFP=1.9417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,824B, BPFP=1.6162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,980B, BPFP=1.8626 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,340B, BPFP=0.5348 +⌛️ [2/4] FRONTEND: Frontend time: 0.285s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.441s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13960611 52.23155581 + layer.0.v_cache 0.00001491 0.00750467 + layer.1.k_cache 0.68890666 4.62523406 + layer.1.v_cache 0.00000583 0.00316047 + layer.2.k_cache 0.00940153 0.67917369 + layer.2.v_cache 0.00001882 0.00841279 + layer.3.k_cache 0.04667800 2.74830819 + layer.3.v_cache 0.00001939 0.00951036 + layer.4.k_cache 0.00070827 0.17486701 + layer.4.v_cache 0.00005291 0.01842239 + layer.4.output 0.04427218 165.19928189 + ------------------------------------------------------------------------------------- + TOTAL 0.07031281 71.58241898 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 389304 +BPFP 1.0942 bits/point +EBPFP 2.1885 equivalent bits/point +MSE 71.582419 +---------------------- -------------------------------------------------------- +Time: 0.740s Load: 0.014s, Pack+Encode: 0.285s, Decode+Unpack: 0.441s +---------------------- -------------------------------------------------------- +💾 Converting with 71.5824 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,260B, BPFP=0.4839 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,992B, BPFP=1.9854 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,864B, BPFP=0.8813 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,952B, BPFP=1.9310 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,108B, BPFP=1.1031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,224B, BPFP=1.8930 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,948B, BPFP=1.0424 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,656B, BPFP=1.9156 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,664B, BPFP=1.6024 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,264B, BPFP=1.8428 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,528B, BPFP=0.5414 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15817952 51.74116848 + layer.0.v_cache 0.00001474 0.00723116 + layer.1.k_cache 0.61609703 4.30652713 + layer.1.v_cache 0.00000566 0.00304164 + layer.2.k_cache 0.00844678 0.70851426 + layer.2.v_cache 0.00001876 0.00852273 + layer.3.k_cache 0.04682518 3.02132638 + layer.3.v_cache 0.00001896 0.00908844 + layer.4.k_cache 0.00069859 0.17469435 + layer.4.v_cache 0.00005216 0.01849645 + layer.4.output 0.04649196 180.19989250 + ------------------------------------------------------------------------------------- + TOTAL 0.06798830 77.72928579 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 353460 +BPFP 1.0865 bits/point +EBPFP 2.1731 equivalent bits/point +MSE 77.729286 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 77.7293 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,696B, BPFP=0.4734 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,584B, BPFP=1.9917 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,260B, BPFP=0.8852 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,640B, BPFP=1.9403 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,332B, BPFP=1.1069 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,760B, BPFP=1.8924 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,124B, BPFP=1.0412 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,184B, BPFP=1.9155 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,492B, BPFP=1.6056 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,760B, BPFP=1.8380 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,780B, BPFP=0.5272 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14002854 51.92680545 + layer.0.v_cache 0.00001601 0.00752902 + layer.1.k_cache 0.54217949 4.36967101 + layer.1.v_cache 0.00000579 0.00323508 + layer.2.k_cache 0.01082110 0.74055858 + layer.2.v_cache 0.00001971 0.00880363 + layer.3.k_cache 0.03102738 2.87946939 + layer.3.v_cache 0.00001907 0.00940038 + layer.4.k_cache 0.00066782 0.17374511 + layer.4.v_cache 0.00005141 0.01890762 + layer.4.output 0.00856273 191.14696366 + ------------------------------------------------------------------------------------- + TOTAL 0.04616326 82.24511005 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 337612 +BPFP 1.0812 bits/point +EBPFP 2.1624 equivalent bits/point +MSE 82.245110 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.012s, Pack+Encode: 0.252s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 82.2451 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,168B, BPFP=0.4823 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,708B, BPFP=1.9838 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,596B, BPFP=0.8731 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,696B, BPFP=1.9306 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,780B, BPFP=1.0932 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,036B, BPFP=1.8958 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,788B, BPFP=1.0410 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,572B, BPFP=1.9240 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,292B, BPFP=1.5936 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,056B, BPFP=1.8443 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,556B, BPFP=0.5303 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14705232 53.07816446 + layer.0.v_cache 0.00001374 0.00721851 + layer.1.k_cache 0.65666707 4.60859202 + layer.1.v_cache 0.00000579 0.00301581 + layer.2.k_cache 0.01062412 0.67545444 + layer.2.v_cache 0.00001918 0.00828987 + layer.3.k_cache 0.03205616 3.14519565 + layer.3.v_cache 0.00001999 0.00954026 + layer.4.k_cache 0.00071739 0.17321536 + layer.4.v_cache 0.00005111 0.01794645 + layer.4.output 0.04976816 180.97509319 + ------------------------------------------------------------------------------------- + TOTAL 0.07032965 78.15013442 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 349248 +BPFP 1.0808 bits/point +EBPFP 2.1616 equivalent bits/point +MSE 78.150134 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 78.1501 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,996B, BPFP=0.5002 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,880B, BPFP=1.9951 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,952B, BPFP=0.8870 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,840B, BPFP=1.9373 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,992B, BPFP=1.1117 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,140B, BPFP=1.8984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,876B, BPFP=1.0496 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,624B, BPFP=1.9253 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,748B, BPFP=1.5985 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,196B, BPFP=1.8459 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,740B, BPFP=0.5460 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131647 51.54050128 + layer.0.v_cache 0.00001392 0.00711717 + layer.1.k_cache 0.51967005 4.24989661 + layer.1.v_cache 0.00000557 0.00280450 + layer.2.k_cache 0.01013049 0.74432786 + layer.2.v_cache 0.00001908 0.00786809 + layer.3.k_cache 0.05186962 2.89218759 + layer.3.v_cache 0.00001891 0.00876477 + layer.4.k_cache 0.00069958 0.16833922 + layer.4.v_cache 0.00005017 0.01748223 + layer.4.output 0.01073789 195.91009469 + ------------------------------------------------------------------------------------- + TOTAL 0.04758583 84.17705601 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 333984 +BPFP 1.0924 bits/point +EBPFP 2.1848 equivalent bits/point +MSE 84.177056 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 84.1771 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,668B, BPFP=0.4703 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,444B, BPFP=1.9772 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,032B, BPFP=0.8698 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,460B, BPFP=1.9238 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,248B, BPFP=1.0985 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,748B, BPFP=1.8852 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,316B, BPFP=1.0480 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,212B, BPFP=1.9104 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,428B, BPFP=1.5966 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,812B, BPFP=1.8344 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,440B, BPFP=0.5459 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15566496 51.91364543 + layer.0.v_cache 0.00001437 0.00724087 + layer.1.k_cache 0.62643353 4.16232469 + layer.1.v_cache 0.00000573 0.00305027 + layer.2.k_cache 0.01010078 0.73520708 + layer.2.v_cache 0.00001868 0.00834804 + layer.3.k_cache 0.06271097 2.98429574 + layer.3.v_cache 0.00001958 0.00924003 + layer.4.k_cache 0.00073784 0.17287160 + layer.4.v_cache 0.00005294 0.01841933 + layer.4.output 0.00677609 191.00633991 + ------------------------------------------------------------------------------------- + TOTAL 0.05312894 82.17994250 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 339808 +BPFP 1.0845 bits/point +EBPFP 2.1689 equivalent bits/point +MSE 82.179942 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 82.1799 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,776B, BPFP=0.4795 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,332B, BPFP=1.9849 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,184B, BPFP=0.8842 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,304B, BPFP=1.9288 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,292B, BPFP=1.1086 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,652B, BPFP=1.8931 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,956B, BPFP=1.0356 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,064B, BPFP=1.9156 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,312B, BPFP=1.6014 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,732B, BPFP=1.8429 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,896B, BPFP=0.5221 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13918158 52.03577087 + layer.0.v_cache 0.00001566 0.00755488 + layer.1.k_cache 0.57997750 4.40572032 + layer.1.v_cache 0.00000571 0.00324269 + layer.2.k_cache 0.01637170 0.73339310 + layer.2.v_cache 0.00001887 0.00898743 + layer.3.k_cache 0.04393929 3.07386812 + layer.3.v_cache 0.00001903 0.00951047 + layer.4.k_cache 0.00071104 0.18200898 + layer.4.v_cache 0.00005175 0.01967939 + layer.4.output 0.01065216 193.21953047 + ------------------------------------------------------------------------------------- + TOTAL 0.05028572 83.11861468 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 335500 +BPFP 1.0782 bits/point +EBPFP 2.1564 equivalent bits/point +MSE 83.118615 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 83.1186 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,608B, BPFP=0.4843 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,568B, BPFP=1.9440 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,984B, BPFP=0.8560 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,496B, BPFP=1.8899 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,488B, BPFP=1.0831 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,756B, BPFP=1.8526 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,388B, BPFP=1.0276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,244B, BPFP=1.8772 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,184B, BPFP=1.5718 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,724B, BPFP=1.8006 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 73,172B, BPFP=0.5269 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16496690 51.43757560 + layer.0.v_cache 0.00001433 0.00749995 + layer.1.k_cache 0.65265572 4.51702605 + layer.1.v_cache 0.00000578 0.00324501 + layer.2.k_cache 0.01241827 0.72363685 + layer.2.v_cache 0.00001920 0.00852925 + layer.3.k_cache 0.08444468 2.90973806 + layer.3.v_cache 0.00001975 0.00940931 + layer.4.k_cache 0.00068928 0.17546389 + layer.4.v_cache 0.00005259 0.01903711 + layer.4.output 0.04785494 173.72903226 + ------------------------------------------------------------------------------------- + TOTAL 0.07354536 75.05378746 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 358612 +BPFP 1.0632 bits/point +EBPFP 2.1265 equivalent bits/point +MSE 75.053787 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.010s, Pack+Encode: 0.257s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 75.0538 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,120B, BPFP=0.4847 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,588B, BPFP=1.9977 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,568B, BPFP=0.8805 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,504B, BPFP=1.9401 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,704B, BPFP=1.1003 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,792B, BPFP=1.9022 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,488B, BPFP=1.0357 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,156B, BPFP=1.9216 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,180B, BPFP=1.6040 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,888B, BPFP=1.8542 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,912B, BPFP=0.5384 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15001445 52.30898903 + layer.0.v_cache 0.00001441 0.00748535 + layer.1.k_cache 0.57475359 4.47888640 + layer.1.v_cache 0.00000584 0.00327459 + layer.2.k_cache 0.01434258 0.74758122 + layer.2.v_cache 0.00001954 0.00863883 + layer.3.k_cache 0.02649183 3.24812556 + layer.3.v_cache 0.00001945 0.00948696 + layer.4.k_cache 0.00069414 0.18254166 + layer.4.v_cache 0.00005306 0.01930280 + layer.4.output 0.05035525 183.60469813 + ------------------------------------------------------------------------------------- + TOTAL 0.06581739 79.19101172 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 347900 +BPFP 1.0876 bits/point +EBPFP 2.1752 equivalent bits/point +MSE 79.191012 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.011s, Pack+Encode: 0.254s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 79.1910 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,888B, BPFP=0.4925 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,988B, BPFP=1.9940 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,052B, BPFP=0.8894 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,076B, BPFP=1.9435 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,236B, BPFP=1.1212 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,288B, BPFP=1.8998 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,968B, BPFP=1.0510 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,764B, BPFP=1.9262 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,100B, BPFP=1.6124 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,432B, BPFP=1.8524 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,424B, BPFP=0.5258 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11854591 53.11984015 + layer.0.v_cache 0.00001566 0.00729402 + layer.1.k_cache 0.56985614 4.35816079 + layer.1.v_cache 0.00000555 0.00310533 + layer.2.k_cache 0.01267573 0.69780699 + layer.2.v_cache 0.00001902 0.00854788 + layer.3.k_cache 0.04477839 2.80836866 + layer.3.v_cache 0.00001896 0.00969576 + layer.4.k_cache 0.00071754 0.17615374 + layer.4.v_cache 0.00005321 0.01910093 + layer.4.output 0.00761333 195.04657421 + ------------------------------------------------------------------------------------- + TOTAL 0.04705761 83.91377022 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 333216 +BPFP 1.0860 bits/point +EBPFP 2.1721 equivalent bits/point +MSE 83.913770 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.012s, Pack+Encode: 0.252s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,952B, BPFP=0.4758 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,744B, BPFP=2.0060 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,576B, BPFP=0.8810 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,644B, BPFP=1.9475 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,888B, BPFP=1.1101 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,004B, BPFP=1.9135 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,784B, BPFP=1.0514 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,420B, BPFP=1.9356 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,492B, BPFP=1.6205 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,084B, BPFP=1.8646 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,356B, BPFP=0.5342 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582189 53.42725606 + layer.0.v_cache 0.00001509 0.00770541 + layer.1.k_cache 0.58311327 4.47464674 + layer.1.v_cache 0.00000606 0.00321803 + layer.2.k_cache 0.00657600 0.73139435 + layer.2.v_cache 0.00001909 0.00880321 + layer.3.k_cache 0.02234380 3.04150038 + layer.3.v_cache 0.00001905 0.00960280 + layer.4.k_cache 0.00071828 0.18094992 + layer.4.v_cache 0.00005370 0.01923591 + layer.4.output 0.05044473 183.40942359 + ------------------------------------------------------------------------------------- + TOTAL 0.06363526 79.16295753 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 348944 +BPFP 1.0909 bits/point +EBPFP 2.1818 equivalent bits/point +MSE 79.162958 +---------------------- -------------------------------------------------------- +Time: 0.651s Load: 0.012s, Pack+Encode: 0.254s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 79.1630 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 292, 128) +Output shape: (1, 292, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.output: torch.Size([1, 292, 3584]) -> torch.Size([1, 1, 292, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,936B, BPFP=0.4782 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,308B, BPFP=1.9964 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,344B, BPFP=0.8746 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,204B, BPFP=1.9373 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,780B, BPFP=1.1119 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,504B, BPFP=1.8998 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,596B, BPFP=1.0486 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,044B, BPFP=1.9287 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,052B, BPFP=1.6081 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,600B, BPFP=1.8515 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,552B, BPFP=0.5317 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887579 52.67003224 + layer.0.v_cache 0.00001384 0.00749062 + layer.1.k_cache 0.57277983 4.36813522 + layer.1.v_cache 0.00000568 0.00310692 + layer.2.k_cache 0.01314942 0.76188074 + layer.2.v_cache 0.00002105 0.00870104 + layer.3.k_cache 0.03125376 2.73977954 + layer.3.v_cache 0.00001978 0.00942299 + layer.4.k_cache 0.00068035 0.17466908 + layer.4.v_cache 0.00005123 0.01855954 + layer.4.output 0.04859251 185.14235262 + ------------------------------------------------------------------------------------- + TOTAL 0.06511755 79.80930860 + (elements=2,541,568) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2541568 +Total Bytes 344920 +BPFP 1.0857 bits/point +EBPFP 2.1714 equivalent bits/point +MSE 79.809309 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.250s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 79.8093 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 338, 128) +Output shape: (1, 338, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.output: torch.Size([1, 338, 3584]) -> torch.Size([1, 1, 338, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,608B, BPFP=0.4904 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 43,228B, BPFP=1.9983 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,844B, BPFP=0.8711 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,748B, BPFP=1.9299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,616B, BPFP=1.0917 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,884B, BPFP=1.8900 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,348B, BPFP=1.0331 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,416B, BPFP=1.9146 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,388B, BPFP=1.5897 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,940B, BPFP=1.8463 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,096B, BPFP=0.5488 +⌛️ [2/4] FRONTEND: Frontend time: 0.284s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.444s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13976314 51.30246857 + layer.0.v_cache 0.00001472 0.00756417 + layer.1.k_cache 0.73750278 4.25987903 + layer.1.v_cache 0.00000589 0.00317510 + layer.2.k_cache 0.02011468 0.70298162 + layer.2.v_cache 0.00002041 0.00823964 + layer.3.k_cache 0.02722352 2.94427924 + layer.3.v_cache 0.00002029 0.00905592 + layer.4.k_cache 0.00072609 0.17711312 + layer.4.v_cache 0.00005477 0.01905796 + layer.4.output 0.04270584 159.40087437 + ------------------------------------------------------------------------------------- + TOTAL 0.07202277 69.13176088 + (elements=2,941,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2941952 +Total Bytes 400116 +BPFP 1.0880 bits/point +EBPFP 2.1761 equivalent bits/point +MSE 69.131761 +---------------------- -------------------------------------------------------- +Time: 0.740s Load: 0.012s, Pack+Encode: 0.284s, Decode+Unpack: 0.444s +---------------------- -------------------------------------------------------- +💾 Converting with 69.1318 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,720B, BPFP=0.4682 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,956B, BPFP=1.9843 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,252B, BPFP=0.8726 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,908B, BPFP=1.9280 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,328B, BPFP=1.0915 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,076B, BPFP=1.8834 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,416B, BPFP=1.0425 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,552B, BPFP=1.9089 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,788B, BPFP=1.5994 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,272B, BPFP=1.8402 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,132B, BPFP=0.5303 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150811 52.47749544 + layer.0.v_cache 0.00001374 0.00738460 + layer.1.k_cache 0.66021251 4.80241473 + layer.1.v_cache 0.00000574 0.00308007 + layer.2.k_cache 0.00963001 0.69582606 + layer.2.v_cache 0.00001988 0.00852254 + layer.3.k_cache 0.08055985 2.64591534 + layer.3.v_cache 0.00001942 0.00902315 + layer.4.k_cache 0.00069026 0.17303317 + layer.4.v_cache 0.00005126 0.01830128 + layer.4.output 0.05093874 185.62556762 + ------------------------------------------------------------------------------------- + TOTAL 0.07289894 80.01293940 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 341400 +BPFP 1.0783 bits/point +EBPFP 2.1566 equivalent bits/point +MSE 80.012939 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 80.0129 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,504B, BPFP=0.4901 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,016B, BPFP=1.9604 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,860B, BPFP=0.8694 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,904B, BPFP=1.9031 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,976B, BPFP=1.0817 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,264B, BPFP=1.8700 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,896B, BPFP=1.0260 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,680B, BPFP=1.8915 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,564B, BPFP=1.5761 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,320B, BPFP=1.8214 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,124B, BPFP=0.5240 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14893842 51.38733434 + layer.0.v_cache 0.00001415 0.00720012 + layer.1.k_cache 0.65110169 4.39265548 + layer.1.v_cache 0.00000575 0.00313503 + layer.2.k_cache 0.01411552 0.68487972 + layer.2.v_cache 0.00001904 0.00834784 + layer.3.k_cache 0.03939451 2.94222025 + layer.3.v_cache 0.00001876 0.00879073 + layer.4.k_cache 0.00070558 0.17684925 + layer.4.v_cache 0.00005345 0.01855327 + layer.4.output 0.04819407 176.50368340 + ------------------------------------------------------------------------------------- + TOTAL 0.07010149 76.18563234 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 352108 +BPFP 1.0681 bits/point +EBPFP 2.1362 equivalent bits/point +MSE 76.185632 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 76.1856 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,256B, BPFP=0.4905 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,228B, BPFP=2.0335 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,168B, BPFP=0.9011 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,312B, BPFP=1.9791 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,460B, BPFP=1.1561 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,724B, BPFP=1.9442 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,056B, BPFP=1.0727 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,120B, BPFP=1.9677 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,584B, BPFP=1.6388 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,724B, BPFP=1.8847 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,928B, BPFP=0.5256 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14492303 52.40332774 + layer.0.v_cache 0.00001375 0.00722443 + layer.1.k_cache 0.50395557 4.52180324 + layer.1.v_cache 0.00000564 0.00307608 + layer.2.k_cache 0.00856250 0.70278101 + layer.2.v_cache 0.00001902 0.00835045 + layer.3.k_cache 0.02692018 2.95037308 + layer.3.v_cache 0.00002118 0.00928255 + layer.4.k_cache 0.00070790 0.17491059 + layer.4.v_cache 0.00005130 0.01855074 + layer.4.output 0.01131245 209.92218903 + ------------------------------------------------------------------------------------- + TOTAL 0.04496278 90.01500018 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 315560 +BPFP 1.1028 bits/point +EBPFP 2.2056 equivalent bits/point +MSE 90.015000 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.009s, Pack+Encode: 0.253s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 90.0150 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,712B, BPFP=0.4932 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,508B, BPFP=2.0102 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,636B, BPFP=0.8852 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,504B, BPFP=1.9534 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,724B, BPFP=1.1166 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,760B, BPFP=1.9112 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,820B, BPFP=1.0654 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,304B, BPFP=1.9420 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,516B, BPFP=1.6144 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,928B, BPFP=1.8641 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,712B, BPFP=0.5476 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14113761 50.37849227 + layer.0.v_cache 0.00001369 0.00730565 + layer.1.k_cache 0.59212339 4.47600080 + layer.1.v_cache 0.00000552 0.00302188 + layer.2.k_cache 0.01129283 0.73611262 + layer.2.v_cache 0.00001853 0.00833729 + layer.3.k_cache 0.01358872 2.93513268 + layer.3.v_cache 0.00001883 0.00900684 + layer.4.k_cache 0.00070763 0.17569542 + layer.4.v_cache 0.00005153 0.01845904 + layer.4.output 0.01019159 200.40194746 + ------------------------------------------------------------------------------------- + TOTAL 0.04884114 85.97418804 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 330124 +BPFP 1.0994 bits/point +EBPFP 2.1987 equivalent bits/point +MSE 85.974188 +---------------------- -------------------------------------------------------- +Time: 0.639s Load: 0.009s, Pack+Encode: 0.250s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 85.9742 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,756B, BPFP=0.4767 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,532B, BPFP=1.9889 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,116B, BPFP=0.8774 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,448B, BPFP=1.9299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,272B, BPFP=1.1037 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,680B, BPFP=1.8881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,104B, BPFP=1.0401 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,160B, BPFP=1.9142 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,536B, BPFP=1.6080 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,868B, BPFP=1.8439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,980B, BPFP=0.5520 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11028925 53.63305069 + layer.0.v_cache 0.00001466 0.00726278 + layer.1.k_cache 0.56032453 4.64004102 + layer.1.v_cache 0.00000595 0.00295571 + layer.2.k_cache 0.00809042 0.71902705 + layer.2.v_cache 0.00001929 0.00815809 + layer.3.k_cache 0.03501172 3.12213645 + layer.3.v_cache 0.00001861 0.00894430 + layer.4.k_cache 0.00070074 0.17200837 + layer.4.v_cache 0.00005246 0.01841792 + layer.4.output 0.00889761 192.30252924 + ------------------------------------------------------------------------------------- + TOTAL 0.04569476 82.84998277 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 340452 +BPFP 1.0903 bits/point +EBPFP 2.1806 equivalent bits/point +MSE 82.849983 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 82.8500 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 311, 128) +Output shape: (1, 311, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.output: torch.Size([1, 311, 3584]) -> torch.Size([1, 1, 311, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,472B, BPFP=0.4759 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,568B, BPFP=1.9377 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,144B, BPFP=0.8613 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,620B, BPFP=1.8901 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,680B, BPFP=1.0892 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,856B, BPFP=1.8517 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,752B, BPFP=1.0426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,480B, BPFP=1.8830 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,344B, BPFP=1.5748 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,888B, BPFP=1.8031 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 73,196B, BPFP=0.5254 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14227538 52.46955700 + layer.0.v_cache 0.00001500 0.00756334 + layer.1.k_cache 0.64922478 4.40161094 + layer.1.v_cache 0.00000572 0.00313254 + layer.2.k_cache 0.00828172 0.71351992 + layer.2.v_cache 0.00001922 0.00842174 + layer.3.k_cache 0.02636586 2.81411027 + layer.3.v_cache 0.00001996 0.00929829 + layer.4.k_cache 0.00069494 0.17253141 + layer.4.v_cache 0.00005248 0.01888247 + layer.4.output 0.04686997 169.62731109 + ------------------------------------------------------------------------------------- + TOTAL 0.06794382 73.41234150 + (elements=2,706,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2706944 +Total Bytes 360000 +BPFP 1.0639 bits/point +EBPFP 2.1279 equivalent bits/point +MSE 73.412342 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.010s, Pack+Encode: 0.252s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 73.4123 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,976B, BPFP=0.4887 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,076B, BPFP=1.9042 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,208B, BPFP=0.8706 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,224B, BPFP=1.8520 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,012B, BPFP=1.1037 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,796B, BPFP=1.8257 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,036B, BPFP=1.0439 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,168B, BPFP=1.8485 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,444B, BPFP=1.5591 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,980B, BPFP=1.7757 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,484B, BPFP=0.5557 +⌛️ [2/4] FRONTEND: Frontend time: 0.285s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13149661 51.04331342 + layer.0.v_cache 0.00001370 0.00742270 + layer.1.k_cache 0.50214604 4.34990043 + layer.1.v_cache 0.00000625 0.00324377 + layer.2.k_cache 0.00807812 0.68139493 + layer.2.v_cache 0.00001969 0.00857163 + layer.3.k_cache 0.06061829 2.87056574 + layer.3.v_cache 0.00001868 0.00920734 + layer.4.k_cache 0.00068906 0.17058629 + layer.4.v_cache 0.00005044 0.01869095 + layer.4.output 1.20679682 179.08697479 + ------------------------------------------------------------------------------------- + TOTAL 0.53827733 77.22186592 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 296404 +BPFP 1.0684 bits/point +EBPFP 2.1367 equivalent bits/point +MSE 77.221866 +---------------------- -------------------------------------------------------- +Time: 0.618s Load: 0.008s, Pack+Encode: 0.285s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 77.2219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 300, 128) +Output shape: (1, 300, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.output: torch.Size([1, 300, 3584]) -> torch.Size([1, 1, 300, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,236B, BPFP=0.4810 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,072B, BPFP=1.9829 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,564B, BPFP=0.8627 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,908B, BPFP=1.9223 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,980B, BPFP=1.0927 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,256B, BPFP=1.8883 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,864B, BPFP=1.0346 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,840B, BPFP=1.9187 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,572B, BPFP=1.5923 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,252B, BPFP=1.8360 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,400B, BPFP=0.5238 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13491969 51.91427083 + layer.0.v_cache 0.00001558 0.00706975 + layer.1.k_cache 0.64590566 4.29039836 + layer.1.v_cache 0.00000556 0.00291442 + layer.2.k_cache 0.01395306 0.65325846 + layer.2.v_cache 0.00001972 0.00828093 + layer.3.k_cache 0.02379198 2.93511007 + layer.3.v_cache 0.00001942 0.00912138 + layer.4.k_cache 0.00073123 0.17783110 + layer.4.v_cache 0.00005274 0.01829513 + layer.4.output 0.05011183 179.34136905 + ------------------------------------------------------------------------------------- + TOTAL 0.06883514 77.37683140 + (elements=2,611,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2611200 +Total Bytes 350944 +BPFP 1.0752 bits/point +EBPFP 2.1504 equivalent bits/point +MSE 77.376831 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.009s, Pack+Encode: 0.258s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 77.3768 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,804B, BPFP=0.4727 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,884B, BPFP=1.9805 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,204B, BPFP=0.8701 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,864B, BPFP=1.9257 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,180B, BPFP=1.0835 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,996B, BPFP=1.8791 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,192B, BPFP=1.0305 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,448B, BPFP=1.9034 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,564B, BPFP=1.5874 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,020B, BPFP=1.8267 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,572B, BPFP=0.5260 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14826349 53.58190037 + layer.0.v_cache 0.00001415 0.00750393 + layer.1.k_cache 0.61440683 4.61896486 + layer.1.v_cache 0.00000562 0.00323646 + layer.2.k_cache 0.00551110 0.72002566 + layer.2.v_cache 0.00001912 0.00850497 + layer.3.k_cache 0.01294636 3.04900704 + layer.3.v_cache 0.00002032 0.00930350 + layer.4.k_cache 0.00070633 0.18047350 + layer.4.v_cache 0.00005074 0.01881160 + layer.4.output 0.04972342 185.84905805 + ------------------------------------------------------------------------------------- + TOTAL 0.06647106 80.18477284 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 339728 +BPFP 1.0730 bits/point +EBPFP 2.1460 equivalent bits/point +MSE 80.184773 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.010s, Pack+Encode: 0.253s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 80.1848 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,860B, BPFP=0.4998 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,524B, BPFP=2.0038 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,632B, BPFP=0.8818 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,472B, BPFP=1.9445 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,576B, BPFP=1.1042 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,716B, BPFP=1.9019 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,640B, BPFP=1.0514 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,240B, BPFP=1.9314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,440B, BPFP=1.6042 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,844B, BPFP=1.8527 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 65,876B, BPFP=0.5308 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13861307 51.34700756 + layer.0.v_cache 0.00001419 0.00738597 + layer.1.k_cache 0.53867822 4.61250256 + layer.1.v_cache 0.00000573 0.00309897 + layer.2.k_cache 0.00871976 0.71670532 + layer.2.v_cache 0.00001877 0.00819566 + layer.3.k_cache 0.02195022 2.92866070 + layer.3.v_cache 0.00001967 0.00931274 + layer.4.k_cache 0.00069227 0.17225421 + layer.4.v_cache 0.00004961 0.01757064 + layer.4.output 0.00954030 199.45090897 + ------------------------------------------------------------------------------------- + TOTAL 0.04562021 85.64582689 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 327820 +BPFP 1.0877 bits/point +EBPFP 2.1755 equivalent bits/point +MSE 85.645827 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.011s, Pack+Encode: 0.253s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 85.6458 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,744B, BPFP=0.4777 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,192B, BPFP=1.9773 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,208B, BPFP=0.8855 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,268B, BPFP=1.9268 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,260B, BPFP=1.1069 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,632B, BPFP=1.8920 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,208B, BPFP=1.0494 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,292B, BPFP=1.9281 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,416B, BPFP=1.6071 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,764B, BPFP=1.8446 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,860B, BPFP=0.5296 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13928204 52.75695203 + layer.0.v_cache 0.00001373 0.00729713 + layer.1.k_cache 0.60860016 4.31151533 + layer.1.v_cache 0.00000555 0.00304308 + layer.2.k_cache 0.01026906 0.76754163 + layer.2.v_cache 0.00001872 0.00842584 + layer.3.k_cache 0.05166252 2.98056393 + layer.3.v_cache 0.00001937 0.00939544 + layer.4.k_cache 0.00073448 0.17474661 + layer.4.v_cache 0.00005026 0.01844897 + layer.4.output 0.00697322 192.77366384 + ------------------------------------------------------------------------------------- + TOTAL 0.05055697 82.96785746 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 336844 +BPFP 1.0825 bits/point +EBPFP 2.1650 equivalent bits/point +MSE 82.967857 +---------------------- -------------------------------------------------------- +Time: 0.641s Load: 0.009s, Pack+Encode: 0.251s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 82.9679 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,192B, BPFP=0.4869 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,628B, BPFP=1.9930 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,548B, BPFP=0.8765 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,584B, BPFP=1.9377 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,680B, BPFP=1.0953 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,884B, BPFP=1.9006 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,608B, BPFP=1.0386 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,340B, BPFP=1.9248 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,532B, BPFP=1.6172 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,984B, BPFP=1.8530 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,624B, BPFP=0.5268 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13820392 53.50052304 + layer.0.v_cache 0.00001454 0.00742081 + layer.1.k_cache 0.61684219 4.55006704 + layer.1.v_cache 0.00000570 0.00312041 + layer.2.k_cache 0.01285210 0.71535308 + layer.2.v_cache 0.00002011 0.00848384 + layer.3.k_cache 0.05980752 3.21323615 + layer.3.v_cache 0.00002263 0.00935865 + layer.4.k_cache 0.00071023 0.17761505 + layer.4.v_cache 0.00005833 0.01891775 + layer.4.output 0.05086143 182.99114709 + ------------------------------------------------------------------------------------- + TOTAL 0.06968043 79.00836032 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 347604 +BPFP 1.0830 bits/point +EBPFP 2.1660 equivalent bits/point +MSE 79.008360 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.378s +---------------------- -------------------------------------------------------- +💾 Converting with 79.0084 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,312B, BPFP=0.4834 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,260B, BPFP=1.9861 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,780B, BPFP=0.8711 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,180B, BPFP=1.9300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,060B, BPFP=1.0932 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,456B, BPFP=1.8924 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,100B, BPFP=1.0434 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,972B, BPFP=1.9192 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,808B, BPFP=1.5993 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,472B, BPFP=1.8414 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,292B, BPFP=0.5287 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14825445 51.34733181 + layer.0.v_cache 0.00001396 0.00773612 + layer.1.k_cache 0.63210902 4.44889796 + layer.1.v_cache 0.00000581 0.00330851 + layer.2.k_cache 0.01184132 0.67850725 + layer.2.v_cache 0.00001890 0.00873048 + layer.3.k_cache 0.01767834 2.88935076 + layer.3.v_cache 0.00002009 0.00958681 + layer.4.k_cache 0.00068272 0.18247086 + layer.4.v_cache 0.00005247 0.01938999 + layer.4.output 0.04867053 178.78301495 + ------------------------------------------------------------------------------------- + TOTAL 0.06772769 77.12214207 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 353692 +BPFP 1.0800 bits/point +EBPFP 2.1600 equivalent bits/point +MSE 77.122142 +---------------------- -------------------------------------------------------- +Time: 0.650s Load: 0.011s, Pack+Encode: 0.255s, Decode+Unpack: 0.384s +---------------------- -------------------------------------------------------- +💾 Converting with 77.1221 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,368B, BPFP=0.4953 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,332B, BPFP=2.0320 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,448B, BPFP=0.9143 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,352B, BPFP=1.9740 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,232B, BPFP=1.1383 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 32,776B, BPFP=1.9399 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,296B, BPFP=1.0829 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,316B, BPFP=1.9718 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,764B, BPFP=1.6432 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 31,880B, BPFP=1.8868 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 61,328B, BPFP=0.5185 +⌛️ [2/4] FRONTEND: Frontend time: 0.264s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14315909 53.50671757 + layer.0.v_cache 0.00001391 0.00773264 + layer.1.k_cache 0.49711875 4.24203769 + layer.1.v_cache 0.00000560 0.00331439 + layer.2.k_cache 0.00646762 0.68874301 + layer.2.v_cache 0.00001905 0.00867633 + layer.3.k_cache 0.04707547 2.88633913 + layer.3.v_cache 0.00001924 0.00962914 + layer.4.k_cache 0.00071833 0.17630161 + layer.4.v_cache 0.00005197 0.01861577 + layer.4.output 0.00997522 209.16002097 + ------------------------------------------------------------------------------------- + TOTAL 0.04496915 89.74519141 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 316092 +BPFP 1.1005 bits/point +EBPFP 2.2010 equivalent bits/point +MSE 89.745191 +---------------------- -------------------------------------------------------- +Time: 0.657s Load: 0.011s, Pack+Encode: 0.264s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 89.7452 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,752B, BPFP=0.4901 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,680B, BPFP=1.9982 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,708B, BPFP=0.8797 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,672B, BPFP=1.9418 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,080B, BPFP=1.1246 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,076B, BPFP=1.9084 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,836B, BPFP=1.0549 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,592B, BPFP=1.9373 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,740B, BPFP=1.6095 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,200B, BPFP=1.8593 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,808B, BPFP=0.5425 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15581964 52.18827705 + layer.0.v_cache 0.00001703 0.00747989 + layer.1.k_cache 0.56298232 4.20858622 + layer.1.v_cache 0.00000550 0.00311436 + layer.2.k_cache 0.00979712 0.72729503 + layer.2.v_cache 0.00002023 0.00868373 + layer.3.k_cache 0.04960944 2.69608824 + layer.3.v_cache 0.00001971 0.00933008 + layer.4.k_cache 0.00068561 0.17194149 + layer.4.v_cache 0.00005245 0.01823947 + layer.4.output 0.01122868 197.57699693 + ------------------------------------------------------------------------------------- + TOTAL 0.05044764 84.88694200 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 332144 +BPFP 1.0942 bits/point +EBPFP 2.1884 equivalent bits/point +MSE 84.886942 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.010s, Pack+Encode: 0.250s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 84.8869 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,976B, BPFP=0.4956 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,036B, BPFP=1.9896 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,952B, BPFP=0.8807 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,996B, BPFP=1.9322 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,032B, BPFP=1.1060 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,276B, BPFP=1.8924 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,812B, BPFP=1.0386 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,708B, BPFP=1.9163 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,940B, BPFP=1.5978 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,344B, BPFP=1.8410 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,068B, BPFP=0.5448 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13599485 52.58568905 + layer.0.v_cache 0.00001418 0.00735099 + layer.1.k_cache 0.56373192 4.35577457 + layer.1.v_cache 0.00000588 0.00310036 + layer.2.k_cache 0.01549433 0.70984917 + layer.2.v_cache 0.00001867 0.00845346 + layer.3.k_cache 0.03475572 3.11676565 + layer.3.v_cache 0.00001907 0.00917450 + layer.4.k_cache 0.00072111 0.17679568 + layer.4.v_cache 0.00005295 0.01857686 + layer.4.output 0.00820773 195.13165699 + ------------------------------------------------------------------------------------- + TOTAL 0.04754487 83.93606642 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 335140 +BPFP 1.0885 bits/point +EBPFP 2.1769 equivalent bits/point +MSE 83.936066 +---------------------- -------------------------------------------------------- +Time: 0.648s Load: 0.011s, Pack+Encode: 0.251s, Decode+Unpack: 0.386s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,464B, BPFP=0.4801 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,296B, BPFP=1.9428 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,968B, BPFP=0.8608 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,248B, BPFP=1.8896 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,504B, BPFP=1.0909 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,688B, BPFP=1.8612 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,404B, BPFP=1.0351 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,280B, BPFP=1.8912 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,036B, BPFP=1.5745 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,792B, BPFP=1.8157 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,288B, BPFP=0.5239 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13331154 51.07029031 + layer.0.v_cache 0.00001460 0.00737264 + layer.1.k_cache 0.64535681 4.50892461 + layer.1.v_cache 0.00000563 0.00320952 + layer.2.k_cache 0.00762887 0.72316806 + layer.2.v_cache 0.00001910 0.00870532 + layer.3.k_cache 0.03538683 3.01810327 + layer.3.v_cache 0.00002103 0.00959739 + layer.4.k_cache 0.00072257 0.17680792 + layer.4.v_cache 0.00005300 0.01906776 + layer.4.output 0.04522907 175.11230288 + ------------------------------------------------------------------------------------- + TOTAL 0.06700726 75.60772747 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 356968 +BPFP 1.0652 bits/point +EBPFP 2.1305 equivalent bits/point +MSE 75.607727 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.010s, Pack+Encode: 0.254s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 75.6077 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,868B, BPFP=0.4879 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,332B, BPFP=1.9989 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,236B, BPFP=0.8933 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,368B, BPFP=1.9459 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,320B, BPFP=1.1180 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,528B, BPFP=1.8996 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,156B, BPFP=1.0539 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,104B, BPFP=1.9313 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,344B, BPFP=1.6144 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,808B, BPFP=1.8600 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,652B, BPFP=0.5474 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14806804 53.32032625 + layer.0.v_cache 0.00001376 0.00718695 + layer.1.k_cache 0.60662186 4.63692163 + layer.1.v_cache 0.00000571 0.00302170 + layer.2.k_cache 0.01239752 0.68664631 + layer.2.v_cache 0.00002027 0.00786551 + layer.3.k_cache 0.02456893 3.06178198 + layer.3.v_cache 0.00001975 0.00877609 + layer.4.k_cache 0.00069718 0.17086276 + layer.4.v_cache 0.00005379 0.01809254 + layer.4.output 0.00818182 194.27060802 + ------------------------------------------------------------------------------------- + TOTAL 0.04998468 83.63621988 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 338716 +BPFP 1.0962 bits/point +EBPFP 2.1924 equivalent bits/point +MSE 83.636220 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.012s, Pack+Encode: 0.252s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 83.6362 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,008B, BPFP=0.4973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,240B, BPFP=2.0009 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,312B, BPFP=0.9006 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,368B, BPFP=1.9527 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,360B, BPFP=1.1241 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,388B, BPFP=1.8986 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,216B, BPFP=1.0610 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,908B, BPFP=1.9273 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,216B, BPFP=1.6131 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,632B, BPFP=1.8569 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,952B, BPFP=0.5517 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14592592 52.78378285 + layer.0.v_cache 0.00001391 0.00739308 + layer.1.k_cache 0.53254910 4.46721442 + layer.1.v_cache 0.00000583 0.00315912 + layer.2.k_cache 0.01029040 0.72223159 + layer.2.v_cache 0.00002009 0.00863419 + layer.3.k_cache 0.03848921 2.99781136 + layer.3.v_cache 0.00001972 0.00949718 + layer.4.k_cache 0.00069417 0.17976067 + layer.4.v_cache 0.00005427 0.01907969 + layer.4.output 0.01004909 194.58333859 + ------------------------------------------------------------------------------------- + TOTAL 0.04696507 83.72246672 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 338600 +BPFP 1.0997 bits/point +EBPFP 2.1994 equivalent bits/point +MSE 83.722467 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.012s, Pack+Encode: 0.251s, Decode+Unpack: 0.382s +---------------------- -------------------------------------------------------- +💾 Converting with 83.7225 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,684B, BPFP=0.4989 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,248B, BPFP=2.0248 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,676B, BPFP=0.9005 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,136B, BPFP=1.9609 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,484B, BPFP=1.1193 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,508B, BPFP=1.9249 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,400B, BPFP=1.0570 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,884B, BPFP=1.9465 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,360B, BPFP=1.6291 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,592B, BPFP=1.8722 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 64,900B, BPFP=0.5326 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15384994 52.83259493 + layer.0.v_cache 0.00001404 0.00751862 + layer.1.k_cache 0.48768111 4.31821217 + layer.1.v_cache 0.00000584 0.00332204 + layer.2.k_cache 0.00538594 0.68669089 + layer.2.v_cache 0.00001948 0.00847937 + layer.3.k_cache 0.02347840 3.00675381 + layer.3.v_cache 0.00001899 0.00903379 + layer.4.k_cache 0.00068094 0.17416157 + layer.4.v_cache 0.00005116 0.01816027 + layer.4.output 0.00909345 202.47321429 + ------------------------------------------------------------------------------------- + TOTAL 0.04322588 86.96337809 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 324872 +BPFP 1.0978 bits/point +EBPFP 2.1956 equivalent bits/point +MSE 86.963378 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.009s, Pack+Encode: 0.254s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9634 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,036B, BPFP=0.4802 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,504B, BPFP=1.9932 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,580B, BPFP=0.8812 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,692B, BPFP=1.9500 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,912B, BPFP=1.1114 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,916B, BPFP=1.9088 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,916B, BPFP=1.0585 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,256B, BPFP=1.9269 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,488B, BPFP=1.6203 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,816B, BPFP=1.8503 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,168B, BPFP=0.5176 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13453574 53.32574670 + layer.0.v_cache 0.00001386 0.00745413 + layer.1.k_cache 0.70462929 4.45108966 + layer.1.v_cache 0.00000580 0.00316121 + layer.2.k_cache 0.00896267 0.77372565 + layer.2.v_cache 0.00001881 0.00856691 + layer.3.k_cache 0.02468209 2.83740712 + layer.3.v_cache 0.00001940 0.00935559 + layer.4.k_cache 0.00068809 0.17656790 + layer.4.v_cache 0.00005209 0.01810377 + layer.4.output 0.04953035 183.52010447 + ------------------------------------------------------------------------------------- + TOTAL 0.07178355 79.19128882 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 346284 +BPFP 1.0826 bits/point +EBPFP 2.1651 equivalent bits/point +MSE 79.191289 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.011s, Pack+Encode: 0.252s, Decode+Unpack: 0.380s +---------------------- -------------------------------------------------------- +💾 Converting with 79.1913 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,596B, BPFP=0.5012 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,884B, BPFP=2.0338 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,320B, BPFP=0.8932 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 33,812B, BPFP=1.9713 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,576B, BPFP=1.1413 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,236B, BPFP=1.9377 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,416B, BPFP=1.0737 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 33,744B, BPFP=1.9674 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,152B, BPFP=1.6413 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 32,380B, BPFP=1.8878 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 63,796B, BPFP=0.5313 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150386 52.24454874 + layer.0.v_cache 0.00001401 0.00742598 + layer.1.k_cache 0.57442873 4.73793964 + layer.1.v_cache 0.00000554 0.00298802 + layer.2.k_cache 0.00779028 0.73612931 + layer.2.v_cache 0.00001875 0.00837677 + layer.3.k_cache 0.05176373 2.92332162 + layer.3.v_cache 0.00002027 0.00911793 + layer.4.k_cache 0.00069872 0.16853335 + layer.4.v_cache 0.00005023 0.01852737 + layer.4.output 0.01044900 205.38955890 + ------------------------------------------------------------------------------------- + TOTAL 0.04937866 88.15198947 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 321912 +BPFP 1.1040 bits/point +EBPFP 2.2080 equivalent bits/point +MSE 88.151989 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.009s, Pack+Encode: 0.253s, Decode+Unpack: 0.381s +---------------------- -------------------------------------------------------- +💾 Converting with 88.1520 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.0876 bits/point +Avg EBPFP 2.1751 equivalent bits/point +Avg MSE 82.161105 +Avg Time 0.653s +------------------------ ---------------------------- diff --git a/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..5b724dcdebced471ee1811de2569596a8d421ba4 --- /dev/null +++ b/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 599 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.01_epochs600_lr0.0001_bs360_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa +Output output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,220B, BPFP=0.5990 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,876B, BPFP=2.2091 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,336B, BPFP=0.9926 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,392B, BPFP=2.1190 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,432B, BPFP=1.1964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,956B, BPFP=2.0379 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,156B, BPFP=1.1451 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,296B, BPFP=2.1012 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,472B, BPFP=1.7619 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,012B, BPFP=2.0484 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,680B, BPFP=0.6027 +⌛️ [2/4] FRONTEND: Frontend time: 0.397s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.251s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14073507 56.25787063 + layer.0.v_cache 0.00001426 0.00725259 + layer.1.k_cache 0.05042761 5.19575682 + layer.1.v_cache 0.00000517 0.00260725 + layer.2.k_cache 0.00244098 0.87477984 + layer.2.v_cache 0.00001715 0.00744132 + layer.3.k_cache 0.02726505 3.57943980 + layer.3.v_cache 0.00001796 0.00847526 + layer.4.k_cache 0.00069123 0.18381093 + layer.4.v_cache 0.00005084 0.01868421 + layer.4.output 0.16186065 643.05187075 + ------------------------------------------------------------------------------------- + TOTAL 0.07968764 268.67642435 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 109828 +BPFP 1.2017 bits/point +EBPFP 2.4034 equivalent bits/point +MSE 268.676424 +---------------------- -------------------------------------------------------- +Time: 0.652s Load: 0.005s, Pack+Encode: 0.397s, Decode+Unpack: 0.251s +---------------------- -------------------------------------------------------- +💾 Converting with 268.6764 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,236B, BPFP=0.6092 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,048B, BPFP=2.0798 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,272B, BPFP=0.9925 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,524B, BPFP=1.9812 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,072B, BPFP=1.1431 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,192B, BPFP=1.9187 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,856B, BPFP=1.1024 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,592B, BPFP=1.9940 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,740B, BPFP=1.6453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,168B, BPFP=1.9142 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,328B, BPFP=0.6005 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11392649 57.82127141 + layer.0.v_cache 0.00001670 0.00745717 + layer.1.k_cache 0.05249626 5.04514800 + layer.1.v_cache 0.00000506 0.00249677 + layer.2.k_cache 0.00416583 0.87492260 + layer.2.v_cache 0.00001711 0.00697673 + layer.3.k_cache 0.05931453 3.60737849 + layer.3.v_cache 0.00001789 0.00833488 + layer.4.k_cache 0.00069248 0.17238893 + layer.4.v_cache 0.00004729 0.01724032 + layer.4.output 0.16383441 651.22837780 + ------------------------------------------------------------------------------------- + TOTAL 0.08103179 272.12719176 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 104028 +BPFP 1.1520 bits/point +EBPFP 2.3040 equivalent bits/point +MSE 272.127192 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.005s, Pack+Encode: 0.164s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 272.1272 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,232B, BPFP=0.5372 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,592B, BPFP=2.0931 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,796B, BPFP=0.9634 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,992B, BPFP=1.9934 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,852B, BPFP=1.1390 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,632B, BPFP=1.9335 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,484B, BPFP=1.0778 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,972B, BPFP=1.9900 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,036B, BPFP=1.6682 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,588B, BPFP=1.9262 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,204B, BPFP=0.5273 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17268721 54.72621966 + layer.0.v_cache 0.00001453 0.00723000 + layer.1.k_cache 0.08276152 4.67956965 + layer.1.v_cache 0.00000545 0.00292966 + layer.2.k_cache 0.00739249 0.75506860 + layer.2.v_cache 0.00001625 0.00694530 + layer.3.k_cache 0.05929063 3.34319744 + layer.3.v_cache 0.00001766 0.00853621 + layer.4.k_cache 0.00068371 0.17429646 + layer.4.v_cache 0.00005055 0.01719081 + layer.4.output 0.14466690 574.40985942 + ------------------------------------------------------------------------------------- + TOTAL 0.07856402 240.27001175 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 114380 +BPFP 1.1184 bits/point +EBPFP 2.2368 equivalent bits/point +MSE 240.270012 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 240.2700 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 112, 128) +Output shape: (1, 112, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.output: torch.Size([1, 112, 3584]) -> torch.Size([1, 1, 112, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,856B, BPFP=0.5379 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,648B, BPFP=2.0435 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,560B, BPFP=0.9152 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,100B, BPFP=1.9671 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,864B, BPFP=1.0971 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,584B, BPFP=1.8951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,608B, BPFP=1.0614 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,996B, BPFP=1.9526 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,440B, BPFP=1.5960 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,512B, BPFP=1.8850 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,952B, BPFP=0.5969 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17321212 53.40850830 + layer.0.v_cache 0.00001450 0.00697586 + layer.1.k_cache 0.08643453 4.48177283 + layer.1.v_cache 0.00000659 0.00288388 + layer.2.k_cache 0.00690480 0.72489371 + layer.2.v_cache 0.00001899 0.00712602 + layer.3.k_cache 0.01205242 3.69648388 + layer.3.v_cache 0.00001923 0.00847296 + layer.4.k_cache 0.00071999 0.16315968 + layer.4.v_cache 0.00004947 0.01680535 + layer.4.output 10.17892331 477.90884088 + ------------------------------------------------------------------------------------- + TOTAL 4.20775858 200.46346874 + (elements=974,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 974848 +Total Bytes 137120 +BPFP 1.1253 bits/point +EBPFP 2.2505 equivalent bits/point +MSE 200.463469 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.006s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 200.4635 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,372B, BPFP=0.5987 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,108B, BPFP=2.1499 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,660B, BPFP=1.0050 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,644B, BPFP=2.0675 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,572B, BPFP=1.1669 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,224B, BPFP=1.9929 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,260B, BPFP=1.1115 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,592B, BPFP=2.0582 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,636B, BPFP=1.7109 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,152B, BPFP=1.9801 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,052B, BPFP=0.5594 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11687154 57.64905895 + layer.0.v_cache 0.00001423 0.00749251 + layer.1.k_cache 0.05324238 5.24545392 + layer.1.v_cache 0.00000539 0.00276507 + layer.2.k_cache 0.00397935 0.93397401 + layer.2.v_cache 0.00001766 0.00724199 + layer.3.k_cache 0.01210797 3.31524381 + layer.3.v_cache 0.00001803 0.00877287 + layer.4.k_cache 0.00075037 0.17874791 + layer.4.v_cache 0.00004481 0.01728913 + layer.4.output 0.15448949 613.81899351 + ------------------------------------------------------------------------------------- + TOTAL 0.07461636 256.71170557 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 111272 +BPFP 1.1622 bits/point +EBPFP 2.3244 equivalent bits/point +MSE 256.711706 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 256.7117 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,148B, BPFP=0.6073 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,624B, BPFP=2.2423 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,292B, BPFP=1.0208 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,172B, BPFP=2.1551 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,324B, BPFP=1.2199 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,736B, BPFP=2.0710 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,044B, BPFP=1.1659 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,860B, BPFP=2.0949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,960B, BPFP=1.7284 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,524B, BPFP=2.0301 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,248B, BPFP=0.6407 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203883 60.17366536 + layer.0.v_cache 0.00001438 0.00799984 + layer.1.k_cache 0.05633916 5.29820873 + layer.1.v_cache 0.00000602 0.00333524 + layer.2.k_cache 0.00244657 0.87570285 + layer.2.v_cache 0.00001847 0.00843944 + layer.3.k_cache 0.05492775 3.78336965 + layer.3.v_cache 0.00001883 0.00957282 + layer.4.k_cache 0.00062984 0.18642969 + layer.4.v_cache 0.00005178 0.01942223 + layer.4.output 0.18538215 665.59325397 + ------------------------------------------------------------------------------------- + TOTAL 0.09142157 278.20699551 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 107932 +BPFP 1.2247 bits/point +EBPFP 2.4494 equivalent bits/point +MSE 278.206996 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 278.2070 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,388B, BPFP=0.5882 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,240B, BPFP=2.1250 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,616B, BPFP=0.9750 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,504B, BPFP=1.9972 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,576B, BPFP=1.1417 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,092B, BPFP=1.9257 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,344B, BPFP=1.1014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,588B, BPFP=2.0118 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,620B, BPFP=1.6701 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,248B, BPFP=1.9528 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,512B, BPFP=0.5831 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13028670 59.69691298 + layer.0.v_cache 0.00001462 0.00699338 + layer.1.k_cache 0.04946813 4.69083286 + layer.1.v_cache 0.00000624 0.00239617 + layer.2.k_cache 0.00411291 0.78542616 + layer.2.v_cache 0.00001729 0.00670396 + layer.3.k_cache 0.01793825 4.10310330 + layer.3.v_cache 0.00001895 0.00842619 + layer.4.k_cache 0.00077120 0.17594982 + layer.4.v_cache 0.00004532 0.01661639 + layer.4.output 0.15109633 598.15029762 + ------------------------------------------------------------------------------------- + TOTAL 0.07413847 250.38502615 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 112728 +BPFP 1.1512 bits/point +EBPFP 2.3025 equivalent bits/point +MSE 250.385026 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 250.3850 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,288B, BPFP=0.6272 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,980B, BPFP=2.1875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,776B, BPFP=1.0351 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,720B, BPFP=2.1162 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,472B, BPFP=1.2259 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,400B, BPFP=2.0285 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,324B, BPFP=1.1853 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,516B, BPFP=2.0603 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,176B, BPFP=1.6930 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,204B, BPFP=1.9748 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,928B, BPFP=0.6629 +⌛️ [2/4] FRONTEND: Frontend time: 0.195s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.153s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11683119 66.69361637 + layer.0.v_cache 0.00001514 0.00964079 + layer.1.k_cache 0.01218004 4.94089602 + layer.1.v_cache 0.00000521 0.00331347 + layer.2.k_cache 0.00491684 0.90991077 + layer.2.v_cache 0.00001706 0.00861618 + layer.3.k_cache 0.10823302 4.38749668 + layer.3.v_cache 0.00002043 0.01094910 + layer.4.k_cache 0.00062874 0.20694258 + layer.4.v_cache 0.00004794 0.02111466 + layer.4.output 0.23833110 943.53344298 + ------------------------------------------------------------------------------------- + TOTAL 0.11242431 393.05450573 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 75784 +BPFP 1.2220 bits/point +EBPFP 2.4440 equivalent bits/point +MSE 393.054506 +---------------------- -------------------------------------------------------- +Time: 0.351s Load: 0.003s, Pack+Encode: 0.195s, Decode+Unpack: 0.153s +---------------------- -------------------------------------------------------- +💾 Converting with 393.0545 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,144B, BPFP=0.6700 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,316B, BPFP=2.2862 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,456B, BPFP=1.0800 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,944B, BPFP=2.1700 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,076B, BPFP=1.2737 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,716B, BPFP=2.0987 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,960B, BPFP=1.2375 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,996B, BPFP=2.1862 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,684B, BPFP=1.7763 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,696B, BPFP=2.0925 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,144B, BPFP=0.6761 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14598725 63.80286133 + layer.0.v_cache 0.00001449 0.00954952 + layer.1.k_cache 0.01347294 6.07656433 + layer.1.v_cache 0.00000521 0.00343541 + layer.2.k_cache 0.00488209 0.91817078 + layer.2.v_cache 0.00001649 0.00873174 + layer.3.k_cache 0.04421538 3.86170410 + layer.3.v_cache 0.00001905 0.01087928 + layer.4.k_cache 0.00073779 0.21706345 + layer.4.v_cache 0.00005089 0.02037295 + layer.4.output 0.27158585 1078.42982143 + ------------------------------------------------------------------------------------- + TOTAL 0.12414721 448.46694605 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 69132 +BPFP 1.2708 bits/point +EBPFP 2.5416 equivalent bits/point +MSE 448.466946 +---------------------- -------------------------------------------------------- +Time: 0.281s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 448.4669 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 46, 128) +Output shape: (1, 46, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.output: torch.Size([1, 46, 3584]) -> torch.Size([1, 1, 46, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,068B, BPFP=0.7024 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,124B, BPFP=2.4198 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,428B, BPFP=1.1644 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,864B, BPFP=2.3315 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,956B, BPFP=1.3438 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,532B, BPFP=2.2188 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,924B, BPFP=1.3329 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,860B, BPFP=2.3302 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,596B, BPFP=1.9008 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,652B, BPFP=2.2595 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,780B, BPFP=0.7172 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14129612 70.95508874 + layer.0.v_cache 0.00001532 0.00925671 + layer.1.k_cache 0.01315293 6.00766721 + layer.1.v_cache 0.00000553 0.00344033 + layer.2.k_cache 0.00779452 0.96424045 + layer.2.v_cache 0.00001723 0.00823899 + layer.3.k_cache 0.01828384 4.69048210 + layer.3.v_cache 0.00001944 0.01188395 + layer.4.k_cache 0.00060559 0.20487848 + layer.4.v_cache 0.00005113 0.02174283 + layer.4.output 0.29517000 1170.99136258 + ------------------------------------------------------------------------------------- + TOTAL 0.13220186 487.04802693 + (elements=400,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 400384 +Total Bytes 67784 +BPFP 1.3544 bits/point +EBPFP 2.7088 equivalent bits/point +MSE 487.048027 +---------------------- -------------------------------------------------------- +Time: 0.279s Load: 0.002s, Pack+Encode: 0.131s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 487.0480 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 110, 128) +Output shape: (1, 110, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.output: torch.Size([1, 110, 3584]) -> torch.Size([1, 1, 110, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,788B, BPFP=0.5381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,688B, BPFP=2.0864 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,508B, BPFP=0.9244 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,932B, BPFP=1.9790 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,768B, BPFP=1.1034 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,512B, BPFP=1.9193 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,504B, BPFP=1.0659 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,920B, BPFP=1.9773 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,568B, BPFP=1.6432 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,260B, BPFP=1.8835 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,832B, BPFP=0.5851 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14476244 54.33098366 + layer.0.v_cache 0.00001476 0.00749394 + layer.1.k_cache 0.05665370 4.62185475 + layer.1.v_cache 0.00000525 0.00250721 + layer.2.k_cache 0.00786287 0.76211423 + layer.2.v_cache 0.00001801 0.00678186 + layer.3.k_cache 0.01410385 3.48608121 + layer.3.v_cache 0.00002027 0.00853193 + layer.4.k_cache 0.00066873 0.16315301 + layer.4.v_cache 0.00004626 0.01512577 + layer.4.output 10.36401748 486.02609578 + ------------------------------------------------------------------------------------- + TOTAL 4.28072226 203.85807635 + (elements=957,440) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 957440 +Total Bytes 135280 +BPFP 1.1303 bits/point +EBPFP 2.2607 equivalent bits/point +MSE 203.858076 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 203.8581 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,196B, BPFP=0.6165 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,968B, BPFP=2.3086 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,404B, BPFP=1.0424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,584B, BPFP=2.2346 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,696B, BPFP=1.2917 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,272B, BPFP=2.1744 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,324B, BPFP=1.2199 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,564B, BPFP=2.2307 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,644B, BPFP=1.8603 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,156B, BPFP=2.1520 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,672B, BPFP=0.7901 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10533771 58.72148679 + layer.0.v_cache 0.00001610 0.00864897 + layer.1.k_cache 0.05553475 4.98670074 + layer.1.v_cache 0.00000748 0.00330600 + layer.2.k_cache 0.02166171 0.87291567 + layer.2.v_cache 0.00002000 0.00848770 + layer.3.k_cache 0.06386927 3.92909599 + layer.3.v_cache 0.00001944 0.00992500 + layer.4.k_cache 0.00064315 0.19227659 + layer.4.v_cache 0.00005385 0.01918507 + layer.4.output 0.16801396 667.12830688 + ------------------------------------------------------------------------------------- + TOTAL 0.08372125 278.74412804 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 117480 +BPFP 1.3331 bits/point +EBPFP 2.6661 equivalent bits/point +MSE 278.744128 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 278.7441 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 49, 128) +Output shape: (1, 49, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.output: torch.Size([1, 49, 3584]) -> torch.Size([1, 1, 49, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,120B, BPFP=0.6760 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,312B, BPFP=2.3316 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,456B, BPFP=1.1020 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,016B, BPFP=2.2372 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,164B, BPFP=1.3278 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,712B, BPFP=2.1403 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,924B, BPFP=1.2513 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,008B, BPFP=2.2347 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,756B, BPFP=1.8355 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,784B, BPFP=2.1633 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,000B, BPFP=0.7744 +⌛️ [2/4] FRONTEND: Frontend time: 0.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12410851 65.40717674 + layer.0.v_cache 0.00001740 0.00948796 + layer.1.k_cache 0.01401279 5.72080028 + layer.1.v_cache 0.00000579 0.00363680 + layer.2.k_cache 0.00545398 1.05842474 + layer.2.v_cache 0.00001955 0.00958120 + layer.3.k_cache 0.07546502 4.09191614 + layer.3.v_cache 0.00001978 0.01130844 + layer.4.k_cache 0.00063206 0.21276536 + layer.4.v_cache 0.00005205 0.02018542 + layer.4.output 0.27729562 1099.67374271 + ------------------------------------------------------------------------------------- + TOTAL 0.12710919 457.30949894 + (elements=426,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 426496 +Total Bytes 71252 +BPFP 1.3365 bits/point +EBPFP 2.6730 equivalent bits/point +MSE 457.309499 +---------------------- -------------------------------------------------------- +Time: 0.281s Load: 0.003s, Pack+Encode: 0.132s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 457.3095 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,304B, BPFP=0.6316 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,728B, BPFP=2.1184 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,736B, BPFP=1.0241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,420B, BPFP=2.0340 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,380B, BPFP=1.2007 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,128B, BPFP=1.9539 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,216B, BPFP=1.1557 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,356B, BPFP=2.0164 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,992B, BPFP=1.6425 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,096B, BPFP=1.9452 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,180B, BPFP=0.6728 +⌛️ [2/4] FRONTEND: Frontend time: 0.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19197876 64.67635091 + layer.0.v_cache 0.00001449 0.00908410 + layer.1.k_cache 0.01466701 5.40163007 + layer.1.v_cache 0.00000519 0.00332932 + layer.2.k_cache 0.00720649 0.86707681 + layer.2.v_cache 0.00001785 0.00855048 + layer.3.k_cache 0.10851213 4.03600646 + layer.3.v_cache 0.00001996 0.01043494 + layer.4.k_cache 0.00062085 0.19538659 + layer.4.v_cache 0.00004770 0.01948945 + layer.4.output 0.23832821 943.08771930 + ------------------------------------------------------------------------------------- + TOTAL 0.11714047 392.75537495 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 74536 +BPFP 1.2019 bits/point +EBPFP 2.4038 equivalent bits/point +MSE 392.755375 +---------------------- -------------------------------------------------------- +Time: 0.280s Load: 0.002s, Pack+Encode: 0.132s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 392.7554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,160B, BPFP=0.6618 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,332B, BPFP=2.2463 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,460B, BPFP=1.0600 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,040B, BPFP=2.1569 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,096B, BPFP=1.2549 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,820B, BPFP=2.0895 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,956B, BPFP=1.2120 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,060B, BPFP=2.1630 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,788B, BPFP=1.7733 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,768B, BPFP=2.0735 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,844B, BPFP=0.6935 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.147s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12665685 64.33449180 + layer.0.v_cache 0.00001436 0.00906775 + layer.1.k_cache 0.01497617 5.91475782 + layer.1.v_cache 0.00000537 0.00323389 + layer.2.k_cache 0.00986420 1.00531185 + layer.2.v_cache 0.00001802 0.00923034 + layer.3.k_cache 0.09846177 4.46714663 + layer.3.v_cache 0.00001923 0.01129131 + layer.4.k_cache 0.00062397 0.20507005 + layer.4.v_cache 0.00005014 0.02065721 + layer.4.output 0.26631049 1058.65896359 + ------------------------------------------------------------------------------------- + TOTAL 0.12440374 440.38782375 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 70324 +BPFP 1.2674 bits/point +EBPFP 2.5347 equivalent bits/point +MSE 440.387824 +---------------------- -------------------------------------------------------- +Time: 0.280s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.147s +---------------------- -------------------------------------------------------- +💾 Converting with 440.3878 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,116B, BPFP=0.6888 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,428B, BPFP=2.4180 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,492B, BPFP=1.1367 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,164B, BPFP=2.3320 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,240B, BPFP=1.3802 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,836B, BPFP=2.2253 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,020B, BPFP=1.3086 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,124B, BPFP=2.3190 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,836B, BPFP=1.8997 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,884B, BPFP=2.2409 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,412B, BPFP=0.8097 +⌛️ [2/4] FRONTEND: Frontend time: 0.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.148s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18531116 66.37510173 + layer.0.v_cache 0.00001874 0.01060284 + layer.1.k_cache 0.01459585 6.11308225 + layer.1.v_cache 0.00000674 0.00430649 + layer.2.k_cache 0.01042949 0.94651437 + layer.2.v_cache 0.00001841 0.01063417 + layer.3.k_cache 0.15260293 4.07812214 + layer.3.v_cache 0.00002390 0.01373320 + layer.4.k_cache 0.00061422 0.22780724 + layer.4.v_cache 0.00004881 0.02364132 + layer.4.output 0.28297247 1126.49665179 + ------------------------------------------------------------------------------------- + TOTAL 0.13791044 468.42824166 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 72552 +BPFP 1.3892 bits/point +EBPFP 2.7785 equivalent bits/point +MSE 468.428242 +---------------------- -------------------------------------------------------- +Time: 0.283s Load: 0.003s, Pack+Encode: 0.132s, Decode+Unpack: 0.148s +---------------------- -------------------------------------------------------- +💾 Converting with 468.4282 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,392B, BPFP=0.6335 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,036B, BPFP=2.1282 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,848B, BPFP=1.0191 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,728B, BPFP=2.0466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,688B, BPFP=1.2415 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,536B, BPFP=1.9958 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,536B, BPFP=1.2013 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,824B, BPFP=2.0720 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,456B, BPFP=1.7097 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,444B, BPFP=1.9714 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,896B, BPFP=0.7527 +⌛️ [2/4] FRONTEND: Frontend time: 0.133s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.150s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12052289 64.72414592 + layer.0.v_cache 0.00001546 0.00936204 + layer.1.k_cache 0.01434235 4.80225644 + layer.1.v_cache 0.00000574 0.00357886 + layer.2.k_cache 0.00254285 0.80262672 + layer.2.v_cache 0.00001710 0.00901350 + layer.3.k_cache 0.03737745 3.79433894 + layer.3.v_cache 0.00001946 0.01207421 + layer.4.k_cache 0.00063766 0.20495103 + layer.4.v_cache 0.00005039 0.02113062 + layer.4.output 0.23032015 910.74250908 + ------------------------------------------------------------------------------------- + TOTAL 0.10516308 379.38712011 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 80384 +BPFP 1.2522 bits/point +EBPFP 2.5045 equivalent bits/point +MSE 379.387120 +---------------------- -------------------------------------------------------- +Time: 0.286s Load: 0.003s, Pack+Encode: 0.133s, Decode+Unpack: 0.150s +---------------------- -------------------------------------------------------- +💾 Converting with 379.3871 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,008B, BPFP=0.6351 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,920B, BPFP=2.3057 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,924B, BPFP=1.0397 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,336B, BPFP=2.1824 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,036B, BPFP=1.2745 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,036B, BPFP=2.1191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,644B, BPFP=1.1917 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,448B, BPFP=2.2061 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,376B, BPFP=1.7686 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,956B, BPFP=2.1022 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,628B, BPFP=0.6826 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11335979 61.30646247 + layer.0.v_cache 0.00001513 0.00920384 + layer.1.k_cache 0.01655370 5.59175192 + layer.1.v_cache 0.00000667 0.00350061 + layer.2.k_cache 0.00428703 0.84459078 + layer.2.v_cache 0.00001789 0.00906703 + layer.3.k_cache 0.04938591 3.27468542 + layer.3.v_cache 0.00002111 0.01120172 + layer.4.k_cache 0.00066117 0.21428375 + layer.4.v_cache 0.00004716 0.02063963 + layer.4.output 0.18372514 729.77183880 + ------------------------------------------------------------------------------------- + TOTAL 0.08649597 304.68754463 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 102312 +BPFP 1.2708 bits/point +EBPFP 2.5415 equivalent bits/point +MSE 304.687545 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 304.6875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,004B, BPFP=0.7116 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,992B, BPFP=2.4830 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,340B, BPFP=1.1861 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,684B, BPFP=2.3736 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,980B, BPFP=1.4134 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,492B, BPFP=2.3054 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,792B, BPFP=1.3466 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,748B, BPFP=2.3963 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,600B, BPFP=1.9886 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,548B, BPFP=2.3253 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,964B, BPFP=0.8606 +⌛️ [2/4] FRONTEND: Frontend time: 0.133s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.148s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14777968 69.69672186 + layer.0.v_cache 0.00001531 0.01089318 + layer.1.k_cache 0.01728138 6.41690202 + layer.1.v_cache 0.00000547 0.00376341 + layer.2.k_cache 0.00823456 0.95293791 + layer.2.v_cache 0.00001932 0.01076140 + layer.3.k_cache 0.16770924 4.69397874 + layer.3.v_cache 0.00002046 0.01267652 + layer.4.k_cache 0.00061569 0.21636727 + layer.4.v_cache 0.00005592 0.02483212 + layer.4.output 0.30867824 1227.35836039 + ------------------------------------------------------------------------------------- + TOTAL 0.14720498 510.20872689 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 69144 +BPFP 1.4444 bits/point +EBPFP 2.8887 equivalent bits/point +MSE 510.208727 +---------------------- -------------------------------------------------------- +Time: 0.284s Load: 0.003s, Pack+Encode: 0.133s, Decode+Unpack: 0.148s +---------------------- -------------------------------------------------------- +💾 Converting with 510.2087 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,116B, BPFP=0.6242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,640B, BPFP=2.3317 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,220B, BPFP=1.0457 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,984B, BPFP=2.2003 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,404B, BPFP=1.2829 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,684B, BPFP=2.1402 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,116B, BPFP=1.2252 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,092B, BPFP=2.2220 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,112B, BPFP=1.8253 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,700B, BPFP=2.1434 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,080B, BPFP=0.6605 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14408422 58.03743490 + layer.0.v_cache 0.00002097 0.01008642 + layer.1.k_cache 0.05735342 5.34100459 + layer.1.v_cache 0.00000575 0.00383575 + layer.2.k_cache 0.01678689 0.94019181 + layer.2.v_cache 0.00001894 0.00993260 + layer.3.k_cache 0.03029843 3.79941344 + layer.3.v_cache 0.00002125 0.01228004 + layer.4.k_cache 0.00062582 0.21655046 + layer.4.v_cache 0.00005050 0.02282329 + layer.4.output 0.17433185 690.13890797 + ------------------------------------------------------------------------------------- + TOTAL 0.08644642 288.19799465 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 108148 +BPFP 1.2744 bits/point +EBPFP 2.5487 equivalent bits/point +MSE 288.197995 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 288.1980 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,020B, BPFP=0.7173 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,892B, BPFP=2.4474 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,332B, BPFP=1.1832 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,552B, BPFP=2.3267 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,852B, BPFP=1.3679 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,500B, BPFP=2.3082 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,648B, BPFP=1.2955 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,640B, BPFP=2.3580 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,420B, BPFP=1.9247 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,380B, BPFP=2.2656 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,448B, BPFP=0.7330 +⌛️ [2/4] FRONTEND: Frontend time: 0.135s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.147s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13243668 66.46248002 + layer.0.v_cache 0.00001576 0.00975651 + layer.1.k_cache 0.01981235 6.54646856 + layer.1.v_cache 0.00000577 0.00377036 + layer.2.k_cache 0.01756107 1.01184819 + layer.2.v_cache 0.00001728 0.01005413 + layer.3.k_cache 0.04920139 4.14295821 + layer.3.v_cache 0.00001837 0.01159996 + layer.4.k_cache 0.00061593 0.22534609 + layer.4.v_cache 0.00004907 0.02358376 + layer.4.output 0.30854983 1224.71225649 + ------------------------------------------------------------------------------------- + TOTAL 0.13997544 508.90786243 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 65684 +BPFP 1.3721 bits/point +EBPFP 2.7442 equivalent bits/point +MSE 508.907862 +---------------------- -------------------------------------------------------- +Time: 0.285s Load: 0.003s, Pack+Encode: 0.135s, Decode+Unpack: 0.147s +---------------------- -------------------------------------------------------- +💾 Converting with 508.9079 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,148B, BPFP=0.6713 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,324B, BPFP=2.2887 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,492B, BPFP=1.0913 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,032B, BPFP=2.1975 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,212B, BPFP=1.3162 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,916B, BPFP=2.1612 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,024B, BPFP=1.2575 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,100B, BPFP=2.2188 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,812B, BPFP=1.8162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,832B, BPFP=2.1350 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,064B, BPFP=0.7171 +⌛️ [2/4] FRONTEND: Frontend time: 0.133s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.148s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20666794 65.06666016 + layer.0.v_cache 0.00002565 0.01036829 + layer.1.k_cache 0.01518770 6.02848755 + layer.1.v_cache 0.00000581 0.00367354 + layer.2.k_cache 0.01031687 0.98464737 + layer.2.v_cache 0.00001873 0.01001189 + layer.3.k_cache 0.09638391 4.57136292 + layer.3.v_cache 0.00001967 0.01261373 + layer.4.k_cache 0.00062172 0.21322844 + layer.4.v_cache 0.00004827 0.02276895 + layer.4.output 0.27162739 1079.05428571 + ------------------------------------------------------------------------------------- + TOTAL 0.13121694 448.84140134 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 70956 +BPFP 1.3043 bits/point +EBPFP 2.6087 equivalent bits/point +MSE 448.841401 +---------------------- -------------------------------------------------------- +Time: 0.284s Load: 0.003s, Pack+Encode: 0.133s, Decode+Unpack: 0.148s +---------------------- -------------------------------------------------------- +💾 Converting with 448.8414 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,824B, BPFP=0.6489 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,140B, BPFP=2.3300 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,688B, BPFP=1.0772 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,632B, BPFP=2.2132 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,620B, BPFP=1.2914 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,176B, BPFP=2.1085 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,308B, BPFP=1.2197 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,596B, BPFP=2.2050 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,764B, BPFP=1.7840 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,324B, BPFP=2.1425 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,688B, BPFP=0.7447 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13431002 61.52396168 + layer.0.v_cache 0.00001720 0.00992800 + layer.1.k_cache 0.01179634 5.92387659 + layer.1.v_cache 0.00000537 0.00344272 + layer.2.k_cache 0.00798416 0.87473550 + layer.2.v_cache 0.00001783 0.00934683 + layer.3.k_cache 0.05795463 3.91574007 + layer.3.v_cache 0.00001974 0.01123021 + layer.4.k_cache 0.00062059 0.20671629 + layer.4.v_cache 0.00005204 0.02154125 + layer.4.output 0.19989206 795.50932248 + ------------------------------------------------------------------------------------- + TOTAL 0.09482484 331.82739862 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 96760 +BPFP 1.3079 bits/point +EBPFP 2.6157 equivalent bits/point +MSE 331.827399 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.004s, Pack+Encode: 0.161s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 331.8274 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,144B, BPFP=0.6700 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,368B, BPFP=2.3025 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,480B, BPFP=1.0875 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,060B, BPFP=2.2062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,252B, BPFP=1.3288 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,896B, BPFP=2.1550 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,024B, BPFP=1.2575 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,116B, BPFP=2.2237 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,896B, BPFP=1.8425 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,900B, BPFP=2.1562 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,976B, BPFP=0.7132 +⌛️ [2/4] FRONTEND: Frontend time: 0.133s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.148s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15467889 62.99649902 + layer.0.v_cache 0.00001741 0.01021006 + layer.1.k_cache 0.01620592 6.27350464 + layer.1.v_cache 0.00000548 0.00352744 + layer.2.k_cache 0.00723796 0.91950836 + layer.2.v_cache 0.00001737 0.00965427 + layer.3.k_cache 0.06661982 4.06319214 + layer.3.v_cache 0.00002040 0.01187014 + layer.4.k_cache 0.00058776 0.21339977 + layer.4.v_cache 0.00004656 0.02208894 + layer.4.output 0.27162074 1079.63339286 + ------------------------------------------------------------------------------------- + TOTAL 0.12628134 448.93865911 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 71112 +BPFP 1.3072 bits/point +EBPFP 2.6144 equivalent bits/point +MSE 448.938659 +---------------------- -------------------------------------------------------- +Time: 0.284s Load: 0.003s, Pack+Encode: 0.133s, Decode+Unpack: 0.148s +---------------------- -------------------------------------------------------- +💾 Converting with 448.9387 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,184B, BPFP=0.6691 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,380B, BPFP=2.2610 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,456B, BPFP=1.0588 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,052B, BPFP=2.1605 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,032B, BPFP=1.2353 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,772B, BPFP=2.0748 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,908B, BPFP=1.1973 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,000B, BPFP=2.1446 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,696B, BPFP=1.7451 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,772B, BPFP=2.0748 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,028B, BPFP=0.6577 +⌛️ [2/4] FRONTEND: Frontend time: 0.134s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.148s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11261671 64.56695198 + layer.0.v_cache 0.00001829 0.00931231 + layer.1.k_cache 0.01190703 5.68520161 + layer.1.v_cache 0.00000528 0.00335548 + layer.2.k_cache 0.00973116 0.95476031 + layer.2.v_cache 0.00001668 0.00876629 + layer.3.k_cache 0.04200072 4.22424915 + layer.3.v_cache 0.00001874 0.01168990 + layer.4.k_cache 0.00062040 0.20545743 + layer.4.v_cache 0.00004776 0.02154321 + layer.4.output 0.26626884 1058.90266106 + ------------------------------------------------------------------------------------- + TOTAL 0.12005086 440.47117148 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 69280 +BPFP 1.2486 bits/point +EBPFP 2.4971 equivalent bits/point +MSE 440.471171 +---------------------- -------------------------------------------------------- +Time: 0.284s Load: 0.002s, Pack+Encode: 0.134s, Decode+Unpack: 0.148s +---------------------- -------------------------------------------------------- +💾 Converting with 440.4712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,188B, BPFP=0.6575 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,624B, BPFP=2.2909 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,560B, BPFP=1.0697 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,288B, BPFP=2.1899 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,332B, BPFP=1.3017 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,088B, BPFP=2.1298 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,152B, BPFP=1.2476 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,288B, BPFP=2.1899 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,976B, BPFP=1.7957 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,032B, BPFP=2.1130 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,832B, BPFP=0.7655 +⌛️ [2/4] FRONTEND: Frontend time: 0.135s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.149s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12932784 58.85035119 + layer.0.v_cache 0.00001491 0.01021397 + layer.1.k_cache 0.01488123 6.10690014 + layer.1.v_cache 0.00000563 0.00374831 + layer.2.k_cache 0.00939683 0.90441792 + layer.2.v_cache 0.00001981 0.01015177 + layer.3.k_cache 0.07051506 4.22477546 + layer.3.v_cache 0.00001927 0.01216748 + layer.4.k_cache 0.00064413 0.22206248 + layer.4.v_cache 0.00005939 0.02313292 + layer.4.output 0.26127321 1038.32718063 + ------------------------------------------------------------------------------------- + TOTAL 0.12081156 431.68577565 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 74360 +BPFP 1.3143 bits/point +EBPFP 2.6287 equivalent bits/point +MSE 431.685776 +---------------------- -------------------------------------------------------- +Time: 0.287s Load: 0.003s, Pack+Encode: 0.135s, Decode+Unpack: 0.149s +---------------------- -------------------------------------------------------- +💾 Converting with 431.6858 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,188B, BPFP=0.6575 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,616B, BPFP=2.2885 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,596B, BPFP=1.0805 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,316B, BPFP=2.1983 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,308B, BPFP=1.2945 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,048B, BPFP=2.1178 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,140B, BPFP=1.2440 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,320B, BPFP=2.1995 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,004B, BPFP=1.8041 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,004B, BPFP=2.1046 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,476B, BPFP=0.7502 +⌛️ [2/4] FRONTEND: Frontend time: 0.138s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.147s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13190371 58.01374230 + layer.0.v_cache 0.00001552 0.01031771 + layer.1.k_cache 0.01668521 6.35531910 + layer.1.v_cache 0.00000581 0.00395631 + layer.2.k_cache 0.00490881 0.92169439 + layer.2.v_cache 0.00001873 0.01000164 + layer.3.k_cache 0.11824917 3.83307794 + layer.3.v_cache 0.00002115 0.01209947 + layer.4.k_cache 0.00061206 0.21558883 + layer.4.v_cache 0.00004844 0.02075671 + layer.4.output 0.26122982 1038.55030907 + ------------------------------------------------------------------------------------- + TOTAL 0.12359279 431.72051281 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 74016 +BPFP 1.3083 bits/point +EBPFP 2.6165 equivalent bits/point +MSE 431.720513 +---------------------- -------------------------------------------------------- +Time: 0.287s Load: 0.002s, Pack+Encode: 0.138s, Decode+Unpack: 0.147s +---------------------- -------------------------------------------------------- +💾 Converting with 431.7205 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,288B, BPFP=0.5584 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,768B, BPFP=2.1685 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,896B, BPFP=1.0014 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,156B, BPFP=2.0645 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,912B, BPFP=1.1739 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,720B, BPFP=1.9905 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,612B, BPFP=1.1230 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,204B, BPFP=2.0727 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,948B, BPFP=1.6895 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,684B, BPFP=1.9844 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,476B, BPFP=0.6181 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10326390 62.44422979 + layer.0.v_cache 0.00001541 0.00798430 + layer.1.k_cache 0.04976616 4.98468449 + layer.1.v_cache 0.00000534 0.00259766 + layer.2.k_cache 0.01168035 0.73533829 + layer.2.v_cache 0.00001798 0.00749692 + layer.3.k_cache 0.02907468 3.48684825 + layer.3.v_cache 0.00001946 0.00916316 + layer.4.k_cache 0.00073861 0.17502407 + layer.4.v_cache 0.00005077 0.01698166 + layer.4.output 0.14788401 586.37257376 + ------------------------------------------------------------------------------------- + TOTAL 0.07234240 245.67519794 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 118664 +BPFP 1.1855 bits/point +EBPFP 2.3710 equivalent bits/point +MSE 245.675198 +---------------------- -------------------------------------------------------- +Time: 0.390s Load: 0.003s, Pack+Encode: 0.168s, Decode+Unpack: 0.219s +---------------------- -------------------------------------------------------- +💾 Converting with 245.6752 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 64, 128) +Output shape: (1, 64, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.output: torch.Size([1, 64, 3584]) -> torch.Size([1, 1, 64, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,776B, BPFP=0.4336 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,680B, BPFP=1.8750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,292B, BPFP=0.8037 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,412B, BPFP=1.8096 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,172B, BPFP=1.0186 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,216B, BPFP=1.7617 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,960B, BPFP=0.9668 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,372B, BPFP=1.7998 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,144B, BPFP=1.5000 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,084B, BPFP=1.7295 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,796B, BPFP=0.6207 +⌛️ [2/4] FRONTEND: Frontend time: 0.135s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.149s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09722494 44.64468384 + layer.0.v_cache 0.00001391 0.00761832 + layer.1.k_cache 0.01806639 3.68119168 + layer.1.v_cache 0.00000557 0.00312444 + layer.2.k_cache 0.00423524 0.66070575 + layer.2.v_cache 0.00001868 0.00823375 + layer.3.k_cache 0.03759564 2.99431872 + layer.3.v_cache 0.00001825 0.00899295 + layer.4.k_cache 0.00061753 0.16751052 + layer.4.v_cache 0.00004894 0.01851678 + layer.4.output 0.21408452 752.59863281 + ------------------------------------------------------------------------------------- + TOTAL 0.09743746 312.96384273 + (elements=557,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 557056 +Total Bytes 73904 +BPFP 1.0614 bits/point +EBPFP 2.1227 equivalent bits/point +MSE 312.963843 +---------------------- -------------------------------------------------------- +Time: 0.287s Load: 0.004s, Pack+Encode: 0.135s, Decode+Unpack: 0.149s +---------------------- -------------------------------------------------------- +💾 Converting with 312.9638 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,548B, BPFP=0.5435 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,744B, BPFP=2.1054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,208B, BPFP=0.9510 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,112B, BPFP=2.0086 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,360B, BPFP=1.1275 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,548B, BPFP=1.9222 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,036B, BPFP=1.0778 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,016B, BPFP=1.9939 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,728B, BPFP=1.6434 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,600B, BPFP=1.9301 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,592B, BPFP=0.5600 +⌛️ [2/4] FRONTEND: Frontend time: 0.170s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.216s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11648470 58.97238817 + layer.0.v_cache 0.00001454 0.00727661 + layer.1.k_cache 0.02721847 4.62551940 + layer.1.v_cache 0.00000531 0.00262099 + layer.2.k_cache 0.01045542 0.83259104 + layer.2.v_cache 0.00001699 0.00673856 + layer.3.k_cache 0.01744144 3.65419934 + layer.3.v_cache 0.00001839 0.00827043 + layer.4.k_cache 0.00076342 0.17353963 + layer.4.v_cache 0.00004795 0.01687337 + layer.4.output 11.17676328 525.57506127 + ------------------------------------------------------------------------------------- + TOTAL 4.61234174 220.43090861 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 125492 +BPFP 1.1308 bits/point +EBPFP 2.2616 equivalent bits/point +MSE 220.430909 +---------------------- -------------------------------------------------------- +Time: 0.392s Load: 0.005s, Pack+Encode: 0.170s, Decode+Unpack: 0.216s +---------------------- -------------------------------------------------------- +💾 Converting with 220.4309 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,172B, BPFP=0.6195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,092B, BPFP=2.1664 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,156B, BPFP=1.0070 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,512B, BPFP=2.0531 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,036B, BPFP=1.1789 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,064B, BPFP=1.9656 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,804B, BPFP=1.1336 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,604B, BPFP=2.0711 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,768B, BPFP=1.7125 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,196B, BPFP=1.9914 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,484B, BPFP=0.5994 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11433966 59.43680420 + layer.0.v_cache 0.00001437 0.00729981 + layer.1.k_cache 0.05315235 5.64149361 + layer.1.v_cache 0.00000524 0.00244252 + layer.2.k_cache 0.00678293 0.77012367 + layer.2.v_cache 0.00001704 0.00694491 + layer.3.k_cache 0.01951189 3.49149513 + layer.3.v_cache 0.00001758 0.00831441 + layer.4.k_cache 0.00076624 0.17337253 + layer.4.v_cache 0.00004856 0.01650907 + layer.4.output 0.16992235 675.79073661 + ------------------------------------------------------------------------------------- + TOTAL 0.08141837 282.35823271 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 102888 +BPFP 1.1821 bits/point +EBPFP 2.3642 equivalent bits/point +MSE 282.358233 +---------------------- -------------------------------------------------------- +Time: 0.372s Load: 0.003s, Pack+Encode: 0.164s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 282.3582 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,392B, BPFP=0.5889 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,340B, BPFP=2.1424 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,616B, BPFP=0.9750 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,828B, BPFP=2.0535 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,712B, BPFP=1.1653 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,260B, BPFP=1.9549 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,500B, BPFP=1.1285 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,804B, BPFP=2.0493 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,884B, BPFP=1.7160 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,392B, BPFP=1.9778 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,844B, BPFP=0.6162 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11570581 61.54861654 + layer.0.v_cache 0.00001517 0.00745022 + layer.1.k_cache 0.03236232 4.59715712 + layer.1.v_cache 0.00000662 0.00260983 + layer.2.k_cache 0.00641356 0.79513872 + layer.2.v_cache 0.00001676 0.00688606 + layer.3.k_cache 0.02648412 3.69632433 + layer.3.v_cache 0.00001896 0.00871342 + layer.4.k_cache 0.00074524 0.18640527 + layer.4.v_cache 0.00004843 0.01746262 + layer.4.output 0.15113260 598.39945437 + ------------------------------------------------------------------------------------- + TOTAL 0.07292619 250.56840851 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 115572 +BPFP 1.1803 bits/point +EBPFP 2.3605 equivalent bits/point +MSE 250.568409 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.005s, Pack+Encode: 0.164s, Decode+Unpack: 0.210s +---------------------- -------------------------------------------------------- +💾 Converting with 250.5684 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 114, 128) +Output shape: (1, 114, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.output: torch.Size([1, 114, 3584]) -> torch.Size([1, 1, 114, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,840B, BPFP=0.5263 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,568B, BPFP=1.9967 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,444B, BPFP=0.8832 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,792B, BPFP=1.8904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,676B, BPFP=1.0521 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,340B, BPFP=1.8284 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,284B, BPFP=0.9984 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,752B, BPFP=1.8849 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,512B, BPFP=1.5779 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,272B, BPFP=1.8191 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,576B, BPFP=0.5595 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.218s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10081583 55.88046104 + layer.0.v_cache 0.00001633 0.00728684 + layer.1.k_cache 0.08473034 4.34851904 + layer.1.v_cache 0.00000563 0.00258094 + layer.2.k_cache 0.00792928 0.77109300 + layer.2.v_cache 0.00001730 0.00684641 + layer.3.k_cache 0.01019665 3.59011546 + layer.3.v_cache 0.00001810 0.00768621 + layer.4.k_cache 0.00077074 0.16660822 + layer.4.v_cache 0.00005022 0.01708940 + layer.4.output 10.00037814 469.33137531 + ------------------------------------------------------------------------------------- + TOTAL 4.12983514 197.06575963 + (elements=992,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 992256 +Total Bytes 134056 +BPFP 1.0808 bits/point +EBPFP 2.1616 equivalent bits/point +MSE 197.065760 +---------------------- -------------------------------------------------------- +Time: 0.387s Load: 0.006s, Pack+Encode: 0.164s, Decode+Unpack: 0.218s +---------------------- -------------------------------------------------------- +💾 Converting with 197.0658 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 98, 128) +Output shape: (1, 98, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.output: torch.Size([1, 98, 3584]) -> torch.Size([1, 1, 98, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,164B, BPFP=0.5045 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,916B, BPFP=2.0593 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,592B, BPFP=0.8916 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,372B, BPFP=1.9726 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,980B, BPFP=1.1129 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,912B, BPFP=1.8992 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,508B, BPFP=1.0376 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,264B, BPFP=1.9554 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,224B, BPFP=1.6301 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,900B, BPFP=1.8973 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,584B, BPFP=0.5372 +⌛️ [2/4] FRONTEND: Frontend time: 0.177s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.222s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11008905 58.63187580 + layer.0.v_cache 0.00001718 0.00730994 + layer.1.k_cache 0.06240322 5.45785896 + layer.1.v_cache 0.00000548 0.00264927 + layer.2.k_cache 0.00473963 0.80577259 + layer.2.v_cache 0.00001718 0.00720321 + layer.3.k_cache 0.01141277 3.20525874 + layer.3.v_cache 0.00001769 0.00813931 + layer.4.k_cache 0.00076610 0.17610357 + layer.4.v_cache 0.00004814 0.01725960 + layer.4.output 0.02426045 566.93467566 + ------------------------------------------------------------------------------------- + TOTAL 0.02113762 237.46248003 + (elements=852,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 852992 +Total Bytes 117416 +BPFP 1.1012 bits/point +EBPFP 2.2024 equivalent bits/point +MSE 237.462480 +---------------------- -------------------------------------------------------- +Time: 0.404s Load: 0.005s, Pack+Encode: 0.177s, Decode+Unpack: 0.222s +---------------------- -------------------------------------------------------- +💾 Converting with 237.4625 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,320B, BPFP=0.5639 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,592B, BPFP=2.1386 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,948B, BPFP=1.0102 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,968B, BPFP=2.0326 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,996B, BPFP=1.1882 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,684B, BPFP=1.9844 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,696B, BPFP=1.1372 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,036B, BPFP=2.0442 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,124B, BPFP=1.7194 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,632B, BPFP=1.9755 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,544B, BPFP=0.6198 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.206s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08919133 61.35388714 + layer.0.v_cache 0.00001414 0.00689718 + layer.1.k_cache 0.08591949 5.13268114 + layer.1.v_cache 0.00000528 0.00242557 + layer.2.k_cache 0.00389855 0.81781860 + layer.2.v_cache 0.00001738 0.00710717 + layer.3.k_cache 0.07679920 3.64815156 + layer.3.v_cache 0.00002028 0.00879247 + layer.4.k_cache 0.00072407 0.17620180 + layer.4.v_cache 0.00004779 0.01780792 + layer.4.output 0.14786680 586.18395769 + ------------------------------------------------------------------------------------- + TOTAL 0.07598265 245.55643967 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 118540 +BPFP 1.1843 bits/point +EBPFP 2.3685 equivalent bits/point +MSE 245.556440 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.005s, Pack+Encode: 0.162s, Decode+Unpack: 0.206s +---------------------- -------------------------------------------------------- +💾 Converting with 245.5564 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,700B, BPFP=0.5403 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,640B, BPFP=2.1379 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,444B, BPFP=0.9410 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,944B, BPFP=2.0362 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,920B, BPFP=1.1565 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,560B, BPFP=1.9801 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,472B, BPFP=1.0911 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,904B, BPFP=2.0304 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,520B, BPFP=1.6822 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,396B, BPFP=1.9562 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,228B, BPFP=0.5889 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.217s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11353527 55.72720137 + layer.0.v_cache 0.00001674 0.00725352 + layer.1.k_cache 0.04276301 4.32742994 + layer.1.v_cache 0.00000542 0.00242398 + layer.2.k_cache 0.00247095 0.77331721 + layer.2.v_cache 0.00001885 0.00707269 + layer.3.k_cache 0.02722364 3.35455607 + layer.3.v_cache 0.00001942 0.00812263 + layer.4.k_cache 0.00068256 0.16875698 + layer.4.v_cache 0.00004843 0.01678210 + layer.4.output 10.65454330 499.58636515 + ------------------------------------------------------------------------------------- + TOTAL 4.39815220 209.49985133 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 134728 +BPFP 1.1573 bits/point +EBPFP 2.3146 equivalent bits/point +MSE 209.499851 +---------------------- -------------------------------------------------------- +Time: 0.389s Load: 0.006s, Pack+Encode: 0.166s, Decode+Unpack: 0.217s +---------------------- -------------------------------------------------------- +💾 Converting with 209.4999 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,232B, BPFP=0.6012 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,572B, BPFP=2.1525 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,256B, BPFP=0.9777 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,776B, BPFP=2.0045 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,120B, BPFP=1.1384 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,468B, BPFP=1.9472 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,844B, BPFP=1.0871 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,988B, BPFP=2.0439 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,080B, BPFP=1.6890 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,544B, BPFP=1.9613 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,992B, BPFP=0.5844 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12550225 56.79163179 + layer.0.v_cache 0.00001468 0.00802286 + layer.1.k_cache 0.05234316 5.24518258 + layer.1.v_cache 0.00000508 0.00242136 + layer.2.k_cache 0.00733058 0.78266071 + layer.2.v_cache 0.00001689 0.00677662 + layer.3.k_cache 0.01592888 3.30463664 + layer.3.v_cache 0.00001847 0.00864871 + layer.4.k_cache 0.00069460 0.17341504 + layer.4.v_cache 0.00005205 0.01706144 + layer.4.output 0.16184913 643.44451531 + ------------------------------------------------------------------------------------- + TOTAL 0.07852062 268.85012146 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 105872 +BPFP 1.1584 bits/point +EBPFP 2.3169 equivalent bits/point +MSE 268.850121 +---------------------- -------------------------------------------------------- +Time: 0.387s Load: 0.005s, Pack+Encode: 0.168s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 268.8501 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,116B, BPFP=0.6242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,092B, BPFP=2.2220 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,072B, BPFP=1.0160 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,448B, BPFP=2.0929 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,884B, BPFP=1.1787 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,096B, BPFP=2.0224 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,720B, BPFP=1.1458 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,544B, BPFP=2.1122 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,692B, BPFP=1.7412 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,164B, BPFP=2.0361 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,980B, BPFP=0.6004 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11879376 56.91431290 + layer.0.v_cache 0.00001438 0.00793179 + layer.1.k_cache 0.05518859 5.15569247 + layer.1.v_cache 0.00000514 0.00289155 + layer.2.k_cache 0.00243519 0.88467955 + layer.2.v_cache 0.00001742 0.00727636 + layer.3.k_cache 0.07959401 3.71367000 + layer.3.v_cache 0.00001812 0.00877334 + layer.4.k_cache 0.00069358 0.18766409 + layer.4.v_cache 0.00004620 0.01753911 + layer.4.output 0.17426339 691.16483516 + ------------------------------------------------------------------------------------- + TOTAL 0.08686177 288.53260455 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 101808 +BPFP 1.1997 bits/point +EBPFP 2.3993 equivalent bits/point +MSE 288.532605 +---------------------- -------------------------------------------------------- +Time: 0.387s Load: 0.004s, Pack+Encode: 0.167s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 288.5326 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,232B, BPFP=0.6159 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,548B, BPFP=2.2005 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,244B, BPFP=0.9992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,024B, BPFP=2.1006 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,256B, BPFP=1.1921 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,704B, BPFP=2.0396 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,968B, BPFP=1.1372 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,140B, BPFP=2.1227 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,164B, BPFP=1.7462 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,660B, BPFP=2.0312 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,316B, BPFP=0.5802 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11325797 57.71360518 + layer.0.v_cache 0.00001565 0.00787392 + layer.1.k_cache 0.07651102 5.28048483 + layer.1.v_cache 0.00000530 0.00299088 + layer.2.k_cache 0.00872486 0.79729852 + layer.2.v_cache 0.00001696 0.00774738 + layer.3.k_cache 0.12876377 3.53054586 + layer.3.v_cache 0.00001905 0.00925694 + layer.4.k_cache 0.00069190 0.18306842 + layer.4.v_cache 0.00004968 0.01978967 + layer.4.output 0.16578155 658.26317509 + ------------------------------------------------------------------------------------- + TOTAL 0.08756041 275.02322866 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 106256 +BPFP 1.1910 bits/point +EBPFP 2.3820 equivalent bits/point +MSE 275.023229 +---------------------- -------------------------------------------------------- +Time: 0.387s Load: 0.004s, Pack+Encode: 0.167s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 275.0232 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,060B, BPFP=0.6209 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,840B, BPFP=2.1997 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,032B, BPFP=1.0211 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,324B, BPFP=2.0950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,876B, BPFP=1.1924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,852B, BPFP=1.9992 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,552B, BPFP=1.1266 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,208B, BPFP=2.0714 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,188B, BPFP=1.6615 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,772B, BPFP=1.9830 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,924B, BPFP=0.5776 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12835005 60.21902902 + layer.0.v_cache 0.00001616 0.00836677 + layer.1.k_cache 0.05672838 5.30770319 + layer.1.v_cache 0.00000535 0.00346754 + layer.2.k_cache 0.00240846 0.77746978 + layer.2.v_cache 0.00001644 0.00825976 + layer.3.k_cache 0.04570918 3.71272575 + layer.3.v_cache 0.00001733 0.00958379 + layer.4.k_cache 0.00072090 0.19896165 + layer.4.v_cache 0.00004908 0.02031562 + layer.4.output 0.18444509 698.76472635 + ------------------------------------------------------------------------------------- + TOTAL 0.08971394 291.85993925 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 98628 +BPFP 1.1773 bits/point +EBPFP 2.3546 equivalent bits/point +MSE 291.859939 +---------------------- -------------------------------------------------------- +Time: 0.386s Load: 0.004s, Pack+Encode: 0.168s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 291.8599 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,932B, BPFP=0.6452 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,988B, BPFP=2.1981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,732B, BPFP=1.0414 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,500B, BPFP=2.0907 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,396B, BPFP=1.1875 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,996B, BPFP=1.9798 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,136B, BPFP=1.1303 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,436B, BPFP=2.0766 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,644B, BPFP=1.6822 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,124B, BPFP=2.0079 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,488B, BPFP=0.6127 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10085608 60.86929192 + layer.0.v_cache 0.00001554 0.00742563 + layer.1.k_cache 0.06589249 5.43264728 + layer.1.v_cache 0.00000499 0.00284172 + layer.2.k_cache 0.00422055 0.87146614 + layer.2.v_cache 0.00001579 0.00694959 + layer.3.k_cache 0.05337993 3.82381342 + layer.3.v_cache 0.00001747 0.00902021 + layer.4.k_cache 0.00075412 0.19587290 + layer.4.v_cache 0.00004712 0.01924833 + layer.4.output 0.19135183 761.63839286 + ------------------------------------------------------------------------------------- + TOTAL 0.09203923 317.80631336 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 92372 +BPFP 1.1958 bits/point +EBPFP 2.3916 equivalent bits/point +MSE 317.806313 +---------------------- -------------------------------------------------------- +Time: 0.386s Load: 0.004s, Pack+Encode: 0.168s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 317.8063 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,268B, BPFP=0.6079 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,504B, BPFP=2.1399 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,252B, BPFP=0.9769 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,060B, BPFP=2.0573 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,292B, BPFP=1.1704 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,640B, BPFP=1.9792 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,016B, BPFP=1.1190 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,028B, BPFP=2.0513 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,108B, BPFP=1.6942 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,496B, BPFP=1.9524 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,236B, BPFP=0.5643 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.216s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09057366 52.67581613 + layer.0.v_cache 0.00001789 0.00705191 + layer.1.k_cache 0.03614600 5.30221485 + layer.1.v_cache 0.00000481 0.00264607 + layer.2.k_cache 0.00554357 0.77133851 + layer.2.v_cache 0.00001714 0.00690902 + layer.3.k_cache 0.02742965 3.36590830 + layer.3.v_cache 0.00001729 0.00817866 + layer.4.k_cache 0.00093027 0.17049033 + layer.4.v_cache 0.00004906 0.01701669 + layer.4.output 0.16441821 643.57743410 + ------------------------------------------------------------------------------------- + TOTAL 0.07715628 268.66880054 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 105900 +BPFP 1.1587 bits/point +EBPFP 2.3175 equivalent bits/point +MSE 268.668801 +---------------------- -------------------------------------------------------- +Time: 0.386s Load: 0.004s, Pack+Encode: 0.167s, Decode+Unpack: 0.216s +---------------------- -------------------------------------------------------- +💾 Converting with 268.6688 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,068B, BPFP=0.6226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,848B, BPFP=2.2013 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,064B, BPFP=1.0276 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,316B, BPFP=2.0933 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,884B, BPFP=1.1940 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,884B, BPFP=2.0057 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,692B, BPFP=1.1550 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,352B, BPFP=2.1006 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,444B, BPFP=1.7135 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,964B, BPFP=2.0219 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,896B, BPFP=0.6058 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09325875 57.54348874 + layer.0.v_cache 0.00001406 0.00757077 + layer.1.k_cache 0.05775181 5.52286203 + layer.1.v_cache 0.00000524 0.00286470 + layer.2.k_cache 0.00413042 0.81187320 + layer.2.v_cache 0.00001743 0.00763449 + layer.3.k_cache 0.04387648 3.76098752 + layer.3.v_cache 0.00002497 0.00967353 + layer.4.k_cache 0.00071512 0.18907725 + layer.4.v_cache 0.00005147 0.01948580 + layer.4.output 0.17652170 701.02939471 + ------------------------------------------------------------------------------------- + TOTAL 0.08444104 292.65184006 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 100412 +BPFP 1.1986 bits/point +EBPFP 2.3972 equivalent bits/point +MSE 292.651840 +---------------------- -------------------------------------------------------- +Time: 0.387s Load: 0.004s, Pack+Encode: 0.168s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 292.6518 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,380B, BPFP=0.6001 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,272B, BPFP=2.0014 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,440B, BPFP=0.9659 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,748B, BPFP=1.9084 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,328B, BPFP=1.1236 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,388B, BPFP=1.8445 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,008B, BPFP=1.0668 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,996B, BPFP=1.9524 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,080B, BPFP=1.6122 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,296B, BPFP=1.8281 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,904B, BPFP=0.5049 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212358 60.74235396 + layer.0.v_cache 0.00001682 0.00621450 + layer.1.k_cache 0.05918076 4.74505546 + layer.1.v_cache 0.00000473 0.00226480 + layer.2.k_cache 0.00260097 0.71610793 + layer.2.v_cache 0.00001544 0.00655610 + layer.3.k_cache 0.02838851 3.12830873 + layer.3.v_cache 0.00001657 0.00773125 + layer.4.k_cache 0.00106615 0.16777281 + layer.4.v_cache 0.00004456 0.01576514 + layer.4.output 0.17173813 611.56894278 + ------------------------------------------------------------------------------------- + TOTAL 0.08091912 255.91298412 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 103840 +BPFP 1.0846 bits/point +EBPFP 2.1691 equivalent bits/point +MSE 255.912984 +---------------------- -------------------------------------------------------- +Time: 0.387s Load: 0.004s, Pack+Encode: 0.168s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 255.9130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,352B, BPFP=0.5952 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,036B, BPFP=2.1371 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,648B, BPFP=1.0028 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,528B, BPFP=2.0469 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,648B, BPFP=1.1804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,220B, BPFP=1.9922 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,280B, BPFP=1.1151 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,528B, BPFP=2.0469 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,580B, BPFP=1.7010 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,128B, BPFP=1.9759 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,328B, BPFP=0.5917 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10731998 61.93241744 + layer.0.v_cache 0.00001618 0.00698117 + layer.1.k_cache 0.07375650 4.83991172 + layer.1.v_cache 0.00000540 0.00259557 + layer.2.k_cache 0.00246326 0.72740910 + layer.2.v_cache 0.00001746 0.00714604 + layer.3.k_cache 0.01642703 3.04084466 + layer.3.v_cache 0.00001778 0.00831304 + layer.4.k_cache 0.00066607 0.17110970 + layer.4.v_cache 0.00004728 0.01681870 + layer.4.output 0.15452130 613.52470576 + ------------------------------------------------------------------------------------- + TOTAL 0.07543447 256.78979338 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 112276 +BPFP 1.1727 bits/point +EBPFP 2.3453 equivalent bits/point +MSE 256.789793 +---------------------- -------------------------------------------------------- +Time: 0.387s Load: 0.005s, Pack+Encode: 0.167s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 256.7898 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 99, 128) +Output shape: (1, 99, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.output: torch.Size([1, 99, 3584]) -> torch.Size([1, 1, 99, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,292B, BPFP=0.5196 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,060B, BPFP=2.0612 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,660B, BPFP=0.8933 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,616B, BPFP=1.9912 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,856B, BPFP=1.0821 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,220B, BPFP=1.9287 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,508B, BPFP=1.0271 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,696B, BPFP=2.0038 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,268B, BPFP=1.6206 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,176B, BPFP=1.9217 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,556B, BPFP=0.5762 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09061131 58.07721749 + layer.0.v_cache 0.00001461 0.00675965 + layer.1.k_cache 0.05062923 4.77832463 + layer.1.v_cache 0.00000709 0.00244220 + layer.2.k_cache 0.00392854 0.81781499 + layer.2.v_cache 0.00001793 0.00667846 + layer.3.k_cache 0.05082532 3.43417697 + layer.3.v_cache 0.00001820 0.00818883 + layer.4.k_cache 0.00080398 0.17064819 + layer.4.v_cache 0.00005053 0.01607071 + layer.4.output 0.02403493 561.28300866 + ------------------------------------------------------------------------------------- + TOTAL 0.02147948 235.07643428 + (elements=861,696) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 861696 +Total Bytes 120908 +BPFP 1.1225 bits/point +EBPFP 2.2450 equivalent bits/point +MSE 235.076434 +---------------------- -------------------------------------------------------- +Time: 0.388s Load: 0.005s, Pack+Encode: 0.168s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 235.0764 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,900B, BPFP=0.6382 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,520B, BPFP=2.3151 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,868B, BPFP=1.0713 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,964B, BPFP=2.1928 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,628B, BPFP=1.2386 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,632B, BPFP=2.1197 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,344B, BPFP=1.1761 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,848B, BPFP=2.1673 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,752B, BPFP=1.7060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,252B, BPFP=2.0361 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,608B, BPFP=0.6793 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11136492 60.95384023 + layer.0.v_cache 0.00001955 0.00838753 + layer.1.k_cache 0.06280328 5.41299825 + layer.1.v_cache 0.00000544 0.00311097 + layer.2.k_cache 0.00235174 0.81249570 + layer.2.v_cache 0.00001859 0.00801167 + layer.3.k_cache 0.03326307 3.45370913 + layer.3.v_cache 0.00001926 0.00951025 + layer.4.k_cache 0.00064473 0.18271488 + layer.4.v_cache 0.00005254 0.01827091 + layer.4.output 0.19494371 761.27081237 + ------------------------------------------------------------------------------------- + TOTAL 0.09265583 317.63286683 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 97316 +BPFP 1.2598 bits/point +EBPFP 2.5196 equivalent bits/point +MSE 317.632867 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.003s, Pack+Encode: 0.166s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 317.6329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,984B, BPFP=0.6301 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,844B, BPFP=2.2897 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,888B, BPFP=1.0321 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,348B, BPFP=2.1850 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,820B, BPFP=1.2289 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,936B, BPFP=2.0980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,512B, BPFP=1.1639 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,376B, BPFP=2.1909 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,240B, BPFP=1.7399 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,936B, BPFP=2.0980 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,040B, BPFP=0.6045 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08829598 63.00094357 + layer.0.v_cache 0.00001455 0.00848946 + layer.1.k_cache 0.05950350 5.44282490 + layer.1.v_cache 0.00000529 0.00324220 + layer.2.k_cache 0.00411868 0.79213395 + layer.2.v_cache 0.00001680 0.00808389 + layer.3.k_cache 0.04877868 3.87656671 + layer.3.v_cache 0.00001896 0.01041664 + layer.4.k_cache 0.00063741 0.20361581 + layer.4.v_cache 0.00004764 0.02053592 + layer.4.output 0.18365649 728.99921573 + ------------------------------------------------------------------------------------- + TOTAL 0.08747253 304.49184489 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 98924 +BPFP 1.2287 bits/point +EBPFP 2.4574 equivalent bits/point +MSE 304.491845 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 304.4918 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,048B, BPFP=0.6185 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,968B, BPFP=2.2256 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,084B, BPFP=1.0317 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,540B, BPFP=2.1388 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,972B, BPFP=1.2119 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,280B, BPFP=2.0860 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,648B, BPFP=1.1461 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,524B, BPFP=2.1356 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,612B, BPFP=1.7476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,200B, BPFP=2.0698 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,252B, BPFP=0.5871 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11195674 59.46055068 + layer.0.v_cache 0.00001346 0.00751082 + layer.1.k_cache 0.03755230 5.49507874 + layer.1.v_cache 0.00000535 0.00300857 + layer.2.k_cache 0.00239600 0.85385380 + layer.2.v_cache 0.00001753 0.00846668 + layer.3.k_cache 0.02867049 3.37645533 + layer.3.v_cache 0.00001751 0.00867312 + layer.4.k_cache 0.00071811 0.18649859 + layer.4.v_cache 0.00005486 0.01985432 + layer.4.output 0.19495771 700.33586503 + ------------------------------------------------------------------------------------- + TOTAL 0.09094743 292.45711799 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 101128 +BPFP 1.2071 bits/point +EBPFP 2.4142 equivalent bits/point +MSE 292.457118 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 292.4571 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,428B, BPFP=0.5951 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,908B, BPFP=2.0674 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,568B, BPFP=0.9667 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,348B, BPFP=1.9701 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,548B, BPFP=1.1368 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,940B, BPFP=1.8993 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,268B, BPFP=1.0882 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,468B, BPFP=1.9910 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,572B, BPFP=1.6618 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,108B, BPFP=1.9285 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,504B, BPFP=0.5333 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08657128 58.54219835 + layer.0.v_cache 0.00001574 0.00768704 + layer.1.k_cache 0.05618976 4.82210320 + layer.1.v_cache 0.00000506 0.00270772 + layer.2.k_cache 0.00392605 0.79624634 + layer.2.v_cache 0.00001570 0.00694850 + layer.3.k_cache 0.02849737 3.58795437 + layer.3.v_cache 0.00001769 0.00926359 + layer.4.k_cache 0.00099738 0.17877388 + layer.4.v_cache 0.00004751 0.01861248 + layer.4.output 0.15107292 598.55367063 + ------------------------------------------------------------------------------------- + TOTAL 0.07257612 250.46165823 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 109660 +BPFP 1.1199 bits/point +EBPFP 2.2398 equivalent bits/point +MSE 250.461658 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.005s, Pack+Encode: 0.166s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 250.4617 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,172B, BPFP=0.6195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,152B, BPFP=2.1781 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,120B, BPFP=1.0000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,508B, BPFP=2.0523 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,020B, BPFP=1.1758 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,252B, BPFP=2.0023 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,700B, BPFP=1.1133 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,596B, BPFP=2.0695 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,696B, BPFP=1.6984 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,216B, BPFP=1.9953 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,960B, BPFP=0.5569 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09254595 57.27202148 + layer.0.v_cache 0.00001776 0.00721865 + layer.1.k_cache 0.05775445 5.29803925 + layer.1.v_cache 0.00000504 0.00255661 + layer.2.k_cache 0.00891839 0.77850933 + layer.2.v_cache 0.00001732 0.00751184 + layer.3.k_cache 0.02836140 3.73209534 + layer.3.v_cache 0.00001672 0.00875349 + layer.4.k_cache 0.00083566 0.17655332 + layer.4.v_cache 0.00004876 0.01878720 + layer.4.output 0.16993865 674.04268973 + ------------------------------------------------------------------------------------- + TOTAL 0.08106423 281.50593380 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 101392 +BPFP 1.1649 bits/point +EBPFP 2.3298 equivalent bits/point +MSE 281.505934 +---------------------- -------------------------------------------------------- +Time: 0.386s Load: 0.005s, Pack+Encode: 0.167s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 281.5059 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,140B, BPFP=0.5058 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,352B, BPFP=1.9897 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,448B, BPFP=0.8776 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,764B, BPFP=1.8950 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,596B, BPFP=1.0625 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,352B, BPFP=1.8286 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,268B, BPFP=1.0097 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,748B, BPFP=1.8924 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,716B, BPFP=1.5651 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,252B, BPFP=1.8125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,132B, BPFP=0.5093 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13093194 60.26195031 + layer.0.v_cache 0.00001635 0.00743886 + layer.1.k_cache 0.05235285 5.39479238 + layer.1.v_cache 0.00000496 0.00262102 + layer.2.k_cache 0.00512381 0.75913286 + layer.2.v_cache 0.00001647 0.00686393 + layer.3.k_cache 0.01138228 3.54926858 + layer.3.v_cache 0.00001850 0.00812043 + layer.4.k_cache 0.00081880 0.16695420 + layer.4.v_cache 0.00004796 0.01558094 + layer.4.output 0.02447439 572.10852356 + ------------------------------------------------------------------------------------- + TOTAL 0.02188439 239.70190520 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 111768 +BPFP 1.0591 bits/point +EBPFP 2.1181 equivalent bits/point +MSE 239.701905 +---------------------- -------------------------------------------------------- +Time: 0.387s Load: 0.005s, Pack+Encode: 0.166s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 239.7019 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,116B, BPFP=0.6163 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,252B, BPFP=2.2255 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,100B, BPFP=1.0087 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,572B, BPFP=2.0910 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,008B, BPFP=1.1883 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,224B, BPFP=2.0222 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,664B, BPFP=1.1203 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,568B, BPFP=2.0902 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,576B, BPFP=1.6962 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,116B, BPFP=2.0008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,284B, BPFP=0.6014 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510638 58.76169403 + layer.0.v_cache 0.00001399 0.00751304 + layer.1.k_cache 0.05467150 5.63644448 + layer.1.v_cache 0.00000511 0.00272947 + layer.2.k_cache 0.00545186 0.76675492 + layer.2.v_cache 0.00001652 0.00758732 + layer.3.k_cache 0.02986256 3.65211313 + layer.3.v_cache 0.00001747 0.00799184 + layer.4.k_cache 0.00068266 0.17426311 + layer.4.v_cache 0.00005277 0.01836810 + layer.4.output 0.17642128 682.32154159 + ------------------------------------------------------------------------------------- + TOTAL 0.08357822 285.01683827 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 102480 +BPFP 1.1923 bits/point +EBPFP 2.3846 equivalent bits/point +MSE 285.016838 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.003s, Pack+Encode: 0.167s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 285.0168 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,068B, BPFP=0.6226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,904B, BPFP=2.2127 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,016B, BPFP=1.0179 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,452B, BPFP=2.1209 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,864B, BPFP=1.1899 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,020B, BPFP=2.0333 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,576B, BPFP=1.1315 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,432B, BPFP=2.1169 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,316B, BPFP=1.6875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,840B, BPFP=1.9968 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,264B, BPFP=0.5874 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08730992 61.00462916 + layer.0.v_cache 0.00001543 0.00796145 + layer.1.k_cache 0.05635426 5.67809226 + layer.1.v_cache 0.00000520 0.00319931 + layer.2.k_cache 0.00390536 0.89513278 + layer.2.v_cache 0.00001661 0.00788054 + layer.3.k_cache 0.06483487 3.58728979 + layer.3.v_cache 0.00002010 0.00940307 + layer.4.k_cache 0.00072652 0.18912450 + layer.4.v_cache 0.00004842 0.01861699 + layer.4.output 0.18845482 696.14726345 + ------------------------------------------------------------------------------------- + TOTAL 0.09014238 290.84895141 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 99752 +BPFP 1.1907 bits/point +EBPFP 2.3814 equivalent bits/point +MSE 290.848951 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 290.8490 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,848B, BPFP=0.6544 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,784B, BPFP=2.2482 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,588B, BPFP=1.0542 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,348B, BPFP=2.1480 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,240B, BPFP=1.2040 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,840B, BPFP=2.0312 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,024B, BPFP=1.1544 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,108B, BPFP=2.0928 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,364B, BPFP=1.6921 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,740B, BPFP=2.0083 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,608B, BPFP=0.6765 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10784986 60.09630630 + layer.0.v_cache 0.00001387 0.00727566 + layer.1.k_cache 0.01409798 5.72603562 + layer.1.v_cache 0.00000489 0.00243538 + layer.2.k_cache 0.00254294 0.79515524 + layer.2.v_cache 0.00001653 0.00735608 + layer.3.k_cache 0.05709259 4.41875772 + layer.3.v_cache 0.00001736 0.00825685 + layer.4.k_cache 0.00062715 0.17207000 + layer.4.v_cache 0.00004758 0.01654773 + layer.4.output 0.22376562 795.26890756 + ------------------------------------------------------------------------------------- + TOTAL 0.10286295 331.65485585 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 91492 +BPFP 1.2366 bits/point +EBPFP 2.4733 equivalent bits/point +MSE 331.654856 +---------------------- -------------------------------------------------------- +Time: 0.381s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.212s +---------------------- -------------------------------------------------------- +💾 Converting with 331.6549 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,108B, BPFP=0.6147 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,788B, BPFP=2.1337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,040B, BPFP=0.9968 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,332B, BPFP=2.0435 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,864B, BPFP=1.1598 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,012B, BPFP=1.9802 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,600B, BPFP=1.1076 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,392B, BPFP=2.0554 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,424B, BPFP=1.6661 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,884B, BPFP=1.9549 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,300B, BPFP=0.6018 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14563359 58.10748368 + layer.0.v_cache 0.00001368 0.00675092 + layer.1.k_cache 0.05947259 5.09342705 + layer.1.v_cache 0.00000496 0.00248101 + layer.2.k_cache 0.00553216 0.79489073 + layer.2.v_cache 0.00001557 0.00698417 + layer.3.k_cache 0.04377548 3.90968574 + layer.3.v_cache 0.00001730 0.00836577 + layer.4.k_cache 0.00079035 0.17176398 + layer.4.v_cache 0.00005153 0.01692892 + layer.4.output 0.17347779 681.76740506 + ------------------------------------------------------------------------------------- + TOTAL 0.08645010 284.73474103 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 100744 +BPFP 1.1721 bits/point +EBPFP 2.3442 equivalent bits/point +MSE 284.734741 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.004s, Pack+Encode: 0.165s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 284.7347 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,632B, BPFP=0.6231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,532B, BPFP=2.2566 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,240B, BPFP=1.0038 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,912B, BPFP=2.1098 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,012B, BPFP=1.1866 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,576B, BPFP=2.0303 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,712B, BPFP=1.1155 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,976B, BPFP=2.1250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,032B, BPFP=1.6648 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,608B, BPFP=2.0379 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,476B, BPFP=0.6249 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.202s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12373332 62.78690962 + layer.0.v_cache 0.00001436 0.00823656 + layer.1.k_cache 0.01270404 6.33212743 + layer.1.v_cache 0.00000490 0.00277203 + layer.2.k_cache 0.00252672 0.86680753 + layer.2.v_cache 0.00001629 0.00814458 + layer.3.k_cache 0.05185594 4.25026865 + layer.3.v_cache 0.00001849 0.00945062 + layer.4.k_cache 0.00068458 0.18842597 + layer.4.v_cache 0.00004975 0.01827426 + layer.4.output 0.22283003 818.54613095 + ------------------------------------------------------------------------------------- + TOTAL 0.10302462 341.42907847 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 86708 +BPFP 1.2075 bits/point +EBPFP 2.4150 equivalent bits/point +MSE 341.429078 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.202s +---------------------- -------------------------------------------------------- +💾 Converting with 341.4291 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,240B, BPFP=0.6027 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,932B, BPFP=2.2195 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,336B, BPFP=0.9926 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,268B, BPFP=2.0960 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,476B, BPFP=1.2046 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,860B, BPFP=2.0201 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,140B, BPFP=1.1421 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,424B, BPFP=2.1250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,384B, BPFP=1.7455 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,988B, BPFP=2.0439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,044B, BPFP=0.6389 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12002775 53.97024972 + layer.0.v_cache 0.00001466 0.00765871 + layer.1.k_cache 0.05271155 5.20729356 + layer.1.v_cache 0.00000524 0.00302138 + layer.2.k_cache 0.00808975 0.81437011 + layer.2.v_cache 0.00001826 0.00766548 + layer.3.k_cache 0.07501186 3.42688424 + layer.3.v_cache 0.00001895 0.00910447 + layer.4.k_cache 0.00071216 0.18850332 + layer.4.v_cache 0.00005170 0.01830851 + layer.4.output 0.17053346 643.28433248 + ------------------------------------------------------------------------------------- + TOTAL 0.08531742 268.62608158 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 111092 +BPFP 1.2156 bits/point +EBPFP 2.4311 equivalent bits/point +MSE 268.626082 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 268.6261 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 75, 128) +Output shape: (1, 75, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.output: torch.Size([1, 75, 3584]) -> torch.Size([1, 1, 75, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,012B, BPFP=0.6275 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,632B, BPFP=2.2150 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,924B, BPFP=1.0258 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,108B, BPFP=2.1058 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,824B, BPFP=1.2133 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,744B, BPFP=2.0300 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,432B, BPFP=1.1317 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,144B, BPFP=2.1133 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,100B, BPFP=1.6875 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,644B, BPFP=2.0092 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,768B, BPFP=0.6181 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08149677 59.36526042 + layer.0.v_cache 0.00001391 0.00787700 + layer.1.k_cache 0.05841089 5.40197347 + layer.1.v_cache 0.00000541 0.00302949 + layer.2.k_cache 0.00252637 0.82575958 + layer.2.v_cache 0.00001693 0.00803259 + layer.3.k_cache 0.02876982 3.83361898 + layer.3.v_cache 0.00001932 0.00909029 + layer.4.k_cache 0.00075246 0.18789688 + layer.4.v_cache 0.00005041 0.01941289 + layer.4.output 0.19266937 718.98785714 + ------------------------------------------------------------------------------------- + TOTAL 0.08945576 300.15158539 + (elements=652,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 652800 +Total Bytes 98332 +BPFP 1.2050 bits/point +EBPFP 2.4101 equivalent bits/point +MSE 300.151585 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 300.1516 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,836B, BPFP=0.6517 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,724B, BPFP=2.2344 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,528B, BPFP=1.0404 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,048B, BPFP=2.0790 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,220B, BPFP=1.1994 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,648B, BPFP=1.9871 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,924B, BPFP=1.1314 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,196B, BPFP=2.1131 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,380B, BPFP=1.6958 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,704B, BPFP=2.0000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,976B, BPFP=0.6229 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10404026 61.03726735 + layer.0.v_cache 0.00001395 0.00775511 + layer.1.k_cache 0.01458383 6.20755095 + layer.1.v_cache 0.00000499 0.00266830 + layer.2.k_cache 0.00236740 0.88891276 + layer.2.v_cache 0.00001641 0.00788254 + layer.3.k_cache 0.03645969 4.24597347 + layer.3.v_cache 0.00002420 0.00931371 + layer.4.k_cache 0.00069996 0.18831747 + layer.4.v_cache 0.00005119 0.01978521 + layer.4.output 0.20980707 795.88458508 + ------------------------------------------------------------------------------------- + TOTAL 0.09570067 331.98867779 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 89184 +BPFP 1.2054 bits/point +EBPFP 2.4109 equivalent bits/point +MSE 331.988678 +---------------------- -------------------------------------------------------- +Time: 0.364s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 331.9887 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,852B, BPFP=0.6366 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,880B, BPFP=2.2054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,680B, BPFP=1.0446 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,172B, BPFP=2.0473 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,412B, BPFP=1.2080 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,900B, BPFP=1.9866 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,080B, BPFP=1.1339 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,072B, BPFP=2.0250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,480B, BPFP=1.6696 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,832B, BPFP=1.9714 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,356B, BPFP=0.6810 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.204s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11725240 64.67843890 + layer.0.v_cache 0.00001567 0.00846202 + layer.1.k_cache 0.03693308 5.42845546 + layer.1.v_cache 0.00000533 0.00333528 + layer.2.k_cache 0.00240566 0.86377449 + layer.2.v_cache 0.00001850 0.00909853 + layer.3.k_cache 0.06751672 3.45582711 + layer.3.v_cache 0.00001909 0.00970651 + layer.4.k_cache 0.00067984 0.19433744 + layer.4.v_cache 0.00004901 0.02008026 + layer.4.output 0.20094273 768.69470663 + ------------------------------------------------------------------------------------- + TOTAL 0.09597026 320.91379191 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 92716 +BPFP 1.2174 bits/point +EBPFP 2.4348 equivalent bits/point +MSE 320.913792 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.158s, Decode+Unpack: 0.204s +---------------------- -------------------------------------------------------- +💾 Converting with 320.9138 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,856B, BPFP=0.6375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,848B, BPFP=2.1982 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,664B, BPFP=1.0411 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,992B, BPFP=2.0071 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,372B, BPFP=1.1991 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,704B, BPFP=1.9429 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,096B, BPFP=1.1375 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,152B, BPFP=2.0429 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,364B, BPFP=1.6438 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,556B, BPFP=1.9098 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,192B, BPFP=0.6120 +⌛️ [2/4] FRONTEND: Frontend time: 0.158s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07685234 60.97275391 + layer.0.v_cache 0.00001390 0.00822835 + layer.1.k_cache 0.01187498 5.29514247 + layer.1.v_cache 0.00000934 0.00311687 + layer.2.k_cache 0.00239603 0.87554605 + layer.2.v_cache 0.00001621 0.00741563 + layer.3.k_cache 0.07241479 3.58262002 + layer.3.v_cache 0.00001716 0.00919994 + layer.4.k_cache 0.00072067 0.19044805 + layer.4.v_cache 0.00004958 0.01799179 + layer.4.output 0.21417375 770.08775510 + ------------------------------------------------------------------------------------- + TOTAL 0.09785772 321.26922052 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 89796 +BPFP 1.1790 bits/point +EBPFP 2.3581 equivalent bits/point +MSE 321.269221 +---------------------- -------------------------------------------------------- +Time: 0.365s Load: 0.003s, Pack+Encode: 0.158s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 321.2692 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,044B, BPFP=0.6258 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,956B, BPFP=2.2525 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,008B, BPFP=1.0296 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,432B, BPFP=2.1447 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,908B, BPFP=1.2146 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,000B, BPFP=2.0559 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,592B, BPFP=1.1497 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,444B, BPFP=2.1472 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,336B, BPFP=1.7138 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,840B, BPFP=2.0230 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,776B, BPFP=0.6396 +⌛️ [2/4] FRONTEND: Frontend time: 0.159s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09264422 59.21542198 + layer.0.v_cache 0.00001680 0.00847481 + layer.1.k_cache 0.05606840 5.47899949 + layer.1.v_cache 0.00000542 0.00322191 + layer.2.k_cache 0.00241068 0.90958886 + layer.2.v_cache 0.00001702 0.00854615 + layer.3.k_cache 0.15270803 3.54792063 + layer.3.v_cache 0.00001849 0.01063357 + layer.4.k_cache 0.00070970 0.19863756 + layer.4.v_cache 0.00005539 0.02055768 + layer.4.output 0.18838861 706.69114192 + ------------------------------------------------------------------------------------- + TOTAL 0.09549261 295.07294095 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 101336 +BPFP 1.2255 bits/point +EBPFP 2.4510 equivalent bits/point +MSE 295.072941 +---------------------- -------------------------------------------------------- +Time: 0.366s Load: 0.004s, Pack+Encode: 0.159s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 295.0729 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,068B, BPFP=0.6226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,900B, BPFP=2.2119 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,952B, BPFP=1.0049 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,228B, BPFP=2.0755 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,868B, BPFP=1.1907 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,908B, BPFP=2.0106 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,520B, BPFP=1.1201 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,348B, BPFP=2.0998 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,384B, BPFP=1.7013 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,880B, BPFP=2.0049 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,284B, BPFP=0.5590 +⌛️ [2/4] FRONTEND: Frontend time: 0.202s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10144231 59.89067573 + layer.0.v_cache 0.00001330 0.00710079 + layer.1.k_cache 0.09893422 5.25014347 + layer.1.v_cache 0.00000513 0.00281220 + layer.2.k_cache 0.00368564 0.86971283 + layer.2.v_cache 0.00001777 0.00761884 + layer.3.k_cache 0.06341467 3.83554909 + layer.3.v_cache 0.00001739 0.00893562 + layer.4.k_cache 0.00073693 0.18303818 + layer.4.v_cache 0.00004994 0.01832871 + layer.4.output 0.18261438 698.15816327 + ------------------------------------------------------------------------------------- + TOTAL 0.09097753 291.59888578 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 98340 +BPFP 1.1738 bits/point +EBPFP 2.3477 equivalent bits/point +MSE 291.598886 +---------------------- -------------------------------------------------------- +Time: 0.421s Load: 0.004s, Pack+Encode: 0.202s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 291.5989 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,320B, BPFP=0.5963 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,416B, BPFP=2.0503 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,392B, BPFP=0.9684 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,888B, BPFP=1.9555 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,196B, BPFP=1.1128 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,432B, BPFP=1.8736 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,908B, BPFP=1.0611 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,908B, BPFP=1.9591 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,096B, BPFP=1.6336 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,444B, BPFP=1.8757 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,996B, BPFP=0.5387 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08947531 59.12500561 + layer.0.v_cache 0.00001432 0.00676116 + layer.1.k_cache 0.05320675 5.02586417 + layer.1.v_cache 0.00000497 0.00255946 + layer.2.k_cache 0.00562141 0.77513070 + layer.2.v_cache 0.00001574 0.00632314 + layer.3.k_cache 0.02742601 2.94466356 + layer.3.v_cache 0.00001759 0.00785376 + layer.4.k_cache 0.00114861 0.18048574 + layer.4.v_cache 0.00004734 0.01596389 + layer.4.output 0.15628038 620.20592159 + ------------------------------------------------------------------------------------- + TOTAL 0.07476122 259.38423896 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 104996 +BPFP 1.1092 bits/point +EBPFP 2.2185 equivalent bits/point +MSE 259.384239 +---------------------- -------------------------------------------------------- +Time: 0.386s Load: 0.005s, Pack+Encode: 0.166s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 259.3842 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,316B, BPFP=0.5632 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,056B, BPFP=2.0476 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,664B, BPFP=0.9620 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,480B, BPFP=1.9497 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,632B, BPFP=1.1264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,124B, BPFP=1.8893 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,276B, BPFP=1.0659 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,536B, BPFP=1.9592 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,368B, BPFP=1.5910 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,852B, BPFP=1.8431 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,660B, BPFP=0.5013 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12311893 59.17393427 + layer.0.v_cache 0.00001357 0.00675370 + layer.1.k_cache 0.03743907 4.69470414 + layer.1.v_cache 0.00000515 0.00271731 + layer.2.k_cache 0.00511221 0.80624871 + layer.2.v_cache 0.00001545 0.00656647 + layer.3.k_cache 0.04225797 3.65697745 + layer.3.v_cache 0.00001799 0.00808338 + layer.4.k_cache 0.00119426 0.16925012 + layer.4.v_cache 0.00004621 0.01607210 + layer.4.output 0.14783020 585.58778144 + ------------------------------------------------------------------------------------- + TOTAL 0.07317837 245.15622222 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 108964 +BPFP 1.0886 bits/point +EBPFP 2.1772 equivalent bits/point +MSE 245.156222 +---------------------- -------------------------------------------------------- +Time: 0.386s Load: 0.005s, Pack+Encode: 0.167s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 245.1562 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,560B, BPFP=0.5453 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,784B, BPFP=2.1115 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,068B, BPFP=0.9295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,980B, BPFP=1.9884 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,080B, BPFP=1.0846 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,660B, BPFP=1.9393 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,640B, BPFP=1.0172 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,164B, BPFP=2.0165 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,616B, BPFP=1.6262 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,588B, BPFP=1.9283 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,596B, BPFP=0.6039 +⌛️ [2/4] FRONTEND: Frontend time: 0.169s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09293266 58.70118145 + layer.0.v_cache 0.00001503 0.00733377 + layer.1.k_cache 0.07854857 5.00769402 + layer.1.v_cache 0.00000559 0.00254122 + layer.2.k_cache 0.00650115 0.79278415 + layer.2.v_cache 0.00001763 0.00682125 + layer.3.k_cache 0.03715762 3.25617173 + layer.3.v_cache 0.00001788 0.00817760 + layer.4.k_cache 0.00073365 0.15977216 + layer.4.v_cache 0.00004630 0.01565303 + layer.4.output 11.17680568 525.12434349 + ------------------------------------------------------------------------------------- + TOTAL 4.61491858 220.22520793 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 126736 +BPFP 1.1420 bits/point +EBPFP 2.2840 equivalent bits/point +MSE 220.225208 +---------------------- -------------------------------------------------------- +Time: 0.388s Load: 0.004s, Pack+Encode: 0.169s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 220.2252 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 96, 128) +Output shape: (1, 96, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.output: torch.Size([1, 96, 3584]) -> torch.Size([1, 1, 96, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,100B, BPFP=0.5046 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,984B, BPFP=2.1133 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,560B, BPFP=0.9049 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,372B, BPFP=2.0137 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,848B, BPFP=1.1146 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,132B, BPFP=1.9746 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,616B, BPFP=1.0768 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,528B, BPFP=2.0391 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,488B, BPFP=1.7070 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,968B, BPFP=1.9479 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,408B, BPFP=0.5675 +⌛️ [2/4] FRONTEND: Frontend time: 0.167s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10158994 57.71727498 + layer.0.v_cache 0.00001467 0.00806050 + layer.1.k_cache 0.06416908 4.50751813 + layer.1.v_cache 0.00000542 0.00282517 + layer.2.k_cache 0.00400108 0.83296506 + layer.2.v_cache 0.00001872 0.00776419 + layer.3.k_cache 0.01491791 3.46027788 + layer.3.v_cache 0.00001877 0.00940881 + layer.4.k_cache 0.00070988 0.18564443 + layer.4.v_cache 0.00004832 0.01882006 + layer.4.output 0.14169503 562.55743118 + ------------------------------------------------------------------------------------- + TOTAL 0.06925641 235.56779867 + (elements=835,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 835584 +Total Bytes 119004 +BPFP 1.1394 bits/point +EBPFP 2.2787 equivalent bits/point +MSE 235.567799 +---------------------- -------------------------------------------------------- +Time: 0.385s Load: 0.004s, Pack+Encode: 0.167s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 235.5678 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,980B, BPFP=0.6467 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,408B, BPFP=2.2587 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,824B, BPFP=1.0469 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,760B, BPFP=2.1181 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,612B, BPFP=1.2179 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,376B, BPFP=2.0347 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,296B, BPFP=1.1493 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,840B, BPFP=2.1354 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,896B, BPFP=1.7135 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,380B, BPFP=2.0356 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,828B, BPFP=0.6147 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10321087 60.90833198 + layer.0.v_cache 0.00001557 0.00786282 + layer.1.k_cache 0.06175087 5.58107249 + layer.1.v_cache 0.00000571 0.00288688 + layer.2.k_cache 0.01329466 0.86442905 + layer.2.v_cache 0.00001784 0.00816489 + layer.3.k_cache 0.05039416 3.84921307 + layer.3.v_cache 0.00001745 0.00926588 + layer.4.k_cache 0.00079786 0.18844133 + layer.4.v_cache 0.00005071 0.01942348 + layer.4.output 0.18870974 750.34362599 + ------------------------------------------------------------------------------------- + TOTAL 0.09120729 313.16732199 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 95200 +BPFP 1.2153 bits/point +EBPFP 2.4306 equivalent bits/point +MSE 313.167322 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.003s, Pack+Encode: 0.166s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 313.1673 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,972B, BPFP=0.6450 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,960B, BPFP=2.1615 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,788B, BPFP=1.0391 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,324B, BPFP=2.0234 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,344B, BPFP=1.1597 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,800B, BPFP=1.9097 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,068B, BPFP=1.0998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,336B, BPFP=2.0260 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,424B, BPFP=1.6111 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,872B, BPFP=1.9253 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,644B, BPFP=0.6400 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09666174 61.33734809 + layer.0.v_cache 0.00001715 0.00724124 + layer.1.k_cache 0.08476152 5.43973287 + layer.1.v_cache 0.00000506 0.00257915 + layer.2.k_cache 0.00238475 0.80527104 + layer.2.v_cache 0.00001567 0.00700837 + layer.3.k_cache 0.07332428 3.65324487 + layer.3.v_cache 0.00001796 0.00837707 + layer.4.k_cache 0.00072290 0.16939640 + layer.4.v_cache 0.00005092 0.01587711 + layer.4.output 0.18875189 750.08767361 + ------------------------------------------------------------------------------------- + TOTAL 0.09289560 313.06234068 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 92532 +BPFP 1.1812 bits/point +EBPFP 2.3624 equivalent bits/point +MSE 313.062341 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 313.0623 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,852B, BPFP=0.6553 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,944B, BPFP=2.2849 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,524B, BPFP=1.0395 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,356B, BPFP=2.1498 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,408B, BPFP=1.2426 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,004B, BPFP=2.0689 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,036B, BPFP=1.1572 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,472B, BPFP=2.1765 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,392B, BPFP=1.6985 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,048B, BPFP=2.0790 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,752B, BPFP=0.6812 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.213s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10601801 64.02579274 + layer.0.v_cache 0.00001817 0.00798651 + layer.1.k_cache 0.01696859 5.90647215 + layer.1.v_cache 0.00000501 0.00299009 + layer.2.k_cache 0.00274148 0.88281542 + layer.2.v_cache 0.00001757 0.00798133 + layer.3.k_cache 0.01706722 4.28129353 + layer.3.v_cache 0.00001843 0.00926888 + layer.4.k_cache 0.00080906 0.18653621 + layer.4.v_cache 0.00005076 0.01866869 + layer.4.output 0.19983917 795.30114233 + ------------------------------------------------------------------------------------- + TOTAL 0.09074050 331.90810599 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 92788 +BPFP 1.2542 bits/point +EBPFP 2.5083 equivalent bits/point +MSE 331.908106 +---------------------- -------------------------------------------------------- +Time: 0.383s Load: 0.004s, Pack+Encode: 0.166s, Decode+Unpack: 0.213s +---------------------- -------------------------------------------------------- +💾 Converting with 331.9081 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 65, 128) +Output shape: (1, 65, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.output: torch.Size([1, 65, 3584]) -> torch.Size([1, 1, 65, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,596B, BPFP=0.6240 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,432B, BPFP=2.2673 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,064B, BPFP=0.9769 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 8,912B, BPFP=2.1423 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,968B, BPFP=1.1942 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,652B, BPFP=2.0798 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,692B, BPFP=1.1279 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 8,944B, BPFP=2.1500 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,204B, BPFP=1.7317 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,680B, BPFP=2.0865 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,036B, BPFP=0.6194 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.214s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09579474 62.83121620 + layer.0.v_cache 0.00001312 0.00779320 + layer.1.k_cache 0.01648773 5.89322604 + layer.1.v_cache 0.00000712 0.00282126 + layer.2.k_cache 0.00431219 0.86116826 + layer.2.v_cache 0.00001803 0.00831833 + layer.3.k_cache 0.03460605 4.26579684 + layer.3.v_cache 0.00001721 0.00996055 + layer.4.k_cache 0.00068162 0.18536502 + layer.4.v_cache 0.00004727 0.02058090 + layer.4.output 0.20898957 831.32657967 + ------------------------------------------------------------------------------------- + TOTAL 0.09499483 346.66895908 + (elements=565,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 565760 +Total Bytes 86180 +BPFP 1.2186 bits/point +EBPFP 2.4372 equivalent bits/point +MSE 346.668959 +---------------------- -------------------------------------------------------- +Time: 0.384s Load: 0.003s, Pack+Encode: 0.166s, Decode+Unpack: 0.214s +---------------------- -------------------------------------------------------- +💾 Converting with 346.6690 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 41, 128) +Output shape: (1, 41, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.output: torch.Size([1, 41, 3584]) -> torch.Size([1, 1, 41, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 1,956B, BPFP=0.7454 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 6,308B, BPFP=2.4040 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,188B, BPFP=1.2149 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,036B, BPFP=2.3003 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 3,664B, BPFP=1.3963 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 5,944B, BPFP=2.2652 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,392B, BPFP=1.2927 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,028B, BPFP=2.2973 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,028B, BPFP=1.9162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 5,956B, BPFP=2.2698 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 14,304B, BPFP=0.7787 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10098721 75.87734018 + layer.0.v_cache 0.00001362 0.01021710 + layer.1.k_cache 0.01708067 6.60620936 + layer.1.v_cache 0.00000533 0.00364751 + layer.2.k_cache 0.00250472 1.02595920 + layer.2.v_cache 0.00001755 0.01089840 + layer.3.k_cache 0.08199174 4.67837115 + layer.3.v_cache 0.00001815 0.01125958 + layer.4.k_cache 0.00059437 0.22606059 + layer.4.v_cache 0.00005232 0.02568647 + layer.4.output 0.35770382 1308.13959059 + ------------------------------------------------------------------------------------- + TOTAL 0.15924661 543.85016375 + (elements=356,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 356864 +Total Bytes 61804 +BPFP 1.3855 bits/point +EBPFP 2.7710 equivalent bits/point +MSE 543.850164 +---------------------- -------------------------------------------------------- +Time: 0.279s Load: 0.002s, Pack+Encode: 0.131s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 543.8502 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,156B, BPFP=0.6478 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,428B, BPFP=2.2320 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,544B, BPFP=1.0649 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,152B, BPFP=2.1490 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,208B, BPFP=1.2644 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,904B, BPFP=2.0745 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,928B, BPFP=1.1803 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,032B, BPFP=2.1130 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,876B, BPFP=1.7656 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,912B, BPFP=2.0769 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,640B, BPFP=0.7143 +⌛️ [2/4] FRONTEND: Frontend time: 0.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.147s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09210638 58.52118389 + layer.0.v_cache 0.00001369 0.00841577 + layer.1.k_cache 0.01682195 6.24323625 + layer.1.v_cache 0.00000520 0.00334732 + layer.2.k_cache 0.00487203 0.98513324 + layer.2.v_cache 0.00001755 0.00933620 + layer.3.k_cache 0.04965206 3.83035983 + layer.3.v_cache 0.00001944 0.01035005 + layer.4.k_cache 0.00060442 0.19933398 + layer.4.v_cache 0.00005207 0.02074305 + layer.4.output 0.26182200 1037.57692308 + ------------------------------------------------------------------------------------- + TOTAL 0.11746581 431.34528830 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 71780 +BPFP 1.2687 bits/point +EBPFP 2.5375 equivalent bits/point +MSE 431.345288 +---------------------- -------------------------------------------------------- +Time: 0.281s Load: 0.002s, Pack+Encode: 0.132s, Decode+Unpack: 0.147s +---------------------- -------------------------------------------------------- +💾 Converting with 431.3453 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 47, 128) +Output shape: (1, 47, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.output: torch.Size([1, 47, 3584]) -> torch.Size([1, 1, 47, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,076B, BPFP=0.6902 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,076B, BPFP=2.3524 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,436B, BPFP=1.1423 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,904B, BPFP=2.2952 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,048B, BPFP=1.3457 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,656B, BPFP=2.2128 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,828B, BPFP=1.2726 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,748B, BPFP=2.2434 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,628B, BPFP=1.8710 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,608B, BPFP=2.1968 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 15,160B, BPFP=0.7200 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08461995 63.99186544 + layer.0.v_cache 0.00001366 0.00956016 + layer.1.k_cache 0.01838745 6.28535624 + layer.1.v_cache 0.00000609 0.00427805 + layer.2.k_cache 0.00780689 0.97515309 + layer.2.v_cache 0.00001932 0.01013171 + layer.3.k_cache 0.04292001 4.24492012 + layer.3.v_cache 0.00001942 0.01106911 + layer.4.k_cache 0.00059704 0.20991025 + layer.4.v_cache 0.00004850 0.02210321 + layer.4.output 0.29441516 1141.27954027 + ------------------------------------------------------------------------------------- + TOTAL 0.13031438 474.39536055 + (elements=409,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 409088 +Total Bytes 68168 +BPFP 1.3331 bits/point +EBPFP 2.6661 equivalent bits/point +MSE 474.395361 +---------------------- -------------------------------------------------------- +Time: 0.279s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 474.3954 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,292B, BPFP=0.6283 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,952B, BPFP=2.1798 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,804B, BPFP=1.0428 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,780B, BPFP=2.1327 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,572B, BPFP=1.2533 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,468B, BPFP=2.0471 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,356B, BPFP=1.1941 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,544B, BPFP=2.0680 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,268B, BPFP=1.7182 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,404B, BPFP=2.0296 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,156B, BPFP=0.7110 +⌛️ [2/4] FRONTEND: Frontend time: 0.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10653571 63.88635040 + layer.0.v_cache 0.00001396 0.00803166 + layer.1.k_cache 0.01513677 5.04215388 + layer.1.v_cache 0.00000574 0.00314109 + layer.2.k_cache 0.00241650 0.80615502 + layer.2.v_cache 0.00001997 0.00861135 + layer.3.k_cache 0.04250543 4.12211662 + layer.3.v_cache 0.00001889 0.00927459 + layer.4.k_cache 0.00060099 0.18675393 + layer.4.v_cache 0.00005317 0.02045527 + layer.4.output 0.25650722 938.56923559 + ------------------------------------------------------------------------------------- + TOTAL 0.11546222 390.82809958 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 77596 +BPFP 1.2512 bits/point +EBPFP 2.5025 equivalent bits/point +MSE 390.828100 +---------------------- -------------------------------------------------------- +Time: 0.280s Load: 0.003s, Pack+Encode: 0.132s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 390.8281 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,092B, BPFP=0.6810 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,208B, BPFP=2.3464 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,472B, BPFP=1.1302 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 6,992B, BPFP=2.2760 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,144B, BPFP=1.3490 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 6,784B, BPFP=2.2083 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 3,864B, BPFP=1.2578 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 6,888B, BPFP=2.2422 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 5,784B, BPFP=1.8828 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 6,844B, BPFP=2.2279 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 16,188B, BPFP=0.7528 +⌛️ [2/4] FRONTEND: Frontend time: 0.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10088900 63.86385091 + layer.0.v_cache 0.00001321 0.00767226 + layer.1.k_cache 0.01601078 6.76665815 + layer.1.v_cache 0.00000529 0.00306482 + layer.2.k_cache 0.00240732 0.99847062 + layer.2.v_cache 0.00001858 0.00874776 + layer.3.k_cache 0.04236705 3.84622637 + layer.3.v_cache 0.00001968 0.00925771 + layer.4.k_cache 0.00061965 0.19812230 + layer.4.v_cache 0.00005235 0.02035967 + layer.4.output 0.30498679 1116.12453497 + ------------------------------------------------------------------------------------- + TOTAL 0.13513591 464.03495149 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 70260 +BPFP 1.3454 bits/point +EBPFP 2.6907 equivalent bits/point +MSE 464.034951 +---------------------- -------------------------------------------------------- +Time: 0.281s Load: 0.003s, Pack+Encode: 0.132s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 464.0350 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 53, 128) +Output shape: (1, 53, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.output: torch.Size([1, 53, 3584]) -> torch.Size([1, 1, 53, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,216B, BPFP=0.6533 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,576B, BPFP=2.2335 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,664B, BPFP=1.0802 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,484B, BPFP=2.2064 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,396B, BPFP=1.2960 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,276B, BPFP=2.1450 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,164B, BPFP=1.2276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,336B, BPFP=2.1627 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,112B, BPFP=1.8019 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,204B, BPFP=2.1238 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,104B, BPFP=0.7204 +⌛️ [2/4] FRONTEND: Frontend time: 0.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.147s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10150213 58.66104990 + layer.0.v_cache 0.00001370 0.00959366 + layer.1.k_cache 0.01753690 6.00891689 + layer.1.v_cache 0.00000536 0.00372013 + layer.2.k_cache 0.00234856 0.96211063 + layer.2.v_cache 0.00001946 0.01065550 + layer.3.k_cache 0.03772523 4.55636165 + layer.3.v_cache 0.00001792 0.01072294 + layer.4.k_cache 0.00061365 0.20624929 + layer.4.v_cache 0.00006478 0.02277558 + layer.4.output 0.25806696 1014.38982480 + ------------------------------------------------------------------------------------- + TOTAL 0.11566567 421.83417234 + (elements=461,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 461312 +Total Bytes 74532 +BPFP 1.2925 bits/point +EBPFP 2.5850 equivalent bits/point +MSE 421.834172 +---------------------- -------------------------------------------------------- +Time: 0.281s Load: 0.002s, Pack+Encode: 0.132s, Decode+Unpack: 0.147s +---------------------- -------------------------------------------------------- +💾 Converting with 421.8342 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,508B, BPFP=0.6531 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,980B, BPFP=2.0781 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,104B, BPFP=1.0688 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,908B, BPFP=2.0594 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,836B, BPFP=1.2594 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,696B, BPFP=2.0042 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,792B, BPFP=1.2479 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,780B, BPFP=2.0260 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,604B, BPFP=1.7198 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,524B, BPFP=1.9594 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,380B, BPFP=0.7582 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.146s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11850098 59.79433187 + layer.0.v_cache 0.00001385 0.00822346 + layer.1.k_cache 0.01468890 4.56483663 + layer.1.v_cache 0.00000523 0.00317990 + layer.2.k_cache 0.01477869 0.95059007 + layer.2.v_cache 0.00001915 0.00925148 + layer.3.k_cache 0.14661819 3.50677643 + layer.3.v_cache 0.00001925 0.01035404 + layer.4.k_cache 0.00060183 0.18241224 + layer.4.v_cache 0.00005752 0.02103233 + layer.4.output 0.22966646 892.75372024 + ------------------------------------------------------------------------------------- + TOTAL 0.11193934 371.66629589 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 82112 +BPFP 1.2578 bits/point +EBPFP 2.5157 equivalent bits/point +MSE 371.666296 +---------------------- -------------------------------------------------------- +Time: 0.281s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.146s +---------------------- -------------------------------------------------------- +💾 Converting with 371.6663 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,324B, BPFP=0.5645 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,628B, BPFP=2.1447 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,820B, BPFP=0.9885 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,088B, BPFP=2.0530 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,004B, BPFP=1.1895 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,840B, BPFP=2.0109 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,492B, BPFP=1.1026 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,184B, BPFP=2.0693 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,096B, BPFP=1.7147 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,804B, BPFP=2.0048 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,952B, BPFP=0.5811 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10836371 60.92029870 + layer.0.v_cache 0.00001445 0.00791925 + layer.1.k_cache 0.06846340 5.05393119 + layer.1.v_cache 0.00000523 0.00283630 + layer.2.k_cache 0.00506382 0.80259978 + layer.2.v_cache 0.00001740 0.00736858 + layer.3.k_cache 0.03973729 3.37223484 + layer.3.v_cache 0.00001777 0.00895310 + layer.4.k_cache 0.00066797 0.17841016 + layer.4.v_cache 0.00005099 0.01852370 + layer.4.output 0.14786712 587.02130241 + ------------------------------------------------------------------------------------- + TOTAL 0.07396893 245.85424661 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 117232 +BPFP 1.1712 bits/point +EBPFP 2.3424 equivalent bits/point +MSE 245.854247 +---------------------- -------------------------------------------------------- +Time: 0.387s Load: 0.005s, Pack+Encode: 0.168s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 245.8542 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,656B, BPFP=0.6288 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,756B, BPFP=2.3097 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,304B, BPFP=1.0189 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,272B, BPFP=2.1951 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,164B, BPFP=1.2225 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,800B, BPFP=2.0833 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,976B, BPFP=1.1780 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,368B, BPFP=2.2178 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,372B, BPFP=1.7453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,828B, BPFP=2.0900 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,252B, BPFP=0.6511 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08728886 63.93232126 + layer.0.v_cache 0.00001734 0.00875800 + layer.1.k_cache 0.01499850 5.82813055 + layer.1.v_cache 0.00000528 0.00296405 + layer.2.k_cache 0.00240521 0.89625226 + layer.2.v_cache 0.00001739 0.00817007 + layer.3.k_cache 0.03126823 4.20601354 + layer.3.v_cache 0.00001843 0.00945802 + layer.4.k_cache 0.00061395 0.19851635 + layer.4.v_cache 0.00004830 0.02004699 + layer.4.output 0.21953388 818.84611742 + ------------------------------------------------------------------------------------- + TOTAL 0.09843639 341.59020312 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 89748 +BPFP 1.2498 bits/point +EBPFP 2.4997 equivalent bits/point +MSE 341.590203 +---------------------- -------------------------------------------------------- +Time: 0.443s Load: 0.003s, Pack+Encode: 0.168s, Decode+Unpack: 0.272s +---------------------- -------------------------------------------------------- +💾 Converting with 341.5902 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,620B, BPFP=0.6203 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,788B, BPFP=2.3172 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,340B, BPFP=1.0275 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,160B, BPFP=2.1686 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,224B, BPFP=1.2367 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 8,916B, BPFP=2.1108 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,048B, BPFP=1.1951 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,368B, BPFP=2.2178 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,560B, BPFP=1.7898 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,012B, BPFP=2.1335 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,980B, BPFP=0.6757 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.216s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08223097 62.51886541 + layer.0.v_cache 0.00001442 0.00865189 + layer.1.k_cache 0.01480860 5.92017619 + layer.1.v_cache 0.00000502 0.00283807 + layer.2.k_cache 0.00576455 0.87633855 + layer.2.v_cache 0.00002304 0.00801964 + layer.3.k_cache 0.03215047 4.22616161 + layer.3.v_cache 0.00001933 0.01005412 + layer.4.k_cache 0.00061510 0.19449601 + layer.4.v_cache 0.00004916 0.01964658 + layer.4.output 0.22077460 818.06642316 + ------------------------------------------------------------------------------------- + TOTAL 0.09888840 341.19118884 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 91016 +BPFP 1.2675 bits/point +EBPFP 2.5350 equivalent bits/point +MSE 341.191189 +---------------------- -------------------------------------------------------- +Time: 0.388s Load: 0.004s, Pack+Encode: 0.168s, Decode+Unpack: 0.216s +---------------------- -------------------------------------------------------- +💾 Converting with 341.1912 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,964B, BPFP=0.6344 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,320B, BPFP=2.2089 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,828B, BPFP=1.0334 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,664B, BPFP=2.0685 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,616B, BPFP=1.2021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,212B, BPFP=1.9717 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,328B, BPFP=1.1404 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,636B, BPFP=2.0625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,872B, BPFP=1.6849 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,272B, BPFP=1.9846 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,140B, BPFP=0.6158 +⌛️ [2/4] FRONTEND: Frontend time: 0.170s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10114760 61.61206389 + layer.0.v_cache 0.00001420 0.00800137 + layer.1.k_cache 0.03683303 5.51562834 + layer.1.v_cache 0.00000544 0.00317905 + layer.2.k_cache 0.00248627 0.88842125 + layer.2.v_cache 0.00001713 0.00814613 + layer.3.k_cache 0.02909129 3.99473467 + layer.3.v_cache 0.00001766 0.00978104 + layer.4.k_cache 0.00072613 0.19187245 + layer.4.v_cache 0.00005154 0.01995882 + layer.4.output 0.19731793 738.99712573 + ------------------------------------------------------------------------------------- + TOTAL 0.09127152 308.54303924 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 94852 +BPFP 1.1942 bits/point +EBPFP 2.3885 equivalent bits/point +MSE 308.543039 +---------------------- -------------------------------------------------------- +Time: 0.389s Load: 0.003s, Pack+Encode: 0.170s, Decode+Unpack: 0.215s +---------------------- -------------------------------------------------------- +💾 Converting with 308.5430 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,044B, BPFP=0.6258 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,852B, BPFP=2.2311 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,032B, BPFP=1.0345 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,248B, BPFP=2.1069 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,880B, BPFP=1.2089 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,852B, BPFP=2.0255 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,576B, BPFP=1.1464 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,328B, BPFP=2.1234 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,300B, BPFP=1.7064 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,812B, BPFP=2.0173 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,428B, BPFP=0.6293 +⌛️ [2/4] FRONTEND: Frontend time: 0.168s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.233s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10230959 57.66345857 + layer.0.v_cache 0.00001409 0.00800077 + layer.1.k_cache 0.07973289 5.28721739 + layer.1.v_cache 0.00000506 0.00268637 + layer.2.k_cache 0.00553133 0.85165988 + layer.2.v_cache 0.00001928 0.00754029 + layer.3.k_cache 0.03946657 3.87208758 + layer.3.v_cache 0.00001980 0.00920542 + layer.4.k_cache 0.00066525 0.18309349 + layer.4.v_cache 0.00005119 0.01863080 + layer.4.output 0.17985767 710.51110197 + ------------------------------------------------------------------------------------- + TOTAL 0.08745993 296.55772320 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 100352 +BPFP 1.2136 bits/point +EBPFP 2.4272 equivalent bits/point +MSE 296.557723 +---------------------- -------------------------------------------------------- +Time: 0.405s Load: 0.004s, Pack+Encode: 0.168s, Decode+Unpack: 0.233s +---------------------- -------------------------------------------------------- +💾 Converting with 296.5577 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,080B, BPFP=0.6250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,796B, BPFP=2.1907 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,920B, BPFP=0.9984 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,152B, BPFP=2.0601 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,716B, BPFP=1.1599 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,800B, BPFP=1.9886 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,356B, BPFP=1.0869 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,172B, BPFP=2.0641 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,228B, BPFP=1.6696 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,720B, BPFP=1.9724 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,128B, BPFP=0.5835 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633976 60.90209644 + layer.0.v_cache 0.00001461 0.00773150 + layer.1.k_cache 0.07748085 5.08172607 + layer.1.v_cache 0.00000515 0.00267585 + layer.2.k_cache 0.00505460 0.87821792 + layer.2.v_cache 0.00001651 0.00720384 + layer.3.k_cache 0.02793909 3.44014354 + layer.3.v_cache 0.00001809 0.00850007 + layer.4.k_cache 0.00070819 0.18106620 + layer.4.v_cache 0.00004716 0.01748700 + layer.4.output 0.18765952 697.91575835 + ------------------------------------------------------------------------------------- + TOTAL 0.09007298 291.52571511 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 98068 +BPFP 1.1706 bits/point +EBPFP 2.3412 equivalent bits/point +MSE 291.525715 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 291.5257 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,880B, BPFP=0.6429 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,260B, BPFP=2.2902 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,696B, BPFP=1.0482 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,304B, BPFP=2.0768 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,484B, BPFP=1.2241 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,164B, BPFP=2.0455 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,200B, BPFP=1.1607 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,476B, BPFP=2.1152 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,608B, BPFP=1.6982 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,012B, BPFP=2.0116 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,148B, BPFP=0.6106 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08399124 61.92724609 + layer.0.v_cache 0.00001671 0.00818676 + layer.1.k_cache 0.03849935 5.47667759 + layer.1.v_cache 0.00000503 0.00287291 + layer.2.k_cache 0.00738276 0.87277511 + layer.2.v_cache 0.00001692 0.00747792 + layer.3.k_cache 0.04914829 3.94813581 + layer.3.v_cache 0.00001713 0.00890219 + layer.4.k_cache 0.00063731 0.17966945 + layer.4.v_cache 0.00004711 0.01811958 + layer.4.output 0.21328995 768.96798469 + ------------------------------------------------------------------------------------- + TOTAL 0.09839950 320.89564449 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 92232 +BPFP 1.2110 bits/point +EBPFP 2.4221 equivalent bits/point +MSE 320.895644 +---------------------- -------------------------------------------------------- +Time: 0.369s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 320.8956 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 56, 128) +Output shape: (1, 56, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.output: torch.Size([1, 56, 3584]) -> torch.Size([1, 1, 56, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,284B, BPFP=0.6373 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,756B, BPFP=2.1641 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,716B, BPFP=1.0368 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,348B, BPFP=2.0502 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,356B, BPFP=1.2154 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,112B, BPFP=1.9844 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,228B, BPFP=1.1797 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,344B, BPFP=2.0491 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,072B, BPFP=1.6942 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,064B, BPFP=1.9710 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 17,176B, BPFP=0.6846 +⌛️ [2/4] FRONTEND: Frontend time: 0.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.147s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13768739 67.06219046 + layer.0.v_cache 0.00001432 0.00852022 + layer.1.k_cache 0.01638686 5.41195406 + layer.1.v_cache 0.00000533 0.00305835 + layer.2.k_cache 0.00251956 0.83958796 + layer.2.v_cache 0.00001745 0.00839407 + layer.3.k_cache 0.08599430 3.97156143 + layer.3.v_cache 0.00002012 0.01068049 + layer.4.k_cache 0.00061995 0.19263213 + layer.4.v_cache 0.00004798 0.01933283 + layer.4.output 0.24256272 962.13392857 + ------------------------------------------------------------------------------------- + TOTAL 0.11419131 400.73325953 + (elements=487,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 487424 +Total Bytes 74456 +BPFP 1.2220 bits/point +EBPFP 2.4441 equivalent bits/point +MSE 400.733260 +---------------------- -------------------------------------------------------- +Time: 0.282s Load: 0.003s, Pack+Encode: 0.132s, Decode+Unpack: 0.147s +---------------------- -------------------------------------------------------- +💾 Converting with 400.7333 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,152B, BPFP=0.6156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,200B, BPFP=2.1875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,104B, BPFP=0.9969 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 10,664B, BPFP=2.0828 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,036B, BPFP=1.1789 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,392B, BPFP=2.0297 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,784B, BPFP=1.1297 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 10,732B, BPFP=2.0961 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,944B, BPFP=1.7469 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,332B, BPFP=2.0180 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,932B, BPFP=0.6119 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.203s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09278157 60.04577026 + layer.0.v_cache 0.00001504 0.00764730 + layer.1.k_cache 0.07523044 5.36544647 + layer.1.v_cache 0.00000517 0.00295282 + layer.2.k_cache 0.00381728 0.81857977 + layer.2.v_cache 0.00001789 0.00775046 + layer.3.k_cache 0.03327554 4.17104492 + layer.3.v_cache 0.00001939 0.00890891 + layer.4.k_cache 0.00067710 0.18253994 + layer.4.v_cache 0.00005192 0.01857298 + layer.4.output 0.16994183 676.11300223 + ------------------------------------------------------------------------------------- + TOTAL 0.08208730 282.55413114 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 104272 +BPFP 1.1980 bits/point +EBPFP 2.3960 equivalent bits/point +MSE 282.554131 +---------------------- -------------------------------------------------------- +Time: 0.367s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.203s +---------------------- -------------------------------------------------------- +💾 Converting with 282.5541 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,380B, BPFP=0.6303 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 7,956B, BPFP=2.1070 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 3,924B, BPFP=1.0392 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,720B, BPFP=2.0445 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,640B, BPFP=1.2288 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,476B, BPFP=1.9799 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,484B, BPFP=1.1875 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,636B, BPFP=2.0222 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,260B, BPFP=1.6578 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,364B, BPFP=1.9502 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 19,316B, BPFP=0.7308 +⌛️ [2/4] FRONTEND: Frontend time: 0.166s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.150s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12468526 63.78383210 + layer.0.v_cache 0.00001451 0.00919409 + layer.1.k_cache 0.01313243 4.83815752 + layer.1.v_cache 0.00000580 0.00326720 + layer.2.k_cache 0.00906488 0.77199613 + layer.2.v_cache 0.00002099 0.00880170 + layer.3.k_cache 0.01569995 4.03906664 + layer.3.v_cache 0.00002002 0.01064287 + layer.4.k_cache 0.00064964 0.18586210 + layer.4.v_cache 0.00005478 0.02089861 + layer.4.output 0.23031524 909.96988499 + ------------------------------------------------------------------------------------- + TOTAL 0.10444441 379.02711258 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 79156 +BPFP 1.2331 bits/point +EBPFP 2.4662 equivalent bits/point +MSE 379.027113 +---------------------- -------------------------------------------------------- +Time: 0.319s Load: 0.003s, Pack+Encode: 0.166s, Decode+Unpack: 0.150s +---------------------- -------------------------------------------------------- +💾 Converting with 379.0271 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,516B, BPFP=0.6552 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 8,076B, BPFP=2.1031 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,108B, BPFP=1.0698 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 7,872B, BPFP=2.0500 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 4,792B, BPFP=1.2479 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 7,728B, BPFP=2.0125 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,692B, BPFP=1.2219 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 7,848B, BPFP=2.0438 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 6,592B, BPFP=1.7167 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 7,572B, BPFP=1.9719 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 20,544B, BPFP=0.7643 +⌛️ [2/4] FRONTEND: Frontend time: 0.131s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.145s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15164433 60.89583333 + layer.0.v_cache 0.00001435 0.00866719 + layer.1.k_cache 0.01704682 4.64834696 + layer.1.v_cache 0.00000573 0.00333811 + layer.2.k_cache 0.00240009 0.85982335 + layer.2.v_cache 0.00001845 0.00867240 + layer.3.k_cache 0.02161660 4.26651866 + layer.3.v_cache 0.00001985 0.01004416 + layer.4.k_cache 0.00061459 0.19362512 + layer.4.v_cache 0.00005275 0.02156537 + layer.4.output 0.22653077 896.45215774 + ------------------------------------------------------------------------------------- + TOTAL 0.10465582 373.29891405 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 82340 +BPFP 1.2613 bits/point +EBPFP 2.5227 equivalent bits/point +MSE 373.298914 +---------------------- -------------------------------------------------------- +Time: 0.280s Load: 0.003s, Pack+Encode: 0.131s, Decode+Unpack: 0.145s +---------------------- -------------------------------------------------------- +💾 Converting with 373.2989 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,012B, BPFP=0.6447 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 9,896B, BPFP=2.1182 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,784B, BPFP=1.0240 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,520B, BPFP=2.0377 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,448B, BPFP=1.1661 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,012B, BPFP=1.9289 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,236B, BPFP=1.1207 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,548B, BPFP=2.0437 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,840B, BPFP=1.6781 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,172B, BPFP=1.9632 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,884B, BPFP=0.5774 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08869729 62.70685734 + layer.0.v_cache 0.00001594 0.00712880 + layer.1.k_cache 0.08105310 5.37152016 + layer.1.v_cache 0.00000494 0.00261328 + layer.2.k_cache 0.00256802 0.85061881 + layer.2.v_cache 0.00001661 0.00708667 + layer.3.k_cache 0.01835501 3.99554694 + layer.3.v_cache 0.00001861 0.00827790 + layer.4.k_cache 0.00078324 0.19162364 + layer.4.v_cache 0.00004769 0.01676755 + layer.4.output 0.18611241 739.82283513 + ------------------------------------------------------------------------------------- + TOTAL 0.08790278 308.93634629 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 92352 +BPFP 1.1628 bits/point +EBPFP 2.3255 equivalent bits/point +MSE 308.936346 +---------------------- -------------------------------------------------------- +Time: 0.368s Load: 0.004s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 308.9363 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,136B, BPFP=0.5052 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,668B, BPFP=2.0406 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,468B, BPFP=0.8808 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,076B, BPFP=1.9452 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,848B, BPFP=1.1031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,776B, BPFP=1.8969 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,476B, BPFP=1.0432 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,144B, BPFP=1.9562 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,052B, BPFP=1.6192 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,588B, BPFP=1.8666 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,784B, BPFP=0.5703 +⌛️ [2/4] FRONTEND: Frontend time: 0.160s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17228427 58.54305433 + layer.0.v_cache 0.00001543 0.00718370 + layer.1.k_cache 0.04621154 5.05571440 + layer.1.v_cache 0.00000545 0.00249705 + layer.2.k_cache 0.00739691 0.79047417 + layer.2.v_cache 0.00001797 0.00693278 + layer.3.k_cache 0.03096496 3.29383016 + layer.3.v_cache 0.00001890 0.00859263 + layer.4.k_cache 0.00071091 0.17187718 + layer.4.v_cache 0.00004900 0.01601687 + layer.4.output 0.02450962 572.59282953 + ------------------------------------------------------------------------------------- + TOTAL 0.02524957 239.76741059 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 117016 +BPFP 1.1088 bits/point +EBPFP 2.2176 equivalent bits/point +MSE 239.767411 +---------------------- -------------------------------------------------------- +Time: 0.370s Load: 0.005s, Pack+Encode: 0.160s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 239.7674 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,288B, BPFP=0.5524 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,800B, BPFP=2.1505 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,924B, BPFP=0.9953 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 12,392B, BPFP=2.0820 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,976B, BPFP=1.1720 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,976B, BPFP=2.0121 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,728B, BPFP=1.1304 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,332B, BPFP=2.0719 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,348B, BPFP=1.7386 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,852B, BPFP=1.9913 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,572B, BPFP=0.5658 +⌛️ [2/4] FRONTEND: Frontend time: 0.205s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16174046 57.87829196 + layer.0.v_cache 0.00001612 0.00769031 + layer.1.k_cache 0.03352992 4.67344255 + layer.1.v_cache 0.00000563 0.00285750 + layer.2.k_cache 0.00496730 0.73422750 + layer.2.v_cache 0.00001740 0.00763866 + layer.3.k_cache 0.01521110 3.75219891 + layer.3.v_cache 0.00001939 0.00931597 + layer.4.k_cache 0.00066431 0.18939225 + layer.4.v_cache 0.00005067 0.01915233 + layer.4.output 0.14624557 581.03398618 + ------------------------------------------------------------------------------------- + TOTAL 0.07293772 243.20659477 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 118188 +BPFP 1.1681 bits/point +EBPFP 2.3361 equivalent bits/point +MSE 243.206595 +---------------------- -------------------------------------------------------- +Time: 0.417s Load: 0.005s, Pack+Encode: 0.205s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 243.2066 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,308B, BPFP=0.6010 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,868B, BPFP=2.1562 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,432B, BPFP=0.9869 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,236B, BPFP=2.0414 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,304B, BPFP=1.1453 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,764B, BPFP=1.9557 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,008B, BPFP=1.0916 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,216B, BPFP=2.0378 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,200B, BPFP=1.6715 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,892B, BPFP=1.9789 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,796B, BPFP=0.5917 +⌛️ [2/4] FRONTEND: Frontend time: 0.161s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07121617 58.18899323 + layer.0.v_cache 0.00001681 0.00742258 + layer.1.k_cache 0.06877367 5.05152467 + layer.1.v_cache 0.00000530 0.00250420 + layer.2.k_cache 0.00486504 0.89235847 + layer.2.v_cache 0.00001749 0.00685270 + layer.3.k_cache 0.08833314 3.24553450 + layer.3.v_cache 0.00001723 0.00819304 + layer.4.k_cache 0.00077278 0.16956529 + layer.4.v_cache 0.00004595 0.01668575 + layer.4.output 0.15812202 628.31597799 + ------------------------------------------------------------------------------------- + TOTAL 0.07887751 262.69420473 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 109024 +BPFP 1.1652 bits/point +EBPFP 2.3304 equivalent bits/point +MSE 262.694205 +---------------------- -------------------------------------------------------- +Time: 0.371s Load: 0.005s, Pack+Encode: 0.161s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 262.6942 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,384B, BPFP=0.5235 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,736B, BPFP=2.1250 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,040B, BPFP=0.9344 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,044B, BPFP=2.0179 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,156B, BPFP=1.1071 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,528B, BPFP=1.9381 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,884B, BPFP=1.0650 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 12,968B, BPFP=2.0062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,632B, BPFP=1.6448 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,472B, BPFP=1.9295 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,032B, BPFP=0.5311 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13964315 58.06042118 + layer.0.v_cache 0.00001432 0.00795503 + layer.1.k_cache 0.04810369 5.03705197 + layer.1.v_cache 0.00000572 0.00327680 + layer.2.k_cache 0.00492308 0.80385023 + layer.2.v_cache 0.00001777 0.00802860 + layer.3.k_cache 0.04255116 3.94074030 + layer.3.v_cache 0.00001880 0.00944611 + layer.4.k_cache 0.00073971 0.19347559 + layer.4.v_cache 0.00004573 0.01952301 + layer.4.output 11.28739116 530.71291549 + ------------------------------------------------------------------------------------- + TOTAL 4.66163537 222.53377513 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 122876 +BPFP 1.1182 bits/point +EBPFP 2.2364 equivalent bits/point +MSE 222.533775 +---------------------- -------------------------------------------------------- +Time: 0.377s Load: 0.005s, Pack+Encode: 0.165s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 222.5338 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,356B, BPFP=0.5959 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,172B, BPFP=2.1612 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,540B, BPFP=0.9837 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,544B, BPFP=2.0497 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,560B, BPFP=1.1648 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,196B, BPFP=1.9879 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,248B, BPFP=1.1094 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,548B, BPFP=2.0504 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,560B, BPFP=1.6974 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,008B, BPFP=1.9545 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,848B, BPFP=0.5795 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14145413 60.48337624 + layer.0.v_cache 0.00001516 0.00729153 + layer.1.k_cache 0.10647225 4.57426349 + layer.1.v_cache 0.00000560 0.00269610 + layer.2.k_cache 0.00244946 0.79213871 + layer.2.v_cache 0.00001685 0.00691467 + layer.3.k_cache 0.01581429 2.73955865 + layer.3.v_cache 0.00001841 0.00867751 + layer.4.k_cache 0.00072023 0.17239035 + layer.4.v_cache 0.00004652 0.01563100 + layer.4.output 0.15451367 614.14752435 + ------------------------------------------------------------------------------------- + TOTAL 0.07932992 256.93150639 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 111580 +BPFP 1.1654 bits/point +EBPFP 2.3308 equivalent bits/point +MSE 256.931506 +---------------------- -------------------------------------------------------- +Time: 0.374s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 256.9315 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,968B, BPFP=0.6267 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,348B, BPFP=2.1850 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,876B, BPFP=1.0296 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,868B, BPFP=2.0836 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,548B, BPFP=1.1715 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,328B, BPFP=1.9696 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,344B, BPFP=1.1284 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,660B, BPFP=2.0397 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,820B, BPFP=1.6512 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 9,308B, BPFP=1.9654 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 18,604B, BPFP=0.5612 +⌛️ [2/4] FRONTEND: Frontend time: 0.162s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10237078 62.29788455 + layer.0.v_cache 0.00001391 0.00774519 + layer.1.k_cache 0.05953315 5.31150447 + layer.1.v_cache 0.00000517 0.00287356 + layer.2.k_cache 0.00728044 0.81019757 + layer.2.v_cache 0.00001648 0.00733085 + layer.3.k_cache 0.03101032 3.98583902 + layer.3.v_cache 0.00001666 0.00904380 + layer.4.k_cache 0.00072555 0.17696578 + layer.4.v_cache 0.00004870 0.01789244 + layer.4.output 0.18364459 727.23497828 + ------------------------------------------------------------------------------------- + TOTAL 0.08744314 303.72188972 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 93672 +BPFP 1.1635 bits/point +EBPFP 2.3269 equivalent bits/point +MSE 303.721890 +---------------------- -------------------------------------------------------- +Time: 0.373s Load: 0.004s, Pack+Encode: 0.162s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 303.7219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 111, 128) +Output shape: (1, 111, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.output: torch.Size([1, 111, 3584]) -> torch.Size([1, 1, 111, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,744B, BPFP=0.5270 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,816B, BPFP=2.0856 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,508B, BPFP=0.9161 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,148B, BPFP=1.9916 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,788B, BPFP=1.0963 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,632B, BPFP=1.9189 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,488B, BPFP=1.0541 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,012B, BPFP=1.9724 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,588B, BPFP=1.6312 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,524B, BPFP=1.9037 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,280B, BPFP=0.5486 +⌛️ [2/4] FRONTEND: Frontend time: 0.164s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.205s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11107565 53.65239654 + layer.0.v_cache 0.00001597 0.00697155 + layer.1.k_cache 0.04333284 4.85063653 + layer.1.v_cache 0.00000558 0.00259372 + layer.2.k_cache 0.00477623 0.78272502 + layer.2.v_cache 0.00001758 0.00707771 + layer.3.k_cache 0.01020195 3.27218930 + layer.3.v_cache 0.00001870 0.00776758 + layer.4.k_cache 0.00079488 0.16508776 + layer.4.v_cache 0.00005041 0.01661470 + layer.4.output 10.27057757 482.50004022 + ------------------------------------------------------------------------------------- + TOTAL 4.23907840 202.36849070 + (elements=966,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 966144 +Total Bytes 134528 +BPFP 1.1139 bits/point +EBPFP 2.2279 equivalent bits/point +MSE 202.368491 +---------------------- -------------------------------------------------------- +Time: 0.375s Load: 0.005s, Pack+Encode: 0.164s, Decode+Unpack: 0.205s +---------------------- -------------------------------------------------------- +💾 Converting with 202.3685 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,908B, BPFP=0.6400 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 10,076B, BPFP=2.2174 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,812B, BPFP=1.0590 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 9,392B, BPFP=2.0669 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,612B, BPFP=1.2350 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 9,100B, BPFP=2.0026 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,200B, BPFP=1.1444 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 9,512B, BPFP=2.0933 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 7,624B, BPFP=1.6778 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 8,940B, BPFP=1.9674 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,568B, BPFP=0.6781 +⌛️ [2/4] FRONTEND: Frontend time: 0.163s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09375787 61.21753961 + layer.0.v_cache 0.00001622 0.00746284 + layer.1.k_cache 0.05984993 5.13152732 + layer.1.v_cache 0.00000512 0.00277893 + layer.2.k_cache 0.00243208 0.83601755 + layer.2.v_cache 0.00001865 0.00810610 + layer.3.k_cache 0.04944374 3.22206653 + layer.3.v_cache 0.00001908 0.00919009 + layer.4.k_cache 0.00084877 0.18769673 + layer.4.v_cache 0.00005139 0.01881363 + layer.4.output 0.19371969 761.17593058 + ------------------------------------------------------------------------------------- + TOTAL 0.09191063 317.58074785 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 94744 +BPFP 1.2265 bits/point +EBPFP 2.4530 equivalent bits/point +MSE 317.580748 +---------------------- -------------------------------------------------------- +Time: 0.375s Load: 0.004s, Pack+Encode: 0.163s, Decode+Unpack: 0.208s +---------------------- -------------------------------------------------------- +💾 Converting with 317.5807 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 109, 128) +Output shape: (1, 109, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.output: torch.Size([1, 109, 3584]) -> torch.Size([1, 1, 109, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,724B, BPFP=0.5338 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,432B, BPFP=2.0688 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,556B, BPFP=0.9398 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,000B, BPFP=2.0069 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,784B, BPFP=1.1158 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,492B, BPFP=1.9341 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,408B, BPFP=1.0619 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,944B, BPFP=1.9989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,380B, BPFP=1.6313 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,412B, BPFP=1.9226 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,876B, BPFP=0.5913 +⌛️ [2/4] FRONTEND: Frontend time: 0.165s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.207s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860953 54.55834289 + layer.0.v_cache 0.00001500 0.00720956 + layer.1.k_cache 0.05917679 4.88107160 + layer.1.v_cache 0.00000528 0.00276222 + layer.2.k_cache 0.01253135 0.71401922 + layer.2.v_cache 0.00001869 0.00721855 + layer.3.k_cache 0.07406004 3.44674263 + layer.3.v_cache 0.00001854 0.00824209 + layer.4.k_cache 0.00071234 0.16620454 + layer.4.v_cache 0.00004916 0.01695480 + layer.4.output 10.45909253 490.14322575 + ------------------------------------------------------------------------------------- + TOTAL 4.32287320 205.57713814 + (elements=948,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 948736 +Total Bytes 135008 +BPFP 1.1384 bits/point +EBPFP 2.2768 equivalent bits/point +MSE 205.577138 +---------------------- -------------------------------------------------------- +Time: 0.378s Load: 0.006s, Pack+Encode: 0.165s, Decode+Unpack: 0.207s +---------------------- -------------------------------------------------------- +💾 Converting with 205.5771 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst + to output-fixed/kimiaudio/lambda0.01/hyperprior-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.2083 bits/point +Avg EBPFP 2.4167 equivalent bits/point +Avg MSE 316.722616 +Avg Time 0.359s +------------------------ ---------------------------- diff --git a/lambda0.02/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.02/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..75810b7efc0a2a4f3b51c2acfca816929696eff1 --- /dev/null +++ b/lambda0.02/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 520 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench +Output output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench +---------------- ------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,748B, BPFP=0.9159 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,284B, BPFP=3.7199 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,868B, BPFP=1.5177 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,068B, BPFP=3.6782 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,024B, BPFP=1.7407 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,108B, BPFP=3.4931 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,400B, BPFP=1.6204 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,612B, BPFP=3.5903 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,692B, BPFP=2.6412 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,072B, BPFP=3.4861 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,004B, BPFP=0.8819 +⌛️ [2/4] FRONTEND: Frontend time: 1.931s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.025s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12089630 29.03925842 + layer.0.v_cache 0.00001394 0.00231033 + layer.1.k_cache 0.01174029 2.12436233 + layer.1.v_cache 0.00000579 0.00094770 + layer.2.k_cache 0.00835781 0.66938089 + layer.2.v_cache 0.00001852 0.00302055 + layer.3.k_cache 0.02896961 3.04913971 + layer.3.v_cache 0.00001872 0.00312345 + layer.4.k_cache 0.00063445 0.07750489 + layer.4.v_cache 0.00005063 0.00660665 + layer.4.output 0.17246709 594.62924383 + ------------------------------------------------------------------------------------- + TOTAL 0.08105739 246.90472716 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 168880 +BPFP 1.9163 bits/point +EBPFP 3.8326 equivalent bits/point +MSE 246.904727 +---------------------- -------------------------------------------------------- +Time: 2.960s Load: 0.004s, Pack+Encode: 1.931s, Decode+Unpack: 1.025s +---------------------- -------------------------------------------------------- +💾 Converting with 246.9047 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,732B, BPFP=0.9242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,652B, BPFP=3.8383 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,980B, BPFP=1.5586 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,692B, BPFP=3.6508 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,372B, BPFP=1.8305 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,400B, BPFP=3.5938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,576B, BPFP=1.6750 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,840B, BPFP=3.6797 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,888B, BPFP=2.7125 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,936B, BPFP=3.5031 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,900B, BPFP=0.8901 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08029544 27.62721863 + layer.0.v_cache 0.00001376 0.00220121 + layer.1.k_cache 0.03632395 2.06322269 + layer.1.v_cache 0.00000563 0.00101627 + layer.2.k_cache 0.00522719 0.65839887 + layer.2.v_cache 0.00001997 0.00319521 + layer.3.k_cache 0.02707962 2.81674843 + layer.3.v_cache 0.00001843 0.00317194 + layer.4.k_cache 0.00063057 0.07865396 + layer.4.v_cache 0.00005085 0.00709719 + layer.4.output 0.18451144 466.43777902 + ------------------------------------------------------------------------------------- + TOTAL 0.08477915 194.01913985 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 169968 +BPFP 1.9528 bits/point +EBPFP 3.9055 equivalent bits/point +MSE 194.019140 +---------------------- -------------------------------------------------------- +Time: 2.502s Load: 0.005s, Pack+Encode: 1.490s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 194.0191 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,800B, BPFP=0.8721 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,508B, BPFP=3.7260 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,108B, BPFP=1.4731 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,428B, BPFP=3.5298 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,600B, BPFP=1.7442 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,060B, BPFP=3.4629 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,792B, BPFP=1.5974 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,424B, BPFP=3.5291 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,884B, BPFP=2.7042 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,124B, BPFP=3.4746 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,532B, BPFP=0.8444 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10874708 26.55297852 + layer.0.v_cache 0.00001608 0.00223245 + layer.1.k_cache 0.03270756 1.74879686 + layer.1.v_cache 0.00000572 0.00095461 + layer.2.k_cache 0.00777350 0.69125508 + layer.2.v_cache 0.00002080 0.00301866 + layer.3.k_cache 0.05920834 3.06920251 + layer.3.v_cache 0.00001846 0.00310451 + layer.4.k_cache 0.00062493 0.08047399 + layer.4.v_cache 0.00005036 0.00636903 + layer.4.output 0.17137358 566.96864618 + ------------------------------------------------------------------------------------- + TOTAL 0.08286988 235.34934762 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 176260 +BPFP 1.8838 bits/point +EBPFP 3.7675 equivalent bits/point +MSE 235.349348 +---------------------- -------------------------------------------------------- +Time: 2.518s Load: 0.003s, Pack+Encode: 1.503s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 235.3493 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,624B, BPFP=0.9263 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,136B, BPFP=3.8333 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,736B, BPFP=1.5497 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,676B, BPFP=3.7412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,024B, BPFP=1.8077 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,268B, BPFP=3.6595 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,296B, BPFP=1.6619 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,632B, BPFP=3.7324 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,632B, BPFP=2.7308 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,752B, BPFP=3.5561 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,856B, BPFP=0.9116 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12073612 26.06916222 + layer.0.v_cache 0.00001348 0.00222114 + layer.1.k_cache 0.03279715 1.88386927 + layer.1.v_cache 0.00000568 0.00100601 + layer.2.k_cache 0.00696802 0.71956126 + layer.2.v_cache 0.00001833 0.00296548 + layer.3.k_cache 0.03227850 2.97890922 + layer.3.v_cache 0.00001834 0.00324567 + layer.4.k_cache 0.00060115 0.07927970 + layer.4.v_cache 0.00005191 0.00669164 + layer.4.output 0.18320683 643.46274038 + ------------------------------------------------------------------------------------- + TOTAL 0.08681979 266.82271143 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 167632 +BPFP 1.9753 bits/point +EBPFP 3.9506 equivalent bits/point +MSE 266.822711 +---------------------- -------------------------------------------------------- +Time: 2.508s Load: 0.005s, Pack+Encode: 1.490s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 266.8227 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,636B, BPFP=0.9287 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,912B, BPFP=3.7885 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,740B, BPFP=1.5505 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,260B, BPFP=3.6579 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,908B, BPFP=1.7845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,216B, BPFP=3.4487 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,268B, BPFP=1.6562 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,980B, BPFP=3.6018 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,636B, BPFP=2.7316 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,384B, BPFP=3.4824 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,592B, BPFP=0.9041 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10595027 25.35942039 + layer.0.v_cache 0.00001376 0.00223613 + layer.1.k_cache 0.03307601 2.15540333 + layer.1.v_cache 0.00000550 0.00091692 + layer.2.k_cache 0.00695119 0.73931689 + layer.2.v_cache 0.00001929 0.00285348 + layer.3.k_cache 0.03450640 3.07349846 + layer.3.v_cache 0.00001844 0.00311360 + layer.4.k_cache 0.00061739 0.07536617 + layer.4.v_cache 0.00005297 0.00666167 + layer.4.output 0.18589628 638.16277473 + ------------------------------------------------------------------------------------- + TOTAL 0.08720501 264.62107118 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 164532 +BPFP 1.9388 bits/point +EBPFP 3.8775 equivalent bits/point +MSE 264.621071 +---------------------- -------------------------------------------------------- +Time: 2.518s Load: 0.005s, Pack+Encode: 1.496s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 264.6211 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,684B, BPFP=0.9264 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,584B, BPFP=3.8734 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,832B, BPFP=1.5491 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,048B, BPFP=3.7674 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,192B, BPFP=1.8180 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,584B, BPFP=3.6756 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,448B, BPFP=1.6709 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,160B, BPFP=3.7896 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,892B, BPFP=2.7476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,304B, BPFP=3.6203 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,228B, BPFP=0.8823 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07772056 24.25076796 + layer.0.v_cache 0.00001351 0.00232224 + layer.1.k_cache 0.03518496 1.98481152 + layer.1.v_cache 0.00000556 0.00104656 + layer.2.k_cache 0.00230807 0.74927545 + layer.2.v_cache 0.00001833 0.00291183 + layer.3.k_cache 0.04514034 2.42115262 + layer.3.v_cache 0.00001901 0.00327263 + layer.4.k_cache 0.00062178 0.07986191 + layer.4.v_cache 0.00004990 0.00679478 + layer.4.output 0.18426369 580.03752260 + ------------------------------------------------------------------------------------- + TOTAL 0.08534870 240.57440445 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 169956 +BPFP 1.9773 bits/point +EBPFP 3.9547 equivalent bits/point +MSE 240.574404 +---------------------- -------------------------------------------------------- +Time: 2.524s Load: 0.005s, Pack+Encode: 1.503s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 240.5744 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,600B, BPFP=0.9334 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,996B, BPFP=3.8547 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,756B, BPFP=1.5739 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,420B, BPFP=3.7378 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,980B, BPFP=1.8222 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,080B, BPFP=3.6688 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,260B, BPFP=1.6761 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,372B, BPFP=3.7281 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,424B, BPFP=2.7240 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,948B, BPFP=3.6420 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,832B, BPFP=0.9228 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14568528 27.25152826 + layer.0.v_cache 0.00001322 0.00223768 + layer.1.k_cache 0.01379078 2.00900427 + layer.1.v_cache 0.00000548 0.00092163 + layer.2.k_cache 0.00727733 0.63906741 + layer.2.v_cache 0.00001786 0.00304409 + layer.3.k_cache 0.03082054 2.92600914 + layer.3.v_cache 0.00001879 0.00315081 + layer.4.k_cache 0.00061820 0.07775299 + layer.4.v_cache 0.00005377 0.00676174 + layer.4.output 0.19662610 649.29330937 + ------------------------------------------------------------------------------------- + TOTAL 0.09262847 269.29250845 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 166668 +BPFP 1.9894 bits/point +EBPFP 3.9789 equivalent bits/point +MSE 269.292508 +---------------------- -------------------------------------------------------- +Time: 2.516s Load: 0.004s, Pack+Encode: 1.503s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 269.2925 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,716B, BPFP=0.9211 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,500B, BPFP=3.8086 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,988B, BPFP=1.5602 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,240B, BPFP=3.7578 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,216B, BPFP=1.8000 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,692B, BPFP=3.6508 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,480B, BPFP=1.6562 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,660B, BPFP=3.6445 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,708B, BPFP=2.6773 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,864B, BPFP=3.4891 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,756B, BPFP=0.9140 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.008s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10149734 24.45495605 + layer.0.v_cache 0.00001421 0.00240324 + layer.1.k_cache 0.01324784 1.97681541 + layer.1.v_cache 0.00000586 0.00105698 + layer.2.k_cache 0.00706157 0.67414036 + layer.2.v_cache 0.00001822 0.00303455 + layer.3.k_cache 0.02712160 3.16719360 + layer.3.v_cache 0.00002016 0.00330762 + layer.4.k_cache 0.00063687 0.07699375 + layer.4.v_cache 0.00005451 0.00674748 + layer.4.output 0.17858309 582.07890625 + ------------------------------------------------------------------------------------- + TOTAL 0.08233881 241.46582311 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 170820 +BPFP 1.9625 bits/point +EBPFP 3.9251 equivalent bits/point +MSE 241.465823 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.004s, Pack+Encode: 1.491s, Decode+Unpack: 1.008s +---------------------- -------------------------------------------------------- +💾 Converting with 241.4658 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,884B, BPFP=0.8772 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,380B, BPFP=3.6602 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,484B, BPFP=1.5237 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,080B, BPFP=3.6063 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,868B, BPFP=1.7723 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,516B, BPFP=3.5050 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,184B, BPFP=1.6494 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,324B, BPFP=3.6501 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,156B, BPFP=2.7220 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,772B, BPFP=3.5510 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,720B, BPFP=0.8908 +⌛️ [2/4] FRONTEND: Frontend time: 1.499s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10772817 26.67753233 + layer.0.v_cache 0.00001348 0.00230585 + layer.1.k_cache 0.03276526 1.90924774 + layer.1.v_cache 0.00000562 0.00101319 + layer.2.k_cache 0.01020141 0.66346535 + layer.2.v_cache 0.00001934 0.00293990 + layer.3.k_cache 0.04694761 3.20656682 + layer.3.v_cache 0.00001910 0.00324772 + layer.4.k_cache 0.00062564 0.07684870 + layer.4.v_cache 0.00005323 0.00656473 + layer.4.output 0.17166949 395.83938834 + ------------------------------------------------------------------------------------- + TOTAL 0.08235678 164.90737945 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 182368 +BPFP 1.9266 bits/point +EBPFP 3.8533 equivalent bits/point +MSE 164.907379 +---------------------- -------------------------------------------------------- +Time: 2.517s Load: 0.005s, Pack+Encode: 1.499s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 164.9074 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,800B, BPFP=0.9036 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,648B, BPFP=3.6988 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,072B, BPFP=1.5196 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,788B, BPFP=3.7252 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,496B, BPFP=1.7877 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,100B, BPFP=3.5956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,820B, BPFP=1.6604 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,724B, BPFP=3.7131 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,604B, BPFP=2.7492 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,796B, BPFP=3.5384 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,364B, BPFP=0.8704 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12335577 25.45546051 + layer.0.v_cache 0.00001400 0.00226746 + layer.1.k_cache 0.01519604 1.96497391 + layer.1.v_cache 0.00000547 0.00100698 + layer.2.k_cache 0.01098718 0.70546732 + layer.2.v_cache 0.00001779 0.00306184 + layer.3.k_cache 0.07416311 3.44185740 + layer.3.v_cache 0.00001944 0.00320942 + layer.4.k_cache 0.00061719 0.08058109 + layer.4.v_cache 0.00006444 0.00675763 + layer.4.output 0.17504946 593.93965146 + ------------------------------------------------------------------------------------- + TOTAL 0.08528157 246.42601199 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 175212 +BPFP 1.9402 bits/point +EBPFP 3.8805 equivalent bits/point +MSE 246.426012 +---------------------- -------------------------------------------------------- +Time: 2.511s Load: 0.003s, Pack+Encode: 1.498s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 246.4260 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,584B, BPFP=0.9302 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,140B, BPFP=3.8839 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,752B, BPFP=1.5731 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,772B, BPFP=3.8093 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,892B, BPFP=1.8044 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,072B, BPFP=3.6672 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,236B, BPFP=1.6713 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,380B, BPFP=3.7297 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,396B, BPFP=2.7183 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,844B, BPFP=3.6209 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,804B, BPFP=0.8930 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13955514 24.57796330 + layer.0.v_cache 0.00001392 0.00228368 + layer.1.k_cache 0.03454666 1.69626568 + layer.1.v_cache 0.00000526 0.00098352 + layer.2.k_cache 0.00236268 0.74919327 + layer.2.v_cache 0.00001660 0.00289075 + layer.3.k_cache 0.02874016 3.06687055 + layer.3.v_cache 0.00001899 0.00329423 + layer.4.k_cache 0.00062191 0.07874393 + layer.4.v_cache 0.00005265 0.00694576 + layer.4.output 0.18728454 659.23805659 + ------------------------------------------------------------------------------------- + TOTAL 0.08923092 273.22657828 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 165872 +BPFP 1.9799 bits/point +EBPFP 3.9599 equivalent bits/point +MSE 273.226578 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.004s, Pack+Encode: 1.490s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 273.2266 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,600B, BPFP=0.9334 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,948B, BPFP=3.8450 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,808B, BPFP=1.5844 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,952B, BPFP=3.8458 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,048B, BPFP=1.8360 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,128B, BPFP=3.6786 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,348B, BPFP=1.6940 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,864B, BPFP=3.8279 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,392B, BPFP=2.7175 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,504B, BPFP=3.5519 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,740B, BPFP=0.9201 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11454884 25.68530908 + layer.0.v_cache 0.00001351 0.00238155 + layer.1.k_cache 0.01665594 2.21202622 + layer.1.v_cache 0.00000570 0.00103301 + layer.2.k_cache 0.00836552 0.75473508 + layer.2.v_cache 0.00001974 0.00313944 + layer.3.k_cache 0.03124422 3.00524268 + layer.3.v_cache 0.00001876 0.00350330 + layer.4.k_cache 0.00062231 0.07882212 + layer.4.v_cache 0.00005487 0.00723993 + layer.4.output 0.19663499 663.49379638 + ------------------------------------------------------------------------------------- + TOTAL 0.09105849 275.07117689 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 167332 +BPFP 1.9974 bits/point +EBPFP 3.9947 equivalent bits/point +MSE 275.071177 +---------------------- -------------------------------------------------------- +Time: 2.504s Load: 0.004s, Pack+Encode: 1.492s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 275.0712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,836B, BPFP=0.8996 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,836B, BPFP=3.6897 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,164B, BPFP=1.5186 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,004B, BPFP=3.5350 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,468B, BPFP=1.7612 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,864B, BPFP=3.5089 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,740B, BPFP=1.6257 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,768B, BPFP=3.4911 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,380B, BPFP=2.6749 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,632B, BPFP=3.4658 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,348B, BPFP=0.8862 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12290487 24.79676165 + layer.0.v_cache 0.00001422 0.00232412 + layer.1.k_cache 0.01283375 1.68120593 + layer.1.v_cache 0.00000579 0.00106636 + layer.2.k_cache 0.01100613 0.70213754 + layer.2.v_cache 0.00001929 0.00309416 + layer.3.k_cache 0.03011658 3.07971736 + layer.3.v_cache 0.00001858 0.00331022 + layer.4.k_cache 0.00063020 0.08013490 + layer.4.v_cache 0.00005223 0.00680824 + layer.4.output 0.17536174 576.43744685 + ------------------------------------------------------------------------------------- + TOTAL 0.08265493 239.14227579 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 174040 +BPFP 1.9043 bits/point +EBPFP 3.8086 equivalent bits/point +MSE 239.142276 +---------------------- -------------------------------------------------------- +Time: 2.515s Load: 0.005s, Pack+Encode: 1.498s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 239.1423 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,808B, BPFP=0.8838 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,592B, BPFP=3.7853 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,144B, BPFP=1.4971 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,808B, BPFP=3.6412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,604B, BPFP=1.7654 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,924B, BPFP=3.4787 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,832B, BPFP=1.6235 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,188B, BPFP=3.5272 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,272B, BPFP=2.6235 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,120B, BPFP=3.5147 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,812B, BPFP=0.8617 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11218034 28.51455078 + layer.0.v_cache 0.00001438 0.00224874 + layer.1.k_cache 0.03157643 2.03378906 + layer.1.v_cache 0.00000545 0.00095003 + layer.2.k_cache 0.00802605 0.74563195 + layer.2.v_cache 0.00001826 0.00293601 + layer.3.k_cache 0.04519741 3.42277760 + layer.3.v_cache 0.00001794 0.00302050 + layer.4.k_cache 0.00061137 0.07380584 + layer.4.v_cache 0.00005913 0.00678368 + layer.4.output 0.17010492 566.89049370 + ------------------------------------------------------------------------------------- + TOTAL 0.08167301 235.47293824 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 176104 +BPFP 1.9042 bits/point +EBPFP 3.8085 equivalent bits/point +MSE 235.472938 +---------------------- -------------------------------------------------------- +Time: 2.511s Load: 0.004s, Pack+Encode: 1.491s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 235.4729 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,788B, BPFP=0.8699 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,500B, BPFP=3.7246 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,228B, BPFP=1.4949 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,828B, BPFP=3.6025 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,604B, BPFP=1.7449 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,228B, BPFP=3.4935 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,892B, BPFP=1.6156 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,796B, BPFP=3.5967 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,504B, BPFP=2.8169 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,320B, BPFP=3.5102 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,444B, BPFP=0.8680 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10146660 27.61807322 + layer.0.v_cache 0.00001344 0.00221560 + layer.1.k_cache 0.03090366 1.93070789 + layer.1.v_cache 0.00000566 0.00096722 + layer.2.k_cache 0.01188999 0.70875412 + layer.2.v_cache 0.00001845 0.00295999 + layer.3.k_cache 0.02542871 3.03853093 + layer.3.v_cache 0.00001884 0.00326818 + layer.4.k_cache 0.00060243 0.08279457 + layer.4.v_cache 0.00005224 0.00666007 + layer.4.output 0.16839122 580.38818522 + ------------------------------------------------------------------------------------- + TOTAL 0.07936109 240.94777813 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 179132 +BPFP 1.9145 bits/point +EBPFP 3.8289 equivalent bits/point +MSE 240.947778 +---------------------- -------------------------------------------------------- +Time: 2.509s Load: 0.005s, Pack+Encode: 1.492s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 240.9478 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,032B, BPFP=0.8546 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,584B, BPFP=3.6658 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,576B, BPFP=1.4565 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,724B, BPFP=3.5197 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,148B, BPFP=1.7235 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,400B, BPFP=3.4647 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,352B, BPFP=1.5883 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,864B, BPFP=3.5435 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,980B, BPFP=2.7140 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,288B, BPFP=3.4457 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,628B, BPFP=0.8402 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12981709 27.59945546 + layer.0.v_cache 0.00001628 0.00222706 + layer.1.k_cache 0.06761242 1.80027373 + layer.1.v_cache 0.00000547 0.00091252 + layer.2.k_cache 0.00384689 0.65364937 + layer.2.v_cache 0.00001812 0.00291324 + layer.3.k_cache 0.05327560 3.11880626 + layer.3.v_cache 0.00001896 0.00313713 + layer.4.k_cache 0.00060800 0.07562247 + layer.4.v_cache 0.00005116 0.00639068 + layer.4.output 0.15518735 528.18133734 + ------------------------------------------------------------------------------------- + TOTAL 0.07891655 219.44310290 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 187576 +BPFP 1.8740 bits/point +EBPFP 3.7479 equivalent bits/point +MSE 219.443103 +---------------------- -------------------------------------------------------- +Time: 2.505s Load: 0.004s, Pack+Encode: 1.489s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 219.4431 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,756B, BPFP=0.9174 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,808B, BPFP=3.8210 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,888B, BPFP=1.5216 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,236B, BPFP=3.7106 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,244B, BPFP=1.7832 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,524B, BPFP=3.5733 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,584B, BPFP=1.6559 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,604B, BPFP=3.5887 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,056B, BPFP=2.7114 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,088B, BPFP=3.4892 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,828B, BPFP=0.9047 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887507 26.85829897 + layer.0.v_cache 0.00001345 0.00226438 + layer.1.k_cache 0.03259810 2.03471167 + layer.1.v_cache 0.00000523 0.00091885 + layer.2.k_cache 0.00244350 0.79038286 + layer.2.v_cache 0.00001834 0.00293735 + layer.3.k_cache 0.10864470 3.26720136 + layer.3.v_cache 0.00001904 0.00331617 + layer.4.k_cache 0.00060864 0.07813969 + layer.4.v_cache 0.00004940 0.00655220 + layer.4.output 0.18628448 597.32291667 + ------------------------------------------------------------------------------------- + TOTAL 0.09395687 247.90030236 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 171616 +BPFP 1.9473 bits/point +EBPFP 3.8947 equivalent bits/point +MSE 247.900302 +---------------------- -------------------------------------------------------- +Time: 2.510s Load: 0.005s, Pack+Encode: 1.495s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 247.9003 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,900B, BPFP=0.8603 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,600B, BPFP=3.6166 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,452B, BPFP=1.4838 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,944B, BPFP=3.5014 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,724B, BPFP=1.7072 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,476B, BPFP=3.4192 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,988B, BPFP=1.5779 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,080B, BPFP=3.5253 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,712B, BPFP=2.5829 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,468B, BPFP=3.4178 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,692B, BPFP=0.8701 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422350 27.09402432 + layer.0.v_cache 0.00001493 0.00223588 + layer.1.k_cache 0.01422887 2.10815567 + layer.1.v_cache 0.00000595 0.00099088 + layer.2.k_cache 0.00668612 0.78705228 + layer.2.v_cache 0.00001902 0.00288624 + layer.3.k_cache 0.02497406 3.35735158 + layer.3.v_cache 0.00001869 0.00311348 + layer.4.k_cache 0.00061935 0.07704980 + layer.4.v_cache 0.00005212 0.00668139 + layer.4.output 0.16899524 513.15609952 + ------------------------------------------------------------------------------------- + TOTAL 0.07904761 213.26660224 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 181036 +BPFP 1.8696 bits/point +EBPFP 3.7392 equivalent bits/point +MSE 213.266602 +---------------------- -------------------------------------------------------- +Time: 2.509s Load: 0.005s, Pack+Encode: 1.493s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 213.2666 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 113, 128) +Output shape: (1, 113, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.output: torch.Size([1, 113, 3584]) -> torch.Size([1, 1, 113, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,796B, BPFP=0.8014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,644B, BPFP=3.2694 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,752B, BPFP=1.4867 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,096B, BPFP=3.1936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,164B, BPFP=1.8202 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,420B, BPFP=3.1001 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,152B, BPFP=1.6803 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,840B, BPFP=3.1582 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,716B, BPFP=2.5879 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,128B, BPFP=3.0597 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,296B, BPFP=0.7960 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10214825 28.28947300 + layer.0.v_cache 0.00001637 0.00227271 + layer.1.k_cache 0.04284562 2.12426488 + layer.1.v_cache 0.00000578 0.00094732 + layer.2.k_cache 0.00913706 0.53360856 + layer.2.v_cache 0.00001928 0.00292819 + layer.3.k_cache 0.05585717 3.94351493 + layer.3.v_cache 0.00001870 0.00305478 + layer.4.k_cache 0.00065968 0.08957039 + layer.4.v_cache 0.00005753 0.00640725 + layer.4.output 10.09706179 424.50643963 + ------------------------------------------------------------------------------------- + TOTAL 4.17001165 176.85535997 + (elements=983,552) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 983552 +Total Bytes 215004 +BPFP 1.7488 bits/point +EBPFP 3.4976 equivalent bits/point +MSE 176.855360 +---------------------- -------------------------------------------------------- +Time: 2.517s Load: 0.005s, Pack+Encode: 1.495s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 176.8554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,004B, BPFP=0.8592 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,180B, BPFP=3.6367 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,392B, BPFP=1.4409 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,464B, BPFP=3.5137 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,072B, BPFP=1.7294 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,220B, BPFP=3.4718 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,228B, BPFP=1.5845 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,556B, BPFP=3.5295 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,640B, BPFP=2.6854 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,036B, BPFP=3.4402 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,480B, BPFP=0.8948 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11481400 27.39566342 + layer.0.v_cache 0.00001396 0.00227368 + layer.1.k_cache 0.03135788 1.78396439 + layer.1.v_cache 0.00000545 0.00098169 + layer.2.k_cache 0.00370715 0.72395174 + layer.2.v_cache 0.00001965 0.00300404 + layer.3.k_cache 0.03930106 3.22580325 + layer.3.v_cache 0.00001886 0.00316911 + layer.4.k_cache 0.00060819 0.07812044 + layer.4.v_cache 0.00005437 0.00677076 + layer.4.output 0.16591480 533.90649529 + ------------------------------------------------------------------------------------- + TOTAL 0.07948848 221.79818644 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 187272 +BPFP 1.8915 bits/point +EBPFP 3.7830 equivalent bits/point +MSE 221.798186 +---------------------- -------------------------------------------------------- +Time: 2.508s Load: 0.005s, Pack+Encode: 1.492s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 221.7982 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,880B, BPFP=0.8567 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,928B, BPFP=3.6742 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,352B, BPFP=1.4663 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,260B, BPFP=3.5569 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,888B, BPFP=1.7360 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,672B, BPFP=3.4537 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,940B, BPFP=1.5695 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,680B, BPFP=3.4551 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,232B, BPFP=2.6742 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,376B, BPFP=3.4017 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,668B, BPFP=0.7942 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09788724 25.61230743 + layer.0.v_cache 0.00001308 0.00224501 + layer.1.k_cache 0.03299916 2.06940966 + layer.1.v_cache 0.00000544 0.00097562 + layer.2.k_cache 0.00367894 0.66128986 + layer.2.v_cache 0.00002012 0.00310117 + layer.3.k_cache 0.03939044 3.08466879 + layer.3.v_cache 0.00001875 0.00325492 + layer.4.k_cache 0.00061788 0.08044311 + layer.4.v_cache 0.00005010 0.00669626 + layer.4.output 0.17021053 543.27829053 + ------------------------------------------------------------------------------------- + TOTAL 0.08036205 225.55720150 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 178876 +BPFP 1.8473 bits/point +EBPFP 3.6946 equivalent bits/point +MSE 225.557202 +---------------------- -------------------------------------------------------- +Time: 2.510s Load: 0.005s, Pack+Encode: 1.496s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 225.5572 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,676B, BPFP=0.8212 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,372B, BPFP=3.3814 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,768B, BPFP=1.5579 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,720B, BPFP=3.2870 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,000B, BPFP=1.8808 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,112B, BPFP=3.1991 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,740B, BPFP=1.6985 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,636B, BPFP=3.2749 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,308B, BPFP=2.6487 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,832B, BPFP=3.1586 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 45,700B, BPFP=0.9445 +⌛️ [2/4] FRONTEND: Frontend time: 1.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.015s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12672405 25.98620153 + layer.0.v_cache 0.00001401 0.00225996 + layer.1.k_cache 0.12044146 2.36135751 + layer.1.v_cache 0.00000568 0.00096361 + layer.2.k_cache 0.00570256 0.61709397 + layer.2.v_cache 0.00001866 0.00296059 + layer.3.k_cache 0.03611955 4.21760559 + layer.3.v_cache 0.00001983 0.00327236 + layer.4.k_cache 0.00066233 0.08185764 + layer.4.v_cache 0.00005331 0.00645772 + layer.4.output 10.56037249 366.10069444 + ------------------------------------------------------------------------------------- + TOTAL 4.36543346 152.70499362 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 217864 +BPFP 1.8541 bits/point +EBPFP 3.7082 equivalent bits/point +MSE 152.704994 +---------------------- -------------------------------------------------------- +Time: 2.509s Load: 0.006s, Pack+Encode: 1.488s, Decode+Unpack: 1.015s +---------------------- -------------------------------------------------------- +💾 Converting with 152.7050 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,868B, BPFP=0.8643 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,836B, BPFP=3.6996 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,392B, BPFP=1.4901 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,076B, BPFP=3.5646 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,952B, BPFP=1.7670 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,392B, BPFP=3.4432 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,072B, BPFP=1.6108 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,036B, BPFP=3.5575 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,980B, BPFP=2.6598 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,584B, BPFP=3.4773 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,844B, BPFP=0.8331 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13733016 26.57006281 + layer.0.v_cache 0.00001373 0.00250479 + layer.1.k_cache 0.03155638 2.03573036 + layer.1.v_cache 0.00000617 0.00116997 + layer.2.k_cache 0.00751131 0.74995821 + layer.2.v_cache 0.00001754 0.00330281 + layer.3.k_cache 0.02548880 2.80784798 + layer.3.v_cache 0.00001765 0.00365101 + layer.4.k_cache 0.00061011 0.08040099 + layer.4.v_cache 0.00005045 0.00729571 + layer.4.output 0.16239834 544.30179586 + ------------------------------------------------------------------------------------- + TOTAL 0.07878769 226.02202916 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 180032 +BPFP 1.8803 bits/point +EBPFP 3.7607 equivalent bits/point +MSE 226.022029 +---------------------- -------------------------------------------------------- +Time: 2.512s Load: 0.004s, Pack+Encode: 1.492s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 226.0220 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,692B, BPFP=0.9280 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,604B, BPFP=3.8774 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,856B, BPFP=1.5538 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,808B, BPFP=3.7199 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,996B, BPFP=1.7793 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,472B, BPFP=3.6535 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,376B, BPFP=1.6566 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,484B, BPFP=3.6559 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,400B, BPFP=2.6503 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,032B, BPFP=3.5665 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,088B, BPFP=0.8784 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.019s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10979201 28.57829497 + layer.0.v_cache 0.00001339 0.00228770 + layer.1.k_cache 0.01160004 1.94705992 + layer.1.v_cache 0.00000567 0.00099321 + layer.2.k_cache 0.00714037 0.73564433 + layer.2.v_cache 0.00001829 0.00297813 + layer.3.k_cache 0.04622446 2.89171166 + layer.3.v_cache 0.00001896 0.00333508 + layer.4.k_cache 0.00064378 0.07406034 + layer.4.v_cache 0.00005251 0.00688022 + layer.4.output 0.18415536 590.08290009 + ------------------------------------------------------------------------------------- + TOTAL 0.08615276 244.98962036 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 167808 +BPFP 1.9523 bits/point +EBPFP 3.9047 equivalent bits/point +MSE 244.989620 +---------------------- -------------------------------------------------------- +Time: 2.512s Load: 0.003s, Pack+Encode: 1.490s, Decode+Unpack: 1.019s +---------------------- -------------------------------------------------------- +💾 Converting with 244.9896 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,588B, BPFP=0.9310 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,156B, BPFP=3.8872 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,688B, BPFP=1.5601 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,704B, BPFP=3.7955 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,880B, BPFP=1.8019 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,804B, BPFP=3.6128 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,204B, BPFP=1.6648 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,476B, BPFP=3.7492 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,592B, BPFP=2.7581 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,280B, BPFP=3.5065 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,696B, BPFP=0.9478 +⌛️ [2/4] FRONTEND: Frontend time: 1.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11290005 27.18695465 + layer.0.v_cache 0.00001570 0.00224866 + layer.1.k_cache 0.03434501 2.00545313 + layer.1.v_cache 0.00000534 0.00089803 + layer.2.k_cache 0.00238675 0.68186901 + layer.2.v_cache 0.00001838 0.00287867 + layer.3.k_cache 0.02838124 2.89269663 + layer.3.v_cache 0.00002047 0.00320304 + layer.4.k_cache 0.00062757 0.07867531 + layer.4.v_cache 0.00005455 0.00645222 + layer.4.output 0.18420308 636.02974258 + ------------------------------------------------------------------------------------- + TOTAL 0.08636333 263.82761926 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 167068 +BPFP 1.9942 bits/point +EBPFP 3.9884 equivalent bits/point +MSE 263.827619 +---------------------- -------------------------------------------------------- +Time: 2.506s Load: 0.004s, Pack+Encode: 1.488s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 263.8276 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,664B, BPFP=0.9343 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,436B, BPFP=3.8934 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,840B, BPFP=1.5705 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,752B, BPFP=3.7564 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,924B, BPFP=1.7877 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,220B, BPFP=3.6498 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,360B, BPFP=1.6747 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,828B, BPFP=3.7716 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,616B, BPFP=2.7276 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,036B, BPFP=3.6130 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,756B, BPFP=0.9374 +⌛️ [2/4] FRONTEND: Frontend time: 1.488s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09752406 26.27185998 + layer.0.v_cache 0.00001393 0.00243259 + layer.1.k_cache 0.01493649 1.92301804 + layer.1.v_cache 0.00000588 0.00100615 + layer.2.k_cache 0.01003601 0.66349768 + layer.2.v_cache 0.00001805 0.00295675 + layer.3.k_cache 0.04451986 2.84939966 + layer.3.v_cache 0.00001914 0.00350033 + layer.4.k_cache 0.00061869 0.07784111 + layer.4.v_cache 0.00005251 0.00689215 + layer.4.output 0.18409089 639.71823489 + ------------------------------------------------------------------------------------- + TOTAL 0.08566946 265.28412051 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 169432 +BPFP 1.9965 bits/point +EBPFP 3.9930 equivalent bits/point +MSE 265.284121 +---------------------- -------------------------------------------------------- +Time: 2.506s Load: 0.004s, Pack+Encode: 1.488s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 265.2841 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,776B, BPFP=0.8991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,868B, BPFP=3.7402 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,964B, BPFP=1.4992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,492B, BPFP=3.6694 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,204B, BPFP=1.7327 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,760B, BPFP=3.5316 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,616B, BPFP=1.6220 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,924B, BPFP=3.5625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,392B, BPFP=2.7093 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,420B, BPFP=3.4676 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,608B, BPFP=0.8769 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13658243 25.56618858 + layer.0.v_cache 0.00001379 0.00238020 + layer.1.k_cache 0.01234693 1.55949071 + layer.1.v_cache 0.00000566 0.00104456 + layer.2.k_cache 0.00375735 0.68427277 + layer.2.v_cache 0.00001852 0.00286195 + layer.3.k_cache 0.02700986 2.96008687 + layer.3.v_cache 0.00001930 0.00321829 + layer.4.k_cache 0.00062007 0.07960574 + layer.4.v_cache 0.00005146 0.00700034 + layer.4.output 0.17614663 558.45605637 + ------------------------------------------------------------------------------------- + TOTAL 0.08314423 231.76814968 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 173024 +BPFP 1.9160 bits/point +EBPFP 3.8320 equivalent bits/point +MSE 231.768150 +---------------------- -------------------------------------------------------- +Time: 2.518s Load: 0.004s, Pack+Encode: 1.495s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 231.7681 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,348B, BPFP=0.9307 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,088B, BPFP=3.8716 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,396B, BPFP=1.5830 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,956B, BPFP=3.6293 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,484B, BPFP=1.8159 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,448B, BPFP=3.7346 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,828B, BPFP=1.6755 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,048B, BPFP=3.6490 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,112B, BPFP=2.8065 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,424B, BPFP=3.5154 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,012B, BPFP=0.9788 +⌛️ [2/4] FRONTEND: Frontend time: 1.508s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08738674 24.21110639 + layer.0.v_cache 0.00001368 0.00239725 + layer.1.k_cache 0.01286346 2.09918297 + layer.1.v_cache 0.00000597 0.00096877 + layer.2.k_cache 0.00251921 0.69580820 + layer.2.v_cache 0.00001769 0.00296874 + layer.3.k_cache 0.08324167 2.79774789 + layer.3.v_cache 0.00002021 0.00327312 + layer.4.k_cache 0.00062021 0.07803581 + layer.4.v_cache 0.00005375 0.00676212 + layer.4.output 0.19567458 625.50947896 + ------------------------------------------------------------------------------------- + TOTAL 0.09155674 259.32144729 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 159144 +BPFP 2.0037 bits/point +EBPFP 4.0075 equivalent bits/point +MSE 259.321447 +---------------------- -------------------------------------------------------- +Time: 2.528s Load: 0.004s, Pack+Encode: 1.508s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 259.3214 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,804B, BPFP=0.9044 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,168B, BPFP=3.7967 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,052B, BPFP=1.5158 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,644B, BPFP=3.6980 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,352B, BPFP=1.7605 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,952B, BPFP=3.5678 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,644B, BPFP=1.6273 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,140B, BPFP=3.6032 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,852B, BPFP=2.7959 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,628B, BPFP=3.5068 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,368B, BPFP=0.8974 +⌛️ [2/4] FRONTEND: Frontend time: 1.499s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12176956 25.46020214 + layer.0.v_cache 0.00001396 0.00230679 + layer.1.k_cache 0.01120206 1.94516929 + layer.1.v_cache 0.00000562 0.00098850 + layer.2.k_cache 0.00678912 0.55570340 + layer.2.v_cache 0.00001849 0.00293926 + layer.3.k_cache 0.02622183 3.43383421 + layer.3.v_cache 0.00001965 0.00348631 + layer.4.k_cache 0.00062413 0.08104265 + layer.4.v_cache 0.00005248 0.00700917 + layer.4.output 0.17940697 550.03227194 + ------------------------------------------------------------------------------------- + TOTAL 0.08368034 228.33638737 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 175604 +BPFP 1.9446 bits/point +EBPFP 3.8892 equivalent bits/point +MSE 228.336387 +---------------------- -------------------------------------------------------- +Time: 2.514s Load: 0.005s, Pack+Encode: 1.499s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 228.3364 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,548B, BPFP=0.8256 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,460B, BPFP=3.4911 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,236B, BPFP=1.5232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,832B, BPFP=3.3976 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,768B, BPFP=1.7512 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,180B, BPFP=3.3006 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,940B, BPFP=1.6280 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,752B, BPFP=3.3857 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,444B, BPFP=2.7446 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,936B, BPFP=3.2643 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,448B, BPFP=0.9236 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120816 26.58510975 + layer.0.v_cache 0.00001607 0.00214645 + layer.1.k_cache 0.10678518 2.54689593 + layer.1.v_cache 0.00000585 0.00089899 + layer.2.k_cache 0.00813519 0.60907346 + layer.2.v_cache 0.00001919 0.00275052 + layer.3.k_cache 0.06560528 4.08122442 + layer.3.v_cache 0.00001973 0.00303951 + layer.4.k_cache 0.00062300 0.08235432 + layer.4.v_cache 0.00005287 0.00637991 + layer.4.output 10.86395687 433.02942177 + ------------------------------------------------------------------------------------- + TOTAL 4.49059815 180.30151915 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 213544 +BPFP 1.8693 bits/point +EBPFP 3.7385 equivalent bits/point +MSE 180.301519 +---------------------- -------------------------------------------------------- +Time: 2.515s Load: 0.005s, Pack+Encode: 1.497s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 180.3015 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,840B, BPFP=0.8693 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,608B, BPFP=3.7011 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,320B, BPFP=1.4943 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,160B, BPFP=3.6207 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,716B, BPFP=1.7450 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,344B, BPFP=3.4741 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,924B, BPFP=1.6027 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,816B, BPFP=3.5589 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,600B, BPFP=2.8017 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,548B, BPFP=3.5108 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,500B, BPFP=0.8852 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12848519 24.45508374 + layer.0.v_cache 0.00001371 0.00231037 + layer.1.k_cache 0.03007136 1.71423094 + layer.1.v_cache 0.00000570 0.00102632 + layer.2.k_cache 0.00936044 0.73235694 + layer.2.v_cache 0.00001842 0.00308364 + layer.3.k_cache 0.04065019 2.99529643 + layer.3.v_cache 0.00001877 0.00335143 + layer.4.k_cache 0.00062204 0.08291614 + layer.4.v_cache 0.00010837 0.00700230 + layer.4.output 0.16426703 541.31573276 + ------------------------------------------------------------------------------------- + TOTAL 0.07995432 224.65922280 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 181376 +BPFP 1.9162 bits/point +EBPFP 3.8323 equivalent bits/point +MSE 224.659223 +---------------------- -------------------------------------------------------- +Time: 2.512s Load: 0.005s, Pack+Encode: 1.496s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 224.6592 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,680B, BPFP=0.9141 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,276B, BPFP=3.7648 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,820B, BPFP=1.5273 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,940B, BPFP=3.6992 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,132B, BPFP=1.7836 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,236B, BPFP=3.5617 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,420B, BPFP=1.6445 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,504B, BPFP=3.6141 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,708B, BPFP=2.6773 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,776B, BPFP=3.4719 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,196B, BPFP=0.8983 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09257430 26.64373169 + layer.0.v_cache 0.00001329 0.00224262 + layer.1.k_cache 0.01202901 2.14294930 + layer.1.v_cache 0.00000544 0.00090647 + layer.2.k_cache 0.00549439 0.72185402 + layer.2.v_cache 0.00001849 0.00295963 + layer.3.k_cache 0.11014032 3.38876190 + layer.3.v_cache 0.00001916 0.00305831 + layer.4.k_cache 0.00063444 0.07415214 + layer.4.v_cache 0.00005494 0.00628895 + layer.4.output 0.19143520 578.50396205 + ------------------------------------------------------------------------------------- + TOTAL 0.09182531 240.14791996 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 168688 +BPFP 1.9381 bits/point +EBPFP 3.8761 equivalent bits/point +MSE 240.147920 +---------------------- -------------------------------------------------------- +Time: 2.515s Load: 0.004s, Pack+Encode: 1.498s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 240.1479 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,836B, BPFP=0.8996 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,560B, BPFP=3.8244 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,240B, BPFP=1.5327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,904B, BPFP=3.7024 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,448B, BPFP=1.7574 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,628B, BPFP=3.4650 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,792B, BPFP=1.6354 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,388B, BPFP=3.6064 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,560B, BPFP=2.7083 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,004B, BPFP=3.5350 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,376B, BPFP=0.8603 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12365316 24.57426816 + layer.0.v_cache 0.00001332 0.00239148 + layer.1.k_cache 0.03511774 1.89625004 + layer.1.v_cache 0.00000569 0.00110504 + layer.2.k_cache 0.01236608 0.69093854 + layer.2.v_cache 0.00001835 0.00290605 + layer.3.k_cache 0.02851222 3.00939287 + layer.3.v_cache 0.00001844 0.00344689 + layer.4.k_cache 0.00063207 0.07700249 + layer.4.v_cache 0.00005066 0.00728618 + layer.4.output 0.17964420 438.20110544 + ------------------------------------------------------------------------------------- + TOTAL 0.08575865 182.21604270 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 175736 +BPFP 1.9229 bits/point +EBPFP 3.8458 equivalent bits/point +MSE 182.216043 +---------------------- -------------------------------------------------------- +Time: 2.524s Load: 0.005s, Pack+Encode: 1.503s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 182.2160 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,616B, BPFP=0.9247 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,492B, BPFP=3.9046 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,792B, BPFP=1.5609 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,700B, BPFP=3.7460 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,928B, BPFP=1.7885 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,484B, BPFP=3.7027 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,356B, BPFP=1.6739 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,596B, BPFP=3.7252 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,752B, BPFP=2.7548 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,284B, BPFP=3.6627 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,760B, BPFP=0.9375 +⌛️ [2/4] FRONTEND: Frontend time: 1.501s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.010s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12096892 27.85128706 + layer.0.v_cache 0.00001393 0.00229456 + layer.1.k_cache 0.03247837 1.88234965 + layer.1.v_cache 0.00000613 0.00097599 + layer.2.k_cache 0.00394396 0.78756479 + layer.2.v_cache 0.00001920 0.00295590 + layer.3.k_cache 0.02751121 3.01161429 + layer.3.v_cache 0.00001919 0.00321795 + layer.4.k_cache 0.00062481 0.07679044 + layer.4.v_cache 0.00005293 0.00664457 + layer.4.output 0.18564256 626.62883471 + ------------------------------------------------------------------------------------- + TOTAL 0.08736098 260.00161989 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 169760 +BPFP 2.0004 bits/point +EBPFP 4.0008 equivalent bits/point +MSE 260.001620 +---------------------- -------------------------------------------------------- +Time: 2.515s Load: 0.005s, Pack+Encode: 1.501s, Decode+Unpack: 1.010s +---------------------- -------------------------------------------------------- +💾 Converting with 260.0016 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,612B, BPFP=0.9359 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,716B, BPFP=3.7979 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,736B, BPFP=1.5698 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,180B, BPFP=3.6891 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,972B, BPFP=1.8206 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,960B, BPFP=3.6445 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,312B, BPFP=1.6867 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,848B, BPFP=3.8247 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,736B, BPFP=2.7873 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,536B, BPFP=3.5584 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,088B, BPFP=0.9592 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08564486 25.72746138 + layer.0.v_cache 0.00001408 0.00230517 + layer.1.k_cache 0.03507188 2.04110757 + layer.1.v_cache 0.00000566 0.00094736 + layer.2.k_cache 0.00239554 0.68920344 + layer.2.v_cache 0.00001938 0.00302874 + layer.3.k_cache 0.03313774 3.39770587 + layer.3.v_cache 0.00001993 0.00341160 + layer.4.k_cache 0.00061714 0.07926680 + layer.4.v_cache 0.00005134 0.00680304 + layer.4.output 0.18226704 639.02261132 + ------------------------------------------------------------------------------------- + TOTAL 0.08428511 265.00644236 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 167696 +BPFP 2.0017 bits/point +EBPFP 4.0034 equivalent bits/point +MSE 265.006442 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.005s, Pack+Encode: 1.503s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 265.0064 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 117, 128) +Output shape: (1, 117, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.output: torch.Size([1, 117, 3584]) -> torch.Size([1, 1, 117, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,920B, BPFP=0.7906 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,884B, BPFP=3.1896 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,500B, BPFP=1.5358 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,360B, BPFP=3.1197 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,760B, BPFP=1.8376 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,756B, BPFP=3.0390 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,588B, BPFP=1.6811 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,152B, BPFP=3.0919 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,776B, BPFP=2.5075 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,120B, BPFP=2.9541 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,872B, BPFP=0.8370 +⌛️ [2/4] FRONTEND: Frontend time: 1.501s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11900647 27.93066824 + layer.0.v_cache 0.00001428 0.00226319 + layer.1.k_cache 0.12535267 2.70461254 + layer.1.v_cache 0.00000579 0.00096725 + layer.2.k_cache 0.00517510 0.53728133 + layer.2.v_cache 0.00001992 0.00298736 + layer.3.k_cache 0.09315677 4.07672223 + layer.3.v_cache 0.00001978 0.00321233 + layer.4.k_cache 0.00061896 0.08114353 + layer.4.v_cache 0.00005133 0.00643775 + layer.4.output 9.75369281 419.46722375 + ------------------------------------------------------------------------------------- + TOTAL 4.03642769 174.80099188 + (elements=1,018,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1018368 +Total Bytes 221688 +BPFP 1.7415 bits/point +EBPFP 3.4830 equivalent bits/point +MSE 174.800992 +---------------------- -------------------------------------------------------- +Time: 2.520s Load: 0.004s, Pack+Encode: 1.501s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 174.8010 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,528B, BPFP=0.8226 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,508B, BPFP=3.4982 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,204B, BPFP=1.5185 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,920B, BPFP=3.4107 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,140B, BPFP=1.8065 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,152B, BPFP=3.2964 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,192B, BPFP=1.6655 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,656B, BPFP=3.3714 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,316B, BPFP=2.7256 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,976B, BPFP=3.2702 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,624B, BPFP=0.8211 +⌛️ [2/4] FRONTEND: Frontend time: 1.510s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14131550 27.76170480 + layer.0.v_cache 0.00001383 0.00217207 + layer.1.k_cache 0.02978072 2.40379493 + layer.1.v_cache 0.00000537 0.00091296 + layer.2.k_cache 0.00578868 0.65396133 + layer.2.v_cache 0.00001872 0.00272265 + layer.3.k_cache 0.04020810 3.85638341 + layer.3.v_cache 0.00001850 0.00296944 + layer.4.k_cache 0.00063358 0.08188469 + layer.4.v_cache 0.00005294 0.00664471 + layer.4.output 10.86925024 459.38333333 + ------------------------------------------------------------------------------------- + TOTAL 4.48838751 191.20332261 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 209216 +BPFP 1.8314 bits/point +EBPFP 3.6627 equivalent bits/point +MSE 191.203323 +---------------------- -------------------------------------------------------- +Time: 2.535s Load: 0.006s, Pack+Encode: 1.510s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 191.2033 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,816B, BPFP=0.8958 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,168B, BPFP=3.7515 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,120B, BPFP=1.5104 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,444B, BPFP=3.6168 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,416B, BPFP=1.7515 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,612B, BPFP=3.4621 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,808B, BPFP=1.6384 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,572B, BPFP=3.6406 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,428B, BPFP=2.6838 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,396B, BPFP=3.4219 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,192B, BPFP=0.8554 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09514355 25.09536307 + layer.0.v_cache 0.00001423 0.00235592 + layer.1.k_cache 0.03283414 1.97157469 + layer.1.v_cache 0.00000553 0.00092531 + layer.2.k_cache 0.00489674 0.70783324 + layer.2.v_cache 0.00001883 0.00279361 + layer.3.k_cache 0.07199075 2.85739063 + layer.3.v_cache 0.00001932 0.00306183 + layer.4.k_cache 0.00061592 0.07678066 + layer.4.v_cache 0.00006816 0.00648757 + layer.4.output 0.17253765 582.66443452 + ------------------------------------------------------------------------------------- + TOTAL 0.08313946 241.72797695 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 173972 +BPFP 1.9036 bits/point +EBPFP 3.8072 equivalent bits/point +MSE 241.727977 +---------------------- -------------------------------------------------------- +Time: 2.507s Load: 0.005s, Pack+Encode: 1.493s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 241.7280 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,880B, BPFP=0.8764 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,756B, BPFP=3.7277 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,408B, BPFP=1.5101 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,116B, BPFP=3.6128 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,668B, BPFP=1.7364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,432B, BPFP=3.4899 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,068B, BPFP=1.6286 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,032B, BPFP=3.5977 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,008B, BPFP=2.6954 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,460B, BPFP=3.4950 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,932B, BPFP=0.8449 +⌛️ [2/4] FRONTEND: Frontend time: 1.487s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11660385 27.51335197 + layer.0.v_cache 0.00001381 0.00238183 + layer.1.k_cache 0.05175667 1.80189216 + layer.1.v_cache 0.00000569 0.00106540 + layer.2.k_cache 0.00227456 0.68735877 + layer.2.v_cache 0.00001743 0.00295922 + layer.3.k_cache 0.02781237 2.96357990 + layer.3.v_cache 0.00001883 0.00323799 + layer.4.k_cache 0.00061846 0.08012059 + layer.4.v_cache 0.00005344 0.00655171 + layer.4.output 0.17410791 521.16122742 + ------------------------------------------------------------------------------------- + TOTAL 0.08340767 216.54065244 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 179760 +BPFP 1.8991 bits/point +EBPFP 3.7982 equivalent bits/point +MSE 216.540652 +---------------------- -------------------------------------------------------- +Time: 2.499s Load: 0.005s, Pack+Encode: 1.487s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 216.5407 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,788B, BPFP=0.9014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,176B, BPFP=3.7982 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,012B, BPFP=1.5083 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,480B, BPFP=3.6672 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,284B, BPFP=1.7477 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,436B, BPFP=3.4706 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,604B, BPFP=1.6197 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,728B, BPFP=3.5256 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,648B, BPFP=2.7575 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,320B, BPFP=3.4488 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,036B, BPFP=0.8616 +⌛️ [2/4] FRONTEND: Frontend time: 1.487s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212080 24.36178111 + layer.0.v_cache 0.00001360 0.00234465 + layer.1.k_cache 0.03288655 1.76326044 + layer.1.v_cache 0.00000536 0.00099373 + layer.2.k_cache 0.00973156 0.75866598 + layer.2.v_cache 0.00001762 0.00281557 + layer.3.k_cache 0.02906220 2.82262246 + layer.3.v_cache 0.00001891 0.00309950 + layer.4.k_cache 0.00063934 0.08070052 + layer.4.v_cache 0.00004866 0.00660966 + layer.4.output 0.18425958 557.38710198 + ------------------------------------------------------------------------------------- + TOTAL 0.08496245 231.26544750 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 172512 +BPFP 1.9103 bits/point +EBPFP 3.8207 equivalent bits/point +MSE 231.265447 +---------------------- -------------------------------------------------------- +Time: 2.511s Load: 0.005s, Pack+Encode: 1.487s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 231.2654 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,800B, BPFP=0.8824 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,468B, BPFP=3.7625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,196B, BPFP=1.5066 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,596B, BPFP=3.6022 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,568B, BPFP=1.7588 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,916B, BPFP=3.4772 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,020B, BPFP=1.6581 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,488B, BPFP=3.5824 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,888B, BPFP=2.7368 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,700B, BPFP=3.4375 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,324B, BPFP=0.9014 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.022s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10758395 28.06798886 + layer.0.v_cache 0.00001397 0.00234083 + layer.1.k_cache 0.03293324 1.76636586 + layer.1.v_cache 0.00000575 0.00093984 + layer.2.k_cache 0.00823421 0.72182150 + layer.2.v_cache 0.00001745 0.00284295 + layer.3.k_cache 0.04116619 3.64713637 + layer.3.v_cache 0.00001985 0.00319667 + layer.4.k_cache 0.00062779 0.07857877 + layer.4.v_cache 0.00005042 0.00652726 + layer.4.output 0.16954060 579.90950630 + ------------------------------------------------------------------------------------- + TOTAL 0.08102571 240.80378135 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 177964 +BPFP 1.9244 bits/point +EBPFP 3.8487 equivalent bits/point +MSE 240.803781 +---------------------- -------------------------------------------------------- +Time: 2.529s Load: 0.005s, Pack+Encode: 1.502s, Decode+Unpack: 1.022s +---------------------- -------------------------------------------------------- +💾 Converting with 240.8038 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,948B, BPFP=0.8590 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,008B, BPFP=3.6472 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,460B, BPFP=1.4688 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,464B, BPFP=3.5528 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,000B, BPFP=1.7361 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,220B, BPFP=3.5104 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,196B, BPFP=1.5965 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,400B, BPFP=3.5417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,708B, BPFP=2.7271 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,884B, BPFP=3.4521 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,492B, BPFP=0.8555 +⌛️ [2/4] FRONTEND: Frontend time: 1.504s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12307767 27.57610677 + layer.0.v_cache 0.00001410 0.00228522 + layer.1.k_cache 0.03088827 1.90839166 + layer.1.v_cache 0.00000584 0.00097596 + layer.2.k_cache 0.00371391 0.67129428 + layer.2.v_cache 0.00001950 0.00310104 + layer.3.k_cache 0.03079298 3.26584642 + layer.3.v_cache 0.00001950 0.00323726 + layer.4.k_cache 0.00063014 0.08523141 + layer.4.v_cache 0.00005282 0.00682755 + layer.4.output 0.15767360 530.50595238 + ------------------------------------------------------------------------------------- + TOTAL 0.07605470 220.41558613 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 184780 +BPFP 1.8871 bits/point +EBPFP 3.7741 equivalent bits/point +MSE 220.415586 +---------------------- -------------------------------------------------------- +Time: 2.525s Load: 0.004s, Pack+Encode: 1.504s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 220.4156 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,728B, BPFP=0.9009 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,740B, BPFP=3.7614 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,900B, BPFP=1.5053 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,356B, BPFP=3.6883 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,216B, BPFP=1.7561 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,584B, BPFP=3.5412 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,504B, BPFP=1.6204 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,920B, BPFP=3.6052 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,832B, BPFP=2.6357 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,364B, BPFP=3.4992 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,840B, BPFP=0.8667 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10295349 28.15090272 + layer.0.v_cache 0.00001348 0.00219252 + layer.1.k_cache 0.03334363 1.80185011 + layer.1.v_cache 0.00000550 0.00099479 + layer.2.k_cache 0.01126997 0.75501372 + layer.2.v_cache 0.00001933 0.00293862 + layer.3.k_cache 0.09131574 2.75450916 + layer.3.v_cache 0.00001769 0.00314905 + layer.4.k_cache 0.00062275 0.07802378 + layer.4.v_cache 0.00006541 0.00673332 + layer.4.output 0.17512874 601.48132622 + ------------------------------------------------------------------------------------- + TOTAL 0.08620754 249.64268184 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 170984 +BPFP 1.9165 bits/point +EBPFP 3.8330 equivalent bits/point +MSE 249.642682 +---------------------- -------------------------------------------------------- +Time: 2.525s Load: 0.003s, Pack+Encode: 1.502s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 249.6427 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,872B, BPFP=0.8750 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,428B, BPFP=3.6688 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,352B, BPFP=1.5000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,356B, BPFP=3.6559 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,760B, BPFP=1.7529 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,244B, BPFP=3.4562 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,040B, BPFP=1.6236 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,804B, BPFP=3.5568 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,468B, BPFP=2.5984 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,216B, BPFP=3.4511 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,888B, BPFP=0.8438 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12314748 27.50454607 + layer.0.v_cache 0.00001636 0.00229419 + layer.1.k_cache 0.01190477 1.67068008 + layer.1.v_cache 0.00000588 0.00097729 + layer.2.k_cache 0.01614088 0.71848157 + layer.2.v_cache 0.00001834 0.00283508 + layer.3.k_cache 0.05604688 3.58613709 + layer.3.v_cache 0.00001882 0.00321802 + layer.4.k_cache 0.00063774 0.07706524 + layer.4.v_cache 0.00005162 0.00627037 + layer.4.output 0.17348555 561.41666667 + ------------------------------------------------------------------------------------- + TOTAL 0.08366986 233.14642186 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 178428 +BPFP 1.8850 bits/point +EBPFP 3.7700 equivalent bits/point +MSE 233.146422 +---------------------- -------------------------------------------------------- +Time: 2.520s Load: 0.005s, Pack+Encode: 1.500s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 233.1464 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,744B, BPFP=0.9151 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,808B, BPFP=3.8210 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,940B, BPFP=1.5316 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,792B, BPFP=3.6250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,224B, BPFP=1.7793 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,228B, BPFP=3.5162 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,452B, BPFP=1.6304 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,420B, BPFP=3.5532 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,964B, BPFP=2.6937 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,920B, BPFP=3.4568 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,160B, BPFP=0.9138 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11560643 26.27073688 + layer.0.v_cache 0.00001402 0.00234606 + layer.1.k_cache 0.03272506 2.10027925 + layer.1.v_cache 0.00000563 0.00101968 + layer.2.k_cache 0.00645547 0.73433059 + layer.2.v_cache 0.00001830 0.00300177 + layer.3.k_cache 0.02731901 2.80865215 + layer.3.v_cache 0.00001960 0.00318985 + layer.4.k_cache 0.00061302 0.07880057 + layer.4.v_cache 0.00004978 0.00639595 + layer.4.output 0.17251868 602.08476631 + ------------------------------------------------------------------------------------- + TOTAL 0.08179159 249.80012453 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 170652 +BPFP 1.9364 bits/point +EBPFP 3.8728 equivalent bits/point +MSE 249.800125 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.004s, Pack+Encode: 1.500s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 249.8001 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,916B, BPFP=0.8729 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,972B, BPFP=3.7237 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,436B, BPFP=1.4979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,404B, BPFP=3.6229 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,740B, BPFP=1.7294 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,392B, BPFP=3.4432 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,876B, BPFP=1.5760 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,196B, BPFP=3.5859 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,708B, BPFP=2.6115 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,632B, BPFP=3.4858 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,308B, BPFP=0.8702 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11690623 26.58112127 + layer.0.v_cache 0.00001418 0.00233204 + layer.1.k_cache 0.03305150 1.78568892 + layer.1.v_cache 0.00000545 0.00101836 + layer.2.k_cache 0.01095151 0.77945120 + layer.2.v_cache 0.00001854 0.00302033 + layer.3.k_cache 0.04293454 2.98770176 + layer.3.v_cache 0.00002007 0.00324580 + layer.4.k_cache 0.00062489 0.07739583 + layer.4.v_cache 0.00005451 0.00651250 + layer.4.output 0.16672201 544.11926745 + ------------------------------------------------------------------------------------- + TOTAL 0.08068444 225.94484472 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 181580 +BPFP 1.8965 bits/point +EBPFP 3.7930 equivalent bits/point +MSE 225.944845 +---------------------- -------------------------------------------------------- +Time: 2.520s Load: 0.005s, Pack+Encode: 1.502s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 225.9448 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,976B, BPFP=0.8544 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,220B, BPFP=3.6435 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,436B, BPFP=1.4485 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,568B, BPFP=3.5316 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,884B, BPFP=1.6971 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,896B, BPFP=3.4162 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,192B, BPFP=1.5783 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,516B, BPFP=3.5227 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,752B, BPFP=2.7047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,844B, BPFP=3.4073 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,916B, BPFP=0.8074 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11067300 28.15495149 + layer.0.v_cache 0.00001409 0.00234637 + layer.1.k_cache 0.05152405 1.82447480 + layer.1.v_cache 0.00000569 0.00102373 + layer.2.k_cache 0.01362713 0.77439696 + layer.2.v_cache 0.00001889 0.00300383 + layer.3.k_cache 0.08227781 3.26940583 + layer.3.v_cache 0.00001889 0.00328551 + layer.4.k_cache 0.00062414 0.07776840 + layer.4.v_cache 0.00005257 0.00657534 + layer.4.output 0.15746453 526.07967033 + ------------------------------------------------------------------------------------- + TOTAL 0.08006400 218.62793674 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 183200 +BPFP 1.8504 bits/point +EBPFP 3.7007 equivalent bits/point +MSE 218.627937 +---------------------- -------------------------------------------------------- +Time: 2.526s Load: 0.005s, Pack+Encode: 1.503s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 218.6279 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 116, 128) +Output shape: (1, 116, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.output: torch.Size([1, 116, 3584]) -> torch.Size([1, 1, 116, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,864B, BPFP=0.7899 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,768B, BPFP=3.2015 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,128B, BPFP=1.4989 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,320B, BPFP=3.1412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,720B, BPFP=1.8481 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,692B, BPFP=3.0566 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,592B, BPFP=1.6961 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,884B, BPFP=3.0824 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,008B, BPFP=2.5603 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,236B, BPFP=2.9952 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,988B, BPFP=0.8080 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.019s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16314862 26.97795158 + layer.0.v_cache 0.00001373 0.00222415 + layer.1.k_cache 0.05481785 2.33498540 + layer.1.v_cache 0.00000557 0.00097430 + layer.2.k_cache 0.00968372 0.57609216 + layer.2.v_cache 0.00001954 0.00311929 + layer.3.k_cache 0.02418142 4.24870826 + layer.3.v_cache 0.00001860 0.00321315 + layer.4.k_cache 0.00062333 0.08136683 + layer.4.v_cache 0.00005467 0.00675125 + layer.4.output 9.83417319 404.50292488 + ------------------------------------------------------------------------------------- + TOTAL 4.06422232 168.57387415 + (elements=1,009,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1009664 +Total Bytes 219200 +BPFP 1.7368 bits/point +EBPFP 3.4736 equivalent bits/point +MSE 168.573874 +---------------------- -------------------------------------------------------- +Time: 2.528s Load: 0.006s, Pack+Encode: 1.503s, Decode+Unpack: 1.019s +---------------------- -------------------------------------------------------- +💾 Converting with 168.5739 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,600B, BPFP=0.9215 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,320B, BPFP=3.8702 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,720B, BPFP=1.5465 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,916B, BPFP=3.7893 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,808B, BPFP=1.7644 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,440B, BPFP=3.6939 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,236B, BPFP=1.6498 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,364B, BPFP=3.6787 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,320B, BPFP=2.6683 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,352B, BPFP=3.4760 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,520B, BPFP=0.9020 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11749315 28.46498773 + layer.0.v_cache 0.00001389 0.00232688 + layer.1.k_cache 0.03399578 2.12490982 + layer.1.v_cache 0.00000555 0.00099182 + layer.2.k_cache 0.00854440 0.67183235 + layer.2.v_cache 0.00001777 0.00281534 + layer.3.k_cache 0.02740619 2.92699569 + layer.3.v_cache 0.00001935 0.00343321 + layer.4.k_cache 0.00063531 0.07640798 + layer.4.v_cache 0.00005166 0.00690054 + layer.4.output 0.18490290 638.72550366 + ------------------------------------------------------------------------------------- + TOTAL 0.08720608 265.02118394 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 166596 +BPFP 1.9631 bits/point +EBPFP 3.9262 equivalent bits/point +MSE 265.021184 +---------------------- -------------------------------------------------------- +Time: 2.524s Load: 0.004s, Pack+Encode: 1.503s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 265.0212 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,008B, BPFP=0.8324 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,424B, BPFP=3.5612 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,584B, BPFP=1.4269 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,736B, BPFP=3.4468 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,360B, BPFP=1.7221 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,272B, BPFP=3.3697 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,272B, BPFP=1.5412 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,216B, BPFP=3.3604 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,812B, BPFP=2.6283 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,940B, BPFP=3.3145 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,368B, BPFP=0.8161 +⌛️ [2/4] FRONTEND: Frontend time: 1.505s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11492289 26.08959701 + layer.0.v_cache 0.00001440 0.00220383 + layer.1.k_cache 0.01112092 1.74915622 + layer.1.v_cache 0.00000555 0.00093054 + layer.2.k_cache 0.01096075 0.65354652 + layer.2.v_cache 0.00001882 0.00294570 + layer.3.k_cache 0.03781782 2.99496460 + layer.3.v_cache 0.00001905 0.00309944 + layer.4.k_cache 0.00062716 0.07860823 + layer.4.v_cache 0.00006343 0.00648057 + layer.4.output 0.15224552 534.82009878 + ------------------------------------------------------------------------------------- + TOTAL 0.07301703 222.07777789 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 185992 +BPFP 1.8186 bits/point +EBPFP 3.6372 equivalent bits/point +MSE 222.077778 +---------------------- -------------------------------------------------------- +Time: 2.529s Load: 0.005s, Pack+Encode: 1.505s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 222.0778 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,564B, BPFP=0.9261 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,052B, BPFP=3.8661 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,732B, BPFP=1.5690 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,760B, BPFP=3.8068 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,960B, BPFP=1.8182 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,072B, BPFP=3.6672 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,244B, BPFP=1.6729 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,368B, BPFP=3.7273 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,448B, BPFP=2.7289 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,484B, BPFP=3.5479 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,380B, BPFP=0.9097 +⌛️ [2/4] FRONTEND: Frontend time: 1.505s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10405910 25.09310319 + layer.0.v_cache 0.00001453 0.00227825 + layer.1.k_cache 0.03397077 1.88635294 + layer.1.v_cache 0.00000547 0.00093169 + layer.2.k_cache 0.00568739 0.78655748 + layer.2.v_cache 0.00001916 0.00286742 + layer.3.k_cache 0.09581309 2.90594760 + layer.3.v_cache 0.00001862 0.00310549 + layer.4.k_cache 0.00061896 0.07632849 + layer.4.v_cache 0.00005653 0.00651034 + layer.4.output 0.19123089 645.64187152 + ------------------------------------------------------------------------------------- + TOTAL 0.09287528 267.66218138 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 166064 +BPFP 1.9822 bits/point +EBPFP 3.9645 equivalent bits/point +MSE 267.662181 +---------------------- -------------------------------------------------------- +Time: 2.528s Load: 0.005s, Pack+Encode: 1.505s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 267.6622 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,768B, BPFP=0.8765 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,972B, BPFP=3.6713 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,320B, BPFP=1.5294 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,644B, BPFP=3.6110 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,644B, BPFP=1.7728 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,728B, BPFP=3.4426 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,884B, BPFP=1.6331 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,392B, BPFP=3.5647 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,792B, BPFP=2.7191 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,752B, BPFP=3.4471 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,788B, BPFP=0.8873 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09384689 27.59315257 + layer.0.v_cache 0.00001364 0.00226210 + layer.1.k_cache 0.01313462 2.03902265 + layer.1.v_cache 0.00000592 0.00102918 + layer.2.k_cache 0.00813576 0.66551648 + layer.2.v_cache 0.00001762 0.00289023 + layer.3.k_cache 0.02628646 3.08075274 + layer.3.v_cache 0.00002044 0.00329262 + layer.4.k_cache 0.00061596 0.07785033 + layer.4.v_cache 0.00005334 0.00684373 + layer.4.output 0.17343949 579.28576681 + ------------------------------------------------------------------------------------- + TOTAL 0.07977689 240.49841060 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 176684 +BPFP 1.9105 bits/point +EBPFP 3.8210 equivalent bits/point +MSE 240.498411 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.003s, Pack+Encode: 1.500s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 240.4984 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 126, 128) +Output shape: (1, 126, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.output: torch.Size([1, 126, 3584]) -> torch.Size([1, 1, 126, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,016B, BPFP=0.7460 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,920B, BPFP=2.9663 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,572B, BPFP=1.4350 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,360B, BPFP=2.8968 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,568B, BPFP=1.6825 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,688B, BPFP=2.8135 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,868B, BPFP=1.5957 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 23,016B, BPFP=2.8542 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 19,132B, BPFP=2.3725 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 22,188B, BPFP=2.7515 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,264B, BPFP=0.7842 +⌛️ [2/4] FRONTEND: Frontend time: 1.508s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13141363 30.95614769 + layer.0.v_cache 0.00001429 0.00218253 + layer.1.k_cache 0.06496696 2.68557885 + layer.1.v_cache 0.00000583 0.00104801 + layer.2.k_cache 0.01057190 0.53358847 + layer.2.v_cache 0.00002020 0.00311671 + layer.3.k_cache 0.05250616 4.52207390 + layer.3.v_cache 0.00001853 0.00327980 + layer.4.k_cache 0.00062861 0.08313578 + layer.4.v_cache 0.00004930 0.00666597 + layer.4.output 9.05470577 364.04372166 + ------------------------------------------------------------------------------------- + TOTAL 3.74371387 152.18252172 + (elements=1,096,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1096704 +Total Bytes 222592 +BPFP 1.6237 bits/point +EBPFP 3.2474 equivalent bits/point +MSE 152.182522 +---------------------- -------------------------------------------------------- +Time: 2.533s Load: 0.005s, Pack+Encode: 1.508s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 152.1825 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,900B, BPFP=0.8800 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,632B, BPFP=3.7055 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,336B, BPFP=1.4971 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,716B, BPFP=3.5409 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,732B, BPFP=1.7478 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,624B, BPFP=3.5244 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,068B, BPFP=1.6286 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,768B, BPFP=3.5503 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,216B, BPFP=2.7328 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,536B, BPFP=3.5086 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,900B, BPFP=0.8441 +⌛️ [2/4] FRONTEND: Frontend time: 1.507s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13059476 26.94292441 + layer.0.v_cache 0.00001380 0.00228285 + layer.1.k_cache 0.03034586 1.55353274 + layer.1.v_cache 0.00000531 0.00097801 + layer.2.k_cache 0.00374181 0.72798657 + layer.2.v_cache 0.00001886 0.00296631 + layer.3.k_cache 0.04454902 2.61633213 + layer.3.v_cache 0.00001968 0.00346043 + layer.4.k_cache 0.00060938 0.08015385 + layer.4.v_cache 0.00007316 0.00682796 + layer.4.output 0.16485390 523.43113711 + ------------------------------------------------------------------------------------- + TOTAL 0.08023229 217.40914147 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 179428 +BPFP 1.8956 bits/point +EBPFP 3.7912 equivalent bits/point +MSE 217.409141 +---------------------- -------------------------------------------------------- +Time: 2.522s Load: 0.003s, Pack+Encode: 1.507s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 217.4091 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,036B, BPFP=0.8553 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,300B, BPFP=3.6175 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,568B, BPFP=1.4552 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,972B, BPFP=3.5618 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,944B, BPFP=1.6889 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,360B, BPFP=3.4579 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,244B, BPFP=1.5700 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,772B, BPFP=3.5279 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,468B, BPFP=2.7969 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,180B, BPFP=3.4273 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,480B, BPFP=0.8851 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15436328 25.13230830 + layer.0.v_cache 0.00001586 0.00223169 + layer.1.k_cache 0.01080767 1.73587434 + layer.1.v_cache 0.00000575 0.00101117 + layer.2.k_cache 0.00786338 0.64377030 + layer.2.v_cache 0.00001961 0.00295229 + layer.3.k_cache 0.05652578 3.18466784 + layer.3.v_cache 0.00002042 0.00329888 + layer.4.k_cache 0.00062828 0.08459892 + layer.4.v_cache 0.00005674 0.00687127 + layer.4.output 0.15192759 537.01999224 + ------------------------------------------------------------------------------------- + TOTAL 0.07610588 222.93750180 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 189324 +BPFP 1.8914 bits/point +EBPFP 3.7828 equivalent bits/point +MSE 222.937502 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.004s, Pack+Encode: 1.502s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 222.9375 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,628B, BPFP=0.9271 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,404B, BPFP=3.8870 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,752B, BPFP=1.5529 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,744B, BPFP=3.7548 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,092B, BPFP=1.8213 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,140B, BPFP=3.6338 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,288B, BPFP=1.6603 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,052B, BPFP=3.8165 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,960B, BPFP=2.7965 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,340B, BPFP=3.6739 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,648B, BPFP=0.9057 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10254851 28.21659655 + layer.0.v_cache 0.00001399 0.00229407 + layer.1.k_cache 0.05476840 1.92045163 + layer.1.v_cache 0.00000542 0.00095673 + layer.2.k_cache 0.00244555 0.77123666 + layer.2.v_cache 0.00001844 0.00289006 + layer.3.k_cache 0.02811479 2.88710726 + layer.3.v_cache 0.00001905 0.00332277 + layer.4.k_cache 0.00060111 0.07612711 + layer.4.v_cache 0.00005313 0.00688252 + layer.4.output 0.19635826 643.35439560 + ------------------------------------------------------------------------------------- + TOTAL 0.09194683 266.90403733 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 169048 +BPFP 1.9920 bits/point +EBPFP 3.9840 equivalent bits/point +MSE 266.904037 +---------------------- -------------------------------------------------------- +Time: 2.520s Load: 0.003s, Pack+Encode: 1.500s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 266.9040 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,768B, BPFP=0.9085 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,388B, BPFP=3.6944 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,960B, BPFP=1.5168 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,032B, BPFP=3.6265 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,236B, BPFP=1.7599 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,384B, BPFP=3.5030 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,564B, BPFP=1.6319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,868B, BPFP=3.5953 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,196B, BPFP=2.7050 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,260B, BPFP=3.4794 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,576B, BPFP=0.8868 +⌛️ [2/4] FRONTEND: Frontend time: 1.504s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11987135 26.68785132 + layer.0.v_cache 0.00001390 0.00216529 + layer.1.k_cache 0.03139273 2.10015385 + layer.1.v_cache 0.00000562 0.00094415 + layer.2.k_cache 0.00816743 0.68224121 + layer.2.v_cache 0.00001857 0.00291565 + layer.3.k_cache 0.02640573 3.34919255 + layer.3.v_cache 0.00001968 0.00314496 + layer.4.k_cache 0.00062091 0.07901214 + layer.4.v_cache 0.00004972 0.00668366 + layer.4.output 0.16983549 576.27362805 + ------------------------------------------------------------------------------------- + TOTAL 0.08090671 239.22527654 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 171232 +BPFP 1.9193 bits/point +EBPFP 3.8386 equivalent bits/point +MSE 239.225277 +---------------------- -------------------------------------------------------- +Time: 2.525s Load: 0.005s, Pack+Encode: 1.504s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 239.2253 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,880B, BPFP=0.8764 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,552B, BPFP=3.6911 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,440B, BPFP=1.5158 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,064B, BPFP=3.6034 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,780B, BPFP=1.7565 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,276B, BPFP=3.4619 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,148B, BPFP=1.6430 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,020B, BPFP=3.5955 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,616B, BPFP=2.8046 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,632B, BPFP=3.5259 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,640B, BPFP=0.8631 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.021s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14089697 26.06855020 + layer.0.v_cache 0.00001393 0.00231099 + layer.1.k_cache 0.01618495 1.94606948 + layer.1.v_cache 0.00000582 0.00100868 + layer.2.k_cache 0.00661465 0.68121500 + layer.2.v_cache 0.00002083 0.00300321 + layer.3.k_cache 0.07087009 3.18971866 + layer.3.v_cache 0.00001822 0.00347448 + layer.4.k_cache 0.00061891 0.08354910 + layer.4.v_cache 0.00005311 0.00701793 + layer.4.output 0.16425455 414.89927135 + ------------------------------------------------------------------------------------- + TOTAL 0.08147525 172.72240101 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 181048 +BPFP 1.9127 bits/point +EBPFP 3.8254 equivalent bits/point +MSE 172.722401 +---------------------- -------------------------------------------------------- +Time: 2.528s Load: 0.004s, Pack+Encode: 1.503s, Decode+Unpack: 1.021s +---------------------- -------------------------------------------------------- +💾 Converting with 172.7224 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,848B, BPFP=0.8511 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,176B, BPFP=3.7177 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,424B, BPFP=1.4789 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,448B, BPFP=3.5899 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,912B, BPFP=1.7402 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,840B, BPFP=3.4831 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,184B, BPFP=1.6124 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,548B, BPFP=3.6074 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,340B, BPFP=2.6931 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,516B, BPFP=3.4263 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,968B, BPFP=0.8770 +⌛️ [2/4] FRONTEND: Frontend time: 1.499s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.017s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12786581 26.37659926 + layer.0.v_cache 0.00001426 0.00229865 + layer.1.k_cache 0.03183105 1.79248441 + layer.1.v_cache 0.00000593 0.00099814 + layer.2.k_cache 0.00241907 0.79531826 + layer.2.v_cache 0.00001785 0.00276865 + layer.3.k_cache 0.02946699 2.77637645 + layer.3.v_cache 0.00001957 0.00326005 + layer.4.k_cache 0.00063747 0.08067771 + layer.4.v_cache 0.00005003 0.00674064 + layer.4.output 0.16568538 530.98244382 + ------------------------------------------------------------------------------------- + TOTAL 0.07953681 220.51262523 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 184204 +BPFP 1.9023 bits/point +EBPFP 3.8046 equivalent bits/point +MSE 220.512625 +---------------------- -------------------------------------------------------- +Time: 2.520s Load: 0.005s, Pack+Encode: 1.499s, Decode+Unpack: 1.017s +---------------------- -------------------------------------------------------- +💾 Converting with 220.5126 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,636B, BPFP=0.9287 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,780B, BPFP=3.9623 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,840B, BPFP=1.5705 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,056B, BPFP=3.8173 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,112B, BPFP=1.8253 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,588B, BPFP=3.7236 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,456B, BPFP=1.6939 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,228B, BPFP=3.8518 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,472B, BPFP=2.6987 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,496B, BPFP=3.5048 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,708B, BPFP=0.9646 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12225728 25.90606220 + layer.0.v_cache 0.00001337 0.00227821 + layer.1.k_cache 0.03480693 1.90313056 + layer.1.v_cache 0.00000545 0.00097200 + layer.2.k_cache 0.00853881 0.68629915 + layer.2.v_cache 0.00001924 0.00306046 + layer.3.k_cache 0.06130744 3.13138737 + layer.3.v_cache 0.00001928 0.00332623 + layer.4.k_cache 0.00058936 0.07481103 + layer.4.v_cache 0.00007197 0.00669719 + layer.4.output 0.18657821 635.36011905 + ------------------------------------------------------------------------------------- + TOTAL 0.09021627 263.48463869 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 171372 +BPFP 2.0194 bits/point +EBPFP 4.0387 equivalent bits/point +MSE 263.484639 +---------------------- -------------------------------------------------------- +Time: 2.523s Load: 0.003s, Pack+Encode: 1.502s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 263.4846 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,596B, BPFP=0.9326 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,900B, BPFP=3.8352 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,720B, BPFP=1.5666 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,324B, BPFP=3.7183 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,928B, BPFP=1.8117 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,452B, BPFP=3.7443 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,324B, BPFP=1.6891 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,252B, BPFP=3.7037 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,588B, BPFP=2.7573 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,472B, BPFP=3.5455 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,356B, BPFP=0.9090 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582910 25.87014097 + layer.0.v_cache 0.00001386 0.00230252 + layer.1.k_cache 0.03408930 1.93823599 + layer.1.v_cache 0.00000538 0.00100823 + layer.2.k_cache 0.00228146 0.71318877 + layer.2.v_cache 0.00001859 0.00315293 + layer.3.k_cache 0.06276476 3.05569815 + layer.3.v_cache 0.00001913 0.00349923 + layer.4.k_cache 0.00060520 0.07977166 + layer.4.v_cache 0.00005125 0.00700156 + layer.4.output 0.19356344 627.00434833 + ------------------------------------------------------------------------------------- + TOTAL 0.09238954 260.04143755 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 165912 +BPFP 1.9804 bits/point +EBPFP 3.9608 equivalent bits/point +MSE 260.041438 +---------------------- -------------------------------------------------------- +Time: 2.520s Load: 0.003s, Pack+Encode: 1.503s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 260.0414 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,796B, BPFP=0.9029 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,060B, BPFP=3.7764 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,024B, BPFP=1.5105 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,412B, BPFP=3.6544 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,392B, BPFP=1.7681 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,312B, BPFP=3.6355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,708B, BPFP=1.6393 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,552B, BPFP=3.6807 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,572B, BPFP=2.7432 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,788B, BPFP=3.5369 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,424B, BPFP=0.8989 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09700135 27.33356081 + layer.0.v_cache 0.00001355 0.00232731 + layer.1.k_cache 0.01103399 1.85728877 + layer.1.v_cache 0.00000598 0.00098232 + layer.2.k_cache 0.00794396 0.60467805 + layer.2.v_cache 0.00001935 0.00307707 + layer.3.k_cache 0.04207947 2.84196674 + layer.3.v_cache 0.00002010 0.00346920 + layer.4.k_cache 0.00062246 0.07896489 + layer.4.v_cache 0.00005226 0.00689790 + layer.4.output 0.17535226 534.84262048 + ------------------------------------------------------------------------------------- + TOTAL 0.08154460 222.15479744 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 176040 +BPFP 1.9494 bits/point +EBPFP 3.8988 equivalent bits/point +MSE 222.154797 +---------------------- -------------------------------------------------------- +Time: 2.523s Load: 0.003s, Pack+Encode: 1.502s, Decode+Unpack: 1.018s +---------------------- -------------------------------------------------------- +💾 Converting with 222.1548 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,788B, BPFP=0.8801 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,340B, BPFP=3.7390 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,244B, BPFP=1.5154 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,540B, BPFP=3.5919 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,568B, BPFP=1.7588 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,560B, BPFP=3.4118 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,912B, BPFP=1.6382 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,520B, BPFP=3.5882 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,848B, BPFP=2.7294 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,748B, BPFP=3.4463 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,224B, BPFP=0.8725 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.019s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11944316 25.74043256 + layer.0.v_cache 0.00001525 0.00222352 + layer.1.k_cache 0.01376067 2.04987578 + layer.1.v_cache 0.00000561 0.00092488 + layer.2.k_cache 0.00522648 0.73681982 + layer.2.v_cache 0.00001966 0.00287124 + layer.3.k_cache 0.04600189 3.18027703 + layer.3.v_cache 0.00001912 0.00312430 + layer.4.k_cache 0.00062079 0.07716296 + layer.4.v_cache 0.00005114 0.00665315 + layer.4.output 0.17175110 572.41622899 + ------------------------------------------------------------------------------------- + TOTAL 0.08161303 237.57140989 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 176292 +BPFP 1.9063 bits/point +EBPFP 3.8125 equivalent bits/point +MSE 237.571410 +---------------------- -------------------------------------------------------- +Time: 2.512s Load: 0.004s, Pack+Encode: 1.489s, Decode+Unpack: 1.019s +---------------------- -------------------------------------------------------- +💾 Converting with 237.5714 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,636B, BPFP=0.9287 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,452B, BPFP=3.8966 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,760B, BPFP=1.5545 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,932B, BPFP=3.7925 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,896B, BPFP=1.7821 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,080B, BPFP=3.6218 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,344B, BPFP=1.6715 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,688B, BPFP=3.7436 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,408B, BPFP=2.6859 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,724B, BPFP=3.5505 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,348B, BPFP=0.9257 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120220 28.75220353 + layer.0.v_cache 0.00001368 0.00232307 + layer.1.k_cache 0.03273370 1.85443585 + layer.1.v_cache 0.00000573 0.00094639 + layer.2.k_cache 0.00892877 0.70631687 + layer.2.v_cache 0.00001953 0.00294766 + layer.3.k_cache 0.02974504 2.91772363 + layer.3.v_cache 0.00001879 0.00321893 + layer.4.k_cache 0.00063874 0.07868898 + layer.4.v_cache 0.00005200 0.00656017 + layer.4.output 0.17853031 641.03531364 + ------------------------------------------------------------------------------------- + TOTAL 0.08429825 265.97485651 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 168268 +BPFP 1.9828 bits/point +EBPFP 3.9656 equivalent bits/point +MSE 265.974857 +---------------------- -------------------------------------------------------- +Time: 2.510s Load: 0.004s, Pack+Encode: 1.490s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 265.9749 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,796B, BPFP=0.9029 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,184B, BPFP=3.7997 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,188B, BPFP=1.5414 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,872B, BPFP=3.7410 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,444B, BPFP=1.7779 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,004B, BPFP=3.5776 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,900B, BPFP=1.6755 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,344B, BPFP=3.6416 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,440B, BPFP=2.7184 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,000B, BPFP=3.5768 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,652B, BPFP=0.9319 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09802409 26.88892484 + layer.0.v_cache 0.00001402 0.00235237 + layer.1.k_cache 0.03453328 2.16402132 + layer.1.v_cache 0.00000594 0.00097650 + layer.2.k_cache 0.01027090 0.75977353 + layer.2.v_cache 0.00001933 0.00291496 + layer.3.k_cache 0.02613793 3.23529788 + layer.3.v_cache 0.00001931 0.00312471 + layer.4.k_cache 0.00063818 0.07766828 + layer.4.v_cache 0.00005393 0.00648261 + layer.4.output 0.18084440 556.38812392 + ------------------------------------------------------------------------------------- + TOTAL 0.08444869 231.05049438 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 177824 +BPFP 1.9692 bits/point +EBPFP 3.9383 equivalent bits/point +MSE 231.050494 +---------------------- -------------------------------------------------------- +Time: 2.507s Load: 0.003s, Pack+Encode: 1.490s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 231.0505 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,496B, BPFP=0.8337 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,132B, BPFP=3.5091 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,064B, BPFP=1.5267 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,904B, BPFP=3.4745 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,760B, BPFP=1.7840 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,080B, BPFP=3.3495 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,728B, BPFP=1.6274 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,396B, BPFP=3.3975 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,240B, BPFP=2.7670 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,756B, BPFP=3.3004 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,668B, BPFP=0.8813 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13030120 28.97085008 + layer.0.v_cache 0.00001385 0.00224653 + layer.1.k_cache 0.02890817 1.98499957 + layer.1.v_cache 0.00000611 0.00101506 + layer.2.k_cache 0.00721959 0.71372904 + layer.2.v_cache 0.00001830 0.00296236 + layer.3.k_cache 0.02195997 3.69478096 + layer.3.v_cache 0.00001903 0.00320083 + layer.4.k_cache 0.00063766 0.08388938 + layer.4.v_cache 0.00005061 0.00650458 + layer.4.output 11.07892277 440.04195562 + ------------------------------------------------------------------------------------- + TOTAL 4.57303494 183.27987457 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 209224 +BPFP 1.8670 bits/point +EBPFP 3.7340 equivalent bits/point +MSE 183.279875 +---------------------- -------------------------------------------------------- +Time: 2.518s Load: 0.005s, Pack+Encode: 1.497s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 183.2799 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,776B, BPFP=0.8991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,856B, BPFP=3.7380 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,064B, BPFP=1.5181 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,444B, BPFP=3.6604 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,404B, BPFP=1.7703 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,984B, BPFP=3.5738 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,776B, BPFP=1.6521 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,760B, BPFP=3.7199 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,012B, BPFP=2.8261 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,956B, BPFP=3.5685 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,832B, BPFP=0.8830 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12633393 27.36135166 + layer.0.v_cache 0.00001372 0.00237141 + layer.1.k_cache 0.01411889 2.03714283 + layer.1.v_cache 0.00000549 0.00102725 + layer.2.k_cache 0.00833277 0.75422976 + layer.2.v_cache 0.00001851 0.00307781 + layer.3.k_cache 0.05797521 2.93317570 + layer.3.v_cache 0.00001879 0.00335924 + layer.4.k_cache 0.00064672 0.08269543 + layer.4.v_cache 0.00005349 0.00706744 + layer.4.output 0.17447430 553.63694062 + ------------------------------------------------------------------------------------- + TOTAL 0.08404927 229.92024017 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 175864 +BPFP 1.9475 bits/point +EBPFP 3.8949 equivalent bits/point +MSE 229.920240 +---------------------- -------------------------------------------------------- +Time: 2.514s Load: 0.003s, Pack+Encode: 1.498s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 229.9202 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,284B, BPFP=0.9563 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,048B, BPFP=3.8054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,028B, BPFP=1.5688 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,424B, BPFP=3.6661 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,184B, BPFP=1.8268 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,840B, BPFP=3.5357 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,544B, BPFP=1.6839 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,884B, BPFP=3.7687 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,160B, BPFP=2.7143 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,512B, BPFP=3.4625 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,148B, BPFP=0.9295 +⌛️ [2/4] FRONTEND: Frontend time: 1.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08086312 25.86740897 + layer.0.v_cache 0.00001331 0.00232595 + layer.1.k_cache 0.01551799 1.88338863 + layer.1.v_cache 0.00000513 0.00092280 + layer.2.k_cache 0.00232868 0.64945635 + layer.2.v_cache 0.00002262 0.00319293 + layer.3.k_cache 0.10623371 2.70720912 + layer.3.v_cache 0.00001836 0.00338005 + layer.4.k_cache 0.00061897 0.07910551 + layer.4.v_cache 0.00005922 0.00681853 + layer.4.output 0.20964142 655.25076531 + ------------------------------------------------------------------------------------- + TOTAL 0.09842183 271.64462153 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 150056 +BPFP 1.9703 bits/point +EBPFP 3.9405 equivalent bits/point +MSE 271.644622 +---------------------- -------------------------------------------------------- +Time: 2.513s Load: 0.004s, Pack+Encode: 1.497s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 271.6446 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,984B, BPFP=0.8465 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,188B, BPFP=3.5985 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,552B, BPFP=1.4524 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,556B, BPFP=3.4912 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,876B, BPFP=1.6773 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,852B, BPFP=3.3716 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,092B, BPFP=1.5442 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,372B, BPFP=3.4599 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,124B, BPFP=2.7385 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,724B, BPFP=3.3499 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,792B, BPFP=0.8199 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13670269 26.81908118 + layer.0.v_cache 0.00001568 0.00233269 + layer.1.k_cache 0.03139175 1.88966104 + layer.1.v_cache 0.00000602 0.00106183 + layer.2.k_cache 0.00629326 0.63314736 + layer.2.v_cache 0.00001871 0.00285579 + layer.3.k_cache 0.03837063 3.07074339 + layer.3.v_cache 0.00001860 0.00306590 + layer.4.k_cache 0.00062597 0.07875996 + layer.4.v_cache 0.00005354 0.00633794 + layer.4.output 0.15974679 517.97248641 + ------------------------------------------------------------------------------------- + TOTAL 0.07833673 215.19496776 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 184112 +BPFP 1.8394 bits/point +EBPFP 3.6787 equivalent bits/point +MSE 215.194968 +---------------------- -------------------------------------------------------- +Time: 2.514s Load: 0.003s, Pack+Encode: 1.495s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 215.1950 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,676B, BPFP=0.8289 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,444B, BPFP=3.4235 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,648B, BPFP=1.5549 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,908B, BPFP=3.3452 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,636B, BPFP=1.8452 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,572B, BPFP=3.2961 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,872B, BPFP=1.7336 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,796B, BPFP=3.3289 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,548B, BPFP=2.7085 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,952B, BPFP=3.2056 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,300B, BPFP=0.8407 +⌛️ [2/4] FRONTEND: Frontend time: 1.499s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.016s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12281527 25.67119506 + layer.0.v_cache 0.00001417 0.00225545 + layer.1.k_cache 0.04442539 2.09596552 + layer.1.v_cache 0.00000572 0.00095488 + layer.2.k_cache 0.01033308 0.79595562 + layer.2.v_cache 0.00001923 0.00289939 + layer.3.k_cache 0.04555132 4.12939539 + layer.3.v_cache 0.00001841 0.00308350 + layer.4.k_cache 0.00060580 0.07894283 + layer.4.v_cache 0.00007068 0.00660605 + layer.4.output 10.66341841 449.74240654 + ------------------------------------------------------------------------------------- + TOTAL 4.40398753 187.11671173 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 213352 +BPFP 1.8327 bits/point +EBPFP 3.6653 equivalent bits/point +MSE 187.116712 +---------------------- -------------------------------------------------------- +Time: 2.521s Load: 0.006s, Pack+Encode: 1.499s, Decode+Unpack: 1.016s +---------------------- -------------------------------------------------------- +💾 Converting with 187.1167 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,796B, BPFP=0.8714 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,680B, BPFP=3.7573 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,416B, BPFP=1.5291 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,900B, BPFP=3.6156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,724B, BPFP=1.7667 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,352B, BPFP=3.5160 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,024B, BPFP=1.6395 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,000B, BPFP=3.6337 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,024B, BPFP=2.7297 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,168B, BPFP=3.4826 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,912B, BPFP=0.8542 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.019s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11362342 26.96359466 + layer.0.v_cache 0.00001418 0.00233018 + layer.1.k_cache 0.05429294 1.93333577 + layer.1.v_cache 0.00000580 0.00102100 + layer.2.k_cache 0.00513919 0.71225033 + layer.2.v_cache 0.00001816 0.00301455 + layer.3.k_cache 0.05885953 2.86414053 + layer.3.v_cache 0.00001841 0.00333010 + layer.4.k_cache 0.00061763 0.07966378 + layer.4.v_cache 0.00005080 0.00677455 + layer.4.output 0.16727475 577.70322882 + ------------------------------------------------------------------------------------- + TOTAL 0.08256255 239.79365042 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 178996 +BPFP 1.9130 bits/point +EBPFP 3.8260 equivalent bits/point +MSE 239.793650 +---------------------- -------------------------------------------------------- +Time: 2.527s Load: 0.005s, Pack+Encode: 1.503s, Decode+Unpack: 1.019s +---------------------- -------------------------------------------------------- +💾 Converting with 239.7937 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,012B, BPFP=0.8421 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,252B, BPFP=3.5706 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,680B, BPFP=1.4583 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,576B, BPFP=3.4570 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,480B, BPFP=1.7608 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,940B, BPFP=3.3501 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,208B, BPFP=1.5470 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,336B, BPFP=3.4167 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,816B, BPFP=2.6573 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,008B, BPFP=3.3616 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,508B, BPFP=0.8282 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702597 25.74687605 + layer.0.v_cache 0.00001402 0.00231793 + layer.1.k_cache 0.01332537 2.15123872 + layer.1.v_cache 0.00000581 0.00098267 + layer.2.k_cache 0.00507878 0.76550293 + layer.2.v_cache 0.00001944 0.00288369 + layer.3.k_cache 0.04051446 2.99807641 + layer.3.v_cache 0.00001931 0.00314888 + layer.4.k_cache 0.00062870 0.08028440 + layer.4.v_cache 0.00006514 0.00657534 + layer.4.output 0.16549673 527.47638249 + ------------------------------------------------------------------------------------- + TOTAL 0.07795142 219.06426850 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 185816 +BPFP 1.8364 bits/point +EBPFP 3.6728 equivalent bits/point +MSE 219.064268 +---------------------- -------------------------------------------------------- +Time: 2.493s Load: 0.005s, Pack+Encode: 1.483s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 219.0643 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,772B, BPFP=0.8772 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,492B, BPFP=3.7669 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,164B, BPFP=1.5007 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,192B, BPFP=3.7118 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,428B, BPFP=1.7331 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,988B, BPFP=3.4904 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,712B, BPFP=1.6015 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,592B, BPFP=3.6015 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,804B, BPFP=2.7213 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,624B, BPFP=3.4235 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,824B, BPFP=0.8620 +⌛️ [2/4] FRONTEND: Frontend time: 1.485s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.021s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13925676 29.97656824 + layer.0.v_cache 0.00001387 0.00221958 + layer.1.k_cache 0.03290265 1.71608851 + layer.1.v_cache 0.00000589 0.00095439 + layer.2.k_cache 0.00800494 0.65431765 + layer.2.v_cache 0.00001921 0.00286472 + layer.3.k_cache 0.08301921 3.56676313 + layer.3.v_cache 0.00001931 0.00304525 + layer.4.k_cache 0.00065893 0.07871852 + layer.4.v_cache 0.00004891 0.00654241 + layer.4.output 0.17367665 587.32983193 + ------------------------------------------------------------------------------------- + TOTAL 0.08704037 243.95981800 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 176592 +BPFP 1.9095 bits/point +EBPFP 3.8190 equivalent bits/point +MSE 243.959818 +---------------------- -------------------------------------------------------- +Time: 2.512s Load: 0.005s, Pack+Encode: 1.485s, Decode+Unpack: 1.021s +---------------------- -------------------------------------------------------- +💾 Converting with 243.9598 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,476B, BPFP=0.8307 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,132B, BPFP=3.5091 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,796B, BPFP=1.4860 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,596B, BPFP=3.4278 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,560B, BPFP=1.7536 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,928B, BPFP=3.3265 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,664B, BPFP=1.6177 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,192B, BPFP=3.3665 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,312B, BPFP=2.7779 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,720B, BPFP=3.2949 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,572B, BPFP=0.8142 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10583916 28.38412564 + layer.0.v_cache 0.00001406 0.00229539 + layer.1.k_cache 0.01351924 2.38320686 + layer.1.v_cache 0.00000536 0.00098425 + layer.2.k_cache 0.00567923 0.68160581 + layer.2.v_cache 0.00001856 0.00295926 + layer.3.k_cache 0.05293757 3.50838611 + layer.3.v_cache 0.00001841 0.00319736 + layer.4.k_cache 0.00064880 0.08518028 + layer.4.v_cache 0.00004836 0.00649781 + layer.4.output 11.07886250 433.77565881 + ------------------------------------------------------------------------------------- + TOTAL 4.57239802 180.67576767 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 204948 +BPFP 1.8288 bits/point +EBPFP 3.6577 equivalent bits/point +MSE 180.675768 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.006s, Pack+Encode: 1.500s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 180.6758 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,784B, BPFP=0.9006 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,992B, BPFP=3.7636 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,116B, BPFP=1.5279 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,280B, BPFP=3.6295 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,388B, BPFP=1.7673 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,844B, BPFP=3.5474 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,748B, BPFP=1.6468 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,848B, BPFP=3.5482 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,268B, BPFP=2.6860 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,604B, BPFP=3.5023 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,328B, BPFP=0.8694 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.009s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860243 25.04687794 + layer.0.v_cache 0.00001403 0.00236466 + layer.1.k_cache 0.01400426 1.85640854 + layer.1.v_cache 0.00000584 0.00106013 + layer.2.k_cache 0.01277778 0.73396237 + layer.2.v_cache 0.00001980 0.00317189 + layer.3.k_cache 0.07376247 2.79838654 + layer.3.v_cache 0.00001877 0.00338530 + layer.4.k_cache 0.00062152 0.07659521 + layer.4.v_cache 0.00005140 0.00708607 + layer.4.output 0.18179212 579.34773021 + ------------------------------------------------------------------------------------- + TOTAL 0.08837783 240.35078883 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 173200 +BPFP 1.9180 bits/point +EBPFP 3.8359 equivalent bits/point +MSE 240.350789 +---------------------- -------------------------------------------------------- +Time: 2.506s Load: 0.004s, Pack+Encode: 1.493s, Decode+Unpack: 1.009s +---------------------- -------------------------------------------------------- +💾 Converting with 240.3508 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,964B, BPFP=0.8618 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,096B, BPFP=3.6625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,508B, BPFP=1.4771 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,436B, BPFP=3.5479 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,936B, BPFP=1.7250 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,108B, BPFP=3.4910 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,144B, BPFP=1.5875 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,448B, BPFP=3.5500 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,364B, BPFP=2.8410 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,984B, BPFP=3.4694 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,544B, BPFP=0.8815 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.008s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13781073 26.70361328 + layer.0.v_cache 0.00001414 0.00230070 + layer.1.k_cache 0.01252146 1.70240072 + layer.1.v_cache 0.00000561 0.00097706 + layer.2.k_cache 0.00238173 0.70931892 + layer.2.v_cache 0.00001849 0.00289953 + layer.3.k_cache 0.04360520 3.05137193 + layer.3.v_cache 0.00001973 0.00320177 + layer.4.k_cache 0.00061668 0.08174603 + layer.4.v_cache 0.00005448 0.00680952 + layer.4.output 0.16772948 529.62767857 + ------------------------------------------------------------------------------------- + TOTAL 0.08065615 219.97990526 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 186532 +BPFP 1.9049 bits/point +EBPFP 3.8099 equivalent bits/point +MSE 219.979905 +---------------------- -------------------------------------------------------- +Time: 2.509s Load: 0.005s, Pack+Encode: 1.496s, Decode+Unpack: 1.008s +---------------------- -------------------------------------------------------- +💾 Converting with 219.9799 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,672B, BPFP=0.9241 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,288B, BPFP=3.8149 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,764B, BPFP=1.5356 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,988B, BPFP=3.7555 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,132B, BPFP=1.8062 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,504B, BPFP=3.6598 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,388B, BPFP=1.6590 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,908B, BPFP=3.7397 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,280B, BPFP=2.8244 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,284B, BPFP=3.6163 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,064B, BPFP=0.9060 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07550411 26.99440331 + layer.0.v_cache 0.00001381 0.00234975 + layer.1.k_cache 0.01271453 1.83471448 + layer.1.v_cache 0.00000531 0.00099880 + layer.2.k_cache 0.00681068 0.87476532 + layer.2.v_cache 0.00001904 0.00323790 + layer.3.k_cache 0.02844424 2.54955533 + layer.3.v_cache 0.00001798 0.00330120 + layer.4.k_cache 0.00061127 0.08102166 + layer.4.v_cache 0.00009323 0.00710414 + layer.4.output 0.18255964 606.48321655 + ------------------------------------------------------------------------------------- + TOTAL 0.08247951 251.63140987 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 170272 +BPFP 1.9810 bits/point +EBPFP 3.9620 equivalent bits/point +MSE 251.631410 +---------------------- -------------------------------------------------------- +Time: 2.502s Load: 0.004s, Pack+Encode: 1.490s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 251.6314 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,724B, BPFP=0.9227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,336B, BPFP=3.7766 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,952B, BPFP=1.5531 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,420B, BPFP=3.5977 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,132B, BPFP=1.7836 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,268B, BPFP=3.5680 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,428B, BPFP=1.6461 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,000B, BPFP=3.5156 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,788B, BPFP=2.6930 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,772B, BPFP=3.4711 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,064B, BPFP=0.9225 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08590809 26.63934631 + layer.0.v_cache 0.00001388 0.00230768 + layer.1.k_cache 0.01594085 2.00410290 + layer.1.v_cache 0.00000552 0.00096669 + layer.2.k_cache 0.00791870 0.68610935 + layer.2.v_cache 0.00001801 0.00298797 + layer.3.k_cache 0.02757127 2.67440224 + layer.3.v_cache 0.00001862 0.00314479 + layer.4.k_cache 0.00063773 0.07709004 + layer.4.v_cache 0.00005492 0.00647324 + layer.4.output 0.18408036 562.52784598 + ------------------------------------------------------------------------------------- + TOTAL 0.08392059 233.51716783 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 168884 +BPFP 1.9403 bits/point +EBPFP 3.8806 equivalent bits/point +MSE 233.517168 +---------------------- -------------------------------------------------------- +Time: 2.498s Load: 0.005s, Pack+Encode: 1.489s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 233.5172 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,560B, BPFP=0.9375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,944B, BPFP=3.8947 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,548B, BPFP=1.5518 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,160B, BPFP=3.7336 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,788B, BPFP=1.8067 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,848B, BPFP=3.6694 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,032B, BPFP=1.6513 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,012B, BPFP=3.7031 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,120B, BPFP=2.6974 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,748B, BPFP=3.4433 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,864B, BPFP=0.9065 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.008s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08991440 26.45822947 + layer.0.v_cache 0.00001347 0.00229947 + layer.1.k_cache 0.03345199 1.97197924 + layer.1.v_cache 0.00000518 0.00098732 + layer.2.k_cache 0.00398020 0.68380336 + layer.2.v_cache 0.00001815 0.00293212 + layer.3.k_cache 0.04607732 3.25806146 + layer.3.v_cache 0.00001896 0.00317049 + layer.4.k_cache 0.00062119 0.07508912 + layer.4.v_cache 0.00005040 0.00629040 + layer.4.output 0.18612516 498.04957707 + ------------------------------------------------------------------------------------- + TOTAL 0.08688396 206.98881658 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 162624 +BPFP 1.9667 bits/point +EBPFP 3.9334 equivalent bits/point +MSE 206.988817 +---------------------- -------------------------------------------------------- +Time: 2.506s Load: 0.005s, Pack+Encode: 1.493s, Decode+Unpack: 1.008s +---------------------- -------------------------------------------------------- +💾 Converting with 206.9888 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,008B, BPFP=0.8599 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,272B, BPFP=3.6525 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,504B, BPFP=1.4602 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,596B, BPFP=3.5364 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,592B, BPFP=1.8187 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,020B, BPFP=3.4375 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,280B, BPFP=1.5934 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,768B, BPFP=3.5659 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,988B, BPFP=2.7452 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,096B, BPFP=3.4505 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,136B, BPFP=0.8619 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11186004 27.06490921 + layer.0.v_cache 0.00001364 0.00225175 + layer.1.k_cache 0.01386243 1.87043276 + layer.1.v_cache 0.00000608 0.00098094 + layer.2.k_cache 0.00506424 0.70635249 + layer.2.v_cache 0.00001872 0.00294107 + layer.3.k_cache 0.02496567 2.71094890 + layer.3.v_cache 0.00001876 0.00325671 + layer.4.k_cache 0.00064518 0.08012243 + layer.4.v_cache 0.00005436 0.00676040 + layer.4.output 0.16682192 525.15163854 + ------------------------------------------------------------------------------------- + TOTAL 0.07789780 218.14767214 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 187260 +BPFP 1.8914 bits/point +EBPFP 3.7827 equivalent bits/point +MSE 218.147672 +---------------------- -------------------------------------------------------- +Time: 2.495s Load: 0.003s, Pack+Encode: 1.484s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 218.1477 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,804B, BPFP=0.9044 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,184B, BPFP=3.7997 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,984B, BPFP=1.5030 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,776B, BPFP=3.7229 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,240B, BPFP=1.7395 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,056B, BPFP=3.5873 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,628B, BPFP=1.6242 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,156B, BPFP=3.6062 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,352B, BPFP=2.7018 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,612B, BPFP=3.5038 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,768B, BPFP=0.9081 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.019s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08542032 25.18810594 + layer.0.v_cache 0.00001402 0.00232497 + layer.1.k_cache 0.01083013 1.60661022 + layer.1.v_cache 0.00000556 0.00105600 + layer.2.k_cache 0.00382075 0.70246478 + layer.2.v_cache 0.00001860 0.00301363 + layer.3.k_cache 0.02610962 2.62078269 + layer.3.v_cache 0.00002051 0.00337362 + layer.4.k_cache 0.00063312 0.07531927 + layer.4.v_cache 0.00005614 0.00718516 + layer.4.output 0.16811641 582.80394793 + ------------------------------------------------------------------------------------- + TOTAL 0.07669080 241.75516893 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 175560 +BPFP 1.9441 bits/point +EBPFP 3.8882 equivalent bits/point +MSE 241.755169 +---------------------- -------------------------------------------------------- +Time: 2.514s Load: 0.005s, Pack+Encode: 1.491s, Decode+Unpack: 1.019s +---------------------- -------------------------------------------------------- +💾 Converting with 241.7552 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 39.735s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,720B, BPFP=0.9105 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,544B, BPFP=3.7701 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,012B, BPFP=1.5455 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,100B, BPFP=3.6844 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,292B, BPFP=1.7924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,484B, BPFP=3.5656 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,624B, BPFP=1.6636 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,880B, BPFP=3.6420 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,060B, BPFP=2.7122 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,512B, BPFP=3.5710 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,880B, BPFP=0.9061 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.018s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10540552 23.20414677 + layer.0.v_cache 0.00001541 0.00227613 + layer.1.k_cache 0.03402772 2.08658327 + layer.1.v_cache 0.00000555 0.00101314 + layer.2.k_cache 0.00394259 0.68744575 + layer.2.v_cache 0.00001868 0.00300848 + layer.3.k_cache 0.02873192 2.52709113 + layer.3.v_cache 0.00001983 0.00327704 + layer.4.k_cache 0.00063741 0.08161177 + layer.4.v_cache 0.00005806 0.00724503 + layer.4.output 0.17961645 462.57947531 + ------------------------------------------------------------------------------------- + TOTAL 0.08412811 192.15647210 + (elements=705,024) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 705024 +Total Bytes 172108 +BPFP 1.9529 bits/point +EBPFP 3.9059 equivalent bits/point +MSE 192.156472 +---------------------- --------------------------------------------------------- +Time: 42.248s Load: 39.735s, Pack+Encode: 1.496s, Decode+Unpack: 1.018s +---------------------- --------------------------------------------------------- +💾 Converting with 192.1565 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,676B, BPFP=0.9133 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,316B, BPFP=3.7727 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,896B, BPFP=1.5422 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,256B, BPFP=3.7609 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,284B, BPFP=1.8133 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,400B, BPFP=3.5938 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,544B, BPFP=1.6687 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,812B, BPFP=3.6742 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,876B, BPFP=2.7102 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,456B, BPFP=3.6047 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,128B, BPFP=0.9522 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10022384 25.10536194 + layer.0.v_cache 0.00001350 0.00234309 + layer.1.k_cache 0.03445883 2.29166584 + layer.1.v_cache 0.00000556 0.00105505 + layer.2.k_cache 0.00550650 0.69915142 + layer.2.v_cache 0.00001848 0.00300578 + layer.3.k_cache 0.03016714 2.71247177 + layer.3.v_cache 0.00001937 0.00365202 + layer.4.k_cache 0.00063435 0.07741007 + layer.4.v_cache 0.00006510 0.00700587 + layer.4.output 0.18593751 562.35189732 + ------------------------------------------------------------------------------------- + TOTAL 0.08662796 233.37449436 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 172644 +BPFP 1.9835 bits/point +EBPFP 3.9670 equivalent bits/point +MSE 233.374494 +---------------------- -------------------------------------------------------- +Time: 2.506s Load: 0.005s, Pack+Encode: 1.496s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 233.3745 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,952B, BPFP=0.7977 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,112B, BPFP=3.5619 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,684B, BPFP=1.3988 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,724B, BPFP=3.4994 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,648B, BPFP=1.7152 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,080B, BPFP=3.3956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,524B, BPFP=1.5341 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,060B, BPFP=3.3924 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,060B, BPFP=2.7481 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,596B, BPFP=3.3177 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,268B, BPFP=0.8346 +⌛️ [2/4] FRONTEND: Frontend time: 1.484s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.007s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11898830 28.25408747 + layer.0.v_cache 0.00001403 0.00223581 + layer.1.k_cache 0.01243949 1.82665465 + layer.1.v_cache 0.00000551 0.00097177 + layer.2.k_cache 0.00346032 0.62681922 + layer.2.v_cache 0.00001901 0.00296438 + layer.3.k_cache 0.02403709 2.71697935 + layer.3.v_cache 0.00001994 0.00316766 + layer.4.k_cache 0.00063949 0.08507740 + layer.4.v_cache 0.00006591 0.00693978 + layer.4.output 0.03773703 390.20337813 + ------------------------------------------------------------------------------------- + TOTAL 0.02493225 162.64409085 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 193708 +BPFP 1.8355 bits/point +EBPFP 3.6709 equivalent bits/point +MSE 162.644091 +---------------------- -------------------------------------------------------- +Time: 2.495s Load: 0.005s, Pack+Encode: 1.484s, Decode+Unpack: 1.007s +---------------------- -------------------------------------------------------- +💾 Converting with 162.6441 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,732B, BPFP=0.9017 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,396B, BPFP=3.6959 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,936B, BPFP=1.5122 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,456B, BPFP=3.7073 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,320B, BPFP=1.7759 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,372B, BPFP=3.5008 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,600B, BPFP=1.6387 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,556B, BPFP=3.5358 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,176B, BPFP=2.7012 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,904B, BPFP=3.4116 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,372B, BPFP=0.9084 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.005s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10693085 26.61477587 + layer.0.v_cache 0.00001341 0.00224523 + layer.1.k_cache 0.01172146 2.10642038 + layer.1.v_cache 0.00000553 0.00105236 + layer.2.k_cache 0.00236719 0.65642124 + layer.2.v_cache 0.00001835 0.00303689 + layer.3.k_cache 0.02622769 3.12775589 + layer.3.v_cache 0.00001864 0.00306678 + layer.4.k_cache 0.00061713 0.07802392 + layer.4.v_cache 0.00004943 0.00688973 + layer.4.output 0.18273007 578.21668118 + ------------------------------------------------------------------------------------- + TOTAL 0.08394589 240.00685039 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 171820 +BPFP 1.9259 bits/point +EBPFP 3.8518 equivalent bits/point +MSE 240.006850 +---------------------- -------------------------------------------------------- +Time: 2.491s Load: 0.004s, Pack+Encode: 1.481s, Decode+Unpack: 1.005s +---------------------- -------------------------------------------------------- +💾 Converting with 240.0069 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,408B, BPFP=0.8366 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,860B, BPFP=3.5365 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,108B, BPFP=1.5637 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,936B, BPFP=3.5483 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,688B, BPFP=1.9629 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,960B, BPFP=3.3973 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,172B, BPFP=1.5736 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,148B, BPFP=3.4264 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,928B, BPFP=2.7735 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,400B, BPFP=3.3106 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,060B, BPFP=0.8190 +⌛️ [2/4] FRONTEND: Frontend time: 1.483s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.006s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13156438 28.54706838 + layer.0.v_cache 0.00001398 0.00222377 + layer.1.k_cache 0.04507210 2.57198039 + layer.1.v_cache 0.00000603 0.00103835 + layer.2.k_cache 0.00848899 0.63542553 + layer.2.v_cache 0.00001983 0.00311290 + layer.3.k_cache 0.06093804 3.79531196 + layer.3.v_cache 0.00001891 0.00331310 + layer.4.k_cache 0.00062755 0.08400611 + layer.4.v_cache 0.00004781 0.00660089 + layer.4.output 11.29576791 481.17896924 + ------------------------------------------------------------------------------------- + TOTAL 4.66571606 200.22958036 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 204668 +BPFP 1.8625 bits/point +EBPFP 3.7250 equivalent bits/point +MSE 200.229580 +---------------------- -------------------------------------------------------- +Time: 2.494s Load: 0.005s, Pack+Encode: 1.483s, Decode+Unpack: 1.006s +---------------------- -------------------------------------------------------- +💾 Converting with 200.2296 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,776B, BPFP=0.8991 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,252B, BPFP=3.8125 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,096B, BPFP=1.5241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,328B, BPFP=3.6386 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,576B, BPFP=1.8027 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,016B, BPFP=3.5798 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,812B, BPFP=1.6589 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,096B, BPFP=3.5949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,856B, BPFP=2.7967 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,940B, BPFP=3.5655 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,836B, BPFP=0.9100 +⌛️ [2/4] FRONTEND: Frontend time: 1.481s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.012s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12340720 25.48197772 + layer.0.v_cache 0.00001379 0.00226987 + layer.1.k_cache 0.01365945 1.91618457 + layer.1.v_cache 0.00000586 0.00097530 + layer.2.k_cache 0.00515996 0.74074858 + layer.2.v_cache 0.00002118 0.00305910 + layer.3.k_cache 0.04682968 3.09264365 + layer.3.v_cache 0.00001899 0.00324484 + layer.4.k_cache 0.00062596 0.08065799 + layer.4.v_cache 0.00004975 0.00671624 + layer.4.output 0.18155291 576.22369836 + ------------------------------------------------------------------------------------- + TOTAL 0.08592130 239.11143332 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 176584 +BPFP 1.9554 bits/point +EBPFP 3.9109 equivalent bits/point +MSE 239.111433 +---------------------- -------------------------------------------------------- +Time: 2.497s Load: 0.005s, Pack+Encode: 1.481s, Decode+Unpack: 1.012s +---------------------- -------------------------------------------------------- +💾 Converting with 239.1114 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,852B, BPFP=0.9025 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,208B, BPFP=3.7589 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,172B, BPFP=1.5201 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,380B, BPFP=3.6049 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,356B, BPFP=1.7403 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,232B, BPFP=3.3914 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,796B, BPFP=1.6362 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,544B, BPFP=3.6354 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,380B, BPFP=2.6749 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,700B, BPFP=3.4784 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,928B, BPFP=0.9281 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.014s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10400932 24.98915899 + layer.0.v_cache 0.00001381 0.00224207 + layer.1.k_cache 0.01300498 2.00693149 + layer.1.v_cache 0.00000559 0.00095343 + layer.2.k_cache 0.00535372 0.71120135 + layer.2.v_cache 0.00001862 0.00286812 + layer.3.k_cache 0.04415199 2.98048710 + layer.3.v_cache 0.00001841 0.00296048 + layer.4.k_cache 0.00060249 0.07318679 + layer.4.v_cache 0.00005274 0.00641713 + layer.4.output 0.16694923 582.77465986 + ------------------------------------------------------------------------------------- + TOTAL 0.07858096 241.77641329 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 176548 +BPFP 1.9318 bits/point +EBPFP 3.8635 equivalent bits/point +MSE 241.776413 +---------------------- -------------------------------------------------------- +Time: 2.509s Load: 0.004s, Pack+Encode: 1.491s, Decode+Unpack: 1.014s +---------------------- -------------------------------------------------------- +💾 Converting with 241.7764 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,656B, BPFP=0.9094 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,968B, BPFP=3.7047 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,772B, BPFP=1.5180 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,692B, BPFP=3.6508 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,224B, BPFP=1.8016 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,348B, BPFP=3.5836 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,556B, BPFP=1.6711 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,180B, BPFP=3.7461 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,956B, BPFP=2.7258 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,372B, BPFP=3.5883 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,036B, BPFP=0.8939 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09321334 26.02268372 + layer.0.v_cache 0.00001862 0.00241649 + layer.1.k_cache 0.03257937 2.11552696 + layer.1.v_cache 0.00000532 0.00097070 + layer.2.k_cache 0.01442847 0.78774109 + layer.2.v_cache 0.00001865 0.00301442 + layer.3.k_cache 0.04385926 3.14982948 + layer.3.v_cache 0.00001828 0.00340763 + layer.4.k_cache 0.00061064 0.07860177 + layer.4.v_cache 0.00005353 0.00688456 + layer.4.output 0.18500509 579.49514509 + ------------------------------------------------------------------------------------- + TOTAL 0.08704948 240.50806426 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 169760 +BPFP 1.9504 bits/point +EBPFP 3.9007 equivalent bits/point +MSE 240.508064 +---------------------- -------------------------------------------------------- +Time: 2.511s Load: 0.004s, Pack+Encode: 1.495s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 240.5081 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,532B, BPFP=0.8392 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,516B, BPFP=3.5674 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,836B, BPFP=1.4921 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,832B, BPFP=3.4636 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,844B, BPFP=1.7967 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,032B, BPFP=3.3422 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,656B, BPFP=1.6165 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,352B, BPFP=3.3908 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,360B, BPFP=2.7852 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,888B, BPFP=3.3204 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,816B, BPFP=0.8412 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.011s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11527186 27.66055446 + layer.0.v_cache 0.00001510 0.00225909 + layer.1.k_cache 0.02920971 1.81716000 + layer.1.v_cache 0.00000551 0.00096843 + layer.2.k_cache 0.00857845 0.51773256 + layer.2.v_cache 0.00001837 0.00291127 + layer.3.k_cache 0.03550095 3.09390822 + layer.3.v_cache 0.00001950 0.00315227 + layer.4.k_cache 0.00063242 0.08374122 + layer.4.v_cache 0.00005150 0.00654339 + layer.4.output 11.08123415 445.16387829 + ------------------------------------------------------------------------------------- + TOTAL 4.57399661 185.25506347 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 207664 +BPFP 1.8531 bits/point +EBPFP 3.7062 equivalent bits/point +MSE 185.255063 +---------------------- -------------------------------------------------------- +Time: 2.511s Load: 0.005s, Pack+Encode: 1.495s, Decode+Unpack: 1.011s +---------------------- -------------------------------------------------------- +💾 Converting with 185.2551 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,016B, BPFP=0.8519 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,180B, BPFP=3.5971 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,644B, BPFP=1.4681 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,692B, BPFP=3.5143 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,072B, BPFP=1.7106 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,272B, BPFP=3.4429 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,300B, BPFP=1.5795 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,668B, BPFP=3.5102 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,284B, BPFP=2.7656 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,204B, BPFP=3.4314 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,792B, BPFP=0.8199 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.022s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09673090 26.82723336 + layer.0.v_cache 0.00001344 0.00221138 + layer.1.k_cache 0.04918879 1.69935409 + layer.1.v_cache 0.00000572 0.00101378 + layer.2.k_cache 0.00880511 0.66208574 + layer.2.v_cache 0.00001944 0.00293958 + layer.3.k_cache 0.06586690 2.85554239 + layer.3.v_cache 0.00001823 0.00311853 + layer.4.k_cache 0.00062484 0.08382955 + layer.4.v_cache 0.00006325 0.00656864 + layer.4.output 0.15294319 395.06558133 + ------------------------------------------------------------------------------------- + TOTAL 0.07599641 164.56488037 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 186124 +BPFP 1.8595 bits/point +EBPFP 3.7189 equivalent bits/point +MSE 164.564880 +---------------------- -------------------------------------------------------- +Time: 2.519s Load: 0.005s, Pack+Encode: 1.492s, Decode+Unpack: 1.022s +---------------------- -------------------------------------------------------- +💾 Converting with 164.5649 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,004B, BPFP=0.8499 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,204B, BPFP=3.6012 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,640B, BPFP=1.4674 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,688B, BPFP=3.5136 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,780B, BPFP=1.8308 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,228B, BPFP=3.4355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,360B, BPFP=1.5897 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,596B, BPFP=3.4980 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,308B, BPFP=2.7697 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,144B, BPFP=3.4212 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,840B, BPFP=0.8453 +⌛️ [2/4] FRONTEND: Frontend time: 1.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.022s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12676482 27.85567574 + layer.0.v_cache 0.00001339 0.00226503 + layer.1.k_cache 0.03133728 1.90418940 + layer.1.v_cache 0.00000590 0.00098130 + layer.2.k_cache 0.00372887 0.60257219 + layer.2.v_cache 0.00001820 0.00298078 + layer.3.k_cache 0.02418065 3.40796496 + layer.3.v_cache 0.00002034 0.00356582 + layer.4.k_cache 0.00062996 0.08032658 + layer.4.v_cache 0.00006053 0.00695686 + layer.4.output 0.15389095 520.73476320 + ------------------------------------------------------------------------------------- + TOTAL 0.07435274 216.41240124 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 187792 +BPFP 1.8761 bits/point +EBPFP 3.7522 equivalent bits/point +MSE 216.412401 +---------------------- -------------------------------------------------------- +Time: 2.517s Load: 0.005s, Pack+Encode: 1.489s, Decode+Unpack: 1.022s +---------------------- -------------------------------------------------------- +💾 Converting with 216.4124 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,940B, BPFP=0.8576 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,124B, BPFP=3.6674 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,444B, BPFP=1.4660 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,388B, BPFP=3.5396 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,768B, BPFP=1.6958 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,720B, BPFP=3.4236 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,024B, BPFP=1.5667 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,224B, BPFP=3.5111 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,192B, BPFP=2.6375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,476B, BPFP=3.3813 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,956B, BPFP=0.8422 +⌛️ [2/4] FRONTEND: Frontend time: 1.491s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.020s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12863920 27.42141385 + layer.0.v_cache 0.00001425 0.00224023 + layer.1.k_cache 0.01389826 1.76525879 + layer.1.v_cache 0.00000571 0.00097312 + layer.2.k_cache 0.00659207 0.64608366 + layer.2.v_cache 0.00001883 0.00291282 + layer.3.k_cache 0.05540012 2.79796024 + layer.3.v_cache 0.00001922 0.00313262 + layer.4.k_cache 0.00063279 0.07764477 + layer.4.v_cache 0.00005416 0.00655849 + layer.4.output 0.16196846 544.95615079 + ------------------------------------------------------------------------------------- + TOTAL 0.07876787 226.31866083 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 182256 +BPFP 1.8613 bits/point +EBPFP 3.7225 equivalent bits/point +MSE 226.318661 +---------------------- -------------------------------------------------------- +Time: 2.516s Load: 0.005s, Pack+Encode: 1.491s, Decode+Unpack: 1.020s +---------------------- -------------------------------------------------------- +💾 Converting with 226.3187 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,628B, BPFP=0.8142 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,552B, BPFP=3.4074 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,008B, BPFP=1.5926 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 23,036B, BPFP=3.3328 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,072B, BPFP=1.8912 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 22,436B, BPFP=3.2459 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,948B, BPFP=1.7286 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,776B, BPFP=3.2951 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,844B, BPFP=2.7263 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,872B, BPFP=3.1644 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,224B, BPFP=0.8313 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.022s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13160179 26.67813675 + layer.0.v_cache 0.00001577 0.00228757 + layer.1.k_cache 0.05775581 2.36541593 + layer.1.v_cache 0.00000549 0.00095396 + layer.2.k_cache 0.01027601 0.62602601 + layer.2.v_cache 0.00002197 0.00295155 + layer.3.k_cache 0.07329086 3.71501725 + layer.3.v_cache 0.00001983 0.00332880 + layer.4.k_cache 0.00064477 0.08134546 + layer.4.v_cache 0.00005260 0.00651439 + layer.4.output 10.56448284 433.37562004 + ------------------------------------------------------------------------------------- + TOTAL 4.36618028 180.41831282 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 214396 +BPFP 1.8246 bits/point +EBPFP 3.6492 equivalent bits/point +MSE 180.418313 +---------------------- -------------------------------------------------------- +Time: 2.518s Load: 0.006s, Pack+Encode: 1.490s, Decode+Unpack: 1.022s +---------------------- -------------------------------------------------------- +💾 Converting with 180.4183 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,784B, BPFP=0.9006 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,752B, BPFP=3.7184 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,028B, BPFP=1.5113 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,396B, BPFP=3.6514 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,352B, BPFP=1.7605 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,604B, BPFP=3.5023 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,604B, BPFP=1.6197 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,096B, BPFP=3.5949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,500B, BPFP=2.7297 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,624B, BPFP=3.5060 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,336B, BPFP=0.8696 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.031s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10139442 27.43487034 + layer.0.v_cache 0.00001346 0.00230685 + layer.1.k_cache 0.03266963 1.78718732 + layer.1.v_cache 0.00000526 0.00099048 + layer.2.k_cache 0.01283502 0.70291257 + layer.2.v_cache 0.00001884 0.00302278 + layer.3.k_cache 0.02666297 3.05397493 + layer.3.v_cache 0.00001780 0.00313778 + layer.4.k_cache 0.00061512 0.07643147 + layer.4.v_cache 0.00004947 0.00655872 + layer.4.output 0.18393325 573.07820568 + ------------------------------------------------------------------------------------- + TOTAL 0.08598910 237.91875488 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 173076 +BPFP 1.9166 bits/point +EBPFP 3.8332 equivalent bits/point +MSE 237.918755 +---------------------- -------------------------------------------------------- +Time: 2.533s Load: 0.005s, Pack+Encode: 1.498s, Decode+Unpack: 1.031s +---------------------- -------------------------------------------------------- +💾 Converting with 237.9188 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,436B, BPFP=0.8410 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,124B, BPFP=3.5774 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,052B, BPFP=1.5551 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,812B, BPFP=3.5291 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,432B, BPFP=1.9233 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,936B, BPFP=3.3936 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,476B, BPFP=1.6207 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,208B, BPFP=3.4356 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 18,052B, BPFP=2.7927 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,396B, BPFP=3.3100 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,304B, BPFP=0.8465 +⌛️ [2/4] FRONTEND: Frontend time: 1.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.027s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11149573 27.94323368 + layer.0.v_cache 0.00001357 0.00222571 + layer.1.k_cache 0.08092193 2.27333930 + layer.1.v_cache 0.00000571 0.00093463 + layer.2.k_cache 0.01109533 0.67986366 + layer.2.v_cache 0.00001920 0.00289780 + layer.3.k_cache 0.03617269 3.97068741 + layer.3.v_cache 0.00001858 0.00295832 + layer.4.k_cache 0.00062559 0.08198206 + layer.4.v_cache 0.00005684 0.00644832 + layer.4.output 11.30219499 489.37716584 + ------------------------------------------------------------------------------------- + TOTAL 4.66798765 203.56498422 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 206228 +BPFP 1.8767 bits/point +EBPFP 3.7534 equivalent bits/point +MSE 203.564984 +---------------------- -------------------------------------------------------- +Time: 2.531s Load: 0.006s, Pack+Encode: 1.498s, Decode+Unpack: 1.027s +---------------------- -------------------------------------------------------- +💾 Converting with 203.5650 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,616B, BPFP=0.9247 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,472B, BPFP=3.9006 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,896B, BPFP=1.5817 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,028B, BPFP=3.8117 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,060B, BPFP=1.8149 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,504B, BPFP=3.7067 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,336B, BPFP=1.6699 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,968B, BPFP=3.7997 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,048B, BPFP=2.8141 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,376B, BPFP=3.6811 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,496B, BPFP=0.9013 +⌛️ [2/4] FRONTEND: Frontend time: 1.492s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.022s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10167113 26.60821690 + layer.0.v_cache 0.00001386 0.00240643 + layer.1.k_cache 0.03545215 1.96077963 + layer.1.v_cache 0.00000569 0.00101437 + layer.2.k_cache 0.01047806 0.66485948 + layer.2.v_cache 0.00001755 0.00300386 + layer.3.k_cache 0.06687863 2.85124813 + layer.3.v_cache 0.00001960 0.00336253 + layer.4.k_cache 0.00059181 0.07761541 + layer.4.v_cache 0.00004882 0.00692195 + layer.4.output 0.19633365 621.27392399 + ------------------------------------------------------------------------------------- + TOTAL 0.09350076 257.71158216 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 169800 +BPFP 2.0008 bits/point +EBPFP 4.0017 equivalent bits/point +MSE 257.711582 +---------------------- -------------------------------------------------------- +Time: 2.518s Load: 0.005s, Pack+Encode: 1.492s, Decode+Unpack: 1.022s +---------------------- -------------------------------------------------------- +💾 Converting with 257.7116 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,844B, BPFP=0.8700 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,428B, BPFP=3.6688 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,308B, BPFP=1.4921 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,904B, BPFP=3.5747 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,712B, BPFP=1.7443 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,108B, BPFP=3.4318 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,992B, BPFP=1.6149 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,932B, BPFP=3.5797 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,988B, BPFP=2.6918 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,400B, BPFP=3.4842 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,568B, BPFP=0.8869 +⌛️ [2/4] FRONTEND: Frontend time: 1.493s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.021s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11445767 28.19917946 + layer.0.v_cache 0.00001583 0.00225584 + layer.1.k_cache 0.03092440 1.71034995 + layer.1.v_cache 0.00000551 0.00098372 + layer.2.k_cache 0.00639992 0.64606230 + layer.2.v_cache 0.00001897 0.00307545 + layer.3.k_cache 0.05613507 3.35587214 + layer.3.v_cache 0.00002026 0.00333009 + layer.4.k_cache 0.00063356 0.07787503 + layer.4.v_cache 0.00005123 0.00675952 + layer.4.output 0.16055210 555.66184319 + ------------------------------------------------------------------------------------- + TOTAL 0.07838395 230.80227328 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 180184 +BPFP 1.9036 bits/point +EBPFP 3.8071 equivalent bits/point +MSE 230.802273 +---------------------- -------------------------------------------------------- +Time: 2.518s Load: 0.004s, Pack+Encode: 1.493s, Decode+Unpack: 1.021s +---------------------- -------------------------------------------------------- +💾 Converting with 230.8023 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,808B, BPFP=0.8943 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,528B, BPFP=3.8185 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,200B, BPFP=1.5253 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,740B, BPFP=3.6719 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,576B, BPFP=1.7812 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,888B, BPFP=3.5134 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,828B, BPFP=1.6421 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,620B, BPFP=3.6496 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,916B, BPFP=2.7746 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,312B, BPFP=3.5923 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,072B, BPFP=0.8788 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.032s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13956541 25.69266474 + layer.0.v_cache 0.00001349 0.00230836 + layer.1.k_cache 0.03336124 1.88667733 + layer.1.v_cache 0.00000558 0.00109399 + layer.2.k_cache 0.00698305 0.64512916 + layer.2.v_cache 0.00001831 0.00297415 + layer.3.k_cache 0.07289400 2.66111501 + layer.3.v_cache 0.00001855 0.00329840 + layer.4.k_cache 0.00061006 0.07893712 + layer.4.v_cache 0.00005429 0.00705694 + layer.4.output 0.17891866 531.26424320 + ------------------------------------------------------------------------------------- + TOTAL 0.08858556 220.57829162 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 177488 +BPFP 1.9421 bits/point +EBPFP 3.8841 equivalent bits/point +MSE 220.578292 +---------------------- -------------------------------------------------------- +Time: 2.540s Load: 0.005s, Pack+Encode: 1.503s, Decode+Unpack: 1.032s +---------------------- -------------------------------------------------------- +💾 Converting with 220.5783 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,696B, BPFP=0.9059 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,480B, BPFP=3.7577 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,804B, BPFP=1.5054 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,328B, BPFP=3.7284 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,028B, BPFP=1.7415 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,420B, BPFP=3.5532 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,324B, BPFP=1.6057 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,396B, BPFP=3.5486 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,540B, BPFP=2.6119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,888B, BPFP=3.4506 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,504B, BPFP=0.8682 +⌛️ [2/4] FRONTEND: Frontend time: 1.490s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.048s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10837562 27.47575171 + layer.0.v_cache 0.00001644 0.00223947 + layer.1.k_cache 0.01241452 2.11769669 + layer.1.v_cache 0.00000532 0.00096424 + layer.2.k_cache 0.01724591 0.69038551 + layer.2.v_cache 0.00001958 0.00289269 + layer.3.k_cache 0.04641535 2.73009387 + layer.3.v_cache 0.00001847 0.00297615 + layer.4.k_cache 0.00063678 0.07490646 + layer.4.v_cache 0.00004974 0.00633203 + layer.4.output 0.18014439 572.34005732 + ------------------------------------------------------------------------------------- + TOTAL 0.08507109 237.61674353 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 168408 +BPFP 1.9109 bits/point +EBPFP 3.8219 equivalent bits/point +MSE 237.616744 +---------------------- -------------------------------------------------------- +Time: 2.542s Load: 0.004s, Pack+Encode: 1.490s, Decode+Unpack: 1.048s +---------------------- -------------------------------------------------------- +💾 Converting with 237.6167 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.9118 bits/point +Avg EBPFP 3.8237 equivalent bits/point +Avg MSE 226.875663 +Avg Time 2.917s +------------------------ ---------------------------- diff --git a/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..6b58a1c1930a2e00b3b037868e8104f5b7f04910 --- /dev/null +++ b/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 520 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean +Output output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean +---------------- ------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,292B, BPFP=0.7343 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,968B, BPFP=3.4224 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,928B, BPFP=1.4932 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,840B, BPFP=3.2705 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,796B, BPFP=1.7691 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,468B, BPFP=3.2440 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,540B, BPFP=1.6082 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,380B, BPFP=3.3091 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,076B, BPFP=2.5739 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,996B, BPFP=3.2103 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,216B, BPFP=0.8482 +⌛️ [2/4] FRONTEND: Frontend time: 2.125s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.287s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12261864 29.02932586 + layer.0.v_cache 0.00001806 0.00247311 + layer.1.k_cache 0.30198812 2.70846655 + layer.1.v_cache 0.00000607 0.00098449 + layer.2.k_cache 0.01966049 0.76546301 + layer.2.v_cache 0.00002021 0.00286793 + layer.3.k_cache 0.01413035 3.97508428 + layer.3.v_cache 0.00002054 0.00312150 + layer.4.k_cache 0.00067868 0.08208560 + layer.4.v_cache 0.00004987 0.00609860 + layer.4.output 1.39792533 222.95036285 + ------------------------------------------------------------------------------------- + TOTAL 0.60262755 93.95461829 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 428500 +BPFP 1.7984 bits/point +EBPFP 3.5967 equivalent bits/point +MSE 93.954618 +---------------------- -------------------------------------------------------- +Time: 3.420s Load: 0.008s, Pack+Encode: 2.125s, Decode+Unpack: 1.287s +---------------------- -------------------------------------------------------- +💾 Converting with 93.9546 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,124B, BPFP=0.7323 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,608B, BPFP=3.3715 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,108B, BPFP=1.4546 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,368B, BPFP=3.2818 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,324B, BPFP=1.8319 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,608B, BPFP=3.2269 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,860B, BPFP=1.6536 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,776B, BPFP=3.3113 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,676B, BPFP=2.5807 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,112B, BPFP=3.1910 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 79,536B, BPFP=0.8219 +⌛️ [2/4] FRONTEND: Frontend time: 1.731s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.230s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11787021 30.61494502 + layer.0.v_cache 0.00001738 0.00237605 + layer.1.k_cache 0.25265321 2.44658887 + layer.1.v_cache 0.00000586 0.00098153 + layer.2.k_cache 0.01010228 0.62718208 + layer.2.v_cache 0.00002017 0.00288870 + layer.3.k_cache 0.01159736 3.85238506 + layer.3.v_cache 0.00002197 0.00333598 + layer.4.k_cache 0.00067244 0.08276325 + layer.4.v_cache 0.00004845 0.00614849 + layer.4.output 1.41728832 218.28811177 + ------------------------------------------------------------------------------------- + TOTAL 0.60670750 92.09743397 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 420100 +BPFP 1.7876 bits/point +EBPFP 3.5752 equivalent bits/point +MSE 92.097434 +---------------------- -------------------------------------------------------- +Time: 2.968s Load: 0.007s, Pack+Encode: 1.731s, Decode+Unpack: 1.230s +---------------------- -------------------------------------------------------- +💾 Converting with 92.0974 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,240B, BPFP=0.7175 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,564B, BPFP=3.3327 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,184B, BPFP=1.4843 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,376B, BPFP=3.2494 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,076B, BPFP=1.7570 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,432B, BPFP=3.1833 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,936B, BPFP=1.6071 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,312B, BPFP=3.2450 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,444B, BPFP=2.5535 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,860B, BPFP=3.1432 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 76,588B, BPFP=0.7666 +⌛️ [2/4] FRONTEND: Frontend time: 1.722s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.231s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13311632 29.55298837 + layer.0.v_cache 0.00001526 0.00231320 + layer.1.k_cache 0.36474110 2.60419741 + layer.1.v_cache 0.00000573 0.00096369 + layer.2.k_cache 0.01296863 0.68275547 + layer.2.v_cache 0.00001984 0.00288408 + layer.3.k_cache 0.01490937 3.75510807 + layer.3.v_cache 0.00002017 0.00317797 + layer.4.k_cache 0.00067370 0.08119657 + layer.4.v_cache 0.00005228 0.00609481 + layer.4.output 1.37278975 231.46004164 + ------------------------------------------------------------------------------------- + TOTAL 0.59623828 97.46541007 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 423012 +BPFP 1.7435 bits/point +EBPFP 3.4870 equivalent bits/point +MSE 97.465410 +---------------------- -------------------------------------------------------- +Time: 2.961s Load: 0.008s, Pack+Encode: 1.722s, Decode+Unpack: 1.231s +---------------------- -------------------------------------------------------- +💾 Converting with 97.4654 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,992B, BPFP=0.7039 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,668B, BPFP=3.1165 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,116B, BPFP=1.4803 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,164B, BPFP=3.0202 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,680B, BPFP=1.7085 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,208B, BPFP=2.9590 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,924B, BPFP=1.5961 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,188B, BPFP=3.0218 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,136B, BPFP=2.4421 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,248B, BPFP=2.8975 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,684B, BPFP=0.8204 +⌛️ [2/4] FRONTEND: Frontend time: 1.718s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.236s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11263794 30.92880699 + layer.0.v_cache 0.00001639 0.00237002 + layer.1.k_cache 0.42101050 2.89984406 + layer.1.v_cache 0.00000577 0.00093649 + layer.2.k_cache 0.00499437 0.54829188 + layer.2.v_cache 0.00001973 0.00283605 + layer.3.k_cache 0.03029772 4.12463904 + layer.3.v_cache 0.00002073 0.00318318 + layer.4.k_cache 0.00068836 0.08148079 + layer.4.v_cache 0.00005112 0.00600074 + layer.4.output 1.25471843 71.61796509 + ------------------------------------------------------------------------------------- + TOTAL 0.55016304 31.76024382 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 448008 +BPFP 1.6876 bits/point +EBPFP 3.3752 equivalent bits/point +MSE 31.760244 +---------------------- -------------------------------------------------------- +Time: 2.964s Load: 0.010s, Pack+Encode: 1.718s, Decode+Unpack: 1.236s +---------------------- -------------------------------------------------------- +💾 Converting with 31.7602 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,260B, BPFP=0.7354 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,560B, BPFP=3.4088 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,300B, BPFP=1.4550 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,084B, BPFP=3.3030 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,968B, BPFP=1.7896 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,516B, BPFP=3.2623 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,856B, BPFP=1.6382 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,288B, BPFP=3.3177 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,052B, BPFP=2.5840 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,860B, BPFP=3.2153 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 80,816B, BPFP=0.8275 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.266s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10233516 28.52986355 + layer.0.v_cache 0.00001822 0.00243733 + layer.1.k_cache 0.31275415 2.67988348 + layer.1.v_cache 0.00000587 0.00093626 + layer.2.k_cache 0.01225604 0.64704503 + layer.2.v_cache 0.00002018 0.00291250 + layer.3.k_cache 0.02847941 3.57072071 + layer.3.v_cache 0.00002101 0.00317792 + layer.4.k_cache 0.00067306 0.08206971 + layer.4.v_cache 0.00005947 0.00624548 + layer.4.output 1.40430263 228.43274902 + ------------------------------------------------------------------------------------- + TOTAL 0.60510241 96.15026677 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 425560 +BPFP 1.7942 bits/point +EBPFP 3.5884 equivalent bits/point +MSE 96.150267 +---------------------- -------------------------------------------------------- +Time: 3.000s Load: 0.008s, Pack+Encode: 1.726s, Decode+Unpack: 1.266s +---------------------- -------------------------------------------------------- +💾 Converting with 96.1503 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,256B, BPFP=0.7045 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,844B, BPFP=3.2336 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,672B, BPFP=1.5770 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 59,284B, BPFP=3.1507 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,408B, BPFP=1.7224 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,652B, BPFP=3.0640 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,280B, BPFP=1.5561 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 59,160B, BPFP=3.1441 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,688B, BPFP=2.4813 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,600B, BPFP=3.0081 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 110,028B, BPFP=0.8354 +⌛️ [2/4] FRONTEND: Frontend time: 2.093s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16298727 31.70887808 + layer.0.v_cache 0.00001770 0.00231146 + layer.1.k_cache 0.61736183 3.01669457 + layer.1.v_cache 0.00000612 0.00093562 + layer.2.k_cache 0.02593859 0.58910479 + layer.2.v_cache 0.00002024 0.00275537 + layer.3.k_cache 0.01326758 4.08701371 + layer.3.v_cache 0.00002270 0.00309812 + layer.4.k_cache 0.00075427 0.08005406 + layer.4.v_cache 0.00005043 0.00579923 + layer.4.output 0.04538776 166.59224672 + ------------------------------------------------------------------------------------- + TOTAL 0.06694947 70.92013953 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 554872 +BPFP 1.7347 bits/point +EBPFP 3.4693 equivalent bits/point +MSE 70.920140 +---------------------- -------------------------------------------------------- +Time: 3.470s Load: 0.011s, Pack+Encode: 2.093s, Decode+Unpack: 1.366s +---------------------- -------------------------------------------------------- +💾 Converting with 70.9201 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 239, 128) +Output shape: (1, 239, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.output: torch.Size([1, 239, 3584]) -> torch.Size([1, 1, 239, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,068B, BPFP=0.7236 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,620B, BPFP=3.1786 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,512B, BPFP=1.4718 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,380B, BPFP=3.0975 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,616B, BPFP=1.7401 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,172B, BPFP=3.0186 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,680B, BPFP=1.6135 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,132B, BPFP=3.0813 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,824B, BPFP=2.4728 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,288B, BPFP=2.9608 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,888B, BPFP=0.8769 +⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.235s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11863919 30.19462196 + layer.0.v_cache 0.00001702 0.00245378 + layer.1.k_cache 0.43747612 2.86004409 + layer.1.v_cache 0.00000797 0.00100748 + layer.2.k_cache 0.02286501 0.53891125 + layer.2.v_cache 0.00002174 0.00284874 + layer.3.k_cache 0.01903599 4.08589358 + layer.3.v_cache 0.00002123 0.00312339 + layer.4.k_cache 0.00068597 0.08396975 + layer.4.v_cache 0.00006921 0.00610206 + layer.4.output 1.28101462 199.64323072 + ------------------------------------------------------------------------------------- + TOTAL 0.56270246 84.42832889 + (elements=2,080,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2080256 +Total Bytes 451180 +BPFP 1.7351 bits/point +EBPFP 3.4702 equivalent bits/point +MSE 84.428329 +---------------------- -------------------------------------------------------- +Time: 2.978s Load: 0.008s, Pack+Encode: 1.735s, Decode+Unpack: 1.235s +---------------------- -------------------------------------------------------- +💾 Converting with 84.4283 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,184B, BPFP=0.7239 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,012B, BPFP=3.4465 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,544B, BPFP=1.5176 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,472B, BPFP=3.2956 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,712B, BPFP=1.7652 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,856B, BPFP=3.3184 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,764B, BPFP=1.5901 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,836B, BPFP=3.3767 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,616B, BPFP=2.5913 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,268B, BPFP=3.2835 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,780B, BPFP=0.8638 +⌛️ [2/4] FRONTEND: Frontend time: 1.852s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15151563 30.06434173 + layer.0.v_cache 0.00001631 0.00236318 + layer.1.k_cache 0.46755480 2.63001556 + layer.1.v_cache 0.00000667 0.00097741 + layer.2.k_cache 0.02171374 0.69201399 + layer.2.v_cache 0.00002115 0.00295918 + layer.3.k_cache 0.01385514 3.29554499 + layer.3.v_cache 0.00002107 0.00314259 + layer.4.k_cache 0.00069081 0.08177864 + layer.4.v_cache 0.00005638 0.00624570 + layer.4.output 0.00504555 190.18743210 + ------------------------------------------------------------------------------------- + TOTAL 0.04063356 80.47596516 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 521044 +BPFP 1.8209 bits/point +EBPFP 3.6418 equivalent bits/point +MSE 80.475965 +---------------------- -------------------------------------------------------- +Time: 3.232s Load: 0.009s, Pack+Encode: 1.852s, Decode+Unpack: 1.371s +---------------------- -------------------------------------------------------- +💾 Converting with 80.4760 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,356B, BPFP=0.7192 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,328B, BPFP=3.3561 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,552B, BPFP=1.5661 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,256B, BPFP=3.2817 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,296B, BPFP=1.7567 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,156B, BPFP=3.2053 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,764B, BPFP=1.5808 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,704B, BPFP=3.2433 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,464B, BPFP=2.5322 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,904B, BPFP=3.1183 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,464B, BPFP=0.8479 +⌛️ [2/4] FRONTEND: Frontend time: 1.734s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.234s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14529029 29.06115451 + layer.0.v_cache 0.00001764 0.00237562 + layer.1.k_cache 0.31089128 3.07568224 + layer.1.v_cache 0.00000630 0.00098361 + layer.2.k_cache 0.01355181 0.63868123 + layer.2.v_cache 0.00002032 0.00287503 + layer.3.k_cache 0.02426839 3.74139730 + layer.3.v_cache 0.00002043 0.00312749 + layer.4.k_cache 0.00068925 0.08208245 + layer.4.v_cache 0.00005165 0.00587842 + layer.4.output 1.36066019 216.28486111 + ------------------------------------------------------------------------------------- + TOTAL 0.58937816 91.21225092 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 436244 +BPFP 1.7820 bits/point +EBPFP 3.5641 equivalent bits/point +MSE 91.212251 +---------------------- -------------------------------------------------------- +Time: 2.976s Load: 0.008s, Pack+Encode: 1.734s, Decode+Unpack: 1.234s +---------------------- -------------------------------------------------------- +💾 Converting with 91.2123 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 262, 128) +Output shape: (1, 262, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.output: torch.Size([1, 262, 3584]) -> torch.Size([1, 1, 262, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,076B, BPFP=0.7202 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,816B, BPFP=3.4480 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,028B, BPFP=1.6119 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,928B, BPFP=3.3354 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,300B, BPFP=1.7474 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,584B, BPFP=3.2552 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,676B, BPFP=1.5909 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,952B, BPFP=3.3965 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,960B, BPFP=2.5620 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,764B, BPFP=3.2660 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,236B, BPFP=0.8540 +⌛️ [2/4] FRONTEND: Frontend time: 1.850s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12917162 30.20616464 + layer.0.v_cache 0.00001830 0.00235072 + layer.1.k_cache 0.40634074 3.00494897 + layer.1.v_cache 0.00000633 0.00095590 + layer.2.k_cache 0.02363680 0.76783013 + layer.2.v_cache 0.00002248 0.00294563 + layer.3.k_cache 0.02343226 3.53538828 + layer.3.v_cache 0.00002042 0.00316475 + layer.4.k_cache 0.00071170 0.08248649 + layer.4.v_cache 0.00005093 0.00611974 + layer.4.output 0.00505940 200.20733370 + ------------------------------------------------------------------------------------- + TOTAL 0.03640161 84.65080536 + (elements=2,280,448) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2280448 +Total Bytes 518320 +BPFP 1.8183 bits/point +EBPFP 3.6366 equivalent bits/point +MSE 84.650805 +---------------------- -------------------------------------------------------- +Time: 3.220s Load: 0.009s, Pack+Encode: 1.850s, Decode+Unpack: 1.362s +---------------------- -------------------------------------------------------- +💾 Converting with 84.6508 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,300B, BPFP=0.7153 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,416B, BPFP=3.3622 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,284B, BPFP=1.5475 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,744B, BPFP=3.2461 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,632B, BPFP=1.7800 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,820B, BPFP=3.1819 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,312B, BPFP=1.6189 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,192B, BPFP=3.2772 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,720B, BPFP=2.5500 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,004B, BPFP=3.1253 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,024B, BPFP=0.8633 +⌛️ [2/4] FRONTEND: Frontend time: 1.736s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.228s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18074451 30.81110894 + layer.0.v_cache 0.00001764 0.00239927 + layer.1.k_cache 0.33421082 3.11891005 + layer.1.v_cache 0.00000655 0.00099308 + layer.2.k_cache 0.01487562 0.56473555 + layer.2.v_cache 0.00002082 0.00296352 + layer.3.k_cache 0.01319740 3.46576172 + layer.3.v_cache 0.00002096 0.00318524 + layer.4.k_cache 0.00067848 0.08180823 + layer.4.v_cache 0.00005239 0.00619963 + layer.4.output 1.36066546 220.04263889 + ------------------------------------------------------------------------------------- + TOTAL 0.59226373 92.84450220 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 438448 +BPFP 1.7910 bits/point +EBPFP 3.5821 equivalent bits/point +MSE 92.844502 +---------------------- -------------------------------------------------------- +Time: 2.974s Load: 0.010s, Pack+Encode: 1.736s, Decode+Unpack: 1.228s +---------------------- -------------------------------------------------------- +💾 Converting with 92.8445 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 232, 128) +Output shape: (1, 232, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.output: torch.Size([1, 232, 3584]) -> torch.Size([1, 1, 232, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,756B, BPFP=0.7244 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,376B, BPFP=3.2581 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,856B, BPFP=1.6740 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,996B, BPFP=3.1651 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,212B, BPFP=1.7654 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,316B, BPFP=3.1193 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,560B, BPFP=1.5867 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,148B, BPFP=3.1754 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,836B, BPFP=2.6156 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,724B, BPFP=3.0795 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,656B, BPFP=0.8626 +⌛️ [2/4] FRONTEND: Frontend time: 1.719s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.237s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11978169 28.20587369 + layer.0.v_cache 0.00001746 0.00235972 + layer.1.k_cache 0.32299256 2.96566010 + layer.1.v_cache 0.00000596 0.00092403 + layer.2.k_cache 0.02091339 0.53535800 + layer.2.v_cache 0.00002092 0.00286112 + layer.3.k_cache 0.01769928 3.75173371 + layer.3.v_cache 0.00002180 0.00307250 + layer.4.k_cache 0.00071208 0.08407826 + layer.4.v_cache 0.00005763 0.00632546 + layer.4.output 1.31963615 213.67268319 + ------------------------------------------------------------------------------------- + TOTAL 0.57174564 90.07453111 + (elements=2,019,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2019328 +Total Bytes 448436 +BPFP 1.7766 bits/point +EBPFP 3.5532 equivalent bits/point +MSE 90.074531 +---------------------- -------------------------------------------------------- +Time: 2.964s Load: 0.008s, Pack+Encode: 1.719s, Decode+Unpack: 1.237s +---------------------- -------------------------------------------------------- +💾 Converting with 90.0745 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,316B, BPFP=0.7327 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,436B, BPFP=3.3690 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,220B, BPFP=1.5071 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,972B, BPFP=3.2651 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,804B, BPFP=1.7616 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,304B, BPFP=3.2176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,968B, BPFP=1.6313 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,600B, BPFP=3.3097 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,280B, BPFP=2.5767 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,784B, BPFP=3.1807 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,696B, BPFP=0.8999 +⌛️ [2/4] FRONTEND: Frontend time: 1.725s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.234s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15639258 28.04955167 + layer.0.v_cache 0.00001700 0.00245374 + layer.1.k_cache 0.33150226 2.76856135 + layer.1.v_cache 0.00000613 0.00098277 + layer.2.k_cache 0.01694729 0.72098708 + layer.2.v_cache 0.00002208 0.00289808 + layer.3.k_cache 0.03858858 3.82118170 + layer.3.v_cache 0.00002090 0.00319319 + layer.4.k_cache 0.00067984 0.08182794 + layer.4.v_cache 0.00005217 0.00622096 + layer.4.output 1.39161012 231.27660308 + ------------------------------------------------------------------------------------- + TOTAL 0.60502939 97.31729883 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 434380 +BPFP 1.8148 bits/point +EBPFP 3.6295 equivalent bits/point +MSE 97.317299 +---------------------- -------------------------------------------------------- +Time: 2.968s Load: 0.008s, Pack+Encode: 1.725s, Decode+Unpack: 1.234s +---------------------- -------------------------------------------------------- +💾 Converting with 97.3173 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,368B, BPFP=0.7079 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,224B, BPFP=3.3897 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,248B, BPFP=1.6168 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,472B, BPFP=3.2894 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,632B, BPFP=1.7532 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,320B, BPFP=3.2234 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,592B, BPFP=1.5792 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,588B, BPFP=3.2960 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,544B, BPFP=2.5495 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,732B, BPFP=3.1898 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 105,048B, BPFP=0.8589 +⌛️ [2/4] FRONTEND: Frontend time: 1.845s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13605902 29.36075578 + layer.0.v_cache 0.00001757 0.00231522 + layer.1.k_cache 0.58683352 2.96757602 + layer.1.v_cache 0.00000623 0.00096624 + layer.2.k_cache 0.01500385 0.62952376 + layer.2.v_cache 0.00002229 0.00291953 + layer.3.k_cache 0.00978726 3.58707727 + layer.3.v_cache 0.00001998 0.00306300 + layer.4.k_cache 0.00067570 0.08158126 + layer.4.v_cache 0.00005472 0.00608181 + layer.4.output 0.00490606 68.30483467 + ------------------------------------------------------------------------------------- + TOTAL 0.04604839 30.28092368 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 534768 +BPFP 1.8004 bits/point +EBPFP 3.6008 equivalent bits/point +MSE 30.280924 +---------------------- -------------------------------------------------------- +Time: 3.219s Load: 0.010s, Pack+Encode: 1.845s, Decode+Unpack: 1.364s +---------------------- -------------------------------------------------------- +💾 Converting with 30.2809 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,084B, BPFP=0.6792 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,684B, BPFP=2.9831 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,108B, BPFP=1.4772 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,296B, BPFP=2.8980 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,300B, BPFP=1.6115 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,144B, BPFP=2.8275 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,388B, BPFP=1.4944 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,236B, BPFP=2.8944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,100B, BPFP=2.3346 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,220B, BPFP=2.7708 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 103,308B, BPFP=0.9043 +⌛️ [2/4] FRONTEND: Frontend time: 1.728s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.240s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12967140 37.72151501 + layer.0.v_cache 0.00001601 0.00224661 + layer.1.k_cache 0.49353925 3.01409338 + layer.1.v_cache 0.00000620 0.00090519 + layer.2.k_cache 0.03189284 0.59203264 + layer.2.v_cache 0.00002040 0.00269558 + layer.3.k_cache 0.04285583 4.16838283 + layer.3.v_cache 0.00002221 0.00301760 + layer.4.k_cache 0.00071132 0.07703302 + layer.4.v_cache 0.00005107 0.00563133 + layer.4.output 1.20063613 198.57387955 + ------------------------------------------------------------------------------------- + TOTAL 0.53548467 84.44733589 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 461868 +BPFP 1.6647 bits/point +EBPFP 3.3295 equivalent bits/point +MSE 84.447336 +---------------------- -------------------------------------------------------- +Time: 2.976s Load: 0.008s, Pack+Encode: 1.728s, Decode+Unpack: 1.240s +---------------------- -------------------------------------------------------- +💾 Converting with 84.4473 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,104B, BPFP=0.6968 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,644B, BPFP=3.0525 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,640B, BPFP=1.5462 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,372B, BPFP=2.9726 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,108B, BPFP=1.7011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,292B, BPFP=2.9049 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,000B, BPFP=1.5688 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,208B, BPFP=2.9623 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,556B, BPFP=2.4194 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,008B, BPFP=2.8243 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,428B, BPFP=0.7927 +⌛️ [2/4] FRONTEND: Frontend time: 1.734s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.235s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16529650 30.79883009 + layer.0.v_cache 0.00001613 0.00245565 + layer.1.k_cache 0.54577502 3.01800513 + layer.1.v_cache 0.00000662 0.00101665 + layer.2.k_cache 0.01453520 0.52264478 + layer.2.v_cache 0.00002096 0.00305417 + layer.3.k_cache 0.02409829 4.24634279 + layer.3.v_cache 0.00002071 0.00336128 + layer.4.k_cache 0.00068699 0.08537194 + layer.4.v_cache 0.00004971 0.00615864 + layer.4.output 1.22953407 205.03741753 + ------------------------------------------------------------------------------------- + TOTAL 0.55042615 86.70289199 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 449360 +BPFP 1.6587 bits/point +EBPFP 3.3174 equivalent bits/point +MSE 86.702892 +---------------------- -------------------------------------------------------- +Time: 2.977s Load: 0.008s, Pack+Encode: 1.734s, Decode+Unpack: 1.235s +---------------------- -------------------------------------------------------- +💾 Converting with 86.7029 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,792B, BPFP=0.7068 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 70,092B, BPFP=3.3492 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 33,068B, BPFP=1.5801 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 68,048B, BPFP=3.2515 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 36,644B, BPFP=1.7510 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 66,376B, BPFP=3.1716 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 32,536B, BPFP=1.5547 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 68,116B, BPFP=3.2548 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 52,608B, BPFP=2.5138 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 65,752B, BPFP=3.1418 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 128,160B, BPFP=0.8748 +⌛️ [2/4] FRONTEND: Frontend time: 2.202s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.481s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19043710 30.65384891 + layer.0.v_cache 0.00001603 0.00233734 + layer.1.k_cache 0.64205214 2.91462526 + layer.1.v_cache 0.00000630 0.00097719 + layer.2.k_cache 0.01513053 0.65605304 + layer.2.v_cache 0.00002191 0.00295871 + layer.3.k_cache 0.02775778 3.38667918 + layer.3.v_cache 0.00002125 0.00308521 + layer.4.k_cache 0.00071623 0.08004779 + layer.4.v_cache 0.00005146 0.00601377 + layer.4.output 0.04089398 142.40921254 + ------------------------------------------------------------------------------------- + TOTAL 0.06838051 60.85712436 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 636192 +BPFP 1.7882 bits/point +EBPFP 3.5764 equivalent bits/point +MSE 60.857124 +---------------------- -------------------------------------------------------- +Time: 3.694s Load: 0.011s, Pack+Encode: 2.202s, Decode+Unpack: 1.481s +---------------------- -------------------------------------------------------- +💾 Converting with 60.8571 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,912B, BPFP=0.7129 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,612B, BPFP=3.2913 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,424B, BPFP=1.5693 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,884B, BPFP=3.1959 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,796B, BPFP=1.7555 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,976B, BPFP=3.1458 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,332B, BPFP=1.5643 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,416B, BPFP=3.2253 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,416B, BPFP=2.5075 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,504B, BPFP=3.1197 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 106,700B, BPFP=0.8416 +⌛️ [2/4] FRONTEND: Frontend time: 1.843s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14392723 29.73920605 + layer.0.v_cache 0.00001682 0.00236140 + layer.1.k_cache 0.53349789 2.83862413 + layer.1.v_cache 0.00000638 0.00096740 + layer.2.k_cache 0.01523476 0.61350889 + layer.2.v_cache 0.00002109 0.00276088 + layer.3.k_cache 0.03867486 3.69949179 + layer.3.v_cache 0.00001994 0.00298960 + layer.4.k_cache 0.00070525 0.07963044 + layer.4.v_cache 0.00005078 0.00585734 + layer.4.output 0.00472849 174.10304139 + ------------------------------------------------------------------------------------- + TOTAL 0.04501497 73.86509928 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 542972 +BPFP 1.7634 bits/point +EBPFP 3.5269 equivalent bits/point +MSE 73.865099 +---------------------- -------------------------------------------------------- +Time: 3.210s Load: 0.011s, Pack+Encode: 1.843s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 73.8651 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,268B, BPFP=0.7131 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,624B, BPFP=3.3767 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,820B, BPFP=1.4458 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,804B, BPFP=3.2503 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,308B, BPFP=1.7575 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,964B, BPFP=3.1919 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,376B, BPFP=1.6233 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,964B, BPFP=3.2614 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,104B, BPFP=2.5767 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,768B, BPFP=3.1089 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,348B, BPFP=0.8368 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14673459 29.65016710 + layer.0.v_cache 0.00001655 0.00238105 + layer.1.k_cache 0.37581726 2.70551270 + layer.1.v_cache 0.00000603 0.00094337 + layer.2.k_cache 0.01062915 0.68779460 + layer.2.v_cache 0.00002006 0.00296760 + layer.3.k_cache 0.04027582 3.95260742 + layer.3.v_cache 0.00002066 0.00315218 + layer.4.k_cache 0.00066373 0.07987108 + layer.4.v_cache 0.00004861 0.00619529 + layer.4.output 1.36064577 222.06577381 + ------------------------------------------------------------------------------------- + TOTAL 0.59404429 93.62070642 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 434348 +BPFP 1.7743 bits/point +EBPFP 3.5486 equivalent bits/point +MSE 93.620706 +---------------------- -------------------------------------------------------- +Time: 2.960s Load: 0.008s, Pack+Encode: 1.730s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 93.6207 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,880B, BPFP=0.7265 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,032B, BPFP=3.2740 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,816B, BPFP=1.7238 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,540B, BPFP=3.1744 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,840B, BPFP=1.7922 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,436B, BPFP=3.1007 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,792B, BPFP=1.6554 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,476B, BPFP=3.1701 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,976B, BPFP=2.5358 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,716B, BPFP=3.0526 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,332B, BPFP=0.9189 +⌛️ [2/4] FRONTEND: Frontend time: 1.717s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12385076 30.20445212 + layer.0.v_cache 0.00002004 0.00252840 + layer.1.k_cache 0.38583452 3.03878941 + layer.1.v_cache 0.00000646 0.00101559 + layer.2.k_cache 0.02500911 0.58853123 + layer.2.v_cache 0.00002014 0.00284122 + layer.3.k_cache 0.03425702 3.92239093 + layer.3.v_cache 0.00002057 0.00318978 + layer.4.k_cache 0.00068928 0.08186750 + layer.4.v_cache 0.00004882 0.00601985 + layer.4.output 1.30836928 207.74301740 + ------------------------------------------------------------------------------------- + TOTAL 0.57225539 87.76780870 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 458836 +BPFP 1.8022 bits/point +EBPFP 3.6045 equivalent bits/point +MSE 87.767809 +---------------------- -------------------------------------------------------- +Time: 2.948s Load: 0.008s, Pack+Encode: 1.717s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 87.7678 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 247, 128) +Output shape: (1, 247, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.output: torch.Size([1, 247, 3584]) -> torch.Size([1, 1, 247, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,120B, BPFP=0.7034 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,796B, BPFP=3.0868 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,560B, BPFP=1.5536 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,316B, BPFP=2.9932 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,588B, BPFP=1.6819 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,072B, BPFP=2.9145 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,152B, BPFP=1.5278 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,936B, BPFP=2.9691 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,684B, BPFP=2.3839 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,024B, BPFP=2.8482 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,532B, BPFP=0.8181 +⌛️ [2/4] FRONTEND: Frontend time: 1.715s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13907220 30.35831146 + layer.0.v_cache 0.00001735 0.00238025 + layer.1.k_cache 0.47208408 3.03343145 + layer.1.v_cache 0.00000611 0.00095130 + layer.2.k_cache 0.03408500 0.60329162 + layer.2.v_cache 0.00001989 0.00271942 + layer.3.k_cache 0.02387486 3.70527556 + layer.3.v_cache 0.00002031 0.00299954 + layer.4.k_cache 0.00069623 0.08100930 + layer.4.v_cache 0.00004922 0.00562311 + layer.4.output 1.23951819 142.94011170 + ------------------------------------------------------------------------------------- + TOTAL 0.54979721 61.08098676 + (elements=2,149,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2149888 +Total Bytes 448780 +BPFP 1.6700 bits/point +EBPFP 3.3399 equivalent bits/point +MSE 61.080987 +---------------------- -------------------------------------------------------- +Time: 2.946s Load: 0.008s, Pack+Encode: 1.715s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 61.0810 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 427, 128) +Output shape: (1, 427, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.output: torch.Size([1, 427, 3584]) -> torch.Size([1, 1, 427, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 18,516B, BPFP=0.6775 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 84,620B, BPFP=3.0965 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 44,068B, BPFP=1.6126 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 82,292B, BPFP=3.0113 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 45,316B, BPFP=1.6582 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 79,548B, BPFP=2.9109 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 39,920B, BPFP=1.4608 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 81,452B, BPFP=2.9805 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 64,668B, BPFP=2.3664 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 78,224B, BPFP=2.8624 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 160,032B, BPFP=0.8366 +⌛️ [2/4] FRONTEND: Frontend time: 2.528s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.587s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15738957 33.09710279 + layer.0.v_cache 0.00001693 0.00218457 + layer.1.k_cache 1.00627419 3.12533819 + layer.1.v_cache 0.00000663 0.00090135 + layer.2.k_cache 0.03590691 0.61079310 + layer.2.v_cache 0.00002108 0.00258241 + layer.3.k_cache 0.02176077 3.98056109 + layer.3.v_cache 0.00002004 0.00277697 + layer.4.k_cache 0.00078997 0.08149486 + layer.4.v_cache 0.00005212 0.00556216 + layer.4.output 0.00621442 114.91209435 + ------------------------------------------------------------------------------------- + TOTAL 0.07445524 49.72317400 + (elements=3,716,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3716608 +Total Bytes 778656 +BPFP 1.6761 bits/point +EBPFP 3.3521 equivalent bits/point +MSE 49.723174 +---------------------- -------------------------------------------------------- +Time: 4.129s Load: 0.014s, Pack+Encode: 2.528s, Decode+Unpack: 1.587s +---------------------- -------------------------------------------------------- +💾 Converting with 49.7232 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,496B, BPFP=0.7205 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,468B, BPFP=3.3711 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,396B, BPFP=1.5796 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,680B, BPFP=3.2680 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,712B, BPFP=1.7131 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,220B, BPFP=3.1838 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,496B, BPFP=1.5277 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,424B, BPFP=3.3109 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,736B, BPFP=2.5217 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,016B, BPFP=3.1720 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,804B, BPFP=0.7891 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.347s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15504150 28.18985132 + layer.0.v_cache 0.00001776 0.00221417 + layer.1.k_cache 0.59434689 2.96842996 + layer.1.v_cache 0.00000612 0.00089347 + layer.2.k_cache 0.01687181 0.76789056 + layer.2.v_cache 0.00002031 0.00275414 + layer.3.k_cache 0.01346795 3.73175567 + layer.3.v_cache 0.00002023 0.00289888 + layer.4.k_cache 0.00071492 0.07773475 + layer.4.v_cache 0.00006004 0.00578503 + layer.4.output 0.00488892 170.36984383 + ------------------------------------------------------------------------------------- + TOTAL 0.04792882 72.25524205 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 518448 +BPFP 1.7584 bits/point +EBPFP 3.5167 equivalent bits/point +MSE 72.255242 +---------------------- -------------------------------------------------------- +Time: 3.198s Load: 0.010s, Pack+Encode: 1.841s, Decode+Unpack: 1.347s +---------------------- -------------------------------------------------------- +💾 Converting with 72.2552 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 361, 128) +Output shape: (1, 361, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.output: torch.Size([1, 361, 3584]) -> torch.Size([1, 1, 361, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,832B, BPFP=0.6852 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 71,872B, BPFP=3.1108 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 36,860B, BPFP=1.5954 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 69,904B, BPFP=3.0256 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 37,900B, BPFP=1.6404 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 68,192B, BPFP=2.9515 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 34,696B, BPFP=1.5017 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 69,912B, BPFP=3.0260 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 55,612B, BPFP=2.4070 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 67,360B, BPFP=2.9155 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 120,756B, BPFP=0.7467 +⌛️ [2/4] FRONTEND: Frontend time: 1.968s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.477s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21057716 31.06341434 + layer.0.v_cache 0.00001554 0.00208972 + layer.1.k_cache 0.82271198 3.13073629 + layer.1.v_cache 0.00000592 0.00085083 + layer.2.k_cache 0.01184860 0.61313306 + layer.2.v_cache 0.00001987 0.00268195 + layer.3.k_cache 0.01345186 3.99374094 + layer.3.v_cache 0.00002111 0.00286479 + layer.4.k_cache 0.00076976 0.07918782 + layer.4.v_cache 0.00005077 0.00559534 + layer.4.output 0.03703259 135.91944499 + ------------------------------------------------------------------------------------- + TOTAL 0.07757063 58.25473000 + (elements=3,142,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3142144 +Total Bytes 648896 +BPFP 1.6521 bits/point +EBPFP 3.3042 equivalent bits/point +MSE 58.254730 +---------------------- -------------------------------------------------------- +Time: 3.457s Load: 0.012s, Pack+Encode: 1.968s, Decode+Unpack: 1.477s +---------------------- -------------------------------------------------------- +💾 Converting with 58.2547 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,796B, BPFP=0.7090 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,268B, BPFP=3.2285 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,028B, BPFP=1.6638 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,364B, BPFP=3.1784 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,008B, BPFP=1.6627 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,188B, BPFP=3.0578 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,980B, BPFP=1.5503 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,148B, BPFP=3.1664 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,500B, BPFP=2.4656 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,208B, BPFP=3.0590 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 97,684B, BPFP=0.7732 +⌛️ [2/4] FRONTEND: Frontend time: 1.844s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21757852 28.74019628 + layer.0.v_cache 0.00001544 0.00202912 + layer.1.k_cache 0.59705023 2.80834420 + layer.1.v_cache 0.00000603 0.00085444 + layer.2.k_cache 0.00787334 0.63609942 + layer.2.v_cache 0.00001970 0.00270340 + layer.3.k_cache 0.01914395 3.78286332 + layer.3.v_cache 0.00001916 0.00281601 + layer.4.k_cache 0.00071861 0.07715538 + layer.4.v_cache 0.00005003 0.00568754 + layer.4.output 0.00471288 179.62319529 + ------------------------------------------------------------------------------------- + TOTAL 0.05149795 76.08359507 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 526172 +BPFP 1.7149 bits/point +EBPFP 3.4299 equivalent bits/point +MSE 76.083595 +---------------------- -------------------------------------------------------- +Time: 3.207s Load: 0.010s, Pack+Encode: 1.844s, Decode+Unpack: 1.353s +---------------------- -------------------------------------------------------- +💾 Converting with 76.0836 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 372, 128) +Output shape: (1, 372, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.output: torch.Size([1, 372, 3584]) -> torch.Size([1, 1, 372, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,120B, BPFP=0.6771 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 73,216B, BPFP=3.0753 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 37,388B, BPFP=1.5704 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 70,436B, BPFP=2.9585 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 38,760B, BPFP=1.6280 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 68,396B, BPFP=2.8728 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 35,476B, BPFP=1.4901 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 69,808B, BPFP=2.9321 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 55,820B, BPFP=2.3446 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 67,552B, BPFP=2.8374 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 144,856B, BPFP=0.8692 +⌛️ [2/4] FRONTEND: Frontend time: 1.991s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.480s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18192377 31.28802031 + layer.0.v_cache 0.00001786 0.00215486 + layer.1.k_cache 0.82695811 3.06849851 + layer.1.v_cache 0.00000687 0.00088772 + layer.2.k_cache 0.01216373 0.59832025 + layer.2.v_cache 0.00002073 0.00263041 + layer.3.k_cache 0.01743993 4.39668061 + layer.3.v_cache 0.00002103 0.00282338 + layer.4.k_cache 0.00069676 0.07816782 + layer.4.v_cache 0.00005544 0.00571427 + layer.4.output 0.03606073 121.63906490 + ------------------------------------------------------------------------------------- + TOTAL 0.07598408 52.40690308 + (elements=3,237,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3237888 +Total Bytes 677828 +BPFP 1.6747 bits/point +EBPFP 3.3495 equivalent bits/point +MSE 52.406903 +---------------------- -------------------------------------------------------- +Time: 3.484s Load: 0.013s, Pack+Encode: 1.991s, Decode+Unpack: 1.480s +---------------------- -------------------------------------------------------- +💾 Converting with 52.4069 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,392B, BPFP=0.6962 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 68,148B, BPFP=3.2966 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 35,764B, BPFP=1.7301 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 66,024B, BPFP=3.1939 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,976B, BPFP=1.6436 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 64,816B, BPFP=3.1354 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,068B, BPFP=1.5029 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 67,368B, BPFP=3.2589 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 51,304B, BPFP=2.4818 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 64,948B, BPFP=3.1418 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 116,872B, BPFP=0.8077 +⌛️ [2/4] FRONTEND: Frontend time: 1.960s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.479s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16470559 29.96376137 + layer.0.v_cache 0.00001558 0.00217626 + layer.1.k_cache 0.70577214 3.28256518 + layer.1.v_cache 0.00000663 0.00089660 + layer.2.k_cache 0.01487649 0.63655931 + layer.2.v_cache 0.00002061 0.00272241 + layer.3.k_cache 0.03012228 3.54186747 + layer.3.v_cache 0.00002030 0.00288707 + layer.4.k_cache 0.00074575 0.07821699 + layer.4.v_cache 0.00005065 0.00560347 + layer.4.output 0.04139163 137.57131800 + ------------------------------------------------------------------------------------- + TOTAL 0.07094573 58.85391071 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 614680 +BPFP 1.7491 bits/point +EBPFP 3.4982 equivalent bits/point +MSE 58.853911 +---------------------- -------------------------------------------------------- +Time: 3.450s Load: 0.011s, Pack+Encode: 1.960s, Decode+Unpack: 1.479s +---------------------- -------------------------------------------------------- +💾 Converting with 58.8539 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,216B, BPFP=0.7203 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,028B, BPFP=3.4215 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,492B, BPFP=1.4441 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,276B, BPFP=3.3182 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,528B, BPFP=1.7410 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,992B, BPFP=3.2425 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,420B, BPFP=1.5578 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,628B, BPFP=3.3389 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,864B, BPFP=2.5863 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,704B, BPFP=3.2255 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,032B, BPFP=0.8426 +⌛️ [2/4] FRONTEND: Frontend time: 1.860s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.352s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16585089 30.08653818 + layer.0.v_cache 0.00001549 0.00219746 + layer.1.k_cache 0.52552974 2.38119772 + layer.1.v_cache 0.00000676 0.00089674 + layer.2.k_cache 0.01473994 0.66486955 + layer.2.v_cache 0.00002188 0.00283685 + layer.3.k_cache 0.05129043 3.58531932 + layer.3.v_cache 0.00002012 0.00294772 + layer.4.k_cache 0.00075129 0.07994378 + layer.4.v_cache 0.00005717 0.00564410 + layer.4.output 0.00501330 189.23438342 + ------------------------------------------------------------------------------------- + TOTAL 0.04666922 80.08547502 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 517180 +BPFP 1.7938 bits/point +EBPFP 3.5875 equivalent bits/point +MSE 80.085475 +---------------------- -------------------------------------------------------- +Time: 3.224s Load: 0.011s, Pack+Encode: 1.860s, Decode+Unpack: 1.352s +---------------------- -------------------------------------------------------- +💾 Converting with 80.0855 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 368, 128) +Output shape: (1, 368, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.output: torch.Size([1, 368, 3584]) -> torch.Size([1, 1, 368, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,120B, BPFP=0.6844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 72,428B, BPFP=3.0752 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 39,064B, BPFP=1.6586 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 70,356B, BPFP=2.9873 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 38,456B, BPFP=1.6328 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 68,156B, BPFP=2.8939 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 35,060B, BPFP=1.4886 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 70,572B, BPFP=2.9964 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 55,600B, BPFP=2.3607 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 68,236B, BPFP=2.8972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 132,412B, BPFP=0.8032 +⌛️ [2/4] FRONTEND: Frontend time: 1.972s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.480s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14147143 32.72378673 + layer.0.v_cache 0.00001611 0.00212598 + layer.1.k_cache 0.89062865 3.32681042 + layer.1.v_cache 0.00000637 0.00086118 + layer.2.k_cache 0.02201829 0.65940044 + layer.2.v_cache 0.00002088 0.00265412 + layer.3.k_cache 0.00965744 4.10871555 + layer.3.v_cache 0.00002101 0.00293870 + layer.4.k_cache 0.00077426 0.07840845 + layer.4.v_cache 0.00005366 0.00589812 + layer.4.output 0.03638816 125.72817595 + ------------------------------------------------------------------------------------- + TOTAL 0.07761090 54.17699008 + (elements=3,203,072) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3203072 +Total Bytes 666460 +BPFP 1.6646 bits/point +EBPFP 3.3291 equivalent bits/point +MSE 54.176990 +---------------------- -------------------------------------------------------- +Time: 3.465s Load: 0.013s, Pack+Encode: 1.972s, Decode+Unpack: 1.480s +---------------------- -------------------------------------------------------- +💾 Converting with 54.1770 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 453, 128) +Output shape: (1, 453, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.output: torch.Size([1, 453, 3584]) -> torch.Size([1, 1, 453, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 19,656B, BPFP=0.6780 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 93,656B, BPFP=3.2304 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 49,820B, BPFP=1.7184 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 90,152B, BPFP=3.1095 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 47,104B, BPFP=1.6247 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 88,076B, BPFP=3.0379 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 43,196B, BPFP=1.4899 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 91,436B, BPFP=3.1538 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 70,304B, BPFP=2.4249 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 88,748B, BPFP=3.0611 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 160,076B, BPFP=0.7888 +⌛️ [2/4] FRONTEND: Frontend time: 2.455s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.746s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13665763 31.96859047 + layer.0.v_cache 0.00001897 0.00215838 + layer.1.k_cache 1.07282356 3.49122199 + layer.1.v_cache 0.00000630 0.00086104 + layer.2.k_cache 0.02241544 0.67749495 + layer.2.v_cache 0.00001979 0.00259960 + layer.3.k_cache 0.01806638 3.74879412 + layer.3.v_cache 0.00002079 0.00292537 + layer.4.k_cache 0.00076567 0.07858346 + layer.4.v_cache 0.00005104 0.00567021 + layer.4.output 0.00584071 110.72515571 + ------------------------------------------------------------------------------------- + TOTAL 0.07598415 47.94441115 + (elements=3,942,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3942912 +Total Bytes 842224 +BPFP 1.7088 bits/point +EBPFP 3.4177 equivalent bits/point +MSE 47.944411 +---------------------- -------------------------------------------------------- +Time: 4.217s Load: 0.015s, Pack+Encode: 2.455s, Decode+Unpack: 1.746s +---------------------- -------------------------------------------------------- +💾 Converting with 47.9444 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 402, 128) +Output shape: (1, 402, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.output: torch.Size([1, 402, 3584]) -> torch.Size([1, 1, 402, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 17,732B, BPFP=0.6892 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 82,864B, BPFP=3.2208 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 42,816B, BPFP=1.6642 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 80,016B, BPFP=3.1101 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 42,384B, BPFP=1.6474 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 77,376B, BPFP=3.0075 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 37,604B, BPFP=1.4616 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 80,612B, BPFP=3.1332 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 61,516B, BPFP=2.3910 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 77,292B, BPFP=3.0042 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 132,104B, BPFP=0.7335 +⌛️ [2/4] FRONTEND: Frontend time: 2.126s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.580s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18062694 30.06832779 + layer.0.v_cache 0.00001752 0.00215443 + layer.1.k_cache 0.89169403 3.25976441 + layer.1.v_cache 0.00000593 0.00081824 + layer.2.k_cache 0.03854401 0.67291875 + layer.2.v_cache 0.00001967 0.00261649 + layer.3.k_cache 0.02336847 3.75412215 + layer.3.v_cache 0.00001924 0.00285461 + layer.4.k_cache 0.00085529 0.07817018 + layer.4.v_cache 0.00004959 0.00554863 + layer.4.output 0.00648942 59.28776319 + ------------------------------------------------------------------------------------- + TOTAL 0.06944863 26.63891988 + (elements=3,499,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3499008 +Total Bytes 732316 +BPFP 1.6743 bits/point +EBPFP 3.3487 equivalent bits/point +MSE 26.638920 +---------------------- -------------------------------------------------------- +Time: 3.719s Load: 0.013s, Pack+Encode: 2.126s, Decode+Unpack: 1.580s +---------------------- -------------------------------------------------------- +💾 Converting with 26.6389 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,868B, BPFP=0.7207 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,176B, BPFP=3.3141 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,368B, BPFP=1.5887 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,476B, BPFP=3.2189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,600B, BPFP=1.7697 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,280B, BPFP=3.1519 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,356B, BPFP=1.5880 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,400B, BPFP=3.2146 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,272B, BPFP=2.5354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,572B, BPFP=3.1122 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 107,888B, BPFP=0.8632 +⌛️ [2/4] FRONTEND: Frontend time: 1.843s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.360s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13958113 31.13584019 + layer.0.v_cache 0.00001688 0.00241370 + layer.1.k_cache 0.56864935 2.82360752 + layer.1.v_cache 0.00000621 0.00100718 + layer.2.k_cache 0.01018704 0.58962482 + layer.2.v_cache 0.00002046 0.00294097 + layer.3.k_cache 0.04374025 3.72891465 + layer.3.v_cache 0.00002178 0.00322216 + layer.4.k_cache 0.00068741 0.08200103 + layer.4.v_cache 0.00005272 0.00620990 + layer.4.output 0.00479460 222.43922811 + ------------------------------------------------------------------------------------- + TOTAL 0.04685444 93.85002229 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 540256 +BPFP 1.7798 bits/point +EBPFP 3.5596 equivalent bits/point +MSE 93.850022 +---------------------- -------------------------------------------------------- +Time: 3.214s Load: 0.011s, Pack+Encode: 1.843s, Decode+Unpack: 1.360s +---------------------- -------------------------------------------------------- +💾 Converting with 93.8500 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 381, 128) +Output shape: (1, 381, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.output: torch.Size([1, 381, 3584]) -> torch.Size([1, 1, 381, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,332B, BPFP=0.6698 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 72,868B, BPFP=2.9884 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 38,636B, BPFP=1.5845 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 70,832B, BPFP=2.9049 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 39,432B, BPFP=1.6171 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 69,296B, BPFP=2.8419 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 36,488B, BPFP=1.4964 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 70,628B, BPFP=2.8965 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 56,908B, BPFP=2.3338 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 67,652B, BPFP=2.7744 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 145,892B, BPFP=0.8547 +⌛️ [2/4] FRONTEND: Frontend time: 1.981s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.487s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15713709 34.13507064 + layer.0.v_cache 0.00001719 0.00224560 + layer.1.k_cache 0.86767426 3.16074353 + layer.1.v_cache 0.00000622 0.00093982 + layer.2.k_cache 0.01635739 0.57799344 + layer.2.v_cache 0.00002098 0.00283532 + layer.3.k_cache 0.01709203 4.21039635 + layer.3.v_cache 0.00002073 0.00297108 + layer.4.k_cache 0.00073405 0.08189578 + layer.4.v_cache 0.00005083 0.00571955 + layer.4.output 0.03519400 128.23299822 + ------------------------------------------------------------------------------------- + TOTAL 0.07679228 55.28304698 + (elements=3,316,224) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3316224 +Total Bytes 684964 +BPFP 1.6524 bits/point +EBPFP 3.3048 equivalent bits/point +MSE 55.283047 +---------------------- -------------------------------------------------------- +Time: 3.482s Load: 0.015s, Pack+Encode: 1.981s, Decode+Unpack: 1.487s +---------------------- -------------------------------------------------------- +💾 Converting with 55.2830 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,816B, BPFP=0.7222 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,532B, BPFP=3.2407 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,860B, BPFP=1.6600 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,288B, BPFP=3.1576 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,456B, BPFP=1.7666 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,256B, BPFP=3.0887 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,412B, BPFP=1.6301 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,104B, BPFP=3.1453 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,192B, BPFP=2.5502 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,420B, BPFP=3.0329 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,828B, BPFP=0.8855 +⌛️ [2/4] FRONTEND: Frontend time: 1.734s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.228s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422791 30.97356604 + layer.0.v_cache 0.00001930 0.00244966 + layer.1.k_cache 0.38928963 2.90011492 + layer.1.v_cache 0.00000631 0.00099159 + layer.2.k_cache 0.01623082 0.62283945 + layer.2.v_cache 0.00002095 0.00294007 + layer.3.k_cache 0.01313608 3.83952892 + layer.3.v_cache 0.00002100 0.00322284 + layer.4.k_cache 0.00066994 0.08261744 + layer.4.v_cache 0.00005661 0.00613368 + layer.4.output 1.30836332 160.30818834 + ------------------------------------------------------------------------------------- + TOTAL 0.57013069 68.27010135 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 452164 +BPFP 1.7760 bits/point +EBPFP 3.5521 equivalent bits/point +MSE 68.270101 +---------------------- -------------------------------------------------------- +Time: 2.970s Load: 0.008s, Pack+Encode: 1.734s, Decode+Unpack: 1.228s +---------------------- -------------------------------------------------------- +💾 Converting with 68.2701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,868B, BPFP=0.7195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,800B, BPFP=3.2309 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,092B, BPFP=1.5951 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,036B, BPFP=3.1141 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,632B, BPFP=1.7632 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,944B, BPFP=3.0418 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,236B, BPFP=1.6708 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,076B, BPFP=3.1168 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,388B, BPFP=2.5416 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,908B, BPFP=2.9733 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,376B, BPFP=0.8359 +⌛️ [2/4] FRONTEND: Frontend time: 1.744s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.232s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13591524 30.13766635 + layer.0.v_cache 0.00001656 0.00239786 + layer.1.k_cache 0.34520291 2.86189839 + layer.1.v_cache 0.00000598 0.00097991 + layer.2.k_cache 0.01250870 0.54878920 + layer.2.v_cache 0.00001982 0.00282182 + layer.3.k_cache 0.02633334 4.04856226 + layer.3.v_cache 0.00002177 0.00333871 + layer.4.k_cache 0.00067446 0.08313847 + layer.4.v_cache 0.00005295 0.00604259 + layer.4.output 1.29725540 210.86435003 + ------------------------------------------------------------------------------------- + TOTAL 0.56479644 89.04388740 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 447356 +BPFP 1.7423 bits/point +EBPFP 3.4845 equivalent bits/point +MSE 89.043887 +---------------------- -------------------------------------------------------- +Time: 2.985s Load: 0.008s, Pack+Encode: 1.744s, Decode+Unpack: 1.232s +---------------------- -------------------------------------------------------- +💾 Converting with 89.0439 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,148B, BPFP=0.7307 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,292B, BPFP=3.4052 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,908B, BPFP=1.5055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,172B, BPFP=3.3246 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,924B, BPFP=1.7946 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,288B, BPFP=3.2609 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,552B, BPFP=1.6238 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,364B, BPFP=3.3384 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,680B, BPFP=2.5691 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,840B, BPFP=3.2287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,392B, BPFP=0.8064 +⌛️ [2/4] FRONTEND: Frontend time: 1.725s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.236s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150146 28.83911020 + layer.0.v_cache 0.00001564 0.00237142 + layer.1.k_cache 0.36060091 2.79196040 + layer.1.v_cache 0.00000598 0.00095409 + layer.2.k_cache 0.01419523 0.63499310 + layer.2.v_cache 0.00002081 0.00282494 + layer.3.k_cache 0.03680107 3.63586370 + layer.3.v_cache 0.00002092 0.00319183 + layer.4.k_cache 0.00067164 0.08064519 + layer.4.v_cache 0.00004838 0.00603629 + layer.4.output 1.41076765 234.75769421 + ------------------------------------------------------------------------------------- + TOTAL 0.61289739 98.78245945 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 422560 +BPFP 1.7898 bits/point +EBPFP 3.5796 equivalent bits/point +MSE 98.782459 +---------------------- -------------------------------------------------------- +Time: 2.970s Load: 0.008s, Pack+Encode: 1.725s, Decode+Unpack: 1.236s +---------------------- -------------------------------------------------------- +💾 Converting with 98.7825 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 343, 128) +Output shape: (1, 343, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.output: torch.Size([1, 343, 3584]) -> torch.Size([1, 1, 343, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,180B, BPFP=0.6915 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 71,364B, BPFP=3.2509 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 37,620B, BPFP=1.7137 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 69,240B, BPFP=3.1542 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 37,612B, BPFP=1.7134 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 67,588B, BPFP=3.0789 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 34,036B, BPFP=1.5505 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 69,200B, BPFP=3.1523 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 54,200B, BPFP=2.4690 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 66,992B, BPFP=3.0517 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 133,128B, BPFP=0.8664 +⌛️ [2/4] FRONTEND: Frontend time: 1.975s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.484s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16677478 30.96263154 + layer.0.v_cache 0.00001698 0.00226790 + layer.1.k_cache 0.72144124 3.36870644 + layer.1.v_cache 0.00000659 0.00096002 + layer.2.k_cache 0.01351720 0.57202326 + layer.2.v_cache 0.00002117 0.00285424 + layer.3.k_cache 0.02347133 4.06392534 + layer.3.v_cache 0.00002006 0.00301114 + layer.4.k_cache 0.00069292 0.07947865 + layer.4.v_cache 0.00005144 0.00600534 + layer.4.output 0.03905215 145.73146606 + ------------------------------------------------------------------------------------- + TOTAL 0.07055169 62.30483096 + (elements=2,985,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2985472 +Total Bytes 656160 +BPFP 1.7583 bits/point +EBPFP 3.5165 equivalent bits/point +MSE 62.304831 +---------------------- -------------------------------------------------------- +Time: 3.470s Load: 0.011s, Pack+Encode: 1.975s, Decode+Unpack: 1.484s +---------------------- -------------------------------------------------------- +💾 Converting with 62.3048 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,564B, BPFP=0.7139 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,296B, BPFP=3.3691 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,560B, BPFP=1.5659 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,260B, BPFP=3.1966 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,096B, BPFP=1.7100 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,540B, BPFP=3.1557 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,460B, BPFP=1.5602 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,420B, BPFP=3.2625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,868B, BPFP=2.4925 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,484B, BPFP=3.1525 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,812B, BPFP=0.8183 +⌛️ [2/4] FRONTEND: Frontend time: 1.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13317820 29.45721236 + layer.0.v_cache 0.00001784 0.00237247 + layer.1.k_cache 0.51782770 2.82413952 + layer.1.v_cache 0.00000612 0.00094774 + layer.2.k_cache 0.01694398 0.64378113 + layer.2.v_cache 0.00002016 0.00276651 + layer.3.k_cache 0.01920085 3.87444735 + layer.3.v_cache 0.00002016 0.00307213 + layer.4.k_cache 0.00071488 0.08097115 + layer.4.v_cache 0.00005349 0.00602308 + layer.4.output 0.00484358 176.27167208 + ------------------------------------------------------------------------------------- + TOTAL 0.04246403 74.75279047 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 526360 +BPFP 1.7592 bits/point +EBPFP 3.5184 equivalent bits/point +MSE 74.752790 +---------------------- -------------------------------------------------------- +Time: 3.215s Load: 0.010s, Pack+Encode: 1.847s, Decode+Unpack: 1.359s +---------------------- -------------------------------------------------------- +💾 Converting with 74.7528 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 374, 128) +Output shape: (1, 374, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.output: torch.Size([1, 374, 3584]) -> torch.Size([1, 1, 374, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,188B, BPFP=0.6763 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 72,528B, BPFP=3.0301 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 36,860B, BPFP=1.5399 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 70,448B, BPFP=2.9432 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 38,792B, BPFP=1.6207 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 68,188B, BPFP=2.8488 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 35,988B, BPFP=1.5035 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 69,636B, BPFP=2.9093 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 56,028B, BPFP=2.3407 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 66,880B, BPFP=2.7941 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 127,568B, BPFP=0.7614 +⌛️ [2/4] FRONTEND: Frontend time: 1.974s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.494s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18781370 31.67337901 + layer.0.v_cache 0.00001670 0.00222662 + layer.1.k_cache 0.85217587 3.12621581 + layer.1.v_cache 0.00000605 0.00092437 + layer.2.k_cache 0.02919085 0.60632230 + layer.2.v_cache 0.00002159 0.00265558 + layer.3.k_cache 0.02976600 4.44444071 + layer.3.v_cache 0.00002022 0.00300375 + layer.4.k_cache 0.00072409 0.08078086 + layer.4.v_cache 0.00004949 0.00560955 + layer.4.output 0.03578218 124.17374905 + ------------------------------------------------------------------------------------- + TOTAL 0.07942705 53.48010599 + (elements=3,255,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3255296 +Total Bytes 659104 +BPFP 1.6198 bits/point +EBPFP 3.2395 equivalent bits/point +MSE 53.480106 +---------------------- -------------------------------------------------------- +Time: 3.482s Load: 0.014s, Pack+Encode: 1.974s, Decode+Unpack: 1.494s +---------------------- -------------------------------------------------------- +💾 Converting with 53.4801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,992B, BPFP=0.7127 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,128B, BPFP=3.1203 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,048B, BPFP=1.4295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,040B, BPFP=3.0498 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,088B, BPFP=1.6914 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,832B, BPFP=2.9715 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,040B, BPFP=1.5586 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,840B, BPFP=3.0368 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,624B, BPFP=2.4393 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,124B, BPFP=2.9256 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,052B, BPFP=0.8063 +⌛️ [2/4] FRONTEND: Frontend time: 1.729s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.238s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13979621 30.29468483 + layer.0.v_cache 0.00001743 0.00227735 + layer.1.k_cache 0.46959303 2.91004076 + layer.1.v_cache 0.00000602 0.00095846 + layer.2.k_cache 0.01642122 0.63014038 + layer.2.v_cache 0.00002053 0.00283323 + layer.3.k_cache 0.03876302 4.08210520 + layer.3.v_cache 0.00002036 0.00313487 + layer.4.k_cache 0.00069112 0.08196723 + layer.4.v_cache 0.00005399 0.00611043 + layer.4.output 1.27033806 90.56036974 + ------------------------------------------------------------------------------------- + TOTAL 0.56222055 39.52569652 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 440808 +BPFP 1.6811 bits/point +EBPFP 3.3623 equivalent bits/point +MSE 39.525697 +---------------------- -------------------------------------------------------- +Time: 2.978s Load: 0.011s, Pack+Encode: 1.729s, Decode+Unpack: 1.238s +---------------------- -------------------------------------------------------- +💾 Converting with 39.5257 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,520B, BPFP=0.7241 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,600B, BPFP=3.3453 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,264B, BPFP=1.5325 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,380B, BPFP=3.2613 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,220B, BPFP=1.8048 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,548B, BPFP=3.2040 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,260B, BPFP=1.6010 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,432B, BPFP=3.2649 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,256B, BPFP=2.5644 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,632B, BPFP=3.1410 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,548B, BPFP=0.8707 +⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.234s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13893990 29.31612016 + layer.0.v_cache 0.00001711 0.00237556 + layer.1.k_cache 0.36820191 2.76065238 + layer.1.v_cache 0.00000637 0.00097378 + layer.2.k_cache 0.01566778 0.57742316 + layer.2.v_cache 0.00002161 0.00292277 + layer.3.k_cache 0.01884081 3.99443316 + layer.3.v_cache 0.00002076 0.00316641 + layer.4.k_cache 0.00070054 0.08457872 + layer.4.v_cache 0.00005210 0.00614639 + layer.4.output 1.34866800 216.91173694 + ------------------------------------------------------------------------------------- + TOTAL 0.58724382 91.47829124 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 443660 +BPFP 1.7964 bits/point +EBPFP 3.5927 equivalent bits/point +MSE 91.478291 +---------------------- -------------------------------------------------------- +Time: 2.978s Load: 0.009s, Pack+Encode: 1.735s, Decode+Unpack: 1.234s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4783 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,392B, BPFP=0.7252 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,628B, BPFP=3.3724 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,284B, BPFP=1.4796 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,204B, BPFP=3.2891 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,616B, BPFP=1.7331 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,064B, BPFP=3.2224 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,084B, BPFP=1.5850 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,688B, BPFP=3.3174 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,744B, BPFP=2.5599 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,892B, BPFP=3.2123 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,012B, BPFP=0.8277 +⌛️ [2/4] FRONTEND: Frontend time: 1.844s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11943251 28.49657472 + layer.0.v_cache 0.00001561 0.00227331 + layer.1.k_cache 0.49619702 2.71551902 + layer.1.v_cache 0.00000608 0.00095213 + layer.2.k_cache 0.01349212 0.68784124 + layer.2.v_cache 0.00002006 0.00282222 + layer.3.k_cache 0.01080775 3.57514594 + layer.3.v_cache 0.00002073 0.00301458 + layer.4.k_cache 0.00068886 0.07946612 + layer.4.v_cache 0.00007406 0.00632612 + layer.4.output 0.00494179 69.83734617 + ------------------------------------------------------------------------------------- + TOTAL 0.03972631 30.84890345 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 517608 +BPFP 1.7818 bits/point +EBPFP 3.5636 equivalent bits/point +MSE 30.848903 +---------------------- -------------------------------------------------------- +Time: 3.210s Load: 0.009s, Pack+Encode: 1.844s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 30.8489 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,212B, BPFP=0.7220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,344B, BPFP=3.3473 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,664B, BPFP=1.5317 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,204B, BPFP=3.2667 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,068B, BPFP=1.7723 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,512B, BPFP=3.2178 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,056B, BPFP=1.6301 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,400B, BPFP=3.2805 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,656B, BPFP=2.5916 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,880B, BPFP=3.1731 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,060B, BPFP=0.8187 +⌛️ [2/4] FRONTEND: Frontend time: 1.716s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11815372 30.71419197 + layer.0.v_cache 0.00001627 0.00232857 + layer.1.k_cache 0.37623358 2.98333409 + layer.1.v_cache 0.00000616 0.00098487 + layer.2.k_cache 0.01429399 0.67084624 + layer.2.v_cache 0.00002051 0.00289850 + layer.3.k_cache 0.01320818 3.70749053 + layer.3.v_cache 0.00002004 0.00323656 + layer.4.k_cache 0.00067885 0.08213854 + layer.4.v_cache 0.00005318 0.00630723 + layer.4.output 1.38524649 220.99727295 + ------------------------------------------------------------------------------------- + TOTAL 0.60114176 93.24439222 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 428056 +BPFP 1.7802 bits/point +EBPFP 3.5605 equivalent bits/point +MSE 93.244392 +---------------------- -------------------------------------------------------- +Time: 2.946s Load: 0.008s, Pack+Encode: 1.716s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 93.2444 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,748B, BPFP=0.7165 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,992B, BPFP=3.2594 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,376B, BPFP=1.5387 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,112B, BPFP=3.1538 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,400B, BPFP=1.7086 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,036B, BPFP=3.0933 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,944B, BPFP=1.5706 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,940B, BPFP=3.2003 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,352B, BPFP=2.4928 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,636B, BPFP=3.1270 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,296B, BPFP=0.7652 +⌛️ [2/4] FRONTEND: Frontend time: 1.840s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.348s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09820673 28.96509843 + layer.0.v_cache 0.00001666 0.00221211 + layer.1.k_cache 0.50544519 2.90629742 + layer.1.v_cache 0.00000613 0.00093265 + layer.2.k_cache 0.01930715 0.64721422 + layer.2.v_cache 0.00002033 0.00282918 + layer.3.k_cache 0.01337509 3.47771383 + layer.3.v_cache 0.00001987 0.00298522 + layer.4.k_cache 0.00074230 0.08162063 + layer.4.v_cache 0.00004869 0.00581536 + layer.4.output 0.00473525 177.13119861 + ------------------------------------------------------------------------------------- + TOTAL 0.03943146 75.05947702 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 519832 +BPFP 1.7187 bits/point +EBPFP 3.4373 equivalent bits/point +MSE 75.059477 +---------------------- -------------------------------------------------------- +Time: 3.197s Load: 0.009s, Pack+Encode: 1.840s, Decode+Unpack: 1.348s +---------------------- -------------------------------------------------------- +💾 Converting with 75.0595 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,264B, BPFP=0.7357 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,092B, BPFP=3.3753 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,996B, BPFP=1.4332 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,520B, BPFP=3.2626 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,728B, BPFP=1.7724 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,400B, BPFP=3.1823 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,360B, BPFP=1.6026 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,252B, BPFP=3.3151 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,872B, BPFP=2.5711 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,320B, BPFP=3.1766 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,872B, BPFP=0.7769 +⌛️ [2/4] FRONTEND: Frontend time: 1.767s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.246s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13853801 28.96884855 + layer.0.v_cache 0.00001803 0.00245959 + layer.1.k_cache 0.23740530 2.33722974 + layer.1.v_cache 0.00000581 0.00099058 + layer.2.k_cache 0.01896443 0.67200519 + layer.2.v_cache 0.00001977 0.00295019 + layer.3.k_cache 0.00876656 3.07101636 + layer.3.v_cache 0.00002034 0.00341424 + layer.4.k_cache 0.00073522 0.08375138 + layer.4.v_cache 0.00004914 0.00626369 + layer.4.output 1.40427464 235.38593545 + ------------------------------------------------------------------------------------- + TOTAL 0.60202618 98.99120457 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 416676 +BPFP 1.7568 bits/point +EBPFP 3.5135 equivalent bits/point +MSE 98.991205 +---------------------- -------------------------------------------------------- +Time: 3.021s Load: 0.008s, Pack+Encode: 1.767s, Decode+Unpack: 1.246s +---------------------- -------------------------------------------------------- +💾 Converting with 98.9912 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,648B, BPFP=0.7160 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,512B, BPFP=3.3125 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,884B, BPFP=1.7484 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,268B, BPFP=3.1855 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,048B, BPFP=1.7011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,468B, BPFP=3.1402 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,388B, BPFP=1.5505 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,356B, BPFP=3.3037 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,376B, BPFP=2.5122 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,928B, BPFP=3.1662 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,080B, BPFP=0.8094 +⌛️ [2/4] FRONTEND: Frontend time: 1.892s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10827675 28.44487375 + layer.0.v_cache 0.00001706 0.00225941 + layer.1.k_cache 0.56919524 3.02045607 + layer.1.v_cache 0.00000592 0.00091142 + layer.2.k_cache 0.01911232 0.59908051 + layer.2.v_cache 0.00002018 0.00289017 + layer.3.k_cache 0.01342056 3.70330811 + layer.3.v_cache 0.00002333 0.00315883 + layer.4.k_cache 0.00074072 0.08225981 + layer.4.v_cache 0.00005062 0.00602249 + layer.4.output 0.00480139 170.86652433 + ------------------------------------------------------------------------------------- + TOTAL 0.04379250 72.46652299 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 529956 +BPFP 1.7648 bits/point +EBPFP 3.5297 equivalent bits/point +MSE 72.466523 +---------------------- -------------------------------------------------------- +Time: 3.273s Load: 0.009s, Pack+Encode: 1.892s, Decode+Unpack: 1.372s +---------------------- -------------------------------------------------------- +💾 Converting with 72.4665 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,488B, BPFP=0.7308 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,456B, BPFP=3.4209 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,872B, BPFP=1.4555 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,388B, BPFP=3.2999 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,444B, BPFP=1.7816 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,700B, BPFP=3.2596 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,964B, BPFP=1.5779 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,896B, BPFP=3.3881 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,464B, BPFP=2.5435 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,308B, BPFP=3.2367 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,120B, BPFP=0.8370 +⌛️ [2/4] FRONTEND: Frontend time: 1.842s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15963619 29.31097115 + layer.0.v_cache 0.00001721 0.00243723 + layer.1.k_cache 0.49929055 2.64573504 + layer.1.v_cache 0.00000624 0.00095079 + layer.2.k_cache 0.03270629 0.74778936 + layer.2.v_cache 0.00002062 0.00274112 + layer.3.k_cache 0.00946706 3.38040150 + layer.3.v_cache 0.00002056 0.00300142 + layer.4.k_cache 0.00068536 0.07766374 + layer.4.v_cache 0.00004941 0.00581989 + layer.4.output 0.00495045 168.47819690 + ------------------------------------------------------------------------------------- + TOTAL 0.04332662 71.50146409 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 522100 +BPFP 1.7973 bits/point +EBPFP 3.5945 equivalent bits/point +MSE 71.501464 +---------------------- -------------------------------------------------------- +Time: 3.205s Load: 0.008s, Pack+Encode: 1.842s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 71.5015 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,868B, BPFP=0.7195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,440B, BPFP=3.2071 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,016B, BPFP=1.5900 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,272B, BPFP=3.1298 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,200B, BPFP=1.7346 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,848B, BPFP=3.0355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,164B, BPFP=1.5998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,228B, BPFP=3.1269 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,756B, BPFP=2.4997 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,344B, BPFP=3.0021 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,112B, BPFP=0.8050 +⌛️ [2/4] FRONTEND: Frontend time: 1.718s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.443s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14874478 31.33243263 + layer.0.v_cache 0.00001714 0.00234152 + layer.1.k_cache 0.45078779 2.89862992 + layer.1.v_cache 0.00000615 0.00092228 + layer.2.k_cache 0.01724745 0.54643211 + layer.2.v_cache 0.00001987 0.00267011 + layer.3.k_cache 0.02849353 4.09906833 + layer.3.v_cache 0.00002056 0.00306811 + layer.4.k_cache 0.00066912 0.08146427 + layer.4.v_cache 0.00005068 0.00588188 + layer.4.output 1.29722614 198.21309398 + ------------------------------------------------------------------------------------- + TOTAL 0.57215530 83.90968053 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 442248 +BPFP 1.7224 bits/point +EBPFP 3.4447 equivalent bits/point +MSE 83.909681 +---------------------- -------------------------------------------------------- +Time: 3.169s Load: 0.008s, Pack+Encode: 1.718s, Decode+Unpack: 1.443s +---------------------- -------------------------------------------------------- +💾 Converting with 83.9097 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 341, 128) +Output shape: (1, 341, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.output: torch.Size([1, 341, 3584]) -> torch.Size([1, 1, 341, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,136B, BPFP=0.6935 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 70,928B, BPFP=3.2500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 34,748B, BPFP=1.5922 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 68,848B, BPFP=3.1547 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 37,556B, BPFP=1.7209 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 67,820B, BPFP=3.1076 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 33,600B, BPFP=1.5396 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 69,524B, BPFP=3.1857 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 53,756B, BPFP=2.4632 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 66,668B, BPFP=3.0548 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 120,540B, BPFP=0.7890 +⌛️ [2/4] FRONTEND: Frontend time: 2.046s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.483s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17009405 31.85829820 + layer.0.v_cache 0.00001631 0.00221618 + layer.1.k_cache 0.79816426 3.13250929 + layer.1.v_cache 0.00000611 0.00093104 + layer.2.k_cache 0.02878924 0.66588505 + layer.2.v_cache 0.00002016 0.00272756 + layer.3.k_cache 0.01035325 3.90188446 + layer.3.v_cache 0.00002082 0.00294106 + layer.4.k_cache 0.00072160 0.08075842 + layer.4.v_cache 0.00005837 0.00575074 + layer.4.output 0.03919861 137.81010421 + ------------------------------------------------------------------------------------- + TOTAL 0.07544909 59.07791950 + (elements=2,968,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2968064 +Total Bytes 639124 +BPFP 1.7227 bits/point +EBPFP 3.4453 equivalent bits/point +MSE 59.077919 +---------------------- -------------------------------------------------------- +Time: 3.542s Load: 0.013s, Pack+Encode: 2.046s, Decode+Unpack: 1.483s +---------------------- -------------------------------------------------------- +💾 Converting with 59.0779 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 335, 128) +Output shape: (1, 335, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.output: torch.Size([1, 335, 3584]) -> torch.Size([1, 1, 335, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,056B, BPFP=0.7022 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 70,628B, BPFP=3.2942 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 36,356B, BPFP=1.6957 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 68,464B, BPFP=3.1933 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 36,292B, BPFP=1.6927 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 66,496B, BPFP=3.1015 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 32,632B, BPFP=1.5220 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 68,596B, BPFP=3.1994 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 53,048B, BPFP=2.4743 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 66,220B, BPFP=3.0886 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 122,492B, BPFP=0.8162 +⌛️ [2/4] FRONTEND: Frontend time: 2.086s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.534s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15195324 30.25933710 + layer.0.v_cache 0.00001666 0.00229120 + layer.1.k_cache 0.76788093 3.11691312 + layer.1.v_cache 0.00000624 0.00095450 + layer.2.k_cache 0.02686723 0.63748916 + layer.2.v_cache 0.00001960 0.00273379 + layer.3.k_cache 0.01080669 3.78767381 + layer.3.v_cache 0.00002024 0.00301501 + layer.4.k_cache 0.00071790 0.08145628 + layer.4.v_cache 0.00005106 0.00601639 + layer.4.output 3.38143948 140.58534115 + ------------------------------------------------------------------------------------- + TOTAL 1.44873036 60.11736873 + (elements=2,915,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2915840 +Total Bytes 636280 +BPFP 1.7457 bits/point +EBPFP 3.4914 equivalent bits/point +MSE 60.117369 +---------------------- -------------------------------------------------------- +Time: 3.632s Load: 0.012s, Pack+Encode: 2.086s, Decode+Unpack: 1.534s +---------------------- -------------------------------------------------------- +💾 Converting with 60.1174 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,272B, BPFP=0.7236 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,460B, BPFP=3.4469 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,992B, BPFP=1.5325 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,748B, BPFP=3.3460 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,428B, BPFP=1.7351 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,520B, BPFP=3.2736 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,796B, BPFP=1.5800 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,588B, BPFP=3.3955 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,844B, BPFP=2.5262 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,708B, BPFP=3.2257 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 103,288B, BPFP=0.8700 +⌛️ [2/4] FRONTEND: Frontend time: 1.884s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17469128 30.37714291 + layer.0.v_cache 0.00001775 0.00232761 + layer.1.k_cache 0.51265420 2.96152183 + layer.1.v_cache 0.00000620 0.00093513 + layer.2.k_cache 0.03428403 0.70095802 + layer.2.v_cache 0.00002024 0.00276980 + layer.3.k_cache 0.01443097 3.27606593 + layer.3.v_cache 0.00002182 0.00314630 + layer.4.k_cache 0.00070119 0.07994794 + layer.4.v_cache 0.00005097 0.00576425 + layer.4.output 0.00503371 189.92109164 + ------------------------------------------------------------------------------------- + TOTAL 0.04541851 80.40342478 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 523644 +BPFP 1.8162 bits/point +EBPFP 3.6324 equivalent bits/point +MSE 80.403425 +---------------------- -------------------------------------------------------- +Time: 3.266s Load: 0.010s, Pack+Encode: 1.884s, Decode+Unpack: 1.372s +---------------------- -------------------------------------------------------- +💾 Converting with 80.4034 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,228B, BPFP=0.7210 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,164B, BPFP=3.3705 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,900B, BPFP=1.4092 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,164B, BPFP=3.3116 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,352B, BPFP=1.7307 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,028B, BPFP=3.2446 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,612B, BPFP=1.5691 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,292B, BPFP=3.3781 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,312B, BPFP=2.5538 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,480B, BPFP=3.2123 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 97,492B, BPFP=0.8212 +⌛️ [2/4] FRONTEND: Frontend time: 1.843s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18043976 30.10373305 + layer.0.v_cache 0.00001814 0.00226063 + layer.1.k_cache 0.58196952 2.41215014 + layer.1.v_cache 0.00000608 0.00094935 + layer.2.k_cache 0.02409397 0.73225789 + layer.2.v_cache 0.00002020 0.00278967 + layer.3.k_cache 0.01875559 3.32440992 + layer.3.v_cache 0.00002085 0.00312409 + layer.4.k_cache 0.00069073 0.07947144 + layer.4.v_cache 0.00005389 0.00616085 + layer.4.output 0.00498411 185.67885782 + ------------------------------------------------------------------------------------- + TOTAL 0.04946809 78.61290069 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 513024 +BPFP 1.7794 bits/point +EBPFP 3.5587 equivalent bits/point +MSE 78.612901 +---------------------- -------------------------------------------------------- +Time: 3.221s Load: 0.009s, Pack+Encode: 1.843s, Decode+Unpack: 1.369s +---------------------- -------------------------------------------------------- +💾 Converting with 78.6129 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 386, 128) +Output shape: (1, 386, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.output: torch.Size([1, 386, 3584]) -> torch.Size([1, 1, 386, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,964B, BPFP=0.6867 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 81,296B, BPFP=3.2908 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 42,220B, BPFP=1.7090 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 79,184B, BPFP=3.2053 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 42,056B, BPFP=1.7024 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 77,592B, BPFP=3.1409 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 38,348B, BPFP=1.5523 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 79,248B, BPFP=3.2079 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 61,832B, BPFP=2.5029 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 76,612B, BPFP=3.1012 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 147,752B, BPFP=0.8544 +⌛️ [2/4] FRONTEND: Frontend time: 2.635s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.802s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16808204 30.44295205 + layer.0.v_cache 0.00001827 0.00234664 + layer.1.k_cache 0.89383947 3.43074091 + layer.1.v_cache 0.00000605 0.00094983 + layer.2.k_cache 0.02643975 0.69666104 + layer.2.v_cache 0.00002094 0.00287523 + layer.3.k_cache 0.04368651 3.68983072 + layer.3.v_cache 0.00002149 0.00312464 + layer.4.k_cache 0.00075087 0.08077562 + layer.4.v_cache 0.00005208 0.00613052 + layer.4.output 0.03469837 122.53763416 + ------------------------------------------------------------------------------------- + TOTAL 0.08092977 52.71293096 + (elements=3,359,744) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3359744 +Total Bytes 743104 +BPFP 1.7694 bits/point +EBPFP 3.5389 equivalent bits/point +MSE 52.712931 +---------------------- -------------------------------------------------------- +Time: 4.450s Load: 0.014s, Pack+Encode: 2.635s, Decode+Unpack: 1.802s +---------------------- -------------------------------------------------------- +💾 Converting with 52.7129 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,780B, BPFP=0.6934 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,156B, BPFP=3.2637 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,232B, BPFP=1.4232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,784B, BPFP=3.1892 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,292B, BPFP=1.7520 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,540B, BPFP=3.1217 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,592B, BPFP=1.5512 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 59,100B, BPFP=3.2064 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,296B, BPFP=2.5117 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,088B, BPFP=3.0430 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 109,348B, BPFP=0.8475 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14364293 29.62355211 + layer.0.v_cache 0.00001600 0.00236312 + layer.1.k_cache 0.62366660 2.92010901 + layer.1.v_cache 0.00000628 0.00099334 + layer.2.k_cache 0.01981349 0.57321845 + layer.2.v_cache 0.00002090 0.00293337 + layer.3.k_cache 0.02158341 3.79271528 + layer.3.v_cache 0.00002097 0.00315195 + layer.4.k_cache 0.00071696 0.08189601 + layer.4.v_cache 0.00005883 0.00620474 + layer.4.output 0.00463133 172.27441406 + ------------------------------------------------------------------------------------- + TOTAL 0.04952739 73.11341387 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 547208 +BPFP 1.7463 bits/point +EBPFP 3.4927 equivalent bits/point +MSE 73.113414 +---------------------- -------------------------------------------------------- +Time: 3.240s Load: 0.009s, Pack+Encode: 1.855s, Decode+Unpack: 1.376s +---------------------- -------------------------------------------------------- +💾 Converting with 73.1134 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,588B, BPFP=0.7101 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,360B, BPFP=3.2920 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,952B, BPFP=1.6331 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,464B, BPFP=3.1850 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,444B, BPFP=1.7173 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,728B, BPFP=3.1435 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,756B, BPFP=1.5657 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,032B, BPFP=3.2735 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,476B, BPFP=2.5088 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,956B, BPFP=3.1564 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 108,740B, BPFP=0.8763 +⌛️ [2/4] FRONTEND: Frontend time: 2.147s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.601s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12300906 30.44215718 + layer.0.v_cache 0.00001720 0.00210369 + layer.1.k_cache 0.58060381 3.05462272 + layer.1.v_cache 0.00000644 0.00085632 + layer.2.k_cache 0.02305070 0.63274601 + layer.2.v_cache 0.00002053 0.00268440 + layer.3.k_cache 0.05279791 3.93825357 + layer.3.v_cache 0.00002021 0.00291621 + layer.4.k_cache 0.00074809 0.08244093 + layer.4.v_cache 0.00005348 0.00595314 + layer.4.output 0.00483522 176.28618167 + ------------------------------------------------------------------------------------- + TOTAL 0.04789259 74.83341211 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 537496 +BPFP 1.7835 bits/point +EBPFP 3.5669 equivalent bits/point +MSE 74.833412 +---------------------- -------------------------------------------------------- +Time: 3.757s Load: 0.009s, Pack+Encode: 2.147s, Decode+Unpack: 1.601s +---------------------- -------------------------------------------------------- +💾 Converting with 74.8334 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,292B, BPFP=0.7211 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,180B, BPFP=3.3758 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,908B, BPFP=1.5350 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,336B, BPFP=3.2466 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,372B, BPFP=1.7777 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,548B, BPFP=3.1914 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,032B, BPFP=1.6138 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,776B, BPFP=3.2775 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,604B, BPFP=2.5647 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,100B, BPFP=3.1600 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,984B, BPFP=0.8507 +⌛️ [2/4] FRONTEND: Frontend time: 2.181s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.559s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16177035 28.62162801 + layer.0.v_cache 0.00001782 0.00251163 + layer.1.k_cache 0.31952496 2.74000830 + layer.1.v_cache 0.00000626 0.00097684 + layer.2.k_cache 0.01271096 0.55300452 + layer.2.v_cache 0.00002040 0.00287608 + layer.3.k_cache 0.01363577 4.12447614 + layer.3.v_cache 0.00002033 0.00314986 + layer.4.k_cache 0.00066992 0.08218832 + layer.4.v_cache 0.00005030 0.00589843 + layer.4.output 1.37284199 226.96706839 + ------------------------------------------------------------------------------------- + TOTAL 0.59519535 95.58271746 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 434132 +BPFP 1.7893 bits/point +EBPFP 3.5786 equivalent bits/point +MSE 95.582717 +---------------------- -------------------------------------------------------- +Time: 3.749s Load: 0.009s, Pack+Encode: 2.181s, Decode+Unpack: 1.559s +---------------------- -------------------------------------------------------- +💾 Converting with 95.5827 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,552B, BPFP=0.7302 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,244B, BPFP=3.1800 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,616B, BPFP=1.4187 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,544B, BPFP=3.0348 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,672B, BPFP=1.7650 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,084B, BPFP=2.9956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,296B, BPFP=1.6475 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,828B, BPFP=3.0591 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,108B, BPFP=2.4853 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,360B, BPFP=2.9337 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 73,692B, BPFP=0.8989 +⌛️ [2/4] FRONTEND: Frontend time: 2.400s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.395s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09962382 28.14014472 + layer.0.v_cache 0.00001686 0.00256133 + layer.1.k_cache 0.10988800 2.46897446 + layer.1.v_cache 0.00000590 0.00100224 + layer.2.k_cache 0.01214521 0.53286385 + layer.2.v_cache 0.00002216 0.00303876 + layer.3.k_cache 0.02260447 3.83524110 + layer.3.v_cache 0.00002006 0.00342938 + layer.4.k_cache 0.00069723 0.08443536 + layer.4.v_cache 0.00005242 0.00647440 + layer.4.output 0.00882482 269.36560792 + ------------------------------------------------------------------------------------- + TOTAL 0.01804999 112.97867183 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 345996 +BPFP 1.7378 bits/point +EBPFP 3.4755 equivalent bits/point +MSE 112.978672 +---------------------- -------------------------------------------------------- +Time: 3.803s Load: 0.008s, Pack+Encode: 2.400s, Decode+Unpack: 1.395s +---------------------- -------------------------------------------------------- +💾 Converting with 112.9787 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,796B, BPFP=0.7115 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,932B, BPFP=3.2769 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,480B, BPFP=1.5836 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,252B, BPFP=3.1835 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,040B, BPFP=1.7260 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,420B, BPFP=3.0816 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,208B, BPFP=1.5685 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,124B, BPFP=3.2320 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,796B, BPFP=2.4909 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,848B, BPFP=3.1054 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,956B, BPFP=0.8178 +⌛️ [2/4] FRONTEND: Frontend time: 2.281s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.757s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11148910 30.42463440 + layer.0.v_cache 0.00001684 0.00228425 + layer.1.k_cache 0.60671574 2.69711619 + layer.1.v_cache 0.00000630 0.00092019 + layer.2.k_cache 0.04149003 0.70046943 + layer.2.v_cache 0.00001976 0.00264993 + layer.3.k_cache 0.02200458 3.58643881 + layer.3.v_cache 0.00001954 0.00298246 + layer.4.k_cache 0.00078895 0.07949932 + layer.4.v_cache 0.00004864 0.00571717 + layer.4.output 0.00472744 184.81205516 + ------------------------------------------------------------------------------------- + TOTAL 0.04798185 78.30512343 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 533852 +BPFP 1.7462 bits/point +EBPFP 3.4923 equivalent bits/point +MSE 78.305123 +---------------------- -------------------------------------------------------- +Time: 4.048s Load: 0.010s, Pack+Encode: 2.281s, Decode+Unpack: 1.757s +---------------------- -------------------------------------------------------- +💾 Converting with 78.3051 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,632B, BPFP=0.7525 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,360B, BPFP=3.5438 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,204B, BPFP=1.5784 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,400B, BPFP=3.4688 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,548B, BPFP=1.8397 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,412B, BPFP=3.3916 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,008B, BPFP=1.6413 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,084B, BPFP=3.5222 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,352B, BPFP=2.6837 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,688B, BPFP=3.4131 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,492B, BPFP=0.8314 +⌛️ [2/4] FRONTEND: Frontend time: 2.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.554s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14565937 29.59301270 + layer.0.v_cache 0.00001550 0.00232472 + layer.1.k_cache 0.22559690 2.76881897 + layer.1.v_cache 0.00000590 0.00098554 + layer.2.k_cache 0.00679684 0.74713226 + layer.2.v_cache 0.00002314 0.00297496 + layer.3.k_cache 0.06866968 3.70562195 + layer.3.v_cache 0.00002119 0.00334217 + layer.4.k_cache 0.00065507 0.08028465 + layer.4.v_cache 0.00005191 0.00622188 + layer.4.output 1.53062134 233.10397321 + ------------------------------------------------------------------------------------- + TOTAL 0.65657911 98.15520778 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 405180 +BPFP 1.8620 bits/point +EBPFP 3.7241 equivalent bits/point +MSE 98.155208 +---------------------- -------------------------------------------------------- +Time: 3.815s Load: 0.008s, Pack+Encode: 2.253s, Decode+Unpack: 1.554s +---------------------- -------------------------------------------------------- +💾 Converting with 98.1552 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,624B, BPFP=0.7481 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,756B, BPFP=3.4792 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,912B, BPFP=1.4701 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,024B, BPFP=3.4223 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,456B, BPFP=1.8234 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,552B, BPFP=3.3856 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,268B, BPFP=1.6533 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,820B, BPFP=3.4841 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,480B, BPFP=2.6803 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,108B, BPFP=3.3511 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,896B, BPFP=0.8428 +⌛️ [2/4] FRONTEND: Frontend time: 2.476s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.835s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10663509 28.11359705 + layer.0.v_cache 0.00001613 0.00245868 + layer.1.k_cache 0.26620468 2.25250138 + layer.1.v_cache 0.00000602 0.00104797 + layer.2.k_cache 0.00649704 0.73363180 + layer.2.v_cache 0.00002002 0.00322708 + layer.3.k_cache 0.02565914 3.50986280 + layer.3.v_cache 0.00002186 0.00358162 + layer.4.k_cache 0.00070101 0.08410552 + layer.4.v_cache 0.00005421 0.00672559 + layer.4.output 1.52303164 239.27036692 + ------------------------------------------------------------------------------------- + TOTAL 0.65100216 100.56490046 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 403896 +BPFP 1.8469 bits/point +EBPFP 3.6938 equivalent bits/point +MSE 100.564900 +---------------------- -------------------------------------------------------- +Time: 4.318s Load: 0.007s, Pack+Encode: 2.476s, Decode+Unpack: 1.835s +---------------------- -------------------------------------------------------- +💾 Converting with 100.5649 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,344B, BPFP=0.7224 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,560B, BPFP=3.4270 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,368B, BPFP=1.6601 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,404B, BPFP=3.3008 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,760B, BPFP=1.8001 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,220B, BPFP=3.2315 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,860B, BPFP=1.6304 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,044B, BPFP=3.3382 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,164B, BPFP=2.5845 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,084B, BPFP=3.2235 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 105,488B, BPFP=0.8819 +⌛️ [2/4] FRONTEND: Frontend time: 2.801s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.782s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258070 28.84857612 + layer.0.v_cache 0.00001713 0.00245305 + layer.1.k_cache 0.49703808 3.02468289 + layer.1.v_cache 0.00000670 0.00103033 + layer.2.k_cache 0.01458295 0.61089711 + layer.2.v_cache 0.00002181 0.00290422 + layer.3.k_cache 0.02940617 3.52031968 + layer.3.v_cache 0.00002120 0.00326604 + layer.4.k_cache 0.00070157 0.08326969 + layer.4.v_cache 0.00005095 0.00625007 + layer.4.output 0.00497170 178.50660447 + ------------------------------------------------------------------------------------- + TOTAL 0.04171936 75.62646356 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 531296 +BPFP 1.8289 bits/point +EBPFP 3.6579 equivalent bits/point +MSE 75.626464 +---------------------- -------------------------------------------------------- +Time: 4.594s Load: 0.011s, Pack+Encode: 2.801s, Decode+Unpack: 1.782s +---------------------- -------------------------------------------------------- +💾 Converting with 75.6265 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,380B, BPFP=0.7176 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,272B, BPFP=3.3374 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,988B, BPFP=1.5202 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,944B, BPFP=3.2456 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,364B, BPFP=1.8227 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,260B, BPFP=3.1983 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,984B, BPFP=1.6582 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,412B, BPFP=3.2779 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,416B, BPFP=2.5868 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,240B, BPFP=3.1278 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,432B, BPFP=0.8537 +⌛️ [2/4] FRONTEND: Frontend time: 2.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.761s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15433978 29.91390694 + layer.0.v_cache 0.00001575 0.00243246 + layer.1.k_cache 0.29362690 2.48342274 + layer.1.v_cache 0.00000607 0.00100173 + layer.2.k_cache 0.00940348 0.50193800 + layer.2.v_cache 0.00002159 0.00305048 + layer.3.k_cache 0.02320663 3.90530152 + layer.3.v_cache 0.00002067 0.00339304 + layer.4.k_cache 0.00067785 0.08397248 + layer.4.v_cache 0.00005016 0.00623927 + layer.4.output 1.35461534 218.56947298 + ------------------------------------------------------------------------------------- + TOTAL 0.58609860 92.17005703 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 440692 +BPFP 1.7922 bits/point +EBPFP 3.5845 equivalent bits/point +MSE 92.170057 +---------------------- -------------------------------------------------------- +Time: 4.259s Load: 0.009s, Pack+Encode: 2.489s, Decode+Unpack: 1.761s +---------------------- -------------------------------------------------------- +💾 Converting with 92.1701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,848B, BPFP=0.7044 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,132B, BPFP=3.2967 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,704B, BPFP=1.6285 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,328B, BPFP=3.1978 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,088B, BPFP=1.7044 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,128B, BPFP=3.1320 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,108B, BPFP=1.5410 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,816B, BPFP=3.1697 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,500B, BPFP=2.4945 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,312B, BPFP=3.0873 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 111,000B, BPFP=0.8694 +⌛️ [2/4] FRONTEND: Frontend time: 2.802s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.768s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14751465 31.35731908 + layer.0.v_cache 0.00001719 0.00226452 + layer.1.k_cache 0.52245269 3.16688040 + layer.1.v_cache 0.00000618 0.00090693 + layer.2.k_cache 0.02707210 0.72414567 + layer.2.v_cache 0.00002227 0.00270858 + layer.3.k_cache 0.03364062 3.99073808 + layer.3.v_cache 0.00002083 0.00295832 + layer.4.k_cache 0.00069748 0.07806379 + layer.4.v_cache 0.00005305 0.00572673 + layer.4.output 0.00471531 174.45241228 + ------------------------------------------------------------------------------------- + TOTAL 0.04497084 74.14697636 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 547964 +BPFP 1.7672 bits/point +EBPFP 3.5343 equivalent bits/point +MSE 74.146976 +---------------------- -------------------------------------------------------- +Time: 4.580s Load: 0.010s, Pack+Encode: 2.802s, Decode+Unpack: 1.768s +---------------------- -------------------------------------------------------- +💾 Converting with 74.1470 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,984B, BPFP=0.6730 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,744B, BPFP=3.0480 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,948B, BPFP=1.5900 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,876B, BPFP=2.8723 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,636B, BPFP=1.5708 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,360B, BPFP=2.7794 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,064B, BPFP=1.4745 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,388B, BPFP=2.8424 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,656B, BPFP=2.3074 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,648B, BPFP=2.7358 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 108,696B, BPFP=0.9515 +⌛️ [2/4] FRONTEND: Frontend time: 2.612s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.881s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17432048 39.03150276 + layer.0.v_cache 0.00001704 0.00215195 + layer.1.k_cache 0.39057378 3.11426810 + layer.1.v_cache 0.00000668 0.00092834 + layer.2.k_cache 0.01299327 0.55955751 + layer.2.v_cache 0.00002114 0.00258602 + layer.3.k_cache 0.02499930 4.24143497 + layer.3.v_cache 0.00002096 0.00287074 + layer.4.k_cache 0.00069121 0.07728703 + layer.4.v_cache 0.00005416 0.00554171 + layer.4.output 1.20070616 198.14788165 + ------------------------------------------------------------------------------------- + TOTAL 0.52992007 84.35725298 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 466000 +BPFP 1.6796 bits/point +EBPFP 3.3593 equivalent bits/point +MSE 84.357253 +---------------------- -------------------------------------------------------- +Time: 4.505s Load: 0.011s, Pack+Encode: 2.612s, Decode+Unpack: 1.881s +---------------------- -------------------------------------------------------- +💾 Converting with 84.3573 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,552B, BPFP=0.7200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,472B, BPFP=3.3073 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,648B, BPFP=1.7500 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,600B, BPFP=3.2478 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,320B, BPFP=1.7276 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,116B, BPFP=3.1466 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,348B, BPFP=1.5931 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,064B, BPFP=3.2112 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,232B, BPFP=2.5404 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,164B, BPFP=3.0816 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,028B, BPFP=0.8678 +⌛️ [2/4] FRONTEND: Frontend time: 2.498s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.627s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16239141 29.07073255 + layer.0.v_cache 0.00001645 0.00232157 + layer.1.k_cache 0.33751179 2.90428195 + layer.1.v_cache 0.00000668 0.00096988 + layer.2.k_cache 0.00660169 0.60023452 + layer.2.v_cache 0.00002045 0.00275514 + layer.3.k_cache 0.02213682 3.94389776 + layer.3.v_cache 0.00002049 0.00311206 + layer.4.k_cache 0.00067824 0.08234545 + layer.4.v_cache 0.00005183 0.00607213 + layer.4.output 1.33690934 210.50647224 + ------------------------------------------------------------------------------------- + TOTAL 0.58163537 88.83306051 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 445544 +BPFP 1.7882 bits/point +EBPFP 3.5765 equivalent bits/point +MSE 88.833061 +---------------------- -------------------------------------------------------- +Time: 4.135s Load: 0.010s, Pack+Encode: 2.498s, Decode+Unpack: 1.627s +---------------------- -------------------------------------------------------- +💾 Converting with 88.8331 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,356B, BPFP=0.7204 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,276B, BPFP=3.3976 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,944B, BPFP=1.3960 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,404B, BPFP=3.2885 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,316B, BPFP=1.7092 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,124B, BPFP=3.2139 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,508B, BPFP=1.5455 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,260B, BPFP=3.3384 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,712B, BPFP=2.5485 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,860B, BPFP=3.1985 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,784B, BPFP=0.7728 +⌛️ [2/4] FRONTEND: Frontend time: 2.731s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.988s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15796844 30.58823679 + layer.0.v_cache 0.00001527 0.00221184 + layer.1.k_cache 0.50235566 2.34641733 + layer.1.v_cache 0.00000626 0.00097328 + layer.2.k_cache 0.02709456 0.72616856 + layer.2.v_cache 0.00002063 0.00267887 + layer.3.k_cache 0.00856834 3.30320125 + layer.3.v_cache 0.00001945 0.00307427 + layer.4.k_cache 0.00071483 0.07939930 + layer.4.v_cache 0.00005464 0.00609366 + layer.4.output 0.00491706 214.77608609 + ------------------------------------------------------------------------------------- + TOTAL 0.04301397 90.61712104 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 510544 +BPFP 1.7509 bits/point +EBPFP 3.5019 equivalent bits/point +MSE 90.617121 +---------------------- -------------------------------------------------------- +Time: 4.731s Load: 0.012s, Pack+Encode: 2.731s, Decode+Unpack: 1.988s +---------------------- -------------------------------------------------------- +💾 Converting with 90.6171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,220B, BPFP=0.7069 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,944B, BPFP=3.0837 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,096B, BPFP=1.5181 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,668B, BPFP=3.0033 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,748B, BPFP=1.6852 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,340B, BPFP=2.9196 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,868B, BPFP=1.5668 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,364B, BPFP=2.9841 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,912B, BPFP=2.3886 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,440B, BPFP=2.8629 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 97,916B, BPFP=0.8813 +⌛️ [2/4] FRONTEND: Frontend time: 2.497s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.604s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13184587 29.65811256 + layer.0.v_cache 0.00001706 0.00234888 + layer.1.k_cache 0.45546043 2.89121369 + layer.1.v_cache 0.00000678 0.00098055 + layer.2.k_cache 0.02272066 0.59003676 + layer.2.v_cache 0.00001996 0.00270604 + layer.3.k_cache 0.01643563 4.08388938 + layer.3.v_cache 0.00002069 0.00312586 + layer.4.k_cache 0.00071781 0.08166790 + layer.4.v_cache 0.00005262 0.00592174 + layer.4.output 1.23455937 150.82879104 + ------------------------------------------------------------------------------------- + TOTAL 0.54524783 64.30126710 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 458516 +BPFP 1.6993 bits/point +EBPFP 3.3986 equivalent bits/point +MSE 64.301267 +---------------------- -------------------------------------------------------- +Time: 4.110s Load: 0.009s, Pack+Encode: 2.497s, Decode+Unpack: 1.604s +---------------------- -------------------------------------------------------- +💾 Converting with 64.3013 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,136B, BPFP=0.7183 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,700B, BPFP=3.4150 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,448B, BPFP=1.4470 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,572B, BPFP=3.2891 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,804B, BPFP=1.7048 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,728B, BPFP=3.2391 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,724B, BPFP=1.5817 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,192B, BPFP=3.3849 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,348B, BPFP=2.6248 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,628B, BPFP=3.2332 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,192B, BPFP=0.8049 +⌛️ [2/4] FRONTEND: Frontend time: 2.798s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 2.026s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17705467 29.73423813 + layer.0.v_cache 0.00001633 0.00217660 + layer.1.k_cache 0.50406676 2.65113923 + layer.1.v_cache 0.00000599 0.00089476 + layer.2.k_cache 0.00944760 0.63664812 + layer.2.v_cache 0.00001883 0.00260363 + layer.3.k_cache 0.04120483 3.43159855 + layer.3.v_cache 0.00002392 0.00311315 + layer.4.k_cache 0.00070100 0.07899225 + layer.4.v_cache 0.00005002 0.00592164 + layer.4.output 0.00499618 181.82457386 + ------------------------------------------------------------------------------------- + TOTAL 0.04515078 77.01878489 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 511472 +BPFP 1.7807 bits/point +EBPFP 3.5614 equivalent bits/point +MSE 77.018785 +---------------------- -------------------------------------------------------- +Time: 4.838s Load: 0.014s, Pack+Encode: 2.798s, Decode+Unpack: 2.026s +---------------------- -------------------------------------------------------- +💾 Converting with 77.0188 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,460B, BPFP=0.7211 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,344B, BPFP=3.3764 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,260B, BPFP=1.6354 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,180B, BPFP=3.2512 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,888B, BPFP=1.7296 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,528B, BPFP=3.2134 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,992B, BPFP=1.5620 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,060B, BPFP=3.3021 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,416B, BPFP=2.6282 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,652B, BPFP=3.1627 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,028B, BPFP=0.7608 +⌛️ [2/4] FRONTEND: Frontend time: 2.559s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.812s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20501094 30.25103624 + layer.0.v_cache 0.00001516 0.00220594 + layer.1.k_cache 0.57085758 2.95637162 + layer.1.v_cache 0.00000574 0.00092482 + layer.2.k_cache 0.01259345 0.67164205 + layer.2.v_cache 0.00002099 0.00273062 + layer.3.k_cache 0.01648921 3.71885851 + layer.3.v_cache 0.00001970 0.00297981 + layer.4.k_cache 0.00070444 0.07810836 + layer.4.v_cache 0.00005121 0.00600941 + layer.4.output 0.00487421 180.57395833 + ------------------------------------------------------------------------------------- + TOTAL 0.04940517 76.57109269 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 516808 +BPFP 1.7593 bits/point +EBPFP 3.5186 equivalent bits/point +MSE 76.571093 +---------------------- -------------------------------------------------------- +Time: 4.380s Load: 0.010s, Pack+Encode: 2.559s, Decode+Unpack: 1.812s +---------------------- -------------------------------------------------------- +💾 Converting with 76.5711 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,568B, BPFP=0.7246 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,348B, BPFP=3.3642 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,168B, BPFP=1.6241 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,096B, BPFP=3.2920 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,564B, BPFP=1.7622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,644B, BPFP=3.2659 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,840B, BPFP=1.6052 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,056B, BPFP=3.3473 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,372B, BPFP=2.5583 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,204B, BPFP=3.1829 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,104B, BPFP=0.8410 +⌛️ [2/4] FRONTEND: Frontend time: 2.786s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.881s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14466163 30.25877104 + layer.0.v_cache 0.00001631 0.00231968 + layer.1.k_cache 0.53789613 2.96104696 + layer.1.v_cache 0.00000653 0.00096809 + layer.2.k_cache 0.02095415 0.76433009 + layer.2.v_cache 0.00002169 0.00279259 + layer.3.k_cache 0.03083320 3.75695711 + layer.3.v_cache 0.00002083 0.00306416 + layer.4.k_cache 0.00069737 0.08229312 + layer.4.v_cache 0.00004943 0.00585991 + layer.4.output 0.00490546 172.18847193 + ------------------------------------------------------------------------------------- + TOTAL 0.04526444 73.12692390 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 530964 +BPFP 1.8008 bits/point +EBPFP 3.6016 equivalent bits/point +MSE 73.126924 +---------------------- -------------------------------------------------------- +Time: 4.680s Load: 0.013s, Pack+Encode: 2.786s, Decode+Unpack: 1.881s +---------------------- -------------------------------------------------------- +💾 Converting with 73.1269 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,392B, BPFP=0.6998 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,932B, BPFP=3.1319 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,540B, BPFP=1.5437 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,168B, BPFP=3.0397 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,872B, BPFP=1.6656 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,420B, BPFP=2.9484 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,756B, BPFP=1.5027 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,720B, BPFP=3.0163 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,044B, BPFP=2.4061 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,052B, BPFP=2.9291 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 108,352B, BPFP=0.8089 +⌛️ [2/4] FRONTEND: Frontend time: 1.970s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14240686 30.38970723 + layer.0.v_cache 0.00001673 0.00220100 + layer.1.k_cache 0.61438330 3.06257145 + layer.1.v_cache 0.00000638 0.00090998 + layer.2.k_cache 0.02027981 0.62991078 + layer.2.v_cache 0.00002012 0.00269077 + layer.3.k_cache 0.02199794 4.18684270 + layer.3.v_cache 0.00002003 0.00292196 + layer.4.k_cache 0.00078937 0.07988960 + layer.4.v_cache 0.00005038 0.00571117 + layer.4.output 0.04464151 163.46315098 + ------------------------------------------------------------------------------------- + TOTAL 0.06543891 69.56502432 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 546248 +BPFP 1.6792 bits/point +EBPFP 3.3583 equivalent bits/point +MSE 69.565024 +---------------------- -------------------------------------------------------- +Time: 3.349s Load: 0.013s, Pack+Encode: 1.970s, Decode+Unpack: 1.366s +---------------------- -------------------------------------------------------- +💾 Converting with 69.5650 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 399, 128) +Output shape: (1, 399, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.output: torch.Size([1, 399, 3584]) -> torch.Size([1, 1, 399, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 17,532B, BPFP=0.6866 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 82,208B, BPFP=3.2193 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 44,772B, BPFP=1.7533 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 79,968B, BPFP=3.1316 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 42,820B, BPFP=1.6768 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 79,360B, BPFP=3.1078 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 39,324B, BPFP=1.5399 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 81,064B, BPFP=3.1745 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 63,000B, BPFP=2.4671 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 77,724B, BPFP=3.0437 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 148,088B, BPFP=0.8285 +⌛️ [2/4] FRONTEND: Frontend time: 2.132s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.596s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17889012 29.52016761 + layer.0.v_cache 0.00001635 0.00214673 + layer.1.k_cache 0.94984295 3.73012763 + layer.1.v_cache 0.00000590 0.00086149 + layer.2.k_cache 0.02305104 0.58393982 + layer.2.v_cache 0.00002114 0.00279574 + layer.3.k_cache 0.01620252 3.64306365 + layer.3.v_cache 0.00002049 0.00299563 + layer.4.k_cache 0.00073175 0.07858460 + layer.4.v_cache 0.00005255 0.00574534 + layer.4.output 0.00656733 116.38052721 + ------------------------------------------------------------------------------------- + TOTAL 0.07145918 50.13141875 + (elements=3,472,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3472896 +Total Bytes 755860 +BPFP 1.7412 bits/point +EBPFP 3.4823 equivalent bits/point +MSE 50.131419 +---------------------- -------------------------------------------------------- +Time: 3.740s Load: 0.012s, Pack+Encode: 2.132s, Decode+Unpack: 1.596s +---------------------- -------------------------------------------------------- +💾 Converting with 50.1314 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,408B, BPFP=0.7196 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,236B, BPFP=3.3349 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,596B, BPFP=1.4239 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,128B, BPFP=3.2583 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,120B, BPFP=1.8059 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,328B, BPFP=3.2030 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,396B, BPFP=1.6175 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,132B, BPFP=3.2586 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,212B, BPFP=2.5727 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,396B, BPFP=3.1386 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,212B, BPFP=0.8811 +⌛️ [2/4] FRONTEND: Frontend time: 1.771s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.266s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10670873 29.74945987 + layer.0.v_cache 0.00001672 0.00245973 + layer.1.k_cache 0.29393762 2.52159834 + layer.1.v_cache 0.00000627 0.00102176 + layer.2.k_cache 0.01243114 0.51034033 + layer.2.v_cache 0.00002105 0.00301012 + layer.3.k_cache 0.00919804 3.60181798 + layer.3.v_cache 0.00002002 0.00331300 + layer.4.k_cache 0.00070102 0.08433119 + layer.4.v_cache 0.00005024 0.00647745 + layer.4.output 1.35463108 217.52082016 + ------------------------------------------------------------------------------------- + TOTAL 0.58267697 91.71350417 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 441164 +BPFP 1.7942 bits/point +EBPFP 3.5883 equivalent bits/point +MSE 91.713504 +---------------------- -------------------------------------------------------- +Time: 3.046s Load: 0.008s, Pack+Encode: 1.771s, Decode+Unpack: 1.266s +---------------------- -------------------------------------------------------- +💾 Converting with 91.7135 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,288B, BPFP=0.6944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,400B, BPFP=3.0389 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,232B, BPFP=1.3676 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,040B, BPFP=2.8937 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,404B, BPFP=1.6243 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,020B, BPFP=2.8310 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,540B, BPFP=1.5096 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,204B, BPFP=2.9038 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,992B, BPFP=2.3986 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,216B, BPFP=2.7815 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,224B, BPFP=0.8105 +⌛️ [2/4] FRONTEND: Frontend time: 1.772s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.253s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633087 36.35772945 + layer.0.v_cache 0.00001891 0.00236308 + layer.1.k_cache 0.36012592 2.85552882 + layer.1.v_cache 0.00000581 0.00089778 + layer.2.k_cache 0.01407751 0.55605827 + layer.2.v_cache 0.00002023 0.00283304 + layer.3.k_cache 0.01230825 4.53031573 + layer.3.v_cache 0.00002000 0.00306152 + layer.4.k_cache 0.00071033 0.08064080 + layer.4.v_cache 0.00004856 0.00549776 + layer.4.output 1.20534739 187.70797596 + ------------------------------------------------------------------------------------- + TOTAL 0.52535871 79.90298576 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 450560 +BPFP 1.6304 bits/point +EBPFP 3.2608 equivalent bits/point +MSE 79.902986 +---------------------- -------------------------------------------------------- +Time: 3.033s Load: 0.008s, Pack+Encode: 1.772s, Decode+Unpack: 1.253s +---------------------- -------------------------------------------------------- +💾 Converting with 79.9030 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,972B, BPFP=0.7114 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,484B, BPFP=3.1434 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,960B, BPFP=1.4238 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,040B, BPFP=3.0498 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,560B, BPFP=1.7220 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,012B, BPFP=2.9831 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,860B, BPFP=1.6118 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,684B, BPFP=3.0267 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,816B, BPFP=2.4518 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,160B, BPFP=2.9279 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,048B, BPFP=0.8155 +⌛️ [2/4] FRONTEND: Frontend time: 1.764s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.248s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11142051 32.11098369 + layer.0.v_cache 0.00001650 0.00232267 + layer.1.k_cache 0.42032433 2.64317404 + layer.1.v_cache 0.00000619 0.00093865 + layer.2.k_cache 0.01443043 0.58308404 + layer.2.v_cache 0.00002030 0.00284664 + layer.3.k_cache 0.02895402 4.45270206 + layer.3.v_cache 0.00001914 0.00308900 + layer.4.k_cache 0.00067703 0.08180462 + layer.4.v_cache 0.00005693 0.00626090 + layer.4.output 1.27033357 80.00117627 + ------------------------------------------------------------------------------------- + TOTAL 0.55695649 35.28796707 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 443596 +BPFP 1.6918 bits/point +EBPFP 3.3835 equivalent bits/point +MSE 35.287967 +---------------------- -------------------------------------------------------- +Time: 3.020s Load: 0.008s, Pack+Encode: 1.764s, Decode+Unpack: 1.248s +---------------------- -------------------------------------------------------- +💾 Converting with 35.2880 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 302, 128) +Output shape: (1, 302, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.output: torch.Size([1, 302, 3584]) -> torch.Size([1, 1, 302, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,540B, BPFP=0.7005 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,880B, BPFP=3.0981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,712B, BPFP=1.5373 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,456B, BPFP=3.0244 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,964B, BPFP=1.7055 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,124B, BPFP=2.9555 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,548B, BPFP=1.5805 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,364B, BPFP=3.0197 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,368B, BPFP=2.4507 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,088B, BPFP=2.9019 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,660B, BPFP=0.7440 +⌛️ [2/4] FRONTEND: Frontend time: 1.882s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15012774 29.33583079 + layer.0.v_cache 0.00001504 0.00220603 + layer.1.k_cache 0.60100308 3.12361600 + layer.1.v_cache 0.00000575 0.00089798 + layer.2.k_cache 0.00814273 0.55442098 + layer.2.v_cache 0.00002032 0.00280724 + layer.3.k_cache 0.01797006 4.24282089 + layer.3.v_cache 0.00001899 0.00297629 + layer.4.k_cache 0.00070977 0.07915779 + layer.4.v_cache 0.00004868 0.00577754 + layer.4.output 0.04414053 162.80672008 + ------------------------------------------------------------------------------------- + TOTAL 0.06394387 69.23515012 + (elements=2,628,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2628608 +Total Bytes 544704 +BPFP 1.6578 bits/point +EBPFP 3.3155 equivalent bits/point +MSE 69.235150 +---------------------- -------------------------------------------------------- +Time: 3.275s Load: 0.010s, Pack+Encode: 1.882s, Decode+Unpack: 1.383s +---------------------- -------------------------------------------------------- +💾 Converting with 69.2352 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,436B, BPFP=0.7250 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,116B, BPFP=3.3300 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,808B, BPFP=1.5047 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,812B, BPFP=3.2540 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,356B, BPFP=1.7115 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,016B, BPFP=3.2076 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,812B, BPFP=1.5632 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,080B, BPFP=3.3279 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,544B, BPFP=2.5387 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,852B, BPFP=3.1980 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,448B, BPFP=0.8449 +⌛️ [2/4] FRONTEND: Frontend time: 1.890s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14350806 30.70111357 + layer.0.v_cache 0.00001655 0.00217466 + layer.1.k_cache 0.46812200 2.77969269 + layer.1.v_cache 0.00000598 0.00091773 + layer.2.k_cache 0.01096032 0.59481345 + layer.2.v_cache 0.00002019 0.00274203 + layer.3.k_cache 0.00658789 3.79089492 + layer.3.v_cache 0.00002014 0.00297295 + layer.4.k_cache 0.00067528 0.07970002 + layer.4.v_cache 0.00005785 0.00614960 + layer.4.output 0.00496594 173.20670642 + ------------------------------------------------------------------------------------- + TOTAL 0.03910211 73.55341862 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 519280 +BPFP 1.7809 bits/point +EBPFP 3.5618 equivalent bits/point +MSE 73.553419 +---------------------- -------------------------------------------------------- +Time: 3.275s Load: 0.010s, Pack+Encode: 1.890s, Decode+Unpack: 1.375s +---------------------- -------------------------------------------------------- +💾 Converting with 73.5534 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 260, 128) +Output shape: (1, 260, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.output: torch.Size([1, 260, 3584]) -> torch.Size([1, 1, 260, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,980B, BPFP=0.7200 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,608B, BPFP=3.4019 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,508B, BPFP=1.4728 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 54,360B, BPFP=3.2668 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,808B, BPFP=1.7312 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,940B, BPFP=3.2416 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,132B, BPFP=1.5704 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,804B, BPFP=3.3536 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,964B, BPFP=2.5820 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,464B, BPFP=3.2130 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,804B, BPFP=0.7967 +⌛️ [2/4] FRONTEND: Frontend time: 1.884s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14548651 29.34910983 + layer.0.v_cache 0.00001584 0.00228828 + layer.1.k_cache 0.47062912 2.74453266 + layer.1.v_cache 0.00000604 0.00096459 + layer.2.k_cache 0.00619981 0.61060773 + layer.2.v_cache 0.00002003 0.00294558 + layer.3.k_cache 0.01299995 3.46789645 + layer.3.v_cache 0.00001981 0.00316970 + layer.4.k_cache 0.00067777 0.07966311 + layer.4.v_cache 0.00006192 0.00636326 + layer.4.output 0.00504520 293.78049451 + ------------------------------------------------------------------------------------- + TOTAL 0.03949607 123.10182369 + (elements=2,263,040) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2263040 +Total Bytes 501372 +BPFP 1.7724 bits/point +EBPFP 3.5448 equivalent bits/point +MSE 123.101824 +---------------------- -------------------------------------------------------- +Time: 3.285s Load: 0.011s, Pack+Encode: 1.884s, Decode+Unpack: 1.390s +---------------------- -------------------------------------------------------- +💾 Converting with 123.1018 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,568B, BPFP=0.7220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,480B, BPFP=3.3594 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,604B, BPFP=1.5857 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,072B, BPFP=3.2785 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,320B, BPFP=1.7417 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,272B, BPFP=3.2325 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,628B, BPFP=1.5871 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,792B, BPFP=3.3199 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,100B, BPFP=2.5333 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,020B, BPFP=3.2181 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,664B, BPFP=0.8179 +⌛️ [2/4] FRONTEND: Frontend time: 1.886s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14555165 29.75224573 + layer.0.v_cache 0.00001661 0.00228973 + layer.1.k_cache 0.55468144 2.96493800 + layer.1.v_cache 0.00000621 0.00093358 + layer.2.k_cache 0.02243149 0.70148939 + layer.2.v_cache 0.00002103 0.00273515 + layer.3.k_cache 0.02028967 3.66097686 + layer.3.v_cache 0.00002036 0.00301093 + layer.4.k_cache 0.00072975 0.08204968 + layer.4.v_cache 0.00005124 0.00588575 + layer.4.output 0.00486695 177.89106815 + ------------------------------------------------------------------------------------- + TOTAL 0.04575695 75.43611952 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 527520 +BPFP 1.7825 bits/point +EBPFP 3.5651 equivalent bits/point +MSE 75.436120 +---------------------- -------------------------------------------------------- +Time: 3.282s Load: 0.009s, Pack+Encode: 1.886s, Decode+Unpack: 1.386s +---------------------- -------------------------------------------------------- +💾 Converting with 75.4361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 356, 128) +Output shape: (1, 356, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.output: torch.Size([1, 356, 3584]) -> torch.Size([1, 1, 356, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,744B, BPFP=0.6910 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 72,136B, BPFP=3.1661 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 34,668B, BPFP=1.5216 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 70,068B, BPFP=3.0753 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 37,620B, BPFP=1.6512 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 68,080B, BPFP=2.9881 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 33,856B, BPFP=1.4860 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 69,412B, BPFP=3.0465 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 54,604B, BPFP=2.3966 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 67,564B, BPFP=2.9654 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 128,844B, BPFP=0.8079 +⌛️ [2/4] FRONTEND: Frontend time: 2.009s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.505s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15223843 31.89673400 + layer.0.v_cache 0.00001740 0.00213940 + layer.1.k_cache 0.82972649 3.03700342 + layer.1.v_cache 0.00000631 0.00089995 + layer.2.k_cache 0.03086275 0.61764269 + layer.2.v_cache 0.00001960 0.00261036 + layer.3.k_cache 0.01100341 4.18127201 + layer.3.v_cache 0.00002062 0.00283424 + layer.4.k_cache 0.00073857 0.08072109 + layer.4.v_cache 0.00005279 0.00569711 + layer.4.output 0.03758925 127.29804123 + ------------------------------------------------------------------------------------- + TOTAL 0.07575360 54.75963782 + (elements=3,098,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3098624 +Total Bytes 652596 +BPFP 1.6849 bits/point +EBPFP 3.3697 equivalent bits/point +MSE 54.759638 +---------------------- -------------------------------------------------------- +Time: 3.525s Load: 0.011s, Pack+Encode: 2.009s, Decode+Unpack: 1.505s +---------------------- -------------------------------------------------------- +💾 Converting with 54.7596 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,288B, BPFP=0.7218 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,996B, BPFP=3.4067 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,052B, BPFP=1.4128 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,008B, BPFP=3.2899 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,160B, BPFP=1.7716 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,104B, BPFP=3.2956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,460B, BPFP=1.5543 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,816B, BPFP=3.3961 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,968B, BPFP=2.5827 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,264B, BPFP=3.2462 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 104,584B, BPFP=0.8776 +⌛️ [2/4] FRONTEND: Frontend time: 1.887s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12347319 28.14513518 + layer.0.v_cache 0.00001696 0.00225259 + layer.1.k_cache 0.44284580 2.39610073 + layer.1.v_cache 0.00000660 0.00095511 + layer.2.k_cache 0.00882101 0.64553202 + layer.2.v_cache 0.00002210 0.00292381 + layer.3.k_cache 0.01145097 3.63432209 + layer.3.v_cache 0.00001984 0.00301133 + layer.4.k_cache 0.00068754 0.07911347 + layer.4.v_cache 0.00006483 0.00616246 + layer.4.output 0.00500964 187.87343918 + ------------------------------------------------------------------------------------- + TOTAL 0.03661626 79.41350488 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 524700 +BPFP 1.8130 bits/point +EBPFP 3.6260 equivalent bits/point +MSE 79.413505 +---------------------- -------------------------------------------------------- +Time: 3.273s Load: 0.011s, Pack+Encode: 1.887s, Decode+Unpack: 1.376s +---------------------- -------------------------------------------------------- +💾 Converting with 79.4135 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,856B, BPFP=0.7048 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,876B, BPFP=3.2827 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,444B, BPFP=1.6691 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,332B, BPFP=3.1980 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,964B, BPFP=1.7524 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,892B, BPFP=3.1191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,792B, BPFP=1.5785 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,204B, BPFP=3.1910 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,828B, BPFP=2.5125 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,636B, BPFP=3.1050 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 105,184B, BPFP=0.8238 +⌛️ [2/4] FRONTEND: Frontend time: 1.877s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13002123 29.42322163 + layer.0.v_cache 0.00001833 0.00235219 + layer.1.k_cache 0.50397591 3.25676955 + layer.1.v_cache 0.00000608 0.00093131 + layer.2.k_cache 0.01926375 0.55485481 + layer.2.v_cache 0.00002060 0.00277209 + layer.3.k_cache 0.04992905 3.99272761 + layer.3.v_cache 0.00002096 0.00313325 + layer.4.k_cache 0.00069540 0.08016214 + layer.4.v_cache 0.00005467 0.00616277 + layer.4.output 0.00466491 172.00805138 + ------------------------------------------------------------------------------------- + TOTAL 0.04333296 73.02232041 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 545008 +BPFP 1.7576 bits/point +EBPFP 3.5153 equivalent bits/point +MSE 73.022320 +---------------------- -------------------------------------------------------- +Time: 3.266s Load: 0.011s, Pack+Encode: 1.877s, Decode+Unpack: 1.378s +---------------------- -------------------------------------------------------- +💾 Converting with 73.0223 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,820B, BPFP=0.7154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,892B, BPFP=3.2864 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,740B, BPFP=1.5480 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,184B, BPFP=3.1911 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,400B, BPFP=1.6964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,328B, BPFP=3.0875 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,588B, BPFP=1.5395 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,264B, BPFP=3.2513 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,100B, BPFP=2.4609 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,752B, BPFP=3.1112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 110,336B, BPFP=0.8796 +⌛️ [2/4] FRONTEND: Frontend time: 1.879s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15974760 29.85539551 + layer.0.v_cache 0.00001739 0.00219640 + layer.1.k_cache 0.45644542 3.05074594 + layer.1.v_cache 0.00000643 0.00088108 + layer.2.k_cache 0.01631587 0.64342172 + layer.2.v_cache 0.00002005 0.00263142 + layer.3.k_cache 0.01511352 3.72560904 + layer.3.v_cache 0.00002063 0.00303160 + layer.4.k_cache 0.00073517 0.07940769 + layer.4.v_cache 0.00005091 0.00589335 + layer.4.output 0.00481773 178.00656888 + ------------------------------------------------------------------------------------- + TOTAL 0.04012924 75.49501152 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 538404 +BPFP 1.7673 bits/point +EBPFP 3.5347 equivalent bits/point +MSE 75.495012 +---------------------- -------------------------------------------------------- +Time: 3.267s Load: 0.009s, Pack+Encode: 1.879s, Decode+Unpack: 1.379s +---------------------- -------------------------------------------------------- +💾 Converting with 75.4950 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,308B, BPFP=0.7203 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,552B, BPFP=3.3680 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,912B, BPFP=1.5164 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,048B, BPFP=3.2800 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,792B, BPFP=1.7434 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,792B, BPFP=3.2065 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,352B, BPFP=1.6007 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,212B, BPFP=3.3481 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,784B, BPFP=2.5623 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,848B, BPFP=3.2097 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,916B, BPFP=0.7851 +⌛️ [2/4] FRONTEND: Frontend time: 1.877s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14148096 29.12641912 + layer.0.v_cache 0.00001588 0.00223544 + layer.1.k_cache 0.55266568 2.92149873 + layer.1.v_cache 0.00000632 0.00092927 + layer.2.k_cache 0.00687984 0.72856494 + layer.2.v_cache 0.00001885 0.00269544 + layer.3.k_cache 0.01675770 3.68726615 + layer.3.v_cache 0.00001964 0.00298569 + layer.4.k_cache 0.00068991 0.07842108 + layer.4.v_cache 0.00004859 0.00591449 + layer.4.output 0.00492290 282.53738630 + ------------------------------------------------------------------------------------- + TOTAL 0.04429669 118.48933144 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 513516 +BPFP 1.7677 bits/point +EBPFP 3.5354 equivalent bits/point +MSE 118.489331 +---------------------- -------------------------------------------------------- +Time: 3.263s Load: 0.009s, Pack+Encode: 1.877s, Decode+Unpack: 1.378s +---------------------- -------------------------------------------------------- +💾 Converting with 118.4893 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,312B, BPFP=0.6923 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 69,616B, BPFP=3.3676 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 33,772B, BPFP=1.6337 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 66,356B, BPFP=3.2099 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 34,708B, BPFP=1.6790 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 65,580B, BPFP=3.1724 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,348B, BPFP=1.5164 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 67,400B, BPFP=3.2604 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 51,036B, BPFP=2.4688 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 64,412B, BPFP=3.1159 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 115,004B, BPFP=0.7948 +⌛️ [2/4] FRONTEND: Frontend time: 2.021s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.499s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13144862 31.49746638 + layer.0.v_cache 0.00001769 0.00233991 + layer.1.k_cache 0.64513603 3.06216478 + layer.1.v_cache 0.00000607 0.00092148 + layer.2.k_cache 0.02726801 0.71227862 + layer.2.v_cache 0.00001956 0.00271868 + layer.3.k_cache 0.02271994 3.80171284 + layer.3.v_cache 0.00002018 0.00296663 + layer.4.k_cache 0.00073057 0.08042134 + layer.4.v_cache 0.00004940 0.00581478 + layer.4.output 0.04135688 140.01599126 + ------------------------------------------------------------------------------------- + TOTAL 0.06570084 59.95769084 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 613544 +BPFP 1.7459 bits/point +EBPFP 3.4918 equivalent bits/point +MSE 59.957691 +---------------------- -------------------------------------------------------- +Time: 3.532s Load: 0.012s, Pack+Encode: 2.021s, Decode+Unpack: 1.499s +---------------------- -------------------------------------------------------- +💾 Converting with 59.9577 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,208B, BPFP=0.7419 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,872B, BPFP=3.4064 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,712B, BPFP=1.4326 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,600B, BPFP=3.3140 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,020B, BPFP=1.8183 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,152B, BPFP=3.2814 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,644B, BPFP=1.6456 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,208B, BPFP=3.3581 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,220B, BPFP=2.6323 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,568B, BPFP=3.2390 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,572B, BPFP=0.7846 +⌛️ [2/4] FRONTEND: Frontend time: 1.757s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.248s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12396691 29.71874319 + layer.0.v_cache 0.00001598 0.00238537 + layer.1.k_cache 0.24854917 2.49071244 + layer.1.v_cache 0.00000584 0.00094244 + layer.2.k_cache 0.02290044 0.73140387 + layer.2.v_cache 0.00002299 0.00284267 + layer.3.k_cache 0.01294550 3.40249364 + layer.3.v_cache 0.00002077 0.00315531 + layer.4.k_cache 0.00065144 0.08122742 + layer.4.v_cache 0.00005054 0.00597620 + layer.4.output 1.42385976 229.13471761 + ------------------------------------------------------------------------------------- + TOTAL 0.61036164 96.49311211 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 417776 +BPFP 1.7860 bits/point +EBPFP 3.5720 equivalent bits/point +MSE 96.493112 +---------------------- -------------------------------------------------------- +Time: 3.013s Load: 0.009s, Pack+Encode: 1.757s, Decode+Unpack: 1.248s +---------------------- -------------------------------------------------------- +💾 Converting with 96.4931 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,748B, BPFP=0.6892 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,688B, BPFP=3.2271 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,748B, BPFP=1.3921 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,080B, BPFP=3.1401 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,808B, BPFP=1.7197 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,992B, BPFP=3.0813 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,196B, BPFP=1.5785 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,612B, BPFP=3.1689 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,124B, BPFP=2.4937 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,172B, BPFP=3.0370 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 106,300B, BPFP=0.8210 +⌛️ [2/4] FRONTEND: Frontend time: 1.879s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14783513 30.86441325 + layer.0.v_cache 0.00001600 0.00233505 + layer.1.k_cache 0.54442752 2.54440234 + layer.1.v_cache 0.00000628 0.00095240 + layer.2.k_cache 0.01848349 0.70055239 + layer.2.v_cache 0.00002061 0.00278759 + layer.3.k_cache 0.01777397 3.96096960 + layer.3.v_cache 0.00002092 0.00307643 + layer.4.k_cache 0.00070447 0.08242540 + layer.4.v_cache 0.00005049 0.00594410 + layer.4.output 0.04615090 171.63406760 + ------------------------------------------------------------------------------------- + TOTAL 0.06190560 72.91801951 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 541468 +BPFP 1.7221 bits/point +EBPFP 3.4441 equivalent bits/point +MSE 72.918020 +---------------------- -------------------------------------------------------- +Time: 3.260s Load: 0.011s, Pack+Encode: 1.879s, Decode+Unpack: 1.370s +---------------------- -------------------------------------------------------- +💾 Converting with 72.9180 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,356B, BPFP=0.7003 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,572B, BPFP=3.1760 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,852B, BPFP=1.6177 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,620B, BPFP=3.0736 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,512B, BPFP=1.7047 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,256B, BPFP=3.0021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,452B, BPFP=1.5443 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,544B, BPFP=3.0696 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,488B, BPFP=2.4375 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,452B, BPFP=2.9599 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 105,568B, BPFP=0.7907 +⌛️ [2/4] FRONTEND: Frontend time: 1.886s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16274841 30.34360253 + layer.0.v_cache 0.00001794 0.00234397 + layer.1.k_cache 0.56737488 3.05330888 + layer.1.v_cache 0.00000586 0.00089987 + layer.2.k_cache 0.02027899 0.60960357 + layer.2.v_cache 0.00002117 0.00274034 + layer.3.k_cache 0.01891868 4.25861620 + layer.3.v_cache 0.00002140 0.00297136 + layer.4.k_cache 0.00069658 0.07955094 + layer.4.v_cache 0.00005005 0.00575100 + layer.4.output 0.04476916 128.57138363 + ------------------------------------------------------------------------------------- + TOTAL 0.06373636 55.19759259 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 549672 +BPFP 1.6953 bits/point +EBPFP 3.3907 equivalent bits/point +MSE 55.197593 +---------------------- -------------------------------------------------------- +Time: 3.263s Load: 0.010s, Pack+Encode: 1.886s, Decode+Unpack: 1.368s +---------------------- -------------------------------------------------------- +💾 Converting with 55.1976 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,396B, BPFP=0.7095 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,596B, BPFP=3.3537 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,848B, BPFP=1.6511 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,512B, BPFP=3.2344 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,468B, BPFP=1.7438 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,584B, BPFP=3.1813 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,788B, BPFP=1.5332 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,012B, BPFP=3.2630 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,784B, BPFP=2.5060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,128B, BPFP=3.1552 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,732B, BPFP=0.7827 +⌛️ [2/4] FRONTEND: Frontend time: 1.876s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13366390 29.07126223 + layer.0.v_cache 0.00001719 0.00231671 + layer.1.k_cache 0.57618574 3.30988890 + layer.1.v_cache 0.00000658 0.00093278 + layer.2.k_cache 0.02858028 0.73636423 + layer.2.v_cache 0.00002021 0.00278579 + layer.3.k_cache 0.00873641 3.62005056 + layer.3.v_cache 0.00002023 0.00314966 + layer.4.k_cache 0.00067337 0.07981033 + layer.4.v_cache 0.00005011 0.00596622 + layer.4.output 0.00484872 71.11272730 + ------------------------------------------------------------------------------------- + TOTAL 0.04599383 31.44833050 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 520848 +BPFP 1.7536 bits/point +EBPFP 3.5071 equivalent bits/point +MSE 31.448331 +---------------------- -------------------------------------------------------- +Time: 3.260s Load: 0.010s, Pack+Encode: 1.876s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 31.4483 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,304B, BPFP=0.7285 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,920B, BPFP=3.3880 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,396B, BPFP=1.5834 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,252B, BPFP=3.2701 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,016B, BPFP=1.8394 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,960B, BPFP=3.2494 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,232B, BPFP=1.6425 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,148B, BPFP=3.3334 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,160B, BPFP=2.6273 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,584B, BPFP=3.2229 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,576B, BPFP=0.8946 +⌛️ [2/4] FRONTEND: Frontend time: 1.763s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.237s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13057878 30.36842256 + layer.0.v_cache 0.00001797 0.00253423 + layer.1.k_cache 0.31658528 2.81536313 + layer.1.v_cache 0.00000647 0.00099219 + layer.2.k_cache 0.02849985 0.69647203 + layer.2.v_cache 0.00001991 0.00283800 + layer.3.k_cache 0.03599559 3.83089948 + layer.3.v_cache 0.00002239 0.00335174 + layer.4.k_cache 0.00078811 0.08544285 + layer.4.v_cache 0.00005540 0.00629769 + layer.4.output 1.38530015 224.30114738 + ------------------------------------------------------------------------------------- + TOTAL 0.60056887 94.58356739 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 440548 +BPFP 1.8322 bits/point +EBPFP 3.6644 equivalent bits/point +MSE 94.583567 +---------------------- -------------------------------------------------------- +Time: 3.009s Load: 0.009s, Pack+Encode: 1.763s, Decode+Unpack: 1.237s +---------------------- -------------------------------------------------------- +💾 Converting with 94.5836 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,176B, BPFP=0.7327 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,472B, BPFP=3.4182 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,548B, BPFP=1.4796 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,256B, BPFP=3.3306 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,544B, BPFP=1.8393 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,568B, BPFP=3.2811 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,732B, BPFP=1.6368 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,116B, BPFP=3.3926 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,920B, BPFP=2.6584 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,104B, BPFP=3.2477 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,096B, BPFP=0.8856 +⌛️ [2/4] FRONTEND: Frontend time: 1.747s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.236s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12725546 28.98212486 + layer.0.v_cache 0.00001736 0.00251788 + layer.1.k_cache 0.32262681 2.75901548 + layer.1.v_cache 0.00000608 0.00097363 + layer.2.k_cache 0.02157613 0.69645670 + layer.2.v_cache 0.00002208 0.00297739 + layer.3.k_cache 0.01760322 3.68094449 + layer.3.v_cache 0.00002192 0.00343411 + layer.4.k_cache 0.00069785 0.08452225 + layer.4.v_cache 0.00005172 0.00620351 + layer.4.output 1.41081570 237.94616113 + ------------------------------------------------------------------------------------- + TOTAL 0.60974050 100.10837048 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 433532 +BPFP 1.8363 bits/point +EBPFP 3.6725 equivalent bits/point +MSE 100.108370 +---------------------- -------------------------------------------------------- +Time: 2.991s Load: 0.008s, Pack+Encode: 1.747s, Decode+Unpack: 1.236s +---------------------- -------------------------------------------------------- +💾 Converting with 100.1084 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,040B, BPFP=0.7365 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,636B, BPFP=3.4944 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,512B, BPFP=1.5047 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,892B, BPFP=3.3665 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,600B, BPFP=1.8779 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,844B, BPFP=3.3630 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,108B, BPFP=1.6951 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,464B, BPFP=3.4085 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,016B, BPFP=2.6420 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,020B, BPFP=3.3025 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,372B, BPFP=0.9156 +⌛️ [2/4] FRONTEND: Frontend time: 1.750s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.237s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13788442 29.49776720 + layer.0.v_cache 0.00001717 0.00261297 + layer.1.k_cache 0.30723858 2.77668039 + layer.1.v_cache 0.00000664 0.00104372 + layer.2.k_cache 0.01337219 0.66985991 + layer.2.v_cache 0.00002159 0.00313525 + layer.3.k_cache 0.03522867 3.61753537 + layer.3.v_cache 0.00002126 0.00335883 + layer.4.k_cache 0.00073009 0.08485497 + layer.4.v_cache 0.00005119 0.00639125 + layer.4.output 1.43730973 220.48933182 + ------------------------------------------------------------------------------------- + TOTAL 0.62092588 92.94638604 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 433504 +BPFP 1.8706 bits/point +EBPFP 3.7412 equivalent bits/point +MSE 92.946386 +---------------------- -------------------------------------------------------- +Time: 2.994s Load: 0.007s, Pack+Encode: 1.750s, Decode+Unpack: 1.237s +---------------------- -------------------------------------------------------- +💾 Converting with 92.9464 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,796B, BPFP=0.7503 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,516B, BPFP=3.5628 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,948B, BPFP=1.4513 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,152B, BPFP=3.4583 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,604B, BPFP=1.8845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,384B, BPFP=3.3995 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,392B, BPFP=1.7151 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,768B, BPFP=3.5055 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,512B, BPFP=2.7966 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,840B, BPFP=3.3578 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,404B, BPFP=0.9345 +⌛️ [2/4] FRONTEND: Frontend time: 1.749s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.239s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11546621 28.63432521 + layer.0.v_cache 0.00001823 0.00251763 + layer.1.k_cache 0.26924565 1.56416291 + layer.1.v_cache 0.00000737 0.00105591 + layer.2.k_cache 0.01761903 0.71535829 + layer.2.v_cache 0.00002145 0.00295106 + layer.3.k_cache 0.02374693 3.79465559 + layer.3.v_cache 0.00002104 0.00321512 + layer.4.k_cache 0.00067272 0.08345353 + layer.4.v_cache 0.00005153 0.00620307 + layer.4.output 1.50071327 223.92329744 + ------------------------------------------------------------------------------------- + TOTAL 0.64305077 94.25123414 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 423316 +BPFP 1.9072 bits/point +EBPFP 3.8145 equivalent bits/point +MSE 94.251234 +---------------------- -------------------------------------------------------- +Time: 2.995s Load: 0.007s, Pack+Encode: 1.749s, Decode+Unpack: 1.239s +---------------------- -------------------------------------------------------- +💾 Converting with 94.2512 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,632B, BPFP=0.6916 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 61,064B, BPFP=3.0978 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 34,052B, BPFP=1.7275 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 59,160B, BPFP=3.0012 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,440B, BPFP=1.6964 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,900B, BPFP=2.9373 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,632B, BPFP=1.5540 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,840B, BPFP=2.9850 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,812B, BPFP=2.4255 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,584B, BPFP=2.8705 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 115,596B, BPFP=0.8377 +⌛️ [2/4] FRONTEND: Frontend time: 1.883s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18206066 32.72122311 + layer.0.v_cache 0.00001652 0.00232932 + layer.1.k_cache 0.62405802 3.45473926 + layer.1.v_cache 0.00000616 0.00091893 + layer.2.k_cache 0.02036299 0.53473391 + layer.2.v_cache 0.00002087 0.00279258 + layer.3.k_cache 0.02546444 4.35907815 + layer.3.v_cache 0.00002064 0.00299216 + layer.4.k_cache 0.00069880 0.08080076 + layer.4.v_cache 0.00007146 0.00610991 + layer.4.output 0.04336799 58.98191457 + ------------------------------------------------------------------------------------- + TOTAL 0.06802097 26.70818353 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 568712 +BPFP 1.6971 bits/point +EBPFP 3.3942 equivalent bits/point +MSE 26.708184 +---------------------- -------------------------------------------------------- +Time: 3.270s Load: 0.010s, Pack+Encode: 1.883s, Decode+Unpack: 1.376s +---------------------- -------------------------------------------------------- +💾 Converting with 26.7082 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,596B, BPFP=0.7230 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,828B, BPFP=3.3316 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,432B, BPFP=1.5988 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,672B, BPFP=3.2527 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,240B, BPFP=1.7904 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,284B, BPFP=3.1580 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,404B, BPFP=1.5969 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,356B, BPFP=3.2312 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,016B, BPFP=2.5257 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,488B, BPFP=3.1037 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,652B, BPFP=0.8739 +⌛️ [2/4] FRONTEND: Frontend time: 1.755s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.235s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14018573 29.50775067 + layer.0.v_cache 0.00001653 0.00233383 + layer.1.k_cache 0.40671090 3.05378450 + layer.1.v_cache 0.00000662 0.00095386 + layer.2.k_cache 0.03036383 0.60594244 + layer.2.v_cache 0.00002088 0.00273913 + layer.3.k_cache 0.05940984 3.91504439 + layer.3.v_cache 0.00002025 0.00294605 + layer.4.k_cache 0.00071161 0.08161118 + layer.4.v_cache 0.00005094 0.00580367 + layer.4.output 1.33692960 209.92533531 + ------------------------------------------------------------------------------------- + TOTAL 0.58800025 88.62683864 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 445968 +BPFP 1.7899 bits/point +EBPFP 3.5799 equivalent bits/point +MSE 88.626839 +---------------------- -------------------------------------------------------- +Time: 2.999s Load: 0.009s, Pack+Encode: 1.755s, Decode+Unpack: 1.235s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6268 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,324B, BPFP=0.7366 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,708B, BPFP=3.4038 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,100B, BPFP=1.5768 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,436B, BPFP=3.3131 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,664B, BPFP=1.8311 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,692B, BPFP=3.2600 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,092B, BPFP=1.6475 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,392B, BPFP=3.3099 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,732B, BPFP=2.6207 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,136B, BPFP=3.2203 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,940B, BPFP=0.8963 +⌛️ [2/4] FRONTEND: Frontend time: 1.755s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.239s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351684 28.65787537 + layer.0.v_cache 0.00001872 0.00261988 + layer.1.k_cache 0.28219827 2.82735635 + layer.1.v_cache 0.00000627 0.00102193 + layer.2.k_cache 0.01639774 0.58466489 + layer.2.v_cache 0.00002172 0.00314433 + layer.3.k_cache 0.02752649 3.67697910 + layer.3.v_cache 0.00002090 0.00331581 + layer.4.k_cache 0.00067264 0.08332657 + layer.4.v_cache 0.00005058 0.00634798 + layer.4.output 1.39794763 213.53445042 + ------------------------------------------------------------------------------------- + TOTAL 0.60212139 90.03457678 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 437216 +BPFP 1.8349 bits/point +EBPFP 3.6699 equivalent bits/point +MSE 90.034577 +---------------------- -------------------------------------------------------- +Time: 3.002s Load: 0.007s, Pack+Encode: 1.755s, Decode+Unpack: 1.239s +---------------------- -------------------------------------------------------- +💾 Converting with 90.0346 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,352B, BPFP=0.7201 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,368B, BPFP=3.4030 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,592B, BPFP=1.4921 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,752B, BPFP=3.3088 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,852B, BPFP=1.7404 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,404B, BPFP=3.2302 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,928B, BPFP=1.5700 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,796B, BPFP=3.3696 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,644B, BPFP=2.5445 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,060B, BPFP=3.2101 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,212B, BPFP=0.8513 +⌛️ [2/4] FRONTEND: Frontend time: 1.870s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15646637 29.53478822 + layer.0.v_cache 0.00001727 0.00237610 + layer.1.k_cache 0.53079696 2.64552535 + layer.1.v_cache 0.00000635 0.00098233 + layer.2.k_cache 0.02340859 0.72691869 + layer.2.v_cache 0.00002061 0.00282810 + layer.3.k_cache 0.01050839 3.69852345 + layer.3.v_cache 0.00002156 0.00306067 + layer.4.k_cache 0.00068596 0.08094902 + layer.4.v_cache 0.00006072 0.00601152 + layer.4.output 0.00497423 175.88472814 + ------------------------------------------------------------------------------------- + TOTAL 0.04451838 74.58206238 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 523960 +BPFP 1.7969 bits/point +EBPFP 3.5939 equivalent bits/point +MSE 74.582062 +---------------------- -------------------------------------------------------- +Time: 3.248s Load: 0.010s, Pack+Encode: 1.870s, Decode+Unpack: 1.368s +---------------------- -------------------------------------------------------- +💾 Converting with 74.5821 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,640B, BPFP=0.6920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,900B, BPFP=3.0895 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 32,052B, BPFP=1.6260 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 59,284B, BPFP=3.0075 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,528B, BPFP=1.7009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,628B, BPFP=2.9235 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,196B, BPFP=1.5319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,600B, BPFP=2.9728 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,204B, BPFP=2.3947 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,364B, BPFP=2.8594 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 118,140B, BPFP=0.8562 +⌛️ [2/4] FRONTEND: Frontend time: 1.877s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15292851 33.61286272 + layer.0.v_cache 0.00001614 0.00242219 + layer.1.k_cache 0.61736927 3.25439433 + layer.1.v_cache 0.00000669 0.00101637 + layer.2.k_cache 0.02003183 0.57730276 + layer.2.v_cache 0.00002344 0.00294119 + layer.3.k_cache 0.01028479 4.01930138 + layer.3.v_cache 0.00002212 0.00305779 + layer.4.k_cache 0.00070434 0.08363677 + layer.4.v_cache 0.00006490 0.00594205 + layer.4.output 0.04342050 57.90120231 + ------------------------------------------------------------------------------------- + TOTAL 0.06502327 26.28654669 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 567536 +BPFP 1.6936 bits/point +EBPFP 3.3872 equivalent bits/point +MSE 26.286547 +---------------------- -------------------------------------------------------- +Time: 3.265s Load: 0.011s, Pack+Encode: 1.877s, Decode+Unpack: 1.376s +---------------------- -------------------------------------------------------- +💾 Converting with 26.2865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,452B, BPFP=0.7233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,260B, BPFP=3.3841 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,440B, BPFP=1.5939 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,612B, BPFP=3.2883 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,664B, BPFP=1.7811 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,600B, BPFP=3.2876 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,488B, BPFP=1.5967 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,792B, BPFP=3.3569 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,476B, BPFP=2.5834 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,260B, BPFP=3.2098 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,760B, BPFP=0.7863 +⌛️ [2/4] FRONTEND: Frontend time: 1.897s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12950333 28.27744903 + layer.0.v_cache 0.00001690 0.00235109 + layer.1.k_cache 0.52417032 2.97572412 + layer.1.v_cache 0.00000602 0.00095577 + layer.2.k_cache 0.01578268 0.70340941 + layer.2.v_cache 0.00002268 0.00294702 + layer.3.k_cache 0.03345619 3.67449520 + layer.3.v_cache 0.00002007 0.00315358 + layer.4.k_cache 0.00067301 0.08173665 + layer.4.v_cache 0.00005077 0.00599546 + layer.4.output 0.00489311 179.17380842 + ------------------------------------------------------------------------------------- + TOTAL 0.04340905 75.87911037 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 521804 +BPFP 1.7829 bits/point +EBPFP 3.5658 equivalent bits/point +MSE 75.879110 +---------------------- -------------------------------------------------------- +Time: 3.280s Load: 0.010s, Pack+Encode: 1.897s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 75.8791 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,296B, BPFP=0.7279 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,520B, BPFP=3.3597 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,948B, BPFP=1.4811 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,120B, BPFP=3.2607 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,792B, BPFP=1.7528 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,104B, BPFP=3.1889 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,280B, BPFP=1.5752 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,444B, BPFP=3.2837 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,088B, BPFP=2.5515 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,888B, BPFP=3.1736 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 80,620B, BPFP=0.8143 +⌛️ [2/4] FRONTEND: Frontend time: 1.761s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.245s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13888381 29.00273305 + layer.0.v_cache 0.00001731 0.00238691 + layer.1.k_cache 0.27925331 3.05668509 + layer.1.v_cache 0.00000589 0.00093741 + layer.2.k_cache 0.03172064 0.74931260 + layer.2.v_cache 0.00001980 0.00285510 + layer.3.k_cache 0.01738505 3.64376403 + layer.3.v_cache 0.00001961 0.00307348 + layer.4.k_cache 0.00068340 0.08163506 + layer.4.v_cache 0.00005063 0.00620062 + layer.4.output 1.38525280 226.98842518 + ------------------------------------------------------------------------------------- + TOTAL 0.59792994 95.61579762 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 425100 +BPFP 1.7679 bits/point +EBPFP 3.5359 equivalent bits/point +MSE 95.615798 +---------------------- -------------------------------------------------------- +Time: 3.013s Load: 0.007s, Pack+Encode: 1.761s, Decode+Unpack: 1.245s +---------------------- -------------------------------------------------------- +💾 Converting with 95.6158 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.7558 bits/point +Avg EBPFP 3.5115 equivalent bits/point +Avg MSE 75.171625 +Avg Time 3.403s +------------------------ ---------------------------- diff --git a/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..fe13cd01f3bfeb93fee5d3936e2a29f78365a70a --- /dev/null +++ b/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 520 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other +Output output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other +---------------- ------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,648B, BPFP=0.7500 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,968B, BPFP=3.5734 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,616B, BPFP=1.4471 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,408B, BPFP=3.4521 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,820B, BPFP=1.8517 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,084B, BPFP=3.4269 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,692B, BPFP=1.6863 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,692B, BPFP=3.5519 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,804B, BPFP=2.7055 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,580B, BPFP=3.3877 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,656B, BPFP=0.9512 +⌛️ [2/4] FRONTEND: Frontend time: 2.273s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.281s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09475397 28.44131394 + layer.0.v_cache 0.00001927 0.00255906 + layer.1.k_cache 0.17401673 1.91791470 + layer.1.v_cache 0.00000670 0.00103633 + layer.2.k_cache 0.00814445 0.66226371 + layer.2.v_cache 0.00002157 0.00305908 + layer.3.k_cache 0.01691350 3.26522053 + layer.3.v_cache 0.00002161 0.00356774 + layer.4.k_cache 0.00069969 0.08598544 + layer.4.v_cache 0.00005477 0.00645965 + layer.4.output 1.52309445 241.66302416 + ------------------------------------------------------------------------------------- + TOTAL 0.64448903 101.53120878 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 417968 +BPFP 1.9113 bits/point +EBPFP 3.8225 equivalent bits/point +MSE 101.531209 +---------------------- -------------------------------------------------------- +Time: 3.561s Load: 0.007s, Pack+Encode: 2.273s, Decode+Unpack: 1.281s +---------------------- -------------------------------------------------------- +💾 Converting with 101.5312 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 212, 128) +Output shape: (1, 212, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.output: torch.Size([1, 212, 3584]) -> torch.Size([1, 1, 212, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,140B, BPFP=0.7473 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,180B, BPFP=3.4773 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,308B, BPFP=1.4968 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,632B, BPFP=3.3632 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,900B, BPFP=1.8352 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,944B, BPFP=3.3125 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,832B, BPFP=1.6828 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,516B, BPFP=3.4284 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,176B, BPFP=2.6663 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,716B, BPFP=3.2957 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,128B, BPFP=0.8753 +⌛️ [2/4] FRONTEND: Frontend time: 1.736s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.234s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09403566 29.18430775 + layer.0.v_cache 0.00001799 0.00262218 + layer.1.k_cache 0.24957257 2.91614345 + layer.1.v_cache 0.00000621 0.00102497 + layer.2.k_cache 0.01169950 0.53220418 + layer.2.v_cache 0.00002031 0.00305502 + layer.3.k_cache 0.03645871 3.83647703 + layer.3.v_cache 0.00002058 0.00352601 + layer.4.k_cache 0.00069161 0.08433348 + layer.4.v_cache 0.00005130 0.00652722 + layer.4.output 1.44404146 216.13169643 + ------------------------------------------------------------------------------------- + TOTAL 0.61769792 91.14659390 + (elements=1,845,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1845248 +Total Bytes 426472 +BPFP 1.8490 bits/point +EBPFP 3.6979 equivalent bits/point +MSE 91.146594 +---------------------- -------------------------------------------------------- +Time: 2.978s Load: 0.008s, Pack+Encode: 1.736s, Decode+Unpack: 1.234s +---------------------- -------------------------------------------------------- +💾 Converting with 91.1466 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 250, 128) +Output shape: (1, 250, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.output: torch.Size([1, 250, 3584]) -> torch.Size([1, 1, 250, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,192B, BPFP=0.6995 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,044B, BPFP=3.0652 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,624B, BPFP=1.4140 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,104B, BPFP=2.9440 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,720B, BPFP=1.6700 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,036B, BPFP=2.8773 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,808B, BPFP=1.5505 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,868B, BPFP=2.9293 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,980B, BPFP=2.6237 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,172B, BPFP=2.8232 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,220B, BPFP=0.8145 +⌛️ [2/4] FRONTEND: Frontend time: 1.763s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.230s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09121272 31.92903516 + layer.0.v_cache 0.00001793 0.00247184 + layer.1.k_cache 0.34710925 3.05027002 + layer.1.v_cache 0.00000615 0.00098314 + layer.2.k_cache 0.02211576 0.57662061 + layer.2.v_cache 0.00002179 0.00306051 + layer.3.k_cache 0.02174057 4.22597363 + layer.3.v_cache 0.00001938 0.00309502 + layer.4.k_cache 0.00073173 0.08551938 + layer.4.v_cache 0.00005145 0.00598075 + layer.4.output 1.22463198 201.52755357 + ------------------------------------------------------------------------------------- + TOTAL 0.53267356 85.32799324 + (elements=2,176,000) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2176000 +Total Bytes 452768 +BPFP 1.6646 bits/point +EBPFP 3.3292 equivalent bits/point +MSE 85.327993 +---------------------- -------------------------------------------------------- +Time: 3.001s Load: 0.008s, Pack+Encode: 1.763s, Decode+Unpack: 1.230s +---------------------- -------------------------------------------------------- +💾 Converting with 85.3280 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,576B, BPFP=0.7280 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,480B, BPFP=3.3370 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,452B, BPFP=1.4078 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,152B, BPFP=3.2456 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,500B, BPFP=1.7552 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,200B, BPFP=3.1801 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,008B, BPFP=1.5837 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,332B, BPFP=3.2580 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,280B, BPFP=2.5661 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,240B, BPFP=3.1140 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,664B, BPFP=0.8325 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.235s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12780563 29.27237705 + layer.0.v_cache 0.00001798 0.00244322 + layer.1.k_cache 0.36413978 2.64053009 + layer.1.v_cache 0.00000619 0.00095980 + layer.2.k_cache 0.01529831 0.62032046 + layer.2.v_cache 0.00002052 0.00285955 + layer.3.k_cache 0.01269315 3.88183701 + layer.3.v_cache 0.00002023 0.00314698 + layer.4.k_cache 0.00069261 0.08259029 + layer.4.v_cache 0.00005545 0.00615052 + layer.4.output 1.34865724 214.88209959 + ------------------------------------------------------------------------------------- + TOTAL 0.58596180 90.62870071 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 435884 +BPFP 1.7649 bits/point +EBPFP 3.5298 equivalent bits/point +MSE 90.628701 +---------------------- -------------------------------------------------------- +Time: 2.969s Load: 0.008s, Pack+Encode: 1.726s, Decode+Unpack: 1.235s +---------------------- -------------------------------------------------------- +💾 Converting with 90.6287 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 214, 128) +Output shape: (1, 214, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.output: torch.Size([1, 214, 3584]) -> torch.Size([1, 1, 214, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,148B, BPFP=0.7409 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,972B, BPFP=3.4296 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,768B, BPFP=1.4433 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,456B, BPFP=3.3189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,436B, BPFP=1.7842 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,776B, BPFP=3.2693 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,272B, BPFP=1.6262 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,596B, BPFP=3.3291 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,012B, BPFP=2.6294 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,332B, BPFP=3.2369 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 77,000B, BPFP=0.8032 +⌛️ [2/4] FRONTEND: Frontend time: 1.721s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12820032 29.16879472 + layer.0.v_cache 0.00001693 0.00238801 + layer.1.k_cache 0.24875343 2.74566280 + layer.1.v_cache 0.00000587 0.00092445 + layer.2.k_cache 0.01453762 0.70345656 + layer.2.v_cache 0.00001988 0.00283333 + layer.3.k_cache 0.02276520 3.62949550 + layer.3.v_cache 0.00001932 0.00299692 + layer.4.k_cache 0.00071399 0.08203403 + layer.4.v_cache 0.00004895 0.00595001 + layer.4.output 1.43051836 223.08315254 + ------------------------------------------------------------------------------------- + TOTAL 0.61345353 93.99568259 + (elements=1,862,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1862656 +Total Bytes 416768 +BPFP 1.7900 bits/point +EBPFP 3.5800 equivalent bits/point +MSE 93.995683 +---------------------- -------------------------------------------------------- +Time: 2.947s Load: 0.008s, Pack+Encode: 1.721s, Decode+Unpack: 1.219s +---------------------- -------------------------------------------------------- +💾 Converting with 93.9957 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,616B, BPFP=0.7116 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,708B, BPFP=3.3116 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,712B, BPFP=1.6760 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,180B, BPFP=3.2254 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,020B, BPFP=1.7498 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,056B, BPFP=3.1620 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,792B, BPFP=1.5677 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,920B, BPFP=3.2671 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,576B, BPFP=2.6837 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,240B, BPFP=3.1724 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 104,040B, BPFP=0.8384 +⌛️ [2/4] FRONTEND: Frontend time: 2.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.356s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11875103 29.96157209 + layer.0.v_cache 0.00001689 0.00234989 + layer.1.k_cache 0.49171178 3.11598134 + layer.1.v_cache 0.00000616 0.00094174 + layer.2.k_cache 0.02532451 0.70420821 + layer.2.v_cache 0.00002086 0.00290650 + layer.3.k_cache 0.02370731 3.77250192 + layer.3.v_cache 0.00002129 0.00310990 + layer.4.k_cache 0.00073268 0.08393063 + layer.4.v_cache 0.00004879 0.00591873 + layer.4.output 0.00480151 175.20355209 + ------------------------------------------------------------------------------------- + TOTAL 0.04082070 74.35754621 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 538860 +BPFP 1.7880 bits/point +EBPFP 3.5760 equivalent bits/point +MSE 74.357546 +---------------------- -------------------------------------------------------- +Time: 3.595s Load: 0.009s, Pack+Encode: 2.229s, Decode+Unpack: 1.356s +---------------------- -------------------------------------------------------- +💾 Converting with 74.3575 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,064B, BPFP=0.7056 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,332B, BPFP=3.1462 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,404B, BPFP=1.5564 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,284B, BPFP=3.0156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,968B, BPFP=1.7199 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,272B, BPFP=2.9510 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,968B, BPFP=1.5923 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,176B, BPFP=3.0087 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,192B, BPFP=2.4357 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,180B, BPFP=2.8814 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,940B, BPFP=0.8468 +⌛️ [2/4] FRONTEND: Frontend time: 1.725s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10657067 31.75683594 + layer.0.v_cache 0.00001604 0.00240466 + layer.1.k_cache 0.39270948 2.82986263 + layer.1.v_cache 0.00000611 0.00095147 + layer.2.k_cache 0.01541229 0.54783549 + layer.2.v_cache 0.00002285 0.00300730 + layer.3.k_cache 0.01321528 3.93833631 + layer.3.v_cache 0.00001985 0.00306559 + layer.4.k_cache 0.00076706 0.08359358 + layer.4.v_cache 0.00005205 0.00586753 + layer.4.output 1.24960368 201.79959913 + ------------------------------------------------------------------------------------- + TOTAL 0.54564808 85.39817379 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 453780 +BPFP 1.7024 bits/point +EBPFP 3.4047 equivalent bits/point +MSE 85.398174 +---------------------- -------------------------------------------------------- +Time: 2.959s Load: 0.009s, Pack+Encode: 1.725s, Decode+Unpack: 1.225s +---------------------- -------------------------------------------------------- +💾 Converting with 85.3982 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 197, 128) +Output shape: (1, 197, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.output: torch.Size([1, 197, 3584]) -> torch.Size([1, 1, 197, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,428B, BPFP=0.7478 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,496B, BPFP=3.6085 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,056B, BPFP=1.5907 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,196B, BPFP=3.5054 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,648B, BPFP=1.7963 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,176B, BPFP=3.5038 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,768B, BPFP=1.6472 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,700B, BPFP=3.6247 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,160B, BPFP=2.7094 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,736B, BPFP=3.4689 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,608B, BPFP=0.9247 +⌛️ [2/4] FRONTEND: Frontend time: 1.721s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.231s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09219642 28.81001646 + layer.0.v_cache 0.00001621 0.00241718 + layer.1.k_cache 0.17840836 2.56387376 + layer.1.v_cache 0.00000639 0.00096806 + layer.2.k_cache 0.00922235 0.63408700 + layer.2.v_cache 0.00002142 0.00303110 + layer.3.k_cache 0.01957878 3.61637514 + layer.3.v_cache 0.00002124 0.00320538 + layer.4.k_cache 0.00068177 0.08181296 + layer.4.v_cache 0.00005086 0.00649433 + layer.4.output 1.55401793 248.80728789 + ------------------------------------------------------------------------------------- + TOTAL 0.65754878 104.55137039 + (elements=1,714,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1714688 +Total Bytes 411972 +BPFP 1.9221 bits/point +EBPFP 3.8442 equivalent bits/point +MSE 104.551370 +---------------------- -------------------------------------------------------- +Time: 2.959s Load: 0.008s, Pack+Encode: 1.721s, Decode+Unpack: 1.231s +---------------------- -------------------------------------------------------- +💾 Converting with 104.5514 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,704B, BPFP=0.6907 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,244B, BPFP=3.0365 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,492B, BPFP=1.5369 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,416B, BPFP=2.9444 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,876B, BPFP=1.6571 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,312B, BPFP=2.8887 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,192B, BPFP=1.5218 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,244B, BPFP=2.9357 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,372B, BPFP=2.3877 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,264B, BPFP=2.8359 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 108,084B, BPFP=0.7783 +⌛️ [2/4] FRONTEND: Frontend time: 1.859s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11169702 32.60486706 + layer.0.v_cache 0.00001595 0.00231699 + layer.1.k_cache 0.54955656 3.07622188 + layer.1.v_cache 0.00000603 0.00093874 + layer.2.k_cache 0.01751511 0.56096285 + layer.2.v_cache 0.00002011 0.00286726 + layer.3.k_cache 0.01650532 4.09985194 + layer.3.v_cache 0.00001967 0.00299207 + layer.4.k_cache 0.00077751 0.08176347 + layer.4.v_cache 0.00005008 0.00589141 + layer.4.output 0.04306356 150.83397177 + ------------------------------------------------------------------------------------- + TOTAL 0.05868284 64.48685154 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 553200 +BPFP 1.6402 bits/point +EBPFP 3.2804 equivalent bits/point +MSE 64.486852 +---------------------- -------------------------------------------------------- +Time: 3.241s Load: 0.012s, Pack+Encode: 1.859s, Decode+Unpack: 1.371s +---------------------- -------------------------------------------------------- +💾 Converting with 64.4869 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,288B, BPFP=0.7374 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,156B, BPFP=3.4515 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,036B, BPFP=1.4361 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,248B, BPFP=3.3148 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,900B, BPFP=1.7847 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,868B, BPFP=3.2159 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,680B, BPFP=1.6256 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,688B, BPFP=3.3463 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,676B, BPFP=2.5571 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,608B, BPFP=3.1972 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,956B, BPFP=0.8801 +⌛️ [2/4] FRONTEND: Frontend time: 1.731s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.242s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09832749 30.03021968 + layer.0.v_cache 0.00001658 0.00246419 + layer.1.k_cache 0.33493462 2.46001273 + layer.1.v_cache 0.00000656 0.00095418 + layer.2.k_cache 0.02762392 0.64825474 + layer.2.v_cache 0.00002012 0.00289317 + layer.3.k_cache 0.01647770 3.60160422 + layer.3.v_cache 0.00001951 0.00315494 + layer.4.k_cache 0.00067889 0.08083737 + layer.4.v_cache 0.00004874 0.00616428 + layer.4.output 1.40435175 235.63202408 + ------------------------------------------------------------------------------------- + TOTAL 0.60638920 99.19180753 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 430104 +BPFP 1.8134 bits/point +EBPFP 3.6268 equivalent bits/point +MSE 99.191808 +---------------------- -------------------------------------------------------- +Time: 2.981s Load: 0.009s, Pack+Encode: 1.731s, Decode+Unpack: 1.242s +---------------------- -------------------------------------------------------- +💾 Converting with 99.1918 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,456B, BPFP=0.7229 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,280B, BPFP=3.3379 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,756B, BPFP=1.5733 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,708B, BPFP=3.2293 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,612B, BPFP=1.7707 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,808B, BPFP=3.1670 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,600B, BPFP=1.5625 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,648B, BPFP=3.2251 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,000B, BPFP=2.5581 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,344B, BPFP=3.1350 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,716B, BPFP=0.8663 +⌛️ [2/4] FRONTEND: Frontend time: 1.733s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.242s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102250 30.56284136 + layer.0.v_cache 0.00001721 0.00232413 + layer.1.k_cache 0.37371607 3.14841373 + layer.1.v_cache 0.00000658 0.00094181 + layer.2.k_cache 0.01127678 0.54789947 + layer.2.v_cache 0.00001998 0.00288143 + layer.3.k_cache 0.02136800 3.79257283 + layer.3.v_cache 0.00001963 0.00307707 + layer.4.k_cache 0.00069450 0.08262329 + layer.4.v_cache 0.00005147 0.00613305 + layer.4.output 1.35464589 221.43967288 + ------------------------------------------------------------------------------------- + TOTAL 0.58945376 93.42514226 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 438928 +BPFP 1.7851 bits/point +EBPFP 3.5701 equivalent bits/point +MSE 93.425142 +---------------------- -------------------------------------------------------- +Time: 2.983s Load: 0.008s, Pack+Encode: 1.733s, Decode+Unpack: 1.242s +---------------------- -------------------------------------------------------- +💾 Converting with 93.4251 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,156B, BPFP=0.6945 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,944B, BPFP=3.0468 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,020B, BPFP=1.4330 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,124B, BPFP=2.9335 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,256B, BPFP=1.6345 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,904B, BPFP=2.8576 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,496B, BPFP=1.5249 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,100B, BPFP=2.9320 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,604B, BPFP=2.3409 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,072B, BPFP=2.8058 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,460B, BPFP=0.7956 +⌛️ [2/4] FRONTEND: Frontend time: 1.736s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.237s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11871604 33.02928909 + layer.0.v_cache 0.00001755 0.00226301 + layer.1.k_cache 0.46315337 3.06870686 + layer.1.v_cache 0.00000588 0.00091198 + layer.2.k_cache 0.02202094 0.64793001 + layer.2.v_cache 0.00001989 0.00280686 + layer.3.k_cache 0.01278766 4.22673223 + layer.3.v_cache 0.00002092 0.00308695 + layer.4.k_cache 0.00074254 0.08016933 + layer.4.v_cache 0.00004925 0.00569747 + layer.4.output 1.21975157 190.35143355 + ------------------------------------------------------------------------------------- + TOTAL 0.53857618 80.79574286 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 446136 +BPFP 1.6337 bits/point +EBPFP 3.2673 equivalent bits/point +MSE 80.795743 +---------------------- -------------------------------------------------------- +Time: 2.983s Load: 0.010s, Pack+Encode: 1.736s, Decode+Unpack: 1.237s +---------------------- -------------------------------------------------------- +💾 Converting with 80.7957 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 195, 128) +Output shape: (1, 195, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.output: torch.Size([1, 195, 3584]) -> torch.Size([1, 1, 195, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,280B, BPFP=0.7436 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 44,544B, BPFP=3.5692 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,468B, BPFP=1.5599 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,220B, BPFP=3.5433 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,112B, BPFP=1.7718 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,184B, BPFP=3.4603 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,544B, BPFP=1.6462 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,160B, BPFP=3.6186 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,236B, BPFP=2.6631 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 42,572B, BPFP=3.4112 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,848B, BPFP=0.8568 +⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.241s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605874 29.31970653 + layer.0.v_cache 0.00001676 0.00242963 + layer.1.k_cache 0.13872500 2.85717398 + layer.1.v_cache 0.00000593 0.00095517 + layer.2.k_cache 0.01222654 0.74039447 + layer.2.v_cache 0.00001968 0.00278364 + layer.3.k_cache 0.01415822 3.39805908 + layer.3.v_cache 0.00002127 0.00327439 + layer.4.k_cache 0.00071127 0.08239431 + layer.4.v_cache 0.00005419 0.00624329 + layer.4.output 1.56987251 239.68976648 + ------------------------------------------------------------------------------------- + TOTAL 0.66300619 100.83775176 + (elements=1,697,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1697280 +Total Bytes 399168 +BPFP 1.8814 bits/point +EBPFP 3.7629 equivalent bits/point +MSE 100.837752 +---------------------- -------------------------------------------------------- +Time: 2.985s Load: 0.009s, Pack+Encode: 1.735s, Decode+Unpack: 1.241s +---------------------- -------------------------------------------------------- +💾 Converting with 100.8378 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,572B, BPFP=0.7359 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,724B, BPFP=3.1528 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,552B, BPFP=1.4210 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,612B, BPFP=3.0573 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,348B, BPFP=1.7469 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,948B, BPFP=3.0003 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,092B, BPFP=1.6391 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,500B, BPFP=3.0477 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,764B, BPFP=2.4694 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,092B, BPFP=2.9269 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,248B, BPFP=0.8738 +⌛️ [2/4] FRONTEND: Frontend time: 1.861s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.130s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11176265 31.62915307 + layer.0.v_cache 0.00001802 0.00256594 + layer.1.k_cache 0.18491925 2.49482476 + layer.1.v_cache 0.00000608 0.00101190 + layer.2.k_cache 0.01999094 0.48562396 + layer.2.v_cache 0.00002176 0.00311287 + layer.3.k_cache 0.03910406 4.04888882 + layer.3.v_cache 0.00002004 0.00326399 + layer.4.k_cache 0.00067358 0.08657905 + layer.4.v_cache 0.00005212 0.00620643 + layer.4.output 0.00886795 258.84056122 + ------------------------------------------------------------------------------------- + TOTAL 0.02462613 108.86147996 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 341452 +BPFP 1.7244 bits/point +EBPFP 3.4487 equivalent bits/point +MSE 108.861480 +---------------------- -------------------------------------------------------- +Time: 2.999s Load: 0.008s, Pack+Encode: 1.861s, Decode+Unpack: 1.130s +---------------------- -------------------------------------------------------- +💾 Converting with 108.8615 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,024B, BPFP=0.7353 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,152B, BPFP=3.4589 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,804B, BPFP=1.5261 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,700B, BPFP=3.3524 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,928B, BPFP=1.8286 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,912B, BPFP=3.2946 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,592B, BPFP=1.6573 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,580B, BPFP=3.4170 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,576B, BPFP=2.6097 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,752B, BPFP=3.2829 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 79,900B, BPFP=0.8373 +⌛️ [2/4] FRONTEND: Frontend time: 1.736s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.236s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15089922 29.32992224 + layer.0.v_cache 0.00001686 0.00250304 + layer.1.k_cache 0.31217133 3.05492935 + layer.1.v_cache 0.00000593 0.00099894 + layer.2.k_cache 0.01305961 0.71058734 + layer.2.v_cache 0.00001999 0.00300350 + layer.3.k_cache 0.01178671 3.61033981 + layer.3.v_cache 0.00002039 0.00326908 + layer.4.k_cache 0.00067606 0.08263856 + layer.4.v_cache 0.00005285 0.00620187 + layer.4.output 1.43724915 230.11701459 + ------------------------------------------------------------------------------------- + TOTAL 0.62055606 96.91902917 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 422920 +BPFP 1.8249 bits/point +EBPFP 3.6499 equivalent bits/point +MSE 96.919029 +---------------------- -------------------------------------------------------- +Time: 2.979s Load: 0.007s, Pack+Encode: 1.736s, Decode+Unpack: 1.236s +---------------------- -------------------------------------------------------- +💾 Converting with 96.9190 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 187, 128) +Output shape: (1, 187, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.output: torch.Size([1, 187, 3584]) -> torch.Size([1, 1, 187, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,572B, BPFP=0.7162 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,948B, BPFP=3.0872 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,088B, BPFP=1.5114 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,620B, BPFP=2.9763 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,880B, BPFP=1.7447 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,296B, BPFP=2.9492 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,712B, BPFP=1.6471 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,188B, BPFP=3.0237 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,184B, BPFP=2.4385 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,576B, BPFP=2.8890 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,232B, BPFP=0.9338 +⌛️ [2/4] FRONTEND: Frontend time: 1.614s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.130s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08258281 32.05992543 + layer.0.v_cache 0.00001733 0.00259932 + layer.1.k_cache 0.10049684 2.75886935 + layer.1.v_cache 0.00000647 0.00102748 + layer.2.k_cache 0.01472783 0.53428878 + layer.2.v_cache 0.00002140 0.00319753 + layer.3.k_cache 0.02125903 4.03038115 + layer.3.v_cache 0.00002180 0.00352615 + layer.4.k_cache 0.00070831 0.08796671 + layer.4.v_cache 0.00005280 0.00655044 + layer.4.output 0.00866398 260.19573625 + ------------------------------------------------------------------------------------- + TOTAL 0.01650250 109.46226389 + (elements=1,627,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1627648 +Total Bytes 353296 +BPFP 1.7365 bits/point +EBPFP 3.4729 equivalent bits/point +MSE 109.462264 +---------------------- -------------------------------------------------------- +Time: 2.750s Load: 0.006s, Pack+Encode: 1.614s, Decode+Unpack: 1.130s +---------------------- -------------------------------------------------------- +💾 Converting with 109.4623 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,060B, BPFP=0.7450 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,064B, BPFP=3.4852 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,076B, BPFP=1.5607 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,612B, BPFP=3.3777 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,152B, BPFP=1.7885 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,280B, BPFP=3.3531 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,300B, BPFP=1.6514 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,040B, BPFP=3.4094 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,212B, BPFP=2.6075 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,964B, BPFP=3.3297 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,224B, BPFP=0.9016 +⌛️ [2/4] FRONTEND: Frontend time: 1.736s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.224s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11071491 28.81455957 + layer.0.v_cache 0.00001804 0.00246092 + layer.1.k_cache 0.26784508 2.65951986 + layer.1.v_cache 0.00000610 0.00096443 + layer.2.k_cache 0.01369454 0.73495975 + layer.2.v_cache 0.00002102 0.00291390 + layer.3.k_cache 0.01444715 3.47736203 + layer.3.v_cache 0.00002041 0.00318973 + layer.4.k_cache 0.00067564 0.08407379 + layer.4.v_cache 0.00005248 0.00616741 + layer.4.output 1.45091435 225.16168754 + ------------------------------------------------------------------------------------- + TOTAL 0.62140564 94.81870495 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 426984 +BPFP 1.8599 bits/point +EBPFP 3.7199 equivalent bits/point +MSE 94.818705 +---------------------- -------------------------------------------------------- +Time: 2.968s Load: 0.007s, Pack+Encode: 1.736s, Decode+Unpack: 1.224s +---------------------- -------------------------------------------------------- +💾 Converting with 94.8187 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,304B, BPFP=0.7220 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,760B, BPFP=3.3464 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,856B, BPFP=1.4613 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,164B, BPFP=3.2346 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,016B, BPFP=1.7528 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,536B, BPFP=3.1906 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,912B, BPFP=1.6054 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,944B, BPFP=3.2892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,028B, BPFP=2.5244 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,280B, BPFP=3.1726 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 82,604B, BPFP=0.8268 +⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.236s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13389712 28.86133688 + layer.0.v_cache 0.00001748 0.00235489 + layer.1.k_cache 0.36644789 2.84363176 + layer.1.v_cache 0.00000587 0.00091208 + layer.2.k_cache 0.02260907 0.69497776 + layer.2.v_cache 0.00001979 0.00273900 + layer.3.k_cache 0.02550721 4.05834140 + layer.3.v_cache 0.00001968 0.00300483 + layer.4.k_cache 0.00067529 0.08109908 + layer.4.v_cache 0.00006705 0.00616185 + layer.4.output 1.37284074 219.15875240 + ------------------------------------------------------------------------------------- + TOTAL 0.59759716 92.39210743 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 429404 +BPFP 1.7698 bits/point +EBPFP 3.5397 equivalent bits/point +MSE 92.392107 +---------------------- -------------------------------------------------------- +Time: 2.980s Load: 0.008s, Pack+Encode: 1.735s, Decode+Unpack: 1.236s +---------------------- -------------------------------------------------------- +💾 Converting with 92.3921 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 406, 128) +Output shape: (1, 406, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.output: torch.Size([1, 406, 3584]) -> torch.Size([1, 1, 406, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 17,804B, BPFP=0.6852 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 85,112B, BPFP=3.2756 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 43,588B, BPFP=1.6775 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 81,280B, BPFP=3.1281 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 44,572B, BPFP=1.7154 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 80,040B, BPFP=3.0804 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 39,692B, BPFP=1.5276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 81,168B, BPFP=3.1238 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 64,744B, BPFP=2.4917 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 78,288B, BPFP=3.0129 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 155,580B, BPFP=0.8554 +⌛️ [2/4] FRONTEND: Frontend time: 2.340s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.598s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10474718 30.70763258 + layer.0.v_cache 0.00001775 0.00237213 + layer.1.k_cache 0.89873824 3.38600407 + layer.1.v_cache 0.00000687 0.00094349 + layer.2.k_cache 0.02602503 0.66162951 + layer.2.v_cache 0.00002164 0.00290681 + layer.3.k_cache 0.01558245 3.72192804 + layer.3.v_cache 0.00002201 0.00308273 + layer.4.k_cache 0.00072958 0.08075955 + layer.4.v_cache 0.00005327 0.00587771 + layer.4.output 0.00651461 120.54389514 + ------------------------------------------------------------------------------------- + TOTAL 0.06420861 51.90472957 + (elements=3,533,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3533824 +Total Bytes 771868 +BPFP 1.7474 bits/point +EBPFP 3.4948 equivalent bits/point +MSE 51.904730 +---------------------- -------------------------------------------------------- +Time: 3.951s Load: 0.013s, Pack+Encode: 2.340s, Decode+Unpack: 1.598s +---------------------- -------------------------------------------------------- +💾 Converting with 51.9047 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,140B, BPFP=0.7105 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,988B, BPFP=3.1242 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,016B, BPFP=1.4679 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,508B, BPFP=3.0298 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,444B, BPFP=1.6865 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,316B, BPFP=2.9538 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,320B, BPFP=1.5510 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,288B, BPFP=3.0158 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,032B, BPFP=2.4255 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,216B, BPFP=2.8837 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,228B, BPFP=0.8585 +⌛️ [2/4] FRONTEND: Frontend time: 1.749s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.237s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14414167 30.92461535 + layer.0.v_cache 0.00001717 0.00240630 + layer.1.k_cache 0.44179790 3.04689593 + layer.1.v_cache 0.00000667 0.00097855 + layer.2.k_cache 0.03016641 0.62740858 + layer.2.v_cache 0.00002043 0.00289468 + layer.3.k_cache 0.03726396 4.31590402 + layer.3.v_cache 0.00002099 0.00322216 + layer.4.k_cache 0.00070653 0.08177544 + layer.4.v_cache 0.00004910 0.00581186 + layer.4.output 1.24963187 190.18534985 + ------------------------------------------------------------------------------------- + TOTAL 0.55303611 80.60643305 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 452496 +BPFP 1.6975 bits/point +EBPFP 3.3951 equivalent bits/point +MSE 80.606433 +---------------------- -------------------------------------------------------- +Time: 2.995s Load: 0.009s, Pack+Encode: 1.749s, Decode+Unpack: 1.237s +---------------------- -------------------------------------------------------- +💾 Converting with 80.6064 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,596B, BPFP=0.7535 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,964B, BPFP=3.6090 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,440B, BPFP=1.4479 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 43,948B, BPFP=3.4507 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,172B, BPFP=1.8194 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,124B, BPFP=3.4645 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,424B, BPFP=1.6822 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,548B, BPFP=3.5763 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,408B, BPFP=2.7016 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,388B, BPFP=3.4067 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 76,888B, BPFP=0.8624 +⌛️ [2/4] FRONTEND: Frontend time: 1.746s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.231s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15985105 28.60826349 + layer.0.v_cache 0.00001905 0.00260124 + layer.1.k_cache 0.22745673 2.02554061 + layer.1.v_cache 0.00000590 0.00102798 + layer.2.k_cache 0.01549998 0.73071741 + layer.2.v_cache 0.00002003 0.00292111 + layer.3.k_cache 0.01333464 3.61559875 + layer.3.v_cache 0.00002034 0.00338192 + layer.4.k_cache 0.00067076 0.08512351 + layer.4.v_cache 0.00004949 0.00623030 + layer.4.output 1.53832061 237.49243988 + ------------------------------------------------------------------------------------- + TOTAL 0.65795131 99.85461679 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 406900 +BPFP 1.8793 bits/point +EBPFP 3.7587 equivalent bits/point +MSE 99.854617 +---------------------- -------------------------------------------------------- +Time: 2.984s Load: 0.008s, Pack+Encode: 1.746s, Decode+Unpack: 1.231s +---------------------- -------------------------------------------------------- +💾 Converting with 99.8546 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 237, 128) +Output shape: (1, 237, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.output: torch.Size([1, 237, 3584]) -> torch.Size([1, 1, 237, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,780B, BPFP=0.7107 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,336B, BPFP=3.1867 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,988B, BPFP=1.5815 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,164B, BPFP=3.1094 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,408B, BPFP=1.7410 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,300B, BPFP=3.0525 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,332B, BPFP=1.6042 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,944B, BPFP=3.0949 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,796B, BPFP=2.4918 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,092B, BPFP=2.9728 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,820B, BPFP=0.8554 +⌛️ [2/4] FRONTEND: Frontend time: 1.750s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.238s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351868 29.47342679 + layer.0.v_cache 0.00001655 0.00237994 + layer.1.k_cache 0.34375901 2.86546718 + layer.1.v_cache 0.00000619 0.00095667 + layer.2.k_cache 0.01282014 0.54803898 + layer.2.v_cache 0.00002104 0.00287604 + layer.3.k_cache 0.03241006 4.29492033 + layer.3.v_cache 0.00001985 0.00306334 + layer.4.k_cache 0.00067667 0.08183009 + layer.4.v_cache 0.00005279 0.00592688 + layer.4.output 1.29178675 201.72716998 + ------------------------------------------------------------------------------------- + TOTAL 0.56210637 85.25700448 + (elements=2,062,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2062848 +Total Bytes 447960 +BPFP 1.7372 bits/point +EBPFP 3.4745 equivalent bits/point +MSE 85.257004 +---------------------- -------------------------------------------------------- +Time: 2.996s Load: 0.008s, Pack+Encode: 1.750s, Decode+Unpack: 1.238s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2570 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,132B, BPFP=0.7363 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,896B, BPFP=3.4081 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,948B, BPFP=1.5224 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,920B, BPFP=3.3372 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,812B, BPFP=1.8032 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,132B, BPFP=3.2799 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,392B, BPFP=1.6273 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,300B, BPFP=3.3648 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,588B, BPFP=2.5863 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,388B, BPFP=3.2259 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 80,016B, BPFP=0.8307 +⌛️ [2/4] FRONTEND: Frontend time: 1.746s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.239s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11023432 29.14779251 + layer.0.v_cache 0.00001677 0.00235646 + layer.1.k_cache 0.30291741 2.62076643 + layer.1.v_cache 0.00000597 0.00094416 + layer.2.k_cache 0.03464115 0.69344795 + layer.2.v_cache 0.00001998 0.00280109 + layer.3.k_cache 0.01330193 3.64688579 + layer.3.v_cache 0.00002112 0.00308184 + layer.4.k_cache 0.00069379 0.08214024 + layer.4.v_cache 0.00005021 0.00608012 + layer.4.output 1.42388847 219.43768688 + ------------------------------------------------------------------------------------- + TOTAL 0.61347776 92.48647675 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 422524 +BPFP 1.8063 bits/point +EBPFP 3.6126 equivalent bits/point +MSE 92.486477 +---------------------- -------------------------------------------------------- +Time: 2.995s Load: 0.009s, Pack+Encode: 1.746s, Decode+Unpack: 1.239s +---------------------- -------------------------------------------------------- +💾 Converting with 92.4865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 231, 128) +Output shape: (1, 231, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.output: torch.Size([1, 231, 3584]) -> torch.Size([1, 1, 231, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,760B, BPFP=0.7278 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,676B, BPFP=3.2925 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,528B, BPFP=1.5915 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,400B, BPFP=3.2062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,956B, BPFP=1.7557 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,456B, BPFP=3.1423 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,904B, BPFP=1.6169 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,324B, BPFP=3.2010 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,144B, BPFP=2.5124 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,436B, BPFP=3.0733 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,252B, BPFP=0.9108 +⌛️ [2/4] FRONTEND: Frontend time: 1.751s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.243s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13257201 30.20835236 + layer.0.v_cache 0.00001701 0.00244605 + layer.1.k_cache 0.40462213 2.80701952 + layer.1.v_cache 0.00000658 0.00097550 + layer.2.k_cache 0.02649644 0.58435310 + layer.2.v_cache 0.00002056 0.00280953 + layer.3.k_cache 0.06147716 3.71189953 + layer.3.v_cache 0.00002105 0.00324904 + layer.4.k_cache 0.00069437 0.08282945 + layer.4.v_cache 0.00005136 0.00611722 + layer.4.output 1.32536713 212.17988559 + ------------------------------------------------------------------------------------- + TOTAL 0.58256168 89.56877944 + (elements=2,010,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2010624 +Total Bytes 450836 +BPFP 1.7938 bits/point +EBPFP 3.5876 equivalent bits/point +MSE 89.568779 +---------------------- -------------------------------------------------------- +Time: 3.002s Load: 0.009s, Pack+Encode: 1.751s, Decode+Unpack: 1.243s +---------------------- -------------------------------------------------------- +💾 Converting with 89.5688 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,176B, BPFP=0.6848 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,108B, BPFP=3.0091 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,644B, BPFP=1.4488 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,436B, BPFP=2.9066 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,872B, BPFP=1.6466 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,940B, BPFP=2.8762 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,024B, BPFP=1.5333 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,676B, BPFP=2.9213 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,548B, BPFP=2.4846 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,784B, BPFP=2.8054 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 106,076B, BPFP=0.9285 +⌛️ [2/4] FRONTEND: Frontend time: 1.757s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.231s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12841698 39.81332721 + layer.0.v_cache 0.00001718 0.00243919 + layer.1.k_cache 0.45450200 2.97424149 + layer.1.v_cache 0.00000627 0.00099226 + layer.2.k_cache 0.01434461 0.52378636 + layer.2.v_cache 0.00002215 0.00299681 + layer.3.k_cache 0.03530993 4.22767502 + layer.3.v_cache 0.00002139 0.00311740 + layer.4.k_cache 0.00068344 0.08062276 + layer.4.v_cache 0.00005073 0.00596033 + layer.4.output 1.20064817 194.62235644 + ------------------------------------------------------------------------------------- + TOTAL 0.53164187 82.94068552 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 470284 +BPFP 1.6951 bits/point +EBPFP 3.3902 equivalent bits/point +MSE 82.940686 +---------------------- -------------------------------------------------------- +Time: 2.998s Load: 0.011s, Pack+Encode: 1.757s, Decode+Unpack: 1.231s +---------------------- -------------------------------------------------------- +💾 Converting with 82.9407 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,840B, BPFP=0.7114 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,432B, BPFP=3.2930 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,848B, BPFP=1.4876 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,168B, BPFP=3.2230 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,480B, BPFP=1.7442 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,060B, BPFP=3.1616 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,604B, BPFP=1.5849 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,324B, BPFP=3.1762 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,580B, BPFP=2.5255 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,196B, BPFP=3.1137 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 106,168B, BPFP=0.8404 +⌛️ [2/4] FRONTEND: Frontend time: 1.858s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13731835 30.42884253 + layer.0.v_cache 0.00001709 0.00232613 + layer.1.k_cache 0.55879569 2.70857120 + layer.1.v_cache 0.00000637 0.00092953 + layer.2.k_cache 0.01348169 0.66807123 + layer.2.v_cache 0.00002375 0.00286376 + layer.3.k_cache 0.01306405 3.78940317 + layer.3.v_cache 0.00002005 0.00305945 + layer.4.k_cache 0.00072533 0.08146225 + layer.4.v_cache 0.00005384 0.00617572 + layer.4.output 0.00472113 187.12224544 + ------------------------------------------------------------------------------------- + TOTAL 0.04450318 79.26749547 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 539700 +BPFP 1.7590 bits/point +EBPFP 3.5181 equivalent bits/point +MSE 79.267495 +---------------------- -------------------------------------------------------- +Time: 3.222s Load: 0.011s, Pack+Encode: 1.858s, Decode+Unpack: 1.353s +---------------------- -------------------------------------------------------- +💾 Converting with 79.2675 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,452B, BPFP=0.6983 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,832B, BPFP=3.1578 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,256B, BPFP=1.6225 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 59,196B, BPFP=3.0729 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,900B, BPFP=1.7598 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 58,056B, BPFP=3.0137 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,692B, BPFP=1.5932 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 59,272B, BPFP=3.0768 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,572B, BPFP=2.4695 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,500B, BPFP=2.9329 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 117,412B, BPFP=0.8707 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13972237 32.43143623 + layer.0.v_cache 0.00001689 0.00240930 + layer.1.k_cache 0.51719427 3.15730524 + layer.1.v_cache 0.00000613 0.00097413 + layer.2.k_cache 0.02019965 0.54029669 + layer.2.v_cache 0.00002053 0.00293445 + layer.3.k_cache 0.01759115 4.25202207 + layer.3.v_cache 0.00002125 0.00319657 + layer.4.k_cache 0.00069604 0.08299854 + layer.4.v_cache 0.00005872 0.00610303 + layer.4.output 0.04434862 155.69878678 + ------------------------------------------------------------------------------------- + TOTAL 0.05917455 66.49242257 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 568140 +BPFP 1.7348 bits/point +EBPFP 3.4697 equivalent bits/point +MSE 66.492423 +---------------------- -------------------------------------------------------- +Time: 3.223s Load: 0.010s, Pack+Encode: 1.856s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 66.4924 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,568B, BPFP=0.7513 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,900B, BPFP=3.6040 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,840B, BPFP=1.5578 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,592B, BPFP=3.5013 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,400B, BPFP=1.8373 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,240B, BPFP=3.3951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,000B, BPFP=1.6489 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,716B, BPFP=3.5110 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,188B, BPFP=2.6844 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,304B, BPFP=3.4001 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 76,604B, BPFP=0.8593 +⌛️ [2/4] FRONTEND: Frontend time: 1.720s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.221s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13044639 28.20876845 + layer.0.v_cache 0.00001737 0.00250168 + layer.1.k_cache 0.22565776 2.68131735 + layer.1.v_cache 0.00000620 0.00101853 + layer.2.k_cache 0.01998205 0.74313753 + layer.2.v_cache 0.00001989 0.00297823 + layer.3.k_cache 0.00629339 3.59585035 + layer.3.v_cache 0.00002050 0.00331063 + layer.4.k_cache 0.00068118 0.08221310 + layer.4.v_cache 0.00005692 0.00614518 + layer.4.output 1.53833376 238.09040739 + ------------------------------------------------------------------------------------- + TOTAL 0.65597165 100.11529958 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 406352 +BPFP 1.8768 bits/point +EBPFP 3.7536 equivalent bits/point +MSE 100.115300 +---------------------- -------------------------------------------------------- +Time: 2.948s Load: 0.007s, Pack+Encode: 1.720s, Decode+Unpack: 1.221s +---------------------- -------------------------------------------------------- +💾 Converting with 100.1153 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 246, 128) +Output shape: (1, 246, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.output: torch.Size([1, 246, 3584]) -> torch.Size([1, 1, 246, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,064B, BPFP=0.7027 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,736B, BPFP=3.0955 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,588B, BPFP=1.4347 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,148B, BPFP=2.9947 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,944B, BPFP=1.7114 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,088B, BPFP=2.9273 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,008B, BPFP=1.5884 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,092B, BPFP=2.9911 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,292B, BPFP=2.4322 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,972B, BPFP=2.8565 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,064B, BPFP=0.8172 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.227s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131485 31.48406337 + layer.0.v_cache 0.00001613 0.00245111 + layer.1.k_cache 0.49701033 2.70802022 + layer.1.v_cache 0.00000620 0.00098972 + layer.2.k_cache 0.01298508 0.54144479 + layer.2.v_cache 0.00002120 0.00291827 + layer.3.k_cache 0.02762915 4.21789849 + layer.3.v_cache 0.00002136 0.00328675 + layer.4.k_cache 0.00068392 0.08179725 + layer.4.v_cache 0.00006372 0.00587994 + layer.4.output 1.24453647 200.26702236 + ------------------------------------------------------------------------------------- + TOTAL 0.55302984 84.75987685 + (elements=2,141,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2141184 +Total Bytes 447996 +BPFP 1.6738 bits/point +EBPFP 3.3477 equivalent bits/point +MSE 84.759877 +---------------------- -------------------------------------------------------- +Time: 2.963s Load: 0.010s, Pack+Encode: 1.726s, Decode+Unpack: 1.227s +---------------------- -------------------------------------------------------- +💾 Converting with 84.7599 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,632B, BPFP=0.7525 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,048B, BPFP=3.5975 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,260B, BPFP=1.5047 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,708B, BPFP=3.4928 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,476B, BPFP=1.8341 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,116B, BPFP=3.4466 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,304B, BPFP=1.6644 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,316B, BPFP=3.5403 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,800B, BPFP=2.7188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,572B, BPFP=3.4041 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,272B, BPFP=0.8289 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.224s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11599884 28.83702637 + layer.0.v_cache 0.00001671 0.00237867 + layer.1.k_cache 0.22559120 2.61399994 + layer.1.v_cache 0.00000595 0.00099033 + layer.2.k_cache 0.01146694 0.73366028 + layer.2.v_cache 0.00002201 0.00304859 + layer.3.k_cache 0.03668072 3.89318176 + layer.3.v_cache 0.00002131 0.00321745 + layer.4.k_cache 0.00066762 0.08198095 + layer.4.v_cache 0.00005146 0.00619826 + layer.4.output 1.53061454 231.30089286 + ------------------------------------------------------------------------------------- + TOTAL 0.65322497 97.36952545 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 406504 +BPFP 1.8681 bits/point +EBPFP 3.7363 equivalent bits/point +MSE 97.369525 +---------------------- -------------------------------------------------------- +Time: 2.959s Load: 0.008s, Pack+Encode: 1.726s, Decode+Unpack: 1.224s +---------------------- -------------------------------------------------------- +💾 Converting with 97.3695 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,580B, BPFP=0.7219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,224B, BPFP=3.3586 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,748B, BPFP=1.6204 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,264B, BPFP=3.2249 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,072B, BPFP=1.7789 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,384B, BPFP=3.1648 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,420B, BPFP=1.5980 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,360B, BPFP=3.2314 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,512B, BPFP=2.5595 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,652B, BPFP=3.1149 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,744B, BPFP=0.9333 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.225s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605523 29.73825566 + layer.0.v_cache 0.00001866 0.00239797 + layer.1.k_cache 0.31881074 3.10366701 + layer.1.v_cache 0.00000653 0.00098639 + layer.2.k_cache 0.02091764 0.55230020 + layer.2.v_cache 0.00002113 0.00295777 + layer.3.k_cache 0.03221333 3.62200208 + layer.3.v_cache 0.00002185 0.00320923 + layer.4.k_cache 0.00068277 0.08136163 + layer.4.v_cache 0.00005596 0.00629567 + layer.4.output 1.33695214 210.62962024 + ------------------------------------------------------------------------------------- + TOTAL 0.57926287 88.91298678 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 452960 +BPFP 1.8180 bits/point +EBPFP 3.6360 equivalent bits/point +MSE 88.912987 +---------------------- -------------------------------------------------------- +Time: 2.961s Load: 0.009s, Pack+Encode: 1.726s, Decode+Unpack: 1.225s +---------------------- -------------------------------------------------------- +💾 Converting with 88.9130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 192, 128) +Output shape: (1, 192, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.output: torch.Size([1, 192, 3584]) -> torch.Size([1, 1, 192, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,076B, BPFP=0.6572 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,964B, BPFP=2.9268 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,200B, BPFP=1.3997 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,056B, BPFP=2.8529 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,496B, BPFP=1.5866 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,504B, BPFP=2.8079 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,108B, BPFP=1.4736 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,200B, BPFP=2.8646 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,992B, BPFP=2.3594 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,828B, BPFP=2.7529 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,328B, BPFP=1.1083 +⌛️ [2/4] FRONTEND: Frontend time: 1.613s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.117s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12444629 35.48327128 + layer.0.v_cache 0.00001540 0.00215353 + layer.1.k_cache 0.18411521 2.94217809 + layer.1.v_cache 0.00000585 0.00088911 + layer.2.k_cache 0.01124177 0.48932135 + layer.2.v_cache 0.00002179 0.00281712 + layer.3.k_cache 0.01616090 4.01426379 + layer.3.v_cache 0.00002117 0.00308835 + layer.4.k_cache 0.00068507 0.07973532 + layer.4.v_cache 0.00005275 0.00570298 + layer.4.output 0.00839530 261.80459449 + ------------------------------------------------------------------------------------- + TOTAL 0.02326666 110.33268132 + (elements=1,671,168) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1671168 +Total Bytes 361752 +BPFP 1.7317 bits/point +EBPFP 3.4635 equivalent bits/point +MSE 110.332681 +---------------------- -------------------------------------------------------- +Time: 2.737s Load: 0.007s, Pack+Encode: 1.613s, Decode+Unpack: 1.117s +---------------------- -------------------------------------------------------- +💾 Converting with 110.3327 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,096B, BPFP=0.7620 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,232B, BPFP=3.4104 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,220B, BPFP=1.4326 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,468B, BPFP=3.3385 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,428B, BPFP=1.8287 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,748B, BPFP=3.2707 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,576B, BPFP=1.6544 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,312B, BPFP=3.3238 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,408B, BPFP=2.6739 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,888B, BPFP=3.1898 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,136B, BPFP=0.8355 +⌛️ [2/4] FRONTEND: Frontend time: 1.615s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.113s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09999813 29.28384730 + layer.0.v_cache 0.00001899 0.00252559 + layer.1.k_cache 0.12989934 2.15495043 + layer.1.v_cache 0.00000610 0.00098254 + layer.2.k_cache 0.00966061 0.54225885 + layer.2.v_cache 0.00001994 0.00304580 + layer.3.k_cache 0.03384685 3.71275789 + layer.3.v_cache 0.00002040 0.00324163 + layer.4.k_cache 0.00066502 0.08151385 + layer.4.v_cache 0.00005200 0.00637943 + layer.4.output 0.00959846 284.33565512 + ------------------------------------------------------------------------------------- + TOTAL 0.02008098 119.18476995 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 326512 +BPFP 1.8078 bits/point +EBPFP 3.6157 equivalent bits/point +MSE 119.184770 +---------------------- -------------------------------------------------------- +Time: 2.733s Load: 0.006s, Pack+Encode: 1.615s, Decode+Unpack: 1.113s +---------------------- -------------------------------------------------------- +💾 Converting with 119.1848 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 428, 128) +Output shape: (1, 428, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.output: torch.Size([1, 428, 3584]) -> torch.Size([1, 1, 428, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 18,608B, BPFP=0.6793 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 85,240B, BPFP=3.1119 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 42,972B, BPFP=1.5688 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 82,160B, BPFP=2.9994 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 46,720B, BPFP=1.7056 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 80,640B, BPFP=2.9439 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 41,296B, BPFP=1.5076 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 81,960B, BPFP=2.9921 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 66,684B, BPFP=2.4344 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 79,088B, BPFP=2.8873 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 159,484B, BPFP=0.8318 +⌛️ [2/4] FRONTEND: Frontend time: 2.099s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.586s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13760449 33.13707926 + layer.0.v_cache 0.00001624 0.00222944 + layer.1.k_cache 0.89392254 3.24244533 + layer.1.v_cache 0.00000615 0.00088283 + layer.2.k_cache 0.03917003 0.62155237 + layer.2.v_cache 0.00002113 0.00282218 + layer.3.k_cache 0.03749864 4.03375016 + layer.3.v_cache 0.00002059 0.00287788 + layer.4.k_cache 0.00086919 0.08213537 + layer.4.v_cache 0.00005177 0.00581607 + layer.4.output 0.00617707 114.10656918 + ------------------------------------------------------------------------------------- + TOTAL 0.06778943 49.40456324 + (elements=3,725,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3725312 +Total Bytes 784852 +BPFP 1.6854 bits/point +EBPFP 3.3709 equivalent bits/point +MSE 49.404563 +---------------------- -------------------------------------------------------- +Time: 3.699s Load: 0.014s, Pack+Encode: 2.099s, Decode+Unpack: 1.586s +---------------------- -------------------------------------------------------- +💾 Converting with 49.4046 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,772B, BPFP=0.7077 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,248B, BPFP=3.2828 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,736B, BPFP=1.5922 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,660B, BPFP=3.1948 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,708B, BPFP=1.7569 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,256B, BPFP=3.1724 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,132B, BPFP=1.5587 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,520B, BPFP=3.2425 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,364B, BPFP=2.6243 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,440B, BPFP=3.1272 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 107,560B, BPFP=0.8514 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.367s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385377 29.75867132 + layer.0.v_cache 0.00001705 0.00231099 + layer.1.k_cache 0.49372572 2.93630397 + layer.1.v_cache 0.00000626 0.00091897 + layer.2.k_cache 0.02422354 0.56939838 + layer.2.v_cache 0.00002046 0.00280194 + layer.3.k_cache 0.04153839 3.58442850 + layer.3.v_cache 0.00002075 0.00298542 + layer.4.k_cache 0.00074884 0.08375030 + layer.4.v_cache 0.00004954 0.00589309 + layer.4.output 0.00474379 179.68336816 + ------------------------------------------------------------------------------------- + TOTAL 0.04161240 76.16064941 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 545396 +BPFP 1.7776 bits/point +EBPFP 3.5552 equivalent bits/point +MSE 76.160649 +---------------------- -------------------------------------------------------- +Time: 3.238s Load: 0.014s, Pack+Encode: 1.857s, Decode+Unpack: 1.367s +---------------------- -------------------------------------------------------- +💾 Converting with 76.1606 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,436B, BPFP=0.6983 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 68,860B, BPFP=3.3311 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 33,772B, BPFP=1.6337 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 66,724B, BPFP=3.2277 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 34,620B, BPFP=1.6747 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 65,388B, BPFP=3.1631 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,404B, BPFP=1.5192 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 67,872B, BPFP=3.2833 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 51,944B, BPFP=2.5128 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 65,216B, BPFP=3.1548 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 127,000B, BPFP=0.8777 +⌛️ [2/4] FRONTEND: Frontend time: 2.205s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.487s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12985682 31.93502685 + layer.0.v_cache 0.00001645 0.00224890 + layer.1.k_cache 0.61935935 3.23869130 + layer.1.v_cache 0.00000656 0.00092347 + layer.2.k_cache 0.01492825 0.67446465 + layer.2.v_cache 0.00002178 0.00275618 + layer.3.k_cache 0.02579401 3.61869618 + layer.3.v_cache 0.00002068 0.00291976 + layer.4.k_cache 0.00071613 0.08148611 + layer.4.v_cache 0.00005057 0.00577482 + layer.4.output 0.04141478 146.38251880 + ------------------------------------------------------------------------------------- + TOTAL 0.06356906 62.60238940 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 627236 +BPFP 1.7848 bits/point +EBPFP 3.5697 equivalent bits/point +MSE 62.602389 +---------------------- -------------------------------------------------------- +Time: 3.705s Load: 0.013s, Pack+Encode: 2.205s, Decode+Unpack: 1.487s +---------------------- -------------------------------------------------------- +💾 Converting with 62.6024 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,784B, BPFP=0.6984 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,792B, BPFP=3.2666 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,020B, BPFP=1.5854 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,048B, BPFP=3.1713 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,596B, BPFP=1.7262 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,008B, BPFP=3.1145 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,960B, BPFP=1.5822 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,460B, BPFP=3.1938 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,228B, BPFP=2.5256 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,184B, BPFP=3.0695 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 109,904B, BPFP=0.8578 +⌛️ [2/4] FRONTEND: Frontend time: 1.884s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12359551 30.60317417 + layer.0.v_cache 0.00001766 0.00239424 + layer.1.k_cache 0.52692803 3.10666640 + layer.1.v_cache 0.00000637 0.00094400 + layer.2.k_cache 0.02102186 0.57449704 + layer.2.v_cache 0.00002187 0.00288454 + layer.3.k_cache 0.03868174 3.80937099 + layer.3.v_cache 0.00002097 0.00314977 + layer.4.k_cache 0.00070799 0.08333629 + layer.4.v_cache 0.00005204 0.00602723 + layer.4.output 0.00469152 177.36746066 + ------------------------------------------------------------------------------------- + TOTAL 0.04375851 75.28027467 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 547984 +BPFP 1.7611 bits/point +EBPFP 3.5221 equivalent bits/point +MSE 75.280275 +---------------------- -------------------------------------------------------- +Time: 3.270s Load: 0.011s, Pack+Encode: 1.884s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 75.2803 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,548B, BPFP=0.7497 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,172B, BPFP=3.6253 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,264B, BPFP=1.4340 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,516B, BPFP=3.4953 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,952B, BPFP=1.8021 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,704B, BPFP=3.4315 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,060B, BPFP=1.6536 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,924B, BPFP=3.5273 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,272B, BPFP=2.6910 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,204B, BPFP=3.3923 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 79,108B, BPFP=0.8873 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.219s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10671347 28.56180316 + layer.0.v_cache 0.00001788 0.00246594 + layer.1.k_cache 0.15257932 2.18075025 + layer.1.v_cache 0.00000642 0.00096850 + layer.2.k_cache 0.02197224 0.65610945 + layer.2.v_cache 0.00002094 0.00289779 + layer.3.k_cache 0.03325213 3.43994018 + layer.3.v_cache 0.00002126 0.00328065 + layer.4.k_cache 0.00067229 0.08346374 + layer.4.v_cache 0.00005023 0.00593856 + layer.4.output 1.53834953 243.57898869 + ------------------------------------------------------------------------------------- + TOTAL 0.65198546 102.35238465 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 407724 +BPFP 1.8831 bits/point +EBPFP 3.7663 equivalent bits/point +MSE 102.352385 +---------------------- -------------------------------------------------------- +Time: 2.953s Load: 0.008s, Pack+Encode: 1.726s, Decode+Unpack: 1.219s +---------------------- -------------------------------------------------------- +💾 Converting with 102.3524 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,228B, BPFP=0.6990 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 50,004B, BPFP=3.1128 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,456B, BPFP=1.4602 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,736B, BPFP=2.9716 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,736B, BPFP=1.7266 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,884B, BPFP=2.9186 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,676B, BPFP=1.5984 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,844B, BPFP=2.9783 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,172B, BPFP=2.5007 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,756B, BPFP=2.8484 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,828B, BPFP=0.8878 +⌛️ [2/4] FRONTEND: Frontend time: 1.729s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.231s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15214830 34.20997261 + layer.0.v_cache 0.00001722 0.00253213 + layer.1.k_cache 0.49754258 3.00861884 + layer.1.v_cache 0.00000639 0.00104674 + layer.2.k_cache 0.01862443 0.54864502 + layer.2.v_cache 0.00002219 0.00311947 + layer.3.k_cache 0.01200796 4.09530871 + layer.3.v_cache 0.00002210 0.00336756 + layer.4.k_cache 0.00075226 0.08602351 + layer.4.v_cache 0.00005409 0.00634307 + layer.4.output 1.21979868 200.30860487 + ------------------------------------------------------------------------------------- + TOTAL 0.54234049 84.94854187 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 466320 +BPFP 1.7076 bits/point +EBPFP 3.4152 equivalent bits/point +MSE 84.948542 +---------------------- -------------------------------------------------------- +Time: 2.968s Load: 0.008s, Pack+Encode: 1.729s, Decode+Unpack: 1.231s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9485 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,492B, BPFP=0.6912 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,840B, BPFP=3.1168 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 32,200B, BPFP=1.6496 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 59,136B, BPFP=3.0295 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,632B, BPFP=1.7230 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,860B, BPFP=2.9641 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,576B, BPFP=1.6176 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 59,132B, BPFP=3.0293 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,888B, BPFP=2.4533 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,788B, BPFP=2.9092 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 114,372B, BPFP=0.8370 +⌛️ [2/4] FRONTEND: Frontend time: 1.863s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.360s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18432382 31.96496222 + layer.0.v_cache 0.00001629 0.00239599 + layer.1.k_cache 0.53789903 3.14270780 + layer.1.v_cache 0.00000607 0.00096405 + layer.2.k_cache 0.01514517 0.52405250 + layer.2.v_cache 0.00002151 0.00288647 + layer.3.k_cache 0.04182221 4.02368644 + layer.3.v_cache 0.00002017 0.00310429 + layer.4.k_cache 0.00068627 0.08145462 + layer.4.v_cache 0.00005063 0.00606262 + layer.4.output 0.04375917 57.79598946 + ------------------------------------------------------------------------------------- + TOTAL 0.06390032 26.13671784 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 566916 +BPFP 1.7084 bits/point +EBPFP 3.4168 equivalent bits/point +MSE 26.136718 +---------------------- -------------------------------------------------------- +Time: 3.236s Load: 0.013s, Pack+Encode: 1.863s, Decode+Unpack: 1.360s +---------------------- -------------------------------------------------------- +💾 Converting with 26.1367 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 371, 128) +Output shape: (1, 371, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.output: torch.Size([1, 371, 3584]) -> torch.Size([1, 1, 371, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,196B, BPFP=0.6821 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 73,092B, BPFP=3.0783 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 40,188B, BPFP=1.6926 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 71,112B, BPFP=2.9949 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 40,524B, BPFP=1.7067 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 69,880B, BPFP=2.9431 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 37,304B, BPFP=1.5711 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 71,060B, BPFP=2.9928 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 57,524B, BPFP=2.4227 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 68,028B, BPFP=2.8651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 145,076B, BPFP=0.8729 +⌛️ [2/4] FRONTEND: Frontend time: 1.982s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.482s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16952868 32.34107038 + layer.0.v_cache 0.00001698 0.00240657 + layer.1.k_cache 0.82091874 3.22812704 + layer.1.v_cache 0.00000666 0.00098899 + layer.2.k_cache 0.02418445 0.56514763 + layer.2.v_cache 0.00002238 0.00295956 + layer.3.k_cache 0.03055777 4.30106849 + layer.3.v_cache 0.00002117 0.00312939 + layer.4.k_cache 0.00071103 0.08229740 + layer.4.v_cache 0.00005915 0.00606457 + layer.4.output 0.03610923 50.48399596 + ------------------------------------------------------------------------------------- + TOTAL 0.07639951 23.17183716 + (elements=3,229,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3229184 +Total Bytes 689984 +BPFP 1.7094 bits/point +EBPFP 3.4187 equivalent bits/point +MSE 23.171837 +---------------------- -------------------------------------------------------- +Time: 3.478s Load: 0.014s, Pack+Encode: 1.982s, Decode+Unpack: 1.482s +---------------------- -------------------------------------------------------- +💾 Converting with 23.1718 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,148B, BPFP=0.7274 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,832B, BPFP=3.4283 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,604B, BPFP=1.4768 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,240B, BPFP=3.3142 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,488B, BPFP=1.8268 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,720B, BPFP=3.2769 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,508B, BPFP=1.6849 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,908B, BPFP=3.3621 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,500B, BPFP=2.6161 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,008B, BPFP=3.2259 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,252B, BPFP=0.8832 +⌛️ [2/4] FRONTEND: Frontend time: 1.731s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.241s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13674593 29.47760626 + layer.0.v_cache 0.00001768 0.00261637 + layer.1.k_cache 0.26206151 2.42576865 + layer.1.v_cache 0.00000624 0.00102988 + layer.2.k_cache 0.02938290 0.76796232 + layer.2.v_cache 0.00002158 0.00303895 + layer.3.k_cache 0.05653211 3.67536268 + layer.3.v_cache 0.00002129 0.00334635 + layer.4.k_cache 0.00070776 0.08360194 + layer.4.v_cache 0.00005193 0.00636125 + layer.4.output 1.40434551 229.94055128 + ------------------------------------------------------------------------------------- + TOTAL 0.60682162 96.82532668 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 434208 +BPFP 1.8307 bits/point +EBPFP 3.6614 equivalent bits/point +MSE 96.825327 +---------------------- -------------------------------------------------------- +Time: 2.981s Load: 0.009s, Pack+Encode: 1.731s, Decode+Unpack: 1.241s +---------------------- -------------------------------------------------------- +💾 Converting with 96.8253 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 193, 128) +Output shape: (1, 193, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.output: torch.Size([1, 193, 3584]) -> torch.Size([1, 1, 193, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,948B, BPFP=0.7244 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,200B, BPFP=3.6593 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,836B, BPFP=1.4440 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,004B, BPFP=3.5625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,156B, BPFP=1.7937 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,324B, BPFP=3.5074 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,272B, BPFP=1.6412 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,644B, BPFP=3.6143 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,368B, BPFP=2.7014 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,188B, BPFP=3.4964 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,612B, BPFP=0.9092 +⌛️ [2/4] FRONTEND: Frontend time: 1.733s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.229s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14305212 28.57908132 + layer.0.v_cache 0.00001770 0.00257122 + layer.1.k_cache 0.14179654 2.61312028 + layer.1.v_cache 0.00000608 0.00099754 + layer.2.k_cache 0.01555968 0.74702952 + layer.2.v_cache 0.00002047 0.00303426 + layer.3.k_cache 0.02655400 3.32969737 + layer.3.v_cache 0.00002031 0.00343878 + layer.4.k_cache 0.00069483 0.08371387 + layer.4.v_cache 0.00005004 0.00627987 + layer.4.output 1.58615237 253.91927276 + ------------------------------------------------------------------------------------- + TOTAL 0.67240225 106.63552196 + (elements=1,679,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1679872 +Total Bytes 401552 +BPFP 1.9123 bits/point +EBPFP 3.8246 equivalent bits/point +MSE 106.635522 +---------------------- -------------------------------------------------------- +Time: 2.968s Load: 0.007s, Pack+Encode: 1.733s, Decode+Unpack: 1.229s +---------------------- -------------------------------------------------------- +💾 Converting with 106.6355 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,164B, BPFP=0.7006 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,544B, BPFP=3.0462 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,240B, BPFP=1.3956 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,936B, BPFP=2.9453 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,628B, BPFP=1.6709 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,344B, BPFP=2.9081 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,384B, BPFP=1.5301 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,196B, BPFP=2.9616 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,648B, BPFP=2.3624 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,140B, BPFP=2.8326 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,816B, BPFP=0.8589 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.231s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10821181 30.44112779 + layer.0.v_cache 0.00001942 0.00233645 + layer.1.k_cache 0.30751975 3.09182457 + layer.1.v_cache 0.00000599 0.00090128 + layer.2.k_cache 0.01421910 0.55334730 + layer.2.v_cache 0.00002020 0.00282726 + layer.3.k_cache 0.01730646 4.46528350 + layer.3.v_cache 0.00002068 0.00320173 + layer.4.k_cache 0.00075739 0.08282625 + layer.4.v_cache 0.00005209 0.00610415 + layer.4.output 1.22958295 202.24777682 + ------------------------------------------------------------------------------------- + TOTAL 0.53265962 85.55201282 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 452040 +BPFP 1.6686 bits/point +EBPFP 3.3372 equivalent bits/point +MSE 85.552013 +---------------------- -------------------------------------------------------- +Time: 2.973s Load: 0.011s, Pack+Encode: 1.730s, Decode+Unpack: 1.231s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5520 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 202, 128) +Output shape: (1, 202, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.output: torch.Size([1, 202, 3584]) -> torch.Size([1, 1, 202, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,688B, BPFP=0.7494 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,400B, BPFP=3.5891 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,704B, BPFP=1.4468 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,292B, BPFP=3.5034 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,888B, BPFP=1.8478 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,236B, BPFP=3.4217 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,768B, BPFP=1.6838 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,048B, BPFP=3.5619 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,968B, BPFP=2.7048 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,660B, BPFP=3.3772 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,720B, BPFP=0.9251 +⌛️ [2/4] FRONTEND: Frontend time: 1.727s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.226s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15282421 27.36194210 + layer.0.v_cache 0.00001801 0.00266312 + layer.1.k_cache 0.08597872 1.71095155 + layer.1.v_cache 0.00000649 0.00103806 + layer.2.k_cache 0.01425038 0.56926727 + layer.2.v_cache 0.00002047 0.00298954 + layer.3.k_cache 0.05704349 3.53470513 + layer.3.v_cache 0.00002256 0.00347502 + layer.4.k_cache 0.00069767 0.08389113 + layer.4.v_cache 0.00005056 0.00635896 + layer.4.output 1.51553867 247.16104579 + ------------------------------------------------------------------------------------- + TOTAL 0.64233431 103.72968250 + (elements=1,758,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1758208 +Total Bytes 418372 +BPFP 1.9036 bits/point +EBPFP 3.8073 equivalent bits/point +MSE 103.729682 +---------------------- -------------------------------------------------------- +Time: 2.959s Load: 0.007s, Pack+Encode: 1.727s, Decode+Unpack: 1.226s +---------------------- -------------------------------------------------------- +💾 Converting with 103.7297 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,816B, BPFP=0.7076 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,528B, BPFP=3.2314 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,948B, BPFP=1.6535 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,584B, BPFP=3.1793 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,752B, BPFP=1.6979 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,920B, BPFP=3.0875 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,544B, BPFP=1.5208 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,456B, BPFP=3.1723 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,852B, BPFP=2.4764 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,496B, BPFP=3.0640 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,384B, BPFP=0.7839 +⌛️ [2/4] FRONTEND: Frontend time: 1.861s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14837397 30.12601797 + layer.0.v_cache 0.00001603 0.00214703 + layer.1.k_cache 0.60911565 3.03842088 + layer.1.v_cache 0.00000597 0.00085302 + layer.2.k_cache 0.02477714 0.65523877 + layer.2.v_cache 0.00001938 0.00269463 + layer.3.k_cache 0.01052417 3.81117017 + layer.3.v_cache 0.00002039 0.00292114 + layer.4.k_cache 0.00073933 0.07817800 + layer.4.v_cache 0.00005963 0.00610372 + layer.4.output 0.00470031 178.84390775 + ------------------------------------------------------------------------------------- + TOTAL 0.04862081 75.86065292 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 530280 +BPFP 1.7222 bits/point +EBPFP 3.4445 equivalent bits/point +MSE 75.860653 +---------------------- -------------------------------------------------------- +Time: 3.234s Load: 0.009s, Pack+Encode: 1.861s, Decode+Unpack: 1.364s +---------------------- -------------------------------------------------------- +💾 Converting with 75.8607 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 210, 128) +Output shape: (1, 210, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.output: torch.Size([1, 210, 3584]) -> torch.Size([1, 1, 210, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,004B, BPFP=0.7443 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,908B, BPFP=3.4902 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,160B, BPFP=1.7232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,160B, BPFP=3.3601 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,552B, BPFP=1.8268 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,616B, BPFP=3.3196 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,136B, BPFP=1.6470 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,236B, BPFP=3.4402 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,660B, BPFP=2.6533 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,432B, BPFP=3.3060 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,136B, BPFP=0.8943 +⌛️ [2/4] FRONTEND: Frontend time: 1.744s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.227s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11415643 27.68166155 + layer.0.v_cache 0.00001882 0.00241667 + layer.1.k_cache 0.27145578 2.74577143 + layer.1.v_cache 0.00000623 0.00100749 + layer.2.k_cache 0.01201470 0.65613156 + layer.2.v_cache 0.00002208 0.00297485 + layer.3.k_cache 0.00887682 3.38546288 + layer.3.v_cache 0.00002165 0.00332691 + layer.4.k_cache 0.00072611 0.08338501 + layer.4.v_cache 0.00005233 0.00632666 + layer.4.output 1.45781997 107.57918793 + ------------------------------------------------------------------------------------- + TOTAL 0.62424063 46.33075179 + (elements=1,827,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1827840 +Total Bytes 427000 +BPFP 1.8689 bits/point +EBPFP 3.7377 equivalent bits/point +MSE 46.330752 +---------------------- -------------------------------------------------------- +Time: 2.979s Load: 0.008s, Pack+Encode: 1.744s, Decode+Unpack: 1.227s +---------------------- -------------------------------------------------------- +💾 Converting with 46.3308 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 165, 128) +Output shape: (1, 165, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.output: torch.Size([1, 165, 3584]) -> torch.Size([1, 1, 165, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,040B, BPFP=0.7614 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,668B, BPFP=3.4723 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,700B, BPFP=1.4867 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,676B, BPFP=3.3784 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,540B, BPFP=1.8504 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,852B, BPFP=3.3004 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,852B, BPFP=1.6905 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,968B, BPFP=3.4061 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,584B, BPFP=2.7068 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,212B, BPFP=3.2398 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,500B, BPFP=0.9402 +⌛️ [2/4] FRONTEND: Frontend time: 1.610s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.112s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10245823 29.18649976 + layer.0.v_cache 0.00002070 0.00268435 + layer.1.k_cache 0.13127365 2.71073312 + layer.1.v_cache 0.00000615 0.00101235 + layer.2.k_cache 0.01458293 0.51887748 + layer.2.v_cache 0.00002074 0.00303176 + layer.3.k_cache 0.02375244 3.70733458 + layer.3.v_cache 0.00002214 0.00348425 + layer.4.k_cache 0.00067489 0.08661666 + layer.4.v_cache 0.00005196 0.00659286 + layer.4.output 0.00973386 291.03360390 + ------------------------------------------------------------------------------------- + TOTAL 0.02005887 121.96835850 + (elements=1,436,160) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1436160 +Total Bytes 336592 +BPFP 1.8750 bits/point +EBPFP 3.7499 equivalent bits/point +MSE 121.968358 +---------------------- -------------------------------------------------------- +Time: 2.728s Load: 0.005s, Pack+Encode: 1.610s, Decode+Unpack: 1.112s +---------------------- -------------------------------------------------------- +💾 Converting with 121.9684 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,596B, BPFP=0.7339 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,992B, BPFP=3.1585 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,004B, BPFP=1.5372 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,908B, BPFP=3.0659 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,944B, BPFP=1.7883 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,152B, BPFP=3.0014 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,452B, BPFP=1.6609 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,864B, BPFP=3.0622 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,480B, BPFP=2.5171 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,396B, BPFP=2.9368 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 76,276B, BPFP=0.9304 +⌛️ [2/4] FRONTEND: Frontend time: 1.614s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.113s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12273688 31.10233361 + layer.0.v_cache 0.00001959 0.00273605 + layer.1.k_cache 0.11126613 2.94180181 + layer.1.v_cache 0.00000674 0.00113318 + layer.2.k_cache 0.01416154 0.54229695 + layer.2.v_cache 0.00002164 0.00326503 + layer.3.k_cache 0.04491856 3.88859950 + layer.3.v_cache 0.00002114 0.00372115 + layer.4.k_cache 0.00068813 0.08872089 + layer.4.v_cache 0.00006151 0.00662968 + layer.4.output 0.00884247 411.47199454 + ------------------------------------------------------------------------------------- + TOTAL 0.02092936 171.69912939 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 351064 +BPFP 1.7632 bits/point +EBPFP 3.5264 equivalent bits/point +MSE 171.699129 +---------------------- -------------------------------------------------------- +Time: 2.733s Load: 0.006s, Pack+Encode: 1.614s, Decode+Unpack: 1.113s +---------------------- -------------------------------------------------------- +💾 Converting with 171.6991 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 257, 128) +Output shape: (1, 257, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.output: torch.Size([1, 257, 3584]) -> torch.Size([1, 1, 257, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,524B, BPFP=0.7006 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,096B, BPFP=3.4713 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,908B, BPFP=1.5143 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 54,896B, BPFP=3.3375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,552B, BPFP=1.7359 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,720B, BPFP=3.2661 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,844B, BPFP=1.5713 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,468B, BPFP=3.4331 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,716B, BPFP=2.5362 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,264B, BPFP=3.2383 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,652B, BPFP=0.8568 +⌛️ [2/4] FRONTEND: Frontend time: 1.844s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12694686 29.92977110 + layer.0.v_cache 0.00002055 0.00237182 + layer.1.k_cache 0.54844232 2.90272872 + layer.1.v_cache 0.00000676 0.00094693 + layer.2.k_cache 0.01835482 0.69805469 + layer.2.v_cache 0.00002099 0.00285862 + layer.3.k_cache 0.02141366 3.54598227 + layer.3.v_cache 0.00002047 0.00307267 + layer.4.k_cache 0.00070767 0.07978097 + layer.4.v_cache 0.00005046 0.00593325 + layer.4.output 0.00513101 193.51987215 + ------------------------------------------------------------------------------------- + TOTAL 0.04422951 81.87121212 + (elements=2,236,928) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2236928 +Total Bytes 506640 +BPFP 1.8119 bits/point +EBPFP 3.6238 equivalent bits/point +MSE 81.871212 +---------------------- -------------------------------------------------------- +Time: 3.206s Load: 0.008s, Pack+Encode: 1.844s, Decode+Unpack: 1.353s +---------------------- -------------------------------------------------------- +💾 Converting with 81.8712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,672B, BPFP=0.7445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,708B, BPFP=3.1514 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,684B, BPFP=1.4323 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,428B, BPFP=3.0416 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,208B, BPFP=1.7349 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,776B, BPFP=2.9856 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,068B, BPFP=1.6370 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,552B, BPFP=3.0522 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,888B, BPFP=2.4801 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,004B, BPFP=2.9193 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,608B, BPFP=0.8537 +⌛️ [2/4] FRONTEND: Frontend time: 1.611s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.116s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12731336 32.75034341 + layer.0.v_cache 0.00001750 0.00269613 + layer.1.k_cache 0.10052182 2.63932348 + layer.1.v_cache 0.00000572 0.00096457 + layer.2.k_cache 0.01141533 0.55205460 + layer.2.v_cache 0.00001926 0.00287039 + layer.3.k_cache 0.03745206 4.49117204 + layer.3.v_cache 0.00002083 0.00343883 + layer.4.k_cache 0.00066429 0.08765908 + layer.4.v_cache 0.00005101 0.00627328 + layer.4.output 0.00884527 270.66652767 + ------------------------------------------------------------------------------------- + TOTAL 0.01996459 113.83544056 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 339596 +BPFP 1.7150 bits/point +EBPFP 3.4300 equivalent bits/point +MSE 113.835441 +---------------------- -------------------------------------------------------- +Time: 2.734s Load: 0.006s, Pack+Encode: 1.611s, Decode+Unpack: 1.116s +---------------------- -------------------------------------------------------- +💾 Converting with 113.8354 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,300B, BPFP=0.7249 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,428B, BPFP=3.4085 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,968B, BPFP=1.6166 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,384B, BPFP=3.2646 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,064B, BPFP=1.7641 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,780B, BPFP=3.2221 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,944B, BPFP=1.6149 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,164B, BPFP=3.3195 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,992B, BPFP=2.5332 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,404B, BPFP=3.1957 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,648B, BPFP=0.8712 +⌛️ [2/4] FRONTEND: Frontend time: 1.725s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.230s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12752426 30.22662848 + layer.0.v_cache 0.00002004 0.00246168 + layer.1.k_cache 0.30660921 3.03371890 + layer.1.v_cache 0.00000613 0.00091628 + layer.2.k_cache 0.01808885 0.57686759 + layer.2.v_cache 0.00002107 0.00281557 + layer.3.k_cache 0.01678786 3.75698715 + layer.3.v_cache 0.00002050 0.00299591 + layer.4.k_cache 0.00068290 0.07981314 + layer.4.v_cache 0.00004851 0.00587780 + layer.4.output 1.37905047 224.74167471 + ------------------------------------------------------------------------------------- + TOTAL 0.59548016 94.75769444 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 437076 +BPFP 1.8096 bits/point +EBPFP 3.6191 equivalent bits/point +MSE 94.757694 +---------------------- -------------------------------------------------------- +Time: 2.962s Load: 0.007s, Pack+Encode: 1.725s, Decode+Unpack: 1.230s +---------------------- -------------------------------------------------------- +💾 Converting with 94.7577 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,132B, BPFP=0.7296 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,280B, BPFP=3.4044 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,872B, BPFP=1.4309 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,096B, BPFP=3.3191 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,044B, BPFP=1.8033 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,288B, BPFP=3.2609 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,812B, BPFP=1.6426 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,908B, BPFP=3.3776 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,644B, BPFP=2.5665 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,876B, BPFP=3.2313 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,224B, BPFP=0.9178 +⌛️ [2/4] FRONTEND: Frontend time: 1.727s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.228s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12766143 29.48473502 + layer.0.v_cache 0.00001653 0.00244145 + layer.1.k_cache 0.34631467 2.49694065 + layer.1.v_cache 0.00000659 0.00095830 + layer.2.k_cache 0.01786991 0.73568170 + layer.2.v_cache 0.00002041 0.00280892 + layer.3.k_cache 0.06874437 3.57601493 + layer.3.v_cache 0.00002276 0.00326469 + layer.4.k_cache 0.00067882 0.07990970 + layer.4.v_cache 0.00005621 0.00666211 + layer.4.output 1.41085444 233.36983624 + ------------------------------------------------------------------------------------- + TOTAL 0.61396310 98.23401595 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 433176 +BPFP 1.8347 bits/point +EBPFP 3.6695 equivalent bits/point +MSE 98.234016 +---------------------- -------------------------------------------------------- +Time: 2.962s Load: 0.007s, Pack+Encode: 1.727s, Decode+Unpack: 1.228s +---------------------- -------------------------------------------------------- +💾 Converting with 98.2340 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,420B, BPFP=0.7204 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,944B, BPFP=3.3838 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,212B, BPFP=1.6048 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,464B, BPFP=3.2815 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,188B, BPFP=1.8106 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,716B, BPFP=3.2298 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,220B, BPFP=1.6054 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,744B, BPFP=3.3009 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,520B, BPFP=2.5940 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,628B, BPFP=3.1546 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,948B, BPFP=0.9872 +⌛️ [2/4] FRONTEND: Frontend time: 1.728s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.230s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12070838 30.07133443 + layer.0.v_cache 0.00001763 0.00247567 + layer.1.k_cache 0.31015487 3.06070892 + layer.1.v_cache 0.00000677 0.00099890 + layer.2.k_cache 0.03544329 0.54497278 + layer.2.v_cache 0.00002212 0.00289225 + layer.3.k_cache 0.05682086 3.93516986 + layer.3.v_cache 0.00002136 0.00328278 + layer.4.k_cache 0.00073636 0.08679814 + layer.4.v_cache 0.00005031 0.00612267 + layer.4.output 1.35470685 211.55667272 + ------------------------------------------------------------------------------------- + TOTAL 0.58864294 89.33008620 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 457004 +BPFP 1.8586 bits/point +EBPFP 3.7172 equivalent bits/point +MSE 89.330086 +---------------------- -------------------------------------------------------- +Time: 2.965s Load: 0.008s, Pack+Encode: 1.728s, Decode+Unpack: 1.230s +---------------------- -------------------------------------------------------- +💾 Converting with 89.3301 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,076B, BPFP=0.7391 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,148B, BPFP=3.4586 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,176B, BPFP=1.4800 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,292B, BPFP=3.3958 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,292B, BPFP=1.8553 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,276B, BPFP=3.3213 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,752B, BPFP=1.6690 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,696B, BPFP=3.4255 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,880B, BPFP=2.6320 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,876B, BPFP=3.2920 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,532B, BPFP=0.9487 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.228s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13227103 29.35551689 + layer.0.v_cache 0.00002050 0.00244729 + layer.1.k_cache 0.22864332 2.49514985 + layer.1.v_cache 0.00000662 0.00100852 + layer.2.k_cache 0.02950178 0.69461246 + layer.2.v_cache 0.00002095 0.00291723 + layer.3.k_cache 0.01835468 3.54987289 + layer.3.v_cache 0.00002123 0.00331139 + layer.4.k_cache 0.00070999 0.08594778 + layer.4.v_cache 0.00005457 0.00618052 + layer.4.output 1.43734575 220.67165493 + ------------------------------------------------------------------------------------- + TOTAL 0.61594264 92.99403231 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 434996 +BPFP 1.8771 bits/point +EBPFP 3.7541 equivalent bits/point +MSE 92.994032 +---------------------- -------------------------------------------------------- +Time: 2.963s Load: 0.008s, Pack+Encode: 1.726s, Decode+Unpack: 1.228s +---------------------- -------------------------------------------------------- +💾 Converting with 92.9940 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,304B, BPFP=0.7252 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,412B, BPFP=3.3370 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,996B, BPFP=1.5481 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,396B, BPFP=3.2655 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,716B, BPFP=1.8100 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,680B, BPFP=3.2151 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,752B, BPFP=1.6014 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,924B, BPFP=3.3026 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,524B, BPFP=2.5707 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,660B, BPFP=3.1433 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,720B, BPFP=0.8719 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.229s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12518616 29.39070418 + layer.0.v_cache 0.00001676 0.00244738 + layer.1.k_cache 0.33007166 2.86755344 + layer.1.v_cache 0.00000630 0.00100472 + layer.2.k_cache 0.01131002 0.63699650 + layer.2.v_cache 0.00002125 0.00295581 + layer.3.k_cache 0.02232737 3.90847971 + layer.3.v_cache 0.00001983 0.00314905 + layer.4.k_cache 0.00071526 0.08392671 + layer.4.v_cache 0.00005006 0.00607082 + layer.4.output 1.37904434 225.03048584 + ------------------------------------------------------------------------------------- + TOTAL 0.59664912 94.83039348 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 435084 +BPFP 1.8013 bits/point +EBPFP 3.6026 equivalent bits/point +MSE 94.830393 +---------------------- -------------------------------------------------------- +Time: 2.963s Load: 0.009s, Pack+Encode: 1.726s, Decode+Unpack: 1.229s +---------------------- -------------------------------------------------------- +💾 Converting with 94.8304 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,076B, BPFP=0.7093 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,316B, BPFP=3.0940 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,952B, BPFP=1.4698 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,000B, BPFP=3.0097 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,612B, BPFP=1.7041 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,016B, BPFP=2.9467 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,624B, BPFP=1.5768 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,856B, BPFP=3.0005 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,692B, BPFP=2.4137 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,040B, BPFP=2.8842 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,992B, BPFP=0.8690 +⌛️ [2/4] FRONTEND: Frontend time: 1.729s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.228s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14655781 30.30209200 + layer.0.v_cache 0.00001709 0.00238510 + layer.1.k_cache 0.36683986 3.08563783 + layer.1.v_cache 0.00000651 0.00096612 + layer.2.k_cache 0.01755252 0.54479280 + layer.2.v_cache 0.00002060 0.00282874 + layer.3.k_cache 0.02713263 4.27038074 + layer.3.v_cache 0.00001997 0.00301214 + layer.4.k_cache 0.00071102 0.08223266 + layer.4.v_cache 0.00005162 0.00583919 + layer.4.output 1.25479176 77.53366511 + ------------------------------------------------------------------------------------- + TOTAL 0.54955600 34.17857783 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 451176 +BPFP 1.6995 bits/point +EBPFP 3.3990 equivalent bits/point +MSE 34.178578 +---------------------- -------------------------------------------------------- +Time: 2.965s Load: 0.008s, Pack+Encode: 1.729s, Decode+Unpack: 1.228s +---------------------- -------------------------------------------------------- +💾 Converting with 34.1786 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 155, 128) +Output shape: (1, 155, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.output: torch.Size([1, 155, 3584]) -> torch.Size([1, 1, 155, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,708B, BPFP=0.7770 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,208B, BPFP=3.5492 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,712B, BPFP=1.4831 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,200B, BPFP=3.4476 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,720B, BPFP=1.8871 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,760B, BPFP=3.4032 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,080B, BPFP=1.7218 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,444B, BPFP=3.4722 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,544B, BPFP=2.7766 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,492B, BPFP=3.3762 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,084B, BPFP=0.8941 +⌛️ [2/4] FRONTEND: Frontend time: 1.610s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.113s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102877 26.95098286 + layer.0.v_cache 0.00001822 0.00259908 + layer.1.k_cache 0.12773433 1.95920646 + layer.1.v_cache 0.00000585 0.00103829 + layer.2.k_cache 0.00931232 0.55694216 + layer.2.v_cache 0.00002110 0.00312292 + layer.3.k_cache 0.01537714 3.32806830 + layer.3.v_cache 0.00002091 0.00351925 + layer.4.k_cache 0.00066471 0.08876653 + layer.4.v_cache 0.00005180 0.00683850 + layer.4.output 0.01742802 309.71854839 + ------------------------------------------------------------------------------------- + TOTAL 0.02389596 129.46652489 + (elements=1,349,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1349120 +Total Bytes 318952 +BPFP 1.8913 bits/point +EBPFP 3.7826 equivalent bits/point +MSE 129.466525 +---------------------- -------------------------------------------------------- +Time: 2.728s Load: 0.006s, Pack+Encode: 1.610s, Decode+Unpack: 1.113s +---------------------- -------------------------------------------------------- +💾 Converting with 129.4665 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,208B, BPFP=0.7419 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,604B, BPFP=3.4596 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,320B, BPFP=1.4767 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,952B, BPFP=3.3395 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,084B, BPFP=1.8230 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,396B, BPFP=3.2991 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,716B, BPFP=1.6509 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,008B, BPFP=3.4163 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,808B, BPFP=2.6023 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,928B, BPFP=3.2651 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,124B, BPFP=0.8734 +⌛️ [2/4] FRONTEND: Frontend time: 1.737s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.239s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11659910 30.79673192 + layer.0.v_cache 0.00001673 0.00248168 + layer.1.k_cache 0.29571861 2.58021723 + layer.1.v_cache 0.00000615 0.00097481 + layer.2.k_cache 0.04707494 0.72762600 + layer.2.v_cache 0.00002002 0.00280162 + layer.3.k_cache 0.03274950 3.53275061 + layer.3.v_cache 0.00002078 0.00320025 + layer.4.k_cache 0.00068608 0.08383723 + layer.4.v_cache 0.00005136 0.00619636 + layer.4.output 1.42391751 223.95980066 + ------------------------------------------------------------------------------------- + TOTAL 0.61531564 94.43855426 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 429148 +BPFP 1.8346 bits/point +EBPFP 3.6692 equivalent bits/point +MSE 94.438554 +---------------------- -------------------------------------------------------- +Time: 2.984s Load: 0.008s, Pack+Encode: 1.737s, Decode+Unpack: 1.239s +---------------------- -------------------------------------------------------- +💾 Converting with 94.4386 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,176B, BPFP=0.7041 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,096B, BPFP=3.0932 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,688B, BPFP=1.5554 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,664B, BPFP=3.0030 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,356B, BPFP=1.7235 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,656B, BPFP=2.9395 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,100B, BPFP=1.5814 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,720B, BPFP=3.0066 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,464B, BPFP=2.4234 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,656B, BPFP=2.8765 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,420B, BPFP=0.9128 +⌛️ [2/4] FRONTEND: Frontend time: 1.742s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.238s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10631144 30.97640302 + layer.0.v_cache 0.00001780 0.00244132 + layer.1.k_cache 0.46192560 3.14365485 + layer.1.v_cache 0.00000619 0.00099978 + layer.2.k_cache 0.02890189 0.53211434 + layer.2.v_cache 0.00002115 0.00290249 + layer.3.k_cache 0.04994748 4.02962395 + layer.3.v_cache 0.00002140 0.00329601 + layer.4.k_cache 0.00068365 0.08396948 + layer.4.v_cache 0.00005040 0.00597660 + layer.4.output 1.23456514 231.78702837 + ------------------------------------------------------------------------------------- + TOTAL 0.54646135 97.72297532 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 464996 +BPFP 1.7233 bits/point +EBPFP 3.4467 equivalent bits/point +MSE 97.722975 +---------------------- -------------------------------------------------------- +Time: 2.989s Load: 0.009s, Pack+Encode: 1.742s, Decode+Unpack: 1.238s +---------------------- -------------------------------------------------------- +💾 Converting with 97.7230 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 194, 128) +Output shape: (1, 194, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.output: torch.Size([1, 194, 3584]) -> torch.Size([1, 1, 194, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,148B, BPFP=0.7368 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,464B, BPFP=3.6617 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,220B, BPFP=1.4675 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,532B, BPFP=3.5867 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,648B, BPFP=1.8241 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 43,388B, BPFP=3.4945 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,508B, BPFP=1.6517 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 44,820B, BPFP=3.6099 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,008B, BPFP=2.7390 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,164B, BPFP=3.4765 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,928B, BPFP=0.8621 +⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.230s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15027005 29.63442584 + layer.0.v_cache 0.00001782 0.00267612 + layer.1.k_cache 0.21402168 2.58366536 + layer.1.v_cache 0.00000588 0.00105370 + layer.2.k_cache 0.02177384 0.77686892 + layer.2.v_cache 0.00002019 0.00318512 + layer.3.k_cache 0.04024272 3.43000872 + layer.3.v_cache 0.00002076 0.00362586 + layer.4.k_cache 0.00069591 0.08515696 + layer.4.v_cache 0.00005073 0.00663876 + layer.4.output 1.57794378 213.21769146 + ------------------------------------------------------------------------------------- + TOTAL 0.67486624 89.94418503 + (elements=1,688,576) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1688576 +Total Bytes 400828 +BPFP 1.8990 bits/point +EBPFP 3.7980 equivalent bits/point +MSE 89.944185 +---------------------- -------------------------------------------------------- +Time: 2.974s Load: 0.009s, Pack+Encode: 1.735s, Decode+Unpack: 1.230s +---------------------- -------------------------------------------------------- +💾 Converting with 89.9442 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 258, 128) +Output shape: (1, 258, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.output: torch.Size([1, 258, 3584]) -> torch.Size([1, 1, 258, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,744B, BPFP=0.7112 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,176B, BPFP=3.4627 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,728B, BPFP=1.5581 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,652B, BPFP=3.3704 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,192B, BPFP=1.7074 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,576B, BPFP=3.3052 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,624B, BPFP=1.5518 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,884B, BPFP=3.4450 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,712B, BPFP=2.5867 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,920B, BPFP=3.2655 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,272B, BPFP=0.7983 +⌛️ [2/4] FRONTEND: Frontend time: 1.869s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09035230 29.90787571 + layer.0.v_cache 0.00001726 0.00242212 + layer.1.k_cache 0.42455262 2.95861272 + layer.1.v_cache 0.00000600 0.00093751 + layer.2.k_cache 0.01891134 0.67575032 + layer.2.v_cache 0.00001996 0.00286384 + layer.3.k_cache 0.02093011 3.38804875 + layer.3.v_cache 0.00001928 0.00310807 + layer.4.k_cache 0.00072082 0.07982414 + layer.4.v_cache 0.00004951 0.00616776 + layer.4.output 0.00505962 194.04940130 + ------------------------------------------------------------------------------------- + TOTAL 0.03476450 82.08067177 + (elements=2,245,632) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2245632 +Total Bytes 504480 +BPFP 1.7972 bits/point +EBPFP 3.5944 equivalent bits/point +MSE 82.080672 +---------------------- -------------------------------------------------------- +Time: 3.260s Load: 0.009s, Pack+Encode: 1.869s, Decode+Unpack: 1.382s +---------------------- -------------------------------------------------------- +💾 Converting with 82.0807 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,660B, BPFP=0.7242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,628B, BPFP=3.3035 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,664B, BPFP=1.5397 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,260B, BPFP=3.2106 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,548B, BPFP=1.7356 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,220B, BPFP=3.1399 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,544B, BPFP=1.5995 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,532B, BPFP=3.2291 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,788B, BPFP=2.4992 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,436B, BPFP=3.0867 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,308B, BPFP=0.9347 +⌛️ [2/4] FRONTEND: Frontend time: 1.742s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.239s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15008528 30.31172512 + layer.0.v_cache 0.00001847 0.00237944 + layer.1.k_cache 0.25612781 2.64799539 + layer.1.v_cache 0.00000678 0.00093284 + layer.2.k_cache 0.03931687 0.62753939 + layer.2.v_cache 0.00002107 0.00286137 + layer.3.k_cache 0.03786138 3.83397482 + layer.3.v_cache 0.00002177 0.00317815 + layer.4.k_cache 0.00069674 0.08430381 + layer.4.v_cache 0.00005115 0.00591573 + layer.4.output 1.33117570 211.06184006 + ------------------------------------------------------------------------------------- + TOTAL 0.57661395 89.11492274 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 450588 +BPFP 1.8006 bits/point +EBPFP 3.6012 equivalent bits/point +MSE 89.114923 +---------------------- -------------------------------------------------------- +Time: 2.989s Load: 0.008s, Pack+Encode: 1.742s, Decode+Unpack: 1.239s +---------------------- -------------------------------------------------------- +💾 Converting with 89.1149 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,608B, BPFP=0.7472 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,024B, BPFP=3.2139 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,304B, BPFP=1.5021 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,704B, BPFP=3.0993 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,408B, BPFP=1.7715 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,988B, BPFP=3.0372 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,036B, BPFP=1.6524 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,760B, BPFP=3.1042 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,776B, BPFP=2.4979 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,264B, BPFP=2.9743 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,684B, BPFP=0.9385 +⌛️ [2/4] FRONTEND: Frontend time: 1.631s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.123s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15161699 29.24027778 + layer.0.v_cache 0.00001786 0.00265189 + layer.1.k_cache 0.08662784 2.69975891 + layer.1.v_cache 0.00000652 0.00099948 + layer.2.k_cache 0.02063684 0.51742736 + layer.2.v_cache 0.00002327 0.00306924 + layer.3.k_cache 0.08087930 4.14015774 + layer.3.v_cache 0.00002116 0.00337653 + layer.4.k_cache 0.00074161 0.08793278 + layer.4.v_cache 0.00006096 0.00635662 + layer.4.output 0.00898438 255.26817956 + ------------------------------------------------------------------------------------- + TOTAL 0.02373665 107.26936855 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 347556 +BPFP 1.7747 bits/point +EBPFP 3.5494 equivalent bits/point +MSE 107.269369 +---------------------- -------------------------------------------------------- +Time: 2.760s Load: 0.006s, Pack+Encode: 1.631s, Decode+Unpack: 1.123s +---------------------- -------------------------------------------------------- +💾 Converting with 107.2694 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,360B, BPFP=0.7358 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,752B, BPFP=3.3915 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,144B, BPFP=1.5727 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,768B, BPFP=3.3216 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,764B, BPFP=1.8298 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,648B, BPFP=3.2420 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,244B, BPFP=1.6509 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,628B, BPFP=3.3116 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,748B, BPFP=2.6099 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,276B, BPFP=3.2156 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,520B, BPFP=0.8880 +⌛️ [2/4] FRONTEND: Frontend time: 1.756s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.234s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14327769 28.53140092 + layer.0.v_cache 0.00001763 0.00252952 + layer.1.k_cache 0.34925704 2.77360119 + layer.1.v_cache 0.00000642 0.00100260 + layer.2.k_cache 0.02352169 0.58870787 + layer.2.v_cache 0.00002271 0.00297256 + layer.3.k_cache 0.00845094 3.38183094 + layer.3.v_cache 0.00002026 0.00325972 + layer.4.k_cache 0.00069169 0.08068054 + layer.4.v_cache 0.00005086 0.00629091 + layer.4.output 1.39158238 224.73100649 + ------------------------------------------------------------------------------------- + TOTAL 0.60390550 94.61701895 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 437852 +BPFP 1.8293 bits/point +EBPFP 3.6585 equivalent bits/point +MSE 94.617019 +---------------------- -------------------------------------------------------- +Time: 2.998s Load: 0.007s, Pack+Encode: 1.756s, Decode+Unpack: 1.234s +---------------------- -------------------------------------------------------- +💾 Converting with 94.6170 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,628B, BPFP=0.7175 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,112B, BPFP=3.3586 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,608B, BPFP=1.4550 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,800B, BPFP=3.2273 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,512B, BPFP=1.7905 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,856B, BPFP=3.2305 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,144B, BPFP=1.5991 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,864B, BPFP=3.2877 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,028B, BPFP=2.5584 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,060B, BPFP=3.1852 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,416B, BPFP=0.8232 +⌛️ [2/4] FRONTEND: Frontend time: 1.868s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12716211 30.04894531 + layer.0.v_cache 0.00001677 0.00237262 + layer.1.k_cache 0.51313377 2.68253462 + layer.1.v_cache 0.00000586 0.00093704 + layer.2.k_cache 0.02995576 0.70033736 + layer.2.v_cache 0.00002133 0.00289207 + layer.3.k_cache 0.03339291 3.74155407 + layer.3.v_cache 0.00002012 0.00304164 + layer.4.k_cache 0.00068476 0.07975981 + layer.4.v_cache 0.00005875 0.00606403 + layer.4.output 0.00481428 68.92456169 + ------------------------------------------------------------------------------------- + TOTAL 0.04342071 30.57296296 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 531028 +BPFP 1.7748 bits/point +EBPFP 3.5497 equivalent bits/point +MSE 30.572963 +---------------------- -------------------------------------------------------- +Time: 3.252s Load: 0.009s, Pack+Encode: 1.868s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 30.5730 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,216B, BPFP=0.6900 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,420B, BPFP=3.0401 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,972B, BPFP=1.4131 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,500B, BPFP=2.9220 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,580B, BPFP=1.6351 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,440B, BPFP=2.8568 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,808B, BPFP=1.5261 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,416B, BPFP=2.9168 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,712B, BPFP=2.3814 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,448B, BPFP=2.7958 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,304B, BPFP=0.8727 +⌛️ [2/4] FRONTEND: Frontend time: 1.759s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.245s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13396681 37.30268824 + layer.0.v_cache 0.00001641 0.00231336 + layer.1.k_cache 0.47111181 3.06344028 + layer.1.v_cache 0.00000608 0.00092860 + layer.2.k_cache 0.02933135 0.57500266 + layer.2.v_cache 0.00002255 0.00288401 + layer.3.k_cache 0.03744583 4.28939171 + layer.3.v_cache 0.00002260 0.00304990 + layer.4.k_cache 0.00071777 0.08216249 + layer.4.v_cache 0.00004917 0.00565783 + layer.4.output 1.20538323 93.87541303 + ------------------------------------------------------------------------------------- + TOTAL 0.53590429 41.32090649 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 459816 +BPFP 1.6639 bits/point +EBPFP 3.3278 equivalent bits/point +MSE 41.320906 +---------------------- -------------------------------------------------------- +Time: 3.014s Load: 0.009s, Pack+Encode: 1.759s, Decode+Unpack: 1.245s +---------------------- -------------------------------------------------------- +💾 Converting with 41.3209 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,728B, BPFP=0.6905 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,492B, BPFP=3.2819 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,724B, BPFP=1.5041 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,952B, BPFP=3.1441 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,628B, BPFP=1.7159 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,432B, BPFP=3.1159 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,100B, BPFP=1.5245 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,592B, BPFP=3.1788 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,880B, BPFP=2.4891 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,340B, BPFP=3.0566 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 107,536B, BPFP=0.8335 +⌛️ [2/4] FRONTEND: Frontend time: 1.903s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14060297 29.15551758 + layer.0.v_cache 0.00001598 0.00227410 + layer.1.k_cache 0.58610762 2.98274846 + layer.1.v_cache 0.00000669 0.00091050 + layer.2.k_cache 0.02834670 0.66318872 + layer.2.v_cache 0.00002256 0.00280053 + layer.3.k_cache 0.01962393 3.83293660 + layer.3.v_cache 0.00002003 0.00301367 + layer.4.k_cache 0.00067797 0.08064526 + layer.4.v_cache 0.00005035 0.00585140 + layer.4.output 0.00463834 174.23513455 + ------------------------------------------------------------------------------------- + TOTAL 0.04752607 73.90446051 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 544404 +BPFP 1.7374 bits/point +EBPFP 3.4748 equivalent bits/point +MSE 73.904461 +---------------------- -------------------------------------------------------- +Time: 3.287s Load: 0.010s, Pack+Encode: 1.903s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 73.9045 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,600B, BPFP=0.6990 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,656B, BPFP=3.1176 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,948B, BPFP=1.5393 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 58,944B, BPFP=3.0296 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,340B, BPFP=1.7136 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,900B, BPFP=2.9759 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,488B, BPFP=1.5670 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,996B, BPFP=3.0323 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,988B, BPFP=2.4151 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,348B, BPFP=2.8962 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 113,848B, BPFP=0.8359 +⌛️ [2/4] FRONTEND: Frontend time: 1.897s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12404394 30.68872070 + layer.0.v_cache 0.00001645 0.00223920 + layer.1.k_cache 0.57408182 3.09555676 + layer.1.v_cache 0.00000616 0.00089344 + layer.2.k_cache 0.03522489 0.62247939 + layer.2.v_cache 0.00002326 0.00280827 + layer.3.k_cache 0.02908128 4.18369012 + layer.3.v_cache 0.00002078 0.00301248 + layer.4.k_cache 0.00072493 0.08258160 + layer.4.v_cache 0.00004967 0.00571407 + layer.4.output 0.04390477 154.95556273 + ------------------------------------------------------------------------------------- + TOTAL 0.06297686 66.08097854 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 561056 +BPFP 1.6963 bits/point +EBPFP 3.3926 equivalent bits/point +MSE 66.080979 +---------------------- -------------------------------------------------------- +Time: 3.288s Load: 0.010s, Pack+Encode: 1.897s, Decode+Unpack: 1.381s +---------------------- -------------------------------------------------------- +💾 Converting with 66.0810 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 242, 128) +Output shape: (1, 242, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.output: torch.Size([1, 242, 3584]) -> torch.Size([1, 1, 242, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,040B, BPFP=0.7128 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,752B, BPFP=3.1477 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,920B, BPFP=1.6090 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,224B, BPFP=3.0491 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,668B, BPFP=1.7218 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,356B, BPFP=2.9930 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,092B, BPFP=1.6201 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,252B, BPFP=3.0509 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,896B, BPFP=2.4468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,336B, BPFP=2.9272 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,676B, BPFP=0.8733 +⌛️ [2/4] FRONTEND: Frontend time: 1.767s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.258s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11466674 30.99226215 + layer.0.v_cache 0.00001794 0.00237761 + layer.1.k_cache 0.39062487 3.15966343 + layer.1.v_cache 0.00000616 0.00094594 + layer.2.k_cache 0.02414889 0.59492865 + layer.2.v_cache 0.00002046 0.00285890 + layer.3.k_cache 0.03495786 4.19940059 + layer.3.v_cache 0.00002053 0.00317469 + layer.4.k_cache 0.00069125 0.08160265 + layer.4.v_cache 0.00005034 0.00605969 + layer.4.output 1.26514320 91.36574860 + ------------------------------------------------------------------------------------- + TOTAL 0.55418868 39.91785379 + (elements=2,106,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2106368 +Total Bytes 455212 +BPFP 1.7289 bits/point +EBPFP 3.4578 equivalent bits/point +MSE 39.917854 +---------------------- -------------------------------------------------------- +Time: 3.033s Load: 0.008s, Pack+Encode: 1.767s, Decode+Unpack: 1.258s +---------------------- -------------------------------------------------------- +💾 Converting with 39.9179 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 189, 128) +Output shape: (1, 189, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.output: torch.Size([1, 189, 3584]) -> torch.Size([1, 1, 189, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,656B, BPFP=0.7156 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,888B, BPFP=3.0496 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,576B, BPFP=1.4530 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,820B, BPFP=2.9613 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,580B, BPFP=1.7014 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,228B, BPFP=2.9124 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,504B, BPFP=1.6124 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,956B, BPFP=2.9726 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,112B, BPFP=2.4067 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,412B, BPFP=2.8449 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,916B, BPFP=0.9320 +⌛️ [2/4] FRONTEND: Frontend time: 1.663s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.134s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14361043 30.44042969 + layer.0.v_cache 0.00001648 0.00244180 + layer.1.k_cache 0.18932698 2.87811182 + layer.1.v_cache 0.00000663 0.00102907 + layer.2.k_cache 0.00768016 0.54192712 + layer.2.v_cache 0.00002453 0.00312730 + layer.3.k_cache 0.01421528 4.48132260 + layer.3.v_cache 0.00002024 0.00324662 + layer.4.k_cache 0.00072804 0.08744955 + layer.4.v_cache 0.00005086 0.00613925 + layer.4.output 0.00858909 248.05713813 + ------------------------------------------------------------------------------------- + TOTAL 0.02445902 104.40265834 + (elements=1,645,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1645056 +Total Bytes 352648 +BPFP 1.7149 bits/point +EBPFP 3.4299 equivalent bits/point +MSE 104.402658 +---------------------- -------------------------------------------------------- +Time: 2.804s Load: 0.007s, Pack+Encode: 1.663s, Decode+Unpack: 1.134s +---------------------- -------------------------------------------------------- +💾 Converting with 104.4027 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,092B, BPFP=0.7403 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,740B, BPFP=3.5021 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,568B, BPFP=1.4354 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,324B, BPFP=3.3982 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,616B, BPFP=1.8058 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,432B, BPFP=3.3327 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,520B, BPFP=1.6520 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,784B, BPFP=3.4319 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,864B, BPFP=2.6309 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,036B, BPFP=3.3037 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,708B, BPFP=0.8563 +⌛️ [2/4] FRONTEND: Frontend time: 1.778s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.252s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16894778 28.69503970 + layer.0.v_cache 0.00001779 0.00245429 + layer.1.k_cache 0.37628758 2.28553793 + layer.1.v_cache 0.00000652 0.00097978 + layer.2.k_cache 0.02181128 0.68090835 + layer.2.v_cache 0.00002030 0.00292877 + layer.3.k_cache 0.01815432 3.39489546 + layer.3.v_cache 0.00002063 0.00322140 + layer.4.k_cache 0.00071780 0.07987937 + layer.4.v_cache 0.00005059 0.00613530 + layer.4.output 1.43727796 224.96457914 + ------------------------------------------------------------------------------------- + TOTAL 0.62629296 94.70023731 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 425684 +BPFP 1.8369 bits/point +EBPFP 3.6737 equivalent bits/point +MSE 94.700237 +---------------------- -------------------------------------------------------- +Time: 3.039s Load: 0.009s, Pack+Encode: 1.778s, Decode+Unpack: 1.252s +---------------------- -------------------------------------------------------- +💾 Converting with 94.7002 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,160B, BPFP=0.7350 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,632B, BPFP=3.4456 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,048B, BPFP=1.4502 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,356B, BPFP=3.3533 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,808B, BPFP=1.7946 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,136B, BPFP=3.2650 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,636B, BPFP=1.6374 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,520B, BPFP=3.3652 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,500B, BPFP=2.5680 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,632B, BPFP=3.2286 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,596B, BPFP=0.8122 +⌛️ [2/4] FRONTEND: Frontend time: 1.780s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.251s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09877178 30.32780852 + layer.0.v_cache 0.00001652 0.00237958 + layer.1.k_cache 0.28824160 2.38000912 + layer.1.v_cache 0.00000619 0.00097042 + layer.2.k_cache 0.02137197 0.64313726 + layer.2.v_cache 0.00001995 0.00284317 + layer.3.k_cache 0.03811082 3.66551293 + layer.3.v_cache 0.00002058 0.00311029 + layer.4.k_cache 0.00070031 0.08173145 + layer.4.v_cache 0.00004695 0.00581204 + layer.4.output 1.41729720 222.83395337 + ------------------------------------------------------------------------------------- + TOTAL 0.60990512 93.93829344 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 422024 +BPFP 1.7958 bits/point +EBPFP 3.5916 equivalent bits/point +MSE 93.938293 +---------------------- -------------------------------------------------------- +Time: 3.038s Load: 0.008s, Pack+Encode: 1.780s, Decode+Unpack: 1.251s +---------------------- -------------------------------------------------------- +💾 Converting with 93.9383 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,372B, BPFP=0.7171 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,692B, BPFP=3.3664 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,156B, BPFP=1.4627 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,116B, BPFP=3.2575 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,420B, BPFP=1.7575 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,908B, BPFP=3.1739 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,016B, BPFP=1.5913 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,740B, BPFP=3.2315 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,940B, BPFP=2.5539 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,292B, BPFP=3.1314 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,384B, BPFP=0.8828 +⌛️ [2/4] FRONTEND: Frontend time: 1.777s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.245s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12161526 29.43585151 + layer.0.v_cache 0.00001916 0.00241372 + layer.1.k_cache 0.26028682 2.82413071 + layer.1.v_cache 0.00000628 0.00093775 + layer.2.k_cache 0.01296707 0.56462631 + layer.2.v_cache 0.00002040 0.00279317 + layer.3.k_cache 0.04595671 4.05228984 + layer.3.v_cache 0.00002073 0.00308895 + layer.4.k_cache 0.00067352 0.08194113 + layer.4.v_cache 0.00005047 0.00601553 + layer.4.output 1.35465500 216.13590392 + ------------------------------------------------------------------------------------- + TOTAL 0.58377655 91.17208330 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 440036 +BPFP 1.7896 bits/point +EBPFP 3.5792 equivalent bits/point +MSE 91.172083 +---------------------- -------------------------------------------------------- +Time: 3.031s Load: 0.008s, Pack+Encode: 1.777s, Decode+Unpack: 1.245s +---------------------- -------------------------------------------------------- +💾 Converting with 91.1721 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,444B, BPFP=0.7221 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,264B, BPFP=3.3368 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,208B, BPFP=1.4663 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,720B, BPFP=3.2301 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,364B, BPFP=1.7536 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,988B, BPFP=3.1795 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,628B, BPFP=1.5644 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,240B, BPFP=3.2660 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,332B, BPFP=2.5119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,352B, BPFP=3.1355 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,572B, BPFP=0.8353 +⌛️ [2/4] FRONTEND: Frontend time: 1.769s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.246s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10229862 28.57418634 + layer.0.v_cache 0.00001727 0.00235328 + layer.1.k_cache 0.34669285 2.86393130 + layer.1.v_cache 0.00000643 0.00095442 + layer.2.k_cache 0.01639456 0.66987110 + layer.2.v_cache 0.00002164 0.00285008 + layer.3.k_cache 0.03343437 3.87339337 + layer.3.v_cache 0.00002043 0.00308129 + layer.4.k_cache 0.00071218 0.07920844 + layer.4.v_cache 0.00005088 0.00591380 + layer.4.output 1.35462105 207.55459861 + ------------------------------------------------------------------------------------- + TOTAL 0.58717627 87.58576081 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 434112 +BPFP 1.7655 bits/point +EBPFP 3.5310 equivalent bits/point +MSE 87.585761 +---------------------- -------------------------------------------------------- +Time: 3.023s Load: 0.008s, Pack+Encode: 1.769s, Decode+Unpack: 1.246s +---------------------- -------------------------------------------------------- +💾 Converting with 87.5858 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,592B, BPFP=0.7291 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,532B, BPFP=3.3406 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,276B, BPFP=1.4645 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,132B, BPFP=3.2442 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,268B, BPFP=1.7393 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,272B, BPFP=3.1850 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,040B, BPFP=1.5859 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,232B, BPFP=3.2511 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,664B, BPFP=2.5237 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,212B, BPFP=3.1121 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,604B, BPFP=0.8319 +⌛️ [2/4] FRONTEND: Frontend time: 1.765s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.237s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11800324 31.04086936 + layer.0.v_cache 0.00001819 0.00238182 + layer.1.k_cache 0.35011090 2.86633516 + layer.1.v_cache 0.00000612 0.00089554 + layer.2.k_cache 0.02017111 0.60426230 + layer.2.v_cache 0.00002006 0.00271218 + layer.3.k_cache 0.03026835 3.84527507 + layer.3.v_cache 0.00001975 0.00303524 + layer.4.k_cache 0.00074701 0.08112109 + layer.4.v_cache 0.00005118 0.00576763 + layer.4.output 1.34863711 214.57439821 + ------------------------------------------------------------------------------------- + TOTAL 0.58587504 90.61608487 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 435824 +BPFP 1.7646 bits/point +EBPFP 3.5293 equivalent bits/point +MSE 90.616085 +---------------------- -------------------------------------------------------- +Time: 3.010s Load: 0.008s, Pack+Encode: 1.765s, Decode+Unpack: 1.237s +---------------------- -------------------------------------------------------- +💾 Converting with 90.6161 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,592B, BPFP=0.7227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,400B, BPFP=3.3024 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,556B, BPFP=1.6073 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,472B, BPFP=3.2391 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,636B, BPFP=1.8174 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,876B, BPFP=3.1984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,304B, BPFP=1.6583 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,004B, BPFP=3.2754 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,728B, BPFP=2.5742 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,028B, BPFP=3.1406 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,996B, BPFP=0.8870 +⌛️ [2/4] FRONTEND: Frontend time: 1.771s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.242s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10957497 29.65393226 + layer.0.v_cache 0.00001697 0.00239972 + layer.1.k_cache 0.31262030 3.05860388 + layer.1.v_cache 0.00000629 0.00098584 + layer.2.k_cache 0.01311818 0.50489264 + layer.2.v_cache 0.00002149 0.00292570 + layer.3.k_cache 0.06227440 3.91862041 + layer.3.v_cache 0.00002139 0.00333137 + layer.4.k_cache 0.00068925 0.08374452 + layer.4.v_cache 0.00005302 0.00630624 + layer.4.output 1.33688878 213.07758890 + ------------------------------------------------------------------------------------- + TOTAL 0.57980104 89.92816852 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 450592 +BPFP 1.8085 bits/point +EBPFP 3.6170 equivalent bits/point +MSE 89.928169 +---------------------- -------------------------------------------------------- +Time: 3.023s Load: 0.010s, Pack+Encode: 1.771s, Decode+Unpack: 1.242s +---------------------- -------------------------------------------------------- +💾 Converting with 89.9282 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,068B, BPFP=0.7386 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,356B, BPFP=3.4739 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,636B, BPFP=1.4404 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,280B, BPFP=3.3216 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,708B, BPFP=1.8125 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,640B, BPFP=3.2746 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,396B, BPFP=1.6429 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,808B, BPFP=3.3603 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,560B, BPFP=2.6086 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,400B, BPFP=3.2570 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,184B, BPFP=0.8508 +⌛️ [2/4] FRONTEND: Frontend time: 1.766s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.242s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13987277 29.90566773 + layer.0.v_cache 0.00001887 0.00245696 + layer.1.k_cache 0.23498858 2.73131636 + layer.1.v_cache 0.00000603 0.00092622 + layer.2.k_cache 0.02028183 0.71003021 + layer.2.v_cache 0.00001992 0.00279153 + layer.3.k_cache 0.05552547 3.83438841 + layer.3.v_cache 0.00002044 0.00319910 + layer.4.k_cache 0.00067773 0.08202819 + layer.4.v_cache 0.00004777 0.00588480 + layer.4.output 1.43727485 232.61894282 + ------------------------------------------------------------------------------------- + TOTAL 0.61837549 97.97713466 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 421036 +BPFP 1.8168 bits/point +EBPFP 3.6336 equivalent bits/point +MSE 97.977135 +---------------------- -------------------------------------------------------- +Time: 3.016s Load: 0.007s, Pack+Encode: 1.766s, Decode+Unpack: 1.242s +---------------------- -------------------------------------------------------- +💾 Converting with 97.9771 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,976B, BPFP=0.7387 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,820B, BPFP=3.4671 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,116B, BPFP=1.5637 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,436B, BPFP=3.3646 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,056B, BPFP=1.7814 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,524B, BPFP=3.2971 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,532B, BPFP=1.6685 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,576B, BPFP=3.3750 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,572B, BPFP=2.6342 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,644B, BPFP=3.2319 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,816B, BPFP=0.8338 +⌛️ [2/4] FRONTEND: Frontend time: 1.766s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.241s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17004660 29.04243187 + layer.0.v_cache 0.00001740 0.00245497 + layer.1.k_cache 0.16740718 2.69629572 + layer.1.v_cache 0.00000654 0.00100217 + layer.2.k_cache 0.01191964 0.68827827 + layer.2.v_cache 0.00001991 0.00296845 + layer.3.k_cache 0.00660476 3.79465286 + layer.3.v_cache 0.00002076 0.00322939 + layer.4.k_cache 0.00070551 0.08309638 + layer.4.v_cache 0.00005179 0.00635606 + layer.4.output 1.45088344 220.86930856 + ------------------------------------------------------------------------------------- + TOTAL 0.61841083 93.08270153 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 418068 +BPFP 1.8211 bits/point +EBPFP 3.6422 equivalent bits/point +MSE 93.082702 +---------------------- -------------------------------------------------------- +Time: 3.013s Load: 0.007s, Pack+Encode: 1.766s, Decode+Unpack: 1.241s +---------------------- -------------------------------------------------------- +💾 Converting with 93.0827 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,644B, BPFP=0.7231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,192B, BPFP=3.2739 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,044B, BPFP=1.4976 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,976B, BPFP=3.1913 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,600B, BPFP=1.7391 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,484B, BPFP=3.1579 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,908B, BPFP=1.6242 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,948B, BPFP=3.1894 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,168B, BPFP=2.5250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,352B, BPFP=3.0810 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 82,888B, BPFP=0.8044 +⌛️ [2/4] FRONTEND: Frontend time: 1.775s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.250s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14712874 31.15711405 + layer.0.v_cache 0.00001826 0.00228748 + layer.1.k_cache 0.36788522 2.71456883 + layer.1.v_cache 0.00000598 0.00094964 + layer.2.k_cache 0.01156470 0.54636300 + layer.2.v_cache 0.00002030 0.00283049 + layer.3.k_cache 0.01315050 4.27833358 + layer.3.v_cache 0.00002017 0.00296004 + layer.4.k_cache 0.00068370 0.08103626 + layer.4.v_cache 0.00004817 0.00591067 + layer.4.output 1.33105469 209.00359084 + ------------------------------------------------------------------------------------- + TOTAL 0.57987697 88.34220529 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 436204 +BPFP 1.7431 bits/point +EBPFP 3.4863 equivalent bits/point +MSE 88.342205 +---------------------- -------------------------------------------------------- +Time: 3.032s Load: 0.008s, Pack+Encode: 1.775s, Decode+Unpack: 1.250s +---------------------- -------------------------------------------------------- +💾 Converting with 88.3422 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 188, 128) +Output shape: (1, 188, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.output: torch.Size([1, 188, 3584]) -> torch.Size([1, 1, 188, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,648B, BPFP=0.7188 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,564B, BPFP=3.0389 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,160B, BPFP=1.5924 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,492B, BPFP=2.9498 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,408B, BPFP=1.6961 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,884B, BPFP=2.8993 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,088B, BPFP=1.5864 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,460B, BPFP=2.9471 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,100B, BPFP=2.4186 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,176B, BPFP=2.8404 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,956B, BPFP=0.8543 +⌛️ [2/4] FRONTEND: Frontend time: 1.643s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.125s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005002 30.07694326 + layer.0.v_cache 0.00001663 0.00240582 + layer.1.k_cache 0.14149688 2.93953818 + layer.1.v_cache 0.00000584 0.00094163 + layer.2.k_cache 0.00881058 0.51973383 + layer.2.v_cache 0.00002285 0.00302136 + layer.3.k_cache 0.03192479 4.19830030 + layer.3.v_cache 0.00001958 0.00320928 + layer.4.k_cache 0.00066612 0.08248560 + layer.4.v_cache 0.00005162 0.00620561 + layer.4.output 0.00857985 258.85175247 + ------------------------------------------------------------------------------------- + TOTAL 0.02253670 108.81147366 + (elements=1,636,352) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1636352 +Total Bytes 344936 +BPFP 1.6864 bits/point +EBPFP 3.3727 equivalent bits/point +MSE 108.811474 +---------------------- -------------------------------------------------------- +Time: 2.776s Load: 0.008s, Pack+Encode: 1.643s, Decode+Unpack: 1.125s +---------------------- -------------------------------------------------------- +💾 Converting with 108.8115 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,564B, BPFP=0.7509 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 45,768B, BPFP=3.5936 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,012B, BPFP=1.4143 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 44,468B, BPFP=3.4915 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,348B, BPFP=1.8332 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,864B, BPFP=3.5226 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,832B, BPFP=1.6357 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,556B, BPFP=3.5769 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,076B, BPFP=2.6756 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,612B, BPFP=3.4243 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 77,768B, BPFP=0.8723 +⌛️ [2/4] FRONTEND: Frontend time: 1.764s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.241s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11435634 28.98531966 + layer.0.v_cache 0.00001764 0.00260037 + layer.1.k_cache 0.25025516 2.19978310 + layer.1.v_cache 0.00000690 0.00099556 + layer.2.k_cache 0.01528106 0.70533047 + layer.2.v_cache 0.00002105 0.00308819 + layer.3.k_cache 0.02533426 3.17481266 + layer.3.v_cache 0.00002127 0.00342539 + layer.4.k_cache 0.00070809 0.08289237 + layer.4.v_cache 0.00005072 0.00622845 + layer.4.output 1.53833902 235.53098080 + ------------------------------------------------------------------------------------- + TOTAL 0.65731915 99.05184364 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 407868 +BPFP 1.8838 bits/point +EBPFP 3.7676 equivalent bits/point +MSE 99.051844 +---------------------- -------------------------------------------------------- +Time: 3.011s Load: 0.007s, Pack+Encode: 1.764s, Decode+Unpack: 1.241s +---------------------- -------------------------------------------------------- +💾 Converting with 99.0518 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,516B, BPFP=0.7392 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,844B, BPFP=3.1983 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,560B, BPFP=1.5243 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,528B, BPFP=3.0840 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,384B, BPFP=1.7694 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,116B, BPFP=3.0483 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,100B, BPFP=1.6580 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,904B, BPFP=3.1167 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,980B, BPFP=2.5156 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,408B, BPFP=2.9868 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,552B, BPFP=0.9245 +⌛️ [2/4] FRONTEND: Frontend time: 1.656s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.132s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258550 29.50190972 + layer.0.v_cache 0.00001767 0.00259232 + layer.1.k_cache 0.14790291 3.03974542 + layer.1.v_cache 0.00000629 0.00101356 + layer.2.k_cache 0.00999225 0.52790184 + layer.2.v_cache 0.00002110 0.00294847 + layer.3.k_cache 0.02117847 4.53990546 + layer.3.v_cache 0.00002045 0.00347070 + layer.4.k_cache 0.00067024 0.08493145 + layer.4.v_cache 0.00006105 0.00644080 + layer.4.output 0.00897616 118.81216518 + ------------------------------------------------------------------------------------- + TOTAL 0.02207583 51.14094212 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 346892 +BPFP 1.7713 bits/point +EBPFP 3.5426 equivalent bits/point +MSE 51.140942 +---------------------- -------------------------------------------------------- +Time: 2.794s Load: 0.006s, Pack+Encode: 1.656s, Decode+Unpack: 1.132s +---------------------- -------------------------------------------------------- +💾 Converting with 51.1409 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,804B, BPFP=0.7509 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,776B, BPFP=3.5827 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,380B, BPFP=1.4078 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,316B, BPFP=3.4709 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,584B, BPFP=1.8064 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,572B, BPFP=3.4139 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,440B, BPFP=1.6422 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,188B, BPFP=3.5377 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,556B, BPFP=2.6468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,332B, BPFP=3.3955 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,396B, BPFP=0.8906 +⌛️ [2/4] FRONTEND: Frontend time: 1.768s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.234s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09364704 28.87075387 + layer.0.v_cache 0.00001853 0.00251152 + layer.1.k_cache 0.25163280 1.89426976 + layer.1.v_cache 0.00000592 0.00097589 + layer.2.k_cache 0.01595125 0.66978993 + layer.2.v_cache 0.00002050 0.00290306 + layer.3.k_cache 0.04325477 3.77423634 + layer.3.v_cache 0.00002016 0.00326003 + layer.4.k_cache 0.00069210 0.08139336 + layer.4.v_cache 0.00005026 0.00617817 + layer.4.output 1.50065788 164.00608368 + ------------------------------------------------------------------------------------- + TOTAL 0.64175873 69.60875634 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 416344 +BPFP 1.8758 bits/point +EBPFP 3.7517 equivalent bits/point +MSE 69.608756 +---------------------- -------------------------------------------------------- +Time: 3.009s Load: 0.007s, Pack+Encode: 1.768s, Decode+Unpack: 1.234s +---------------------- -------------------------------------------------------- +💾 Converting with 69.6088 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,408B, BPFP=0.7196 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,096B, BPFP=3.3252 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,440B, BPFP=1.4132 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,960B, BPFP=3.2467 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,936B, BPFP=1.7240 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,640B, BPFP=3.1554 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,988B, BPFP=1.5893 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,888B, BPFP=3.2417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 36,228B, BPFP=2.5047 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,732B, BPFP=3.0926 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,120B, BPFP=0.8802 +⌛️ [2/4] FRONTEND: Frontend time: 1.776s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.245s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510114 30.34364413 + layer.0.v_cache 0.00001717 0.00232492 + layer.1.k_cache 0.31797088 2.68023790 + layer.1.v_cache 0.00000673 0.00094931 + layer.2.k_cache 0.01711024 0.62212325 + layer.2.v_cache 0.00002022 0.00285189 + layer.3.k_cache 0.01307247 3.82913127 + layer.3.v_cache 0.00002089 0.00306747 + layer.4.k_cache 0.00069413 0.08352524 + layer.4.v_cache 0.00005461 0.00636952 + layer.4.output 1.35466395 221.49753082 + ------------------------------------------------------------------------------------- + TOTAL 0.58392448 93.41511415 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 436436 +BPFP 1.7749 bits/point +EBPFP 3.5499 equivalent bits/point +MSE 93.415114 +---------------------- -------------------------------------------------------- +Time: 3.028s Load: 0.008s, Pack+Encode: 1.776s, Decode+Unpack: 1.245s +---------------------- -------------------------------------------------------- +💾 Converting with 93.4151 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,152B, BPFP=0.7026 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,728B, BPFP=3.0701 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,496B, BPFP=1.5433 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,484B, BPFP=2.9917 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,724B, BPFP=1.6837 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,588B, BPFP=2.9352 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,840B, BPFP=1.5650 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,420B, BPFP=2.9877 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,968B, BPFP=2.3921 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,440B, BPFP=2.8629 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,100B, BPFP=0.8830 +⌛️ [2/4] FRONTEND: Frontend time: 1.777s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.253s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14860459 30.09885333 + layer.0.v_cache 0.00001629 0.00226155 + layer.1.k_cache 0.38266554 3.10821238 + layer.1.v_cache 0.00000669 0.00095317 + layer.2.k_cache 0.01628031 0.58016365 + layer.2.v_cache 0.00002106 0.00278654 + layer.3.k_cache 0.01842327 4.22247708 + layer.3.v_cache 0.00002289 0.00320632 + layer.4.k_cache 0.00072690 0.08115657 + layer.4.v_cache 0.00004923 0.00573922 + layer.4.output 1.23454798 149.70177491 + ------------------------------------------------------------------------------------- + TOTAL 0.54168545 63.88342554 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 458940 +BPFP 1.7009 bits/point +EBPFP 3.4018 equivalent bits/point +MSE 63.883426 +---------------------- -------------------------------------------------------- +Time: 3.039s Load: 0.008s, Pack+Encode: 1.777s, Decode+Unpack: 1.253s +---------------------- -------------------------------------------------------- +💾 Converting with 63.8834 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 435, 128) +Output shape: (1, 435, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.output: torch.Size([1, 435, 3584]) -> torch.Size([1, 1, 435, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 18,648B, BPFP=0.6698 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 85,092B, BPFP=3.0565 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 44,176B, BPFP=1.5868 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 82,488B, BPFP=2.9629 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 45,740B, BPFP=1.6430 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 80,812B, BPFP=2.9027 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 42,272B, BPFP=1.5184 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 82,244B, BPFP=2.9542 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 65,360B, BPFP=2.3477 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 79,520B, BPFP=2.8563 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 159,444B, BPFP=0.8182 +⌛️ [2/4] FRONTEND: Frontend time: 2.148s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.605s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15864419 33.07735722 + layer.0.v_cache 0.00001607 0.00224495 + layer.1.k_cache 0.94329666 3.22199791 + layer.1.v_cache 0.00000639 0.00091441 + layer.2.k_cache 0.02505282 0.57513992 + layer.2.v_cache 0.00002131 0.00273972 + layer.3.k_cache 0.01724711 4.31093806 + layer.3.v_cache 0.00002032 0.00290077 + layer.4.k_cache 0.00074513 0.08003899 + layer.4.v_cache 0.00005173 0.00578159 + layer.4.output 0.00608059 53.24358067 + ------------------------------------------------------------------------------------- + TOTAL 0.06986270 24.35206578 + (elements=3,786,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3786240 +Total Bytes 785796 +BPFP 1.6603 bits/point +EBPFP 3.3206 equivalent bits/point +MSE 24.352066 +---------------------- -------------------------------------------------------- +Time: 3.769s Load: 0.015s, Pack+Encode: 2.148s, Decode+Unpack: 1.605s +---------------------- -------------------------------------------------------- +💾 Converting with 24.3521 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 437, 128) +Output shape: (1, 437, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.output: torch.Size([1, 437, 3584]) -> torch.Size([1, 1, 437, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 18,664B, BPFP=0.6673 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 85,124B, BPFP=3.0436 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 43,692B, BPFP=1.5622 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 82,592B, BPFP=2.9531 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 45,504B, BPFP=1.6270 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 81,016B, BPFP=2.8967 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 42,020B, BPFP=1.5024 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 82,716B, BPFP=2.9575 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 65,884B, BPFP=2.3557 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 79,292B, BPFP=2.8351 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 181,052B, BPFP=0.9248 +⌛️ [2/4] FRONTEND: Frontend time: 2.149s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.612s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14454122 31.84215443 + layer.0.v_cache 0.00001702 0.00222612 + layer.1.k_cache 0.94267751 3.09493830 + layer.1.v_cache 0.00000668 0.00092518 + layer.2.k_cache 0.01808156 0.55366397 + layer.2.v_cache 0.00002154 0.00273074 + layer.3.k_cache 0.03034366 4.16751790 + layer.3.v_cache 0.00002210 0.00311933 + layer.4.k_cache 0.00074160 0.08223987 + layer.4.v_cache 0.00005552 0.00596970 + layer.4.output 0.00611241 110.56988599 + ------------------------------------------------------------------------------------- + TOTAL 0.06937031 47.86733456 + (elements=3,803,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3803648 +Total Bytes 807556 +BPFP 1.6985 bits/point +EBPFP 3.3970 equivalent bits/point +MSE 47.867335 +---------------------- -------------------------------------------------------- +Time: 3.776s Load: 0.016s, Pack+Encode: 2.149s, Decode+Unpack: 1.612s +---------------------- -------------------------------------------------------- +💾 Converting with 47.8673 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,668B, BPFP=0.6934 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,948B, BPFP=3.0919 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,996B, BPFP=1.6232 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 59,056B, BPFP=2.9959 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,176B, BPFP=1.6830 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,928B, BPFP=2.9387 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,380B, BPFP=1.5412 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 59,068B, BPFP=2.9966 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,964B, BPFP=2.3825 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,708B, BPFP=2.8768 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 125,236B, BPFP=0.9076 +⌛️ [2/4] FRONTEND: Frontend time: 1.897s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12827574 31.09155908 + layer.0.v_cache 0.00001866 0.00232902 + layer.1.k_cache 0.53653341 3.24751222 + layer.1.v_cache 0.00000626 0.00090215 + layer.2.k_cache 0.01950284 0.58583881 + layer.2.v_cache 0.00002083 0.00271887 + layer.3.k_cache 0.04340698 4.34321337 + layer.3.v_cache 0.00002181 0.00312298 + layer.4.k_cache 0.00072091 0.08173751 + layer.4.v_cache 0.00005393 0.00604522 + layer.4.output 0.04342448 75.23856389 + ------------------------------------------------------------------------------------- + TOTAL 0.06073722 33.29617214 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 575128 +BPFP 1.7163 bits/point +EBPFP 3.4325 equivalent bits/point +MSE 33.296172 +---------------------- -------------------------------------------------------- +Time: 3.289s Load: 0.011s, Pack+Encode: 1.897s, Decode+Unpack: 1.381s +---------------------- -------------------------------------------------------- +💾 Converting with 33.2962 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,300B, BPFP=0.7225 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,400B, BPFP=3.4305 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,532B, BPFP=1.4410 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,072B, BPFP=3.3524 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,328B, BPFP=1.7815 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,624B, BPFP=3.3261 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,564B, BPFP=1.6191 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,092B, BPFP=3.4124 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,156B, BPFP=2.5938 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,576B, BPFP=3.2646 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 106,644B, BPFP=0.8949 +⌛️ [2/4] FRONTEND: Frontend time: 1.894s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702504 29.49507313 + layer.0.v_cache 0.00001641 0.00239402 + layer.1.k_cache 0.51130361 2.46181104 + layer.1.v_cache 0.00000656 0.00099937 + layer.2.k_cache 0.01810413 0.60987292 + layer.2.v_cache 0.00002122 0.00291759 + layer.3.k_cache 0.02476560 3.53514845 + layer.3.v_cache 0.00002197 0.00317661 + layer.4.k_cache 0.00071356 0.08204745 + layer.4.v_cache 0.00005290 0.00636488 + layer.4.output 0.00501064 186.57500336 + ------------------------------------------------------------------------------------- + TOTAL 0.04100620 78.95440170 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 531288 +BPFP 1.8358 bits/point +EBPFP 3.6716 equivalent bits/point +MSE 78.954402 +---------------------- -------------------------------------------------------- +Time: 3.279s Load: 0.009s, Pack+Encode: 1.894s, Decode+Unpack: 1.377s +---------------------- -------------------------------------------------------- +💾 Converting with 78.9544 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,864B, BPFP=0.7254 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,712B, BPFP=3.2527 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,016B, BPFP=1.7372 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,248B, BPFP=3.1549 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,412B, BPFP=1.7636 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,568B, BPFP=3.1095 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,676B, BPFP=1.5809 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,524B, BPFP=3.1733 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,688B, BPFP=2.5833 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,312B, BPFP=3.0256 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,696B, BPFP=0.9129 +⌛️ [2/4] FRONTEND: Frontend time: 1.793s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.255s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11892304 29.99983307 + layer.0.v_cache 0.00001803 0.00245419 + layer.1.k_cache 0.31174104 2.82432830 + layer.1.v_cache 0.00000657 0.00096488 + layer.2.k_cache 0.01452858 0.57432628 + layer.2.v_cache 0.00002161 0.00286061 + layer.3.k_cache 0.02263702 3.72393199 + layer.3.v_cache 0.00002141 0.00327997 + layer.4.k_cache 0.00068312 0.08311448 + layer.4.v_cache 0.00005251 0.00615596 + layer.4.output 1.30838121 206.97754502 + ------------------------------------------------------------------------------------- + TOTAL 0.56631185 87.41553323 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 456716 +BPFP 1.7939 bits/point +EBPFP 3.5878 equivalent bits/point +MSE 87.415533 +---------------------- -------------------------------------------------------- +Time: 3.057s Load: 0.010s, Pack+Encode: 1.793s, Decode+Unpack: 1.255s +---------------------- -------------------------------------------------------- +💾 Converting with 87.4155 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 233, 128) +Output shape: (1, 233, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.output: torch.Size([1, 233, 3584]) -> torch.Size([1, 1, 233, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,768B, BPFP=0.7221 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,940B, BPFP=3.2819 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,992B, BPFP=1.7430 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,648B, BPFP=3.1953 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,352B, BPFP=1.7672 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,524B, BPFP=3.1199 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,080B, BPFP=1.6148 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,192B, BPFP=3.1647 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,872B, BPFP=2.5397 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,204B, BPFP=3.0314 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,588B, BPFP=0.8774 +⌛️ [2/4] FRONTEND: Frontend time: 1.785s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.242s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13365312 29.66184323 + layer.0.v_cache 0.00001735 0.00246240 + layer.1.k_cache 0.40872042 2.98787863 + layer.1.v_cache 0.00000792 0.00104065 + layer.2.k_cache 0.01308113 0.54365474 + layer.2.v_cache 0.00002137 0.00297315 + layer.3.k_cache 0.00601081 4.09968587 + layer.3.v_cache 0.00002094 0.00321176 + layer.4.k_cache 0.00068546 0.08564251 + layer.4.v_cache 0.00005338 0.00627092 + layer.4.output 1.31399239 205.32177345 + ------------------------------------------------------------------------------------- + TOTAL 0.57413051 86.74394577 + (elements=2,028,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2028032 +Total Bytes 452160 +BPFP 1.7836 bits/point +EBPFP 3.5673 equivalent bits/point +MSE 86.743946 +---------------------- -------------------------------------------------------- +Time: 3.037s Load: 0.010s, Pack+Encode: 1.785s, Decode+Unpack: 1.242s +---------------------- -------------------------------------------------------- +💾 Converting with 86.7439 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 148, 128) +Output shape: (1, 148, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.output: torch.Size([1, 148, 3584]) -> torch.Size([1, 1, 148, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,560B, BPFP=0.7981 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 34,984B, BPFP=3.6934 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 14,324B, BPFP=1.5122 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 34,268B, BPFP=3.6178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,068B, BPFP=1.9075 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 33,840B, BPFP=3.5726 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,268B, BPFP=1.7175 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 34,268B, BPFP=3.6178 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,232B, BPFP=2.7694 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,248B, BPFP=3.5101 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 58,232B, BPFP=0.8783 +⌛️ [2/4] FRONTEND: Frontend time: 1.673s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.138s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08608804 29.43753299 + layer.0.v_cache 0.00001767 0.00259947 + layer.1.k_cache 0.12276106 1.82575741 + layer.1.v_cache 0.00000594 0.00102750 + layer.2.k_cache 0.00886028 0.66925765 + layer.2.v_cache 0.00002088 0.00306139 + layer.3.k_cache 0.01513371 3.49547433 + layer.3.v_cache 0.00002085 0.00347043 + layer.4.k_cache 0.00068430 0.08431092 + layer.4.v_cache 0.00005331 0.00646750 + layer.4.output 0.08957551 315.85928451 + ------------------------------------------------------------------------------------- + TOTAL 0.05062792 132.14964419 + (elements=1,288,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1288192 +Total Bytes 311292 +BPFP 1.9332 bits/point +EBPFP 3.8664 equivalent bits/point +MSE 132.149644 +---------------------- -------------------------------------------------------- +Time: 2.816s Load: 0.005s, Pack+Encode: 1.673s, Decode+Unpack: 1.138s +---------------------- -------------------------------------------------------- +💾 Converting with 132.1496 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 170, 128) +Output shape: (1, 170, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.output: torch.Size([1, 170, 3584]) -> torch.Size([1, 1, 170, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,348B, BPFP=0.7673 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,588B, BPFP=3.3629 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,824B, BPFP=1.4544 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,656B, BPFP=3.2772 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,848B, BPFP=1.8243 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,776B, BPFP=3.1963 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,700B, BPFP=1.7188 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,532B, BPFP=3.2658 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,204B, BPFP=2.5923 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,832B, BPFP=3.1096 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 68,956B, BPFP=0.9054 +⌛️ [2/4] FRONTEND: Frontend time: 1.654s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.135s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13519069 27.68610983 + layer.0.v_cache 0.00001779 0.00275047 + layer.1.k_cache 0.14803200 2.60218290 + layer.1.v_cache 0.00000678 0.00109632 + layer.2.k_cache 0.02267574 0.54965946 + layer.2.v_cache 0.00002005 0.00301698 + layer.3.k_cache 0.04618364 4.25891113 + layer.3.v_cache 0.00002282 0.00346229 + layer.4.k_cache 0.00077641 0.08725622 + layer.4.v_cache 0.00005264 0.00663347 + layer.4.output 0.00945594 285.89041492 + ------------------------------------------------------------------------------------- + TOTAL 0.02465707 119.79023432 + (elements=1,479,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1479680 +Total Bytes 336264 +BPFP 1.8180 bits/point +EBPFP 3.6361 equivalent bits/point +MSE 119.790234 +---------------------- -------------------------------------------------------- +Time: 2.795s Load: 0.006s, Pack+Encode: 1.654s, Decode+Unpack: 1.135s +---------------------- -------------------------------------------------------- +💾 Converting with 119.7902 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,212B, BPFP=0.7730 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,796B, BPFP=3.4635 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,308B, BPFP=1.4409 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,968B, BPFP=3.3855 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,500B, BPFP=1.8355 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,156B, BPFP=3.3091 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,892B, BPFP=1.6841 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,980B, BPFP=3.3867 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,292B, BPFP=2.6630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,240B, BPFP=3.2229 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,716B, BPFP=0.9374 +⌛️ [2/4] FRONTEND: Frontend time: 1.667s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.142s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12520085 27.86213997 + layer.0.v_cache 0.00001716 0.00270627 + layer.1.k_cache 0.15152623 2.09657527 + layer.1.v_cache 0.00000633 0.00104803 + layer.2.k_cache 0.01605446 0.64687085 + layer.2.v_cache 0.00002200 0.00321291 + layer.3.k_cache 0.04332644 4.05983798 + layer.3.v_cache 0.00002349 0.00343647 + layer.4.k_cache 0.00069336 0.08755579 + layer.4.v_cache 0.00005614 0.00640815 + layer.4.output 0.00966471 284.59020009 + ------------------------------------------------------------------------------------- + TOTAL 0.02379879 119.22948190 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 337060 +BPFP 1.8663 bits/point +EBPFP 3.7325 equivalent bits/point +MSE 119.229482 +---------------------- -------------------------------------------------------- +Time: 2.817s Load: 0.008s, Pack+Encode: 1.667s, Decode+Unpack: 1.142s +---------------------- -------------------------------------------------------- +💾 Converting with 119.2295 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 181, 128) +Output shape: (1, 181, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.output: torch.Size([1, 181, 3584]) -> torch.Size([1, 1, 181, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,560B, BPFP=0.7390 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,156B, BPFP=3.2075 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,288B, BPFP=1.4924 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,944B, BPFP=3.1029 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,652B, BPFP=1.7828 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,852B, BPFP=3.0086 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,436B, BPFP=1.6778 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,640B, BPFP=3.0767 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,956B, BPFP=2.4997 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,936B, BPFP=2.9296 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,028B, BPFP=0.8759 +⌛️ [2/4] FRONTEND: Frontend time: 1.664s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.137s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13566811 34.36939421 + layer.0.v_cache 0.00002016 0.00269981 + layer.1.k_cache 0.14782082 2.75512527 + layer.1.v_cache 0.00000634 0.00108533 + layer.2.k_cache 0.01546242 0.53632991 + layer.2.v_cache 0.00002121 0.00306603 + layer.3.k_cache 0.02895848 3.93289741 + layer.3.v_cache 0.00002197 0.00350993 + layer.4.k_cache 0.00067433 0.08493533 + layer.4.v_cache 0.00005229 0.00650567 + layer.4.output 0.00892717 266.37906965 + ------------------------------------------------------------------------------------- + TOTAL 0.02301155 112.13817862 + (elements=1,575,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1575424 +Total Bytes 343448 +BPFP 1.7440 bits/point +EBPFP 3.4881 equivalent bits/point +MSE 112.138179 +---------------------- -------------------------------------------------------- +Time: 2.807s Load: 0.006s, Pack+Encode: 1.664s, Decode+Unpack: 1.137s +---------------------- -------------------------------------------------------- +💾 Converting with 112.1382 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 186, 128) +Output shape: (1, 186, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.output: torch.Size([1, 186, 3584]) -> torch.Size([1, 1, 186, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,576B, BPFP=0.7204 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,848B, BPFP=3.0954 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,876B, BPFP=1.5857 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,604B, BPFP=2.9909 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,516B, BPFP=1.7235 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,760B, BPFP=2.9200 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,312B, BPFP=1.6223 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,768B, BPFP=3.0047 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,012B, BPFP=2.4372 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,060B, BPFP=2.8612 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 73,476B, BPFP=0.8818 +⌛️ [2/4] FRONTEND: Frontend time: 1.670s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.138s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09795722 29.67161773 + layer.0.v_cache 0.00001815 0.00266347 + layer.1.k_cache 0.13430934 3.08996155 + layer.1.v_cache 0.00000632 0.00106211 + layer.2.k_cache 0.01830538 0.55940977 + layer.2.v_cache 0.00002120 0.00315253 + layer.3.k_cache 0.04892764 4.42096341 + layer.3.v_cache 0.00002113 0.00359211 + layer.4.k_cache 0.00067235 0.08765074 + layer.4.v_cache 0.00008605 0.00663053 + layer.4.output 0.00869856 275.97273425 + ------------------------------------------------------------------------------------- + TOTAL 0.02124792 115.86210846 + (elements=1,618,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1618944 +Total Bytes 346808 +BPFP 1.7137 bits/point +EBPFP 3.4275 equivalent bits/point +MSE 115.862108 +---------------------- -------------------------------------------------------- +Time: 2.815s Load: 0.008s, Pack+Encode: 1.670s, Decode+Unpack: 1.138s +---------------------- -------------------------------------------------------- +💾 Converting with 115.8621 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 163, 128) +Output shape: (1, 163, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.output: torch.Size([1, 163, 3584]) -> torch.Size([1, 1, 163, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 7,892B, BPFP=0.7565 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,536B, BPFP=3.5023 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,088B, BPFP=1.4463 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,668B, BPFP=3.4191 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,088B, BPFP=1.8298 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,440B, BPFP=3.3014 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,380B, BPFP=1.6660 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,552B, BPFP=3.4080 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 27,872B, BPFP=2.6718 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,024B, BPFP=3.2615 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 66,328B, BPFP=0.9083 +⌛️ [2/4] FRONTEND: Frontend time: 1.640s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.134s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11530171 29.26954623 + layer.0.v_cache 0.00001813 0.00264194 + layer.1.k_cache 0.14319220 1.95961270 + layer.1.v_cache 0.00000612 0.00103573 + layer.2.k_cache 0.02368733 0.73561873 + layer.2.v_cache 0.00002106 0.00304291 + layer.3.k_cache 0.02525310 3.82425133 + layer.3.v_cache 0.00002161 0.00347907 + layer.4.k_cache 0.00069984 0.08513019 + layer.4.v_cache 0.00005709 0.00640208 + layer.4.output 0.00982084 308.87075482 + ------------------------------------------------------------------------------------- + TOTAL 0.02217671 129.29329674 + (elements=1,418,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1418752 +Total Bytes 329868 +BPFP 1.8600 bits/point +EBPFP 3.7201 equivalent bits/point +MSE 129.293297 +---------------------- -------------------------------------------------------- +Time: 2.780s Load: 0.006s, Pack+Encode: 1.640s, Decode+Unpack: 1.134s +---------------------- -------------------------------------------------------- +💾 Converting with 129.2933 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 191, 128) +Output shape: (1, 191, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.output: torch.Size([1, 191, 3584]) -> torch.Size([1, 1, 191, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,604B, BPFP=0.7039 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,788B, BPFP=3.0095 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,968B, BPFP=1.3881 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,672B, BPFP=2.9182 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,484B, BPFP=1.6757 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,056B, BPFP=2.8678 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,040B, BPFP=1.5576 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,844B, BPFP=2.9323 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 29,468B, BPFP=2.4107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,444B, BPFP=2.8177 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 77,640B, BPFP=0.9073 +⌛️ [2/4] FRONTEND: Frontend time: 1.667s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.143s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15112510 36.10381729 + layer.0.v_cache 0.00002069 0.00256873 + layer.1.k_cache 0.18491925 2.39720657 + layer.1.v_cache 0.00000629 0.00100066 + layer.2.k_cache 0.02122488 0.47433851 + layer.2.v_cache 0.00002327 0.00315246 + layer.3.k_cache 0.01259250 4.04985933 + layer.3.v_cache 0.00002182 0.00343504 + layer.4.k_cache 0.00068006 0.08131435 + layer.4.v_cache 0.00005885 0.00613929 + layer.4.output 0.00845948 267.03206806 + ------------------------------------------------------------------------------------- + TOTAL 0.02528759 112.49101816 + (elements=1,662,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1662464 +Total Bytes 350008 +BPFP 1.6843 bits/point +EBPFP 3.3686 equivalent bits/point +MSE 112.491018 +---------------------- -------------------------------------------------------- +Time: 2.820s Load: 0.009s, Pack+Encode: 1.667s, Decode+Unpack: 1.143s +---------------------- -------------------------------------------------------- +💾 Converting with 112.4910 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 174, 128) +Output shape: (1, 174, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.output: torch.Size([1, 174, 3584]) -> torch.Size([1, 1, 174, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,420B, BPFP=0.7561 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,600B, BPFP=3.2866 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,072B, BPFP=1.4432 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,528B, BPFP=3.1904 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,164B, BPFP=1.8107 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,952B, BPFP=3.1386 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,072B, BPFP=1.7126 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,700B, BPFP=3.2058 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 28,852B, BPFP=2.5909 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,236B, BPFP=3.0744 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 70,540B, BPFP=0.9049 +⌛️ [2/4] FRONTEND: Frontend time: 1.669s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.143s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14146894 30.21402714 + layer.0.v_cache 0.00001664 0.00257026 + layer.1.k_cache 0.16292955 2.13463601 + layer.1.v_cache 0.00000602 0.00100525 + layer.2.k_cache 0.01834387 0.50760887 + layer.2.v_cache 0.00002001 0.00306173 + layer.3.k_cache 0.03015787 4.00410970 + layer.3.v_cache 0.00002180 0.00349928 + layer.4.k_cache 0.00067479 0.08683530 + layer.4.v_cache 0.00005130 0.00642533 + layer.4.output 0.00923280 268.76937089 + ------------------------------------------------------------------------------------- + TOTAL 0.02460708 112.84408089 + (elements=1,514,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1514496 +Total Bytes 340136 +BPFP 1.7967 bits/point +EBPFP 3.5934 equivalent bits/point +MSE 112.844081 +---------------------- -------------------------------------------------------- +Time: 2.820s Load: 0.009s, Pack+Encode: 1.669s, Decode+Unpack: 1.143s +---------------------- -------------------------------------------------------- +💾 Converting with 112.8441 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.7826 bits/point +Avg EBPFP 3.5653 equivalent bits/point +Avg MSE 87.724079 +Avg Time 3.042s +------------------------ ---------------------------- diff --git a/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..e88a847191d25eb9c47fde0799ca4dfc2a9303b4 --- /dev/null +++ b/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 520 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa +Output output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa +---------------- ------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,576B, BPFP=0.6823 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,788B, BPFP=3.1352 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,816B, BPFP=1.4549 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,992B, BPFP=3.0378 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,968B, BPFP=1.6259 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,784B, BPFP=2.9722 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,936B, BPFP=1.4614 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,192B, BPFP=3.0486 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,348B, BPFP=2.4060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,112B, BPFP=2.9900 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,964B, BPFP=0.6895 +⌛️ [2/4] FRONTEND: Frontend time: 2.343s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.423s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14134069 30.80447388 + layer.0.v_cache 0.00001405 0.00186935 + layer.1.k_cache 0.62039534 2.95353720 + layer.1.v_cache 0.00000582 0.00081585 + layer.2.k_cache 0.00676800 0.70363077 + layer.2.v_cache 0.00001893 0.00259665 + layer.3.k_cache 0.03896382 4.10915205 + layer.3.v_cache 0.00001942 0.00283891 + layer.4.k_cache 0.00069280 0.07595543 + layer.4.v_cache 0.00005238 0.00583126 + layer.4.output 0.00804578 169.52219742 + ------------------------------------------------------------------------------------- + TOTAL 0.05085834 72.07741666 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 509476 +BPFP 1.6259 bits/point +EBPFP 3.2519 equivalent bits/point +MSE 72.077417 +---------------------- -------------------------------------------------------- +Time: 3.776s Load: 0.010s, Pack+Encode: 2.343s, Decode+Unpack: 1.423s +---------------------- -------------------------------------------------------- +💾 Converting with 72.0774 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-105.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,124B, BPFP=0.6999 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,916B, BPFP=3.1419 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,364B, BPFP=1.4593 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,376B, BPFP=3.0597 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,324B, BPFP=1.6704 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,144B, BPFP=2.9940 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,224B, BPFP=1.5051 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,504B, BPFP=3.0666 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,392B, BPFP=2.4206 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,976B, BPFP=2.9851 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 97,364B, BPFP=0.7417 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15682776 30.09264345 + layer.0.v_cache 0.00001501 0.00194375 + layer.1.k_cache 0.54447479 2.87312999 + layer.1.v_cache 0.00000580 0.00086019 + layer.2.k_cache 0.00853899 0.60014494 + layer.2.v_cache 0.00001972 0.00265693 + layer.3.k_cache 0.02401569 4.36664756 + layer.3.v_cache 0.00001939 0.00283019 + layer.4.k_cache 0.00071059 0.07562198 + layer.4.v_cache 0.00005329 0.00588100 + layer.4.output 0.05121508 168.85590870 + ------------------------------------------------------------------------------------- + TOTAL 0.06430510 71.76551299 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 528708 +BPFP 1.6585 bits/point +EBPFP 3.3170 equivalent bits/point +MSE 71.765513 +---------------------- -------------------------------------------------------- +Time: 3.234s Load: 0.010s, Pack+Encode: 1.855s, Decode+Unpack: 1.369s +---------------------- -------------------------------------------------------- +💾 Converting with 71.7655 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-106.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,160B, BPFP=0.6994 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,820B, BPFP=3.1261 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,832B, BPFP=1.4792 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,544B, BPFP=3.0582 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,984B, BPFP=1.6467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,036B, BPFP=2.9781 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,120B, BPFP=1.4945 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,496B, BPFP=3.0557 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,896B, BPFP=2.3861 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,664B, BPFP=2.9583 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,012B, BPFP=0.7441 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15363285 29.64546463 + layer.0.v_cache 0.00001412 0.00183726 + layer.1.k_cache 0.60852461 2.93327363 + layer.1.v_cache 0.00000580 0.00079599 + layer.2.k_cache 0.01182756 0.60417689 + layer.2.v_cache 0.00001863 0.00246610 + layer.3.k_cache 0.03467803 4.20923890 + layer.3.v_cache 0.00001919 0.00267767 + layer.4.k_cache 0.00072319 0.07574495 + layer.4.v_cache 0.00005301 0.00556739 + layer.4.output 0.05035121 161.65470117 + ------------------------------------------------------------------------------------- + TOTAL 0.06835032 68.76847950 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 528564 +BPFP 1.6524 bits/point +EBPFP 3.3048 equivalent bits/point +MSE 68.768480 +---------------------- -------------------------------------------------------- +Time: 3.239s Load: 0.010s, Pack+Encode: 1.856s, Decode+Unpack: 1.373s +---------------------- -------------------------------------------------------- +💾 Converting with 68.7685 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-108.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-108.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,716B, BPFP=0.7021 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,300B, BPFP=3.1636 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,704B, BPFP=1.5848 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,268B, BPFP=3.0515 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,184B, BPFP=1.6665 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,888B, BPFP=3.0305 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,332B, BPFP=1.5091 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,884B, BPFP=3.1407 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,044B, BPFP=2.4318 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,744B, BPFP=3.0777 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,540B, BPFP=0.7220 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12197856 30.17343474 + layer.0.v_cache 0.00001365 0.00191102 + layer.1.k_cache 0.62524759 2.80322503 + layer.1.v_cache 0.00000551 0.00080169 + layer.2.k_cache 0.01486346 0.73043791 + layer.2.v_cache 0.00001981 0.00268722 + layer.3.k_cache 0.07387851 3.54599290 + layer.3.v_cache 0.00001943 0.00282541 + layer.4.k_cache 0.00069560 0.07415083 + layer.4.v_cache 0.00005405 0.00590089 + layer.4.output 0.01098318 172.48023410 + ------------------------------------------------------------------------------------- + TOTAL 0.05374461 73.21782390 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 514604 +BPFP 1.6713 bits/point +EBPFP 3.3426 equivalent bits/point +MSE 73.217824 +---------------------- -------------------------------------------------------- +Time: 3.233s Load: 0.012s, Pack+Encode: 1.856s, Decode+Unpack: 1.366s +---------------------- -------------------------------------------------------- +💾 Converting with 73.2178 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-110.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-110.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,728B, BPFP=0.6978 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,320B, BPFP=3.1425 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 32,596B, BPFP=1.7871 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,008B, BPFP=3.0706 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,824B, BPFP=1.6351 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,684B, BPFP=2.9980 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,884B, BPFP=1.4739 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,280B, BPFP=3.0855 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,248B, BPFP=2.4259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,244B, BPFP=3.0287 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,940B, BPFP=0.7201 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15362121 29.97821752 + layer.0.v_cache 0.00001420 0.00188961 + layer.1.k_cache 0.52952420 3.43861127 + layer.1.v_cache 0.00000567 0.00081697 + layer.2.k_cache 0.00939641 0.61191856 + layer.2.v_cache 0.00001891 0.00248234 + layer.3.k_cache 0.04339437 4.16395842 + layer.3.v_cache 0.00002732 0.00285195 + layer.4.k_cache 0.00068326 0.07458473 + layer.4.v_cache 0.00005156 0.00574363 + layer.4.output 0.00776138 178.23881579 + ------------------------------------------------------------------------------------- + TOTAL 0.04653334 75.64428150 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 517756 +BPFP 1.6697 bits/point +EBPFP 3.3395 equivalent bits/point +MSE 75.644282 +---------------------- -------------------------------------------------------- +Time: 3.231s Load: 0.010s, Pack+Encode: 1.856s, Decode+Unpack: 1.365s +---------------------- -------------------------------------------------------- +💾 Converting with 75.6443 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-111.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,148B, BPFP=0.6988 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,644B, BPFP=3.1167 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,836B, BPFP=1.4794 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,160B, BPFP=3.0378 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,028B, BPFP=1.6490 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,252B, BPFP=2.9896 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,200B, BPFP=1.4987 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,460B, BPFP=3.0538 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,412B, BPFP=2.4135 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,628B, BPFP=2.9564 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,932B, BPFP=0.7663 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14921317 31.69118038 + layer.0.v_cache 0.00001395 0.00191280 + layer.1.k_cache 0.64083395 3.00982645 + layer.1.v_cache 0.00000574 0.00084160 + layer.2.k_cache 0.00578804 0.65617474 + layer.2.v_cache 0.00001912 0.00254812 + layer.3.k_cache 0.02848778 3.87040119 + layer.3.v_cache 0.00001956 0.00277694 + layer.4.k_cache 0.00070939 0.07550246 + layer.4.v_cache 0.00005206 0.00570625 + layer.4.output 0.04916091 166.48234026 + ------------------------------------------------------------------------------------- + TOTAL 0.06878054 70.86430898 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 531700 +BPFP 1.6622 bits/point +EBPFP 3.3245 equivalent bits/point +MSE 70.864309 +---------------------- -------------------------------------------------------- +Time: 3.234s Load: 0.011s, Pack+Encode: 1.855s, Decode+Unpack: 1.369s +---------------------- -------------------------------------------------------- +💾 Converting with 70.8643 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-12.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-12.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 293, 128) +Output shape: (1, 293, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) -> torch.Size([1, 1, 293, 512]) + layer.4.output: torch.Size([1, 293, 3584]) -> torch.Size([1, 1, 293, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,032B, BPFP=0.6950 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,940B, BPFP=3.1431 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,836B, BPFP=1.4844 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,236B, BPFP=3.0523 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,576B, BPFP=1.6305 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,896B, BPFP=2.9808 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,876B, BPFP=1.4866 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,468B, BPFP=3.0646 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,044B, BPFP=2.4021 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,556B, BPFP=2.9627 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,432B, BPFP=0.7270 +⌛️ [2/4] FRONTEND: Frontend time: 1.882s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 293, 128]) + layer.0.v_cache: torch.Size([1, 4, 293, 128]) + layer.1.k_cache: torch.Size([1, 4, 293, 128]) + layer.1.v_cache: torch.Size([1, 4, 293, 128]) + layer.2.k_cache: torch.Size([1, 4, 293, 128]) + layer.2.v_cache: torch.Size([1, 4, 293, 128]) + layer.3.k_cache: torch.Size([1, 4, 293, 128]) + layer.3.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.k_cache: torch.Size([1, 4, 293, 128]) + layer.4.v_cache: torch.Size([1, 4, 293, 128]) + layer.4.output: torch.Size([1, 293, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12806672 31.96093750 + layer.0.v_cache 0.00001388 0.00185591 + layer.1.k_cache 0.56701790 2.89640105 + layer.1.v_cache 0.00000557 0.00080297 + layer.2.k_cache 0.01020011 0.60160036 + layer.2.v_cache 0.00001881 0.00259009 + layer.3.k_cache 0.04301871 4.27275716 + layer.3.v_cache 0.00002051 0.00273614 + layer.4.k_cache 0.00071755 0.07390899 + layer.4.v_cache 0.00005284 0.00563749 + layer.4.output 0.04940383 169.76225012 + ------------------------------------------------------------------------------------- + TOTAL 0.06440938 72.24441050 + (elements=2,550,272) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2550272 +Total Bytes 524892 +BPFP 1.6465 bits/point +EBPFP 3.2931 equivalent bits/point +MSE 72.244411 +---------------------- -------------------------------------------------------- +Time: 3.259s Load: 0.011s, Pack+Encode: 1.882s, Decode+Unpack: 1.366s +---------------------- -------------------------------------------------------- +💾 Converting with 72.2444 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-120.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,732B, BPFP=0.6980 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,536B, BPFP=3.1544 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,932B, BPFP=1.5862 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,116B, BPFP=3.0765 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,068B, BPFP=1.6485 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,764B, BPFP=3.0024 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,440B, BPFP=1.5044 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,316B, BPFP=3.0875 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,296B, BPFP=2.4285 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,864B, BPFP=3.0079 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,884B, BPFP=0.7118 +⌛️ [2/4] FRONTEND: Frontend time: 1.853s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12903612 29.58760280 + layer.0.v_cache 0.00001458 0.00189920 + layer.1.k_cache 0.59860808 3.12871565 + layer.1.v_cache 0.00000581 0.00081536 + layer.2.k_cache 0.00940504 0.78267233 + layer.2.v_cache 0.00001954 0.00261454 + layer.3.k_cache 0.02644065 4.09095138 + layer.3.v_cache 0.00002029 0.00281981 + layer.4.k_cache 0.00068693 0.07651517 + layer.4.v_cache 0.00005376 0.00588985 + layer.4.output 0.00862506 180.47500000 + ------------------------------------------------------------------------------------- + TOTAL 0.04850978 76.52973506 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 513948 +BPFP 1.6575 bits/point +EBPFP 3.3149 equivalent bits/point +MSE 76.529735 +---------------------- -------------------------------------------------------- +Time: 3.237s Load: 0.011s, Pack+Encode: 1.853s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 76.5297 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-129.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,528B, BPFP=0.6863 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,864B, BPFP=2.9862 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,460B, BPFP=1.5453 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,304B, BPFP=2.9071 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,976B, BPFP=1.6222 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,200B, BPFP=2.8511 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,136B, BPFP=1.4781 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,468B, BPFP=2.9154 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,700B, BPFP=2.3184 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,596B, BPFP=2.8204 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,784B, BPFP=0.6942 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16469599 30.97082044 + layer.0.v_cache 0.00001357 0.00186512 + layer.1.k_cache 0.70445866 3.07219032 + layer.1.v_cache 0.00000558 0.00080852 + layer.2.k_cache 0.01447187 0.62070648 + layer.2.v_cache 0.00001924 0.00254791 + layer.3.k_cache 0.02617162 4.39154608 + layer.3.v_cache 0.00001974 0.00273443 + layer.4.k_cache 0.00070375 0.07393148 + layer.4.v_cache 0.00005076 0.00561617 + layer.4.output 0.04814868 158.54506320 + ------------------------------------------------------------------------------------- + TOTAL 0.07339127 67.58577702 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 532016 +BPFP 1.5876 bits/point +EBPFP 3.1752 equivalent bits/point +MSE 67.585777 +---------------------- -------------------------------------------------------- +Time: 3.240s Load: 0.013s, Pack+Encode: 1.856s, Decode+Unpack: 1.371s +---------------------- -------------------------------------------------------- +💾 Converting with 67.5858 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-13.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-13.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,268B, BPFP=0.7126 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,404B, BPFP=3.2763 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,560B, BPFP=1.5428 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,248B, BPFP=3.2091 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,772B, BPFP=1.6712 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,668B, BPFP=3.1173 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,684B, BPFP=1.4919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,488B, BPFP=3.2230 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,240B, BPFP=2.4535 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,608B, BPFP=3.1138 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 82,760B, BPFP=0.6867 +⌛️ [2/4] FRONTEND: Frontend time: 1.853s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12132761 28.75325279 + layer.0.v_cache 0.00001366 0.00186339 + layer.1.k_cache 0.57241889 2.91863333 + layer.1.v_cache 0.00000556 0.00077333 + layer.2.k_cache 0.00926234 0.76147370 + layer.2.v_cache 0.00002145 0.00263686 + layer.3.k_cache 0.03109839 3.59500496 + layer.3.v_cache 0.00002257 0.00280927 + layer.4.k_cache 0.00070701 0.07317612 + layer.4.v_cache 0.00004920 0.00560883 + layer.4.output 0.01031505 179.55184546 + ------------------------------------------------------------------------------------- + TOTAL 0.04747836 76.05753828 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 492700 +BPFP 1.6835 bits/point +EBPFP 3.3669 equivalent bits/point +MSE 76.057538 +---------------------- -------------------------------------------------------- +Time: 3.232s Load: 0.011s, Pack+Encode: 1.853s, Decode+Unpack: 1.368s +---------------------- -------------------------------------------------------- +💾 Converting with 76.0575 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-133.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,364B, BPFP=0.7007 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,872B, BPFP=3.0868 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,460B, BPFP=1.5447 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,244B, BPFP=3.0015 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,532B, BPFP=1.6533 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,808B, BPFP=2.9262 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,044B, BPFP=1.4704 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,300B, BPFP=3.0044 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,544B, BPFP=2.3880 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,820B, BPFP=2.9268 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,036B, BPFP=0.7343 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.359s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14357947 30.47891870 + layer.0.v_cache 0.00001420 0.00196835 + layer.1.k_cache 0.63723908 3.17776346 + layer.1.v_cache 0.00000573 0.00082330 + layer.2.k_cache 0.00765126 0.59689679 + layer.2.v_cache 0.00002073 0.00256724 + layer.3.k_cache 0.02843528 4.25104497 + layer.3.v_cache 0.00002112 0.00284945 + layer.4.k_cache 0.00069171 0.07475365 + layer.4.v_cache 0.00005234 0.00578511 + layer.4.output 0.04871205 151.53093540 + ------------------------------------------------------------------------------------- + TOTAL 0.06815855 64.66528934 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 531024 +BPFP 1.6378 bits/point +EBPFP 3.2757 equivalent bits/point +MSE 64.665289 +---------------------- -------------------------------------------------------- +Time: 3.229s Load: 0.012s, Pack+Encode: 1.857s, Decode+Unpack: 1.359s +---------------------- -------------------------------------------------------- +💾 Converting with 64.6653 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-135.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,488B, BPFP=0.6910 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,952B, BPFP=3.0201 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,736B, BPFP=1.4721 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,344B, BPFP=2.9377 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,004B, BPFP=1.6395 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,364B, BPFP=2.8875 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,800B, BPFP=1.5266 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,292B, BPFP=2.9350 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,936B, BPFP=2.3533 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,664B, BPFP=2.8516 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,140B, BPFP=0.7402 +⌛️ [2/4] FRONTEND: Frontend time: 1.858s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.363s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16857396 31.55144403 + layer.0.v_cache 0.00001478 0.00193334 + layer.1.k_cache 0.58716421 3.16538326 + layer.1.v_cache 0.00000600 0.00086537 + layer.2.k_cache 0.00827454 0.59979628 + layer.2.v_cache 0.00002177 0.00263452 + layer.3.k_cache 0.02531170 4.37541624 + layer.3.v_cache 0.00001983 0.00289529 + layer.4.k_cache 0.00069463 0.07542666 + layer.4.v_cache 0.00005316 0.00583714 + layer.4.output 0.04836898 158.53780738 + ------------------------------------------------------------------------------------- + TOTAL 0.06639514 67.62036963 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 536720 +BPFP 1.6174 bits/point +EBPFP 3.2348 equivalent bits/point +MSE 67.620370 +---------------------- -------------------------------------------------------- +Time: 3.233s Load: 0.012s, Pack+Encode: 1.858s, Decode+Unpack: 1.363s +---------------------- -------------------------------------------------------- +💾 Converting with 67.6204 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-143.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,304B, BPFP=0.6976 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,680B, BPFP=3.0768 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,356B, BPFP=1.5392 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,388B, BPFP=3.0090 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,312B, BPFP=1.6418 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,056B, BPFP=2.9392 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,184B, BPFP=1.4778 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,184B, BPFP=2.9983 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,412B, BPFP=2.4335 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,896B, BPFP=2.9308 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,984B, BPFP=0.7414 +⌛️ [2/4] FRONTEND: Frontend time: 1.853s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.357s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14028978 30.05235096 + layer.0.v_cache 0.00001397 0.00192641 + layer.1.k_cache 0.59966084 3.05469672 + layer.1.v_cache 0.00000578 0.00084110 + layer.2.k_cache 0.01765286 0.61424378 + layer.2.v_cache 0.00001884 0.00254244 + layer.3.k_cache 0.02653811 4.43006805 + layer.3.v_cache 0.00001935 0.00279690 + layer.4.k_cache 0.00071527 0.07823766 + layer.4.v_cache 0.00005402 0.00587617 + layer.4.output 0.04907703 163.26697327 + ------------------------------------------------------------------------------------- + TOTAL 0.06638282 69.47719959 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 532756 +BPFP 1.6432 bits/point +EBPFP 3.2863 equivalent bits/point +MSE 69.477200 +---------------------- -------------------------------------------------------- +Time: 3.220s Load: 0.010s, Pack+Encode: 1.853s, Decode+Unpack: 1.357s +---------------------- -------------------------------------------------------- +💾 Converting with 69.4772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-146.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-146.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,396B, BPFP=0.7121 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,144B, BPFP=3.2826 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,032B, BPFP=1.6103 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,368B, BPFP=3.1806 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,864B, BPFP=1.6581 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,932B, BPFP=3.0981 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,308B, BPFP=1.5113 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,884B, BPFP=3.2102 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,232B, BPFP=2.4260 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,616B, BPFP=3.1374 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,664B, BPFP=0.7358 +⌛️ [2/4] FRONTEND: Frontend time: 1.841s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13255837 30.25016515 + layer.0.v_cache 0.00001394 0.00183261 + layer.1.k_cache 0.54921044 3.02767137 + layer.1.v_cache 0.00000578 0.00080572 + layer.2.k_cache 0.01292971 0.79391575 + layer.2.v_cache 0.00001901 0.00265551 + layer.3.k_cache 0.07190662 3.84875174 + layer.3.v_cache 0.00001944 0.00281640 + layer.4.k_cache 0.00068159 0.07318389 + layer.4.v_cache 0.00005125 0.00588238 + layer.4.output 0.00790709 153.39925814 + ------------------------------------------------------------------------------------- + TOTAL 0.04839681 65.40014632 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 504440 +BPFP 1.7046 bits/point +EBPFP 3.4091 equivalent bits/point +MSE 65.400146 +---------------------- -------------------------------------------------------- +Time: 3.206s Load: 0.011s, Pack+Encode: 1.841s, Decode+Unpack: 1.354s +---------------------- -------------------------------------------------------- +💾 Converting with 65.4001 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-148.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,556B, BPFP=0.6957 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,752B, BPFP=3.1999 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,044B, BPFP=1.6647 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,220B, BPFP=3.1150 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,444B, BPFP=1.6868 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,216B, BPFP=3.0594 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,476B, BPFP=1.5224 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,076B, BPFP=3.1625 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,092B, BPFP=2.4430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,580B, BPFP=3.0796 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,664B, BPFP=0.7651 +⌛️ [2/4] FRONTEND: Frontend time: 1.860s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 32.233s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13695534 30.21034187 + layer.0.v_cache 0.00001452 0.00195714 + layer.1.k_cache 0.52272531 2.89610648 + layer.1.v_cache 0.00000574 0.00084514 + layer.2.k_cache 0.00793556 0.66334902 + layer.2.v_cache 0.00001907 0.00263768 + layer.3.k_cache 0.04682703 3.89168294 + layer.3.v_cache 0.00002008 0.00291249 + layer.4.k_cache 0.00068945 0.07540486 + layer.4.v_cache 0.00005221 0.00582652 + layer.4.output 0.00941004 179.51543503 + ------------------------------------------------------------------------------------- + TOTAL 0.04594792 76.13877114 + (elements=2,454,528) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2454528 +Total Bytes 523120 +BPFP 1.7050 bits/point +EBPFP 3.4100 equivalent bits/point +MSE 76.138771 +---------------------- --------------------------------------------------------- +Time: 34.105s Load: 0.011s, Pack+Encode: 1.860s, Decode+Unpack: 32.233s +---------------------- --------------------------------------------------------- +💾 Converting with 76.1388 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-151.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-151.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,616B, BPFP=0.7040 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,040B, BPFP=3.2388 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,560B, BPFP=1.4821 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,012B, BPFP=3.1257 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,468B, BPFP=1.6444 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,772B, BPFP=3.0565 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,216B, BPFP=1.5188 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,376B, BPFP=3.1460 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,368B, BPFP=2.4201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,320B, BPFP=3.0871 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,104B, BPFP=0.7103 +⌛️ [2/4] FRONTEND: Frontend time: 1.849s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14340984 30.45676270 + layer.0.v_cache 0.00001397 0.00187794 + layer.1.k_cache 0.51622396 2.99717058 + layer.1.v_cache 0.00000557 0.00081350 + layer.2.k_cache 0.00797994 0.65036289 + layer.2.v_cache 0.00001813 0.00265239 + layer.3.k_cache 0.02756977 3.90995789 + layer.3.v_cache 0.00001890 0.00282083 + layer.4.k_cache 0.00068956 0.07266622 + layer.4.v_cache 0.00005177 0.00566903 + layer.4.output 0.00947078 171.24035395 + ------------------------------------------------------------------------------------- + TOTAL 0.04483982 72.75195480 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 508852 +BPFP 1.6703 bits/point +EBPFP 3.3407 equivalent bits/point +MSE 72.751955 +---------------------- -------------------------------------------------------- +Time: 3.239s Load: 0.012s, Pack+Encode: 1.849s, Decode+Unpack: 1.378s +---------------------- -------------------------------------------------------- +💾 Converting with 72.7520 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-160.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,100B, BPFP=0.6962 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,460B, BPFP=3.1069 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,692B, BPFP=1.4186 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,984B, BPFP=3.0285 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,060B, BPFP=1.6507 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,792B, BPFP=2.9651 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,532B, BPFP=1.4632 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,852B, BPFP=3.0215 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,160B, BPFP=2.4001 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,384B, BPFP=2.9435 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,980B, BPFP=0.7135 +⌛️ [2/4] FRONTEND: Frontend time: 1.851s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13622742 29.96121652 + layer.0.v_cache 0.00001380 0.00190596 + layer.1.k_cache 0.61520739 2.84278050 + layer.1.v_cache 0.00000584 0.00085190 + layer.2.k_cache 0.00900942 0.57225649 + layer.2.v_cache 0.00001916 0.00262176 + layer.3.k_cache 0.02667915 4.34858880 + layer.3.v_cache 0.00001881 0.00279668 + layer.4.k_cache 0.00071466 0.07504367 + layer.4.v_cache 0.00005760 0.00593864 + layer.4.output 0.04972813 170.32481475 + ------------------------------------------------------------------------------------- + TOTAL 0.06682648 72.35810024 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 520996 +BPFP 1.6288 bits/point +EBPFP 3.2575 equivalent bits/point +MSE 72.358100 +---------------------- -------------------------------------------------------- +Time: 3.237s Load: 0.009s, Pack+Encode: 1.851s, Decode+Unpack: 1.377s +---------------------- -------------------------------------------------------- +💾 Converting with 72.3581 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-163.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-163.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,344B, BPFP=0.6973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,712B, BPFP=3.0681 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,896B, BPFP=1.6145 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,124B, BPFP=2.9852 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,740B, BPFP=1.6587 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,068B, BPFP=2.9300 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,548B, BPFP=1.4918 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,632B, BPFP=3.0117 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,472B, BPFP=2.3763 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,292B, BPFP=2.9417 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,108B, BPFP=0.7324 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11478914 31.54887058 + layer.0.v_cache 0.00001377 0.00184660 + layer.1.k_cache 0.64218915 3.24067285 + layer.1.v_cache 0.00000556 0.00076706 + layer.2.k_cache 0.00898698 0.60311726 + layer.2.v_cache 0.00001980 0.00252690 + layer.3.k_cache 0.05442990 4.33069909 + layer.3.v_cache 0.00001964 0.00270441 + layer.4.k_cache 0.00070461 0.07269449 + layer.4.v_cache 0.00005015 0.00558451 + layer.4.output 0.05017195 159.22479694 + ------------------------------------------------------------------------------------- + TOTAL 0.06896543 67.90488602 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 533936 +BPFP 1.6413 bits/point +EBPFP 3.2826 equivalent bits/point +MSE 67.904886 +---------------------- -------------------------------------------------------- +Time: 3.253s Load: 0.012s, Pack+Encode: 1.856s, Decode+Unpack: 1.386s +---------------------- -------------------------------------------------------- +💾 Converting with 67.9049 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-164.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-164.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,144B, BPFP=0.7160 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,944B, BPFP=3.2986 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,920B, BPFP=1.3514 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 54,004B, BPFP=3.1842 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,272B, BPFP=1.6670 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 52,816B, BPFP=3.1142 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,804B, BPFP=1.5215 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,104B, BPFP=3.2491 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,856B, BPFP=2.4679 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,048B, BPFP=3.1278 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,092B, BPFP=0.6831 +⌛️ [2/4] FRONTEND: Frontend time: 1.852s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14755792 28.54480579 + layer.0.v_cache 0.00001375 0.00190744 + layer.1.k_cache 0.51940866 2.15399723 + layer.1.v_cache 0.00000586 0.00088034 + layer.2.k_cache 0.00816260 0.74553919 + layer.2.v_cache 0.00001900 0.00268610 + layer.3.k_cache 0.02921419 3.87815356 + layer.3.v_cache 0.00001979 0.00286924 + layer.4.k_cache 0.00071017 0.07626858 + layer.4.v_cache 0.00005507 0.00571453 + layer.4.output 0.00937099 186.50341981 + ------------------------------------------------------------------------------------- + TOTAL 0.04533906 78.87863298 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 483004 +BPFP 1.6752 bits/point +EBPFP 3.3505 equivalent bits/point +MSE 78.878633 +---------------------- -------------------------------------------------------- +Time: 3.228s Load: 0.011s, Pack+Encode: 1.852s, Decode+Unpack: 1.366s +---------------------- -------------------------------------------------------- +💾 Converting with 78.8786 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-172.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-172.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,164B, BPFP=0.7118 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,376B, BPFP=3.2992 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,608B, BPFP=1.5571 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 54,200B, BPFP=3.1718 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,712B, BPFP=1.6802 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,360B, BPFP=3.1227 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,168B, BPFP=1.5314 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,084B, BPFP=3.2235 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,680B, BPFP=2.4977 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,376B, BPFP=3.1236 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,012B, BPFP=0.6940 +⌛️ [2/4] FRONTEND: Frontend time: 1.848s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12840328 29.03907896 + layer.0.v_cache 0.00001372 0.00190004 + layer.1.k_cache 0.60081110 3.05074353 + layer.1.v_cache 0.00000562 0.00081761 + layer.2.k_cache 0.00586274 0.87137058 + layer.2.v_cache 0.00001839 0.00261939 + layer.3.k_cache 0.02491910 3.80005157 + layer.3.v_cache 0.00001957 0.00286922 + layer.4.k_cache 0.00072047 0.07369833 + layer.4.v_cache 0.00005229 0.00585673 + layer.4.output 0.01050496 179.23970037 + ------------------------------------------------------------------------------------- + TOTAL 0.04908006 75.97217109 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 491740 +BPFP 1.6928 bits/point +EBPFP 3.3855 equivalent bits/point +MSE 75.972171 +---------------------- -------------------------------------------------------- +Time: 3.227s Load: 0.009s, Pack+Encode: 1.848s, Decode+Unpack: 1.370s +---------------------- -------------------------------------------------------- +💾 Converting with 75.9722 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-174.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-174.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,704B, BPFP=0.7014 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,924B, BPFP=3.1981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,292B, BPFP=1.5068 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,480B, BPFP=3.1184 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,344B, BPFP=1.6201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,500B, BPFP=3.0091 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,148B, BPFP=1.4989 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,620B, BPFP=3.1261 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,772B, BPFP=2.4167 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,196B, BPFP=3.0475 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,412B, BPFP=0.6895 +⌛️ [2/4] FRONTEND: Frontend time: 1.854s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14595435 30.08534742 + layer.0.v_cache 0.00001364 0.00184462 + layer.1.k_cache 0.59759904 3.01182012 + layer.1.v_cache 0.00000566 0.00079668 + layer.2.k_cache 0.01409306 0.68665451 + layer.2.v_cache 0.00002047 0.00260549 + layer.3.k_cache 0.06079736 3.80304624 + layer.3.v_cache 0.00002215 0.00283641 + layer.4.k_cache 0.00070776 0.07354239 + layer.4.v_cache 0.00005221 0.00585809 + layer.4.output 0.00769303 79.84884528 + ------------------------------------------------------------------------------------- + TOTAL 0.05135982 35.09507464 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 508392 +BPFP 1.6511 bits/point +EBPFP 3.3023 equivalent bits/point +MSE 35.095075 +---------------------- -------------------------------------------------------- +Time: 3.233s Load: 0.009s, Pack+Encode: 1.854s, Decode+Unpack: 1.370s +---------------------- -------------------------------------------------------- +💾 Converting with 35.0951 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-183.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-183.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 315, 128) +Output shape: (1, 315, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) -> torch.Size([1, 1, 315, 512]) + layer.4.output: torch.Size([1, 315, 3584]) -> torch.Size([1, 1, 315, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,580B, BPFP=0.6736 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,164B, BPFP=2.9347 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,592B, BPFP=1.4679 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,588B, BPFP=2.8565 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,264B, BPFP=1.6004 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,372B, BPFP=2.7962 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,788B, BPFP=1.4776 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,672B, BPFP=2.8607 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,668B, BPFP=2.3149 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,464B, BPFP=2.7512 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,540B, BPFP=0.6983 +⌛️ [2/4] FRONTEND: Frontend time: 1.858s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 315, 128]) + layer.0.v_cache: torch.Size([1, 4, 315, 128]) + layer.1.k_cache: torch.Size([1, 4, 315, 128]) + layer.1.v_cache: torch.Size([1, 4, 315, 128]) + layer.2.k_cache: torch.Size([1, 4, 315, 128]) + layer.2.v_cache: torch.Size([1, 4, 315, 128]) + layer.3.k_cache: torch.Size([1, 4, 315, 128]) + layer.3.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.k_cache: torch.Size([1, 4, 315, 128]) + layer.4.v_cache: torch.Size([1, 4, 315, 128]) + layer.4.output: torch.Size([1, 315, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15500355 32.91135293 + layer.0.v_cache 0.00001366 0.00193990 + layer.1.k_cache 0.68167124 3.20127631 + layer.1.v_cache 0.00000571 0.00084355 + layer.2.k_cache 0.01007247 0.61350970 + layer.2.v_cache 0.00001936 0.00265331 + layer.3.k_cache 0.03272899 4.58376659 + layer.3.v_cache 0.00002006 0.00277309 + layer.4.k_cache 0.00071175 0.07614252 + layer.4.v_cache 0.00005427 0.00562592 + layer.4.output 0.04584261 147.92439059 + ------------------------------------------------------------------------------------- + TOTAL 0.07065879 63.34533047 + (elements=2,741,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2741760 +Total Bytes 536692 +BPFP 1.5660 bits/point +EBPFP 3.1320 equivalent bits/point +MSE 63.345330 +---------------------- -------------------------------------------------------- +Time: 3.246s Load: 0.011s, Pack+Encode: 1.858s, Decode+Unpack: 1.377s +---------------------- -------------------------------------------------------- +💾 Converting with 63.3453 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-186.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,436B, BPFP=0.7170 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,720B, BPFP=3.3280 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,152B, BPFP=1.5655 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,216B, BPFP=3.1836 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,640B, BPFP=1.6513 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,324B, BPFP=3.0745 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,852B, BPFP=1.4905 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,036B, BPFP=3.2309 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,880B, BPFP=2.4723 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,068B, BPFP=3.1174 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 82,028B, BPFP=0.6756 +⌛️ [2/4] FRONTEND: Frontend time: 1.851s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12889319 29.95625288 + layer.0.v_cache 0.00001350 0.00193269 + layer.1.k_cache 0.58775566 2.78456797 + layer.1.v_cache 0.00000553 0.00081835 + layer.2.k_cache 0.00723753 0.71623799 + layer.2.v_cache 0.00001924 0.00270843 + layer.3.k_cache 0.01761036 3.62530855 + layer.3.v_cache 0.00001847 0.00290612 + layer.4.k_cache 0.00071814 0.07095025 + layer.4.v_cache 0.00005270 0.00582542 + layer.4.output 0.01035140 176.36457235 + ------------------------------------------------------------------------------------- + TOTAL 0.04792848 74.80703030 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 495352 +BPFP 1.6800 bits/point +EBPFP 3.3600 equivalent bits/point +MSE 74.807030 +---------------------- -------------------------------------------------------- +Time: 3.225s Load: 0.010s, Pack+Encode: 1.851s, Decode+Unpack: 1.364s +---------------------- -------------------------------------------------------- +💾 Converting with 74.8070 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-188.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-188.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,756B, BPFP=0.7018 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,340B, BPFP=3.1547 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,260B, BPFP=1.6098 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,704B, BPFP=3.0647 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,636B, BPFP=1.6305 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,172B, BPFP=2.9804 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,068B, BPFP=1.4892 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,860B, BPFP=3.0733 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,060B, BPFP=2.4241 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,568B, BPFP=3.0022 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,044B, BPFP=0.6684 +⌛️ [2/4] FRONTEND: Frontend time: 1.876s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13033964 29.78123281 + layer.0.v_cache 0.00001368 0.00187482 + layer.1.k_cache 0.63029007 2.87663226 + layer.1.v_cache 0.00000551 0.00080492 + layer.2.k_cache 0.00637632 0.75464415 + layer.2.v_cache 0.00001960 0.00261048 + layer.3.k_cache 0.02368569 3.89526496 + layer.3.v_cache 0.00001974 0.00285426 + layer.4.k_cache 0.00068050 0.07499728 + layer.4.v_cache 0.00005209 0.00584497 + layer.4.output 0.00765416 177.62334947 + ------------------------------------------------------------------------------------- + TOTAL 0.04970952 75.33883572 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 505468 +BPFP 1.6359 bits/point +EBPFP 3.2717 equivalent bits/point +MSE 75.338836 +---------------------- -------------------------------------------------------- +Time: 3.254s Load: 0.009s, Pack+Encode: 1.876s, Decode+Unpack: 1.369s +---------------------- -------------------------------------------------------- +💾 Converting with 75.3388 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-194.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,476B, BPFP=0.7167 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,676B, BPFP=3.2557 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,384B, BPFP=1.6880 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 54,668B, BPFP=3.1404 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,208B, BPFP=1.6778 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,796B, BPFP=3.0903 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,648B, BPFP=1.5308 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,896B, BPFP=3.2109 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,124B, BPFP=2.4773 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,324B, BPFP=3.1206 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,148B, BPFP=0.7070 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15644295 29.13899859 + layer.0.v_cache 0.00001405 0.00192290 + layer.1.k_cache 0.56163917 3.08319720 + layer.1.v_cache 0.00000554 0.00082993 + layer.2.k_cache 0.01059573 0.78626678 + layer.2.v_cache 0.00001807 0.00259260 + layer.3.k_cache 0.04944291 3.80897971 + layer.3.v_cache 0.00001950 0.00283815 + layer.4.k_cache 0.00071267 0.07527784 + layer.4.v_cache 0.00005817 0.00594331 + layer.4.output 0.00956290 184.64454766 + ------------------------------------------------------------------------------------- + TOTAL 0.04975818 78.20109886 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 502348 +BPFP 1.6975 bits/point +EBPFP 3.3950 equivalent bits/point +MSE 78.201099 +---------------------- -------------------------------------------------------- +Time: 3.251s Load: 0.010s, Pack+Encode: 1.855s, Decode+Unpack: 1.387s +---------------------- -------------------------------------------------------- +💾 Converting with 78.2011 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-20.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-20.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 312, 128) +Output shape: (1, 312, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) -> torch.Size([1, 1, 312, 512]) + layer.4.output: torch.Size([1, 312, 3584]) -> torch.Size([1, 1, 312, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,532B, BPFP=0.6777 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,204B, BPFP=2.9649 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,404B, BPFP=1.5226 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,580B, BPFP=2.8836 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,056B, BPFP=1.6054 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,440B, BPFP=2.8265 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,760B, BPFP=1.4904 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,596B, BPFP=2.8844 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,180B, BPFP=2.3127 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,824B, BPFP=2.7957 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,840B, BPFP=0.7357 +⌛️ [2/4] FRONTEND: Frontend time: 1.864s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 312, 128]) + layer.0.v_cache: torch.Size([1, 4, 312, 128]) + layer.1.k_cache: torch.Size([1, 4, 312, 128]) + layer.1.v_cache: torch.Size([1, 4, 312, 128]) + layer.2.k_cache: torch.Size([1, 4, 312, 128]) + layer.2.v_cache: torch.Size([1, 4, 312, 128]) + layer.3.k_cache: torch.Size([1, 4, 312, 128]) + layer.3.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.k_cache: torch.Size([1, 4, 312, 128]) + layer.4.v_cache: torch.Size([1, 4, 312, 128]) + layer.4.output: torch.Size([1, 312, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14984696 30.43659856 + layer.0.v_cache 0.00001474 0.00192813 + layer.1.k_cache 0.67859346 3.20518376 + layer.1.v_cache 0.00000581 0.00084835 + layer.2.k_cache 0.01626646 0.57122460 + layer.2.v_cache 0.00001951 0.00265389 + layer.3.k_cache 0.02914289 4.43049739 + layer.3.v_cache 0.00001919 0.00279362 + layer.4.k_cache 0.00071665 0.07679774 + layer.4.v_cache 0.00005096 0.00559574 + layer.4.output 0.04642455 60.30805718 + ------------------------------------------------------------------------------------- + TOTAL 0.07056756 27.11120718 + (elements=2,715,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2715648 +Total Bytes 541416 +BPFP 1.5950 bits/point +EBPFP 3.1899 equivalent bits/point +MSE 27.111207 +---------------------- -------------------------------------------------------- +Time: 3.260s Load: 0.012s, Pack+Encode: 1.864s, Decode+Unpack: 1.384s +---------------------- -------------------------------------------------------- +💾 Converting with 27.1112 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-204.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,288B, BPFP=0.7138 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,568B, BPFP=3.3439 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,984B, BPFP=1.5674 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,732B, BPFP=3.2372 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,168B, BPFP=1.6942 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,328B, BPFP=3.1557 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,528B, BPFP=1.4828 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,680B, BPFP=3.2923 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,464B, BPFP=2.4665 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,280B, BPFP=3.1529 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,424B, BPFP=0.7088 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.391s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15823836 29.70366048 + layer.0.v_cache 0.00001502 0.00187308 + layer.1.k_cache 0.53668020 3.05656189 + layer.1.v_cache 0.00000576 0.00081528 + layer.2.k_cache 0.01018995 0.71130564 + layer.2.v_cache 0.00002218 0.00267378 + layer.3.k_cache 0.05410197 3.91620803 + layer.3.v_cache 0.00002001 0.00286391 + layer.4.k_cache 0.00069143 0.07482138 + layer.4.v_cache 0.00005304 0.00579249 + layer.4.output 0.01041225 183.57368561 + ------------------------------------------------------------------------------------- + TOTAL 0.04899433 77.79366913 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 500444 +BPFP 1.7099 bits/point +EBPFP 3.4198 equivalent bits/point +MSE 77.793669 +---------------------- -------------------------------------------------------- +Time: 3.259s Load: 0.011s, Pack+Encode: 1.857s, Decode+Unpack: 1.391s +---------------------- -------------------------------------------------------- +💾 Converting with 77.7937 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-207.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 329, 128) +Output shape: (1, 329, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) -> torch.Size([1, 1, 329, 512]) + layer.4.output: torch.Size([1, 329, 3584]) -> torch.Size([1, 1, 329, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,640B, BPFP=0.6953 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 68,664B, BPFP=3.2610 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,928B, BPFP=1.5163 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 66,736B, BPFP=3.1695 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 34,756B, BPFP=1.6506 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 64,980B, BPFP=3.0861 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,300B, BPFP=1.4865 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 67,472B, BPFP=3.2044 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 51,292B, BPFP=2.4360 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 65,392B, BPFP=3.1056 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 121,996B, BPFP=0.8277 +⌛️ [2/4] FRONTEND: Frontend time: 2.205s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.508s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 329, 128]) + layer.0.v_cache: torch.Size([1, 4, 329, 128]) + layer.1.k_cache: torch.Size([1, 4, 329, 128]) + layer.1.v_cache: torch.Size([1, 4, 329, 128]) + layer.2.k_cache: torch.Size([1, 4, 329, 128]) + layer.2.v_cache: torch.Size([1, 4, 329, 128]) + layer.3.k_cache: torch.Size([1, 4, 329, 128]) + layer.3.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.k_cache: torch.Size([1, 4, 329, 128]) + layer.4.v_cache: torch.Size([1, 4, 329, 128]) + layer.4.output: torch.Size([1, 329, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11391156 30.51215805 + layer.0.v_cache 0.00001522 0.00188808 + layer.1.k_cache 0.63933222 3.00662296 + layer.1.v_cache 0.00000652 0.00088201 + layer.2.k_cache 0.01643764 0.62180205 + layer.2.v_cache 0.00001873 0.00259217 + layer.3.k_cache 0.02826017 3.52871658 + layer.3.v_cache 0.00001876 0.00274188 + layer.4.k_cache 0.00074607 0.07557397 + layer.4.v_cache 0.00005483 0.00598628 + layer.4.output 3.44632065 141.62954570 + ------------------------------------------------------------------------------------- + TOTAL 1.46606155 60.53916376 + (elements=2,863,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2863616 +Total Bytes 619156 +BPFP 1.7297 bits/point +EBPFP 3.4594 equivalent bits/point +MSE 60.539164 +---------------------- -------------------------------------------------------- +Time: 3.724s Load: 0.011s, Pack+Encode: 2.205s, Decode+Unpack: 1.508s +---------------------- -------------------------------------------------------- +💾 Converting with 60.5392 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-211.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,596B, BPFP=0.7054 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,932B, BPFP=3.1884 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,808B, BPFP=1.6134 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,148B, BPFP=3.0885 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,184B, BPFP=1.6904 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,856B, BPFP=3.0721 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,036B, BPFP=1.5141 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,960B, BPFP=3.1340 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,888B, BPFP=2.4579 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,956B, BPFP=3.0777 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,936B, BPFP=0.6715 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13536828 29.43013553 + layer.0.v_cache 0.00001433 0.00190528 + layer.1.k_cache 0.57185238 2.88819803 + layer.1.v_cache 0.00000558 0.00081346 + layer.2.k_cache 0.00766427 0.67164344 + layer.2.v_cache 0.00001891 0.00263660 + layer.3.k_cache 0.04143807 4.00478656 + layer.3.v_cache 0.00001947 0.00284411 + layer.4.k_cache 0.00070404 0.07373735 + layer.4.v_cache 0.00005498 0.00581937 + layer.4.output 0.00948056 171.48607911 + ------------------------------------------------------------------------------------- + TOTAL 0.04844142 72.79323962 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 504300 +BPFP 1.6613 bits/point +EBPFP 3.3227 equivalent bits/point +MSE 72.793240 +---------------------- -------------------------------------------------------- +Time: 3.237s Load: 0.011s, Pack+Encode: 1.855s, Decode+Unpack: 1.371s +---------------------- -------------------------------------------------------- +💾 Converting with 72.7932 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-215.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,532B, BPFP=0.7095 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,472B, BPFP=3.3102 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,460B, BPFP=1.6678 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,416B, BPFP=3.1938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,676B, BPFP=1.6800 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,192B, BPFP=3.1245 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,740B, BPFP=1.5138 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,484B, BPFP=3.1977 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,340B, BPFP=2.5102 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,712B, BPFP=3.0974 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,912B, BPFP=0.7029 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15132187 30.43742923 + layer.0.v_cache 0.00001456 0.00189111 + layer.1.k_cache 0.61798344 3.24089979 + layer.1.v_cache 0.00000569 0.00080526 + layer.2.k_cache 0.01686249 0.68813180 + layer.2.v_cache 0.00001898 0.00268010 + layer.3.k_cache 0.04401700 3.72975269 + layer.3.v_cache 0.00001935 0.00280169 + layer.4.k_cache 0.00071189 0.07383510 + layer.4.v_cache 0.00005111 0.00567994 + layer.4.output 0.01058244 177.73235313 + ------------------------------------------------------------------------------------- + TOTAL 0.05324020 75.43002227 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 510936 +BPFP 1.7015 bits/point +EBPFP 3.4030 equivalent bits/point +MSE 75.430022 +---------------------- -------------------------------------------------------- +Time: 3.240s Load: 0.010s, Pack+Encode: 1.855s, Decode+Unpack: 1.375s +---------------------- -------------------------------------------------------- +💾 Converting with 75.4300 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-216.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,540B, BPFP=0.7099 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,056B, BPFP=3.2301 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,588B, BPFP=1.6750 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,264B, BPFP=3.1286 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,700B, BPFP=1.6814 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,612B, BPFP=3.0917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,816B, BPFP=1.5181 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,728B, BPFP=3.1549 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,296B, BPFP=2.4511 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,940B, BPFP=3.1103 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,684B, BPFP=0.7091 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12581717 28.65948752 + layer.0.v_cache 0.00001375 0.00196202 + layer.1.k_cache 0.57814745 3.18082771 + layer.1.v_cache 0.00000592 0.00083731 + layer.2.k_cache 0.00733882 0.69157039 + layer.2.v_cache 0.00002058 0.00278238 + layer.3.k_cache 0.02353939 3.73766426 + layer.3.v_cache 0.00002077 0.00303039 + layer.4.k_cache 0.00069182 0.07523122 + layer.4.v_cache 0.00005027 0.00600508 + layer.4.output 0.01015823 179.84015916 + ------------------------------------------------------------------------------------- + TOTAL 0.04745609 76.19061838 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 507224 +BPFP 1.6891 bits/point +EBPFP 3.3783 equivalent bits/point +MSE 76.190618 +---------------------- -------------------------------------------------------- +Time: 3.235s Load: 0.011s, Pack+Encode: 1.856s, Decode+Unpack: 1.368s +---------------------- -------------------------------------------------------- +💾 Converting with 76.1906 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-22.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-22.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,396B, BPFP=0.7069 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,916B, BPFP=3.2457 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,116B, BPFP=1.6604 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,040B, BPFP=3.1387 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,112B, BPFP=1.7172 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,764B, BPFP=3.1229 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,920B, BPFP=1.5351 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,700B, BPFP=3.2333 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,896B, BPFP=2.4462 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,428B, BPFP=3.1038 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,332B, BPFP=0.7196 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 32.118s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13625057 28.36002367 + layer.0.v_cache 0.00001362 0.00186013 + layer.1.k_cache 0.57512581 3.16020258 + layer.1.v_cache 0.00000584 0.00081392 + layer.2.k_cache 0.00872732 0.71933574 + layer.2.v_cache 0.00001851 0.00269029 + layer.3.k_cache 0.04362779 3.38319141 + layer.3.v_cache 0.00001925 0.00293574 + layer.4.k_cache 0.00073619 0.07433461 + layer.4.v_cache 0.00005148 0.00589267 + layer.4.output 0.00783937 179.67619265 + ------------------------------------------------------------------------------------- + TOTAL 0.04820306 76.08497819 + (elements=2,384,896) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2384896 +Total Bytes 507620 +BPFP 1.7028 bits/point +EBPFP 3.4056 equivalent bits/point +MSE 76.084978 +---------------------- --------------------------------------------------------- +Time: 33.986s Load: 0.010s, Pack+Encode: 1.857s, Decode+Unpack: 32.118s +---------------------- --------------------------------------------------------- +💾 Converting with 76.0850 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-221.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-221.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,572B, BPFP=0.6976 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,832B, BPFP=3.0238 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,688B, BPFP=1.4745 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,172B, BPFP=2.9385 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,592B, BPFP=1.6238 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,880B, BPFP=2.8721 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,824B, BPFP=1.4815 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,588B, BPFP=2.9599 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,356B, BPFP=2.3312 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,508B, BPFP=2.8530 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,108B, BPFP=0.7497 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16730670 29.37109375 + layer.0.v_cache 0.00001388 0.00187752 + layer.1.k_cache 0.64158324 3.07025367 + layer.1.v_cache 0.00000571 0.00081607 + layer.2.k_cache 0.02207795 0.61540563 + layer.2.v_cache 0.00001922 0.00252196 + layer.3.k_cache 0.04503629 4.45051655 + layer.3.v_cache 0.00001955 0.00275066 + layer.4.k_cache 0.00070478 0.07658674 + layer.4.v_cache 0.00005260 0.00569597 + layer.4.output 0.04763387 153.72061501 + ------------------------------------------------------------------------------------- + TOTAL 0.07119159 65.50834257 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 535120 +BPFP 1.6179 bits/point +EBPFP 3.2358 equivalent bits/point +MSE 65.508343 +---------------------- -------------------------------------------------------- +Time: 3.243s Load: 0.010s, Pack+Encode: 1.857s, Decode+Unpack: 1.375s +---------------------- -------------------------------------------------------- +💾 Converting with 65.5083 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-222.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,580B, BPFP=0.7096 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,816B, BPFP=3.2613 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,236B, BPFP=1.7056 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,088B, BPFP=3.2202 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,716B, BPFP=1.6762 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,404B, BPFP=3.0688 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,716B, BPFP=1.5070 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,872B, BPFP=3.2080 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,956B, BPFP=2.4231 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,000B, BPFP=3.1024 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,448B, BPFP=0.7047 +⌛️ [2/4] FRONTEND: Frontend time: 1.852s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15225300 29.26373533 + layer.0.v_cache 0.00001427 0.00184389 + layer.1.k_cache 0.52731114 3.03889636 + layer.1.v_cache 0.00000556 0.00079913 + layer.2.k_cache 0.01483519 0.71129594 + layer.2.v_cache 0.00001864 0.00269316 + layer.3.k_cache 0.04699480 4.01774735 + layer.3.v_cache 0.00001968 0.00277655 + layer.4.k_cache 0.00070142 0.07511005 + layer.4.v_cache 0.00005177 0.00573331 + layer.4.output 0.01077793 183.46228726 + ------------------------------------------------------------------------------------- + TOTAL 0.04809712 77.72686129 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 510832 +BPFP 1.6950 bits/point +EBPFP 3.3900 equivalent bits/point +MSE 77.726861 +---------------------- -------------------------------------------------------- +Time: 3.237s Load: 0.011s, Pack+Encode: 1.852s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 77.7269 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-223.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,544B, BPFP=0.7076 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,076B, BPFP=3.2195 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,396B, BPFP=1.6018 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,484B, BPFP=3.1297 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,676B, BPFP=1.6740 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,064B, BPFP=3.0496 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,804B, BPFP=1.5120 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,632B, BPFP=3.1381 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,004B, BPFP=2.4258 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,768B, BPFP=3.0894 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,636B, BPFP=0.7223 +⌛️ [2/4] FRONTEND: Frontend time: 1.858s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15700974 29.24322047 + layer.0.v_cache 0.00001374 0.00180798 + layer.1.k_cache 0.52255304 3.19061345 + layer.1.v_cache 0.00000566 0.00078101 + layer.2.k_cache 0.01360668 0.82463922 + layer.2.v_cache 0.00001926 0.00272804 + layer.3.k_cache 0.08085050 3.81344836 + layer.3.v_cache 0.00001943 0.00282172 + layer.4.k_cache 0.00071029 0.07517834 + layer.4.v_cache 0.00005450 0.00580676 + layer.4.output 0.01014708 180.74110366 + ------------------------------------------------------------------------------------- + TOTAL 0.04975720 76.60875123 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 507084 +BPFP 1.6826 bits/point +EBPFP 3.3651 equivalent bits/point +MSE 76.608751 +---------------------- -------------------------------------------------------- +Time: 3.242s Load: 0.010s, Pack+Encode: 1.858s, Decode+Unpack: 1.373s +---------------------- -------------------------------------------------------- +💾 Converting with 76.6088 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-23.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-23.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,736B, BPFP=0.6982 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,196B, BPFP=3.1357 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,736B, BPFP=1.6303 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,728B, BPFP=3.0553 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,172B, BPFP=1.6542 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,512B, BPFP=2.9886 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,012B, BPFP=1.4809 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,996B, BPFP=3.0700 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,048B, BPFP=2.4149 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,268B, BPFP=3.0300 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,812B, BPFP=0.6878 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16436556 30.39015214 + layer.0.v_cache 0.00001441 0.00192667 + layer.1.k_cache 0.62741763 3.25427075 + layer.1.v_cache 0.00000560 0.00082143 + layer.2.k_cache 0.01487184 0.67308574 + layer.2.v_cache 0.00002167 0.00275061 + layer.3.k_cache 0.02142572 4.06271973 + layer.3.v_cache 0.00001943 0.00280773 + layer.4.k_cache 0.00070776 0.07519230 + layer.4.v_cache 0.00005251 0.00574951 + layer.4.output 0.00894435 168.66231203 + ------------------------------------------------------------------------------------- + TOTAL 0.05244192 71.71209770 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 510216 +BPFP 1.6454 bits/point +EBPFP 3.2909 equivalent bits/point +MSE 71.712098 +---------------------- -------------------------------------------------------- +Time: 3.240s Load: 0.012s, Pack+Encode: 1.857s, Decode+Unpack: 1.372s +---------------------- -------------------------------------------------------- +💾 Converting with 71.7121 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-232.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-232.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,516B, BPFP=0.6960 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,976B, BPFP=3.2238 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,220B, BPFP=1.5136 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,784B, BPFP=3.1019 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,700B, BPFP=1.6515 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,836B, BPFP=3.0492 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,424B, BPFP=1.5249 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,112B, BPFP=3.1201 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,116B, BPFP=2.3975 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,408B, BPFP=3.0254 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,492B, BPFP=0.6712 +⌛️ [2/4] FRONTEND: Frontend time: 1.861s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11705062 30.34362836 + layer.0.v_cache 0.00001445 0.00189241 + layer.1.k_cache 0.56620099 2.71226246 + layer.1.v_cache 0.00000566 0.00080724 + layer.2.k_cache 0.01299965 0.67827077 + layer.2.v_cache 0.00001972 0.00266867 + layer.3.k_cache 0.05069733 3.87978594 + layer.3.v_cache 0.00001944 0.00284545 + layer.4.k_cache 0.00068883 0.07303482 + layer.4.v_cache 0.00005144 0.00561489 + layer.4.output 0.00838925 175.84597420 + ------------------------------------------------------------------------------------- + TOTAL 0.04743958 74.62486061 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 503584 +BPFP 1.6472 bits/point +EBPFP 3.2943 equivalent bits/point +MSE 74.624861 +---------------------- -------------------------------------------------------- +Time: 3.242s Load: 0.010s, Pack+Encode: 1.861s, Decode+Unpack: 1.371s +---------------------- -------------------------------------------------------- +💾 Converting with 74.6249 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-236.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,708B, BPFP=0.7016 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,392B, BPFP=3.1687 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,348B, BPFP=1.6204 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,676B, BPFP=3.0740 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,784B, BPFP=1.6444 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,496B, BPFP=3.0088 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,296B, BPFP=1.5071 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,152B, BPFP=3.1003 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,752B, BPFP=2.4156 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,532B, BPFP=3.0108 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,584B, BPFP=0.6987 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14370025 30.42138499 + layer.0.v_cache 0.00001382 0.00190339 + layer.1.k_cache 0.61448195 3.15515309 + layer.1.v_cache 0.00000570 0.00083001 + layer.2.k_cache 0.01208573 0.70970790 + layer.2.v_cache 0.00001823 0.00261814 + layer.3.k_cache 0.01319265 3.74061094 + layer.3.v_cache 0.00002242 0.00290755 + layer.4.k_cache 0.00071920 0.07454447 + layer.4.v_cache 0.00005145 0.00602623 + layer.4.output 0.00992508 137.94130174 + ------------------------------------------------------------------------------------- + TOTAL 0.05022159 59.04145876 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 509720 +BPFP 1.6555 bits/point +EBPFP 3.3109 equivalent bits/point +MSE 59.041459 +---------------------- -------------------------------------------------------- +Time: 3.237s Load: 0.009s, Pack+Encode: 1.857s, Decode+Unpack: 1.371s +---------------------- -------------------------------------------------------- +💾 Converting with 59.0415 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-24.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 259, 128) +Output shape: (1, 259, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) -> torch.Size([1, 1, 259, 512]) + layer.4.output: torch.Size([1, 259, 3584]) -> torch.Size([1, 1, 259, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,676B, BPFP=0.7044 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,344B, BPFP=3.3388 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,020B, BPFP=1.6301 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 53,248B, BPFP=3.2124 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,284B, BPFP=1.6460 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 51,348B, BPFP=3.0977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,700B, BPFP=1.4901 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 53,560B, BPFP=3.2312 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,084B, BPFP=2.4785 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,408B, BPFP=3.1617 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,140B, BPFP=0.7338 +⌛️ [2/4] FRONTEND: Frontend time: 1.858s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.369s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 259, 128]) + layer.0.v_cache: torch.Size([1, 4, 259, 128]) + layer.1.k_cache: torch.Size([1, 4, 259, 128]) + layer.1.v_cache: torch.Size([1, 4, 259, 128]) + layer.2.k_cache: torch.Size([1, 4, 259, 128]) + layer.2.v_cache: torch.Size([1, 4, 259, 128]) + layer.3.k_cache: torch.Size([1, 4, 259, 128]) + layer.3.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.k_cache: torch.Size([1, 4, 259, 128]) + layer.4.v_cache: torch.Size([1, 4, 259, 128]) + layer.4.output: torch.Size([1, 259, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13349269 29.14696323 + layer.0.v_cache 0.00001436 0.00199290 + layer.1.k_cache 0.50562990 3.08939954 + layer.1.v_cache 0.00000590 0.00086700 + layer.2.k_cache 0.00820730 0.82604197 + layer.2.v_cache 0.00001794 0.00265176 + layer.3.k_cache 0.03685973 3.34469946 + layer.3.v_cache 0.00002061 0.00289783 + layer.4.k_cache 0.00069539 0.07360915 + layer.4.v_cache 0.00005422 0.00591234 + layer.4.output 0.01018517 147.84247449 + ------------------------------------------------------------------------------------- + TOTAL 0.04448790 63.02307980 + (elements=2,254,336) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2254336 +Total Bytes 482812 +BPFP 1.7134 bits/point +EBPFP 3.4267 equivalent bits/point +MSE 63.023080 +---------------------- -------------------------------------------------------- +Time: 3.238s Load: 0.011s, Pack+Encode: 1.858s, Decode+Unpack: 1.369s +---------------------- -------------------------------------------------------- +💾 Converting with 63.0231 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-245.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-245.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,344B, BPFP=0.7144 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,716B, BPFP=3.3400 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,956B, BPFP=1.6178 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,116B, BPFP=3.2475 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,796B, BPFP=1.6664 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,512B, BPFP=3.1546 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,080B, BPFP=1.5093 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,244B, BPFP=3.3127 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,504B, BPFP=2.6333 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,156B, BPFP=3.1919 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,152B, BPFP=0.7205 +⌛️ [2/4] FRONTEND: Frontend time: 1.860s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12649573 29.35730975 + layer.0.v_cache 0.00001373 0.00188143 + layer.1.k_cache 0.51574391 3.02629553 + layer.1.v_cache 0.00000589 0.00083737 + layer.2.k_cache 0.01194659 0.73160768 + layer.2.v_cache 0.00001984 0.00261095 + layer.3.k_cache 0.04962628 3.76314064 + layer.3.v_cache 0.00002059 0.00282518 + layer.4.k_cache 0.00068541 0.07552107 + layer.4.v_cache 0.00005704 0.00599216 + layer.4.output 0.00772497 183.50653108 + ------------------------------------------------------------------------------------- + TOTAL 0.04462881 77.73610231 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 508576 +BPFP 1.7313 bits/point +EBPFP 3.4625 equivalent bits/point +MSE 77.736102 +---------------------- -------------------------------------------------------- +Time: 3.249s Load: 0.012s, Pack+Encode: 1.860s, Decode+Unpack: 1.377s +---------------------- -------------------------------------------------------- +💾 Converting with 77.7361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-246.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,236B, BPFP=0.7107 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,292B, BPFP=3.3278 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,244B, BPFP=1.5825 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,248B, BPFP=3.2091 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,752B, BPFP=1.6701 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,168B, BPFP=3.1464 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,928B, BPFP=1.5060 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,472B, BPFP=3.2802 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,768B, BPFP=2.4842 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,240B, BPFP=3.1506 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,128B, BPFP=0.6898 +⌛️ [2/4] FRONTEND: Frontend time: 1.879s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12210708 30.13789680 + layer.0.v_cache 0.00001342 0.00184925 + layer.1.k_cache 0.54942129 3.18366273 + layer.1.v_cache 0.00000547 0.00081103 + layer.2.k_cache 0.01053270 0.69256728 + layer.2.v_cache 0.00001820 0.00265107 + layer.3.k_cache 0.02850239 3.50564087 + layer.3.v_cache 0.00002491 0.00285574 + layer.4.k_cache 0.00067987 0.07355414 + layer.4.v_cache 0.00005179 0.00575394 + layer.4.output 0.01126663 159.01156731 + ------------------------------------------------------------------------------------- + TOTAL 0.04648373 67.68754200 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 497476 +BPFP 1.6998 bits/point +EBPFP 3.3995 equivalent bits/point +MSE 67.687542 +---------------------- -------------------------------------------------------- +Time: 3.264s Load: 0.008s, Pack+Encode: 1.879s, Decode+Unpack: 1.376s +---------------------- -------------------------------------------------------- +💾 Converting with 67.6875 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-249.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-249.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,496B, BPFP=0.6960 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,844B, BPFP=3.0344 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,036B, BPFP=1.4973 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,224B, BPFP=2.9509 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,520B, BPFP=1.6254 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,072B, BPFP=2.8915 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,992B, BPFP=1.4950 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,356B, BPFP=2.9577 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,676B, BPFP=2.3554 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,012B, BPFP=2.8884 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,712B, BPFP=0.6977 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14443027 31.41937075 + layer.0.v_cache 0.00001375 0.00191784 + layer.1.k_cache 0.62837083 3.00369273 + layer.1.v_cache 0.00000570 0.00081359 + layer.2.k_cache 0.01557094 0.60952482 + layer.2.v_cache 0.00001871 0.00256191 + layer.3.k_cache 0.04676386 4.40522187 + layer.3.v_cache 0.00001981 0.00272619 + layer.4.k_cache 0.00070582 0.07644040 + layer.4.v_cache 0.00005437 0.00569576 + layer.4.output 0.04838977 72.80841879 + ------------------------------------------------------------------------------------- + TOTAL 0.06909897 32.30511161 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 528940 +BPFP 1.6045 bits/point +EBPFP 3.2090 equivalent bits/point +MSE 32.305112 +---------------------- -------------------------------------------------------- +Time: 3.243s Load: 0.011s, Pack+Encode: 1.857s, Decode+Unpack: 1.375s +---------------------- -------------------------------------------------------- +💾 Converting with 32.3051 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-250.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,720B, BPFP=0.7023 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,400B, BPFP=3.1692 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,932B, BPFP=1.5974 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,788B, BPFP=3.0802 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,832B, BPFP=1.6471 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,560B, BPFP=3.0124 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,904B, BPFP=1.4854 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,196B, BPFP=3.1027 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,240B, BPFP=2.4426 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,796B, BPFP=3.0254 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,252B, BPFP=0.6882 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.375s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13293209 30.42218212 + layer.0.v_cache 0.00001388 0.00190479 + layer.1.k_cache 0.56267701 2.84430536 + layer.1.v_cache 0.00000566 0.00079661 + layer.2.k_cache 0.01145046 0.67850896 + layer.2.v_cache 0.00001921 0.00260859 + layer.3.k_cache 0.02674564 3.97213519 + layer.3.v_cache 0.00001926 0.00280052 + layer.4.k_cache 0.00069135 0.07501194 + layer.4.v_cache 0.00005493 0.00589510 + layer.4.output 0.00973303 166.01383455 + ------------------------------------------------------------------------------------- + TOTAL 0.04722004 70.59429359 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 508620 +BPFP 1.6519 bits/point +EBPFP 3.3038 equivalent bits/point +MSE 70.594294 +---------------------- -------------------------------------------------------- +Time: 3.241s Load: 0.011s, Pack+Encode: 1.855s, Decode+Unpack: 1.375s +---------------------- -------------------------------------------------------- +💾 Converting with 70.5943 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-252.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-252.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,588B, BPFP=0.7101 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,908B, BPFP=3.2101 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,708B, BPFP=1.5630 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,572B, BPFP=3.1347 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,516B, BPFP=1.6649 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,664B, BPFP=3.0835 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,084B, BPFP=1.5278 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,024B, BPFP=3.1602 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,748B, BPFP=2.4677 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,204B, BPFP=3.0575 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,712B, BPFP=0.6907 +⌛️ [2/4] FRONTEND: Frontend time: 1.861s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15182159 29.81847219 + layer.0.v_cache 0.00001420 0.00189478 + layer.1.k_cache 0.56224567 3.11322925 + layer.1.v_cache 0.00000588 0.00083948 + layer.2.k_cache 0.00652889 0.71588939 + layer.2.v_cache 0.00001844 0.00261010 + layer.3.k_cache 0.06011557 3.90232436 + layer.3.v_cache 0.00001921 0.00284478 + layer.4.k_cache 0.00068317 0.07467687 + layer.4.v_cache 0.00005096 0.00564097 + layer.4.output 0.01130176 177.87554796 + ------------------------------------------------------------------------------------- + TOTAL 0.05062446 75.45689752 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 503728 +BPFP 1.6714 bits/point +EBPFP 3.3429 equivalent bits/point +MSE 75.456898 +---------------------- -------------------------------------------------------- +Time: 3.245s Load: 0.010s, Pack+Encode: 1.861s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 75.4569 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-253.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-253.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,764B, BPFP=0.6998 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,528B, BPFP=3.1539 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,624B, BPFP=1.5693 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,200B, BPFP=3.0811 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,356B, BPFP=1.6643 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,372B, BPFP=3.0357 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,456B, BPFP=1.5053 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,124B, BPFP=3.0770 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,776B, BPFP=2.4548 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,112B, BPFP=3.0215 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,756B, BPFP=0.7421 +⌛️ [2/4] FRONTEND: Frontend time: 1.852s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13053865 28.91434348 + layer.0.v_cache 0.00001414 0.00196324 + layer.1.k_cache 0.56363723 3.08830245 + layer.1.v_cache 0.00000590 0.00084462 + layer.2.k_cache 0.01033439 0.74227113 + layer.2.v_cache 0.00002015 0.00267316 + layer.3.k_cache 0.03655195 4.13579316 + layer.3.v_cache 0.00002026 0.00278268 + layer.4.k_cache 0.00068516 0.07472410 + layer.4.v_cache 0.00005378 0.00587379 + layer.4.output 0.00863618 178.53204887 + ------------------------------------------------------------------------------------- + TOTAL 0.04719499 75.68787729 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 519068 +BPFP 1.6740 bits/point +EBPFP 3.3480 equivalent bits/point +MSE 75.687877 +---------------------- -------------------------------------------------------- +Time: 3.239s Load: 0.009s, Pack+Encode: 1.852s, Decode+Unpack: 1.377s +---------------------- -------------------------------------------------------- +💾 Converting with 75.6879 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-254.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-254.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,512B, BPFP=0.7058 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,240B, BPFP=3.2288 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,380B, BPFP=1.5444 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,632B, BPFP=3.1381 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,528B, BPFP=1.6656 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,544B, BPFP=3.0767 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,884B, BPFP=1.5165 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,312B, BPFP=3.1764 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,388B, BPFP=2.4474 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,920B, BPFP=3.0979 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,308B, BPFP=0.7116 +⌛️ [2/4] FRONTEND: Frontend time: 1.854s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15908699 29.60639243 + layer.0.v_cache 0.00001433 0.00191522 + layer.1.k_cache 0.53822481 3.07856635 + layer.1.v_cache 0.00000570 0.00082831 + layer.2.k_cache 0.01139215 0.65473751 + layer.2.v_cache 0.00001845 0.00265348 + layer.3.k_cache 0.02907990 3.70808581 + layer.3.v_cache 0.00001963 0.00292065 + layer.4.k_cache 0.00069750 0.07484415 + layer.4.v_cache 0.00005397 0.00574134 + layer.4.output 0.01069233 179.61929474 + ------------------------------------------------------------------------------------- + TOTAL 0.04784940 76.14539697 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 506648 +BPFP 1.6811 bits/point +EBPFP 3.3622 equivalent bits/point +MSE 76.145397 +---------------------- -------------------------------------------------------- +Time: 3.240s Load: 0.009s, Pack+Encode: 1.854s, Decode+Unpack: 1.377s +---------------------- -------------------------------------------------------- +💾 Converting with 76.1454 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-255.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-255.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,764B, BPFP=0.6973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,636B, BPFP=3.1488 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,036B, BPFP=1.6410 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,432B, BPFP=3.0830 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,196B, BPFP=1.6497 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,088B, BPFP=3.0096 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,276B, BPFP=1.4902 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,700B, BPFP=3.0977 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,200B, BPFP=2.4148 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,528B, BPFP=3.0337 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,896B, BPFP=0.7250 +⌛️ [2/4] FRONTEND: Frontend time: 1.859s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16084659 29.95849951 + layer.0.v_cache 0.00001440 0.00187929 + layer.1.k_cache 0.59028764 2.92261382 + layer.1.v_cache 0.00000559 0.00080450 + layer.2.k_cache 0.00929020 0.60026945 + layer.2.v_cache 0.00002205 0.00259953 + layer.3.k_cache 0.04382113 3.96560690 + layer.3.v_cache 0.00002103 0.00275125 + layer.4.k_cache 0.00069271 0.07431074 + layer.4.v_cache 0.00005052 0.00568362 + layer.4.output 0.00984751 172.13494318 + ------------------------------------------------------------------------------------- + TOTAL 0.05141085 73.08703652 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 518752 +BPFP 1.6671 bits/point +EBPFP 3.3342 equivalent bits/point +MSE 73.087037 +---------------------- -------------------------------------------------------- +Time: 3.247s Load: 0.012s, Pack+Encode: 1.859s, Decode+Unpack: 1.376s +---------------------- -------------------------------------------------------- +💾 Converting with 73.0870 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-258.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,636B, BPFP=0.6855 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,936B, BPFP=3.1432 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,792B, BPFP=1.5078 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,568B, BPFP=3.0690 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,200B, BPFP=1.6385 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,912B, BPFP=2.9792 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,368B, BPFP=1.4848 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,892B, BPFP=3.0866 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,920B, BPFP=2.3828 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,692B, BPFP=3.0215 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 97,336B, BPFP=0.7544 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14211286 31.21293471 + layer.0.v_cache 0.00001516 0.00190615 + layer.1.k_cache 0.60291523 3.26947615 + layer.1.v_cache 0.00000570 0.00083345 + layer.2.k_cache 0.01041983 0.65533537 + layer.2.v_cache 0.00001835 0.00259136 + layer.3.k_cache 0.03896446 3.98000505 + layer.3.v_cache 0.00001926 0.00282777 + layer.4.k_cache 0.00071470 0.07520225 + layer.4.v_cache 0.00005236 0.00583362 + layer.4.output 0.00754456 162.37647259 + ------------------------------------------------------------------------------------- + TOTAL 0.04988528 69.16719141 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 521252 +BPFP 1.6635 bits/point +EBPFP 3.3270 equivalent bits/point +MSE 69.167191 +---------------------- -------------------------------------------------------- +Time: 3.239s Load: 0.010s, Pack+Encode: 1.856s, Decode+Unpack: 1.373s +---------------------- -------------------------------------------------------- +💾 Converting with 69.1672 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-260.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-260.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 296, 128) +Output shape: (1, 296, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) -> torch.Size([1, 1, 296, 512]) + layer.4.output: torch.Size([1, 296, 3584]) -> torch.Size([1, 1, 296, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,172B, BPFP=0.6953 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,028B, BPFP=3.1159 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,564B, BPFP=1.5606 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,196B, BPFP=3.0192 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,048B, BPFP=1.6389 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,036B, BPFP=2.9580 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,688B, BPFP=1.4616 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,516B, BPFP=3.0361 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,368B, BPFP=2.3948 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,904B, BPFP=2.9510 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,820B, BPFP=0.7075 +⌛️ [2/4] FRONTEND: Frontend time: 34.344s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 296, 128]) + layer.0.v_cache: torch.Size([1, 4, 296, 128]) + layer.1.k_cache: torch.Size([1, 4, 296, 128]) + layer.1.v_cache: torch.Size([1, 4, 296, 128]) + layer.2.k_cache: torch.Size([1, 4, 296, 128]) + layer.2.v_cache: torch.Size([1, 4, 296, 128]) + layer.3.k_cache: torch.Size([1, 4, 296, 128]) + layer.3.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.k_cache: torch.Size([1, 4, 296, 128]) + layer.4.v_cache: torch.Size([1, 4, 296, 128]) + layer.4.output: torch.Size([1, 296, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14616069 31.03361223 + layer.0.v_cache 0.00001366 0.00187094 + layer.1.k_cache 0.61022733 3.23603945 + layer.1.v_cache 0.00000576 0.00080209 + layer.2.k_cache 0.01378242 0.58028402 + layer.2.v_cache 0.00001907 0.00263844 + layer.3.k_cache 0.03184220 4.22303731 + layer.3.v_cache 0.00001934 0.00280640 + layer.4.k_cache 0.00072724 0.07527918 + layer.4.v_cache 0.00005080 0.00558131 + layer.4.output 0.04784788 150.58516832 + ------------------------------------------------------------------------------------- + TOTAL 0.06692845 64.30930174 + (elements=2,576,384) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2576384 +Total Bytes 526340 +BPFP 1.6344 bits/point +EBPFP 3.2687 equivalent bits/point +MSE 64.309302 +---------------------- --------------------------------------------------------- +Time: 35.739s Load: 0.012s, Pack+Encode: 34.344s, Decode+Unpack: 1.384s +---------------------- --------------------------------------------------------- +💾 Converting with 64.3093 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-264.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,324B, BPFP=0.6986 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,860B, BPFP=3.0862 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,496B, BPFP=1.4417 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,520B, BPFP=3.0159 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,540B, BPFP=1.6537 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,284B, BPFP=2.9511 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,604B, BPFP=1.4998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,728B, BPFP=3.0268 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,300B, BPFP=2.3752 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,044B, BPFP=2.9385 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,448B, BPFP=0.7524 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.374s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15387250 30.59458237 + layer.0.v_cache 0.00001382 0.00187883 + layer.1.k_cache 0.63844893 3.05626581 + layer.1.v_cache 0.00000583 0.00082927 + layer.2.k_cache 0.01105453 0.63286191 + layer.2.v_cache 0.00001885 0.00254682 + layer.3.k_cache 0.03893430 4.15438945 + layer.3.v_cache 0.00001923 0.00275058 + layer.4.k_cache 0.00070257 0.07566725 + layer.4.v_cache 0.00005146 0.00562175 + layer.4.output 0.04873301 159.92551534 + ------------------------------------------------------------------------------------- + TOTAL 0.06966195 68.11800008 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 533148 +BPFP 1.6444 bits/point +EBPFP 3.2888 equivalent bits/point +MSE 68.118000 +---------------------- -------------------------------------------------------- +Time: 3.240s Load: 0.010s, Pack+Encode: 1.855s, Decode+Unpack: 1.374s +---------------------- -------------------------------------------------------- +💾 Converting with 68.1180 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-265.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,448B, BPFP=0.7177 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,800B, BPFP=3.2749 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,180B, BPFP=1.5671 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,352B, BPFP=3.1914 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,104B, BPFP=1.6780 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,964B, BPFP=3.1114 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,176B, BPFP=1.5092 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,748B, BPFP=3.2143 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,680B, BPFP=2.5761 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,260B, BPFP=3.1285 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,244B, BPFP=0.7433 +⌛️ [2/4] FRONTEND: Frontend time: 1.854s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12629252 29.90364463 + layer.0.v_cache 0.00001451 0.00191234 + layer.1.k_cache 0.49097243 3.07520072 + layer.1.v_cache 0.00000577 0.00084560 + layer.2.k_cache 0.01060468 0.76314643 + layer.2.v_cache 0.00001909 0.00260621 + layer.3.k_cache 0.02748298 3.87374822 + layer.3.v_cache 0.00001876 0.00271029 + layer.4.k_cache 0.00069967 0.07488743 + layer.4.v_cache 0.00005414 0.00569805 + layer.4.output 0.01047103 180.94310095 + ------------------------------------------------------------------------------------- + TOTAL 0.04290952 76.72388862 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 505956 +BPFP 1.7160 bits/point +EBPFP 3.4320 equivalent bits/point +MSE 76.723889 +---------------------- -------------------------------------------------------- +Time: 3.237s Load: 0.011s, Pack+Encode: 1.854s, Decode+Unpack: 1.372s +---------------------- -------------------------------------------------------- +💾 Converting with 76.7239 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-266.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,612B, BPFP=0.7063 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,004B, BPFP=3.2484 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,920B, BPFP=1.5636 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,984B, BPFP=3.1353 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,716B, BPFP=1.6642 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,520B, BPFP=3.0533 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,136B, BPFP=1.5197 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,240B, BPFP=3.1496 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,236B, BPFP=2.4214 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,180B, BPFP=3.0903 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,856B, BPFP=0.7189 +⌛️ [2/4] FRONTEND: Frontend time: 1.861s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15022702 29.48492804 + layer.0.v_cache 0.00001395 0.00190186 + layer.1.k_cache 0.54271635 2.80442367 + layer.1.v_cache 0.00000573 0.00082715 + layer.2.k_cache 0.01332666 0.66763038 + layer.2.v_cache 0.00001832 0.00263335 + layer.3.k_cache 0.02579262 3.83012669 + layer.3.v_cache 0.00001838 0.00275789 + layer.4.k_cache 0.00069325 0.07311565 + layer.4.v_cache 0.00005155 0.00573630 + layer.4.output 0.01006869 171.07683692 + ------------------------------------------------------------------------------------- + TOTAL 0.04725557 72.61246702 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 510404 +BPFP 1.6814 bits/point +EBPFP 3.3629 equivalent bits/point +MSE 72.612467 +---------------------- -------------------------------------------------------- +Time: 3.243s Load: 0.009s, Pack+Encode: 1.861s, Decode+Unpack: 1.373s +---------------------- -------------------------------------------------------- +💾 Converting with 72.6125 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-267.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,136B, BPFP=0.6958 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,592B, BPFP=3.1034 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,248B, BPFP=1.4962 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,224B, BPFP=3.0309 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,972B, BPFP=1.6405 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,608B, BPFP=2.9453 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,872B, BPFP=1.4763 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,360B, BPFP=3.0381 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,216B, BPFP=2.3949 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,076B, BPFP=2.9701 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,296B, BPFP=0.7513 +⌛️ [2/4] FRONTEND: Frontend time: 1.862s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12416320 29.67664857 + layer.0.v_cache 0.00001370 0.00185372 + layer.1.k_cache 0.62725913 2.80199347 + layer.1.v_cache 0.00000565 0.00081940 + layer.2.k_cache 0.01102555 0.63986915 + layer.2.v_cache 0.00002033 0.00265254 + layer.3.k_cache 0.04118381 3.89940413 + layer.3.v_cache 0.00001949 0.00280124 + layer.4.k_cache 0.00070768 0.07615825 + layer.4.v_cache 0.00005397 0.00569585 + layer.4.output 0.05105524 159.16849274 + ------------------------------------------------------------------------------------- + TOTAL 0.06834348 67.72278503 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 529600 +BPFP 1.6500 bits/point +EBPFP 3.3001 equivalent bits/point +MSE 67.722785 +---------------------- -------------------------------------------------------- +Time: 3.256s Load: 0.012s, Pack+Encode: 1.862s, Decode+Unpack: 1.382s +---------------------- -------------------------------------------------------- +💾 Converting with 67.7228 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-269.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,672B, BPFP=0.6851 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,912B, BPFP=3.1311 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,988B, BPFP=1.4051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,516B, BPFP=3.0556 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,708B, BPFP=1.6603 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,028B, BPFP=2.9751 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,528B, BPFP=1.4883 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,652B, BPFP=3.0629 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,884B, BPFP=2.4267 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,616B, BPFP=3.0069 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,696B, BPFP=0.7314 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15516999 29.95979536 + layer.0.v_cache 0.00001361 0.00194132 + layer.1.k_cache 0.56212566 2.79096958 + layer.1.v_cache 0.00000565 0.00084497 + layer.2.k_cache 0.00839357 0.61718370 + layer.2.v_cache 0.00001980 0.00275478 + layer.3.k_cache 0.02402391 4.26059900 + layer.3.v_cache 0.00001987 0.00285627 + layer.4.k_cache 0.00071212 0.07715540 + layer.4.v_cache 0.00005188 0.00590698 + layer.4.output 0.05073151 144.01289854 + ------------------------------------------------------------------------------------- + TOTAL 0.06503863 61.51825277 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 518200 +BPFP 1.6481 bits/point +EBPFP 3.2961 equivalent bits/point +MSE 61.518253 +---------------------- -------------------------------------------------------- +Time: 3.245s Load: 0.010s, Pack+Encode: 1.856s, Decode+Unpack: 1.379s +---------------------- -------------------------------------------------------- +💾 Converting with 61.5183 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-271.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,280B, BPFP=0.6987 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,888B, BPFP=3.0981 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,688B, BPFP=1.6671 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,356B, BPFP=3.0175 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,004B, BPFP=1.6311 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,224B, BPFP=2.9579 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,164B, BPFP=1.4817 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,204B, BPFP=3.0621 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,460B, BPFP=2.3916 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,292B, BPFP=2.9615 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,384B, BPFP=0.7695 +⌛️ [2/4] FRONTEND: Frontend time: 1.857s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005571 30.94787063 + layer.0.v_cache 0.00001459 0.00184484 + layer.1.k_cache 0.64038677 3.18936024 + layer.1.v_cache 0.00000545 0.00078096 + layer.2.k_cache 0.01623841 0.63185669 + layer.2.v_cache 0.00002030 0.00262636 + layer.3.k_cache 0.01774169 4.08357090 + layer.3.v_cache 0.00001948 0.00279730 + layer.4.k_cache 0.00069419 0.07500247 + layer.4.v_cache 0.00005057 0.00579919 + layer.4.output 0.05052170 157.73358586 + ------------------------------------------------------------------------------------- + TOTAL 0.06875759 67.23980062 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 538944 +BPFP 1.6679 bits/point +EBPFP 3.3357 equivalent bits/point +MSE 67.239801 +---------------------- -------------------------------------------------------- +Time: 3.253s Load: 0.010s, Pack+Encode: 1.857s, Decode+Unpack: 1.386s +---------------------- -------------------------------------------------------- +💾 Converting with 67.2398 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-278.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,620B, BPFP=0.7068 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,492B, BPFP=3.2198 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,708B, BPFP=1.5517 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,332B, BPFP=3.1548 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,572B, BPFP=1.6561 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,916B, BPFP=3.0755 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,276B, BPFP=1.5276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,636B, BPFP=3.1718 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,436B, BPFP=2.4326 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,472B, BPFP=3.1066 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,828B, BPFP=0.7027 +⌛️ [2/4] FRONTEND: Frontend time: 1.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14010120 28.52673821 + layer.0.v_cache 0.00001437 0.00192753 + layer.1.k_cache 0.58692467 3.05158411 + layer.1.v_cache 0.00000569 0.00085184 + layer.2.k_cache 0.00806839 0.72916628 + layer.2.v_cache 0.00001934 0.00268160 + layer.3.k_cache 0.03832107 3.75854361 + layer.3.v_cache 0.00001992 0.00296951 + layer.4.k_cache 0.00069804 0.07536473 + layer.4.v_cache 0.00005131 0.00601413 + layer.4.output 0.00759059 170.27638569 + ------------------------------------------------------------------------------------- + TOTAL 0.04866812 72.24062008 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 509288 +BPFP 1.6778 bits/point +EBPFP 3.3555 equivalent bits/point +MSE 72.240620 +---------------------- -------------------------------------------------------- +Time: 3.242s Load: 0.010s, Pack+Encode: 1.847s, Decode+Unpack: 1.385s +---------------------- -------------------------------------------------------- +💾 Converting with 72.2406 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-28.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,308B, BPFP=0.7176 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,228B, BPFP=3.3365 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,576B, BPFP=1.4911 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,648B, BPFP=3.2444 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,960B, BPFP=1.6884 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,396B, BPFP=3.1714 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,000B, BPFP=1.5159 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,288B, BPFP=3.2817 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,024B, BPFP=2.5084 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,344B, BPFP=3.1684 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,164B, BPFP=0.7093 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13199043 30.58043887 + layer.0.v_cache 0.00001386 0.00190756 + layer.1.k_cache 0.56989129 2.82812318 + layer.1.v_cache 0.00000577 0.00084508 + layer.2.k_cache 0.00892691 0.72993196 + layer.2.v_cache 0.00001816 0.00269062 + layer.3.k_cache 0.04626510 3.67072900 + layer.3.v_cache 0.00001927 0.00285276 + layer.4.k_cache 0.00069392 0.07259034 + layer.4.v_cache 0.00005195 0.00572074 + layer.4.output 0.00911696 177.44026519 + ------------------------------------------------------------------------------------- + TOTAL 0.04833502 75.29280508 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 498936 +BPFP 1.7111 bits/point +EBPFP 3.4222 equivalent bits/point +MSE 75.292805 +---------------------- -------------------------------------------------------- +Time: 3.245s Load: 0.010s, Pack+Encode: 1.855s, Decode+Unpack: 1.380s +---------------------- -------------------------------------------------------- +💾 Converting with 75.2928 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-280.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 274, 128) +Output shape: (1, 274, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) -> torch.Size([1, 1, 274, 512]) + layer.4.output: torch.Size([1, 274, 3584]) -> torch.Size([1, 1, 274, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,504B, BPFP=0.7130 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,748B, BPFP=3.2361 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,948B, BPFP=1.7648 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,120B, BPFP=3.1432 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,292B, BPFP=1.6704 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,004B, BPFP=3.0796 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,448B, BPFP=1.5082 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,184B, BPFP=3.2039 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,972B, BPFP=2.4505 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,228B, BPFP=3.0924 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,548B, BPFP=0.6969 +⌛️ [2/4] FRONTEND: Frontend time: 1.889s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 274, 128]) + layer.0.v_cache: torch.Size([1, 4, 274, 128]) + layer.1.k_cache: torch.Size([1, 4, 274, 128]) + layer.1.v_cache: torch.Size([1, 4, 274, 128]) + layer.2.k_cache: torch.Size([1, 4, 274, 128]) + layer.2.v_cache: torch.Size([1, 4, 274, 128]) + layer.3.k_cache: torch.Size([1, 4, 274, 128]) + layer.3.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.k_cache: torch.Size([1, 4, 274, 128]) + layer.4.v_cache: torch.Size([1, 4, 274, 128]) + layer.4.output: torch.Size([1, 274, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16500157 30.42090557 + layer.0.v_cache 0.00001391 0.00189453 + layer.1.k_cache 0.57858722 3.17891476 + layer.1.v_cache 0.00000575 0.00083553 + layer.2.k_cache 0.01018238 0.79028181 + layer.2.v_cache 0.00001877 0.00270745 + layer.3.k_cache 0.03425542 3.75178205 + layer.3.v_cache 0.00002021 0.00290351 + layer.4.k_cache 0.00069807 0.07340526 + layer.4.v_cache 0.00005194 0.00587351 + layer.4.output 0.01023613 179.33832443 + ------------------------------------------------------------------------------------- + TOTAL 0.05061695 76.09398676 + (elements=2,384,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2384896 +Total Bytes 503996 +BPFP 1.6906 bits/point +EBPFP 3.3813 equivalent bits/point +MSE 76.093987 +---------------------- -------------------------------------------------------- +Time: 3.278s Load: 0.008s, Pack+Encode: 1.889s, Decode+Unpack: 1.381s +---------------------- -------------------------------------------------------- +💾 Converting with 76.0940 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-285.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,356B, BPFP=0.7150 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,104B, BPFP=3.3046 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,528B, BPFP=1.5931 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,312B, BPFP=3.2009 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,476B, BPFP=1.6479 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,960B, BPFP=3.1227 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,880B, BPFP=1.4977 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,520B, BPFP=3.2708 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,560B, BPFP=2.4630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,844B, BPFP=3.1160 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,360B, BPFP=0.7222 +⌛️ [2/4] FRONTEND: Frontend time: 1.850s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.366s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385483 29.58237124 + layer.0.v_cache 0.00001387 0.00187189 + layer.1.k_cache 0.55871571 3.10206231 + layer.1.v_cache 0.00000565 0.00079476 + layer.2.k_cache 0.00818202 0.71712844 + layer.2.v_cache 0.00001875 0.00259576 + layer.3.k_cache 0.04697755 3.90900065 + layer.3.v_cache 0.00001978 0.00276058 + layer.4.k_cache 0.00071636 0.07438524 + layer.4.v_cache 0.00005146 0.00585972 + layer.4.output 0.01106052 184.72628968 + ------------------------------------------------------------------------------------- + TOTAL 0.04741057 78.26369755 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 500900 +BPFP 1.7051 bits/point +EBPFP 3.4103 equivalent bits/point +MSE 78.263698 +---------------------- -------------------------------------------------------- +Time: 3.225s Load: 0.009s, Pack+Encode: 1.850s, Decode+Unpack: 1.366s +---------------------- -------------------------------------------------------- +💾 Converting with 78.2637 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-291.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,544B, BPFP=0.7127 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,708B, BPFP=3.3357 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,536B, BPFP=1.7918 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,664B, BPFP=3.2195 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,276B, BPFP=1.6634 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,572B, BPFP=3.1007 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,348B, BPFP=1.4970 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,696B, BPFP=3.2214 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,436B, BPFP=2.4680 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,952B, BPFP=3.1223 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,172B, BPFP=0.6913 +⌛️ [2/4] FRONTEND: Frontend time: 1.851s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203608 28.27250888 + layer.0.v_cache 0.00001420 0.00190371 + layer.1.k_cache 0.56673284 3.24146928 + layer.1.v_cache 0.00000552 0.00080167 + layer.2.k_cache 0.00530624 0.68881914 + layer.2.v_cache 0.00001825 0.00266731 + layer.3.k_cache 0.02188354 3.80047763 + layer.3.v_cache 0.00001854 0.00281561 + layer.4.k_cache 0.00069888 0.07425910 + layer.4.v_cache 0.00004918 0.00565158 + layer.4.output 0.00799176 180.74332792 + ------------------------------------------------------------------------------------- + TOTAL 0.04662974 76.54674526 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 509904 +BPFP 1.7042 bits/point +EBPFP 3.4084 equivalent bits/point +MSE 76.546745 +---------------------- -------------------------------------------------------- +Time: 3.223s Load: 0.009s, Pack+Encode: 1.851s, Decode+Unpack: 1.362s +---------------------- -------------------------------------------------------- +💾 Converting with 76.5467 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-292.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-292.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,660B, BPFP=0.6990 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,368B, BPFP=3.1674 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,580B, BPFP=1.5227 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,932B, BPFP=3.0881 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,972B, BPFP=1.5996 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,460B, BPFP=3.0068 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,824B, BPFP=1.4810 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,768B, BPFP=3.0791 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,500B, BPFP=2.4017 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,944B, BPFP=3.0336 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,056B, BPFP=0.7103 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.368s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12732459 30.16032534 + layer.0.v_cache 0.00001480 0.00186435 + layer.1.k_cache 0.57322790 2.71895618 + layer.1.v_cache 0.00000557 0.00081555 + layer.2.k_cache 0.00950604 0.61614834 + layer.2.v_cache 0.00002039 0.00272927 + layer.3.k_cache 0.01510038 3.86411417 + layer.3.v_cache 0.00001916 0.00276814 + layer.4.k_cache 0.00068141 0.07201407 + layer.4.v_cache 0.00005165 0.00573738 + layer.4.output 0.00936729 165.95665068 + ------------------------------------------------------------------------------------- + TOTAL 0.04656017 70.53776633 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 508064 +BPFP 1.6501 bits/point +EBPFP 3.3001 equivalent bits/point +MSE 70.537766 +---------------------- -------------------------------------------------------- +Time: 3.236s Load: 0.012s, Pack+Encode: 1.856s, Decode+Unpack: 1.368s +---------------------- -------------------------------------------------------- +💾 Converting with 70.5378 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-30.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-30.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,536B, BPFP=0.7097 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,816B, BPFP=3.2731 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,332B, BPFP=1.6606 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,360B, BPFP=3.1907 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,652B, BPFP=1.6787 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,272B, BPFP=3.1291 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,816B, BPFP=1.5181 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,856B, BPFP=3.2188 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,328B, BPFP=2.4529 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,240B, BPFP=3.1273 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,260B, BPFP=0.7138 +⌛️ [2/4] FRONTEND: Frontend time: 1.858s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13630490 28.60342363 + layer.0.v_cache 0.00001544 0.00190210 + layer.1.k_cache 0.58776402 3.15610029 + layer.1.v_cache 0.00000559 0.00081227 + layer.2.k_cache 0.01000112 0.80706102 + layer.2.v_cache 0.00001940 0.00270565 + layer.3.k_cache 0.04424953 3.62359066 + layer.3.v_cache 0.00002139 0.00289137 + layer.4.k_cache 0.00071093 0.07345925 + layer.4.v_cache 0.00005158 0.00587096 + layer.4.output 0.01101471 174.75784485 + ------------------------------------------------------------------------------------- + TOTAL 0.05036746 74.09310183 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 511468 +BPFP 1.7033 bits/point +EBPFP 3.4065 equivalent bits/point +MSE 74.093102 +---------------------- -------------------------------------------------------- +Time: 3.241s Load: 0.010s, Pack+Encode: 1.858s, Decode+Unpack: 1.373s +---------------------- -------------------------------------------------------- +💾 Converting with 74.0931 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-300.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,564B, BPFP=0.7062 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,336B, BPFP=3.2226 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,112B, BPFP=1.5238 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,656B, BPFP=3.1281 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,384B, BPFP=1.6515 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,680B, BPFP=3.0733 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,156B, BPFP=1.5263 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,784B, BPFP=3.1353 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,396B, BPFP=2.4391 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,796B, BPFP=3.0798 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,780B, BPFP=0.7128 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12958164 28.89765590 + layer.0.v_cache 0.00001430 0.00188857 + layer.1.k_cache 0.54121377 2.89530143 + layer.1.v_cache 0.00000577 0.00080416 + layer.2.k_cache 0.01080061 0.72823048 + layer.2.v_cache 0.00001903 0.00262778 + layer.3.k_cache 0.01783211 3.80516876 + layer.3.v_cache 0.00002028 0.00279886 + layer.4.k_cache 0.00068935 0.07452454 + layer.4.v_cache 0.00005053 0.00572223 + layer.4.output 0.00833506 176.73527428 + ------------------------------------------------------------------------------------- + TOTAL 0.04462193 74.91539075 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 506644 +BPFP 1.6751 bits/point +EBPFP 3.3501 equivalent bits/point +MSE 74.915391 +---------------------- -------------------------------------------------------- +Time: 3.236s Load: 0.010s, Pack+Encode: 1.856s, Decode+Unpack: 1.370s +---------------------- -------------------------------------------------------- +💾 Converting with 74.9154 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-308.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,084B, BPFP=0.7179 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 55,968B, BPFP=3.3251 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,432B, BPFP=1.4515 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 54,588B, BPFP=3.2431 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,388B, BPFP=1.6865 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,392B, BPFP=3.1721 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,012B, BPFP=1.5454 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,288B, BPFP=3.2847 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,160B, BPFP=2.5048 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,060B, BPFP=3.1523 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,456B, BPFP=0.7592 +⌛️ [2/4] FRONTEND: Frontend time: 1.858s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.370s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12547956 29.79211844 + layer.0.v_cache 0.00001492 0.00192206 + layer.1.k_cache 0.50446328 2.56265062 + layer.1.v_cache 0.00000633 0.00081912 + layer.2.k_cache 0.00671430 0.80245084 + layer.2.v_cache 0.00001992 0.00273891 + layer.3.k_cache 0.02142675 3.56737283 + layer.3.v_cache 0.00001986 0.00296019 + layer.4.k_cache 0.00069859 0.07252220 + layer.4.v_cache 0.00005414 0.00598956 + layer.4.output 0.01038299 149.68931627 + ------------------------------------------------------------------------------------- + TOTAL 0.04303403 63.80216227 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 494828 +BPFP 1.7293 bits/point +EBPFP 3.4586 equivalent bits/point +MSE 63.802162 +---------------------- -------------------------------------------------------- +Time: 3.239s Load: 0.010s, Pack+Encode: 1.858s, Decode+Unpack: 1.370s +---------------------- -------------------------------------------------------- +💾 Converting with 63.8022 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-309.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,432B, BPFP=0.6973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,848B, BPFP=3.0548 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,600B, BPFP=1.5365 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,252B, BPFP=2.9720 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,436B, BPFP=1.6319 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,980B, BPFP=2.9059 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,984B, BPFP=1.5046 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,164B, BPFP=2.9674 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,480B, BPFP=2.3609 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,288B, BPFP=2.8700 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,972B, BPFP=0.7562 +⌛️ [2/4] FRONTEND: Frontend time: 1.850s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13970643 30.61669111 + layer.0.v_cache 0.00001441 0.00186700 + layer.1.k_cache 0.60015088 3.11698554 + layer.1.v_cache 0.00000584 0.00080485 + layer.2.k_cache 0.00648033 0.61030908 + layer.2.v_cache 0.00001840 0.00253330 + layer.3.k_cache 0.03288257 4.27800723 + layer.3.v_cache 0.00002024 0.00274287 + layer.4.k_cache 0.00070287 0.07602437 + layer.4.v_cache 0.00005060 0.00559734 + layer.4.output 0.04874230 155.58740211 + ------------------------------------------------------------------------------------- + TOTAL 0.06595463 66.34255162 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 535436 +BPFP 1.6350 bits/point +EBPFP 3.2700 equivalent bits/point +MSE 66.342552 +---------------------- -------------------------------------------------------- +Time: 3.242s Load: 0.011s, Pack+Encode: 1.850s, Decode+Unpack: 1.381s +---------------------- -------------------------------------------------------- +💾 Converting with 66.3426 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-31.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-31.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,560B, BPFP=0.6957 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 68,704B, BPFP=3.2829 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,540B, BPFP=1.5071 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 66,328B, BPFP=3.1693 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 34,380B, BPFP=1.6428 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 65,252B, BPFP=3.1179 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,240B, BPFP=1.4927 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 67,316B, BPFP=3.2166 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 51,268B, BPFP=2.4497 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 64,624B, BPFP=3.0879 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 108,244B, BPFP=0.7389 +⌛️ [2/4] FRONTEND: Frontend time: 34.922s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.494s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13960611 29.92003536 + layer.0.v_cache 0.00001491 0.00192727 + layer.1.k_cache 0.68890666 2.92080362 + layer.1.v_cache 0.00000583 0.00081674 + layer.2.k_cache 0.00940153 0.71662721 + layer.2.v_cache 0.00001882 0.00261758 + layer.3.k_cache 0.04667800 3.70060968 + layer.3.v_cache 0.00001939 0.00283283 + layer.4.k_cache 0.00070827 0.07181553 + layer.4.v_cache 0.00005291 0.00574435 + layer.4.output 0.04427218 139.66582023 + ------------------------------------------------------------------------------------- + TOTAL 0.07031281 59.70615128 + (elements=2,846,208) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2846208 +Total Bytes 603456 +BPFP 1.6962 bits/point +EBPFP 3.3923 equivalent bits/point +MSE 59.706151 +---------------------- --------------------------------------------------------- +Time: 36.429s Load: 0.012s, Pack+Encode: 34.922s, Decode+Unpack: 1.494s +---------------------- --------------------------------------------------------- +💾 Converting with 59.7062 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-313.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,360B, BPFP=0.6982 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,616B, BPFP=3.0631 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,240B, BPFP=1.5803 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,212B, BPFP=2.9898 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,372B, BPFP=1.6394 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,824B, BPFP=2.9172 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,264B, BPFP=1.4770 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,264B, BPFP=2.9925 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,504B, BPFP=2.3779 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,888B, BPFP=2.9206 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,572B, BPFP=0.7359 +⌛️ [2/4] FRONTEND: Frontend time: 1.855s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15817952 29.41089439 + layer.0.v_cache 0.00001474 0.00188845 + layer.1.k_cache 0.61609703 3.16916210 + layer.1.v_cache 0.00000566 0.00080261 + layer.2.k_cache 0.00844678 0.60311165 + layer.2.v_cache 0.00001876 0.00257914 + layer.3.k_cache 0.04682518 4.37707234 + layer.3.v_cache 0.00001896 0.00272230 + layer.4.k_cache 0.00069859 0.07621608 + layer.4.v_cache 0.00005216 0.00577772 + layer.4.output 0.04649196 154.44783206 + ------------------------------------------------------------------------------------- + TOTAL 0.06798830 65.81088536 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 532116 +BPFP 1.6357 bits/point +EBPFP 3.2714 equivalent bits/point +MSE 65.810885 +---------------------- -------------------------------------------------------- +Time: 3.246s Load: 0.011s, Pack+Encode: 1.855s, Decode+Unpack: 1.380s +---------------------- -------------------------------------------------------- +💾 Converting with 65.8109 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-343.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,656B, BPFP=0.6890 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,380B, BPFP=3.1239 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,856B, BPFP=1.5710 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,864B, BPFP=3.0414 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,928B, BPFP=1.6294 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,900B, BPFP=2.9889 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,556B, BPFP=1.5002 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,860B, BPFP=3.0956 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,000B, BPFP=2.5588 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,864B, BPFP=2.9869 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,844B, BPFP=0.7143 +⌛️ [2/4] FRONTEND: Frontend time: 1.864s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14002854 29.61555082 + layer.0.v_cache 0.00001601 0.00194136 + layer.1.k_cache 0.54217949 3.17147944 + layer.1.v_cache 0.00000579 0.00083397 + layer.2.k_cache 0.01082110 0.66060030 + layer.2.v_cache 0.00001971 0.00271324 + layer.3.k_cache 0.03102738 4.29865799 + layer.3.v_cache 0.00001907 0.00283293 + layer.4.k_cache 0.00066782 0.07629296 + layer.4.v_cache 0.00005141 0.00577921 + layer.4.output 0.00856273 167.92773146 + ------------------------------------------------------------------------------------- + TOTAL 0.04616326 71.37240014 + (elements=2,498,048) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2498048 +Total Bytes 517708 +BPFP 1.6580 bits/point +EBPFP 3.3159 equivalent bits/point +MSE 71.372400 +---------------------- -------------------------------------------------------- +Time: 3.256s Load: 0.012s, Pack+Encode: 1.864s, Decode+Unpack: 1.380s +---------------------- -------------------------------------------------------- +💾 Converting with 71.3724 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-344.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 297, 128) +Output shape: (1, 297, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) -> torch.Size([1, 1, 297, 512]) + layer.4.output: torch.Size([1, 297, 3584]) -> torch.Size([1, 1, 297, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,224B, BPFP=0.6957 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,672B, BPFP=3.0867 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,452B, BPFP=1.4968 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,332B, BPFP=3.0162 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,476B, BPFP=1.6033 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,036B, BPFP=2.9480 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,528B, BPFP=1.5008 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,484B, BPFP=3.0242 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,384B, BPFP=2.3876 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,128B, BPFP=2.9529 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,252B, BPFP=0.7234 +⌛️ [2/4] FRONTEND: Frontend time: 1.873s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 297, 128]) + layer.0.v_cache: torch.Size([1, 4, 297, 128]) + layer.1.k_cache: torch.Size([1, 4, 297, 128]) + layer.1.v_cache: torch.Size([1, 4, 297, 128]) + layer.2.k_cache: torch.Size([1, 4, 297, 128]) + layer.2.v_cache: torch.Size([1, 4, 297, 128]) + layer.3.k_cache: torch.Size([1, 4, 297, 128]) + layer.3.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.k_cache: torch.Size([1, 4, 297, 128]) + layer.4.v_cache: torch.Size([1, 4, 297, 128]) + layer.4.output: torch.Size([1, 297, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14705232 30.26537839 + layer.0.v_cache 0.00001374 0.00184951 + layer.1.k_cache 0.65666707 3.10821143 + layer.1.v_cache 0.00000579 0.00079681 + layer.2.k_cache 0.01062412 0.64564488 + layer.2.v_cache 0.00001918 0.00259012 + layer.3.k_cache 0.03205616 4.34012981 + layer.3.v_cache 0.00001999 0.00282527 + layer.4.k_cache 0.00071739 0.07633304 + layer.4.v_cache 0.00005111 0.00560717 + layer.4.output 0.04976816 117.74773028 + ------------------------------------------------------------------------------------- + TOTAL 0.07032965 50.74608696 + (elements=2,585,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2585088 +Total Bytes 527968 +BPFP 1.6339 bits/point +EBPFP 3.2678 equivalent bits/point +MSE 50.746087 +---------------------- -------------------------------------------------------- +Time: 3.275s Load: 0.012s, Pack+Encode: 1.873s, Decode+Unpack: 1.390s +---------------------- -------------------------------------------------------- +💾 Converting with 50.7461 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-350.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-350.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,636B, BPFP=0.7026 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,280B, BPFP=3.1851 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,356B, BPFP=1.5767 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,676B, BPFP=3.0959 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,672B, BPFP=1.6499 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,312B, BPFP=3.0200 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,240B, BPFP=1.5147 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,712B, BPFP=3.0979 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,916B, BPFP=2.3863 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,660B, BPFP=3.0394 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,908B, BPFP=0.7221 +⌛️ [2/4] FRONTEND: Frontend time: 1.864s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131647 30.00408697 + layer.0.v_cache 0.00001392 0.00180609 + layer.1.k_cache 0.51967005 2.79715607 + layer.1.v_cache 0.00000557 0.00077583 + layer.2.k_cache 0.01013049 0.64260191 + layer.2.v_cache 0.00001908 0.00253148 + layer.3.k_cache 0.05186962 3.69406226 + layer.3.v_cache 0.00001891 0.00270787 + layer.4.k_cache 0.00069958 0.07378218 + layer.4.v_cache 0.00005017 0.00551924 + layer.4.output 0.01073789 174.73945094 + ------------------------------------------------------------------------------------- + TOTAL 0.04758583 74.14124626 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 509368 +BPFP 1.6661 bits/point +EBPFP 3.3322 equivalent bits/point +MSE 74.141246 +---------------------- -------------------------------------------------------- +Time: 3.262s Load: 0.011s, Pack+Encode: 1.864s, Decode+Unpack: 1.388s +---------------------- -------------------------------------------------------- +💾 Converting with 74.1412 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-351.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-351.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,608B, BPFP=0.6840 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,144B, BPFP=3.1545 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,104B, BPFP=1.4705 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,656B, BPFP=3.0738 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,796B, BPFP=1.6165 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,940B, BPFP=2.9807 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,648B, BPFP=1.5000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,660B, BPFP=3.0740 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,472B, BPFP=2.4128 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,132B, BPFP=2.9911 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,164B, BPFP=0.7221 +⌛️ [2/4] FRONTEND: Frontend time: 1.869s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.385s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15566496 30.53669230 + layer.0.v_cache 0.00001437 0.00183271 + layer.1.k_cache 0.62643353 2.77392069 + layer.1.v_cache 0.00000573 0.00080859 + layer.2.k_cache 0.01010078 0.67295074 + layer.2.v_cache 0.00001868 0.00264549 + layer.3.k_cache 0.06271097 3.97049120 + layer.3.v_cache 0.00001958 0.00284922 + layer.4.k_cache 0.00073784 0.07404003 + layer.4.v_cache 0.00005294 0.00578662 + layer.4.output 0.00677609 169.70576017 + ------------------------------------------------------------------------------------- + TOTAL 0.05312894 72.11660816 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 516324 +BPFP 1.6478 bits/point +EBPFP 3.2956 equivalent bits/point +MSE 72.116608 +---------------------- -------------------------------------------------------- +Time: 3.266s Load: 0.012s, Pack+Encode: 1.869s, Decode+Unpack: 1.385s +---------------------- -------------------------------------------------------- +💾 Converting with 72.1166 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-360.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-360.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,680B, BPFP=0.6927 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,928B, BPFP=3.1648 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,400B, BPFP=1.4423 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,476B, BPFP=3.0854 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,544B, BPFP=1.6687 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,668B, BPFP=3.0413 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,336B, BPFP=1.4934 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,420B, BPFP=3.0824 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,520B, BPFP=2.4323 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,520B, BPFP=3.0332 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,352B, BPFP=0.7052 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13918158 30.46646566 + layer.0.v_cache 0.00001566 0.00192525 + layer.1.k_cache 0.57997750 2.95342932 + layer.1.v_cache 0.00000571 0.00084065 + layer.2.k_cache 0.01637170 0.69122677 + layer.2.v_cache 0.00001887 0.00279325 + layer.3.k_cache 0.04393929 4.16344831 + layer.3.v_cache 0.00001903 0.00287155 + layer.4.k_cache 0.00071104 0.07589404 + layer.4.v_cache 0.00005175 0.00598659 + layer.4.output 0.01065216 174.19805195 + ------------------------------------------------------------------------------------- + TOTAL 0.05028572 73.98536736 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 513844 +BPFP 1.6513 bits/point +EBPFP 3.3027 equivalent bits/point +MSE 73.985367 +---------------------- -------------------------------------------------------- +Time: 3.255s Load: 0.012s, Pack+Encode: 1.856s, Decode+Unpack: 1.387s +---------------------- -------------------------------------------------------- +💾 Converting with 73.9854 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-367.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-367.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,620B, BPFP=0.6865 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,068B, BPFP=2.9772 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,744B, BPFP=1.4992 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,452B, BPFP=2.8958 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,032B, BPFP=1.6145 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,340B, BPFP=2.8397 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,124B, BPFP=1.4679 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,504B, BPFP=2.8984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,188B, BPFP=2.3280 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,748B, BPFP=2.8099 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,960B, BPFP=0.7270 +⌛️ [2/4] FRONTEND: Frontend time: 1.861s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16496690 32.42443611 + layer.0.v_cache 0.00001433 0.00191940 + layer.1.k_cache 0.65265572 3.12961603 + layer.1.v_cache 0.00000578 0.00083099 + layer.2.k_cache 0.01241827 0.61437181 + layer.2.v_cache 0.00001920 0.00256528 + layer.3.k_cache 0.08444468 4.29913054 + layer.3.v_cache 0.00001975 0.00279207 + layer.4.k_cache 0.00068928 0.07651391 + layer.4.v_cache 0.00005259 0.00580592 + layer.4.output 0.04785494 159.29156106 + ------------------------------------------------------------------------------------- + TOTAL 0.07354536 67.97640644 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 537780 +BPFP 1.5945 bits/point +EBPFP 3.1889 equivalent bits/point +MSE 67.976406 +---------------------- -------------------------------------------------------- +Time: 3.258s Load: 0.010s, Pack+Encode: 1.861s, Decode+Unpack: 1.388s +---------------------- -------------------------------------------------------- +💾 Converting with 67.9764 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-375.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-375.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,088B, BPFP=0.6956 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,664B, BPFP=3.1178 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,872B, BPFP=1.4813 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,232B, BPFP=3.0417 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,036B, BPFP=1.6494 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,680B, BPFP=2.9592 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,652B, BPFP=1.4696 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,260B, BPFP=3.0432 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,156B, BPFP=2.3999 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,652B, BPFP=2.9577 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,752B, BPFP=0.7498 +⌛️ [2/4] FRONTEND: Frontend time: 1.859s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.386s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15001445 32.07898863 + layer.0.v_cache 0.00001441 0.00193960 + layer.1.k_cache 0.57475359 3.00249061 + layer.1.v_cache 0.00000584 0.00084369 + layer.2.k_cache 0.01434258 0.62653496 + layer.2.v_cache 0.00001954 0.00269459 + layer.3.k_cache 0.02649183 4.02580583 + layer.3.v_cache 0.00001945 0.00281296 + layer.4.k_cache 0.00069414 0.07599000 + layer.4.v_cache 0.00005306 0.00597611 + layer.4.output 0.05035525 172.19890367 + ------------------------------------------------------------------------------------- + TOTAL 0.06581739 73.24802369 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 528044 +BPFP 1.6508 bits/point +EBPFP 3.3016 equivalent bits/point +MSE 73.248024 +---------------------- -------------------------------------------------------- +Time: 3.256s Load: 0.011s, Pack+Encode: 1.859s, Decode+Unpack: 1.386s +---------------------- -------------------------------------------------------- +💾 Converting with 73.2480 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-38.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-38.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,636B, BPFP=0.7001 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,872B, BPFP=3.2066 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,664B, BPFP=1.5882 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,928B, BPFP=3.0988 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,944B, BPFP=1.6591 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,844B, BPFP=3.0388 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,264B, BPFP=1.5106 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,924B, BPFP=3.1540 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,880B, BPFP=2.4313 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,532B, BPFP=3.0769 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,248B, BPFP=0.6906 +⌛️ [2/4] FRONTEND: Frontend time: 1.870s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11854591 29.46376676 + layer.0.v_cache 0.00001566 0.00190409 + layer.1.k_cache 0.56985614 2.73718132 + layer.1.v_cache 0.00000555 0.00079702 + layer.2.k_cache 0.01267573 0.71387152 + layer.2.v_cache 0.00001902 0.00262957 + layer.3.k_cache 0.04477839 3.51702448 + layer.3.v_cache 0.00001896 0.00288896 + layer.4.k_cache 0.00071754 0.07630290 + layer.4.v_cache 0.00005321 0.00600405 + layer.4.output 0.00761333 180.75487589 + ------------------------------------------------------------------------------------- + TOTAL 0.04705761 76.57685305 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 510736 +BPFP 1.6646 bits/point +EBPFP 3.3293 equivalent bits/point +MSE 76.576853 +---------------------- -------------------------------------------------------- +Time: 3.262s Load: 0.008s, Pack+Encode: 1.870s, Decode+Unpack: 1.384s +---------------------- -------------------------------------------------------- +💾 Converting with 76.5769 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-386.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,104B, BPFP=0.6964 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,768B, BPFP=3.1233 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,140B, BPFP=1.4955 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,216B, BPFP=3.0408 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,460B, BPFP=1.6720 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,092B, BPFP=2.9811 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,636B, BPFP=1.5219 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,348B, BPFP=3.0478 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,316B, BPFP=2.4084 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,144B, BPFP=2.9838 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,912B, BPFP=0.7662 +⌛️ [2/4] FRONTEND: Frontend time: 1.866s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.393s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582189 30.55426565 + layer.0.v_cache 0.00001509 0.00192892 + layer.1.k_cache 0.58311327 2.96655502 + layer.1.v_cache 0.00000606 0.00083254 + layer.2.k_cache 0.00657600 0.65253734 + layer.2.v_cache 0.00001909 0.00260348 + layer.3.k_cache 0.02234380 4.23632522 + layer.3.v_cache 0.00001905 0.00275705 + layer.4.k_cache 0.00071828 0.07620503 + layer.4.v_cache 0.00005370 0.00575321 + layer.4.output 0.05044473 168.18593598 + ------------------------------------------------------------------------------------- + TOTAL 0.06363526 71.51772443 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 533136 +BPFP 1.6667 bits/point +EBPFP 3.3334 equivalent bits/point +MSE 71.517724 +---------------------- -------------------------------------------------------- +Time: 3.271s Load: 0.011s, Pack+Encode: 1.866s, Decode+Unpack: 1.393s +---------------------- -------------------------------------------------------- +💾 Converting with 71.5177 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-402.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 292, 128) +Output shape: (1, 292, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) -> torch.Size([1, 1, 292, 512]) + layer.4.output: torch.Size([1, 292, 3584]) -> torch.Size([1, 1, 292, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,016B, BPFP=0.6965 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,420B, BPFP=3.1261 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,564B, BPFP=1.5285 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,892B, BPFP=3.0443 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,416B, BPFP=1.6811 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,408B, BPFP=2.9649 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,136B, BPFP=1.5056 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,100B, BPFP=3.0554 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,208B, BPFP=2.4191 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,556B, BPFP=2.9728 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,100B, BPFP=0.7346 +⌛️ [2/4] FRONTEND: Frontend time: 1.873s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.394s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 292, 128]) + layer.0.v_cache: torch.Size([1, 4, 292, 128]) + layer.1.k_cache: torch.Size([1, 4, 292, 128]) + layer.1.v_cache: torch.Size([1, 4, 292, 128]) + layer.2.k_cache: torch.Size([1, 4, 292, 128]) + layer.2.v_cache: torch.Size([1, 4, 292, 128]) + layer.3.k_cache: torch.Size([1, 4, 292, 128]) + layer.3.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.k_cache: torch.Size([1, 4, 292, 128]) + layer.4.v_cache: torch.Size([1, 4, 292, 128]) + layer.4.output: torch.Size([1, 292, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887579 31.58203459 + layer.0.v_cache 0.00001384 0.00191372 + layer.1.k_cache 0.57277983 2.87223753 + layer.1.v_cache 0.00000568 0.00081397 + layer.2.k_cache 0.01314942 0.57702067 + layer.2.v_cache 0.00002105 0.00271237 + layer.3.k_cache 0.03125376 4.55813515 + layer.3.v_cache 0.00001978 0.00285198 + layer.4.k_cache 0.00068035 0.07686360 + layer.4.v_cache 0.00005123 0.00578942 + layer.4.output 0.04859251 61.62046692 + ------------------------------------------------------------------------------------- + TOTAL 0.06511755 27.70727302 + (elements=2,541,568) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2541568 +Total Bytes 525816 +BPFP 1.6551 bits/point +EBPFP 3.3102 equivalent bits/point +MSE 27.707273 +---------------------- -------------------------------------------------------- +Time: 3.279s Load: 0.012s, Pack+Encode: 1.873s, Decode+Unpack: 1.394s +---------------------- -------------------------------------------------------- +💾 Converting with 27.7073 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-413.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 338, 128) +Output shape: (1, 338, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) -> torch.Size([1, 1, 338, 512]) + layer.4.output: torch.Size([1, 338, 3584]) -> torch.Size([1, 1, 338, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,952B, BPFP=0.6912 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 69,576B, BPFP=3.2163 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 37,748B, BPFP=1.7450 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 67,240B, BPFP=3.1084 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 35,084B, BPFP=1.6219 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 65,656B, BPFP=3.0351 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,748B, BPFP=1.4676 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 68,140B, BPFP=3.1500 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 52,212B, BPFP=2.4136 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 65,844B, BPFP=3.0438 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 114,140B, BPFP=0.7538 +⌛️ [2/4] FRONTEND: Frontend time: 2.005s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.522s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 338, 128]) + layer.0.v_cache: torch.Size([1, 4, 338, 128]) + layer.1.k_cache: torch.Size([1, 4, 338, 128]) + layer.1.v_cache: torch.Size([1, 4, 338, 128]) + layer.2.k_cache: torch.Size([1, 4, 338, 128]) + layer.2.v_cache: torch.Size([1, 4, 338, 128]) + layer.3.k_cache: torch.Size([1, 4, 338, 128]) + layer.3.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.k_cache: torch.Size([1, 4, 338, 128]) + layer.4.v_cache: torch.Size([1, 4, 338, 128]) + layer.4.output: torch.Size([1, 338, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13976314 29.07184957 + layer.0.v_cache 0.00001472 0.00192415 + layer.1.k_cache 0.73750278 3.30866959 + layer.1.v_cache 0.00000589 0.00084121 + layer.2.k_cache 0.02011468 0.65377627 + layer.2.v_cache 0.00002041 0.00263103 + layer.3.k_cache 0.02722352 3.80521298 + layer.3.v_cache 0.00002029 0.00279320 + layer.4.k_cache 0.00072609 0.07624626 + layer.4.v_cache 0.00005477 0.00579294 + layer.4.output 0.04270584 144.85266536 + ------------------------------------------------------------------------------------- + TOTAL 0.07202277 61.81755263 + (elements=2,941,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2941952 +Total Bytes 622340 +BPFP 1.6923 bits/point +EBPFP 3.3846 equivalent bits/point +MSE 61.817553 +---------------------- -------------------------------------------------------- +Time: 3.539s Load: 0.013s, Pack+Encode: 2.005s, Decode+Unpack: 1.522s +---------------------- -------------------------------------------------------- +💾 Converting with 61.8176 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-417.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,960B, BPFP=0.6959 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,300B, BPFP=3.1304 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,912B, BPFP=1.5524 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,764B, BPFP=3.0479 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,488B, BPFP=1.6370 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,136B, BPFP=2.9605 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,340B, BPFP=1.4680 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,472B, BPFP=3.0322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,112B, BPFP=2.4223 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,136B, BPFP=2.9605 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,308B, BPFP=0.7157 +⌛️ [2/4] FRONTEND: Frontend time: 1.881s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150811 30.60394853 + layer.0.v_cache 0.00001374 0.00187242 + layer.1.k_cache 0.66021251 2.98236881 + layer.1.v_cache 0.00000574 0.00081806 + layer.2.k_cache 0.00963001 0.63473825 + layer.2.v_cache 0.00001988 0.00260977 + layer.3.k_cache 0.08055985 3.97072871 + layer.3.v_cache 0.00001942 0.00266428 + layer.4.k_cache 0.00069026 0.07473724 + layer.4.v_cache 0.00005126 0.00550717 + layer.4.output 0.05093874 151.98640771 + ------------------------------------------------------------------------------------- + TOTAL 0.07289894 64.83440278 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 519928 +BPFP 1.6422 bits/point +EBPFP 3.2844 equivalent bits/point +MSE 64.834403 +---------------------- -------------------------------------------------------- +Time: 3.282s Load: 0.011s, Pack+Encode: 1.881s, Decode+Unpack: 1.390s +---------------------- -------------------------------------------------------- +💾 Converting with 64.8344 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-42.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-42.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 303, 128) +Output shape: (1, 303, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) -> torch.Size([1, 1, 303, 512]) + layer.4.output: torch.Size([1, 303, 3584]) -> torch.Size([1, 1, 303, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,564B, BPFP=0.6995 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,696B, BPFP=3.0268 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,324B, BPFP=1.4090 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,028B, BPFP=2.9408 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,472B, BPFP=1.6229 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,932B, BPFP=2.8843 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,212B, BPFP=1.5064 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,064B, BPFP=2.9427 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,696B, BPFP=2.3564 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,336B, BPFP=2.8535 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,332B, BPFP=0.7318 +⌛️ [2/4] FRONTEND: Frontend time: 1.889s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.392s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 303, 128]) + layer.0.v_cache: torch.Size([1, 4, 303, 128]) + layer.1.k_cache: torch.Size([1, 4, 303, 128]) + layer.1.v_cache: torch.Size([1, 4, 303, 128]) + layer.2.k_cache: torch.Size([1, 4, 303, 128]) + layer.2.v_cache: torch.Size([1, 4, 303, 128]) + layer.3.k_cache: torch.Size([1, 4, 303, 128]) + layer.3.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.k_cache: torch.Size([1, 4, 303, 128]) + layer.4.v_cache: torch.Size([1, 4, 303, 128]) + layer.4.output: torch.Size([1, 303, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14893842 30.43933387 + layer.0.v_cache 0.00001415 0.00189075 + layer.1.k_cache 0.65110169 3.04631299 + layer.1.v_cache 0.00000575 0.00082357 + layer.2.k_cache 0.01411552 0.59715895 + layer.2.v_cache 0.00001904 0.00262074 + layer.3.k_cache 0.03939451 4.35400471 + layer.3.v_cache 0.00001876 0.00269049 + layer.4.k_cache 0.00070558 0.07776021 + layer.4.v_cache 0.00005345 0.00572770 + layer.4.output 0.04819407 156.04719177 + ------------------------------------------------------------------------------------- + TOTAL 0.07010149 66.52109802 + (elements=2,637,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2637312 +Total Bytes 530656 +BPFP 1.6097 bits/point +EBPFP 3.2194 equivalent bits/point +MSE 66.521098 +---------------------- -------------------------------------------------------- +Time: 3.292s Load: 0.011s, Pack+Encode: 1.889s, Decode+Unpack: 1.392s +---------------------- -------------------------------------------------------- +💾 Converting with 66.5211 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-429.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-429.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,032B, BPFP=0.7148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,324B, BPFP=3.3462 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,660B, BPFP=1.4651 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 54,116B, BPFP=3.2151 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,200B, BPFP=1.6754 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,132B, BPFP=3.1566 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,552B, BPFP=1.5181 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,224B, BPFP=3.2809 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,600B, BPFP=2.4715 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,064B, BPFP=3.1526 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,304B, BPFP=0.7070 +⌛️ [2/4] FRONTEND: Frontend time: 1.887s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.391s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14492303 28.86870618 + layer.0.v_cache 0.00001375 0.00192231 + layer.1.k_cache 0.50395557 2.36874819 + layer.1.v_cache 0.00000564 0.00082346 + layer.2.k_cache 0.00856250 0.74968728 + layer.2.v_cache 0.00001902 0.00267550 + layer.3.k_cache 0.02692018 3.64237947 + layer.3.v_cache 0.00002118 0.00285990 + layer.4.k_cache 0.00070790 0.07252282 + layer.4.v_cache 0.00005130 0.00585000 + layer.4.output 0.01131245 179.84378395 + ------------------------------------------------------------------------------------- + TOTAL 0.04496278 76.15427428 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 487208 +BPFP 1.7027 bits/point +EBPFP 3.4053 equivalent bits/point +MSE 76.154274 +---------------------- -------------------------------------------------------- +Time: 3.287s Load: 0.009s, Pack+Encode: 1.887s, Decode+Unpack: 1.391s +---------------------- -------------------------------------------------------- +💾 Converting with 76.1543 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-431.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-431.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,528B, BPFP=0.7092 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,480B, BPFP=3.2541 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,964B, BPFP=1.5831 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,652B, BPFP=3.2072 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,564B, BPFP=1.6737 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,096B, BPFP=3.1191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,984B, BPFP=1.5276 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,880B, BPFP=3.2201 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,160B, BPFP=2.5000 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,880B, BPFP=3.1069 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,768B, BPFP=0.7179 +⌛️ [2/4] FRONTEND: Frontend time: 1.875s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.392s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14113761 28.53981261 + layer.0.v_cache 0.00001369 0.00187258 + layer.1.k_cache 0.59212339 3.03218587 + layer.1.v_cache 0.00000552 0.00080664 + layer.2.k_cache 0.01129283 0.68891885 + layer.2.v_cache 0.00001853 0.00277429 + layer.3.k_cache 0.01358872 3.52418584 + layer.3.v_cache 0.00001883 0.00281541 + layer.4.k_cache 0.00070763 0.07542447 + layer.4.v_cache 0.00005153 0.00571609 + layer.4.output 0.01019159 211.48408385 + ------------------------------------------------------------------------------------- + TOTAL 0.04884114 89.19194704 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 510956 +BPFP 1.7016 bits/point +EBPFP 3.4031 equivalent bits/point +MSE 89.191947 +---------------------- -------------------------------------------------------- +Time: 3.277s Load: 0.011s, Pack+Encode: 1.875s, Decode+Unpack: 1.392s +---------------------- -------------------------------------------------------- +💾 Converting with 89.1919 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-432.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 287, 128) +Output shape: (1, 287, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) -> torch.Size([1, 1, 287, 512]) + layer.4.output: torch.Size([1, 287, 3584]) -> torch.Size([1, 1, 287, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,648B, BPFP=0.6886 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,548B, BPFP=3.1331 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,868B, BPFP=1.4083 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,832B, BPFP=3.0396 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,048B, BPFP=1.6359 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,268B, BPFP=2.9545 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,012B, BPFP=1.4706 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,432B, BPFP=3.0723 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,248B, BPFP=2.4090 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,500B, BPFP=3.0216 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,684B, BPFP=0.7286 +⌛️ [2/4] FRONTEND: Frontend time: 34.852s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.380s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 287, 128]) + layer.0.v_cache: torch.Size([1, 4, 287, 128]) + layer.1.k_cache: torch.Size([1, 4, 287, 128]) + layer.1.v_cache: torch.Size([1, 4, 287, 128]) + layer.2.k_cache: torch.Size([1, 4, 287, 128]) + layer.2.v_cache: torch.Size([1, 4, 287, 128]) + layer.3.k_cache: torch.Size([1, 4, 287, 128]) + layer.3.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.k_cache: torch.Size([1, 4, 287, 128]) + layer.4.v_cache: torch.Size([1, 4, 287, 128]) + layer.4.output: torch.Size([1, 287, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11028925 29.92476045 + layer.0.v_cache 0.00001466 0.00186244 + layer.1.k_cache 0.56032453 3.06052221 + layer.1.v_cache 0.00000595 0.00078810 + layer.2.k_cache 0.00809042 0.60412970 + layer.2.v_cache 0.00001929 0.00257250 + layer.3.k_cache 0.03501172 4.25626514 + layer.3.v_cache 0.00001861 0.00267616 + layer.4.k_cache 0.00070074 0.07540082 + layer.4.v_cache 0.00005246 0.00554331 + layer.4.output 0.00889761 168.37826655 + ------------------------------------------------------------------------------------- + TOTAL 0.04569476 71.56366980 + (elements=2,498,048) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2498048 +Total Bytes 513088 +BPFP 1.6432 bits/point +EBPFP 3.2863 equivalent bits/point +MSE 71.563670 +---------------------- --------------------------------------------------------- +Time: 36.243s Load: 0.012s, Pack+Encode: 34.852s, Decode+Unpack: 1.380s +---------------------- --------------------------------------------------------- +💾 Converting with 71.5637 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-434.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-434.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 311, 128) +Output shape: (1, 311, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) -> torch.Size([1, 1, 311, 512]) + layer.4.output: torch.Size([1, 311, 3584]) -> torch.Size([1, 1, 311, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,680B, BPFP=0.6873 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,028B, BPFP=2.9656 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,748B, BPFP=1.4946 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,356B, BPFP=2.8816 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,904B, BPFP=1.6029 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,240B, BPFP=2.8256 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,900B, BPFP=1.5022 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,644B, BPFP=2.8961 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,212B, BPFP=2.3217 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,532B, BPFP=2.7900 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,636B, BPFP=0.7079 +⌛️ [2/4] FRONTEND: Frontend time: 1.875s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.378s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 311, 128]) + layer.0.v_cache: torch.Size([1, 4, 311, 128]) + layer.1.k_cache: torch.Size([1, 4, 311, 128]) + layer.1.v_cache: torch.Size([1, 4, 311, 128]) + layer.2.k_cache: torch.Size([1, 4, 311, 128]) + layer.2.v_cache: torch.Size([1, 4, 311, 128]) + layer.3.k_cache: torch.Size([1, 4, 311, 128]) + layer.3.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.k_cache: torch.Size([1, 4, 311, 128]) + layer.4.v_cache: torch.Size([1, 4, 311, 128]) + layer.4.output: torch.Size([1, 311, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14227538 30.28886468 + layer.0.v_cache 0.00001500 0.00187071 + layer.1.k_cache 0.64922478 3.19800040 + layer.1.v_cache 0.00000572 0.00081790 + layer.2.k_cache 0.00828172 0.59526101 + layer.2.v_cache 0.00001922 0.00256855 + layer.3.k_cache 0.02636586 4.41468031 + layer.3.v_cache 0.00001996 0.00277291 + layer.4.k_cache 0.00069494 0.07537964 + layer.4.v_cache 0.00005248 0.00567234 + layer.4.output 0.04686997 73.93058825 + ------------------------------------------------------------------------------------- + TOTAL 0.06794382 32.71176507 + (elements=2,706,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2706944 +Total Bytes 535880 +BPFP 1.5837 bits/point +EBPFP 3.1674 equivalent bits/point +MSE 32.711765 +---------------------- -------------------------------------------------------- +Time: 3.264s Load: 0.010s, Pack+Encode: 1.875s, Decode+Unpack: 1.378s +---------------------- -------------------------------------------------------- +💾 Converting with 32.7118 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-440.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,936B, BPFP=0.6701 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,288B, BPFP=2.8975 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,328B, BPFP=1.6745 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,632B, BPFP=2.8574 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,512B, BPFP=1.5632 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,964B, BPFP=2.7551 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,640B, BPFP=1.4485 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,916B, BPFP=2.8135 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 37,136B, BPFP=2.2755 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,592B, BPFP=2.7324 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,756B, BPFP=0.7419 +⌛️ [2/4] FRONTEND: Frontend time: 1.988s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.272s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13149661 35.18461244 + layer.0.v_cache 0.00001370 0.00188499 + layer.1.k_cache 0.50214604 3.19936140 + layer.1.v_cache 0.00000625 0.00082207 + layer.2.k_cache 0.00807812 0.53468436 + layer.2.v_cache 0.00001969 0.00265733 + layer.3.k_cache 0.06061829 4.41544118 + layer.3.v_cache 0.00001868 0.00281301 + layer.4.k_cache 0.00068906 0.07351154 + layer.4.v_cache 0.00005044 0.00558500 + layer.4.output 1.20679682 191.85544468 + ------------------------------------------------------------------------------------- + TOTAL 0.53827733 81.55349918 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 438700 +BPFP 1.5812 bits/point +EBPFP 3.1625 equivalent bits/point +MSE 81.553499 +---------------------- -------------------------------------------------------- +Time: 3.270s Load: 0.010s, Pack+Encode: 1.988s, Decode+Unpack: 1.272s +---------------------- -------------------------------------------------------- +💾 Converting with 81.5535 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-442.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 300, 128) +Output shape: (1, 300, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) -> torch.Size([1, 1, 300, 512]) + layer.4.output: torch.Size([1, 300, 3584]) -> torch.Size([1, 1, 300, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,388B, BPFP=0.6973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,828B, BPFP=3.0640 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,768B, BPFP=1.4983 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,124B, BPFP=2.9752 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,408B, BPFP=1.6358 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,872B, BPFP=2.9100 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,104B, BPFP=1.5158 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,392B, BPFP=2.9892 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,936B, BPFP=2.3925 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,232B, BPFP=2.8767 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,832B, BPFP=0.7130 +⌛️ [2/4] FRONTEND: Frontend time: 1.889s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.384s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 300, 128]) + layer.0.v_cache: torch.Size([1, 4, 300, 128]) + layer.1.k_cache: torch.Size([1, 4, 300, 128]) + layer.1.v_cache: torch.Size([1, 4, 300, 128]) + layer.2.k_cache: torch.Size([1, 4, 300, 128]) + layer.2.v_cache: torch.Size([1, 4, 300, 128]) + layer.3.k_cache: torch.Size([1, 4, 300, 128]) + layer.3.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.k_cache: torch.Size([1, 4, 300, 128]) + layer.4.v_cache: torch.Size([1, 4, 300, 128]) + layer.4.output: torch.Size([1, 300, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13491969 31.14322266 + layer.0.v_cache 0.00001558 0.00181770 + layer.1.k_cache 0.64590566 3.17249634 + layer.1.v_cache 0.00000556 0.00078813 + layer.2.k_cache 0.01395306 0.59408198 + layer.2.v_cache 0.00001972 0.00253986 + layer.3.k_cache 0.02379198 4.48078979 + layer.3.v_cache 0.00001942 0.00277545 + layer.4.k_cache 0.00073123 0.07511351 + layer.4.v_cache 0.00005274 0.00548216 + layer.4.output 0.05011183 164.99287202 + ------------------------------------------------------------------------------------- + TOTAL 0.06883514 70.26054187 + (elements=2,611,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2611200 +Total Bytes 528884 +BPFP 1.6204 bits/point +EBPFP 3.2407 equivalent bits/point +MSE 70.260542 +---------------------- -------------------------------------------------------- +Time: 3.285s Load: 0.011s, Pack+Encode: 1.889s, Decode+Unpack: 1.384s +---------------------- -------------------------------------------------------- +💾 Converting with 70.2605 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-453.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 291, 128) +Output shape: (1, 291, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) -> torch.Size([1, 1, 291, 512]) + layer.4.output: torch.Size([1, 291, 3584]) -> torch.Size([1, 1, 291, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,932B, BPFP=0.6944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,888B, BPFP=3.1619 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,340B, BPFP=1.5217 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,928B, BPFP=3.0567 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,608B, BPFP=1.6435 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,196B, BPFP=2.9637 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,452B, BPFP=1.4740 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,748B, BPFP=3.0470 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,848B, BPFP=2.4081 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,424B, BPFP=2.9759 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,560B, BPFP=0.7100 +⌛️ [2/4] FRONTEND: Frontend time: 1.883s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 291, 128]) + layer.0.v_cache: torch.Size([1, 4, 291, 128]) + layer.1.k_cache: torch.Size([1, 4, 291, 128]) + layer.1.v_cache: torch.Size([1, 4, 291, 128]) + layer.2.k_cache: torch.Size([1, 4, 291, 128]) + layer.2.v_cache: torch.Size([1, 4, 291, 128]) + layer.3.k_cache: torch.Size([1, 4, 291, 128]) + layer.3.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.k_cache: torch.Size([1, 4, 291, 128]) + layer.4.v_cache: torch.Size([1, 4, 291, 128]) + layer.4.output: torch.Size([1, 291, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14826349 30.62413754 + layer.0.v_cache 0.00001415 0.00189879 + layer.1.k_cache 0.61440683 3.06360534 + layer.1.v_cache 0.00000562 0.00082307 + layer.2.k_cache 0.00551110 0.67231053 + layer.2.v_cache 0.00001912 0.00254020 + layer.3.k_cache 0.01294636 4.37041922 + layer.3.v_cache 0.00002032 0.00272500 + layer.4.k_cache 0.00070633 0.07556059 + layer.4.v_cache 0.00005074 0.00569757 + layer.4.output 0.04972342 154.01511107 + ------------------------------------------------------------------------------------- + TOTAL 0.06647106 65.70149973 + (elements=2,532,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2532864 +Total Bytes 519924 +BPFP 1.6422 bits/point +EBPFP 3.2843 equivalent bits/point +MSE 65.701500 +---------------------- -------------------------------------------------------- +Time: 3.281s Load: 0.010s, Pack+Encode: 1.883s, Decode+Unpack: 1.388s +---------------------- -------------------------------------------------------- +💾 Converting with 65.7015 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-47.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-47.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,544B, BPFP=0.7076 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,900B, BPFP=3.2096 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,684B, BPFP=1.6180 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,936B, BPFP=3.1552 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,328B, BPFP=1.6543 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,120B, BPFP=3.0528 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,064B, BPFP=1.5266 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,376B, BPFP=3.1801 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,392B, BPFP=2.4477 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,744B, BPFP=3.0880 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,312B, BPFP=0.6955 +⌛️ [2/4] FRONTEND: Frontend time: 1.883s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.376s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13861307 29.60306789 + layer.0.v_cache 0.00001419 0.00188068 + layer.1.k_cache 0.53867822 3.02175958 + layer.1.v_cache 0.00000573 0.00081516 + layer.2.k_cache 0.00871976 0.68998013 + layer.2.v_cache 0.00001877 0.00261696 + layer.3.k_cache 0.02195022 3.94664646 + layer.3.v_cache 0.00001967 0.00295034 + layer.4.k_cache 0.00069227 0.07518631 + layer.4.v_cache 0.00004961 0.00574876 + layer.4.output 0.00954030 177.55447396 + ------------------------------------------------------------------------------------- + TOTAL 0.04562021 75.30776294 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 505400 +BPFP 1.6770 bits/point +EBPFP 3.3539 equivalent bits/point +MSE 75.307763 +---------------------- -------------------------------------------------------- +Time: 3.270s Load: 0.011s, Pack+Encode: 1.883s, Decode+Unpack: 1.376s +---------------------- -------------------------------------------------------- +💾 Converting with 75.3078 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-5.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,596B, BPFP=0.6882 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,924B, BPFP=3.1646 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,420B, BPFP=1.4980 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,592B, BPFP=3.0918 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,400B, BPFP=1.6608 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,048B, BPFP=3.0074 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,628B, BPFP=1.5094 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,896B, BPFP=3.1084 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,404B, BPFP=2.4259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,204B, BPFP=3.0160 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,288B, BPFP=0.6969 +⌛️ [2/4] FRONTEND: Frontend time: 1.881s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.399s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13928204 31.57397290 + layer.0.v_cache 0.00001373 0.00187772 + layer.1.k_cache 0.60860016 2.90930368 + layer.1.v_cache 0.00000555 0.00079658 + layer.2.k_cache 0.01026906 0.72574279 + layer.2.v_cache 0.00001872 0.00276252 + layer.3.k_cache 0.05166252 4.08806989 + layer.3.v_cache 0.00001937 0.00287287 + layer.4.k_cache 0.00073448 0.07638880 + layer.4.v_cache 0.00005026 0.00578809 + layer.4.output 0.00697322 178.19862950 + ------------------------------------------------------------------------------------- + TOTAL 0.05055697 75.69282249 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 513400 +BPFP 1.6499 bits/point +EBPFP 3.2998 equivalent bits/point +MSE 75.692822 +---------------------- -------------------------------------------------------- +Time: 3.290s Load: 0.010s, Pack+Encode: 1.881s, Decode+Unpack: 1.399s +---------------------- -------------------------------------------------------- +💾 Converting with 75.6928 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-61.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 295, 128) +Output shape: (1, 295, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) -> torch.Size([1, 1, 295, 512]) + layer.4.output: torch.Size([1, 295, 3584]) -> torch.Size([1, 1, 295, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,204B, BPFP=0.6994 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,672B, BPFP=3.1076 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,592B, BPFP=1.4614 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,132B, BPFP=3.0261 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,380B, BPFP=1.6091 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,100B, BPFP=2.9714 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,712B, BPFP=1.4678 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,404B, BPFP=3.0405 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,424B, BPFP=2.4059 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,272B, BPFP=2.9805 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,556B, BPFP=0.7306 +⌛️ [2/4] FRONTEND: Frontend time: 1.887s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 295, 128]) + layer.0.v_cache: torch.Size([1, 4, 295, 128]) + layer.1.k_cache: torch.Size([1, 4, 295, 128]) + layer.1.v_cache: torch.Size([1, 4, 295, 128]) + layer.2.k_cache: torch.Size([1, 4, 295, 128]) + layer.2.v_cache: torch.Size([1, 4, 295, 128]) + layer.3.k_cache: torch.Size([1, 4, 295, 128]) + layer.3.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.k_cache: torch.Size([1, 4, 295, 128]) + layer.4.v_cache: torch.Size([1, 4, 295, 128]) + layer.4.output: torch.Size([1, 295, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13820392 30.56194055 + layer.0.v_cache 0.00001454 0.00187519 + layer.1.k_cache 0.61684219 2.83025564 + layer.1.v_cache 0.00000570 0.00079058 + layer.2.k_cache 0.01285210 0.63453245 + layer.2.v_cache 0.00002011 0.00263159 + layer.3.k_cache 0.05980752 4.13937202 + layer.3.v_cache 0.00002263 0.00287215 + layer.4.k_cache 0.00071023 0.07318806 + layer.4.v_cache 0.00005833 0.00582798 + layer.4.output 0.05086143 154.10799031 + ------------------------------------------------------------------------------------- + TOTAL 0.06968043 65.70642461 + (elements=2,567,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2567680 +Total Bytes 526448 +BPFP 1.6402 bits/point +EBPFP 3.2805 equivalent bits/point +MSE 65.706425 +---------------------- -------------------------------------------------------- +Time: 3.286s Load: 0.012s, Pack+Encode: 1.887s, Decode+Unpack: 1.387s +---------------------- -------------------------------------------------------- +💾 Converting with 65.7064 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-62.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-62.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,360B, BPFP=0.6935 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,940B, BPFP=3.0596 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,832B, BPFP=1.6005 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,380B, BPFP=2.9786 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,856B, BPFP=1.6537 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,208B, BPFP=2.9178 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,516B, BPFP=1.5322 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,556B, BPFP=2.9877 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,268B, BPFP=2.4018 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,492B, BPFP=2.8806 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 104,092B, BPFP=0.7719 +⌛️ [2/4] FRONTEND: Frontend time: 1.892s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.388s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14825445 31.42394492 + layer.0.v_cache 0.00001396 0.00193117 + layer.1.k_cache 0.63210902 3.19334396 + layer.1.v_cache 0.00000581 0.00084851 + layer.2.k_cache 0.01184132 0.65689822 + layer.2.v_cache 0.00001890 0.00258791 + layer.3.k_cache 0.01767834 4.21288130 + layer.3.v_cache 0.00002009 0.00274496 + layer.4.k_cache 0.00068272 0.07494321 + layer.4.v_cache 0.00005247 0.00578884 + layer.4.output 0.04867053 156.46659943 + ------------------------------------------------------------------------------------- + TOTAL 0.06772769 66.75541818 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 541500 +BPFP 1.6535 bits/point +EBPFP 3.3070 equivalent bits/point +MSE 66.755418 +---------------------- -------------------------------------------------------- +Time: 3.290s Load: 0.010s, Pack+Encode: 1.892s, Decode+Unpack: 1.388s +---------------------- -------------------------------------------------------- +💾 Converting with 66.7554 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-63.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,032B, BPFP=0.7121 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,848B, BPFP=3.3646 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,300B, BPFP=1.3790 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 54,404B, BPFP=3.2199 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,552B, BPFP=1.6899 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 53,132B, BPFP=3.1446 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,892B, BPFP=1.5324 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,004B, BPFP=3.2554 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,944B, BPFP=2.4825 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,424B, BPFP=3.1619 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,836B, BPFP=0.6919 +⌛️ [2/4] FRONTEND: Frontend time: 1.898s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.400s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14315909 29.14047704 + layer.0.v_cache 0.00001391 0.00196182 + layer.1.k_cache 0.49711875 2.33127155 + layer.1.v_cache 0.00000560 0.00085763 + layer.2.k_cache 0.00646762 0.79045908 + layer.2.v_cache 0.00001905 0.00267315 + layer.3.k_cache 0.04707547 3.72028790 + layer.3.v_cache 0.00001924 0.00292046 + layer.4.k_cache 0.00071833 0.07504350 + layer.4.v_cache 0.00005197 0.00571094 + layer.4.output 0.00997522 172.42571361 + ------------------------------------------------------------------------------------- + TOTAL 0.04496915 73.12068578 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 486368 +BPFP 1.6933 bits/point +EBPFP 3.3866 equivalent bits/point +MSE 73.120686 +---------------------- -------------------------------------------------------- +Time: 3.308s Load: 0.009s, Pack+Encode: 1.898s, Decode+Unpack: 1.400s +---------------------- -------------------------------------------------------- +💾 Converting with 73.1207 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-66.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-66.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,644B, BPFP=0.7081 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,764B, BPFP=3.2350 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,516B, BPFP=1.6530 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,932B, BPFP=3.1324 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,040B, BPFP=1.6823 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,140B, BPFP=3.0880 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,020B, BPFP=1.5132 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,644B, BPFP=3.1723 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,884B, BPFP=2.4577 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,368B, BPFP=3.1008 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,112B, BPFP=0.7129 +⌛️ [2/4] FRONTEND: Frontend time: 1.890s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.395s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15581964 29.53895049 + layer.0.v_cache 0.00001703 0.00190300 + layer.1.k_cache 0.56298232 2.85368342 + layer.1.v_cache 0.00000550 0.00079922 + layer.2.k_cache 0.00979712 0.63324368 + layer.2.v_cache 0.00002023 0.00268030 + layer.3.k_cache 0.04960944 3.62088724 + layer.3.v_cache 0.00001971 0.00286152 + layer.4.k_cache 0.00068561 0.07388431 + layer.4.v_cache 0.00005245 0.00575913 + layer.4.output 0.01122868 180.05435548 + ------------------------------------------------------------------------------------- + TOTAL 0.05044764 76.30089063 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 513064 +BPFP 1.6902 bits/point +EBPFP 3.3804 equivalent bits/point +MSE 76.300891 +---------------------- -------------------------------------------------------- +Time: 3.295s Load: 0.010s, Pack+Encode: 1.890s, Decode+Unpack: 1.395s +---------------------- -------------------------------------------------------- +💾 Converting with 76.3009 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-75.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,596B, BPFP=0.6955 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,276B, BPFP=3.1623 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,528B, BPFP=1.5199 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,444B, BPFP=3.0612 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,344B, BPFP=1.6201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,304B, BPFP=2.9982 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,988B, BPFP=1.4901 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,576B, BPFP=3.0685 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,840B, BPFP=2.4205 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,232B, BPFP=2.9943 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,992B, BPFP=0.7256 +⌛️ [2/4] FRONTEND: Frontend time: 1.882s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.406s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13599485 30.59145525 + layer.0.v_cache 0.00001418 0.00190415 + layer.1.k_cache 0.56373192 2.80494372 + layer.1.v_cache 0.00000588 0.00082976 + layer.2.k_cache 0.01549433 0.71370241 + layer.2.v_cache 0.00001867 0.00267971 + layer.3.k_cache 0.03475572 3.74102675 + layer.3.v_cache 0.00001907 0.00281696 + layer.4.k_cache 0.00072111 0.07506966 + layer.4.v_cache 0.00005295 0.00575342 + layer.4.output 0.00820773 167.99600896 + ------------------------------------------------------------------------------------- + TOTAL 0.04754487 71.40660262 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 509120 +BPFP 1.6535 bits/point +EBPFP 3.3070 equivalent bits/point +MSE 71.406603 +---------------------- -------------------------------------------------------- +Time: 3.297s Load: 0.010s, Pack+Encode: 1.882s, Decode+Unpack: 1.406s +---------------------- -------------------------------------------------------- +💾 Converting with 71.4066 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-76.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-76.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,532B, BPFP=0.6865 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,832B, BPFP=2.9846 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,908B, BPFP=1.5172 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,196B, BPFP=2.9016 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,012B, BPFP=1.6240 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,316B, BPFP=2.8569 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,688B, BPFP=1.5061 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,556B, BPFP=2.9198 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,088B, BPFP=2.3381 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,820B, BPFP=2.8318 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,092B, BPFP=0.7109 +⌛️ [2/4] FRONTEND: Frontend time: 1.896s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13331154 30.24677861 + layer.0.v_cache 0.00001460 0.00187854 + layer.1.k_cache 0.64535681 3.11392014 + layer.1.v_cache 0.00000563 0.00081011 + layer.2.k_cache 0.00762887 0.62592955 + layer.2.v_cache 0.00001910 0.00263679 + layer.3.k_cache 0.03538683 4.41648092 + layer.3.v_cache 0.00002103 0.00282260 + layer.4.k_cache 0.00072257 0.07442055 + layer.4.v_cache 0.00005300 0.00568607 + layer.4.output 0.04522907 110.15976490 + ------------------------------------------------------------------------------------- + TOTAL 0.06700726 47.62410107 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 535040 +BPFP 1.5966 bits/point +EBPFP 3.1933 equivalent bits/point +MSE 47.624101 +---------------------- -------------------------------------------------------- +Time: 3.297s Load: 0.012s, Pack+Encode: 1.896s, Decode+Unpack: 1.390s +---------------------- -------------------------------------------------------- +💾 Converting with 47.6241 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-79.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-79.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 284, 128) +Output shape: (1, 284, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) -> torch.Size([1, 1, 284, 512]) + layer.4.output: torch.Size([1, 284, 3584]) -> torch.Size([1, 1, 284, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,692B, BPFP=0.6983 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,496B, BPFP=3.1633 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,488B, BPFP=1.6224 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,044B, BPFP=3.0834 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,672B, BPFP=1.6325 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,768B, BPFP=3.0132 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,876B, BPFP=1.4787 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,272B, BPFP=3.0960 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,160B, BPFP=2.4296 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,624B, BPFP=3.0603 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,796B, BPFP=0.7215 +⌛️ [2/4] FRONTEND: Frontend time: 1.885s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.381s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 284, 128]) + layer.0.v_cache: torch.Size([1, 4, 284, 128]) + layer.1.k_cache: torch.Size([1, 4, 284, 128]) + layer.1.v_cache: torch.Size([1, 4, 284, 128]) + layer.2.k_cache: torch.Size([1, 4, 284, 128]) + layer.2.v_cache: torch.Size([1, 4, 284, 128]) + layer.3.k_cache: torch.Size([1, 4, 284, 128]) + layer.3.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.k_cache: torch.Size([1, 4, 284, 128]) + layer.4.v_cache: torch.Size([1, 4, 284, 128]) + layer.4.output: torch.Size([1, 284, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14806804 29.62516505 + layer.0.v_cache 0.00001376 0.00187422 + layer.1.k_cache 0.60662186 2.95682407 + layer.1.v_cache 0.00000571 0.00079803 + layer.2.k_cache 0.01239752 0.66442232 + layer.2.v_cache 0.00002027 0.00263674 + layer.3.k_cache 0.02456893 3.88571124 + layer.3.v_cache 0.00001975 0.00273368 + layer.4.k_cache 0.00069718 0.07549114 + layer.4.v_cache 0.00005379 0.00556996 + layer.4.output 0.00818182 173.34316839 + ------------------------------------------------------------------------------------- + TOTAL 0.04998468 73.56608266 + (elements=2,471,936) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2471936 +Total Bytes 514888 +BPFP 1.6663 bits/point +EBPFP 3.3327 equivalent bits/point +MSE 73.566083 +---------------------- -------------------------------------------------------- +Time: 3.277s Load: 0.010s, Pack+Encode: 1.885s, Decode+Unpack: 1.381s +---------------------- -------------------------------------------------------- +💾 Converting with 73.5661 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-83.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-83.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,684B, BPFP=0.7003 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,552B, BPFP=3.1776 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,388B, BPFP=1.6778 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,764B, BPFP=3.0788 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,168B, BPFP=1.6656 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,740B, BPFP=3.0223 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,308B, BPFP=1.5077 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,332B, BPFP=3.1102 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 44,332B, BPFP=2.4477 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,324B, BPFP=3.0545 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,524B, BPFP=0.7377 +⌛️ [2/4] FRONTEND: Frontend time: 1.894s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.382s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14592592 28.65717825 + layer.0.v_cache 0.00001391 0.00187763 + layer.1.k_cache 0.53254910 3.04303302 + layer.1.v_cache 0.00000583 0.00083662 + layer.2.k_cache 0.01029040 0.66781848 + layer.2.v_cache 0.00002009 0.00268717 + layer.3.k_cache 0.03848921 4.01278762 + layer.3.v_cache 0.00001972 0.00281635 + layer.4.k_cache 0.00069417 0.07594002 + layer.4.v_cache 0.00005427 0.00591791 + layer.4.output 0.01004909 181.39099571 + ------------------------------------------------------------------------------------- + TOTAL 0.04696507 76.83575665 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 518116 +BPFP 1.6827 bits/point +EBPFP 3.3654 equivalent bits/point +MSE 76.835757 +---------------------- -------------------------------------------------------- +Time: 3.285s Load: 0.009s, Pack+Encode: 1.894s, Decode+Unpack: 1.382s +---------------------- -------------------------------------------------------- +💾 Converting with 76.8358 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-87.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-87.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,384B, BPFP=0.7114 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,444B, BPFP=3.2999 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,624B, BPFP=1.6443 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,380B, BPFP=3.1813 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,872B, BPFP=1.6585 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,400B, BPFP=3.1250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,108B, BPFP=1.4998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,816B, BPFP=3.2063 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,496B, BPFP=2.4412 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,252B, BPFP=3.1165 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,064B, BPFP=0.7063 +⌛️ [2/4] FRONTEND: Frontend time: 1.881s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.390s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15384994 30.00995232 + layer.0.v_cache 0.00001404 0.00196241 + layer.1.k_cache 0.48768111 3.14705748 + layer.1.v_cache 0.00000584 0.00086102 + layer.2.k_cache 0.00538594 0.78617500 + layer.2.v_cache 0.00001948 0.00269303 + layer.3.k_cache 0.02347840 3.87494435 + layer.3.v_cache 0.00001899 0.00273347 + layer.4.k_cache 0.00068094 0.07458580 + layer.4.v_cache 0.00005116 0.00570545 + layer.4.output 0.00909345 181.35221901 + ------------------------------------------------------------------------------------- + TOTAL 0.04322588 76.90424726 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 501840 +BPFP 1.6958 bits/point +EBPFP 3.3915 equivalent bits/point +MSE 76.904247 +---------------------- -------------------------------------------------------- +Time: 3.280s Load: 0.009s, Pack+Encode: 1.881s, Decode+Unpack: 1.390s +---------------------- -------------------------------------------------------- +💾 Converting with 76.9042 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-94.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-94.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,092B, BPFP=0.6958 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,420B, BPFP=3.1048 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,616B, BPFP=1.5208 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,164B, BPFP=3.0381 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,440B, BPFP=1.6709 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,112B, BPFP=2.9821 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,840B, BPFP=1.5327 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,056B, BPFP=3.0323 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 45,352B, BPFP=2.4103 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,316B, BPFP=2.9398 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,568B, BPFP=0.6952 +⌛️ [2/4] FRONTEND: Frontend time: 1.908s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.397s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13453574 30.55662402 + layer.0.v_cache 0.00001386 0.00186879 + layer.1.k_cache 0.70462929 3.07249648 + layer.1.v_cache 0.00000580 0.00081812 + layer.2.k_cache 0.00896267 0.63700119 + layer.2.v_cache 0.00001881 0.00258028 + layer.3.k_cache 0.02468209 4.18952537 + layer.3.v_cache 0.00001940 0.00271755 + layer.4.k_cache 0.00068809 0.07512961 + layer.4.v_cache 0.00005209 0.00559549 + layer.4.output 0.04953035 164.22998664 + ------------------------------------------------------------------------------------- + TOTAL 0.07178355 69.89142726 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 522976 +BPFP 1.6350 bits/point +EBPFP 3.2699 equivalent bits/point +MSE 69.891427 +---------------------- -------------------------------------------------------- +Time: 3.315s Load: 0.010s, Pack+Encode: 1.908s, Decode+Unpack: 1.397s +---------------------- -------------------------------------------------------- +💾 Converting with 69.8914 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-97.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-97.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,288B, BPFP=0.7164 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,132B, BPFP=3.3309 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,436B, BPFP=1.4247 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,380B, BPFP=3.2288 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,796B, BPFP=1.6789 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,232B, BPFP=3.1618 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,864B, BPFP=1.5079 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,200B, BPFP=3.2766 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,732B, BPFP=2.4914 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,104B, BPFP=3.1544 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,984B, BPFP=0.7162 +⌛️ [2/4] FRONTEND: Frontend time: 34.987s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.389s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150386 30.65658524 + layer.0.v_cache 0.00001401 0.00188574 + layer.1.k_cache 0.57442873 2.73976796 + layer.1.v_cache 0.00000554 0.00079629 + layer.2.k_cache 0.00779028 0.69971352 + layer.2.v_cache 0.00001875 0.00269226 + layer.3.k_cache 0.05176373 3.52521207 + layer.3.v_cache 0.00002027 0.00289693 + layer.4.k_cache 0.00069872 0.07304241 + layer.4.v_cache 0.00005023 0.00553929 + layer.4.output 0.01044900 182.46628465 + ------------------------------------------------------------------------------------- + TOTAL 0.04937866 77.35130143 + (elements=2,332,672) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 2332672 +Total Bytes 497148 +BPFP 1.7050 bits/point +EBPFP 3.4100 equivalent bits/point +MSE 77.351301 +---------------------- --------------------------------------------------------- +Time: 36.385s Load: 0.009s, Pack+Encode: 34.987s, Decode+Unpack: 1.389s +---------------------- --------------------------------------------------------- +💾 Converting with 77.3513 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/openbookqa/OpenBookQA-openbookqa-99.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/openbookqa/OpenBookQA-openbookqa-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.6626 bits/point +Avg EBPFP 3.3251 equivalent bits/point +Avg MSE 69.328181 +Avg Time 5.200s +------------------------ ---------------------------- diff --git a/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log b/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..c0e77e92896d8797492590f8baded2fa83d2d6de --- /dev/null +++ b/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_elic-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/dtufc_elic-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: elic-featurecoding + handler: kimiaudio + checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 520 +Loaded elic-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +Checkpoint codec_weights/elic_hybrid/elic2022-official_lambda0.02_epochs600_lr0.0001_bs60_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa +Output output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa +---------------- ------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,840B, BPFP=0.9003 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,752B, BPFP=3.6741 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,764B, BPFP=1.4442 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,588B, BPFP=3.4576 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,896B, BPFP=1.6548 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,096B, BPFP=3.1801 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,304B, BPFP=1.5446 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,236B, BPFP=3.3921 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,640B, BPFP=2.5372 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,980B, BPFP=3.3445 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,300B, BPFP=0.8317 +⌛️ [2/4] FRONTEND: Frontend time: 2.150s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.046s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14073507 26.01675851 + layer.0.v_cache 0.00001426 0.00218895 + layer.1.k_cache 0.05042761 2.01155744 + layer.1.v_cache 0.00000517 0.00078670 + layer.2.k_cache 0.00244098 0.64953127 + layer.2.v_cache 0.00001715 0.00241894 + layer.3.k_cache 0.02726505 2.79899742 + layer.3.v_cache 0.00001796 0.00271920 + layer.4.k_cache 0.00069123 0.07498468 + layer.4.v_cache 0.00005084 0.00614428 + layer.4.output 0.16186065 551.67788053 + ------------------------------------------------------------------------------------- + TOTAL 0.07968764 229.01830889 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 166396 +BPFP 1.8207 bits/point +EBPFP 3.6414 equivalent bits/point +MSE 229.018309 +---------------------- -------------------------------------------------------- +Time: 3.201s Load: 0.004s, Pack+Encode: 2.150s, Decode+Unpack: 1.046s +---------------------- -------------------------------------------------------- +💾 Converting with 229.0183 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-1.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,748B, BPFP=0.8938 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,344B, BPFP=3.6416 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,604B, BPFP=1.4315 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,716B, BPFP=3.3351 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,392B, BPFP=1.5798 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,208B, BPFP=3.0512 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,936B, BPFP=1.4940 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,496B, BPFP=3.2937 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,784B, BPFP=2.4066 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,448B, BPFP=3.0964 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,732B, BPFP=0.8265 +⌛️ [2/4] FRONTEND: Frontend time: 1.513s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.029s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11392649 26.87976515 + layer.0.v_cache 0.00001670 0.00218790 + layer.1.k_cache 0.05249626 1.74171282 + layer.1.v_cache 0.00000506 0.00079515 + layer.2.k_cache 0.00416583 0.71715812 + layer.2.v_cache 0.00001711 0.00240040 + layer.3.k_cache 0.05931453 2.90805936 + layer.3.v_cache 0.00001789 0.00263284 + layer.4.k_cache 0.00069248 0.07029540 + layer.4.v_cache 0.00004729 0.00611721 + layer.4.output 0.16383441 574.99499785 + ------------------------------------------------------------------------------------- + TOTAL 0.08103179 238.66447702 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 159408 +BPFP 1.7652 bits/point +EBPFP 3.5305 equivalent bits/point +MSE 238.664477 +---------------------- -------------------------------------------------------- +Time: 2.547s Load: 0.005s, Pack+Encode: 1.513s, Decode+Unpack: 1.029s +---------------------- -------------------------------------------------------- +💾 Converting with 238.6645 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-10.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-10.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,964B, BPFP=0.8251 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,372B, BPFP=3.3863 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,100B, BPFP=1.3464 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,536B, BPFP=3.2473 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,240B, BPFP=1.5359 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,884B, BPFP=3.1390 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,572B, BPFP=1.4249 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,584B, BPFP=3.2553 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,004B, BPFP=2.4940 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,980B, BPFP=3.1549 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,648B, BPFP=0.7278 +⌛️ [2/4] FRONTEND: Frontend time: 1.508s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.027s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17268721 26.54314536 + layer.0.v_cache 0.00001453 0.00220462 + layer.1.k_cache 0.08276152 1.97179039 + layer.1.v_cache 0.00000545 0.00080070 + layer.2.k_cache 0.00739249 0.68893473 + layer.2.v_cache 0.00001625 0.00239246 + layer.3.k_cache 0.05929063 3.50775082 + layer.3.v_cache 0.00001766 0.00282095 + layer.4.k_cache 0.00068371 0.08028790 + layer.4.v_cache 0.00005055 0.00630425 + layer.4.output 0.14466690 518.43583777 + ------------------------------------------------------------------------------------- + TOTAL 0.07856402 215.40337039 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 173884 +BPFP 1.7002 bits/point +EBPFP 3.4004 equivalent bits/point +MSE 215.403370 +---------------------- -------------------------------------------------------- +Time: 2.539s Load: 0.004s, Pack+Encode: 1.508s, Decode+Unpack: 1.027s +---------------------- -------------------------------------------------------- +💾 Converting with 215.4034 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-102.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-102.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 112, 128) +Output shape: (1, 112, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) -> torch.Size([1, 1, 112, 512]) + layer.4.output: torch.Size([1, 112, 3584]) -> torch.Size([1, 1, 112, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,736B, BPFP=0.8002 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,100B, BPFP=3.2227 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,268B, BPFP=1.4325 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,204B, BPFP=3.0977 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,184B, BPFP=1.6998 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,232B, BPFP=2.9621 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,820B, BPFP=1.5095 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,004B, BPFP=3.0698 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,596B, BPFP=2.4548 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,204B, BPFP=2.9581 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,732B, BPFP=0.8317 +⌛️ [2/4] FRONTEND: Frontend time: 1.522s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.035s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 112, 128]) + layer.0.v_cache: torch.Size([1, 4, 112, 128]) + layer.1.k_cache: torch.Size([1, 4, 112, 128]) + layer.1.v_cache: torch.Size([1, 4, 112, 128]) + layer.2.k_cache: torch.Size([1, 4, 112, 128]) + layer.2.v_cache: torch.Size([1, 4, 112, 128]) + layer.3.k_cache: torch.Size([1, 4, 112, 128]) + layer.3.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.k_cache: torch.Size([1, 4, 112, 128]) + layer.4.v_cache: torch.Size([1, 4, 112, 128]) + layer.4.output: torch.Size([1, 112, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17321212 26.84586661 + layer.0.v_cache 0.00001450 0.00207850 + layer.1.k_cache 0.08643453 2.37692669 + layer.1.v_cache 0.00000659 0.00080849 + layer.2.k_cache 0.00690480 0.59200191 + layer.2.v_cache 0.00001899 0.00249897 + layer.3.k_cache 0.01205242 3.76318196 + layer.3.v_cache 0.00001923 0.00293400 + layer.4.k_cache 0.00071999 0.08522247 + layer.4.v_cache 0.00004947 0.00576630 + layer.4.output 10.17892331 356.24107143 + ------------------------------------------------------------------------------------- + TOTAL 4.20775858 148.66851682 + (elements=974,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 974848 +Total Bytes 208080 +BPFP 1.7076 bits/point +EBPFP 3.4152 equivalent bits/point +MSE 148.668517 +---------------------- -------------------------------------------------------- +Time: 2.562s Load: 0.005s, Pack+Encode: 1.522s, Decode+Unpack: 1.035s +---------------------- -------------------------------------------------------- +💾 Converting with 148.6685 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-106.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,888B, BPFP=0.8679 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,128B, BPFP=3.5739 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,012B, BPFP=1.4226 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,596B, BPFP=3.3018 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,948B, BPFP=1.5888 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,616B, BPFP=3.1278 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,336B, BPFP=1.4801 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,936B, BPFP=3.3622 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,644B, BPFP=2.4226 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,740B, BPFP=3.1499 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,384B, BPFP=0.7707 +⌛️ [2/4] FRONTEND: Frontend time: 1.512s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.028s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11687154 28.33461692 + layer.0.v_cache 0.00001423 0.00215750 + layer.1.k_cache 0.05324238 1.79533091 + layer.1.v_cache 0.00000539 0.00083595 + layer.2.k_cache 0.00397935 0.72458250 + layer.2.v_cache 0.00001766 0.00242812 + layer.3.k_cache 0.01210797 2.73756738 + layer.3.v_cache 0.00001803 0.00284290 + layer.4.k_cache 0.00075037 0.07242042 + layer.4.v_cache 0.00004481 0.00575843 + layer.4.output 0.15448949 535.34669237 + ------------------------------------------------------------------------------------- + TOTAL 0.07461636 222.41796398 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 167228 +BPFP 1.7466 bits/point +EBPFP 3.4932 equivalent bits/point +MSE 222.417964 +---------------------- -------------------------------------------------------- +Time: 2.544s Load: 0.003s, Pack+Encode: 1.512s, Decode+Unpack: 1.028s +---------------------- -------------------------------------------------------- +💾 Converting with 222.4180 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-117.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-117.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,716B, BPFP=0.9097 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,888B, BPFP=3.6435 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,756B, BPFP=1.4961 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,284B, BPFP=3.5270 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,740B, BPFP=1.6860 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,852B, BPFP=3.4437 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,116B, BPFP=1.5656 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,252B, BPFP=3.5208 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,140B, BPFP=2.5347 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,864B, BPFP=3.4460 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,752B, BPFP=0.9026 +⌛️ [2/4] FRONTEND: Frontend time: 1.505s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.028s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14203883 28.29343292 + layer.0.v_cache 0.00001438 0.00212444 + layer.1.k_cache 0.05633916 1.93002866 + layer.1.v_cache 0.00000602 0.00091135 + layer.2.k_cache 0.00244657 0.61036875 + layer.2.v_cache 0.00001847 0.00259020 + layer.3.k_cache 0.05492775 2.90844745 + layer.3.v_cache 0.00001883 0.00288187 + layer.4.k_cache 0.00062984 0.07467732 + layer.4.v_cache 0.00005178 0.00652011 + layer.4.output 0.18538215 610.63938492 + ------------------------------------------------------------------------------------- + TOTAL 0.09142157 253.42986338 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 166360 +BPFP 1.8877 bits/point +EBPFP 3.7754 equivalent bits/point +MSE 253.429863 +---------------------- -------------------------------------------------------- +Time: 2.535s Load: 0.003s, Pack+Encode: 1.505s, Decode+Unpack: 1.028s +---------------------- -------------------------------------------------------- +💾 Converting with 253.4299 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-130.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-130.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,868B, BPFP=0.8451 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,412B, BPFP=3.5438 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,924B, BPFP=1.3757 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,180B, BPFP=3.3299 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,084B, BPFP=1.5771 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,688B, BPFP=3.2444 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,408B, BPFP=1.4597 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,528B, BPFP=3.3903 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,032B, BPFP=2.4361 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,080B, BPFP=3.3125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,208B, BPFP=0.7988 +⌛️ [2/4] FRONTEND: Frontend time: 1.507s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.027s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13028670 28.13809136 + layer.0.v_cache 0.00001462 0.00207108 + layer.1.k_cache 0.04946813 1.75721724 + layer.1.v_cache 0.00000624 0.00079913 + layer.2.k_cache 0.00411291 0.60006332 + layer.2.v_cache 0.00001729 0.00258471 + layer.3.k_cache 0.01793825 3.41398722 + layer.3.v_cache 0.00001895 0.00294439 + layer.4.k_cache 0.00077120 0.07705246 + layer.4.v_cache 0.00004532 0.00620637 + layer.4.output 0.15109633 556.90505952 + ------------------------------------------------------------------------------------- + TOTAL 0.07413847 231.31390788 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 173412 +BPFP 1.7710 bits/point +EBPFP 3.5419 equivalent bits/point +MSE 231.313908 +---------------------- -------------------------------------------------------- +Time: 2.538s Load: 0.004s, Pack+Encode: 1.507s, Decode+Unpack: 1.027s +---------------------- -------------------------------------------------------- +💾 Converting with 231.3139 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-139.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-139.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 34.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,420B, BPFP=0.9375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,136B, BPFP=3.3268 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,848B, BPFP=1.6031 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,680B, BPFP=3.2018 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,408B, BPFP=1.7566 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,332B, BPFP=3.1064 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,224B, BPFP=1.7061 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,652B, BPFP=3.1941 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,048B, BPFP=2.7544 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,196B, BPFP=3.0691 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,808B, BPFP=0.9715 +⌛️ [2/4] FRONTEND: Frontend time: 1.946s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.987s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11683119 25.85314941 + layer.0.v_cache 0.00001514 0.00260410 + layer.1.k_cache 0.01218004 3.01961986 + layer.1.v_cache 0.00000521 0.00097333 + layer.2.k_cache 0.00491684 0.57969953 + layer.2.v_cache 0.00001706 0.00291695 + layer.3.k_cache 0.10823302 3.41159272 + layer.3.v_cache 0.00002043 0.00350148 + layer.4.k_cache 0.00062874 0.09022712 + layer.4.v_cache 0.00004794 0.00720092 + layer.4.output 0.23833110 857.10447995 + ------------------------------------------------------------------------------------- + TOTAL 0.11242431 354.86487324 + (elements=496,128) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 496128 +Total Bytes 114752 +BPFP 1.8504 bits/point +EBPFP 3.7007 equivalent bits/point +MSE 354.864873 +---------------------- --------------------------------------------------------- +Time: 36.948s Load: 34.016s, Pack+Encode: 1.946s, Decode+Unpack: 0.987s +---------------------- --------------------------------------------------------- +💾 Converting with 354.8649 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-143.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,316B, BPFP=1.0362 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,924B, BPFP=3.7262 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,876B, BPFP=1.8362 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,496B, BPFP=3.5925 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,604B, BPFP=2.0638 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,296B, BPFP=3.5300 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,876B, BPFP=1.8362 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,504B, BPFP=3.5950 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,376B, BPFP=2.9300 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,100B, BPFP=3.4688 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,148B, BPFP=1.0334 +⌛️ [2/4] FRONTEND: Frontend time: 1.508s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.981s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14598725 26.76259521 + layer.0.v_cache 0.00001449 0.00270070 + layer.1.k_cache 0.01347294 2.80609222 + layer.1.v_cache 0.00000521 0.00098027 + layer.2.k_cache 0.00488209 0.59651741 + layer.2.v_cache 0.00001649 0.00281500 + layer.3.k_cache 0.04421538 4.19194824 + layer.3.v_cache 0.00001905 0.00360204 + layer.4.k_cache 0.00073779 0.09470222 + layer.4.v_cache 0.00005089 0.00727545 + layer.4.output 0.27158585 970.55785714 + ------------------------------------------------------------------------------------- + TOTAL 0.12414721 401.66907228 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 111516 +BPFP 2.0499 bits/point +EBPFP 4.0999 equivalent bits/point +MSE 401.669072 +---------------------- -------------------------------------------------------- +Time: 2.492s Load: 0.003s, Pack+Encode: 1.508s, Decode+Unpack: 0.981s +---------------------- -------------------------------------------------------- +💾 Converting with 401.6691 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-145.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-145.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 46, 128) +Output shape: (1, 46, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) -> torch.Size([1, 1, 46, 512]) + layer.4.output: torch.Size([1, 46, 3584]) -> torch.Size([1, 1, 46, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,224B, BPFP=1.0951 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,996B, BPFP=4.0747 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,208B, BPFP=1.7690 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,652B, BPFP=3.9579 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,188B, BPFP=2.1019 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,320B, BPFP=3.8451 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,404B, BPFP=1.8356 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,684B, BPFP=3.9688 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,460B, BPFP=3.2133 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,196B, BPFP=3.8030 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,916B, BPFP=1.0635 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.985s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 46, 128]) + layer.0.v_cache: torch.Size([1, 4, 46, 128]) + layer.1.k_cache: torch.Size([1, 4, 46, 128]) + layer.1.v_cache: torch.Size([1, 4, 46, 128]) + layer.2.k_cache: torch.Size([1, 4, 46, 128]) + layer.2.v_cache: torch.Size([1, 4, 46, 128]) + layer.3.k_cache: torch.Size([1, 4, 46, 128]) + layer.3.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.k_cache: torch.Size([1, 4, 46, 128]) + layer.4.v_cache: torch.Size([1, 4, 46, 128]) + layer.4.output: torch.Size([1, 46, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14129612 25.50719684 + layer.0.v_cache 0.00001532 0.00282019 + layer.1.k_cache 0.01315293 2.45866991 + layer.1.v_cache 0.00000553 0.00098814 + layer.2.k_cache 0.00779452 0.56207852 + layer.2.v_cache 0.00001723 0.00283165 + layer.3.k_cache 0.01828384 3.33092698 + layer.3.v_cache 0.00001944 0.00375215 + layer.4.k_cache 0.00060559 0.09220173 + layer.4.v_cache 0.00005113 0.00762653 + layer.4.output 0.29517000 1044.61578028 + ------------------------------------------------------------------------------------- + TOTAL 0.13220186 432.01644439 + (elements=400,384) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 400384 +Total Bytes 109248 +BPFP 2.1829 bits/point +EBPFP 4.3657 equivalent bits/point +MSE 432.016444 +---------------------- -------------------------------------------------------- +Time: 2.489s Load: 0.002s, Pack+Encode: 1.502s, Decode+Unpack: 0.985s +---------------------- -------------------------------------------------------- +💾 Converting with 432.0164 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-148.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-148.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 110, 128) +Output shape: (1, 110, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) -> torch.Size([1, 1, 110, 512]) + layer.4.output: torch.Size([1, 110, 3584]) -> torch.Size([1, 1, 110, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,704B, BPFP=0.8102 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,240B, BPFP=3.3011 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,372B, BPFP=1.4733 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,296B, BPFP=3.1670 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,920B, BPFP=1.6932 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,284B, BPFP=3.0233 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,272B, BPFP=1.6011 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,980B, BPFP=3.1222 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,652B, BPFP=2.5074 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,044B, BPFP=2.9892 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,388B, BPFP=0.8196 +⌛️ [2/4] FRONTEND: Frontend time: 1.510s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.030s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 110, 128]) + layer.0.v_cache: torch.Size([1, 4, 110, 128]) + layer.1.k_cache: torch.Size([1, 4, 110, 128]) + layer.1.v_cache: torch.Size([1, 4, 110, 128]) + layer.2.k_cache: torch.Size([1, 4, 110, 128]) + layer.2.v_cache: torch.Size([1, 4, 110, 128]) + layer.3.k_cache: torch.Size([1, 4, 110, 128]) + layer.3.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.k_cache: torch.Size([1, 4, 110, 128]) + layer.4.v_cache: torch.Size([1, 4, 110, 128]) + layer.4.output: torch.Size([1, 110, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14476244 27.28904031 + layer.0.v_cache 0.00001476 0.00210916 + layer.1.k_cache 0.05665370 1.89033259 + layer.1.v_cache 0.00000525 0.00077835 + layer.2.k_cache 0.00786287 0.62012239 + layer.2.v_cache 0.00001801 0.00243859 + layer.3.k_cache 0.01410385 3.91354204 + layer.3.v_cache 0.00002027 0.00281301 + layer.4.k_cache 0.00066873 0.08563109 + layer.4.v_cache 0.00004626 0.00567579 + layer.4.output 10.36401748 441.15308442 + ------------------------------------------------------------------------------------- + TOTAL 4.28072226 183.64023966 + (elements=957,440) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 957440 +Total Bytes 207152 +BPFP 1.7309 bits/point +EBPFP 3.4618 equivalent bits/point +MSE 183.640240 +---------------------- -------------------------------------------------------- +Time: 2.545s Load: 0.005s, Pack+Encode: 1.510s, Decode+Unpack: 1.030s +---------------------- -------------------------------------------------------- +💾 Converting with 183.6402 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-15.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-15.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,724B, BPFP=0.9113 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,176B, BPFP=3.8920 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,792B, BPFP=1.5031 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,368B, BPFP=3.7361 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,196B, BPFP=1.7739 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,948B, BPFP=3.6551 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,516B, BPFP=1.6427 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,628B, BPFP=3.7863 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,072B, BPFP=2.7145 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,824B, BPFP=3.6312 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,564B, BPFP=1.0903 +⌛️ [2/4] FRONTEND: Frontend time: 1.506s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.022s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10533771 24.54881908 + layer.0.v_cache 0.00001610 0.00238049 + layer.1.k_cache 0.05553475 1.95173702 + layer.1.v_cache 0.00000748 0.00098681 + layer.2.k_cache 0.02166171 0.60311946 + layer.2.v_cache 0.00002000 0.00285812 + layer.3.k_cache 0.06386927 3.05103293 + layer.3.v_cache 0.00001944 0.00314027 + layer.4.k_cache 0.00064315 0.07883936 + layer.4.v_cache 0.00005385 0.00659347 + layer.4.output 0.16801396 551.69334215 + ------------------------------------------------------------------------------------- + TOTAL 0.08372125 228.94722953 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 180808 +BPFP 2.0517 bits/point +EBPFP 4.1033 equivalent bits/point +MSE 228.947230 +---------------------- -------------------------------------------------------- +Time: 2.532s Load: 0.004s, Pack+Encode: 1.506s, Decode+Unpack: 1.022s +---------------------- -------------------------------------------------------- +💾 Converting with 228.9472 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-156.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-156.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 49, 128) +Output shape: (1, 49, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) -> torch.Size([1, 1, 49, 512]) + layer.4.output: torch.Size([1, 49, 3584]) -> torch.Size([1, 1, 49, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,284B, BPFP=1.0472 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,980B, BPFP=3.8202 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,716B, BPFP=1.8227 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,624B, BPFP=3.7066 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,228B, BPFP=1.9860 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,368B, BPFP=3.6250 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,712B, BPFP=1.8214 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,572B, BPFP=3.6901 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,496B, BPFP=3.0281 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,232B, BPFP=3.5816 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,980B, BPFP=1.1379 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.983s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 49, 128]) + layer.0.v_cache: torch.Size([1, 4, 49, 128]) + layer.1.k_cache: torch.Size([1, 4, 49, 128]) + layer.1.v_cache: torch.Size([1, 4, 49, 128]) + layer.2.k_cache: torch.Size([1, 4, 49, 128]) + layer.2.v_cache: torch.Size([1, 4, 49, 128]) + layer.3.k_cache: torch.Size([1, 4, 49, 128]) + layer.3.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.k_cache: torch.Size([1, 4, 49, 128]) + layer.4.v_cache: torch.Size([1, 4, 49, 128]) + layer.4.output: torch.Size([1, 49, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12410851 23.49839316 + layer.0.v_cache 0.00001740 0.00268354 + layer.1.k_cache 0.01401279 3.25513287 + layer.1.v_cache 0.00000579 0.00100530 + layer.2.k_cache 0.00545398 0.51507008 + layer.2.v_cache 0.00001955 0.00292245 + layer.3.k_cache 0.07546502 3.86587431 + layer.3.v_cache 0.00001978 0.00364144 + layer.4.k_cache 0.00063206 0.09787163 + layer.4.v_cache 0.00005205 0.00691132 + layer.4.output 0.27729562 987.91572522 + ------------------------------------------------------------------------------------- + TOTAL 0.12710919 408.62703427 + (elements=426,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 426496 +Total Bytes 113192 +BPFP 2.1232 bits/point +EBPFP 4.2464 equivalent bits/point +MSE 408.627034 +---------------------- -------------------------------------------------------- +Time: 2.485s Load: 0.003s, Pack+Encode: 1.500s, Decode+Unpack: 0.983s +---------------------- -------------------------------------------------------- +💾 Converting with 408.6270 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-158.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-158.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,416B, BPFP=0.9364 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,048B, BPFP=3.3026 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,704B, BPFP=1.5636 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,596B, BPFP=3.1787 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,928B, BPFP=1.8991 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,248B, BPFP=3.0833 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,108B, BPFP=1.6743 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,584B, BPFP=3.1754 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,616B, BPFP=2.6360 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,140B, BPFP=3.0537 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,628B, BPFP=0.9644 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.982s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19197876 26.87616074 + layer.0.v_cache 0.00001449 0.00251331 + layer.1.k_cache 0.01466701 2.31431339 + layer.1.v_cache 0.00000519 0.00092483 + layer.2.k_cache 0.00720649 0.68613795 + layer.2.v_cache 0.00001785 0.00285994 + layer.3.k_cache 0.10851213 2.49162159 + layer.3.v_cache 0.00001996 0.00323616 + layer.4.k_cache 0.00062085 0.09250600 + layer.4.v_cache 0.00004770 0.00700426 + layer.4.output 0.23832821 875.42606516 + ------------------------------------------------------------------------------------- + TOTAL 0.11714047 362.37998437 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 114016 +BPFP 1.8385 bits/point +EBPFP 3.6770 equivalent bits/point +MSE 362.379984 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.002s, Pack+Encode: 1.495s, Decode+Unpack: 0.982s +---------------------- -------------------------------------------------------- +💾 Converting with 362.3800 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-161.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-161.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,364B, BPFP=1.0306 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,856B, BPFP=3.6324 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,300B, BPFP=1.6238 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,432B, BPFP=3.5025 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,372B, BPFP=1.9522 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,232B, BPFP=3.4412 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,148B, BPFP=1.8836 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,488B, BPFP=3.5196 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,412B, BPFP=2.8836 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,088B, BPFP=3.3971 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,304B, BPFP=1.0200 +⌛️ [2/4] FRONTEND: Frontend time: 1.507s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.980s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12665685 26.05120251 + layer.0.v_cache 0.00001436 0.00267673 + layer.1.k_cache 0.01497617 2.88841427 + layer.1.v_cache 0.00000537 0.00098028 + layer.2.k_cache 0.00986420 0.58461993 + layer.2.v_cache 0.00001802 0.00284354 + layer.3.k_cache 0.09846177 4.39827294 + layer.3.v_cache 0.00001923 0.00346472 + layer.4.k_cache 0.00062397 0.08776072 + layer.4.v_cache 0.00005014 0.00710052 + layer.4.output 0.26631049 300.11368610 + ------------------------------------------------------------------------------------- + TOTAL 0.12440374 125.57783170 + (elements=443,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 443904 +Total Bytes 110996 +BPFP 2.0004 bits/point +EBPFP 4.0007 equivalent bits/point +MSE 125.577832 +---------------------- -------------------------------------------------------- +Time: 2.490s Load: 0.003s, Pack+Encode: 1.507s, Decode+Unpack: 0.980s +---------------------- -------------------------------------------------------- +💾 Converting with 125.5778 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-166.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-166.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,228B, BPFP=1.0508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,124B, BPFP=3.9466 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,592B, BPFP=1.8203 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,908B, BPFP=3.8763 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,128B, BPFP=2.3203 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,612B, BPFP=3.7799 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,032B, BPFP=1.9635 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,820B, BPFP=3.8477 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,856B, BPFP=3.2083 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,424B, BPFP=3.7188 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,836B, BPFP=1.2015 +⌛️ [2/4] FRONTEND: Frontend time: 1.503s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.983s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18531116 23.71522776 + layer.0.v_cache 0.00001874 0.00305484 + layer.1.k_cache 0.01459585 3.06698132 + layer.1.v_cache 0.00000674 0.00122006 + layer.2.k_cache 0.01042949 0.64436849 + layer.2.v_cache 0.00001841 0.00312443 + layer.3.k_cache 0.15260293 4.11640040 + layer.3.v_cache 0.00002390 0.00430049 + layer.4.k_cache 0.00061422 0.09280107 + layer.4.v_cache 0.00004881 0.00753170 + layer.4.output 0.28297247 968.70675223 + ------------------------------------------------------------------------------------- + TOTAL 0.13791044 400.74131036 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 116560 +BPFP 2.2319 bits/point +EBPFP 4.4638 equivalent bits/point +MSE 400.741310 +---------------------- -------------------------------------------------------- +Time: 2.489s Load: 0.003s, Pack+Encode: 1.503s, Decode+Unpack: 0.983s +---------------------- -------------------------------------------------------- +💾 Converting with 400.7413 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-182.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,484B, BPFP=0.9227 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,200B, BPFP=3.2309 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,776B, BPFP=1.5297 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,732B, BPFP=3.1070 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,216B, BPFP=1.9110 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,400B, BPFP=3.0191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,688B, BPFP=1.7712 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,856B, BPFP=3.1398 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,888B, BPFP=2.6186 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,416B, BPFP=3.0233 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,252B, BPFP=1.0310 +⌛️ [2/4] FRONTEND: Frontend time: 1.500s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.979s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12052289 26.06010825 + layer.0.v_cache 0.00001546 0.00276844 + layer.1.k_cache 0.01434235 2.39599480 + layer.1.v_cache 0.00000574 0.00104393 + layer.2.k_cache 0.00254285 0.61068751 + layer.2.v_cache 0.00001710 0.00301422 + layer.3.k_cache 0.03737745 3.48447612 + layer.3.v_cache 0.00001946 0.00388361 + layer.4.k_cache 0.00063766 0.09776627 + layer.4.v_cache 0.00005039 0.00745246 + layer.4.output 0.23032015 784.81946126 + ------------------------------------------------------------------------------------- + TOTAL 0.10516308 325.08255438 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 118908 +BPFP 1.8524 bits/point +EBPFP 3.7048 equivalent bits/point +MSE 325.082554 +---------------------- -------------------------------------------------------- +Time: 2.482s Load: 0.003s, Pack+Encode: 1.500s, Decode+Unpack: 0.979s +---------------------- -------------------------------------------------------- +💾 Converting with 325.0826 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-192.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-192.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,432B, BPFP=0.9358 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,684B, BPFP=3.7340 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,208B, BPFP=1.5220 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,652B, BPFP=3.5160 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,432B, BPFP=1.7804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,716B, BPFP=3.3184 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,860B, BPFP=1.6596 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,476B, BPFP=3.6900 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,656B, BPFP=2.6723 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,944B, BPFP=3.3666 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,488B, BPFP=0.9800 +⌛️ [2/4] FRONTEND: Frontend time: 1.505s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.013s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11335979 26.68247532 + layer.0.v_cache 0.00001513 0.00246465 + layer.1.k_cache 0.01655370 1.78789190 + layer.1.v_cache 0.00000667 0.00098085 + layer.2.k_cache 0.00428703 0.62617534 + layer.2.v_cache 0.00001789 0.00280813 + layer.3.k_cache 0.04938591 2.85819719 + layer.3.v_cache 0.00002111 0.00335546 + layer.4.k_cache 0.00066117 0.08561521 + layer.4.v_cache 0.00004716 0.00655638 + layer.4.output 0.18372514 791.00959218 + ------------------------------------------------------------------------------------- + TOTAL 0.08649597 327.59550975 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 156548 +BPFP 1.9444 bits/point +EBPFP 3.8888 equivalent bits/point +MSE 327.595510 +---------------------- -------------------------------------------------------- +Time: 2.521s Load: 0.003s, Pack+Encode: 1.505s, Decode+Unpack: 1.013s +---------------------- -------------------------------------------------------- +💾 Converting with 327.5955 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-207.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-207.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,108B, BPFP=1.1037 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,080B, BPFP=4.2898 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,020B, BPFP=1.7827 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,728B, BPFP=4.1648 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,272B, BPFP=2.2273 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,324B, BPFP=4.0213 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,444B, BPFP=1.9332 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,724B, BPFP=4.1634 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,356B, BPFP=3.3224 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,200B, BPFP=3.9773 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,136B, BPFP=1.2244 +⌛️ [2/4] FRONTEND: Frontend time: 1.511s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.985s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14777968 22.90544406 + layer.0.v_cache 0.00001531 0.00308124 + layer.1.k_cache 0.01728138 2.11001639 + layer.1.v_cache 0.00000547 0.00107052 + layer.2.k_cache 0.00823456 0.52356091 + layer.2.v_cache 0.00001932 0.00326983 + layer.3.k_cache 0.16770924 3.17608851 + layer.3.v_cache 0.00002046 0.00385987 + layer.4.k_cache 0.00061569 0.09247810 + layer.4.v_cache 0.00005592 0.00793813 + layer.4.output 0.30867824 1016.81990666 + ------------------------------------------------------------------------------------- + TOTAL 0.14720498 420.38624436 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 111392 +BPFP 2.3269 bits/point +EBPFP 4.6537 equivalent bits/point +MSE 420.386244 +---------------------- -------------------------------------------------------- +Time: 2.498s Load: 0.003s, Pack+Encode: 1.511s, Decode+Unpack: 0.985s +---------------------- -------------------------------------------------------- +💾 Converting with 420.3862 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-208.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,656B, BPFP=0.9327 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,332B, BPFP=4.0729 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,740B, BPFP=1.5505 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,964B, BPFP=3.7989 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,112B, BPFP=1.8253 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,612B, BPFP=3.7284 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,492B, BPFP=1.7011 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,688B, BPFP=3.9439 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,576B, BPFP=2.7196 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,636B, BPFP=3.7332 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,288B, BPFP=0.9812 +⌛️ [2/4] FRONTEND: Frontend time: 1.502s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.022s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14408422 25.72647016 + layer.0.v_cache 0.00002097 0.00265681 + layer.1.k_cache 0.05735342 1.97742012 + layer.1.v_cache 0.00000575 0.00103541 + layer.2.k_cache 0.01678689 0.61460847 + layer.2.v_cache 0.00001894 0.00286848 + layer.3.k_cache 0.03029843 2.80980311 + layer.3.v_cache 0.00002125 0.00358205 + layer.4.k_cache 0.00062582 0.07962077 + layer.4.v_cache 0.00005050 0.00692635 + layer.4.output 0.17433185 532.98786630 + ------------------------------------------------------------------------------------- + TOTAL 0.08644642 221.30235623 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 174096 +BPFP 2.0515 bits/point +EBPFP 4.1029 equivalent bits/point +MSE 221.302356 +---------------------- -------------------------------------------------------- +Time: 2.527s Load: 0.003s, Pack+Encode: 1.502s, Decode+Unpack: 1.022s +---------------------- -------------------------------------------------------- +💾 Converting with 221.3024 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-212.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-212.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 44, 128) +Output shape: (1, 44, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) -> torch.Size([1, 1, 44, 512]) + layer.4.output: torch.Size([1, 44, 3584]) -> torch.Size([1, 1, 44, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,128B, BPFP=1.1108 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,876B, BPFP=4.2173 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,952B, BPFP=1.7585 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,556B, BPFP=4.1037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,816B, BPFP=2.0653 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 10,988B, BPFP=3.9020 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,188B, BPFP=1.8423 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,544B, BPFP=4.0994 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,432B, BPFP=3.3494 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,116B, BPFP=3.9474 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,216B, BPFP=1.0763 +⌛️ [2/4] FRONTEND: Frontend time: 1.554s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.979s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 44, 128]) + layer.0.v_cache: torch.Size([1, 4, 44, 128]) + layer.1.k_cache: torch.Size([1, 4, 44, 128]) + layer.1.v_cache: torch.Size([1, 4, 44, 128]) + layer.2.k_cache: torch.Size([1, 4, 44, 128]) + layer.2.v_cache: torch.Size([1, 4, 44, 128]) + layer.3.k_cache: torch.Size([1, 4, 44, 128]) + layer.3.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.k_cache: torch.Size([1, 4, 44, 128]) + layer.4.v_cache: torch.Size([1, 4, 44, 128]) + layer.4.output: torch.Size([1, 44, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13243668 21.24966708 + layer.0.v_cache 0.00001576 0.00288951 + layer.1.k_cache 0.01981235 2.47192452 + layer.1.v_cache 0.00000577 0.00102516 + layer.2.k_cache 0.01756107 0.60048324 + layer.2.v_cache 0.00001728 0.00308347 + layer.3.k_cache 0.04920139 2.31061363 + layer.3.v_cache 0.00001837 0.00340455 + layer.4.k_cache 0.00061593 0.09173581 + layer.4.v_cache 0.00004907 0.00738111 + layer.4.output 0.30854983 886.75537744 + ------------------------------------------------------------------------------------- + TOTAL 0.13997544 366.70763824 + (elements=382,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 382976 +Total Bytes 106812 +BPFP 2.2312 bits/point +EBPFP 4.4624 equivalent bits/point +MSE 366.707638 +---------------------- -------------------------------------------------------- +Time: 2.536s Load: 0.003s, Pack+Encode: 1.554s, Decode+Unpack: 0.979s +---------------------- -------------------------------------------------------- +💾 Converting with 366.7076 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-215.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,328B, BPFP=1.0400 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,100B, BPFP=3.7812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,884B, BPFP=1.8388 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,752B, BPFP=3.6725 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,860B, BPFP=2.1437 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,428B, BPFP=3.5713 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,536B, BPFP=2.0425 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,772B, BPFP=3.6787 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,384B, BPFP=2.9325 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,240B, BPFP=3.5125 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,320B, BPFP=1.0857 +⌛️ [2/4] FRONTEND: Frontend time: 1.495s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.981s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20666794 26.46999512 + layer.0.v_cache 0.00002565 0.00288967 + layer.1.k_cache 0.01518770 2.81772919 + layer.1.v_cache 0.00000581 0.00103734 + layer.2.k_cache 0.01031687 0.60697140 + layer.2.v_cache 0.00001873 0.00306266 + layer.3.k_cache 0.09638391 4.00221222 + layer.3.v_cache 0.00001967 0.00371139 + layer.4.k_cache 0.00062172 0.09309103 + layer.4.v_cache 0.00004827 0.00751316 + layer.4.output 0.27162739 985.64491071 + ------------------------------------------------------------------------------------- + TOTAL 0.13121694 407.85426989 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 114604 +BPFP 2.1067 bits/point +EBPFP 4.2134 equivalent bits/point +MSE 407.854270 +---------------------- -------------------------------------------------------- +Time: 2.479s Load: 0.002s, Pack+Encode: 1.495s, Decode+Unpack: 0.981s +---------------------- -------------------------------------------------------- +💾 Converting with 407.8543 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-217.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-217.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,012B, BPFP=0.9219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,256B, BPFP=3.9651 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,780B, BPFP=1.5579 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,556B, BPFP=3.8042 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,760B, BPFP=1.7831 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,412B, BPFP=3.5414 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,216B, BPFP=1.6581 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,176B, BPFP=3.7169 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,484B, BPFP=2.6388 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,768B, BPFP=3.6232 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,904B, BPFP=1.0473 +⌛️ [2/4] FRONTEND: Frontend time: 1.582s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.023s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13431002 23.97280345 + layer.0.v_cache 0.00001720 0.00258098 + layer.1.k_cache 0.01179634 2.02593613 + layer.1.v_cache 0.00000537 0.00093692 + layer.2.k_cache 0.00798416 0.58555390 + layer.2.v_cache 0.00001783 0.00273131 + layer.3.k_cache 0.05795463 2.86954947 + layer.3.v_cache 0.00001974 0.00321099 + layer.4.k_cache 0.00062059 0.07881274 + layer.4.v_cache 0.00005204 0.00633259 + layer.4.output 0.19989206 591.96028099 + ------------------------------------------------------------------------------------- + TOTAL 0.09482484 245.48649502 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 150324 +BPFP 2.0318 bits/point +EBPFP 4.0637 equivalent bits/point +MSE 245.486495 +---------------------- -------------------------------------------------------- +Time: 2.608s Load: 0.003s, Pack+Encode: 1.582s, Decode+Unpack: 1.023s +---------------------- -------------------------------------------------------- +💾 Converting with 245.4865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-219.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 50, 128) +Output shape: (1, 50, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) -> torch.Size([1, 1, 50, 512]) + layer.4.output: torch.Size([1, 50, 3584]) -> torch.Size([1, 1, 50, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,320B, BPFP=1.0375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,960B, BPFP=3.7375 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,776B, BPFP=1.8050 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,664B, BPFP=3.6450 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,856B, BPFP=2.1425 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,444B, BPFP=3.5762 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,028B, BPFP=1.8838 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,724B, BPFP=3.6637 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,616B, BPFP=3.0050 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,340B, BPFP=3.5438 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,128B, BPFP=1.0771 +⌛️ [2/4] FRONTEND: Frontend time: 1.496s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.979s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 50, 128]) + layer.0.v_cache: torch.Size([1, 4, 50, 128]) + layer.1.k_cache: torch.Size([1, 4, 50, 128]) + layer.1.v_cache: torch.Size([1, 4, 50, 128]) + layer.2.k_cache: torch.Size([1, 4, 50, 128]) + layer.2.v_cache: torch.Size([1, 4, 50, 128]) + layer.3.k_cache: torch.Size([1, 4, 50, 128]) + layer.3.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.k_cache: torch.Size([1, 4, 50, 128]) + layer.4.v_cache: torch.Size([1, 4, 50, 128]) + layer.4.output: torch.Size([1, 50, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15467889 23.47854004 + layer.0.v_cache 0.00001741 0.00291921 + layer.1.k_cache 0.01620592 2.60449829 + layer.1.v_cache 0.00000548 0.00104502 + layer.2.k_cache 0.00723796 0.57569679 + layer.2.v_cache 0.00001737 0.00298482 + layer.3.k_cache 0.06661982 4.27295227 + layer.3.v_cache 0.00002040 0.00377721 + layer.4.k_cache 0.00058776 0.09429190 + layer.4.v_cache 0.00004656 0.00706421 + layer.4.output 0.27162074 961.70794643 + ------------------------------------------------------------------------------------- + TOTAL 0.12628134 397.82349381 + (elements=435,200) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 435200 +Total Bytes 113856 +BPFP 2.0929 bits/point +EBPFP 4.1859 equivalent bits/point +MSE 397.823494 +---------------------- -------------------------------------------------------- +Time: 2.478s Load: 0.003s, Pack+Encode: 1.496s, Decode+Unpack: 0.979s +---------------------- -------------------------------------------------------- +💾 Converting with 397.8235 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-224.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-224.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 51, 128) +Output shape: (1, 51, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) -> torch.Size([1, 1, 51, 512]) + layer.4.output: torch.Size([1, 51, 3584]) -> torch.Size([1, 1, 51, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,364B, BPFP=1.0306 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,804B, BPFP=3.6164 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,372B, BPFP=1.6458 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,372B, BPFP=3.4841 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,168B, BPFP=1.8897 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,140B, BPFP=3.4130 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,156B, BPFP=1.8860 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,500B, BPFP=3.5233 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,292B, BPFP=2.8468 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,136B, BPFP=3.4118 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 22,456B, BPFP=0.9828 +⌛️ [2/4] FRONTEND: Frontend time: 41.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.981s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 51, 128]) + layer.0.v_cache: torch.Size([1, 4, 51, 128]) + layer.1.k_cache: torch.Size([1, 4, 51, 128]) + layer.1.v_cache: torch.Size([1, 4, 51, 128]) + layer.2.k_cache: torch.Size([1, 4, 51, 128]) + layer.2.v_cache: torch.Size([1, 4, 51, 128]) + layer.3.k_cache: torch.Size([1, 4, 51, 128]) + layer.3.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.k_cache: torch.Size([1, 4, 51, 128]) + layer.4.v_cache: torch.Size([1, 4, 51, 128]) + layer.4.output: torch.Size([1, 51, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11261671 26.23448510 + layer.0.v_cache 0.00001829 0.00284272 + layer.1.k_cache 0.01190703 2.36701576 + layer.1.v_cache 0.00000528 0.00097106 + layer.2.k_cache 0.00973116 0.59351850 + layer.2.v_cache 0.00001668 0.00276611 + layer.3.k_cache 0.04200072 4.22937790 + layer.3.v_cache 0.00001874 0.00363694 + layer.4.k_cache 0.00062040 0.08724918 + layer.4.v_cache 0.00004776 0.00699862 + layer.4.output 0.26626884 903.42191877 + ------------------------------------------------------------------------------------- + TOTAL 0.12005086 373.96954666 + (elements=443,904) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 443904 +Total Bytes 109760 +BPFP 1.9781 bits/point +EBPFP 3.9562 equivalent bits/point +MSE 373.969547 +---------------------- --------------------------------------------------------- +Time: 42.211s Load: 0.003s, Pack+Encode: 41.227s, Decode+Unpack: 0.981s +---------------------- --------------------------------------------------------- +💾 Converting with 373.9695 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-225.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-225.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,364B, BPFP=1.0108 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,204B, BPFP=3.6671 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,588B, BPFP=1.6791 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,628B, BPFP=3.4940 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,060B, BPFP=2.1214 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,464B, BPFP=3.4447 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,344B, BPFP=1.9062 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,680B, BPFP=3.5096 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,700B, BPFP=2.9147 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,360B, BPFP=3.4135 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,404B, BPFP=1.1334 +⌛️ [2/4] FRONTEND: Frontend time: 1.585s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.042s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12932784 24.95493727 + layer.0.v_cache 0.00001491 0.00273758 + layer.1.k_cache 0.01488123 2.59368691 + layer.1.v_cache 0.00000563 0.00103756 + layer.2.k_cache 0.00939683 0.53795961 + layer.2.v_cache 0.00001981 0.00308602 + layer.3.k_cache 0.07051506 4.58780993 + layer.3.v_cache 0.00001927 0.00369882 + layer.4.k_cache 0.00064413 0.09215082 + layer.4.v_cache 0.00005939 0.00714248 + layer.4.output 0.26127321 675.21205357 + ------------------------------------------------------------------------------------- + TOTAL 0.12081156 279.95697776 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 116796 +BPFP 2.0644 bits/point +EBPFP 4.1288 equivalent bits/point +MSE 279.956978 +---------------------- -------------------------------------------------------- +Time: 2.630s Load: 0.002s, Pack+Encode: 1.585s, Decode+Unpack: 1.042s +---------------------- -------------------------------------------------------- +💾 Converting with 279.9570 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-229.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-229.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,376B, BPFP=1.0144 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,156B, BPFP=3.6526 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,456B, BPFP=1.6394 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,812B, BPFP=3.5493 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,236B, BPFP=2.1743 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,476B, BPFP=3.4483 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,404B, BPFP=1.9243 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,744B, BPFP=3.5288 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,852B, BPFP=2.9603 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,264B, BPFP=3.3846 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,652B, BPFP=1.1011 +⌛️ [2/4] FRONTEND: Frontend time: 1.611s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.050s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13190371 23.64801495 + layer.0.v_cache 0.00001552 0.00292570 + layer.1.k_cache 0.01668521 2.35429676 + layer.1.v_cache 0.00000581 0.00108028 + layer.2.k_cache 0.00490881 0.57327623 + layer.2.v_cache 0.00001873 0.00296783 + layer.3.k_cache 0.11824917 4.02896148 + layer.3.v_cache 0.00002115 0.00375383 + layer.4.k_cache 0.00061206 0.08716910 + layer.4.v_cache 0.00004844 0.00665375 + layer.4.output 0.26122982 871.10302198 + ------------------------------------------------------------------------------------- + TOTAL 0.12359279 360.49589728 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 116428 +BPFP 2.0579 bits/point +EBPFP 4.1158 equivalent bits/point +MSE 360.495897 +---------------------- -------------------------------------------------------- +Time: 2.663s Load: 0.003s, Pack+Encode: 1.611s, Decode+Unpack: 1.050s +---------------------- -------------------------------------------------------- +💾 Converting with 360.4959 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-235.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-235.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,000B, BPFP=0.8492 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,068B, BPFP=3.5781 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,140B, BPFP=1.3825 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,800B, BPFP=3.3628 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,264B, BPFP=1.5734 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,212B, BPFP=3.2629 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,688B, BPFP=1.4755 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,000B, BPFP=3.3967 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,360B, BPFP=2.4389 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,944B, BPFP=3.2174 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,484B, BPFP=0.8367 +⌛️ [2/4] FRONTEND: Frontend time: 1.838s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.175s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10326390 26.62999692 + layer.0.v_cache 0.00001541 0.00220781 + layer.1.k_cache 0.04976616 1.84108220 + layer.1.v_cache 0.00000534 0.00079997 + layer.2.k_cache 0.01168035 0.66733219 + layer.2.v_cache 0.00001798 0.00248538 + layer.3.k_cache 0.02907468 3.42297032 + layer.3.v_cache 0.00001946 0.00292754 + layer.4.k_cache 0.00073861 0.07614538 + layer.4.v_cache 0.00005077 0.00623036 + layer.4.output 0.14788401 499.23956716 + ------------------------------------------------------------------------------------- + TOTAL 0.07234240 207.48994989 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 178960 +BPFP 1.7879 bits/point +EBPFP 3.5758 equivalent bits/point +MSE 207.489950 +---------------------- -------------------------------------------------------- +Time: 3.017s Load: 0.004s, Pack+Encode: 1.838s, Decode+Unpack: 1.175s +---------------------- -------------------------------------------------------- +💾 Converting with 207.4899 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-238.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 64, 128) +Output shape: (1, 64, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) -> torch.Size([1, 1, 64, 512]) + layer.4.output: torch.Size([1, 64, 3584]) -> torch.Size([1, 1, 64, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 2,984B, BPFP=0.7285 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,856B, BPFP=2.8945 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,240B, BPFP=1.2793 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,400B, BPFP=2.7832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,160B, BPFP=1.5039 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,012B, BPFP=2.6885 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,772B, BPFP=1.4092 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,376B, BPFP=2.7773 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,160B, BPFP=2.2363 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,032B, BPFP=2.6934 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,484B, BPFP=0.8888 +⌛️ [2/4] FRONTEND: Frontend time: 1.829s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.144s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 64, 128]) + layer.0.v_cache: torch.Size([1, 4, 64, 128]) + layer.1.k_cache: torch.Size([1, 4, 64, 128]) + layer.1.v_cache: torch.Size([1, 4, 64, 128]) + layer.2.k_cache: torch.Size([1, 4, 64, 128]) + layer.2.v_cache: torch.Size([1, 4, 64, 128]) + layer.3.k_cache: torch.Size([1, 4, 64, 128]) + layer.3.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.k_cache: torch.Size([1, 4, 64, 128]) + layer.4.v_cache: torch.Size([1, 4, 64, 128]) + layer.4.output: torch.Size([1, 64, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09722494 30.58337593 + layer.0.v_cache 0.00001391 0.00187666 + layer.1.k_cache 0.01806639 1.59798169 + layer.1.v_cache 0.00000557 0.00080578 + layer.2.k_cache 0.00423524 0.49209797 + layer.2.v_cache 0.00001868 0.00268041 + layer.3.k_cache 0.03759564 4.37092781 + layer.3.v_cache 0.00001825 0.00263359 + layer.4.k_cache 0.00061753 0.06794105 + layer.4.v_cache 0.00004894 0.00526828 + layer.4.output 0.21408452 765.25648717 + ------------------------------------------------------------------------------------- + TOTAL 0.09743746 317.28947055 + (elements=557,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 557056 +Total Bytes 111476 +BPFP 1.6009 bits/point +EBPFP 3.2019 equivalent bits/point +MSE 317.289471 +---------------------- -------------------------------------------------------- +Time: 2.976s Load: 0.003s, Pack+Encode: 1.829s, Decode+Unpack: 1.144s +---------------------- -------------------------------------------------------- +💾 Converting with 317.2895 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-24.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-24.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,368B, BPFP=0.8223 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,600B, BPFP=3.4620 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,428B, BPFP=1.4442 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,628B, BPFP=3.3131 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,536B, BPFP=1.6140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,380B, BPFP=3.1219 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,904B, BPFP=1.5172 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,092B, BPFP=3.2310 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,200B, BPFP=2.6348 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,472B, BPFP=3.1360 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,724B, BPFP=0.7818 +⌛️ [2/4] FRONTEND: Frontend time: 1.729s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.305s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11648470 28.24277392 + layer.0.v_cache 0.00001454 0.00204434 + layer.1.k_cache 0.02721847 2.20059548 + layer.1.v_cache 0.00000531 0.00080249 + layer.2.k_cache 0.01045542 0.67407451 + layer.2.v_cache 0.00001699 0.00247691 + layer.3.k_cache 0.01744144 4.12096240 + layer.3.v_cache 0.00001839 0.00279828 + layer.4.k_cache 0.00076342 0.09267007 + layer.4.v_cache 0.00004795 0.00609117 + layer.4.output 11.17676328 466.33552171 + ------------------------------------------------------------------------------------- + TOTAL 4.61234174 194.09964362 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 194332 +BPFP 1.7511 bits/point +EBPFP 3.5022 equivalent bits/point +MSE 194.099644 +---------------------- -------------------------------------------------------- +Time: 3.040s Load: 0.005s, Pack+Encode: 1.729s, Decode+Unpack: 1.305s +---------------------- -------------------------------------------------------- +💾 Converting with 194.0996 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-263.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,708B, BPFP=0.9195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,136B, BPFP=3.7375 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,352B, BPFP=1.4359 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,336B, BPFP=3.3859 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,284B, BPFP=1.6180 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,992B, BPFP=3.1234 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,852B, BPFP=1.5336 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,612B, BPFP=3.2445 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,624B, BPFP=2.4656 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,268B, BPFP=3.1773 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,120B, BPFP=0.8404 +⌛️ [2/4] FRONTEND: Frontend time: 1.659s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.196s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11433966 27.24613037 + layer.0.v_cache 0.00001437 0.00211746 + layer.1.k_cache 0.05315235 2.20502701 + layer.1.v_cache 0.00000524 0.00076998 + layer.2.k_cache 0.00678293 0.66301656 + layer.2.v_cache 0.00001704 0.00239932 + layer.3.k_cache 0.01951189 3.08289089 + layer.3.v_cache 0.00001758 0.00270272 + layer.4.k_cache 0.00076624 0.07383350 + layer.4.v_cache 0.00004856 0.00581594 + layer.4.output 0.16992235 489.85000000 + ------------------------------------------------------------------------------------- + TOTAL 0.08141837 203.66086493 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 156284 +BPFP 1.7955 bits/point +EBPFP 3.5911 equivalent bits/point +MSE 203.660865 +---------------------- -------------------------------------------------------- +Time: 2.859s Load: 0.004s, Pack+Encode: 1.659s, Decode+Unpack: 1.196s +---------------------- -------------------------------------------------------- +💾 Converting with 203.6609 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-264.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-264.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,864B, BPFP=0.8444 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,616B, BPFP=3.5792 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,952B, BPFP=1.3806 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,564B, BPFP=3.3965 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,112B, BPFP=1.5819 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,772B, BPFP=3.2590 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,616B, BPFP=1.4958 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,948B, BPFP=3.4632 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,484B, BPFP=2.5146 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,848B, BPFP=3.2722 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,392B, BPFP=0.8282 +⌛️ [2/4] FRONTEND: Frontend time: 1.932s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.173s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11570581 28.39061415 + layer.0.v_cache 0.00001517 0.00212995 + layer.1.k_cache 0.03236232 1.67766656 + layer.1.v_cache 0.00000662 0.00079316 + layer.2.k_cache 0.00641356 0.72801141 + layer.2.v_cache 0.00001676 0.00239342 + layer.3.k_cache 0.02648412 2.94858331 + layer.3.v_cache 0.00001896 0.00290506 + layer.4.k_cache 0.00074524 0.07758673 + layer.4.v_cache 0.00004843 0.00620720 + layer.4.output 0.15113260 555.41651786 + ------------------------------------------------------------------------------------- + TOTAL 0.07292619 230.69132447 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 176168 +BPFP 1.7991 bits/point +EBPFP 3.5982 equivalent bits/point +MSE 230.691324 +---------------------- -------------------------------------------------------- +Time: 3.110s Load: 0.005s, Pack+Encode: 1.932s, Decode+Unpack: 1.173s +---------------------- -------------------------------------------------------- +💾 Converting with 230.6913 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-271.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-271.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 114, 128) +Output shape: (1, 114, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) -> torch.Size([1, 1, 114, 512]) + layer.4.output: torch.Size([1, 114, 3584]) -> torch.Size([1, 1, 114, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,848B, BPFP=0.8015 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,020B, BPFP=3.1552 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,940B, BPFP=1.3624 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,864B, BPFP=2.9967 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,944B, BPFP=1.6371 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,904B, BPFP=2.8651 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,636B, BPFP=1.4578 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,540B, BPFP=2.9523 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,544B, BPFP=2.4046 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,972B, BPFP=2.8745 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,244B, BPFP=0.7880 +⌛️ [2/4] FRONTEND: Frontend time: 2.079s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.542s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 114, 128]) + layer.0.v_cache: torch.Size([1, 4, 114, 128]) + layer.1.k_cache: torch.Size([1, 4, 114, 128]) + layer.1.v_cache: torch.Size([1, 4, 114, 128]) + layer.2.k_cache: torch.Size([1, 4, 114, 128]) + layer.2.v_cache: torch.Size([1, 4, 114, 128]) + layer.3.k_cache: torch.Size([1, 4, 114, 128]) + layer.3.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.k_cache: torch.Size([1, 4, 114, 128]) + layer.4.v_cache: torch.Size([1, 4, 114, 128]) + layer.4.output: torch.Size([1, 114, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10081583 27.09540330 + layer.0.v_cache 0.00001633 0.00213230 + layer.1.k_cache 0.08473034 2.07401235 + layer.1.v_cache 0.00000563 0.00078771 + layer.2.k_cache 0.00792928 0.65000046 + layer.2.v_cache 0.00001730 0.00256336 + layer.3.k_cache 0.01019665 4.17524880 + layer.3.v_cache 0.00001810 0.00269756 + layer.4.k_cache 0.00077074 0.08810120 + layer.4.v_cache 0.00005022 0.00605052 + layer.4.output 10.00037814 416.05987625 + ------------------------------------------------------------------------------------- + TOTAL 4.12983514 173.32447831 + (elements=992,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 992256 +Total Bytes 204456 +BPFP 1.6484 bits/point +EBPFP 3.2968 equivalent bits/point +MSE 173.324478 +---------------------- -------------------------------------------------------- +Time: 3.626s Load: 0.004s, Pack+Encode: 2.079s, Decode+Unpack: 1.542s +---------------------- -------------------------------------------------------- +💾 Converting with 173.3245 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-275.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-275.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 98, 128) +Output shape: (1, 98, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) -> torch.Size([1, 1, 98, 512]) + layer.4.output: torch.Size([1, 98, 3584]) -> torch.Size([1, 1, 98, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,120B, BPFP=0.8163 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,812B, BPFP=3.4777 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,512B, BPFP=1.3571 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,788B, BPFP=3.3144 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,040B, BPFP=1.6008 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,060B, BPFP=3.1983 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,040B, BPFP=1.4413 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,620B, BPFP=3.2876 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,664B, BPFP=2.6569 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,100B, BPFP=3.2047 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,984B, BPFP=0.7968 +⌛️ [2/4] FRONTEND: Frontend time: 1.765s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.244s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 98, 128]) + layer.0.v_cache: torch.Size([1, 4, 98, 128]) + layer.1.k_cache: torch.Size([1, 4, 98, 128]) + layer.1.v_cache: torch.Size([1, 4, 98, 128]) + layer.2.k_cache: torch.Size([1, 4, 98, 128]) + layer.2.v_cache: torch.Size([1, 4, 98, 128]) + layer.3.k_cache: torch.Size([1, 4, 98, 128]) + layer.3.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.k_cache: torch.Size([1, 4, 98, 128]) + layer.4.v_cache: torch.Size([1, 4, 98, 128]) + layer.4.output: torch.Size([1, 98, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11008905 26.89502202 + layer.0.v_cache 0.00001718 0.00202143 + layer.1.k_cache 0.06240322 1.68820829 + layer.1.v_cache 0.00000548 0.00078154 + layer.2.k_cache 0.00473963 0.77258924 + layer.2.v_cache 0.00001718 0.00241687 + layer.3.k_cache 0.01141277 3.13123399 + layer.3.v_cache 0.00001769 0.00271285 + layer.4.k_cache 0.00076610 0.08830471 + layer.4.v_cache 0.00004814 0.00636149 + layer.4.output 0.02426045 509.61579810 + ------------------------------------------------------------------------------------- + TOTAL 0.02113762 211.75883760 + (elements=852,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 852992 +Total Bytes 187740 +BPFP 1.7608 bits/point +EBPFP 3.5215 equivalent bits/point +MSE 211.758838 +---------------------- -------------------------------------------------------- +Time: 3.014s Load: 0.005s, Pack+Encode: 1.765s, Decode+Unpack: 1.244s +---------------------- -------------------------------------------------------- +💾 Converting with 211.7588 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-276.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,004B, BPFP=0.8499 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,816B, BPFP=3.5353 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,256B, BPFP=1.4022 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,676B, BPFP=3.3417 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,480B, BPFP=1.6101 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,796B, BPFP=3.1923 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,912B, BPFP=1.5136 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,620B, BPFP=3.3322 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,664B, BPFP=2.4905 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,180B, BPFP=3.2575 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,232B, BPFP=0.8548 +⌛️ [2/4] FRONTEND: Frontend time: 1.927s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.265s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08919133 26.52727475 + layer.0.v_cache 0.00001414 0.00210478 + layer.1.k_cache 0.08591949 1.85463184 + layer.1.v_cache 0.00000528 0.00076937 + layer.2.k_cache 0.00389855 0.69257363 + layer.2.v_cache 0.00001738 0.00240601 + layer.3.k_cache 0.07679920 3.20774642 + layer.3.v_cache 0.00002028 0.00273075 + layer.4.k_cache 0.00072407 0.07939848 + layer.4.v_cache 0.00004779 0.00621738 + layer.4.output 0.14786680 524.30109666 + ------------------------------------------------------------------------------------- + TOTAL 0.07598265 217.79314883 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 179636 +BPFP 1.7946 bits/point +EBPFP 3.5893 equivalent bits/point +MSE 217.793149 +---------------------- -------------------------------------------------------- +Time: 3.197s Load: 0.004s, Pack+Encode: 1.927s, Decode+Unpack: 1.265s +---------------------- -------------------------------------------------------- +💾 Converting with 217.7931 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-280.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,580B, BPFP=0.8148 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,156B, BPFP=3.3814 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,860B, BPFP=1.4398 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,384B, BPFP=3.2687 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,876B, BPFP=1.7342 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,344B, BPFP=3.1168 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,992B, BPFP=1.6051 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 22,128B, BPFP=3.2313 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,924B, BPFP=2.6174 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,456B, BPFP=3.1332 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,520B, BPFP=0.8244 +⌛️ [2/4] FRONTEND: Frontend time: 2.201s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.223s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11353527 27.77472437 + layer.0.v_cache 0.00001674 0.00209146 + layer.1.k_cache 0.04276301 2.18052716 + layer.1.v_cache 0.00000542 0.00076996 + layer.2.k_cache 0.00247095 0.53882463 + layer.2.v_cache 0.00001885 0.00259180 + layer.3.k_cache 0.02722364 4.21913047 + layer.3.v_cache 0.00001942 0.00271221 + layer.4.k_cache 0.00068256 0.08252168 + layer.4.v_cache 0.00004843 0.00564949 + layer.4.output 10.65454330 336.65399700 + ------------------------------------------------------------------------------------- + TOTAL 4.39815220 140.66985425 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 206220 +BPFP 1.7714 bits/point +EBPFP 3.5428 equivalent bits/point +MSE 140.669854 +---------------------- -------------------------------------------------------- +Time: 3.429s Load: 0.005s, Pack+Encode: 2.201s, Decode+Unpack: 1.223s +---------------------- -------------------------------------------------------- +💾 Converting with 140.6699 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-284.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,848B, BPFP=0.9018 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,728B, BPFP=3.6696 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,596B, BPFP=1.4129 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,312B, BPFP=3.2202 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,560B, BPFP=1.5923 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,424B, BPFP=3.0551 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,988B, BPFP=1.4859 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,872B, BPFP=3.3244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,148B, BPFP=2.4457 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,464B, BPFP=3.0625 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,768B, BPFP=0.8176 +⌛️ [2/4] FRONTEND: Frontend time: 1.842s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.268s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12550225 26.55411784 + layer.0.v_cache 0.00001468 0.00221346 + layer.1.k_cache 0.05234316 2.05547987 + layer.1.v_cache 0.00000508 0.00074184 + layer.2.k_cache 0.00733058 0.65042446 + layer.2.v_cache 0.00001689 0.00234584 + layer.3.k_cache 0.01592888 2.69622730 + layer.3.v_cache 0.00001847 0.00275615 + layer.4.k_cache 0.00069460 0.07134115 + layer.4.v_cache 0.00005205 0.00591855 + layer.4.output 0.16184913 280.46088435 + ------------------------------------------------------------------------------------- + TOTAL 0.07852062 117.36869158 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 160708 +BPFP 1.7584 bits/point +EBPFP 3.5169 equivalent bits/point +MSE 117.368692 +---------------------- -------------------------------------------------------- +Time: 3.115s Load: 0.005s, Pack+Encode: 1.842s, Decode+Unpack: 1.268s +---------------------- -------------------------------------------------------- +💾 Converting with 117.3687 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-285.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-285.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,620B, BPFP=0.9255 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,344B, BPFP=3.6747 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,396B, BPFP=1.4816 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,312B, BPFP=3.4679 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,212B, BPFP=1.6450 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,060B, BPFP=3.2171 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,636B, BPFP=1.5296 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,460B, BPFP=3.4976 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,608B, BPFP=2.5256 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,768B, BPFP=3.3590 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,124B, BPFP=0.8334 +⌛️ [2/4] FRONTEND: Frontend time: 2.064s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.420s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11879376 26.55943572 + layer.0.v_cache 0.00001438 0.00220376 + layer.1.k_cache 0.05518859 2.17813482 + layer.1.v_cache 0.00000514 0.00083962 + layer.2.k_cache 0.00243519 0.71163094 + layer.2.v_cache 0.00001742 0.00252242 + layer.3.k_cache 0.07959401 3.13918304 + layer.3.v_cache 0.00001812 0.00280754 + layer.4.k_cache 0.00069358 0.07635729 + layer.4.v_cache 0.00004620 0.00593203 + layer.4.output 0.17426339 451.15911172 + ------------------------------------------------------------------------------------- + TOTAL 0.08686177 187.69369584 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 155540 +BPFP 1.8328 bits/point +EBPFP 3.6656 equivalent bits/point +MSE 187.693696 +---------------------- -------------------------------------------------------- +Time: 3.488s Load: 0.005s, Pack+Encode: 2.064s, Decode+Unpack: 1.420s +---------------------- -------------------------------------------------------- +💾 Converting with 187.6937 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-293.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,740B, BPFP=0.9032 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,916B, BPFP=3.6044 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,672B, BPFP=1.4619 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,228B, BPFP=3.4733 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,636B, BPFP=1.6456 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,428B, BPFP=3.3209 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,036B, BPFP=1.5312 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,588B, BPFP=3.5419 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,360B, BPFP=2.5457 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,804B, BPFP=3.3925 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,800B, BPFP=0.8112 +⌛️ [2/4] FRONTEND: Frontend time: 1.741s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.286s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11325797 25.11391244 + layer.0.v_cache 0.00001565 0.00215564 + layer.1.k_cache 0.07651102 1.93827280 + layer.1.v_cache 0.00000530 0.00086456 + layer.2.k_cache 0.00872486 0.69495066 + layer.2.v_cache 0.00001696 0.00251683 + layer.3.k_cache 0.12876377 2.70805787 + layer.3.v_cache 0.00001905 0.00286462 + layer.4.k_cache 0.00069190 0.08029293 + layer.4.v_cache 0.00004968 0.00651180 + layer.4.output 0.16578155 567.53576873 + ------------------------------------------------------------------------------------- + TOTAL 0.08756041 235.48828125 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 163208 +BPFP 1.8294 bits/point +EBPFP 3.6587 equivalent bits/point +MSE 235.488281 +---------------------- -------------------------------------------------------- +Time: 3.030s Load: 0.003s, Pack+Encode: 1.741s, Decode+Unpack: 1.286s +---------------------- -------------------------------------------------------- +💾 Converting with 235.4883 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-299.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-299.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,612B, BPFP=0.9359 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,064B, BPFP=3.4627 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,468B, BPFP=1.5154 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,272B, BPFP=3.3019 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,300B, BPFP=1.6843 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,452B, BPFP=3.1356 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,692B, BPFP=1.5609 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,164B, BPFP=3.2800 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,448B, BPFP=2.5260 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,128B, BPFP=3.2727 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,520B, BPFP=0.8268 +⌛️ [2/4] FRONTEND: Frontend time: 1.996s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.499s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12835005 26.45978655 + layer.0.v_cache 0.00001616 0.00224230 + layer.1.k_cache 0.05672838 2.02595401 + layer.1.v_cache 0.00000535 0.00092092 + layer.2.k_cache 0.00240846 0.71368448 + layer.2.v_cache 0.00001644 0.00263255 + layer.3.k_cache 0.04570918 2.90138443 + layer.3.v_cache 0.00001733 0.00305417 + layer.4.k_cache 0.00072090 0.08003220 + layer.4.v_cache 0.00004908 0.00635600 + layer.4.output 0.18444509 485.66039541 + ------------------------------------------------------------------------------------- + TOTAL 0.08971394 201.87169503 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 150120 +BPFP 1.7919 bits/point +EBPFP 3.5838 equivalent bits/point +MSE 201.871695 +---------------------- -------------------------------------------------------- +Time: 3.498s Load: 0.003s, Pack+Encode: 1.996s, Decode+Unpack: 1.499s +---------------------- -------------------------------------------------------- +💾 Converting with 201.8717 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-301.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-301.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,312B, BPFP=0.9489 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,780B, BPFP=3.4727 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,848B, BPFP=1.5070 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,700B, BPFP=3.2350 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,568B, BPFP=1.6655 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,248B, BPFP=3.1356 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,044B, BPFP=1.5502 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,980B, BPFP=3.2967 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,488B, BPFP=2.5282 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,216B, BPFP=3.1285 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,796B, BPFP=0.8424 +⌛️ [2/4] FRONTEND: Frontend time: 2.109s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.503s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10085608 26.49323283 + layer.0.v_cache 0.00001554 0.00222214 + layer.1.k_cache 0.06589249 1.97846362 + layer.1.v_cache 0.00000499 0.00086364 + layer.2.k_cache 0.00422055 0.69718444 + layer.2.v_cache 0.00001579 0.00242597 + layer.3.k_cache 0.05337993 2.42050257 + layer.3.v_cache 0.00001747 0.00275018 + layer.4.k_cache 0.00075412 0.07711503 + layer.4.v_cache 0.00004712 0.00619887 + layer.4.output 0.19135183 695.87424547 + ------------------------------------------------------------------------------------- + TOTAL 0.09203923 288.40003986 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 137980 +BPFP 1.7862 bits/point +EBPFP 3.5724 equivalent bits/point +MSE 288.400040 +---------------------- -------------------------------------------------------- +Time: 3.616s Load: 0.004s, Pack+Encode: 2.109s, Decode+Unpack: 1.503s +---------------------- -------------------------------------------------------- +💾 Converting with 288.4000 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-313.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-313.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,856B, BPFP=0.9033 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,692B, BPFP=3.4769 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,736B, BPFP=1.4390 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,484B, BPFP=3.2522 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,704B, BPFP=1.6190 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,780B, BPFP=3.1213 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,152B, BPFP=1.5164 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,812B, BPFP=3.3132 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,172B, BPFP=2.4501 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,224B, BPFP=3.2039 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,476B, BPFP=0.7833 +⌛️ [2/4] FRONTEND: Frontend time: 34.405s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.241s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09057366 27.08527483 + layer.0.v_cache 0.00001789 0.00209422 + layer.1.k_cache 0.03614600 1.84286899 + layer.1.v_cache 0.00000481 0.00074274 + layer.2.k_cache 0.00554357 0.69391955 + layer.2.v_cache 0.00001714 0.00235104 + layer.3.k_cache 0.02742965 2.99485216 + layer.3.v_cache 0.00001729 0.00270829 + layer.4.k_cache 0.00093027 0.07243368 + layer.4.v_cache 0.00004906 0.00597326 + layer.4.output 0.16441821 551.98676658 + ------------------------------------------------------------------------------------- + TOTAL 0.07715628 229.21238734 + (elements=731,136) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 731136 +Total Bytes 160088 +BPFP 1.7517 bits/point +EBPFP 3.5033 equivalent bits/point +MSE 229.212387 +---------------------- --------------------------------------------------------- +Time: 35.650s Load: 0.005s, Pack+Encode: 34.405s, Decode+Unpack: 1.241s +---------------------- --------------------------------------------------------- +💾 Converting with 229.2124 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-315.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-315.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,568B, BPFP=0.9269 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,384B, BPFP=3.5276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,420B, BPFP=1.5057 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,156B, BPFP=3.4813 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,184B, BPFP=1.6607 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,924B, BPFP=3.2313 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,752B, BPFP=1.5731 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,988B, BPFP=3.4472 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,504B, BPFP=2.5373 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,544B, BPFP=3.3571 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,184B, BPFP=0.8460 +⌛️ [2/4] FRONTEND: Frontend time: 1.850s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.287s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09325875 25.12846394 + layer.0.v_cache 0.00001406 0.00215739 + layer.1.k_cache 0.05775181 1.86769639 + layer.1.v_cache 0.00000524 0.00080613 + layer.2.k_cache 0.00413042 0.59070354 + layer.2.v_cache 0.00001743 0.00237848 + layer.3.k_cache 0.04387648 3.24213082 + layer.3.v_cache 0.00002497 0.00297296 + layer.4.k_cache 0.00071512 0.07414220 + layer.4.v_cache 0.00005147 0.00661396 + layer.4.output 0.17652170 602.54017857 + ------------------------------------------------------------------------------------- + TOTAL 0.08444104 249.92348917 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 153608 +BPFP 1.8336 bits/point +EBPFP 3.6671 equivalent bits/point +MSE 249.923489 +---------------------- -------------------------------------------------------- +Time: 3.141s Load: 0.004s, Pack+Encode: 1.850s, Decode+Unpack: 1.287s +---------------------- -------------------------------------------------------- +💾 Converting with 249.9235 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-316.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,872B, BPFP=0.8651 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,128B, BPFP=3.2188 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,732B, BPFP=1.3729 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,820B, BPFP=2.9865 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,672B, BPFP=1.5398 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,204B, BPFP=2.8771 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,940B, BPFP=1.4098 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,468B, BPFP=3.1016 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,092B, BPFP=2.3246 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,432B, BPFP=2.9176 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,808B, BPFP=0.6800 +⌛️ [2/4] FRONTEND: Frontend time: 1.707s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.429s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212358 26.03571389 + layer.0.v_cache 0.00001682 0.00199866 + layer.1.k_cache 0.05918076 1.58078905 + layer.1.v_cache 0.00000473 0.00073435 + layer.2.k_cache 0.00260097 0.67701084 + layer.2.v_cache 0.00001544 0.00230473 + layer.3.k_cache 0.02838851 2.98011572 + layer.3.v_cache 0.00001657 0.00263395 + layer.4.k_cache 0.00106615 0.07975663 + layer.4.v_cache 0.00004456 0.00570456 + layer.4.output 0.17173813 539.21058239 + ------------------------------------------------------------------------------------- + TOTAL 0.08091912 223.87299053 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 154168 +BPFP 1.6102 bits/point +EBPFP 3.2204 equivalent bits/point +MSE 223.872991 +---------------------- -------------------------------------------------------- +Time: 3.140s Load: 0.004s, Pack+Encode: 1.707s, Decode+Unpack: 1.429s +---------------------- -------------------------------------------------------- +💾 Converting with 223.8730 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-324.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,876B, BPFP=0.8658 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,448B, BPFP=3.4531 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,060B, BPFP=1.4311 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,264B, BPFP=3.2429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,040B, BPFP=1.6051 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,720B, BPFP=3.1463 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,248B, BPFP=1.4645 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,160B, BPFP=3.4020 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,556B, BPFP=2.4070 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,352B, BPFP=3.2585 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,292B, BPFP=0.7937 +⌛️ [2/4] FRONTEND: Frontend time: 1.794s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.664s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10731998 26.25578169 + layer.0.v_cache 0.00001618 0.00203661 + layer.1.k_cache 0.07375650 2.07967862 + layer.1.v_cache 0.00000540 0.00079401 + layer.2.k_cache 0.00246326 0.67655126 + layer.2.v_cache 0.00001746 0.00242287 + layer.3.k_cache 0.01642703 3.26688177 + layer.3.v_cache 0.00001778 0.00261802 + layer.4.k_cache 0.00066607 0.07070251 + layer.4.v_cache 0.00004728 0.00575350 + layer.4.output 0.15452130 416.74107143 + ------------------------------------------------------------------------------------- + TOTAL 0.07543447 173.50298358 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 168016 +BPFP 1.7548 bits/point +EBPFP 3.5097 equivalent bits/point +MSE 173.502984 +---------------------- -------------------------------------------------------- +Time: 3.461s Load: 0.003s, Pack+Encode: 1.794s, Decode+Unpack: 1.664s +---------------------- -------------------------------------------------------- +💾 Converting with 173.5030 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-325.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-325.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 99, 128) +Output shape: (1, 99, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) -> torch.Size([1, 1, 99, 512]) + layer.4.output: torch.Size([1, 99, 3584]) -> torch.Size([1, 1, 99, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,224B, BPFP=0.8245 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,860B, BPFP=3.4501 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,744B, BPFP=1.3801 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,464B, BPFP=3.2298 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,520B, BPFP=1.6604 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,716B, BPFP=3.1117 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,088B, BPFP=1.4343 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,712B, BPFP=3.2689 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,528B, BPFP=2.6086 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,052B, BPFP=3.1648 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,436B, BPFP=0.7990 +⌛️ [2/4] FRONTEND: Frontend time: 2.314s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 99, 128]) + layer.0.v_cache: torch.Size([1, 4, 99, 128]) + layer.1.k_cache: torch.Size([1, 4, 99, 128]) + layer.1.v_cache: torch.Size([1, 4, 99, 128]) + layer.2.k_cache: torch.Size([1, 4, 99, 128]) + layer.2.v_cache: torch.Size([1, 4, 99, 128]) + layer.3.k_cache: torch.Size([1, 4, 99, 128]) + layer.3.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.k_cache: torch.Size([1, 4, 99, 128]) + layer.4.v_cache: torch.Size([1, 4, 99, 128]) + layer.4.output: torch.Size([1, 99, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09061131 27.91951744 + layer.0.v_cache 0.00001461 0.00201734 + layer.1.k_cache 0.05062923 2.02765170 + layer.1.v_cache 0.00000709 0.00081282 + layer.2.k_cache 0.00392854 0.60087632 + layer.2.v_cache 0.00001793 0.00242631 + layer.3.k_cache 0.05082532 3.64083030 + layer.3.v_cache 0.00001820 0.00277986 + layer.4.k_cache 0.00080398 0.09170925 + layer.4.v_cache 0.00005053 0.00572817 + layer.4.output 0.02403493 493.14042208 + ------------------------------------------------------------------------------------- + TOTAL 0.02147948 205.07513553 + (elements=861,696) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 861696 +Total Bytes 188344 +BPFP 1.7486 bits/point +EBPFP 3.4972 equivalent bits/point +MSE 205.075136 +---------------------- -------------------------------------------------------- +Time: 3.640s Load: 0.006s, Pack+Encode: 2.314s, Decode+Unpack: 1.320s +---------------------- -------------------------------------------------------- +💾 Converting with 205.0751 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-328.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-328.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,304B, BPFP=0.9472 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,244B, BPFP=3.7949 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,992B, BPFP=1.5387 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,492B, BPFP=3.6294 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,884B, BPFP=1.7350 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,768B, BPFP=3.4701 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,272B, BPFP=1.6004 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,400B, BPFP=3.6092 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,580B, BPFP=2.5484 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,420B, BPFP=3.3935 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,648B, BPFP=0.9321 +⌛️ [2/4] FRONTEND: Frontend time: 2.169s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.180s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11136492 27.51777756 + layer.0.v_cache 0.00001955 0.00235122 + layer.1.k_cache 0.06280328 2.01765549 + layer.1.v_cache 0.00000544 0.00087846 + layer.2.k_cache 0.00235174 0.68725478 + layer.2.v_cache 0.00001859 0.00259204 + layer.3.k_cache 0.03326307 2.79266744 + layer.3.v_cache 0.00001926 0.00311892 + layer.4.k_cache 0.00064473 0.07582532 + layer.4.v_cache 0.00005254 0.00635340 + layer.4.output 0.19494371 642.48578974 + ------------------------------------------------------------------------------------- + TOTAL 0.09265583 266.50041193 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 149004 +BPFP 1.9289 bits/point +EBPFP 3.8578 equivalent bits/point +MSE 266.500412 +---------------------- -------------------------------------------------------- +Time: 3.353s Load: 0.004s, Pack+Encode: 2.169s, Decode+Unpack: 1.180s +---------------------- -------------------------------------------------------- +💾 Converting with 266.5004 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-33.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-33.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,420B, BPFP=0.9333 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,048B, BPFP=3.8108 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,120B, BPFP=1.5034 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,616B, BPFP=3.5084 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,128B, BPFP=1.7162 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,516B, BPFP=3.4873 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,632B, BPFP=1.6115 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,632B, BPFP=3.5118 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,360B, BPFP=2.6098 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,064B, BPFP=3.3919 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,896B, BPFP=0.8716 +⌛️ [2/4] FRONTEND: Frontend time: 1.914s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.252s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08829598 27.29333826 + layer.0.v_cache 0.00001455 0.00239084 + layer.1.k_cache 0.05950350 1.88137776 + layer.1.v_cache 0.00000529 0.00086659 + layer.2.k_cache 0.00411868 0.69083621 + layer.2.v_cache 0.00001680 0.00260481 + layer.3.k_cache 0.04877868 3.04569471 + layer.3.v_cache 0.00001896 0.00310681 + layer.4.k_cache 0.00063741 0.07801264 + layer.4.v_cache 0.00004764 0.00611230 + layer.4.output 0.18365649 655.22623069 + ------------------------------------------------------------------------------------- + TOTAL 0.08747253 271.74046799 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 152432 +BPFP 1.8933 bits/point +EBPFP 3.7866 equivalent bits/point +MSE 271.740468 +---------------------- -------------------------------------------------------- +Time: 3.171s Load: 0.004s, Pack+Encode: 1.914s, Decode+Unpack: 1.252s +---------------------- -------------------------------------------------------- +💾 Converting with 271.7405 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-332.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-332.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,616B, BPFP=0.9367 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,380B, BPFP=3.7297 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,468B, BPFP=1.5154 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,236B, BPFP=3.4976 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,396B, BPFP=1.7037 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,756B, BPFP=3.4002 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,736B, BPFP=1.5698 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,368B, BPFP=3.5244 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,664B, BPFP=2.5698 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,156B, BPFP=3.4813 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,456B, BPFP=0.8249 +⌛️ [2/4] FRONTEND: Frontend time: 1.737s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.342s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11195674 26.73382331 + layer.0.v_cache 0.00001346 0.00205987 + layer.1.k_cache 0.03755230 2.04605261 + layer.1.v_cache 0.00000535 0.00084600 + layer.2.k_cache 0.00239600 0.71284752 + layer.2.v_cache 0.00001753 0.00278722 + layer.3.k_cache 0.02867049 2.95948058 + layer.3.v_cache 0.00001751 0.00287631 + layer.4.k_cache 0.00071811 0.07775338 + layer.4.v_cache 0.00005486 0.00668183 + layer.4.output 0.19495771 617.69515306 + ------------------------------------------------------------------------------------- + TOTAL 0.09094743 256.25948706 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 156232 +BPFP 1.8649 bits/point +EBPFP 3.7298 equivalent bits/point +MSE 256.259487 +---------------------- -------------------------------------------------------- +Time: 3.083s Load: 0.003s, Pack+Encode: 1.737s, Decode+Unpack: 1.342s +---------------------- -------------------------------------------------------- +💾 Converting with 256.2595 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-336.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-336.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,948B, BPFP=0.8590 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,820B, BPFP=3.4410 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,964B, BPFP=1.3826 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,816B, BPFP=3.2667 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,932B, BPFP=1.5507 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,980B, BPFP=3.1215 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,368B, BPFP=1.4528 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,520B, BPFP=3.3889 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,144B, BPFP=2.4556 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,608B, BPFP=3.2306 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,340B, BPFP=0.7277 +⌛️ [2/4] FRONTEND: Frontend time: 1.873s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.285s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08657128 28.05518934 + layer.0.v_cache 0.00001574 0.00220353 + layer.1.k_cache 0.05618976 1.76836124 + layer.1.v_cache 0.00000506 0.00079393 + layer.2.k_cache 0.00392605 0.68572566 + layer.2.v_cache 0.00001570 0.00240687 + layer.3.k_cache 0.02849737 3.21330227 + layer.3.v_cache 0.00001769 0.00284295 + layer.4.k_cache 0.00099738 0.08089177 + layer.4.v_cache 0.00004751 0.00631806 + layer.4.output 0.15107292 569.60600198 + ------------------------------------------------------------------------------------- + TOTAL 0.07257612 236.53294409 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 168440 +BPFP 1.7202 bits/point +EBPFP 3.4404 equivalent bits/point +MSE 236.532944 +---------------------- -------------------------------------------------------- +Time: 3.164s Load: 0.006s, Pack+Encode: 1.873s, Decode+Unpack: 1.285s +---------------------- -------------------------------------------------------- +💾 Converting with 236.5329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-342.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,704B, BPFP=0.9187 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,628B, BPFP=3.4430 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,528B, BPFP=1.4703 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,976B, BPFP=3.3156 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,324B, BPFP=1.6258 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,844B, BPFP=3.2898 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,764B, BPFP=1.5164 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,652B, BPFP=3.4477 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,656B, BPFP=2.4719 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,812B, BPFP=3.2836 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,128B, BPFP=0.7848 +⌛️ [2/4] FRONTEND: Frontend time: 1.888s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.178s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09254595 26.51261292 + layer.0.v_cache 0.00001776 0.00210228 + layer.1.k_cache 0.05775445 1.85851555 + layer.1.v_cache 0.00000504 0.00076892 + layer.2.k_cache 0.00891839 0.70126510 + layer.2.v_cache 0.00001732 0.00252385 + layer.3.k_cache 0.02836140 2.53440323 + layer.3.v_cache 0.00001672 0.00273889 + layer.4.k_cache 0.00083566 0.07641141 + layer.4.v_cache 0.00004876 0.00598375 + layer.4.output 0.16993865 599.30546875 + ------------------------------------------------------------------------------------- + TOTAL 0.08106423 248.63738866 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 155016 +BPFP 1.7810 bits/point +EBPFP 3.5619 equivalent bits/point +MSE 248.637389 +---------------------- -------------------------------------------------------- +Time: 3.072s Load: 0.005s, Pack+Encode: 1.888s, Decode+Unpack: 1.178s +---------------------- -------------------------------------------------------- +💾 Converting with 248.6374 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-346.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-346.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,976B, BPFP=0.8015 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,668B, BPFP=3.3293 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,212B, BPFP=1.3228 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,384B, BPFP=3.1224 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,332B, BPFP=1.5032 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,724B, BPFP=3.0161 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,556B, BPFP=1.3782 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,484B, BPFP=3.1385 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,804B, BPFP=2.3847 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,696B, BPFP=3.0116 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,204B, BPFP=0.7411 +⌛️ [2/4] FRONTEND: Frontend time: 1.753s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.308s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13093194 26.31918996 + layer.0.v_cache 0.00001635 0.00212582 + layer.1.k_cache 0.05235285 2.08114593 + layer.1.v_cache 0.00000496 0.00081732 + layer.2.k_cache 0.00512381 0.70997926 + layer.2.v_cache 0.00001647 0.00251220 + layer.3.k_cache 0.01138228 3.54485392 + layer.3.v_cache 0.00001850 0.00278138 + layer.4.k_cache 0.00081880 0.08447527 + layer.4.v_cache 0.00004796 0.00627819 + layer.4.output 0.02447439 516.10843152 + ------------------------------------------------------------------------------------- + TOTAL 0.02188439 214.44195176 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 175040 +BPFP 1.6586 bits/point +EBPFP 3.3172 equivalent bits/point +MSE 214.441952 +---------------------- -------------------------------------------------------- +Time: 3.065s Load: 0.004s, Pack+Encode: 1.753s, Decode+Unpack: 1.308s +---------------------- -------------------------------------------------------- +💾 Converting with 214.4420 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-348.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,648B, BPFP=0.9193 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,680B, BPFP=3.6946 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,440B, BPFP=1.4715 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,416B, BPFP=3.4446 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,336B, BPFP=1.6487 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,352B, BPFP=3.2342 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,668B, BPFP=1.5166 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,504B, BPFP=3.2642 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,472B, BPFP=2.4668 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,808B, BPFP=3.1266 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,036B, BPFP=0.8487 +⌛️ [2/4] FRONTEND: Frontend time: 1.798s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.263s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510638 26.03975784 + layer.0.v_cache 0.00001399 0.00215324 + layer.1.k_cache 0.05467150 1.75816403 + layer.1.v_cache 0.00000511 0.00080811 + layer.2.k_cache 0.00545186 0.70418071 + layer.2.v_cache 0.00001652 0.00268524 + layer.3.k_cache 0.02986256 2.36428234 + layer.3.v_cache 0.00001747 0.00275156 + layer.4.k_cache 0.00068266 0.07254853 + layer.4.v_cache 0.00005277 0.00604215 + layer.4.output 0.17642128 580.38601944 + ------------------------------------------------------------------------------------- + TOTAL 0.08357822 240.80326528 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 155360 +BPFP 1.8075 bits/point +EBPFP 3.6150 equivalent bits/point +MSE 240.803265 +---------------------- -------------------------------------------------------- +Time: 3.068s Load: 0.006s, Pack+Encode: 1.798s, Decode+Unpack: 1.263s +---------------------- -------------------------------------------------------- +💾 Converting with 240.8033 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-356.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-356.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,556B, BPFP=0.9245 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,920B, BPFP=3.6364 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,488B, BPFP=1.5195 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,888B, BPFP=3.4269 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,284B, BPFP=1.6810 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,380B, BPFP=3.3239 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,636B, BPFP=1.5495 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,816B, BPFP=3.4123 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,604B, BPFP=2.5576 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,304B, BPFP=3.3084 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,436B, BPFP=0.8243 +⌛️ [2/4] FRONTEND: Frontend time: 1.821s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.193s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08730992 25.24813407 + layer.0.v_cache 0.00001543 0.00222708 + layer.1.k_cache 0.05635426 2.29412069 + layer.1.v_cache 0.00000520 0.00085915 + layer.2.k_cache 0.00390536 0.65871558 + layer.2.v_cache 0.00001661 0.00251905 + layer.3.k_cache 0.06483487 2.83288277 + layer.3.v_cache 0.00002010 0.00292081 + layer.4.k_cache 0.00072652 0.07749906 + layer.4.v_cache 0.00004842 0.00594223 + layer.4.output 0.18845482 615.61450603 + ------------------------------------------------------------------------------------- + TOTAL 0.09014238 255.31925663 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 153312 +BPFP 1.8300 bits/point +EBPFP 3.6600 equivalent bits/point +MSE 255.319257 +---------------------- -------------------------------------------------------- +Time: 3.018s Load: 0.003s, Pack+Encode: 1.821s, Decode+Unpack: 1.193s +---------------------- -------------------------------------------------------- +💾 Converting with 255.3193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-377.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,024B, BPFP=0.9246 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,484B, BPFP=3.5579 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,508B, BPFP=1.4954 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,976B, BPFP=3.2114 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,072B, BPFP=1.6250 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,328B, BPFP=3.0625 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,568B, BPFP=1.5092 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,996B, BPFP=3.2160 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,632B, BPFP=2.4430 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,456B, BPFP=3.0919 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,368B, BPFP=0.8984 +⌛️ [2/4] FRONTEND: Frontend time: 1.789s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.215s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10784986 26.67228788 + layer.0.v_cache 0.00001387 0.00211077 + layer.1.k_cache 0.01409798 2.12561843 + layer.1.v_cache 0.00000489 0.00074707 + layer.2.k_cache 0.00254294 0.55514639 + layer.2.v_cache 0.00001653 0.00248989 + layer.3.k_cache 0.05709259 3.16010733 + layer.3.v_cache 0.00001736 0.00271840 + layer.4.k_cache 0.00062715 0.06982732 + layer.4.v_cache 0.00004758 0.00590890 + layer.4.output 0.22376562 659.16898634 + ------------------------------------------------------------------------------------- + TOTAL 0.10286295 273.33999216 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 132412 +BPFP 1.7897 bits/point +EBPFP 3.5795 equivalent bits/point +MSE 273.339992 +---------------------- -------------------------------------------------------- +Time: 3.008s Load: 0.004s, Pack+Encode: 1.789s, Decode+Unpack: 1.215s +---------------------- -------------------------------------------------------- +💾 Converting with 273.3400 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-383.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-383.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,692B, BPFP=0.9280 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,664B, BPFP=3.2959 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,384B, BPFP=1.4604 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,312B, BPFP=3.2263 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,160B, BPFP=1.6139 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,508B, BPFP=3.0672 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,492B, BPFP=1.4818 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,824B, BPFP=3.3275 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,224B, BPFP=2.4177 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,416B, BPFP=3.0491 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,860B, BPFP=0.8437 +⌛️ [2/4] FRONTEND: Frontend time: 1.876s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14563359 26.14101130 + layer.0.v_cache 0.00001368 0.00194646 + layer.1.k_cache 0.05947259 2.16072759 + layer.1.v_cache 0.00000496 0.00077287 + layer.2.k_cache 0.00553216 0.65350878 + layer.2.v_cache 0.00001557 0.00238647 + layer.3.k_cache 0.04377548 2.85921739 + layer.3.v_cache 0.00001730 0.00272706 + layer.4.k_cache 0.00079035 0.07820564 + layer.4.v_cache 0.00005153 0.00617522 + layer.4.output 0.17347779 607.57781420 + ------------------------------------------------------------------------------------- + TOTAL 0.08645010 252.05596342 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 150536 +BPFP 1.7514 bits/point +EBPFP 3.5028 equivalent bits/point +MSE 252.055963 +---------------------- -------------------------------------------------------- +Time: 3.223s Load: 0.005s, Pack+Encode: 1.876s, Decode+Unpack: 1.343s +---------------------- -------------------------------------------------------- +💾 Converting with 252.0560 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-386.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-386.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,792B, BPFP=0.8977 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,676B, BPFP=3.4744 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,116B, BPFP=1.4479 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,156B, BPFP=3.1146 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,904B, BPFP=1.6345 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,712B, BPFP=3.0095 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,300B, BPFP=1.4915 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,476B, BPFP=3.1903 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,192B, BPFP=2.4129 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,544B, BPFP=2.9697 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,684B, BPFP=0.9025 +⌛️ [2/4] FRONTEND: Frontend time: 1.760s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.210s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12373332 26.62793709 + layer.0.v_cache 0.00001436 0.00222390 + layer.1.k_cache 0.01270404 1.93224312 + layer.1.v_cache 0.00000490 0.00080490 + layer.2.k_cache 0.00252672 0.65508952 + layer.2.v_cache 0.00001629 0.00262670 + layer.3.k_cache 0.05185594 2.71361380 + layer.3.v_cache 0.00001849 0.00285070 + layer.4.k_cache 0.00068458 0.07528558 + layer.4.v_cache 0.00004975 0.00591837 + layer.4.output 0.22283003 728.75196158 + ------------------------------------------------------------------------------------- + TOTAL 0.10302462 301.95778381 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 126552 +BPFP 1.7624 bits/point +EBPFP 3.5247 equivalent bits/point +MSE 301.957784 +---------------------- -------------------------------------------------------- +Time: 2.974s Load: 0.004s, Pack+Encode: 1.760s, Decode+Unpack: 1.210s +---------------------- -------------------------------------------------------- +💾 Converting with 301.9578 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-389.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,860B, BPFP=0.9040 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,708B, BPFP=3.6659 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,868B, BPFP=1.4635 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,840B, BPFP=3.5045 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,104B, BPFP=1.6935 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,752B, BPFP=3.3021 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,396B, BPFP=1.5618 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,780B, BPFP=3.4933 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,832B, BPFP=2.5729 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,308B, BPFP=3.4055 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,236B, BPFP=0.8832 +⌛️ [2/4] FRONTEND: Frontend time: 1.949s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12002775 27.43219285 + layer.0.v_cache 0.00001466 0.00221178 + layer.1.k_cache 0.05271155 1.95133246 + layer.1.v_cache 0.00000524 0.00083632 + layer.2.k_cache 0.00808975 0.73519184 + layer.2.v_cache 0.00001826 0.00261894 + layer.3.k_cache 0.07501186 2.74657422 + layer.3.v_cache 0.00001895 0.00298723 + layer.4.k_cache 0.00071216 0.07783877 + layer.4.v_cache 0.00005170 0.00615528 + layer.4.output 0.17053346 561.88196216 + ------------------------------------------------------------------------------------- + TOTAL 0.08531742 233.30186322 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 170684 +BPFP 1.8676 bits/point +EBPFP 3.7352 equivalent bits/point +MSE 233.301863 +---------------------- -------------------------------------------------------- +Time: 3.160s Load: 0.003s, Pack+Encode: 1.949s, Decode+Unpack: 1.208s +---------------------- -------------------------------------------------------- +💾 Converting with 233.3019 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-390.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-390.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 75, 128) +Output shape: (1, 75, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) -> torch.Size([1, 1, 75, 512]) + layer.4.output: torch.Size([1, 75, 3584]) -> torch.Size([1, 1, 75, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,476B, BPFP=0.9325 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,464B, BPFP=3.6383 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,208B, BPFP=1.5017 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,892B, BPFP=3.3108 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,044B, BPFP=1.6758 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,232B, BPFP=3.1733 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,400B, BPFP=1.5417 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,168B, BPFP=3.3683 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,156B, BPFP=2.5325 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,388B, BPFP=3.2058 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,912B, BPFP=0.8605 +⌛️ [2/4] FRONTEND: Frontend time: 1.721s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) +⌛️ [3/4] BACKEND: Backend time: 33.346s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 75, 128]) + layer.0.v_cache: torch.Size([1, 4, 75, 128]) + layer.1.k_cache: torch.Size([1, 4, 75, 128]) + layer.1.v_cache: torch.Size([1, 4, 75, 128]) + layer.2.k_cache: torch.Size([1, 4, 75, 128]) + layer.2.v_cache: torch.Size([1, 4, 75, 128]) + layer.3.k_cache: torch.Size([1, 4, 75, 128]) + layer.3.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.k_cache: torch.Size([1, 4, 75, 128]) + layer.4.v_cache: torch.Size([1, 4, 75, 128]) + layer.4.output: torch.Size([1, 75, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08149677 24.87530762 + layer.0.v_cache 0.00001391 0.00224882 + layer.1.k_cache 0.05841089 1.85051270 + layer.1.v_cache 0.00000541 0.00082185 + layer.2.k_cache 0.00252637 0.63826284 + layer.2.v_cache 0.00001693 0.00244696 + layer.3.k_cache 0.02876982 2.53929179 + layer.3.v_cache 0.00001932 0.00292446 + layer.4.k_cache 0.00075246 0.07646200 + layer.4.v_cache 0.00005041 0.00625710 + layer.4.output 0.19266937 556.88529762 + ------------------------------------------------------------------------------------- + TOTAL 0.08945576 231.07009526 + (elements=652,800) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 652800 +Total Bytes 148340 +BPFP 1.8179 bits/point +EBPFP 3.6358 equivalent bits/point +MSE 231.070095 +---------------------- --------------------------------------------------------- +Time: 35.070s Load: 0.003s, Pack+Encode: 1.721s, Decode+Unpack: 33.346s +---------------------- --------------------------------------------------------- +💾 Converting with 231.0701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-395.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-395.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,008B, BPFP=0.9210 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,336B, BPFP=3.5239 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,500B, BPFP=1.4936 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,476B, BPFP=3.0965 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,168B, BPFP=1.6471 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,848B, BPFP=2.9522 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,576B, BPFP=1.5110 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,736B, BPFP=3.1562 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,820B, BPFP=2.4862 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,292B, BPFP=3.0542 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,120B, BPFP=0.8574 +⌛️ [2/4] FRONTEND: Frontend time: 2.012s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.294s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10404026 26.83342249 + layer.0.v_cache 0.00001395 0.00220169 + layer.1.k_cache 0.01458383 2.08114893 + layer.1.v_cache 0.00000499 0.00079665 + layer.2.k_cache 0.00236740 0.58607932 + layer.2.v_cache 0.00001641 0.00248655 + layer.3.k_cache 0.03645969 2.81769382 + layer.3.v_cache 0.00002420 0.00295457 + layer.4.k_cache 0.00069996 0.07244358 + layer.4.v_cache 0.00005119 0.00634112 + layer.4.output 0.20980707 683.26818540 + ------------------------------------------------------------------------------------- + TOTAL 0.09570067 283.25193333 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 129880 +BPFP 1.7555 bits/point +EBPFP 3.5110 equivalent bits/point +MSE 283.251933 +---------------------- -------------------------------------------------------- +Time: 3.311s Load: 0.004s, Pack+Encode: 2.012s, Decode+Unpack: 1.294s +---------------------- -------------------------------------------------------- +💾 Converting with 283.2519 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-396.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-396.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,236B, BPFP=0.9455 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,036B, BPFP=3.5795 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,852B, BPFP=1.5295 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,896B, BPFP=3.3250 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,576B, BPFP=1.6911 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,244B, BPFP=3.1795 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,996B, BPFP=1.5616 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,496B, BPFP=3.2357 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,288B, BPFP=2.5196 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,856B, BPFP=3.3161 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,224B, BPFP=0.9319 +⌛️ [2/4] FRONTEND: Frontend time: 1.763s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.292s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11725240 26.66067592 + layer.0.v_cache 0.00001567 0.00223607 + layer.1.k_cache 0.03693308 1.98492562 + layer.1.v_cache 0.00000533 0.00090055 + layer.2.k_cache 0.00240566 0.57227413 + layer.2.v_cache 0.00001850 0.00286942 + layer.3.k_cache 0.06751672 2.99325997 + layer.3.v_cache 0.00001909 0.00302729 + layer.4.k_cache 0.00067984 0.07467525 + layer.4.v_cache 0.00004901 0.00661597 + layer.4.output 0.20094273 684.46505102 + ------------------------------------------------------------------------------------- + TOTAL 0.09597026 283.73863631 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 140700 +BPFP 1.8474 bits/point +EBPFP 3.6949 equivalent bits/point +MSE 283.738636 +---------------------- -------------------------------------------------------- +Time: 3.058s Load: 0.003s, Pack+Encode: 1.763s, Decode+Unpack: 1.292s +---------------------- -------------------------------------------------------- +💾 Converting with 283.7386 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-401.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-401.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,248B, BPFP=0.9482 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,224B, BPFP=3.3982 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,800B, BPFP=1.5179 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,408B, BPFP=3.2161 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,440B, BPFP=1.6607 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,312B, BPFP=2.9714 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,924B, BPFP=1.5455 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,084B, BPFP=3.1437 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,004B, BPFP=2.4562 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,328B, BPFP=2.9750 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,452B, BPFP=0.8435 +⌛️ [2/4] FRONTEND: Frontend time: 1.787s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.149s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07685234 26.58529227 + layer.0.v_cache 0.00001390 0.00222461 + layer.1.k_cache 0.01187498 2.17370714 + layer.1.v_cache 0.00000934 0.00092627 + layer.2.k_cache 0.00239603 0.73646251 + layer.2.v_cache 0.00001621 0.00243487 + layer.3.k_cache 0.07241479 2.96916983 + layer.3.v_cache 0.00001716 0.00294413 + layer.4.k_cache 0.00072067 0.07517591 + layer.4.v_cache 0.00004958 0.00606322 + layer.4.output 0.21417375 697.11358418 + ------------------------------------------------------------------------------------- + TOTAL 0.09785772 288.96173471 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 133224 +BPFP 1.7493 bits/point +EBPFP 3.4985 equivalent bits/point +MSE 288.961735 +---------------------- -------------------------------------------------------- +Time: 2.940s Load: 0.003s, Pack+Encode: 1.787s, Decode+Unpack: 1.149s +---------------------- -------------------------------------------------------- +💾 Converting with 288.9617 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-408.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-408.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,496B, BPFP=0.9243 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,680B, BPFP=3.6349 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,388B, BPFP=1.5189 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,336B, BPFP=3.5641 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,344B, BPFP=1.7155 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,664B, BPFP=3.4260 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,708B, BPFP=1.5847 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,176B, BPFP=3.5312 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,568B, BPFP=2.5839 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,672B, BPFP=3.4276 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,840B, BPFP=0.9058 +⌛️ [2/4] FRONTEND: Frontend time: 1.911s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.230s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09264422 27.31886051 + layer.0.v_cache 0.00001680 0.00224402 + layer.1.k_cache 0.05606840 1.69380007 + layer.1.v_cache 0.00000542 0.00087080 + layer.2.k_cache 0.00241068 0.58966667 + layer.2.v_cache 0.00001702 0.00261564 + layer.3.k_cache 0.15270803 2.71233288 + layer.3.v_cache 0.00001849 0.00304447 + layer.4.k_cache 0.00070970 0.07541284 + layer.4.v_cache 0.00005539 0.00656004 + layer.4.output 0.18838861 472.46687030 + ------------------------------------------------------------------------------------- + TOTAL 0.09549261 196.45138236 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 156872 +BPFP 1.8972 bits/point +EBPFP 3.7943 equivalent bits/point +MSE 196.451382 +---------------------- -------------------------------------------------------- +Time: 3.144s Load: 0.003s, Pack+Encode: 1.911s, Decode+Unpack: 1.230s +---------------------- -------------------------------------------------------- +💾 Converting with 196.4514 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-417.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-417.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,544B, BPFP=0.9221 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,500B, BPFP=3.5511 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,348B, BPFP=1.4911 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,336B, BPFP=3.3149 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,144B, BPFP=1.6526 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,016B, BPFP=3.2500 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,504B, BPFP=1.5227 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,088B, BPFP=3.4675 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,392B, BPFP=2.5146 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,344B, BPFP=3.3166 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,924B, BPFP=0.7805 +⌛️ [2/4] FRONTEND: Frontend time: 2.186s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.485s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10144231 26.66100916 + layer.0.v_cache 0.00001330 0.00203498 + layer.1.k_cache 0.09893422 2.08336748 + layer.1.v_cache 0.00000513 0.00079550 + layer.2.k_cache 0.00368564 0.74079063 + layer.2.v_cache 0.00001777 0.00253485 + layer.3.k_cache 0.06341467 2.64408458 + layer.3.v_cache 0.00001739 0.00276205 + layer.4.k_cache 0.00073693 0.08092062 + layer.4.v_cache 0.00004994 0.00589022 + layer.4.output 0.18261438 632.71770640 + ------------------------------------------------------------------------------------- + TOTAL 0.09097753 262.42636088 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 150140 +BPFP 1.7922 bits/point +EBPFP 3.5843 equivalent bits/point +MSE 262.426361 +---------------------- -------------------------------------------------------- +Time: 3.675s Load: 0.004s, Pack+Encode: 2.186s, Decode+Unpack: 1.485s +---------------------- -------------------------------------------------------- +💾 Converting with 262.4264 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-422.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,852B, BPFP=0.8714 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,448B, BPFP=3.3132 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,768B, BPFP=1.3951 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,508B, BPFP=3.1444 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,548B, BPFP=1.5352 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,728B, BPFP=3.0043 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,016B, BPFP=1.4397 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,912B, BPFP=3.2170 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,232B, BPFP=2.3764 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,484B, BPFP=3.1401 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,804B, BPFP=0.7390 +⌛️ [2/4] FRONTEND: Frontend time: 1.692s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.248s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08947531 25.37238461 + layer.0.v_cache 0.00001432 0.00211509 + layer.1.k_cache 0.05320675 1.65800774 + layer.1.v_cache 0.00000497 0.00077247 + layer.2.k_cache 0.00562141 0.64658856 + layer.2.v_cache 0.00001574 0.00240910 + layer.3.k_cache 0.02742601 3.19312365 + layer.3.v_cache 0.00001759 0.00279038 + layer.4.k_cache 0.00114861 0.07993178 + layer.4.v_cache 0.00004734 0.00613808 + layer.4.output 0.15628038 536.87961823 + ------------------------------------------------------------------------------------- + TOTAL 0.07476122 222.88950524 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 159300 +BPFP 1.6829 bits/point +EBPFP 3.3659 equivalent bits/point +MSE 222.889505 +---------------------- -------------------------------------------------------- +Time: 2.943s Load: 0.004s, Pack+Encode: 1.692s, Decode+Unpack: 1.248s +---------------------- -------------------------------------------------------- +💾 Converting with 222.8895 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-432.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-432.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,020B, BPFP=0.8526 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,800B, BPFP=3.3628 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,032B, BPFP=1.3641 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,784B, BPFP=3.1902 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,916B, BPFP=1.5143 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,052B, BPFP=2.8961 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,284B, BPFP=1.4069 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,112B, BPFP=3.0761 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,792B, BPFP=2.3424 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,972B, BPFP=3.0523 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,760B, BPFP=0.6735 +⌛️ [2/4] FRONTEND: Frontend time: 2.277s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12311893 27.31948454 + layer.0.v_cache 0.00001357 0.00205809 + layer.1.k_cache 0.03743907 1.89736972 + layer.1.v_cache 0.00000515 0.00079091 + layer.2.k_cache 0.00511221 0.66208740 + layer.2.v_cache 0.00001545 0.00226960 + layer.3.k_cache 0.04225797 3.24134031 + layer.3.v_cache 0.00001799 0.00264949 + layer.4.k_cache 0.00119426 0.07792323 + layer.4.v_cache 0.00004621 0.00579136 + layer.4.output 0.14783020 388.75019410 + ------------------------------------------------------------------------------------- + TOTAL 0.07317837 162.02724255 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 163524 +BPFP 1.6337 bits/point +EBPFP 3.2673 equivalent bits/point +MSE 162.027243 +---------------------- -------------------------------------------------------- +Time: 3.600s Load: 0.004s, Pack+Encode: 2.277s, Decode+Unpack: 1.319s +---------------------- -------------------------------------------------------- +💾 Converting with 162.0272 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-435.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 102, 128) +Output shape: (1, 102, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) -> torch.Size([1, 1, 102, 512]) + layer.4.output: torch.Size([1, 102, 3584]) -> torch.Size([1, 1, 102, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,392B, BPFP=0.8260 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,256B, BPFP=3.4093 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,148B, BPFP=1.4013 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,868B, BPFP=3.1967 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,396B, BPFP=1.5925 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,992B, BPFP=3.0625 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,416B, BPFP=1.4424 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,020B, BPFP=3.2200 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,616B, BPFP=2.5453 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,372B, BPFP=3.1207 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,792B, BPFP=0.8270 +⌛️ [2/4] FRONTEND: Frontend time: 1.901s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.087s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 102, 128]) + layer.0.v_cache: torch.Size([1, 4, 102, 128]) + layer.1.k_cache: torch.Size([1, 4, 102, 128]) + layer.1.v_cache: torch.Size([1, 4, 102, 128]) + layer.2.k_cache: torch.Size([1, 4, 102, 128]) + layer.2.v_cache: torch.Size([1, 4, 102, 128]) + layer.3.k_cache: torch.Size([1, 4, 102, 128]) + layer.3.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.k_cache: torch.Size([1, 4, 102, 128]) + layer.4.v_cache: torch.Size([1, 4, 102, 128]) + layer.4.output: torch.Size([1, 102, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09293266 29.64452646 + layer.0.v_cache 0.00001503 0.00208213 + layer.1.k_cache 0.07854857 2.26485279 + layer.1.v_cache 0.00000559 0.00080090 + layer.2.k_cache 0.00650115 0.59062726 + layer.2.v_cache 0.00001763 0.00245202 + layer.3.k_cache 0.03715762 3.44669417 + layer.3.v_cache 0.00001788 0.00269523 + layer.4.k_cache 0.00073365 0.08376683 + layer.4.v_cache 0.00004630 0.00551165 + layer.4.output 11.17680568 468.71161590 + ------------------------------------------------------------------------------------- + TOTAL 4.61491858 195.11913651 + (elements=887,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 887808 +Total Bytes 193268 +BPFP 1.7415 bits/point +EBPFP 3.4831 equivalent bits/point +MSE 195.119137 +---------------------- -------------------------------------------------------- +Time: 2.993s Load: 0.005s, Pack+Encode: 1.901s, Decode+Unpack: 1.087s +---------------------- -------------------------------------------------------- +💾 Converting with 195.1191 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-438.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 33.978s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 96, 128) +Output shape: (1, 96, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) -> torch.Size([1, 1, 96, 512]) + layer.4.output: torch.Size([1, 96, 3584]) -> torch.Size([1, 1, 96, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,920B, BPFP=0.8008 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,716B, BPFP=3.5345 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,152B, BPFP=1.3268 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,504B, BPFP=3.3372 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,836B, BPFP=1.6009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,848B, BPFP=3.2305 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,148B, BPFP=1.4889 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,728B, BPFP=3.3737 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,576B, BPFP=2.5352 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,924B, BPFP=3.2428 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,012B, BPFP=0.8141 +⌛️ [2/4] FRONTEND: Frontend time: 2.227s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.476s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 96, 128]) + layer.0.v_cache: torch.Size([1, 4, 96, 128]) + layer.1.k_cache: torch.Size([1, 4, 96, 128]) + layer.1.v_cache: torch.Size([1, 4, 96, 128]) + layer.2.k_cache: torch.Size([1, 4, 96, 128]) + layer.2.v_cache: torch.Size([1, 4, 96, 128]) + layer.3.k_cache: torch.Size([1, 4, 96, 128]) + layer.3.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.k_cache: torch.Size([1, 4, 96, 128]) + layer.4.v_cache: torch.Size([1, 4, 96, 128]) + layer.4.output: torch.Size([1, 96, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10158994 27.53300985 + layer.0.v_cache 0.00001467 0.00227949 + layer.1.k_cache 0.06416908 1.87511428 + layer.1.v_cache 0.00000542 0.00083341 + layer.2.k_cache 0.00400108 0.72091524 + layer.2.v_cache 0.00001872 0.00262187 + layer.3.k_cache 0.01491791 3.02407869 + layer.3.v_cache 0.00001877 0.00301105 + layer.4.k_cache 0.00070988 0.08558730 + layer.4.v_cache 0.00004832 0.00635745 + layer.4.output 0.14169503 501.71098400 + ------------------------------------------------------------------------------------- + TOTAL 0.06925641 208.54298216 + (elements=835,584) +---------------------- --------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- --------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- --------------------------------------------------------- +Total Elements 835584 +Total Bytes 185364 +BPFP 1.7747 bits/point +EBPFP 3.5494 equivalent bits/point +MSE 208.542982 +---------------------- --------------------------------------------------------- +Time: 37.681s Load: 33.978s, Pack+Encode: 2.227s, Decode+Unpack: 1.476s +---------------------- --------------------------------------------------------- +💾 Converting with 208.5430 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-44.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-44.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,340B, BPFP=0.9418 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,016B, BPFP=3.6927 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,956B, BPFP=1.5095 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,968B, BPFP=3.2483 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,836B, BPFP=1.7005 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,608B, BPFP=3.1701 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,172B, BPFP=1.5564 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,912B, BPFP=3.2361 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,668B, BPFP=2.5321 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,932B, BPFP=3.2405 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,536B, BPFP=0.8537 +⌛️ [2/4] FRONTEND: Frontend time: 2.136s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.230s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10321087 27.29733615 + layer.0.v_cache 0.00001557 0.00226932 + layer.1.k_cache 0.06175087 2.08916389 + layer.1.v_cache 0.00000571 0.00082483 + layer.2.k_cache 0.01329466 0.71558073 + layer.2.v_cache 0.00001784 0.00251433 + layer.3.k_cache 0.05039416 2.86118698 + layer.3.v_cache 0.00001745 0.00279867 + layer.4.k_cache 0.00079786 0.07332112 + layer.4.v_cache 0.00005071 0.00603006 + layer.4.output 0.18870974 663.11303323 + ------------------------------------------------------------------------------------- + TOTAL 0.09120729 274.99072110 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 141944 +BPFP 1.8120 bits/point +EBPFP 3.6240 equivalent bits/point +MSE 274.990721 +---------------------- -------------------------------------------------------- +Time: 3.373s Load: 0.006s, Pack+Encode: 2.136s, Decode+Unpack: 1.230s +---------------------- -------------------------------------------------------- +💾 Converting with 274.9907 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-440.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 72, 128) +Output shape: (1, 72, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) -> torch.Size([1, 1, 72, 512]) + layer.4.output: torch.Size([1, 72, 3584]) -> torch.Size([1, 1, 72, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,388B, BPFP=0.9523 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,864B, BPFP=3.4427 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,924B, BPFP=1.5026 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,100B, BPFP=3.0599 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,400B, BPFP=1.6059 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,376B, BPFP=2.9028 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,856B, BPFP=1.4878 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,392B, BPFP=3.1233 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,156B, BPFP=2.4210 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,508B, BPFP=2.9314 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,780B, BPFP=0.8612 +⌛️ [2/4] FRONTEND: Frontend time: 1.853s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.275s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 72, 128]) + layer.0.v_cache: torch.Size([1, 4, 72, 128]) + layer.1.k_cache: torch.Size([1, 4, 72, 128]) + layer.1.v_cache: torch.Size([1, 4, 72, 128]) + layer.2.k_cache: torch.Size([1, 4, 72, 128]) + layer.2.v_cache: torch.Size([1, 4, 72, 128]) + layer.3.k_cache: torch.Size([1, 4, 72, 128]) + layer.3.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.k_cache: torch.Size([1, 4, 72, 128]) + layer.4.v_cache: torch.Size([1, 4, 72, 128]) + layer.4.output: torch.Size([1, 72, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09666174 26.02468533 + layer.0.v_cache 0.00001715 0.00222466 + layer.1.k_cache 0.08476152 1.99437926 + layer.1.v_cache 0.00000506 0.00076155 + layer.2.k_cache 0.00238475 0.72008853 + layer.2.v_cache 0.00001567 0.00233516 + layer.3.k_cache 0.07332428 2.60768021 + layer.3.v_cache 0.00001796 0.00267844 + layer.4.k_cache 0.00072290 0.07723998 + layer.4.v_cache 0.00005092 0.00572587 + layer.4.output 0.18875189 695.11966766 + ------------------------------------------------------------------------------------- + TOTAL 0.09289560 288.07502780 + (elements=626,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 626688 +Total Bytes 135744 +BPFP 1.7328 bits/point +EBPFP 3.4657 equivalent bits/point +MSE 288.075028 +---------------------- -------------------------------------------------------- +Time: 3.133s Load: 0.005s, Pack+Encode: 1.853s, Decode+Unpack: 1.275s +---------------------- -------------------------------------------------------- +💾 Converting with 288.0750 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-443.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 68, 128) +Output shape: (1, 68, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) -> torch.Size([1, 1, 68, 512]) + layer.4.output: torch.Size([1, 68, 3584]) -> torch.Size([1, 1, 68, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,048B, BPFP=0.9301 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,564B, BPFP=3.5763 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,592B, BPFP=1.5147 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,484B, BPFP=3.3281 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,308B, BPFP=1.6792 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,296B, BPFP=3.0551 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,740B, BPFP=1.5487 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,336B, BPFP=3.2941 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,844B, BPFP=2.4917 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,684B, BPFP=3.1443 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,328B, BPFP=0.8971 +⌛️ [2/4] FRONTEND: Frontend time: 1.772s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.323s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 68, 128]) + layer.0.v_cache: torch.Size([1, 4, 68, 128]) + layer.1.k_cache: torch.Size([1, 4, 68, 128]) + layer.1.v_cache: torch.Size([1, 4, 68, 128]) + layer.2.k_cache: torch.Size([1, 4, 68, 128]) + layer.2.v_cache: torch.Size([1, 4, 68, 128]) + layer.3.k_cache: torch.Size([1, 4, 68, 128]) + layer.3.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.k_cache: torch.Size([1, 4, 68, 128]) + layer.4.v_cache: torch.Size([1, 4, 68, 128]) + layer.4.output: torch.Size([1, 68, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10601801 24.32851634 + layer.0.v_cache 0.00001817 0.00226153 + layer.1.k_cache 0.01696859 2.29509331 + layer.1.v_cache 0.00000501 0.00084198 + layer.2.k_cache 0.00274148 0.67252294 + layer.2.v_cache 0.00001757 0.00248477 + layer.3.k_cache 0.01706722 3.35971024 + layer.3.v_cache 0.00001843 0.00284144 + layer.4.k_cache 0.00080906 0.07268589 + layer.4.v_cache 0.00005076 0.00621913 + layer.4.output 0.19983917 669.33836660 + ------------------------------------------------------------------------------------- + TOTAL 0.09074050 277.41833787 + (elements=591,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 591872 +Total Bytes 134224 +BPFP 1.8142 bits/point +EBPFP 3.6285 equivalent bits/point +MSE 277.418338 +---------------------- -------------------------------------------------------- +Time: 3.098s Load: 0.004s, Pack+Encode: 1.772s, Decode+Unpack: 1.323s +---------------------- -------------------------------------------------------- +💾 Converting with 277.4183 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-444.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-444.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 65, 128) +Output shape: (1, 65, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) -> torch.Size([1, 1, 65, 512]) + layer.4.output: torch.Size([1, 65, 3584]) -> torch.Size([1, 1, 65, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,632B, BPFP=0.8731 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 13,892B, BPFP=3.3394 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,792B, BPFP=1.3923 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,028B, BPFP=3.1317 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,668B, BPFP=1.6029 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 12,424B, BPFP=2.9865 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,180B, BPFP=1.4856 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,276B, BPFP=3.1913 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,292B, BPFP=2.4740 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 12,860B, BPFP=3.0913 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,992B, BPFP=0.8926 +⌛️ [2/4] FRONTEND: Frontend time: 1.762s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.292s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 65, 128]) + layer.0.v_cache: torch.Size([1, 4, 65, 128]) + layer.1.k_cache: torch.Size([1, 4, 65, 128]) + layer.1.v_cache: torch.Size([1, 4, 65, 128]) + layer.2.k_cache: torch.Size([1, 4, 65, 128]) + layer.2.v_cache: torch.Size([1, 4, 65, 128]) + layer.3.k_cache: torch.Size([1, 4, 65, 128]) + layer.3.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.k_cache: torch.Size([1, 4, 65, 128]) + layer.4.v_cache: torch.Size([1, 4, 65, 128]) + layer.4.output: torch.Size([1, 65, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09579474 23.76775841 + layer.0.v_cache 0.00001312 0.00210531 + layer.1.k_cache 0.01648773 2.09558082 + layer.1.v_cache 0.00000712 0.00081475 + layer.2.k_cache 0.00431219 0.63942049 + layer.2.v_cache 0.00001803 0.00273936 + layer.3.k_cache 0.03460605 2.51811500 + layer.3.v_cache 0.00001721 0.00286860 + layer.4.k_cache 0.00068162 0.07923225 + layer.4.v_cache 0.00004727 0.00638679 + layer.4.output 0.20898957 579.38186813 + ------------------------------------------------------------------------------------- + TOTAL 0.09499483 240.28165287 + (elements=565,760) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 565760 +Total Bytes 124036 +BPFP 1.7539 bits/point +EBPFP 3.5078 equivalent bits/point +MSE 240.281653 +---------------------- -------------------------------------------------------- +Time: 3.057s Load: 0.003s, Pack+Encode: 1.762s, Decode+Unpack: 1.292s +---------------------- -------------------------------------------------------- +💾 Converting with 240.2817 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-452.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 41, 128) +Output shape: (1, 41, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) -> torch.Size([1, 1, 41, 512]) + layer.4.output: torch.Size([1, 41, 3584]) -> torch.Size([1, 1, 41, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,012B, BPFP=1.1479 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,864B, BPFP=4.5213 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 4,696B, BPFP=1.7896 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,056B, BPFP=4.2134 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 5,296B, BPFP=2.0183 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,260B, BPFP=4.2912 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 4,752B, BPFP=1.8110 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,264B, BPFP=4.2927 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 8,680B, BPFP=3.3079 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,864B, BPFP=4.1402 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 21,208B, BPFP=1.1546 +⌛️ [2/4] FRONTEND: Frontend time: 1.811s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.063s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 41, 128]) + layer.0.v_cache: torch.Size([1, 4, 41, 128]) + layer.1.k_cache: torch.Size([1, 4, 41, 128]) + layer.1.v_cache: torch.Size([1, 4, 41, 128]) + layer.2.k_cache: torch.Size([1, 4, 41, 128]) + layer.2.v_cache: torch.Size([1, 4, 41, 128]) + layer.3.k_cache: torch.Size([1, 4, 41, 128]) + layer.3.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.k_cache: torch.Size([1, 4, 41, 128]) + layer.4.v_cache: torch.Size([1, 4, 41, 128]) + layer.4.output: torch.Size([1, 41, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10098721 24.07329280 + layer.0.v_cache 0.00001362 0.00267799 + layer.1.k_cache 0.01708067 2.24595865 + layer.1.v_cache 0.00000533 0.00100085 + layer.2.k_cache 0.00250472 0.64680495 + layer.2.v_cache 0.00001755 0.00313564 + layer.3.k_cache 0.08199174 2.78344243 + layer.3.v_cache 0.00001815 0.00331195 + layer.4.k_cache 0.00059437 0.09549143 + layer.4.v_cache 0.00005232 0.00819694 + layer.4.output 0.35770382 1149.96178136 + ------------------------------------------------------------------------------------- + TOTAL 0.15924661 475.27034019 + (elements=356,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 356864 +Total Bytes 103952 +BPFP 2.3303 bits/point +EBPFP 4.6607 equivalent bits/point +MSE 475.270340 +---------------------- -------------------------------------------------------- +Time: 2.877s Load: 0.002s, Pack+Encode: 1.811s, Decode+Unpack: 1.063s +---------------------- -------------------------------------------------------- +💾 Converting with 475.2703 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-462.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 52, 128) +Output shape: (1, 52, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) -> torch.Size([1, 1, 52, 512]) + layer.4.output: torch.Size([1, 52, 3584]) -> torch.Size([1, 1, 52, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,312B, BPFP=0.9952 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,808B, BPFP=3.5481 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,348B, BPFP=1.6070 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,556B, BPFP=3.4724 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,548B, BPFP=1.9675 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,224B, BPFP=3.3726 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,728B, BPFP=1.7212 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,520B, BPFP=3.4615 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,392B, BPFP=2.8221 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 10,952B, BPFP=3.2909 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,124B, BPFP=1.0355 +⌛️ [2/4] FRONTEND: Frontend time: 1.730s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.189s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 52, 128]) + layer.0.v_cache: torch.Size([1, 4, 52, 128]) + layer.1.k_cache: torch.Size([1, 4, 52, 128]) + layer.1.v_cache: torch.Size([1, 4, 52, 128]) + layer.2.k_cache: torch.Size([1, 4, 52, 128]) + layer.2.v_cache: torch.Size([1, 4, 52, 128]) + layer.3.k_cache: torch.Size([1, 4, 52, 128]) + layer.3.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.k_cache: torch.Size([1, 4, 52, 128]) + layer.4.v_cache: torch.Size([1, 4, 52, 128]) + layer.4.output: torch.Size([1, 52, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09210638 24.36107459 + layer.0.v_cache 0.00001369 0.00244473 + layer.1.k_cache 0.01682195 2.52669877 + layer.1.v_cache 0.00000520 0.00098464 + layer.2.k_cache 0.00487203 0.65045929 + layer.2.v_cache 0.00001755 0.00318707 + layer.3.k_cache 0.04965206 3.24184506 + layer.3.v_cache 0.00001944 0.00330120 + layer.4.k_cache 0.00060442 0.09092361 + layer.4.v_cache 0.00005207 0.00705821 + layer.4.output 0.26182200 872.46720467 + ------------------------------------------------------------------------------------- + TOTAL 0.11746581 361.06814176 + (elements=452,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 452608 +Total Bytes 111512 +BPFP 1.9710 bits/point +EBPFP 3.9420 equivalent bits/point +MSE 361.068142 +---------------------- -------------------------------------------------------- +Time: 2.922s Load: 0.003s, Pack+Encode: 1.730s, Decode+Unpack: 1.189s +---------------------- -------------------------------------------------------- +💾 Converting with 361.0681 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-464.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.002s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 47, 128) +Output shape: (1, 47, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) -> torch.Size([1, 1, 47, 512]) + layer.4.output: torch.Size([1, 47, 3584]) -> torch.Size([1, 1, 47, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,252B, BPFP=1.0811 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,944B, BPFP=3.9707 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,360B, BPFP=1.7819 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,828B, BPFP=3.9322 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,544B, BPFP=2.1755 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,468B, BPFP=3.8125 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,436B, BPFP=1.8072 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,648B, BPFP=3.8723 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,660B, BPFP=3.2114 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,292B, BPFP=3.7540 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,004B, BPFP=1.0925 +⌛️ [2/4] FRONTEND: Frontend time: 1.726s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.208s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 47, 128]) + layer.0.v_cache: torch.Size([1, 4, 47, 128]) + layer.1.k_cache: torch.Size([1, 4, 47, 128]) + layer.1.v_cache: torch.Size([1, 4, 47, 128]) + layer.2.k_cache: torch.Size([1, 4, 47, 128]) + layer.2.v_cache: torch.Size([1, 4, 47, 128]) + layer.3.k_cache: torch.Size([1, 4, 47, 128]) + layer.3.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.k_cache: torch.Size([1, 4, 47, 128]) + layer.4.v_cache: torch.Size([1, 4, 47, 128]) + layer.4.output: torch.Size([1, 47, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08461995 23.30598872 + layer.0.v_cache 0.00001366 0.00279808 + layer.1.k_cache 0.01838745 2.10650326 + layer.1.v_cache 0.00000609 0.00122023 + layer.2.k_cache 0.00780689 0.59445105 + layer.2.v_cache 0.00001932 0.00342030 + layer.3.k_cache 0.04292001 2.54348300 + layer.3.v_cache 0.00001942 0.00332742 + layer.4.k_cache 0.00059704 0.09103151 + layer.4.v_cache 0.00004850 0.00764389 + layer.4.output 0.29441516 1034.74116641 + ------------------------------------------------------------------------------------- + TOTAL 0.13031438 427.75576661 + (elements=409,088) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 409088 +Total Bytes 111436 +BPFP 2.1792 bits/point +EBPFP 4.3584 equivalent bits/point +MSE 427.755767 +---------------------- -------------------------------------------------------- +Time: 2.936s Load: 0.002s, Pack+Encode: 1.726s, Decode+Unpack: 1.208s +---------------------- -------------------------------------------------------- +💾 Converting with 427.7558 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-467.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 57, 128) +Output shape: (1, 57, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) -> torch.Size([1, 1, 57, 512]) + layer.4.output: torch.Size([1, 57, 3584]) -> torch.Size([1, 1, 57, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,408B, BPFP=0.9342 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,024B, BPFP=3.2961 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,676B, BPFP=1.5559 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,724B, BPFP=3.2138 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,464B, BPFP=1.7719 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,468B, BPFP=3.1436 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,920B, BPFP=1.6228 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,636B, BPFP=3.1897 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,548B, BPFP=2.6173 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,364B, BPFP=3.1151 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,732B, BPFP=1.0077 +⌛️ [2/4] FRONTEND: Frontend time: 1.702s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.101s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 57, 128]) + layer.0.v_cache: torch.Size([1, 4, 57, 128]) + layer.1.k_cache: torch.Size([1, 4, 57, 128]) + layer.1.v_cache: torch.Size([1, 4, 57, 128]) + layer.2.k_cache: torch.Size([1, 4, 57, 128]) + layer.2.v_cache: torch.Size([1, 4, 57, 128]) + layer.3.k_cache: torch.Size([1, 4, 57, 128]) + layer.3.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.k_cache: torch.Size([1, 4, 57, 128]) + layer.4.v_cache: torch.Size([1, 4, 57, 128]) + layer.4.output: torch.Size([1, 57, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10653571 26.26488187 + layer.0.v_cache 0.00001396 0.00227751 + layer.1.k_cache 0.01513677 2.62463941 + layer.1.v_cache 0.00000574 0.00094156 + layer.2.k_cache 0.00241650 0.60274686 + layer.2.v_cache 0.00001997 0.00312339 + layer.3.k_cache 0.04250543 2.95346551 + layer.3.v_cache 0.00001889 0.00314313 + layer.4.k_cache 0.00060099 0.09331395 + layer.4.v_cache 0.00005317 0.00739229 + layer.4.output 0.25650722 843.46436404 + ------------------------------------------------------------------------------------- + TOTAL 0.11546222 349.22391022 + (elements=496,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 496128 +Total Bytes 114964 +BPFP 1.8538 bits/point +EBPFP 3.7076 equivalent bits/point +MSE 349.223910 +---------------------- -------------------------------------------------------- +Time: 2.807s Load: 0.003s, Pack+Encode: 1.702s, Decode+Unpack: 1.101s +---------------------- -------------------------------------------------------- +💾 Converting with 349.2239 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-470.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-470.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 48, 128) +Output shape: (1, 48, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) -> torch.Size([1, 1, 48, 512]) + layer.4.output: torch.Size([1, 48, 3584]) -> torch.Size([1, 1, 48, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,228B, BPFP=1.0508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,808B, BPFP=3.8438 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,312B, BPFP=1.7292 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,484B, BPFP=3.7383 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,072B, BPFP=1.9766 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,424B, BPFP=3.7188 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,812B, BPFP=1.8919 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,488B, BPFP=3.7396 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,204B, BPFP=2.9961 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,168B, BPFP=3.6354 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 23,240B, BPFP=1.0807 +⌛️ [2/4] FRONTEND: Frontend time: 1.771s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.068s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 48, 128]) + layer.0.v_cache: torch.Size([1, 4, 48, 128]) + layer.1.k_cache: torch.Size([1, 4, 48, 128]) + layer.1.v_cache: torch.Size([1, 4, 48, 128]) + layer.2.k_cache: torch.Size([1, 4, 48, 128]) + layer.2.v_cache: torch.Size([1, 4, 48, 128]) + layer.3.k_cache: torch.Size([1, 4, 48, 128]) + layer.3.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.k_cache: torch.Size([1, 4, 48, 128]) + layer.4.v_cache: torch.Size([1, 4, 48, 128]) + layer.4.output: torch.Size([1, 48, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10088900 21.23037084 + layer.0.v_cache 0.00001321 0.00239971 + layer.1.k_cache 0.01601078 2.86588001 + layer.1.v_cache 0.00000529 0.00098739 + layer.2.k_cache 0.00240732 0.59132441 + layer.2.v_cache 0.00001858 0.00324951 + layer.3.k_cache 0.04236705 3.35614141 + layer.3.v_cache 0.00001968 0.00344399 + layer.4.k_cache 0.00061965 0.09185655 + layer.4.v_cache 0.00005235 0.00757171 + layer.4.output 0.30498679 1011.77111235 + ------------------------------------------------------------------------------------- + TOTAL 0.13513591 418.26770659 + (elements=417,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 417792 +Total Bytes 110240 +BPFP 2.1109 bits/point +EBPFP 4.2218 equivalent bits/point +MSE 418.267707 +---------------------- -------------------------------------------------------- +Time: 2.841s Load: 0.003s, Pack+Encode: 1.771s, Decode+Unpack: 1.068s +---------------------- -------------------------------------------------------- +💾 Converting with 418.2677 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-471.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-471.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 53, 128) +Output shape: (1, 53, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) -> torch.Size([1, 1, 53, 512]) + layer.4.output: torch.Size([1, 53, 3584]) -> torch.Size([1, 1, 53, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,384B, BPFP=0.9976 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,956B, BPFP=3.5248 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,628B, BPFP=1.6592 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,672B, BPFP=3.4410 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,048B, BPFP=2.0778 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,504B, BPFP=3.3915 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,028B, BPFP=1.7771 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,536B, BPFP=3.4009 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,524B, BPFP=2.8078 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,256B, BPFP=3.3184 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 24,992B, BPFP=1.0526 +⌛️ [2/4] FRONTEND: Frontend time: 1.835s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.154s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 53, 128]) + layer.0.v_cache: torch.Size([1, 4, 53, 128]) + layer.1.k_cache: torch.Size([1, 4, 53, 128]) + layer.1.v_cache: torch.Size([1, 4, 53, 128]) + layer.2.k_cache: torch.Size([1, 4, 53, 128]) + layer.2.v_cache: torch.Size([1, 4, 53, 128]) + layer.3.k_cache: torch.Size([1, 4, 53, 128]) + layer.3.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.k_cache: torch.Size([1, 4, 53, 128]) + layer.4.v_cache: torch.Size([1, 4, 53, 128]) + layer.4.output: torch.Size([1, 53, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10150213 24.40641122 + layer.0.v_cache 0.00001370 0.00259619 + layer.1.k_cache 0.01753690 2.71467360 + layer.1.v_cache 0.00000536 0.00108929 + layer.2.k_cache 0.00234856 0.63966003 + layer.2.v_cache 0.00001946 0.00334779 + layer.3.k_cache 0.03772523 3.06865476 + layer.3.v_cache 0.00001792 0.00342382 + layer.4.k_cache 0.00061365 0.09053827 + layer.4.v_cache 0.00006478 0.00800707 + layer.4.output 0.25806696 899.76229784 + ------------------------------------------------------------------------------------- + TOTAL 0.11566567 372.31026394 + (elements=461,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 461312 +Total Bytes 114528 +BPFP 1.9861 bits/point +EBPFP 3.9723 equivalent bits/point +MSE 372.310264 +---------------------- -------------------------------------------------------- +Time: 2.991s Load: 0.003s, Pack+Encode: 1.835s, Decode+Unpack: 1.154s +---------------------- -------------------------------------------------------- +💾 Converting with 372.3103 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-496.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-496.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,472B, BPFP=0.9042 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,008B, BPFP=3.1271 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,784B, BPFP=1.5063 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,664B, BPFP=3.0375 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,228B, BPFP=1.8823 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,416B, BPFP=2.9729 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,036B, BPFP=1.8323 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,744B, BPFP=3.0583 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,776B, BPFP=2.5458 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,404B, BPFP=2.9698 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,452B, BPFP=0.9841 +⌛️ [2/4] FRONTEND: Frontend time: 1.745s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.126s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11850098 27.55071004 + layer.0.v_cache 0.00001385 0.00233248 + layer.1.k_cache 0.01468890 2.19979197 + layer.1.v_cache 0.00000523 0.00093785 + layer.2.k_cache 0.01477869 0.53369306 + layer.2.v_cache 0.00001915 0.00313132 + layer.3.k_cache 0.14661819 4.31130778 + layer.3.v_cache 0.00001925 0.00319237 + layer.4.k_cache 0.00060183 0.08424826 + layer.4.v_cache 0.00005752 0.00714959 + layer.4.output 0.22966646 772.74538690 + ------------------------------------------------------------------------------------- + TOTAL 0.11193934 320.23024724 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 117984 +BPFP 1.8074 bits/point +EBPFP 3.6147 equivalent bits/point +MSE 320.230247 +---------------------- -------------------------------------------------------- +Time: 2.875s Load: 0.004s, Pack+Encode: 1.745s, Decode+Unpack: 1.126s +---------------------- -------------------------------------------------------- +💾 Converting with 320.2302 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-498.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-498.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,992B, BPFP=0.8478 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,732B, BPFP=3.5211 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,120B, BPFP=1.3791 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,716B, BPFP=3.3485 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,452B, BPFP=1.6053 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,108B, BPFP=3.2452 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,724B, BPFP=1.4817 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,264B, BPFP=3.4416 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,880B, BPFP=2.5272 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,688B, BPFP=3.3438 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,784B, BPFP=0.7954 +⌛️ [2/4] FRONTEND: Frontend time: 1.735s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.190s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10836371 25.87611456 + layer.0.v_cache 0.00001445 0.00221770 + layer.1.k_cache 0.06846340 2.04489202 + layer.1.v_cache 0.00000523 0.00083190 + layer.2.k_cache 0.00506382 0.60273573 + layer.2.v_cache 0.00001740 0.00253300 + layer.3.k_cache 0.03973729 3.18676293 + layer.3.v_cache 0.00001777 0.00288544 + layer.4.k_cache 0.00066797 0.07807094 + layer.4.v_cache 0.00005099 0.00652553 + layer.4.output 0.14786712 442.04425466 + ------------------------------------------------------------------------------------- + TOTAL 0.07396893 183.88902073 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 178460 +BPFP 1.7829 bits/point +EBPFP 3.5658 equivalent bits/point +MSE 183.889021 +---------------------- -------------------------------------------------------- +Time: 2.930s Load: 0.004s, Pack+Encode: 1.735s, Decode+Unpack: 1.190s +---------------------- -------------------------------------------------------- +💾 Converting with 183.8890 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-504.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-504.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,796B, BPFP=0.8987 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,676B, BPFP=3.7112 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,172B, BPFP=1.4612 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,084B, BPFP=3.5710 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,060B, BPFP=1.6714 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,180B, BPFP=3.1203 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,620B, BPFP=1.5672 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,188B, BPFP=3.3589 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,664B, BPFP=2.5246 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,496B, BPFP=3.1951 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,168B, BPFP=0.9527 +⌛️ [2/4] FRONTEND: Frontend time: 1.892s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.212s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08728886 26.01300974 + layer.0.v_cache 0.00001734 0.00232245 + layer.1.k_cache 0.01499850 2.07628169 + layer.1.v_cache 0.00000528 0.00082511 + layer.2.k_cache 0.00240521 0.67778674 + layer.2.v_cache 0.00001739 0.00260502 + layer.3.k_cache 0.03126823 2.79219818 + layer.3.v_cache 0.00001843 0.00300204 + layer.4.k_cache 0.00061395 0.07660924 + layer.4.v_cache 0.00004830 0.00620449 + layer.4.output 0.21953388 566.71732955 + ------------------------------------------------------------------------------------- + TOTAL 0.09843639 235.21600891 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 134104 +BPFP 1.8675 bits/point +EBPFP 3.7351 equivalent bits/point +MSE 235.216009 +---------------------- -------------------------------------------------------- +Time: 3.107s Load: 0.003s, Pack+Encode: 1.892s, Decode+Unpack: 1.212s +---------------------- -------------------------------------------------------- +💾 Converting with 235.2160 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-518.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-518.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 66, 128) +Output shape: (1, 66, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) -> torch.Size([1, 1, 66, 512]) + layer.4.output: torch.Size([1, 66, 3584]) -> torch.Size([1, 1, 66, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,792B, BPFP=0.8977 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,848B, BPFP=3.7519 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,256B, BPFP=1.4811 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,460B, BPFP=3.4233 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,140B, BPFP=1.6903 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,488B, BPFP=3.1932 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,684B, BPFP=1.5824 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,088B, BPFP=3.3352 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 10,800B, BPFP=2.5568 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,112B, BPFP=3.3409 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,076B, BPFP=0.9834 +⌛️ [2/4] FRONTEND: Frontend time: 1.845s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.351s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 66, 128]) + layer.0.v_cache: torch.Size([1, 4, 66, 128]) + layer.1.k_cache: torch.Size([1, 4, 66, 128]) + layer.1.v_cache: torch.Size([1, 4, 66, 128]) + layer.2.k_cache: torch.Size([1, 4, 66, 128]) + layer.2.v_cache: torch.Size([1, 4, 66, 128]) + layer.3.k_cache: torch.Size([1, 4, 66, 128]) + layer.3.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.k_cache: torch.Size([1, 4, 66, 128]) + layer.4.v_cache: torch.Size([1, 4, 66, 128]) + layer.4.output: torch.Size([1, 66, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08223097 25.44959236 + layer.0.v_cache 0.00001442 0.00234227 + layer.1.k_cache 0.01480860 2.40382616 + layer.1.v_cache 0.00000502 0.00080974 + layer.2.k_cache 0.00576455 0.72824201 + layer.2.v_cache 0.00002304 0.00258850 + layer.3.k_cache 0.03215047 2.84192634 + layer.3.v_cache 0.00001933 0.00311660 + layer.4.k_cache 0.00061510 0.07684822 + layer.4.v_cache 0.00004916 0.00633631 + layer.4.output 0.22077460 571.76440747 + ------------------------------------------------------------------------------------- + TOTAL 0.09888840 237.28626358 + (elements=574,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 574464 +Total Bytes 135744 +BPFP 1.8904 bits/point +EBPFP 3.7807 equivalent bits/point +MSE 237.286264 +---------------------- -------------------------------------------------------- +Time: 3.199s Load: 0.003s, Pack+Encode: 1.845s, Decode+Unpack: 1.351s +---------------------- -------------------------------------------------------- +💾 Converting with 237.2863 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-522.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-522.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,324B, BPFP=0.9255 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,764B, BPFP=3.5882 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,064B, BPFP=1.5120 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,380B, BPFP=3.5060 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,880B, BPFP=1.6866 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,104B, BPFP=3.2329 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,244B, BPFP=1.5505 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,260B, BPFP=3.4803 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,716B, BPFP=2.5077 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,032B, BPFP=3.2175 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 28,336B, BPFP=0.8664 +⌛️ [2/4] FRONTEND: Frontend time: 1.723s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10114760 26.30452195 + layer.0.v_cache 0.00001420 0.00211442 + layer.1.k_cache 0.03683303 1.91777101 + layer.1.v_cache 0.00000544 0.00085774 + layer.2.k_cache 0.00248627 0.73661977 + layer.2.v_cache 0.00001713 0.00255469 + layer.3.k_cache 0.02909129 2.62508027 + layer.3.v_cache 0.00001766 0.00287450 + layer.4.k_cache 0.00072613 0.07471534 + layer.4.v_cache 0.00005154 0.00650419 + layer.4.output 0.19731793 666.37396037 + ------------------------------------------------------------------------------------- + TOTAL 0.09127152 276.25243156 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 146104 +BPFP 1.8395 bits/point +EBPFP 3.6791 equivalent bits/point +MSE 276.252432 +---------------------- -------------------------------------------------------- +Time: 3.082s Load: 0.003s, Pack+Encode: 1.723s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 276.2524 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-523.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-523.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,536B, BPFP=0.9326 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,956B, BPFP=3.6916 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,308B, BPFP=1.5025 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,828B, BPFP=3.4597 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,216B, BPFP=1.6891 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,692B, BPFP=3.2262 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,688B, BPFP=1.5806 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,900B, BPFP=3.4745 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,404B, BPFP=2.5502 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,008B, BPFP=3.2911 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 30,316B, BPFP=0.8904 +⌛️ [2/4] FRONTEND: Frontend time: 2.012s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.173s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10230959 23.82004587 + layer.0.v_cache 0.00001409 0.00217100 + layer.1.k_cache 0.07973289 1.73580772 + layer.1.v_cache 0.00000506 0.00078329 + layer.2.k_cache 0.00553133 0.66754261 + layer.2.v_cache 0.00001928 0.00252534 + layer.3.k_cache 0.03946657 2.51844266 + layer.3.v_cache 0.00001980 0.00283808 + layer.4.k_cache 0.00066525 0.07381410 + layer.4.v_cache 0.00005119 0.00599675 + layer.4.output 0.17985767 628.07242716 + ------------------------------------------------------------------------------------- + TOTAL 0.08745993 260.31393868 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 153852 +BPFP 1.8606 bits/point +EBPFP 3.7213 equivalent bits/point +MSE 260.313939 +---------------------- -------------------------------------------------------- +Time: 3.191s Load: 0.006s, Pack+Encode: 2.012s, Decode+Unpack: 1.173s +---------------------- -------------------------------------------------------- +💾 Converting with 260.3139 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-531.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-531.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,568B, BPFP=0.9269 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,464B, BPFP=3.5438 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,224B, BPFP=1.4659 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,532B, BPFP=3.1518 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,972B, BPFP=1.6177 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,932B, BPFP=3.0300 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,488B, BPFP=1.5195 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,500B, BPFP=3.3482 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,036B, BPFP=2.4424 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,972B, BPFP=3.0381 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,756B, BPFP=0.8046 +⌛️ [2/4] FRONTEND: Frontend time: 2.058s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.177s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633976 26.99727958 + layer.0.v_cache 0.00001461 0.00218173 + layer.1.k_cache 0.07748085 2.12633683 + layer.1.v_cache 0.00000515 0.00078295 + layer.2.k_cache 0.00505460 0.67338919 + layer.2.v_cache 0.00001651 0.00236354 + layer.3.k_cache 0.02793909 2.51413083 + layer.3.v_cache 0.00001809 0.00278506 + layer.4.k_cache 0.00070819 0.07293368 + layer.4.v_cache 0.00004716 0.00594371 + layer.4.output 0.18765952 627.12331865 + ------------------------------------------------------------------------------------- + TOTAL 0.09007298 260.13302104 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 146444 +BPFP 1.7480 bits/point +EBPFP 3.4961 equivalent bits/point +MSE 260.133021 +---------------------- -------------------------------------------------------- +Time: 3.240s Load: 0.004s, Pack+Encode: 2.058s, Decode+Unpack: 1.177s +---------------------- -------------------------------------------------------- +💾 Converting with 260.1330 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-533.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,232B, BPFP=0.9446 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,408B, BPFP=3.6625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,820B, BPFP=1.5223 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,448B, BPFP=3.4482 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,544B, BPFP=1.6839 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,276B, BPFP=3.1866 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,016B, BPFP=1.5661 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,872B, BPFP=3.3196 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,312B, BPFP=2.5250 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,564B, BPFP=3.2509 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,684B, BPFP=0.8509 +⌛️ [2/4] FRONTEND: Frontend time: 1.883s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.325s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08399124 25.99707380 + layer.0.v_cache 0.00001671 0.00218095 + layer.1.k_cache 0.03849935 1.94845843 + layer.1.v_cache 0.00000503 0.00081227 + layer.2.k_cache 0.00738276 0.70402205 + layer.2.v_cache 0.00001692 0.00246836 + layer.3.k_cache 0.04914829 3.22899453 + layer.3.v_cache 0.00001713 0.00273573 + layer.4.k_cache 0.00063731 0.07459277 + layer.4.v_cache 0.00004711 0.00602911 + layer.4.output 0.21328995 703.76320153 + ------------------------------------------------------------------------------------- + TOTAL 0.09839950 291.66528110 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 139176 +BPFP 1.8274 bits/point +EBPFP 3.6548 equivalent bits/point +MSE 291.665281 +---------------------- -------------------------------------------------------- +Time: 3.212s Load: 0.003s, Pack+Encode: 1.883s, Decode+Unpack: 1.325s +---------------------- -------------------------------------------------------- +💾 Converting with 291.6653 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-537.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-537.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 56, 128) +Output shape: (1, 56, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) -> torch.Size([1, 1, 56, 512]) + layer.4.output: torch.Size([1, 56, 3584]) -> torch.Size([1, 1, 56, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,460B, BPFP=0.9654 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 11,956B, BPFP=3.3359 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,700B, BPFP=1.5904 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,480B, BPFP=3.2031 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 6,452B, BPFP=1.8002 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,276B, BPFP=3.1462 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 5,920B, BPFP=1.6518 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,620B, BPFP=3.2422 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,448B, BPFP=2.6362 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,148B, BPFP=3.1105 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,236B, BPFP=1.0059 +⌛️ [2/4] FRONTEND: Frontend time: 2.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 56, 128]) + layer.0.v_cache: torch.Size([1, 4, 56, 128]) + layer.1.k_cache: torch.Size([1, 4, 56, 128]) + layer.1.v_cache: torch.Size([1, 4, 56, 128]) + layer.2.k_cache: torch.Size([1, 4, 56, 128]) + layer.2.v_cache: torch.Size([1, 4, 56, 128]) + layer.3.k_cache: torch.Size([1, 4, 56, 128]) + layer.3.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.k_cache: torch.Size([1, 4, 56, 128]) + layer.4.v_cache: torch.Size([1, 4, 56, 128]) + layer.4.output: torch.Size([1, 56, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13768739 24.49425616 + layer.0.v_cache 0.00001432 0.00242967 + layer.1.k_cache 0.01638686 2.72513417 + layer.1.v_cache 0.00000533 0.00094365 + layer.2.k_cache 0.00251956 0.72817169 + layer.2.v_cache 0.00001745 0.00287304 + layer.3.k_cache 0.08599430 2.67086247 + layer.3.v_cache 0.00002012 0.00334062 + layer.4.k_cache 0.00061995 0.09223371 + layer.4.v_cache 0.00004798 0.00665605 + layer.4.output 0.24256272 276.75159439 + ------------------------------------------------------------------------------------- + TOTAL 0.11419131 115.76400364 + (elements=487,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 487424 +Total Bytes 113696 +BPFP 1.8661 bits/point +EBPFP 3.7321 equivalent bits/point +MSE 115.764004 +---------------------- -------------------------------------------------------- +Time: 3.681s Load: 0.004s, Pack+Encode: 2.267s, Decode+Unpack: 1.411s +---------------------- -------------------------------------------------------- +💾 Converting with 115.7640 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-54.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-54.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,692B, BPFP=0.9164 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,736B, BPFP=3.6594 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,432B, BPFP=1.4516 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,852B, BPFP=3.4867 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,348B, BPFP=1.6305 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,552B, BPFP=3.2328 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,812B, BPFP=1.5258 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 17,200B, BPFP=3.3594 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,040B, BPFP=2.5469 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,628B, BPFP=3.2477 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,184B, BPFP=0.8701 +⌛️ [2/4] FRONTEND: Frontend time: 1.913s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.396s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09278157 26.04541321 + layer.0.v_cache 0.00001504 0.00218199 + layer.1.k_cache 0.07523044 1.87675247 + layer.1.v_cache 0.00000517 0.00086702 + layer.2.k_cache 0.00381728 0.65393395 + layer.2.v_cache 0.00001789 0.00244421 + layer.3.k_cache 0.03327554 2.43125896 + layer.3.v_cache 0.00001939 0.00283857 + layer.4.k_cache 0.00067710 0.07654592 + layer.4.v_cache 0.00005192 0.00614091 + layer.4.output 0.16994183 490.04603795 + ------------------------------------------------------------------------------------- + TOTAL 0.08208730 203.61297899 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 159476 +BPFP 1.8322 bits/point +EBPFP 3.6644 equivalent bits/point +MSE 203.612979 +---------------------- -------------------------------------------------------- +Time: 3.315s Load: 0.005s, Pack+Encode: 1.913s, Decode+Unpack: 1.396s +---------------------- -------------------------------------------------------- +💾 Converting with 203.6130 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-540.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-540.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 59, 128) +Output shape: (1, 59, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) -> torch.Size([1, 1, 59, 512]) + layer.4.output: torch.Size([1, 59, 3584]) -> torch.Size([1, 1, 59, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,468B, BPFP=0.9184 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,192B, BPFP=3.2288 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,108B, BPFP=1.6176 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,712B, BPFP=3.1017 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,312B, BPFP=1.9364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,496B, BPFP=3.0445 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,568B, BPFP=1.7394 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,736B, BPFP=3.1081 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,772B, BPFP=2.5879 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,372B, BPFP=3.0117 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,968B, BPFP=0.9824 +⌛️ [2/4] FRONTEND: Frontend time: 1.675s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.141s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 59, 128]) + layer.0.v_cache: torch.Size([1, 4, 59, 128]) + layer.1.k_cache: torch.Size([1, 4, 59, 128]) + layer.1.v_cache: torch.Size([1, 4, 59, 128]) + layer.2.k_cache: torch.Size([1, 4, 59, 128]) + layer.2.v_cache: torch.Size([1, 4, 59, 128]) + layer.3.k_cache: torch.Size([1, 4, 59, 128]) + layer.3.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.k_cache: torch.Size([1, 4, 59, 128]) + layer.4.v_cache: torch.Size([1, 4, 59, 128]) + layer.4.output: torch.Size([1, 59, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12468526 27.99177784 + layer.0.v_cache 0.00001451 0.00275893 + layer.1.k_cache 0.01313243 2.49544047 + layer.1.v_cache 0.00000580 0.00104662 + layer.2.k_cache 0.00906488 0.56630552 + layer.2.v_cache 0.00002099 0.00307492 + layer.3.k_cache 0.01569995 4.23225067 + layer.3.v_cache 0.00002002 0.00352885 + layer.4.k_cache 0.00064964 0.09351971 + layer.4.v_cache 0.00005478 0.00717371 + layer.4.output 0.23031524 203.67562046 + ------------------------------------------------------------------------------------- + TOTAL 0.10444441 85.94860120 + (elements=513,536) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 513536 +Total Bytes 117704 +BPFP 1.8336 bits/point +EBPFP 3.6672 equivalent bits/point +MSE 85.948601 +---------------------- -------------------------------------------------------- +Time: 2.820s Load: 0.005s, Pack+Encode: 1.675s, Decode+Unpack: 1.141s +---------------------- -------------------------------------------------------- +💾 Converting with 85.9486 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-57.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-57.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 60, 128) +Output shape: (1, 60, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) -> torch.Size([1, 1, 60, 512]) + layer.4.output: torch.Size([1, 60, 3584]) -> torch.Size([1, 1, 60, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 3,468B, BPFP=0.9031 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 12,180B, BPFP=3.1719 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 5,752B, BPFP=1.4979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 11,764B, BPFP=3.0635 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,188B, BPFP=1.8719 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 11,456B, BPFP=2.9833 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 6,464B, BPFP=1.6833 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 11,740B, BPFP=3.0573 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 9,740B, BPFP=2.5365 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 11,460B, BPFP=2.9844 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 27,404B, BPFP=1.0195 +⌛️ [2/4] FRONTEND: Frontend time: 1.787s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.104s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 60, 128]) + layer.0.v_cache: torch.Size([1, 4, 60, 128]) + layer.1.k_cache: torch.Size([1, 4, 60, 128]) + layer.1.v_cache: torch.Size([1, 4, 60, 128]) + layer.2.k_cache: torch.Size([1, 4, 60, 128]) + layer.2.v_cache: torch.Size([1, 4, 60, 128]) + layer.3.k_cache: torch.Size([1, 4, 60, 128]) + layer.3.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.k_cache: torch.Size([1, 4, 60, 128]) + layer.4.v_cache: torch.Size([1, 4, 60, 128]) + layer.4.output: torch.Size([1, 60, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15164433 26.35222371 + layer.0.v_cache 0.00001435 0.00252726 + layer.1.k_cache 0.01704682 1.85202599 + layer.1.v_cache 0.00000573 0.00100550 + layer.2.k_cache 0.00240009 0.57673213 + layer.2.v_cache 0.00001845 0.00300955 + layer.3.k_cache 0.02161660 4.53598684 + layer.3.v_cache 0.00001985 0.00343343 + layer.4.k_cache 0.00061459 0.08785538 + layer.4.v_cache 0.00005275 0.00761105 + layer.4.output 0.22653077 676.16711310 + ------------------------------------------------------------------------------------- + TOTAL 0.10465582 280.38777662 + (elements=522,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 522240 +Total Bytes 118616 +BPFP 1.8170 bits/point +EBPFP 3.6341 equivalent bits/point +MSE 280.387777 +---------------------- -------------------------------------------------------- +Time: 2.895s Load: 0.003s, Pack+Encode: 1.787s, Decode+Unpack: 1.104s +---------------------- -------------------------------------------------------- +💾 Converting with 280.3878 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-59.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-59.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,336B, BPFP=0.9281 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,388B, BPFP=3.5077 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,912B, BPFP=1.4795 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,244B, BPFP=3.2628 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,624B, BPFP=1.6318 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,772B, BPFP=2.9478 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,096B, BPFP=1.5188 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,824B, BPFP=3.1729 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,744B, BPFP=2.5137 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,408B, BPFP=3.0839 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 25,952B, BPFP=0.7935 +⌛️ [2/4] FRONTEND: Frontend time: 1.890s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.355s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08869729 26.43195165 + layer.0.v_cache 0.00001594 0.00210823 + layer.1.k_cache 0.08105310 1.72621991 + layer.1.v_cache 0.00000494 0.00075335 + layer.2.k_cache 0.00256802 0.69243622 + layer.2.v_cache 0.00001661 0.00236975 + layer.3.k_cache 0.01835501 2.45051387 + layer.3.v_cache 0.00001861 0.00266022 + layer.4.k_cache 0.00078324 0.07448046 + layer.4.v_cache 0.00004769 0.00562698 + layer.4.output 0.18611241 628.12414384 + ------------------------------------------------------------------------------------- + TOTAL 0.08790278 260.48577220 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 138300 +BPFP 1.7413 bits/point +EBPFP 3.4826 equivalent bits/point +MSE 260.485772 +---------------------- -------------------------------------------------------- +Time: 3.248s Load: 0.004s, Pack+Encode: 1.890s, Decode+Unpack: 1.355s +---------------------- -------------------------------------------------------- +💾 Converting with 260.4858 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-63.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,952B, BPFP=0.7977 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,584B, BPFP=3.4768 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,248B, BPFP=1.3286 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,244B, BPFP=3.2610 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,528B, BPFP=1.5348 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,536B, BPFP=3.1469 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,844B, BPFP=1.4246 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,436B, BPFP=3.2919 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,760B, BPFP=2.5387 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,628B, BPFP=3.1617 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,468B, BPFP=0.8392 +⌛️ [2/4] FRONTEND: Frontend time: 1.812s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.317s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17228427 26.38842270 + layer.0.v_cache 0.00001543 0.00208166 + layer.1.k_cache 0.04621154 2.01043921 + layer.1.v_cache 0.00000545 0.00080722 + layer.2.k_cache 0.00739691 0.66647803 + layer.2.v_cache 0.00001797 0.00249176 + layer.3.k_cache 0.03096496 3.01073778 + layer.3.v_cache 0.00001890 0.00273005 + layer.4.k_cache 0.00071091 0.08952166 + layer.4.v_cache 0.00004900 0.00595607 + layer.4.output 0.02450962 508.62762334 + ------------------------------------------------------------------------------------- + TOTAL 0.02524957 211.32782527 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 185228 +BPFP 1.7551 bits/point +EBPFP 3.5102 equivalent bits/point +MSE 211.327825 +---------------------- -------------------------------------------------------- +Time: 3.132s Load: 0.004s, Pack+Encode: 1.812s, Decode+Unpack: 1.317s +---------------------- -------------------------------------------------------- +💾 Converting with 211.3278 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-65.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-65.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,004B, BPFP=0.8407 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 21,024B, BPFP=3.5323 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,304B, BPFP=1.3952 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,216B, BPFP=3.3965 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,540B, BPFP=1.6028 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,600B, BPFP=3.2930 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,928B, BPFP=1.5000 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 20,320B, BPFP=3.4140 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,440B, BPFP=2.5941 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,612B, BPFP=3.2950 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,696B, BPFP=0.8088 +⌛️ [2/4] FRONTEND: Frontend time: 1.989s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.271s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16174046 27.12463248 + layer.0.v_cache 0.00001612 0.00214021 + layer.1.k_cache 0.03352992 1.97014331 + layer.1.v_cache 0.00000563 0.00083099 + layer.2.k_cache 0.00496730 0.68806892 + layer.2.v_cache 0.00001740 0.00257411 + layer.3.k_cache 0.01521110 3.47874861 + layer.3.v_cache 0.00001939 0.00291955 + layer.4.k_cache 0.00066431 0.08063157 + layer.4.v_cache 0.00005067 0.00632351 + layer.4.output 0.14624557 507.96807796 + ------------------------------------------------------------------------------------- + TOTAL 0.07293772 211.12550347 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 181684 +BPFP 1.7956 bits/point +EBPFP 3.5912 equivalent bits/point +MSE 211.125503 +---------------------- -------------------------------------------------------- +Time: 3.265s Load: 0.005s, Pack+Encode: 1.989s, Decode+Unpack: 1.271s +---------------------- -------------------------------------------------------- +💾 Converting with 211.1255 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-67.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-67.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,776B, BPFP=0.8677 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,156B, BPFP=3.6621 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,692B, BPFP=1.3975 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,956B, BPFP=3.4440 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,712B, BPFP=1.5828 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,620B, BPFP=3.2013 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,104B, BPFP=1.4724 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,676B, BPFP=3.3932 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,320B, BPFP=2.4201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,696B, BPFP=3.2151 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,548B, BPFP=0.8188 +⌛️ [2/4] FRONTEND: Frontend time: 1.866s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07121617 28.24392203 + layer.0.v_cache 0.00001681 0.00216538 + layer.1.k_cache 0.06877367 1.85109551 + layer.1.v_cache 0.00000530 0.00079326 + layer.2.k_cache 0.00486504 0.73154352 + layer.2.v_cache 0.00001749 0.00242576 + layer.3.k_cache 0.08833314 2.84289498 + layer.3.v_cache 0.00001723 0.00268072 + layer.4.k_cache 0.00077278 0.07430965 + layer.4.v_cache 0.00004595 0.00581237 + layer.4.output 0.15812202 546.85439161 + ------------------------------------------------------------------------------------- + TOTAL 0.07887751 227.16108144 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 167256 +BPFP 1.7875 bits/point +EBPFP 3.5751 equivalent bits/point +MSE 227.161081 +---------------------- -------------------------------------------------------- +Time: 3.189s Load: 0.004s, Pack+Encode: 1.866s, Decode+Unpack: 1.319s +---------------------- -------------------------------------------------------- +💾 Converting with 227.1611 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-7.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-7.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,364B, BPFP=0.8298 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,400B, BPFP=3.4653 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,872B, BPFP=1.5272 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,416B, BPFP=3.3131 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,832B, BPFP=1.6757 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,388B, BPFP=3.1541 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,760B, BPFP=1.5099 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,292B, BPFP=3.2939 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,292B, BPFP=2.6751 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 20,572B, BPFP=3.1825 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,144B, BPFP=0.7546 +⌛️ [2/4] FRONTEND: Frontend time: 1.817s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13964315 29.52594659 + layer.0.v_cache 0.00001432 0.00214358 + layer.1.k_cache 0.04810369 2.48569360 + layer.1.v_cache 0.00000572 0.00088639 + layer.2.k_cache 0.00492308 0.64425055 + layer.2.v_cache 0.00001777 0.00254059 + layer.3.k_cache 0.04255116 3.59188208 + layer.3.v_cache 0.00001880 0.00300349 + layer.4.k_cache 0.00073971 0.08897589 + layer.4.v_cache 0.00004573 0.00592048 + layer.4.output 11.28739116 471.63251414 + ------------------------------------------------------------------------------------- + TOTAL 4.66163537 196.33993190 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 193332 +BPFP 1.7594 bits/point +EBPFP 3.5187 equivalent bits/point +MSE 196.339932 +---------------------- -------------------------------------------------------- +Time: 3.136s Load: 0.004s, Pack+Encode: 1.817s, Decode+Unpack: 1.314s +---------------------- -------------------------------------------------------- +💾 Converting with 196.3399 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-72.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-72.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,864B, BPFP=0.8636 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,036B, BPFP=3.5575 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,012B, BPFP=1.4226 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,936B, BPFP=3.3622 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 8,916B, BPFP=1.5831 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 17,848B, BPFP=3.1690 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,400B, BPFP=1.4915 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,588B, BPFP=3.3004 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,888B, BPFP=2.4659 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,816B, BPFP=3.1634 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,564B, BPFP=0.8006 +⌛️ [2/4] FRONTEND: Frontend time: 1.856s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.406s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14145413 25.78230424 + layer.0.v_cache 0.00001516 0.00219781 + layer.1.k_cache 0.10647225 1.99646395 + layer.1.v_cache 0.00000560 0.00084038 + layer.2.k_cache 0.00244946 0.72963871 + layer.2.v_cache 0.00001685 0.00238709 + layer.3.k_cache 0.01581429 2.74080086 + layer.3.v_cache 0.00001841 0.00284198 + layer.4.k_cache 0.00072023 0.07949266 + layer.4.v_cache 0.00004652 0.00559355 + layer.4.output 0.15451367 394.48135653 + ------------------------------------------------------------------------------------- + TOTAL 0.07932992 164.27717982 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 168868 +BPFP 1.7637 bits/point +EBPFP 3.5275 equivalent bits/point +MSE 164.277180 +---------------------- -------------------------------------------------------- +Time: 3.266s Load: 0.004s, Pack+Encode: 1.856s, Decode+Unpack: 1.406s +---------------------- -------------------------------------------------------- +💾 Converting with 164.2772 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-75.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-75.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 74, 128) +Output shape: (1, 74, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) -> torch.Size([1, 1, 74, 512]) + layer.4.output: torch.Size([1, 74, 3584]) -> torch.Size([1, 1, 74, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,376B, BPFP=0.9240 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,804B, BPFP=3.5481 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 7,128B, BPFP=1.5051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,112B, BPFP=3.4020 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,804B, BPFP=1.6478 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,888B, BPFP=3.1436 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,184B, BPFP=1.5169 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,980B, BPFP=3.3742 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,652B, BPFP=2.4603 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,296B, BPFP=3.2297 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 26,456B, BPFP=0.7980 +⌛️ [2/4] FRONTEND: Frontend time: 1.888s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.446s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 74, 128]) + layer.0.v_cache: torch.Size([1, 4, 74, 128]) + layer.1.k_cache: torch.Size([1, 4, 74, 128]) + layer.1.v_cache: torch.Size([1, 4, 74, 128]) + layer.2.k_cache: torch.Size([1, 4, 74, 128]) + layer.2.v_cache: torch.Size([1, 4, 74, 128]) + layer.3.k_cache: torch.Size([1, 4, 74, 128]) + layer.3.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.k_cache: torch.Size([1, 4, 74, 128]) + layer.4.v_cache: torch.Size([1, 4, 74, 128]) + layer.4.output: torch.Size([1, 74, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10237078 25.46066202 + layer.0.v_cache 0.00001391 0.00204342 + layer.1.k_cache 0.05953315 1.90721708 + layer.1.v_cache 0.00000517 0.00078765 + layer.2.k_cache 0.00728044 0.69444203 + layer.2.v_cache 0.00001648 0.00234058 + layer.3.k_cache 0.03101032 2.57467218 + layer.3.v_cache 0.00001666 0.00262958 + layer.4.k_cache 0.00072555 0.07412593 + layer.4.v_cache 0.00004870 0.00597197 + layer.4.output 0.18364459 654.81370656 + ------------------------------------------------------------------------------------- + TOTAL 0.08744314 271.43651991 + (elements=644,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 644096 +Total Bytes 143680 +BPFP 1.7846 bits/point +EBPFP 3.5692 equivalent bits/point +MSE 271.436520 +---------------------- -------------------------------------------------------- +Time: 3.338s Load: 0.003s, Pack+Encode: 1.888s, Decode+Unpack: 1.446s +---------------------- -------------------------------------------------------- +💾 Converting with 271.4365 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-77.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-77.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 111, 128) +Output shape: (1, 111, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) -> torch.Size([1, 1, 111, 512]) + layer.4.output: torch.Size([1, 111, 3584]) -> torch.Size([1, 1, 111, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,764B, BPFP=0.8114 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 23,264B, BPFP=3.2748 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,176B, BPFP=1.4324 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 22,456B, BPFP=3.1610 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,336B, BPFP=1.7365 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 21,228B, BPFP=2.9882 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,912B, BPFP=1.5360 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,912B, BPFP=3.0845 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,776B, BPFP=2.5023 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,280B, BPFP=2.9955 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,340B, BPFP=0.7710 +⌛️ [2/4] FRONTEND: Frontend time: 2.544s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.119s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 111, 128]) + layer.0.v_cache: torch.Size([1, 4, 111, 128]) + layer.1.k_cache: torch.Size([1, 4, 111, 128]) + layer.1.v_cache: torch.Size([1, 4, 111, 128]) + layer.2.k_cache: torch.Size([1, 4, 111, 128]) + layer.2.v_cache: torch.Size([1, 4, 111, 128]) + layer.3.k_cache: torch.Size([1, 4, 111, 128]) + layer.3.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.k_cache: torch.Size([1, 4, 111, 128]) + layer.4.v_cache: torch.Size([1, 4, 111, 128]) + layer.4.output: torch.Size([1, 111, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11107565 27.25192453 + layer.0.v_cache 0.00001597 0.00210574 + layer.1.k_cache 0.04333284 2.19559266 + layer.1.v_cache 0.00000558 0.00078950 + layer.2.k_cache 0.00477623 0.65195389 + layer.2.v_cache 0.00001758 0.00241790 + layer.3.k_cache 0.01020195 3.62760747 + layer.3.v_cache 0.00001870 0.00266546 + layer.4.k_cache 0.00079488 0.08774087 + layer.4.v_cache 0.00005041 0.00613994 + layer.4.output 10.27057757 371.81258044 + ------------------------------------------------------------------------------------- + TOTAL 4.23907840 155.08923535 + (elements=966,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 966144 +Total Bytes 205444 +BPFP 1.7011 bits/point +EBPFP 3.4023 equivalent bits/point +MSE 155.089235 +---------------------- -------------------------------------------------------- +Time: 3.670s Load: 0.006s, Pack+Encode: 2.544s, Decode+Unpack: 1.119s +---------------------- -------------------------------------------------------- +💾 Converting with 155.0892 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-8.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 71, 128) +Output shape: (1, 71, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) -> torch.Size([1, 1, 71, 512]) + layer.4.output: torch.Size([1, 71, 3584]) -> torch.Size([1, 1, 71, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,280B, BPFP=0.9419 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,888B, BPFP=3.7165 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 6,872B, BPFP=1.5123 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,592B, BPFP=3.2113 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 7,752B, BPFP=1.7060 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,100B, BPFP=3.1030 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 7,144B, BPFP=1.5722 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,784B, BPFP=3.2535 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 11,432B, BPFP=2.5158 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,452B, BPFP=3.1805 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 29,768B, BPFP=0.9359 +⌛️ [2/4] FRONTEND: Frontend time: 1.847s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.289s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 71, 128]) + layer.0.v_cache: torch.Size([1, 4, 71, 128]) + layer.1.k_cache: torch.Size([1, 4, 71, 128]) + layer.1.v_cache: torch.Size([1, 4, 71, 128]) + layer.2.k_cache: torch.Size([1, 4, 71, 128]) + layer.2.v_cache: torch.Size([1, 4, 71, 128]) + layer.3.k_cache: torch.Size([1, 4, 71, 128]) + layer.3.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.k_cache: torch.Size([1, 4, 71, 128]) + layer.4.v_cache: torch.Size([1, 4, 71, 128]) + layer.4.output: torch.Size([1, 71, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09375787 25.20476693 + layer.0.v_cache 0.00001622 0.00217326 + layer.1.k_cache 0.05984993 1.81081444 + layer.1.v_cache 0.00000512 0.00080254 + layer.2.k_cache 0.00243208 0.76157304 + layer.2.v_cache 0.00001865 0.00253194 + layer.3.k_cache 0.04944374 3.35250854 + layer.3.v_cache 0.00001908 0.00285475 + layer.4.k_cache 0.00084877 0.07336038 + layer.4.v_cache 0.00005139 0.00613127 + layer.4.output 0.19371969 690.32381791 + ------------------------------------------------------------------------------------- + TOTAL 0.09191063 286.08730838 + (elements=617,984) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 617984 +Total Bytes 142064 +BPFP 1.8391 bits/point +EBPFP 3.6781 equivalent bits/point +MSE 286.087308 +---------------------- -------------------------------------------------------- +Time: 3.139s Load: 0.003s, Pack+Encode: 1.847s, Decode+Unpack: 1.289s +---------------------- -------------------------------------------------------- +💾 Converting with 286.0873 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-90.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 109, 128) +Output shape: (1, 109, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) -> torch.Size([1, 1, 109, 512]) + layer.4.output: torch.Size([1, 109, 3584]) -> torch.Size([1, 1, 109, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,668B, BPFP=0.8125 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 22,684B, BPFP=3.2517 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,168B, BPFP=1.4576 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 21,768B, BPFP=3.1204 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,980B, BPFP=1.7173 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 20,860B, BPFP=2.9903 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,696B, BPFP=1.5333 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 21,612B, BPFP=3.0981 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,572B, BPFP=2.5189 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 21,008B, BPFP=3.0115 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,764B, BPFP=0.7938 +⌛️ [2/4] FRONTEND: Frontend time: 1.705s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) +⌛️ [3/4] BACKEND: Backend time: 1.290s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 109, 128]) + layer.0.v_cache: torch.Size([1, 4, 109, 128]) + layer.1.k_cache: torch.Size([1, 4, 109, 128]) + layer.1.v_cache: torch.Size([1, 4, 109, 128]) + layer.2.k_cache: torch.Size([1, 4, 109, 128]) + layer.2.v_cache: torch.Size([1, 4, 109, 128]) + layer.3.k_cache: torch.Size([1, 4, 109, 128]) + layer.3.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.k_cache: torch.Size([1, 4, 109, 128]) + layer.4.v_cache: torch.Size([1, 4, 109, 128]) + layer.4.output: torch.Size([1, 109, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860953 25.44149136 + layer.0.v_cache 0.00001500 0.00203955 + layer.1.k_cache 0.05917679 2.18870334 + layer.1.v_cache 0.00000528 0.00079511 + layer.2.k_cache 0.01253135 0.61220334 + layer.2.v_cache 0.00001869 0.00248704 + layer.3.k_cache 0.07406004 3.64416224 + layer.3.v_cache 0.00001854 0.00275621 + layer.4.k_cache 0.00071234 0.08541672 + layer.4.v_cache 0.00004916 0.00588510 + layer.4.output 10.45909253 435.30307995 + ------------------------------------------------------------------------------------- + TOTAL 4.32287320 181.12397057 + (elements=948,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture elic-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 948736 +Total Bytes 202780 +BPFP 1.7099 bits/point +EBPFP 3.4198 equivalent bits/point +MSE 181.123971 +---------------------- -------------------------------------------------------- +Time: 3.000s Load: 0.005s, Pack+Encode: 1.705s, Decode+Unpack: 1.290s +---------------------- -------------------------------------------------------- +💾 Converting with 181.1240 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/sd-qa/SD-QA-sd-qa-98.zst + to output-fixed/kimiaudio/lambda0.02/elic-featurecoding-8bit-individual/sd-qa/SD-QA-sd-qa-98.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.8511 bits/point +Avg EBPFP 3.7021 equivalent bits/point +Avg MSE 259.584139 +Avg Time 4.728s +------------------------ ---------------------------- diff --git a/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..16aa34dfb95231d66de40ca82c181a6b8f174ed9 --- /dev/null +++ b/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.02_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.02_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 598 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.02_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench +Output output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,072B, BPFP=0.9784 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,004B, BPFP=3.0872 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,992B, BPFP=1.7346 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,624B, BPFP=3.0139 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,176B, BPFP=1.9630 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,028B, BPFP=2.8989 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,696B, BPFP=1.8704 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,292B, BPFP=2.9498 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,420B, BPFP=2.5887 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,892B, BPFP=2.8727 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,864B, BPFP=0.9332 +⌛️ [2/4] FRONTEND: Frontend time: 0.514s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.371s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12089630 29.84821687 + layer.0.v_cache 0.00001394 0.00554111 + layer.1.k_cache 0.01174029 3.80598921 + layer.1.v_cache 0.00000579 0.00226055 + layer.2.k_cache 0.00835781 0.47753741 + layer.2.v_cache 0.00001852 0.00585882 + layer.3.k_cache 0.02896961 2.25518950 + layer.3.v_cache 0.00001872 0.00654643 + layer.4.k_cache 0.00063445 0.12582871 + layer.4.v_cache 0.00005063 0.01322465 + layer.4.output 0.17246709 660.09468695 + ------------------------------------------------------------------------------------- + TOTAL 0.08105739 273.95347070 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 158060 +BPFP 1.7935 bits/point +EBPFP 3.5871 equivalent bits/point +MSE 273.953471 +---------------------- -------------------------------------------------------- +Time: 0.890s Load: 0.005s, Pack+Encode: 0.514s, Decode+Unpack: 0.371s +---------------------- -------------------------------------------------------- +💾 Converting with 273.9535 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-1.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-1.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,988B, BPFP=0.9742 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,952B, BPFP=3.1156 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,980B, BPFP=1.7539 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,492B, BPFP=3.0258 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,408B, BPFP=2.0328 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,024B, BPFP=2.9344 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,804B, BPFP=1.9148 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,272B, BPFP=2.9828 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,360B, BPFP=2.6094 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,944B, BPFP=2.9188 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,184B, BPFP=0.9538 +⌛️ [2/4] FRONTEND: Frontend time: 0.292s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.293s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08029544 29.26207275 + layer.0.v_cache 0.00001376 0.00570057 + layer.1.k_cache 0.03632395 3.59058838 + layer.1.v_cache 0.00000563 0.00247503 + layer.2.k_cache 0.00522719 0.50518332 + layer.2.v_cache 0.00001997 0.00676397 + layer.3.k_cache 0.02707962 2.13107281 + layer.3.v_cache 0.00001843 0.00780402 + layer.4.k_cache 0.00063057 0.13161319 + layer.4.v_cache 0.00005085 0.01469726 + layer.4.output 0.18451144 662.74447545 + ------------------------------------------------------------------------------------- + TOTAL 0.08477915 274.99231173 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 158408 +BPFP 1.8199 bits/point +EBPFP 3.6399 equivalent bits/point +MSE 274.992312 +---------------------- -------------------------------------------------------- +Time: 0.589s Load: 0.004s, Pack+Encode: 0.292s, Decode+Unpack: 0.293s +---------------------- -------------------------------------------------------- +💾 Converting with 274.9923 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-103.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-103.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,160B, BPFP=0.9375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,616B, BPFP=3.0189 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,212B, BPFP=1.6737 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,960B, BPFP=2.8997 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,524B, BPFP=1.9121 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,712B, BPFP=2.8547 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,136B, BPFP=1.8416 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,012B, BPFP=2.9092 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,828B, BPFP=2.5124 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,452B, BPFP=2.8074 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,356B, BPFP=0.8917 +⌛️ [2/4] FRONTEND: Frontend time: 0.265s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.296s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10874708 29.17087005 + layer.0.v_cache 0.00001608 0.00551374 + layer.1.k_cache 0.03270756 3.53729035 + layer.1.v_cache 0.00000572 0.00231741 + layer.2.k_cache 0.00777350 0.47127835 + layer.2.v_cache 0.00002080 0.00586643 + layer.3.k_cache 0.05920834 2.13542850 + layer.3.v_cache 0.00001846 0.00645699 + layer.4.k_cache 0.00062493 0.11951961 + layer.4.v_cache 0.00005036 0.01254706 + layer.4.output 0.17137358 624.78742733 + ------------------------------------------------------------------------------------- + TOTAL 0.08286988 259.35171057 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 162968 +BPFP 1.7417 bits/point +EBPFP 3.4834 equivalent bits/point +MSE 259.351711 +---------------------- -------------------------------------------------------- +Time: 0.565s Load: 0.003s, Pack+Encode: 0.265s, Decode+Unpack: 0.296s +---------------------- -------------------------------------------------------- +💾 Converting with 259.3517 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-106.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-106.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,972B, BPFP=0.9960 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,688B, BPFP=3.1426 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,680B, BPFP=1.7388 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,436B, BPFP=3.0921 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,188B, BPFP=2.0409 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,648B, BPFP=2.9343 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,400B, BPFP=1.8830 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,904B, BPFP=2.9856 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,168B, BPFP=2.6378 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,732B, BPFP=2.9511 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,556B, BPFP=0.9603 +⌛️ [2/4] FRONTEND: Frontend time: 0.241s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.299s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12073612 30.73515124 + layer.0.v_cache 0.00001348 0.00555983 + layer.1.k_cache 0.03279715 3.72012290 + layer.1.v_cache 0.00000568 0.00261617 + layer.2.k_cache 0.00696802 0.49590536 + layer.2.v_cache 0.00001833 0.00667430 + layer.3.k_cache 0.03227850 2.19789750 + layer.3.v_cache 0.00001834 0.00736837 + layer.4.k_cache 0.00060115 0.13420056 + layer.4.v_cache 0.00005191 0.01455481 + layer.4.output 0.18320683 688.53039148 + ------------------------------------------------------------------------------------- + TOTAL 0.08681979 285.70781126 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 155372 +BPFP 1.8308 bits/point +EBPFP 3.6617 equivalent bits/point +MSE 285.707811 +---------------------- -------------------------------------------------------- +Time: 0.543s Load: 0.004s, Pack+Encode: 0.241s, Decode+Unpack: 0.299s +---------------------- -------------------------------------------------------- +💾 Converting with 285.7078 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-111.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,948B, BPFP=0.9912 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,812B, BPFP=3.1675 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,712B, BPFP=1.7452 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,308B, BPFP=3.0665 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,088B, BPFP=2.0208 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,672B, BPFP=2.9391 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,540B, BPFP=1.9111 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,860B, BPFP=2.9768 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,060B, BPFP=2.6162 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,576B, BPFP=2.9199 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,256B, BPFP=0.9517 +⌛️ [2/4] FRONTEND: Frontend time: 0.240s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.271s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10595027 31.85469251 + layer.0.v_cache 0.00001376 0.00545437 + layer.1.k_cache 0.03307601 3.35678335 + layer.1.v_cache 0.00000550 0.00231803 + layer.2.k_cache 0.00695119 0.51145348 + layer.2.v_cache 0.00001929 0.00595729 + layer.3.k_cache 0.03450640 2.29206868 + layer.3.v_cache 0.00001844 0.00645480 + layer.4.k_cache 0.00061739 0.11867716 + layer.4.v_cache 0.00005297 0.01249696 + layer.4.output 0.18589628 688.82039835 + ------------------------------------------------------------------------------------- + TOTAL 0.08720501 285.87700854 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 154832 +BPFP 1.8245 bits/point +EBPFP 3.6489 equivalent bits/point +MSE 285.877009 +---------------------- -------------------------------------------------------- +Time: 0.514s Load: 0.004s, Pack+Encode: 0.240s, Decode+Unpack: 0.271s +---------------------- -------------------------------------------------------- +💾 Converting with 285.8770 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-114.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,072B, BPFP=1.0032 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,984B, BPFP=3.1614 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,824B, BPFP=1.7453 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,456B, BPFP=3.0570 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,400B, BPFP=2.0570 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,984B, BPFP=2.9636 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,776B, BPFP=1.9335 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,060B, BPFP=2.9786 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,328B, BPFP=2.6361 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,092B, BPFP=2.9850 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,048B, BPFP=0.9338 +⌛️ [2/4] FRONTEND: Frontend time: 0.242s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07772056 29.62151713 + layer.0.v_cache 0.00001351 0.00539934 + layer.1.k_cache 0.03518496 3.27709111 + layer.1.v_cache 0.00000556 0.00226629 + layer.2.k_cache 0.00230807 0.48677913 + layer.2.v_cache 0.00001833 0.00608306 + layer.3.k_cache 0.04514034 2.04634094 + layer.3.v_cache 0.00001901 0.00691290 + layer.4.k_cache 0.00062178 0.12838848 + layer.4.v_cache 0.00004990 0.01436210 + layer.4.output 0.18426369 669.65788879 + ------------------------------------------------------------------------------------- + TOTAL 0.08534870 277.83531541 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 157024 +BPFP 1.8269 bits/point +EBPFP 3.6538 equivalent bits/point +MSE 277.835315 +---------------------- -------------------------------------------------------- +Time: 0.557s Load: 0.004s, Pack+Encode: 0.242s, Decode+Unpack: 0.311s +---------------------- -------------------------------------------------------- +💾 Converting with 277.8353 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-115.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-115.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,024B, BPFP=1.0195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,704B, BPFP=3.1867 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,816B, BPFP=1.7890 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,240B, BPFP=3.0925 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,300B, BPFP=2.0901 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,800B, BPFP=3.0032 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,516B, BPFP=1.9310 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,076B, BPFP=3.0593 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,152B, BPFP=2.6688 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,888B, BPFP=3.0211 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,976B, BPFP=0.9849 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14568528 31.51834542 + layer.0.v_cache 0.00001322 0.00546765 + layer.1.k_cache 0.01379078 3.42922141 + layer.1.v_cache 0.00000548 0.00249468 + layer.2.k_cache 0.00727733 0.50851490 + layer.2.v_cache 0.00001786 0.00669769 + layer.3.k_cache 0.03082054 2.75868106 + layer.3.v_cache 0.00001879 0.00760974 + layer.4.k_cache 0.00061820 0.12939162 + layer.4.v_cache 0.00005377 0.01415226 + layer.4.output 0.19662610 698.20680659 + ------------------------------------------------------------------------------------- + TOTAL 0.09262847 289.75460133 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 156492 +BPFP 1.8680 bits/point +EBPFP 3.7360 equivalent bits/point +MSE 289.754601 +---------------------- -------------------------------------------------------- +Time: 0.587s Load: 0.004s, Pack+Encode: 0.259s, Decode+Unpack: 0.324s +---------------------- -------------------------------------------------------- +💾 Converting with 289.7546 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-119.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-119.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,988B, BPFP=0.9742 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,040B, BPFP=3.1328 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,032B, BPFP=1.7641 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,664B, BPFP=3.0594 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,440B, BPFP=2.0391 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,220B, BPFP=2.9727 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,824B, BPFP=1.9187 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,300B, BPFP=2.9883 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,384B, BPFP=2.6141 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,036B, BPFP=2.9367 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,620B, BPFP=0.9660 +⌛️ [2/4] FRONTEND: Frontend time: 0.264s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.327s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10149734 29.20462036 + layer.0.v_cache 0.00001421 0.00584559 + layer.1.k_cache 0.01324784 3.64120331 + layer.1.v_cache 0.00000586 0.00243553 + layer.2.k_cache 0.00706157 0.44856586 + layer.2.v_cache 0.00001822 0.00593914 + layer.3.k_cache 0.02712160 2.57072906 + layer.3.v_cache 0.00002016 0.00691139 + layer.4.k_cache 0.00063687 0.12683907 + layer.4.v_cache 0.00005451 0.01363429 + layer.4.output 0.17858309 660.72706473 + ------------------------------------------------------------------------------------- + TOTAL 0.08233881 274.18330451 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 159548 +BPFP 1.8330 bits/point +EBPFP 3.6661 equivalent bits/point +MSE 274.183305 +---------------------- -------------------------------------------------------- +Time: 0.595s Load: 0.004s, Pack+Encode: 0.264s, Decode+Unpack: 0.327s +---------------------- -------------------------------------------------------- +💾 Converting with 274.1833 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-120.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-120.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,280B, BPFP=0.9483 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,604B, BPFP=2.9820 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,496B, BPFP=1.7055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,284B, BPFP=2.9246 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,836B, BPFP=1.9461 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,924B, BPFP=2.8599 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,432B, BPFP=1.8736 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,152B, BPFP=2.9009 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,048B, BPFP=2.5230 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,744B, BPFP=2.8276 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,164B, BPFP=0.9279 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.296s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10772817 27.28247351 + layer.0.v_cache 0.00001348 0.00581029 + layer.1.k_cache 0.03276526 3.40646397 + layer.1.v_cache 0.00000562 0.00249523 + layer.2.k_cache 0.01020141 0.45289033 + layer.2.v_cache 0.00001934 0.00634014 + layer.3.k_cache 0.04694761 2.37676055 + layer.3.v_cache 0.00001910 0.00702955 + layer.4.k_cache 0.00062564 0.12496356 + layer.4.v_cache 0.00005323 0.01372008 + layer.4.output 0.17166949 607.97465107 + ------------------------------------------------------------------------------------- + TOTAL 0.08235678 252.32361792 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 166964 +BPFP 1.7639 bits/point +EBPFP 3.5278 equivalent bits/point +MSE 252.323618 +---------------------- -------------------------------------------------------- +Time: 0.559s Load: 0.004s, Pack+Encode: 0.259s, Decode+Unpack: 0.296s +---------------------- -------------------------------------------------------- +💾 Converting with 252.3236 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-128.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-128.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,164B, BPFP=0.9721 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,316B, BPFP=3.0715 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,184B, BPFP=1.7289 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,800B, BPFP=2.9744 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,504B, BPFP=1.9774 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,332B, BPFP=2.8863 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,932B, BPFP=1.8697 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,404B, BPFP=2.8998 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,480B, BPFP=2.5377 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,996B, BPFP=2.8230 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,052B, BPFP=0.9427 +⌛️ [2/4] FRONTEND: Frontend time: 0.240s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.294s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12335577 28.04308347 + layer.0.v_cache 0.00001400 0.00566582 + layer.1.k_cache 0.01519604 3.52017543 + layer.1.v_cache 0.00000547 0.00234778 + layer.2.k_cache 0.01098718 0.52312047 + layer.2.v_cache 0.00001779 0.00682049 + layer.3.k_cache 0.07416311 2.21842865 + layer.3.v_cache 0.00001944 0.00778131 + layer.4.k_cache 0.00061719 0.12870203 + layer.4.v_cache 0.00006444 0.01414938 + layer.4.output 0.17504946 631.12725904 + ------------------------------------------------------------------------------------- + TOTAL 0.08528157 261.90359342 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 161164 +BPFP 1.7847 bits/point +EBPFP 3.5694 equivalent bits/point +MSE 261.903593 +---------------------- -------------------------------------------------------- +Time: 0.538s Load: 0.004s, Pack+Encode: 0.240s, Decode+Unpack: 0.294s +---------------------- -------------------------------------------------------- +💾 Converting with 261.9036 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-132.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,004B, BPFP=1.0154 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,608B, BPFP=3.1672 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,820B, BPFP=1.7898 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,976B, BPFP=3.0390 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,004B, BPFP=2.0300 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,460B, BPFP=2.9343 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,484B, BPFP=1.9245 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,688B, BPFP=2.9805 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,040B, BPFP=2.6461 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,412B, BPFP=2.9245 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,912B, BPFP=0.9541 +⌛️ [2/4] FRONTEND: Frontend time: 0.233s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.282s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13955514 30.59634360 + layer.0.v_cache 0.00001392 0.00600933 + layer.1.k_cache 0.03454666 3.84971797 + layer.1.v_cache 0.00000526 0.00255282 + layer.2.k_cache 0.00236268 0.51038271 + layer.2.v_cache 0.00001660 0.00644738 + layer.3.k_cache 0.02874016 2.46678380 + layer.3.v_cache 0.00001899 0.00737188 + layer.4.k_cache 0.00062191 0.13390170 + layer.4.v_cache 0.00005265 0.01406097 + layer.4.output 0.18728454 694.30212199 + ------------------------------------------------------------------------------------- + TOTAL 0.08923092 288.10049565 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 153408 +BPFP 1.8312 bits/point +EBPFP 3.6623 equivalent bits/point +MSE 288.100496 +---------------------- -------------------------------------------------------- +Time: 0.520s Load: 0.005s, Pack+Encode: 0.233s, Decode+Unpack: 0.282s +---------------------- -------------------------------------------------------- +💾 Converting with 288.1005 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-133.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-133.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,024B, BPFP=1.0195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,652B, BPFP=3.1761 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,908B, BPFP=1.8076 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,360B, BPFP=3.1169 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,072B, BPFP=2.0438 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,644B, BPFP=2.9716 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,524B, BPFP=1.9326 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,816B, BPFP=3.0065 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,092B, BPFP=2.6567 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,660B, BPFP=2.9748 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,824B, BPFP=0.9805 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.302s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11454884 31.62459099 + layer.0.v_cache 0.00001351 0.00619021 + layer.1.k_cache 0.01665594 3.43574702 + layer.1.v_cache 0.00000570 0.00277274 + layer.2.k_cache 0.00836552 0.50660502 + layer.2.v_cache 0.00001974 0.00669016 + layer.3.k_cache 0.03124422 2.46508631 + layer.3.v_cache 0.00001876 0.00750963 + layer.4.k_cache 0.00062231 0.13069242 + layer.4.v_cache 0.00005487 0.01481863 + layer.4.output 0.19663499 697.78942486 + ------------------------------------------------------------------------------------- + TOTAL 0.09105849 289.57215748 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 155576 +BPFP 1.8570 bits/point +EBPFP 3.7141 equivalent bits/point +MSE 289.572157 +---------------------- -------------------------------------------------------- +Time: 0.558s Load: 0.004s, Pack+Encode: 0.252s, Decode+Unpack: 0.302s +---------------------- -------------------------------------------------------- +💾 Converting with 289.5722 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-135.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-135.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,188B, BPFP=0.9650 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,332B, BPFP=3.0379 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,200B, BPFP=1.7113 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,872B, BPFP=2.9524 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,384B, BPFP=1.9315 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,412B, BPFP=2.8668 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,028B, BPFP=1.8653 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,708B, BPFP=2.9219 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,596B, BPFP=2.5290 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,232B, BPFP=2.8333 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,048B, BPFP=0.9579 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.283s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12290487 28.30181013 + layer.0.v_cache 0.00001422 0.00571526 + layer.1.k_cache 0.01283375 3.77207547 + layer.1.v_cache 0.00000579 0.00262959 + layer.2.k_cache 0.01100613 0.48503004 + layer.2.v_cache 0.00001929 0.00619581 + layer.3.k_cache 0.03011658 2.13349896 + layer.3.v_cache 0.00001858 0.00686659 + layer.4.k_cache 0.00063020 0.12674373 + layer.4.v_cache 0.00005223 0.01430755 + layer.4.output 0.17536174 615.31494473 + ------------------------------------------------------------------------------------- + TOTAL 0.08265493 255.41526390 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 163000 +BPFP 1.7835 bits/point +EBPFP 3.5671 equivalent bits/point +MSE 255.415264 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.004s, Pack+Encode: 0.255s, Decode+Unpack: 0.283s +---------------------- -------------------------------------------------------- +💾 Converting with 255.4153 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-144.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,152B, BPFP=0.9471 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,404B, BPFP=3.0154 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,200B, BPFP=1.6912 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,816B, BPFP=2.9074 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,436B, BPFP=1.9184 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,516B, BPFP=2.8522 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,012B, BPFP=1.8404 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,672B, BPFP=2.8809 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,692B, BPFP=2.5169 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,376B, BPFP=2.8265 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,928B, BPFP=0.9172 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.291s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11218034 29.66394761 + layer.0.v_cache 0.00001438 0.00569631 + layer.1.k_cache 0.03157643 3.56606158 + layer.1.v_cache 0.00000545 0.00259071 + layer.2.k_cache 0.00802605 0.48532998 + layer.2.v_cache 0.00001826 0.00628366 + layer.3.k_cache 0.04519741 2.24837287 + layer.3.v_cache 0.00001794 0.00694955 + layer.4.k_cache 0.00061137 0.12193914 + layer.4.v_cache 0.00005913 0.01372366 + layer.4.output 0.17010492 631.26586134 + ------------------------------------------------------------------------------------- + TOTAL 0.08167301 262.05776026 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 162204 +BPFP 1.7539 bits/point +EBPFP 3.5079 equivalent bits/point +MSE 262.057760 +---------------------- -------------------------------------------------------- +Time: 0.544s Load: 0.004s, Pack+Encode: 0.248s, Decode+Unpack: 0.291s +---------------------- -------------------------------------------------------- +💾 Converting with 262.0578 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-147.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,160B, BPFP=0.9375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,536B, BPFP=3.0044 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,312B, BPFP=1.6919 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,160B, BPFP=2.9360 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,540B, BPFP=1.9150 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,776B, BPFP=2.8663 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,140B, BPFP=1.8423 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,076B, BPFP=2.9208 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,812B, BPFP=2.5094 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,532B, BPFP=2.8219 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,944B, BPFP=0.9070 +⌛️ [2/4] FRONTEND: Frontend time: 0.265s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10146660 27.52410747 + layer.0.v_cache 0.00001344 0.00549465 + layer.1.k_cache 0.03090366 3.52845161 + layer.1.v_cache 0.00000566 0.00240165 + layer.2.k_cache 0.01188999 0.48424623 + layer.2.v_cache 0.00001845 0.00623024 + layer.3.k_cache 0.02542871 2.31976407 + layer.3.v_cache 0.00001884 0.00691394 + layer.4.k_cache 0.00060243 0.12132238 + layer.4.v_cache 0.00005224 0.01326940 + layer.4.output 0.16839122 624.85070598 + ------------------------------------------------------------------------------------- + TOTAL 0.07936109 259.29218491 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 163988 +BPFP 1.7526 bits/point +EBPFP 3.5052 equivalent bits/point +MSE 259.292185 +---------------------- -------------------------------------------------------- +Time: 0.580s Load: 0.004s, Pack+Encode: 0.265s, Decode+Unpack: 0.311s +---------------------- -------------------------------------------------------- +💾 Converting with 259.2922 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-150.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-150.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,664B, BPFP=0.9620 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,448B, BPFP=2.9633 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,064B, BPFP=1.7092 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,072B, BPFP=2.8995 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,368B, BPFP=1.9307 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,684B, BPFP=2.8336 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,732B, BPFP=1.8227 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,940B, BPFP=2.8770 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,980B, BPFP=2.5442 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,460B, BPFP=2.7955 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,632B, BPFP=0.9858 +⌛️ [2/4] FRONTEND: Frontend time: 0.268s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12981709 28.23642897 + layer.0.v_cache 0.00001628 0.00579023 + layer.1.k_cache 0.06761242 3.57761847 + layer.1.v_cache 0.00000547 0.00229678 + layer.2.k_cache 0.00384689 0.47627101 + layer.2.v_cache 0.00001812 0.00625307 + layer.3.k_cache 0.05327560 2.19879830 + layer.3.v_cache 0.00001896 0.00725807 + layer.4.k_cache 0.00060800 0.12484861 + layer.4.v_cache 0.00005116 0.01348370 + layer.4.output 0.15518735 572.06740101 + ------------------------------------------------------------------------------------- + TOTAL 0.07891655 237.59534437 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 178044 +BPFP 1.7787 bits/point +EBPFP 3.5575 equivalent bits/point +MSE 237.595344 +---------------------- -------------------------------------------------------- +Time: 0.603s Load: 0.005s, Pack+Encode: 0.268s, Decode+Unpack: 0.330s +---------------------- -------------------------------------------------------- +💾 Converting with 237.5953 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-154.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-154.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,044B, BPFP=0.9730 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,084B, BPFP=3.1026 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,996B, BPFP=1.7353 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,632B, BPFP=3.0154 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,252B, BPFP=1.9776 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,184B, BPFP=2.9290 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,732B, BPFP=1.8773 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,548B, BPFP=2.9992 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,408B, BPFP=2.5864 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,096B, BPFP=2.9120 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,284B, BPFP=0.9448 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.340s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14887507 30.33813175 + layer.0.v_cache 0.00001345 0.00570372 + layer.1.k_cache 0.03259810 3.73725759 + layer.1.v_cache 0.00000523 0.00234397 + layer.2.k_cache 0.00244350 0.51688131 + layer.2.v_cache 0.00001834 0.00641574 + layer.3.k_cache 0.10864470 2.35695884 + layer.3.v_cache 0.00001904 0.00708680 + layer.4.k_cache 0.00060864 0.12951055 + layer.4.v_cache 0.00004940 0.01380564 + layer.4.output 0.18628448 660.79106041 + ------------------------------------------------------------------------------------- + TOTAL 0.09395687 274.27361875 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 159260 +BPFP 1.8071 bits/point +EBPFP 3.6143 equivalent bits/point +MSE 274.273619 +---------------------- -------------------------------------------------------- +Time: 0.606s Load: 0.005s, Pack+Encode: 0.261s, Decode+Unpack: 0.340s +---------------------- -------------------------------------------------------- +💾 Converting with 274.2736 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-155.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-155.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,396B, BPFP=0.9473 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,880B, BPFP=2.9635 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,588B, BPFP=1.6833 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,408B, BPFP=2.8806 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,764B, BPFP=1.8897 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,084B, BPFP=2.8237 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,268B, BPFP=1.8027 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,228B, BPFP=2.8490 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,160B, BPFP=2.4860 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,784B, BPFP=2.7711 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,784B, BPFP=0.9226 +⌛️ [2/4] FRONTEND: Frontend time: 0.285s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.293s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422350 28.66092708 + layer.0.v_cache 0.00001493 0.00581416 + layer.1.k_cache 0.01422887 3.58034387 + layer.1.v_cache 0.00000595 0.00245231 + layer.2.k_cache 0.00668612 0.47722184 + layer.2.v_cache 0.00001902 0.00604548 + layer.3.k_cache 0.02497406 2.31563817 + layer.3.v_cache 0.00001869 0.00669423 + layer.4.k_cache 0.00061935 0.11964622 + layer.4.v_cache 0.00005212 0.01328111 + layer.4.output 0.16899524 603.13483146 + ------------------------------------------------------------------------------------- + TOTAL 0.07904761 250.41952263 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 168344 +BPFP 1.7385 bits/point +EBPFP 3.4770 equivalent bits/point +MSE 250.419523 +---------------------- -------------------------------------------------------- +Time: 0.582s Load: 0.004s, Pack+Encode: 0.285s, Decode+Unpack: 0.293s +---------------------- -------------------------------------------------------- +💾 Converting with 250.4195 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-157.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-157.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 113, 128) +Output shape: (1, 113, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) -> torch.Size([1, 1, 113, 512]) + layer.4.output: torch.Size([1, 113, 3584]) -> torch.Size([1, 1, 113, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,008B, BPFP=0.8308 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,000B, BPFP=2.7655 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,496B, BPFP=1.5896 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,480B, BPFP=2.6936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,744B, BPFP=1.7622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,088B, BPFP=2.6394 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,056B, BPFP=1.6670 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,264B, BPFP=2.6637 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,856B, BPFP=2.3308 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,776B, BPFP=2.5962 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,664B, BPFP=0.8230 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.334s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 113, 128]) + layer.0.v_cache: torch.Size([1, 4, 113, 128]) + layer.1.k_cache: torch.Size([1, 4, 113, 128]) + layer.1.v_cache: torch.Size([1, 4, 113, 128]) + layer.2.k_cache: torch.Size([1, 4, 113, 128]) + layer.2.v_cache: torch.Size([1, 4, 113, 128]) + layer.3.k_cache: torch.Size([1, 4, 113, 128]) + layer.3.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.k_cache: torch.Size([1, 4, 113, 128]) + layer.4.v_cache: torch.Size([1, 4, 113, 128]) + layer.4.output: torch.Size([1, 113, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10214825 27.68570459 + layer.0.v_cache 0.00001637 0.00564084 + layer.1.k_cache 0.04284562 3.41055325 + layer.1.v_cache 0.00000578 0.00248218 + layer.2.k_cache 0.00913706 0.45986365 + layer.2.v_cache 0.00001928 0.00592329 + layer.3.k_cache 0.05585717 2.28685065 + layer.3.v_cache 0.00001870 0.00661707 + layer.4.k_cache 0.00065968 0.12641466 + layer.4.v_cache 0.00005753 0.01401373 + layer.4.output 10.09706179 469.41411188 + ------------------------------------------------------------------------------------- + TOTAL 4.17001165 195.28840277 + (elements=983,552) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 983552 +Total Bytes 197432 +BPFP 1.6059 bits/point +EBPFP 3.2117 equivalent bits/point +MSE 195.288403 +---------------------- -------------------------------------------------------- +Time: 0.599s Load: 0.005s, Pack+Encode: 0.260s, Decode+Unpack: 0.334s +---------------------- -------------------------------------------------------- +💾 Converting with 195.2884 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-159.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-159.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,512B, BPFP=0.9464 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,196B, BPFP=2.9526 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,056B, BPFP=1.7266 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,852B, BPFP=2.8935 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,408B, BPFP=1.9588 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,640B, BPFP=2.8571 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,684B, BPFP=1.8345 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,756B, BPFP=2.8771 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,752B, BPFP=2.5330 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,288B, BPFP=2.7967 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,440B, BPFP=0.9920 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.295s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11481400 27.26925492 + layer.0.v_cache 0.00001396 0.00560159 + layer.1.k_cache 0.03135788 3.41102751 + layer.1.v_cache 0.00000545 0.00246635 + layer.2.k_cache 0.00370715 0.47997334 + layer.2.v_cache 0.00001965 0.00608466 + layer.3.k_cache 0.03930106 2.60978447 + layer.3.v_cache 0.00001886 0.00640871 + layer.4.k_cache 0.00060819 0.12421515 + layer.4.v_cache 0.00005437 0.01341823 + layer.4.output 0.16591480 576.19809655 + ------------------------------------------------------------------------------------- + TOTAL 0.07948848 239.25381828 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 176584 +BPFP 1.7835 bits/point +EBPFP 3.5671 equivalent bits/point +MSE 239.253818 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.004s, Pack+Encode: 0.253s, Decode+Unpack: 0.295s +---------------------- -------------------------------------------------------- +💾 Converting with 239.2538 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-160.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,380B, BPFP=0.9445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,664B, BPFP=2.9256 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,588B, BPFP=1.6833 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,320B, BPFP=2.8652 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,756B, BPFP=1.8883 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,024B, BPFP=2.8132 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,084B, BPFP=1.7704 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,052B, BPFP=2.8181 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,116B, BPFP=2.4782 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,712B, BPFP=2.7584 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,064B, BPFP=0.8543 +⌛️ [2/4] FRONTEND: Frontend time: 0.279s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09788724 29.59066670 + layer.0.v_cache 0.00001308 0.00566646 + layer.1.k_cache 0.03299916 3.70704565 + layer.1.v_cache 0.00000544 0.00249627 + layer.2.k_cache 0.00367894 0.46195230 + layer.2.v_cache 0.00002012 0.00668171 + layer.3.k_cache 0.03939044 2.15357200 + layer.3.v_cache 0.00001875 0.00667274 + layer.4.k_cache 0.00061788 0.12792604 + layer.4.v_cache 0.00005010 0.01356042 + layer.4.output 0.17021053 602.77116774 + ------------------------------------------------------------------------------------- + TOTAL 0.08036205 250.32202438 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 164760 +BPFP 1.7015 bits/point +EBPFP 3.4030 equivalent bits/point +MSE 250.322024 +---------------------- -------------------------------------------------------- +Time: 0.609s Load: 0.004s, Pack+Encode: 0.279s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 250.3220 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-167.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-167.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,072B, BPFP=0.8785 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,700B, BPFP=2.8501 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,008B, BPFP=1.5926 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,200B, BPFP=2.7778 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,448B, BPFP=1.8009 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,768B, BPFP=2.7153 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,848B, BPFP=1.7141 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,156B, BPFP=2.7714 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,516B, BPFP=2.3895 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,532B, BPFP=2.6811 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 49,996B, BPFP=1.0333 +⌛️ [2/4] FRONTEND: Frontend time: 0.254s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12672405 27.12658239 + layer.0.v_cache 0.00001401 0.00526945 + layer.1.k_cache 0.12044146 3.28284511 + layer.1.v_cache 0.00000568 0.00231541 + layer.2.k_cache 0.00570256 0.42939048 + layer.2.v_cache 0.00001866 0.00555695 + layer.3.k_cache 0.03611955 2.12043126 + layer.3.v_cache 0.00001983 0.00642105 + layer.4.k_cache 0.00066233 0.11840883 + layer.4.v_cache 0.00005331 0.01305998 + layer.4.output 10.56037249 482.93923611 + ------------------------------------------------------------------------------------- + TOTAL 4.36543346 200.80499610 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 203244 +BPFP 1.7297 bits/point +EBPFP 3.4594 equivalent bits/point +MSE 200.804996 +---------------------- -------------------------------------------------------- +Time: 0.582s Load: 0.006s, Pack+Encode: 0.254s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 200.8050 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-169.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-169.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,392B, BPFP=0.9574 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,732B, BPFP=2.9709 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,540B, BPFP=1.6939 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,356B, BPFP=2.9041 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,864B, BPFP=1.9290 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,892B, BPFP=2.8217 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,276B, BPFP=1.8246 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,140B, BPFP=2.8658 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,052B, BPFP=2.4950 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,720B, BPFP=2.7912 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,280B, BPFP=0.8949 +⌛️ [2/4] FRONTEND: Frontend time: 0.286s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13733016 27.24138295 + layer.0.v_cache 0.00001373 0.00596564 + layer.1.k_cache 0.03155638 3.82552615 + layer.1.v_cache 0.00000617 0.00280650 + layer.2.k_cache 0.00751131 0.50254406 + layer.2.v_cache 0.00001754 0.00674861 + layer.3.k_cache 0.02548880 2.05874166 + layer.3.v_cache 0.00001765 0.00729664 + layer.4.k_cache 0.00061011 0.12979915 + layer.4.v_cache 0.00005045 0.01437003 + layer.4.output 0.16239834 602.19434862 + ------------------------------------------------------------------------------------- + TOTAL 0.07878769 249.95033069 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 166244 +BPFP 1.7363 bits/point +EBPFP 3.4727 equivalent bits/point +MSE 249.950331 +---------------------- -------------------------------------------------------- +Time: 0.632s Load: 0.005s, Pack+Encode: 0.286s, Decode+Unpack: 0.341s +---------------------- -------------------------------------------------------- +💾 Converting with 249.9503 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-178.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-178.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,000B, BPFP=0.9889 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,876B, BPFP=3.1400 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,936B, BPFP=1.7674 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,376B, BPFP=3.0411 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,240B, BPFP=2.0253 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,900B, BPFP=2.9470 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,680B, BPFP=1.9146 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,152B, BPFP=2.9968 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,380B, BPFP=2.6464 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,924B, BPFP=2.9517 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,432B, BPFP=0.9446 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10979201 29.39363504 + layer.0.v_cache 0.00001339 0.00533553 + layer.1.k_cache 0.01160004 3.29194815 + layer.1.v_cache 0.00000567 0.00241950 + layer.2.k_cache 0.00714037 0.50621438 + layer.2.v_cache 0.00001829 0.00598708 + layer.3.k_cache 0.04622446 2.21267913 + layer.3.v_cache 0.00001896 0.00675938 + layer.4.k_cache 0.00064378 0.11881703 + layer.4.v_cache 0.00005251 0.01352478 + layer.4.output 0.18415536 667.07402803 + ------------------------------------------------------------------------------------- + TOTAL 0.08615276 276.76914801 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 156896 +BPFP 1.8254 bits/point +EBPFP 3.6508 equivalent bits/point +MSE 276.769148 +---------------------- -------------------------------------------------------- +Time: 0.579s Load: 0.004s, Pack+Encode: 0.262s, Decode+Unpack: 0.312s +---------------------- -------------------------------------------------------- +💾 Converting with 276.7691 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-181.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-181.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,024B, BPFP=1.0195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,720B, BPFP=3.1899 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,776B, BPFP=1.7808 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,180B, BPFP=3.0804 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,220B, BPFP=2.0739 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,644B, BPFP=2.9716 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,492B, BPFP=1.9261 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,880B, BPFP=3.0195 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,180B, BPFP=2.6745 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,896B, BPFP=3.0227 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,692B, BPFP=1.0057 +⌛️ [2/4] FRONTEND: Frontend time: 0.265s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11290005 30.90390371 + layer.0.v_cache 0.00001570 0.00561793 + layer.1.k_cache 0.03434501 3.50524268 + layer.1.v_cache 0.00000534 0.00219646 + layer.2.k_cache 0.00238675 0.50886318 + layer.2.v_cache 0.00001838 0.00615299 + layer.3.k_cache 0.02838124 2.69120868 + layer.3.v_cache 0.00002047 0.00704147 + layer.4.k_cache 0.00062757 0.12406909 + layer.4.v_cache 0.00005455 0.01353395 + layer.4.output 0.18420308 699.55931122 + ------------------------------------------------------------------------------------- + TOTAL 0.08636333 290.27547110 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 156704 +BPFP 1.8705 bits/point +EBPFP 3.7410 equivalent bits/point +MSE 290.275471 +---------------------- -------------------------------------------------------- +Time: 0.595s Load: 0.005s, Pack+Encode: 0.265s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 290.2755 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-184.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-184.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,004B, BPFP=1.0024 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,676B, BPFP=3.1402 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,844B, BPFP=1.7716 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,380B, BPFP=3.0809 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,052B, BPFP=2.0136 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,704B, BPFP=2.9455 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,584B, BPFP=1.9199 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,132B, BPFP=3.0312 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,168B, BPFP=2.6378 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,800B, BPFP=2.9647 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,540B, BPFP=0.9884 +⌛️ [2/4] FRONTEND: Frontend time: 0.279s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09752406 30.66482622 + layer.0.v_cache 0.00001393 0.00584384 + layer.1.k_cache 0.01493649 3.79973896 + layer.1.v_cache 0.00000588 0.00243814 + layer.2.k_cache 0.01003601 0.50501178 + layer.2.v_cache 0.00001805 0.00608501 + layer.3.k_cache 0.04451986 2.35030619 + layer.3.v_cache 0.00001914 0.00697346 + layer.4.k_cache 0.00061869 0.12158824 + layer.4.v_cache 0.00005251 0.01360155 + layer.4.output 0.18409089 688.38993819 + ------------------------------------------------------------------------------------- + TOTAL 0.08566946 285.65917534 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 156884 +BPFP 1.8487 bits/point +EBPFP 3.6973 equivalent bits/point +MSE 285.659175 +---------------------- -------------------------------------------------------- +Time: 0.602s Load: 0.004s, Pack+Encode: 0.279s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 285.6592 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-187.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-187.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,148B, BPFP=0.9691 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,280B, BPFP=3.0648 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,104B, BPFP=1.7139 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,724B, BPFP=2.9601 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,188B, BPFP=1.9179 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,112B, BPFP=2.8449 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,900B, BPFP=1.8637 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,540B, BPFP=2.9255 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,468B, BPFP=2.5354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,068B, BPFP=2.8366 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,516B, BPFP=0.9551 +⌛️ [2/4] FRONTEND: Frontend time: 0.264s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.288s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13658243 27.61769637 + layer.0.v_cache 0.00001379 0.00550814 + layer.1.k_cache 0.01234693 3.65415127 + layer.1.v_cache 0.00000566 0.00241135 + layer.2.k_cache 0.00375735 0.44859718 + layer.2.v_cache 0.00001852 0.00585762 + layer.3.k_cache 0.02700986 2.02034309 + layer.3.v_cache 0.00001930 0.00658475 + layer.4.k_cache 0.00062007 0.12545195 + layer.4.v_cache 0.00005146 0.01311782 + layer.4.output 0.17614663 628.16593158 + ------------------------------------------------------------------------------------- + TOTAL 0.08314423 260.65066121 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 161048 +BPFP 1.7834 bits/point +EBPFP 3.5668 equivalent bits/point +MSE 260.650661 +---------------------- -------------------------------------------------------- +Time: 0.557s Load: 0.004s, Pack+Encode: 0.264s, Decode+Unpack: 0.288s +---------------------- -------------------------------------------------------- +💾 Converting with 260.6507 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-193.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-193.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 73, 128) +Output shape: (1, 73, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) -> torch.Size([1, 1, 73, 512]) + layer.4.output: torch.Size([1, 73, 3584]) -> torch.Size([1, 1, 73, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,876B, BPFP=1.0437 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,028B, BPFP=3.2166 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,616B, BPFP=1.8442 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 14,192B, BPFP=3.0377 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,932B, BPFP=2.1259 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,816B, BPFP=2.9572 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,068B, BPFP=1.9409 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,952B, BPFP=2.9863 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,468B, BPFP=2.6687 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,092B, BPFP=3.0163 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,556B, BPFP=1.0261 +⌛️ [2/4] FRONTEND: Frontend time: 0.240s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.247s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 73, 128]) + layer.0.v_cache: torch.Size([1, 4, 73, 128]) + layer.1.k_cache: torch.Size([1, 4, 73, 128]) + layer.1.v_cache: torch.Size([1, 4, 73, 128]) + layer.2.k_cache: torch.Size([1, 4, 73, 128]) + layer.2.v_cache: torch.Size([1, 4, 73, 128]) + layer.3.k_cache: torch.Size([1, 4, 73, 128]) + layer.3.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.k_cache: torch.Size([1, 4, 73, 128]) + layer.4.v_cache: torch.Size([1, 4, 73, 128]) + layer.4.output: torch.Size([1, 73, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08738674 31.35889341 + layer.0.v_cache 0.00001368 0.00594450 + layer.1.k_cache 0.01286346 4.04195958 + layer.1.v_cache 0.00000597 0.00253733 + layer.2.k_cache 0.00251921 0.47704305 + layer.2.v_cache 0.00001769 0.00661633 + layer.3.k_cache 0.08324167 2.30931217 + layer.3.v_cache 0.00002021 0.00755662 + layer.4.k_cache 0.00062021 0.13123801 + layer.4.v_cache 0.00005375 0.01479779 + layer.4.output 0.19567458 731.28014922 + ------------------------------------------------------------------------------------- + TOTAL 0.09155674 303.37158490 + (elements=635,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 635392 +Total Bytes 149596 +BPFP 1.8835 bits/point +EBPFP 3.7670 equivalent bits/point +MSE 303.371585 +---------------------- -------------------------------------------------------- +Time: 0.490s Load: 0.003s, Pack+Encode: 0.240s, Decode+Unpack: 0.247s +---------------------- -------------------------------------------------------- +💾 Converting with 303.3716 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-194.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-194.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,156B, BPFP=0.9706 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,292B, BPFP=3.0670 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,204B, BPFP=1.7327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,820B, BPFP=2.9782 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,356B, BPFP=1.9495 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,344B, BPFP=2.8886 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,828B, BPFP=1.8502 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,520B, BPFP=2.9217 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,456B, BPFP=2.5331 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,216B, BPFP=2.8645 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,472B, BPFP=0.9809 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.289s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12176956 27.36542851 + layer.0.v_cache 0.00001396 0.00534101 + layer.1.k_cache 0.01120206 3.21580561 + layer.1.v_cache 0.00000562 0.00235211 + layer.2.k_cache 0.00678912 0.42571856 + layer.2.v_cache 0.00001849 0.00573385 + layer.3.k_cache 0.02622183 2.25636034 + layer.3.v_cache 0.00001965 0.00614404 + layer.4.k_cache 0.00062413 0.11743084 + layer.4.v_cache 0.00005248 0.01242113 + layer.4.output 0.17940697 636.90807874 + ------------------------------------------------------------------------------------- + TOTAL 0.08368034 264.22172278 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 162664 +BPFP 1.8013 bits/point +EBPFP 3.6026 equivalent bits/point +MSE 264.221723 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.004s, Pack+Encode: 0.260s, Decode+Unpack: 0.289s +---------------------- -------------------------------------------------------- +💾 Converting with 264.2217 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-197.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,940B, BPFP=0.8839 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,296B, BPFP=2.8714 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,072B, BPFP=1.6476 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,856B, BPFP=2.8060 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,124B, BPFP=1.8042 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,476B, BPFP=2.7494 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,592B, BPFP=1.7250 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,732B, BPFP=2.7875 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,312B, BPFP=2.4274 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,224B, BPFP=2.7119 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 45,108B, BPFP=0.9589 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.283s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120816 27.89453590 + layer.0.v_cache 0.00001607 0.00549190 + layer.1.k_cache 0.10678518 3.39134463 + layer.1.v_cache 0.00000585 0.00226993 + layer.2.k_cache 0.00813519 0.51066586 + layer.2.v_cache 0.00001919 0.00629572 + layer.3.k_cache 0.06560528 2.13115569 + layer.3.v_cache 0.00001973 0.00710643 + layer.4.k_cache 0.00062300 0.11857397 + layer.4.v_cache 0.00005287 0.01332055 + layer.4.output 10.86395687 505.89022109 + ------------------------------------------------------------------------------------- + TOTAL 4.49059815 210.31248872 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 195732 +BPFP 1.7133 bits/point +EBPFP 3.4267 equivalent bits/point +MSE 210.312489 +---------------------- -------------------------------------------------------- +Time: 0.540s Load: 0.005s, Pack+Encode: 0.251s, Decode+Unpack: 0.283s +---------------------- -------------------------------------------------------- +💾 Converting with 210.3125 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-206.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-206.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,236B, BPFP=0.9404 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,536B, BPFP=2.9698 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,344B, BPFP=1.6782 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,012B, BPFP=2.8757 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,624B, BPFP=1.9080 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,712B, BPFP=2.8218 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,004B, BPFP=1.7967 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,740B, BPFP=2.8269 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,888B, BPFP=2.4943 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,516B, BPFP=2.7866 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,520B, BPFP=0.9370 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.283s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12848519 27.00415039 + layer.0.v_cache 0.00001371 0.00521821 + layer.1.k_cache 0.03007136 3.39601574 + layer.1.v_cache 0.00000570 0.00229208 + layer.2.k_cache 0.00936044 0.51378360 + layer.2.v_cache 0.00001842 0.00603586 + layer.3.k_cache 0.04065019 2.18448069 + layer.3.v_cache 0.00001877 0.00697012 + layer.4.k_cache 0.00062204 0.12415735 + layer.4.v_cache 0.00010837 0.01305873 + layer.4.output 0.16426703 606.79069171 + ------------------------------------------------------------------------------------- + TOTAL 0.07995432 251.81123557 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 165132 +BPFP 1.7445 bits/point +EBPFP 3.4891 equivalent bits/point +MSE 251.811236 +---------------------- -------------------------------------------------------- +Time: 0.533s Load: 0.005s, Pack+Encode: 0.246s, Decode+Unpack: 0.283s +---------------------- -------------------------------------------------------- +💾 Converting with 251.8112 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-211.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-211.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,012B, BPFP=0.9789 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,964B, BPFP=3.1180 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,848B, BPFP=1.7281 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,568B, BPFP=3.0406 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,336B, BPFP=2.0187 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,156B, BPFP=2.9602 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,796B, BPFP=1.9133 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,272B, BPFP=2.9828 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,404B, BPFP=2.6180 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,032B, BPFP=2.9359 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,288B, BPFP=0.9567 +⌛️ [2/4] FRONTEND: Frontend time: 0.236s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.306s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09257430 28.43540649 + layer.0.v_cache 0.00001329 0.00523177 + layer.1.k_cache 0.01202901 3.36513824 + layer.1.v_cache 0.00000544 0.00204479 + layer.2.k_cache 0.00549439 0.49963627 + layer.2.v_cache 0.00001849 0.00584326 + layer.3.k_cache 0.11014032 2.54143867 + layer.3.v_cache 0.00001916 0.00612394 + layer.4.k_cache 0.00063444 0.11742120 + layer.4.v_cache 0.00005494 0.01253867 + layer.4.output 0.19143520 655.09464286 + ------------------------------------------------------------------------------------- + TOTAL 0.09182531 271.80313667 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 158676 +BPFP 1.8230 bits/point +EBPFP 3.6460 equivalent bits/point +MSE 271.803137 +---------------------- -------------------------------------------------------- +Time: 0.546s Load: 0.003s, Pack+Encode: 0.236s, Decode+Unpack: 0.306s +---------------------- -------------------------------------------------------- +💾 Converting with 271.8031 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-213.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,200B, BPFP=0.9673 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,272B, BPFP=3.0268 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,216B, BPFP=1.7143 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,868B, BPFP=2.9516 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,476B, BPFP=1.9487 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,424B, BPFP=2.8690 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,116B, BPFP=1.8817 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,692B, BPFP=2.9189 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,584B, BPFP=2.5268 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,268B, BPFP=2.8400 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,124B, BPFP=0.9334 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.300s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12365316 25.80410621 + layer.0.v_cache 0.00001332 0.00568449 + layer.1.k_cache 0.03511774 3.44041479 + layer.1.v_cache 0.00000569 0.00279788 + layer.2.k_cache 0.01236608 0.41230356 + layer.2.v_cache 0.00001835 0.00600692 + layer.3.k_cache 0.02851222 2.14195542 + layer.3.v_cache 0.00001844 0.00676250 + layer.4.k_cache 0.00063207 0.12976376 + layer.4.v_cache 0.00005066 0.01381382 + layer.4.output 0.17964420 611.94308036 + ------------------------------------------------------------------------------------- + TOTAL 0.08575865 253.85677481 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 162240 +BPFP 1.7752 bits/point +EBPFP 3.5504 equivalent bits/point +MSE 253.856775 +---------------------- -------------------------------------------------------- +Time: 0.560s Load: 0.004s, Pack+Encode: 0.256s, Decode+Unpack: 0.300s +---------------------- -------------------------------------------------------- +💾 Converting with 253.8568 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-215.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-215.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,912B, BPFP=0.9840 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,780B, BPFP=3.1611 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,772B, BPFP=1.7572 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,372B, BPFP=3.0793 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,340B, BPFP=2.0713 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,860B, BPFP=2.9768 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,740B, BPFP=1.9511 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,032B, BPFP=3.0112 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,176B, BPFP=2.6394 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,844B, BPFP=2.9736 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,236B, BPFP=0.9797 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12096892 30.20650541 + layer.0.v_cache 0.00001393 0.00568664 + layer.1.k_cache 0.03247837 3.75600492 + layer.1.v_cache 0.00000613 0.00236602 + layer.2.k_cache 0.00394396 0.53418732 + layer.2.v_cache 0.00001920 0.00609269 + layer.3.k_cache 0.02751121 2.45112023 + layer.3.v_cache 0.00001919 0.00670735 + layer.4.k_cache 0.00062481 0.11943796 + layer.4.v_cache 0.00005293 0.01277504 + layer.4.output 0.18564256 688.18875916 + ------------------------------------------------------------------------------------- + TOTAL 0.08736098 285.55424692 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 157064 +BPFP 1.8508 bits/point +EBPFP 3.7015 equivalent bits/point +MSE 285.554247 +---------------------- -------------------------------------------------------- +Time: 0.564s Load: 0.004s, Pack+Encode: 0.251s, Decode+Unpack: 0.309s +---------------------- -------------------------------------------------------- +💾 Converting with 285.5542 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-222.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-222.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,064B, BPFP=1.0276 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,796B, BPFP=3.2054 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,788B, BPFP=1.7833 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,172B, BPFP=3.0787 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,232B, BPFP=2.0763 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,624B, BPFP=2.9675 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,596B, BPFP=1.9472 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,036B, BPFP=3.0511 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,076B, BPFP=2.6534 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,852B, BPFP=3.0138 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,072B, BPFP=1.0167 +⌛️ [2/4] FRONTEND: Frontend time: 0.242s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.295s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08564486 30.36465097 + layer.0.v_cache 0.00001408 0.00617053 + layer.1.k_cache 0.03507188 3.59974651 + layer.1.v_cache 0.00000566 0.00254615 + layer.2.k_cache 0.00239554 0.53968018 + layer.2.v_cache 0.00001938 0.00677772 + layer.3.k_cache 0.03313774 2.27520435 + layer.3.v_cache 0.00001993 0.00807985 + layer.4.k_cache 0.00061714 0.13052195 + layer.4.v_cache 0.00005134 0.01473262 + layer.4.output 0.18226704 697.99466605 + ------------------------------------------------------------------------------------- + TOTAL 0.08428511 289.58298666 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 157308 +BPFP 1.8777 bits/point +EBPFP 3.7554 equivalent bits/point +MSE 289.582987 +---------------------- -------------------------------------------------------- +Time: 0.541s Load: 0.003s, Pack+Encode: 0.242s, Decode+Unpack: 0.295s +---------------------- -------------------------------------------------------- +💾 Converting with 289.5830 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-223.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 117, 128) +Output shape: (1, 117, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) -> torch.Size([1, 1, 117, 512]) + layer.4.output: torch.Size([1, 117, 3584]) -> torch.Size([1, 1, 117, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,176B, BPFP=0.8248 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,340B, BPFP=2.7163 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,512B, BPFP=1.5374 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,016B, BPFP=2.6731 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,336B, BPFP=1.7810 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,604B, BPFP=2.6181 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,608B, BPFP=1.6838 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,808B, BPFP=2.6453 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,324B, BPFP=2.3136 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,164B, BPFP=2.5593 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 45,732B, BPFP=0.8725 +⌛️ [2/4] FRONTEND: Frontend time: 0.240s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 117, 128]) + layer.0.v_cache: torch.Size([1, 4, 117, 128]) + layer.1.k_cache: torch.Size([1, 4, 117, 128]) + layer.1.v_cache: torch.Size([1, 4, 117, 128]) + layer.2.k_cache: torch.Size([1, 4, 117, 128]) + layer.2.v_cache: torch.Size([1, 4, 117, 128]) + layer.3.k_cache: torch.Size([1, 4, 117, 128]) + layer.3.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.k_cache: torch.Size([1, 4, 117, 128]) + layer.4.v_cache: torch.Size([1, 4, 117, 128]) + layer.4.output: torch.Size([1, 117, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11900647 26.31598474 + layer.0.v_cache 0.00001428 0.00568102 + layer.1.k_cache 0.12535267 3.51947465 + layer.1.v_cache 0.00000579 0.00243391 + layer.2.k_cache 0.00517510 0.42279764 + layer.2.v_cache 0.00001992 0.00654454 + layer.3.k_cache 0.09315677 2.09284491 + layer.3.v_cache 0.00001978 0.00696977 + layer.4.k_cache 0.00061896 0.12251511 + layer.4.v_cache 0.00005133 0.01283986 + layer.4.output 9.75369281 454.82001679 + ------------------------------------------------------------------------------------- + TOTAL 4.03642769 189.19107080 + (elements=1,018,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1018368 +Total Bytes 205620 +BPFP 1.6153 bits/point +EBPFP 3.2306 equivalent bits/point +MSE 189.191071 +---------------------- -------------------------------------------------------- +Time: 0.553s Load: 0.006s, Pack+Encode: 0.240s, Decode+Unpack: 0.307s +---------------------- -------------------------------------------------------- +💾 Converting with 189.1911 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-227.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-227.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 105, 128) +Output shape: (1, 105, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) -> torch.Size([1, 1, 105, 512]) + layer.4.output: torch.Size([1, 105, 3584]) -> torch.Size([1, 1, 105, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,016B, BPFP=0.8952 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,312B, BPFP=2.8738 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,972B, BPFP=1.6327 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,000B, BPFP=2.8274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,280B, BPFP=1.8274 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,484B, BPFP=2.7506 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,868B, BPFP=1.7661 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,664B, BPFP=2.7774 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,316B, BPFP=2.4280 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,260B, BPFP=2.7173 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,172B, BPFP=0.8540 +⌛️ [2/4] FRONTEND: Frontend time: 0.232s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.305s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 105, 128]) + layer.0.v_cache: torch.Size([1, 4, 105, 128]) + layer.1.k_cache: torch.Size([1, 4, 105, 128]) + layer.1.v_cache: torch.Size([1, 4, 105, 128]) + layer.2.k_cache: torch.Size([1, 4, 105, 128]) + layer.2.v_cache: torch.Size([1, 4, 105, 128]) + layer.3.k_cache: torch.Size([1, 4, 105, 128]) + layer.3.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.k_cache: torch.Size([1, 4, 105, 128]) + layer.4.v_cache: torch.Size([1, 4, 105, 128]) + layer.4.output: torch.Size([1, 105, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14131550 27.74842820 + layer.0.v_cache 0.00001383 0.00571626 + layer.1.k_cache 0.02978072 3.53312174 + layer.1.v_cache 0.00000537 0.00236903 + layer.2.k_cache 0.00578868 0.47719872 + layer.2.v_cache 0.00001872 0.00609623 + layer.3.k_cache 0.04020810 2.25719096 + layer.3.v_cache 0.00001850 0.00693315 + layer.4.k_cache 0.00063358 0.12528989 + layer.4.v_cache 0.00005294 0.01383501 + layer.4.output 10.86925024 505.00714286 + ------------------------------------------------------------------------------------- + TOTAL 4.48838751 209.95448113 + (elements=913,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 913920 +Total Bytes 191344 +BPFP 1.6749 bits/point +EBPFP 3.3499 equivalent bits/point +MSE 209.954481 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.004s, Pack+Encode: 0.232s, Decode+Unpack: 0.305s +---------------------- -------------------------------------------------------- +💾 Converting with 209.9545 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-233.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-233.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,172B, BPFP=0.9621 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,400B, BPFP=3.0506 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,240B, BPFP=1.7188 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,916B, BPFP=2.9606 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,456B, BPFP=1.9449 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,524B, BPFP=2.8876 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,160B, BPFP=1.8899 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,852B, BPFP=2.9487 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,536B, BPFP=2.5179 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,244B, BPFP=2.8356 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,796B, BPFP=0.9246 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.314s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09514355 27.22051711 + layer.0.v_cache 0.00001423 0.00552175 + layer.1.k_cache 0.03283414 3.39248003 + layer.1.v_cache 0.00000553 0.00230786 + layer.2.k_cache 0.00489674 0.48609007 + layer.2.v_cache 0.00001883 0.00615254 + layer.3.k_cache 0.07199075 2.08058893 + layer.3.v_cache 0.00001932 0.00690248 + layer.4.k_cache 0.00061592 0.12189120 + layer.4.v_cache 0.00006816 0.01313226 + layer.4.output 0.17253765 621.37372449 + ------------------------------------------------------------------------------------- + TOTAL 0.08313946 257.82068563 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 162296 +BPFP 1.7758 bits/point +EBPFP 3.5516 equivalent bits/point +MSE 257.820686 +---------------------- -------------------------------------------------------- +Time: 0.571s Load: 0.004s, Pack+Encode: 0.253s, Decode+Unpack: 0.314s +---------------------- -------------------------------------------------------- +💾 Converting with 257.8207 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-238.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-238.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,292B, BPFP=0.9504 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,580B, BPFP=2.9777 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,468B, BPFP=1.7004 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,212B, BPFP=2.9116 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,688B, BPFP=1.9195 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,692B, BPFP=2.8182 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,256B, BPFP=1.8420 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,052B, BPFP=2.8829 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,960B, BPFP=2.5072 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,680B, BPFP=2.8161 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,332B, BPFP=0.9065 +⌛️ [2/4] FRONTEND: Frontend time: 0.264s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.273s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11660385 26.91331605 + layer.0.v_cache 0.00001381 0.00576866 + layer.1.k_cache 0.05175667 3.29802485 + layer.1.v_cache 0.00000569 0.00256098 + layer.2.k_cache 0.00227456 0.50731159 + layer.2.v_cache 0.00001743 0.00631968 + layer.3.k_cache 0.02781237 2.06117441 + layer.3.v_cache 0.00001883 0.00694187 + layer.4.k_cache 0.00061846 0.12483848 + layer.4.v_cache 0.00005344 0.01363620 + layer.4.output 0.17410791 603.67138752 + ------------------------------------------------------------------------------------- + TOTAL 0.08340767 250.50821208 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 165212 +BPFP 1.7454 bits/point +EBPFP 3.4908 equivalent bits/point +MSE 250.508212 +---------------------- -------------------------------------------------------- +Time: 0.542s Load: 0.004s, Pack+Encode: 0.264s, Decode+Unpack: 0.273s +---------------------- -------------------------------------------------------- +💾 Converting with 250.5082 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-239.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,152B, BPFP=0.9699 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,188B, BPFP=3.0474 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,128B, BPFP=1.7184 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,596B, BPFP=2.9360 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,364B, BPFP=1.9511 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,104B, BPFP=2.8434 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,880B, BPFP=1.8599 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,388B, BPFP=2.8968 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,536B, BPFP=2.5482 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,904B, BPFP=2.8057 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,860B, BPFP=0.9375 +⌛️ [2/4] FRONTEND: Frontend time: 0.237s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.289s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08212080 25.85225374 + layer.0.v_cache 0.00001360 0.00525468 + layer.1.k_cache 0.03288655 3.69117976 + layer.1.v_cache 0.00000536 0.00233225 + layer.2.k_cache 0.00973156 0.45539061 + layer.2.v_cache 0.00001762 0.00601438 + layer.3.k_cache 0.02906220 2.22525521 + layer.3.v_cache 0.00001891 0.00654676 + layer.4.k_cache 0.00063934 0.11785114 + layer.4.v_cache 0.00004866 0.01333412 + layer.4.output 0.18425958 613.59552496 + ------------------------------------------------------------------------------------- + TOTAL 0.08496245 254.56141690 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 160100 +BPFP 1.7729 bits/point +EBPFP 3.5458 equivalent bits/point +MSE 254.561417 +---------------------- -------------------------------------------------------- +Time: 0.530s Load: 0.004s, Pack+Encode: 0.237s, Decode+Unpack: 0.289s +---------------------- -------------------------------------------------------- +💾 Converting with 254.5614 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-243.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-243.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,132B, BPFP=0.9434 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,424B, BPFP=3.0191 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,220B, BPFP=1.6949 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,132B, BPFP=2.9654 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,424B, BPFP=1.9162 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,552B, BPFP=2.8588 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,396B, BPFP=1.9110 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,848B, BPFP=2.9132 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,680B, BPFP=2.5147 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,400B, BPFP=2.8309 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,212B, BPFP=0.9509 +⌛️ [2/4] FRONTEND: Frontend time: 0.232s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.291s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10758395 27.66617647 + layer.0.v_cache 0.00001397 0.00566147 + layer.1.k_cache 0.03293324 3.91536434 + layer.1.v_cache 0.00000575 0.00235978 + layer.2.k_cache 0.00823421 0.45937087 + layer.2.v_cache 0.00001745 0.00586814 + layer.3.k_cache 0.04116619 2.08118322 + layer.3.v_cache 0.00001985 0.00725054 + layer.4.k_cache 0.00062779 0.11829946 + layer.4.v_cache 0.00005042 0.01343632 + layer.4.output 0.16954060 633.06386555 + ------------------------------------------------------------------------------------- + TOTAL 0.08102571 262.68953114 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 164420 +BPFP 1.7779 bits/point +EBPFP 3.5558 equivalent bits/point +MSE 262.689531 +---------------------- -------------------------------------------------------- +Time: 0.527s Load: 0.004s, Pack+Encode: 0.232s, Decode+Unpack: 0.291s +---------------------- -------------------------------------------------------- +💾 Converting with 262.6895 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-257.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-257.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,496B, BPFP=0.9542 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,012B, BPFP=2.9535 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,892B, BPFP=1.7174 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,608B, BPFP=2.8833 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,096B, BPFP=1.9264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,344B, BPFP=2.8375 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,652B, BPFP=1.8493 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,512B, BPFP=2.8667 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,492B, BPFP=2.5160 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,096B, BPFP=2.7944 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,304B, BPFP=0.9500 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.315s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12307767 27.88722059 + layer.0.v_cache 0.00001410 0.00564142 + layer.1.k_cache 0.03088827 3.67263862 + layer.1.v_cache 0.00000584 0.00253406 + layer.2.k_cache 0.00371391 0.47138210 + layer.2.v_cache 0.00001950 0.00681749 + layer.3.k_cache 0.03079298 2.34153358 + layer.3.v_cache 0.00001950 0.00688960 + layer.4.k_cache 0.00063014 0.12713275 + layer.4.v_cache 0.00005282 0.01352905 + layer.4.output 0.15767360 596.21939484 + ------------------------------------------------------------------------------------- + TOTAL 0.07605470 247.53359313 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 172504 +BPFP 1.7617 bits/point +EBPFP 3.5234 equivalent bits/point +MSE 247.533593 +---------------------- -------------------------------------------------------- +Time: 0.580s Load: 0.005s, Pack+Encode: 0.261s, Decode+Unpack: 0.315s +---------------------- -------------------------------------------------------- +💾 Converting with 247.5336 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-258.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-258.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,132B, BPFP=0.9779 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,092B, BPFP=3.0663 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,008B, BPFP=1.7165 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,728B, BPFP=2.9970 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,364B, BPFP=1.9748 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,204B, BPFP=2.8971 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,760B, BPFP=1.8598 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,328B, BPFP=2.9207 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,284B, BPFP=2.5312 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,020B, BPFP=2.8620 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,860B, BPFP=0.9217 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.295s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10295349 29.24692145 + layer.0.v_cache 0.00001348 0.00521815 + layer.1.k_cache 0.03334363 3.41325713 + layer.1.v_cache 0.00000550 0.00238883 + layer.2.k_cache 0.01126997 0.46162633 + layer.2.v_cache 0.00001933 0.00609387 + layer.3.k_cache 0.09131574 2.34594094 + layer.3.v_cache 0.00001769 0.00642972 + layer.4.k_cache 0.00062275 0.12526377 + layer.4.v_cache 0.00006541 0.01260797 + layer.4.output 0.17512874 650.99319469 + ------------------------------------------------------------------------------------- + TOTAL 0.08620754 270.15165359 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 158780 +BPFP 1.7797 bits/point +EBPFP 3.5595 equivalent bits/point +MSE 270.151654 +---------------------- -------------------------------------------------------- +Time: 0.556s Load: 0.004s, Pack+Encode: 0.257s, Decode+Unpack: 0.295s +---------------------- -------------------------------------------------------- +💾 Converting with 270.1517 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-259.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,252B, BPFP=0.9432 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,624B, BPFP=2.9856 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,312B, BPFP=1.6724 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,140B, BPFP=2.8987 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,556B, BPFP=1.8958 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,624B, BPFP=2.8060 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,084B, BPFP=1.8111 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,952B, BPFP=2.8649 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,872B, BPFP=2.4914 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,516B, BPFP=2.7866 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,532B, BPFP=0.8860 +⌛️ [2/4] FRONTEND: Frontend time: 0.252s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.299s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12314748 27.81768869 + layer.0.v_cache 0.00001636 0.00555690 + layer.1.k_cache 0.01190477 3.27966589 + layer.1.v_cache 0.00000588 0.00234405 + layer.2.k_cache 0.01614088 0.45729626 + layer.2.v_cache 0.00001834 0.00573201 + layer.3.k_cache 0.05604688 2.17665083 + layer.3.v_cache 0.00001882 0.00695250 + layer.4.k_cache 0.00063774 0.11793584 + layer.4.v_cache 0.00005162 0.01281359 + layer.4.output 0.17348555 606.96762110 + ------------------------------------------------------------------------------------- + TOTAL 0.08366986 251.92094025 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 163464 +BPFP 1.7269 bits/point +EBPFP 3.4539 equivalent bits/point +MSE 251.920940 +---------------------- -------------------------------------------------------- +Time: 0.555s Load: 0.004s, Pack+Encode: 0.252s, Decode+Unpack: 0.299s +---------------------- -------------------------------------------------------- +💾 Converting with 251.9209 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-265.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,080B, BPFP=0.9799 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,108B, BPFP=3.1073 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,092B, BPFP=1.7539 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,632B, BPFP=3.0154 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,332B, BPFP=1.9931 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,108B, BPFP=2.9144 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,860B, BPFP=1.9020 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,192B, BPFP=2.9306 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,360B, BPFP=2.5772 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,952B, BPFP=2.8843 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,844B, BPFP=0.9602 +⌛️ [2/4] FRONTEND: Frontend time: 0.265s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.318s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11560643 31.08781226 + layer.0.v_cache 0.00001402 0.00608448 + layer.1.k_cache 0.03272506 3.73233673 + layer.1.v_cache 0.00000563 0.00251355 + layer.2.k_cache 0.00645547 0.48492224 + layer.2.v_cache 0.00001830 0.00699840 + layer.3.k_cache 0.02731901 2.29838769 + layer.3.v_cache 0.00001960 0.00762327 + layer.4.k_cache 0.00061302 0.12953308 + layer.4.v_cache 0.00004978 0.01399436 + layer.4.output 0.17251868 658.56161817 + ------------------------------------------------------------------------------------- + TOTAL 0.08179159 273.39420784 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 159560 +BPFP 1.8105 bits/point +EBPFP 3.6211 equivalent bits/point +MSE 273.394208 +---------------------- -------------------------------------------------------- +Time: 0.587s Load: 0.004s, Pack+Encode: 0.265s, Decode+Unpack: 0.318s +---------------------- -------------------------------------------------------- +💾 Converting with 273.3942 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-267.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-267.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 88, 128) +Output shape: (1, 88, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) -> torch.Size([1, 1, 88, 512]) + layer.4.output: torch.Size([1, 88, 3584]) -> torch.Size([1, 1, 88, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,356B, BPFP=0.9510 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,696B, BPFP=2.9645 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,504B, BPFP=1.6875 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,296B, BPFP=2.8935 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,740B, BPFP=1.9070 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,848B, BPFP=2.8139 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,168B, BPFP=1.8054 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,020B, BPFP=2.8445 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,980B, BPFP=2.4822 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,544B, BPFP=2.7599 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,740B, BPFP=0.9319 +⌛️ [2/4] FRONTEND: Frontend time: 0.255s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.330s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 88, 128]) + layer.0.v_cache: torch.Size([1, 4, 88, 128]) + layer.1.k_cache: torch.Size([1, 4, 88, 128]) + layer.1.v_cache: torch.Size([1, 4, 88, 128]) + layer.2.k_cache: torch.Size([1, 4, 88, 128]) + layer.2.v_cache: torch.Size([1, 4, 88, 128]) + layer.3.k_cache: torch.Size([1, 4, 88, 128]) + layer.3.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.k_cache: torch.Size([1, 4, 88, 128]) + layer.4.v_cache: torch.Size([1, 4, 88, 128]) + layer.4.output: torch.Size([1, 88, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11690623 28.15015758 + layer.0.v_cache 0.00001418 0.00574604 + layer.1.k_cache 0.03305150 3.64088163 + layer.1.v_cache 0.00000545 0.00241127 + layer.2.k_cache 0.01095151 0.45437787 + layer.2.v_cache 0.00001854 0.00595097 + layer.3.k_cache 0.04293454 2.12154146 + layer.3.v_cache 0.00002007 0.00683876 + layer.4.k_cache 0.00062489 0.12357372 + layer.4.v_cache 0.00005451 0.01332678 + layer.4.output 0.16672201 602.99908685 + ------------------------------------------------------------------------------------- + TOTAL 0.08068444 250.32461259 + (elements=765,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 765952 +Total Bytes 166892 +BPFP 1.7431 bits/point +EBPFP 3.4862 equivalent bits/point +MSE 250.324613 +---------------------- -------------------------------------------------------- +Time: 0.588s Load: 0.004s, Pack+Encode: 0.255s, Decode+Unpack: 0.330s +---------------------- -------------------------------------------------------- +💾 Converting with 250.3246 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-268.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-268.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,568B, BPFP=0.9560 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,152B, BPFP=2.9451 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,968B, BPFP=1.7115 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,780B, BPFP=2.8812 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,232B, BPFP=1.9286 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,404B, BPFP=2.8166 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,628B, BPFP=1.8249 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,596B, BPFP=2.8496 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,596B, BPFP=2.5062 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,052B, BPFP=2.7562 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,192B, BPFP=0.8878 +⌛️ [2/4] FRONTEND: Frontend time: 0.269s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.288s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11067300 28.99409770 + layer.0.v_cache 0.00001409 0.00559583 + layer.1.k_cache 0.05152405 3.38550677 + layer.1.v_cache 0.00000569 0.00240660 + layer.2.k_cache 0.01362713 0.47498950 + layer.2.v_cache 0.00001889 0.00636754 + layer.3.k_cache 0.08227781 2.11904740 + layer.3.v_cache 0.00001889 0.00690641 + layer.4.k_cache 0.00062414 0.12389825 + layer.4.v_cache 0.00005257 0.01263577 + layer.4.output 0.15746453 577.65006868 + ------------------------------------------------------------------------------------- + TOTAL 0.08006400 239.92246662 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 171168 +BPFP 1.7288 bits/point +EBPFP 3.4577 equivalent bits/point +MSE 239.922467 +---------------------- -------------------------------------------------------- +Time: 0.562s Load: 0.005s, Pack+Encode: 0.269s, Decode+Unpack: 0.288s +---------------------- -------------------------------------------------------- +💾 Converting with 239.9225 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-28.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-28.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 116, 128) +Output shape: (1, 116, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) -> torch.Size([1, 1, 116, 512]) + layer.4.output: torch.Size([1, 116, 3584]) -> torch.Size([1, 1, 116, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,240B, BPFP=0.8405 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,228B, BPFP=2.7247 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,532B, BPFP=1.5533 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,948B, BPFP=2.6870 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,152B, BPFP=1.7716 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,544B, BPFP=2.6325 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,456B, BPFP=1.6778 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,668B, BPFP=2.6492 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,264B, BPFP=2.3254 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,068B, BPFP=2.5684 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 44,096B, BPFP=0.8485 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.293s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 116, 128]) + layer.0.v_cache: torch.Size([1, 4, 116, 128]) + layer.1.k_cache: torch.Size([1, 4, 116, 128]) + layer.1.v_cache: torch.Size([1, 4, 116, 128]) + layer.2.k_cache: torch.Size([1, 4, 116, 128]) + layer.2.v_cache: torch.Size([1, 4, 116, 128]) + layer.3.k_cache: torch.Size([1, 4, 116, 128]) + layer.3.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.k_cache: torch.Size([1, 4, 116, 128]) + layer.4.v_cache: torch.Size([1, 4, 116, 128]) + layer.4.output: torch.Size([1, 116, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16314862 25.57566676 + layer.0.v_cache 0.00001373 0.00555054 + layer.1.k_cache 0.05481785 3.23523212 + layer.1.v_cache 0.00000557 0.00247819 + layer.2.k_cache 0.00968372 0.46184408 + layer.2.v_cache 0.00001954 0.00648312 + layer.3.k_cache 0.02418142 1.99296202 + layer.3.v_cache 0.00001860 0.00677511 + layer.4.k_cache 0.00062333 0.12744268 + layer.4.v_cache 0.00005467 0.01434267 + layer.4.output 9.83417319 435.98972445 + ------------------------------------------------------------------------------------- + TOTAL 4.06422232 181.37393226 + (elements=1,009,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1009664 +Total Bytes 203196 +BPFP 1.6100 bits/point +EBPFP 3.2200 equivalent bits/point +MSE 181.373932 +---------------------- -------------------------------------------------------- +Time: 0.547s Load: 0.005s, Pack+Encode: 0.249s, Decode+Unpack: 0.293s +---------------------- -------------------------------------------------------- +💾 Converting with 181.3739 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-281.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-281.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,964B, BPFP=0.9944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,900B, BPFP=3.1851 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,764B, BPFP=1.7556 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,344B, BPFP=3.0737 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,996B, BPFP=2.0024 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,624B, BPFP=2.9295 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,388B, BPFP=1.8806 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,860B, BPFP=2.9768 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,996B, BPFP=2.6034 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,696B, BPFP=2.9439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,560B, BPFP=0.9604 +⌛️ [2/4] FRONTEND: Frontend time: 0.239s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.289s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11749315 31.12202649 + layer.0.v_cache 0.00001389 0.00557616 + layer.1.k_cache 0.03399578 3.33891805 + layer.1.v_cache 0.00000555 0.00251016 + layer.2.k_cache 0.00854440 0.49696942 + layer.2.v_cache 0.00001777 0.00589080 + layer.3.k_cache 0.02740619 2.68166703 + layer.3.v_cache 0.00001935 0.00686011 + layer.4.k_cache 0.00063531 0.12406151 + layer.4.v_cache 0.00005166 0.01353232 + layer.4.output 0.18490290 688.96474359 + ------------------------------------------------------------------------------------- + TOTAL 0.08720608 285.91477748 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 155092 +BPFP 1.8275 bits/point +EBPFP 3.6551 equivalent bits/point +MSE 285.914777 +---------------------- -------------------------------------------------------- +Time: 0.532s Load: 0.004s, Pack+Encode: 0.239s, Decode+Unpack: 0.289s +---------------------- -------------------------------------------------------- +💾 Converting with 285.9148 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-284.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-284.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 94, 128) +Output shape: (1, 94, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) -> torch.Size([1, 1, 94, 512]) + layer.4.output: torch.Size([1, 94, 3584]) -> torch.Size([1, 1, 94, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,544B, BPFP=0.9215 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,564B, BPFP=2.9195 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,300B, BPFP=1.7121 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,060B, BPFP=2.8358 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,280B, BPFP=1.8750 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,684B, BPFP=2.7733 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,828B, BPFP=1.7999 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,940B, BPFP=2.8158 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,884B, BPFP=2.4741 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,372B, BPFP=2.7214 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,728B, BPFP=0.8722 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.293s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 94, 128]) + layer.0.v_cache: torch.Size([1, 4, 94, 128]) + layer.1.k_cache: torch.Size([1, 4, 94, 128]) + layer.1.v_cache: torch.Size([1, 4, 94, 128]) + layer.2.k_cache: torch.Size([1, 4, 94, 128]) + layer.2.v_cache: torch.Size([1, 4, 94, 128]) + layer.3.k_cache: torch.Size([1, 4, 94, 128]) + layer.3.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.k_cache: torch.Size([1, 4, 94, 128]) + layer.4.v_cache: torch.Size([1, 4, 94, 128]) + layer.4.output: torch.Size([1, 94, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11492289 30.26770539 + layer.0.v_cache 0.00001440 0.00560856 + layer.1.k_cache 0.01112092 3.42044295 + layer.1.v_cache 0.00000555 0.00236336 + layer.2.k_cache 0.01096075 0.50125240 + layer.2.v_cache 0.00001882 0.00623968 + layer.3.k_cache 0.03781782 2.40343751 + layer.3.v_cache 0.00001905 0.00657905 + layer.4.k_cache 0.00062716 0.12135194 + layer.4.v_cache 0.00006343 0.01278359 + layer.4.output 0.15224552 570.83358663 + ------------------------------------------------------------------------------------- + TOTAL 0.07301703 237.21075711 + (elements=818,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 818176 +Total Bytes 174184 +BPFP 1.7031 bits/point +EBPFP 3.4063 equivalent bits/point +MSE 237.210757 +---------------------- -------------------------------------------------------- +Time: 0.557s Load: 0.005s, Pack+Encode: 0.259s, Decode+Unpack: 0.293s +---------------------- -------------------------------------------------------- +💾 Converting with 237.2108 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-286.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-286.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,000B, BPFP=1.0146 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,536B, BPFP=3.1526 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,816B, BPFP=1.7890 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,160B, BPFP=3.0763 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,300B, BPFP=2.0901 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,656B, BPFP=2.9740 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,644B, BPFP=1.9570 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,772B, BPFP=2.9976 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,132B, BPFP=2.6648 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,708B, BPFP=2.9846 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,536B, BPFP=0.9722 +⌛️ [2/4] FRONTEND: Frontend time: 0.266s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10405910 31.26705180 + layer.0.v_cache 0.00001453 0.00564752 + layer.1.k_cache 0.03397077 3.70346862 + layer.1.v_cache 0.00000547 0.00240234 + layer.2.k_cache 0.00568739 0.56528349 + layer.2.v_cache 0.00001916 0.00643156 + layer.3.k_cache 0.09581309 2.34545304 + layer.3.v_cache 0.00001862 0.00742194 + layer.4.k_cache 0.00061896 0.12751770 + layer.4.v_cache 0.00005653 0.01364468 + layer.4.output 0.19123089 697.97773655 + ------------------------------------------------------------------------------------- + TOTAL 0.09287528 289.64049874 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 155260 +BPFP 1.8533 bits/point +EBPFP 3.7066 equivalent bits/point +MSE 289.640499 +---------------------- -------------------------------------------------------- +Time: 0.579s Load: 0.003s, Pack+Encode: 0.266s, Decode+Unpack: 0.309s +---------------------- -------------------------------------------------------- +💾 Converting with 289.6405 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-288.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,180B, BPFP=0.9522 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,244B, BPFP=2.9860 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,280B, BPFP=1.7059 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,860B, BPFP=2.9154 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,472B, BPFP=1.9250 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,428B, BPFP=2.8360 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,928B, BPFP=1.8250 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,660B, BPFP=2.8787 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,588B, BPFP=2.4978 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,328B, BPFP=2.8176 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,940B, BPFP=0.9438 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.304s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09384689 29.69870749 + layer.0.v_cache 0.00001364 0.00578317 + layer.1.k_cache 0.01313462 3.52433579 + layer.1.v_cache 0.00000592 0.00280898 + layer.2.k_cache 0.00813576 0.49095917 + layer.2.v_cache 0.00001762 0.00661982 + layer.3.k_cache 0.02628646 2.26868071 + layer.3.v_cache 0.00002044 0.00761153 + layer.4.k_cache 0.00061596 0.13508295 + layer.4.v_cache 0.00005334 0.01386631 + layer.4.output 0.17343949 632.58655462 + ------------------------------------------------------------------------------------- + TOTAL 0.07977689 262.60354931 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 162908 +BPFP 1.7615 bits/point +EBPFP 3.5231 equivalent bits/point +MSE 262.603549 +---------------------- -------------------------------------------------------- +Time: 0.570s Load: 0.004s, Pack+Encode: 0.261s, Decode+Unpack: 0.304s +---------------------- -------------------------------------------------------- +💾 Converting with 262.6035 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-293.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-293.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 126, 128) +Output shape: (1, 126, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) -> torch.Size([1, 1, 126, 512]) + layer.4.output: torch.Size([1, 126, 3584]) -> torch.Size([1, 1, 126, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,628B, BPFP=0.8219 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 20,640B, BPFP=2.5595 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,968B, BPFP=1.4841 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 20,276B, BPFP=2.5144 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 13,548B, BPFP=1.6801 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 19,872B, BPFP=2.4643 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 12,948B, BPFP=1.6057 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,996B, BPFP=2.4797 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 17,580B, BPFP=2.1801 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 19,356B, BPFP=2.4003 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 48,860B, BPFP=0.8656 +⌛️ [2/4] FRONTEND: Frontend time: 0.299s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.358s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 126, 128]) + layer.0.v_cache: torch.Size([1, 4, 126, 128]) + layer.1.k_cache: torch.Size([1, 4, 126, 128]) + layer.1.v_cache: torch.Size([1, 4, 126, 128]) + layer.2.k_cache: torch.Size([1, 4, 126, 128]) + layer.2.v_cache: torch.Size([1, 4, 126, 128]) + layer.3.k_cache: torch.Size([1, 4, 126, 128]) + layer.3.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.k_cache: torch.Size([1, 4, 126, 128]) + layer.4.v_cache: torch.Size([1, 4, 126, 128]) + layer.4.output: torch.Size([1, 126, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13141363 24.43224516 + layer.0.v_cache 0.00001429 0.00565081 + layer.1.k_cache 0.06496696 3.09724693 + layer.1.v_cache 0.00000583 0.00267292 + layer.2.k_cache 0.01057190 0.47044233 + layer.2.v_cache 0.00002020 0.00653437 + layer.3.k_cache 0.05250616 1.95030128 + layer.3.v_cache 0.00001853 0.00702403 + layer.4.k_cache 0.00062861 0.12836017 + layer.4.v_cache 0.00004930 0.01365882 + layer.4.output 9.05470577 386.08482143 + ------------------------------------------------------------------------------------- + TOTAL 3.74371387 160.74752275 + (elements=1,096,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1096704 +Total Bytes 211672 +BPFP 1.5441 bits/point +EBPFP 3.0881 equivalent bits/point +MSE 160.747523 +---------------------- -------------------------------------------------------- +Time: 0.663s Load: 0.006s, Pack+Encode: 0.299s, Decode+Unpack: 0.358s +---------------------- -------------------------------------------------------- +💾 Converting with 160.7475 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-302.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,288B, BPFP=0.9497 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,704B, BPFP=3.0000 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,368B, BPFP=1.6825 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,300B, BPFP=2.9274 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,608B, BPFP=1.9052 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,896B, BPFP=2.8549 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,216B, BPFP=1.8348 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,988B, BPFP=2.8714 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,020B, BPFP=2.5180 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,640B, BPFP=2.8089 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,788B, BPFP=0.8925 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.298s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13059476 26.70853538 + layer.0.v_cache 0.00001380 0.00541283 + layer.1.k_cache 0.03034586 3.26238628 + layer.1.v_cache 0.00000531 0.00233731 + layer.2.k_cache 0.00374181 0.45841243 + layer.2.v_cache 0.00001886 0.00618337 + layer.3.k_cache 0.04454902 2.18111955 + layer.3.v_cache 0.00001968 0.00685469 + layer.4.k_cache 0.00060938 0.12449267 + layer.4.v_cache 0.00007316 0.01328031 + layer.4.output 0.16485390 610.78438013 + ------------------------------------------------------------------------------------- + TOTAL 0.08023229 253.42703975 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 164816 +BPFP 1.7412 bits/point +EBPFP 3.4824 equivalent bits/point +MSE 253.427040 +---------------------- -------------------------------------------------------- +Time: 0.564s Load: 0.005s, Pack+Encode: 0.261s, Decode+Unpack: 0.298s +---------------------- -------------------------------------------------------- +💾 Converting with 253.4270 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-304.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-304.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,652B, BPFP=0.9599 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,320B, BPFP=2.9416 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,016B, BPFP=1.7011 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,956B, BPFP=2.8798 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,312B, BPFP=1.9212 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,604B, BPFP=2.8200 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,724B, BPFP=1.8213 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,840B, BPFP=2.8601 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,800B, BPFP=2.5136 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,380B, BPFP=2.7819 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,464B, BPFP=1.0060 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.333s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15436328 28.62221096 + layer.0.v_cache 0.00001586 0.00546805 + layer.1.k_cache 0.01080767 3.36353402 + layer.1.v_cache 0.00000575 0.00240623 + layer.2.k_cache 0.00786338 0.47687908 + layer.2.v_cache 0.00001961 0.00611629 + layer.3.k_cache 0.05652578 2.34109315 + layer.3.v_cache 0.00002042 0.00674065 + layer.4.k_cache 0.00062828 0.13026397 + layer.4.v_cache 0.00005674 0.01440739 + layer.4.output 0.15192759 572.08550078 + ------------------------------------------------------------------------------------- + TOTAL 0.07610588 237.62162501 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 178068 +BPFP 1.7790 bits/point +EBPFP 3.5579 equivalent bits/point +MSE 237.621625 +---------------------- -------------------------------------------------------- +Time: 0.596s Load: 0.005s, Pack+Encode: 0.258s, Decode+Unpack: 0.333s +---------------------- -------------------------------------------------------- +💾 Converting with 237.6216 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-309.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-309.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,952B, BPFP=0.9920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,828B, BPFP=3.1707 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,776B, BPFP=1.7580 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,220B, BPFP=3.0489 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,248B, BPFP=2.0529 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,700B, BPFP=2.9447 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,548B, BPFP=1.9127 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,992B, BPFP=3.0032 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,184B, BPFP=2.6410 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,840B, BPFP=2.9728 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 32,924B, BPFP=0.9422 +⌛️ [2/4] FRONTEND: Frontend time: 0.233s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.320s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10254851 30.84728065 + layer.0.v_cache 0.00001399 0.00577003 + layer.1.k_cache 0.05476840 3.55482561 + layer.1.v_cache 0.00000542 0.00238264 + layer.2.k_cache 0.00244555 0.48967987 + layer.2.v_cache 0.00001844 0.00657267 + layer.3.k_cache 0.02811479 2.38610996 + layer.3.v_cache 0.00001905 0.00731271 + layer.4.k_cache 0.00060111 0.12151288 + layer.4.v_cache 0.00005313 0.01403643 + layer.4.output 0.19635826 686.84117445 + ------------------------------------------------------------------------------------- + TOTAL 0.09194683 285.01904145 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 155212 +BPFP 1.8289 bits/point +EBPFP 3.6579 equivalent bits/point +MSE 285.019041 +---------------------- -------------------------------------------------------- +Time: 0.558s Load: 0.004s, Pack+Encode: 0.233s, Decode+Unpack: 0.320s +---------------------- -------------------------------------------------------- +💾 Converting with 285.0190 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-316.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-316.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,068B, BPFP=0.9657 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,140B, BPFP=3.0755 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,036B, BPFP=1.7218 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,724B, BPFP=2.9962 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,268B, BPFP=1.9566 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,188B, BPFP=2.8941 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,820B, BPFP=1.8712 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,232B, BPFP=2.9024 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,464B, BPFP=2.5655 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,976B, BPFP=2.8537 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,508B, BPFP=0.9394 +⌛️ [2/4] FRONTEND: Frontend time: 0.245s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.278s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11987135 28.92491485 + layer.0.v_cache 0.00001390 0.00530234 + layer.1.k_cache 0.03139273 3.52757561 + layer.1.v_cache 0.00000562 0.00232227 + layer.2.k_cache 0.00816743 0.50515468 + layer.2.v_cache 0.00001857 0.00599159 + layer.3.k_cache 0.02640573 2.21854270 + layer.3.v_cache 0.00001968 0.00641697 + layer.4.k_cache 0.00062091 0.11983730 + layer.4.v_cache 0.00004972 0.01221377 + layer.4.output 0.16983549 651.25337544 + ------------------------------------------------------------------------------------- + TOTAL 0.08090671 270.24128824 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 159424 +BPFP 1.7869 bits/point +EBPFP 3.5739 equivalent bits/point +MSE 270.241288 +---------------------- -------------------------------------------------------- +Time: 0.527s Load: 0.004s, Pack+Encode: 0.245s, Decode+Unpack: 0.278s +---------------------- -------------------------------------------------------- +💾 Converting with 270.2413 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-320.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,280B, BPFP=0.9483 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,632B, BPFP=2.9871 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,496B, BPFP=1.7055 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,340B, BPFP=2.9346 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,932B, BPFP=1.9634 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,868B, BPFP=2.8499 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,392B, BPFP=1.8664 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,116B, BPFP=2.8944 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,020B, BPFP=2.5180 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,636B, BPFP=2.8082 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,400B, BPFP=0.9083 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14089697 28.44008733 + layer.0.v_cache 0.00001393 0.00550125 + layer.1.k_cache 0.01618495 3.50450959 + layer.1.v_cache 0.00000582 0.00228360 + layer.2.k_cache 0.00661465 0.50636647 + layer.2.v_cache 0.00002083 0.00588489 + layer.3.k_cache 0.07087009 2.23820776 + layer.3.v_cache 0.00001822 0.00682874 + layer.4.k_cache 0.00061891 0.12585390 + layer.4.v_cache 0.00005311 0.01264399 + layer.4.output 0.16425455 599.49096880 + ------------------------------------------------------------------------------------- + TOTAL 0.08147525 248.89911465 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 166112 +BPFP 1.7549 bits/point +EBPFP 3.5098 equivalent bits/point +MSE 248.899115 +---------------------- -------------------------------------------------------- +Time: 0.578s Load: 0.004s, Pack+Encode: 0.253s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 248.8991 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-344.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-344.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 89, 128) +Output shape: (1, 89, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) -> torch.Size([1, 1, 89, 512]) + layer.4.output: torch.Size([1, 89, 3584]) -> torch.Size([1, 1, 89, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,416B, BPFP=0.9508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,924B, BPFP=2.9712 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,712B, BPFP=1.7051 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,608B, BPFP=2.9157 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,908B, BPFP=1.9150 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,176B, BPFP=2.8399 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,388B, BPFP=1.8237 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,452B, BPFP=2.8883 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,308B, BPFP=2.5119 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,964B, BPFP=2.8027 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,016B, BPFP=0.9535 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.298s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 89, 128]) + layer.0.v_cache: torch.Size([1, 4, 89, 128]) + layer.1.k_cache: torch.Size([1, 4, 89, 128]) + layer.1.v_cache: torch.Size([1, 4, 89, 128]) + layer.2.k_cache: torch.Size([1, 4, 89, 128]) + layer.2.v_cache: torch.Size([1, 4, 89, 128]) + layer.3.k_cache: torch.Size([1, 4, 89, 128]) + layer.3.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.k_cache: torch.Size([1, 4, 89, 128]) + layer.4.v_cache: torch.Size([1, 4, 89, 128]) + layer.4.output: torch.Size([1, 89, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12786581 28.96567218 + layer.0.v_cache 0.00001426 0.00604922 + layer.1.k_cache 0.03183105 3.75401049 + layer.1.v_cache 0.00000593 0.00269475 + layer.2.k_cache 0.00241907 0.49686115 + layer.2.v_cache 0.00001785 0.00610166 + layer.3.k_cache 0.02946699 2.18395722 + layer.3.v_cache 0.00001957 0.00755783 + layer.4.k_cache 0.00063747 0.13045317 + layer.4.v_cache 0.00005003 0.01407669 + layer.4.output 0.16568538 603.15218700 + ------------------------------------------------------------------------------------- + TOTAL 0.07953681 250.44898490 + (elements=774,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 774656 +Total Bytes 170872 +BPFP 1.7646 bits/point +EBPFP 3.5292 equivalent bits/point +MSE 250.448985 +---------------------- -------------------------------------------------------- +Time: 0.564s Load: 0.005s, Pack+Encode: 0.261s, Decode+Unpack: 0.298s +---------------------- -------------------------------------------------------- +💾 Converting with 250.4490 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-347.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-347.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,012B, BPFP=1.0040 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,912B, BPFP=3.1875 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,752B, BPFP=1.7532 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,540B, BPFP=3.1130 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,316B, BPFP=2.0665 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,076B, BPFP=3.0200 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,704B, BPFP=1.9439 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,184B, BPFP=3.0417 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,132B, BPFP=2.6306 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,776B, BPFP=2.9599 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,400B, BPFP=1.0130 +⌛️ [2/4] FRONTEND: Frontend time: 0.259s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.305s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12225728 30.12914413 + layer.0.v_cache 0.00001337 0.00553094 + layer.1.k_cache 0.03480693 3.56750410 + layer.1.v_cache 0.00000545 0.00248260 + layer.2.k_cache 0.00853881 0.49792148 + layer.2.v_cache 0.00001924 0.00648155 + layer.3.k_cache 0.06130744 2.40870412 + layer.3.v_cache 0.00001928 0.00728348 + layer.4.k_cache 0.00058936 0.11874270 + layer.4.v_cache 0.00007197 0.01381170 + layer.4.output 0.18657821 687.75795559 + ------------------------------------------------------------------------------------- + TOTAL 0.09021627 285.35666447 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 158804 +BPFP 1.8713 bits/point +EBPFP 3.7426 equivalent bits/point +MSE 285.356664 +---------------------- -------------------------------------------------------- +Time: 0.567s Load: 0.003s, Pack+Encode: 0.259s, Decode+Unpack: 0.305s +---------------------- -------------------------------------------------------- +💾 Converting with 285.3567 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-355.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-355.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 77, 128) +Output shape: (1, 77, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) -> torch.Size([1, 1, 77, 512]) + layer.4.output: torch.Size([1, 77, 3584]) -> torch.Size([1, 1, 77, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,016B, BPFP=1.0179 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,588B, BPFP=3.1631 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,704B, BPFP=1.7662 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,176B, BPFP=3.0795 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,088B, BPFP=2.0471 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,656B, BPFP=2.9740 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,408B, BPFP=1.9091 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,888B, BPFP=3.0211 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,152B, BPFP=2.6688 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,700B, BPFP=2.9830 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,520B, BPFP=0.9717 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.312s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 77, 128]) + layer.0.v_cache: torch.Size([1, 4, 77, 128]) + layer.1.k_cache: torch.Size([1, 4, 77, 128]) + layer.1.v_cache: torch.Size([1, 4, 77, 128]) + layer.2.k_cache: torch.Size([1, 4, 77, 128]) + layer.2.v_cache: torch.Size([1, 4, 77, 128]) + layer.3.k_cache: torch.Size([1, 4, 77, 128]) + layer.3.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.k_cache: torch.Size([1, 4, 77, 128]) + layer.4.v_cache: torch.Size([1, 4, 77, 128]) + layer.4.output: torch.Size([1, 77, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11582910 31.96219308 + layer.0.v_cache 0.00001386 0.00597820 + layer.1.k_cache 0.03408930 3.31529117 + layer.1.v_cache 0.00000538 0.00251506 + layer.2.k_cache 0.00228146 0.51811298 + layer.2.v_cache 0.00001859 0.00666946 + layer.3.k_cache 0.06276476 2.53113962 + layer.3.v_cache 0.00001913 0.00796950 + layer.4.k_cache 0.00060520 0.13530200 + layer.4.v_cache 0.00005125 0.01519194 + layer.4.output 0.19356344 697.04417904 + ------------------------------------------------------------------------------------- + TOTAL 0.09238954 289.28291860 + (elements=670,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 670208 +Total Bytes 154896 +BPFP 1.8489 bits/point +EBPFP 3.6979 equivalent bits/point +MSE 289.282919 +---------------------- -------------------------------------------------------- +Time: 0.573s Load: 0.004s, Pack+Encode: 0.257s, Decode+Unpack: 0.312s +---------------------- -------------------------------------------------------- +💾 Converting with 289.2829 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-361.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-361.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,152B, BPFP=0.9699 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,300B, BPFP=3.0685 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,112B, BPFP=1.7154 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,852B, BPFP=2.9842 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,484B, BPFP=1.9736 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,472B, BPFP=2.9127 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,020B, BPFP=1.8863 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,828B, BPFP=2.9797 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,468B, BPFP=2.5354 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,136B, BPFP=2.8494 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,444B, BPFP=0.9801 +⌛️ [2/4] FRONTEND: Frontend time: 0.262s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.322s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09700135 25.96454666 + layer.0.v_cache 0.00001355 0.00545157 + layer.1.k_cache 0.01103399 3.31568633 + layer.1.v_cache 0.00000598 0.00248144 + layer.2.k_cache 0.00794396 0.48205125 + layer.2.v_cache 0.00001935 0.00688284 + layer.3.k_cache 0.04207947 2.06233307 + layer.3.v_cache 0.00002010 0.00753403 + layer.4.k_cache 0.00062246 0.12116536 + layer.4.v_cache 0.00005226 0.01419647 + layer.4.output 0.17535226 631.44583692 + ------------------------------------------------------------------------------------- + TOTAL 0.08154460 261.88842279 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 163268 +BPFP 1.8080 bits/point +EBPFP 3.6160 equivalent bits/point +MSE 261.888423 +---------------------- -------------------------------------------------------- +Time: 0.589s Load: 0.005s, Pack+Encode: 0.262s, Decode+Unpack: 0.322s +---------------------- -------------------------------------------------------- +💾 Converting with 261.8884 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-364.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-364.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,152B, BPFP=0.9471 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,368B, BPFP=3.0088 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,148B, BPFP=1.6816 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,860B, BPFP=2.9154 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,500B, BPFP=1.9301 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,428B, BPFP=2.8360 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,084B, BPFP=1.8537 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,920B, BPFP=2.9265 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,652B, BPFP=2.5096 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,352B, BPFP=2.8221 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,344B, BPFP=0.9282 +⌛️ [2/4] FRONTEND: Frontend time: 0.263s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.309s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11944316 28.42616039 + layer.0.v_cache 0.00001525 0.00532126 + layer.1.k_cache 0.01376067 3.64037799 + layer.1.v_cache 0.00000561 0.00241937 + layer.2.k_cache 0.00522648 0.44221263 + layer.2.v_cache 0.00001966 0.00613568 + layer.3.k_cache 0.04600189 2.15435270 + layer.3.v_cache 0.00001912 0.00703840 + layer.4.k_cache 0.00062079 0.11850640 + layer.4.v_cache 0.00005114 0.01296800 + layer.4.output 0.17175110 632.61129202 + ------------------------------------------------------------------------------------- + TOTAL 0.08161303 262.53497276 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 162808 +BPFP 1.7605 bits/point +EBPFP 3.5209 equivalent bits/point +MSE 262.534973 +---------------------- -------------------------------------------------------- +Time: 0.575s Load: 0.003s, Pack+Encode: 0.263s, Decode+Unpack: 0.309s +---------------------- -------------------------------------------------------- +💾 Converting with 262.5350 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-365.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-365.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,964B, BPFP=0.9944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,672B, BPFP=3.1394 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,788B, BPFP=1.7604 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,220B, BPFP=3.0489 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,116B, BPFP=2.0264 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,660B, BPFP=2.9367 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,540B, BPFP=1.9111 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,908B, BPFP=2.9864 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,140B, BPFP=2.6322 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,640B, BPFP=2.9327 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,992B, BPFP=0.9728 +⌛️ [2/4] FRONTEND: Frontend time: 0.246s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.299s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11120220 30.26096755 + layer.0.v_cache 0.00001368 0.00559883 + layer.1.k_cache 0.03273370 3.41626407 + layer.1.v_cache 0.00000573 0.00249684 + layer.2.k_cache 0.00892877 0.50709759 + layer.2.v_cache 0.00001953 0.00613890 + layer.3.k_cache 0.02974504 2.23998652 + layer.3.v_cache 0.00001879 0.00715063 + layer.4.k_cache 0.00063874 0.13017355 + layer.4.v_cache 0.00005200 0.01408729 + layer.4.output 0.17853031 686.09872940 + ------------------------------------------------------------------------------------- + TOTAL 0.08429825 284.66359221 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 155640 +BPFP 1.8340 bits/point +EBPFP 3.6680 equivalent bits/point +MSE 284.663592 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.004s, Pack+Encode: 0.246s, Decode+Unpack: 0.299s +---------------------- -------------------------------------------------------- +💾 Converting with 284.6636 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-381.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-381.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,192B, BPFP=0.9774 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,400B, BPFP=3.0873 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,288B, BPFP=1.7485 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,996B, BPFP=3.0113 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,484B, BPFP=1.9736 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,328B, BPFP=2.8855 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,036B, BPFP=1.8893 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,596B, BPFP=2.9360 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,628B, BPFP=2.5655 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,364B, BPFP=2.8923 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,132B, BPFP=0.9986 +⌛️ [2/4] FRONTEND: Frontend time: 0.257s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.276s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09802409 26.70494576 + layer.0.v_cache 0.00001402 0.00570705 + layer.1.k_cache 0.03453328 3.28891947 + layer.1.v_cache 0.00000594 0.00236545 + layer.2.k_cache 0.01027090 0.46541301 + layer.2.v_cache 0.00001933 0.00572188 + layer.3.k_cache 0.02613793 2.20473324 + layer.3.v_cache 0.00001931 0.00691470 + layer.4.k_cache 0.00063818 0.12246433 + layer.4.v_cache 0.00005393 0.01311849 + layer.4.output 0.18084440 605.13758606 + ------------------------------------------------------------------------------------- + TOTAL 0.08444869 251.10490622 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 164444 +BPFP 1.8210 bits/point +EBPFP 3.6420 equivalent bits/point +MSE 251.104906 +---------------------- -------------------------------------------------------- +Time: 0.537s Load: 0.004s, Pack+Encode: 0.257s, Decode+Unpack: 0.276s +---------------------- -------------------------------------------------------- +💾 Converting with 251.1049 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-385.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-385.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,044B, BPFP=0.9169 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,148B, BPFP=2.9047 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,880B, BPFP=1.6505 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,840B, BPFP=2.8580 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,124B, BPFP=1.8392 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,280B, BPFP=2.7731 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,592B, BPFP=1.7585 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,496B, BPFP=2.8058 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,116B, BPFP=2.4448 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,916B, BPFP=2.7178 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 43,048B, BPFP=0.9329 +⌛️ [2/4] FRONTEND: Frontend time: 0.236s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.296s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13030120 26.92501327 + layer.0.v_cache 0.00001385 0.00549134 + layer.1.k_cache 0.02890817 3.49853130 + layer.1.v_cache 0.00000611 0.00266280 + layer.2.k_cache 0.00721959 0.47658976 + layer.2.v_cache 0.00001830 0.00615714 + layer.3.k_cache 0.02195997 2.37715697 + layer.3.v_cache 0.00001903 0.00707770 + layer.4.k_cache 0.00063766 0.13032646 + layer.4.v_cache 0.00005061 0.01360046 + layer.4.output 11.07892277 508.43593967 + ------------------------------------------------------------------------------------- + TOTAL 4.57303494 211.32318735 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 192484 +BPFP 1.7176 bits/point +EBPFP 3.4353 equivalent bits/point +MSE 211.323187 +---------------------- -------------------------------------------------------- +Time: 0.537s Load: 0.004s, Pack+Encode: 0.236s, Decode+Unpack: 0.296s +---------------------- -------------------------------------------------------- +💾 Converting with 211.3232 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-387.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,160B, BPFP=0.9714 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,256B, BPFP=3.0602 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,164B, BPFP=1.7252 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,916B, BPFP=2.9962 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,352B, BPFP=1.9488 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,528B, BPFP=2.9232 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,000B, BPFP=1.8825 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,668B, BPFP=2.9495 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,548B, BPFP=2.5505 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,344B, BPFP=2.8886 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,824B, BPFP=0.9634 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.304s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12633393 27.07440700 + layer.0.v_cache 0.00001372 0.00550658 + layer.1.k_cache 0.01411889 3.51406088 + layer.1.v_cache 0.00000549 0.00241702 + layer.2.k_cache 0.00833277 0.46505553 + layer.2.v_cache 0.00001851 0.00664188 + layer.3.k_cache 0.05797521 2.59489772 + layer.3.v_cache 0.00001879 0.00728931 + layer.4.k_cache 0.00064672 0.13136126 + layer.4.v_cache 0.00005349 0.01465301 + layer.4.output 0.17447430 618.80163511 + ------------------------------------------------------------------------------------- + TOTAL 0.08404927 256.78986682 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 162760 +BPFP 1.8024 bits/point +EBPFP 3.6047 equivalent bits/point +MSE 256.789867 +---------------------- -------------------------------------------------------- +Time: 0.567s Load: 0.004s, Pack+Encode: 0.258s, Decode+Unpack: 0.304s +---------------------- -------------------------------------------------------- +💾 Converting with 256.7899 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-394.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-394.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 70, 128) +Output shape: (1, 70, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) -> torch.Size([1, 1, 70, 512]) + layer.4.output: torch.Size([1, 70, 3584]) -> torch.Size([1, 1, 70, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,652B, BPFP=1.0384 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 14,676B, BPFP=3.2759 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,384B, BPFP=1.8714 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 13,916B, BPFP=3.1063 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 9,684B, BPFP=2.1616 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 13,528B, BPFP=3.0196 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 8,800B, BPFP=1.9643 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 13,708B, BPFP=3.0598 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,288B, BPFP=2.7429 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 13,852B, BPFP=3.0920 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 31,060B, BPFP=0.9904 +⌛️ [2/4] FRONTEND: Frontend time: 0.238s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.278s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 70, 128]) + layer.0.v_cache: torch.Size([1, 4, 70, 128]) + layer.1.k_cache: torch.Size([1, 4, 70, 128]) + layer.1.v_cache: torch.Size([1, 4, 70, 128]) + layer.2.k_cache: torch.Size([1, 4, 70, 128]) + layer.2.v_cache: torch.Size([1, 4, 70, 128]) + layer.3.k_cache: torch.Size([1, 4, 70, 128]) + layer.3.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.k_cache: torch.Size([1, 4, 70, 128]) + layer.4.v_cache: torch.Size([1, 4, 70, 128]) + layer.4.output: torch.Size([1, 70, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08086312 30.14971052 + layer.0.v_cache 0.00001331 0.00560878 + layer.1.k_cache 0.01551799 4.12443935 + layer.1.v_cache 0.00000513 0.00221936 + layer.2.k_cache 0.00232868 0.52442006 + layer.2.v_cache 0.00002262 0.00670209 + layer.3.k_cache 0.10623371 2.26078317 + layer.3.v_cache 0.00001836 0.00699705 + layer.4.k_cache 0.00061897 0.12807316 + layer.4.v_cache 0.00005922 0.01348443 + layer.4.output 0.20964142 768.45567602 + ------------------------------------------------------------------------------------- + TOTAL 0.09842183 318.61248059 + (elements=609,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 609280 +Total Bytes 144548 +BPFP 1.8980 bits/point +EBPFP 3.7959 equivalent bits/point +MSE 318.612481 +---------------------- -------------------------------------------------------- +Time: 0.519s Load: 0.003s, Pack+Encode: 0.238s, Decode+Unpack: 0.278s +---------------------- -------------------------------------------------------- +💾 Converting with 318.6125 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-397.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,720B, BPFP=0.9715 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,224B, BPFP=2.9253 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,228B, BPFP=1.7371 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,868B, BPFP=2.8648 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,180B, BPFP=1.8988 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,424B, BPFP=2.7894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,580B, BPFP=1.7969 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,544B, BPFP=2.8098 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,688B, BPFP=2.4946 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,132B, BPFP=2.7398 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,388B, BPFP=0.9314 +⌛️ [2/4] FRONTEND: Frontend time: 0.265s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.308s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13670269 29.05312712 + layer.0.v_cache 0.00001568 0.00563797 + layer.1.k_cache 0.03139175 3.15746672 + layer.1.v_cache 0.00000602 0.00260219 + layer.2.k_cache 0.00629326 0.47396299 + layer.2.v_cache 0.00001871 0.00593438 + layer.3.k_cache 0.03837063 2.38175168 + layer.3.v_cache 0.00001860 0.00642202 + layer.4.k_cache 0.00062597 0.12634858 + layer.4.v_cache 0.00005354 0.01331205 + layer.4.output 0.15974679 572.77154503 + ------------------------------------------------------------------------------------- + TOTAL 0.07833673 237.91925770 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 173976 +BPFP 1.7381 bits/point +EBPFP 3.4762 equivalent bits/point +MSE 237.919258 +---------------------- -------------------------------------------------------- +Time: 0.577s Load: 0.004s, Pack+Encode: 0.265s, Decode+Unpack: 0.308s +---------------------- -------------------------------------------------------- +💾 Converting with 237.9193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-402.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-402.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 107, 128) +Output shape: (1, 107, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) -> torch.Size([1, 1, 107, 512]) + layer.4.output: torch.Size([1, 107, 3584]) -> torch.Size([1, 1, 107, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,996B, BPFP=0.8756 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,656B, BPFP=2.8703 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,132B, BPFP=1.6256 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,292B, BPFP=2.8172 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,576B, BPFP=1.8364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,932B, BPFP=2.7646 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,856B, BPFP=1.7313 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,096B, BPFP=2.7886 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,556B, BPFP=2.4176 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,484B, BPFP=2.6992 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 41,980B, BPFP=0.8758 +⌛️ [2/4] FRONTEND: Frontend time: 0.260s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.300s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 107, 128]) + layer.0.v_cache: torch.Size([1, 4, 107, 128]) + layer.1.k_cache: torch.Size([1, 4, 107, 128]) + layer.1.v_cache: torch.Size([1, 4, 107, 128]) + layer.2.k_cache: torch.Size([1, 4, 107, 128]) + layer.2.v_cache: torch.Size([1, 4, 107, 128]) + layer.3.k_cache: torch.Size([1, 4, 107, 128]) + layer.3.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.k_cache: torch.Size([1, 4, 107, 128]) + layer.4.v_cache: torch.Size([1, 4, 107, 128]) + layer.4.output: torch.Size([1, 107, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12281527 26.93439690 + layer.0.v_cache 0.00001417 0.00556124 + layer.1.k_cache 0.04442539 3.45045428 + layer.1.v_cache 0.00000572 0.00247648 + layer.2.k_cache 0.01033308 0.50830435 + layer.2.v_cache 0.00001923 0.00649008 + layer.3.k_cache 0.04555132 2.10482246 + layer.3.v_cache 0.00001841 0.00673451 + layer.4.k_cache 0.00060580 0.12249950 + layer.4.v_cache 0.00007068 0.01356471 + layer.4.output 10.66341841 489.60388852 + ------------------------------------------------------------------------------------- + TOTAL 4.40398753 203.55191319 + (elements=931,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 931328 +Total Bytes 195556 +BPFP 1.6798 bits/point +EBPFP 3.3596 equivalent bits/point +MSE 203.551913 +---------------------- -------------------------------------------------------- +Time: 0.566s Load: 0.005s, Pack+Encode: 0.260s, Decode+Unpack: 0.300s +---------------------- -------------------------------------------------------- +💾 Converting with 203.5519 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-405.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-405.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 86, 128) +Output shape: (1, 86, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) -> torch.Size([1, 1, 86, 512]) + layer.4.output: torch.Size([1, 86, 3584]) -> torch.Size([1, 1, 86, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,188B, BPFP=0.9426 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,628B, BPFP=3.0211 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,380B, BPFP=1.7042 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,228B, BPFP=2.9484 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,800B, BPFP=1.9622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,856B, BPFP=2.8808 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,196B, BPFP=1.8525 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,088B, BPFP=2.9230 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,912B, BPFP=2.5276 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,632B, BPFP=2.8401 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,872B, BPFP=0.9051 +⌛️ [2/4] FRONTEND: Frontend time: 0.281s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.292s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 86, 128]) + layer.0.v_cache: torch.Size([1, 4, 86, 128]) + layer.1.k_cache: torch.Size([1, 4, 86, 128]) + layer.1.v_cache: torch.Size([1, 4, 86, 128]) + layer.2.k_cache: torch.Size([1, 4, 86, 128]) + layer.2.v_cache: torch.Size([1, 4, 86, 128]) + layer.3.k_cache: torch.Size([1, 4, 86, 128]) + layer.3.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.k_cache: torch.Size([1, 4, 86, 128]) + layer.4.v_cache: torch.Size([1, 4, 86, 128]) + layer.4.output: torch.Size([1, 86, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11362342 27.56680937 + layer.0.v_cache 0.00001418 0.00554585 + layer.1.k_cache 0.05429294 3.42395232 + layer.1.v_cache 0.00000580 0.00249555 + layer.2.k_cache 0.00513919 0.49288963 + layer.2.v_cache 0.00001816 0.00603479 + layer.3.k_cache 0.05885953 2.50914516 + layer.3.v_cache 0.00001841 0.00667871 + layer.4.k_cache 0.00061763 0.12399807 + layer.4.v_cache 0.00005080 0.01285528 + layer.4.output 0.16727475 625.59707226 + ------------------------------------------------------------------------------------- + TOTAL 0.08256255 259.60764180 + (elements=748,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 748544 +Total Bytes 164780 +BPFP 1.7611 bits/point +EBPFP 3.5221 equivalent bits/point +MSE 259.607642 +---------------------- -------------------------------------------------------- +Time: 0.577s Load: 0.004s, Pack+Encode: 0.281s, Decode+Unpack: 0.292s +---------------------- -------------------------------------------------------- +💾 Converting with 259.6076 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-406.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-406.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 93, 128) +Output shape: (1, 93, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) -> torch.Size([1, 1, 93, 512]) + layer.4.output: torch.Size([1, 93, 3584]) -> torch.Size([1, 1, 93, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,708B, BPFP=0.9590 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,460B, BPFP=2.9335 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,216B, BPFP=1.7164 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,092B, BPFP=2.8716 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,344B, BPFP=1.9059 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,740B, BPFP=2.8125 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,744B, BPFP=1.8051 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,772B, BPFP=2.8179 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,852B, BPFP=2.4953 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,384B, BPFP=2.7527 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,732B, BPFP=0.9296 +⌛️ [2/4] FRONTEND: Frontend time: 0.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.275s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 93, 128]) + layer.0.v_cache: torch.Size([1, 4, 93, 128]) + layer.1.k_cache: torch.Size([1, 4, 93, 128]) + layer.1.v_cache: torch.Size([1, 4, 93, 128]) + layer.2.k_cache: torch.Size([1, 4, 93, 128]) + layer.2.v_cache: torch.Size([1, 4, 93, 128]) + layer.3.k_cache: torch.Size([1, 4, 93, 128]) + layer.3.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.k_cache: torch.Size([1, 4, 93, 128]) + layer.4.v_cache: torch.Size([1, 4, 93, 128]) + layer.4.output: torch.Size([1, 93, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702597 29.05485026 + layer.0.v_cache 0.00001402 0.00570080 + layer.1.k_cache 0.01332537 3.46137984 + layer.1.v_cache 0.00000581 0.00248725 + layer.2.k_cache 0.00507878 0.50340517 + layer.2.v_cache 0.00001944 0.00605501 + layer.3.k_cache 0.04051446 2.21873950 + layer.3.v_cache 0.00001931 0.00662782 + layer.4.k_cache 0.00062870 0.12002104 + layer.4.v_cache 0.00006514 0.01292116 + layer.4.output 0.16549673 578.10363863 + ------------------------------------------------------------------------------------- + TOTAL 0.07795142 240.12456813 + (elements=809,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 809472 +Total Bytes 176044 +BPFP 1.7398 bits/point +EBPFP 3.4797 equivalent bits/point +MSE 240.124568 +---------------------- -------------------------------------------------------- +Time: 0.546s Load: 0.004s, Pack+Encode: 0.267s, Decode+Unpack: 0.275s +---------------------- -------------------------------------------------------- +💾 Converting with 240.1246 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-413.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-413.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 85, 128) +Output shape: (1, 85, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) -> torch.Size([1, 1, 85, 512]) + layer.4.output: torch.Size([1, 85, 3584]) -> torch.Size([1, 1, 85, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,140B, BPFP=0.9449 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,384B, BPFP=3.0118 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,192B, BPFP=1.6897 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,796B, BPFP=2.9037 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,384B, BPFP=1.9088 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,312B, BPFP=2.8147 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,920B, BPFP=1.8235 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,496B, BPFP=2.8485 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,592B, BPFP=2.4985 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,252B, BPFP=2.8037 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,232B, BPFP=0.9252 +⌛️ [2/4] FRONTEND: Frontend time: 0.275s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 85, 128]) + layer.0.v_cache: torch.Size([1, 4, 85, 128]) + layer.1.k_cache: torch.Size([1, 4, 85, 128]) + layer.1.v_cache: torch.Size([1, 4, 85, 128]) + layer.2.k_cache: torch.Size([1, 4, 85, 128]) + layer.2.v_cache: torch.Size([1, 4, 85, 128]) + layer.3.k_cache: torch.Size([1, 4, 85, 128]) + layer.3.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.k_cache: torch.Size([1, 4, 85, 128]) + layer.4.v_cache: torch.Size([1, 4, 85, 128]) + layer.4.output: torch.Size([1, 85, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13925676 29.77898093 + layer.0.v_cache 0.00001387 0.00568398 + layer.1.k_cache 0.03290265 3.44237707 + layer.1.v_cache 0.00000589 0.00239615 + layer.2.k_cache 0.00800494 0.48080570 + layer.2.v_cache 0.00001921 0.00588953 + layer.3.k_cache 0.08301921 2.15531652 + layer.3.v_cache 0.00001931 0.00658447 + layer.4.k_cache 0.00065893 0.11768399 + layer.4.v_cache 0.00004891 0.01256359 + layer.4.output 0.17367665 630.87310924 + ------------------------------------------------------------------------------------- + TOTAL 0.08704037 261.88941451 + (elements=739,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 739840 +Total Bytes 161700 +BPFP 1.7485 bits/point +EBPFP 3.4970 equivalent bits/point +MSE 261.889415 +---------------------- -------------------------------------------------------- +Time: 0.596s Load: 0.004s, Pack+Encode: 0.275s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 261.8894 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-414.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-414.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,028B, BPFP=0.9144 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,164B, BPFP=2.9072 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,724B, BPFP=1.6268 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,752B, BPFP=2.8447 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,992B, BPFP=1.8192 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,284B, BPFP=2.7737 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,572B, BPFP=1.7555 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,468B, BPFP=2.8016 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,180B, BPFP=2.4545 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,880B, BPFP=2.7124 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,608B, BPFP=0.8584 +⌛️ [2/4] FRONTEND: Frontend time: 0.242s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.298s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10583916 26.68262430 + layer.0.v_cache 0.00001406 0.00538052 + layer.1.k_cache 0.01351924 3.39227947 + layer.1.v_cache 0.00000536 0.00234807 + layer.2.k_cache 0.00567923 0.48175327 + layer.2.v_cache 0.00001856 0.00599345 + layer.3.k_cache 0.05293757 2.00838493 + layer.3.v_cache 0.00001841 0.00666208 + layer.4.k_cache 0.00064880 0.12682246 + layer.4.v_cache 0.00004836 0.01249636 + layer.4.output 11.07886250 509.48105929 + ------------------------------------------------------------------------------------- + TOTAL 4.57239802 211.71130353 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 188652 +BPFP 1.6834 bits/point +EBPFP 3.3669 equivalent bits/point +MSE 211.711304 +---------------------- -------------------------------------------------------- +Time: 0.545s Load: 0.005s, Pack+Encode: 0.242s, Decode+Unpack: 0.298s +---------------------- -------------------------------------------------------- +💾 Converting with 211.7113 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-422.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-422.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,156B, BPFP=0.9706 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,196B, BPFP=3.0489 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,236B, BPFP=1.7387 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,780B, BPFP=2.9706 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,420B, BPFP=1.9616 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,404B, BPFP=2.8998 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,032B, BPFP=1.8886 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,524B, BPFP=2.9224 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,528B, BPFP=2.5467 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,184B, BPFP=2.8584 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,600B, BPFP=0.9574 +⌛️ [2/4] FRONTEND: Frontend time: 0.236s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.294s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12860243 27.27780850 + layer.0.v_cache 0.00001403 0.00546331 + layer.1.k_cache 0.01400426 3.51965847 + layer.1.v_cache 0.00000584 0.00266272 + layer.2.k_cache 0.01277778 0.50401251 + layer.2.v_cache 0.00001980 0.00660061 + layer.3.k_cache 0.07376247 2.09098172 + layer.3.v_cache 0.00001877 0.00663300 + layer.4.k_cache 0.00062152 0.11777454 + layer.4.v_cache 0.00005140 0.01383358 + layer.4.output 0.18179212 626.73434811 + ------------------------------------------------------------------------------------- + TOTAL 0.08837783 260.04034504 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 162060 +BPFP 1.7946 bits/point +EBPFP 3.5892 equivalent bits/point +MSE 260.040345 +---------------------- -------------------------------------------------------- +Time: 0.534s Load: 0.003s, Pack+Encode: 0.236s, Decode+Unpack: 0.294s +---------------------- -------------------------------------------------------- +💾 Converting with 260.0403 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-425.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-425.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,448B, BPFP=0.9458 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,052B, BPFP=2.9604 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,892B, BPFP=1.7174 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,744B, BPFP=2.9069 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,136B, BPFP=1.9333 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,372B, BPFP=2.8424 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,416B, BPFP=1.8083 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,548B, BPFP=2.8729 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,628B, BPFP=2.5396 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,220B, BPFP=2.8160 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,956B, BPFP=0.9414 +⌛️ [2/4] FRONTEND: Frontend time: 0.264s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.263s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13781073 28.90565864 + layer.0.v_cache 0.00001414 0.00586406 + layer.1.k_cache 0.01252146 3.44430508 + layer.1.v_cache 0.00000561 0.00239098 + layer.2.k_cache 0.00238173 0.50191828 + layer.2.v_cache 0.00001849 0.00636828 + layer.3.k_cache 0.04360520 2.07033013 + layer.3.v_cache 0.00001973 0.00741645 + layer.4.k_cache 0.00061668 0.12202409 + layer.4.v_cache 0.00005448 0.01361553 + layer.4.output 0.16772948 589.95942460 + ------------------------------------------------------------------------------------- + TOTAL 0.08065615 244.98799199 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 172412 +BPFP 1.7607 bits/point +EBPFP 3.5215 equivalent bits/point +MSE 244.987992 +---------------------- -------------------------------------------------------- +Time: 0.531s Load: 0.004s, Pack+Encode: 0.264s, Decode+Unpack: 0.263s +---------------------- -------------------------------------------------------- +💾 Converting with 244.9880 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-428.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-428.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 79, 128) +Output shape: (1, 79, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) -> torch.Size([1, 1, 79, 512]) + layer.4.output: torch.Size([1, 79, 3584]) -> torch.Size([1, 1, 79, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,048B, BPFP=0.9984 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,068B, BPFP=3.1780 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,844B, BPFP=1.7492 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,432B, BPFP=3.0522 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,348B, BPFP=2.0467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,988B, BPFP=2.9644 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,716B, BPFP=1.9217 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,244B, BPFP=3.0150 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,268B, BPFP=2.6242 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,976B, BPFP=2.9620 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,276B, BPFP=0.9685 +⌛️ [2/4] FRONTEND: Frontend time: 0.258s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.294s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 79, 128]) + layer.0.v_cache: torch.Size([1, 4, 79, 128]) + layer.1.k_cache: torch.Size([1, 4, 79, 128]) + layer.1.v_cache: torch.Size([1, 4, 79, 128]) + layer.2.k_cache: torch.Size([1, 4, 79, 128]) + layer.2.v_cache: torch.Size([1, 4, 79, 128]) + layer.3.k_cache: torch.Size([1, 4, 79, 128]) + layer.3.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.k_cache: torch.Size([1, 4, 79, 128]) + layer.4.v_cache: torch.Size([1, 4, 79, 128]) + layer.4.output: torch.Size([1, 79, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.07550411 28.35881564 + layer.0.v_cache 0.00001381 0.00534173 + layer.1.k_cache 0.01271453 3.12530266 + layer.1.v_cache 0.00000531 0.00224955 + layer.2.k_cache 0.00681068 0.47530331 + layer.2.v_cache 0.00001904 0.00655172 + layer.3.k_cache 0.02844424 2.32786908 + layer.3.v_cache 0.00001798 0.00683743 + layer.4.k_cache 0.00061127 0.12246929 + layer.4.v_cache 0.00009323 0.01394592 + layer.4.output 0.18255964 663.10007911 + ------------------------------------------------------------------------------------- + TOTAL 0.08247951 275.06736707 + (elements=687,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 687616 +Total Bytes 158208 +BPFP 1.8407 bits/point +EBPFP 3.6813 equivalent bits/point +MSE 275.067367 +---------------------- -------------------------------------------------------- +Time: 0.555s Load: 0.004s, Pack+Encode: 0.258s, Decode+Unpack: 0.294s +---------------------- -------------------------------------------------------- +💾 Converting with 275.0674 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-435.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-435.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,028B, BPFP=0.9820 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,908B, BPFP=3.1070 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,912B, BPFP=1.7406 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,444B, BPFP=3.0164 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,464B, BPFP=2.0438 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,132B, BPFP=2.9555 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,748B, BPFP=1.9039 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,348B, BPFP=2.9977 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,268B, BPFP=2.5914 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,020B, BPFP=2.9336 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,956B, BPFP=0.9753 +⌛️ [2/4] FRONTEND: Frontend time: 0.229s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.286s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08590809 30.44415588 + layer.0.v_cache 0.00001388 0.00552935 + layer.1.k_cache 0.01594085 3.62343216 + layer.1.v_cache 0.00000552 0.00226381 + layer.2.k_cache 0.00791870 0.52525253 + layer.2.v_cache 0.00001801 0.00600532 + layer.3.k_cache 0.02757127 2.42340126 + layer.3.v_cache 0.00001862 0.00676015 + layer.4.k_cache 0.00063773 0.12156837 + layer.4.v_cache 0.00005492 0.01357573 + layer.4.output 0.18408036 642.18504464 + ------------------------------------------------------------------------------------- + TOTAL 0.08392059 266.61572100 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 159228 +BPFP 1.8294 bits/point +EBPFP 3.6587 equivalent bits/point +MSE 266.615721 +---------------------- -------------------------------------------------------- +Time: 0.520s Load: 0.005s, Pack+Encode: 0.229s, Decode+Unpack: 0.286s +---------------------- -------------------------------------------------------- +💾 Converting with 266.6157 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-438.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-438.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 76, 128) +Output shape: (1, 76, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) -> torch.Size([1, 1, 76, 512]) + layer.4.output: torch.Size([1, 76, 3584]) -> torch.Size([1, 1, 76, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,020B, BPFP=1.0321 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,496B, BPFP=3.1859 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,640B, BPFP=1.7763 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,016B, BPFP=3.0872 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,012B, BPFP=2.0584 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,480B, BPFP=2.9770 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,332B, BPFP=1.9186 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 14,840B, BPFP=3.0510 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 12,888B, BPFP=2.6497 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,484B, BPFP=2.9778 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,292B, BPFP=0.9778 +⌛️ [2/4] FRONTEND: Frontend time: 0.253s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.284s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 76, 128]) + layer.0.v_cache: torch.Size([1, 4, 76, 128]) + layer.1.k_cache: torch.Size([1, 4, 76, 128]) + layer.1.v_cache: torch.Size([1, 4, 76, 128]) + layer.2.k_cache: torch.Size([1, 4, 76, 128]) + layer.2.v_cache: torch.Size([1, 4, 76, 128]) + layer.3.k_cache: torch.Size([1, 4, 76, 128]) + layer.3.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.k_cache: torch.Size([1, 4, 76, 128]) + layer.4.v_cache: torch.Size([1, 4, 76, 128]) + layer.4.output: torch.Size([1, 76, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08991440 30.81649620 + layer.0.v_cache 0.00001347 0.00553690 + layer.1.k_cache 0.03345199 3.37942746 + layer.1.v_cache 0.00000518 0.00232145 + layer.2.k_cache 0.00398020 0.48098614 + layer.2.v_cache 0.00001815 0.00575872 + layer.3.k_cache 0.04607732 1.99122098 + layer.3.v_cache 0.00001896 0.00677193 + layer.4.k_cache 0.00062119 0.12588567 + layer.4.v_cache 0.00005040 0.01306239 + layer.4.output 0.18612516 691.15102209 + ------------------------------------------------------------------------------------- + TOTAL 0.08688396 286.75791897 + (elements=661,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 661504 +Total Bytes 153500 +BPFP 1.8564 bits/point +EBPFP 3.7128 equivalent bits/point +MSE 286.757919 +---------------------- -------------------------------------------------------- +Time: 0.541s Load: 0.004s, Pack+Encode: 0.253s, Decode+Unpack: 0.284s +---------------------- -------------------------------------------------------- +💾 Converting with 286.7579 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-439.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-439.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 91, 128) +Output shape: (1, 91, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) -> torch.Size([1, 1, 91, 512]) + layer.4.output: torch.Size([1, 91, 3584]) -> torch.Size([1, 1, 91, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,540B, BPFP=0.9512 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,244B, BPFP=2.9609 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,964B, BPFP=1.7109 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,792B, BPFP=2.8832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,276B, BPFP=1.9361 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,448B, BPFP=2.8242 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,636B, BPFP=1.8262 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,680B, BPFP=2.8640 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,764B, BPFP=2.5350 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,260B, BPFP=2.7919 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,696B, BPFP=0.9737 +⌛️ [2/4] FRONTEND: Frontend time: 0.256s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.316s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 91, 128]) + layer.0.v_cache: torch.Size([1, 4, 91, 128]) + layer.1.k_cache: torch.Size([1, 4, 91, 128]) + layer.1.v_cache: torch.Size([1, 4, 91, 128]) + layer.2.k_cache: torch.Size([1, 4, 91, 128]) + layer.2.v_cache: torch.Size([1, 4, 91, 128]) + layer.3.k_cache: torch.Size([1, 4, 91, 128]) + layer.3.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.k_cache: torch.Size([1, 4, 91, 128]) + layer.4.v_cache: torch.Size([1, 4, 91, 128]) + layer.4.output: torch.Size([1, 91, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11186004 28.64250300 + layer.0.v_cache 0.00001364 0.00543734 + layer.1.k_cache 0.01386243 3.58272812 + layer.1.v_cache 0.00000608 0.00254275 + layer.2.k_cache 0.00506424 0.46458917 + layer.2.v_cache 0.00001872 0.00614436 + layer.3.k_cache 0.02496567 2.10663680 + layer.3.v_cache 0.00001876 0.00667154 + layer.4.k_cache 0.00064518 0.12408764 + layer.4.v_cache 0.00005436 0.01350429 + layer.4.output 0.16682192 579.36970173 + ------------------------------------------------------------------------------------- + TOTAL 0.07789780 240.62016218 + (elements=792,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 792064 +Total Bytes 175300 +BPFP 1.7706 bits/point +EBPFP 3.5411 equivalent bits/point +MSE 240.620162 +---------------------- -------------------------------------------------------- +Time: 0.576s Load: 0.004s, Pack+Encode: 0.256s, Decode+Unpack: 0.316s +---------------------- -------------------------------------------------------- +💾 Converting with 240.6202 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-447.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,144B, BPFP=0.9684 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,256B, BPFP=3.0602 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,112B, BPFP=1.7154 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,780B, BPFP=2.9706 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,344B, BPFP=1.9473 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,388B, BPFP=2.8968 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,876B, BPFP=1.8592 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,464B, BPFP=2.9111 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,528B, BPFP=2.5467 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,064B, BPFP=2.8358 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,408B, BPFP=0.9791 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.275s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08542032 26.16740399 + layer.0.v_cache 0.00001402 0.00550848 + layer.1.k_cache 0.01083013 3.31888038 + layer.1.v_cache 0.00000556 0.00233005 + layer.2.k_cache 0.00382075 0.45413254 + layer.2.v_cache 0.00001860 0.00657444 + layer.3.k_cache 0.02610962 2.21587491 + layer.3.v_cache 0.00002051 0.00720494 + layer.4.k_cache 0.00063312 0.12158496 + layer.4.v_cache 0.00005614 0.01434538 + layer.4.output 0.16811641 637.38941480 + ------------------------------------------------------------------------------------- + TOTAL 0.07669080 264.35527904 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 162364 +BPFP 1.7980 bits/point +EBPFP 3.5959 equivalent bits/point +MSE 264.355279 +---------------------- -------------------------------------------------------- +Time: 0.528s Load: 0.004s, Pack+Encode: 0.249s, Decode+Unpack: 0.275s +---------------------- -------------------------------------------------------- +💾 Converting with 264.3553 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-449.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-449.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,072B, BPFP=0.9784 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,960B, BPFP=3.0787 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,076B, BPFP=1.7508 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,288B, BPFP=2.9491 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,268B, BPFP=1.9807 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,960B, BPFP=2.8858 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,760B, BPFP=1.8827 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,232B, BPFP=2.9383 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,356B, BPFP=2.5764 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,980B, BPFP=2.8897 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,676B, BPFP=0.9556 +⌛️ [2/4] FRONTEND: Frontend time: 0.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.301s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10540552 31.17188404 + layer.0.v_cache 0.00001541 0.00564881 + layer.1.k_cache 0.03402772 3.76348764 + layer.1.v_cache 0.00000555 0.00261757 + layer.2.k_cache 0.00394259 0.48537558 + layer.2.v_cache 0.00001868 0.00661083 + layer.3.k_cache 0.02873192 2.47178405 + layer.3.v_cache 0.00001983 0.00715064 + layer.4.k_cache 0.00063741 0.13364078 + layer.4.v_cache 0.00005806 0.01453625 + layer.4.output 0.17961645 660.76184965 + ------------------------------------------------------------------------------------- + TOTAL 0.08412811 274.31739316 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 158628 +BPFP 1.8000 bits/point +EBPFP 3.5999 equivalent bits/point +MSE 274.317393 +---------------------- -------------------------------------------------------- +Time: 0.573s Load: 0.004s, Pack+Encode: 0.267s, Decode+Unpack: 0.301s +---------------------- -------------------------------------------------------- +💾 Converting with 274.3174 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-45.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-45.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,040B, BPFP=0.9844 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,048B, BPFP=3.1344 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,000B, BPFP=1.7578 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,644B, BPFP=3.0555 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,388B, BPFP=2.0289 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,216B, BPFP=2.9719 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,844B, BPFP=1.9227 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,460B, BPFP=3.0195 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,512B, BPFP=2.6391 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,076B, BPFP=2.9445 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,252B, BPFP=1.0115 +⌛️ [2/4] FRONTEND: Frontend time: 0.247s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.289s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10022384 31.32287292 + layer.0.v_cache 0.00001350 0.00554127 + layer.1.k_cache 0.03445883 3.50318832 + layer.1.v_cache 0.00000556 0.00245139 + layer.2.k_cache 0.00550650 0.50020132 + layer.2.v_cache 0.00001848 0.00607742 + layer.3.k_cache 0.03016714 2.28686562 + layer.3.v_cache 0.00001937 0.00742989 + layer.4.k_cache 0.00063435 0.12464328 + layer.4.v_cache 0.00006510 0.01371376 + layer.4.output 0.18593751 655.08989955 + ------------------------------------------------------------------------------------- + TOTAL 0.08662796 271.96484012 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 161480 +BPFP 1.8552 bits/point +EBPFP 3.7105 equivalent bits/point +MSE 271.964840 +---------------------- -------------------------------------------------------- +Time: 0.540s Load: 0.003s, Pack+Encode: 0.247s, Decode+Unpack: 0.289s +---------------------- -------------------------------------------------------- +💾 Converting with 271.9648 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-453.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-453.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 97, 128) +Output shape: (1, 97, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) -> torch.Size([1, 1, 97, 512]) + layer.4.output: torch.Size([1, 97, 3584]) -> torch.Size([1, 1, 97, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,304B, BPFP=0.8544 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,752B, BPFP=2.8595 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,948B, BPFP=1.6024 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 17,304B, BPFP=2.7874 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,104B, BPFP=1.7887 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,828B, BPFP=2.7107 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,556B, BPFP=1.7004 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,992B, BPFP=2.7371 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,860B, BPFP=2.3937 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,452B, BPFP=2.6501 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,728B, BPFP=0.8682 +⌛️ [2/4] FRONTEND: Frontend time: 0.266s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 97, 128]) + layer.0.v_cache: torch.Size([1, 4, 97, 128]) + layer.1.k_cache: torch.Size([1, 4, 97, 128]) + layer.1.v_cache: torch.Size([1, 4, 97, 128]) + layer.2.k_cache: torch.Size([1, 4, 97, 128]) + layer.2.v_cache: torch.Size([1, 4, 97, 128]) + layer.3.k_cache: torch.Size([1, 4, 97, 128]) + layer.3.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.k_cache: torch.Size([1, 4, 97, 128]) + layer.4.v_cache: torch.Size([1, 4, 97, 128]) + layer.4.output: torch.Size([1, 97, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11898830 27.81759927 + layer.0.v_cache 0.00001403 0.00581718 + layer.1.k_cache 0.01243949 3.52536546 + layer.1.v_cache 0.00000551 0.00247236 + layer.2.k_cache 0.00346032 0.48396600 + layer.2.v_cache 0.00001901 0.00627149 + layer.3.k_cache 0.02403709 2.11344453 + layer.3.v_cache 0.00001994 0.00688725 + layer.4.k_cache 0.00063949 0.12903746 + layer.4.v_cache 0.00006591 0.01361849 + layer.4.output 0.03773703 567.85175810 + ------------------------------------------------------------------------------------- + TOTAL 0.02493225 235.82745801 + (elements=844,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 844288 +Total Bytes 174828 +BPFP 1.6566 bits/point +EBPFP 3.3131 equivalent bits/point +MSE 235.827458 +---------------------- -------------------------------------------------------- +Time: 0.612s Load: 0.004s, Pack+Encode: 0.266s, Decode+Unpack: 0.341s +---------------------- -------------------------------------------------------- +💾 Converting with 235.8275 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-465.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 82, 128) +Output shape: (1, 82, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) -> torch.Size([1, 1, 82, 512]) + layer.4.output: torch.Size([1, 82, 3584]) -> torch.Size([1, 1, 82, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,144B, BPFP=0.9802 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,056B, BPFP=3.0595 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,128B, BPFP=1.7393 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,596B, BPFP=2.9718 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,336B, BPFP=1.9695 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,976B, BPFP=2.8537 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,716B, BPFP=1.8514 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,180B, BPFP=2.8925 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,432B, BPFP=2.5595 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,928B, BPFP=2.8445 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,260B, BPFP=0.9598 +⌛️ [2/4] FRONTEND: Frontend time: 0.271s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.296s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 82, 128]) + layer.0.v_cache: torch.Size([1, 4, 82, 128]) + layer.1.k_cache: torch.Size([1, 4, 82, 128]) + layer.1.v_cache: torch.Size([1, 4, 82, 128]) + layer.2.k_cache: torch.Size([1, 4, 82, 128]) + layer.2.v_cache: torch.Size([1, 4, 82, 128]) + layer.3.k_cache: torch.Size([1, 4, 82, 128]) + layer.3.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.k_cache: torch.Size([1, 4, 82, 128]) + layer.4.v_cache: torch.Size([1, 4, 82, 128]) + layer.4.output: torch.Size([1, 82, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10693085 28.28715106 + layer.0.v_cache 0.00001341 0.00582292 + layer.1.k_cache 0.01172146 3.48937765 + layer.1.v_cache 0.00000553 0.00259229 + layer.2.k_cache 0.00236719 0.47749464 + layer.2.v_cache 0.00001835 0.00664226 + layer.3.k_cache 0.02622769 2.30087113 + layer.3.v_cache 0.00001864 0.00728213 + layer.4.k_cache 0.00061713 0.13060387 + layer.4.v_cache 0.00004943 0.01419744 + layer.4.output 0.18273007 652.64868249 + ------------------------------------------------------------------------------------- + TOTAL 0.08394589 270.78016546 + (elements=713,728) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 713728 +Total Bytes 159752 +BPFP 1.7906 bits/point +EBPFP 3.5812 equivalent bits/point +MSE 270.780165 +---------------------- -------------------------------------------------------- +Time: 0.572s Load: 0.005s, Pack+Encode: 0.271s, Decode+Unpack: 0.296s +---------------------- -------------------------------------------------------- +💾 Converting with 270.7802 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-467.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,800B, BPFP=0.8973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 18,832B, BPFP=2.9134 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,680B, BPFP=1.6522 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,552B, BPFP=2.8700 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,852B, BPFP=1.8335 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,068B, BPFP=2.7952 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,392B, BPFP=1.7624 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,212B, BPFP=2.8175 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 15,920B, BPFP=2.4629 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,608B, BPFP=2.7240 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,228B, BPFP=0.8670 +⌛️ [2/4] FRONTEND: Frontend time: 0.238s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.324s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13156438 28.79756633 + layer.0.v_cache 0.00001398 0.00584924 + layer.1.k_cache 0.04507210 3.74414486 + layer.1.v_cache 0.00000603 0.00257992 + layer.2.k_cache 0.00848899 0.44223914 + layer.2.v_cache 0.00001983 0.00664307 + layer.3.k_cache 0.06093804 1.96633790 + layer.3.v_cache 0.00001891 0.00705882 + layer.4.k_cache 0.00062755 0.13099263 + layer.4.v_cache 0.00004781 0.01296300 + layer.4.output 11.29576791 525.92980905 + ------------------------------------------------------------------------------------- + TOTAL 4.66571606 218.62500225 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 186144 +BPFP 1.6939 bits/point +EBPFP 3.3879 equivalent bits/point +MSE 218.625002 +---------------------- -------------------------------------------------------- +Time: 0.566s Load: 0.004s, Pack+Encode: 0.238s, Decode+Unpack: 0.324s +---------------------- -------------------------------------------------------- +💾 Converting with 218.6250 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-488.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-488.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,184B, BPFP=0.9759 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,312B, BPFP=3.0708 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,180B, BPFP=1.7282 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,872B, BPFP=2.9880 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,628B, BPFP=2.0008 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,608B, BPFP=2.9383 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,124B, BPFP=1.9059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,776B, BPFP=2.9699 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,580B, BPFP=2.5565 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,432B, BPFP=2.9051 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,760B, BPFP=0.9886 +⌛️ [2/4] FRONTEND: Frontend time: 0.270s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.321s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12340720 26.81072925 + layer.0.v_cache 0.00001379 0.00549979 + layer.1.k_cache 0.01365945 3.43604619 + layer.1.v_cache 0.00000586 0.00230151 + layer.2.k_cache 0.00515996 0.43873422 + layer.2.v_cache 0.00002118 0.00651011 + layer.3.k_cache 0.04682968 2.27716763 + layer.3.v_cache 0.00001899 0.00687771 + layer.4.k_cache 0.00062596 0.12096028 + layer.4.v_cache 0.00004975 0.01369728 + layer.4.output 0.18155291 632.73225043 + ------------------------------------------------------------------------------------- + TOTAL 0.08592130 262.48495747 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 164456 +BPFP 1.8211 bits/point +EBPFP 3.6423 equivalent bits/point +MSE 262.484957 +---------------------- -------------------------------------------------------- +Time: 0.596s Load: 0.005s, Pack+Encode: 0.270s, Decode+Unpack: 0.321s +---------------------- -------------------------------------------------------- +💾 Converting with 262.4850 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-491.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-491.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,224B, BPFP=0.9717 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,312B, BPFP=3.0342 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,208B, BPFP=1.7128 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,904B, BPFP=2.9583 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,312B, BPFP=1.9182 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,400B, BPFP=2.8646 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,000B, BPFP=1.8601 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,644B, BPFP=2.9100 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,612B, BPFP=2.5320 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,364B, BPFP=2.8579 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 37,436B, BPFP=0.9948 +⌛️ [2/4] FRONTEND: Frontend time: 0.266s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.326s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10400932 27.40533738 + layer.0.v_cache 0.00001381 0.00539372 + layer.1.k_cache 0.01300498 3.22774578 + layer.1.v_cache 0.00000559 0.00235331 + layer.2.k_cache 0.00535372 0.45658884 + layer.2.v_cache 0.00001862 0.00612441 + layer.3.k_cache 0.04415199 2.05105228 + layer.3.v_cache 0.00001841 0.00706595 + layer.4.k_cache 0.00060249 0.11892925 + layer.4.v_cache 0.00005274 0.01450878 + layer.4.output 0.16694923 616.57886905 + ------------------------------------------------------------------------------------- + TOTAL 0.07858096 255.84395194 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 164416 +BPFP 1.7990 bits/point +EBPFP 3.5980 equivalent bits/point +MSE 255.843952 +---------------------- -------------------------------------------------------- +Time: 0.596s Load: 0.004s, Pack+Encode: 0.266s, Decode+Unpack: 0.326s +---------------------- -------------------------------------------------------- +💾 Converting with 255.8440 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-492.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-492.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 80, 128) +Output shape: (1, 80, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) -> torch.Size([1, 1, 80, 512]) + layer.4.output: torch.Size([1, 80, 3584]) -> torch.Size([1, 1, 80, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,936B, BPFP=0.9641 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,092B, BPFP=3.1430 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,888B, BPFP=1.7359 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,604B, BPFP=3.0477 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,280B, BPFP=2.0078 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,360B, BPFP=3.0000 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,812B, BPFP=1.9164 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,544B, BPFP=3.0359 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,372B, BPFP=2.6117 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,208B, BPFP=2.9703 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 34,316B, BPFP=0.9575 +⌛️ [2/4] FRONTEND: Frontend time: 0.263s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.322s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 80, 128]) + layer.0.v_cache: torch.Size([1, 4, 80, 128]) + layer.1.k_cache: torch.Size([1, 4, 80, 128]) + layer.1.v_cache: torch.Size([1, 4, 80, 128]) + layer.2.k_cache: torch.Size([1, 4, 80, 128]) + layer.2.v_cache: torch.Size([1, 4, 80, 128]) + layer.3.k_cache: torch.Size([1, 4, 80, 128]) + layer.3.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.k_cache: torch.Size([1, 4, 80, 128]) + layer.4.v_cache: torch.Size([1, 4, 80, 128]) + layer.4.output: torch.Size([1, 80, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09321334 30.35870972 + layer.0.v_cache 0.00001862 0.00589841 + layer.1.k_cache 0.03257937 3.48467178 + layer.1.v_cache 0.00000532 0.00238720 + layer.2.k_cache 0.01442847 0.51776533 + layer.2.v_cache 0.00001865 0.00651698 + layer.3.k_cache 0.04385926 2.48806076 + layer.3.v_cache 0.00001828 0.00730399 + layer.4.k_cache 0.00061064 0.12767259 + layer.4.v_cache 0.00005353 0.01484137 + layer.4.output 0.18500509 657.42075893 + ------------------------------------------------------------------------------------- + TOTAL 0.08704948 272.87994945 + (elements=696,320) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 696320 +Total Bytes 159412 +BPFP 1.8315 bits/point +EBPFP 3.6630 equivalent bits/point +MSE 272.879949 +---------------------- -------------------------------------------------------- +Time: 0.589s Load: 0.004s, Pack+Encode: 0.263s, Decode+Unpack: 0.322s +---------------------- -------------------------------------------------------- +💾 Converting with 272.8799 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-5.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-5.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 103, 128) +Output shape: (1, 103, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) -> torch.Size([1, 1, 103, 512]) + layer.4.output: torch.Size([1, 103, 3584]) -> torch.Size([1, 1, 103, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,960B, BPFP=0.9041 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,204B, BPFP=2.9132 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,928B, BPFP=1.6578 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,876B, BPFP=2.8635 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,228B, BPFP=1.8550 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,348B, BPFP=2.7834 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,620B, BPFP=1.7627 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,592B, BPFP=2.8204 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,296B, BPFP=2.4721 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,088B, BPFP=2.7439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,804B, BPFP=0.8843 +⌛️ [2/4] FRONTEND: Frontend time: 0.275s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 103, 128]) + layer.0.v_cache: torch.Size([1, 4, 103, 128]) + layer.1.k_cache: torch.Size([1, 4, 103, 128]) + layer.1.v_cache: torch.Size([1, 4, 103, 128]) + layer.2.k_cache: torch.Size([1, 4, 103, 128]) + layer.2.v_cache: torch.Size([1, 4, 103, 128]) + layer.3.k_cache: torch.Size([1, 4, 103, 128]) + layer.3.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.k_cache: torch.Size([1, 4, 103, 128]) + layer.4.v_cache: torch.Size([1, 4, 103, 128]) + layer.4.output: torch.Size([1, 103, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11527186 26.91742832 + layer.0.v_cache 0.00001510 0.00561922 + layer.1.k_cache 0.02920971 3.65053462 + layer.1.v_cache 0.00000551 0.00254582 + layer.2.k_cache 0.00857845 0.48201122 + layer.2.v_cache 0.00001837 0.00653289 + layer.3.k_cache 0.03550095 1.98615776 + layer.3.v_cache 0.00001950 0.00693050 + layer.4.k_cache 0.00063242 0.12519032 + layer.4.v_cache 0.00005150 0.01472964 + layer.4.output 11.08123415 508.10662275 + ------------------------------------------------------------------------------------- + TOTAL 4.57399661 211.17317880 + (elements=896,512) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 896512 +Total Bytes 190944 +BPFP 1.7039 bits/point +EBPFP 3.4078 equivalent bits/point +MSE 211.173179 +---------------------- -------------------------------------------------------- +Time: 0.598s Load: 0.004s, Pack+Encode: 0.275s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 211.1732 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-503.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,676B, BPFP=0.9640 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,304B, BPFP=2.9389 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,052B, BPFP=1.7072 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,976B, BPFP=2.8832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,304B, BPFP=1.9198 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,592B, BPFP=2.8179 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,796B, BPFP=1.8336 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,712B, BPFP=2.8383 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,800B, BPFP=2.5136 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,352B, BPFP=2.7772 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 38,372B, BPFP=0.9310 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.298s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09673090 27.48009192 + layer.0.v_cache 0.00001344 0.00536837 + layer.1.k_cache 0.04918879 3.38404647 + layer.1.v_cache 0.00000572 0.00248733 + layer.2.k_cache 0.00880511 0.47360495 + layer.2.v_cache 0.00001944 0.00610940 + layer.3.k_cache 0.06586690 2.30593159 + layer.3.v_cache 0.00001823 0.00694088 + layer.4.k_cache 0.00062484 0.12865999 + layer.4.v_cache 0.00006325 0.01352551 + layer.4.output 0.15294319 567.58419061 + ------------------------------------------------------------------------------------- + TOTAL 0.07599641 235.69977063 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 174936 +BPFP 1.7477 bits/point +EBPFP 3.4954 equivalent bits/point +MSE 235.699771 +---------------------- -------------------------------------------------------- +Time: 0.552s Load: 0.004s, Pack+Encode: 0.250s, Decode+Unpack: 0.298s +---------------------- -------------------------------------------------------- +💾 Converting with 235.6998 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-51.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-51.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 92, 128) +Output shape: (1, 92, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) -> torch.Size([1, 1, 92, 512]) + layer.4.output: torch.Size([1, 92, 3584]) -> torch.Size([1, 1, 92, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,708B, BPFP=0.9694 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 17,276B, BPFP=2.9341 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,176B, BPFP=1.7283 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,968B, BPFP=2.8818 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,328B, BPFP=1.9239 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,668B, BPFP=2.8308 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,672B, BPFP=1.8125 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,836B, BPFP=2.8594 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,808B, BPFP=2.5149 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 16,268B, BPFP=2.7629 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 39,292B, BPFP=0.9533 +⌛️ [2/4] FRONTEND: Frontend time: 0.243s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.284s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 92, 128]) + layer.0.v_cache: torch.Size([1, 4, 92, 128]) + layer.1.k_cache: torch.Size([1, 4, 92, 128]) + layer.1.v_cache: torch.Size([1, 4, 92, 128]) + layer.2.k_cache: torch.Size([1, 4, 92, 128]) + layer.2.v_cache: torch.Size([1, 4, 92, 128]) + layer.3.k_cache: torch.Size([1, 4, 92, 128]) + layer.3.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.k_cache: torch.Size([1, 4, 92, 128]) + layer.4.v_cache: torch.Size([1, 4, 92, 128]) + layer.4.output: torch.Size([1, 92, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12676482 27.21300739 + layer.0.v_cache 0.00001339 0.00550745 + layer.1.k_cache 0.03133728 3.19835232 + layer.1.v_cache 0.00000590 0.00241535 + layer.2.k_cache 0.00372887 0.49715639 + layer.2.v_cache 0.00001820 0.00631010 + layer.3.k_cache 0.02418065 2.07210541 + layer.3.v_cache 0.00002034 0.00692640 + layer.4.k_cache 0.00062996 0.12380817 + layer.4.v_cache 0.00006053 0.01302396 + layer.4.output 0.15389095 571.54357531 + ------------------------------------------------------------------------------------- + TOTAL 0.07435274 237.29080236 + (elements=800,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 800768 +Total Bytes 176000 +BPFP 1.7583 bits/point +EBPFP 3.5166 equivalent bits/point +MSE 237.290802 +---------------------- -------------------------------------------------------- +Time: 0.532s Load: 0.006s, Pack+Encode: 0.243s, Decode+Unpack: 0.284s +---------------------- -------------------------------------------------------- +💾 Converting with 237.2908 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-511.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-511.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 90, 128) +Output shape: (1, 90, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) -> torch.Size([1, 1, 90, 512]) + layer.4.output: torch.Size([1, 90, 3584]) -> torch.Size([1, 1, 90, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,480B, BPFP=0.9514 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,900B, BPFP=2.9340 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,880B, BPFP=1.7153 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,552B, BPFP=2.8736 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 11,028B, BPFP=1.9146 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 16,100B, BPFP=2.7951 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,324B, BPFP=1.7924 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,228B, BPFP=2.8174 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 14,344B, BPFP=2.4903 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,924B, BPFP=2.7646 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,616B, BPFP=0.9081 +⌛️ [2/4] FRONTEND: Frontend time: 0.226s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.300s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 90, 128]) + layer.0.v_cache: torch.Size([1, 4, 90, 128]) + layer.1.k_cache: torch.Size([1, 4, 90, 128]) + layer.1.v_cache: torch.Size([1, 4, 90, 128]) + layer.2.k_cache: torch.Size([1, 4, 90, 128]) + layer.2.v_cache: torch.Size([1, 4, 90, 128]) + layer.3.k_cache: torch.Size([1, 4, 90, 128]) + layer.3.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.k_cache: torch.Size([1, 4, 90, 128]) + layer.4.v_cache: torch.Size([1, 4, 90, 128]) + layer.4.output: torch.Size([1, 90, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12863920 28.76035156 + layer.0.v_cache 0.00001425 0.00557608 + layer.1.k_cache 0.01389826 3.65372518 + layer.1.v_cache 0.00000571 0.00246357 + layer.2.k_cache 0.00659207 0.47599703 + layer.2.v_cache 0.00001883 0.00605852 + layer.3.k_cache 0.05540012 2.29106479 + layer.3.v_cache 0.00001922 0.00686939 + layer.4.k_cache 0.00063279 0.12489071 + layer.4.v_cache 0.00005416 0.01307744 + layer.4.output 0.16196846 596.56041667 + ------------------------------------------------------------------------------------- + TOTAL 0.07876787 247.72135241 + (elements=783,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 783360 +Total Bytes 169376 +BPFP 1.7297 bits/point +EBPFP 3.4595 equivalent bits/point +MSE 247.721352 +---------------------- -------------------------------------------------------- +Time: 0.531s Load: 0.004s, Pack+Encode: 0.226s, Decode+Unpack: 0.300s +---------------------- -------------------------------------------------------- +💾 Converting with 247.7214 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-515.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-515.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.004s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 108, 128) +Output shape: (1, 108, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) -> torch.Size([1, 1, 108, 512]) + layer.4.output: torch.Size([1, 108, 3584]) -> torch.Size([1, 1, 108, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 6,024B, BPFP=0.8715 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,728B, BPFP=2.8542 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 11,088B, BPFP=1.6042 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 19,292B, BPFP=2.7911 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,596B, BPFP=1.8223 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,960B, BPFP=2.7431 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,836B, BPFP=1.7124 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 19,096B, BPFP=2.7627 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,592B, BPFP=2.4005 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 18,528B, BPFP=2.6806 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 42,904B, BPFP=0.8867 +⌛️ [2/4] FRONTEND: Frontend time: 0.250s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.297s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 108, 128]) + layer.0.v_cache: torch.Size([1, 4, 108, 128]) + layer.1.k_cache: torch.Size([1, 4, 108, 128]) + layer.1.v_cache: torch.Size([1, 4, 108, 128]) + layer.2.k_cache: torch.Size([1, 4, 108, 128]) + layer.2.v_cache: torch.Size([1, 4, 108, 128]) + layer.3.k_cache: torch.Size([1, 4, 108, 128]) + layer.3.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.k_cache: torch.Size([1, 4, 108, 128]) + layer.4.v_cache: torch.Size([1, 4, 108, 128]) + layer.4.output: torch.Size([1, 108, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13160179 27.55112938 + layer.0.v_cache 0.00001577 0.00566421 + layer.1.k_cache 0.05775581 3.28648094 + layer.1.v_cache 0.00000549 0.00247568 + layer.2.k_cache 0.01027601 0.46034912 + layer.2.v_cache 0.00002197 0.00655192 + layer.3.k_cache 0.07329086 2.16799192 + layer.3.v_cache 0.00001983 0.00742016 + layer.4.k_cache 0.00064477 0.12804068 + layer.4.v_cache 0.00005260 0.01397162 + layer.4.output 10.56448284 484.42154431 + ------------------------------------------------------------------------------------- + TOTAL 4.36618028 201.44593446 + (elements=940,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 940032 +Total Bytes 196644 +BPFP 1.6735 bits/point +EBPFP 3.3470 equivalent bits/point +MSE 201.445934 +---------------------- -------------------------------------------------------- +Time: 0.551s Load: 0.004s, Pack+Encode: 0.250s, Decode+Unpack: 0.297s +---------------------- -------------------------------------------------------- +💾 Converting with 201.4459 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-61.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-61.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 83, 128) +Output shape: (1, 83, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) -> torch.Size([1, 1, 83, 512]) + layer.4.output: torch.Size([1, 83, 3584]) -> torch.Size([1, 1, 83, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,120B, BPFP=0.9639 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,172B, BPFP=3.0444 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,148B, BPFP=1.7221 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,784B, BPFP=2.9714 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,380B, BPFP=1.9541 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,196B, BPFP=2.8607 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,864B, BPFP=1.8569 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,476B, BPFP=2.9134 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,552B, BPFP=2.5512 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,076B, BPFP=2.8381 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,352B, BPFP=0.9507 +⌛️ [2/4] FRONTEND: Frontend time: 0.248s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.311s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 83, 128]) + layer.0.v_cache: torch.Size([1, 4, 83, 128]) + layer.1.k_cache: torch.Size([1, 4, 83, 128]) + layer.1.v_cache: torch.Size([1, 4, 83, 128]) + layer.2.k_cache: torch.Size([1, 4, 83, 128]) + layer.2.v_cache: torch.Size([1, 4, 83, 128]) + layer.3.k_cache: torch.Size([1, 4, 83, 128]) + layer.3.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.k_cache: torch.Size([1, 4, 83, 128]) + layer.4.v_cache: torch.Size([1, 4, 83, 128]) + layer.4.output: torch.Size([1, 83, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10139442 27.24542898 + layer.0.v_cache 0.00001346 0.00520017 + layer.1.k_cache 0.03266963 3.56900135 + layer.1.v_cache 0.00000526 0.00240297 + layer.2.k_cache 0.01283502 0.41340453 + layer.2.v_cache 0.00001884 0.00613185 + layer.3.k_cache 0.02666297 2.14228582 + layer.3.v_cache 0.00001780 0.00643895 + layer.4.k_cache 0.00061512 0.12085411 + layer.4.v_cache 0.00004947 0.01289923 + layer.4.output 0.18393325 630.44320138 + ------------------------------------------------------------------------------------- + TOTAL 0.08598910 261.56626221 + (elements=722,432) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 722432 +Total Bytes 161120 +BPFP 1.7842 bits/point +EBPFP 3.5684 equivalent bits/point +MSE 261.566262 +---------------------- -------------------------------------------------------- +Time: 0.561s Load: 0.003s, Pack+Encode: 0.248s, Decode+Unpack: 0.311s +---------------------- -------------------------------------------------------- +💾 Converting with 261.5663 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-63.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-63.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 101, 128) +Output shape: (1, 101, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) -> torch.Size([1, 1, 101, 512]) + layer.4.output: torch.Size([1, 101, 3584]) -> torch.Size([1, 1, 101, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,816B, BPFP=0.8998 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 19,008B, BPFP=2.9406 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 10,776B, BPFP=1.6671 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 18,672B, BPFP=2.8886 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 12,232B, BPFP=1.8923 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 18,280B, BPFP=2.8280 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 11,600B, BPFP=1.7946 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 18,392B, BPFP=2.8453 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 16,116B, BPFP=2.4932 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 17,768B, BPFP=2.7488 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 40,296B, BPFP=0.8906 +⌛️ [2/4] FRONTEND: Frontend time: 0.261s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.307s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 101, 128]) + layer.0.v_cache: torch.Size([1, 4, 101, 128]) + layer.1.k_cache: torch.Size([1, 4, 101, 128]) + layer.1.v_cache: torch.Size([1, 4, 101, 128]) + layer.2.k_cache: torch.Size([1, 4, 101, 128]) + layer.2.v_cache: torch.Size([1, 4, 101, 128]) + layer.3.k_cache: torch.Size([1, 4, 101, 128]) + layer.3.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.k_cache: torch.Size([1, 4, 101, 128]) + layer.4.v_cache: torch.Size([1, 4, 101, 128]) + layer.4.output: torch.Size([1, 101, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11149573 28.16126334 + layer.0.v_cache 0.00001357 0.00551148 + layer.1.k_cache 0.08092193 3.58712134 + layer.1.v_cache 0.00000571 0.00244617 + layer.2.k_cache 0.01109533 0.49251730 + layer.2.v_cache 0.00001920 0.00624328 + layer.3.k_cache 0.03617269 2.36528287 + layer.3.v_cache 0.00001858 0.00662822 + layer.4.k_cache 0.00062559 0.12162941 + layer.4.v_cache 0.00005684 0.01322459 + layer.4.output 11.30219499 526.14979668 + ------------------------------------------------------------------------------------- + TOTAL 4.66798765 218.69473204 + (elements=879,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 879104 +Total Bytes 188956 +BPFP 1.7195 bits/point +EBPFP 3.4391 equivalent bits/point +MSE 218.694732 +---------------------- -------------------------------------------------------- +Time: 0.573s Load: 0.005s, Pack+Encode: 0.261s, Decode+Unpack: 0.307s +---------------------- -------------------------------------------------------- +💾 Converting with 218.6947 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-9.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-9.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.003s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 78, 128) +Output shape: (1, 78, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) -> torch.Size([1, 1, 78, 512]) + layer.4.output: torch.Size([1, 78, 3584]) -> torch.Size([1, 1, 78, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 4,952B, BPFP=0.9920 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,944B, BPFP=3.1939 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,900B, BPFP=1.7829 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,492B, BPFP=3.1034 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,212B, BPFP=2.0457 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 14,960B, BPFP=2.9968 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,572B, BPFP=1.9175 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,208B, BPFP=3.0465 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,320B, BPFP=2.6683 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,964B, BPFP=2.9976 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,024B, BPFP=0.9451 +⌛️ [2/4] FRONTEND: Frontend time: 0.265s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.306s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 78, 128]) + layer.0.v_cache: torch.Size([1, 4, 78, 128]) + layer.1.k_cache: torch.Size([1, 4, 78, 128]) + layer.1.v_cache: torch.Size([1, 4, 78, 128]) + layer.2.k_cache: torch.Size([1, 4, 78, 128]) + layer.2.v_cache: torch.Size([1, 4, 78, 128]) + layer.3.k_cache: torch.Size([1, 4, 78, 128]) + layer.3.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.k_cache: torch.Size([1, 4, 78, 128]) + layer.4.v_cache: torch.Size([1, 4, 78, 128]) + layer.4.output: torch.Size([1, 78, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10167113 30.29237092 + layer.0.v_cache 0.00001386 0.00575833 + layer.1.k_cache 0.03545215 3.65027442 + layer.1.v_cache 0.00000569 0.00246116 + layer.2.k_cache 0.01047806 0.51011853 + layer.2.v_cache 0.00001755 0.00635857 + layer.3.k_cache 0.06687863 2.51901049 + layer.3.v_cache 0.00001960 0.00697387 + layer.4.k_cache 0.00059181 0.11798568 + layer.4.v_cache 0.00004882 0.01276406 + layer.4.output 0.19633365 690.34546703 + ------------------------------------------------------------------------------------- + TOTAL 0.09350076 286.44366737 + (elements=678,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 678912 +Total Bytes 156548 +BPFP 1.8447 bits/point +EBPFP 3.6894 equivalent bits/point +MSE 286.443667 +---------------------- -------------------------------------------------------- +Time: 0.574s Load: 0.003s, Pack+Encode: 0.265s, Decode+Unpack: 0.306s +---------------------- -------------------------------------------------------- +💾 Converting with 286.4437 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-90.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-90.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 87, 128) +Output shape: (1, 87, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) -> torch.Size([1, 1, 87, 512]) + layer.4.output: torch.Size([1, 87, 3584]) -> torch.Size([1, 1, 87, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,260B, BPFP=0.9447 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,656B, BPFP=2.9914 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,364B, BPFP=1.6818 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,240B, BPFP=2.9167 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,704B, BPFP=1.9224 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,828B, BPFP=2.8427 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,104B, BPFP=1.8147 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 16,136B, BPFP=2.8980 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,888B, BPFP=2.4943 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,580B, BPFP=2.7981 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 36,316B, BPFP=0.9318 +⌛️ [2/4] FRONTEND: Frontend time: 0.267s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.319s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 87, 128]) + layer.0.v_cache: torch.Size([1, 4, 87, 128]) + layer.1.k_cache: torch.Size([1, 4, 87, 128]) + layer.1.v_cache: torch.Size([1, 4, 87, 128]) + layer.2.k_cache: torch.Size([1, 4, 87, 128]) + layer.2.v_cache: torch.Size([1, 4, 87, 128]) + layer.3.k_cache: torch.Size([1, 4, 87, 128]) + layer.3.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.k_cache: torch.Size([1, 4, 87, 128]) + layer.4.v_cache: torch.Size([1, 4, 87, 128]) + layer.4.output: torch.Size([1, 87, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11445767 27.26581302 + layer.0.v_cache 0.00001583 0.00556801 + layer.1.k_cache 0.03092440 3.47675192 + layer.1.v_cache 0.00000551 0.00234885 + layer.2.k_cache 0.00639992 0.44404300 + layer.2.v_cache 0.00001897 0.00624743 + layer.3.k_cache 0.05613507 2.35141605 + layer.3.v_cache 0.00002026 0.00722066 + layer.4.k_cache 0.00063356 0.12233548 + layer.4.v_cache 0.00005123 0.01367351 + layer.4.output 0.16055210 608.32815066 + ------------------------------------------------------------------------------------- + TOTAL 0.07838395 252.47014544 + (elements=757,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 757248 +Total Bytes 166076 +BPFP 1.7545 bits/point +EBPFP 3.5090 equivalent bits/point +MSE 252.470145 +---------------------- -------------------------------------------------------- +Time: 0.591s Load: 0.005s, Pack+Encode: 0.267s, Decode+Unpack: 0.319s +---------------------- -------------------------------------------------------- +💾 Converting with 252.4701 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-95.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-95.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 84, 128) +Output shape: (1, 84, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) -> torch.Size([1, 1, 84, 512]) + layer.4.output: torch.Size([1, 84, 3584]) -> torch.Size([1, 1, 84, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,188B, BPFP=0.9650 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 16,344B, BPFP=3.0402 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 9,280B, BPFP=1.7262 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 16,024B, BPFP=2.9807 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,484B, BPFP=1.9501 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,596B, BPFP=2.9010 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 10,104B, BPFP=1.8795 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,676B, BPFP=2.9159 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,800B, BPFP=2.5670 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 15,356B, BPFP=2.8564 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 35,548B, BPFP=0.9446 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.293s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 84, 128]) + layer.0.v_cache: torch.Size([1, 4, 84, 128]) + layer.1.k_cache: torch.Size([1, 4, 84, 128]) + layer.1.v_cache: torch.Size([1, 4, 84, 128]) + layer.2.k_cache: torch.Size([1, 4, 84, 128]) + layer.2.v_cache: torch.Size([1, 4, 84, 128]) + layer.3.k_cache: torch.Size([1, 4, 84, 128]) + layer.3.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.k_cache: torch.Size([1, 4, 84, 128]) + layer.4.v_cache: torch.Size([1, 4, 84, 128]) + layer.4.output: torch.Size([1, 84, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13956541 28.03517950 + layer.0.v_cache 0.00001349 0.00581943 + layer.1.k_cache 0.03336124 3.60053871 + layer.1.v_cache 0.00000558 0.00265874 + layer.2.k_cache 0.00698305 0.46042710 + layer.2.v_cache 0.00001831 0.00653183 + layer.3.k_cache 0.07289400 1.98007638 + layer.3.v_cache 0.00001855 0.00709004 + layer.4.k_cache 0.00061006 0.12986920 + layer.4.v_cache 0.00005429 0.01507978 + layer.4.output 0.17891866 623.86564626 + ------------------------------------------------------------------------------------- + TOTAL 0.08858556 258.90016438 + (elements=731,136) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 731136 +Total Bytes 163400 +BPFP 1.7879 bits/point +EBPFP 3.5758 equivalent bits/point +MSE 258.900164 +---------------------- -------------------------------------------------------- +Time: 0.550s Load: 0.006s, Pack+Encode: 0.251s, Decode+Unpack: 0.293s +---------------------- -------------------------------------------------------- +💾 Converting with 258.9002 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-96.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-96.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 81, 128) +Output shape: (1, 81, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) -> torch.Size([1, 1, 81, 512]) + layer.4.output: torch.Size([1, 81, 3584]) -> torch.Size([1, 1, 81, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 5,032B, BPFP=0.9707 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 15,988B, BPFP=3.0841 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 8,924B, BPFP=1.7215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 15,520B, BPFP=2.9938 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 10,180B, BPFP=1.9637 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 15,048B, BPFP=2.9028 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 9,536B, BPFP=1.8395 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 15,216B, BPFP=2.9352 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 13,160B, BPFP=2.5386 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 14,932B, BPFP=2.8804 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 33,320B, BPFP=0.9182 +⌛️ [2/4] FRONTEND: Frontend time: 0.249s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.294s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 81, 128]) + layer.0.v_cache: torch.Size([1, 4, 81, 128]) + layer.1.k_cache: torch.Size([1, 4, 81, 128]) + layer.1.v_cache: torch.Size([1, 4, 81, 128]) + layer.2.k_cache: torch.Size([1, 4, 81, 128]) + layer.2.v_cache: torch.Size([1, 4, 81, 128]) + layer.3.k_cache: torch.Size([1, 4, 81, 128]) + layer.3.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.k_cache: torch.Size([1, 4, 81, 128]) + layer.4.v_cache: torch.Size([1, 4, 81, 128]) + layer.4.output: torch.Size([1, 81, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10837562 29.52229215 + layer.0.v_cache 0.00001644 0.00575843 + layer.1.k_cache 0.01241452 3.45697813 + layer.1.v_cache 0.00000532 0.00246130 + layer.2.k_cache 0.01724591 0.51528978 + layer.2.v_cache 0.00001958 0.00605689 + layer.3.k_cache 0.04641535 2.35659809 + layer.3.v_cache 0.00001847 0.00637716 + layer.4.k_cache 0.00063678 0.12044974 + layer.4.v_cache 0.00004974 0.01288507 + layer.4.output 0.18014439 661.53058862 + ------------------------------------------------------------------------------------- + TOTAL 0.08507109 274.51289806 + (elements=705,024) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 705024 +Total Bytes 156856 +BPFP 1.7799 bits/point +EBPFP 3.5597 equivalent bits/point +MSE 274.512898 +---------------------- -------------------------------------------------------- +Time: 0.549s Load: 0.005s, Pack+Encode: 0.249s, Decode+Unpack: 0.294s +---------------------- -------------------------------------------------------- +💾 Converting with 274.5129 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/advbench/AdvBench-advbench-99.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/advbench/AdvBench-advbench-99.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.7751 bits/point +Avg EBPFP 3.5501 equivalent bits/point +Avg MSE 255.766668 +Avg Time 0.566s +------------------------ ---------------------------- diff --git a/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..08a82375fe09729d6e8c39925be4660f080f982a --- /dev/null +++ b/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.02_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.02_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 598 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.02_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean +Output output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,800B, BPFP=0.8419 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,652B, BPFP=2.7577 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,928B, BPFP=1.4932 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,532B, BPFP=2.6778 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,784B, BPFP=1.7683 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,032B, BPFP=2.6421 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,408B, BPFP=1.6701 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,428B, BPFP=2.6704 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,408B, BPFP=2.3122 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,152B, BPFP=2.5793 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,868B, BPFP=0.9262 +⌛️ [2/4] FRONTEND: Frontend time: 0.712s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.466s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12261864 24.34805981 + layer.0.v_cache 0.00001806 0.00667518 + layer.1.k_cache 0.30198812 3.08635973 + layer.1.v_cache 0.00000607 0.00251069 + layer.2.k_cache 0.01966049 0.47461896 + layer.2.v_cache 0.00002021 0.00666631 + layer.3.k_cache 0.01413035 1.86619867 + layer.3.v_cache 0.00002054 0.00737207 + layer.4.k_cache 0.00067868 0.13394499 + layer.4.v_cache 0.00004987 0.01371033 + layer.4.output 1.39792533 235.37589693 + ------------------------------------------------------------------------------------- + TOTAL 0.60262755 98.68102325 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 390992 +BPFP 1.6409 bits/point +EBPFP 3.2819 equivalent bits/point +MSE 98.681023 +---------------------- -------------------------------------------------------- +Time: 1.186s Load: 0.009s, Pack+Encode: 0.712s, Decode+Unpack: 0.466s +---------------------- -------------------------------------------------------- +💾 Converting with 98.6810 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1028.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,700B, BPFP=0.8464 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,000B, BPFP=2.7488 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,452B, BPFP=1.4795 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,168B, BPFP=2.6887 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,388B, BPFP=1.7642 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,700B, BPFP=2.6548 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,060B, BPFP=1.6681 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,204B, BPFP=2.6913 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,276B, BPFP=2.3348 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,816B, BPFP=2.5909 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,424B, BPFP=0.9138 +⌛️ [2/4] FRONTEND: Frontend time: 0.331s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.446s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11787021 22.98466887 + layer.0.v_cache 0.00001738 0.00649901 + layer.1.k_cache 0.25265321 3.26298127 + layer.1.v_cache 0.00000586 0.00261953 + layer.2.k_cache 0.01010228 0.45678584 + layer.2.v_cache 0.00002017 0.00715370 + layer.3.k_cache 0.01159736 2.01444131 + layer.3.v_cache 0.00002197 0.00787341 + layer.4.k_cache 0.00067244 0.14026423 + layer.4.v_cache 0.00004845 0.01391887 + layer.4.output 1.41728832 246.31444279 + ------------------------------------------------------------------------------------- + TOTAL 0.60670750 103.12342974 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 385188 +BPFP 1.6390 bits/point +EBPFP 3.2781 equivalent bits/point +MSE 103.123430 +---------------------- -------------------------------------------------------- +Time: 0.785s Load: 0.009s, Pack+Encode: 0.331s, Decode+Unpack: 0.446s +---------------------- -------------------------------------------------------- +💾 Converting with 103.1234 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-105.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-105.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,056B, BPFP=0.8447 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,836B, BPFP=2.7211 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,152B, BPFP=1.4821 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,056B, BPFP=2.6665 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,012B, BPFP=1.7525 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,340B, BPFP=2.6163 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,792B, BPFP=1.6670 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,900B, BPFP=2.6555 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,944B, BPFP=2.3083 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,572B, BPFP=2.5625 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 87,744B, BPFP=0.8783 +⌛️ [2/4] FRONTEND: Frontend time: 0.298s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.434s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13311632 22.50916567 + layer.0.v_cache 0.00001526 0.00636034 + layer.1.k_cache 0.36474110 3.15377602 + layer.1.v_cache 0.00000573 0.00258361 + layer.2.k_cache 0.01296863 0.45695147 + layer.2.v_cache 0.00001984 0.00690848 + layer.3.k_cache 0.01490937 1.88924186 + layer.3.v_cache 0.00002017 0.00795245 + layer.4.k_cache 0.00067370 0.13751634 + layer.4.v_cache 0.00005228 0.01357278 + layer.4.output 1.37278975 237.90578956 + ------------------------------------------------------------------------------------- + TOTAL 0.59623828 99.61909152 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 391404 +BPFP 1.6132 bits/point +EBPFP 3.2264 equivalent bits/point +MSE 99.619092 +---------------------- -------------------------------------------------------- +Time: 0.738s Load: 0.007s, Pack+Encode: 0.298s, Decode+Unpack: 0.434s +---------------------- -------------------------------------------------------- +💾 Converting with 99.6191 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1065.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,844B, BPFP=0.8225 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,512B, BPFP=2.6583 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,604B, BPFP=1.4475 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,644B, BPFP=2.6027 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,744B, BPFP=1.7126 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,008B, BPFP=2.5620 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,356B, BPFP=1.6237 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,652B, BPFP=2.6032 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,104B, BPFP=2.2480 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,256B, BPFP=2.5138 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,620B, BPFP=0.8839 +⌛️ [2/4] FRONTEND: Frontend time: 0.372s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.459s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11263794 22.48426294 + layer.0.v_cache 0.00001639 0.00638497 + layer.1.k_cache 0.42101050 3.12272669 + layer.1.v_cache 0.00000577 0.00250241 + layer.2.k_cache 0.00499437 0.41693021 + layer.2.v_cache 0.00001973 0.00672775 + layer.3.k_cache 0.03029772 1.77009207 + layer.3.v_cache 0.00002073 0.00773143 + layer.4.k_cache 0.00068836 0.13665670 + layer.4.v_cache 0.00005112 0.01359863 + layer.4.output 1.25471843 213.34135319 + ------------------------------------------------------------------------------------- + TOTAL 0.55016304 89.49159330 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 421344 +BPFP 1.5872 bits/point +EBPFP 3.1743 equivalent bits/point +MSE 89.491593 +---------------------- -------------------------------------------------------- +Time: 0.840s Load: 0.009s, Pack+Encode: 0.372s, Decode+Unpack: 0.459s +---------------------- -------------------------------------------------------- +💾 Converting with 89.4916 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-111.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-111.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,768B, BPFP=0.8435 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,460B, BPFP=2.7566 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,792B, BPFP=1.4903 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,436B, BPFP=2.6832 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,732B, BPFP=1.7726 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,880B, BPFP=2.6433 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,140B, BPFP=1.6585 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,512B, BPFP=2.6886 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,456B, BPFP=2.3263 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,212B, BPFP=2.5955 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,672B, BPFP=0.8875 +⌛️ [2/4] FRONTEND: Frontend time: 0.335s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.464s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10233516 23.36167754 + layer.0.v_cache 0.00001822 0.00650129 + layer.1.k_cache 0.31275415 2.85420493 + layer.1.v_cache 0.00000587 0.00247425 + layer.2.k_cache 0.01225604 0.45263714 + layer.2.v_cache 0.00002018 0.00697185 + layer.3.k_cache 0.02847941 1.85244233 + layer.3.v_cache 0.00002101 0.00744186 + layer.4.k_cache 0.00067306 0.13713630 + layer.4.v_cache 0.00005947 0.01348819 + layer.4.output 1.40430263 243.76429391 + ------------------------------------------------------------------------------------- + TOTAL 0.60510241 102.06147253 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 386060 +BPFP 1.6277 bits/point +EBPFP 3.2554 equivalent bits/point +MSE 102.061473 +---------------------- -------------------------------------------------------- +Time: 0.806s Load: 0.007s, Pack+Encode: 0.335s, Decode+Unpack: 0.464s +---------------------- -------------------------------------------------------- +💾 Converting with 102.0615 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 294, 128) +Output shape: (1, 294, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) -> torch.Size([1, 1, 294, 512]) + layer.4.output: torch.Size([1, 294, 3584]) -> torch.Size([1, 1, 294, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,344B, BPFP=0.8155 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 50,976B, BPFP=2.7092 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,504B, BPFP=1.4617 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 49,856B, BPFP=2.6497 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,456B, BPFP=1.7249 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 48,832B, BPFP=2.5952 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,608B, BPFP=1.6267 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 49,604B, BPFP=2.6363 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,548B, BPFP=2.2613 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 47,552B, BPFP=2.5272 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 113,664B, BPFP=0.8630 +⌛️ [2/4] FRONTEND: Frontend time: 0.433s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.511s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 294, 128]) + layer.0.v_cache: torch.Size([1, 4, 294, 128]) + layer.1.k_cache: torch.Size([1, 4, 294, 128]) + layer.1.v_cache: torch.Size([1, 4, 294, 128]) + layer.2.k_cache: torch.Size([1, 4, 294, 128]) + layer.2.v_cache: torch.Size([1, 4, 294, 128]) + layer.3.k_cache: torch.Size([1, 4, 294, 128]) + layer.3.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.k_cache: torch.Size([1, 4, 294, 128]) + layer.4.v_cache: torch.Size([1, 4, 294, 128]) + layer.4.output: torch.Size([1, 294, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16298727 23.09707496 + layer.0.v_cache 0.00001770 0.00636273 + layer.1.k_cache 0.61736183 3.18756270 + layer.1.v_cache 0.00000612 0.00248250 + layer.2.k_cache 0.02593859 0.43708563 + layer.2.v_cache 0.00002024 0.00652622 + layer.3.k_cache 0.01326758 1.94391866 + layer.3.v_cache 0.00002270 0.00763954 + layer.4.k_cache 0.00075427 0.13673782 + layer.4.v_cache 0.00005043 0.01270991 + layer.4.output 0.04538776 183.07492104 + ------------------------------------------------------------------------------------- + TOTAL 0.06694947 77.08014988 + (elements=2,558,976) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2558976 +Total Bytes 508944 +BPFP 1.5911 bits/point +EBPFP 3.1822 equivalent bits/point +MSE 77.080150 +---------------------- -------------------------------------------------------- +Time: 0.955s Load: 0.011s, Pack+Encode: 0.433s, Decode+Unpack: 0.511s +---------------------- -------------------------------------------------------- +💾 Converting with 77.0801 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1247.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 239, 128) +Output shape: (1, 239, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) -> torch.Size([1, 1, 239, 512]) + layer.4.output: torch.Size([1, 239, 3584]) -> torch.Size([1, 1, 239, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,784B, BPFP=0.8358 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,348B, BPFP=2.7032 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,544B, BPFP=1.4738 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,404B, BPFP=2.6415 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,456B, BPFP=1.7296 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,680B, BPFP=2.5941 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,084B, BPFP=1.6399 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,220B, BPFP=2.6294 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,616B, BPFP=2.2631 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,724B, BPFP=2.5316 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,304B, BPFP=0.9461 +⌛️ [2/4] FRONTEND: Frontend time: 0.308s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.409s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 239, 128]) + layer.0.v_cache: torch.Size([1, 4, 239, 128]) + layer.1.k_cache: torch.Size([1, 4, 239, 128]) + layer.1.v_cache: torch.Size([1, 4, 239, 128]) + layer.2.k_cache: torch.Size([1, 4, 239, 128]) + layer.2.v_cache: torch.Size([1, 4, 239, 128]) + layer.3.k_cache: torch.Size([1, 4, 239, 128]) + layer.3.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.k_cache: torch.Size([1, 4, 239, 128]) + layer.4.v_cache: torch.Size([1, 4, 239, 128]) + layer.4.output: torch.Size([1, 239, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11863919 22.72145904 + layer.0.v_cache 0.00001702 0.00667444 + layer.1.k_cache 0.43747612 3.12528704 + layer.1.v_cache 0.00000797 0.00264075 + layer.2.k_cache 0.02286501 0.44261782 + layer.2.v_cache 0.00002174 0.00644590 + layer.3.k_cache 0.01903599 2.00053144 + layer.3.v_cache 0.00002123 0.00750473 + layer.4.k_cache 0.00068597 0.13558020 + layer.4.v_cache 0.00006921 0.01320571 + layer.4.output 1.28101462 221.43443664 + ------------------------------------------------------------------------------------- + TOTAL 0.56270246 92.85311786 + (elements=2,080,256) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2080256 +Total Bytes 423164 +BPFP 1.6274 bits/point +EBPFP 3.2547 equivalent bits/point +MSE 92.853118 +---------------------- -------------------------------------------------------- +Time: 0.726s Load: 0.009s, Pack+Encode: 0.308s, Decode+Unpack: 0.409s +---------------------- -------------------------------------------------------- +💾 Converting with 92.8531 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1250.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 263, 128) +Output shape: (1, 263, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) -> torch.Size([1, 1, 263, 512]) + layer.4.output: torch.Size([1, 263, 3584]) -> torch.Size([1, 1, 263, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,976B, BPFP=0.8303 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,244B, BPFP=2.8068 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,896B, BPFP=1.4791 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,092B, BPFP=2.7384 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,720B, BPFP=1.7657 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,252B, BPFP=2.6885 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,704B, BPFP=1.6459 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,864B, BPFP=2.7248 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,680B, BPFP=2.3574 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,512B, BPFP=2.6445 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 107,572B, BPFP=0.9130 +⌛️ [2/4] FRONTEND: Frontend time: 0.339s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.486s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 263, 128]) + layer.0.v_cache: torch.Size([1, 4, 263, 128]) + layer.1.k_cache: torch.Size([1, 4, 263, 128]) + layer.1.v_cache: torch.Size([1, 4, 263, 128]) + layer.2.k_cache: torch.Size([1, 4, 263, 128]) + layer.2.v_cache: torch.Size([1, 4, 263, 128]) + layer.3.k_cache: torch.Size([1, 4, 263, 128]) + layer.3.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.k_cache: torch.Size([1, 4, 263, 128]) + layer.4.v_cache: torch.Size([1, 4, 263, 128]) + layer.4.output: torch.Size([1, 263, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15151563 22.56900175 + layer.0.v_cache 0.00001631 0.00645037 + layer.1.k_cache 0.46755480 3.19757475 + layer.1.v_cache 0.00000667 0.00252865 + layer.2.k_cache 0.02171374 0.44953202 + layer.2.v_cache 0.00002115 0.00656327 + layer.3.k_cache 0.01385514 1.88778698 + layer.3.v_cache 0.00002107 0.00737387 + layer.4.k_cache 0.00069081 0.13601889 + layer.4.v_cache 0.00005638 0.01368278 + layer.4.output 0.00504555 207.05672868 + ------------------------------------------------------------------------------------- + TOTAL 0.04063356 86.92197730 + (elements=2,289,152) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2289152 +Total Bytes 472512 +BPFP 1.6513 bits/point +EBPFP 3.3026 equivalent bits/point +MSE 86.921977 +---------------------- -------------------------------------------------------- +Time: 0.834s Load: 0.009s, Pack+Encode: 0.339s, Decode+Unpack: 0.486s +---------------------- -------------------------------------------------------- +💾 Converting with 86.9220 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1263.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,564B, BPFP=0.8031 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,164B, BPFP=2.7197 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,932B, BPFP=1.4536 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,284B, BPFP=2.6586 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,652B, BPFP=1.7119 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,524B, BPFP=2.6058 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,320B, BPFP=1.6194 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,108B, BPFP=2.6464 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,480B, BPFP=2.2556 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,536B, BPFP=2.5372 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,804B, BPFP=0.9207 +⌛️ [2/4] FRONTEND: Frontend time: 0.303s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.391s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14529029 24.36172960 + layer.0.v_cache 0.00001764 0.00643666 + layer.1.k_cache 0.31089128 3.50991645 + layer.1.v_cache 0.00000630 0.00265162 + layer.2.k_cache 0.01355181 0.44858846 + layer.2.v_cache 0.00002032 0.00661166 + layer.3.k_cache 0.02426839 1.96410183 + layer.3.v_cache 0.00002043 0.00740712 + layer.4.k_cache 0.00068925 0.13310242 + layer.4.v_cache 0.00005165 0.01314903 + layer.4.output 1.36066019 236.81694444 + ------------------------------------------------------------------------------------- + TOTAL 0.58937816 99.30425329 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 395368 +BPFP 1.6151 bits/point +EBPFP 3.2301 equivalent bits/point +MSE 99.304253 +---------------------- -------------------------------------------------------- +Time: 0.702s Load: 0.009s, Pack+Encode: 0.303s, Decode+Unpack: 0.391s +---------------------- -------------------------------------------------------- +💾 Converting with 99.3043 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1269.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 262, 128) +Output shape: (1, 262, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) -> torch.Size([1, 1, 262, 512]) + layer.4.output: torch.Size([1, 262, 3584]) -> torch.Size([1, 1, 262, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,000B, BPFP=0.8349 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,120B, BPFP=2.8101 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,692B, BPFP=1.4726 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,960B, BPFP=2.7409 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,344B, BPFP=1.7500 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,164B, BPFP=2.6935 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,584B, BPFP=1.6450 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,968B, BPFP=2.7414 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,300B, BPFP=2.3438 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,460B, BPFP=2.6515 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 106,784B, BPFP=0.9098 +⌛️ [2/4] FRONTEND: Frontend time: 0.358s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.550s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 262, 128]) + layer.0.v_cache: torch.Size([1, 4, 262, 128]) + layer.1.k_cache: torch.Size([1, 4, 262, 128]) + layer.1.v_cache: torch.Size([1, 4, 262, 128]) + layer.2.k_cache: torch.Size([1, 4, 262, 128]) + layer.2.v_cache: torch.Size([1, 4, 262, 128]) + layer.3.k_cache: torch.Size([1, 4, 262, 128]) + layer.3.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.k_cache: torch.Size([1, 4, 262, 128]) + layer.4.v_cache: torch.Size([1, 4, 262, 128]) + layer.4.output: torch.Size([1, 262, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12917162 23.20577886 + layer.0.v_cache 0.00001830 0.00625896 + layer.1.k_cache 0.40634074 3.27389689 + layer.1.v_cache 0.00000633 0.00249795 + layer.2.k_cache 0.02363680 0.46460546 + layer.2.v_cache 0.00002248 0.00661483 + layer.3.k_cache 0.02343226 1.87886292 + layer.3.v_cache 0.00002042 0.00731729 + layer.4.k_cache 0.00071170 0.13613602 + layer.4.v_cache 0.00005093 0.01334944 + layer.4.output 0.00505940 210.19441794 + ------------------------------------------------------------------------------------- + TOTAL 0.03640161 88.25624966 + (elements=2,280,448) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2280448 +Total Bytes 470376 +BPFP 1.6501 bits/point +EBPFP 3.3002 equivalent bits/point +MSE 88.256250 +---------------------- -------------------------------------------------------- +Time: 0.917s Load: 0.009s, Pack+Encode: 0.358s, Decode+Unpack: 0.550s +---------------------- -------------------------------------------------------- +💾 Converting with 88.2562 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1274.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,572B, BPFP=0.8036 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,208B, BPFP=2.7228 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,936B, BPFP=1.4539 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,284B, BPFP=2.6586 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,852B, BPFP=1.7258 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,568B, BPFP=2.6089 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,580B, BPFP=1.6375 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,100B, BPFP=2.6458 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,856B, BPFP=2.2817 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,764B, BPFP=2.5531 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,284B, BPFP=0.9254 +⌛️ [2/4] FRONTEND: Frontend time: 0.317s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.464s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18074451 24.65373264 + layer.0.v_cache 0.00001764 0.00665227 + layer.1.k_cache 0.33421082 3.21772515 + layer.1.v_cache 0.00000655 0.00259708 + layer.2.k_cache 0.01487562 0.45024041 + layer.2.v_cache 0.00002082 0.00669729 + layer.3.k_cache 0.01319740 1.91024048 + layer.3.v_cache 0.00002096 0.00766633 + layer.4.k_cache 0.00067848 0.13860908 + layer.4.v_cache 0.00005239 0.01341050 + layer.4.output 1.36066546 236.87365079 + ------------------------------------------------------------------------------------- + TOTAL 0.59226373 99.32488981 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 397004 +BPFP 1.6217 bits/point +EBPFP 3.2435 equivalent bits/point +MSE 99.324890 +---------------------- -------------------------------------------------------- +Time: 0.789s Load: 0.008s, Pack+Encode: 0.317s, Decode+Unpack: 0.464s +---------------------- -------------------------------------------------------- +💾 Converting with 99.3249 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1276.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 232, 128) +Output shape: (1, 232, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) -> torch.Size([1, 1, 232, 512]) + layer.4.output: torch.Size([1, 232, 3584]) -> torch.Size([1, 1, 232, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,736B, BPFP=0.8578 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,520B, BPFP=2.7290 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,176B, BPFP=1.4935 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,524B, BPFP=2.6619 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,016B, BPFP=1.7522 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,964B, BPFP=2.6242 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,524B, BPFP=1.6517 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,492B, BPFP=2.6598 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,236B, BPFP=2.3058 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,268B, BPFP=2.5773 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,852B, BPFP=0.9126 +⌛️ [2/4] FRONTEND: Frontend time: 0.316s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.401s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 232, 128]) + layer.0.v_cache: torch.Size([1, 4, 232, 128]) + layer.1.k_cache: torch.Size([1, 4, 232, 128]) + layer.1.v_cache: torch.Size([1, 4, 232, 128]) + layer.2.k_cache: torch.Size([1, 4, 232, 128]) + layer.2.v_cache: torch.Size([1, 4, 232, 128]) + layer.3.k_cache: torch.Size([1, 4, 232, 128]) + layer.3.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.k_cache: torch.Size([1, 4, 232, 128]) + layer.4.v_cache: torch.Size([1, 4, 232, 128]) + layer.4.output: torch.Size([1, 232, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11978169 23.44529987 + layer.0.v_cache 0.00001746 0.00643584 + layer.1.k_cache 0.32299256 2.92560156 + layer.1.v_cache 0.00000596 0.00241717 + layer.2.k_cache 0.02091339 0.46617889 + layer.2.v_cache 0.00002092 0.00655971 + layer.3.k_cache 0.01769928 2.02429094 + layer.3.v_cache 0.00002180 0.00738983 + layer.4.k_cache 0.00071208 0.13460013 + layer.4.v_cache 0.00005763 0.01353458 + layer.4.output 1.31963615 228.99994227 + ------------------------------------------------------------------------------------- + TOTAL 0.57174564 96.00187673 + (elements=2,019,328) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2019328 +Total Bytes 411308 +BPFP 1.6295 bits/point +EBPFP 3.2590 equivalent bits/point +MSE 96.001877 +---------------------- -------------------------------------------------------- +Time: 0.725s Load: 0.008s, Pack+Encode: 0.316s, Decode+Unpack: 0.401s +---------------------- -------------------------------------------------------- +💾 Converting with 96.0019 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1277.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,144B, BPFP=0.8625 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,840B, BPFP=2.7585 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,296B, BPFP=1.5125 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,888B, BPFP=2.6909 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,992B, BPFP=1.7750 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,152B, BPFP=2.6386 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,704B, BPFP=1.6835 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,688B, BPFP=2.6767 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,728B, BPFP=2.3244 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,496B, BPFP=2.5920 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 97,076B, BPFP=0.9849 +⌛️ [2/4] FRONTEND: Frontend time: 0.337s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.406s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15639258 24.21873668 + layer.0.v_cache 0.00001700 0.00669606 + layer.1.k_cache 0.33150226 3.29109053 + layer.1.v_cache 0.00000613 0.00253866 + layer.2.k_cache 0.01694729 0.43568396 + layer.2.v_cache 0.00002208 0.00674062 + layer.3.k_cache 0.03858858 1.89555692 + layer.3.v_cache 0.00002090 0.00784678 + layer.4.k_cache 0.00067984 0.13532006 + layer.4.v_cache 0.00005217 0.01392253 + layer.4.output 1.39161012 233.93309659 + ------------------------------------------------------------------------------------- + TOTAL 0.60502939 98.09092994 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 400004 +BPFP 1.6711 bits/point +EBPFP 3.3423 equivalent bits/point +MSE 98.090930 +---------------------- -------------------------------------------------------- +Time: 0.751s Load: 0.008s, Pack+Encode: 0.337s, Decode+Unpack: 0.406s +---------------------- -------------------------------------------------------- +💾 Converting with 98.0909 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1283.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,584B, BPFP=0.8347 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,012B, BPFP=2.7479 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,384B, BPFP=1.4528 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,828B, BPFP=2.6802 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,268B, BPFP=1.7324 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,144B, BPFP=2.6410 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,612B, BPFP=1.6376 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,540B, BPFP=2.6637 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,220B, BPFP=2.3020 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,016B, BPFP=2.5765 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 112,688B, BPFP=0.9214 +⌛️ [2/4] FRONTEND: Frontend time: 0.337s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.505s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13605902 23.67098071 + layer.0.v_cache 0.00001757 0.00635598 + layer.1.k_cache 0.58683352 3.05741910 + layer.1.v_cache 0.00000623 0.00255106 + layer.2.k_cache 0.01500385 0.43577953 + layer.2.v_cache 0.00002229 0.00678072 + layer.3.k_cache 0.00978726 1.96315377 + layer.3.v_cache 0.00001998 0.00721439 + layer.4.k_cache 0.00067570 0.13470656 + layer.4.v_cache 0.00005472 0.01366136 + layer.4.output 0.00490606 198.89217033 + ------------------------------------------------------------------------------------- + TOTAL 0.04604839 83.62022326 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 484296 +BPFP 1.6305 bits/point +EBPFP 3.2610 equivalent bits/point +MSE 83.620223 +---------------------- -------------------------------------------------------- +Time: 0.851s Load: 0.010s, Pack+Encode: 0.337s, Decode+Unpack: 0.505s +---------------------- -------------------------------------------------------- +💾 Converting with 83.6202 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,036B, BPFP=0.7988 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,820B, BPFP=2.5625 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,396B, BPFP=1.3723 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,840B, BPFP=2.5025 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,700B, BPFP=1.6360 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,120B, BPFP=2.4583 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,276B, BPFP=1.5488 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,724B, BPFP=2.4953 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,196B, BPFP=2.1566 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,200B, BPFP=2.4020 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 103,560B, BPFP=0.9065 +⌛️ [2/4] FRONTEND: Frontend time: 0.276s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.421s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12967140 21.59784773 + layer.0.v_cache 0.00001601 0.00636360 + layer.1.k_cache 0.49353925 2.89525649 + layer.1.v_cache 0.00000620 0.00249650 + layer.2.k_cache 0.03189284 0.45712825 + layer.2.v_cache 0.00002040 0.00648488 + layer.3.k_cache 0.04285583 1.92707196 + layer.3.v_cache 0.00002221 0.00742602 + layer.4.k_cache 0.00071132 0.13366286 + layer.4.v_cache 0.00005107 0.01294003 + layer.4.output 1.20063613 177.59924720 + ------------------------------------------------------------------------------------- + TOTAL 0.53548467 74.72008287 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 428868 +BPFP 1.5458 bits/point +EBPFP 3.0916 equivalent bits/point +MSE 74.720083 +---------------------- -------------------------------------------------------- +Time: 0.705s Load: 0.008s, Pack+Encode: 0.276s, Decode+Unpack: 0.421s +---------------------- -------------------------------------------------------- +💾 Converting with 74.7201 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,828B, BPFP=0.8050 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,708B, BPFP=2.6172 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,860B, BPFP=1.4345 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,836B, BPFP=2.5625 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,940B, BPFP=1.6905 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,144B, BPFP=2.5191 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,556B, BPFP=1.6037 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,644B, BPFP=2.5505 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,280B, BPFP=2.2139 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,120B, BPFP=2.4548 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,464B, BPFP=0.8558 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.443s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16529650 22.83679640 + layer.0.v_cache 0.00001613 0.00654225 + layer.1.k_cache 0.54577502 3.19205337 + layer.1.v_cache 0.00000662 0.00265641 + layer.2.k_cache 0.01453520 0.45474283 + layer.2.v_cache 0.00002096 0.00697346 + layer.3.k_cache 0.02409829 1.70057966 + layer.3.v_cache 0.00002071 0.00763343 + layer.4.k_cache 0.00068699 0.14213352 + layer.4.v_cache 0.00004971 0.01373938 + layer.4.output 1.22953407 213.26595668 + ------------------------------------------------------------------------------------- + TOTAL 0.55042615 89.48385573 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 421380 +BPFP 1.5554 bits/point +EBPFP 3.1108 equivalent bits/point +MSE 89.483856 +---------------------- -------------------------------------------------------- +Time: 0.770s Load: 0.008s, Pack+Encode: 0.319s, Decode+Unpack: 0.443s +---------------------- -------------------------------------------------------- +💾 Converting with 89.4839 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1295.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 327, 128) +Output shape: (1, 327, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) -> torch.Size([1, 1, 327, 512]) + layer.4.output: torch.Size([1, 327, 3584]) -> torch.Size([1, 1, 327, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,980B, BPFP=0.8114 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,712B, BPFP=2.7576 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,424B, BPFP=1.4537 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,368B, BPFP=2.6934 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 36,260B, BPFP=1.7326 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,440B, BPFP=2.6491 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 33,932B, BPFP=1.6214 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,016B, BPFP=2.6766 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 48,148B, BPFP=2.3006 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,124B, BPFP=2.5862 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 134,900B, BPFP=0.9208 +⌛️ [2/4] FRONTEND: Frontend time: 0.455s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.635s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 327, 128]) + layer.0.v_cache: torch.Size([1, 4, 327, 128]) + layer.1.k_cache: torch.Size([1, 4, 327, 128]) + layer.1.v_cache: torch.Size([1, 4, 327, 128]) + layer.2.k_cache: torch.Size([1, 4, 327, 128]) + layer.2.v_cache: torch.Size([1, 4, 327, 128]) + layer.3.k_cache: torch.Size([1, 4, 327, 128]) + layer.3.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.k_cache: torch.Size([1, 4, 327, 128]) + layer.4.v_cache: torch.Size([1, 4, 327, 128]) + layer.4.output: torch.Size([1, 327, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.19043710 22.54929251 + layer.0.v_cache 0.00001603 0.00651115 + layer.1.k_cache 0.64205214 3.05772227 + layer.1.v_cache 0.00000630 0.00263525 + layer.2.k_cache 0.01513053 0.42309738 + layer.2.v_cache 0.00002191 0.00707804 + layer.3.k_cache 0.02775778 1.80306568 + layer.3.v_cache 0.00002125 0.00759737 + layer.4.k_cache 0.00071623 0.13362556 + layer.4.v_cache 0.00005146 0.01344833 + layer.4.output 0.04089398 162.91029107 + ------------------------------------------------------------------------------------- + TOTAL 0.06838051 68.72800653 + (elements=2,846,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2846208 +Total Bytes 580304 +BPFP 1.6311 bits/point +EBPFP 3.2622 equivalent bits/point +MSE 68.728007 +---------------------- -------------------------------------------------------- +Time: 1.101s Load: 0.011s, Pack+Encode: 0.455s, Decode+Unpack: 0.635s +---------------------- -------------------------------------------------------- +💾 Converting with 68.7280 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1302.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,296B, BPFP=0.8445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,116B, BPFP=2.7118 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,600B, BPFP=1.4686 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,020B, BPFP=2.6513 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,616B, BPFP=1.7456 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,152B, BPFP=2.6034 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,608B, BPFP=1.6347 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,724B, BPFP=2.6349 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,204B, BPFP=2.2750 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,956B, BPFP=2.5373 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 118,588B, BPFP=0.9354 +⌛️ [2/4] FRONTEND: Frontend time: 0.351s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.511s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14392723 22.06032430 + layer.0.v_cache 0.00001682 0.00641137 + layer.1.k_cache 0.53349789 3.00944341 + layer.1.v_cache 0.00000638 0.00253897 + layer.2.k_cache 0.01523476 0.45330681 + layer.2.v_cache 0.00002109 0.00646155 + layer.3.k_cache 0.03867486 1.90654396 + layer.3.v_cache 0.00001994 0.00727672 + layer.4.k_cache 0.00070525 0.13161649 + layer.4.v_cache 0.00005078 0.01275103 + layer.4.output 0.00472849 189.81813163 + ------------------------------------------------------------------------------------- + TOTAL 0.04501497 79.78374094 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 500880 +BPFP 1.6267 bits/point +EBPFP 3.2535 equivalent bits/point +MSE 79.783741 +---------------------- -------------------------------------------------------- +Time: 0.872s Load: 0.009s, Pack+Encode: 0.351s, Decode+Unpack: 0.511s +---------------------- -------------------------------------------------------- +💾 Converting with 79.7837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1317.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 225, 128) +Output shape: (1, 225, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) -> torch.Size([1, 1, 225, 512]) + layer.4.output: torch.Size([1, 225, 3584]) -> torch.Size([1, 1, 225, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,728B, BPFP=0.8144 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,204B, BPFP=2.7225 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,932B, BPFP=1.4536 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,244B, BPFP=2.6558 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,840B, BPFP=1.7250 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,600B, BPFP=2.6111 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,572B, BPFP=1.6369 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,204B, BPFP=2.6531 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,856B, BPFP=2.2817 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,720B, BPFP=2.5500 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,900B, BPFP=0.9216 +⌛️ [2/4] FRONTEND: Frontend time: 0.329s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.493s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 225, 128]) + layer.0.v_cache: torch.Size([1, 4, 225, 128]) + layer.1.k_cache: torch.Size([1, 4, 225, 128]) + layer.1.v_cache: torch.Size([1, 4, 225, 128]) + layer.2.k_cache: torch.Size([1, 4, 225, 128]) + layer.2.v_cache: torch.Size([1, 4, 225, 128]) + layer.3.k_cache: torch.Size([1, 4, 225, 128]) + layer.3.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.k_cache: torch.Size([1, 4, 225, 128]) + layer.4.v_cache: torch.Size([1, 4, 225, 128]) + layer.4.output: torch.Size([1, 225, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14673459 23.17338542 + layer.0.v_cache 0.00001655 0.00661279 + layer.1.k_cache 0.37581726 3.13873237 + layer.1.v_cache 0.00000603 0.00262454 + layer.2.k_cache 0.01062915 0.43894986 + layer.2.v_cache 0.00002006 0.00692951 + layer.3.k_cache 0.04027582 1.96403293 + layer.3.v_cache 0.00002066 0.00772653 + layer.4.k_cache 0.00066373 0.13635941 + layer.4.v_cache 0.00004861 0.01341118 + layer.4.output 1.36064577 236.75311508 + ------------------------------------------------------------------------------------- + TOTAL 0.59404429 99.18591589 + (elements=1,958,400) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1958400 +Total Bytes 396800 +BPFP 1.6209 bits/point +EBPFP 3.2418 equivalent bits/point +MSE 99.185916 +---------------------- -------------------------------------------------------- +Time: 0.828s Load: 0.007s, Pack+Encode: 0.329s, Decode+Unpack: 0.493s +---------------------- -------------------------------------------------------- +💾 Converting with 99.1859 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1320.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,552B, BPFP=0.8381 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,236B, BPFP=2.7535 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,524B, BPFP=1.5040 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,240B, BPFP=2.6870 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,636B, BPFP=1.7786 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,476B, BPFP=2.6360 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,948B, BPFP=1.6659 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,152B, BPFP=2.6811 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,552B, BPFP=2.3072 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,612B, BPFP=2.5783 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,732B, BPFP=0.9609 +⌛️ [2/4] FRONTEND: Frontend time: 0.333s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12385076 23.08965595 + layer.0.v_cache 0.00002004 0.00718083 + layer.1.k_cache 0.38583452 3.26340635 + layer.1.v_cache 0.00000646 0.00276637 + layer.2.k_cache 0.02500911 0.45921009 + layer.2.v_cache 0.00002014 0.00676756 + layer.3.k_cache 0.03425702 1.90293350 + layer.3.v_cache 0.00002057 0.00792934 + layer.4.k_cache 0.00068928 0.14424828 + layer.4.v_cache 0.00004882 0.01368524 + layer.4.output 1.30836928 225.48737027 + ------------------------------------------------------------------------------------- + TOTAL 0.57225539 94.54761032 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 421660 +BPFP 1.6562 bits/point +EBPFP 3.3124 equivalent bits/point +MSE 94.547610 +---------------------- -------------------------------------------------------- +Time: 0.780s Load: 0.008s, Pack+Encode: 0.333s, Decode+Unpack: 0.439s +---------------------- -------------------------------------------------------- +💾 Converting with 94.5476 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-137.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-137.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 247, 128) +Output shape: (1, 247, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) -> torch.Size([1, 1, 247, 512]) + layer.4.output: torch.Size([1, 247, 3584]) -> torch.Size([1, 1, 247, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,132B, BPFP=0.8307 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,812B, BPFP=2.6450 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,468B, BPFP=1.4213 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,696B, BPFP=2.5744 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,860B, BPFP=1.6991 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,912B, BPFP=2.5248 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,140B, BPFP=1.5903 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,424B, BPFP=2.5572 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,040B, BPFP=2.2166 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,960B, BPFP=2.4646 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,892B, BPFP=0.9027 +⌛️ [2/4] FRONTEND: Frontend time: 0.354s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.480s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 247, 128]) + layer.0.v_cache: torch.Size([1, 4, 247, 128]) + layer.1.k_cache: torch.Size([1, 4, 247, 128]) + layer.1.v_cache: torch.Size([1, 4, 247, 128]) + layer.2.k_cache: torch.Size([1, 4, 247, 128]) + layer.2.v_cache: torch.Size([1, 4, 247, 128]) + layer.3.k_cache: torch.Size([1, 4, 247, 128]) + layer.3.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.k_cache: torch.Size([1, 4, 247, 128]) + layer.4.v_cache: torch.Size([1, 4, 247, 128]) + layer.4.output: torch.Size([1, 247, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13907220 23.03660726 + layer.0.v_cache 0.00001735 0.00639240 + layer.1.k_cache 0.47208408 2.95992906 + layer.1.v_cache 0.00000611 0.00242029 + layer.2.k_cache 0.03408500 0.45493147 + layer.2.v_cache 0.00001989 0.00605771 + layer.3.k_cache 0.02387486 2.07090382 + layer.3.v_cache 0.00002031 0.00681688 + layer.4.k_cache 0.00069623 0.12924037 + layer.4.v_cache 0.00004922 0.01182537 + layer.4.output 1.23951819 214.58286943 + ------------------------------------------------------------------------------------- + TOTAL 0.54979721 90.04501239 + (elements=2,149,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2149888 +Total Bytes 424336 +BPFP 1.5790 bits/point +EBPFP 3.1580 equivalent bits/point +MSE 90.045012 +---------------------- -------------------------------------------------------- +Time: 0.842s Load: 0.009s, Pack+Encode: 0.354s, Decode+Unpack: 0.480s +---------------------- -------------------------------------------------------- +💾 Converting with 90.0450 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1389.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 427, 128) +Output shape: (1, 427, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) -> torch.Size([1, 1, 427, 512]) + layer.4.output: torch.Size([1, 427, 3584]) -> torch.Size([1, 1, 427, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 21,932B, BPFP=0.8025 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 72,032B, BPFP=2.6358 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 38,576B, BPFP=1.4116 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 70,308B, BPFP=2.5727 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 45,336B, BPFP=1.6590 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 68,564B, BPFP=2.5089 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 42,376B, BPFP=1.5506 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 69,492B, BPFP=2.5429 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 60,024B, BPFP=2.1964 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 67,244B, BPFP=2.4606 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 165,356B, BPFP=0.8644 +⌛️ [2/4] FRONTEND: Frontend time: 0.554s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.714s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 427, 128]) + layer.0.v_cache: torch.Size([1, 4, 427, 128]) + layer.1.k_cache: torch.Size([1, 4, 427, 128]) + layer.1.v_cache: torch.Size([1, 4, 427, 128]) + layer.2.k_cache: torch.Size([1, 4, 427, 128]) + layer.2.v_cache: torch.Size([1, 4, 427, 128]) + layer.3.k_cache: torch.Size([1, 4, 427, 128]) + layer.3.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.k_cache: torch.Size([1, 4, 427, 128]) + layer.4.v_cache: torch.Size([1, 4, 427, 128]) + layer.4.output: torch.Size([1, 427, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15738957 21.57370645 + layer.0.v_cache 0.00001693 0.00590917 + layer.1.k_cache 1.00627419 3.11480127 + layer.1.v_cache 0.00000663 0.00234656 + layer.2.k_cache 0.03590691 0.46102244 + layer.2.v_cache 0.00002108 0.00583423 + layer.3.k_cache 0.02176077 1.87412292 + layer.3.v_cache 0.00002004 0.00633752 + layer.4.k_cache 0.00078997 0.12658514 + layer.4.v_cache 0.00005212 0.01188777 + layer.4.output 0.00621442 124.34298678 + ------------------------------------------------------------------------------------- + TOTAL 0.07445524 52.79902712 + (elements=3,716,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3716608 +Total Bytes 721240 +BPFP 1.5525 bits/point +EBPFP 3.1049 equivalent bits/point +MSE 52.799027 +---------------------- -------------------------------------------------------- +Time: 1.283s Load: 0.014s, Pack+Encode: 0.554s, Decode+Unpack: 0.714s +---------------------- -------------------------------------------------------- +💾 Converting with 52.7990 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,292B, BPFP=0.8240 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,676B, BPFP=2.7488 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,208B, BPFP=1.4534 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,608B, BPFP=2.6873 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,728B, BPFP=1.7140 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,512B, BPFP=2.6241 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,836B, BPFP=1.6049 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,996B, BPFP=2.6520 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,628B, BPFP=2.2848 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,300B, BPFP=2.5542 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,828B, BPFP=0.8387 +⌛️ [2/4] FRONTEND: Frontend time: 0.395s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.576s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15504150 22.58418798 + layer.0.v_cache 0.00001776 0.00607764 + layer.1.k_cache 0.59434689 3.00220154 + layer.1.v_cache 0.00000612 0.00241055 + layer.2.k_cache 0.01687181 0.42913635 + layer.2.v_cache 0.00002031 0.00593928 + layer.3.k_cache 0.01346795 1.89376752 + layer.3.v_cache 0.00002023 0.00681171 + layer.4.k_cache 0.00071492 0.12505031 + layer.4.v_cache 0.00006004 0.01176399 + layer.4.output 0.00488892 200.43652807 + ------------------------------------------------------------------------------------- + TOTAL 0.04792882 84.18370843 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 468612 +BPFP 1.5893 bits/point +EBPFP 3.1787 equivalent bits/point +MSE 84.183708 +---------------------- -------------------------------------------------------- +Time: 0.983s Load: 0.011s, Pack+Encode: 0.395s, Decode+Unpack: 0.576s +---------------------- -------------------------------------------------------- +💾 Converting with 84.1837 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1442.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 361, 128) +Output shape: (1, 361, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) -> torch.Size([1, 1, 361, 512]) + layer.4.output: torch.Size([1, 361, 3584]) -> torch.Size([1, 1, 361, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 19,004B, BPFP=0.8225 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,888B, BPFP=2.6354 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 32,856B, BPFP=1.4221 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 59,468B, BPFP=2.5739 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 38,352B, BPFP=1.6600 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 58,368B, BPFP=2.5263 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 36,108B, BPFP=1.5628 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,868B, BPFP=2.5480 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 51,076B, BPFP=2.2107 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,900B, BPFP=2.4628 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 128,596B, BPFP=0.7951 +⌛️ [2/4] FRONTEND: Frontend time: 0.427s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.561s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 361, 128]) + layer.0.v_cache: torch.Size([1, 4, 361, 128]) + layer.1.k_cache: torch.Size([1, 4, 361, 128]) + layer.1.v_cache: torch.Size([1, 4, 361, 128]) + layer.2.k_cache: torch.Size([1, 4, 361, 128]) + layer.2.v_cache: torch.Size([1, 4, 361, 128]) + layer.3.k_cache: torch.Size([1, 4, 361, 128]) + layer.3.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.k_cache: torch.Size([1, 4, 361, 128]) + layer.4.v_cache: torch.Size([1, 4, 361, 128]) + layer.4.output: torch.Size([1, 361, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21057716 22.38987432 + layer.0.v_cache 0.00001554 0.00560524 + layer.1.k_cache 0.82271198 3.20414146 + layer.1.v_cache 0.00000592 0.00222266 + layer.2.k_cache 0.01184860 0.42234396 + layer.2.v_cache 0.00001987 0.00603098 + layer.3.k_cache 0.01345186 1.89241243 + layer.3.v_cache 0.00002111 0.00654647 + layer.4.k_cache 0.00076976 0.12864738 + layer.4.v_cache 0.00005077 0.01181346 + layer.4.output 0.03703259 146.49181342 + ------------------------------------------------------------------------------------- + TOTAL 0.07757063 61.97131366 + (elements=3,142,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3142144 +Total Bytes 600484 +BPFP 1.5289 bits/point +EBPFP 3.0577 equivalent bits/point +MSE 61.971314 +---------------------- -------------------------------------------------------- +Time: 0.999s Load: 0.011s, Pack+Encode: 0.427s, Decode+Unpack: 0.561s +---------------------- -------------------------------------------------------- +💾 Converting with 61.9713 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1452.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,076B, BPFP=0.8353 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,016B, BPFP=2.6605 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,884B, BPFP=1.4342 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,160B, BPFP=2.6130 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,536B, BPFP=1.6919 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,120B, BPFP=2.5554 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,800B, BPFP=1.5957 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,492B, BPFP=2.5760 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,704B, BPFP=2.2553 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,972B, BPFP=2.4918 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 108,572B, BPFP=0.8594 +⌛️ [2/4] FRONTEND: Frontend time: 0.361s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.479s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.21757852 22.92791099 + layer.0.v_cache 0.00001544 0.00545648 + layer.1.k_cache 0.59705023 2.95102773 + layer.1.v_cache 0.00000603 0.00223775 + layer.2.k_cache 0.00787334 0.39918472 + layer.2.v_cache 0.00001970 0.00584582 + layer.3.k_cache 0.01914395 1.95410611 + layer.3.v_cache 0.00001916 0.00624848 + layer.4.k_cache 0.00071861 0.12565890 + layer.4.v_cache 0.00005003 0.01204134 + layer.4.output 0.00471288 192.26644820 + ------------------------------------------------------------------------------------- + TOTAL 0.05149795 80.83852093 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 482332 +BPFP 1.5721 bits/point +EBPFP 3.1441 equivalent bits/point +MSE 80.838521 +---------------------- -------------------------------------------------------- +Time: 0.849s Load: 0.010s, Pack+Encode: 0.361s, Decode+Unpack: 0.479s +---------------------- -------------------------------------------------------- +💾 Converting with 80.8385 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1456.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 372, 128) +Output shape: (1, 372, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) -> torch.Size([1, 1, 372, 512]) + layer.4.output: torch.Size([1, 372, 3584]) -> torch.Size([1, 1, 372, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 18,956B, BPFP=0.7962 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 62,096B, BPFP=2.6082 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 33,456B, BPFP=1.4052 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 60,720B, BPFP=2.5504 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 39,324B, BPFP=1.6517 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 59,460B, BPFP=2.4975 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 36,924B, BPFP=1.5509 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 60,012B, BPFP=2.5207 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 51,808B, BPFP=2.1761 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 58,076B, BPFP=2.4393 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 145,700B, BPFP=0.8743 +⌛️ [2/4] FRONTEND: Frontend time: 0.386s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.562s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 372, 128]) + layer.0.v_cache: torch.Size([1, 4, 372, 128]) + layer.1.k_cache: torch.Size([1, 4, 372, 128]) + layer.1.v_cache: torch.Size([1, 4, 372, 128]) + layer.2.k_cache: torch.Size([1, 4, 372, 128]) + layer.2.v_cache: torch.Size([1, 4, 372, 128]) + layer.3.k_cache: torch.Size([1, 4, 372, 128]) + layer.3.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.k_cache: torch.Size([1, 4, 372, 128]) + layer.4.v_cache: torch.Size([1, 4, 372, 128]) + layer.4.output: torch.Size([1, 372, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18192377 21.17238822 + layer.0.v_cache 0.00001786 0.00579917 + layer.1.k_cache 0.82695811 3.06864060 + layer.1.v_cache 0.00000687 0.00238876 + layer.2.k_cache 0.01216373 0.40685740 + layer.2.v_cache 0.00002073 0.00604825 + layer.3.k_cache 0.01743993 1.92607166 + layer.3.v_cache 0.00002103 0.00652774 + layer.4.k_cache 0.00069676 0.12552942 + layer.4.v_cache 0.00005544 0.01208931 + layer.4.output 0.03606073 138.08333333 + ------------------------------------------------------------------------------------- + TOTAL 0.07598408 58.43033376 + (elements=3,237,888) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3237888 +Total Bytes 626532 +BPFP 1.5480 bits/point +EBPFP 3.0960 equivalent bits/point +MSE 58.430334 +---------------------- -------------------------------------------------------- +Time: 0.961s Load: 0.013s, Pack+Encode: 0.386s, Decode+Unpack: 0.562s +---------------------- -------------------------------------------------------- +💾 Converting with 58.4303 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1457.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,804B, BPFP=0.8129 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,600B, BPFP=2.7380 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,516B, BPFP=1.4278 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,332B, BPFP=2.6767 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 34,676B, BPFP=1.6774 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,176B, BPFP=2.6207 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 32,568B, BPFP=1.5755 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 54,856B, BPFP=2.6536 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,728B, BPFP=2.2604 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,868B, BPFP=2.5575 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 123,612B, BPFP=0.8542 +⌛️ [2/4] FRONTEND: Frontend time: 0.397s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.541s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16470559 21.87124341 + layer.0.v_cache 0.00001558 0.00591525 + layer.1.k_cache 0.70577214 3.08636947 + layer.1.v_cache 0.00000663 0.00234806 + layer.2.k_cache 0.01487649 0.43053363 + layer.2.v_cache 0.00002061 0.00602863 + layer.3.k_cache 0.03012228 1.87660902 + layer.3.v_cache 0.00002030 0.00661268 + layer.4.k_cache 0.00074575 0.12587731 + layer.4.v_cache 0.00005065 0.01157026 + layer.4.output 0.04139163 160.58382629 + ------------------------------------------------------------------------------------- + TOTAL 0.07094573 67.73587599 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 557736 +BPFP 1.5871 bits/point +EBPFP 3.1741 equivalent bits/point +MSE 67.735876 +---------------------- -------------------------------------------------------- +Time: 0.949s Load: 0.011s, Pack+Encode: 0.397s, Decode+Unpack: 0.541s +---------------------- -------------------------------------------------------- +💾 Converting with 67.7359 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1464.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,208B, BPFP=0.8377 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,092B, BPFP=2.7767 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,900B, BPFP=1.4682 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,968B, BPFP=2.7104 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,368B, BPFP=1.7316 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,136B, BPFP=2.6613 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,512B, BPFP=1.6222 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,656B, BPFP=2.6920 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,444B, BPFP=2.3257 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,140B, BPFP=2.6026 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 105,352B, BPFP=0.8874 +⌛️ [2/4] FRONTEND: Frontend time: 0.363s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.482s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16585089 22.75928471 + layer.0.v_cache 0.00001549 0.00590724 + layer.1.k_cache 0.52552974 3.06919567 + layer.1.v_cache 0.00000676 0.00238637 + layer.2.k_cache 0.01473994 0.44893206 + layer.2.v_cache 0.00002188 0.00634171 + layer.3.k_cache 0.05129043 1.89866863 + layer.3.v_cache 0.00002012 0.00687374 + layer.4.k_cache 0.00075129 0.12998148 + layer.4.v_cache 0.00005717 0.01229113 + layer.4.output 0.00501330 205.13273248 + ------------------------------------------------------------------------------------- + TOTAL 0.04666922 86.13347001 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 468776 +BPFP 1.6259 bits/point +EBPFP 3.2518 equivalent bits/point +MSE 86.133470 +---------------------- -------------------------------------------------------- +Time: 0.854s Load: 0.009s, Pack+Encode: 0.363s, Decode+Unpack: 0.482s +---------------------- -------------------------------------------------------- +💾 Converting with 86.1335 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1465.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 368, 128) +Output shape: (1, 368, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) -> torch.Size([1, 1, 368, 512]) + layer.4.output: torch.Size([1, 368, 3584]) -> torch.Size([1, 1, 368, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 19,440B, BPFP=0.8254 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 61,784B, BPFP=2.6233 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 33,416B, BPFP=1.4188 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 60,428B, BPFP=2.5657 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 39,252B, BPFP=1.6666 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 59,212B, BPFP=2.5141 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 36,872B, BPFP=1.5656 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 59,972B, BPFP=2.5464 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 51,616B, BPFP=2.1916 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 58,048B, BPFP=2.4647 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 137,232B, BPFP=0.8324 +⌛️ [2/4] FRONTEND: Frontend time: 0.380s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.545s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 368, 128]) + layer.0.v_cache: torch.Size([1, 4, 368, 128]) + layer.1.k_cache: torch.Size([1, 4, 368, 128]) + layer.1.v_cache: torch.Size([1, 4, 368, 128]) + layer.2.k_cache: torch.Size([1, 4, 368, 128]) + layer.2.v_cache: torch.Size([1, 4, 368, 128]) + layer.3.k_cache: torch.Size([1, 4, 368, 128]) + layer.3.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.k_cache: torch.Size([1, 4, 368, 128]) + layer.4.v_cache: torch.Size([1, 4, 368, 128]) + layer.4.output: torch.Size([1, 368, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14147143 22.22991147 + layer.0.v_cache 0.00001611 0.00570476 + layer.1.k_cache 0.89062865 3.18767382 + layer.1.v_cache 0.00000637 0.00225076 + layer.2.k_cache 0.02201829 0.44484740 + layer.2.v_cache 0.00002088 0.00596651 + layer.3.k_cache 0.00965744 1.82659332 + layer.3.v_cache 0.00002101 0.00662466 + layer.4.k_cache 0.00077426 0.12483810 + layer.4.v_cache 0.00005366 0.01244650 + layer.4.output 0.03638816 143.76891256 + ------------------------------------------------------------------------------------- + TOTAL 0.07761090 60.83701442 + (elements=3,203,072) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3203072 +Total Bytes 617272 +BPFP 1.5417 bits/point +EBPFP 3.0834 equivalent bits/point +MSE 60.837014 +---------------------- -------------------------------------------------------- +Time: 0.937s Load: 0.012s, Pack+Encode: 0.380s, Decode+Unpack: 0.545s +---------------------- -------------------------------------------------------- +💾 Converting with 60.8370 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 453, 128) +Output shape: (1, 453, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) -> torch.Size([1, 1, 453, 512]) + layer.4.output: torch.Size([1, 453, 3584]) -> torch.Size([1, 1, 453, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 23,724B, BPFP=0.8183 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 78,148B, BPFP=2.6955 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 40,820B, BPFP=1.4080 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 75,920B, BPFP=2.6187 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 48,456B, BPFP=1.6714 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 74,664B, BPFP=2.5753 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 45,328B, BPFP=1.5635 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 75,692B, BPFP=2.6108 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 64,836B, BPFP=2.2363 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 73,148B, BPFP=2.5230 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 168,268B, BPFP=0.8291 +⌛️ [2/4] FRONTEND: Frontend time: 0.584s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.778s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 453, 128]) + layer.0.v_cache: torch.Size([1, 4, 453, 128]) + layer.1.k_cache: torch.Size([1, 4, 453, 128]) + layer.1.v_cache: torch.Size([1, 4, 453, 128]) + layer.2.k_cache: torch.Size([1, 4, 453, 128]) + layer.2.v_cache: torch.Size([1, 4, 453, 128]) + layer.3.k_cache: torch.Size([1, 4, 453, 128]) + layer.3.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.k_cache: torch.Size([1, 4, 453, 128]) + layer.4.v_cache: torch.Size([1, 4, 453, 128]) + layer.4.output: torch.Size([1, 453, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13665763 21.95216353 + layer.0.v_cache 0.00001897 0.00608970 + layer.1.k_cache 1.07282356 3.08791218 + layer.1.v_cache 0.00000630 0.00230151 + layer.2.k_cache 0.02241544 0.43296534 + layer.2.v_cache 0.00001979 0.00598545 + layer.3.k_cache 0.01806638 1.84743030 + layer.3.v_cache 0.00002079 0.00703569 + layer.4.k_cache 0.00076567 0.13050015 + layer.4.v_cache 0.00005104 0.01246027 + layer.4.output 0.00584071 121.55027003 + ------------------------------------------------------------------------------------- + TOTAL 0.07598415 51.66686672 + (elements=3,942,912) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3942912 +Total Bytes 769004 +BPFP 1.5603 bits/point +EBPFP 3.1206 equivalent bits/point +MSE 51.666867 +---------------------- -------------------------------------------------------- +Time: 1.377s Load: 0.015s, Pack+Encode: 0.584s, Decode+Unpack: 0.778s +---------------------- -------------------------------------------------------- +💾 Converting with 51.6669 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1517.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 402, 128) +Output shape: (1, 402, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) -> torch.Size([1, 1, 402, 512]) + layer.4.output: torch.Size([1, 402, 3584]) -> torch.Size([1, 1, 402, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 20,832B, BPFP=0.8097 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 68,556B, BPFP=2.6646 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 35,900B, BPFP=1.3954 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 66,612B, BPFP=2.5891 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 42,632B, BPFP=1.6570 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 65,192B, BPFP=2.5339 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 39,820B, BPFP=1.5477 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 66,152B, BPFP=2.5712 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 56,416B, BPFP=2.1928 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 63,532B, BPFP=2.4694 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 141,256B, BPFP=0.7843 +⌛️ [2/4] FRONTEND: Frontend time: 0.489s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.639s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 402, 128]) + layer.0.v_cache: torch.Size([1, 4, 402, 128]) + layer.1.k_cache: torch.Size([1, 4, 402, 128]) + layer.1.v_cache: torch.Size([1, 4, 402, 128]) + layer.2.k_cache: torch.Size([1, 4, 402, 128]) + layer.2.v_cache: torch.Size([1, 4, 402, 128]) + layer.3.k_cache: torch.Size([1, 4, 402, 128]) + layer.3.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.k_cache: torch.Size([1, 4, 402, 128]) + layer.4.v_cache: torch.Size([1, 4, 402, 128]) + layer.4.output: torch.Size([1, 402, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18062694 22.67427025 + layer.0.v_cache 0.00001752 0.00573257 + layer.1.k_cache 0.89169403 3.09790100 + layer.1.v_cache 0.00000593 0.00207222 + layer.2.k_cache 0.03854401 0.44202624 + layer.2.v_cache 0.00001967 0.00556848 + layer.3.k_cache 0.02336847 1.92919816 + layer.3.v_cache 0.00001924 0.00648924 + layer.4.k_cache 0.00085529 0.12386840 + layer.4.v_cache 0.00004959 0.01131948 + layer.4.output 0.00648942 134.47622379 + ------------------------------------------------------------------------------------- + TOTAL 0.06944863 57.03717721 + (elements=3,499,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3499008 +Total Bytes 666900 +BPFP 1.5248 bits/point +EBPFP 3.0496 equivalent bits/point +MSE 57.037177 +---------------------- -------------------------------------------------------- +Time: 1.141s Load: 0.014s, Pack+Encode: 0.489s, Decode+Unpack: 0.639s +---------------------- -------------------------------------------------------- +💾 Converting with 57.0372 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1555.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 279, 128) +Output shape: (1, 279, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) -> torch.Size([1, 1, 279, 512]) + layer.4.output: torch.Size([1, 279, 3584]) -> torch.Size([1, 1, 279, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,808B, BPFP=0.8293 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,656B, BPFP=2.7249 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,204B, BPFP=1.4675 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,592B, BPFP=2.6653 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,156B, BPFP=1.7448 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,760B, BPFP=2.6187 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,172B, BPFP=1.6337 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,388B, BPFP=2.6539 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,964B, BPFP=2.2941 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,756B, BPFP=2.5625 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 113,164B, BPFP=0.9054 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.500s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 279, 128]) + layer.0.v_cache: torch.Size([1, 4, 279, 128]) + layer.1.k_cache: torch.Size([1, 4, 279, 128]) + layer.1.v_cache: torch.Size([1, 4, 279, 128]) + layer.2.k_cache: torch.Size([1, 4, 279, 128]) + layer.2.v_cache: torch.Size([1, 4, 279, 128]) + layer.3.k_cache: torch.Size([1, 4, 279, 128]) + layer.3.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.k_cache: torch.Size([1, 4, 279, 128]) + layer.4.v_cache: torch.Size([1, 4, 279, 128]) + layer.4.output: torch.Size([1, 279, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13958113 21.35625280 + layer.0.v_cache 0.00001688 0.00667768 + layer.1.k_cache 0.56864935 3.10334543 + layer.1.v_cache 0.00000621 0.00270761 + layer.2.k_cache 0.01018704 0.43574551 + layer.2.v_cache 0.00002046 0.00698297 + layer.3.k_cache 0.04374025 2.13804402 + layer.3.v_cache 0.00002178 0.00802549 + layer.4.k_cache 0.00068741 0.13934723 + layer.4.v_cache 0.00005272 0.01390232 + layer.4.output 0.00479460 195.10264657 + ------------------------------------------------------------------------------------- + TOTAL 0.04685444 81.93703277 + (elements=2,428,416) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2428416 +Total Bytes 491620 +BPFP 1.6196 bits/point +EBPFP 3.2391 equivalent bits/point +MSE 81.937033 +---------------------- -------------------------------------------------------- +Time: 0.850s Load: 0.008s, Pack+Encode: 0.342s, Decode+Unpack: 0.500s +---------------------- -------------------------------------------------------- +💾 Converting with 81.9370 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1606.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 381, 128) +Output shape: (1, 381, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) -> torch.Size([1, 1, 381, 512]) + layer.4.output: torch.Size([1, 381, 3584]) -> torch.Size([1, 1, 381, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 19,168B, BPFP=0.7861 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 62,660B, BPFP=2.5697 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 33,512B, BPFP=1.3743 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 61,136B, BPFP=2.5072 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 40,080B, BPFP=1.6437 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 60,192B, BPFP=2.4685 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 37,932B, BPFP=1.5556 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 60,932B, BPFP=2.4989 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 52,564B, BPFP=2.1557 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 58,760B, BPFP=2.4098 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 149,940B, BPFP=0.8784 +⌛️ [2/4] FRONTEND: Frontend time: 0.398s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.550s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 381, 128]) + layer.0.v_cache: torch.Size([1, 4, 381, 128]) + layer.1.k_cache: torch.Size([1, 4, 381, 128]) + layer.1.v_cache: torch.Size([1, 4, 381, 128]) + layer.2.k_cache: torch.Size([1, 4, 381, 128]) + layer.2.v_cache: torch.Size([1, 4, 381, 128]) + layer.3.k_cache: torch.Size([1, 4, 381, 128]) + layer.3.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.k_cache: torch.Size([1, 4, 381, 128]) + layer.4.v_cache: torch.Size([1, 4, 381, 128]) + layer.4.output: torch.Size([1, 381, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15713709 21.23193488 + layer.0.v_cache 0.00001719 0.00637761 + layer.1.k_cache 0.86767426 3.26512166 + layer.1.v_cache 0.00000622 0.00252665 + layer.2.k_cache 0.01635739 0.43265475 + layer.2.v_cache 0.00002098 0.00668352 + layer.3.k_cache 0.01709203 1.83778041 + layer.3.v_cache 0.00002073 0.00734134 + layer.4.k_cache 0.00073405 0.13366108 + layer.4.v_cache 0.00005083 0.01297510 + layer.4.output 0.03519400 135.27724269 + ------------------------------------------------------------------------------------- + TOTAL 0.07679228 57.28692681 + (elements=3,316,224) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3316224 +Total Bytes 636876 +BPFP 1.5364 bits/point +EBPFP 3.0728 equivalent bits/point +MSE 57.286927 +---------------------- -------------------------------------------------------- +Time: 0.960s Load: 0.012s, Pack+Encode: 0.398s, Decode+Unpack: 0.550s +---------------------- -------------------------------------------------------- +💾 Converting with 57.2869 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1611.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,612B, BPFP=0.8421 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,940B, BPFP=2.7337 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,436B, BPFP=1.4981 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,020B, BPFP=2.6723 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,464B, BPFP=1.7671 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,376B, BPFP=2.6293 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,856B, BPFP=1.6597 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,880B, BPFP=2.6629 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,392B, BPFP=2.2965 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,524B, BPFP=2.5724 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 97,724B, BPFP=0.9322 +⌛️ [2/4] FRONTEND: Frontend time: 0.347s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.419s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11422791 22.70640734 + layer.0.v_cache 0.00001930 0.00660920 + layer.1.k_cache 0.38928963 3.21010961 + layer.1.v_cache 0.00000631 0.00262103 + layer.2.k_cache 0.01623082 0.44240055 + layer.2.v_cache 0.00002095 0.00701264 + layer.3.k_cache 0.01313608 1.86214480 + layer.3.v_cache 0.00002100 0.00769646 + layer.4.k_cache 0.00066994 0.13858365 + layer.4.v_cache 0.00005661 0.01389028 + layer.4.output 1.30836332 226.00766941 + ------------------------------------------------------------------------------------- + TOTAL 0.57013069 94.73242126 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 417224 +BPFP 1.6388 bits/point +EBPFP 3.2776 equivalent bits/point +MSE 94.732421 +---------------------- -------------------------------------------------------- +Time: 0.773s Load: 0.008s, Pack+Encode: 0.347s, Decode+Unpack: 0.419s +---------------------- -------------------------------------------------------- +💾 Converting with 94.7324 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1621.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,596B, BPFP=0.8340 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,000B, BPFP=2.7145 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,372B, BPFP=1.4812 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,044B, BPFP=2.6512 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,284B, BPFP=1.7402 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,280B, BPFP=2.6006 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,972B, BPFP=1.6533 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,952B, BPFP=2.6451 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,484B, BPFP=2.2831 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,468B, BPFP=2.5469 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,792B, BPFP=0.9155 +⌛️ [2/4] FRONTEND: Frontend time: 0.333s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.401s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13591524 23.34468932 + layer.0.v_cache 0.00001656 0.00659962 + layer.1.k_cache 0.34520291 3.17748144 + layer.1.v_cache 0.00000598 0.00265528 + layer.2.k_cache 0.01250870 0.41698908 + layer.2.v_cache 0.00001982 0.00670725 + layer.3.k_cache 0.02633334 1.82762146 + layer.3.v_cache 0.00002177 0.00801015 + layer.4.k_cache 0.00067446 0.13853623 + layer.4.v_cache 0.00005295 0.01367456 + layer.4.output 1.29725540 219.64315981 + ------------------------------------------------------------------------------------- + TOTAL 0.56479644 92.14382841 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 416244 +BPFP 1.6211 bits/point +EBPFP 3.2422 equivalent bits/point +MSE 92.143828 +---------------------- -------------------------------------------------------- +Time: 0.741s Load: 0.007s, Pack+Encode: 0.333s, Decode+Unpack: 0.401s +---------------------- -------------------------------------------------------- +💾 Converting with 92.1438 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1635.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,764B, BPFP=0.8471 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,160B, BPFP=2.7477 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,692B, BPFP=1.4899 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,356B, BPFP=2.6898 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,156B, BPFP=1.7393 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,736B, BPFP=2.6452 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,860B, BPFP=1.6460 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,160B, BPFP=2.6757 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,116B, BPFP=2.3125 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,768B, BPFP=2.5755 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 82,984B, BPFP=0.8536 +⌛️ [2/4] FRONTEND: Frontend time: 0.330s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.424s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13150146 23.73949183 + layer.0.v_cache 0.00001564 0.00644920 + layer.1.k_cache 0.36060091 3.15010711 + layer.1.v_cache 0.00000598 0.00254154 + layer.2.k_cache 0.01419523 0.45944597 + layer.2.v_cache 0.00002081 0.00678947 + layer.3.k_cache 0.03680107 1.82799534 + layer.3.v_cache 0.00002092 0.00749589 + layer.4.k_cache 0.00067164 0.13603214 + layer.4.v_cache 0.00004838 0.01317108 + layer.4.output 1.41076765 244.76514154 + ------------------------------------------------------------------------------------- + TOTAL 0.61289739 102.51208884 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 379752 +BPFP 1.6085 bits/point +EBPFP 3.2169 equivalent bits/point +MSE 102.512089 +---------------------- -------------------------------------------------------- +Time: 0.761s Load: 0.007s, Pack+Encode: 0.330s, Decode+Unpack: 0.424s +---------------------- -------------------------------------------------------- +💾 Converting with 102.5121 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 343, 128) +Output shape: (1, 343, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) -> torch.Size([1, 1, 343, 512]) + layer.4.output: torch.Size([1, 343, 3584]) -> torch.Size([1, 1, 343, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 17,788B, BPFP=0.8103 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 59,020B, BPFP=2.6886 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,460B, BPFP=1.4331 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,696B, BPFP=2.6283 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 37,340B, BPFP=1.7010 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,576B, BPFP=2.5773 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 34,992B, BPFP=1.5940 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,408B, BPFP=2.6152 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 49,272B, BPFP=2.2445 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 55,292B, BPFP=2.5188 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 140,292B, BPFP=0.9130 +⌛️ [2/4] FRONTEND: Frontend time: 0.394s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.556s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 343, 128]) + layer.0.v_cache: torch.Size([1, 4, 343, 128]) + layer.1.k_cache: torch.Size([1, 4, 343, 128]) + layer.1.v_cache: torch.Size([1, 4, 343, 128]) + layer.2.k_cache: torch.Size([1, 4, 343, 128]) + layer.2.v_cache: torch.Size([1, 4, 343, 128]) + layer.3.k_cache: torch.Size([1, 4, 343, 128]) + layer.3.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.k_cache: torch.Size([1, 4, 343, 128]) + layer.4.v_cache: torch.Size([1, 4, 343, 128]) + layer.4.output: torch.Size([1, 343, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16677478 22.15830847 + layer.0.v_cache 0.00001698 0.00630730 + layer.1.k_cache 0.72144124 3.19946858 + layer.1.v_cache 0.00000659 0.00259959 + layer.2.k_cache 0.01351720 0.43560631 + layer.2.v_cache 0.00002117 0.00665775 + layer.3.k_cache 0.02347133 1.87082434 + layer.3.v_cache 0.00002006 0.00716203 + layer.4.k_cache 0.00069292 0.13071261 + layer.4.v_cache 0.00005144 0.01319621 + layer.4.output 0.03905215 155.26993961 + ------------------------------------------------------------------------------------- + TOTAL 0.07055169 65.57178944 + (elements=2,985,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2985472 +Total Bytes 597136 +BPFP 1.6001 bits/point +EBPFP 3.2002 equivalent bits/point +MSE 65.571789 +---------------------- -------------------------------------------------------- +Time: 0.960s Load: 0.011s, Pack+Encode: 0.394s, Decode+Unpack: 0.556s +---------------------- -------------------------------------------------------- +💾 Converting with 65.5718 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1657.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,864B, BPFP=0.8445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,180B, BPFP=2.7375 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,672B, BPFP=1.4586 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,804B, BPFP=2.6593 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,256B, BPFP=1.7191 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,896B, BPFP=2.6077 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,344B, BPFP=1.6105 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,516B, BPFP=2.6430 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,052B, BPFP=2.2757 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,852B, BPFP=2.5484 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 108,568B, BPFP=0.8812 +⌛️ [2/4] FRONTEND: Frontend time: 0.392s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.459s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13317820 22.03426136 + layer.0.v_cache 0.00001784 0.00602718 + layer.1.k_cache 0.51782770 3.23916726 + layer.1.v_cache 0.00000612 0.00250202 + layer.2.k_cache 0.01694398 0.42067846 + layer.2.v_cache 0.00002016 0.00626462 + layer.3.k_cache 0.01920085 1.84419101 + layer.3.v_cache 0.00002016 0.00687267 + layer.4.k_cache 0.00071488 0.12947673 + layer.4.v_cache 0.00005349 0.01260950 + layer.4.output 0.00484358 190.60629870 + ------------------------------------------------------------------------------------- + TOTAL 0.04246403 80.11447892 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 480004 +BPFP 1.6043 bits/point +EBPFP 3.2086 equivalent bits/point +MSE 80.114479 +---------------------- -------------------------------------------------------- +Time: 0.860s Load: 0.009s, Pack+Encode: 0.392s, Decode+Unpack: 0.459s +---------------------- -------------------------------------------------------- +💾 Converting with 80.1145 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1677.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.016s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 374, 128) +Output shape: (1, 374, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) -> torch.Size([1, 1, 374, 512]) + layer.4.output: torch.Size([1, 374, 3584]) -> torch.Size([1, 1, 374, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 19,656B, BPFP=0.8212 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 62,244B, BPFP=2.6004 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 33,360B, BPFP=1.3937 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 60,788B, BPFP=2.5396 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 39,316B, BPFP=1.6425 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 59,324B, BPFP=2.4784 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 36,936B, BPFP=1.5431 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 60,148B, BPFP=2.5129 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 51,700B, BPFP=2.1599 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 57,968B, BPFP=2.4218 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 138,416B, BPFP=0.8261 +⌛️ [2/4] FRONTEND: Frontend time: 0.423s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.597s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 374, 128]) + layer.0.v_cache: torch.Size([1, 4, 374, 128]) + layer.1.k_cache: torch.Size([1, 4, 374, 128]) + layer.1.v_cache: torch.Size([1, 4, 374, 128]) + layer.2.k_cache: torch.Size([1, 4, 374, 128]) + layer.2.v_cache: torch.Size([1, 4, 374, 128]) + layer.3.k_cache: torch.Size([1, 4, 374, 128]) + layer.3.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.k_cache: torch.Size([1, 4, 374, 128]) + layer.4.v_cache: torch.Size([1, 4, 374, 128]) + layer.4.output: torch.Size([1, 374, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18781370 22.34937960 + layer.0.v_cache 0.00001670 0.00605317 + layer.1.k_cache 0.85217587 3.15165082 + layer.1.v_cache 0.00000605 0.00245888 + layer.2.k_cache 0.02919085 0.44200914 + layer.2.v_cache 0.00002159 0.00603366 + layer.3.k_cache 0.02976600 1.88076896 + layer.3.v_cache 0.00002022 0.00673328 + layer.4.k_cache 0.00072409 0.12721138 + layer.4.v_cache 0.00004949 0.01202101 + layer.4.output 0.03578218 143.58588379 + ------------------------------------------------------------------------------------- + TOTAL 0.07942705 60.76973567 + (elements=3,255,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3255296 +Total Bytes 619856 +BPFP 1.5233 bits/point +EBPFP 3.0466 equivalent bits/point +MSE 60.769736 +---------------------- -------------------------------------------------------- +Time: 1.035s Load: 0.016s, Pack+Encode: 0.423s, Decode+Unpack: 0.597s +---------------------- -------------------------------------------------------- +💾 Converting with 60.7697 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1688.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,984B, BPFP=0.8418 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,024B, BPFP=2.6598 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,756B, BPFP=1.4754 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,204B, BPFP=2.6066 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,428B, BPFP=1.7134 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,608B, BPFP=2.5679 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,916B, BPFP=1.6154 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,084B, BPFP=2.5988 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,700B, BPFP=2.2497 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,796B, BPFP=2.5153 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,024B, BPFP=0.8616 +⌛️ [2/4] FRONTEND: Frontend time: 0.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.423s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13979621 22.91566917 + layer.0.v_cache 0.00001743 0.00606632 + layer.1.k_cache 0.46959303 3.23913118 + layer.1.v_cache 0.00000602 0.00251499 + layer.2.k_cache 0.01642122 0.43396461 + layer.2.v_cache 0.00002053 0.00644182 + layer.3.k_cache 0.03876302 1.91772106 + layer.3.v_cache 0.00002036 0.00742007 + layer.4.k_cache 0.00069112 0.13405546 + layer.4.v_cache 0.00005399 0.01348891 + layer.4.output 1.27033806 220.15134114 + ------------------------------------------------------------------------------------- + TOTAL 0.56222055 92.33740362 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 414524 +BPFP 1.5809 bits/point +EBPFP 3.1618 equivalent bits/point +MSE 92.337404 +---------------------- -------------------------------------------------------- +Time: 0.765s Load: 0.008s, Pack+Encode: 0.334s, Decode+Unpack: 0.423s +---------------------- -------------------------------------------------------- +💾 Converting with 92.3374 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1692.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,136B, BPFP=0.8354 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,996B, BPFP=2.7530 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,576B, BPFP=1.4851 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,072B, BPFP=2.6894 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,720B, BPFP=1.7704 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,440B, BPFP=2.6459 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,312B, BPFP=1.6735 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,024B, BPFP=2.6861 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,616B, BPFP=2.3139 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,460B, BPFP=2.5785 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,936B, BPFP=0.9237 +⌛️ [2/4] FRONTEND: Frontend time: 0.329s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.442s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13893990 23.07034692 + layer.0.v_cache 0.00001711 0.00631562 + layer.1.k_cache 0.36820191 3.16020545 + layer.1.v_cache 0.00000637 0.00255776 + layer.2.k_cache 0.01566778 0.43065042 + layer.2.v_cache 0.00002161 0.00689834 + layer.3.k_cache 0.01884081 1.85683693 + layer.3.v_cache 0.00002076 0.00759024 + layer.4.k_cache 0.00070054 0.13598393 + layer.4.v_cache 0.00005210 0.01354283 + layer.4.output 1.34866800 227.41313326 + ------------------------------------------------------------------------------------- + TOTAL 0.58724382 95.32840360 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 405288 +BPFP 1.6410 bits/point +EBPFP 3.2820 equivalent bits/point +MSE 95.328404 +---------------------- -------------------------------------------------------- +Time: 0.779s Load: 0.009s, Pack+Encode: 0.329s, Decode+Unpack: 0.442s +---------------------- -------------------------------------------------------- +💾 Converting with 95.3284 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-17.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-17.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,364B, BPFP=0.8406 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,404B, BPFP=2.7741 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,136B, BPFP=1.4710 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,316B, BPFP=2.7104 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,660B, BPFP=1.7357 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,540B, BPFP=2.6650 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,740B, BPFP=1.6234 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,092B, BPFP=2.6973 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,960B, BPFP=2.3385 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,736B, BPFP=2.6180 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 103,592B, BPFP=0.8660 +⌛️ [2/4] FRONTEND: Frontend time: 0.357s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.581s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11943251 22.58027929 + layer.0.v_cache 0.00001561 0.00608878 + layer.1.k_cache 0.49619702 3.01152856 + layer.1.v_cache 0.00000608 0.00243968 + layer.2.k_cache 0.01349212 0.42832495 + layer.2.v_cache 0.00002006 0.00649907 + layer.3.k_cache 0.01080775 1.72146961 + layer.3.v_cache 0.00002073 0.00695576 + layer.4.k_cache 0.00068886 0.13220296 + layer.4.v_cache 0.00007406 0.01312437 + layer.4.output 0.00494179 196.12036851 + ------------------------------------------------------------------------------------- + TOTAL 0.03972631 82.39714662 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 470540 +BPFP 1.6198 bits/point +EBPFP 3.2396 equivalent bits/point +MSE 82.397147 +---------------------- -------------------------------------------------------- +Time: 0.948s Load: 0.010s, Pack+Encode: 0.357s, Decode+Unpack: 0.581s +---------------------- -------------------------------------------------------- +💾 Converting with 82.3971 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1718.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,288B, BPFP=0.8688 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,800B, BPFP=2.7432 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,396B, BPFP=1.5127 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,000B, BPFP=2.6867 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,064B, BPFP=1.7721 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,280B, BPFP=2.6357 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,740B, BPFP=1.6785 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,844B, BPFP=2.6756 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,652B, BPFP=2.3085 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,512B, BPFP=2.5814 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,292B, BPFP=0.9322 +⌛️ [2/4] FRONTEND: Frontend time: 0.337s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.424s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11815372 24.29914407 + layer.0.v_cache 0.00001627 0.00648662 + layer.1.k_cache 0.37623358 3.40130035 + layer.1.v_cache 0.00000616 0.00271534 + layer.2.k_cache 0.01429399 0.44284651 + layer.2.v_cache 0.00002051 0.00716593 + layer.3.k_cache 0.01320818 1.97046879 + layer.3.v_cache 0.00002004 0.00815766 + layer.4.k_cache 0.00067885 0.14121711 + layer.4.v_cache 0.00005318 0.01478192 + layer.4.output 1.38524649 242.72002262 + ------------------------------------------------------------------------------------- + TOTAL 0.60114176 101.72555545 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 395868 +BPFP 1.6464 bits/point +EBPFP 3.2928 equivalent bits/point +MSE 101.725555 +---------------------- -------------------------------------------------------- +Time: 0.769s Load: 0.008s, Pack+Encode: 0.337s, Decode+Unpack: 0.424s +---------------------- -------------------------------------------------------- +💾 Converting with 101.7256 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1730.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 278, 128) +Output shape: (1, 278, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) -> torch.Size([1, 1, 278, 512]) + layer.4.output: torch.Size([1, 278, 3584]) -> torch.Size([1, 1, 278, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,496B, BPFP=0.8147 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,964B, BPFP=2.6958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,596B, BPFP=1.4386 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,040B, BPFP=2.6439 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,304B, BPFP=1.7032 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,064B, BPFP=2.5890 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,464B, BPFP=1.5998 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,800B, BPFP=2.6304 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,568B, BPFP=2.2801 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,432B, BPFP=2.5535 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 103,264B, BPFP=0.8291 +⌛️ [2/4] FRONTEND: Frontend time: 0.374s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.473s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 278, 128]) + layer.0.v_cache: torch.Size([1, 4, 278, 128]) + layer.1.k_cache: torch.Size([1, 4, 278, 128]) + layer.1.v_cache: torch.Size([1, 4, 278, 128]) + layer.2.k_cache: torch.Size([1, 4, 278, 128]) + layer.2.v_cache: torch.Size([1, 4, 278, 128]) + layer.3.k_cache: torch.Size([1, 4, 278, 128]) + layer.3.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.k_cache: torch.Size([1, 4, 278, 128]) + layer.4.v_cache: torch.Size([1, 4, 278, 128]) + layer.4.output: torch.Size([1, 278, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09820673 22.35094600 + layer.0.v_cache 0.00001666 0.00586389 + layer.1.k_cache 0.50544519 3.11428526 + layer.1.v_cache 0.00000613 0.00240998 + layer.2.k_cache 0.01930715 0.43009145 + layer.2.v_cache 0.00002033 0.00623374 + layer.3.k_cache 0.01337509 1.87029678 + layer.3.v_cache 0.00001987 0.00730539 + layer.4.k_cache 0.00074230 0.13299194 + layer.4.v_cache 0.00004869 0.01261099 + layer.4.output 0.00473525 197.77033016 + ------------------------------------------------------------------------------------- + TOTAL 0.03943146 83.07796156 + (elements=2,419,712) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2419712 +Total Bytes 475992 +BPFP 1.5737 bits/point +EBPFP 3.1474 equivalent bits/point +MSE 83.077962 +---------------------- -------------------------------------------------------- +Time: 0.857s Load: 0.010s, Pack+Encode: 0.374s, Decode+Unpack: 0.473s +---------------------- -------------------------------------------------------- +💾 Converting with 83.0780 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1775.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,884B, BPFP=0.8518 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,152B, BPFP=2.7345 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,732B, BPFP=1.4860 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,212B, BPFP=2.6671 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,636B, BPFP=1.7658 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,568B, BPFP=2.6210 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,120B, BPFP=1.6571 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,344B, BPFP=2.6766 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,384B, BPFP=2.3211 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,852B, BPFP=2.5697 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,704B, BPFP=0.8878 +⌛️ [2/4] FRONTEND: Frontend time: 0.331s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.455s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13853801 23.51179490 + layer.0.v_cache 0.00001803 0.00654854 + layer.1.k_cache 0.23740530 2.97414125 + layer.1.v_cache 0.00000581 0.00256391 + layer.2.k_cache 0.01896443 0.44990988 + layer.2.v_cache 0.00001977 0.00687980 + layer.3.k_cache 0.00876656 1.83792142 + layer.3.v_cache 0.00002034 0.00833530 + layer.4.k_cache 0.00073522 0.13912263 + layer.4.v_cache 0.00004914 0.01397389 + layer.4.output 1.40427464 243.86281537 + ------------------------------------------------------------------------------------- + TOTAL 0.60202618 102.11711171 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 384588 +BPFP 1.6215 bits/point +EBPFP 3.2430 equivalent bits/point +MSE 102.117112 +---------------------- -------------------------------------------------------- +Time: 0.794s Load: 0.007s, Pack+Encode: 0.331s, Decode+Unpack: 0.455s +---------------------- -------------------------------------------------------- +💾 Converting with 102.1171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1784.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 276, 128) +Output shape: (1, 276, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) -> torch.Size([1, 1, 276, 512]) + layer.4.output: torch.Size([1, 276, 3584]) -> torch.Size([1, 1, 276, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,028B, BPFP=0.8508 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,800B, BPFP=2.7061 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,396B, BPFP=1.4377 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,648B, BPFP=2.6409 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,448B, BPFP=1.7237 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,912B, BPFP=2.5992 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,568B, BPFP=1.6173 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,756B, BPFP=2.6470 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,320B, BPFP=2.2826 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,188B, BPFP=2.5582 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 109,340B, BPFP=0.8843 +⌛️ [2/4] FRONTEND: Frontend time: 0.389s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.488s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 276, 128]) + layer.0.v_cache: torch.Size([1, 4, 276, 128]) + layer.1.k_cache: torch.Size([1, 4, 276, 128]) + layer.1.v_cache: torch.Size([1, 4, 276, 128]) + layer.2.k_cache: torch.Size([1, 4, 276, 128]) + layer.2.v_cache: torch.Size([1, 4, 276, 128]) + layer.3.k_cache: torch.Size([1, 4, 276, 128]) + layer.3.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.k_cache: torch.Size([1, 4, 276, 128]) + layer.4.v_cache: torch.Size([1, 4, 276, 128]) + layer.4.output: torch.Size([1, 276, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10827675 22.02906689 + layer.0.v_cache 0.00001706 0.00598033 + layer.1.k_cache 0.56919524 2.98822818 + layer.1.v_cache 0.00000592 0.00243567 + layer.2.k_cache 0.01911232 0.42913296 + layer.2.v_cache 0.00002018 0.00644757 + layer.3.k_cache 0.01342056 1.83682439 + layer.3.v_cache 0.00002333 0.00762168 + layer.4.k_cache 0.00074072 0.13262391 + layer.4.v_cache 0.00005062 0.01303163 + layer.4.output 0.00480139 190.44998706 + ------------------------------------------------------------------------------------- + TOTAL 0.04379250 80.03537074 + (elements=2,402,304) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2402304 +Total Bytes 481404 +BPFP 1.6031 bits/point +EBPFP 3.2063 equivalent bits/point +MSE 80.035371 +---------------------- -------------------------------------------------------- +Time: 0.887s Load: 0.010s, Pack+Encode: 0.389s, Decode+Unpack: 0.488s +---------------------- -------------------------------------------------------- +💾 Converting with 80.0354 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,436B, BPFP=0.8448 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,992B, BPFP=2.8085 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,452B, BPFP=1.4895 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,780B, BPFP=2.7376 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,112B, BPFP=1.7622 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,740B, BPFP=2.6767 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,208B, BPFP=1.6507 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,324B, BPFP=2.7109 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,664B, BPFP=2.3212 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,560B, BPFP=2.6077 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 104,996B, BPFP=0.8778 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.466s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15963619 22.90129404 + layer.0.v_cache 0.00001721 0.00661042 + layer.1.k_cache 0.49929055 3.18994324 + layer.1.v_cache 0.00000624 0.00248779 + layer.2.k_cache 0.03270629 0.43876282 + layer.2.v_cache 0.00002062 0.00623316 + layer.3.k_cache 0.00946706 1.89454965 + layer.3.v_cache 0.00002056 0.00714112 + layer.4.k_cache 0.00068536 0.12899789 + layer.4.v_cache 0.00004941 0.01259991 + layer.4.output 0.00495045 197.03213617 + ------------------------------------------------------------------------------------- + TOTAL 0.04332662 82.81256313 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 474264 +BPFP 1.6326 bits/point +EBPFP 3.2652 equivalent bits/point +MSE 82.812563 +---------------------- -------------------------------------------------------- +Time: 0.793s Load: 0.008s, Pack+Encode: 0.319s, Decode+Unpack: 0.466s +---------------------- -------------------------------------------------------- +💾 Converting with 82.8126 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1823.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 236, 128) +Output shape: (1, 236, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) -> torch.Size([1, 1, 236, 512]) + layer.4.output: torch.Size([1, 236, 3584]) -> torch.Size([1, 1, 236, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,668B, BPFP=0.8387 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,064B, BPFP=2.7188 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,504B, BPFP=1.4899 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,204B, BPFP=2.6618 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,288B, BPFP=1.7405 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,204B, BPFP=2.5956 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,860B, BPFP=1.6459 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,892B, BPFP=2.6412 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,576B, BPFP=2.2892 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,508B, BPFP=2.5495 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,932B, BPFP=0.8790 +⌛️ [2/4] FRONTEND: Frontend time: 0.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 236, 128]) + layer.0.v_cache: torch.Size([1, 4, 236, 128]) + layer.1.k_cache: torch.Size([1, 4, 236, 128]) + layer.1.v_cache: torch.Size([1, 4, 236, 128]) + layer.2.k_cache: torch.Size([1, 4, 236, 128]) + layer.2.v_cache: torch.Size([1, 4, 236, 128]) + layer.3.k_cache: torch.Size([1, 4, 236, 128]) + layer.3.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.k_cache: torch.Size([1, 4, 236, 128]) + layer.4.v_cache: torch.Size([1, 4, 236, 128]) + layer.4.output: torch.Size([1, 236, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14874478 23.09843833 + layer.0.v_cache 0.00001714 0.00626117 + layer.1.k_cache 0.45078779 3.05729546 + layer.1.v_cache 0.00000615 0.00252788 + layer.2.k_cache 0.01724745 0.45711149 + layer.2.v_cache 0.00001987 0.00644639 + layer.3.k_cache 0.02849353 1.90529296 + layer.3.v_cache 0.00002056 0.00740077 + layer.4.k_cache 0.00066912 0.13264881 + layer.4.v_cache 0.00005068 0.01275537 + layer.4.output 1.29722614 219.01475484 + ------------------------------------------------------------------------------------- + TOTAL 0.57215530 91.86996838 + (elements=2,054,144) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2054144 +Total Bytes 412700 +BPFP 1.6073 bits/point +EBPFP 3.2146 equivalent bits/point +MSE 91.869968 +---------------------- -------------------------------------------------------- +Time: 0.729s Load: 0.008s, Pack+Encode: 0.309s, Decode+Unpack: 0.411s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8700 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1829.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 341, 128) +Output shape: (1, 341, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) -> torch.Size([1, 1, 341, 512]) + layer.4.output: torch.Size([1, 341, 3584]) -> torch.Size([1, 1, 341, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 17,972B, BPFP=0.8235 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,780B, BPFP=2.6934 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 31,140B, BPFP=1.4269 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 57,424B, BPFP=2.6312 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 37,076B, BPFP=1.6989 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 56,252B, BPFP=2.5775 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 34,644B, BPFP=1.5874 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 57,036B, BPFP=2.6135 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 48,860B, BPFP=2.2388 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,796B, BPFP=2.5108 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 127,392B, BPFP=0.8339 +⌛️ [2/4] FRONTEND: Frontend time: 0.383s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.561s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 341, 128]) + layer.0.v_cache: torch.Size([1, 4, 341, 128]) + layer.1.k_cache: torch.Size([1, 4, 341, 128]) + layer.1.v_cache: torch.Size([1, 4, 341, 128]) + layer.2.k_cache: torch.Size([1, 4, 341, 128]) + layer.2.v_cache: torch.Size([1, 4, 341, 128]) + layer.3.k_cache: torch.Size([1, 4, 341, 128]) + layer.3.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.k_cache: torch.Size([1, 4, 341, 128]) + layer.4.v_cache: torch.Size([1, 4, 341, 128]) + layer.4.output: torch.Size([1, 341, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17009405 22.17762841 + layer.0.v_cache 0.00001631 0.00626776 + layer.1.k_cache 0.79816426 3.24054796 + layer.1.v_cache 0.00000611 0.00251671 + layer.2.k_cache 0.02878924 0.45318697 + layer.2.v_cache 0.00002016 0.00617702 + layer.3.k_cache 0.01035325 1.98080990 + layer.3.v_cache 0.00002082 0.00708460 + layer.4.k_cache 0.00072160 0.13275019 + layer.4.v_cache 0.00005837 0.01232099 + layer.4.output 0.03919861 157.54161866 + ------------------------------------------------------------------------------------- + TOTAL 0.07544909 66.51827183 + (elements=2,968,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2968064 +Total Bytes 581372 +BPFP 1.5670 bits/point +EBPFP 3.1340 equivalent bits/point +MSE 66.518272 +---------------------- -------------------------------------------------------- +Time: 0.956s Load: 0.011s, Pack+Encode: 0.383s, Decode+Unpack: 0.561s +---------------------- -------------------------------------------------------- +💾 Converting with 66.5183 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1834.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 335, 128) +Output shape: (1, 335, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) -> torch.Size([1, 1, 335, 512]) + layer.4.output: torch.Size([1, 335, 3584]) -> torch.Size([1, 1, 335, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 17,748B, BPFP=0.8278 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 58,348B, BPFP=2.7215 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 30,880B, BPFP=1.4403 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 56,748B, BPFP=2.6468 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 36,472B, BPFP=1.7011 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 55,372B, BPFP=2.5826 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 34,236B, BPFP=1.5968 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 56,164B, BPFP=2.6196 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 48,428B, BPFP=2.2588 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 54,320B, BPFP=2.5336 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 127,444B, BPFP=0.8492 +⌛️ [2/4] FRONTEND: Frontend time: 0.345s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.548s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 335, 128]) + layer.0.v_cache: torch.Size([1, 4, 335, 128]) + layer.1.k_cache: torch.Size([1, 4, 335, 128]) + layer.1.v_cache: torch.Size([1, 4, 335, 128]) + layer.2.k_cache: torch.Size([1, 4, 335, 128]) + layer.2.v_cache: torch.Size([1, 4, 335, 128]) + layer.3.k_cache: torch.Size([1, 4, 335, 128]) + layer.3.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.k_cache: torch.Size([1, 4, 335, 128]) + layer.4.v_cache: torch.Size([1, 4, 335, 128]) + layer.4.output: torch.Size([1, 335, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15195324 21.96408874 + layer.0.v_cache 0.00001666 0.00630390 + layer.1.k_cache 0.76788093 3.14248266 + layer.1.v_cache 0.00000624 0.00256191 + layer.2.k_cache 0.02686723 0.42010662 + layer.2.v_cache 0.00001960 0.00608145 + layer.3.k_cache 0.01080669 1.81744476 + layer.3.v_cache 0.00002024 0.00704874 + layer.4.k_cache 0.00071790 0.13035453 + layer.4.v_cache 0.00005106 0.01269869 + layer.4.output 3.38143948 156.65955490 + ------------------------------------------------------------------------------------- + TOTAL 1.44873036 66.12506214 + (elements=2,915,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2915840 +Total Bytes 576160 +BPFP 1.5808 bits/point +EBPFP 3.1615 equivalent bits/point +MSE 66.125062 +---------------------- -------------------------------------------------------- +Time: 0.904s Load: 0.011s, Pack+Encode: 0.345s, Decode+Unpack: 0.548s +---------------------- -------------------------------------------------------- +💾 Converting with 66.1251 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1843.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,108B, BPFP=0.8318 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,424B, BPFP=2.7962 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,040B, BPFP=1.4764 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,232B, BPFP=2.7259 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,476B, BPFP=1.7380 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,140B, BPFP=2.6616 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,692B, BPFP=1.6328 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,948B, BPFP=2.7092 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,236B, BPFP=2.3134 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,188B, BPFP=2.6054 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 109,260B, BPFP=0.9203 +⌛️ [2/4] FRONTEND: Frontend time: 0.384s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.569s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17469128 22.45921838 + layer.0.v_cache 0.00001775 0.00644777 + layer.1.k_cache 0.51265420 3.11050542 + layer.1.v_cache 0.00000620 0.00253304 + layer.2.k_cache 0.03428403 0.44626223 + layer.2.v_cache 0.00002024 0.00611975 + layer.3.k_cache 0.01443097 1.81297780 + layer.3.v_cache 0.00002182 0.00747511 + layer.4.k_cache 0.00070119 0.12590214 + layer.4.v_cache 0.00005097 0.01223858 + layer.4.output 0.00503371 204.50279650 + ------------------------------------------------------------------------------------- + TOTAL 0.04541851 85.85348563 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 473744 +BPFP 1.6431 bits/point +EBPFP 3.2862 equivalent bits/point +MSE 85.853486 +---------------------- -------------------------------------------------------- +Time: 0.961s Load: 0.009s, Pack+Encode: 0.384s, Decode+Unpack: 0.569s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8535 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1844.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 265, 128) +Output shape: (1, 265, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) -> torch.Size([1, 1, 265, 512]) + layer.4.output: torch.Size([1, 265, 3584]) -> torch.Size([1, 1, 265, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,988B, BPFP=0.8248 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,104B, BPFP=2.7774 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,812B, BPFP=1.4630 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,984B, BPFP=2.7113 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,464B, BPFP=1.7373 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,816B, BPFP=2.6425 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,524B, BPFP=1.6229 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,624B, BPFP=2.6901 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,484B, BPFP=2.3281 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,256B, BPFP=2.6094 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,552B, BPFP=0.8638 +⌛️ [2/4] FRONTEND: Frontend time: 0.412s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.590s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 265, 128]) + layer.0.v_cache: torch.Size([1, 4, 265, 128]) + layer.1.k_cache: torch.Size([1, 4, 265, 128]) + layer.1.v_cache: torch.Size([1, 4, 265, 128]) + layer.2.k_cache: torch.Size([1, 4, 265, 128]) + layer.2.v_cache: torch.Size([1, 4, 265, 128]) + layer.3.k_cache: torch.Size([1, 4, 265, 128]) + layer.3.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.k_cache: torch.Size([1, 4, 265, 128]) + layer.4.v_cache: torch.Size([1, 4, 265, 128]) + layer.4.output: torch.Size([1, 265, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18043976 22.62828346 + layer.0.v_cache 0.00001814 0.00631006 + layer.1.k_cache 0.58196952 3.36893840 + layer.1.v_cache 0.00000608 0.00255569 + layer.2.k_cache 0.02409397 0.43113536 + layer.2.v_cache 0.00002020 0.00620665 + layer.3.k_cache 0.01875559 1.82423556 + layer.3.v_cache 0.00002085 0.00728772 + layer.4.k_cache 0.00069073 0.13419345 + layer.4.v_cache 0.00005389 0.01352653 + layer.4.output 0.00498411 204.77552224 + ------------------------------------------------------------------------------------- + TOTAL 0.04946809 85.99125462 + (elements=2,306,560) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2306560 +Total Bytes 465608 +BPFP 1.6149 bits/point +EBPFP 3.2298 equivalent bits/point +MSE 85.991255 +---------------------- -------------------------------------------------------- +Time: 1.010s Load: 0.008s, Pack+Encode: 0.412s, Decode+Unpack: 0.590s +---------------------- -------------------------------------------------------- +💾 Converting with 85.9913 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-1854.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 386, 128) +Output shape: (1, 386, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) -> torch.Size([1, 1, 386, 512]) + layer.4.output: torch.Size([1, 386, 3584]) -> torch.Size([1, 1, 386, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 19,792B, BPFP=0.8012 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 67,564B, BPFP=2.7349 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 34,908B, BPFP=1.4131 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 65,800B, BPFP=2.6635 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 42,544B, BPFP=1.7222 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 64,792B, BPFP=2.6227 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 39,812B, BPFP=1.6116 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 65,760B, BPFP=2.6619 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 56,672B, BPFP=2.2940 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 63,464B, BPFP=2.5690 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 149,516B, BPFP=0.8646 +⌛️ [2/4] FRONTEND: Frontend time: 0.505s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.719s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 386, 128]) + layer.0.v_cache: torch.Size([1, 4, 386, 128]) + layer.1.k_cache: torch.Size([1, 4, 386, 128]) + layer.1.v_cache: torch.Size([1, 4, 386, 128]) + layer.2.k_cache: torch.Size([1, 4, 386, 128]) + layer.2.v_cache: torch.Size([1, 4, 386, 128]) + layer.3.k_cache: torch.Size([1, 4, 386, 128]) + layer.3.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.k_cache: torch.Size([1, 4, 386, 128]) + layer.4.v_cache: torch.Size([1, 4, 386, 128]) + layer.4.output: torch.Size([1, 386, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16808204 21.56906776 + layer.0.v_cache 0.00001827 0.00648166 + layer.1.k_cache 0.89383947 3.17057966 + layer.1.v_cache 0.00000605 0.00252812 + layer.2.k_cache 0.02643975 0.42657131 + layer.2.v_cache 0.00002094 0.00683723 + layer.3.k_cache 0.04368651 1.83265931 + layer.3.v_cache 0.00002149 0.00772602 + layer.4.k_cache 0.00075087 0.13802555 + layer.4.v_cache 0.00005208 0.01338922 + layer.4.output 0.03469837 138.17942728 + ------------------------------------------------------------------------------------- + TOTAL 0.08092977 58.49587393 + (elements=3,359,744) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3359744 +Total Bytes 670624 +BPFP 1.5968 bits/point +EBPFP 3.1937 equivalent bits/point +MSE 58.495874 +---------------------- -------------------------------------------------------- +Time: 1.238s Load: 0.014s, Pack+Encode: 0.505s, Decode+Unpack: 0.719s +---------------------- -------------------------------------------------------- +💾 Converting with 58.4959 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2100.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,100B, BPFP=0.8192 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,716B, BPFP=2.6973 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,800B, BPFP=1.4540 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,768B, BPFP=2.6458 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,624B, BPFP=1.7157 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,852B, BPFP=2.5961 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,900B, BPFP=1.6222 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,340B, BPFP=2.6226 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,712B, BPFP=2.2630 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,672B, BPFP=2.5321 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 115,028B, BPFP=0.8915 +⌛️ [2/4] FRONTEND: Frontend time: 0.379s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.560s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14364293 22.68620131 + layer.0.v_cache 0.00001600 0.00656864 + layer.1.k_cache 0.62366660 3.17808533 + layer.1.v_cache 0.00000628 0.00267163 + layer.2.k_cache 0.01981349 0.46332031 + layer.2.v_cache 0.00002090 0.00694688 + layer.3.k_cache 0.02158341 1.90737089 + layer.3.v_cache 0.00002097 0.00772953 + layer.4.k_cache 0.00071696 0.13686638 + layer.4.v_cache 0.00005883 0.01384826 + layer.4.output 0.00463133 189.23552207 + ------------------------------------------------------------------------------------- + TOTAL 0.04952739 79.59166257 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 501512 +BPFP 1.6005 bits/point +EBPFP 3.2010 equivalent bits/point +MSE 79.591663 +---------------------- -------------------------------------------------------- +Time: 0.950s Load: 0.011s, Pack+Encode: 0.379s, Decode+Unpack: 0.560s +---------------------- -------------------------------------------------------- +💾 Converting with 79.5917 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2141.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,684B, BPFP=0.8283 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,108B, BPFP=2.7137 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,416B, BPFP=1.4337 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,996B, BPFP=2.6509 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,320B, BPFP=1.7103 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,288B, BPFP=2.6110 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,636B, BPFP=1.6153 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,844B, BPFP=2.6424 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,160B, BPFP=2.2653 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,256B, BPFP=2.5528 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 114,264B, BPFP=0.9208 +⌛️ [2/4] FRONTEND: Frontend time: 0.355s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.507s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12300906 21.63494543 + layer.0.v_cache 0.00001720 0.00568695 + layer.1.k_cache 0.58060381 3.19513490 + layer.1.v_cache 0.00000644 0.00217938 + layer.2.k_cache 0.02305070 0.43605871 + layer.2.v_cache 0.00002053 0.00613827 + layer.3.k_cache 0.05279791 1.84354255 + layer.3.v_cache 0.00002021 0.00712848 + layer.4.k_cache 0.00074809 0.12851197 + layer.4.v_cache 0.00005348 0.01273579 + layer.4.output 0.00483522 198.61550735 + ------------------------------------------------------------------------------------- + TOTAL 0.04789259 83.38709493 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 486972 +BPFP 1.6158 bits/point +EBPFP 3.2317 equivalent bits/point +MSE 83.387095 +---------------------- -------------------------------------------------------- +Time: 0.870s Load: 0.009s, Pack+Encode: 0.355s, Decode+Unpack: 0.507s +---------------------- -------------------------------------------------------- +💾 Converting with 83.3871 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2291.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,444B, BPFP=0.8719 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,324B, BPFP=2.7553 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,368B, BPFP=1.4972 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,380B, BPFP=2.6892 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,420B, BPFP=1.7811 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,760B, BPFP=2.6457 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,964B, BPFP=1.6791 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,152B, BPFP=2.6732 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,208B, BPFP=2.3268 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,860B, BPFP=2.5827 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,848B, BPFP=0.9294 +⌛️ [2/4] FRONTEND: Frontend time: 0.339s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.443s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16177035 24.04138567 + layer.0.v_cache 0.00001782 0.00691255 + layer.1.k_cache 0.31952496 2.99771679 + layer.1.v_cache 0.00000626 0.00257431 + layer.2.k_cache 0.01271096 0.42610370 + layer.2.v_cache 0.00002040 0.00696711 + layer.3.k_cache 0.01363577 2.03839577 + layer.3.v_cache 0.00002033 0.00792306 + layer.4.k_cache 0.00066992 0.14020456 + layer.4.v_cache 0.00005030 0.01321825 + layer.4.output 1.37284199 237.78939782 + ------------------------------------------------------------------------------------- + TOTAL 0.59519535 99.65924627 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 399728 +BPFP 1.6475 bits/point +EBPFP 3.2950 equivalent bits/point +MSE 99.659246 +---------------------- -------------------------------------------------------- +Time: 0.789s Load: 0.007s, Pack+Encode: 0.339s, Decode+Unpack: 0.443s +---------------------- -------------------------------------------------------- +💾 Converting with 99.6592 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2340.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,976B, BPFP=0.8518 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,340B, BPFP=2.6759 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,412B, BPFP=1.4867 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,568B, BPFP=2.6100 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,652B, BPFP=1.7633 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,248B, BPFP=2.5827 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,472B, BPFP=1.6626 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,756B, BPFP=2.6260 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,748B, BPFP=2.2838 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,724B, BPFP=2.5379 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 77,592B, BPFP=0.9464 +⌛️ [2/4] FRONTEND: Frontend time: 0.343s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.401s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09962382 23.75142482 + layer.0.v_cache 0.00001686 0.00694238 + layer.1.k_cache 0.10988800 3.18908191 + layer.1.v_cache 0.00000590 0.00266316 + layer.2.k_cache 0.01214521 0.46341243 + layer.2.v_cache 0.00002216 0.00729551 + layer.3.k_cache 0.02260447 1.86419594 + layer.3.v_cache 0.00002006 0.00849769 + layer.4.k_cache 0.00069723 0.14506529 + layer.4.v_cache 0.00005242 0.01517060 + layer.4.output 0.00882482 291.55423009 + ------------------------------------------------------------------------------------- + TOTAL 0.01804999 121.78431532 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 324488 +BPFP 1.6297 bits/point +EBPFP 3.2595 equivalent bits/point +MSE 121.784315 +---------------------- -------------------------------------------------------- +Time: 0.751s Load: 0.006s, Pack+Encode: 0.343s, Decode+Unpack: 0.401s +---------------------- -------------------------------------------------------- +💾 Converting with 121.7843 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2342.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 281, 128) +Output shape: (1, 281, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) -> torch.Size([1, 1, 281, 512]) + layer.4.output: torch.Size([1, 281, 3584]) -> torch.Size([1, 1, 281, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,932B, BPFP=0.8303 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,676B, BPFP=2.7066 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,976B, BPFP=1.4444 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,380B, BPFP=2.6346 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,884B, BPFP=1.7173 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,360B, BPFP=2.5778 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,004B, BPFP=1.6128 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,244B, BPFP=2.6270 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,648B, BPFP=2.2602 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,500B, BPFP=2.5300 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 112,308B, BPFP=0.8921 +⌛️ [2/4] FRONTEND: Frontend time: 0.410s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.576s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 281, 128]) + layer.0.v_cache: torch.Size([1, 4, 281, 128]) + layer.1.k_cache: torch.Size([1, 4, 281, 128]) + layer.1.v_cache: torch.Size([1, 4, 281, 128]) + layer.2.k_cache: torch.Size([1, 4, 281, 128]) + layer.2.v_cache: torch.Size([1, 4, 281, 128]) + layer.3.k_cache: torch.Size([1, 4, 281, 128]) + layer.3.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.k_cache: torch.Size([1, 4, 281, 128]) + layer.4.v_cache: torch.Size([1, 4, 281, 128]) + layer.4.output: torch.Size([1, 281, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11148910 22.46792114 + layer.0.v_cache 0.00001684 0.00644420 + layer.1.k_cache 0.60671574 3.01425399 + layer.1.v_cache 0.00000630 0.00243974 + layer.2.k_cache 0.04149003 0.45851885 + layer.2.v_cache 0.00001976 0.00610955 + layer.3.k_cache 0.02200458 1.93538136 + layer.3.v_cache 0.00001954 0.00698314 + layer.4.k_cache 0.00078895 0.13294652 + layer.4.v_cache 0.00004864 0.01220665 + layer.4.output 0.00472744 192.63591446 + ------------------------------------------------------------------------------------- + TOTAL 0.04798185 80.97027096 + (elements=2,445,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2445824 +Total Bytes 488912 +BPFP 1.5992 bits/point +EBPFP 3.1983 equivalent bits/point +MSE 80.970271 +---------------------- -------------------------------------------------------- +Time: 0.998s Load: 0.011s, Pack+Encode: 0.410s, Decode+Unpack: 0.576s +---------------------- -------------------------------------------------------- +💾 Converting with 80.9703 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-241.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-241.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,900B, BPFP=0.8516 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,592B, BPFP=2.8588 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,504B, BPFP=1.5237 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,852B, BPFP=2.8009 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,280B, BPFP=1.8188 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,336B, BPFP=2.7606 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,580B, BPFP=1.6859 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,792B, BPFP=2.7963 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,052B, BPFP=2.4259 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,740B, BPFP=2.7141 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,840B, BPFP=0.8799 +⌛️ [2/4] FRONTEND: Frontend time: 0.366s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.479s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14565937 23.83738770 + layer.0.v_cache 0.00001550 0.00626695 + layer.1.k_cache 0.22559690 3.30470795 + layer.1.v_cache 0.00000590 0.00263303 + layer.2.k_cache 0.00679684 0.43863400 + layer.2.v_cache 0.00002314 0.00714368 + layer.3.k_cache 0.06866968 1.98849670 + layer.3.v_cache 0.00002119 0.00805610 + layer.4.k_cache 0.00065507 0.13763040 + layer.4.v_cache 0.00005191 0.01392005 + layer.4.output 1.53062134 265.96792411 + ------------------------------------------------------------------------------------- + TOTAL 0.65657911 111.26590266 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 363468 +BPFP 1.6703 bits/point +EBPFP 3.3407 equivalent bits/point +MSE 111.265903 +---------------------- -------------------------------------------------------- +Time: 0.851s Load: 0.006s, Pack+Encode: 0.366s, Decode+Unpack: 0.479s +---------------------- -------------------------------------------------------- +💾 Converting with 111.2659 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2462.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,932B, BPFP=0.8498 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,956B, BPFP=2.8728 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,620B, BPFP=1.5252 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,996B, BPFP=2.7982 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,388B, BPFP=1.8181 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,300B, BPFP=2.7441 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,852B, BPFP=1.6987 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,944B, BPFP=2.7942 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,168B, BPFP=2.4229 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,732B, BPFP=2.6999 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 80,616B, BPFP=0.8953 +⌛️ [2/4] FRONTEND: Frontend time: 0.358s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.483s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10663509 24.20055484 + layer.0.v_cache 0.00001613 0.00683317 + layer.1.k_cache 0.26620468 3.26440187 + layer.1.v_cache 0.00000602 0.00275142 + layer.2.k_cache 0.00649704 0.45525910 + layer.2.v_cache 0.00002002 0.00705394 + layer.3.k_cache 0.02565914 1.94937544 + layer.3.v_cache 0.00002186 0.00799755 + layer.4.k_cache 0.00070101 0.14442512 + layer.4.v_cache 0.00005421 0.01467354 + layer.4.output 1.52303164 264.70153696 + ------------------------------------------------------------------------------------- + TOTAL 0.65100216 110.76259322 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 366504 +BPFP 1.6759 bits/point +EBPFP 3.3518 equivalent bits/point +MSE 110.762593 +---------------------- -------------------------------------------------------- +Time: 0.848s Load: 0.008s, Pack+Encode: 0.358s, Decode+Unpack: 0.483s +---------------------- -------------------------------------------------------- +💾 Converting with 110.7626 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,212B, BPFP=0.8317 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,596B, BPFP=2.7853 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,596B, BPFP=1.4979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,600B, BPFP=2.7271 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,360B, BPFP=1.7767 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,848B, BPFP=2.6831 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,588B, BPFP=1.6730 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,564B, BPFP=2.7250 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,244B, BPFP=2.3551 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,848B, BPFP=2.6245 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 112,152B, BPFP=0.9376 +⌛️ [2/4] FRONTEND: Frontend time: 0.397s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.496s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258070 23.21367151 + layer.0.v_cache 0.00001713 0.00678070 + layer.1.k_cache 0.49703808 3.29078726 + layer.1.v_cache 0.00000670 0.00284668 + layer.2.k_cache 0.01458295 0.43946193 + layer.2.v_cache 0.00002181 0.00691758 + layer.3.k_cache 0.02940617 1.80703747 + layer.3.v_cache 0.00002120 0.00794016 + layer.4.k_cache 0.00070157 0.13967403 + layer.4.v_cache 0.00005095 0.01343668 + layer.4.output 0.00497170 198.09744516 + ------------------------------------------------------------------------------------- + TOTAL 0.04171936 83.27121589 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 482608 +BPFP 1.6613 bits/point +EBPFP 3.3226 equivalent bits/point +MSE 83.271216 +---------------------- -------------------------------------------------------- +Time: 0.903s Load: 0.009s, Pack+Encode: 0.397s, Decode+Unpack: 0.496s +---------------------- -------------------------------------------------------- +💾 Converting with 83.2712 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2469.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,532B, BPFP=0.7973 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,480B, BPFP=2.7295 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,160B, BPFP=1.4629 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,612B, BPFP=2.6695 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,548B, BPFP=1.7663 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,032B, BPFP=2.6294 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,188B, BPFP=1.6723 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,756B, BPFP=2.6795 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,456B, BPFP=2.3131 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,164B, BPFP=2.5694 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,056B, BPFP=0.9092 +⌛️ [2/4] FRONTEND: Frontend time: 0.348s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.454s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15433978 23.46507709 + layer.0.v_cache 0.00001575 0.00653904 + layer.1.k_cache 0.29362690 3.18092954 + layer.1.v_cache 0.00000607 0.00277711 + layer.2.k_cache 0.00940348 0.44088705 + layer.2.v_cache 0.00002159 0.00733153 + layer.3.k_cache 0.02320663 1.70821171 + layer.3.v_cache 0.00002067 0.00830072 + layer.4.k_cache 0.00067785 0.14241808 + layer.4.v_cache 0.00005016 0.01413408 + layer.4.output 1.35461534 235.72240439 + ------------------------------------------------------------------------------------- + TOTAL 0.58609860 98.76667275 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 399984 +BPFP 1.6267 bits/point +EBPFP 3.2534 equivalent bits/point +MSE 98.766673 +---------------------- -------------------------------------------------------- +Time: 0.812s Load: 0.010s, Pack+Encode: 0.348s, Decode+Unpack: 0.454s +---------------------- -------------------------------------------------------- +💾 Converting with 98.7667 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2480.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,416B, BPFP=0.8452 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,372B, BPFP=2.7068 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,596B, BPFP=1.4581 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,204B, BPFP=2.6428 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,672B, BPFP=1.7364 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,376B, BPFP=2.5974 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,736B, BPFP=1.6303 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,136B, BPFP=2.6390 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,208B, BPFP=2.2592 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,136B, BPFP=2.5294 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 119,148B, BPFP=0.9332 +⌛️ [2/4] FRONTEND: Frontend time: 0.357s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.516s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14751465 22.74643298 + layer.0.v_cache 0.00001719 0.00643289 + layer.1.k_cache 0.52245269 3.26539628 + layer.1.v_cache 0.00000618 0.00244151 + layer.2.k_cache 0.02707210 0.43702323 + layer.2.v_cache 0.00002227 0.00647194 + layer.3.k_cache 0.03364062 2.01692601 + layer.3.v_cache 0.00002083 0.00751676 + layer.4.k_cache 0.00069748 0.13531560 + layer.4.v_cache 0.00005305 0.01270856 + layer.4.output 0.00471531 193.12055138 + ------------------------------------------------------------------------------------- + TOTAL 0.04497084 81.20473679 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 503000 +BPFP 1.6222 bits/point +EBPFP 3.2443 equivalent bits/point +MSE 81.204737 +---------------------- -------------------------------------------------------- +Time: 0.884s Load: 0.010s, Pack+Encode: 0.357s, Decode+Unpack: 0.516s +---------------------- -------------------------------------------------------- +💾 Converting with 81.2047 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-2493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,912B, BPFP=0.7912 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,460B, BPFP=2.5404 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,380B, BPFP=1.3713 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,556B, BPFP=2.4850 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,232B, BPFP=1.6074 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,616B, BPFP=2.4275 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,980B, BPFP=1.5306 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,036B, BPFP=2.4532 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,860B, BPFP=2.1360 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,840B, BPFP=2.3799 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 105,844B, BPFP=0.9265 +⌛️ [2/4] FRONTEND: Frontend time: 0.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.420s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17432048 21.47684015 + layer.0.v_cache 0.00001704 0.00590919 + layer.1.k_cache 0.39057378 3.11153253 + layer.1.v_cache 0.00000668 0.00246137 + layer.2.k_cache 0.01299327 0.40434418 + layer.2.v_cache 0.00002114 0.00578221 + layer.3.k_cache 0.02499930 1.75395268 + layer.3.v_cache 0.00002096 0.00678650 + layer.4.k_cache 0.00069121 0.12731958 + layer.4.v_cache 0.00005416 0.01231475 + layer.4.output 1.20070616 175.43067227 + ------------------------------------------------------------------------------------- + TOTAL 0.52992007 73.81893818 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 427716 +BPFP 1.5417 bits/point +EBPFP 3.0833 equivalent bits/point +MSE 73.818938 +---------------------- -------------------------------------------------------- +Time: 0.737s Load: 0.008s, Pack+Encode: 0.309s, Decode+Unpack: 0.420s +---------------------- -------------------------------------------------------- +💾 Converting with 73.8189 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-265.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-265.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,256B, BPFP=0.8362 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,228B, BPFP=2.7448 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,996B, BPFP=1.5008 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,304B, BPFP=2.6818 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,628B, BPFP=1.7486 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,568B, BPFP=2.6316 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,360B, BPFP=1.6621 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,172B, BPFP=2.6728 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,864B, BPFP=2.3106 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,740B, BPFP=2.5751 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,396B, BPFP=0.9104 +⌛️ [2/4] FRONTEND: Frontend time: 0.316s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.424s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16239141 23.30451692 + layer.0.v_cache 0.00001645 0.00637498 + layer.1.k_cache 0.33751179 3.23149968 + layer.1.v_cache 0.00000668 0.00255931 + layer.2.k_cache 0.00660169 0.43631741 + layer.2.v_cache 0.00002045 0.00644515 + layer.3.k_cache 0.02213682 1.82888594 + layer.3.v_cache 0.00002049 0.00767135 + layer.4.k_cache 0.00067824 0.13862065 + layer.4.v_cache 0.00005183 0.01291434 + layer.4.output 1.33690934 234.68506316 + ------------------------------------------------------------------------------------- + TOTAL 0.58163537 98.33948517 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 406512 +BPFP 1.6316 bits/point +EBPFP 3.2632 equivalent bits/point +MSE 98.339485 +---------------------- -------------------------------------------------------- +Time: 0.749s Load: 0.008s, Pack+Encode: 0.316s, Decode+Unpack: 0.424s +---------------------- -------------------------------------------------------- +💾 Converting with 98.3395 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-266.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-266.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,588B, BPFP=0.8505 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,388B, BPFP=2.7628 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,512B, BPFP=1.4874 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,416B, BPFP=2.7062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,504B, BPFP=1.7201 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,164B, BPFP=2.6332 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,592B, BPFP=1.6087 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,692B, BPFP=2.6639 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,752B, BPFP=2.3176 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,364B, BPFP=2.5865 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,904B, BPFP=0.8321 +⌛️ [2/4] FRONTEND: Frontend time: 0.381s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.488s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15796844 23.72267994 + layer.0.v_cache 0.00001527 0.00585790 + layer.1.k_cache 0.50235566 2.91428865 + layer.1.v_cache 0.00000626 0.00255340 + layer.2.k_cache 0.02709456 0.45910303 + layer.2.v_cache 0.00002063 0.00594067 + layer.3.k_cache 0.00856834 1.74743174 + layer.3.v_cache 0.00001945 0.00680608 + layer.4.k_cache 0.00071483 0.12946066 + layer.4.v_cache 0.00005464 0.01261541 + layer.4.output 0.00491706 196.08049041 + ------------------------------------------------------------------------------------- + TOTAL 0.04301397 82.44530414 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 465876 +BPFP 1.5977 bits/point +EBPFP 3.1955 equivalent bits/point +MSE 82.445304 +---------------------- -------------------------------------------------------- +Time: 0.879s Load: 0.010s, Pack+Encode: 0.381s, Decode+Unpack: 0.488s +---------------------- -------------------------------------------------------- +💾 Converting with 82.4453 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-278.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-278.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,348B, BPFP=0.8410 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,848B, BPFP=2.6366 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,900B, BPFP=1.4428 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,068B, BPFP=2.5874 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,884B, BPFP=1.6938 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,168B, BPFP=2.5307 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,560B, BPFP=1.6104 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,764B, BPFP=2.5683 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,156B, BPFP=2.2150 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,280B, BPFP=2.4748 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,052B, BPFP=0.9185 +⌛️ [2/4] FRONTEND: Frontend time: 0.291s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.449s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13184587 22.74085260 + layer.0.v_cache 0.00001706 0.00648435 + layer.1.k_cache 0.45546043 3.12449326 + layer.1.v_cache 0.00000678 0.00257403 + layer.2.k_cache 0.02272066 0.45499888 + layer.2.v_cache 0.00001996 0.00622797 + layer.3.k_cache 0.01643563 1.77887529 + layer.3.v_cache 0.00002069 0.00747223 + layer.4.k_cache 0.00071781 0.13511998 + layer.4.v_cache 0.00005262 0.01303606 + layer.4.output 1.23455937 213.79586694 + ------------------------------------------------------------------------------------- + TOTAL 0.54524783 89.69654136 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 429028 +BPFP 1.5900 bits/point +EBPFP 3.1801 equivalent bits/point +MSE 89.696541 +---------------------- -------------------------------------------------------- +Time: 0.749s Load: 0.009s, Pack+Encode: 0.291s, Decode+Unpack: 0.449s +---------------------- -------------------------------------------------------- +💾 Converting with 89.6965 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-280.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-280.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 264, 128) +Output shape: (1, 264, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) -> torch.Size([1, 1, 264, 512]) + layer.4.output: torch.Size([1, 264, 3584]) -> torch.Size([1, 1, 264, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,976B, BPFP=0.8272 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,892B, BPFP=2.7753 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,808B, BPFP=1.4683 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,852B, BPFP=2.7138 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,200B, BPFP=1.7282 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,836B, BPFP=2.6536 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,524B, BPFP=1.6290 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,756B, BPFP=2.7081 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,448B, BPFP=2.3348 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,228B, BPFP=2.6177 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,872B, BPFP=0.8613 +⌛️ [2/4] FRONTEND: Frontend time: 0.380s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.528s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 264, 128]) + layer.0.v_cache: torch.Size([1, 4, 264, 128]) + layer.1.k_cache: torch.Size([1, 4, 264, 128]) + layer.1.v_cache: torch.Size([1, 4, 264, 128]) + layer.2.k_cache: torch.Size([1, 4, 264, 128]) + layer.2.v_cache: torch.Size([1, 4, 264, 128]) + layer.3.k_cache: torch.Size([1, 4, 264, 128]) + layer.3.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.k_cache: torch.Size([1, 4, 264, 128]) + layer.4.v_cache: torch.Size([1, 4, 264, 128]) + layer.4.output: torch.Size([1, 264, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17705467 22.99082993 + layer.0.v_cache 0.00001633 0.00605153 + layer.1.k_cache 0.50406676 3.06206651 + layer.1.v_cache 0.00000599 0.00239540 + layer.2.k_cache 0.00944760 0.38043421 + layer.2.v_cache 0.00001883 0.00598303 + layer.3.k_cache 0.04120483 1.93484266 + layer.3.v_cache 0.00002392 0.00711683 + layer.4.k_cache 0.00070100 0.13315522 + layer.4.v_cache 0.00005002 0.01274984 + layer.4.output 0.00499618 206.63436824 + ------------------------------------------------------------------------------------- + TOTAL 0.04515078 86.76330605 + (elements=2,297,856) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2297856 +Total Bytes 464392 +BPFP 1.6168 bits/point +EBPFP 3.2336 equivalent bits/point +MSE 86.763306 +---------------------- -------------------------------------------------------- +Time: 0.917s Load: 0.009s, Pack+Encode: 0.380s, Decode+Unpack: 0.528s +---------------------- -------------------------------------------------------- +💾 Converting with 86.7633 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-288.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 270, 128) +Output shape: (1, 270, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) -> torch.Size([1, 1, 270, 512]) + layer.4.output: torch.Size([1, 270, 3584]) -> torch.Size([1, 1, 270, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,196B, BPFP=0.8215 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,596B, BPFP=2.7544 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,440B, BPFP=1.4722 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,476B, BPFP=2.6896 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,920B, BPFP=1.7315 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,504B, BPFP=2.6333 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,344B, BPFP=1.6403 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,100B, BPFP=2.6678 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,092B, BPFP=2.3201 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,592B, BPFP=2.5806 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,488B, BPFP=0.8142 +⌛️ [2/4] FRONTEND: Frontend time: 0.363s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.482s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 270, 128]) + layer.0.v_cache: torch.Size([1, 4, 270, 128]) + layer.1.k_cache: torch.Size([1, 4, 270, 128]) + layer.1.v_cache: torch.Size([1, 4, 270, 128]) + layer.2.k_cache: torch.Size([1, 4, 270, 128]) + layer.2.v_cache: torch.Size([1, 4, 270, 128]) + layer.3.k_cache: torch.Size([1, 4, 270, 128]) + layer.3.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.k_cache: torch.Size([1, 4, 270, 128]) + layer.4.v_cache: torch.Size([1, 4, 270, 128]) + layer.4.output: torch.Size([1, 270, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.20501094 23.82152778 + layer.0.v_cache 0.00001516 0.00594961 + layer.1.k_cache 0.57085758 3.27062898 + layer.1.v_cache 0.00000574 0.00240913 + layer.2.k_cache 0.01259345 0.43847868 + layer.2.v_cache 0.00002099 0.00623879 + layer.3.k_cache 0.01648921 1.86541929 + layer.3.v_cache 0.00001970 0.00731053 + layer.4.k_cache 0.00070444 0.13167874 + layer.4.v_cache 0.00005121 0.01304741 + layer.4.output 0.00487421 203.64178241 + ------------------------------------------------------------------------------------- + TOTAL 0.04940517 85.59148034 + (elements=2,350,080) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2350080 +Total Bytes 466748 +BPFP 1.5889 bits/point +EBPFP 3.1778 equivalent bits/point +MSE 85.591480 +---------------------- -------------------------------------------------------- +Time: 0.855s Load: 0.010s, Pack+Encode: 0.363s, Decode+Unpack: 0.482s +---------------------- -------------------------------------------------------- +💾 Converting with 85.5915 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-294.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-294.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 271, 128) +Output shape: (1, 271, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) -> torch.Size([1, 1, 271, 512]) + layer.4.output: torch.Size([1, 271, 3584]) -> torch.Size([1, 1, 271, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,328B, BPFP=0.8261 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,012B, BPFP=2.7682 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,476B, BPFP=1.4689 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,724B, BPFP=2.6940 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,296B, BPFP=1.7468 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,916B, BPFP=2.6474 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,428B, BPFP=1.6391 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,456B, BPFP=2.6785 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,268B, BPFP=2.3217 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,832B, BPFP=2.5849 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 109,844B, BPFP=0.9048 +⌛️ [2/4] FRONTEND: Frontend time: 0.351s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.473s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 271, 128]) + layer.0.v_cache: torch.Size([1, 4, 271, 128]) + layer.1.k_cache: torch.Size([1, 4, 271, 128]) + layer.1.v_cache: torch.Size([1, 4, 271, 128]) + layer.2.k_cache: torch.Size([1, 4, 271, 128]) + layer.2.v_cache: torch.Size([1, 4, 271, 128]) + layer.3.k_cache: torch.Size([1, 4, 271, 128]) + layer.3.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.k_cache: torch.Size([1, 4, 271, 128]) + layer.4.v_cache: torch.Size([1, 4, 271, 128]) + layer.4.output: torch.Size([1, 271, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14466163 22.73331916 + layer.0.v_cache 0.00001631 0.00625742 + layer.1.k_cache 0.53789613 3.41065444 + layer.1.v_cache 0.00000653 0.00250897 + layer.2.k_cache 0.02095415 0.43872329 + layer.2.v_cache 0.00002169 0.00658941 + layer.3.k_cache 0.03083320 1.92826027 + layer.3.v_cache 0.00002083 0.00745761 + layer.4.k_cache 0.00069737 0.13502263 + layer.4.v_cache 0.00004943 0.01293697 + layer.4.output 0.00490546 199.59493608 + ------------------------------------------------------------------------------------- + TOTAL 0.04526444 83.87331075 + (elements=2,358,784) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2358784 +Total Bytes 480580 +BPFP 1.6299 bits/point +EBPFP 3.2598 equivalent bits/point +MSE 83.873311 +---------------------- -------------------------------------------------------- +Time: 0.833s Load: 0.009s, Pack+Encode: 0.351s, Decode+Unpack: 0.473s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8733 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-300.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-300.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 299, 128) +Output shape: (1, 299, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) -> torch.Size([1, 1, 299, 512]) + layer.4.output: torch.Size([1, 299, 3584]) -> torch.Size([1, 1, 299, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,684B, BPFP=0.8196 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 50,776B, BPFP=2.6534 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,444B, BPFP=1.4342 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 49,536B, BPFP=2.5886 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,928B, BPFP=1.6685 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 48,516B, BPFP=2.5353 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,036B, BPFP=1.5696 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 49,280B, BPFP=2.5753 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,452B, BPFP=2.2184 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 47,588B, BPFP=2.4868 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 113,620B, BPFP=0.8482 +⌛️ [2/4] FRONTEND: Frontend time: 0.356s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.505s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 299, 128]) + layer.0.v_cache: torch.Size([1, 4, 299, 128]) + layer.1.k_cache: torch.Size([1, 4, 299, 128]) + layer.1.v_cache: torch.Size([1, 4, 299, 128]) + layer.2.k_cache: torch.Size([1, 4, 299, 128]) + layer.2.v_cache: torch.Size([1, 4, 299, 128]) + layer.3.k_cache: torch.Size([1, 4, 299, 128]) + layer.3.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.k_cache: torch.Size([1, 4, 299, 128]) + layer.4.v_cache: torch.Size([1, 4, 299, 128]) + layer.4.output: torch.Size([1, 299, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14240686 23.17175252 + layer.0.v_cache 0.00001673 0.00575908 + layer.1.k_cache 0.61438330 3.04900470 + layer.1.v_cache 0.00000638 0.00236882 + layer.2.k_cache 0.02027981 0.46020814 + layer.2.v_cache 0.00002012 0.00598513 + layer.3.k_cache 0.02199794 1.88084958 + layer.3.v_cache 0.00002003 0.00657532 + layer.4.k_cache 0.00078937 0.12471947 + layer.4.v_cache 0.00005038 0.01215261 + layer.4.output 0.04464151 171.72363235 + ------------------------------------------------------------------------------------- + TOTAL 0.06543891 72.39910599 + (elements=2,602,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2602496 +Total Bytes 506860 +BPFP 1.5581 bits/point +EBPFP 3.1161 equivalent bits/point +MSE 72.399106 +---------------------- -------------------------------------------------------- +Time: 0.870s Load: 0.010s, Pack+Encode: 0.356s, Decode+Unpack: 0.505s +---------------------- -------------------------------------------------------- +💾 Converting with 72.3991 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-324.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-324.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 399, 128) +Output shape: (1, 399, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) -> torch.Size([1, 1, 399, 512]) + layer.4.output: torch.Size([1, 399, 3584]) -> torch.Size([1, 1, 399, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 20,844B, BPFP=0.8163 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 68,436B, BPFP=2.6800 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 36,012B, BPFP=1.4102 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 66,624B, BPFP=2.6090 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 43,348B, BPFP=1.6975 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 65,804B, BPFP=2.5769 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 40,524B, BPFP=1.5869 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 66,596B, BPFP=2.6079 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 57,432B, BPFP=2.2491 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 64,272B, BPFP=2.5169 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 156,796B, BPFP=0.8772 +⌛️ [2/4] FRONTEND: Frontend time: 0.444s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.639s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 399, 128]) + layer.0.v_cache: torch.Size([1, 4, 399, 128]) + layer.1.k_cache: torch.Size([1, 4, 399, 128]) + layer.1.v_cache: torch.Size([1, 4, 399, 128]) + layer.2.k_cache: torch.Size([1, 4, 399, 128]) + layer.2.v_cache: torch.Size([1, 4, 399, 128]) + layer.3.k_cache: torch.Size([1, 4, 399, 128]) + layer.3.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.k_cache: torch.Size([1, 4, 399, 128]) + layer.4.v_cache: torch.Size([1, 4, 399, 128]) + layer.4.output: torch.Size([1, 399, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17889012 21.16869077 + layer.0.v_cache 0.00001635 0.00579819 + layer.1.k_cache 0.94984295 3.06862371 + layer.1.v_cache 0.00000590 0.00230388 + layer.2.k_cache 0.02305104 0.45027547 + layer.2.v_cache 0.00002114 0.00655455 + layer.3.k_cache 0.01620252 1.89594286 + layer.3.v_cache 0.00002049 0.00731888 + layer.4.k_cache 0.00073175 0.12931891 + layer.4.v_cache 0.00005255 0.01261637 + layer.4.output 0.00656733 135.79843582 + ------------------------------------------------------------------------------------- + TOTAL 0.07145918 57.49038202 + (elements=3,472,896) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3472896 +Total Bytes 686688 +BPFP 1.5818 bits/point +EBPFP 3.1636 equivalent bits/point +MSE 57.490382 +---------------------- -------------------------------------------------------- +Time: 1.096s Load: 0.013s, Pack+Encode: 0.444s, Decode+Unpack: 0.639s +---------------------- -------------------------------------------------------- +💾 Converting with 57.4904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-348.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-348.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,804B, BPFP=0.8161 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,564B, BPFP=2.7353 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,252B, BPFP=1.4693 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,580B, BPFP=2.6673 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,532B, BPFP=1.7652 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,156B, BPFP=2.6380 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,052B, BPFP=1.6629 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,808B, BPFP=2.6831 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,284B, BPFP=2.3012 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,328B, BPFP=2.5808 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,572B, BPFP=0.9341 +⌛️ [2/4] FRONTEND: Frontend time: 0.346s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.472s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10670873 23.06292347 + layer.0.v_cache 0.00001672 0.00676396 + layer.1.k_cache 0.29393762 2.96734673 + layer.1.v_cache 0.00000627 0.00260077 + layer.2.k_cache 0.01243114 0.44675830 + layer.2.v_cache 0.00002105 0.00721055 + layer.3.k_cache 0.00919804 1.97652766 + layer.3.v_cache 0.00002002 0.00787325 + layer.4.k_cache 0.00070102 0.13705765 + layer.4.v_cache 0.00005024 0.01429435 + layer.4.output 1.35463108 235.31382348 + ------------------------------------------------------------------------------------- + TOTAL 0.58267697 98.57800712 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 402932 +BPFP 1.6387 bits/point +EBPFP 3.2774 equivalent bits/point +MSE 98.578007 +---------------------- -------------------------------------------------------- +Time: 0.826s Load: 0.008s, Pack+Encode: 0.346s, Decode+Unpack: 0.472s +---------------------- -------------------------------------------------------- +💾 Converting with 98.5780 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-399.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-399.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,084B, BPFP=0.8049 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,716B, BPFP=2.5662 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,544B, BPFP=1.3868 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,560B, BPFP=2.4951 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,880B, BPFP=1.6535 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,056B, BPFP=2.4641 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,472B, BPFP=1.5669 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,668B, BPFP=2.5017 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,160B, BPFP=2.1629 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,324B, BPFP=2.4190 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,480B, BPFP=0.8918 +⌛️ [2/4] FRONTEND: Frontend time: 0.331s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.485s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10633087 22.78355300 + layer.0.v_cache 0.00001891 0.00643877 + layer.1.k_cache 0.36012592 3.13914406 + layer.1.v_cache 0.00000581 0.00238047 + layer.2.k_cache 0.01407751 0.43917351 + layer.2.v_cache 0.00002023 0.00623601 + layer.3.k_cache 0.01230825 1.89035058 + layer.3.v_cache 0.00002000 0.00723178 + layer.4.k_cache 0.00071033 0.13315860 + layer.4.v_cache 0.00004856 0.01252836 + layer.4.output 1.20534739 202.35617618 + ------------------------------------------------------------------------------------- + TOTAL 0.52535871 84.99490755 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 426944 +BPFP 1.5449 bits/point +EBPFP 3.0899 equivalent bits/point +MSE 84.994908 +---------------------- -------------------------------------------------------- +Time: 0.825s Load: 0.009s, Pack+Encode: 0.331s, Decode+Unpack: 0.485s +---------------------- -------------------------------------------------------- +💾 Converting with 84.9949 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-410.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-410.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 241, 128) +Output shape: (1, 241, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) -> torch.Size([1, 1, 241, 512]) + layer.4.output: torch.Size([1, 241, 3584]) -> torch.Size([1, 1, 241, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,840B, BPFP=0.8325 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,032B, BPFP=2.6603 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,680B, BPFP=1.4704 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,200B, BPFP=2.6063 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,460B, BPFP=1.7155 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,608B, BPFP=2.5679 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,996B, BPFP=1.6206 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,964B, BPFP=2.5910 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,856B, BPFP=2.2599 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,892B, BPFP=2.5215 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,568B, BPFP=0.8759 +⌛️ [2/4] FRONTEND: Frontend time: 0.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.398s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 241, 128]) + layer.0.v_cache: torch.Size([1, 4, 241, 128]) + layer.1.k_cache: torch.Size([1, 4, 241, 128]) + layer.1.v_cache: torch.Size([1, 4, 241, 128]) + layer.2.k_cache: torch.Size([1, 4, 241, 128]) + layer.2.v_cache: torch.Size([1, 4, 241, 128]) + layer.3.k_cache: torch.Size([1, 4, 241, 128]) + layer.3.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.k_cache: torch.Size([1, 4, 241, 128]) + layer.4.v_cache: torch.Size([1, 4, 241, 128]) + layer.4.output: torch.Size([1, 241, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11142051 23.74556090 + layer.0.v_cache 0.00001650 0.00630754 + layer.1.k_cache 0.42032433 3.01386182 + layer.1.v_cache 0.00000619 0.00250928 + layer.2.k_cache 0.01443043 0.45081931 + layer.2.v_cache 0.00002030 0.00652388 + layer.3.k_cache 0.02895402 1.97449262 + layer.3.v_cache 0.00001914 0.00714153 + layer.4.k_cache 0.00067703 0.13343634 + layer.4.v_cache 0.00005693 0.01334255 + layer.4.output 1.27033357 219.53126852 + ------------------------------------------------------------------------------------- + TOTAL 0.55695649 92.12193385 + (elements=2,097,664) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2097664 +Total Bytes 416096 +BPFP 1.5869 bits/point +EBPFP 3.1738 equivalent bits/point +MSE 92.121934 +---------------------- -------------------------------------------------------- +Time: 0.730s Load: 0.008s, Pack+Encode: 0.324s, Decode+Unpack: 0.398s +---------------------- -------------------------------------------------------- +💾 Converting with 92.1219 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-430.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 302, 128) +Output shape: (1, 302, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) -> torch.Size([1, 1, 302, 512]) + layer.4.output: torch.Size([1, 302, 3584]) -> torch.Size([1, 1, 302, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,424B, BPFP=0.7980 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 51,056B, BPFP=2.6416 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,900B, BPFP=1.4435 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 49,956B, BPFP=2.5846 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,812B, BPFP=1.6976 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 49,260B, BPFP=2.5486 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,808B, BPFP=1.5940 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 49,704B, BPFP=2.5716 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,340B, BPFP=2.2423 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 48,184B, BPFP=2.4930 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 111,796B, BPFP=0.8263 +⌛️ [2/4] FRONTEND: Frontend time: 0.358s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.525s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 302, 128]) + layer.0.v_cache: torch.Size([1, 4, 302, 128]) + layer.1.k_cache: torch.Size([1, 4, 302, 128]) + layer.1.v_cache: torch.Size([1, 4, 302, 128]) + layer.2.k_cache: torch.Size([1, 4, 302, 128]) + layer.2.v_cache: torch.Size([1, 4, 302, 128]) + layer.3.k_cache: torch.Size([1, 4, 302, 128]) + layer.3.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.k_cache: torch.Size([1, 4, 302, 128]) + layer.4.v_cache: torch.Size([1, 4, 302, 128]) + layer.4.output: torch.Size([1, 302, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15012774 23.13692732 + layer.0.v_cache 0.00001504 0.00590639 + layer.1.k_cache 0.60100308 3.18429666 + layer.1.v_cache 0.00000575 0.00249371 + layer.2.k_cache 0.00814273 0.42302229 + layer.2.v_cache 0.00002032 0.00651319 + layer.3.k_cache 0.01797006 1.92970811 + layer.3.v_cache 0.00001899 0.00711154 + layer.4.k_cache 0.00070977 0.13192757 + layer.4.v_cache 0.00004868 0.01271286 + layer.4.output 0.04414053 178.06022351 + ------------------------------------------------------------------------------------- + TOTAL 0.06394387 75.01542260 + (elements=2,628,608) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2628608 +Total Bytes 510240 +BPFP 1.5529 bits/point +EBPFP 3.1058 equivalent bits/point +MSE 75.015423 +---------------------- -------------------------------------------------------- +Time: 0.893s Load: 0.009s, Pack+Encode: 0.358s, Decode+Unpack: 0.525s +---------------------- -------------------------------------------------------- +💾 Converting with 75.0154 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-466.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-466.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,212B, BPFP=0.8286 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,296B, BPFP=2.7575 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,128B, BPFP=1.4650 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,128B, BPFP=2.6894 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,700B, BPFP=1.7316 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,404B, BPFP=2.6472 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,000B, BPFP=1.6325 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,968B, BPFP=2.6800 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,652B, BPFP=2.3118 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,612B, BPFP=2.6010 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 107,980B, BPFP=0.8994 +⌛️ [2/4] FRONTEND: Frontend time: 0.393s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.473s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14350806 23.06025901 + layer.0.v_cache 0.00001655 0.00582810 + layer.1.k_cache 0.46812200 2.92428817 + layer.1.v_cache 0.00000598 0.00231873 + layer.2.k_cache 0.01096032 0.42915336 + layer.2.v_cache 0.00002019 0.00614052 + layer.3.k_cache 0.00658789 1.85065722 + layer.3.v_cache 0.00002014 0.00694262 + layer.4.k_cache 0.00067528 0.12854718 + layer.4.v_cache 0.00005785 0.01314207 + layer.4.output 0.00496594 196.56636461 + ------------------------------------------------------------------------------------- + TOTAL 0.03910211 82.61128407 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 474080 +BPFP 1.6259 bits/point +EBPFP 3.2518 equivalent bits/point +MSE 82.611284 +---------------------- -------------------------------------------------------- +Time: 0.876s Load: 0.010s, Pack+Encode: 0.393s, Decode+Unpack: 0.473s +---------------------- -------------------------------------------------------- +💾 Converting with 82.6113 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-468.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-468.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 260, 128) +Output shape: (1, 260, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) -> torch.Size([1, 1, 260, 512]) + layer.4.output: torch.Size([1, 260, 3584]) -> torch.Size([1, 1, 260, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,436B, BPFP=0.8075 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,428B, BPFP=2.7901 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 24,328B, BPFP=1.4620 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,424B, BPFP=2.7298 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,128B, BPFP=1.7505 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,748B, BPFP=2.6892 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,232B, BPFP=1.6365 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,252B, BPFP=2.7195 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,212B, BPFP=2.3565 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,964B, BPFP=2.6421 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,320B, BPFP=0.8698 +⌛️ [2/4] FRONTEND: Frontend time: 0.323s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.536s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 260, 128]) + layer.0.v_cache: torch.Size([1, 4, 260, 128]) + layer.1.k_cache: torch.Size([1, 4, 260, 128]) + layer.1.v_cache: torch.Size([1, 4, 260, 128]) + layer.2.k_cache: torch.Size([1, 4, 260, 128]) + layer.2.v_cache: torch.Size([1, 4, 260, 128]) + layer.3.k_cache: torch.Size([1, 4, 260, 128]) + layer.3.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.k_cache: torch.Size([1, 4, 260, 128]) + layer.4.v_cache: torch.Size([1, 4, 260, 128]) + layer.4.output: torch.Size([1, 260, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14548651 22.84608248 + layer.0.v_cache 0.00001584 0.00613986 + layer.1.k_cache 0.47062912 3.22814848 + layer.1.v_cache 0.00000604 0.00250442 + layer.2.k_cache 0.00619981 0.42642640 + layer.2.v_cache 0.00002003 0.00674852 + layer.3.k_cache 0.01299995 1.99511907 + layer.3.v_cache 0.00001981 0.00741437 + layer.4.k_cache 0.00067777 0.13624963 + layer.4.v_cache 0.00006192 0.01378845 + layer.4.output 0.00504520 202.33348214 + ------------------------------------------------------------------------------------- + TOTAL 0.03949607 85.00017628 + (elements=2,263,040) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2263040 +Total Bytes 460472 +BPFP 1.6278 bits/point +EBPFP 3.2556 equivalent bits/point +MSE 85.000176 +---------------------- -------------------------------------------------------- +Time: 0.868s Load: 0.009s, Pack+Encode: 0.323s, Decode+Unpack: 0.536s +---------------------- -------------------------------------------------------- +💾 Converting with 85.0002 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-489.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-489.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 272, 128) +Output shape: (1, 272, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) -> torch.Size([1, 1, 272, 512]) + layer.4.output: torch.Size([1, 272, 3584]) -> torch.Size([1, 1, 272, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,432B, BPFP=0.8290 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,892B, BPFP=2.7511 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,560B, BPFP=1.4683 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,856B, BPFP=2.6916 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,236B, BPFP=1.7369 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,820B, BPFP=2.6321 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,532B, BPFP=1.6390 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,432B, BPFP=2.6673 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,948B, BPFP=2.2948 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,932B, BPFP=2.5811 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 106,072B, BPFP=0.8705 +⌛️ [2/4] FRONTEND: Frontend time: 0.347s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.499s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 272, 128]) + layer.0.v_cache: torch.Size([1, 4, 272, 128]) + layer.1.k_cache: torch.Size([1, 4, 272, 128]) + layer.1.v_cache: torch.Size([1, 4, 272, 128]) + layer.2.k_cache: torch.Size([1, 4, 272, 128]) + layer.2.v_cache: torch.Size([1, 4, 272, 128]) + layer.3.k_cache: torch.Size([1, 4, 272, 128]) + layer.3.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.k_cache: torch.Size([1, 4, 272, 128]) + layer.4.v_cache: torch.Size([1, 4, 272, 128]) + layer.4.output: torch.Size([1, 272, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14555165 22.69876996 + layer.0.v_cache 0.00001661 0.00633318 + layer.1.k_cache 0.55468144 3.22007976 + layer.1.v_cache 0.00000621 0.00251221 + layer.2.k_cache 0.02243149 0.45443762 + layer.2.v_cache 0.00002103 0.00630439 + layer.3.k_cache 0.02028967 1.82959658 + layer.3.v_cache 0.00002036 0.00702549 + layer.4.k_cache 0.00072975 0.12913823 + layer.4.v_cache 0.00005124 0.01271916 + layer.4.output 0.00486695 199.59366794 + ------------------------------------------------------------------------------------- + TOTAL 0.04575695 83.85427013 + (elements=2,367,488) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2367488 +Total Bytes 476712 +BPFP 1.6109 bits/point +EBPFP 3.2217 equivalent bits/point +MSE 83.854270 +---------------------- -------------------------------------------------------- +Time: 0.856s Load: 0.010s, Pack+Encode: 0.347s, Decode+Unpack: 0.499s +---------------------- -------------------------------------------------------- +💾 Converting with 83.8543 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-513.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-513.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 356, 128) +Output shape: (1, 356, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) -> torch.Size([1, 1, 356, 512]) + layer.4.output: torch.Size([1, 356, 3584]) -> torch.Size([1, 1, 356, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 18,740B, BPFP=0.8225 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 60,608B, BPFP=2.6601 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 32,488B, BPFP=1.4259 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 59,272B, BPFP=2.6015 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 38,188B, BPFP=1.6761 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 57,812B, BPFP=2.5374 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 35,724B, BPFP=1.5679 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 58,416B, BPFP=2.5639 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 50,192B, BPFP=2.2029 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 56,240B, BPFP=2.4684 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 131,456B, BPFP=0.8242 +⌛️ [2/4] FRONTEND: Frontend time: 0.393s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.545s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 356, 128]) + layer.0.v_cache: torch.Size([1, 4, 356, 128]) + layer.1.k_cache: torch.Size([1, 4, 356, 128]) + layer.1.v_cache: torch.Size([1, 4, 356, 128]) + layer.2.k_cache: torch.Size([1, 4, 356, 128]) + layer.2.v_cache: torch.Size([1, 4, 356, 128]) + layer.3.k_cache: torch.Size([1, 4, 356, 128]) + layer.3.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.k_cache: torch.Size([1, 4, 356, 128]) + layer.4.v_cache: torch.Size([1, 4, 356, 128]) + layer.4.output: torch.Size([1, 356, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15223843 22.30327340 + layer.0.v_cache 0.00001740 0.00577608 + layer.1.k_cache 0.82972649 3.12140750 + layer.1.v_cache 0.00000631 0.00239400 + layer.2.k_cache 0.03086275 0.44633458 + layer.2.v_cache 0.00001960 0.00583632 + layer.3.k_cache 0.01100341 1.86674980 + layer.3.v_cache 0.00002062 0.00653527 + layer.4.k_cache 0.00073857 0.12337312 + layer.4.v_cache 0.00005279 0.01167828 + layer.4.output 0.03758925 145.05472512 + ------------------------------------------------------------------------------------- + TOTAL 0.07575360 61.36920201 + (elements=3,098,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3098624 +Total Bytes 599136 +BPFP 1.5468 bits/point +EBPFP 3.0937 equivalent bits/point +MSE 61.369202 +---------------------- -------------------------------------------------------- +Time: 0.950s Load: 0.012s, Pack+Encode: 0.393s, Decode+Unpack: 0.545s +---------------------- -------------------------------------------------------- +💾 Converting with 61.3692 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-519.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-519.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,032B, BPFP=0.8242 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,376B, BPFP=2.7829 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,092B, BPFP=1.4739 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,348B, BPFP=2.7225 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,008B, BPFP=1.7627 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,532B, BPFP=2.6746 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,740B, BPFP=1.6295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,104B, BPFP=2.7082 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,756B, BPFP=2.3353 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,592B, BPFP=2.6194 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 110,876B, BPFP=0.9304 +⌛️ [2/4] FRONTEND: Frontend time: 0.346s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.479s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12347319 23.03855403 + layer.0.v_cache 0.00001696 0.00606711 + layer.1.k_cache 0.44284580 3.16040452 + layer.1.v_cache 0.00000660 0.00248587 + layer.2.k_cache 0.00882101 0.42385107 + layer.2.v_cache 0.00002210 0.00686660 + layer.3.k_cache 0.01145097 1.95024304 + layer.3.v_cache 0.00001984 0.00707408 + layer.4.k_cache 0.00068754 0.13453620 + layer.4.v_cache 0.00006483 0.01333406 + layer.4.output 0.00500964 204.08349557 + ------------------------------------------------------------------------------------- + TOTAL 0.03661626 85.72516974 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 477456 +BPFP 1.6498 bits/point +EBPFP 3.2995 equivalent bits/point +MSE 85.725170 +---------------------- -------------------------------------------------------- +Time: 0.834s Load: 0.009s, Pack+Encode: 0.346s, Decode+Unpack: 0.479s +---------------------- -------------------------------------------------------- +💾 Converting with 85.7252 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-533.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-533.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 285, 128) +Output shape: (1, 285, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) -> torch.Size([1, 1, 285, 512]) + layer.4.output: torch.Size([1, 285, 3584]) -> torch.Size([1, 1, 285, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,276B, BPFP=0.8375 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,480B, BPFP=2.7127 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,828B, BPFP=1.4708 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,308B, BPFP=2.6485 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,764B, BPFP=1.7414 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,504B, BPFP=2.6044 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,996B, BPFP=1.6445 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,116B, BPFP=2.6379 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,720B, BPFP=2.2873 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,552B, BPFP=2.5522 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 118,024B, BPFP=0.9244 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.487s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 285, 128]) + layer.0.v_cache: torch.Size([1, 4, 285, 128]) + layer.1.k_cache: torch.Size([1, 4, 285, 128]) + layer.1.v_cache: torch.Size([1, 4, 285, 128]) + layer.2.k_cache: torch.Size([1, 4, 285, 128]) + layer.2.v_cache: torch.Size([1, 4, 285, 128]) + layer.3.k_cache: torch.Size([1, 4, 285, 128]) + layer.3.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.k_cache: torch.Size([1, 4, 285, 128]) + layer.4.v_cache: torch.Size([1, 4, 285, 128]) + layer.4.output: torch.Size([1, 285, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13002123 23.28237562 + layer.0.v_cache 0.00001833 0.00655036 + layer.1.k_cache 0.50397591 3.16611457 + layer.1.v_cache 0.00000608 0.00248638 + layer.2.k_cache 0.01926375 0.44793203 + layer.2.v_cache 0.00002060 0.00670765 + layer.3.k_cache 0.04992905 2.11798824 + layer.3.v_cache 0.00002096 0.00789826 + layer.4.k_cache 0.00069540 0.13612079 + layer.4.v_cache 0.00005467 0.01366601 + layer.4.output 0.00466491 193.06870301 + ------------------------------------------------------------------------------------- + TOTAL 0.04333296 81.21580947 + (elements=2,480,640) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2480640 +Total Bytes 503568 +BPFP 1.6240 bits/point +EBPFP 3.2480 equivalent bits/point +MSE 81.215809 +---------------------- -------------------------------------------------------- +Time: 0.815s Load: 0.009s, Pack+Encode: 0.319s, Decode+Unpack: 0.487s +---------------------- -------------------------------------------------------- +💾 Converting with 81.2158 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-547.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-547.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 280, 128) +Output shape: (1, 280, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) -> torch.Size([1, 1, 280, 512]) + layer.4.output: torch.Size([1, 280, 3584]) -> torch.Size([1, 1, 280, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,760B, BPFP=0.8237 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,296B, BPFP=2.6951 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,156B, BPFP=1.4596 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,136B, BPFP=2.6304 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,596B, BPFP=1.7074 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,192B, BPFP=2.5777 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,740B, BPFP=1.6038 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,972B, BPFP=2.6212 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,332B, BPFP=2.2507 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,412B, BPFP=2.5342 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 119,124B, BPFP=0.9496 +⌛️ [2/4] FRONTEND: Frontend time: 0.366s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.471s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 280, 128]) + layer.0.v_cache: torch.Size([1, 4, 280, 128]) + layer.1.k_cache: torch.Size([1, 4, 280, 128]) + layer.1.v_cache: torch.Size([1, 4, 280, 128]) + layer.2.k_cache: torch.Size([1, 4, 280, 128]) + layer.2.v_cache: torch.Size([1, 4, 280, 128]) + layer.3.k_cache: torch.Size([1, 4, 280, 128]) + layer.3.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.k_cache: torch.Size([1, 4, 280, 128]) + layer.4.v_cache: torch.Size([1, 4, 280, 128]) + layer.4.output: torch.Size([1, 280, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15974760 22.80947440 + layer.0.v_cache 0.00001739 0.00582684 + layer.1.k_cache 0.45644542 3.23528966 + layer.1.v_cache 0.00000643 0.00236165 + layer.2.k_cache 0.01631587 0.41208875 + layer.2.v_cache 0.00002005 0.00567798 + layer.3.k_cache 0.01511352 1.90347181 + layer.3.v_cache 0.00002063 0.00686580 + layer.4.k_cache 0.00073517 0.12600686 + layer.4.v_cache 0.00005091 0.01247384 + layer.4.output 0.00481773 193.91868622 + ------------------------------------------------------------------------------------- + TOTAL 0.04012924 81.52649066 + (elements=2,437,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2437120 +Total Bytes 493716 +BPFP 1.6207 bits/point +EBPFP 3.2413 equivalent bits/point +MSE 81.526491 +---------------------- -------------------------------------------------------- +Time: 0.848s Load: 0.010s, Pack+Encode: 0.366s, Decode+Unpack: 0.471s +---------------------- -------------------------------------------------------- +💾 Converting with 81.5265 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-559.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-559.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 267, 128) +Output shape: (1, 267, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) -> torch.Size([1, 1, 267, 512]) + layer.4.output: torch.Size([1, 267, 3584]) -> torch.Size([1, 1, 267, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,176B, BPFP=0.8296 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,300B, BPFP=2.7680 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,316B, BPFP=1.4815 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,232B, BPFP=2.7055 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,804B, BPFP=1.7441 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,268B, BPFP=2.6491 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,772B, BPFP=1.6252 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,820B, BPFP=2.6814 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,736B, BPFP=2.3254 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,364B, BPFP=2.5962 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,224B, BPFP=0.8379 +⌛️ [2/4] FRONTEND: Frontend time: 0.362s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.487s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 267, 128]) + layer.0.v_cache: torch.Size([1, 4, 267, 128]) + layer.1.k_cache: torch.Size([1, 4, 267, 128]) + layer.1.v_cache: torch.Size([1, 4, 267, 128]) + layer.2.k_cache: torch.Size([1, 4, 267, 128]) + layer.2.v_cache: torch.Size([1, 4, 267, 128]) + layer.3.k_cache: torch.Size([1, 4, 267, 128]) + layer.3.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.k_cache: torch.Size([1, 4, 267, 128]) + layer.4.v_cache: torch.Size([1, 4, 267, 128]) + layer.4.output: torch.Size([1, 267, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14148096 23.22195035 + layer.0.v_cache 0.00001588 0.00599720 + layer.1.k_cache 0.55266568 2.99756751 + layer.1.v_cache 0.00000632 0.00251475 + layer.2.k_cache 0.00687984 0.43190020 + layer.2.v_cache 0.00001885 0.00615585 + layer.3.k_cache 0.01675770 1.99659963 + layer.3.v_cache 0.00001964 0.00710485 + layer.4.k_cache 0.00068991 0.12937762 + layer.4.v_cache 0.00004859 0.01266118 + layer.4.output 0.00492290 197.12463216 + ------------------------------------------------------------------------------------- + TOTAL 0.04429669 82.86377966 + (elements=2,323,968) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2323968 +Total Bytes 466012 +BPFP 1.6042 bits/point +EBPFP 3.2084 equivalent bits/point +MSE 82.863780 +---------------------- -------------------------------------------------------- +Time: 0.858s Load: 0.009s, Pack+Encode: 0.362s, Decode+Unpack: 0.487s +---------------------- -------------------------------------------------------- +💾 Converting with 82.8638 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-569.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-569.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,804B, BPFP=0.8129 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 57,076B, BPFP=2.7610 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,256B, BPFP=1.4152 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,256B, BPFP=2.6730 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 34,964B, BPFP=1.6914 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,108B, BPFP=2.6175 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 32,492B, BPFP=1.5718 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,016B, BPFP=2.6614 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 46,704B, BPFP=2.2593 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 52,840B, BPFP=2.5561 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 123,356B, BPFP=0.8525 +⌛️ [2/4] FRONTEND: Frontend time: 0.368s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.578s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13144862 22.50793948 + layer.0.v_cache 0.00001769 0.00637975 + layer.1.k_cache 0.64513603 3.19489451 + layer.1.v_cache 0.00000607 0.00245747 + layer.2.k_cache 0.02726801 0.44308146 + layer.2.v_cache 0.00001956 0.00612376 + layer.3.k_cache 0.02271994 1.75302152 + layer.3.v_cache 0.00002018 0.00673385 + layer.4.k_cache 0.00073057 0.12967733 + layer.4.v_cache 0.00004940 0.01187936 + layer.4.output 0.04135688 160.45939297 + ------------------------------------------------------------------------------------- + TOTAL 0.06570084 67.72223172 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 557872 +BPFP 1.5875 bits/point +EBPFP 3.1749 equivalent bits/point +MSE 67.722232 +---------------------- -------------------------------------------------------- +Time: 0.957s Load: 0.011s, Pack+Encode: 0.368s, Decode+Unpack: 0.578s +---------------------- -------------------------------------------------------- +💾 Converting with 67.7222 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-622.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-622.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,492B, BPFP=0.8352 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,160B, BPFP=2.7733 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,448B, BPFP=1.4860 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,164B, BPFP=2.7009 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,124B, BPFP=1.7532 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,356B, BPFP=2.6422 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,828B, BPFP=1.6590 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,900B, BPFP=2.6817 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,056B, BPFP=2.3297 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,624B, BPFP=2.5890 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,504B, BPFP=0.8462 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.395s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12396691 22.57749364 + layer.0.v_cache 0.00001598 0.00650701 + layer.1.k_cache 0.24854917 3.04892465 + layer.1.v_cache 0.00000584 0.00244918 + layer.2.k_cache 0.02290044 0.46449294 + layer.2.v_cache 0.00002299 0.00650919 + layer.3.k_cache 0.01294550 2.05183475 + layer.3.v_cache 0.00002077 0.00736269 + layer.4.k_cache 0.00065144 0.13390700 + layer.4.v_cache 0.00005054 0.01300610 + layer.4.output 1.42385976 247.07288206 + ------------------------------------------------------------------------------------- + TOTAL 0.61036164 103.40133303 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 376656 +BPFP 1.6102 bits/point +EBPFP 3.2204 equivalent bits/point +MSE 103.401333 +---------------------- -------------------------------------------------------- +Time: 0.720s Load: 0.007s, Pack+Encode: 0.319s, Decode+Unpack: 0.395s +---------------------- -------------------------------------------------------- +💾 Converting with 103.4013 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-641.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-641.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 289, 128) +Output shape: (1, 289, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) -> torch.Size([1, 1, 289, 512]) + layer.4.output: torch.Size([1, 289, 3584]) -> torch.Size([1, 1, 289, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,668B, BPFP=0.7930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,476B, BPFP=2.6750 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,096B, BPFP=1.4109 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,120B, BPFP=2.6016 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,396B, BPFP=1.6974 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,460B, BPFP=2.5660 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,624B, BPFP=1.6016 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,332B, BPFP=2.6131 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,312B, BPFP=2.2336 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,196B, BPFP=2.4976 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 112,920B, BPFP=0.8722 +⌛️ [2/4] FRONTEND: Frontend time: 0.388s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.554s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 289, 128]) + layer.0.v_cache: torch.Size([1, 4, 289, 128]) + layer.1.k_cache: torch.Size([1, 4, 289, 128]) + layer.1.v_cache: torch.Size([1, 4, 289, 128]) + layer.2.k_cache: torch.Size([1, 4, 289, 128]) + layer.2.v_cache: torch.Size([1, 4, 289, 128]) + layer.3.k_cache: torch.Size([1, 4, 289, 128]) + layer.3.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.k_cache: torch.Size([1, 4, 289, 128]) + layer.4.v_cache: torch.Size([1, 4, 289, 128]) + layer.4.output: torch.Size([1, 289, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14783513 22.67956416 + layer.0.v_cache 0.00001600 0.00644042 + layer.1.k_cache 0.54442752 3.20501857 + layer.1.v_cache 0.00000628 0.00251322 + layer.2.k_cache 0.01848349 0.44296180 + layer.2.v_cache 0.00002061 0.00658258 + layer.3.k_cache 0.01777397 1.86418157 + layer.3.v_cache 0.00002092 0.00773837 + layer.4.k_cache 0.00070447 0.13617621 + layer.4.v_cache 0.00005049 0.01322917 + layer.4.output 0.04615090 184.54362333 + ------------------------------------------------------------------------------------- + TOTAL 0.06190560 77.65704526 + (elements=2,515,456) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2515456 +Total Bytes 495600 +BPFP 1.5762 bits/point +EBPFP 3.1524 equivalent bits/point +MSE 77.657045 +---------------------- -------------------------------------------------------- +Time: 0.952s Load: 0.010s, Pack+Encode: 0.388s, Decode+Unpack: 0.554s +---------------------- -------------------------------------------------------- +💾 Converting with 77.6570 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-644.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-644.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 298, 128) +Output shape: (1, 298, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) -> torch.Size([1, 1, 298, 512]) + layer.4.output: torch.Size([1, 298, 3584]) -> torch.Size([1, 1, 298, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,056B, BPFP=0.8419 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 51,232B, BPFP=2.6862 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,432B, BPFP=1.4383 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 49,816B, BPFP=2.6120 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 32,528B, BPFP=1.7055 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 48,880B, BPFP=2.5629 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,432B, BPFP=1.5956 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 49,652B, BPFP=2.6034 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 42,644B, BPFP=2.2359 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 47,792B, BPFP=2.5059 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 110,920B, BPFP=0.8308 +⌛️ [2/4] FRONTEND: Frontend time: 0.369s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.511s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 298, 128]) + layer.0.v_cache: torch.Size([1, 4, 298, 128]) + layer.1.k_cache: torch.Size([1, 4, 298, 128]) + layer.1.v_cache: torch.Size([1, 4, 298, 128]) + layer.2.k_cache: torch.Size([1, 4, 298, 128]) + layer.2.v_cache: torch.Size([1, 4, 298, 128]) + layer.3.k_cache: torch.Size([1, 4, 298, 128]) + layer.3.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.k_cache: torch.Size([1, 4, 298, 128]) + layer.4.v_cache: torch.Size([1, 4, 298, 128]) + layer.4.output: torch.Size([1, 298, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16274841 22.59653877 + layer.0.v_cache 0.00001794 0.00624421 + layer.1.k_cache 0.56737488 3.04101337 + layer.1.v_cache 0.00000586 0.00236428 + layer.2.k_cache 0.02027899 0.44142970 + layer.2.v_cache 0.00002117 0.00626658 + layer.3.k_cache 0.01891868 1.94093353 + layer.3.v_cache 0.00002140 0.00713485 + layer.4.k_cache 0.00069658 0.12966747 + layer.4.v_cache 0.00005005 0.01241657 + layer.4.output 0.04476916 177.55127936 + ------------------------------------------------------------------------------------- + TOTAL 0.06373636 74.76723323 + (elements=2,593,792) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2593792 +Total Bytes 507384 +BPFP 1.5649 bits/point +EBPFP 3.1298 equivalent bits/point +MSE 74.767233 +---------------------- -------------------------------------------------------- +Time: 0.889s Load: 0.009s, Pack+Encode: 0.369s, Decode+Unpack: 0.511s +---------------------- -------------------------------------------------------- +💾 Converting with 74.7672 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-650.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-650.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 273, 128) +Output shape: (1, 273, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) -> torch.Size([1, 1, 273, 512]) + layer.4.output: torch.Size([1, 273, 3584]) -> torch.Size([1, 1, 273, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,308B, BPFP=0.8189 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,868B, BPFP=2.7397 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,332B, BPFP=1.4499 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,628B, BPFP=2.6687 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,168B, BPFP=1.7266 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,652B, BPFP=2.6129 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,040B, BPFP=1.6049 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,356B, BPFP=2.6532 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,928B, BPFP=2.2853 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,596B, BPFP=2.5524 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,672B, BPFP=0.8395 +⌛️ [2/4] FRONTEND: Frontend time: 0.378s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.498s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 273, 128]) + layer.0.v_cache: torch.Size([1, 4, 273, 128]) + layer.1.k_cache: torch.Size([1, 4, 273, 128]) + layer.1.v_cache: torch.Size([1, 4, 273, 128]) + layer.2.k_cache: torch.Size([1, 4, 273, 128]) + layer.2.v_cache: torch.Size([1, 4, 273, 128]) + layer.3.k_cache: torch.Size([1, 4, 273, 128]) + layer.3.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.k_cache: torch.Size([1, 4, 273, 128]) + layer.4.v_cache: torch.Size([1, 4, 273, 128]) + layer.4.output: torch.Size([1, 273, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13366390 23.69228623 + layer.0.v_cache 0.00001719 0.00630274 + layer.1.k_cache 0.57618574 3.41469968 + layer.1.v_cache 0.00000658 0.00241018 + layer.2.k_cache 0.02858028 0.44360430 + layer.2.v_cache 0.00002021 0.00636428 + layer.3.k_cache 0.00873641 1.85881473 + layer.3.v_cache 0.00002023 0.00710153 + layer.4.k_cache 0.00067337 0.13164501 + layer.4.v_cache 0.00005011 0.01262671 + layer.4.output 0.00484872 197.96031201 + ------------------------------------------------------------------------------------- + TOTAL 0.04599383 83.25282585 + (elements=2,376,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2376192 +Total Bytes 471548 +BPFP 1.5876 bits/point +EBPFP 3.1752 equivalent bits/point +MSE 83.252826 +---------------------- -------------------------------------------------------- +Time: 0.884s Load: 0.009s, Pack+Encode: 0.378s, Decode+Unpack: 0.498s +---------------------- -------------------------------------------------------- +💾 Converting with 83.2528 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-660.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-660.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,524B, BPFP=0.8855 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,196B, BPFP=2.7712 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,328B, BPFP=1.5079 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,084B, BPFP=2.6926 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,200B, BPFP=1.7817 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,428B, BPFP=2.6462 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,940B, BPFP=1.6926 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,176B, BPFP=2.6991 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,908B, BPFP=2.3266 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,708B, BPFP=2.5953 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,660B, BPFP=0.9965 +⌛️ [2/4] FRONTEND: Frontend time: 0.308s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.421s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13057878 22.80702285 + layer.0.v_cache 0.00001797 0.00702014 + layer.1.k_cache 0.31658528 3.22605986 + layer.1.v_cache 0.00000647 0.00265321 + layer.2.k_cache 0.02849985 0.47069543 + layer.2.v_cache 0.00001991 0.00689267 + layer.3.k_cache 0.03599559 1.95265867 + layer.3.v_cache 0.00002239 0.00826433 + layer.4.k_cache 0.00078811 0.14406318 + layer.4.v_cache 0.00005540 0.01445959 + layer.4.output 1.38530015 242.86580882 + ------------------------------------------------------------------------------------- + TOTAL 0.60056887 101.68826187 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 404152 +BPFP 1.6808 bits/point +EBPFP 3.3617 equivalent bits/point +MSE 101.688262 +---------------------- -------------------------------------------------------- +Time: 0.738s Load: 0.008s, Pack+Encode: 0.308s, Decode+Unpack: 0.421s +---------------------- -------------------------------------------------------- +💾 Converting with 101.6883 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-666.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-666.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,500B, BPFP=0.8281 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,680B, BPFP=2.7851 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,744B, BPFP=1.4937 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,584B, BPFP=2.7062 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,688B, BPFP=1.7776 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,972B, BPFP=2.6622 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,084B, BPFP=1.6622 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,656B, BPFP=2.7114 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,416B, BPFP=2.3341 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,220B, BPFP=2.6080 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,584B, BPFP=0.9421 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.431s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12725546 22.78103399 + layer.0.v_cache 0.00001736 0.00706024 + layer.1.k_cache 0.32262681 3.08218257 + layer.1.v_cache 0.00000608 0.00259746 + layer.2.k_cache 0.02157613 0.44855390 + layer.2.v_cache 0.00002208 0.00680044 + layer.3.k_cache 0.01760322 1.92571346 + layer.3.v_cache 0.00002192 0.00812639 + layer.4.k_cache 0.00069785 0.14126188 + layer.4.v_cache 0.00005172 0.01417069 + layer.4.output 1.41081570 243.94147054 + ------------------------------------------------------------------------------------- + TOTAL 0.60974050 102.11810558 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 391128 +BPFP 1.6566 bits/point +EBPFP 3.3133 equivalent bits/point +MSE 102.118106 +---------------------- -------------------------------------------------------- +Time: 0.781s Load: 0.008s, Pack+Encode: 0.342s, Decode+Unpack: 0.431s +---------------------- -------------------------------------------------------- +💾 Converting with 102.1181 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-683.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-683.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,372B, BPFP=0.8342 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,404B, BPFP=2.8172 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,576B, BPFP=1.5094 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,404B, BPFP=2.7438 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,736B, BPFP=1.8146 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,736B, BPFP=2.6948 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,028B, BPFP=1.6893 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,396B, BPFP=2.7433 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,192B, BPFP=2.3615 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,996B, BPFP=2.6406 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,592B, BPFP=0.9703 +⌛️ [2/4] FRONTEND: Frontend time: 0.291s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.417s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13788442 23.21257657 + layer.0.v_cache 0.00001717 0.00727653 + layer.1.k_cache 0.30723858 3.15928628 + layer.1.v_cache 0.00000664 0.00274692 + layer.2.k_cache 0.01337219 0.45896990 + layer.2.v_cache 0.00002159 0.00713837 + layer.3.k_cache 0.03522867 1.80890580 + layer.3.v_cache 0.00002126 0.00848643 + layer.4.k_cache 0.00073009 0.14358313 + layer.4.v_cache 0.00005119 0.01447601 + layer.4.output 1.43730973 252.02427062 + ------------------------------------------------------------------------------------- + TOTAL 0.62092588 105.47019649 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 390432 +BPFP 1.6848 bits/point +EBPFP 3.3695 equivalent bits/point +MSE 105.470196 +---------------------- -------------------------------------------------------- +Time: 0.717s Load: 0.008s, Pack+Encode: 0.291s, Decode+Unpack: 0.417s +---------------------- -------------------------------------------------------- +💾 Converting with 105.4702 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-696.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,056B, BPFP=0.8468 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,548B, BPFP=2.8759 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,984B, BPFP=1.5306 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,636B, BPFP=2.8061 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,988B, BPFP=1.8373 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,052B, BPFP=2.7613 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,372B, BPFP=1.7135 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,736B, BPFP=2.8137 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,432B, BPFP=2.4075 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,264B, BPFP=2.7010 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 91,300B, BPFP=0.9990 +⌛️ [2/4] FRONTEND: Frontend time: 0.325s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.430s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11546621 24.21091596 + layer.0.v_cache 0.00001823 0.00681370 + layer.1.k_cache 0.26924565 3.13817073 + layer.1.v_cache 0.00000737 0.00273899 + layer.2.k_cache 0.01761903 0.45605573 + layer.2.v_cache 0.00002145 0.00679929 + layer.3.k_cache 0.02374693 1.75816734 + layer.3.v_cache 0.00002104 0.00776470 + layer.4.k_cache 0.00067272 0.13691996 + layer.4.v_cache 0.00005153 0.01353619 + layer.4.output 1.50071327 253.73328081 + ------------------------------------------------------------------------------------- + TOTAL 0.64305077 106.22769696 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 382368 +BPFP 1.7228 bits/point +EBPFP 3.4455 equivalent bits/point +MSE 106.227697 +---------------------- -------------------------------------------------------- +Time: 0.761s Load: 0.006s, Pack+Encode: 0.325s, Decode+Unpack: 0.430s +---------------------- -------------------------------------------------------- +💾 Converting with 106.2277 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-727.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-727.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,684B, BPFP=0.7957 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 52,204B, BPFP=2.6483 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,192B, BPFP=1.4302 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,964B, BPFP=2.5854 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,576B, BPFP=1.7033 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 50,196B, BPFP=2.5465 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,656B, BPFP=1.6059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 50,652B, BPFP=2.5696 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,908B, BPFP=2.2275 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 49,068B, BPFP=2.4892 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 121,416B, BPFP=0.8799 +⌛️ [2/4] FRONTEND: Frontend time: 0.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.558s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18206066 21.90213132 + layer.0.v_cache 0.00001652 0.00636964 + layer.1.k_cache 0.62405802 3.12695372 + layer.1.v_cache 0.00000616 0.00251656 + layer.2.k_cache 0.02036299 0.42469401 + layer.2.v_cache 0.00002087 0.00651971 + layer.3.k_cache 0.02546444 1.79014409 + layer.3.v_cache 0.00002064 0.00725961 + layer.4.k_cache 0.00069880 0.13208423 + layer.4.v_cache 0.00007146 0.01339131 + layer.4.output 0.04336799 168.26252319 + ------------------------------------------------------------------------------------- + TOTAL 0.06802097 70.89704274 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 527516 +BPFP 1.5742 bits/point +EBPFP 3.1484 equivalent bits/point +MSE 70.897043 +---------------------- -------------------------------------------------------- +Time: 0.903s Load: 0.011s, Pack+Encode: 0.334s, Decode+Unpack: 0.558s +---------------------- -------------------------------------------------------- +💾 Converting with 70.8970 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-8.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-8.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,272B, BPFP=0.8373 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,540B, BPFP=2.7661 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,052B, BPFP=1.5046 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,580B, BPFP=2.7006 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,880B, BPFP=1.7658 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,684B, BPFP=2.6395 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,312B, BPFP=1.6588 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,276B, BPFP=2.6799 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,632B, BPFP=2.2948 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,716B, BPFP=2.5734 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,216B, BPFP=0.9184 +⌛️ [2/4] FRONTEND: Frontend time: 0.333s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.451s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14018573 23.65737795 + layer.0.v_cache 0.00001653 0.00651184 + layer.1.k_cache 0.40671090 3.07145591 + layer.1.v_cache 0.00000662 0.00258082 + layer.2.k_cache 0.03036383 0.46763711 + layer.2.v_cache 0.00002088 0.00642445 + layer.3.k_cache 0.05940984 1.84197145 + layer.3.v_cache 0.00002025 0.00743339 + layer.4.k_cache 0.00071161 0.13691985 + layer.4.v_cache 0.00005094 0.01235858 + layer.4.output 1.33692960 234.69972318 + ------------------------------------------------------------------------------------- + TOTAL 0.58800025 98.35933727 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 408160 +BPFP 1.6382 bits/point +EBPFP 3.2764 equivalent bits/point +MSE 98.359337 +---------------------- -------------------------------------------------------- +Time: 0.792s Load: 0.007s, Pack+Encode: 0.333s, Decode+Unpack: 0.451s +---------------------- -------------------------------------------------------- +💾 Converting with 98.3593 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-806.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 219, 128) +Output shape: (1, 219, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) -> torch.Size([1, 1, 219, 512]) + layer.4.output: torch.Size([1, 219, 3584]) -> torch.Size([1, 1, 219, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,668B, BPFP=0.8325 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,944B, BPFP=2.7785 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,276B, BPFP=1.5180 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,892B, BPFP=2.7035 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,356B, BPFP=1.8091 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,440B, BPFP=2.6712 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,580B, BPFP=1.6824 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,932B, BPFP=2.7063 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,780B, BPFP=2.3388 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,608B, BPFP=2.6119 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,824B, BPFP=0.9869 +⌛️ [2/4] FRONTEND: Frontend time: 0.291s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 219, 128]) + layer.0.v_cache: torch.Size([1, 4, 219, 128]) + layer.1.k_cache: torch.Size([1, 4, 219, 128]) + layer.1.v_cache: torch.Size([1, 4, 219, 128]) + layer.2.k_cache: torch.Size([1, 4, 219, 128]) + layer.2.v_cache: torch.Size([1, 4, 219, 128]) + layer.3.k_cache: torch.Size([1, 4, 219, 128]) + layer.3.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.k_cache: torch.Size([1, 4, 219, 128]) + layer.4.v_cache: torch.Size([1, 4, 219, 128]) + layer.4.output: torch.Size([1, 219, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351684 23.52267944 + layer.0.v_cache 0.00001872 0.00703060 + layer.1.k_cache 0.28219827 3.10858544 + layer.1.v_cache 0.00000627 0.00282678 + layer.2.k_cache 0.01639774 0.46051795 + layer.2.v_cache 0.00002172 0.00770133 + layer.3.k_cache 0.02752649 1.92618564 + layer.3.v_cache 0.00002090 0.00831715 + layer.4.k_cache 0.00067264 0.14317216 + layer.4.v_cache 0.00005058 0.01430048 + layer.4.output 1.39794763 237.48340672 + ------------------------------------------------------------------------------------- + TOTAL 0.60212139 99.50500965 + (elements=1,906,176) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1906176 +Total Bytes 400300 +BPFP 1.6800 bits/point +EBPFP 3.3600 equivalent bits/point +MSE 99.505010 +---------------------- -------------------------------------------------------- +Time: 0.676s Load: 0.008s, Pack+Encode: 0.291s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 99.5050 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-849.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-849.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 268, 128) +Output shape: (1, 268, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) -> torch.Size([1, 1, 268, 512]) + layer.4.output: torch.Size([1, 268, 3584]) -> torch.Size([1, 1, 268, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,328B, BPFP=0.8354 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,552B, BPFP=2.7724 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,132B, BPFP=1.4653 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,472B, BPFP=2.7094 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 29,728B, BPFP=1.7332 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,468B, BPFP=2.6509 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 27,732B, BPFP=1.6168 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,996B, BPFP=2.6817 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 39,564B, BPFP=2.3067 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,488B, BPFP=2.5938 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 109,660B, BPFP=0.9133 +⌛️ [2/4] FRONTEND: Frontend time: 0.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.475s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 268, 128]) + layer.0.v_cache: torch.Size([1, 4, 268, 128]) + layer.1.k_cache: torch.Size([1, 4, 268, 128]) + layer.1.v_cache: torch.Size([1, 4, 268, 128]) + layer.2.k_cache: torch.Size([1, 4, 268, 128]) + layer.2.v_cache: torch.Size([1, 4, 268, 128]) + layer.3.k_cache: torch.Size([1, 4, 268, 128]) + layer.3.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.k_cache: torch.Size([1, 4, 268, 128]) + layer.4.v_cache: torch.Size([1, 4, 268, 128]) + layer.4.output: torch.Size([1, 268, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15646637 23.53654639 + layer.0.v_cache 0.00001727 0.00631660 + layer.1.k_cache 0.53079696 3.08673369 + layer.1.v_cache 0.00000635 0.00267289 + layer.2.k_cache 0.02340859 0.45423673 + layer.2.v_cache 0.00002061 0.00658352 + layer.3.k_cache 0.01050839 1.80767207 + layer.3.v_cache 0.00002156 0.00699355 + layer.4.k_cache 0.00068596 0.13268303 + layer.4.v_cache 0.00006072 0.01300991 + layer.4.output 0.00497423 196.97947761 + ------------------------------------------------------------------------------------- + TOTAL 0.04451838 82.81822304 + (elements=2,332,672) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2332672 +Total Bytes 476120 +BPFP 1.6329 bits/point +EBPFP 3.2657 equivalent bits/point +MSE 82.818223 +---------------------- -------------------------------------------------------- +Time: 0.819s Load: 0.009s, Pack+Encode: 0.334s, Decode+Unpack: 0.475s +---------------------- -------------------------------------------------------- +💾 Converting with 82.8182 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-863.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-863.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,268B, BPFP=0.8253 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 52,128B, BPFP=2.6445 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,492B, BPFP=1.4454 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 51,116B, BPFP=2.5931 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,480B, BPFP=1.6985 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 49,992B, BPFP=2.5361 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,288B, BPFP=1.5873 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 50,452B, BPFP=2.5595 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,580B, BPFP=2.2108 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 48,612B, BPFP=2.4661 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 124,940B, BPFP=0.9055 +⌛️ [2/4] FRONTEND: Frontend time: 0.330s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.479s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15292851 22.18009651 + layer.0.v_cache 0.00001614 0.00668055 + layer.1.k_cache 0.61736927 3.19976826 + layer.1.v_cache 0.00000669 0.00271781 + layer.2.k_cache 0.02003183 0.45182077 + layer.2.v_cache 0.00002344 0.00665333 + layer.3.k_cache 0.01028479 1.85583655 + layer.3.v_cache 0.00002212 0.00704934 + layer.4.k_cache 0.00070434 0.13619418 + layer.4.v_cache 0.00006490 0.01294768 + layer.4.output 0.04342050 167.82877725 + ------------------------------------------------------------------------------------- + TOTAL 0.06502327 70.74477681 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 530348 +BPFP 1.5826 bits/point +EBPFP 3.1653 equivalent bits/point +MSE 70.744777 +---------------------- -------------------------------------------------------- +Time: 0.819s Load: 0.010s, Pack+Encode: 0.330s, Decode+Unpack: 0.479s +---------------------- -------------------------------------------------------- +💾 Converting with 70.7448 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-866.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-866.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 269, 128) +Output shape: (1, 269, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) -> torch.Size([1, 1, 269, 512]) + layer.4.output: torch.Size([1, 269, 3584]) -> torch.Size([1, 1, 269, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,416B, BPFP=0.8374 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,892B, BPFP=2.7818 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,608B, BPFP=1.4875 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,704B, BPFP=2.7128 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,296B, BPFP=1.7598 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,880B, BPFP=2.6650 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,364B, BPFP=1.6475 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,684B, BPFP=2.7117 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,092B, BPFP=2.3288 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,824B, BPFP=2.6036 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,168B, BPFP=0.8478 +⌛️ [2/4] FRONTEND: Frontend time: 0.354s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.510s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 269, 128]) + layer.0.v_cache: torch.Size([1, 4, 269, 128]) + layer.1.k_cache: torch.Size([1, 4, 269, 128]) + layer.1.v_cache: torch.Size([1, 4, 269, 128]) + layer.2.k_cache: torch.Size([1, 4, 269, 128]) + layer.2.v_cache: torch.Size([1, 4, 269, 128]) + layer.3.k_cache: torch.Size([1, 4, 269, 128]) + layer.3.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.k_cache: torch.Size([1, 4, 269, 128]) + layer.4.v_cache: torch.Size([1, 4, 269, 128]) + layer.4.output: torch.Size([1, 269, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12950333 22.94738906 + layer.0.v_cache 0.00001690 0.00665615 + layer.1.k_cache 0.52417032 3.27529442 + layer.1.v_cache 0.00000602 0.00264247 + layer.2.k_cache 0.01578268 0.45202444 + layer.2.v_cache 0.00002268 0.00690344 + layer.3.k_cache 0.03345619 1.90063454 + layer.3.v_cache 0.00002007 0.00784872 + layer.4.k_cache 0.00067301 0.13564784 + layer.4.v_cache 0.00005077 0.01361589 + layer.4.output 0.00489311 204.31220127 + ------------------------------------------------------------------------------------- + TOTAL 0.04340905 85.81965093 + (elements=2,341,376) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2341376 +Total Bytes 472928 +BPFP 1.6159 bits/point +EBPFP 3.2318 equivalent bits/point +MSE 85.819651 +---------------------- -------------------------------------------------------- +Time: 0.873s Load: 0.009s, Pack+Encode: 0.354s, Decode+Unpack: 0.510s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8197 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-897.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-897.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 221, 128) +Output shape: (1, 221, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) -> torch.Size([1, 1, 221, 512]) + layer.4.output: torch.Size([1, 221, 3584]) -> torch.Size([1, 1, 221, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,480B, BPFP=0.8824 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,780B, BPFP=2.7418 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,268B, BPFP=1.5037 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,732B, BPFP=2.6677 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,860B, BPFP=1.7576 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,024B, BPFP=2.6176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,500B, BPFP=1.6615 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,676B, BPFP=2.6637 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,580B, BPFP=2.3035 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,368B, BPFP=2.5713 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,568B, BPFP=0.9148 +⌛️ [2/4] FRONTEND: Frontend time: 0.309s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.424s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 221, 128]) + layer.0.v_cache: torch.Size([1, 4, 221, 128]) + layer.1.k_cache: torch.Size([1, 4, 221, 128]) + layer.1.v_cache: torch.Size([1, 4, 221, 128]) + layer.2.k_cache: torch.Size([1, 4, 221, 128]) + layer.2.v_cache: torch.Size([1, 4, 221, 128]) + layer.3.k_cache: torch.Size([1, 4, 221, 128]) + layer.3.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.k_cache: torch.Size([1, 4, 221, 128]) + layer.4.v_cache: torch.Size([1, 4, 221, 128]) + layer.4.output: torch.Size([1, 221, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13888381 23.39152865 + layer.0.v_cache 0.00001731 0.00634197 + layer.1.k_cache 0.27925331 3.08691517 + layer.1.v_cache 0.00000589 0.00246434 + layer.2.k_cache 0.03172064 0.44299572 + layer.2.v_cache 0.00001980 0.00646747 + layer.3.k_cache 0.01738505 1.84568628 + layer.3.v_cache 0.00001961 0.00734590 + layer.4.k_cache 0.00068340 0.13708463 + layer.4.v_cache 0.00005063 0.01336977 + layer.4.output 1.38525280 242.95618536 + ------------------------------------------------------------------------------------- + TOTAL 0.59792994 101.74314691 + (elements=1,923,584) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1923584 +Total Bytes 392836 +BPFP 1.6338 bits/point +EBPFP 3.2675 equivalent bits/point +MSE 101.743147 +---------------------- -------------------------------------------------------- +Time: 0.741s Load: 0.008s, Pack+Encode: 0.309s, Decode+Unpack: 0.424s +---------------------- -------------------------------------------------------- +💾 Converting with 101.7431 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-clean/LibriSpeech-test_clean-991.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-clean/LibriSpeech-test_clean-991.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.6087 bits/point +Avg EBPFP 3.2174 equivalent bits/point +Avg MSE 84.904617 +Avg Time 0.867s +------------------------ ---------------------------- diff --git a/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log b/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log new file mode 100644 index 0000000000000000000000000000000000000000..a98100beaf57a4ae8430bdef3a99c6d0735bf436 --- /dev/null +++ b/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log @@ -0,0 +1,14258 @@ +Experiment: dtufc_hyperprior-featurecoding_kimiaudio_individual +Log file: output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/dtufc_hyperprior-featurecoding_kimiaudio_individual.log +DTUFCCodecConfig: + arch: hyperprior-featurecoding + handler: kimiaudio + checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.02_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar + transform_type: kmeans + transform_mapping:featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json + bit_depth: 8 + device: cuda:0 +Loading checkpoint: codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.02_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Checkpoint epoch: 598 +Loaded hyperprior-featurecoding (1-channel) on cuda:0 +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json: torch.Size([256]) +Loaded per-key quantization points for key 'output' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_feature.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.0.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_0_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.1.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_1_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.2.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_2_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.3.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_3_v.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.k_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_k.json +Loaded quantization points from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json: torch.Size([256]) +Loaded per-key quantization points for key 'layer.4.v_cache' from featurecoding_utils/transform_mapping/kmeans10samples-8bits/kimiaudio/librispeech_fewshot-8bit_layer_4_v.json +Loaded per-key mappings: model=kimiaudio + Keys: ['output', 'layer.0.k_cache', 'layer.0.v_cache', 'layer.1.k_cache', 'layer.1.v_cache', 'layer.2.k_cache', 'layer.2.v_cache', 'layer.3.k_cache', 'layer.3.v_cache', 'layer.4.k_cache', 'layer.4.v_cache'] +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +Checkpoint codec_weights/hyperprior_hybrid/bmshj2018-hyperprior_lambda0.02_epochs600_lr0.0001_bs180_patch256-256_checkpoint_best.pth.tar +Transform type kmeans +Transform config featurecoding_utils/transform_mapping/kmeans10samples-8bits/mapping.json +Input ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other +Output output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other +---------------- ----------------------------------------------------------------------------------------------------------------------------- +Files found: 100 +---------------------------------------------------------------------- + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst (1/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 201, 128) +Output shape: (1, 201, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) -> torch.Size([1, 1, 201, 512]) + layer.4.output: torch.Size([1, 201, 3584]) -> torch.Size([1, 1, 201, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,940B, BPFP=0.8504 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,004B, BPFP=2.8766 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,572B, BPFP=1.5215 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,116B, BPFP=2.8075 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,360B, BPFP=1.8159 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,552B, BPFP=2.7637 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,880B, BPFP=1.7009 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,304B, BPFP=2.8221 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,140B, BPFP=2.4207 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,096B, BPFP=2.7282 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,868B, BPFP=1.0091 +⌛️ [2/4] FRONTEND: Frontend time: 0.872s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.509s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 201, 128]) + layer.0.v_cache: torch.Size([1, 4, 201, 128]) + layer.1.k_cache: torch.Size([1, 4, 201, 128]) + layer.1.v_cache: torch.Size([1, 4, 201, 128]) + layer.2.k_cache: torch.Size([1, 4, 201, 128]) + layer.2.v_cache: torch.Size([1, 4, 201, 128]) + layer.3.k_cache: torch.Size([1, 4, 201, 128]) + layer.3.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.k_cache: torch.Size([1, 4, 201, 128]) + layer.4.v_cache: torch.Size([1, 4, 201, 128]) + layer.4.output: torch.Size([1, 201, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09475397 23.93494199 + layer.0.v_cache 0.00001927 0.00705497 + layer.1.k_cache 0.17401673 3.24735605 + layer.1.v_cache 0.00000670 0.00269222 + layer.2.k_cache 0.00814445 0.42541307 + layer.2.v_cache 0.00002157 0.00710616 + layer.3.k_cache 0.01691350 1.80243968 + layer.3.v_cache 0.00002161 0.00896660 + layer.4.k_cache 0.00069969 0.15544556 + layer.4.v_cache 0.00005477 0.01496522 + layer.4.output 1.52309445 264.21264215 + ------------------------------------------------------------------------------------- + TOTAL 0.64448903 110.53499274 + (elements=1,749,504) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1749504 +Total Bytes 377832 +BPFP 1.7277 bits/point +EBPFP 3.4554 equivalent bits/point +MSE 110.534993 +---------------------- -------------------------------------------------------- +Time: 1.389s Load: 0.008s, Pack+Encode: 0.872s, Decode+Unpack: 0.509s +---------------------- -------------------------------------------------------- +💾 Converting with 110.5350 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2620.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2620.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst (2/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 212, 128) +Output shape: (1, 212, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) -> torch.Size([1, 1, 212, 512]) + layer.4.output: torch.Size([1, 212, 3584]) -> torch.Size([1, 1, 212, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,492B, BPFP=0.8470 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,320B, BPFP=2.8243 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,244B, BPFP=1.4920 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,216B, BPFP=2.7429 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,648B, BPFP=1.8166 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,780B, BPFP=2.7108 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,140B, BPFP=1.7055 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,516B, BPFP=2.7650 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,168B, BPFP=2.3709 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,028B, BPFP=2.6554 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,808B, BPFP=0.9561 +⌛️ [2/4] FRONTEND: Frontend time: 0.349s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.461s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 212, 128]) + layer.0.v_cache: torch.Size([1, 4, 212, 128]) + layer.1.k_cache: torch.Size([1, 4, 212, 128]) + layer.1.v_cache: torch.Size([1, 4, 212, 128]) + layer.2.k_cache: torch.Size([1, 4, 212, 128]) + layer.2.v_cache: torch.Size([1, 4, 212, 128]) + layer.3.k_cache: torch.Size([1, 4, 212, 128]) + layer.3.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.k_cache: torch.Size([1, 4, 212, 128]) + layer.4.v_cache: torch.Size([1, 4, 212, 128]) + layer.4.output: torch.Size([1, 212, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09403566 23.05754118 + layer.0.v_cache 0.00001799 0.00727510 + layer.1.k_cache 0.24957257 3.15594943 + layer.1.v_cache 0.00000621 0.00273856 + layer.2.k_cache 0.01169950 0.40051957 + layer.2.v_cache 0.00002031 0.00716789 + layer.3.k_cache 0.03645871 1.96206867 + layer.3.v_cache 0.00002058 0.00855439 + layer.4.k_cache 0.00069161 0.14802989 + layer.4.v_cache 0.00005130 0.01408999 + layer.4.output 1.44404146 241.96898164 + ------------------------------------------------------------------------------------- + TOTAL 0.61769792 101.32628271 + (elements=1,845,248) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1845248 +Total Bytes 388360 +BPFP 1.6837 bits/point +EBPFP 3.3674 equivalent bits/point +MSE 101.326283 +---------------------- -------------------------------------------------------- +Time: 0.818s Load: 0.008s, Pack+Encode: 0.349s, Decode+Unpack: 0.461s +---------------------- -------------------------------------------------------- +💾 Converting with 101.3263 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2669.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2669.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst (3/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 250, 128) +Output shape: (1, 250, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) -> torch.Size([1, 1, 250, 512]) + layer.4.output: torch.Size([1, 250, 3584]) -> torch.Size([1, 1, 250, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,824B, BPFP=0.8015 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,668B, BPFP=2.6042 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,700B, BPFP=1.4187 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,660B, BPFP=2.5412 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,952B, BPFP=1.6845 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,020B, BPFP=2.5013 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,376B, BPFP=1.5860 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,520B, BPFP=2.5325 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,208B, BPFP=2.2005 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,252B, BPFP=2.4533 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,880B, BPFP=0.8829 +⌛️ [2/4] FRONTEND: Frontend time: 0.337s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.410s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 250, 128]) + layer.0.v_cache: torch.Size([1, 4, 250, 128]) + layer.1.k_cache: torch.Size([1, 4, 250, 128]) + layer.1.v_cache: torch.Size([1, 4, 250, 128]) + layer.2.k_cache: torch.Size([1, 4, 250, 128]) + layer.2.v_cache: torch.Size([1, 4, 250, 128]) + layer.3.k_cache: torch.Size([1, 4, 250, 128]) + layer.3.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.k_cache: torch.Size([1, 4, 250, 128]) + layer.4.v_cache: torch.Size([1, 4, 250, 128]) + layer.4.output: torch.Size([1, 250, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09121272 22.78169922 + layer.0.v_cache 0.00001793 0.00647244 + layer.1.k_cache 0.34710925 3.14870239 + layer.1.v_cache 0.00000615 0.00250979 + layer.2.k_cache 0.02211576 0.45080649 + layer.2.v_cache 0.00002179 0.00667916 + layer.3.k_cache 0.02174057 1.88204565 + layer.3.v_cache 0.00001938 0.00716153 + layer.4.k_cache 0.00073173 0.13297560 + layer.4.v_cache 0.00005145 0.01291320 + layer.4.output 1.22463198 211.73305357 + ------------------------------------------------------------------------------------- + TOTAL 0.53267356 88.85666709 + (elements=2,176,000) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2176000 +Total Bytes 424060 +BPFP 1.5590 bits/point +EBPFP 3.1181 equivalent bits/point +MSE 88.856667 +---------------------- -------------------------------------------------------- +Time: 0.757s Load: 0.011s, Pack+Encode: 0.337s, Decode+Unpack: 0.410s +---------------------- -------------------------------------------------------- +💾 Converting with 88.8567 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2760.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2760.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst (4/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,080B, BPFP=0.8315 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,008B, BPFP=2.7539 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,688B, BPFP=1.4928 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,060B, BPFP=2.6886 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,592B, BPFP=1.7616 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,356B, BPFP=2.6401 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,140B, BPFP=1.6616 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,856B, BPFP=2.6746 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,688B, BPFP=2.3188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,476B, BPFP=2.5796 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,520B, BPFP=0.8803 +⌛️ [2/4] FRONTEND: Frontend time: 0.328s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.412s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12780563 23.20805298 + layer.0.v_cache 0.00001798 0.00657834 + layer.1.k_cache 0.36413978 3.19396435 + layer.1.v_cache 0.00000619 0.00253620 + layer.2.k_cache 0.01529831 0.42868842 + layer.2.v_cache 0.00002052 0.00670608 + layer.3.k_cache 0.01269315 1.94353226 + layer.3.v_cache 0.00002023 0.00744806 + layer.4.k_cache 0.00069261 0.13432319 + layer.4.v_cache 0.00005545 0.01315235 + layer.4.output 1.34865724 227.24976400 + ------------------------------------------------------------------------------------- + TOTAL 0.58596180 95.27607825 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 400464 +BPFP 1.6215 bits/point +EBPFP 3.2429 equivalent bits/point +MSE 95.276078 +---------------------- -------------------------------------------------------- +Time: 0.748s Load: 0.009s, Pack+Encode: 0.328s, Decode+Unpack: 0.412s +---------------------- -------------------------------------------------------- +💾 Converting with 95.2761 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2765.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2765.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst (5/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 214, 128) +Output shape: (1, 214, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) -> torch.Size([1, 1, 214, 512]) + layer.4.output: torch.Size([1, 214, 3584]) -> torch.Size([1, 1, 214, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,396B, BPFP=0.8321 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,016B, BPFP=2.7757 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,176B, BPFP=1.4731 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,104B, BPFP=2.7091 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,100B, BPFP=1.7596 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,412B, BPFP=2.6586 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,636B, BPFP=1.6527 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,864B, BPFP=2.6916 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,912B, BPFP=2.3300 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,560B, BPFP=2.5964 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,488B, BPFP=0.8500 +⌛️ [2/4] FRONTEND: Frontend time: 0.344s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.475s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 214, 128]) + layer.0.v_cache: torch.Size([1, 4, 214, 128]) + layer.1.k_cache: torch.Size([1, 4, 214, 128]) + layer.1.v_cache: torch.Size([1, 4, 214, 128]) + layer.2.k_cache: torch.Size([1, 4, 214, 128]) + layer.2.v_cache: torch.Size([1, 4, 214, 128]) + layer.3.k_cache: torch.Size([1, 4, 214, 128]) + layer.3.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.k_cache: torch.Size([1, 4, 214, 128]) + layer.4.v_cache: torch.Size([1, 4, 214, 128]) + layer.4.output: torch.Size([1, 214, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12820032 23.22147205 + layer.0.v_cache 0.00001693 0.00629779 + layer.1.k_cache 0.24875343 3.16049964 + layer.1.v_cache 0.00000587 0.00236707 + layer.2.k_cache 0.01453762 0.45202594 + layer.2.v_cache 0.00001988 0.00644579 + layer.3.k_cache 0.02276520 1.77080928 + layer.3.v_cache 0.00001932 0.00740414 + layer.4.k_cache 0.00071399 0.13905913 + layer.4.v_cache 0.00004895 0.01273511 + layer.4.output 1.43051836 250.66453188 + ------------------------------------------------------------------------------------- + TOTAL 0.61345353 104.90769642 + (elements=1,862,656) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1862656 +Total Bytes 375664 +BPFP 1.6135 bits/point +EBPFP 3.2269 equivalent bits/point +MSE 104.907696 +---------------------- -------------------------------------------------------- +Time: 0.827s Load: 0.008s, Pack+Encode: 0.344s, Decode+Unpack: 0.475s +---------------------- -------------------------------------------------------- +💾 Converting with 104.9077 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2771.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2771.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst (6/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 277, 128) +Output shape: (1, 277, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) -> torch.Size([1, 1, 277, 512]) + layer.4.output: torch.Size([1, 277, 3584]) -> torch.Size([1, 1, 277, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,980B, BPFP=0.8450 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,144B, BPFP=2.7157 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,656B, BPFP=1.4472 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,076B, BPFP=2.6555 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,604B, BPFP=1.7263 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,376B, BPFP=2.6160 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,620B, BPFP=1.6144 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,032B, BPFP=2.6530 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,400B, BPFP=2.2789 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,492B, BPFP=2.5661 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 110,848B, BPFP=0.8932 +⌛️ [2/4] FRONTEND: Frontend time: 0.399s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.518s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 277, 128]) + layer.0.v_cache: torch.Size([1, 4, 277, 128]) + layer.1.k_cache: torch.Size([1, 4, 277, 128]) + layer.1.v_cache: torch.Size([1, 4, 277, 128]) + layer.2.k_cache: torch.Size([1, 4, 277, 128]) + layer.2.v_cache: torch.Size([1, 4, 277, 128]) + layer.3.k_cache: torch.Size([1, 4, 277, 128]) + layer.3.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.k_cache: torch.Size([1, 4, 277, 128]) + layer.4.v_cache: torch.Size([1, 4, 277, 128]) + layer.4.output: torch.Size([1, 277, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11875103 22.74419703 + layer.0.v_cache 0.00001689 0.00640142 + layer.1.k_cache 0.49171178 3.20132193 + layer.1.v_cache 0.00000616 0.00240711 + layer.2.k_cache 0.02532451 0.43371345 + layer.2.v_cache 0.00002086 0.00667234 + layer.3.k_cache 0.02370731 1.93663625 + layer.3.v_cache 0.00002129 0.00741056 + layer.4.k_cache 0.00073268 0.13369801 + layer.4.v_cache 0.00004879 0.01317645 + layer.4.output 0.00480151 198.23488267 + ------------------------------------------------------------------------------------- + TOTAL 0.04082070 83.30175372 + (elements=2,411,008) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2411008 +Total Bytes 485228 +BPFP 1.6100 bits/point +EBPFP 3.2201 equivalent bits/point +MSE 83.301754 +---------------------- -------------------------------------------------------- +Time: 0.928s Load: 0.011s, Pack+Encode: 0.399s, Decode+Unpack: 0.518s +---------------------- -------------------------------------------------------- +💾 Converting with 83.3018 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2774.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2774.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst (7/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,020B, BPFP=0.8304 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,496B, BPFP=2.6464 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,620B, BPFP=1.4426 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,608B, BPFP=2.5898 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,832B, BPFP=1.7112 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,064B, BPFP=2.5551 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,504B, BPFP=1.6265 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,592B, BPFP=2.5888 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,176B, BPFP=2.2434 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,224B, BPFP=2.5015 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,036B, BPFP=0.9296 +⌛️ [2/4] FRONTEND: Frontend time: 0.285s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.426s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10657067 22.36764987 + layer.0.v_cache 0.00001604 0.00645251 + layer.1.k_cache 0.39270948 3.13011300 + layer.1.v_cache 0.00000611 0.00258574 + layer.2.k_cache 0.01541229 0.44430623 + layer.2.v_cache 0.00002285 0.00711556 + layer.3.k_cache 0.01321528 1.86963090 + layer.3.v_cache 0.00001985 0.00772655 + layer.4.k_cache 0.00076706 0.13952662 + layer.4.v_cache 0.00005205 0.01343702 + layer.4.output 1.24960368 218.85798105 + ------------------------------------------------------------------------------------- + TOTAL 0.54564808 91.76437714 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 427172 +BPFP 1.6025 bits/point +EBPFP 3.2051 equivalent bits/point +MSE 91.764377 +---------------------- -------------------------------------------------------- +Time: 0.719s Load: 0.008s, Pack+Encode: 0.285s, Decode+Unpack: 0.426s +---------------------- -------------------------------------------------------- +💾 Converting with 91.7644 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2778.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2778.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst (8/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 197, 128) +Output shape: (1, 197, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) -> torch.Size([1, 1, 197, 512]) + layer.4.output: torch.Size([1, 197, 3584]) -> torch.Size([1, 1, 197, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,492B, BPFP=0.8322 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,296B, BPFP=2.8788 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,032B, BPFP=1.5095 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,440B, BPFP=2.8109 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,016B, BPFP=1.8255 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,052B, BPFP=2.7801 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,512B, BPFP=1.7062 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,444B, BPFP=2.8112 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,632B, BPFP=2.4296 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,556B, BPFP=2.7408 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,236B, BPFP=0.9771 +⌛️ [2/4] FRONTEND: Frontend time: 0.326s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.433s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 197, 128]) + layer.0.v_cache: torch.Size([1, 4, 197, 128]) + layer.1.k_cache: torch.Size([1, 4, 197, 128]) + layer.1.v_cache: torch.Size([1, 4, 197, 128]) + layer.2.k_cache: torch.Size([1, 4, 197, 128]) + layer.2.v_cache: torch.Size([1, 4, 197, 128]) + layer.3.k_cache: torch.Size([1, 4, 197, 128]) + layer.3.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.k_cache: torch.Size([1, 4, 197, 128]) + layer.4.v_cache: torch.Size([1, 4, 197, 128]) + layer.4.output: torch.Size([1, 197, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09219642 23.96533947 + layer.0.v_cache 0.00001621 0.00644995 + layer.1.k_cache 0.17840836 3.33355124 + layer.1.v_cache 0.00000639 0.00248474 + layer.2.k_cache 0.00922235 0.41252117 + layer.2.v_cache 0.00002142 0.00675066 + layer.3.k_cache 0.01957878 2.03002449 + layer.3.v_cache 0.00002124 0.00751939 + layer.4.k_cache 0.00068177 0.13986012 + layer.4.v_cache 0.00005086 0.01396819 + layer.4.output 1.55401793 272.86672861 + ------------------------------------------------------------------------------------- + TOTAL 0.65754878 114.11679822 + (elements=1,714,688) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1714688 +Total Bytes 367708 +BPFP 1.7156 bits/point +EBPFP 3.4311 equivalent bits/point +MSE 114.116798 +---------------------- -------------------------------------------------------- +Time: 0.766s Load: 0.007s, Pack+Encode: 0.326s, Decode+Unpack: 0.433s +---------------------- -------------------------------------------------------- +💾 Converting with 114.1168 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2781.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2781.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst (9/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 310, 128) +Output shape: (1, 310, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) -> torch.Size([1, 1, 310, 512]) + layer.4.output: torch.Size([1, 310, 3584]) -> torch.Size([1, 1, 310, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,200B, BPFP=0.8165 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 51,488B, BPFP=2.5952 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 27,688B, BPFP=1.3956 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,112B, BPFP=2.5258 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,048B, BPFP=1.6657 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 49,588B, BPFP=2.4994 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,036B, BPFP=1.5643 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 50,132B, BPFP=2.5268 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,360B, BPFP=2.1855 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 48,736B, BPFP=2.4565 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 114,648B, BPFP=0.8255 +⌛️ [2/4] FRONTEND: Frontend time: 0.358s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.494s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 310, 128]) + layer.0.v_cache: torch.Size([1, 4, 310, 128]) + layer.1.k_cache: torch.Size([1, 4, 310, 128]) + layer.1.v_cache: torch.Size([1, 4, 310, 128]) + layer.2.k_cache: torch.Size([1, 4, 310, 128]) + layer.2.v_cache: torch.Size([1, 4, 310, 128]) + layer.3.k_cache: torch.Size([1, 4, 310, 128]) + layer.3.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.k_cache: torch.Size([1, 4, 310, 128]) + layer.4.v_cache: torch.Size([1, 4, 310, 128]) + layer.4.output: torch.Size([1, 310, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11169702 22.49399572 + layer.0.v_cache 0.00001595 0.00623536 + layer.1.k_cache 0.54955656 3.13578767 + layer.1.v_cache 0.00000603 0.00244115 + layer.2.k_cache 0.01751511 0.42185516 + layer.2.v_cache 0.00002011 0.00659429 + layer.3.k_cache 0.01650532 1.99256218 + layer.3.v_cache 0.00001967 0.00716077 + layer.4.k_cache 0.00077751 0.13639715 + layer.4.v_cache 0.00005008 0.01314632 + layer.4.output 0.04306356 172.97930588 + ------------------------------------------------------------------------------------- + TOTAL 0.05868284 72.88654805 + (elements=2,698,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2698240 +Total Bytes 516036 +BPFP 1.5300 bits/point +EBPFP 3.0600 equivalent bits/point +MSE 72.886548 +---------------------- -------------------------------------------------------- +Time: 0.862s Load: 0.011s, Pack+Encode: 0.358s, Decode+Unpack: 0.494s +---------------------- -------------------------------------------------------- +💾 Converting with 72.8865 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2785.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2785.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst (10/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,860B, BPFP=0.8501 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,344B, BPFP=2.7483 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,980B, BPFP=1.5037 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,420B, BPFP=2.6821 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,636B, BPFP=1.7658 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,644B, BPFP=2.6264 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,036B, BPFP=1.6511 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,452B, BPFP=2.6843 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,268B, BPFP=2.3128 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,040B, BPFP=2.5831 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,212B, BPFP=0.9442 +⌛️ [2/4] FRONTEND: Frontend time: 0.290s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.395s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09832749 23.31882302 + layer.0.v_cache 0.00001658 0.00684210 + layer.1.k_cache 0.33493462 2.91710586 + layer.1.v_cache 0.00000656 0.00266489 + layer.2.k_cache 0.02762392 0.44286560 + layer.2.v_cache 0.00002012 0.00659533 + layer.3.k_cache 0.01647770 2.09237503 + layer.3.v_cache 0.00001951 0.00758190 + layer.4.k_cache 0.00067889 0.13697069 + layer.4.v_cache 0.00004874 0.01331834 + layer.4.output 1.40435175 243.05451343 + ------------------------------------------------------------------------------------- + TOTAL 0.60638920 101.78392569 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 390892 +BPFP 1.6481 bits/point +EBPFP 3.2961 equivalent bits/point +MSE 101.783926 +---------------------- -------------------------------------------------------- +Time: 0.692s Load: 0.007s, Pack+Encode: 0.290s, Decode+Unpack: 0.395s +---------------------- -------------------------------------------------------- +💾 Converting with 101.7839 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2787.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2787.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst (11/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,700B, BPFP=0.8089 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,244B, BPFP=2.7132 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,088B, BPFP=1.4580 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,448B, BPFP=2.6582 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,156B, BPFP=1.7392 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,712B, BPFP=2.6073 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,604B, BPFP=1.6319 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,328B, BPFP=2.6499 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,160B, BPFP=2.2926 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,084B, BPFP=2.5639 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,552B, BPFP=0.9240 +⌛️ [2/4] FRONTEND: Frontend time: 0.320s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102250 23.77948268 + layer.0.v_cache 0.00001721 0.00625714 + layer.1.k_cache 0.37371607 3.04725012 + layer.1.v_cache 0.00000658 0.00241124 + layer.2.k_cache 0.01127678 0.40755824 + layer.2.v_cache 0.00001998 0.00654903 + layer.3.k_cache 0.02136800 1.91384483 + layer.3.v_cache 0.00001963 0.00714491 + layer.4.k_cache 0.00069450 0.13323676 + layer.4.v_cache 0.00005147 0.01334622 + layer.4.output 1.35464589 235.61044169 + ------------------------------------------------------------------------------------- + TOTAL 0.58945376 98.74059841 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 399076 +BPFP 1.6230 bits/point +EBPFP 3.2460 equivalent bits/point +MSE 98.740598 +---------------------- -------------------------------------------------------- +Time: 0.768s Load: 0.008s, Pack+Encode: 0.320s, Decode+Unpack: 0.439s +---------------------- -------------------------------------------------------- +💾 Converting with 98.7406 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2789.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2789.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst (12/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,744B, BPFP=0.7933 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,572B, BPFP=2.5879 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,448B, BPFP=1.3974 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,636B, BPFP=2.5296 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,568B, BPFP=1.6539 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,936B, BPFP=2.4861 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,308B, BPFP=1.5754 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,596B, BPFP=2.5271 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,904B, BPFP=2.1728 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,056B, BPFP=2.4313 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,936B, BPFP=0.8887 +⌛️ [2/4] FRONTEND: Frontend time: 0.328s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.431s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11871604 21.89254311 + layer.0.v_cache 0.00001755 0.00600482 + layer.1.k_cache 0.46315337 2.95203074 + layer.1.v_cache 0.00000588 0.00227691 + layer.2.k_cache 0.02202094 0.43560800 + layer.2.v_cache 0.00001989 0.00608976 + layer.3.k_cache 0.01278766 1.81906395 + layer.3.v_cache 0.00002092 0.00699036 + layer.4.k_cache 0.00074254 0.12828626 + layer.4.v_cache 0.00004925 0.01209835 + layer.4.output 1.21975157 207.03297524 + ------------------------------------------------------------------------------------- + TOTAL 0.53857618 86.85245994 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 423704 +BPFP 1.5515 bits/point +EBPFP 3.1031 equivalent bits/point +MSE 86.852460 +---------------------- -------------------------------------------------------- +Time: 0.768s Load: 0.009s, Pack+Encode: 0.328s, Decode+Unpack: 0.431s +---------------------- -------------------------------------------------------- +💾 Converting with 86.8525 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2816.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2816.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst (13/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 195, 128) +Output shape: (1, 195, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) -> torch.Size([1, 1, 195, 512]) + layer.4.output: torch.Size([1, 195, 3584]) -> torch.Size([1, 1, 195, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,220B, BPFP=0.8189 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,280B, BPFP=2.9071 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,820B, BPFP=1.5080 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,432B, BPFP=2.8391 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,392B, BPFP=1.7942 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,724B, BPFP=2.7824 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,052B, BPFP=1.6869 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,280B, BPFP=2.8269 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,156B, BPFP=2.4163 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,228B, BPFP=2.7426 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 79,764B, BPFP=0.9130 +⌛️ [2/4] FRONTEND: Frontend time: 0.323s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.419s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 195, 128]) + layer.0.v_cache: torch.Size([1, 4, 195, 128]) + layer.1.k_cache: torch.Size([1, 4, 195, 128]) + layer.1.v_cache: torch.Size([1, 4, 195, 128]) + layer.2.k_cache: torch.Size([1, 4, 195, 128]) + layer.2.v_cache: torch.Size([1, 4, 195, 128]) + layer.3.k_cache: torch.Size([1, 4, 195, 128]) + layer.3.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.k_cache: torch.Size([1, 4, 195, 128]) + layer.4.v_cache: torch.Size([1, 4, 195, 128]) + layer.4.output: torch.Size([1, 195, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605874 22.97505008 + layer.0.v_cache 0.00001676 0.00638144 + layer.1.k_cache 0.13872500 3.09616105 + layer.1.v_cache 0.00000593 0.00240478 + layer.2.k_cache 0.01222654 0.40697691 + layer.2.v_cache 0.00001968 0.00634704 + layer.3.k_cache 0.01415822 1.87886525 + layer.3.v_cache 0.00002127 0.00760616 + layer.4.k_cache 0.00071127 0.13396258 + layer.4.v_cache 0.00005419 0.01389831 + layer.4.output 1.56987251 267.23468407 + ------------------------------------------------------------------------------------- + TOTAL 0.66300619 111.71590836 + (elements=1,697,280) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1697280 +Total Bytes 358348 +BPFP 1.6890 bits/point +EBPFP 3.3781 equivalent bits/point +MSE 111.715908 +---------------------- -------------------------------------------------------- +Time: 0.750s Load: 0.007s, Pack+Encode: 0.323s, Decode+Unpack: 0.419s +---------------------- -------------------------------------------------------- +💾 Converting with 111.7159 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2967.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2967.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst (14/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,968B, BPFP=0.8558 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,356B, BPFP=2.6920 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,300B, BPFP=1.4852 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,644B, BPFP=2.6308 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,536B, BPFP=1.7630 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,232B, BPFP=2.5955 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,200B, BPFP=1.6484 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,556B, BPFP=2.6233 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,464B, BPFP=2.2720 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,524B, BPFP=2.5347 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,264B, BPFP=0.9231 +⌛️ [2/4] FRONTEND: Frontend time: 0.368s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.362s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11176265 23.81561481 + layer.0.v_cache 0.00001802 0.00685059 + layer.1.k_cache 0.18491925 3.31644314 + layer.1.v_cache 0.00000608 0.00265789 + layer.2.k_cache 0.01999094 0.43867426 + layer.2.v_cache 0.00002176 0.00685748 + layer.3.k_cache 0.03910406 1.90238500 + layer.3.v_cache 0.00002004 0.00764473 + layer.4.k_cache 0.00067358 0.13703145 + layer.4.v_cache 0.00005212 0.01362828 + layer.4.output 0.00886795 298.63302100 + ------------------------------------------------------------------------------------- + TOTAL 0.02462613 124.71052556 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 321044 +BPFP 1.6213 bits/point +EBPFP 3.2426 equivalent bits/point +MSE 124.710526 +---------------------- -------------------------------------------------------- +Time: 0.737s Load: 0.007s, Pack+Encode: 0.368s, Decode+Unpack: 0.362s +---------------------- -------------------------------------------------------- +💾 Converting with 124.7105 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2971.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst (15/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,844B, BPFP=0.8688 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,156B, BPFP=2.7990 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,356B, BPFP=1.4933 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,160B, BPFP=2.7259 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,336B, BPFP=1.7852 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,632B, BPFP=2.6872 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,704B, BPFP=1.6655 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,284B, BPFP=2.7350 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,912B, BPFP=2.3410 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,760B, BPFP=2.6232 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,848B, BPFP=0.9101 +⌛️ [2/4] FRONTEND: Frontend time: 0.279s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.413s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15089922 23.20276051 + layer.0.v_cache 0.00001686 0.00659094 + layer.1.k_cache 0.31217133 3.11189607 + layer.1.v_cache 0.00000593 0.00259173 + layer.2.k_cache 0.01305961 0.44217542 + layer.2.v_cache 0.00001999 0.00688587 + layer.3.k_cache 0.01178671 1.76574478 + layer.3.v_cache 0.00002039 0.00787648 + layer.4.k_cache 0.00067606 0.13939423 + layer.4.v_cache 0.00005285 0.01352947 + layer.4.output 1.43724915 251.68854795 + ------------------------------------------------------------------------------------- + TOTAL 0.62055606 105.32466360 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 382992 +BPFP 1.6527 bits/point +EBPFP 3.3053 equivalent bits/point +MSE 105.324664 +---------------------- -------------------------------------------------------- +Time: 0.700s Load: 0.007s, Pack+Encode: 0.279s, Decode+Unpack: 0.413s +---------------------- -------------------------------------------------------- +💾 Converting with 105.3247 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-2992.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-2992.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst (16/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 187, 128) +Output shape: (1, 187, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) -> torch.Size([1, 1, 187, 512]) + layer.4.output: torch.Size([1, 187, 3584]) -> torch.Size([1, 1, 187, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,580B, BPFP=0.8005 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,608B, BPFP=2.6410 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,156B, BPFP=1.4335 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,696B, BPFP=2.5648 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,820B, BPFP=1.7396 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,544B, BPFP=2.5521 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,788B, BPFP=1.6534 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,088B, BPFP=2.5976 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,824B, BPFP=2.2413 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 30,012B, BPFP=2.5077 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,900B, BPFP=1.0254 +⌛️ [2/4] FRONTEND: Frontend time: 0.292s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.379s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 187, 128]) + layer.0.v_cache: torch.Size([1, 4, 187, 128]) + layer.1.k_cache: torch.Size([1, 4, 187, 128]) + layer.1.v_cache: torch.Size([1, 4, 187, 128]) + layer.2.k_cache: torch.Size([1, 4, 187, 128]) + layer.2.v_cache: torch.Size([1, 4, 187, 128]) + layer.3.k_cache: torch.Size([1, 4, 187, 128]) + layer.3.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.k_cache: torch.Size([1, 4, 187, 128]) + layer.4.v_cache: torch.Size([1, 4, 187, 128]) + layer.4.output: torch.Size([1, 187, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08258281 23.00985963 + layer.0.v_cache 0.00001733 0.00708465 + layer.1.k_cache 0.10049684 2.98616460 + layer.1.v_cache 0.00000647 0.00264469 + layer.2.k_cache 0.01472783 0.44699798 + layer.2.v_cache 0.00002140 0.00733656 + layer.3.k_cache 0.02125903 2.05256041 + layer.3.v_cache 0.00002180 0.00852144 + layer.4.k_cache 0.00070831 0.14801527 + layer.4.v_cache 0.00005280 0.01470721 + layer.4.output 0.00866398 281.03893717 + ------------------------------------------------------------------------------------- + TOTAL 0.01650250 117.40920309 + (elements=1,627,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1627648 +Total Bytes 334016 +BPFP 1.6417 bits/point +EBPFP 3.2834 equivalent bits/point +MSE 117.409203 +---------------------- -------------------------------------------------------- +Time: 0.677s Load: 0.006s, Pack+Encode: 0.292s, Decode+Unpack: 0.379s +---------------------- -------------------------------------------------------- +💾 Converting with 117.4092 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3008.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3008.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst (17/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,732B, BPFP=0.8688 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,992B, BPFP=2.8134 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,388B, BPFP=1.5098 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,172B, BPFP=2.7527 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,204B, BPFP=1.7924 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,552B, BPFP=2.7068 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,752B, BPFP=1.6848 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,140B, BPFP=2.7503 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,664B, BPFP=2.3448 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,704B, BPFP=2.6440 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,192B, BPFP=0.9541 +⌛️ [2/4] FRONTEND: Frontend time: 0.289s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.409s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11071491 22.75080532 + layer.0.v_cache 0.00001804 0.00657424 + layer.1.k_cache 0.26784508 3.24555861 + layer.1.v_cache 0.00000610 0.00252627 + layer.2.k_cache 0.01369454 0.45086771 + layer.2.v_cache 0.00002102 0.00672969 + layer.3.k_cache 0.01444715 1.83964734 + layer.3.v_cache 0.00002041 0.00756967 + layer.4.k_cache 0.00067564 0.13215898 + layer.4.v_cache 0.00005248 0.01342477 + layer.4.output 1.45091435 244.49365267 + ------------------------------------------------------------------------------------- + TOTAL 0.62140564 102.34773125 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 385492 +BPFP 1.6792 bits/point +EBPFP 3.3584 equivalent bits/point +MSE 102.347731 +---------------------- -------------------------------------------------------- +Time: 0.705s Load: 0.007s, Pack+Encode: 0.289s, Decode+Unpack: 0.409s +---------------------- -------------------------------------------------------- +💾 Converting with 102.3477 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3013.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3013.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst (18/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 223, 128) +Output shape: (1, 223, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) -> torch.Size([1, 1, 223, 512]) + layer.4.output: torch.Size([1, 223, 3584]) -> torch.Size([1, 1, 223, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,088B, BPFP=0.8470 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,212B, BPFP=2.7475 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,944B, BPFP=1.4675 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,248B, BPFP=2.6799 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,004B, BPFP=1.7520 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,556B, BPFP=2.6314 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,656B, BPFP=1.6575 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,140B, BPFP=2.6724 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,864B, BPFP=2.3027 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,780B, BPFP=2.5771 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 90,744B, BPFP=0.9083 +⌛️ [2/4] FRONTEND: Frontend time: 0.322s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.465s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 223, 128]) + layer.0.v_cache: torch.Size([1, 4, 223, 128]) + layer.1.k_cache: torch.Size([1, 4, 223, 128]) + layer.1.v_cache: torch.Size([1, 4, 223, 128]) + layer.2.k_cache: torch.Size([1, 4, 223, 128]) + layer.2.v_cache: torch.Size([1, 4, 223, 128]) + layer.3.k_cache: torch.Size([1, 4, 223, 128]) + layer.3.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.k_cache: torch.Size([1, 4, 223, 128]) + layer.4.v_cache: torch.Size([1, 4, 223, 128]) + layer.4.output: torch.Size([1, 223, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13389712 23.67006639 + layer.0.v_cache 0.00001748 0.00643005 + layer.1.k_cache 0.36644789 3.12786126 + layer.1.v_cache 0.00000587 0.00247491 + layer.2.k_cache 0.02260907 0.44464686 + layer.2.v_cache 0.00001979 0.00651891 + layer.3.k_cache 0.02550721 1.74927688 + layer.3.v_cache 0.00001968 0.00738947 + layer.4.k_cache 0.00067529 0.12987484 + layer.4.v_cache 0.00006705 0.01324869 + layer.4.output 1.37284074 238.55118914 + ------------------------------------------------------------------------------------- + TOTAL 0.59759716 99.94212425 + (elements=1,940,992) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1940992 +Total Bytes 395236 +BPFP 1.6290 bits/point +EBPFP 3.2580 equivalent bits/point +MSE 99.942124 +---------------------- -------------------------------------------------------- +Time: 0.797s Load: 0.009s, Pack+Encode: 0.322s, Decode+Unpack: 0.465s +---------------------- -------------------------------------------------------- +💾 Converting with 99.9421 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3015.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3015.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst (19/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.013s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 406, 128) +Output shape: (1, 406, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) -> torch.Size([1, 1, 406, 512]) + layer.4.output: torch.Size([1, 406, 3584]) -> torch.Size([1, 1, 406, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 20,920B, BPFP=0.8051 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 69,460B, BPFP=2.6732 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 36,648B, BPFP=1.4104 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 67,652B, BPFP=2.6036 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 43,956B, BPFP=1.6917 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 66,724B, BPFP=2.5679 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 41,128B, BPFP=1.5828 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 67,348B, BPFP=2.5919 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 57,584B, BPFP=2.2161 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 64,960B, BPFP=2.5000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 164,592B, BPFP=0.9049 +⌛️ [2/4] FRONTEND: Frontend time: 0.482s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.583s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 406, 128]) + layer.0.v_cache: torch.Size([1, 4, 406, 128]) + layer.1.k_cache: torch.Size([1, 4, 406, 128]) + layer.1.v_cache: torch.Size([1, 4, 406, 128]) + layer.2.k_cache: torch.Size([1, 4, 406, 128]) + layer.2.v_cache: torch.Size([1, 4, 406, 128]) + layer.3.k_cache: torch.Size([1, 4, 406, 128]) + layer.3.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.k_cache: torch.Size([1, 4, 406, 128]) + layer.4.v_cache: torch.Size([1, 4, 406, 128]) + layer.4.output: torch.Size([1, 406, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10474718 21.71608490 + layer.0.v_cache 0.00001775 0.00647292 + layer.1.k_cache 0.89873824 3.19054093 + layer.1.v_cache 0.00000687 0.00251626 + layer.2.k_cache 0.02602503 0.42944971 + layer.2.v_cache 0.00002164 0.00664765 + layer.3.k_cache 0.01558245 1.86111555 + layer.3.v_cache 0.00002201 0.00733684 + layer.4.k_cache 0.00072958 0.13479727 + layer.4.v_cache 0.00005327 0.01286843 + layer.4.output 0.00651461 135.19139251 + ------------------------------------------------------------------------------------- + TOTAL 0.06420861 57.27691635 + (elements=3,533,824) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3533824 +Total Bytes 700972 +BPFP 1.5869 bits/point +EBPFP 3.1738 equivalent bits/point +MSE 57.276916 +---------------------- -------------------------------------------------------- +Time: 1.078s Load: 0.013s, Pack+Encode: 0.482s, Decode+Unpack: 0.583s +---------------------- -------------------------------------------------------- +💾 Converting with 57.2769 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3123.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3123.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst (20/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.014s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 245, 128) +Output shape: (1, 245, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) -> torch.Size([1, 1, 245, 512]) + layer.4.output: torch.Size([1, 245, 3584]) -> torch.Size([1, 1, 245, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,316B, BPFP=0.8492 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,588B, BPFP=2.6523 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,780B, BPFP=1.4528 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,788B, BPFP=2.6013 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,872B, BPFP=1.7138 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,068B, BPFP=2.5554 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,240B, BPFP=1.6097 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,608B, BPFP=2.5898 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,080B, BPFP=2.2372 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,224B, BPFP=2.5015 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,408B, BPFP=0.9057 +⌛️ [2/4] FRONTEND: Frontend time: 0.353s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.452s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 245, 128]) + layer.0.v_cache: torch.Size([1, 4, 245, 128]) + layer.1.k_cache: torch.Size([1, 4, 245, 128]) + layer.1.v_cache: torch.Size([1, 4, 245, 128]) + layer.2.k_cache: torch.Size([1, 4, 245, 128]) + layer.2.v_cache: torch.Size([1, 4, 245, 128]) + layer.3.k_cache: torch.Size([1, 4, 245, 128]) + layer.3.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.k_cache: torch.Size([1, 4, 245, 128]) + layer.4.v_cache: torch.Size([1, 4, 245, 128]) + layer.4.output: torch.Size([1, 245, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14414167 22.80229592 + layer.0.v_cache 0.00001717 0.00658900 + layer.1.k_cache 0.44179790 3.33883854 + layer.1.v_cache 0.00000667 0.00265457 + layer.2.k_cache 0.03016641 0.44345102 + layer.2.v_cache 0.00002043 0.00674964 + layer.3.k_cache 0.03726396 1.82472858 + layer.3.v_cache 0.00002099 0.00774426 + layer.4.k_cache 0.00070653 0.13546607 + layer.4.v_cache 0.00004910 0.01313696 + layer.4.output 1.24963187 218.92906341 + ------------------------------------------------------------------------------------- + TOTAL 0.55303611 91.82853520 + (elements=2,132,480) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2132480 +Total Bytes 424972 +BPFP 1.5943 bits/point +EBPFP 3.1886 equivalent bits/point +MSE 91.828535 +---------------------- -------------------------------------------------------- +Time: 0.819s Load: 0.014s, Pack+Encode: 0.353s, Decode+Unpack: 0.452s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8285 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3132.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3132.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst (21/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,912B, BPFP=0.8568 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,888B, BPFP=2.8964 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,300B, BPFP=1.5154 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,888B, BPFP=2.8178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,180B, BPFP=1.8200 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,260B, BPFP=2.7685 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,832B, BPFP=1.7142 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,964B, BPFP=2.8238 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,936B, BPFP=2.4290 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,676B, BPFP=2.7227 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 81,252B, BPFP=0.9114 +⌛️ [2/4] FRONTEND: Frontend time: 0.310s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.442s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15985105 24.15954529 + layer.0.v_cache 0.00001905 0.00704034 + layer.1.k_cache 0.22745673 3.08094734 + layer.1.v_cache 0.00000590 0.00263481 + layer.2.k_cache 0.01549998 0.43997461 + layer.2.v_cache 0.00002003 0.00683829 + layer.3.k_cache 0.01333464 1.89065015 + layer.3.v_cache 0.00002034 0.00802343 + layer.4.k_cache 0.00067076 0.13961401 + layer.4.v_cache 0.00004949 0.01390021 + layer.4.output 1.53832061 267.07293162 + ------------------------------------------------------------------------------------- + TOTAL 0.65795131 111.72115823 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 366088 +BPFP 1.6908 bits/point +EBPFP 3.3817 equivalent bits/point +MSE 111.721158 +---------------------- -------------------------------------------------------- +Time: 0.759s Load: 0.007s, Pack+Encode: 0.310s, Decode+Unpack: 0.442s +---------------------- -------------------------------------------------------- +💾 Converting with 111.7212 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3138.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3138.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst (22/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 237, 128) +Output shape: (1, 237, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) -> torch.Size([1, 1, 237, 512]) + layer.4.output: torch.Size([1, 237, 3584]) -> torch.Size([1, 1, 237, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,488B, BPFP=0.8233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,088B, BPFP=2.7089 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,456B, BPFP=1.4805 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,148B, BPFP=2.6469 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,328B, BPFP=1.7358 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,544B, BPFP=2.6071 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,764B, BPFP=1.6326 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,916B, BPFP=2.6316 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,520B, BPFP=2.2758 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,632B, BPFP=2.5469 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 96,452B, BPFP=0.9084 +⌛️ [2/4] FRONTEND: Frontend time: 0.298s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.416s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 237, 128]) + layer.0.v_cache: torch.Size([1, 4, 237, 128]) + layer.1.k_cache: torch.Size([1, 4, 237, 128]) + layer.1.v_cache: torch.Size([1, 4, 237, 128]) + layer.2.k_cache: torch.Size([1, 4, 237, 128]) + layer.2.v_cache: torch.Size([1, 4, 237, 128]) + layer.3.k_cache: torch.Size([1, 4, 237, 128]) + layer.3.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.k_cache: torch.Size([1, 4, 237, 128]) + layer.4.v_cache: torch.Size([1, 4, 237, 128]) + layer.4.output: torch.Size([1, 237, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12351868 23.06589531 + layer.0.v_cache 0.00001655 0.00643100 + layer.1.k_cache 0.34375901 3.17354329 + layer.1.v_cache 0.00000619 0.00257242 + layer.2.k_cache 0.01282014 0.45496227 + layer.2.v_cache 0.00002104 0.00660449 + layer.3.k_cache 0.03241006 1.94007185 + layer.3.v_cache 0.00001985 0.00745930 + layer.4.k_cache 0.00067667 0.13623984 + layer.4.v_cache 0.00005279 0.01348980 + layer.4.output 1.29178675 226.70773056 + ------------------------------------------------------------------------------------- + TOTAL 0.56210637 95.04478726 + (elements=2,062,848) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2062848 +Total Bytes 416336 +BPFP 1.6146 bits/point +EBPFP 3.2292 equivalent bits/point +MSE 95.044787 +---------------------- -------------------------------------------------------- +Time: 0.722s Load: 0.007s, Pack+Encode: 0.298s, Decode+Unpack: 0.416s +---------------------- -------------------------------------------------------- +💾 Converting with 95.0448 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3143.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3143.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst (23/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,584B, BPFP=0.8419 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,000B, BPFP=2.7616 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,456B, BPFP=1.4866 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,152B, BPFP=2.7000 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,180B, BPFP=1.7573 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,552B, BPFP=2.6564 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,724B, BPFP=1.6515 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,920B, BPFP=2.6831 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,892B, BPFP=2.3177 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,560B, BPFP=2.5843 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 85,712B, BPFP=0.8899 +⌛️ [2/4] FRONTEND: Frontend time: 0.343s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.372s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11023432 23.31845476 + layer.0.v_cache 0.00001677 0.00634927 + layer.1.k_cache 0.30291741 3.05555590 + layer.1.v_cache 0.00000597 0.00247936 + layer.2.k_cache 0.03464115 0.46214919 + layer.2.v_cache 0.00001998 0.00654428 + layer.3.k_cache 0.01330193 1.85295566 + layer.3.v_cache 0.00002112 0.00725357 + layer.4.k_cache 0.00069379 0.13593399 + layer.4.v_cache 0.00005021 0.01286495 + layer.4.output 1.42388847 247.11634136 + ------------------------------------------------------------------------------------- + TOTAL 0.61347776 103.45146650 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 380732 +BPFP 1.6276 bits/point +EBPFP 3.2552 equivalent bits/point +MSE 103.451466 +---------------------- -------------------------------------------------------- +Time: 0.723s Load: 0.007s, Pack+Encode: 0.343s, Decode+Unpack: 0.372s +---------------------- -------------------------------------------------------- +💾 Converting with 103.4515 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3147.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3147.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst (24/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 231, 128) +Output shape: (1, 231, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) -> torch.Size([1, 1, 231, 512]) + layer.4.output: torch.Size([1, 231, 3584]) -> torch.Size([1, 1, 231, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,420B, BPFP=0.8401 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,616B, BPFP=2.7473 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,192B, BPFP=1.5011 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,736B, BPFP=2.6878 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,100B, BPFP=1.7654 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,076B, BPFP=2.6431 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,628B, BPFP=1.6659 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,588B, BPFP=2.6778 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,092B, BPFP=2.3060 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,320B, BPFP=2.5920 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,708B, BPFP=0.9731 +⌛️ [2/4] FRONTEND: Frontend time: 0.308s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.455s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 231, 128]) + layer.0.v_cache: torch.Size([1, 4, 231, 128]) + layer.1.k_cache: torch.Size([1, 4, 231, 128]) + layer.1.v_cache: torch.Size([1, 4, 231, 128]) + layer.2.k_cache: torch.Size([1, 4, 231, 128]) + layer.2.v_cache: torch.Size([1, 4, 231, 128]) + layer.3.k_cache: torch.Size([1, 4, 231, 128]) + layer.3.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.k_cache: torch.Size([1, 4, 231, 128]) + layer.4.v_cache: torch.Size([1, 4, 231, 128]) + layer.4.output: torch.Size([1, 231, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13257201 23.35044643 + layer.0.v_cache 0.00001701 0.00654399 + layer.1.k_cache 0.40462213 3.10232927 + layer.1.v_cache 0.00000658 0.00253363 + layer.2.k_cache 0.02649644 0.42523629 + layer.2.v_cache 0.00002056 0.00656075 + layer.3.k_cache 0.06147716 2.02162943 + layer.3.v_cache 0.00002105 0.00780462 + layer.4.k_cache 0.00069437 0.13665192 + layer.4.v_cache 0.00005136 0.01344675 + layer.4.output 1.32536713 229.56694496 + ------------------------------------------------------------------------------------- + TOTAL 0.58256168 96.23775281 + (elements=2,010,624) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2010624 +Total Bytes 417476 +BPFP 1.6611 bits/point +EBPFP 3.3222 equivalent bits/point +MSE 96.237753 +---------------------- -------------------------------------------------------- +Time: 0.770s Load: 0.008s, Pack+Encode: 0.308s, Decode+Unpack: 0.455s +---------------------- -------------------------------------------------------- +💾 Converting with 96.2378 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3160.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3160.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst (25/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 255, 128) +Output shape: (1, 255, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) -> torch.Size([1, 1, 255, 512]) + layer.4.output: torch.Size([1, 255, 3584]) -> torch.Size([1, 1, 255, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,064B, BPFP=0.8005 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,984B, BPFP=2.5725 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,436B, BPFP=1.3748 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,016B, BPFP=2.5132 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,364B, BPFP=1.6767 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,664B, BPFP=2.4917 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,832B, BPFP=1.5828 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,064B, BPFP=2.5162 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,556B, BPFP=2.1787 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,744B, BPFP=2.4353 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 108,800B, BPFP=0.9524 +⌛️ [2/4] FRONTEND: Frontend time: 0.313s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.449s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 255, 128]) + layer.0.v_cache: torch.Size([1, 4, 255, 128]) + layer.1.k_cache: torch.Size([1, 4, 255, 128]) + layer.1.v_cache: torch.Size([1, 4, 255, 128]) + layer.2.k_cache: torch.Size([1, 4, 255, 128]) + layer.2.v_cache: torch.Size([1, 4, 255, 128]) + layer.3.k_cache: torch.Size([1, 4, 255, 128]) + layer.3.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.k_cache: torch.Size([1, 4, 255, 128]) + layer.4.v_cache: torch.Size([1, 4, 255, 128]) + layer.4.output: torch.Size([1, 255, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12841698 21.36371975 + layer.0.v_cache 0.00001718 0.00696766 + layer.1.k_cache 0.45450200 3.02730426 + layer.1.v_cache 0.00000627 0.00263684 + layer.2.k_cache 0.01434461 0.41140992 + layer.2.v_cache 0.00002215 0.00739499 + layer.3.k_cache 0.03530993 1.95319106 + layer.3.v_cache 0.00002139 0.00811386 + layer.4.k_cache 0.00068344 0.14208712 + layer.4.v_cache 0.00005073 0.01426733 + layer.4.output 1.20064817 181.40908613 + ------------------------------------------------------------------------------------- + TOTAL 0.53164187 76.28239387 + (elements=2,219,520) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2219520 +Total Bytes 437524 +BPFP 1.5770 bits/point +EBPFP 3.1540 equivalent bits/point +MSE 76.282394 +---------------------- -------------------------------------------------------- +Time: 0.770s Load: 0.008s, Pack+Encode: 0.313s, Decode+Unpack: 0.449s +---------------------- -------------------------------------------------------- +💾 Converting with 76.2824 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3182.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3182.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst (26/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,808B, BPFP=0.8205 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,708B, BPFP=2.6988 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,168B, BPFP=1.4499 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,660B, BPFP=2.6407 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,172B, BPFP=1.7272 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,916B, BPFP=2.5995 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,148B, BPFP=1.6150 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,356B, BPFP=2.6239 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,332B, BPFP=2.2901 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,912B, BPFP=2.5439 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 109,684B, BPFP=0.8682 +⌛️ [2/4] FRONTEND: Frontend time: 0.379s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.538s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13731835 22.76338098 + layer.0.v_cache 0.00001709 0.00626033 + layer.1.k_cache 0.55879569 3.04205452 + layer.1.v_cache 0.00000637 0.00247452 + layer.2.k_cache 0.01348169 0.42118164 + layer.2.v_cache 0.00002375 0.00677040 + layer.3.k_cache 0.01306405 1.85994812 + layer.3.v_cache 0.00002005 0.00723297 + layer.4.k_cache 0.00072533 0.13405754 + layer.4.v_cache 0.00005384 0.01365190 + layer.4.output 0.00472113 192.75253293 + ------------------------------------------------------------------------------------- + TOTAL 0.04450318 81.03086726 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 488864 +BPFP 1.5933 bits/point +EBPFP 3.1867 equivalent bits/point +MSE 81.030867 +---------------------- -------------------------------------------------------- +Time: 0.926s Load: 0.009s, Pack+Encode: 0.379s, Decode+Unpack: 0.538s +---------------------- -------------------------------------------------------- +💾 Converting with 81.0309 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3189.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3189.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst (27/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 301, 128) +Output shape: (1, 301, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) -> torch.Size([1, 1, 301, 512]) + layer.4.output: torch.Size([1, 301, 3584]) -> torch.Size([1, 1, 301, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,856B, BPFP=0.8231 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 51,624B, BPFP=2.6798 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,160B, BPFP=1.4618 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,516B, BPFP=2.6223 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,312B, BPFP=1.7292 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 49,700B, BPFP=2.5799 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,320B, BPFP=1.6258 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 50,364B, BPFP=2.6144 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,420B, BPFP=2.2539 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 48,504B, BPFP=2.5179 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 119,956B, BPFP=0.8896 +⌛️ [2/4] FRONTEND: Frontend time: 0.396s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.557s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 301, 128]) + layer.0.v_cache: torch.Size([1, 4, 301, 128]) + layer.1.k_cache: torch.Size([1, 4, 301, 128]) + layer.1.v_cache: torch.Size([1, 4, 301, 128]) + layer.2.k_cache: torch.Size([1, 4, 301, 128]) + layer.2.v_cache: torch.Size([1, 4, 301, 128]) + layer.3.k_cache: torch.Size([1, 4, 301, 128]) + layer.3.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.k_cache: torch.Size([1, 4, 301, 128]) + layer.4.v_cache: torch.Size([1, 4, 301, 128]) + layer.4.output: torch.Size([1, 301, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13972237 22.83057290 + layer.0.v_cache 0.00001689 0.00669563 + layer.1.k_cache 0.51719427 3.14452780 + layer.1.v_cache 0.00000613 0.00271386 + layer.2.k_cache 0.02019965 0.46741929 + layer.2.v_cache 0.00002053 0.00705557 + layer.3.k_cache 0.01759115 1.87065576 + layer.3.v_cache 0.00002125 0.00798107 + layer.4.k_cache 0.00069604 0.14302542 + layer.4.v_cache 0.00005872 0.01397208 + layer.4.output 0.04434862 178.73420444 + ------------------------------------------------------------------------------------- + TOTAL 0.05917455 75.27259120 + (elements=2,619,904) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2619904 +Total Bytes 522732 +BPFP 1.5962 bits/point +EBPFP 3.1924 equivalent bits/point +MSE 75.272591 +---------------------- -------------------------------------------------------- +Time: 0.963s Load: 0.010s, Pack+Encode: 0.396s, Decode+Unpack: 0.557s +---------------------- -------------------------------------------------------- +💾 Converting with 75.2726 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3191.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3191.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst (28/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,828B, BPFP=0.8502 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,824B, BPFP=2.8913 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,328B, BPFP=1.5176 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,964B, BPFP=2.8238 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,076B, BPFP=1.8119 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,176B, BPFP=2.7619 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,592B, BPFP=1.6954 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,776B, BPFP=2.8090 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,832B, BPFP=2.4209 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,580B, BPFP=2.7151 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 80,920B, BPFP=0.9077 +⌛️ [2/4] FRONTEND: Frontend time: 0.343s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13044639 23.49277147 + layer.0.v_cache 0.00001737 0.00672958 + layer.1.k_cache 0.22565776 3.23089937 + layer.1.v_cache 0.00000620 0.00263454 + layer.2.k_cache 0.01998205 0.42666143 + layer.2.v_cache 0.00001989 0.00678590 + layer.3.k_cache 0.00629339 2.09865382 + layer.3.v_cache 0.00002050 0.00758243 + layer.4.k_cache 0.00068118 0.13725369 + layer.4.v_cache 0.00005692 0.01332350 + layer.4.output 1.53833376 267.40799982 + ------------------------------------------------------------------------------------- + TOTAL 0.65597165 111.83995850 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 364896 +BPFP 1.6853 bits/point +EBPFP 3.3707 equivalent bits/point +MSE 111.839958 +---------------------- -------------------------------------------------------- +Time: 0.822s Load: 0.009s, Pack+Encode: 0.343s, Decode+Unpack: 0.470s +---------------------- -------------------------------------------------------- +💾 Converting with 111.8400 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3208.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3208.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst (29/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 246, 128) +Output shape: (1, 246, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) -> torch.Size([1, 1, 246, 512]) + layer.4.output: torch.Size([1, 246, 3584]) -> torch.Size([1, 1, 246, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,212B, BPFP=0.8392 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,640B, BPFP=2.6448 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,740B, BPFP=1.4444 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,648B, BPFP=2.5818 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,812B, BPFP=1.7030 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,024B, BPFP=2.5422 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,404B, BPFP=1.6136 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,584B, BPFP=2.5777 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,240B, BPFP=2.2383 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,080B, BPFP=2.4822 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,452B, BPFP=0.9024 +⌛️ [2/4] FRONTEND: Frontend time: 0.308s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.428s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 246, 128]) + layer.0.v_cache: torch.Size([1, 4, 246, 128]) + layer.1.k_cache: torch.Size([1, 4, 246, 128]) + layer.1.v_cache: torch.Size([1, 4, 246, 128]) + layer.2.k_cache: torch.Size([1, 4, 246, 128]) + layer.2.v_cache: torch.Size([1, 4, 246, 128]) + layer.3.k_cache: torch.Size([1, 4, 246, 128]) + layer.3.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.k_cache: torch.Size([1, 4, 246, 128]) + layer.4.v_cache: torch.Size([1, 4, 246, 128]) + layer.4.output: torch.Size([1, 246, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15131485 22.36317208 + layer.0.v_cache 0.00001613 0.00651729 + layer.1.k_cache 0.49701033 3.32810862 + layer.1.v_cache 0.00000620 0.00263315 + layer.2.k_cache 0.01298508 0.42930643 + layer.2.v_cache 0.00002120 0.00684053 + layer.3.k_cache 0.02762915 1.80148861 + layer.3.v_cache 0.00002136 0.00783367 + layer.4.k_cache 0.00068392 0.13982091 + layer.4.v_cache 0.00006372 0.01313362 + layer.4.output 1.24453647 217.98263284 + ------------------------------------------------------------------------------------- + TOTAL 0.55302984 91.41042852 + (elements=2,141,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2141184 +Total Bytes 424836 +BPFP 1.5873 bits/point +EBPFP 3.1746 equivalent bits/point +MSE 91.410429 +---------------------- -------------------------------------------------------- +Time: 0.745s Load: 0.009s, Pack+Encode: 0.308s, Decode+Unpack: 0.428s +---------------------- -------------------------------------------------------- +💾 Converting with 91.4104 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3219.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3219.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst (30/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 200, 128) +Output shape: (1, 200, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) -> torch.Size([1, 1, 200, 512]) + layer.4.output: torch.Size([1, 200, 3584]) -> torch.Size([1, 1, 200, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,668B, BPFP=0.8334 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,760B, BPFP=2.8719 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,400B, BPFP=1.5156 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,828B, BPFP=2.7991 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,256B, BPFP=1.8169 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,316B, BPFP=2.7591 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,712B, BPFP=1.6963 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,692B, BPFP=2.7884 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,092B, BPFP=2.4291 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,688B, BPFP=2.7100 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,268B, BPFP=0.8735 +⌛️ [2/4] FRONTEND: Frontend time: 0.295s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.423s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 200, 128]) + layer.0.v_cache: torch.Size([1, 4, 200, 128]) + layer.1.k_cache: torch.Size([1, 4, 200, 128]) + layer.1.v_cache: torch.Size([1, 4, 200, 128]) + layer.2.k_cache: torch.Size([1, 4, 200, 128]) + layer.2.v_cache: torch.Size([1, 4, 200, 128]) + layer.3.k_cache: torch.Size([1, 4, 200, 128]) + layer.3.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.k_cache: torch.Size([1, 4, 200, 128]) + layer.4.v_cache: torch.Size([1, 4, 200, 128]) + layer.4.output: torch.Size([1, 200, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11599884 23.48974609 + layer.0.v_cache 0.00001671 0.00642847 + layer.1.k_cache 0.22559120 3.38623291 + layer.1.v_cache 0.00000595 0.00257088 + layer.2.k_cache 0.01146694 0.48013012 + layer.2.v_cache 0.00002201 0.00696345 + layer.3.k_cache 0.03668072 2.14747482 + layer.3.v_cache 0.00002131 0.00762705 + layer.4.k_cache 0.00066762 0.14081296 + layer.4.v_cache 0.00005146 0.01393545 + layer.4.output 1.53061454 265.77238839 + ------------------------------------------------------------------------------------- + TOTAL 0.65322497 111.18168476 + (elements=1,740,800) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1740800 +Total Bytes 362680 +BPFP 1.6667 bits/point +EBPFP 3.3335 equivalent bits/point +MSE 111.181685 +---------------------- -------------------------------------------------------- +Time: 0.725s Load: 0.007s, Pack+Encode: 0.295s, Decode+Unpack: 0.423s +---------------------- -------------------------------------------------------- +💾 Converting with 111.1817 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3239.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3239.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst (31/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,984B, BPFP=0.8177 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,400B, BPFP=2.7566 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,908B, BPFP=1.4948 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,388B, BPFP=2.6875 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,036B, BPFP=1.7765 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,896B, BPFP=2.6539 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,232B, BPFP=1.6534 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,412B, BPFP=2.6891 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,000B, BPFP=2.3199 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,004B, BPFP=2.5931 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 99,944B, BPFP=0.9742 +⌛️ [2/4] FRONTEND: Frontend time: 0.300s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.433s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11605523 23.07841072 + layer.0.v_cache 0.00001866 0.00668332 + layer.1.k_cache 0.31881074 3.21785766 + layer.1.v_cache 0.00000653 0.00263536 + layer.2.k_cache 0.02091764 0.42862645 + layer.2.v_cache 0.00002113 0.00705282 + layer.3.k_cache 0.03221333 2.06201172 + layer.3.v_cache 0.00002185 0.00805816 + layer.4.k_cache 0.00068277 0.14177917 + layer.4.v_cache 0.00005596 0.01402012 + layer.4.output 1.33695214 234.66527604 + ------------------------------------------------------------------------------------- + TOTAL 0.57926287 98.33082752 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 414204 +BPFP 1.6625 bits/point +EBPFP 3.3249 equivalent bits/point +MSE 98.330828 +---------------------- -------------------------------------------------------- +Time: 0.742s Load: 0.008s, Pack+Encode: 0.300s, Decode+Unpack: 0.433s +---------------------- -------------------------------------------------------- +💾 Converting with 98.3308 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3246.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3246.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst (32/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 192, 128) +Output shape: (1, 192, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) -> torch.Size([1, 1, 192, 512]) + layer.4.output: torch.Size([1, 192, 3584]) -> torch.Size([1, 1, 192, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,496B, BPFP=0.7728 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,952B, BPFP=2.5189 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,480B, BPFP=1.3411 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,320B, BPFP=2.4674 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,820B, BPFP=1.6130 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,956B, BPFP=2.4378 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,676B, BPFP=1.5199 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,316B, BPFP=2.4671 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,480B, BPFP=2.1549 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,412B, BPFP=2.3936 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 79,744B, BPFP=0.9271 +⌛️ [2/4] FRONTEND: Frontend time: 0.278s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.398s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 192, 128]) + layer.0.v_cache: torch.Size([1, 4, 192, 128]) + layer.1.k_cache: torch.Size([1, 4, 192, 128]) + layer.1.v_cache: torch.Size([1, 4, 192, 128]) + layer.2.k_cache: torch.Size([1, 4, 192, 128]) + layer.2.v_cache: torch.Size([1, 4, 192, 128]) + layer.3.k_cache: torch.Size([1, 4, 192, 128]) + layer.3.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.k_cache: torch.Size([1, 4, 192, 128]) + layer.4.v_cache: torch.Size([1, 4, 192, 128]) + layer.4.output: torch.Size([1, 192, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12444629 21.59647878 + layer.0.v_cache 0.00001540 0.00627674 + layer.1.k_cache 0.18411521 2.99551328 + layer.1.v_cache 0.00000585 0.00251397 + layer.2.k_cache 0.01124177 0.38707217 + layer.2.v_cache 0.00002179 0.00721036 + layer.3.k_cache 0.01616090 1.80072784 + layer.3.v_cache 0.00002117 0.00798022 + layer.4.k_cache 0.00068507 0.14182472 + layer.4.v_cache 0.00005275 0.01374230 + layer.4.output 0.00839530 238.60005115 + ------------------------------------------------------------------------------------- + TOTAL 0.02326666 99.83292344 + (elements=1,671,168) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1671168 +Total Bytes 321652 +BPFP 1.5398 bits/point +EBPFP 3.0795 equivalent bits/point +MSE 99.832923 +---------------------- -------------------------------------------------------- +Time: 0.685s Load: 0.008s, Pack+Encode: 0.278s, Decode+Unpack: 0.398s +---------------------- -------------------------------------------------------- +💾 Converting with 99.8329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3248.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3248.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst (33/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,132B, BPFP=0.8596 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,956B, BPFP=2.8197 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,668B, BPFP=1.5689 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,272B, BPFP=2.7553 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,460B, BPFP=1.8317 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,872B, BPFP=2.7176 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,500B, BPFP=1.7413 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,204B, BPFP=2.7489 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,488B, BPFP=2.3991 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,252B, BPFP=2.6593 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 67,812B, BPFP=0.9118 +⌛️ [2/4] FRONTEND: Frontend time: 0.275s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.373s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09999813 24.49602168 + layer.0.v_cache 0.00001899 0.00671019 + layer.1.k_cache 0.12989934 3.23260204 + layer.1.v_cache 0.00000610 0.00268724 + layer.2.k_cache 0.00966061 0.47573742 + layer.2.v_cache 0.00001994 0.00748028 + layer.3.k_cache 0.03384685 1.87502482 + layer.3.v_cache 0.00002040 0.00800619 + layer.4.k_cache 0.00066502 0.14457326 + layer.4.v_cache 0.00005200 0.01398220 + layer.4.output 0.00959846 328.40027969 + ------------------------------------------------------------------------------------- + TOTAL 0.02008098 137.00381077 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 302616 +BPFP 1.6755 bits/point +EBPFP 3.3511 equivalent bits/point +MSE 137.003811 +---------------------- -------------------------------------------------------- +Time: 0.654s Load: 0.006s, Pack+Encode: 0.275s, Decode+Unpack: 0.373s +---------------------- -------------------------------------------------------- +💾 Converting with 137.0038 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3259.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3259.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst (34/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 428, 128) +Output shape: (1, 428, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) -> torch.Size([1, 1, 428, 512]) + layer.4.output: torch.Size([1, 428, 3584]) -> torch.Size([1, 1, 428, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 21,752B, BPFP=0.7941 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 72,272B, BPFP=2.6384 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 38,484B, BPFP=1.4049 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 70,340B, BPFP=2.5679 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 46,412B, BPFP=1.6944 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 69,452B, BPFP=2.5355 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 43,260B, BPFP=1.5793 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 70,208B, BPFP=2.5631 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 60,572B, BPFP=2.2113 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 67,932B, BPFP=2.4800 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 168,264B, BPFP=0.8775 +⌛️ [2/4] FRONTEND: Frontend time: 0.472s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.722s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 428, 128]) + layer.0.v_cache: torch.Size([1, 4, 428, 128]) + layer.1.k_cache: torch.Size([1, 4, 428, 128]) + layer.1.v_cache: torch.Size([1, 4, 428, 128]) + layer.2.k_cache: torch.Size([1, 4, 428, 128]) + layer.2.v_cache: torch.Size([1, 4, 428, 128]) + layer.3.k_cache: torch.Size([1, 4, 428, 128]) + layer.3.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.k_cache: torch.Size([1, 4, 428, 128]) + layer.4.v_cache: torch.Size([1, 4, 428, 128]) + layer.4.output: torch.Size([1, 428, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13760449 21.99661397 + layer.0.v_cache 0.00001624 0.00625431 + layer.1.k_cache 0.89392254 3.19470158 + layer.1.v_cache 0.00000615 0.00239865 + layer.2.k_cache 0.03917003 0.45577404 + layer.2.v_cache 0.00002113 0.00668308 + layer.3.k_cache 0.03749864 1.94627922 + layer.3.v_cache 0.00002059 0.00712276 + layer.4.k_cache 0.00086919 0.14001103 + layer.4.v_cache 0.00005177 0.01322383 + layer.4.output 0.00617707 124.41721253 + ------------------------------------------------------------------------------------- + TOTAL 0.06778943 52.86409119 + (elements=3,725,312) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3725312 +Total Bytes 728948 +BPFP 1.5654 bits/point +EBPFP 3.1308 equivalent bits/point +MSE 52.864091 +---------------------- -------------------------------------------------------- +Time: 1.208s Load: 0.015s, Pack+Encode: 0.472s, Decode+Unpack: 0.722s +---------------------- -------------------------------------------------------- +💾 Converting with 52.8641 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3272.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3272.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst (35/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 282, 128) +Output shape: (1, 282, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) -> torch.Size([1, 1, 282, 512]) + layer.4.output: torch.Size([1, 282, 3584]) -> torch.Size([1, 1, 282, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,164B, BPFP=0.8402 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,756B, BPFP=2.7015 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,292B, BPFP=1.4568 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,624B, BPFP=2.6387 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,316B, BPFP=1.7352 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,964B, BPFP=2.6022 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,172B, BPFP=1.6164 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,476B, BPFP=2.6305 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,876B, BPFP=2.2648 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,844B, BPFP=2.5401 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 116,108B, BPFP=0.9190 +⌛️ [2/4] FRONTEND: Frontend time: 0.369s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.538s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 282, 128]) + layer.0.v_cache: torch.Size([1, 4, 282, 128]) + layer.1.k_cache: torch.Size([1, 4, 282, 128]) + layer.1.v_cache: torch.Size([1, 4, 282, 128]) + layer.2.k_cache: torch.Size([1, 4, 282, 128]) + layer.2.v_cache: torch.Size([1, 4, 282, 128]) + layer.3.k_cache: torch.Size([1, 4, 282, 128]) + layer.3.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.k_cache: torch.Size([1, 4, 282, 128]) + layer.4.v_cache: torch.Size([1, 4, 282, 128]) + layer.4.output: torch.Size([1, 282, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11385377 22.18141380 + layer.0.v_cache 0.00001705 0.00621422 + layer.1.k_cache 0.49372572 2.99608704 + layer.1.v_cache 0.00000626 0.00238393 + layer.2.k_cache 0.02422354 0.45823788 + layer.2.v_cache 0.00002046 0.00627317 + layer.3.k_cache 0.04153839 1.90223055 + layer.3.v_cache 0.00002075 0.00726804 + layer.4.k_cache 0.00074884 0.13555854 + layer.4.v_cache 0.00004954 0.01258158 + layer.4.output 0.00474379 193.18271910 + ------------------------------------------------------------------------------------- + TOTAL 0.04161240 81.17572250 + (elements=2,454,528) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2454528 +Total Bytes 495592 +BPFP 1.6153 bits/point +EBPFP 3.2305 equivalent bits/point +MSE 81.175722 +---------------------- -------------------------------------------------------- +Time: 0.919s Load: 0.012s, Pack+Encode: 0.369s, Decode+Unpack: 0.538s +---------------------- -------------------------------------------------------- +💾 Converting with 81.1757 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3288.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3288.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst (36/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.020s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 323, 128) +Output shape: (1, 323, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) -> torch.Size([1, 1, 323, 512]) + layer.4.output: torch.Size([1, 323, 3584]) -> torch.Size([1, 1, 323, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,724B, BPFP=0.8090 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 56,828B, BPFP=2.7490 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 29,412B, BPFP=1.4228 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 55,500B, BPFP=2.6848 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 35,192B, BPFP=1.7024 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 54,460B, BPFP=2.6345 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 32,944B, BPFP=1.5937 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 55,116B, BPFP=2.6662 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 47,224B, BPFP=2.2844 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 53,444B, BPFP=2.5853 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 134,652B, BPFP=0.9305 +⌛️ [2/4] FRONTEND: Frontend time: 0.512s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.585s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 323, 128]) + layer.0.v_cache: torch.Size([1, 4, 323, 128]) + layer.1.k_cache: torch.Size([1, 4, 323, 128]) + layer.1.v_cache: torch.Size([1, 4, 323, 128]) + layer.2.k_cache: torch.Size([1, 4, 323, 128]) + layer.2.v_cache: torch.Size([1, 4, 323, 128]) + layer.3.k_cache: torch.Size([1, 4, 323, 128]) + layer.3.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.k_cache: torch.Size([1, 4, 323, 128]) + layer.4.v_cache: torch.Size([1, 4, 323, 128]) + layer.4.output: torch.Size([1, 323, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12985682 22.00326226 + layer.0.v_cache 0.00001645 0.00603002 + layer.1.k_cache 0.61935935 3.14019605 + layer.1.v_cache 0.00000656 0.00237977 + layer.2.k_cache 0.01492825 0.42179823 + layer.2.v_cache 0.00002178 0.00614978 + layer.3.k_cache 0.02579401 1.78224716 + layer.3.v_cache 0.00002068 0.00662339 + layer.4.k_cache 0.00071613 0.13323611 + layer.4.v_cache 0.00005057 0.01236208 + layer.4.output 0.04141478 159.68817724 + ------------------------------------------------------------------------------------- + TOTAL 0.06356906 67.37244268 + (elements=2,811,392) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2811392 +Total Bytes 571496 +BPFP 1.6262 bits/point +EBPFP 3.2525 equivalent bits/point +MSE 67.372443 +---------------------- -------------------------------------------------------- +Time: 1.118s Load: 0.020s, Pack+Encode: 0.512s, Decode+Unpack: 0.585s +---------------------- -------------------------------------------------------- +💾 Converting with 67.3724 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3308.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3308.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst (37/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 286, 128) +Output shape: (1, 286, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) -> torch.Size([1, 1, 286, 512]) + layer.4.output: torch.Size([1, 286, 3584]) -> torch.Size([1, 1, 286, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,356B, BPFP=0.8389 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,600B, BPFP=2.7098 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,664B, BPFP=1.4567 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,392B, BPFP=2.6438 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,904B, BPFP=1.7430 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,692B, BPFP=2.6056 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 30,204B, BPFP=1.6501 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,384B, BPFP=2.6434 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,640B, BPFP=2.2749 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,736B, BPFP=2.5533 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 115,852B, BPFP=0.9042 +⌛️ [2/4] FRONTEND: Frontend time: 0.324s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.505s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 286, 128]) + layer.0.v_cache: torch.Size([1, 4, 286, 128]) + layer.1.k_cache: torch.Size([1, 4, 286, 128]) + layer.1.v_cache: torch.Size([1, 4, 286, 128]) + layer.2.k_cache: torch.Size([1, 4, 286, 128]) + layer.2.v_cache: torch.Size([1, 4, 286, 128]) + layer.3.k_cache: torch.Size([1, 4, 286, 128]) + layer.3.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.k_cache: torch.Size([1, 4, 286, 128]) + layer.4.v_cache: torch.Size([1, 4, 286, 128]) + layer.4.output: torch.Size([1, 286, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12359551 22.45664370 + layer.0.v_cache 0.00001766 0.00653114 + layer.1.k_cache 0.52692803 2.94658997 + layer.1.v_cache 0.00000637 0.00255083 + layer.2.k_cache 0.02102186 0.43534990 + layer.2.v_cache 0.00002187 0.00672842 + layer.3.k_cache 0.03868174 2.08286760 + layer.3.v_cache 0.00002097 0.00765004 + layer.4.k_cache 0.00070799 0.13802223 + layer.4.v_cache 0.00005204 0.01335917 + layer.4.output 0.00469152 192.37100400 + ------------------------------------------------------------------------------------- + TOTAL 0.04375851 80.86431300 + (elements=2,489,344) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2489344 +Total Bytes 502424 +BPFP 1.6146 bits/point +EBPFP 3.2293 equivalent bits/point +MSE 80.864313 +---------------------- -------------------------------------------------------- +Time: 0.839s Load: 0.010s, Pack+Encode: 0.324s, Decode+Unpack: 0.505s +---------------------- -------------------------------------------------------- +💾 Converting with 80.8643 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3331.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3331.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst (38/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,864B, BPFP=0.8530 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,780B, BPFP=2.8879 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,320B, BPFP=1.5170 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,888B, BPFP=2.8178 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,148B, BPFP=1.8175 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,428B, BPFP=2.7817 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,768B, BPFP=1.7092 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,916B, BPFP=2.8200 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,844B, BPFP=2.4218 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,676B, BPFP=2.7227 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,708B, BPFP=0.9389 +⌛️ [2/4] FRONTEND: Frontend time: 0.344s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10671347 24.07714844 + layer.0.v_cache 0.00001788 0.00687338 + layer.1.k_cache 0.15257932 3.13408780 + layer.1.v_cache 0.00000642 0.00257011 + layer.2.k_cache 0.02197224 0.44636513 + layer.2.v_cache 0.00002094 0.00673338 + layer.3.k_cache 0.03325213 1.96654691 + layer.3.v_cache 0.00002126 0.00805197 + layer.4.k_cache 0.00067229 0.13453440 + layer.4.v_cache 0.00005023 0.01295650 + layer.4.output 1.53834953 267.33228643 + ------------------------------------------------------------------------------------- + TOTAL 0.65198546 111.83069841 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 368340 +BPFP 1.7012 bits/point +EBPFP 3.4025 equivalent bits/point +MSE 111.830698 +---------------------- -------------------------------------------------------- +Time: 0.820s Load: 0.007s, Pack+Encode: 0.344s, Decode+Unpack: 0.470s +---------------------- -------------------------------------------------------- +💾 Converting with 111.8307 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3343.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3343.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst (39/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 251, 128) +Output shape: (1, 251, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) -> torch.Size([1, 1, 251, 512]) + layer.4.output: torch.Size([1, 251, 3584]) -> torch.Size([1, 1, 251, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,932B, BPFP=0.8050 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 42,052B, BPFP=2.6178 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,972B, BPFP=1.4300 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 41,160B, BPFP=2.5623 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,552B, BPFP=1.7151 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,612B, BPFP=2.5281 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,152B, BPFP=1.6280 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 41,128B, BPFP=2.5603 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,588B, BPFP=2.2154 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,664B, BPFP=2.4691 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 107,808B, BPFP=0.9587 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.436s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 251, 128]) + layer.0.v_cache: torch.Size([1, 4, 251, 128]) + layer.1.k_cache: torch.Size([1, 4, 251, 128]) + layer.1.v_cache: torch.Size([1, 4, 251, 128]) + layer.2.k_cache: torch.Size([1, 4, 251, 128]) + layer.2.v_cache: torch.Size([1, 4, 251, 128]) + layer.3.k_cache: torch.Size([1, 4, 251, 128]) + layer.3.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.k_cache: torch.Size([1, 4, 251, 128]) + layer.4.v_cache: torch.Size([1, 4, 251, 128]) + layer.4.output: torch.Size([1, 251, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15214830 21.90804750 + layer.0.v_cache 0.00001722 0.00690929 + layer.1.k_cache 0.49754258 3.06107260 + layer.1.v_cache 0.00000639 0.00273496 + layer.2.k_cache 0.01862443 0.42827597 + layer.2.v_cache 0.00002219 0.00711341 + layer.3.k_cache 0.01200796 1.91065620 + layer.3.v_cache 0.00002210 0.00814199 + layer.4.k_cache 0.00075226 0.13983464 + layer.4.v_cache 0.00005409 0.01408757 + layer.4.output 1.21979868 207.68186184 + ------------------------------------------------------------------------------------- + TOTAL 0.54234049 87.13293571 + (elements=2,184,704) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2184704 +Total Bytes 437620 +BPFP 1.6025 bits/point +EBPFP 3.2050 equivalent bits/point +MSE 87.132936 +---------------------- -------------------------------------------------------- +Time: 0.765s Load: 0.010s, Pack+Encode: 0.319s, Decode+Unpack: 0.436s +---------------------- -------------------------------------------------------- +💾 Converting with 87.1329 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3362.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3362.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst (40/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 305, 128) +Output shape: (1, 305, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) -> torch.Size([1, 1, 305, 512]) + layer.4.output: torch.Size([1, 305, 3584]) -> torch.Size([1, 1, 305, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,256B, BPFP=0.8328 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 51,792B, BPFP=2.6533 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,412B, BPFP=1.4555 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,628B, BPFP=2.5936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,592B, BPFP=1.7209 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 50,036B, BPFP=2.5633 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,808B, BPFP=1.6295 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 50,720B, BPFP=2.5984 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,940B, BPFP=2.2510 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 48,972B, BPFP=2.5088 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 120,700B, BPFP=0.8833 +⌛️ [2/4] FRONTEND: Frontend time: 0.360s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.502s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 305, 128]) + layer.0.v_cache: torch.Size([1, 4, 305, 128]) + layer.1.k_cache: torch.Size([1, 4, 305, 128]) + layer.1.v_cache: torch.Size([1, 4, 305, 128]) + layer.2.k_cache: torch.Size([1, 4, 305, 128]) + layer.2.v_cache: torch.Size([1, 4, 305, 128]) + layer.3.k_cache: torch.Size([1, 4, 305, 128]) + layer.3.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.k_cache: torch.Size([1, 4, 305, 128]) + layer.4.v_cache: torch.Size([1, 4, 305, 128]) + layer.4.output: torch.Size([1, 305, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.18432382 23.01147861 + layer.0.v_cache 0.00001629 0.00663389 + layer.1.k_cache 0.53789903 3.26784388 + layer.1.v_cache 0.00000607 0.00265615 + layer.2.k_cache 0.01514517 0.43779192 + layer.2.v_cache 0.00002151 0.00704227 + layer.3.k_cache 0.04182221 1.86210057 + layer.3.v_cache 0.00002017 0.00751113 + layer.4.k_cache 0.00068627 0.14009953 + layer.4.v_cache 0.00005063 0.01388037 + layer.4.output 0.04375917 173.69664813 + ------------------------------------------------------------------------------------- + TOTAL 0.06390032 73.21373972 + (elements=2,654,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2654720 +Total Bytes 526856 +BPFP 1.5877 bits/point +EBPFP 3.1754 equivalent bits/point +MSE 73.213740 +---------------------- -------------------------------------------------------- +Time: 0.872s Load: 0.010s, Pack+Encode: 0.360s, Decode+Unpack: 0.502s +---------------------- -------------------------------------------------------- +💾 Converting with 73.2137 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3363.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3363.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst (41/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.011s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 371, 128) +Output shape: (1, 371, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) -> torch.Size([1, 1, 371, 512]) + layer.4.output: torch.Size([1, 371, 3584]) -> torch.Size([1, 1, 371, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 19,612B, BPFP=0.8260 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 62,540B, BPFP=2.6339 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 33,640B, BPFP=1.4168 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 61,132B, BPFP=2.5746 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 40,440B, BPFP=1.7032 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 60,340B, BPFP=2.5413 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 38,120B, BPFP=1.6055 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 61,000B, BPFP=2.5691 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 52,684B, BPFP=2.2188 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 58,856B, BPFP=2.4788 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 152,292B, BPFP=0.9163 +⌛️ [2/4] FRONTEND: Frontend time: 0.413s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.549s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 371, 128]) + layer.0.v_cache: torch.Size([1, 4, 371, 128]) + layer.1.k_cache: torch.Size([1, 4, 371, 128]) + layer.1.v_cache: torch.Size([1, 4, 371, 128]) + layer.2.k_cache: torch.Size([1, 4, 371, 128]) + layer.2.v_cache: torch.Size([1, 4, 371, 128]) + layer.3.k_cache: torch.Size([1, 4, 371, 128]) + layer.3.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.k_cache: torch.Size([1, 4, 371, 128]) + layer.4.v_cache: torch.Size([1, 4, 371, 128]) + layer.4.output: torch.Size([1, 371, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16952868 21.49989603 + layer.0.v_cache 0.00001698 0.00652001 + layer.1.k_cache 0.82091874 3.23653212 + layer.1.v_cache 0.00000666 0.00261672 + layer.2.k_cache 0.02418445 0.43987091 + layer.2.v_cache 0.00002238 0.00687722 + layer.3.k_cache 0.03055777 1.87766087 + layer.3.v_cache 0.00002117 0.00749672 + layer.4.k_cache 0.00071103 0.13705457 + layer.4.v_cache 0.00005915 0.01350524 + layer.4.output 0.03610923 138.80514777 + ------------------------------------------------------------------------------------- + TOTAL 0.07639951 58.75670969 + (elements=3,229,184) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3229184 +Total Bytes 640656 +BPFP 1.5872 bits/point +EBPFP 3.1743 equivalent bits/point +MSE 58.756710 +---------------------- -------------------------------------------------------- +Time: 0.973s Load: 0.011s, Pack+Encode: 0.413s, Decode+Unpack: 0.549s +---------------------- -------------------------------------------------------- +💾 Converting with 58.7567 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3371.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3371.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst (42/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 218, 128) +Output shape: (1, 218, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) -> torch.Size([1, 1, 218, 512]) + layer.4.output: torch.Size([1, 218, 3584]) -> torch.Size([1, 1, 218, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,728B, BPFP=0.8406 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,560B, BPFP=2.7638 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,880B, BPFP=1.4966 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,644B, BPFP=2.6981 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,904B, BPFP=1.7850 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,024B, BPFP=2.6537 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,712B, BPFP=1.6995 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,748B, BPFP=2.7056 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,524B, BPFP=2.3311 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,244B, BPFP=2.5978 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,044B, BPFP=0.9527 +⌛️ [2/4] FRONTEND: Frontend time: 0.334s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.441s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 218, 128]) + layer.0.v_cache: torch.Size([1, 4, 218, 128]) + layer.1.k_cache: torch.Size([1, 4, 218, 128]) + layer.1.v_cache: torch.Size([1, 4, 218, 128]) + layer.2.k_cache: torch.Size([1, 4, 218, 128]) + layer.2.v_cache: torch.Size([1, 4, 218, 128]) + layer.3.k_cache: torch.Size([1, 4, 218, 128]) + layer.3.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.k_cache: torch.Size([1, 4, 218, 128]) + layer.4.v_cache: torch.Size([1, 4, 218, 128]) + layer.4.output: torch.Size([1, 218, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13674593 23.53377876 + layer.0.v_cache 0.00001768 0.00698790 + layer.1.k_cache 0.26206151 2.99394800 + layer.1.v_cache 0.00000624 0.00275174 + layer.2.k_cache 0.02938290 0.45632641 + layer.2.v_cache 0.00002158 0.00701372 + layer.3.k_cache 0.05653211 1.91133748 + layer.3.v_cache 0.00002129 0.00789237 + layer.4.k_cache 0.00070776 0.13914253 + layer.4.v_cache 0.00005193 0.01418639 + layer.4.output 1.40434551 243.53659486 + ------------------------------------------------------------------------------------- + TOTAL 0.60682162 101.98997231 + (elements=1,897,472) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1897472 +Total Bytes 394012 +BPFP 1.6612 bits/point +EBPFP 3.3224 equivalent bits/point +MSE 101.989972 +---------------------- -------------------------------------------------------- +Time: 0.783s Load: 0.008s, Pack+Encode: 0.334s, Decode+Unpack: 0.441s +---------------------- -------------------------------------------------------- +💾 Converting with 101.9900 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3374.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3374.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst (43/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 193, 128) +Output shape: (1, 193, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) -> torch.Size([1, 1, 193, 512]) + layer.4.output: torch.Size([1, 193, 3584]) -> torch.Size([1, 1, 193, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,856B, BPFP=0.7979 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 35,996B, BPFP=2.9142 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,564B, BPFP=1.5029 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,076B, BPFP=2.8397 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,716B, BPFP=1.8391 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,648B, BPFP=2.8051 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 20,856B, BPFP=1.6885 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,260B, BPFP=2.8546 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,172B, BPFP=2.4427 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 33,908B, BPFP=2.7451 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,004B, BPFP=0.9600 +⌛️ [2/4] FRONTEND: Frontend time: 0.311s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.447s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 193, 128]) + layer.0.v_cache: torch.Size([1, 4, 193, 128]) + layer.1.k_cache: torch.Size([1, 4, 193, 128]) + layer.1.v_cache: torch.Size([1, 4, 193, 128]) + layer.2.k_cache: torch.Size([1, 4, 193, 128]) + layer.2.v_cache: torch.Size([1, 4, 193, 128]) + layer.3.k_cache: torch.Size([1, 4, 193, 128]) + layer.3.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.k_cache: torch.Size([1, 4, 193, 128]) + layer.4.v_cache: torch.Size([1, 4, 193, 128]) + layer.4.output: torch.Size([1, 193, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14305212 24.13393074 + layer.0.v_cache 0.00001770 0.00699705 + layer.1.k_cache 0.14179654 3.02752844 + layer.1.v_cache 0.00000608 0.00270281 + layer.2.k_cache 0.01555968 0.43575694 + layer.2.v_cache 0.00002047 0.00732464 + layer.3.k_cache 0.02655400 1.87413159 + layer.3.v_cache 0.00002031 0.00831234 + layer.4.k_cache 0.00069483 0.14707092 + layer.4.v_cache 0.00005004 0.01421482 + layer.4.output 1.58615237 276.15426073 + ------------------------------------------------------------------------------------- + TOTAL 0.67240225 115.45516444 + (elements=1,679,872) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1679872 +Total Bytes 360056 +BPFP 1.7147 bits/point +EBPFP 3.4294 equivalent bits/point +MSE 115.455164 +---------------------- -------------------------------------------------------- +Time: 0.766s Load: 0.007s, Pack+Encode: 0.311s, Decode+Unpack: 0.447s +---------------------- -------------------------------------------------------- +💾 Converting with 115.4552 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3387.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3387.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst (44/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 249, 128) +Output shape: (1, 249, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) -> torch.Size([1, 1, 249, 512]) + layer.4.output: torch.Size([1, 249, 3584]) -> torch.Size([1, 1, 249, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,848B, BPFP=0.8062 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,564B, BPFP=2.6082 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,508B, BPFP=1.4124 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,552B, BPFP=2.5447 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,040B, BPFP=1.6968 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,204B, BPFP=2.5228 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,516B, BPFP=1.6012 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,600B, BPFP=2.5477 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,984B, BPFP=2.1953 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,248B, BPFP=2.4629 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 103,228B, BPFP=0.9254 +⌛️ [2/4] FRONTEND: Frontend time: 0.297s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.458s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 249, 128]) + layer.0.v_cache: torch.Size([1, 4, 249, 128]) + layer.1.k_cache: torch.Size([1, 4, 249, 128]) + layer.1.v_cache: torch.Size([1, 4, 249, 128]) + layer.2.k_cache: torch.Size([1, 4, 249, 128]) + layer.2.v_cache: torch.Size([1, 4, 249, 128]) + layer.3.k_cache: torch.Size([1, 4, 249, 128]) + layer.3.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.k_cache: torch.Size([1, 4, 249, 128]) + layer.4.v_cache: torch.Size([1, 4, 249, 128]) + layer.4.output: torch.Size([1, 249, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10821181 23.54163137 + layer.0.v_cache 0.00001942 0.00631715 + layer.1.k_cache 0.30751975 3.05738579 + layer.1.v_cache 0.00000599 0.00232124 + layer.2.k_cache 0.01421910 0.43925979 + layer.2.v_cache 0.00002020 0.00641798 + layer.3.k_cache 0.01730646 1.91233354 + layer.3.v_cache 0.00002068 0.00756270 + layer.4.k_cache 0.00075739 0.13470248 + layer.4.v_cache 0.00005209 0.01311352 + layer.4.output 1.22958295 212.52952883 + ------------------------------------------------------------------------------------- + TOTAL 0.53265962 89.22516161 + (elements=2,167,296) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2167296 +Total Bytes 428292 +BPFP 1.5809 bits/point +EBPFP 3.1619 equivalent bits/point +MSE 89.225162 +---------------------- -------------------------------------------------------- +Time: 0.764s Load: 0.009s, Pack+Encode: 0.297s, Decode+Unpack: 0.458s +---------------------- -------------------------------------------------------- +💾 Converting with 89.2252 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3467.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3467.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst (45/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 202, 128) +Output shape: (1, 202, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) -> torch.Size([1, 1, 202, 512]) + layer.4.output: torch.Size([1, 202, 3584]) -> torch.Size([1, 1, 202, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,908B, BPFP=0.8438 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,448B, BPFP=2.8967 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,864B, BPFP=1.5365 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,440B, BPFP=2.8187 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,812B, BPFP=1.8419 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,872B, BPFP=2.7748 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,260B, BPFP=1.7218 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,588B, BPFP=2.8301 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,184B, BPFP=2.4121 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,080B, BPFP=2.7135 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,712B, BPFP=0.9803 +⌛️ [2/4] FRONTEND: Frontend time: 0.312s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.426s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 202, 128]) + layer.0.v_cache: torch.Size([1, 4, 202, 128]) + layer.1.k_cache: torch.Size([1, 4, 202, 128]) + layer.1.v_cache: torch.Size([1, 4, 202, 128]) + layer.2.k_cache: torch.Size([1, 4, 202, 128]) + layer.2.v_cache: torch.Size([1, 4, 202, 128]) + layer.3.k_cache: torch.Size([1, 4, 202, 128]) + layer.3.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.k_cache: torch.Size([1, 4, 202, 128]) + layer.4.v_cache: torch.Size([1, 4, 202, 128]) + layer.4.output: torch.Size([1, 202, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15282421 23.92030863 + layer.0.v_cache 0.00001801 0.00723968 + layer.1.k_cache 0.08597872 3.24606776 + layer.1.v_cache 0.00000649 0.00275510 + layer.2.k_cache 0.01425038 0.41644212 + layer.2.v_cache 0.00002047 0.00705536 + layer.3.k_cache 0.05704349 2.00659240 + layer.3.v_cache 0.00002256 0.00813798 + layer.4.k_cache 0.00069767 0.14455092 + layer.4.v_cache 0.00005056 0.01383864 + layer.4.output 1.51553867 263.02917256 + ------------------------------------------------------------------------------------- + TOTAL 0.64233431 110.05748215 + (elements=1,758,208) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1758208 +Total Bytes 378168 +BPFP 1.7207 bits/point +EBPFP 3.4414 equivalent bits/point +MSE 110.057482 +---------------------- -------------------------------------------------------- +Time: 0.746s Load: 0.008s, Pack+Encode: 0.312s, Decode+Unpack: 0.426s +---------------------- -------------------------------------------------------- +💾 Converting with 110.0575 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3475.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3475.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst (46/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 283, 128) +Output shape: (1, 283, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) -> torch.Size([1, 1, 283, 512]) + layer.4.output: torch.Size([1, 283, 3584]) -> torch.Size([1, 1, 283, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,892B, BPFP=0.8222 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,496B, BPFP=2.6776 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,144B, BPFP=1.4435 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,212B, BPFP=2.6067 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,976B, BPFP=1.7102 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,536B, BPFP=2.5693 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,192B, BPFP=1.6117 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,236B, BPFP=2.6080 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,768B, BPFP=2.2509 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,516B, BPFP=2.5130 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 108,824B, BPFP=0.8583 +⌛️ [2/4] FRONTEND: Frontend time: 0.360s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.540s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 283, 128]) + layer.0.v_cache: torch.Size([1, 4, 283, 128]) + layer.1.k_cache: torch.Size([1, 4, 283, 128]) + layer.1.v_cache: torch.Size([1, 4, 283, 128]) + layer.2.k_cache: torch.Size([1, 4, 283, 128]) + layer.2.v_cache: torch.Size([1, 4, 283, 128]) + layer.3.k_cache: torch.Size([1, 4, 283, 128]) + layer.3.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.k_cache: torch.Size([1, 4, 283, 128]) + layer.4.v_cache: torch.Size([1, 4, 283, 128]) + layer.4.output: torch.Size([1, 283, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14837397 22.57309726 + layer.0.v_cache 0.00001603 0.00572686 + layer.1.k_cache 0.60911565 3.23269858 + layer.1.v_cache 0.00000597 0.00214874 + layer.2.k_cache 0.02477714 0.44581981 + layer.2.v_cache 0.00001938 0.00573982 + layer.3.k_cache 0.01052417 1.84183095 + layer.3.v_cache 0.00002039 0.00665834 + layer.4.k_cache 0.00073933 0.12413931 + layer.4.v_cache 0.00005963 0.01253153 + layer.4.output 0.00470031 187.92437532 + ------------------------------------------------------------------------------------- + TOTAL 0.04862081 79.04241285 + (elements=2,463,232) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2463232 +Total Bytes 485792 +BPFP 1.5777 bits/point +EBPFP 3.1555 equivalent bits/point +MSE 79.042413 +---------------------- -------------------------------------------------------- +Time: 0.909s Load: 0.009s, Pack+Encode: 0.360s, Decode+Unpack: 0.540s +---------------------- -------------------------------------------------------- +💾 Converting with 79.0424 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3575.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3575.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst (47/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 210, 128) +Output shape: (1, 210, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) -> torch.Size([1, 1, 210, 512]) + layer.4.output: torch.Size([1, 210, 3584]) -> torch.Size([1, 1, 210, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,464B, BPFP=0.8530 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,808B, BPFP=2.8131 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,160B, BPFP=1.5000 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,892B, BPFP=2.7449 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,264B, BPFP=1.8054 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,320B, BPFP=2.7024 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,640B, BPFP=1.6845 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,868B, BPFP=2.7432 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,768B, BPFP=2.3637 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,564B, BPFP=2.6461 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,152B, BPFP=0.9476 +⌛️ [2/4] FRONTEND: Frontend time: 0.341s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.427s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 210, 128]) + layer.0.v_cache: torch.Size([1, 4, 210, 128]) + layer.1.k_cache: torch.Size([1, 4, 210, 128]) + layer.1.v_cache: torch.Size([1, 4, 210, 128]) + layer.2.k_cache: torch.Size([1, 4, 210, 128]) + layer.2.v_cache: torch.Size([1, 4, 210, 128]) + layer.3.k_cache: torch.Size([1, 4, 210, 128]) + layer.3.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.k_cache: torch.Size([1, 4, 210, 128]) + layer.4.v_cache: torch.Size([1, 4, 210, 128]) + layer.4.output: torch.Size([1, 210, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11415643 23.79568452 + layer.0.v_cache 0.00001882 0.00657515 + layer.1.k_cache 0.27145578 3.19471232 + layer.1.v_cache 0.00000623 0.00261592 + layer.2.k_cache 0.01201470 0.46214087 + layer.2.v_cache 0.00002208 0.00711167 + layer.3.k_cache 0.00887682 2.01973615 + layer.3.v_cache 0.00002165 0.00801739 + layer.4.k_cache 0.00072611 0.13923509 + layer.4.v_cache 0.00005233 0.01401771 + layer.4.output 1.45781997 252.34081633 + ------------------------------------------------------------------------------------- + TOTAL 0.62424063 105.64915065 + (elements=1,827,840) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1827840 +Total Bytes 382900 +BPFP 1.6759 bits/point +EBPFP 3.3517 equivalent bits/point +MSE 105.649151 +---------------------- -------------------------------------------------------- +Time: 0.775s Load: 0.007s, Pack+Encode: 0.341s, Decode+Unpack: 0.427s +---------------------- -------------------------------------------------------- +💾 Converting with 105.6492 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3597.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3597.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst (48/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 165, 128) +Output shape: (1, 165, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) -> torch.Size([1, 1, 165, 512]) + layer.4.output: torch.Size([1, 165, 3584]) -> torch.Size([1, 1, 165, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,108B, BPFP=0.8625 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,096B, BPFP=2.8500 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,708B, BPFP=1.5822 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,404B, BPFP=2.7845 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,584B, BPFP=1.8545 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,996B, BPFP=2.7458 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,596B, BPFP=1.7610 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,492B, BPFP=2.7928 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,680B, BPFP=2.4318 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,384B, BPFP=2.6879 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,456B, BPFP=1.0073 +⌛️ [2/4] FRONTEND: Frontend time: 0.278s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.356s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 165, 128]) + layer.0.v_cache: torch.Size([1, 4, 165, 128]) + layer.1.k_cache: torch.Size([1, 4, 165, 128]) + layer.1.v_cache: torch.Size([1, 4, 165, 128]) + layer.2.k_cache: torch.Size([1, 4, 165, 128]) + layer.2.v_cache: torch.Size([1, 4, 165, 128]) + layer.3.k_cache: torch.Size([1, 4, 165, 128]) + layer.3.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.k_cache: torch.Size([1, 4, 165, 128]) + layer.4.v_cache: torch.Size([1, 4, 165, 128]) + layer.4.output: torch.Size([1, 165, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10245823 24.60990175 + layer.0.v_cache 0.00002070 0.00735174 + layer.1.k_cache 0.13127365 3.29329501 + layer.1.v_cache 0.00000615 0.00279471 + layer.2.k_cache 0.01458293 0.41867741 + layer.2.v_cache 0.00002074 0.00721315 + layer.3.k_cache 0.02375244 1.97743919 + layer.3.v_cache 0.00002214 0.00887179 + layer.4.k_cache 0.00067489 0.15174202 + layer.4.v_cache 0.00005196 0.01475275 + layer.4.output 0.00973386 330.37494589 + ------------------------------------------------------------------------------------- + TOTAL 0.02005887 137.83039181 + (elements=1,436,160) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1436160 +Total Bytes 310504 +BPFP 1.7296 bits/point +EBPFP 3.4593 equivalent bits/point +MSE 137.830392 +---------------------- -------------------------------------------------------- +Time: 0.640s Load: 0.006s, Pack+Encode: 0.278s, Decode+Unpack: 0.356s +---------------------- -------------------------------------------------------- +💾 Converting with 137.8304 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3670.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3670.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst (49/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 183, 128) +Output shape: (1, 183, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) -> torch.Size([1, 1, 183, 512]) + layer.4.output: torch.Size([1, 183, 3584]) -> torch.Size([1, 1, 183, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,924B, BPFP=0.8473 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,456B, BPFP=2.6858 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,648B, BPFP=1.5068 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,840B, BPFP=2.6332 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 21,020B, BPFP=1.7947 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,336B, BPFP=2.5902 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,664B, BPFP=1.6790 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,788B, BPFP=2.6288 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,900B, BPFP=2.2968 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,760B, BPFP=2.5410 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,192B, BPFP=1.0269 +⌛️ [2/4] FRONTEND: Frontend time: 0.272s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 183, 128]) + layer.0.v_cache: torch.Size([1, 4, 183, 128]) + layer.1.k_cache: torch.Size([1, 4, 183, 128]) + layer.1.v_cache: torch.Size([1, 4, 183, 128]) + layer.2.k_cache: torch.Size([1, 4, 183, 128]) + layer.2.v_cache: torch.Size([1, 4, 183, 128]) + layer.3.k_cache: torch.Size([1, 4, 183, 128]) + layer.3.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.k_cache: torch.Size([1, 4, 183, 128]) + layer.4.v_cache: torch.Size([1, 4, 183, 128]) + layer.4.output: torch.Size([1, 183, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12273688 22.95043812 + layer.0.v_cache 0.00001959 0.00724228 + layer.1.k_cache 0.11126613 3.08877747 + layer.1.v_cache 0.00000674 0.00292997 + layer.2.k_cache 0.01416154 0.46901765 + layer.2.v_cache 0.00002164 0.00739803 + layer.3.k_cache 0.04491856 1.85434344 + layer.3.v_cache 0.00002114 0.00882220 + layer.4.k_cache 0.00068813 0.15003225 + layer.4.v_cache 0.00006151 0.01478588 + layer.4.output 0.00884247 292.26341725 + ------------------------------------------------------------------------------------- + TOTAL 0.02092936 122.02339459 + (elements=1,592,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1592832 +Total Bytes 332528 +BPFP 1.6701 bits/point +EBPFP 3.3402 equivalent bits/point +MSE 122.023395 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.006s, Pack+Encode: 0.272s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 122.0234 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3696.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3696.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst (50/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 257, 128) +Output shape: (1, 257, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) -> torch.Size([1, 1, 257, 512]) + layer.4.output: torch.Size([1, 257, 3584]) -> torch.Size([1, 1, 257, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,032B, BPFP=0.7923 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,472B, BPFP=2.8254 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,724B, BPFP=1.4424 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,016B, BPFP=2.7369 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,744B, BPFP=1.7476 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,196B, BPFP=2.6870 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,496B, BPFP=1.6109 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,144B, BPFP=2.7446 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,284B, BPFP=2.3276 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,356B, BPFP=2.6359 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 104,536B, BPFP=0.9079 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.507s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 257, 128]) + layer.0.v_cache: torch.Size([1, 4, 257, 128]) + layer.1.k_cache: torch.Size([1, 4, 257, 128]) + layer.1.v_cache: torch.Size([1, 4, 257, 128]) + layer.2.k_cache: torch.Size([1, 4, 257, 128]) + layer.2.v_cache: torch.Size([1, 4, 257, 128]) + layer.3.k_cache: torch.Size([1, 4, 257, 128]) + layer.3.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.k_cache: torch.Size([1, 4, 257, 128]) + layer.4.v_cache: torch.Size([1, 4, 257, 128]) + layer.4.output: torch.Size([1, 257, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12694686 22.83815471 + layer.0.v_cache 0.00002055 0.00658421 + layer.1.k_cache 0.54844232 3.13485041 + layer.1.v_cache 0.00000676 0.00252372 + layer.2.k_cache 0.01835482 0.45117766 + layer.2.v_cache 0.00002099 0.00659375 + layer.3.k_cache 0.02141366 1.86635949 + layer.3.v_cache 0.00002047 0.00755518 + layer.4.k_cache 0.00070767 0.13409117 + layer.4.v_cache 0.00005046 0.01298643 + layer.4.output 0.00513101 212.72621943 + ------------------------------------------------------------------------------------- + TOTAL 0.04422951 89.26731840 + (elements=2,236,928) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2236928 +Total Bytes 459000 +BPFP 1.6415 bits/point +EBPFP 3.2831 equivalent bits/point +MSE 89.267318 +---------------------- -------------------------------------------------------- +Time: 0.858s Load: 0.009s, Pack+Encode: 0.342s, Decode+Unpack: 0.507s +---------------------- -------------------------------------------------------- +💾 Converting with 89.2673 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3763.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3763.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst (51/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 182, 128) +Output shape: (1, 182, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) -> torch.Size([1, 1, 182, 512]) + layer.4.output: torch.Size([1, 182, 3584]) -> torch.Size([1, 1, 182, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,772B, BPFP=0.8389 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,416B, BPFP=2.6971 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,252B, BPFP=1.4811 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,520B, BPFP=2.6202 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,316B, BPFP=1.7442 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,096B, BPFP=2.5838 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,252B, BPFP=1.6528 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,620B, BPFP=2.6288 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,480B, BPFP=2.2734 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,496B, BPFP=2.5323 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,684B, BPFP=0.9160 +⌛️ [2/4] FRONTEND: Frontend time: 0.290s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.323s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 182, 128]) + layer.0.v_cache: torch.Size([1, 4, 182, 128]) + layer.1.k_cache: torch.Size([1, 4, 182, 128]) + layer.1.v_cache: torch.Size([1, 4, 182, 128]) + layer.2.k_cache: torch.Size([1, 4, 182, 128]) + layer.2.v_cache: torch.Size([1, 4, 182, 128]) + layer.3.k_cache: torch.Size([1, 4, 182, 128]) + layer.3.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.k_cache: torch.Size([1, 4, 182, 128]) + layer.4.v_cache: torch.Size([1, 4, 182, 128]) + layer.4.output: torch.Size([1, 182, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12731336 22.87453586 + layer.0.v_cache 0.00001750 0.00702199 + layer.1.k_cache 0.10052182 3.35337243 + layer.1.v_cache 0.00000572 0.00250995 + layer.2.k_cache 0.01141533 0.43619512 + layer.2.v_cache 0.00001926 0.00661410 + layer.3.k_cache 0.03745206 2.00329070 + layer.3.v_cache 0.00002083 0.00832607 + layer.4.k_cache 0.00066429 0.13770692 + layer.4.v_cache 0.00005101 0.01400586 + layer.4.output 0.00884527 298.90885008 + ------------------------------------------------------------------------------------- + TOTAL 0.01996459 124.77679586 + (elements=1,584,128) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1584128 +Total Bytes 319904 +BPFP 1.6155 bits/point +EBPFP 3.2311 equivalent bits/point +MSE 124.776796 +---------------------- -------------------------------------------------------- +Time: 0.620s Load: 0.007s, Pack+Encode: 0.290s, Decode+Unpack: 0.323s +---------------------- -------------------------------------------------------- +💾 Converting with 124.7768 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3764.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3764.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst (52/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,936B, BPFP=0.8401 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,436B, BPFP=2.7756 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,068B, BPFP=1.4828 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,192B, BPFP=2.6881 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,068B, BPFP=1.7644 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,628B, BPFP=2.6484 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,540B, BPFP=1.6568 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,176B, BPFP=2.6869 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,676B, BPFP=2.2998 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,744B, BPFP=2.5861 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,608B, BPFP=0.9513 +⌛️ [2/4] FRONTEND: Frontend time: 0.307s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.415s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12752426 24.37415541 + layer.0.v_cache 0.00002004 0.00689854 + layer.1.k_cache 0.30660921 3.02862081 + layer.1.v_cache 0.00000613 0.00239530 + layer.2.k_cache 0.01808885 0.45752774 + layer.2.v_cache 0.00002107 0.00661390 + layer.3.k_cache 0.01678786 2.02662947 + layer.3.v_cache 0.00002050 0.00765957 + layer.4.k_cache 0.00068290 0.13478736 + layer.4.v_cache 0.00004851 0.01290296 + layer.4.output 1.37905047 241.83536438 + ------------------------------------------------------------------------------------- + TOTAL 0.59548016 101.34739657 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 399072 +BPFP 1.6522 bits/point +EBPFP 3.3045 equivalent bits/point +MSE 101.347397 +---------------------- -------------------------------------------------------- +Time: 0.730s Load: 0.007s, Pack+Encode: 0.307s, Decode+Unpack: 0.415s +---------------------- -------------------------------------------------------- +💾 Converting with 101.3474 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3806.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3806.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst (53/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 217, 128) +Output shape: (1, 217, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) -> torch.Size([1, 1, 217, 512]) + layer.4.output: torch.Size([1, 217, 3584]) -> torch.Size([1, 1, 217, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,020B, BPFP=0.8655 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,528B, BPFP=2.7742 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,656B, BPFP=1.4873 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,540B, BPFP=2.7031 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,616B, BPFP=1.7725 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,856B, BPFP=2.6538 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,108B, BPFP=1.6639 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,628B, BPFP=2.7094 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,248B, BPFP=2.3220 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,256B, BPFP=2.6106 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,864B, BPFP=0.9758 +⌛️ [2/4] FRONTEND: Frontend time: 0.320s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.413s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 217, 128]) + layer.0.v_cache: torch.Size([1, 4, 217, 128]) + layer.1.k_cache: torch.Size([1, 4, 217, 128]) + layer.1.v_cache: torch.Size([1, 4, 217, 128]) + layer.2.k_cache: torch.Size([1, 4, 217, 128]) + layer.2.v_cache: torch.Size([1, 4, 217, 128]) + layer.3.k_cache: torch.Size([1, 4, 217, 128]) + layer.3.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.k_cache: torch.Size([1, 4, 217, 128]) + layer.4.v_cache: torch.Size([1, 4, 217, 128]) + layer.4.output: torch.Size([1, 217, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12766143 23.89365369 + layer.0.v_cache 0.00001653 0.00661918 + layer.1.k_cache 0.34631467 3.18742771 + layer.1.v_cache 0.00000659 0.00256012 + layer.2.k_cache 0.01786991 0.43537882 + layer.2.v_cache 0.00002041 0.00654601 + layer.3.k_cache 0.06874437 1.87003267 + layer.3.v_cache 0.00002276 0.00824728 + layer.4.k_cache 0.00067882 0.13644138 + layer.4.v_cache 0.00005621 0.01440986 + layer.4.output 1.41085444 244.03423305 + ------------------------------------------------------------------------------------- + TOTAL 0.61396310 102.22358518 + (elements=1,888,768) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1888768 +Total Bytes 394320 +BPFP 1.6702 bits/point +EBPFP 3.3403 equivalent bits/point +MSE 102.223585 +---------------------- -------------------------------------------------------- +Time: 0.740s Load: 0.008s, Pack+Encode: 0.320s, Decode+Unpack: 0.413s +---------------------- -------------------------------------------------------- +💾 Converting with 102.2236 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3817.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3817.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst (54/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,812B, BPFP=0.8166 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,784B, BPFP=2.7506 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,208B, BPFP=1.4663 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,772B, BPFP=2.6806 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,588B, BPFP=1.7691 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,216B, BPFP=2.6421 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,072B, BPFP=1.6643 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,844B, BPFP=2.6856 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,272B, BPFP=2.3003 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,292B, BPFP=2.5783 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 103,592B, BPFP=1.0232 +⌛️ [2/4] FRONTEND: Frontend time: 0.310s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.401s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12070838 23.69947153 + layer.0.v_cache 0.00001763 0.00677381 + layer.1.k_cache 0.31015487 3.26668420 + layer.1.v_cache 0.00000677 0.00265896 + layer.2.k_cache 0.03544329 0.47131672 + layer.2.v_cache 0.00002212 0.00696967 + layer.3.k_cache 0.05682086 1.89023394 + layer.3.v_cache 0.00002136 0.00820152 + layer.4.k_cache 0.00073636 0.14243008 + layer.4.v_cache 0.00005031 0.01374977 + layer.4.output 1.35470685 234.66754899 + ------------------------------------------------------------------------------------- + TOTAL 0.58864294 98.36360783 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 412452 +BPFP 1.6774 bits/point +EBPFP 3.3548 equivalent bits/point +MSE 98.363608 +---------------------- -------------------------------------------------------- +Time: 0.719s Load: 0.008s, Pack+Encode: 0.310s, Decode+Unpack: 0.401s +---------------------- -------------------------------------------------------- +💾 Converting with 98.3636 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3985.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3985.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst (55/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,280B, BPFP=0.8275 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,220B, BPFP=2.8037 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,644B, BPFP=1.5144 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,388B, BPFP=2.7427 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,700B, BPFP=1.8119 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,784B, BPFP=2.6984 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,072B, BPFP=1.6925 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,400B, BPFP=2.7435 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,184B, BPFP=2.3609 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,988B, BPFP=2.6400 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,984B, BPFP=0.9954 +⌛️ [2/4] FRONTEND: Frontend time: 0.332s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.445s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13227103 23.35926038 + layer.0.v_cache 0.00002050 0.00686476 + layer.1.k_cache 0.22864332 3.37188979 + layer.1.v_cache 0.00000662 0.00268665 + layer.2.k_cache 0.02950178 0.45251597 + layer.2.v_cache 0.00002095 0.00703632 + layer.3.k_cache 0.01835468 1.94676072 + layer.3.v_cache 0.00002123 0.00821552 + layer.4.k_cache 0.00070999 0.14372947 + layer.4.v_cache 0.00005457 0.01375488 + layer.4.output 1.43734575 251.89727951 + ------------------------------------------------------------------------------------- + TOTAL 0.61594264 105.44668653 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 392644 +BPFP 1.6943 bits/point +EBPFP 3.3886 equivalent bits/point +MSE 105.446687 +---------------------- -------------------------------------------------------- +Time: 0.785s Load: 0.009s, Pack+Encode: 0.332s, Decode+Unpack: 0.445s +---------------------- -------------------------------------------------------- +💾 Converting with 105.4467 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-3988.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-3988.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst (56/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 222, 128) +Output shape: (1, 222, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) -> torch.Size([1, 1, 222, 512]) + layer.4.output: torch.Size([1, 222, 3584]) -> torch.Size([1, 1, 222, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,000B, BPFP=0.8446 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,124B, BPFP=2.7537 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,388B, BPFP=1.5053 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,192B, BPFP=2.6881 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,296B, BPFP=1.7804 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,712B, BPFP=2.6543 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,980B, BPFP=1.6878 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,964B, BPFP=2.6720 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,020B, BPFP=2.3240 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,680B, BPFP=2.5816 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 93,132B, BPFP=0.9364 +⌛️ [2/4] FRONTEND: Frontend time: 0.337s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.464s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 222, 128]) + layer.0.v_cache: torch.Size([1, 4, 222, 128]) + layer.1.k_cache: torch.Size([1, 4, 222, 128]) + layer.1.v_cache: torch.Size([1, 4, 222, 128]) + layer.2.k_cache: torch.Size([1, 4, 222, 128]) + layer.2.v_cache: torch.Size([1, 4, 222, 128]) + layer.3.k_cache: torch.Size([1, 4, 222, 128]) + layer.3.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.k_cache: torch.Size([1, 4, 222, 128]) + layer.4.v_cache: torch.Size([1, 4, 222, 128]) + layer.4.output: torch.Size([1, 222, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12518616 24.33321016 + layer.0.v_cache 0.00001676 0.00656273 + layer.1.k_cache 0.33007166 3.09207016 + layer.1.v_cache 0.00000630 0.00261647 + layer.2.k_cache 0.01131002 0.45150977 + layer.2.v_cache 0.00002125 0.00724776 + layer.3.k_cache 0.02232737 2.05785940 + layer.3.v_cache 0.00001983 0.00754903 + layer.4.k_cache 0.00071526 0.14136387 + layer.4.v_cache 0.00005006 0.01307264 + layer.4.output 1.37904434 241.63286277 + ------------------------------------------------------------------------------------- + TOTAL 0.59664912 101.26724126 + (elements=1,932,288) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1932288 +Total Bytes 398488 +BPFP 1.6498 bits/point +EBPFP 3.2996 equivalent bits/point +MSE 101.267241 +---------------------- -------------------------------------------------------- +Time: 0.809s Load: 0.009s, Pack+Encode: 0.337s, Decode+Unpack: 0.464s +---------------------- -------------------------------------------------------- +💾 Converting with 101.2672 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4019.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4019.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst (57/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 244, 128) +Output shape: (1, 244, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) -> torch.Size([1, 1, 244, 512]) + layer.4.output: torch.Size([1, 244, 3584]) -> torch.Size([1, 1, 244, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,004B, BPFP=0.8327 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,348B, BPFP=2.6478 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,452B, BPFP=1.4378 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,336B, BPFP=2.5830 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,628B, BPFP=1.7052 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,840B, BPFP=2.5512 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,952B, BPFP=1.5978 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,256B, BPFP=2.5779 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,812B, BPFP=2.2293 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,952B, BPFP=2.4944 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 101,000B, BPFP=0.9240 +⌛️ [2/4] FRONTEND: Frontend time: 0.356s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.420s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 244, 128]) + layer.0.v_cache: torch.Size([1, 4, 244, 128]) + layer.1.k_cache: torch.Size([1, 4, 244, 128]) + layer.1.v_cache: torch.Size([1, 4, 244, 128]) + layer.2.k_cache: torch.Size([1, 4, 244, 128]) + layer.2.v_cache: torch.Size([1, 4, 244, 128]) + layer.3.k_cache: torch.Size([1, 4, 244, 128]) + layer.3.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.k_cache: torch.Size([1, 4, 244, 128]) + layer.4.v_cache: torch.Size([1, 4, 244, 128]) + layer.4.output: torch.Size([1, 244, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14655781 22.60952709 + layer.0.v_cache 0.00001709 0.00618536 + layer.1.k_cache 0.36683986 2.93405502 + layer.1.v_cache 0.00000651 0.00256888 + layer.2.k_cache 0.01755252 0.40923953 + layer.2.v_cache 0.00002060 0.00670285 + layer.3.k_cache 0.02713263 1.93218244 + layer.3.v_cache 0.00001997 0.00724951 + layer.4.k_cache 0.00071102 0.13364523 + layer.4.v_cache 0.00005162 0.01252738 + layer.4.output 1.25479176 211.19653835 + ------------------------------------------------------------------------------------- + TOTAL 0.54955600 88.61350893 + (elements=2,123,776) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2123776 +Total Bytes 423580 +BPFP 1.5956 bits/point +EBPFP 3.1911 equivalent bits/point +MSE 88.613509 +---------------------- -------------------------------------------------------- +Time: 0.785s Load: 0.009s, Pack+Encode: 0.356s, Decode+Unpack: 0.420s +---------------------- -------------------------------------------------------- +💾 Converting with 88.6135 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4034.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4034.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst (58/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 155, 128) +Output shape: (1, 155, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) -> torch.Size([1, 1, 155, 512]) + layer.4.output: torch.Size([1, 155, 3584]) -> torch.Size([1, 1, 155, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,696B, BPFP=0.8766 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 28,364B, BPFP=2.8593 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,768B, BPFP=1.5895 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 27,556B, BPFP=2.7778 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 18,740B, BPFP=1.8891 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 27,268B, BPFP=2.7488 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 17,588B, BPFP=1.7730 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 27,756B, BPFP=2.7980 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 24,280B, BPFP=2.4476 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 26,784B, BPFP=2.7000 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 69,680B, BPFP=1.0035 +⌛️ [2/4] FRONTEND: Frontend time: 0.251s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.356s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 155, 128]) + layer.0.v_cache: torch.Size([1, 4, 155, 128]) + layer.1.k_cache: torch.Size([1, 4, 155, 128]) + layer.1.v_cache: torch.Size([1, 4, 155, 128]) + layer.2.k_cache: torch.Size([1, 4, 155, 128]) + layer.2.v_cache: torch.Size([1, 4, 155, 128]) + layer.3.k_cache: torch.Size([1, 4, 155, 128]) + layer.3.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.k_cache: torch.Size([1, 4, 155, 128]) + layer.4.v_cache: torch.Size([1, 4, 155, 128]) + layer.4.output: torch.Size([1, 155, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13102877 24.55299584 + layer.0.v_cache 0.00001822 0.00666421 + layer.1.k_cache 0.12773433 3.18168965 + layer.1.v_cache 0.00000585 0.00261703 + layer.2.k_cache 0.00931232 0.45213653 + layer.2.v_cache 0.00002110 0.00745822 + layer.3.k_cache 0.01537714 1.95508246 + layer.3.v_cache 0.00002091 0.00881963 + layer.4.k_cache 0.00066471 0.15159181 + layer.4.v_cache 0.00005180 0.01527528 + layer.4.output 0.01742802 342.36491935 + ------------------------------------------------------------------------------------- + TOTAL 0.02389596 142.75816271 + (elements=1,349,120) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1349120 +Total Bytes 292480 +BPFP 1.7343 bits/point +EBPFP 3.4687 equivalent bits/point +MSE 142.758163 +---------------------- -------------------------------------------------------- +Time: 0.613s Load: 0.006s, Pack+Encode: 0.251s, Decode+Unpack: 0.356s +---------------------- -------------------------------------------------------- +💾 Converting with 142.7582 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4081.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4081.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst (59/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 215, 128) +Output shape: (1, 215, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) -> torch.Size([1, 1, 215, 512]) + layer.4.output: torch.Size([1, 215, 3584]) -> torch.Size([1, 1, 215, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,492B, BPFP=0.8352 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,432B, BPFP=2.7930 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,700B, BPFP=1.5044 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,456B, BPFP=2.7221 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,628B, BPFP=1.7898 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,764B, BPFP=2.6718 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,112B, BPFP=1.6797 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,260B, BPFP=2.7078 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,208B, BPFP=2.3407 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,912B, BPFP=2.6099 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,952B, BPFP=0.9339 +⌛️ [2/4] FRONTEND: Frontend time: 0.333s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.402s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 215, 128]) + layer.0.v_cache: torch.Size([1, 4, 215, 128]) + layer.1.k_cache: torch.Size([1, 4, 215, 128]) + layer.1.v_cache: torch.Size([1, 4, 215, 128]) + layer.2.k_cache: torch.Size([1, 4, 215, 128]) + layer.2.v_cache: torch.Size([1, 4, 215, 128]) + layer.3.k_cache: torch.Size([1, 4, 215, 128]) + layer.3.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.k_cache: torch.Size([1, 4, 215, 128]) + layer.4.v_cache: torch.Size([1, 4, 215, 128]) + layer.4.output: torch.Size([1, 215, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11659910 23.79212845 + layer.0.v_cache 0.00001673 0.00682136 + layer.1.k_cache 0.29571861 2.96134885 + layer.1.v_cache 0.00000615 0.00257053 + layer.2.k_cache 0.04707494 0.46835586 + layer.2.v_cache 0.00002002 0.00663529 + layer.3.k_cache 0.03274950 2.09206004 + layer.3.v_cache 0.00002078 0.00766978 + layer.4.k_cache 0.00068608 0.13612029 + layer.4.v_cache 0.00005136 0.01371463 + layer.4.output 1.42391751 247.35263704 + ------------------------------------------------------------------------------------- + TOTAL 0.61531564 103.58564026 + (elements=1,871,360) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1871360 +Total Bytes 387916 +BPFP 1.6583 bits/point +EBPFP 3.3167 equivalent bits/point +MSE 103.585640 +---------------------- -------------------------------------------------------- +Time: 0.744s Load: 0.009s, Pack+Encode: 0.333s, Decode+Unpack: 0.402s +---------------------- -------------------------------------------------------- +💾 Converting with 103.5856 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4094.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4094.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst (60/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,620B, BPFP=0.7951 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,960B, BPFP=2.6436 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,972B, BPFP=1.4473 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,908B, BPFP=2.5774 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,308B, BPFP=1.7205 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,264B, BPFP=2.5368 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,816B, BPFP=1.6265 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,960B, BPFP=2.5806 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,356B, BPFP=2.2276 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,500B, BPFP=2.4887 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 105,584B, BPFP=0.9503 +⌛️ [2/4] FRONTEND: Frontend time: 0.296s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.415s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10631144 22.93950038 + layer.0.v_cache 0.00001780 0.00685573 + layer.1.k_cache 0.46192560 3.03187389 + layer.1.v_cache 0.00000619 0.00266805 + layer.2.k_cache 0.02890189 0.44226468 + layer.2.v_cache 0.00002115 0.00685987 + layer.3.k_cache 0.04994748 1.89125307 + layer.3.v_cache 0.00002140 0.00843732 + layer.4.k_cache 0.00068365 0.14065238 + layer.4.v_cache 0.00005040 0.01402309 + layer.4.output 1.23456514 213.48882128 + ------------------------------------------------------------------------------------- + TOTAL 0.54646135 89.58271397 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 433248 +BPFP 1.6057 bits/point +EBPFP 3.2113 equivalent bits/point +MSE 89.582714 +---------------------- -------------------------------------------------------- +Time: 0.719s Load: 0.009s, Pack+Encode: 0.296s, Decode+Unpack: 0.415s +---------------------- -------------------------------------------------------- +💾 Converting with 89.5827 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4114.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4114.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst (61/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 194, 128) +Output shape: (1, 194, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) -> torch.Size([1, 1, 194, 512]) + layer.4.output: torch.Size([1, 194, 3584]) -> torch.Size([1, 1, 194, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,064B, BPFP=0.8106 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,092B, BPFP=2.9069 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 18,452B, BPFP=1.4861 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,128B, BPFP=2.8293 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 22,836B, BPFP=1.8392 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 34,736B, BPFP=2.7977 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,140B, BPFP=1.7026 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 35,228B, BPFP=2.8373 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,500B, BPFP=2.4565 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,028B, BPFP=2.7407 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 77,764B, BPFP=0.8947 +⌛️ [2/4] FRONTEND: Frontend time: 0.273s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.406s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 194, 128]) + layer.0.v_cache: torch.Size([1, 4, 194, 128]) + layer.1.k_cache: torch.Size([1, 4, 194, 128]) + layer.1.v_cache: torch.Size([1, 4, 194, 128]) + layer.2.k_cache: torch.Size([1, 4, 194, 128]) + layer.2.v_cache: torch.Size([1, 4, 194, 128]) + layer.3.k_cache: torch.Size([1, 4, 194, 128]) + layer.3.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.k_cache: torch.Size([1, 4, 194, 128]) + layer.4.v_cache: torch.Size([1, 4, 194, 128]) + layer.4.output: torch.Size([1, 194, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15027005 23.28367630 + layer.0.v_cache 0.00001782 0.00709160 + layer.1.k_cache 0.21402168 3.13595770 + layer.1.v_cache 0.00000588 0.00274471 + layer.2.k_cache 0.02177384 0.44628946 + layer.2.v_cache 0.00002019 0.00730287 + layer.3.k_cache 0.04024272 2.12255781 + layer.3.v_cache 0.00002076 0.00834610 + layer.4.k_cache 0.00069591 0.14937491 + layer.4.v_cache 0.00005073 0.01436578 + layer.4.output 1.57794378 274.53831462 + ------------------------------------------------------------------------------------- + TOTAL 0.67486624 114.76152409 + (elements=1,688,576) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1688576 +Total Bytes 355968 +BPFP 1.6865 bits/point +EBPFP 3.3730 equivalent bits/point +MSE 114.761524 +---------------------- -------------------------------------------------------- +Time: 0.686s Load: 0.007s, Pack+Encode: 0.273s, Decode+Unpack: 0.406s +---------------------- -------------------------------------------------------- +💾 Converting with 114.7615 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4121.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4121.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst (62/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 258, 128) +Output shape: (1, 258, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) -> torch.Size([1, 1, 258, 512]) + layer.4.output: torch.Size([1, 258, 3584]) -> torch.Size([1, 1, 258, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,220B, BPFP=0.8006 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 46,436B, BPFP=2.8123 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 23,696B, BPFP=1.4351 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 45,072B, BPFP=2.7297 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 28,912B, BPFP=1.7510 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 44,408B, BPFP=2.6894 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 26,516B, BPFP=1.6059 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 45,112B, BPFP=2.7321 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 38,704B, BPFP=2.3440 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 43,700B, BPFP=2.6466 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 97,808B, BPFP=0.8462 +⌛️ [2/4] FRONTEND: Frontend time: 0.359s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.499s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 258, 128]) + layer.0.v_cache: torch.Size([1, 4, 258, 128]) + layer.1.k_cache: torch.Size([1, 4, 258, 128]) + layer.1.v_cache: torch.Size([1, 4, 258, 128]) + layer.2.k_cache: torch.Size([1, 4, 258, 128]) + layer.2.v_cache: torch.Size([1, 4, 258, 128]) + layer.3.k_cache: torch.Size([1, 4, 258, 128]) + layer.3.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.k_cache: torch.Size([1, 4, 258, 128]) + layer.4.v_cache: torch.Size([1, 4, 258, 128]) + layer.4.output: torch.Size([1, 258, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09035230 22.89568761 + layer.0.v_cache 0.00001726 0.00650613 + layer.1.k_cache 0.42455262 3.20305474 + layer.1.v_cache 0.00000600 0.00243553 + layer.2.k_cache 0.01891134 0.41277683 + layer.2.v_cache 0.00001996 0.00654040 + layer.3.k_cache 0.02093011 1.82459371 + layer.3.v_cache 0.00001928 0.00732775 + layer.4.k_cache 0.00072082 0.13624641 + layer.4.v_cache 0.00004951 0.01357516 + layer.4.output 0.00505962 212.02799695 + ------------------------------------------------------------------------------------- + TOTAL 0.03476450 88.98263076 + (elements=2,245,632) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2245632 +Total Bytes 453584 +BPFP 1.6159 bits/point +EBPFP 3.2318 equivalent bits/point +MSE 88.982631 +---------------------- -------------------------------------------------------- +Time: 0.869s Load: 0.010s, Pack+Encode: 0.359s, Decode+Unpack: 0.499s +---------------------- -------------------------------------------------------- +💾 Converting with 88.9826 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4129.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4129.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst (63/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,212B, BPFP=0.8296 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,424B, BPFP=2.7462 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,980B, BPFP=1.4932 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,456B, BPFP=2.6804 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,832B, BPFP=1.7549 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,620B, BPFP=2.6236 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,332B, BPFP=1.6530 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,352B, BPFP=2.6734 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,608B, BPFP=2.2832 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,724B, BPFP=2.5628 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,544B, BPFP=0.9952 +⌛️ [2/4] FRONTEND: Frontend time: 0.320s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.411s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15008528 23.68834282 + layer.0.v_cache 0.00001847 0.00645945 + layer.1.k_cache 0.25612781 3.22896967 + layer.1.v_cache 0.00000678 0.00244464 + layer.2.k_cache 0.03931687 0.45335654 + layer.2.v_cache 0.00002107 0.00646016 + layer.3.k_cache 0.03786138 1.94129612 + layer.3.v_cache 0.00002177 0.00776974 + layer.4.k_cache 0.00069674 0.13516137 + layer.4.v_cache 0.00005115 0.01271608 + layer.4.output 1.33117570 233.68841227 + ------------------------------------------------------------------------------------- + TOTAL 0.57661395 97.95893309 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 416084 +BPFP 1.6627 bits/point +EBPFP 3.3255 equivalent bits/point +MSE 97.958933 +---------------------- -------------------------------------------------------- +Time: 0.739s Load: 0.008s, Pack+Encode: 0.320s, Decode+Unpack: 0.411s +---------------------- -------------------------------------------------------- +💾 Converting with 97.9589 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4223.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4223.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst (64/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,928B, BPFP=0.8618 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,372B, BPFP=2.7233 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,344B, BPFP=1.5056 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,632B, BPFP=2.6590 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,592B, BPFP=1.7875 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,244B, BPFP=2.6253 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,396B, BPFP=1.6837 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,656B, BPFP=2.6611 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,532B, BPFP=2.3031 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,652B, BPFP=2.5740 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 80,812B, BPFP=1.0021 +⌛️ [2/4] FRONTEND: Frontend time: 0.273s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.343s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15161699 22.72326660 + layer.0.v_cache 0.00001786 0.00705934 + layer.1.k_cache 0.08662784 3.17251655 + layer.1.v_cache 0.00000652 0.00268761 + layer.2.k_cache 0.02063684 0.43475071 + layer.2.v_cache 0.00002327 0.00713334 + layer.3.k_cache 0.08087930 1.91607564 + layer.3.v_cache 0.00002116 0.00801794 + layer.4.k_cache 0.00074161 0.14042600 + layer.4.v_cache 0.00006096 0.01461047 + layer.4.output 0.00898438 288.64273313 + ------------------------------------------------------------------------------------- + TOTAL 0.02373665 120.52503977 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 327160 +BPFP 1.6705 bits/point +EBPFP 3.3411 equivalent bits/point +MSE 120.525040 +---------------------- -------------------------------------------------------- +Time: 0.622s Load: 0.006s, Pack+Encode: 0.273s, Decode+Unpack: 0.343s +---------------------- -------------------------------------------------------- +💾 Converting with 120.5250 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4236.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4236.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst (65/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 220, 128) +Output shape: (1, 220, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) -> torch.Size([1, 1, 220, 512]) + layer.4.output: torch.Size([1, 220, 3584]) -> torch.Size([1, 1, 220, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,284B, BPFP=0.8724 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,924B, BPFP=2.7645 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,428B, BPFP=1.5219 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,968B, BPFP=2.6966 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,388B, BPFP=1.8031 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,380B, BPFP=2.6548 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,700B, BPFP=1.6832 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,956B, BPFP=2.6957 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,932B, BPFP=2.3389 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,620B, BPFP=2.6009 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,096B, BPFP=0.9953 +⌛️ [2/4] FRONTEND: Frontend time: 0.302s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.439s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 220, 128]) + layer.0.v_cache: torch.Size([1, 4, 220, 128]) + layer.1.k_cache: torch.Size([1, 4, 220, 128]) + layer.1.v_cache: torch.Size([1, 4, 220, 128]) + layer.2.k_cache: torch.Size([1, 4, 220, 128]) + layer.2.v_cache: torch.Size([1, 4, 220, 128]) + layer.3.k_cache: torch.Size([1, 4, 220, 128]) + layer.3.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.k_cache: torch.Size([1, 4, 220, 128]) + layer.4.v_cache: torch.Size([1, 4, 220, 128]) + layer.4.output: torch.Size([1, 220, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14327769 23.17848455 + layer.0.v_cache 0.00001763 0.00693694 + layer.1.k_cache 0.34925704 3.08818997 + layer.1.v_cache 0.00000642 0.00283489 + layer.2.k_cache 0.02352169 0.48761500 + layer.2.v_cache 0.00002271 0.00735603 + layer.3.k_cache 0.00845094 1.74705450 + layer.3.v_cache 0.00002026 0.00814874 + layer.4.k_cache 0.00069169 0.14175077 + layer.4.v_cache 0.00005086 0.01463027 + layer.4.output 1.39158238 236.50531656 + ------------------------------------------------------------------------------------- + TOTAL 0.60390550 99.07177750 + (elements=1,914,880) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1914880 +Total Bytes 402676 +BPFP 1.6823 bits/point +EBPFP 3.3646 equivalent bits/point +MSE 99.071778 +---------------------- -------------------------------------------------------- +Time: 0.748s Load: 0.007s, Pack+Encode: 0.302s, Decode+Unpack: 0.439s +---------------------- -------------------------------------------------------- +💾 Converting with 99.0718 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4298.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4298.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst (66/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 275, 128) +Output shape: (1, 275, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) -> torch.Size([1, 1, 275, 512]) + layer.4.output: torch.Size([1, 275, 3584]) -> torch.Size([1, 1, 275, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,536B, BPFP=0.8259 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 48,544B, BPFP=2.7582 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,680B, BPFP=1.4591 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 47,332B, BPFP=2.6893 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,036B, BPFP=1.7634 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 46,600B, BPFP=2.6477 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,132B, BPFP=1.6552 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 47,140B, BPFP=2.6784 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,536B, BPFP=2.3032 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 45,456B, BPFP=2.5827 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 110,268B, BPFP=0.8950 +⌛️ [2/4] FRONTEND: Frontend time: 0.344s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.528s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 275, 128]) + layer.0.v_cache: torch.Size([1, 4, 275, 128]) + layer.1.k_cache: torch.Size([1, 4, 275, 128]) + layer.1.v_cache: torch.Size([1, 4, 275, 128]) + layer.2.k_cache: torch.Size([1, 4, 275, 128]) + layer.2.v_cache: torch.Size([1, 4, 275, 128]) + layer.3.k_cache: torch.Size([1, 4, 275, 128]) + layer.3.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.k_cache: torch.Size([1, 4, 275, 128]) + layer.4.v_cache: torch.Size([1, 4, 275, 128]) + layer.4.output: torch.Size([1, 275, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12716211 22.10744673 + layer.0.v_cache 0.00001677 0.00639209 + layer.1.k_cache 0.51313377 3.35766535 + layer.1.v_cache 0.00000586 0.00248621 + layer.2.k_cache 0.02995576 0.44137634 + layer.2.v_cache 0.00002133 0.00695434 + layer.3.k_cache 0.03339291 1.90583163 + layer.3.v_cache 0.00002012 0.00752634 + layer.4.k_cache 0.00068476 0.13313981 + layer.4.v_cache 0.00005875 0.01331017 + layer.4.output 0.00481428 191.75862013 + ------------------------------------------------------------------------------------- + TOTAL 0.04342071 80.60543941 + (elements=2,393,600) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2393600 +Total Bytes 486260 +BPFP 1.6252 bits/point +EBPFP 3.2504 equivalent bits/point +MSE 80.605439 +---------------------- -------------------------------------------------------- +Time: 0.882s Load: 0.010s, Pack+Encode: 0.344s, Decode+Unpack: 0.528s +---------------------- -------------------------------------------------------- +💾 Converting with 80.6054 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4427.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4427.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst (67/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 254, 128) +Output shape: (1, 254, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) -> torch.Size([1, 1, 254, 512]) + layer.4.output: torch.Size([1, 254, 3584]) -> torch.Size([1, 1, 254, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,116B, BPFP=0.8068 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,880B, BPFP=2.5763 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,792B, BPFP=1.4021 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,948B, BPFP=2.5189 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 27,100B, BPFP=1.6671 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,384B, BPFP=2.4843 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,748B, BPFP=1.5839 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,820B, BPFP=2.5111 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,280B, BPFP=2.1703 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,440B, BPFP=2.4262 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 104,812B, BPFP=0.9211 +⌛️ [2/4] FRONTEND: Frontend time: 0.341s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.475s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 254, 128]) + layer.0.v_cache: torch.Size([1, 4, 254, 128]) + layer.1.k_cache: torch.Size([1, 4, 254, 128]) + layer.1.v_cache: torch.Size([1, 4, 254, 128]) + layer.2.k_cache: torch.Size([1, 4, 254, 128]) + layer.2.v_cache: torch.Size([1, 4, 254, 128]) + layer.3.k_cache: torch.Size([1, 4, 254, 128]) + layer.3.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.k_cache: torch.Size([1, 4, 254, 128]) + layer.4.v_cache: torch.Size([1, 4, 254, 128]) + layer.4.output: torch.Size([1, 254, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13396681 22.61200480 + layer.0.v_cache 0.00001641 0.00631922 + layer.1.k_cache 0.47111181 3.18355530 + layer.1.v_cache 0.00000608 0.00250675 + layer.2.k_cache 0.02933135 0.46211282 + layer.2.v_cache 0.00002255 0.00660669 + layer.3.k_cache 0.03744583 1.94724256 + layer.3.v_cache 0.00002260 0.00736485 + layer.4.k_cache 0.00071777 0.13367356 + layer.4.v_cache 0.00004917 0.01314821 + layer.4.output 1.20538323 202.90236572 + ------------------------------------------------------------------------------------- + TOTAL 0.53590429 85.21712322 + (elements=2,210,816) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2210816 +Total Bytes 432320 +BPFP 1.5644 bits/point +EBPFP 3.1288 equivalent bits/point +MSE 85.217123 +---------------------- -------------------------------------------------------- +Time: 0.824s Load: 0.008s, Pack+Encode: 0.341s, Decode+Unpack: 0.475s +---------------------- -------------------------------------------------------- +💾 Converting with 85.2171 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4440.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4440.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst (68/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 288, 128) +Output shape: (1, 288, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) -> torch.Size([1, 1, 288, 512]) + layer.4.output: torch.Size([1, 288, 3584]) -> torch.Size([1, 1, 288, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,520B, BPFP=0.7878 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 49,596B, BPFP=2.6908 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 26,512B, BPFP=1.4384 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 48,364B, BPFP=2.6239 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 31,636B, BPFP=1.7164 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 47,736B, BPFP=2.5898 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 29,936B, BPFP=1.6241 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 48,208B, BPFP=2.6155 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 41,672B, BPFP=2.2609 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 46,476B, BPFP=2.5215 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 113,384B, BPFP=0.8788 +⌛️ [2/4] FRONTEND: Frontend time: 0.377s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.533s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 288, 128]) + layer.0.v_cache: torch.Size([1, 4, 288, 128]) + layer.1.k_cache: torch.Size([1, 4, 288, 128]) + layer.1.v_cache: torch.Size([1, 4, 288, 128]) + layer.2.k_cache: torch.Size([1, 4, 288, 128]) + layer.2.v_cache: torch.Size([1, 4, 288, 128]) + layer.3.k_cache: torch.Size([1, 4, 288, 128]) + layer.3.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.k_cache: torch.Size([1, 4, 288, 128]) + layer.4.v_cache: torch.Size([1, 4, 288, 128]) + layer.4.output: torch.Size([1, 288, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14060297 22.63423326 + layer.0.v_cache 0.00001598 0.00624708 + layer.1.k_cache 0.58610762 2.92709054 + layer.1.v_cache 0.00000669 0.00242476 + layer.2.k_cache 0.02834670 0.44636705 + layer.2.v_cache 0.00002256 0.00655403 + layer.3.k_cache 0.01962393 1.84790887 + layer.3.v_cache 0.00002003 0.00718845 + layer.4.k_cache 0.00067797 0.13322017 + layer.4.v_cache 0.00005035 0.01274279 + layer.4.output 0.00463834 187.90535094 + ------------------------------------------------------------------------------------- + TOTAL 0.04752607 79.02126080 + (elements=2,506,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2506752 +Total Bytes 498040 +BPFP 1.5894 bits/point +EBPFP 3.1789 equivalent bits/point +MSE 79.021261 +---------------------- -------------------------------------------------------- +Time: 0.919s Load: 0.009s, Pack+Encode: 0.377s, Decode+Unpack: 0.533s +---------------------- -------------------------------------------------------- +💾 Converting with 79.0213 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4443.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4443.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst (69/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.010s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 304, 128) +Output shape: (1, 304, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) -> torch.Size([1, 1, 304, 512]) + layer.4.output: torch.Size([1, 304, 3584]) -> torch.Size([1, 1, 304, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 15,944B, BPFP=0.8195 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 51,728B, BPFP=2.6587 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,036B, BPFP=1.4410 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,476B, BPFP=2.5944 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,384B, BPFP=1.7159 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 49,792B, BPFP=2.5592 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,368B, BPFP=1.6123 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 50,364B, BPFP=2.5886 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,304B, BPFP=2.2257 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 48,548B, BPFP=2.4953 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 119,168B, BPFP=0.8750 +⌛️ [2/4] FRONTEND: Frontend time: 0.325s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.508s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 304, 128]) + layer.0.v_cache: torch.Size([1, 4, 304, 128]) + layer.1.k_cache: torch.Size([1, 4, 304, 128]) + layer.1.v_cache: torch.Size([1, 4, 304, 128]) + layer.2.k_cache: torch.Size([1, 4, 304, 128]) + layer.2.v_cache: torch.Size([1, 4, 304, 128]) + layer.3.k_cache: torch.Size([1, 4, 304, 128]) + layer.3.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.k_cache: torch.Size([1, 4, 304, 128]) + layer.4.v_cache: torch.Size([1, 4, 304, 128]) + layer.4.output: torch.Size([1, 304, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12404394 22.44975843 + layer.0.v_cache 0.00001645 0.00611831 + layer.1.k_cache 0.57408182 3.13984781 + layer.1.v_cache 0.00000616 0.00236339 + layer.2.k_cache 0.03522489 0.43219792 + layer.2.v_cache 0.00002326 0.00653850 + layer.3.k_cache 0.02908128 1.86902016 + layer.3.v_cache 0.00002078 0.00735181 + layer.4.k_cache 0.00072493 0.13043848 + layer.4.v_cache 0.00004967 0.01237869 + layer.4.output 0.04390477 174.06492305 + ------------------------------------------------------------------------------------- + TOTAL 0.06297686 73.32414558 + (elements=2,646,016) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2646016 +Total Bytes 522112 +BPFP 1.5786 bits/point +EBPFP 3.1571 equivalent bits/point +MSE 73.324146 +---------------------- -------------------------------------------------------- +Time: 0.843s Load: 0.010s, Pack+Encode: 0.325s, Decode+Unpack: 0.508s +---------------------- -------------------------------------------------------- +💾 Converting with 73.3241 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4447.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4447.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst (70/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 242, 128) +Output shape: (1, 242, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) -> torch.Size([1, 1, 242, 512]) + layer.4.output: torch.Size([1, 242, 3584]) -> torch.Size([1, 1, 242, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,084B, BPFP=0.8448 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,480B, BPFP=2.6782 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,796B, BPFP=1.4718 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,480B, BPFP=2.6136 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,632B, BPFP=1.7195 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,940B, BPFP=2.5788 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,180B, BPFP=1.6258 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,508B, BPFP=2.6154 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,916B, BPFP=2.2544 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,136B, BPFP=2.5269 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 100,124B, BPFP=0.9235 +⌛️ [2/4] FRONTEND: Frontend time: 0.336s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.470s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 242, 128]) + layer.0.v_cache: torch.Size([1, 4, 242, 128]) + layer.1.k_cache: torch.Size([1, 4, 242, 128]) + layer.1.v_cache: torch.Size([1, 4, 242, 128]) + layer.2.k_cache: torch.Size([1, 4, 242, 128]) + layer.2.v_cache: torch.Size([1, 4, 242, 128]) + layer.3.k_cache: torch.Size([1, 4, 242, 128]) + layer.3.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.k_cache: torch.Size([1, 4, 242, 128]) + layer.4.v_cache: torch.Size([1, 4, 242, 128]) + layer.4.output: torch.Size([1, 242, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11466674 22.81946305 + layer.0.v_cache 0.00001794 0.00644692 + layer.1.k_cache 0.39062487 2.98422393 + layer.1.v_cache 0.00000616 0.00243917 + layer.2.k_cache 0.02414889 0.44268449 + layer.2.v_cache 0.00002046 0.00664333 + layer.3.k_cache 0.03495786 1.87173664 + layer.3.v_cache 0.00002053 0.00753310 + layer.4.k_cache 0.00069125 0.13482809 + layer.4.v_cache 0.00005034 0.01361914 + layer.4.output 1.26514320 219.12096001 + ------------------------------------------------------------------------------------- + TOTAL 0.55418868 91.89037282 + (elements=2,106,368) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2106368 +Total Bytes 424276 +BPFP 1.6114 bits/point +EBPFP 3.2228 equivalent bits/point +MSE 91.890373 +---------------------- -------------------------------------------------------- +Time: 0.814s Load: 0.009s, Pack+Encode: 0.336s, Decode+Unpack: 0.470s +---------------------- -------------------------------------------------------- +💾 Converting with 91.8904 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4493.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4493.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst (71/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 189, 128) +Output shape: (1, 189, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) -> torch.Size([1, 1, 189, 512]) + layer.4.output: torch.Size([1, 189, 3584]) -> torch.Size([1, 1, 189, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,592B, BPFP=0.7930 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,636B, BPFP=2.6154 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,332B, BPFP=1.4329 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,920B, BPFP=2.5562 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,856B, BPFP=1.7242 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,612B, BPFP=2.5308 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,772B, BPFP=1.6346 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 31,000B, BPFP=2.5628 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,844B, BPFP=2.2192 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,968B, BPFP=2.4775 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,944B, BPFP=1.0268 +⌛️ [2/4] FRONTEND: Frontend time: 0.325s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.387s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 189, 128]) + layer.0.v_cache: torch.Size([1, 4, 189, 128]) + layer.1.k_cache: torch.Size([1, 4, 189, 128]) + layer.1.v_cache: torch.Size([1, 4, 189, 128]) + layer.2.k_cache: torch.Size([1, 4, 189, 128]) + layer.2.v_cache: torch.Size([1, 4, 189, 128]) + layer.3.k_cache: torch.Size([1, 4, 189, 128]) + layer.3.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.k_cache: torch.Size([1, 4, 189, 128]) + layer.4.v_cache: torch.Size([1, 4, 189, 128]) + layer.4.output: torch.Size([1, 189, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14361043 23.05486576 + layer.0.v_cache 0.00001648 0.00686169 + layer.1.k_cache 0.18932698 3.15333000 + layer.1.v_cache 0.00000663 0.00286046 + layer.2.k_cache 0.00768016 0.45564407 + layer.2.v_cache 0.00002453 0.00730049 + layer.3.k_cache 0.01421528 1.95772573 + layer.3.v_cache 0.00002024 0.00819953 + layer.4.k_cache 0.00072804 0.14794431 + layer.4.v_cache 0.00005086 0.01415471 + layer.4.output 0.00858909 274.65764361 + ------------------------------------------------------------------------------------- + TOTAL 0.02445902 114.78896424 + (elements=1,645,056) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1645056 +Total Bytes 335476 +BPFP 1.6314 bits/point +EBPFP 3.2629 equivalent bits/point +MSE 114.788964 +---------------------- -------------------------------------------------------- +Time: 0.719s Load: 0.007s, Pack+Encode: 0.325s, Decode+Unpack: 0.387s +---------------------- -------------------------------------------------------- +💾 Converting with 114.7890 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4503.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4503.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst (72/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,436B, BPFP=0.8389 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,096B, BPFP=2.7946 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,344B, BPFP=1.4924 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,216B, BPFP=2.7300 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,224B, BPFP=1.7770 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,552B, BPFP=2.6813 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,784B, BPFP=1.6714 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,176B, BPFP=2.7271 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,776B, BPFP=2.3310 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,764B, BPFP=2.6235 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,212B, BPFP=0.9244 +⌛️ [2/4] FRONTEND: Frontend time: 0.345s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.458s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.16894778 22.84701209 + layer.0.v_cache 0.00001779 0.00665557 + layer.1.k_cache 0.37628758 3.28650733 + layer.1.v_cache 0.00000652 0.00259875 + layer.2.k_cache 0.02181128 0.45782442 + layer.2.v_cache 0.00002030 0.00667222 + layer.3.k_cache 0.01815432 1.88296380 + layer.3.v_cache 0.00002063 0.00812209 + layer.4.k_cache 0.00071780 0.13610624 + layer.4.v_cache 0.00005059 0.01334491 + layer.4.output 1.43727796 251.91620557 + ------------------------------------------------------------------------------------- + TOTAL 0.62629296 105.41536743 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 383580 +BPFP 1.6552 bits/point +EBPFP 3.3104 equivalent bits/point +MSE 105.415367 +---------------------- -------------------------------------------------------- +Time: 0.811s Load: 0.007s, Pack+Encode: 0.345s, Decode+Unpack: 0.458s +---------------------- -------------------------------------------------------- +💾 Converting with 105.4154 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4526.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4526.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst (73/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 216, 128) +Output shape: (1, 216, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) -> torch.Size([1, 1, 216, 512]) + layer.4.output: torch.Size([1, 216, 3584]) -> torch.Size([1, 1, 216, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,520B, BPFP=0.8333 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,136B, BPFP=2.7587 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,668B, BPFP=1.4951 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 37,236B, BPFP=2.6936 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,384B, BPFP=1.7639 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,500B, BPFP=2.6403 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,864B, BPFP=1.6539 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 37,152B, BPFP=2.6875 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,916B, BPFP=2.3087 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,580B, BPFP=2.5738 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,688B, BPFP=0.8752 +⌛️ [2/4] FRONTEND: Frontend time: 0.339s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.454s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 216, 128]) + layer.0.v_cache: torch.Size([1, 4, 216, 128]) + layer.1.k_cache: torch.Size([1, 4, 216, 128]) + layer.1.v_cache: torch.Size([1, 4, 216, 128]) + layer.2.k_cache: torch.Size([1, 4, 216, 128]) + layer.2.v_cache: torch.Size([1, 4, 216, 128]) + layer.3.k_cache: torch.Size([1, 4, 216, 128]) + layer.3.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.k_cache: torch.Size([1, 4, 216, 128]) + layer.4.v_cache: torch.Size([1, 4, 216, 128]) + layer.4.output: torch.Size([1, 216, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09877178 23.33855975 + layer.0.v_cache 0.00001652 0.00643339 + layer.1.k_cache 0.28824160 2.89625041 + layer.1.v_cache 0.00000619 0.00256507 + layer.2.k_cache 0.02137197 0.47440169 + layer.2.v_cache 0.00001995 0.00645526 + layer.3.k_cache 0.03811082 1.85317357 + layer.3.v_cache 0.00002058 0.00739037 + layer.4.k_cache 0.00070031 0.13370626 + layer.4.v_cache 0.00004695 0.01213179 + layer.4.output 1.41729720 246.46705522 + ------------------------------------------------------------------------------------- + TOTAL 0.60990512 103.17649730 + (elements=1,880,064) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1880064 +Total Bytes 380644 +BPFP 1.6197 bits/point +EBPFP 3.2394 equivalent bits/point +MSE 103.176497 +---------------------- -------------------------------------------------------- +Time: 0.802s Load: 0.009s, Pack+Encode: 0.339s, Decode+Unpack: 0.454s +---------------------- -------------------------------------------------------- +💾 Converting with 103.1765 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4541.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4541.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst (74/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,500B, BPFP=0.7951 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,592B, BPFP=2.7373 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,152B, BPFP=1.4624 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,680B, BPFP=2.6742 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,180B, BPFP=1.7409 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,816B, BPFP=2.6145 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,856B, BPFP=1.6493 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,256B, BPFP=2.6449 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,080B, BPFP=2.2871 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,008B, BPFP=2.5586 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 92,964B, BPFP=0.9182 +⌛️ [2/4] FRONTEND: Frontend time: 0.344s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.419s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12161526 23.51619322 + layer.0.v_cache 0.00001916 0.00667257 + layer.1.k_cache 0.26028682 3.18057683 + layer.1.v_cache 0.00000628 0.00250298 + layer.2.k_cache 0.01296707 0.41867487 + layer.2.v_cache 0.00002040 0.00656231 + layer.3.k_cache 0.04595671 1.95509217 + layer.3.v_cache 0.00002073 0.00726463 + layer.4.k_cache 0.00067352 0.13152406 + layer.4.v_cache 0.00005047 0.01304339 + layer.4.output 1.35465500 235.82178413 + ------------------------------------------------------------------------------------- + TOTAL 0.58377655 98.82297623 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 399084 +BPFP 1.6230 bits/point +EBPFP 3.2461 equivalent bits/point +MSE 98.822976 +---------------------- -------------------------------------------------------- +Time: 0.772s Load: 0.008s, Pack+Encode: 0.344s, Decode+Unpack: 0.419s +---------------------- -------------------------------------------------------- +💾 Converting with 98.8230 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4564.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4564.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst (75/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,560B, BPFP=0.7992 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,452B, BPFP=2.7276 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,124B, BPFP=1.4605 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,528B, BPFP=2.6637 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,096B, BPFP=1.7351 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,836B, BPFP=2.6159 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,608B, BPFP=1.6322 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,308B, BPFP=2.6485 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,932B, BPFP=2.2768 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,884B, BPFP=2.5501 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 88,660B, BPFP=0.8757 +⌛️ [2/4] FRONTEND: Frontend time: 0.319s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.409s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10229862 23.79447680 + layer.0.v_cache 0.00001727 0.00631598 + layer.1.k_cache 0.34669285 3.13890710 + layer.1.v_cache 0.00000643 0.00241239 + layer.2.k_cache 0.01639456 0.44950576 + layer.2.v_cache 0.00002164 0.00653340 + layer.3.k_cache 0.03343437 2.04218069 + layer.3.v_cache 0.00002043 0.00742587 + layer.4.k_cache 0.00071218 0.13114667 + layer.4.v_cache 0.00005088 0.01261670 + layer.4.output 1.35462105 234.98850348 + ------------------------------------------------------------------------------------- + TOTAL 0.58717627 98.50064975 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 393988 +BPFP 1.6023 bits/point +EBPFP 3.2046 equivalent bits/point +MSE 98.500650 +---------------------- -------------------------------------------------------- +Time: 0.735s Load: 0.007s, Pack+Encode: 0.319s, Decode+Unpack: 0.409s +---------------------- -------------------------------------------------------- +💾 Converting with 98.5006 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4634.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4634.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst (76/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 227, 128) +Output shape: (1, 227, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) -> torch.Size([1, 1, 227, 512]) + layer.4.output: torch.Size([1, 227, 3584]) -> torch.Size([1, 1, 227, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,876B, BPFP=0.8175 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,024B, BPFP=2.7550 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,480B, BPFP=1.4785 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,080B, BPFP=2.6900 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,548B, BPFP=1.7585 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,444B, BPFP=2.6462 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,168B, BPFP=1.6635 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,000B, BPFP=2.6845 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,512B, BPFP=2.3067 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,488B, BPFP=2.5804 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 89,788B, BPFP=0.8829 +⌛️ [2/4] FRONTEND: Frontend time: 0.333s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.447s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 227, 128]) + layer.0.v_cache: torch.Size([1, 4, 227, 128]) + layer.1.k_cache: torch.Size([1, 4, 227, 128]) + layer.1.v_cache: torch.Size([1, 4, 227, 128]) + layer.2.k_cache: torch.Size([1, 4, 227, 128]) + layer.2.v_cache: torch.Size([1, 4, 227, 128]) + layer.3.k_cache: torch.Size([1, 4, 227, 128]) + layer.3.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.k_cache: torch.Size([1, 4, 227, 128]) + layer.4.v_cache: torch.Size([1, 4, 227, 128]) + layer.4.output: torch.Size([1, 227, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11800324 23.67340653 + layer.0.v_cache 0.00001819 0.00632543 + layer.1.k_cache 0.35011090 3.06046218 + layer.1.v_cache 0.00000612 0.00237114 + layer.2.k_cache 0.02017111 0.44357350 + layer.2.v_cache 0.00002006 0.00654902 + layer.3.k_cache 0.03026835 1.96533351 + layer.3.v_cache 0.00001975 0.00732445 + layer.4.k_cache 0.00074701 0.13242953 + layer.4.v_cache 0.00005118 0.01274028 + layer.4.output 1.34863711 226.87061438 + ------------------------------------------------------------------------------------- + TOTAL 0.58587504 95.14145978 + (elements=1,975,808) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1975808 +Total Bytes 400408 +BPFP 1.6212 bits/point +EBPFP 3.2425 equivalent bits/point +MSE 95.141460 +---------------------- -------------------------------------------------------- +Time: 0.788s Load: 0.008s, Pack+Encode: 0.333s, Decode+Unpack: 0.447s +---------------------- -------------------------------------------------------- +💾 Converting with 95.1415 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4638.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4638.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst (77/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 229, 128) +Output shape: (1, 229, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) -> torch.Size([1, 1, 229, 512]) + layer.4.output: torch.Size([1, 229, 3584]) -> torch.Size([1, 1, 229, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,360B, BPFP=0.8433 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,284B, BPFP=2.7486 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,112B, BPFP=1.5087 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,504B, BPFP=2.6954 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,332B, BPFP=1.7967 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,028B, BPFP=2.6629 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,724B, BPFP=1.6870 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,784B, BPFP=2.7145 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,252B, BPFP=2.3371 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,332B, BPFP=2.6154 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 95,876B, BPFP=0.9345 +⌛️ [2/4] FRONTEND: Frontend time: 0.342s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.442s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 229, 128]) + layer.0.v_cache: torch.Size([1, 4, 229, 128]) + layer.1.k_cache: torch.Size([1, 4, 229, 128]) + layer.1.v_cache: torch.Size([1, 4, 229, 128]) + layer.2.k_cache: torch.Size([1, 4, 229, 128]) + layer.2.v_cache: torch.Size([1, 4, 229, 128]) + layer.3.k_cache: torch.Size([1, 4, 229, 128]) + layer.3.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.k_cache: torch.Size([1, 4, 229, 128]) + layer.4.v_cache: torch.Size([1, 4, 229, 128]) + layer.4.output: torch.Size([1, 229, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10957497 24.01676788 + layer.0.v_cache 0.00001697 0.00679685 + layer.1.k_cache 0.31262030 3.04257429 + layer.1.v_cache 0.00000629 0.00268737 + layer.2.k_cache 0.01311818 0.44323244 + layer.2.v_cache 0.00002149 0.00695144 + layer.3.k_cache 0.06227440 1.95187365 + layer.3.v_cache 0.00002139 0.00852635 + layer.4.k_cache 0.00068925 0.14075197 + layer.4.v_cache 0.00005302 0.01457297 + layer.4.output 1.33688878 234.70438241 + ------------------------------------------------------------------------------------- + TOTAL 0.57980104 98.38620071 + (elements=1,993,216) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1993216 +Total Bytes 412588 +BPFP 1.6560 bits/point +EBPFP 3.3119 equivalent bits/point +MSE 98.386201 +---------------------- -------------------------------------------------------- +Time: 0.793s Load: 0.008s, Pack+Encode: 0.342s, Decode+Unpack: 0.442s +---------------------- -------------------------------------------------------- +💾 Converting with 98.3862 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4871.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4871.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst (78/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 213, 128) +Output shape: (1, 213, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) -> torch.Size([1, 1, 213, 512]) + layer.4.output: torch.Size([1, 213, 3584]) -> torch.Size([1, 1, 213, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,512B, BPFP=0.8445 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 38,112B, BPFP=2.7958 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,160B, BPFP=1.4789 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,892B, BPFP=2.7063 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,016B, BPFP=1.7617 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,344B, BPFP=2.6661 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,584B, BPFP=1.6567 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,900B, BPFP=2.7069 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,592B, BPFP=2.3175 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,480B, BPFP=2.6027 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,268B, BPFP=0.9040 +⌛️ [2/4] FRONTEND: Frontend time: 0.305s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.413s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 213, 128]) + layer.0.v_cache: torch.Size([1, 4, 213, 128]) + layer.1.k_cache: torch.Size([1, 4, 213, 128]) + layer.1.v_cache: torch.Size([1, 4, 213, 128]) + layer.2.k_cache: torch.Size([1, 4, 213, 128]) + layer.2.v_cache: torch.Size([1, 4, 213, 128]) + layer.3.k_cache: torch.Size([1, 4, 213, 128]) + layer.3.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.k_cache: torch.Size([1, 4, 213, 128]) + layer.4.v_cache: torch.Size([1, 4, 213, 128]) + layer.4.output: torch.Size([1, 213, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13987277 23.22493719 + layer.0.v_cache 0.00001887 0.00670520 + layer.1.k_cache 0.23498858 3.08470218 + layer.1.v_cache 0.00000603 0.00249753 + layer.2.k_cache 0.02028183 0.43439319 + layer.2.v_cache 0.00001992 0.00682035 + layer.3.k_cache 0.05552547 1.95968957 + layer.3.v_cache 0.00002044 0.00785579 + layer.4.k_cache 0.00067773 0.13837256 + layer.4.v_cache 0.00004777 0.01323548 + layer.4.output 1.43727485 251.78085178 + ------------------------------------------------------------------------------------- + TOTAL 0.61837549 105.37324538 + (elements=1,853,952) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1853952 +Total Bytes 379860 +BPFP 1.6391 bits/point +EBPFP 3.2783 equivalent bits/point +MSE 105.373245 +---------------------- -------------------------------------------------------- +Time: 0.725s Load: 0.007s, Pack+Encode: 0.305s, Decode+Unpack: 0.413s +---------------------- -------------------------------------------------------- +💾 Converting with 105.3732 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4905.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4905.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst (79/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 211, 128) +Output shape: (1, 211, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) -> torch.Size([1, 1, 211, 512]) + layer.4.output: torch.Size([1, 211, 3584]) -> torch.Size([1, 1, 211, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,304B, BPFP=0.8371 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,732B, BPFP=2.7941 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,236B, BPFP=1.4985 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,948B, BPFP=2.7361 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,960B, BPFP=1.7743 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 36,240B, BPFP=2.6836 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,564B, BPFP=1.6709 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,652B, BPFP=2.7142 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,764B, BPFP=2.3522 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,388B, BPFP=2.6206 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 84,884B, BPFP=0.8980 +⌛️ [2/4] FRONTEND: Frontend time: 0.335s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.406s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 211, 128]) + layer.0.v_cache: torch.Size([1, 4, 211, 128]) + layer.1.k_cache: torch.Size([1, 4, 211, 128]) + layer.1.v_cache: torch.Size([1, 4, 211, 128]) + layer.2.k_cache: torch.Size([1, 4, 211, 128]) + layer.2.v_cache: torch.Size([1, 4, 211, 128]) + layer.3.k_cache: torch.Size([1, 4, 211, 128]) + layer.3.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.k_cache: torch.Size([1, 4, 211, 128]) + layer.4.v_cache: torch.Size([1, 4, 211, 128]) + layer.4.output: torch.Size([1, 211, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.17004660 23.03415886 + layer.0.v_cache 0.00001740 0.00641300 + layer.1.k_cache 0.16740718 3.05572712 + layer.1.v_cache 0.00000654 0.00266905 + layer.2.k_cache 0.01191964 0.42261136 + layer.2.v_cache 0.00001991 0.00669843 + layer.3.k_cache 0.00660476 1.80948614 + layer.3.v_cache 0.00002076 0.00775416 + layer.4.k_cache 0.00070551 0.13745954 + layer.4.v_cache 0.00005179 0.01344820 + layer.4.output 1.45088344 245.34510410 + ------------------------------------------------------------------------------------- + TOTAL 0.61841083 102.70071497 + (elements=1,836,544) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1836544 +Total Bytes 377672 +BPFP 1.6451 bits/point +EBPFP 3.2903 equivalent bits/point +MSE 102.700715 +---------------------- -------------------------------------------------------- +Time: 0.749s Load: 0.008s, Pack+Encode: 0.335s, Decode+Unpack: 0.406s +---------------------- -------------------------------------------------------- +💾 Converting with 102.7007 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4945.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4945.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst (80/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 230, 128) +Output shape: (1, 230, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) -> torch.Size([1, 1, 230, 512]) + layer.4.output: torch.Size([1, 230, 3584]) -> torch.Size([1, 1, 230, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,140B, BPFP=0.8247 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,212B, BPFP=2.7318 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 21,984B, BPFP=1.4935 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,272B, BPFP=2.6679 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 25,712B, BPFP=1.7467 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 38,716B, BPFP=2.6302 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,216B, BPFP=1.6451 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,176B, BPFP=2.6614 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 33,804B, BPFP=2.2965 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 37,704B, BPFP=2.5614 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,620B, BPFP=0.8406 +⌛️ [2/4] FRONTEND: Frontend time: 0.297s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.421s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 230, 128]) + layer.0.v_cache: torch.Size([1, 4, 230, 128]) + layer.1.k_cache: torch.Size([1, 4, 230, 128]) + layer.1.v_cache: torch.Size([1, 4, 230, 128]) + layer.2.k_cache: torch.Size([1, 4, 230, 128]) + layer.2.v_cache: torch.Size([1, 4, 230, 128]) + layer.3.k_cache: torch.Size([1, 4, 230, 128]) + layer.3.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.k_cache: torch.Size([1, 4, 230, 128]) + layer.4.v_cache: torch.Size([1, 4, 230, 128]) + layer.4.output: torch.Size([1, 230, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14712874 23.84353983 + layer.0.v_cache 0.00001826 0.00603381 + layer.1.k_cache 0.36788522 3.08234916 + layer.1.v_cache 0.00000598 0.00250053 + layer.2.k_cache 0.01156470 0.41858726 + layer.2.v_cache 0.00002030 0.00671644 + layer.3.k_cache 0.01315050 1.98661791 + layer.3.v_cache 0.00002017 0.00734171 + layer.4.k_cache 0.00068370 0.13596967 + layer.4.v_cache 0.00004817 0.01277409 + layer.4.output 1.33105469 233.63029891 + ------------------------------------------------------------------------------------- + TOTAL 0.57987697 97.93614840 + (elements=2,001,920) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2001920 +Total Bytes 399556 +BPFP 1.5967 bits/point +EBPFP 3.1934 equivalent bits/point +MSE 97.936148 +---------------------- -------------------------------------------------------- +Time: 0.725s Load: 0.008s, Pack+Encode: 0.297s, Decode+Unpack: 0.421s +---------------------- -------------------------------------------------------- +💾 Converting with 97.9361 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-4971.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-4971.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst (81/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 188, 128) +Output shape: (1, 188, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) -> torch.Size([1, 1, 188, 512]) + layer.4.output: torch.Size([1, 188, 3584]) -> torch.Size([1, 1, 188, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,820B, BPFP=0.8162 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,432B, BPFP=2.6124 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,476B, BPFP=1.4525 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,688B, BPFP=2.5505 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,576B, BPFP=1.7101 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,380B, BPFP=2.5249 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,488B, BPFP=1.6197 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,744B, BPFP=2.5552 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,796B, BPFP=2.2271 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,804B, BPFP=2.4771 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 80,204B, BPFP=0.9523 +⌛️ [2/4] FRONTEND: Frontend time: 0.301s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.354s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 188, 128]) + layer.0.v_cache: torch.Size([1, 4, 188, 128]) + layer.1.k_cache: torch.Size([1, 4, 188, 128]) + layer.1.v_cache: torch.Size([1, 4, 188, 128]) + layer.2.k_cache: torch.Size([1, 4, 188, 128]) + layer.2.v_cache: torch.Size([1, 4, 188, 128]) + layer.3.k_cache: torch.Size([1, 4, 188, 128]) + layer.3.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.k_cache: torch.Size([1, 4, 188, 128]) + layer.4.v_cache: torch.Size([1, 4, 188, 128]) + layer.4.output: torch.Size([1, 188, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14005002 23.04686461 + layer.0.v_cache 0.00001663 0.00648073 + layer.1.k_cache 0.14149688 3.15756420 + layer.1.v_cache 0.00000584 0.00248812 + layer.2.k_cache 0.00881058 0.42942826 + layer.2.v_cache 0.00002285 0.00697981 + layer.3.k_cache 0.03192479 1.94879150 + layer.3.v_cache 0.00001958 0.00745074 + layer.4.k_cache 0.00066612 0.13922180 + layer.4.v_cache 0.00005162 0.01403246 + layer.4.output 0.00857985 279.53462196 + ------------------------------------------------------------------------------------- + TOTAL 0.02253670 116.79421506 + (elements=1,636,352) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1636352 +Total Bytes 327408 +BPFP 1.6007 bits/point +EBPFP 3.2013 equivalent bits/point +MSE 116.794215 +---------------------- -------------------------------------------------------- +Time: 0.662s Load: 0.007s, Pack+Encode: 0.301s, Decode+Unpack: 0.354s +---------------------- -------------------------------------------------------- +💾 Converting with 116.7942 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5093.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5093.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst (82/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 199, 128) +Output shape: (1, 199, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) -> torch.Size([1, 1, 199, 512]) + layer.4.output: torch.Size([1, 199, 3584]) -> torch.Size([1, 1, 199, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 10,936B, BPFP=0.8587 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 36,976B, BPFP=2.9033 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,344B, BPFP=1.5188 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 35,908B, BPFP=2.8194 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,160B, BPFP=1.8185 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,448B, BPFP=2.7833 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 21,340B, BPFP=1.6756 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,000B, BPFP=2.8266 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 30,888B, BPFP=2.4253 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 34,804B, BPFP=2.7327 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 82,280B, BPFP=0.9229 +⌛️ [2/4] FRONTEND: Frontend time: 0.357s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.475s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 199, 128]) + layer.0.v_cache: torch.Size([1, 4, 199, 128]) + layer.1.k_cache: torch.Size([1, 4, 199, 128]) + layer.1.v_cache: torch.Size([1, 4, 199, 128]) + layer.2.k_cache: torch.Size([1, 4, 199, 128]) + layer.2.v_cache: torch.Size([1, 4, 199, 128]) + layer.3.k_cache: torch.Size([1, 4, 199, 128]) + layer.3.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.k_cache: torch.Size([1, 4, 199, 128]) + layer.4.v_cache: torch.Size([1, 4, 199, 128]) + layer.4.output: torch.Size([1, 199, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11435634 23.76373322 + layer.0.v_cache 0.00001764 0.00680759 + layer.1.k_cache 0.25025516 3.22955476 + layer.1.v_cache 0.00000690 0.00257430 + layer.2.k_cache 0.01528106 0.44770054 + layer.2.v_cache 0.00002105 0.00695822 + layer.3.k_cache 0.02533426 2.11048099 + layer.3.v_cache 0.00002127 0.00801571 + layer.4.k_cache 0.00070809 0.13917219 + layer.4.v_cache 0.00005072 0.01362796 + layer.4.output 1.53833902 267.30837222 + ------------------------------------------------------------------------------------- + TOTAL 0.65731915 111.81689594 + (elements=1,732,096) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1732096 +Total Bytes 367084 +BPFP 1.6954 bits/point +EBPFP 3.3909 equivalent bits/point +MSE 111.816896 +---------------------- -------------------------------------------------------- +Time: 0.839s Load: 0.007s, Pack+Encode: 0.357s, Decode+Unpack: 0.475s +---------------------- -------------------------------------------------------- +💾 Converting with 111.8169 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5118.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5118.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst (83/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 180, 128) +Output shape: (1, 180, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) -> torch.Size([1, 1, 180, 512]) + layer.4.output: torch.Size([1, 180, 3584]) -> torch.Size([1, 1, 180, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,684B, BPFP=0.8406 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,368B, BPFP=2.7229 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,284B, BPFP=1.5003 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,516B, BPFP=2.6490 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,516B, BPFP=1.7809 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,272B, BPFP=2.6278 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,444B, BPFP=1.6878 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,712B, BPFP=2.6660 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,668B, BPFP=2.3149 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,696B, BPFP=2.5778 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 80,572B, BPFP=0.9992 +⌛️ [2/4] FRONTEND: Frontend time: 0.284s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.345s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 180, 128]) + layer.0.v_cache: torch.Size([1, 4, 180, 128]) + layer.1.k_cache: torch.Size([1, 4, 180, 128]) + layer.1.v_cache: torch.Size([1, 4, 180, 128]) + layer.2.k_cache: torch.Size([1, 4, 180, 128]) + layer.2.v_cache: torch.Size([1, 4, 180, 128]) + layer.3.k_cache: torch.Size([1, 4, 180, 128]) + layer.3.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.k_cache: torch.Size([1, 4, 180, 128]) + layer.4.v_cache: torch.Size([1, 4, 180, 128]) + layer.4.output: torch.Size([1, 180, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13258550 23.15472276 + layer.0.v_cache 0.00001767 0.00693969 + layer.1.k_cache 0.14790291 3.26010946 + layer.1.v_cache 0.00000629 0.00254482 + layer.2.k_cache 0.00999225 0.40475337 + layer.2.v_cache 0.00002110 0.00716634 + layer.3.k_cache 0.02117847 1.95801324 + layer.3.v_cache 0.00002045 0.00851615 + layer.4.k_cache 0.00067024 0.13924690 + layer.4.v_cache 0.00006105 0.01413683 + layer.4.output 0.00897616 288.86123512 + ------------------------------------------------------------------------------------- + TOTAL 0.02207583 120.64616444 + (elements=1,566,720) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1566720 +Total Bytes 326732 +BPFP 1.6684 bits/point +EBPFP 3.3367 equivalent bits/point +MSE 120.646164 +---------------------- -------------------------------------------------------- +Time: 0.636s Load: 0.006s, Pack+Encode: 0.284s, Decode+Unpack: 0.345s +---------------------- -------------------------------------------------------- +💾 Converting with 120.6462 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5144.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5144.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst (84/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 204, 128) +Output shape: (1, 204, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) -> torch.Size([1, 1, 204, 512]) + layer.4.output: torch.Size([1, 204, 3584]) -> torch.Size([1, 1, 204, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,004B, BPFP=0.8428 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 37,304B, BPFP=2.8572 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 19,784B, BPFP=1.5153 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 36,324B, BPFP=2.7822 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 23,732B, BPFP=1.8177 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 35,620B, BPFP=2.7282 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 22,200B, BPFP=1.7004 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 36,268B, BPFP=2.7779 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 31,280B, BPFP=2.3958 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 35,056B, BPFP=2.6850 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 86,596B, BPFP=0.9475 +⌛️ [2/4] FRONTEND: Frontend time: 0.350s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.417s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 204, 128]) + layer.0.v_cache: torch.Size([1, 4, 204, 128]) + layer.1.k_cache: torch.Size([1, 4, 204, 128]) + layer.1.v_cache: torch.Size([1, 4, 204, 128]) + layer.2.k_cache: torch.Size([1, 4, 204, 128]) + layer.2.v_cache: torch.Size([1, 4, 204, 128]) + layer.3.k_cache: torch.Size([1, 4, 204, 128]) + layer.3.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.k_cache: torch.Size([1, 4, 204, 128]) + layer.4.v_cache: torch.Size([1, 4, 204, 128]) + layer.4.output: torch.Size([1, 204, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09364704 24.23380534 + layer.0.v_cache 0.00001853 0.00669575 + layer.1.k_cache 0.25163280 3.15653154 + layer.1.v_cache 0.00000592 0.00252340 + layer.2.k_cache 0.01595125 0.42546717 + layer.2.v_cache 0.00002050 0.00675478 + layer.3.k_cache 0.04325477 1.76212386 + layer.3.v_cache 0.00002016 0.00772286 + layer.4.k_cache 0.00069210 0.13900330 + layer.4.v_cache 0.00005026 0.01354113 + layer.4.output 1.50065788 253.84237132 + ------------------------------------------------------------------------------------- + TOTAL 0.64175873 106.27357461 + (elements=1,775,616) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1775616 +Total Bytes 375168 +BPFP 1.6903 bits/point +EBPFP 3.3806 equivalent bits/point +MSE 106.273575 +---------------------- -------------------------------------------------------- +Time: 0.773s Load: 0.006s, Pack+Encode: 0.350s, Decode+Unpack: 0.417s +---------------------- -------------------------------------------------------- +💾 Converting with 106.2736 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5162.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5162.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst (85/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 226, 128) +Output shape: (1, 226, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) -> torch.Size([1, 1, 226, 512]) + layer.4.output: torch.Size([1, 226, 3584]) -> torch.Size([1, 1, 226, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 11,652B, BPFP=0.8056 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 39,392B, BPFP=2.7235 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 20,928B, BPFP=1.4469 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 38,400B, BPFP=2.6549 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 24,816B, BPFP=1.7157 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 37,744B, BPFP=2.6095 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 23,608B, BPFP=1.6322 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 38,376B, BPFP=2.6532 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 32,836B, BPFP=2.2702 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 36,876B, BPFP=2.5495 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 94,388B, BPFP=0.9322 +⌛️ [2/4] FRONTEND: Frontend time: 0.286s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.414s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 226, 128]) + layer.0.v_cache: torch.Size([1, 4, 226, 128]) + layer.1.k_cache: torch.Size([1, 4, 226, 128]) + layer.1.v_cache: torch.Size([1, 4, 226, 128]) + layer.2.k_cache: torch.Size([1, 4, 226, 128]) + layer.2.v_cache: torch.Size([1, 4, 226, 128]) + layer.3.k_cache: torch.Size([1, 4, 226, 128]) + layer.3.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.k_cache: torch.Size([1, 4, 226, 128]) + layer.4.v_cache: torch.Size([1, 4, 226, 128]) + layer.4.output: torch.Size([1, 226, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09510114 22.75282166 + layer.0.v_cache 0.00001717 0.00632260 + layer.1.k_cache 0.31797088 3.11431480 + layer.1.v_cache 0.00000673 0.00250114 + layer.2.k_cache 0.01711024 0.44445521 + layer.2.v_cache 0.00002022 0.00636828 + layer.3.k_cache 0.01307247 2.01521531 + layer.3.v_cache 0.00002089 0.00733445 + layer.4.k_cache 0.00069413 0.13381231 + layer.4.v_cache 0.00005461 0.01337211 + layer.4.output 1.35466395 235.04673673 + ------------------------------------------------------------------------------------- + TOTAL 0.58392448 98.46021617 + (elements=1,967,104) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1967104 +Total Bytes 399016 +BPFP 1.6228 bits/point +EBPFP 3.2455 equivalent bits/point +MSE 98.460216 +---------------------- -------------------------------------------------------- +Time: 0.708s Load: 0.009s, Pack+Encode: 0.286s, Decode+Unpack: 0.414s +---------------------- -------------------------------------------------------- +💾 Converting with 98.4602 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5173.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5173.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst (86/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 248, 128) +Output shape: (1, 248, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) -> torch.Size([1, 1, 248, 512]) + layer.4.output: torch.Size([1, 248, 3584]) -> torch.Size([1, 1, 248, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 13,068B, BPFP=0.8233 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,680B, BPFP=2.6260 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,684B, BPFP=1.4292 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 40,824B, BPFP=2.5721 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,904B, BPFP=1.6951 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 40,204B, BPFP=2.5330 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 25,444B, BPFP=1.6031 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 40,688B, BPFP=2.5635 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 35,048B, BPFP=2.2082 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 39,296B, BPFP=2.4758 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 102,804B, BPFP=0.9253 +⌛️ [2/4] FRONTEND: Frontend time: 0.327s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.483s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 248, 128]) + layer.0.v_cache: torch.Size([1, 4, 248, 128]) + layer.1.k_cache: torch.Size([1, 4, 248, 128]) + layer.1.v_cache: torch.Size([1, 4, 248, 128]) + layer.2.k_cache: torch.Size([1, 4, 248, 128]) + layer.2.v_cache: torch.Size([1, 4, 248, 128]) + layer.3.k_cache: torch.Size([1, 4, 248, 128]) + layer.3.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.k_cache: torch.Size([1, 4, 248, 128]) + layer.4.v_cache: torch.Size([1, 4, 248, 128]) + layer.4.output: torch.Size([1, 248, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14860459 22.79543772 + layer.0.v_cache 0.00001629 0.00623198 + layer.1.k_cache 0.38266554 3.10525390 + layer.1.v_cache 0.00000669 0.00246668 + layer.2.k_cache 0.01628031 0.43465608 + layer.2.v_cache 0.00002106 0.00650935 + layer.3.k_cache 0.01842327 1.82618689 + layer.3.v_cache 0.00002289 0.00772253 + layer.4.k_cache 0.00072690 0.13348299 + layer.4.v_cache 0.00004923 0.01319906 + layer.4.output 1.23454798 213.75775850 + ------------------------------------------------------------------------------------- + TOTAL 0.54168545 89.68443863 + (elements=2,158,592) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2158592 +Total Bytes 428644 +BPFP 1.5886 bits/point +EBPFP 3.1772 equivalent bits/point +MSE 89.684439 +---------------------- -------------------------------------------------------- +Time: 0.817s Load: 0.008s, Pack+Encode: 0.327s, Decode+Unpack: 0.483s +---------------------- -------------------------------------------------------- +💾 Converting with 89.6844 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5186.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5186.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst (87/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 435, 128) +Output shape: (1, 435, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) -> torch.Size([1, 1, 435, 512]) + layer.4.output: torch.Size([1, 435, 3584]) -> torch.Size([1, 1, 435, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 22,264B, BPFP=0.7997 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 72,840B, BPFP=2.6164 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 38,916B, BPFP=1.3978 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 71,076B, BPFP=2.5530 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 46,300B, BPFP=1.6631 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 69,976B, BPFP=2.5135 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 43,616B, BPFP=1.5667 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 70,596B, BPFP=2.5358 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 60,444B, BPFP=2.1711 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 68,156B, BPFP=2.4481 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 166,212B, BPFP=0.8529 +⌛️ [2/4] FRONTEND: Frontend time: 0.403s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.628s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 435, 128]) + layer.0.v_cache: torch.Size([1, 4, 435, 128]) + layer.1.k_cache: torch.Size([1, 4, 435, 128]) + layer.1.v_cache: torch.Size([1, 4, 435, 128]) + layer.2.k_cache: torch.Size([1, 4, 435, 128]) + layer.2.v_cache: torch.Size([1, 4, 435, 128]) + layer.3.k_cache: torch.Size([1, 4, 435, 128]) + layer.3.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.k_cache: torch.Size([1, 4, 435, 128]) + layer.4.v_cache: torch.Size([1, 4, 435, 128]) + layer.4.output: torch.Size([1, 435, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15864419 20.96359330 + layer.0.v_cache 0.00001607 0.00613244 + layer.1.k_cache 0.94329666 2.98569392 + layer.1.v_cache 0.00000639 0.00242867 + layer.2.k_cache 0.02505282 0.45427523 + layer.2.v_cache 0.00002131 0.00623582 + layer.3.k_cache 0.01724711 1.83622780 + layer.3.v_cache 0.00002032 0.00680229 + layer.4.k_cache 0.00074513 0.12810245 + layer.4.v_cache 0.00005173 0.01247549 + layer.4.output 0.00608059 120.33021346 + ------------------------------------------------------------------------------------- + TOTAL 0.06986270 51.10079186 + (elements=3,786,240) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3786240 +Total Bytes 730396 +BPFP 1.5433 bits/point +EBPFP 3.0865 equivalent bits/point +MSE 51.100792 +---------------------- -------------------------------------------------------- +Time: 1.046s Load: 0.015s, Pack+Encode: 0.403s, Decode+Unpack: 0.628s +---------------------- -------------------------------------------------------- +💾 Converting with 51.1008 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5197.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5197.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst (88/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.015s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 437, 128) +Output shape: (1, 437, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) -> torch.Size([1, 1, 437, 512]) + layer.4.output: torch.Size([1, 437, 3584]) -> torch.Size([1, 1, 437, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 22,576B, BPFP=0.8072 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 72,820B, BPFP=2.6037 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 38,784B, BPFP=1.3867 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 71,092B, BPFP=2.5419 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 46,444B, BPFP=1.6606 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 70,180B, BPFP=2.5093 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 43,624B, BPFP=1.5598 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 71,016B, BPFP=2.5392 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 60,668B, BPFP=2.1692 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 68,468B, BPFP=2.4481 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 178,476B, BPFP=0.9116 +⌛️ [2/4] FRONTEND: Frontend time: 0.421s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.612s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 437, 128]) + layer.0.v_cache: torch.Size([1, 4, 437, 128]) + layer.1.k_cache: torch.Size([1, 4, 437, 128]) + layer.1.v_cache: torch.Size([1, 4, 437, 128]) + layer.2.k_cache: torch.Size([1, 4, 437, 128]) + layer.2.v_cache: torch.Size([1, 4, 437, 128]) + layer.3.k_cache: torch.Size([1, 4, 437, 128]) + layer.3.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.k_cache: torch.Size([1, 4, 437, 128]) + layer.4.v_cache: torch.Size([1, 4, 437, 128]) + layer.4.output: torch.Size([1, 437, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14454122 21.40488237 + layer.0.v_cache 0.00001702 0.00611626 + layer.1.k_cache 0.94267751 3.06821496 + layer.1.v_cache 0.00000668 0.00251453 + layer.2.k_cache 0.01808156 0.41917276 + layer.2.v_cache 0.00002154 0.00648386 + layer.3.k_cache 0.03034366 1.86915976 + layer.3.v_cache 0.00002210 0.00748611 + layer.4.k_cache 0.00074160 0.13471779 + layer.4.v_cache 0.00005552 0.01322486 + layer.4.output 0.00611241 125.58325842 + ------------------------------------------------------------------------------------- + TOTAL 0.06937031 53.29498719 + (elements=3,803,648) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 3803648 +Total Bytes 744148 +BPFP 1.5651 bits/point +EBPFP 3.1302 equivalent bits/point +MSE 53.294987 +---------------------- -------------------------------------------------------- +Time: 1.048s Load: 0.015s, Pack+Encode: 0.421s, Decode+Unpack: 0.612s +---------------------- -------------------------------------------------------- +💾 Converting with 53.2950 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5199.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5199.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst (89/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.012s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 308, 128) +Output shape: (1, 308, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) -> torch.Size([1, 1, 308, 512]) + layer.4.output: torch.Size([1, 308, 3584]) -> torch.Size([1, 1, 308, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 16,172B, BPFP=0.8204 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 52,132B, BPFP=2.6447 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 28,032B, BPFP=1.4221 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 50,756B, BPFP=2.5749 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 33,224B, BPFP=1.6855 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 50,064B, BPFP=2.5398 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 31,412B, BPFP=1.5935 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 50,732B, BPFP=2.5737 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 43,524B, BPFP=2.2080 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 48,892B, BPFP=2.4803 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 128,708B, BPFP=0.9328 +⌛️ [2/4] FRONTEND: Frontend time: 0.352s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.476s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 308, 128]) + layer.0.v_cache: torch.Size([1, 4, 308, 128]) + layer.1.k_cache: torch.Size([1, 4, 308, 128]) + layer.1.v_cache: torch.Size([1, 4, 308, 128]) + layer.2.k_cache: torch.Size([1, 4, 308, 128]) + layer.2.v_cache: torch.Size([1, 4, 308, 128]) + layer.3.k_cache: torch.Size([1, 4, 308, 128]) + layer.3.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.k_cache: torch.Size([1, 4, 308, 128]) + layer.4.v_cache: torch.Size([1, 4, 308, 128]) + layer.4.output: torch.Size([1, 308, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12827574 22.30919776 + layer.0.v_cache 0.00001866 0.00641589 + layer.1.k_cache 0.53653341 3.03643779 + layer.1.v_cache 0.00000626 0.00233857 + layer.2.k_cache 0.01950284 0.43155561 + layer.2.v_cache 0.00002083 0.00625729 + layer.3.k_cache 0.04340698 1.87692617 + layer.3.v_cache 0.00002181 0.00734830 + layer.4.k_cache 0.00072091 0.13141960 + layer.4.v_cache 0.00005393 0.01287599 + layer.4.output 0.04342448 164.13098620 + ------------------------------------------------------------------------------------- + TOTAL 0.06073722 69.21986332 + (elements=2,680,832) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2680832 +Total Bytes 533648 +BPFP 1.5925 bits/point +EBPFP 3.1850 equivalent bits/point +MSE 69.219863 +---------------------- -------------------------------------------------------- +Time: 0.839s Load: 0.012s, Pack+Encode: 0.352s, Decode+Unpack: 0.476s +---------------------- -------------------------------------------------------- +💾 Converting with 69.2199 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5204.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5204.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst (90/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 266, 128) +Output shape: (1, 266, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) -> torch.Size([1, 1, 266, 512]) + layer.4.output: torch.Size([1, 266, 3584]) -> torch.Size([1, 1, 266, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 14,140B, BPFP=0.8306 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 47,708B, BPFP=2.8024 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 25,148B, BPFP=1.4772 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 46,684B, BPFP=2.7422 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 30,104B, BPFP=1.7683 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 45,988B, BPFP=2.7014 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 28,248B, BPFP=1.6593 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 46,496B, BPFP=2.7312 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 40,004B, BPFP=2.3499 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 44,984B, BPFP=2.6424 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 112,924B, BPFP=0.9476 +⌛️ [2/4] FRONTEND: Frontend time: 0.341s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.469s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 266, 128]) + layer.0.v_cache: torch.Size([1, 4, 266, 128]) + layer.1.k_cache: torch.Size([1, 4, 266, 128]) + layer.1.v_cache: torch.Size([1, 4, 266, 128]) + layer.2.k_cache: torch.Size([1, 4, 266, 128]) + layer.2.v_cache: torch.Size([1, 4, 266, 128]) + layer.3.k_cache: torch.Size([1, 4, 266, 128]) + layer.3.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.k_cache: torch.Size([1, 4, 266, 128]) + layer.4.v_cache: torch.Size([1, 4, 266, 128]) + layer.4.output: torch.Size([1, 266, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.10702504 22.43971746 + layer.0.v_cache 0.00001641 0.00661394 + layer.1.k_cache 0.51130361 3.09588784 + layer.1.v_cache 0.00000656 0.00269947 + layer.2.k_cache 0.01810413 0.41907369 + layer.2.v_cache 0.00002122 0.00690567 + layer.3.k_cache 0.02476560 1.93873585 + layer.3.v_cache 0.00002197 0.00767789 + layer.4.k_cache 0.00071356 0.13961095 + layer.4.v_cache 0.00005290 0.01374322 + layer.4.output 0.00501064 204.55686090 + ------------------------------------------------------------------------------------- + TOTAL 0.04100620 85.88051131 + (elements=2,315,264) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2315264 +Total Bytes 482428 +BPFP 1.6669 bits/point +EBPFP 3.3339 equivalent bits/point +MSE 85.880511 +---------------------- -------------------------------------------------------- +Time: 0.819s Load: 0.009s, Pack+Encode: 0.341s, Decode+Unpack: 0.469s +---------------------- -------------------------------------------------------- +💾 Converting with 85.8805 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5213.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5213.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst (91/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.009s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 234, 128) +Output shape: (1, 234, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) -> torch.Size([1, 1, 234, 512]) + layer.4.output: torch.Size([1, 234, 3584]) -> torch.Size([1, 1, 234, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,736B, BPFP=0.8504 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 41,060B, BPFP=2.7417 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,424B, BPFP=1.4973 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,920B, BPFP=2.6656 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,344B, BPFP=1.7591 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,488B, BPFP=2.6368 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,844B, BPFP=1.6589 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,972B, BPFP=2.6691 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,204B, BPFP=2.2839 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,520B, BPFP=2.5721 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 98,832B, BPFP=0.9428 +⌛️ [2/4] FRONTEND: Frontend time: 0.310s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.433s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 234, 128]) + layer.0.v_cache: torch.Size([1, 4, 234, 128]) + layer.1.k_cache: torch.Size([1, 4, 234, 128]) + layer.1.v_cache: torch.Size([1, 4, 234, 128]) + layer.2.k_cache: torch.Size([1, 4, 234, 128]) + layer.2.v_cache: torch.Size([1, 4, 234, 128]) + layer.3.k_cache: torch.Size([1, 4, 234, 128]) + layer.3.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.k_cache: torch.Size([1, 4, 234, 128]) + layer.4.v_cache: torch.Size([1, 4, 234, 128]) + layer.4.output: torch.Size([1, 234, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11892304 23.34445530 + layer.0.v_cache 0.00001803 0.00668830 + layer.1.k_cache 0.31174104 2.91785320 + layer.1.v_cache 0.00000657 0.00262583 + layer.2.k_cache 0.01452858 0.43058907 + layer.2.v_cache 0.00002161 0.00697460 + layer.3.k_cache 0.02263702 1.79192763 + layer.3.v_cache 0.00002141 0.00781657 + layer.4.k_cache 0.00068312 0.13414191 + layer.4.v_cache 0.00005251 0.01356988 + layer.4.output 1.30838121 225.77665217 + ------------------------------------------------------------------------------------- + TOTAL 0.56631185 94.65254161 + (elements=2,036,736) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2036736 +Total Bytes 418344 +BPFP 1.6432 bits/point +EBPFP 3.2864 equivalent bits/point +MSE 94.652542 +---------------------- -------------------------------------------------------- +Time: 0.751s Load: 0.009s, Pack+Encode: 0.310s, Decode+Unpack: 0.433s +---------------------- -------------------------------------------------------- +💾 Converting with 94.6525 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5216.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5216.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst (92/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.008s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 233, 128) +Output shape: (1, 233, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) -> torch.Size([1, 1, 233, 512]) + layer.4.output: torch.Size([1, 233, 3584]) -> torch.Size([1, 1, 233, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 12,576B, BPFP=0.8433 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 40,848B, BPFP=2.7393 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 22,268B, BPFP=1.4933 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 39,892B, BPFP=2.6752 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 26,096B, BPFP=1.7500 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 39,196B, BPFP=2.6285 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 24,588B, BPFP=1.6489 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 39,656B, BPFP=2.6593 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 34,004B, BPFP=2.2803 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 38,268B, BPFP=2.5663 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 97,872B, BPFP=0.9376 +⌛️ [2/4] FRONTEND: Frontend time: 0.321s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.406s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 233, 128]) + layer.0.v_cache: torch.Size([1, 4, 233, 128]) + layer.1.k_cache: torch.Size([1, 4, 233, 128]) + layer.1.v_cache: torch.Size([1, 4, 233, 128]) + layer.2.k_cache: torch.Size([1, 4, 233, 128]) + layer.2.v_cache: torch.Size([1, 4, 233, 128]) + layer.3.k_cache: torch.Size([1, 4, 233, 128]) + layer.3.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.k_cache: torch.Size([1, 4, 233, 128]) + layer.4.v_cache: torch.Size([1, 4, 233, 128]) + layer.4.output: torch.Size([1, 233, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13365312 23.35198163 + layer.0.v_cache 0.00001735 0.00668066 + layer.1.k_cache 0.40872042 3.32159162 + layer.1.v_cache 0.00000792 0.00274576 + layer.2.k_cache 0.01308113 0.42085728 + layer.2.v_cache 0.00002137 0.00691959 + layer.3.k_cache 0.00601081 1.81622524 + layer.3.v_cache 0.00002094 0.00756802 + layer.4.k_cache 0.00068546 0.13679197 + layer.4.v_cache 0.00005338 0.01414428 + layer.4.output 1.31399239 226.63695586 + ------------------------------------------------------------------------------------- + TOTAL 0.57413051 95.03201159 + (elements=2,028,032) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 2028032 +Total Bytes 415264 +BPFP 1.6381 bits/point +EBPFP 3.2762 equivalent bits/point +MSE 95.032012 +---------------------- -------------------------------------------------------- +Time: 0.735s Load: 0.008s, Pack+Encode: 0.321s, Decode+Unpack: 0.406s +---------------------- -------------------------------------------------------- +💾 Converting with 95.0320 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5218.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5218.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst (93/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.005s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 148, 128) +Output shape: (1, 148, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) -> torch.Size([1, 1, 148, 512]) + layer.4.output: torch.Size([1, 148, 3584]) -> torch.Size([1, 1, 148, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,680B, BPFP=0.9164 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 27,580B, BPFP=2.9117 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 15,124B, BPFP=1.5967 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 26,908B, BPFP=2.8408 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 17,576B, BPFP=1.8556 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 26,524B, BPFP=2.8003 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 16,676B, BPFP=1.7606 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 26,796B, BPFP=2.8290 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 23,160B, BPFP=2.4451 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 25,928B, BPFP=2.7373 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 62,740B, BPFP=0.9462 +⌛️ [2/4] FRONTEND: Frontend time: 0.275s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.341s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 148, 128]) + layer.0.v_cache: torch.Size([1, 4, 148, 128]) + layer.1.k_cache: torch.Size([1, 4, 148, 128]) + layer.1.v_cache: torch.Size([1, 4, 148, 128]) + layer.2.k_cache: torch.Size([1, 4, 148, 128]) + layer.2.v_cache: torch.Size([1, 4, 148, 128]) + layer.3.k_cache: torch.Size([1, 4, 148, 128]) + layer.3.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.k_cache: torch.Size([1, 4, 148, 128]) + layer.4.v_cache: torch.Size([1, 4, 148, 128]) + layer.4.output: torch.Size([1, 148, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.08608804 24.50902990 + layer.0.v_cache 0.00001767 0.00695969 + layer.1.k_cache 0.12276106 3.05761389 + layer.1.v_cache 0.00000594 0.00259738 + layer.2.k_cache 0.00886028 0.47708815 + layer.2.v_cache 0.00002088 0.00717280 + layer.3.k_cache 0.01513371 1.93723153 + layer.3.v_cache 0.00002085 0.00841209 + layer.4.k_cache 0.00068430 0.14184096 + layer.4.v_cache 0.00005331 0.01430935 + layer.4.output 0.08957551 352.16825531 + ------------------------------------------------------------------------------------- + TOTAL 0.05062792 146.78470841 + (elements=1,288,192) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1288192 +Total Bytes 277692 +BPFP 1.7245 bits/point +EBPFP 3.4491 equivalent bits/point +MSE 146.784708 +---------------------- -------------------------------------------------------- +Time: 0.621s Load: 0.005s, Pack+Encode: 0.275s, Decode+Unpack: 0.341s +---------------------- -------------------------------------------------------- +💾 Converting with 146.7847 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5357.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5357.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst (94/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 170, 128) +Output shape: (1, 170, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) -> torch.Size([1, 1, 170, 512]) + layer.4.output: torch.Size([1, 170, 3584]) -> torch.Size([1, 1, 170, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,624B, BPFP=0.8846 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,576B, BPFP=2.8103 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,188B, BPFP=1.5798 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,856B, BPFP=2.7441 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,752B, BPFP=1.8154 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,248B, BPFP=2.6882 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,668B, BPFP=1.7158 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,776B, BPFP=2.7368 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,480B, BPFP=2.3419 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,524B, BPFP=2.6217 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 72,984B, BPFP=0.9583 +⌛️ [2/4] FRONTEND: Frontend time: 0.266s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.383s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 170, 128]) + layer.0.v_cache: torch.Size([1, 4, 170, 128]) + layer.1.k_cache: torch.Size([1, 4, 170, 128]) + layer.1.v_cache: torch.Size([1, 4, 170, 128]) + layer.2.k_cache: torch.Size([1, 4, 170, 128]) + layer.2.v_cache: torch.Size([1, 4, 170, 128]) + layer.3.k_cache: torch.Size([1, 4, 170, 128]) + layer.3.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.k_cache: torch.Size([1, 4, 170, 128]) + layer.4.v_cache: torch.Size([1, 4, 170, 128]) + layer.4.output: torch.Size([1, 170, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13519069 24.42837201 + layer.0.v_cache 0.00001779 0.00731240 + layer.1.k_cache 0.14803200 3.26982817 + layer.1.v_cache 0.00000678 0.00288216 + layer.2.k_cache 0.02267574 0.46110194 + layer.2.v_cache 0.00002005 0.00712636 + layer.3.k_cache 0.04618364 1.86131287 + layer.3.v_cache 0.00002282 0.00850895 + layer.4.k_cache 0.00077641 0.14609369 + layer.4.v_cache 0.00005264 0.01454398 + layer.4.output 0.00945594 315.06814601 + ------------------------------------------------------------------------------------- + TOTAL 0.02465707 131.51082968 + (elements=1,479,680) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1479680 +Total Bytes 311676 +BPFP 1.6851 bits/point +EBPFP 3.3702 equivalent bits/point +MSE 131.510830 +---------------------- -------------------------------------------------------- +Time: 0.655s Load: 0.006s, Pack+Encode: 0.266s, Decode+Unpack: 0.383s +---------------------- -------------------------------------------------------- +💾 Converting with 131.5108 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5377.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5377.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst (95/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 166, 128) +Output shape: (1, 166, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) -> torch.Size([1, 1, 166, 512]) + layer.4.output: torch.Size([1, 166, 3584]) -> torch.Size([1, 1, 166, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,304B, BPFP=0.8758 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 30,280B, BPFP=2.8502 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,704B, BPFP=1.5723 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 29,556B, BPFP=2.7820 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,692B, BPFP=1.8535 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,084B, BPFP=2.7376 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,600B, BPFP=1.7508 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,596B, BPFP=2.7858 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,440B, BPFP=2.3946 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 28,360B, BPFP=2.6694 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 74,008B, BPFP=0.9952 +⌛️ [2/4] FRONTEND: Frontend time: 0.274s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 166, 128]) + layer.0.v_cache: torch.Size([1, 4, 166, 128]) + layer.1.k_cache: torch.Size([1, 4, 166, 128]) + layer.1.v_cache: torch.Size([1, 4, 166, 128]) + layer.2.k_cache: torch.Size([1, 4, 166, 128]) + layer.2.v_cache: torch.Size([1, 4, 166, 128]) + layer.3.k_cache: torch.Size([1, 4, 166, 128]) + layer.3.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.k_cache: torch.Size([1, 4, 166, 128]) + layer.4.v_cache: torch.Size([1, 4, 166, 128]) + layer.4.output: torch.Size([1, 166, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.12520085 24.14344732 + layer.0.v_cache 0.00001716 0.00729410 + layer.1.k_cache 0.15152623 3.44936619 + layer.1.v_cache 0.00000633 0.00274266 + layer.2.k_cache 0.01605446 0.47642926 + layer.2.v_cache 0.00002200 0.00775076 + layer.3.k_cache 0.04332644 2.01039841 + layer.3.v_cache 0.00002349 0.00893176 + layer.4.k_cache 0.00069336 0.14891783 + layer.4.v_cache 0.00005614 0.01428744 + layer.4.output 0.00966471 328.08861338 + ------------------------------------------------------------------------------------- + TOTAL 0.02379879 136.87587408 + (elements=1,444,864) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1444864 +Total Bytes 310624 +BPFP 1.7199 bits/point +EBPFP 3.4398 equivalent bits/point +MSE 136.875874 +---------------------- -------------------------------------------------------- +Time: 0.644s Load: 0.006s, Pack+Encode: 0.274s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 136.8759 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5380.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5380.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst (96/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 181, 128) +Output shape: (1, 181, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) -> torch.Size([1, 1, 181, 512]) + layer.4.output: torch.Size([1, 181, 3584]) -> torch.Size([1, 1, 181, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,756B, BPFP=0.8422 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,304B, BPFP=2.7023 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,352B, BPFP=1.4979 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,588B, BPFP=2.6405 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,352B, BPFP=1.7569 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,128B, BPFP=2.6008 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,424B, BPFP=1.6768 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,528B, BPFP=2.6354 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,440B, BPFP=2.2825 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,452B, BPFP=2.5425 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,340B, BPFP=0.9291 +⌛️ [2/4] FRONTEND: Frontend time: 0.272s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.364s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 181, 128]) + layer.0.v_cache: torch.Size([1, 4, 181, 128]) + layer.1.k_cache: torch.Size([1, 4, 181, 128]) + layer.1.v_cache: torch.Size([1, 4, 181, 128]) + layer.2.k_cache: torch.Size([1, 4, 181, 128]) + layer.2.v_cache: torch.Size([1, 4, 181, 128]) + layer.3.k_cache: torch.Size([1, 4, 181, 128]) + layer.3.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.k_cache: torch.Size([1, 4, 181, 128]) + layer.4.v_cache: torch.Size([1, 4, 181, 128]) + layer.4.output: torch.Size([1, 181, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.13566811 23.24245997 + layer.0.v_cache 0.00002016 0.00733285 + layer.1.k_cache 0.14782082 3.29704394 + layer.1.v_cache 0.00000634 0.00300557 + layer.2.k_cache 0.01546242 0.47687868 + layer.2.v_cache 0.00002121 0.00739887 + layer.3.k_cache 0.02895848 1.77171014 + layer.3.v_cache 0.00002197 0.00849615 + layer.4.k_cache 0.00067433 0.14723958 + layer.4.v_cache 0.00005229 0.01405557 + layer.4.output 0.00892717 300.45323599 + ------------------------------------------------------------------------------------- + TOTAL 0.02301155 125.42048666 + (elements=1,575,424) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1575424 +Total Bytes 320664 +BPFP 1.6283 bits/point +EBPFP 3.2567 equivalent bits/point +MSE 125.420487 +---------------------- -------------------------------------------------------- +Time: 0.643s Load: 0.007s, Pack+Encode: 0.272s, Decode+Unpack: 0.364s +---------------------- -------------------------------------------------------- +💾 Converting with 125.4205 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5397.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5397.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst (97/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 186, 128) +Output shape: (1, 186, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) -> torch.Size([1, 1, 186, 512]) + layer.4.output: torch.Size([1, 186, 3584]) -> torch.Size([1, 1, 186, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,456B, BPFP=0.7944 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,492B, BPFP=2.6455 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,320B, BPFP=1.4550 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,692B, BPFP=2.5783 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,484B, BPFP=1.7208 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,180B, BPFP=2.5353 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,364B, BPFP=1.6267 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,720B, BPFP=2.5806 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,592B, BPFP=2.2339 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,656B, BPFP=2.4913 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 78,052B, BPFP=0.9367 +⌛️ [2/4] FRONTEND: Frontend time: 0.276s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.353s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 186, 128]) + layer.0.v_cache: torch.Size([1, 4, 186, 128]) + layer.1.k_cache: torch.Size([1, 4, 186, 128]) + layer.1.v_cache: torch.Size([1, 4, 186, 128]) + layer.2.k_cache: torch.Size([1, 4, 186, 128]) + layer.2.v_cache: torch.Size([1, 4, 186, 128]) + layer.3.k_cache: torch.Size([1, 4, 186, 128]) + layer.3.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.k_cache: torch.Size([1, 4, 186, 128]) + layer.4.v_cache: torch.Size([1, 4, 186, 128]) + layer.4.output: torch.Size([1, 186, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.09795722 23.35105584 + layer.0.v_cache 0.00001815 0.00712584 + layer.1.k_cache 0.13430934 3.09356230 + layer.1.v_cache 0.00000632 0.00270433 + layer.2.k_cache 0.01830538 0.44948389 + layer.2.v_cache 0.00002120 0.00711159 + layer.3.k_cache 0.04892764 1.96489117 + layer.3.v_cache 0.00002113 0.00828606 + layer.4.k_cache 0.00067235 0.14437650 + layer.4.v_cache 0.00008605 0.01388784 + layer.4.output 0.00869856 288.12159178 + ------------------------------------------------------------------------------------- + TOTAL 0.02124792 120.34668399 + (elements=1,618,944) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1618944 +Total Bytes 324008 +BPFP 1.6011 bits/point +EBPFP 3.2022 equivalent bits/point +MSE 120.346684 +---------------------- -------------------------------------------------------- +Time: 0.635s Load: 0.006s, Pack+Encode: 0.276s, Decode+Unpack: 0.353s +---------------------- -------------------------------------------------------- +💾 Converting with 120.3467 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5407.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5407.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst (98/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.007s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 163, 128) +Output shape: (1, 163, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) -> torch.Size([1, 1, 163, 512]) + layer.4.output: torch.Size([1, 163, 3584]) -> torch.Size([1, 1, 163, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 8,988B, BPFP=0.8616 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 29,740B, BPFP=2.8508 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 16,336B, BPFP=1.5660 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 28,968B, BPFP=2.7768 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 19,240B, BPFP=1.8443 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 28,456B, BPFP=2.7278 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 18,208B, BPFP=1.7454 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 29,032B, BPFP=2.7830 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 25,068B, BPFP=2.4030 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 27,788B, BPFP=2.6637 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 71,120B, BPFP=0.9739 +⌛️ [2/4] FRONTEND: Frontend time: 0.277s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.377s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 163, 128]) + layer.0.v_cache: torch.Size([1, 4, 163, 128]) + layer.1.k_cache: torch.Size([1, 4, 163, 128]) + layer.1.v_cache: torch.Size([1, 4, 163, 128]) + layer.2.k_cache: torch.Size([1, 4, 163, 128]) + layer.2.v_cache: torch.Size([1, 4, 163, 128]) + layer.3.k_cache: torch.Size([1, 4, 163, 128]) + layer.3.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.k_cache: torch.Size([1, 4, 163, 128]) + layer.4.v_cache: torch.Size([1, 4, 163, 128]) + layer.4.output: torch.Size([1, 163, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.11530171 23.46855678 + layer.0.v_cache 0.00001813 0.00700249 + layer.1.k_cache 0.14319220 3.30663463 + layer.1.v_cache 0.00000612 0.00267305 + layer.2.k_cache 0.02368733 0.44355736 + layer.2.v_cache 0.00002106 0.00693790 + layer.3.k_cache 0.02525310 2.05503387 + layer.3.v_cache 0.00002161 0.00827755 + layer.4.k_cache 0.00069984 0.13745056 + layer.4.v_cache 0.00005709 0.01384264 + layer.4.output 0.00982084 320.62491784 + ------------------------------------------------------------------------------------- + TOTAL 0.02217671 133.75437598 + (elements=1,418,752) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1418752 +Total Bytes 302944 +BPFP 1.7082 bits/point +EBPFP 3.4165 equivalent bits/point +MSE 133.754376 +---------------------- -------------------------------------------------------- +Time: 0.661s Load: 0.007s, Pack+Encode: 0.277s, Decode+Unpack: 0.377s +---------------------- -------------------------------------------------------- +💾 Converting with 133.7544 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5423.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5423.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst (99/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 191, 128) +Output shape: (1, 191, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) -> torch.Size([1, 1, 191, 512]) + layer.4.output: torch.Size([1, 191, 3584]) -> torch.Size([1, 1, 191, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,824B, BPFP=0.8037 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,552B, BPFP=2.5812 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,372B, BPFP=1.4211 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,780B, BPFP=2.5180 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,632B, BPFP=1.6878 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 30,416B, BPFP=2.4882 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,544B, BPFP=1.5988 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,832B, BPFP=2.5223 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,932B, BPFP=2.2032 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,900B, BPFP=2.4460 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 83,264B, BPFP=0.9731 +⌛️ [2/4] FRONTEND: Frontend time: 0.273s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.365s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 191, 128]) + layer.0.v_cache: torch.Size([1, 4, 191, 128]) + layer.1.k_cache: torch.Size([1, 4, 191, 128]) + layer.1.v_cache: torch.Size([1, 4, 191, 128]) + layer.2.k_cache: torch.Size([1, 4, 191, 128]) + layer.2.v_cache: torch.Size([1, 4, 191, 128]) + layer.3.k_cache: torch.Size([1, 4, 191, 128]) + layer.3.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.k_cache: torch.Size([1, 4, 191, 128]) + layer.4.v_cache: torch.Size([1, 4, 191, 128]) + layer.4.output: torch.Size([1, 191, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.15112510 23.34985735 + layer.0.v_cache 0.00002069 0.00741421 + layer.1.k_cache 0.18491925 3.24657053 + layer.1.v_cache 0.00000629 0.00285426 + layer.2.k_cache 0.02122488 0.43083087 + layer.2.v_cache 0.00002327 0.00769638 + layer.3.k_cache 0.01259250 1.84505539 + layer.3.v_cache 0.00002182 0.00853139 + layer.4.k_cache 0.00068006 0.14604766 + layer.4.v_cache 0.00005885 0.01486951 + layer.4.output 0.00845948 237.29557779 + ------------------------------------------------------------------------------------- + TOTAL 0.02528759 99.41933953 + (elements=1,662,464) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1662464 +Total Bytes 331048 +BPFP 1.5930 bits/point +EBPFP 3.1861 equivalent bits/point +MSE 99.419340 +---------------------- -------------------------------------------------------- +Time: 0.645s Load: 0.006s, Pack+Encode: 0.273s, Decode+Unpack: 0.365s +---------------------- -------------------------------------------------------- +💾 Converting with 99.4193 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5430.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5430.zst + + 💪 Processing: ../datasets/KimiAudio-7B-Instruct-500features-L5wCache/KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst (100/100) + +[1/4] FRONTEND: Loading features from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst... + +Original data structure: +⌛️ [1/4] FRONTEND: Load time: 0.006s + +------------------------------------------------------------ +KimiAudio Features Summary +------------------------------------------------------------ +Number of layers: 5 +Layer indices: [0, 1, 2, 3, 4] +Last layer index: 4 +Cache shape: (1, 4, 174, 128) +Output shape: (1, 174, 3584) +Data type: torch.bfloat16 +Has output output: True +------------------------------------------------------------ + +[2/4] FRONTEND: Pack + Encode (strategy: individual)... + IndividualPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) -> torch.Size([1, 1, 174, 512]) + layer.4.output: torch.Size([1, 174, 3584]) -> torch.Size([1, 1, 174, 3584]) + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + layer.0.k_cache: 9,588B, BPFP=0.8610 + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + layer.0.v_cache: 31,008B, BPFP=2.7845 + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + layer.1.k_cache: 17,092B, BPFP=1.5348 + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + layer.1.v_cache: 30,248B, BPFP=2.7162 + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + layer.2.k_cache: 20,188B, BPFP=1.8129 + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + layer.2.v_cache: 29,920B, BPFP=2.6868 + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + layer.3.k_cache: 19,172B, BPFP=1.7216 + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + layer.3.v_cache: 30,356B, BPFP=2.7259 + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + layer.4.k_cache: 26,372B, BPFP=2.3682 + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + layer.4.v_cache: 29,264B, BPFP=2.6279 + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + layer.4.output: 75,532B, BPFP=0.9690 +⌛️ [2/4] FRONTEND: Frontend time: 0.269s (Pack+Encode) + +[3/4] BACKEND: Decode + Unpack... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) +⌛️ [3/4] BACKEND: Backend time: 0.391s + +[4/4] METRICS: Computing MSE Breakdown... + Using per-key quantization points (layer.0.k_cache: torch.Size([256])) for layer.0.k_cache + Using per-key quantization points (layer.0.v_cache: torch.Size([256])) for layer.0.v_cache + Using per-key quantization points (layer.1.k_cache: torch.Size([256])) for layer.1.k_cache + Using per-key quantization points (layer.1.v_cache: torch.Size([256])) for layer.1.v_cache + Using per-key quantization points (layer.2.k_cache: torch.Size([256])) for layer.2.k_cache + Using per-key quantization points (layer.2.v_cache: torch.Size([256])) for layer.2.v_cache + Using per-key quantization points (layer.3.k_cache: torch.Size([256])) for layer.3.k_cache + Using per-key quantization points (layer.3.v_cache: torch.Size([256])) for layer.3.v_cache + Using per-key quantization points (layer.4.k_cache: torch.Size([256])) for layer.4.k_cache + Using per-key quantization points (layer.4.v_cache: torch.Size([256])) for layer.4.v_cache + Using per-key quantization points (output: torch.Size([256])) for layer.4.output + IndividualUnPacker: + layer.0.k_cache: torch.Size([1, 4, 174, 128]) + layer.0.v_cache: torch.Size([1, 4, 174, 128]) + layer.1.k_cache: torch.Size([1, 4, 174, 128]) + layer.1.v_cache: torch.Size([1, 4, 174, 128]) + layer.2.k_cache: torch.Size([1, 4, 174, 128]) + layer.2.v_cache: torch.Size([1, 4, 174, 128]) + layer.3.k_cache: torch.Size([1, 4, 174, 128]) + layer.3.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.k_cache: torch.Size([1, 4, 174, 128]) + layer.4.v_cache: torch.Size([1, 4, 174, 128]) + layer.4.output: torch.Size([1, 174, 3584]) + Per-key MSE Breakdown (Quant=Truncation+Quantization, Codec=Compression): + Key Quant-MSE Total-MSE + ------------------------------------------------------------------------------------- + layer.0.k_cache 0.14146894 24.13108948 + layer.0.v_cache 0.00001664 0.00681971 + layer.1.k_cache 0.16292955 3.26678993 + layer.1.v_cache 0.00000602 0.00260717 + layer.2.k_cache 0.01834387 0.46872808 + layer.2.v_cache 0.00002001 0.00699990 + layer.3.k_cache 0.03015787 1.84541566 + layer.3.v_cache 0.00002180 0.00835494 + layer.4.k_cache 0.00067479 0.14888762 + layer.4.v_cache 0.00005130 0.01454857 + layer.4.output 0.00923280 312.92467159 + ------------------------------------------------------------------------------------- + TOTAL 0.02460708 130.61017307 + (elements=1,514,496) +---------------------- -------------------------------------------------------- +SAMPLE-WISE STATISTICS +---------------------- -------------------------------------------------------- +Handler kimiaudio +Strategy individual +Architecture hyperprior-featurecoding +---------------------- -------------------------------------------------------- +Total Elements 1514496 +Total Bytes 318740 +BPFP 1.6837 bits/point +EBPFP 3.3674 equivalent bits/point +MSE 130.610173 +---------------------- -------------------------------------------------------- +Time: 0.666s Load: 0.006s, Pack+Encode: 0.269s, Decode+Unpack: 0.391s +---------------------- -------------------------------------------------------- +💾 Converting with 130.6102 MSE: + from ../datasets/KimiAudio-7B-Instruct-500features-L5wCache//KimiAudio-7B-Instruct-500features-L5wCache/librispeech-test-other/LibriSpeech-test_other-5437.zst + to output-fixed/kimiaudio/lambda0.02/hyperprior-featurecoding-8bit-individual/librispeech-test-other/LibriSpeech-test_other-5437.zst +------------------------ ---------------------------- +TOTAL PROCESSING SUMMARY +------------------------ ---------------------------- +Total files 100 +Avg BPFP 1.6362 bits/point +Avg EBPFP 3.2724 equivalent bits/point +Avg MSE 99.404400 +Avg Time 0.784s +------------------------ ----------------------------